Causation and Counterfactuals PDF

CAUSATION AND COUNTERFACTUALS
E D I T E D B Y JOHN COLLI NS, NED HALL, AND L. A. PAUL

One philosophical approach to causation sees counterfactual dependence as the key to the
explanation of causal facts: for example, events c (the cause) and e (the effect) both occur,
but had c not occurred, e would not have occurred either. The counterfactual analysis of
causation became a focus of philosophical debate after the 1973 publication of the late David
Lewiss groundbreaking paper, Causation, which argues against the previously accepted
regularity analysis and in favor of what he called the promising alternative of the counter-
factual analysis. Thirty years after Lewiss paper, this book brings together some of the most
important recent work connectingor, in some cases, disputing the connection between
counterfactuals and causation, including the complete version of Lewiss Whitehead lectures,
Causation as Influence, a major reworking of his original paper. Also included is a more
recent essay by Lewis, Void and Object, on causation by omission. Several of the essays first
appeared in a special issue of the Journal of Philosophy, but most, including the unabridged
version of Causation as Influence, are published for the first time or in updated forms.
Other topics considered include the trumping of one event over another in determining
causation; de facto dependence; challenges to the transitivity of causation; the possibility that
entities other than events are the fundamental causal relata; the distinction between depend-
ence and production in accounts of causation; the distinction between causation and causal
explanation; the context-dependence of causation; probabilistic analyses of causation; and a
singularist theory of causation.
John Collins is Associate Professor of Philosophy at Columbia University. Ned Hall is Associate
Professor of Philosophy at MIT. L. A. Paul is Assistant Professor of Philosophy at the University
of Arizona and Research Fellow at the Australian National University.
REPRESENTATION AND MIND SERIES
A BRADFORD BOOK
C O N T R I B U T O R S
D. M. Armstrong, Helen Beebee, David Coady, John Collins, Ned Hall, Christopher Hitchcock,
Igal Kvart, David Lewis, Cei Maslen, Tim Maudlin, D. H. Mellor, Peter Menzies, L. A. Paul, Murali
Ramachandran, Jonathan Schaffer, Stephen Yablo.
0-262-53256-5
,!7IA2G2-fdcfgd!:t;K;k;K;k
E D I T E D B Y JOHN COLLI NS, NED HALL, AND L. A. PAUL
CAUSATION AND
COUNTERFACTUALS
C
O
L
L
I
N
S
,

H
A
L
L
,

A
N
D

P
A
U
L
,

E
D
I
T
O
R
S
C
A
U
S
A
T
I
O
N

A
N
D

C
O
U
N
T
E
R
F
A
C
T
U
A
L
S
THE MIT PRESS
Massachusetts Institute of Technology
Cambridge, Massachusetts 02142
http://mitpress.mit.edu
Cover image: Bullet Through the Apple, 1964, Harold Edgerton.
Harold & Esther Edgerton Foundation, courtesy of Palm Press, Inc.
1999Collins 6/16/04 2:50 PM Page 1
Causation and Counterfactuals
Representation and Mind
Hilary Putnam and Ned Block, editors
Representation and Reality Hilary Putnam
Explaining Behavior: Reasons in a World of Causes Fred Dretske
The Metaphysics of Meaning Jerrold J. Katz
A Theory of Content and Other Essays Jerry A. Fodor
The Realistic Spirit: Wittgenstein, Philosophy, and the Mind Cora Diamond
The Imagery Debate Michael Tye
A Study of Concepts Christopher Peacocke
The Rediscovery of the Mind John R. Searle
Past, Space, and Self John Campbell
Mental Reality Galen Strawson
Ten Problems of Consciousness: A Representational Theory of the Phenomenal
Mind Michael Tye
Representations, Targets, and Attitudes Robert Cummins
A Logical Journey: From Godel to Philosophy Hao Wang
Brainchildren: Essays on Designing Minds Daniel C. Dennett
Realistic Rationalism Jerrold J. Katz
The Paradox of Self-Consciousness Jose Luis Bermu dez
In Critical Condition: Polemical Essays on Cognitive Science and the Philosophy of
Mind Jerry Fodor
Mind in a Physical World: An Essay on the Mind-Body Problem and Mental
Causation Jaegwon Kim
The Mind Doesnt Work That Way Jerry Fodor
New Essays on Semantic Externalism and Self-Knowledge Susana Nuccetelli
Consciousness and Persons: Unity and Identity Michael Tye
Naturalistic Realism and the Antirealist Challenge Drew Khlentzos
Causation and Counterfactuals John Collins, Ned Hall, and L. A. Paul (eds.)
Causation and Counterfactuals
edited by John Collins, Ned Hall, and L. A. Paul
A Bradford Book
The MIT Press
Cambridge, Massachusetts
London, England
6 2004 Massachusetts Institute of Technology
All rights reserved. No part of this book may be reproduced in any form by any electronic or mechanical
means (including photocopying, recording, or information storage and retrieval) without permission in
writing from the publisher.
This book was set in Times New Roman on 3B2 by Asco Typesetters, Hong Kong, and was printed and
bound in the United States of America.
Library of Congress Cataloging-in-Publication Data
Causation and counterfactuals / edited by John Collins, Ned Hall, and L. A. Paul.
p. cm. (Representation and mind)
A Bradford book.
Includes bibliographical references and index.
ISBN 0-262-03317-8 (hc. : alk. paper) ISBN 0-262-53256-5 (pb. : alk. paper)
1. Causation. 2. Counterfactuals (Logic). I. Collins, John David. II. Hall, Edward J.
(Edward Jonathan), 1966 . III. Paul, L. A. (Laurie Ann), 1966 . IV. Series.
BD541.C192 2004
122dc22 2003060612
10 9 8 7 6 5 4 3 2 1
To the memory of David Lewis
Contents
Contributors ix
1 Counterfactuals and Causation: History, Problems, and Prospects 1
John Collins, Ned Hall, and L. A. Paul
2 Trumping Preemption 59
Jonathan Schaer
3 Causation as Inuence 75
David Lewis
4 Preemptive Prevention 107
John Collins
5 Advertisement for a Sketch of an Outline of a Prototheory of
Causation 119
Stephen Yablo
6 Dierence-making in Context 139
Peter Menzies
7 Causation and the Price of Transitivity 181
Ned Hall
8 Aspect Causation 205
L. A. Paul
9 Two Concepts of Causation 225
Ned Hall
10 Void and Object 277
David Lewis
11 Causing and Nothingness 291
Helen Beebee
12 For Facts as Causes and Eects 309
D. H. Mellor
13 Preempting Preemption 325
David Coady
14 Causes, Contrasts, and the Nontransitivity of Causation 341
Cei Maslen
15 Causation: Probabilistic and Counterfactual Analyses 359
Igal Kvart
16 A Counterfactual Analysis of Indeterministic Causation 387
Murali Ramachandran
17 Do All and Only Causes Raise the Probabilities of Eects? 403
Christopher Hitchcock
18 Causation, Counterfactuals, and the Third Factor 419
Tim Maudlin
19 Going through the Open Door Again: Counterfactual versus Singularist
Theories of Causation 445
D. M. Armstrong
References 459
General Index 469
Index of Examples 479
viii Contents
Contributors
D. M. Armstrong is Emeritus Professor of Philosophy at the University of Sydney.
Helen Beebee is Senior Lecturer in Philosophy at the University of Manchester.
David Coady is Lecturer in Philosophy at the University of Tasmania.
John Collins is Associate Professor of Philosophy at Columbia University.
Ned Hall is Associate Professor of Philosophy at the Massachusetts Institute of
Technology.
Christopher Hitchcock is Professor of Philosophy at the California Institute of
Technology.
Igal Kvart is Professor of Philosophy at the Hebrew University of Jerusalem.
David Lewis was the Class of 1943 University Professor of Philosophy at Princeton
University.
Cei Maslen is Assistant Professor of Philosophy at Florida State University.
Tim Maudlin is Professor of Philosophy at Rutgers University.
D. H. Mellor is Emeritus Professor of Philosophy at Cambridge University.
Peter Menzies is Associate Professor of Philosophy at Macquarie University.
L. A. Paul is Assistant Professor of Philosophy at the University of Arizona and a
Research Fellow in Philosophy at the Australian National Universitys Research
School of Social Sciences.
Murali Ramachandran is Senior Lecturer in Philosophy at the University of Sussex.
Jonathan Schaer is Assistant Professor of Philosophy at the University of
Massachusetts, Amherst.
Stephen Yablo is Professor of Philosophy at the Massachusetts Institute of
Technology.
1Counterfactuals and Causation: History, Problems, and Prospects
John Collins, Ned Hall, and L. A. Paul
Among the many philosophers who hold that causal facts1 are to be explained in
terms ofor more ambitiously, shown to reduce tofacts about what happens,
together with facts about the fundamental laws that govern what happens, the clear
favorite is an approach that sees counterfactual dependence as the key to such
explanation or reduction. The paradigm examples of causation, so advocates of this
approach tell us, are examples in which events c and ethe cause and its eect
both occur, but: Had c not occurred, e would not have occurred either. From this
starting point ideas proliferate in a vast profusion. But the remarkable disparity
among these ideas should not obscure their common foundation. Neither should the
diversity of opinion about the prospects for a philosophical analysis of causation
obscure their importance: Even those philosophers who see these prospects as dim
perhaps because they suer post-Quinean queasiness at the thought of any analysis
of any concept of interestcan often be heard to say such things as that causal
relations between events are somehow a matter of the patterns of counterfactual
dependence to be found in them.
It was not always so. Thirty-odd years ago, so-called regularity analyses (so-called,
presumably, because they traced back to Humes well-known analysis of causation
as constant conjunction) ruled the day, with Mackies Cement of the Universe
embodying a classic statement. But this sort of view fell on hard times, both because
of internal problemswhich we will review in due courseand because dramatic
improvements in philosophical understanding of counterfactuals made possible the
emergence of a serious and potent rival: a counterfactual analysis of causation rest-
ing on foundations rm enough to repel the kind of philosophical suspicion that had
formerly warranted dismissal. One speculates at peril about the specic intellectual
and social forces driving trends in philosophy. Still, it is a safe bet that Lewiss
groundbreaking paper Causation (1973a) helped to turn the tide. Elegant, con-
centrated, and compelling, this paper contained a forceful condemnationmuch
quoted, by this pointof the incumbent view:
It remains to be seen whether any regularity analysis can succeed in distinguishing genuine
causes from eects, epiphenomena, and preempted potential causesand whether it can suc-
ceed without falling victim to worse problems, without piling on the epicycles, and without
departing from the fundamental idea that causation is instantiation of regularities. I have no
proof that regularity analyses are beyond repair, nor any space to review the repairs that have
been tried. Suce it to say that the prospects look dark. I think it is time to give up and try
something else.
A promising alternative is not far to seek. (1973a, p. 160)
At the time, the sentiment Lewis expressed in this passage was entirely tting.
Whetherand to what extentthat sentiment continues to be tting is a judg-
ment that this volume aims to inform. Here we bring together some of the most
important recent work connectingor in some cases, disputing the connection
betweencounterfactuals and causation. Some of these essays (chapters 24, 7, and
8) appeared previously in a special issue (April 2000) of the Journal of Philosophy
devoted to this topic; the rest are published here for the rst time, or in new and
updated forms. (Lewiss groundbreaking Causation as Inuencechapter 3has
been greatly expanded for this volume.) Our aim in this introductory chapter is to set
the stage for the essays that follow by providing a map of the main features of the
philosophical terrain that our authors are navigating. Accordingly, we will discuss
the work on the semantics of counterfactuals that set the stage for counterfactual
analyses of causation (section 1), review the early successes and more recent chal-
lenges to such analyses (sections 2 through 4), probe some important points of
methodology that play a signicant but often tacit role in philosophical investiga-
tions of causation (section 5), and review (in sections 6 and 7) two topics of general
interest that work in the counterfactual tradition has shed a great deal of light on,
and that many of our contributors are concerned to address: the transitivity of cau-
sation, and the nature of the causal relata.
For those who come to our topics for the rst time, our overview should provide
you with the background you need to appreciate the often subtle and intricate dia-
lectical interplay featured in the remaining contributions; for seasoned veterans, we
hope nevertheless to reveal unnoticed connections within and novel perspectives on
what may have seemed boringly familiar material. (Those of you who are impatient
can skip ahead to section 8, where we give capsule summaries of the contributions in
this volume.)
Let us begin by rehearsing the rst half of our topic: counterfactuals.
1 Counterfactuals
A counterfactual is a conditional sentence in the subjunctive mood. The term coun-
terfactual or contrary-to-fact conditional carries the suggestion that the anteced-
ent of such a conditional is false. Consider, for example: If this glass had been
struck, then it would have shattered. The implication is that the glass was actually
not struck and did not shatter. Yet these conditionals do not appear to dier in kind
from those in which the truth value of the antecedent is an open question (e.g., If
this glass were struck, then it would shatter). While this suggests that the expres-
sion subjunctive conditional might more appropriately delineate the topic, the
2 J. Collins, N. Hall, and L. A. Paul
term counterfactual is by now so well entrenched that it will suce to stipulate
that counterfactuals may have true antecedents.
The philosophical signicance of counterfactuals is apparent from the way in
which they have gured so prominently in recent discussions of many other philo-
sophically important concepts: knowledge, perception, and freedom of the will, to
name just a few. An attractive thought is that this conceptual connection is best
understood as being mediated by the concept of cause. Thus, for example, counter-
factuals are relevant to discussions of perception because the most promising theories
of perception are causal theories, and counterfactuals are fundamental to any philo-
sophical understanding of causation.
The thought from which counterfactual theories of causation proceeds is this. A
certain glass is struck, and shatters. To say that the striking of the glass caused the
shattering of the glass is to say that if the glass had not been struck then it would
not have shattered. The striking caused the shattering in virtue of the fact that the
shattering was counterfactually dependent on the striking. The papers in the present
volume are devoted to exploring the prospects of an account of causation based on
that simple idea. But before turning to that discussion, we shall review some of the
basic semantic features of counterfactuals.
Counterfactual conditionals themselves pose a problem that must be solved before
such conditionals are recognized as a t tool for further philosophical analysis. The
problem, rst raised in the recent philosophical literature by Chisholm (1946) and
Goodman (1947), is to provide noncircular truth conditions for the general counter-
factual conditional with antecedent A and consequent C. We shall abbreviate this
conditional as A LC, to be read: If it were the case that A, then it would be the
case that C.
Observe rst of all that the counterfactual must be distinguished from both the
material conditional and strict entailment. The material conditionals truth condi-
tions are too weak; not every subjunctive conditional with a false antecedent is true.
But on the other hand, the strict conditional is too strong; the shattering of the glass
is not logically entailed, or metaphysically necessitated, by the striking of the glass.
The right account is, presumably, located somewhere between these extremes. A
natural next thought is that A LC is true if and only if C is entailed not by A
alone, but by A in conjunction with some other truths: these might include, for
example, fundamental truths about the laws of nature. The problem with this sug-
gestion is that no single set of xed truths will do the job for all A and C. As Lewis
(1973b) notes, specifying such a set would incorrectly include strengthening the
antecedent among the valid forms of inference for counterfactuals, yet counter-
factuals are clearly nonmonotonic. For example, from the fact that a particular match
Counterfactuals and Causation: History, Problems, and Prospects 3
would light if it were struck, it does not follow that if the match were struck and there
were no oxygen present, then it would light.
Counterfactuals are variably strict conditionals. In evaluating a counterfactual
we see what is entailed by the antecedent along with some other truths, but the set
of truths to be held xed varies from case to case in a way that is determined by the
antecedent A. In Goodmans phrase: We hold xed everything that is cotenable with
the truth of the antecedent. The threat of circularity should now be apparent, for
what could it mean for a proposition to be cotenable with A, other than that the
proposition is one that would still be true if A were true? One can hardly explain
counterfactuals in terms of the notion of cotenability and then proceed to give a
counterfactual analysis of cotenable.
A way out of this circle was presented in the work of Stalnaker (1968) and Lewis
(1973b), inspired by the development of possible worlds semantics in modal logic.
Central to the StalnakerLewis approach is the assumption that possible worlds
can be ordered with respect to their similarity to the actual world. Since the relation
of comparative similarity to the actual world is assumed to be a weak order (i.e., a
connected and transitive relation) we may usefully think of it as a relation of com-
parative closeness. Call a world at which the proposition A is true an A-world.
Then the counterfactual A LC is true if and only if C is true at the closest A-
world to the actual world. If, following Lewis (1973b), we want to allow that there
may not be a single A-world, or even any set of A-worlds, closest to the actual world,
the truth condition is more appropriately given like this:
A LC is true if and only if some (A&C)-world is more similar to the actual
world than any (A&@C)-world is.
It is worth noting how much of the logic of counterfactuals follows simply from
the assumption that the similarity relation is a weak order. We can see immediately,
for example, the particular logical form of the fallacy of strengthening the anteced-
ent. It is a fallacy analogous to the mistake made by someone who infers from
There is no hardware store in the closest town to here that has a bank that There
is no hardware store in the closest town to here that has a bank and a restaurant.
Further exercise of the analogy between similarity and closeness leads to the correct
prediction that contraposition is also invalid, and that counterfactuals fail to be
transitive. Appreciation of these points will be key to the discussion that follows.
Although much of the logic of counterfactuals follows just from the assumption
that any reasonable relation of comparative similarity will be a weak ordering, we
will need to say something more substantial about the notion of similarity if our hope
is to provide an account of counterfactuals on which a theory of causation can be
based. The problem is that similarity is multidimensional. If two worlds are each
similar to the actual world in a dierent respect, which is most similar to the actual
world all things considered?
We might in general be content to consign the weighing of respects of similarity
and dierence to the pragmatic rather than the semantic part of the theory. So we
might say: In one context it is appropriate to assert that Caesar would have used the
atom bomb had he been in command in Korea; in another that he would have used
catapults (example from Quine 1960, p. 222). One of the contributors to this volume
(Maslen, chapter 14) is happy to allow that because of this feature of counterfactuals,
the causal relation itself is context dependentthat whether or not it is correct to
describe one event as a cause of another is relative to the contrast situation one has in
mind. But those who hope for a context-independent counterfactual account of cau-
sation must confront the issue of respects of similarity head on.
The issue arises most sharply in connection with temporal considerations. Causes
precede their eects. A defender of a counterfactual theory of causation will likely
think that this temporal asymmetry in causation reects an asymmetry in counter-
factual dependence: Whereas the future depends counterfactually on the past and
present, the past is not counterfactually dependent on the present and future.
Admittedly, we are sometimes willing to talk about how things would have been
dierent in the past, had they been dierent now. (If I had pulled the trigger just
now, there would have to have been no one in the line of re. Im not a homicidal
maniac!) But as Lewis (1979a) notes, such backtracking conditionals are typically
marked by a syntactic peculiaritynote the have to have been construction in
the above example. Backtracking conditionals will have to be set aside if we want to
defend a counterfactual theory of causation. And they will have to be ruled out in
some principled way. This provides a clue to the problem of weighing respects of
similarity.
It is clear, as we noted above, that similarities with respect to the laws of nature
(nomological similarities) should generally outweigh similarities with respect to acci-
dental matters of fact in the all-things-considered evaluation of similarity. But
apparently this cannot be an invariable rule. Thats one of the lessons of temporal
asymmetry: If the actual laws of nature are deterministic toward both past and
future, then any world that shares these laws yet diers from the actual world in
some particular present matter of fact will have not only a dierent future from the
actual world, but a dierent past as well. We then have backtracking on a massive
scale.
So we must allow that certain minor violations of the laws of nature (small
miracles) may be compensated for by agreement in past matters of particular fact.
Yet it would hardly do to secure this agreement by stipulation. That would incor-
rectly (to our minds) deem backward causation an a priori impossibility. (Later in
this section we will see that this issue is more complicated than the present remarks
suggest, as accommodating backward causation turns out to be trickier than it might
seem.) Nor, in the present context, is it an attractive option to appeal to causal
notionsrequiring, for example, that in a counterfactual situation in which some
(actually occurring) event c does not occur, everything either causally antecedent to
or causally independent of c be held xed. Such a stipulation might guarantee the
right truth values for the counterfactuals that feature in a counterfactual analysis
of causation, but it would do so at too high a price in circularity.2 The challenge for
the StalnakerLewis account of counterfactuals is to provide criteria for weighing
respects of comparative similarity that yield temporal asymmetry without doing so
by at.
The challenge has been pressed most vigorously in connection with what is known
as the future similarity objection (see Fine 1975, and references in Lewis 1986a, 43
to seven more published versions of the objection). We may suppose the following
counterfactual is true:
If Nixon had pressed the button, there would have been a nuclear holocaust.
But it seems as though according to the StalnakerLewis theory it will come out
false. Suppose a nuclear war never actually takes place; then any world in which no
nuclear war takes place will be more similar to the actual world than any world in
which a nuclear war does take place. A nuclear war would, after all, make things
very dierent from the way they actually are. Hence, given any world in which the
antecedent and consequent are both true, it is easier to imagine a closer world in
which the antecedent is true and the consequent false. Just imagine some minor
change that prevents the nuclear disaster from occurring.
In response Lewis says:
The thing to do is not to start by deciding, once and for all, what we think about similarity of
worlds, so that we can afterwards use these decisions to test [the analysis]. . . . Rather, we must
use what we know about the truth and falsity of counterfactuals to see if we can nd some sort
of similarity relationnot necessarily the rst one that springs to mindthat combines with
[the analysis] to yield the proper truth conditions. (Lewis 1979a, p. 43)
By pursuing this strategy of reverse engineering the similarity ordering from
our intuitions about particular conditionals, Lewis obtains a set of criteria (1979a,
pp. 4748) that, he argues, suce to rule out backtracking in worlds like ours with-
out making backward causation a conceptual impossibility. In brief, Lewis argues
as follows: Take some event c that occurs at time t. Assuming determinismand
assuming that the world displays certain de facto temporal asymmetries that Lewis
attempts to characterize (it is this assumption that is supposed to leave the door open
for backward causation)the only way to construct a counterfactual scenario in
which (i) c does not occur at t, and (ii) there are no gratuitous departures from actu-
ality, is to insert, shortly before t, a divergence miracle: a localized violation of
actual (though not counterfactual) law that throws history o course just enough to
make it the case that c does not occur. Thus, we get a counterfactual situation in
which history until shortly before t is exactly as it actually is, but whose history after
t diverges substantially from actual history; moreover, only during a relatively short
interval of time, and in a relatively localized region of space, do the events of this
world fail to conform to the actual laws. Lewis argues that we cannot do things in
reverse: There is no similar way to insert a single localized reconvergence miracle
so as to arrive at a counterfactual scenario whose past before t diers from the actual
past, but whose future from shortly after t onward is exactly like the actual future.
It is in this way that his truth conditions secure a nonbacktracking reading of the
conditional.
But worries remain on this score. For example, it is not clear whether future simi-
larity should count for absolutely nothing in the ordering, or whether it should
merely have relatively little weight. Here it helps to think about examples that are
not deterministic, but which involve some genuine chance process. For simplicity,
suppose that a certain coin toss is genuinely chancy. If I havent tossed the coin yet,
and the coin is fair, it seems wrong to say that If I were to toss the coin, then it
would land heads. It seems wrong to say this even if, as a matter of fact, the coin
actually does land heads when I get around to tossing it. This suggests that worlds in
which the coin will land tails when tossed are as close to the actual world as any
world in which the coin will land heads. This means that there are nonactual worlds
that are as similar to the actual world as the actual world is to itself. And of course
thats just what one would expect to be the case if a certain respect of similarity or
dierence is to count for nothing.
But the interplay between counterfactuals and chance involves some subtlety. As
Sid Morgenbesser reminds us, if I oer to bet you that the coin will land tails and
you decline the wager, then I can say later, after the coin lands heads, Look, you
would have won! How can the StalnakerLewis account reconcile these conicting
intuitions?
A second and quite serious diculty has recently been pressed by Adam Elga
(2000).3 Elga argues, very persuasively, that Lewiss attempt to provide truth con-
ditions for a nonbacktracking reading of the counterfactual runs afoul of statistical
mechanics. Recall the dialectic: For any event c occurring at time t in world w, there
are worlds w
0
whose histories match that of w exactly until shortly before t, but
which then exhibit a small divergence miracle that throws history o course just
enough to guarantee cs nonoccurrence. But, Lewis claims (and must claim), pro-
vided w exhibits the appropriate de facto temporal asymmetries, there are no worlds
w
0
in which (i) c does not occur; (ii) history from shortly after t on matches ws his-
tory exactly; and (iii) this perfect match comes about as a result of a small recon-
vergence miracle.
Elga shows that this last claim is falseat least, it is false for many typical choices
for c, in any world that displays the global entropic features of our own. Suppose, to
borrow his example, that Suzy cracks an egg into a hot pan, and it cooks. Let c be
her cracking of the egg. Pick a time t shortly after it has cooked. Now look at what
happens from t backward: A cooked egg sits in a pan, gradually uncooking; it co-
alesces into a raw egg and leaps upward; a shell closes around it, forming a perfect
seal. Surprisingly, the kind of time-reversibility that the laws of our world apparently
exhibit allows that this entropy-diminishing process (ipped over in time, and run
from past to future) is possible (see Albert 2000 for details); but they also guarantee
that it is highly unstable, in the sense that a slight change in the initial conditions will
upset the delicately timed processes that result in the uncooking and sealing of the
egg (it is in establishing this result that statistical mechanics comes into play). Make
such a slight changesay, by altering the positions of the molecules in one corner of
the panand the egg will just sit there, slowly growing colder.
But, as Elga points out, such a slight change, inserted at t in the actual world,
accomplishes exactly what is required of a reconvergence miracle. That is, we
construct a possible world by holding history beyond t xed, making a slight, local-
ized change in the state of the world at t, and running the laws backward: In such
a world, the egg is never cracked, and so c does not occur. The needed asymmetry
between divergence and reconvergence miracles vanishes. There is a real irony here:
Lewis closes his famous paper by observing, I regret that I do not know how to
connect the several asymmetries I have discussed and the famous asymmetry of
entropy (1986a, p. 51).
Though serious, this problem should not be overstated. It is not just that some
way might be found to amend Lewiss account in response to Elgas challenge (for
example, one might construct standards of similarity that give signicant weight to
the global entropic prole of a world).4 It is also that the philosopher who desires
a nonbacktracking reading of the counterfactual solely to serve the purposes of her
counterfactual analysis of causationand not also, as in Lewis (1986a), to provide
an account of the asymmetry between past and futurecan fall back on a recipe for
evaluating counterfactuals that guarantees, more or less by stipulation, that they will
have the needed nonbacktracking reading. This can be done as follows:
The rst step is to require that causes must precede their eects. (More, in a
moment, on whether this requirement carries too high a price.) Then the counter-
factual analyst need only consider conditionals of the form @Oc L@Oe where
c and e both occur, and c before e. (Oc is short for c occurs, etc.) Suppose
we are given such a conditional. To construct the relevant counterfactual situation,
focus rst on the time t at which c actually occurs. Consider the complete physical
state of the world at that time. (Or: the complete state on some spacelike hypersur-
face that includes the region of cs occurrence.) Make changes to that state localized
to the region in which c occurs, and just sucient to make it the case that c does not
occur. (These are the changes that Lewiss account would have brought about by a
miracle.) Evolve the new state forward in time in accordance with the actual laws,
and check to see whether e occurs. If it does, the conditional is false; if it doesnt, true.
Matters might be more complicated: Perhaps there are many equally minimal
ways to change the given state, or an innite succession of ever-more-minimal ways.
If so, we employ the obvious sophistication: the conditional @Oc L@Oe is
true just in case some @Oc-state that yields a @Oe-future is closer to the actual
state (i.e., involves a more minimal change from the actual state) than is any @Oc-
state that yields an Oe-future.
Notice that in following this recipe we have implicitly replaced a similarity metric
on worlds with a similarity metric on states, or if you like, worlds-at-times. Not a
problembut it does make clear (what was perhaps obvious anyway) that the recipe
is limited in scope, working best (perhaps only) when the antecedent of the condi-
tional is a proposition entirely about a particular time or narrow interval of time.5
The counterfactual analyst should not worry: After all, she is doing the metaphysics
of causation, and not the semantics of some fragment of English. And her counter-
factual conditional evidently has enough in common with the ordinary language one
still to deserve the name.
But two other problems plague the counterfactual analyst who takes our sugges-
tion for responding to Elgas challenge. She has stipulated that causes must precede
their eects. Has she not thereby ruled outa prioriboth simultaneous causa-
tion and backward causation? But both are surely possible. As to the rst, we have,
for example, the situation in which the particles presence at a certain location in
the force eld at time t causes its acceleration with a certain magnitude at time t.
As to the second, we have any of a number of seemingly perfectly consistent time-
travel stories, for example, the following: A billiard ball subject to no forces ies
freely through space; call this ball Bill. O to the left is the Time Travel Exit
Portal. O to the right is the Time Travel Entrance Portal. At time 0, a billiard ball
call it Sueemerges from the Exit Portal, ies toward Bill, and strikes it. Bill
goes careening o toward the Entrance Portal, and crosses it at time 1and, of
course, leaves the Exit Portal at time 0: Bill and Sue are one and the same.6
Examples such as these appear to depict genuine possibilities, but if causes must
precede their eects then it seems that this appearance is illusory. In fact, however,
we think matters are not nearly so simple. For reasons of space we will forgo a full
discussion; but we will try to say enough to make it clear that these two problems
should not prevent one from taking seriously the method we have suggested for
guaranteeing the asymmetry of causation.
As to simultaneous causation, one promising strategy is to deny that the examples
allegedly depicting it in fact succeed in doing so. Consider the example we gave,
where the eect supposedly contemporaneous with its cause is the acceleration of
a particle at time t. We suggest that the example gains its intuitive appeal only so
long as we remain naive about the underlying physics: what it is for a particle to be
accelerating with a certain magnitude at time t is for its motionin an interval of
time extending forward from tto display a certain character.7 Put another way, it is
not an intrinsic property of the state of the particleor indeed of the worldas it is
at time t that the particle has the given acceleration, then; rather, it is a relational
property, where the relatum is the trajectory of the particle in the immediate future of
t. We could properly insist, then, that when we loosely speak of the particles position
in the force eld at time t causing its acceleration at that time, what we mean is that
its position in the eld causes a certain aspect of its subsequent motion. But then any
conict with the stipulation that causes must precede their eects disappears. Indeed,
the example begins to look like the following example of simultaneous causation:
Say that a window is doomed at time t just in case it will shatter within ten seconds
of t. Suzy throws a rock at a window at nooninstantaneously causing the event of
its being doomed at noon. That is silly, the kind of example about which no com-
petent philosopher would think twice. We think that careful reection exposes the
example of the particle in the force eld as being equally silly. And, while we will not
pursue the matter here, we suspect that other alleged cases of simultaneous causation
can be given a similar expose.
Backward causation requires, we think, a dierent treatment, for unlike some
authors (e.g., Mellor 1998, chapter 12), we do not wish to deny its possibility.
Accordingly, we suggest a much more defensive maneuver, which is to point out that
the need to accommodate backward causation provides, surprisingly, no reason
whatsoever to return to Lewiss miracle-based account of the truth conditions of the
counterfactual.
To see why, return to our example of the collision of Bill/Sue with itself. Focus on
a time moments before the ball crosses the Time Travel Entrance Portal. Let c be the
event of its ying toward the Entrance, at that time; let e be the earlier event of its
ying out of the Exit. Clearly, c causes e. But whether e counterfactually depends on
c is not at all clear, and that is because it is not at all clear what is going on in a
counterfactual situation in which c does not occur. Try rst to construct this situa-
tion in a at-footed way, by inserting a divergence miracle into the actual course
of events just before c occurs, a miracle just sucient to guarantee that c doesnt
occur. You will quickly nd that limiting yourself to just one such miracle cannot
possibly do, for remember that the rest of what goes onthe totality of events that
follow your miracle, and the totality of events that precede itmust conform to
actual law. Suppose for illustration that your miracle makes the ball vanish, just
before it reaches the Entrance. Then, by law, it cannot leave the Exitnot at time 0,
nor at any other time. And there are no other balls in the picture. So no ball ever
leaves the Exit. So no ball ever collides with Bill/Sue. So Bill/Sue continues on its
original course. So, far from vanishing just before it reaches the Entrance at time 1,
the ball is nowhere near it, then. Contradiction.
We could try to x matters by inserting a second localized miracle: Let the ball
(miraculously) leave the Exit at time 0, even though it does not cross it at time 1.
Then we have a consistent story, most of which is compatible with the actual laws:
The ball is ying through space; a ball (miraculously) ies out of the Exit at time 0;
the two collide, sending the original ball toward the Entrance; just before time 1, it
(miraculously) vanishes. But e occurs, in this scenario. So if this is what would have
happened, had c not occurred, then e does not depend on c. So the counterfactual
analysis gets the causal structure of our case wrong.
We could also try to x matters by insisting that the counterfactual situation in
which c does not occur is simply one in which nothing ever leaves the Exit or crosses
the Entrance: instead, the ball just ies through space, unmolested. So e does not
occur; soif this is what would have happenede depends on c, after all. True, we
do not reach this situation by inserting a miracle into the actual course of events. But
Lewiss ocial criteria for similarity of worlds do not require miracles: One is sup-
posed to introduce them only if doing so is necessary to secure widespread and per-
fect match of particular fact.
The problem should now be obvious, for securing such a perfect match is exactly
what we accomplish by introducing a second miracle, whereby the ball leaves the
Exit, even though it does not cross the Entrance. In fact, the rst counterfactual sit-
uation we described matches the actual situation perfectly, everywhere, except for a
tiny interval before time 1, in a small region of space near the Entrance. By Lewiss
own criteria, it should therefore win out over the second situation. So the account of
the truth conditions for counterfactuals in which these criteria feature does not, after
all, seem so clearly preferable to the asymmetry by at account sketched above.
There is an important open question here, which we must simply note and pass
over: How exactly can any sort of reductive account of causationcounterfactual
or otherwiseaccommodate backward causation? We strongly suspect that the best
answers will end up treating backward causation as a kind of special case. If so, that
may prove to be more good news for our suggested truth conditions for the counter-
factual, for it may be to their credit that they allow for no straightforward treatment
of backward causation within the context of a counterfactual analysis. At any rate,
those truth conditions evidently provide the ingredients necessary for pursuing a
counterfactual account of less exotic varieties of causation. Let us now consider the
prospects for such an account.
2 Reductionism about Causation
It will be helpful to begin by situating counterfactual approaches to causation within
the much larger constellation of philosophical accounts of causation quite generally.
Probably the most useful distinction to make at the outset is that between accounts
that do and accounts that do not attempt to reduce causal facts to facts about what
happens, together with facts about what the laws are that govern what happens. (We
have a permissive sense of what happens in mind: It is to include facts about what
objects exist where and when, and what categorical properties and relations they
instantiate.) Immediately, two questions arise, which we will not pursue in any detail:
First, can facts about the laws themselves be reduced to the totality of categorical
facts? Somenotably, Lewis (e.g., 1983b)will say yes, others no. We wish
merely to note that with respect to the aims of the essays in this volume, almost
nothing hangs on this dispute. (For example, Lewiss account of causation could be
adopted wholesale by one who disagreed with him about whether the laws them-
selves reduce to categorical facts.) Second, is there in fact a legitimate distinction to
be drawn between categorical and noncategorical facts about the world? Put another
way, is it possible to specify, in noncausal terms, the facts about what happens that
form part of our reduction base? The facts about the laws? We will simply proceed
under the assumption of an armative answer; but see Shoemaker (1980) and Cart-
wright (1999) for inuential statements of contrary views.
We will henceforth label reductionist any position according to which causal
facts can be reduced to categorical plus nomological facts, and label antireduc-
tionist any position that denies this claim. Most of the essays in this volume are
written from a reductionist perspectiveor at any rate, from a perspective entirely
congenial to reductionism. (Armstrong, chapter 19 of this volume, Going through
the Open Door Again, is a notable exception.) It is useful to add, however, that
some antireductionists nd strong motivation for their view in certain key thought
experiments. Here, for example, is one provided by Michael Tooley (1990, p. 228,
italics in the original):
Given [the assumption that there is nothing incoherent in the idea of an uncaused event], one
can consider a world where objects sometimes acquire property Q without there being any
cause of their doing so, and similarly for property R, and where, in addition, the following two
statements are true:
(1) It is a law that, for any object x, xs having property P for a temporal interval of length Dt
either causes x to acquire property Q, or else causes x to acquire property R;
(2) It can never be the case, for any object x, that xs having property P for a temporal interval
of length Dt causes x to acquire both property Q and property R.
Suppose, nally, that an object a in such a world, having had property P for the appropriate
interval, acquires both Q and R. In view of the law described in statement (1), either the
acquisition of Q was caused by the possession of P for the relevant interval, or else the acqui-
sition of R was so caused. But, given statement (2), it cannot be the case that the possession
of P for the relevant interval caused both the acquisition of Q and the acquisition of R. So
once again, it must be the case that one of two causal states of aairs obtains, but the totality
of facts concerning, rst, the non-causal properties of, and relations between, events, secondly,
what laws there are, and thirdly, the direction of causation in all potential causal processes,
does not suce to x which causal state of aairs obtains.
In section 5 below, where we take up a number of tricky methodological issues, we
will briey touch on the question of how much probative value thought experiments
like this have (to anticipate: not much). Let us now proceed to draw further distinc-
tions between reductionist approaches to causation.
Once again, a particular distinction stands out as especially helpful: that between
what we will call physical connection accounts of causation and what we will call
nomological entailment accounts. Examples will work better than denitions to
give the idea.
Physical connection accounts, although reductionist, appeal to the fundamental
laws only indirectlytypically, in the specication of some quantity whose trans-
fer is thought to constitute the causal relation.8 Thus, we might say, with Fair
(1979), that causation consists in the transfer of energy. Or, with Ehring (1997), that
it occurs in the transfer of a trope; or, with Dowe (1992) and Salmon (1994), in the
transfer of some conserved quantitywhere appeal is made to fundamental physics
for an inventory of such quantities. This last example shows most clearly how the
fundamental laws can have a place in such accounts.
We think such accounts suer from a lack of ambition, twice over: First, however
successful they are at limning the features of causation, as it relates events in the
purely microphysical realm, there seems little hope that they can succeed in doing so
in the messy macroscopic realm. Suzy kisses Billy, causing him to ush; are we to
suppose that the causal relation between these two events is to be mapped out by
looking at how energy, or some other fundamental physical quantity, is transferred?
Well, it might be romantic to say so. But it is not, we think, particularly enlightening.
Second, assuming, as seems reasonable, that it is a contingent matter what the
fundamental laws are, physical connection accounts cannot (because they are not
designed to) tell us anything about causation as it might have beenin particular, as
it is in worlds with laws very dierent from our own. That limitation seems not
merely unfortunate but deeply misguided; for it seems clear that although the fun-
damental laws play an essential role in xing the causal facts, they do not do so in so
specialized a manner. Put another way, it seems both reasonable and worthwhile to
try to specify the way in which the fundamental laws x the causal facts in terms that
abstract away from the gory details of those lawsthereby to produce an account
that has a hope of proving to be not merely true, but necessarily so.
That is what nomological entailment accounts attempt to do. As a simple
exampleone that shows where the label nomological entailment comes from
consider the view that takes c to be a cause of e just in case it follows, from the
proposition that c occurs, together with the proposition that encapsulates the funda-
mental laws, that e occurs (where the following relation could be spelled out in
terms of deductive entailment, or in terms of some notion of metaphysical necessity:
more on this in the next section). Or we might go the other way: c is a cause of e just
in case it follows, from the proposition that e occurs, together with the proposition
that encapsulates the fundamental laws, that c occurs. Or we might go both ways
at once. (Lesson: do not take the entailment in question invariably to proceed
from cause to eect.) The defects in these accounts are perfectly obvious, but it is also
obvious that they do not suer from the two kinds of constriction that made physical
connection accounts so disappointing.
(The distinction between physical connection accounts and nomological entail-
ment accounts does not exhaust the range of reductionist accounts of causation.
Maudlins Causation, Counterfactuals, and the Third Factor [chapter 18] provides
an excellent illustration of this point: He argues that our ability to apply any dis-
criminating notion of cause is parasitic on our ability to analyze the laws governing
the situation in question into, on the one hand, laws that specify how things will
behave if left undisturbed, and, on the other hand, laws that specify the exact con-
sequences of particular disturbances.)
Of course, more sophisticated and correspondingly more interesting nomological
entailment accounts are easy to nd. For example, we might side with Mackie
(1965): c is a cause of e i, from the proposition that c occurs, together with the
proposition that encapsulates the fundamental laws, together with some suitably
chosen proposition about conditions that obtain at the time of cs occurrence, it fol-
lows that e occurs, where the premise that c occurs is essential to the entailment.
Or we might try building on the idea that c is a cause of e just in case the condi-
tional probability that e occurs, given that c occurs, is greater than the unconditional
probability that e occurs. (For a sophisticated version, see Eells 1991.) Note that in
calling this a nomological entailment account, we are, in the rst place, assuming
that the relevant probabilities are somehow determined by the laws (i.e., they are
not subjective probabilities); and, in the second place, we are permissively counting
probability-raising as a kind of entailment. No matter: consider this a small sacri-
ce of accuracy for the sake of having a handy label.
More to the point, we might endorse the simplest counterfactual analysis: c is a
cause of e i, had c not occurred, e would not have occurredwhere this entailment
relation between the proposition that c does not occur and that e does not occur
counts as nomological because of the central role played by the fundamental laws
in xing the truth conditions of the counterfactual (see section 1). Or we might hold
some more sophisticated counterfactual analysis that builds on the core idea con-
tained in the simple version. Indeed, many philosophers hold out such rm hope for
some sort of counterfactual analysis precisely because they are convinced, in the rst
place, that some sort of nomological entailment account must be correct; and, in the
second place, that no rival to the counterfactual approach has a prayer. (See Lewiss
Causation as Inuence, chapter 3 in this volume, for a particularly forceful defense
of this motivation.) Be that as it may, in attempting such an analysis one should
certainly come armed with a clear view of the major obstacles, a host of which have
emerged over the course of the last twenty-ve years. We will now briey survey the
most important of them.
3 The Original Lewis Analysis
Let us rst take stock of the principal advantages of the counterfactual approach.
Focus on one of the simplest counterfactual analyses: Lewiss original (1973a) anal-
ysis. He took causation to be the ancestral of counterfactual dependence: c causes e i
c and e both occur, c and e are distinct, and there is a (possibly empty) set of events
fd
1
; d
2
; . . . ; d
n
g such that if c had not occurred, d
1
would not have occurred; and if
d
1
had not occurred, d
2
would not have occurred; . . . and if d
n
had not occurred, e
would not have occurred.
The analysis is attractively simple and extremely intuitive, given how obvious it
seems that there must be some intimate connection between causation and counter-
factual dependence. To be sure, the connection need not be explained by supposing
that causation must be given a counterfactual analysis; see Maudlins contribution
(chapter 18) for a rival explanation. But the prima facie plausibility of a counter-
factual analysis is suciently clear that no one can reasonably doubt that exploring
its prospects is worthwhile.
Next, the account makes it exquisitely clear what role the laws of nature play in
xing the causal factsand moreover, it makes it clear that this role does not require
too much of the laws. Let us explain what we mean by this by contrasting the Lewis
analysis with a Davidsonian account of causation. According to this latter kind of
account, what distinguishes a true causal claim of the form c is a cause of e as such
is that, when the names c and e are replaced by suitable uniquely identifying
descriptions of the respective events, the resulting sentence can be seen to be a
deductive consequence of some universal generalization or generalizations that hold
with lawful necessity (together with additional premises specifying that the given
events in fact occur). Morton White (1965, pp. 56104) presents a clear statement of
this sort of approach.9
This Davidsonian account faces a number of diculties, but for our purposes the
one to focus on is that it places severe constraints on an account of laws that can
meet its needs. Are we to suppose, for example, that the lawful universal general-
izations come directly from fundamental physicsso that the uniquely identifying
descriptions have to be descriptions in the language of fundamental physics? Or is it
rather that we can pick out higher-level laws tailored to any domain where we can
nd true causal claims? But then as good reductionists, we should want to know how
these higher-level laws themselves reduce to facts about what happens, together
with facts about the fundamental laws. Even if we are not good reductionists, we
should want to know thisfor unless we are told, we should view with serious skep-
ticism the claim that there even are such suitable higher-level laws (enough of them,
anyway).
Our simple counterfactual analysis elegantly avoids these issues: As we saw in
section 1, all we need by way of an account of laws is an account of which worlds are
nomologically possible. Put another way, that is all that the laws need to contribute to
the truth conditions for the relevant counterfactual conditionals. We do not need
what the Davidsonian requires: for each causal relation, a law that will directly
cover it, at least when the relata are described appropriately. (In hindsight, the
crucial mistake the Davidsonian committed here was to require a deductive relation
between sentences rather than a kind of metaphysical entailment relation between
propositions.)
Next, if we remember that the semantics used for the kind of counterfactual con-
ditional appropriate to counterfactual analyses will guarantee a nonbacktracking
reading of that conditional, then we can appreciate how the Lewis analysis neatly
avoids three problems that were thorns in the side of earlier regularity analyses. First,
in many cases where c is a cause of e, it will turn out that e is lawfully necessary and/
or sucient for c (at least, in the circumstances); on a number of regularity analyses,
it will follow that e is a cause of c. More generally, those analyses tend to face a
serious problem in explaining the asymmetry of the causal relation. One way to
guarantee such asymmetry is to insist that causes must precede their eects, thus
piggybacking causal asymmetry on the asymmetry between past and future. But
doing so seems to rule out rather hastily the possibility of either simultaneous or
backward causation.
Still, our discussion at the end of the last section may show that simultaneous
causation need not be taken too seriously, and that backward causation is everyones
problem. So observe that this maneuver is of no help in solving the second problem,
which arises from cases in which an event c causes two events d and e along dierent
and independent causal routes. As an illustration, consider gure 1.1. Here, c res,
sending a stimulatory signal to d, causing it to re. (Circles represent neurons; arrows
represent stimulatory connections; shaded circles represent neurons that are ring;
the time order of events is left-to-right. In the text, bold letters name neurons, itali-
cized letters the events of their ring.) In addition, cs ring causes e to re, by way of
the intermediate neuron b. Now, given a suitable specication of the circumstances
(e.g., which neurons are connected to which other neurons), it follows from the fact
that d res, together with the laws, that e res; and likewise vice versa. Moreover, the
d
e
c
b
Figure 1.1
ring of d precedes the ring of e. But the one is not a cause of the othertrouble,
once again, for any of a number of regularity analyses of causation. And stipulating
that causes must precede their eects helps not at all.
For the third problem, consider gure 1.2. Here, the line with a blob at the end
represents an inhibitory signal. Thus, c res, stimulating d to re, which in turn
stimulates e to re. Meanwhile, a res, but the stimulatory signal it sends to b is
blocked by an inhibitory signal from c, so b does not re. Again, given a suitable
specication of the circumstances, it follows from the fact that a res, together with
the laws, that e will re. But to conclude from this fact, together, perhaps, with the
fact that the ring of a precedes the ring of e, that a is a cause of e is simply to
confuse guaranteeing an outcome with causing that outcome.
Thanks to the stipulation that the counterfactual conditional is to be given a non-
backtracking reading, Lewiss analysis cleanly avoids each of these three problems
(bracketing, for the moment, the reservations voiced in section 1 about some of the
principal attempts to provide the needed truth conditions for nonbacktracking
counterfactuals). If the eect had not occurred, we are not entitled to conclude that it
must have been the case that the cause did not occur: for to do so would be to impose
the forbidden backtracking reading of the conditional. In gure 1.1, in evaluating
what would have happened if d had not red, we hold xed the occurrence of other
events contemporaneous with din particular, the ring of band hence conclude
that it still would have been the case that e red. Likewise in gure 1.2, if d had not
red, we do not conclude that it must (or even might) have been the case that b red
(on the basis that it must have been the case that c did not re, and so must have
been the case that no signal from c prevented b from ring), so that e would (or
might) have red all the same. Rather, we hold xed the nonring of b, and so
endorse the conditional that if d had not red, e would not have red. (This last
example, by the way, shows that the prohibition on backtracking is needed not
merely to secure the suciency of the Lewis analysis, but also to secure its necessity:
See Hall, Two Concepts of Causation, chapter 9, pp. 233234.)
a b
c d
e
Figure 1.2
Thus, confronted with a suite of extraordinarily simple and straightforward
examples that nevertheless proved stubbornly resistant to the regularity analysts
treatment, the counterfactual analysis appears to succeed quite eortlessly. Whats
more, the Lewis analysis appears to have the resources to handle quite naturally both
causation by omission and causation under probabilistic laws. For the rst, observe
that we can take the failure of c to occur (or perhaps better: the failure of an event of
type C to occur) to cause event e just in case, had c occurred (or: had an event of
type C occurred), e would not have occurred. (Causation of omission can be given a
similar treatment.) For the second, the counterfactual analyst can build on the idea
that even under probabilistic laws, causes typically make a positive dierence to the
chances of their eects. That is, what distinguishes c as a cause of e, in the rst
instance, is that if c had not occurred, the chance of es occurring would have been
much less than in fact it was (Lewis 1986b). Note that in the deterministic case, this
condition reduces to the condition that had c not occurred, e would not have oc-
curred (since the only chances available are 0 and 1). Many have therefore viewed
its apparent ability to provide a uniform analysis of causation under both determin-
ism and indeterminism to be a signal advantage of the counterfactual analysis.
The successes of the Lewis analysis are impressive; little wonder that Lewiss orig-
inal paper had the inuence it did. Still, further developments quickly showed the
need to revise his proposal. We will focus now on the principal sources of pressure
for revision, beginning with problems well known for quite some time, and moving
to problems that have only recently cropped up in the literature.
4 Problems, Old and New
Some problems take the form of clean, simple counterexamples to the Lewis analysis.
Others take the form of challenges to lay out its foundations more precisely. We will
begin with three problems of the latter variety.
First, for reasons already addressed, one might doubt whether an adequate
semantics for the counterfactual can be produced that will yield the needed non-
backtracking reading of that conditional. Observe that insofar as the Lewis analysis
aims to be reductive, the project of giving such a semantics becomes especially dif-
cult. We cannot say, for example, that we are to arrive at the nonbacktracking
reading of if c had not occurred, then e would not have occurred by considering a
counterfactual situation in which c does not occur, but in which all of its causes do
(and then checking to see whether, in this situation, e occurs).
Waiving this problem, one might worry, second, that we do not really have a
handle on what sort of counterfactual situation we are supposed to be envisioning.
That is, we are told that we must consider what would have happened if a particular
event c had not occurred. To see how dicult it might be to do so, consider a simple
example: Billy and Suzy converse, say, between noon and 12:30 one day. Let c be
their conversation. What would have happened if cthat very conversationhad
not occurred? Ohand, it is extremely dicult to say, simply because it is extremely
dicult to discern where the boundary lies between those possible worlds in which
that very conversation occurs, and those possible worlds in which it does not. After
all, to map this boundary we need answers to the following sorts of questions: Would
c have occurred if Billy and Suzy had conversed, but about an entirely dierent
topic? Would it have occurred if they had conversed about the same topic, but an
hour later? A day later? A few yards away from where they in fact conversed? A
hundred miles? Could c have involved entirely dierent peopleor was it essential to
it that it be a conversation between Billy and Suzy? Could their conversation c have
been, in fact, a race they were running together? Presumably notbut why not? To
answer these and other questions, we should need an account of which features of it
are essential to their conversation, and which accidental. Without such an account, it
seems that we are at sea with respect to evaluating the very conditionals that form
the heart of the Lewis analysis.
Unfortunately, this problem has tended to get overlooked in the literature. (See
Bennett 1988 for a notable and welcome exception.) Part of the reason, no doubt, is
that in the sorts of simple neuron world examples that often drive discussion in the
literature, it is clear enough what counterfactual situation the author has in mind:
letting c be the ring of some particular neuron, the counterfactual situation in which
c does not occur is a situation in which that neuron doesnt re, but rather remains
idle. More generally, if event c consists in the j-ing of some particular x at some time
t, then we might naturally construe the counterfactual situation in which c does not
occur as simply a situation in which x does not j at time t. That does not yet solve
the problem, for it is not necessarily clear how we decide what x would be doing
instead. But here too, there might be a natural choice. Thus, if a neuron were not
ring at a certain time, it would be idle instead. (It would not be glowing pink, or
have turned into a dove, or have disappeared, etc.) Similarly, if Billy and Suzy had
not conversed between noon and 12:30, they would have done some other activity
normal for them instead (played, read books, sat quietly, etc.).
But to say that some such rough and ready rules guide our understanding of the
counterfactuals that appear in the Lewis analysis does not help that analysis, so much
as make more vivid one crucial shortcoming in its foundations (namely, that no clear
and precise rules have been laid down for evaluating the counterfactuals)unless,
of course, the analysis is amended to explicitly incorporate these rules (in, one would
hope, a much less rough and ready form). Note, nally, that although Lewis is aware
of this problem, his own remarks are not particularly helpful. Here is what he says, in
Causation as Inuence, chapter 3 (p. 90, italics added):
What is the closest way to actuality for C not to occur?It is for C to be replaced by a
very similar event, one which is almost but not quite C, one that is just barely over the border
that divides versions of C itself from its nearest alternatives. But if C is taken to be fairly
fragile, then if C had not occurred and almost-C had occurred instead, very likely the eects of
almost-C would have been much the same as the actual eects of C. So our causal counter-
factual will not mean what we thought it meant, and it may well not have the truth value we
thought it had. When asked to suppose counterfactually that C does not occur, we dont really
look for the very closest possible world where Cs conditions of occurrence are not quite sat-
ised. Rather, we imagine that C is completely and cleanly excised from history, leaving behind
no fragment or approximation of itself.
We leave it to the reader to try to gure out what it means to cleanly excise an
event. (Fill the spacetime region in which it occurs with vacuum?) That an answer
will likely prove elusive reinforces the suspicion that this issue needs to be taken more
seriously by advocates of counterfactual analyses.
A third and related challenge for the counterfactual analystand, to be fair, for
almost anyone who wishes to come up with a philosophical account of causation that
treats events as the primary causal relatais to provide some sort of account of what
events are. Without such an account it may be dicult to fend o various annoying
counterexamples. For instance:
Suppose that Suzy shuts the door, and in fact slams it. We may have two events
herea shutting and a slammingdistinguished because the rst could have hap-
pened without the second (if, for example, Suzy had shut the door softly). But if she
hadnt shut the door, she couldnt (and so wouldnt) have slammed itso it seems
that the analysis wrongly tells us that the shutting is a cause of the slamming.
Another case: Suzy expertly throws a rock at a glass bottle, shattering it. The
shattering consists at least in part of many smaller and more localized events: rst the
glass fractures, then one shard goes ying o this way, another that way, and so on.
If the shattering hadnt happened then none of its constituent events would have
happenedso it seems that the analysis wrongly tells us that the shattering is a cause
of them all.
A third case: far away, Billy throws another rock, shattering a dierent glass
bottle. Suppose there is a disjunctive event c that, necessarily, occurs if and only if
either Suzys throw or Billys throw occurs. If c hadnt occurred, neither bottle would
have shatteredso it seems that, according to the analysis, we have discovered a
fairly immediate common cause of these widely separated events.
A nal case: Suppose there is a type of extrinsically specied event that, neces-
sarily, has an instance occurring at time t if and only if Suzy sends an email message
exactly two days before t. Let c be such an instance, and let e be the reply (from
Billy, of course), occurring at, say, a time one day after Suzy sent the given email
(and so a day before c). Since e would not have happened if c hadnt, it seems that,
according to the analysis, we have backward causation very much on the cheap.
Each case deserves a dierent diagnosis. In the rst case, we should say that the
events fail to be distinctand so are not eligible to stand in a causal relationship
because of their logical relationship (some would prefer: because there is only one
event, dierently described). In the second case, we should say that the shattering is
not distinct from its constituents, because of their mereological relationship. In the
third case, we should say that the disjunctive event is not an event at all, hence not
apt to cause (or be caused). In the nal case, we should say that genuine events can-
not have such extrinsic essences. We should say all these things, and its up to a
philosophical theory of events to tell us why we are justied in doing so.
Still, it is hardly a peculiarity of the counterfactual analysis that it needs this sort
of supplementing. Suppose, following Mackie (1965), that we analyze a cause of an
event as a distinct event that is sucient, given the laws and relevant circumstances,
for that rst events occurrence. This analysis can be refuted in an equally trivial
fashion by noting the following consequences: The slamming causes the shutting;
the shatterings constituent events jointly cause it (and, perhaps, it causes them as
well); there is still a fairly immediate common cause of the two widely separated
shatterings (namely, the conjunctive event that, necessarily, occurs i both Suzys
and Billys throws occur); and the emailed two days ago event c still causes the
earlier reply e. So we do not think that these (admittedly challenging) issues involving
the nature and individuation of events pose any problem for counterfactual analyses
of causation in particular. (For discussion, see Kim 1971.) Contrast the earlier prob-
lem about evaluating the relevant counterfactuals: that problem does, ohand, seem
idiosyncratic to counterfactual analyses.
Let us turn now to the simple counterexamples, both old and new, that have
forced those pursuing a counterfactual analysis to add successive renements to
Lewiss original version. It will in fact be helpful to rewind slightly, and consider a
version even simpler than Lewiss, namely, the analysis that says that c is a cause of e
if and only if c and e both occur, but had c not occurred, e would not have occurred.
Figure 1.2 already showed why this analysis wont work; for if c had not red, e
would have red all the same, as a result of the backup process initiated by the
ring of a. Hence Lewiss decision to take causation to be the ancestral of counter-
factual dependence: Although the ring of e does not depend on the ring of c, we
can trace a chain of stepwise dependence via the ring of d.
But there are two problems with the simple x. The rst is that it automatically
commits the analysis to the view that causation is transitivea view that, however
intuitive, has recently come under suspicion, thanks to a range of purported coun-
terexamples (see section 6 below). These counterexamples can, perhaps, be resisted.
But a second and more serious problem is that there are simple ways to tweak the
example depicted in gure 1.2 that render Lewiss x inapplicable. For instance,
consider gure 1.3. Here, a and b re at the same time, both sending stimulatory
signals to e. However, the signal from b is a bit slower (a fact represented by drawing
its arrow so that it does not quite connect with e). Hence the signal from a, arriving
rst, is what causes e to re. Countless real-world examples duplicate this structure,
for example: Billy and Suzy both throw rocks at a window; each throw would be
enough, by itself, to break the window; but Suzys rock gets there rst, and hence it is
her throw, not Billys, that is a cause of the subsequent shattering.
Pick the ring of aor indeed, any event in the sequence constituting the passage
of the stimulatory signal from a to eand consider what would have happened if it
had not occurred. Answer: e would have red all the same, as a result of the signal
from b. There is no hope, then, of tracing a chain of stepwise dependence from the
ring of a to the ring of e.
The literature has dubbed this the problem of late preemption (as contrasted
with the problem of early preemption exhibited, for example, by gure 1.2). A
natural enough response to it is to focus attention on the timing of es ring, pointing
out that if a had not red, e would have red moments later, whereas if b had not
red, e would have red at exactly the same time. The literature has seen a number
of attempts to leverage this feature of the case into a solution to the problem of late
preemption. We wont try to review these here (but see the extended version of Lew-
iss Causation as Inuence, chapter 3, for a valuable and comprehensive discus-
sion); we do, however, want to suggest that the pursuit of this kind of solution may
b
e
a
Figure 1.3
be misguided. Suppose, for example, that the signal from a exerts a retarding force
on the signal from b, slowing it down slightly. Suppose that the exact strength of this
force is such that, had a not red, the signal from b would have arrived at e at exactly
the same time that the signal from a in fact arrives. Then not only does the ring of e
not depend on the ring of a, but nothing about the ring of enot its timing, nor
any other feature of its manner of occurrencedepends on the ring of a. (We can
further suppose that the retarding force is of just the right strength that, for exactly
the same reason, e exhibits a total lack of dependence on each of the events that
consist in the passage of the stimulatory signal from a to e.)
At any rate, it is safe to say that the problem of late preemption has proved quite
stubborn and has provoked vigorous discussion in the literature. Several of the con-
tributions in this volume give up-to-the-minute treatments of it.
There is another way to tweak the example of gure 1.2 so as to foil Lewiss orig-
inal x. It involves the controversial assumption that there can be action at a tem-
poral distance. Suppose, as in gure 1.4, that one eect of cs ring is to cut o the
stimulatory signal from a, as before; but suppose further that when cs ring causes e
to re, it does so not by way of any intermediate events, but rather by acting directly
on e (albeit at a spatial and temporal distance; we represent this direct action by a
thick, shaded arrow). Once again, the situation provides no materials for tracing a
chain of dependence from c to e. We leave it to the reader to consider whether this
sort of example is more stubbornly resistant to treatment than the kind of late pre-
emption depicted in gure 1.3.
We noted above that one of the features of the Lewis analysis that made it appear
particularly promising was that it could give a clean and uniform account not only of
ordinary causation under determinism, but of causation by omission, and of causa-
tion under indeterminism. But serious complaints have been raised with respect to
these latter two points. About causation by omissionand, more generally, other
kinds of causal fact that seem to involve omissions (such as prevention)two prob-
lems call for particular attention.
a
b
c
e
Figure 1.4
First, the typical counterfactual condition deemed necessary and sucient for
causation of event e by the omission of an event of type Cnamely, that no event of
type C occurs, that e occurs, and that had an event of type C occurred, e would not
have occurredhas struck many as wildly permissive. Suzy and Billy plan to meet
for lunch, but Billy doesnt show up; consequently, Suzy becomes sad. So far, no
problem: Billy does not show up (i.e., no event of the Billy-showing-up type occurs);
Suzy becomes sad; and, if Billy had shown up (i.e., if an event of the omitted type
had occurred), Suzy would not have become sad. But something else that doesnt
happen is Bill Gatess showing up to present Suzy with a $10 million check. And, if
he had, she would certainly not have become sad (she doesnt like Billy that much).
Are we to conclude that his failure to show up with the check likewise caused her to
become sad? Once you see the trick, the embarrassing examples quickly proliferate.
In general, for any event e there will typically be a vast multitude of ways that it
could, in principle (and compatibly with the laws), have been prevented. Are we to
say that for each such would-be preventer, its failure to occur was among es causes?
The counterfactual analyst seems committed to a yes answer here; but again, to
many, that answer seems absurd.
So posed, the challenge invites the response that it is really everyones problem
that is, that given modest assumptions, it follows that there is either a whole lot
of causation by omission (much more than we would ordinarily feel comfortable
admitting), or that there is none at all. For example, what principled basis is there for
distinguishing Billys failure to show up for lunch from Bills failure to show up with
a check? That Billy promised to, whereas Bill did not? But how could that sort of
normative consideration have any place in a proper account of causation? Beebees
contribution (chapter 11: Causing and Nothingness) presses this argument, and
both she and Lewis (chapter 3: Causation as Inuence, and especially chapter 10:
Void and Object) endorse this disjunctive either an awful lot or none at all
conclusion, albeit for dierent purposes. (See also McGrath 2002 for a sophisticated
treatment of this issue.)
For the second problem, consider how a counterfactual analyst might analyze the
notion of prevention. A natural approach is to say that event c prevents event e from
occurring (or perhaps better: prevents an event of type E from occurring) just in
case c occurs, e does not (or: No event of type E does), and if c had not occurred, e
(or: An event of type E) would have occurred. Thus, Billy throws a baseball at the
window, and Suzy prevents the window from breaking by catching it: That is, her
catch occurs, the window does not break (no event of the window-breaking type
occurs), but if she hadnt caught the baseball, the window would have broken.
So far, so good. But now tweak the example so that it becomes an instance of what
is commonly called preemptive prevention: Suzys friend Sally stands behind her,
ready and able to catch the ball if Suzy does not. In fact, because Suzy catches the
ball, Sally doesnt have to do anything. With that change, we have lost the depen-
dence of the windows failure to break on Suzys catch; but for many, the intuitive
verdict that Suzys catch prevents the window from breaking still holds. After all,
Sally just stood there, doing nothingand since something obviously prevented the
window from breaking, it had to be Suzy (or rather, her catch). On the other hand,
many nd that Sallys presence undermines the original verdictor if Sally is not
enough, then substitute in her place the presence of a high, thick, sturdy brick wall.
After all, how can Suzy get credit for preventing the window from breaking, when
it was never in any danger from Billys throw? The issues here are subtle, for it is
a dicult matter to tease out exactly what is guiding our intuitions in such cases.
(For discussion, see the contributions by Maudlin [chapter 18: Causation, Coun-
terfactuals, and the Third Factor] and Collins [chapter 4: Preemptive Prevention];
we will briey return to this kind of case in section 5, below.)
Still, if, with Sally present, the intuitive verdict that Suzys catch prevents the win-
dows breaking stands, then that is trouble for the counterfactual analysis. It is not at
all clear how any of the standard tools for handling ordinary cases of preemption can
be applied to get Suzys catchand only Suzys catch (not also Sallys presence, or
the presence of the wall, etc.)to connect up in the right way with the windows
failure to break.
A serious problem has also emerged for the standard counterfactual treatment of
causation under probabilistic laws. Recall the key idea: c is a cause of e if c and e
occur, but if c had not occurred, then the chance of es occurring would have been
much lower than in fact it was. (We then add various renements to extend that
sucient condition for causation into a full-blown analysis of causation.) As an
example, consider gure 1.5. Here, a res, sending a stimulatory signal to e, which
res. Let the laws state that e is certain to re if it receives such a signal, but has a
tiny chancesay, 0.01of ring spontaneously, even if it receives no signal. So far,
so good: our sucient condition correctly classies as ring as a cause of es ring,
in this case.
a
e
Figure 1.5
(Some would object, wanting to insist that it remains up in the air whether es
ring was caused by the signal from a, or rather was one of those rare instances
of a spontaneous ring. But no reductionist can endorse this objection, for what
dierenceeither in categorical facts, or in the lawscould distinguish the two sit-
uations? Note, in this regard, the close parallel between the example depicted in
gure 1.5 and Tooleys antireductionist thought experiment discussed above.)
But now consider a variant: We allow that there is some small chanceagain, say,
0.01that the signal from a will die out shortly before it reaches e. Suppose, as in
gure 1.6, that it does so. But e res all the same. Then the trouble is that even
though it is clear that as ring is not a cause of es ring, it still seems that the su-
cient condition is met: a res, e res, and if a had not red, then the chance of es
ring would have been much lower (0.01 instead of slightly greater than 0.99).
A possible x is to be more careful about the time at which we evaluate the chance
of es ring. In the depicted scenario, it is true that immediately after a res, the
chance that e will re when it does is close to 1. But, letting time t be a time imme-
diately before e res but after the signal from a has zzled out, it is also true that the
chance, at t, of es ring is a mere 0.01exactly what it would have been, if a had
not red. So the t-chance of es ring does not depend counterfactually on as ring.
It remains to be seen whether this x could be extended into a proper counterfactual
treatment of causation under probabilistic laws; in particular, it is not clear how the
original sucient condition should be modied, in order to be able correctly to clas-
sify some cases of causation as such, and not merely avoid misclassifying some cases
of noncausation. See the contributions by Kvart (chapter 15: Causation: Proba-
bilistic and Counterfactual Analyses), Ramachandran (chapter 16: A Counter-
factual Analysis of Indeterministic Causation), and Hitchcock (chapter 17: Do All
and Only Causes Raise the Probabilities of Eects?) for highly sophisticated treat-
ments of causation under indeterminism.
Turn now to more novel counterexamples. The rst of these has come to be
called trumping preemption. As an illustration, consider gure 1.7. Let us suppose,
here, that neurons can re with dierent intensities, and with dierent polarities
(positive and negative). Suppose further that a neuron that receives a single
stimulatory signal res in the polarity of that signal (although with an intensity in-
dependent of the intensity of the stimulating signal). Finally, when, as here, a neuron
a e
Figure 1.6
receives two stimulatory signals simultaneously, it res in the polarity of the more
intense signal: in this case, the signal from c (its greater intensity being represented by
a thicker arrow). No problem yetor at least, no serious problem: as long as we are
willing to assert that the positive ring of e that in fact occurs would not have
occurred if e had red negatively, then we can safely conclude that the ring of e
depends on the ring of c.
But now tweak the example in the obvious way, so that c and d re with the same
(positive) polarity; but as before, let c re more intensely. Many consider it just as
intuitive that in gure 1.8 c aloneand not dis a cause of e. But if c had not red,
e would have red in exactly the same manner; and the same is true if we focus not
on c, but rather on any event in the passage of the stimulatory signal from c to e.
(For detailed discussionwhich shows how wide is the range of counterfactual
analyses that this example threatenssee Schaer, Trumping Preemption, chapter
2 in this volume.)
For the next novel counterexample, consider double prevention. We will borrow
Halls example. In gure 1.9, neurons a, b, and c all re simultaneously; cs ring
prevents d from ring. Figure 1.10 depicts what would have happened, had c not
red. Evidently, es ring depends counterfactually on cs ring; for that reason,
almost every counterfactual analysis will count c a cause of e.
+
d
c
+
e
Figure 1.7
+
+
d
c
+
e
Figure 1.8
Some would consider that result a counterexample as it stands, since it strikes
them as intuitively wrong that c is a cause of e, herethe idea being that c is not
connected up to e in the right sort of way. Hall (in Two Concepts of Causation,
chapter 9 in this volume) argues in favor of this intuition by drawing what he takes
to be three damaging conclusions from the claim that c is a cause of e: rst, that the
counterfactual analysis will lose the ability to distinguish genuine from ersatz action
at a distance (for c and e are not connected by a spatiotemporally continuous causal
chain); second, that the analysis will be forced to give up the idea that, roughly, the
causal structure of a process is intrinsic to that process (for the fact that e depends on
c can easily be made to be extrinsic to the processes depicted); and third, that it will
be forced to deny that causation is transitive (for reasons we will take up in section 6,
below). Whether this last consequence is really so devastating is controversial, for the
recent literature has also seen a spate of counterexamples to transitivity itselfand not
merely to the conjunction of transitivity with one or another counterfactual analysis.
d
c
b
a
e
Figure 1.9
d
c
b
a e
Figure 1.10
Let us now turn our attention away from the prospects and problems for a suc-
cessful counterfactual analysis of causation, and consider a number of broader issues
in the philosophy of causation. As will be apparent, tight connections remain: Al-
most every topic we will discuss in the sections ahead has been illuminated in myriad
and interesting ways by work in the counterfactual tradition. We begin with a brief
look at methodology.
5 Methodology
Work on philosophy of causation is, not surprisingly, heavily driven by intuitions
about cases. Standard procedure often seems to be the following: A philosopher
proposes a new analysis of causation, showing how it delivers the intuitively correct
results about a wide range of cases. But then novel cases are proposed, and intuitions
about them exhibited that run counter to the given theoryat which point, either
renements are added to accommodate the recalcitrant data, or its back to the
drawing board.
Its worth taking time out to consider some methodological questions concerning
this procedure. That is what we will do in this section.
One obvious and extremely dicult methodological question is this: What sort of
project are philosophers of causation engaged in, that consulting intuitions about
cases would be an appropriate way to pursue it? We wont have much to say about
that question, not merely because it is so dicult, but also because it does not
have much to do with work on causation per se. The methodology in question is so
widespread in philosophy that the question should really be asked of the eld as a
whole. Still, it will be useful to distinguish two quite dierent aims that a philosopher
working on causation might have. On the one hand, she might, like Lewis, take
herself to be providing a good old-fashioned conceptual analysis of causationa
detailed explanation, that is, of how our ordinary concept works (see especially sec-
tion 1.1 of Lewiss Causation as Inuence, chapter 3 of this volume). (Note that
in doing so, she need not take on the extra commitment to a reductive account
reductionism is an optional, if natural and attractive, extra.) On the other hand, she
might view her account as at least partially stipulativethat is, as providing a
cleaned up, sanitized version of some causal concept that, though it may not track
our ordinary notion of causation precisely, nevertheless can plausibly be argued to
serve some useful theoretical purpose. (Or one might fall somewhere in between:
certainly, the distinction admits of plenty of gradations.)
Obviously, someone who pursues this latter aim ought to say at some point what
such purposes might be. But we think that she is under no obligation to make this
clear at the outset. On the contrary, it strikes us as a perfectly appropriate strategy
for a philosopher working on causation to try to come up with a clean, elegant, the-
oretically attractive account of causation (or of some causal concept), in the reason-
able expectation that such an account will serve some, possibly as-yet undisclosed,
philosophical or perhaps even scientic purpose. And that expectation can be rea-
sonable even if it is clearbecause of recalcitrant intuitions about certain cases
that the account does not earn its keep by providing an explanation of how our
ordinary concept works. This tolerant methodological perspective gains further sup-
port from the observation that causal concepts are used all over the place, both in
philosophy and in the sciences; indeed, just sticking to philosophy, it often seems that
for any philosophically interesting X, there is at least one causal theory of X on
the market. It would be hasty to assume that the causal concept or concepts at work
in any such theory is just our plain old ordinary one. In short, then, there is good
reason to think that there is plenty of work available for those philosophers of cau-
sation who take themselves to be in the business of, as it were, conceptual synthe-
sis, rather than old-fashioned conceptual analysis.
It is clear enoughat least, for present purposeswhy someone interested in
providing a conceptual analysis of our ordinary notion of causation should attend
carefully to intuitions about cases. What we wish to emphasize is that even someone
interested in synthesizing a new and potentially useful causal concept needs to
heed these intuitions, else she risks cutting her project free of any rm mooring.
More specically, a reasonable and cautious approach for her to take is to treat
intuitions about cases as providing a guide to where interesting causal concepts might
be found. Thus, although the account can selectively diverge from these intuitions,
provided there are principled reasons for doing so, it should not diverge from them
wholesale.
It is particularly clear that one should pay attention to intuitions about cases if the
stipulative element in ones analysis consists only in resolving various ambiguities or
indeterminacies in our ordinary notion of cause, or in some other way rening or
precisifying that notion. For example, one might think that our ordinary notion is
context sensitive in its application in various ways; and one might be interested in
looking for a closely related notion from which such context-sensitivity has been
expunged. Or one might suspect that our ordinary notion of causation involves an
unsupportable element of antireductionismas witness the intuitive pull (such as it
is) of thought experiments like Tooleys; one might therefore try to produce a sani-
tized, reductionist-friendly surrogate for our ordinary notion. More provocatively,
one might hold with Hall (Two Concepts of Causation, chapter 9) that there is
no way to provide a univocal analysis of our ordinary notion of causation, and that
therefore the best thing for a philosopher to do is to break it up into two or more
distinct conceptsdistinct, at least, in the sense that they deserve radically dierent
analyses. The list goes on: Indeed, even a cursory survey of work in philosophy
(as well as in the conceptual foundations of other disciplines) will reveal numerous
precedents for the view that there can be legitimate philosophical studies that do not
count as conceptual analysis of some ordinary concept, but that still require close
attention to intuitions involving the application of such a concept.
It is not always clear in the literature what a given authors aim is: conceptual
analysis, conceptual synthesis, or perhaps something else entirely. But it is important
to be aware of the options, for otherwise it can be quite dicult to assess the cogency
of certain typical responses philosophers make when confronted with recalcitrant
intuitions. Above, we observed that one typical responsethe most obvious oneis
to go back and revise the theory. But that is hardly the only response. In addition,
philosophers will often try to argue that the recalcitrant intuitions need not be
respected. There are a number of strategies for doing so, and philosophers are not
always careful to make explicit which strategy or combination of strategies they are
pursuing. So we think it useful to list, and briey comment on, some of the most
prominent ones. As we will see, in several cases the eectiveness of the strategy
depends crucially on whether or not the theory being defended is oered as a con-
ceptual analysis of our ordinary concept of causation.
Suppose, then, that we have some philosophical theory of causation T that issues
some verdict about some case, and that ordinary intuition about the case runs
counter to the verdict. Then there are at least eight distinct ways that we might try to
fend o the recalcitrant intuition.
First way: The intuitions in question are really not so rmso the case is an
example of spoils to the victor.
Lewis has provided a particularly nice statement of the idea behind this strategy:
When common sense delivers a rm and uncontroversial answer about a not-too-far-fetched
case, theory had better agree. If an analysis of causation does not deliver the common-sense
answer, that is bad trouble. But when common sense falls into indecision or controversy, or
when it is reasonable to suspect that far-fetched cases are being judged by false analogy to
commonplace ones, then theory may safely say what it likes. Such cases can be left as spoils to
the victor, in D. M. Armstrongs phrase. We can reasonably accept as true whatever answer
comes from the analysis that does best on the clearer cases. (1986b, p. 194)
As an example, consider a case of perfectly symmetrical overdetermination: Suzy
and Billy both throw rocks at a window; the rocks strike at the same time, with
exactly the same force; the window shatters. Furthermore, each rock strikes with
sucient force to shatter the window all by itself. There is some intuition here that
both Suzys and Billys throws are causes of the shattering. But this intuition is far
from rm. Hence, when a counterfactual analysis saysas it likely willthat neither
Suzys nor Billys throw was a cause of the shattering (because it depended on nei-
ther, etc.), then that verdict can perhaps be considered one of the spoils to which
the analysis is entitled (provided, of course, that it emerges victorious in its competi-
tion with its rivals).
But Lewiss claim makes little sense if one is pursuing a genuine conceptual analy-
sis of our ordinary notion of cause. A successful such analysis will show how our
concept works; so if our concept goes indeterminate in a particular case, then the
analysis had better show that and why it does so. (Obviously, such an expectation
is not appropriate if ones aim is to rene, precisify, or in some other way modify
our ordinary conception of causation.) For comparison, suppose that we oered an
analysis of bald according to which anyone with fewer than 500 hairs on his or her
head counted as bald, and anyone with 500 or more hairs counted as not bald. Never
mind the many other reasons why the analysis is inadequate (it matters how the hair
is distributed, etc.): The one to focus on is that it leaves it completely opaque why
there should ever be a borderline case of bald. That would be reason enough to
count it a failure, if we had oered it as an account of our ordinary notion of
bald.10
At any rate, the example of symmetrical overdetermination also illustrates an im-
portant if unrelated point: Whereas we might happily reject whatever intuition there
is to the eect that each of Billys and Suzys throws is a cause of the shattering, we
should not be happy with the conclusion that the shattering lacks causes altogether
and on the face of it, counterfactual analyses threaten to yield this conclusion.
(Some, e.g., Lewis 1986b, try to defend the view that the mereological sum of the
two throws causes the shattering.) Our point is not to press this problem on those
analyses, but simply to observe that, to handle a dicult case, it may not always be
enough simply to argue away the recalcitrant intuitions about it.
Second way: The case is misleading, because it is easily confused with other,
unproblematic cases.
As an example of a case where this strategy might apply, consider the instance of
action at a temporal distance depicted in gure 1.4. One might try to argue that in
forming intuitive judgments about this casein particular, the intuitive judgment
that c is a cause of eit is extremely hard not to let ones intuitions be guided by the
thought that there is some faint, indetectable signal transmitted from c to e that is
responsible for establishing the causal connection between the rings of the two
neurons.
Or again, consider the example of trumping depicted in gure 1.8. One might
argue that, in judging any such case, it is very hard not to presuppose that the
trumping pattern is present in virtue of some unspecied and hidden physical mech-
anisms. (And when that is the case, examples of trumping become much less prob-
lematic; for further discussion see Schaer, Trumping Preemption, chapter 2.)
Third way: The case is misleading, because it is misleadingly presented.
As an example, let us return to Tooleys antireductionist argument. Observe that
in setting up that argument, he invites us to consider a possible world in which it is
a law that when an object has property P, its possession of this property causes it
to acquire either property Q or property Rbut never both. Note that the word
causes is crucial to the specication of the content of the laws, for we are also
invited to believe that in this possible world an object can, consistently with the laws,
acquire both Q and R after coming to possess P. Thus we cannot consistently replace
causes by is followed by. But to describe the relevant laws in this way prejudices
matters enormously. After all, a good reductionist about causation will hold that the
content of the fundamental laws can, in principle, be specied in noncausal terms
indeed, can perhaps be identied simply with a set of nomologically possible worlds.
(And it is well to remember that insofar as we wish to look to the physics of our
day for paradigm examples of fundamental laws, this reductionist position is by far
the more sensible: Schro dingers equation, for example, makes no mention of causal
relations between physical states.) Tooleys thought experiment presupposes a very
dierent conception of laws; it seems to us that the reductionist is well within her
rights to reject this conception. She should insist that Tooley has either mischarac-
terized the content of the laws, or has not succeeded in describing a genuine possible
world.
Fourth way: The example is misleading, because it naturally draws our attention to
other, related concepts; we therefore mistake the conditions appropriate for apply-
ing these other concepts with the conditions appropriate for applying the concept of
causation.
Beebees contribution (Causing and Nothingness, chapter 11) provides an ex-
cellent case study of this strategy at work. She argues that there is no such thing as
causation by omission, but that we can perfectly well explain the occurrence of some
ordinary event by citing the failure of some other kind of event to occur, where an
occurrence of that kind of event would have prevented the explanandum event. Bee-
bee goes on to argue that the reason that we often mistakenly think that there is
causation by omission is that it is easy for us to confuse causation with explanation.
As another exampleone where this strategy should have been applied, but
wasntconsider the following passage from Lewis (1986d, p. 250):
It is one thing to postpone an event, another to cancel it. A cause without which it would have
occurred later, or sooner, is not a cause without which it would not have occurred at all. Who
would dare be a doctor, if the hypothesis under consideration [that an events time is essential
to it] were right? You might manage to keep your patient alive until 4:12, when otherwise he
would have died at 4:08. You would then have caused his death. For his death was, in fact, his
death at 4:12. If that time is essential, his death is an event that would not have occurred had
he died at 4:08, as he would have done without your action. That will not do.
The case is a bad one, because it is very easy to confuse the question of whether
the doctors action is a cause of the patients death with the question of whether, by
so acting, the doctor killed the patient. The answer to the latter question is quite
clearly no; confuse it with the former, and you will think that the answer to that
question is no as well. But once the questions are rmly separated, it is no longer
so clear that it is unacceptable to count the doctors action as among the causes of
the patients death. (Certainly, the doctors action was one of the things that helped
cause it to be the case that the patient died at 4:12.) At any rate, what is clear, given
the manifest possibility of confusion, is that it is unwise to use this case to generate
substantive conclusions about the metaphysics of causation or of events.
Fifth way: The recalcitrant intuitions conict with sacrosanct general principles
about causation, and so ought to be dismissed.
One such principle commonly appealed to is, roughly, that like cases be treated
alike. Thus, in Causation as Inuence, Lewis has us imagine a case where it is
intuitively correct to say that someones birth is a cause of his death; by appeal to the
foregoing principle, he uses this case to argue against the oft-cited complaint that
counterfactual analyses have the unacceptably unintuitive consequence that births
cause deaths. The example also illustrates how more than one general principle might
be brought into play, for one can certainly appeal to transitivity to connect up an
agents birth with his death, as well.
Or again, one might respond to alleged counterexamples to transitivity simply by
refusing to give up that principle, perhaps on the basis that its intuitive appeal out-
weighs the intuitions driving the counterexample, or on the basis that it plays too
essential a role in ones analysis. (For example, one might think that only by an
appeal to transitivity can an analysis handle certain kinds of preemption cases.)
Someone who pursues this strategyinsisting that an intuition, however rm, be
rejected for the sake of some sacrosanct general principlewill help her case im-
measurably if she can add some explanation for why intuition has been so misled.
Thus, this strategy is most eectively combined with one or more of the other strat-
egies being discussed here. (As a nice example, consider Beebees case, in Causing
and Nothingness [chapter 11], that there is no causation by omission: She supports
this counterintuitive position in part by appeal to the general principles that causa-
tion must relate events, and that omissions cannot be thought of as a species of event,
and in part by providing the explanation already discussed for why our intuitions
about alleged cases of causation by omission should not be considered trustworthy.)
But observe that it may be far less urgent to provide such an explanation if one is
not engaged in conceptual analysis: then one may be able to argue that the given
divergence from intuition is the price one must pay for synthesizing a concept that is
clean, theoretically attractive, and so on.
Sixth way: The unintuitive verdict issued by the theory is not falseits just odd to
say, for pragmatic reasons.
Here, for example, is Lewis, responding to a worry about his account of causation
by omission:
One reason for an aversion to causation by absences is that if there is any of it at all, there is a
lot of itfar more of it than we would normally want to mention. At this very moment, we
are being kept alive by an absence of nerve gas in the air we are breathing. The foe of causa-
tion by absences owes us an explanation of why we sometimes do say that an absence caused
something. The friend of causation by absences owes us an explanation of why we sometimes
refuse to say that an absence caused something, even when we have just the right pattern of
dependence. I think the friend is much better able to pay his debt than the foe is to pay his.
There are ever so many reasons why it might be inappropriate to say something true. It might
be irrelevant to the conversation, it might convey a false hint, it might be known already to all
concerned. . . . (Lewis, Causation as Inuence, chapter 3 in this volume)
We think that care must be taken in executing this gambit. In general, solid reasons
should be given why certain judgments should be explained away by appeal to prag-
matics while others should be incorporated into the analysis. (Observe, in this regard,
how silly it would be to oer as a theory of causation the following: Every event
causes every later event; it is just that in many cases, it is odd to say so, for pragmatic
reasons.) Whats more, it is one thing to wave ones hands toward some pragmatic
accountas Lewis does hereand quite another to provide the specics (including:
an account of exactly which of the many principles governing the pragmatics of
communication should be appealed to). We think that unless they are provided
or unless it is fairly clear how they can be providedthe reader deserves to be sus-
picious of such appeals.
Still, there are other problem cases to which the pragmatic treatment seems ex-
actly what is needed. Suzy is in a room with a match. She strikes the match, lighting
it. Intuitively, it is perfectly clear that among the causes of the lighting is the striking.
It is much less intuitivebut a consequence of most every analysis of causation, all
the samethat among the causes of the lighting is the presence of oxygen next to the
match. But we think it a serious mistake to count this a counterexampleat least,
without rst investigating the possibility that for straightforward pragmatic reasons,
reports of causes are typically expected to be reports of salient causes. Compare the
following scenario: Until recently, all air had been pumped out of the room; only
shortly before Suzy entered it with her match was air pumped back in. Now, when
asked about the causes of the lighting, we nd it quite appropriate to cite the pres-
ence of oxygen. Are we to expect this shift in intuitive judgments about what is ap-
propriate to say to reect a shift in the truth of the claim that the presence of oxygen
is a cause of the lighting? That strikes us as foolish. Much better to suppose that in
altering this scenario, we have altered which features of it are especially salientand
to remember that considerations of salience will weigh heavily on our judgments
about what is appropriate to report as a cause.
Seventh way: Intuitions about the case are too easily bueted about to be taken
seriously.
As an example of how such bueting can work, consider again the case of pre-
emptive prevention in which Billy throws a rock at a window, Suzy blocks it, but
Sally was waiting behind Suzy to block the rock, if necessary. Emphasize Sallys
idleness, and you can easily secure the intuitive judgment that Suzy prevented the
window from breaking. But as Maudlin notes in his contribution (Causation,
Counterfactuals, and the Third Factor, chapter 18), there are other ways to describe
such cases. In particular, think of the window and Sally as constituting a single sys-
tem, remembering that Sally is perfectly able and willing to block the rock. We might
then describe this system as a protected windowwhich, because of the state it is
in, is not under any threat of damage from the rock. Conceived that way, the case
prompts a very dierent intuitive judgment: Suzy does not prevent the window from
breaking, because, after all, the window was never under any threat of being broken.
It seems to us a fascinating phenomenonworth much more attention than it has
been given in the literaturethat some cases evoke intuitions that are easily subject
to such bueting, whereas others do not. For example, when Suzy and Billy both
throw rocks at the window, and Suzys gets there rst, no amount of redescription of
the case will reverse the intuitive judgment that only Suzys throw is a cause of the
shattering. If one is engaged in conceptual analysis, then, this strategy surely back-
res, for one thing we should expect from any successful such analysis is an expla-
nation of what distinguishes cases that are subject to bueting from cases that are
not. But if one is engaged in something more like conceptual synthesis, then this
strategy might well be appropriate.
Eighth way: The case is too outre for intuitions about it to be of much concern.
As examples, one might want to dismiss from consideration any intuitions about
cases that involve action at a temporal distance, or backward causation, or indeter-
minism at the level of the fundamental laws. We have already seen several reasons
why one might set aside such cases: intuitions about them might not be very rm; or
there might be reason to suspect that our intuitive judgments cannot easily distin-
guish such cases from other, less problematic cases; or the judgments about these
outre cases might conict with sacrosanct general principles about causation; and so
on. But we have in mind something dierent here, which is that these cases be dis-
missed simply because they are outre in some readily recognizable respect.
To see what this amounts to, observe that the conceptual analyst has no business
whatsoever pursuing this strategy. If, for example, we construct a case involving
action at a temporal distanceand the case passes all other reasonable method-
ological teststhen any decent account of how our ordinary concept of causation
works must respect the data concerning its application to the given case. On the
other hand, someone who constructs an account of causation with the aim of pre-
cisifying, reforming, or in some other way altering or replacing our ordinary concept
can perfectly well insist, at the outset, that she only intends her account to cover
causation under determinism, or under the assumption that there is no backward
causation, or action at a temporal distance, and so on. Imagine that her account
succeeds brilliantly within its intended domain, and then consider how foolish it
would be to reject it out of hand because it had not been extended to cover a broader
domain of possible laws. We do not mean to suggest that there would be no interest
in trying to extend the account; of course there would. But even if there were in
principle obstacles to doing so, this would not rob the account of its philosophical
interest or utility.
There is, in addition, a second point to makeone that, though modest, is often
overlooked, with unfortunate results. When a philosopher sets out to give an account
of causation, and it is clear at the outset that there will be great diculty in extending
the account to cover a certain range of cases (causation under indeterminism, back-
ward causation, etc.), it is all too tempting to simply x on this limitation as a reason
to dismiss the account out of hand. We consider that attitude mistaken, and urge, in
opposition to it, a modest methodological pluralism: the topic of causation is dicult
enough that it is worth pursuing avenues of investigation even when, at the outset, it
seems clear that they will not give us everything we want (i.e., will not give us an
account that covers a range of cases that fall outside certain well-dened limitations).
A philosopher can therefore reasonably dismiss outre cases from consideration sim-
ply because she meansat least for the momentto limit her ambitions. She is
guilty of no philosophical error in doing so. (As an example in this volume, Kvarts
contribution, Causation: Probabilistic and Counterfactual Analysis [chapter 15],
lays out an account of causation under indeterminism that he quite forthrightly
observes would be dicult to extend to the deterministic case; it strikes us as no less
interesting and important for all that.)
Let us move away now from methodological concerns and back to properly
metaphysical concerns, focusing in the next section on two questions that confront
any account of causation, counterfactual or otherwise: is causation transitive? And
what exactly are the causal relata?
6 Transitivity
We normally think of the causal relation as a transitive relation: if c causes d, and d
causes e, then c causes e. The simplest analyses of causation based on suciency
under laws generate a transitive causal relation immediately: When c, together with
the right laws, is sucient for d, and d, together with the right laws, is sucient for e,
then c, together with the right laws, is sucient for e. Analyses of causation based on
counterfactual dependence typically need to guarantee transitivity by taking causa-
tion to be the ancestral of the dependence relation, as dependence relations (e.g.,
straightforward counterfactual dependence, or the kind of counterfactual covariation
that gures in Lewiss new inuence account) are typically not transitive.
Some (e.g., Hall: See chapter 7, Causation and the Price of Transitivity) suggest
that the claim that causation is transitive should be treated as a sort of bedrock
datum; but that strikes us (Hall now included) as hasty. Certainly the claim has a
great deal of intuitive appeal; but rather than simply endorse it unquestioningly, we
ought to investigate the source of that appeal. Although we have nothing decisive to
oer on this score, we think it suggestive that the way in which causal claims are
often justied seems to presuppose transitivity. Someone claims that c causes e. Why?
Because c causes d, which in turn causes e.
This kind of reasoning is commonplace. But if we could not assume that the causal
relation were always transitive, then it is not clear why we would ever be entitled to
it. It would seem, rather, that to determine whether c causes e, for any c and e that
are not directly linked (i.e., for any c and e that are causally related, if at all, only by
way of intermediate events) we would have to determine of every link in the causal
chain whether it prevented or allowed for transitivitywhether it was the sort
of cde link that licensed the inference from c causes d and d causes e to
c causes e. This would be a serious problem indeed, since (a) it is not clear that
we could always distinguish between the types of causal connections that were
transitive and the types that were not, and (b) we would, on the face of it, need to
have a detailed account of every link in the causal chain. Providing such an account
would apparently require us to determine the minimal units of the causal chain in
questionwhich, among other things, would involve a fuller specication of the
causal relata (usually taken to be events) than theorists have been able to develop.
Hence, it appears that if the causal relation were not transitive, many of our
everyday causal claims would be unjustied. (For related discussion, see Lewis,
chapter 3, Causation as Inuence, sections 2.2 and 2.3.)
Unfortunately, there are some compelling examples that suggest that causation
is not transitive. The examples often exhibit the following abstract structure: c does
something that threatens to prevent e. However, c also causes d, which in turn helps
bring about e in spite of the threat.11 For example, a train rushes toward a fork in
the tracks. If a switch is ipped, the train will take the left track, and if the switch is
left in its original position, the train will take the right track. Further on, the left and
the right tracks merge, and just after they meet, a damsel in distress is tied to the
tracks. If Jill ips the switch, and the train runs over the damsel, should we say that
Jills ip is a cause of the death of the damsel? Obviously not, says intuition. But if
the ipping of the switch is a cause of the train taking the left track, and the train
running on the left track is a cause of the trains merging back on the main track, and
the trains being on the main track is a cause of the death of the damseland,
nally, if causation is transitivethen Jills ip is a cause of the death. Notice the
abstract structure: the ip threatens to prevent the deathby diverting the train from
a track that would have led it to the damselbut simultaneously does something
that helps undo the threatby diverting it onto another track leading to the
damsel.
A similar kind of example12 discussed by several of the papers in this collection
involves a bomb that is placed in front of someones door. The bomb is set to ex-
plode in ve minutes. But just before it explodes, a friend comes along and defuses it.
As a result, the intended victim continues to live. The presence of the live bomb is a
cause of its being defused, and the bombs being defused is a cause of the intended
victims continued existence. If transitivity holds here, then the presence of the live
bomb is a cause of the intended victims continued existence. Note again the abstract
structure: The presence of the live bomb threatens to cut short the victims life, but
simultaneously undoes this threat by attracting the attention of the friend.
The manifest similarity in structure between these two cases should not make us
overlook one dierencea dierence that may be crucial. In the case of Jills ip,
there is, straightforwardly, a spatiotemporally continuous causal chain running from
the ip to the damsels death; it is causal in the sense that each constituent event is,
unproblematically, a cause of the immediately subsequent events. The question is
only whether being linked by such a sequence of causes is enough for the endpoints
to count as causally related. By contrast, in the case of the bomb there is no such
spatiotemporally continuous causal chain linking its presence to the continued life of
the intended victim. Rather, the case is much more like Halls inert neuron net-
work (gure 1.11).
Here a res, stimulating e to re. Meanwhile, c res, stimulating both b and d to
re. The ring of b prevents the signal from d from stimulating f to re, thereby
safeguarding the ring of efor if b had not red, e would not have red. If such
dependence by double prevention suces for causation, then b is a cause of e.
And, obviously, c is a cause of b. So if, in addition, causation is transitive, then c is a
cause of e: the neuron-world analogue of the bombs presence causing the intended
victim to continue living. As in the bomb example, no spatiotemporally continuous
causal chain connects c to e. (Unless, perhaps, it is a chain partially constituted by
omissions. But even if so, that fact does not erase the distinction between this kind of
case and the switching cases of which Jills ip is an instance. And see Hall 2002a
for doubts that one can always interpolate an appropriate sequence of omissions.)
This distinction might matter, because there is some reasonalbeit far from
decisiveto hold that dependence by double-prevention does not suce for causa-
tion. If so, then a whole class of alleged counterexamples to transitivitynamely,
those that exhibit the structure of gure 1.11can be dismissed. Pursuing a divide
and conquer strategy in defense of transitivity, one could then try to nd grounds
for dismissing the remaining counterexamples. Halls Causation and the Price of
Transitivity (chapter 7) takes this approach.
But there are other strategies one can employ to address the problems. David
Lewis, in his Causation As Inuence, chapter 3, denies that these cases are coun-
terexamples to transitivity. Lewis argues that as long as the requisite dependence
f
c
d
b
a e
Figure 1.11
relations hold between links in the chain, transitivity holds, prima facie intuitions to
the contrary. Lewis traces our mistaken rejection of transitivity to misgivings we
might have about accepting preventers as causes, accepting events that initiate devi-
ant or unusual paths to their eects as causes, or to a residual inclination to think
that for c to cause e, e must depend counterfactually on c.
Stephen Yablo, taking the opposite tack, takes the cases that raise problems for
transitivity to be so important as to justify acceptance of a strikingly novel counter-
factual account of causation. In his Advertisement for a Sketch of an Outline of a
Prototheory of Causation, chapter 5, Yablo defends an analysis of causation that
tells us the ip, the bomb or the neuron that initiates the inert neuron network cant
count as causes merely because they bring about threats to the eect but simulta-
neously cancel those threats. There are deeper requirements (such as the eects
depending on the putative cause under conditions that are suitably natural) that an
event must meet in order to be a cause.
Cei Maslen, in The Context-Dependence of Causation, chapter 14, argues that
causation is a three-place relation between a cause, an eect, and a contrast event:
when c causes e, c is a cause of e relative to a contrast event c
. Maslen argues that in

cases where contexts (which x the contrast events) are incompatible, causation is not
transitive. She considers several cases that seem to cause problems for transitivity
and resolves them under the contrast account.
There is a very dierent kind of example that needs discussing. Here is an instance
(thanks to Judith Thomson for providing it): Suzy lights a rework, which shoots up
in the air and explodes in a brilliant shade of red. The light triggers a nearby photo-
detector that is sensitive to that particular shade of red. Meanwhile, young Zeno, a
boy in the crowd, covers his ears in response to the loud explosion. Plausibly, Suzys
lighting the rework counts as a cause both of Zenos action and of the triggering of
the photodetector. But now let us complicate matters: Earlier in the day, Billy added
a chemical to the rework that would guarantee that it would explode in the given
shade of red; if he hadnt, the explosion would have been yellow. What did Billys
action help cause? Well, the triggering of the photodetector, certainly. But, intui-
tively, not Zenos action: Billy had nothing to do with that. The problem is this: How
are we to secure the rst result without undermining the second?
Billys action does not help bring about the triggering in any mysterious, action-
at-a-temporal-distance manner. Rather, there is an ordinary causal chain of events
leading from one to the otherby way of the explosion. It therefore seems that, to
count as a cause of the triggering, Billys action must also count as a cause of the
explosion, since it was only by way of the explosion that he helped bring about the
triggering. But the explosion also caused Zeno to put his hands over his ears. So why
must we not therefore conclude that by adding the chemical, Billy likewise was a
cause of Zenos action?
Looked at a certain way, the case can seem to be another counterexample to
transitivity: Billys action was a cause of the explosionit had to be, else we could
not get it to come out a cause of the triggering of the photodetector. The explosion,
in turn, was a cause of Zenos action. But Billys action was not a cause of Zenos
action. So perhaps we have, here, yet another reason to give up transitivity.
If so, it is an entirely dierent kind of reason from that which applied, say, to
the inert neuron network: Billys action does not both initiate a threat to Zenos
action and simultaneously do something to cancel that threat. Moreover, we cannot
in good philosophical conscience stop with a denial of transitivity, for a thorny puz-
zle remains: why is it that Suzys action counts as a cause of Zenos, by way of the
explosion, but Billys does not?
Some will be impatient for what they consider the obvious response: There is not
just one explosion, but at least twodistinguished by the fact that it is somehow
constitutive or essential to one of them, but not the other, that it be an explosion in
the given shade of red. Call these explosions e and e, the latter being the one
that is constitutively or essentially red in the given shade. Then we have what seems a
transitivity-preserving solution: Billys action causes e but not e; Suzys action
causes both e and e; e causes the triggering but not Zenos action.
An attractive story, perhaps, although many will nd the ontological ination of
events unattractive. We should not accept it without taking several steps back, and
considering with more care than we have so far the question of what, exactly, the
causal relation relates. That is what L. A. Paul does in Aspect Causation, chapter
8. She discusses a case13 that appears to create problems for transitivity, and she
argues that this appearance is illusorycloser inspection shows that in such cases,
transitivity fails to apply. Causation may well be transitive; it is, rather, an underly-
ing tension between extant theories of events coupled with reductive analyses of
causation that gives rise to these seeming counterexamples. Here is the example that
Paul considers: Suzy breaks her right wrist, which causes her to write her philosophy
paper with her left hand. The paper is then sent o to a journal, which accepts it.
If the breaking causes the event of the left-handed paper-writing, and the paper
writing causes the acceptance, and if causation is transitive, it seemscontrary to
intuitionthat the breaking caused the acceptance.
Paul shows that none of the major contenders for a theory of events coupled with
a theory of causation succeeds against such examples, and she argues that this makes
trouble for theories of event causation. She exploits this trouble in order to argue in
favor of taking property instances (what she calls aspects) rather than events to be
the causal relata. (We discuss related issues involving the nature of the causal relata
below.) Observe how this maneuver will handle the forgoing case of the loud, color-
ful explosion: Billys action causes the explosion to have a certain aspectnamely, to
be an explosion in that particular shade of red. This aspect in turn is responsible for
the triggeringbut not for Zenos covering his ears. Transitivity is preserved. Still, as
Hall points out in his Causation and the Price of Transitivity, chapter 7, making
room for aspects is not enough to resolve what he calls the hard cases for tran-
sitivity, so even if a dierent account of the causal relata helps, serious problems re-
main. We will nevertheless set them aside, and focus squarely on the question Pauls
paper brings to the fore: what does causation relate?
7 The Causal Relata
Most contemporary analyses of causation hold that causation is a relation between
events, where these are taken to be particulars, entities that occupy spatiotemporal
regions. But unless we are prepared to take events as unanalyzable primitives, an
acceptable (reductive) analysis of the causal relation as a relation between events
requires an acceptable theory of eventhood. The theory of events thus counts as a
subtheory of a complete theory of causation.
By wide agreement, such a theory will broaden the category of event to include
items never ordinarily so-called. When Suzy throws a rock at the window, breaking
it, we naturally tend to think that there is just one sequence of eventsthe one ini-
tiated by Suzys throwconverging on the eect. But it is far better, at least for the
purposes of systematic metaphysics, to see this eect as standing at the intersection of
two sequences of events: There is the interesting sequence just mentioned, and then
there is the quite boring sequence consisting in the continued presence of the window,
up to the moment it shatters. More generally, a proper theory of events almost
certainly must count as such things that we would ordinarily classify as states, or
standing or background conditions.
Next, there is widespread agreement that a theory of events must have the re-
sources to tell us when two events, though nonidentical, fail to be wholly distinct
for example, because one is part of the other, or they have a common part, or they
stand in some sort of intimate logical relationship. As we saw in section 4, a theory
unable to draw such distinctions will almost certainly leave the account of causation
that makes use of it easy prey for various counterexamples, in the form of not wholly
distinct events that stand in relations of counterfactual dependence.
But within these broad parameters, there is no consensus as to which theory of
eventhood is the correct one. This lack of consensus traces in part to conicting
intuitionsin particular, about when we have just one event, and when (and why)
we have more than one. You walk to class. As you walk, you talk to your companion.
Is your walking one event and your talking a numerically distinct event? Or are they
the same event? There are considerations on both sides. On the one hand, the walk-
ing and the talking occurred at the same time and were performed by the same per-
son, so common sense might dictate that they are the same event. On the other hand,
the walking and the talking seem to stand in dierent causal relations to subsequent
events: your walking made your legs tired, while your talking gave you a sore throat.
In general, we can characterize the individuation conditions put forward in dif-
ferent theories of events along a continuum from coarse grained to ne grained.
Roughly, the more ne grained a theory takes event identity to be, the more events
exist according to the theory. So a theory that held that the walking and talking were
one and the same would be more coarse grained than a theory that distinguished
them; this theory in turn would, if it recognized but one talking, be more coarse
grained than a theory that distinguished your talking loudly from your talking
simpliciter.
Its one thing to insist on such distinctions; its another to explain them. Among
those who want to draw them, probably the most popular explanation appeals to
dierences in the possible conditions of occurrenceor essencesof events.14
You could have walked without talking, and vice versa; so your walking and talking
are distinct events. More ambitiously, perhaps your talking loudly is a dierent event
from your talking, because the latter, but not the former, would have occurred if you
had been talking softly. If so, then the two events are alike in all categorical respects:
Not only do they occupy the same region of spacetime, but they are both loud (one
essentially, one accidentally so), both energetic, and so on. Notice that once the door
is open to events that dier only along such modal dimensions, it will be dicult
to close it; that is, the number of events that one must recognize as being perfectly
categorically coincident will probably be enormous. There is the talking, the talking
that is essentially not quiet, the talking that is essentially loud, the talking that is
essentially exactly that loud, and so onand we have been drawing only crude dis-
tinctions, about only one of the many aspects of these events. This ontological pro-
liferation leads to trouble, as we will shortly see.
Theories of events that do not draw distinctions in this waybecause they draw
them either in some other way,15 or not at allstill face an important question
concerning essences. Consider your walking: what are the possible conditions under
which itand not just some event very much like itcould have occurred? To the
extent that these conditions are very restrictive, your walking is modally fragile
unable, as it were, to survive in all but a few alternate possible worlds.
It is important to recognize that the question of how fragile events are cross-cuts
the question of how many there are. For example, one could hold a highly coarse-
grained theorysay, one according to which there is at most one event per region
of spacetimewithout committing oneself as to how modally fragile these events
are. Or one could hold a highly ne-grained theory while remaining uncommitted.
Of course, the questions can intersect: The account just considered, one that dis-
tinguishes a multitude of perfectly coincident events by their diering conditions of
occurrence, is highly ne grained in a way that automatically commits it to the exis-
tence of plenty of events that are extremely modally fragileand, for that matter,
plenty of events that are not fragile at all, and plenty of events that are somewhere in
between. (For example, your talking loudly is more fragile than your talking sim-
pliciter, etc.)
Among theories that hold that (all) events are fragile, toward one extreme is a view
according to which any dierence in time, place, or manner of occurrence would
make for a numerically dierent event.16 Why adopt such a position? We can rec-
ognize one important reasonand also an unfortunate consequenceif we return to
late preemption. Recall our paradigm example: Billy and Suzy throw rocks at a
window, with Suzys reaching it rst, breaking it. How can we secure the result that
Suzys throw, and not Billys, causes the breaking? Perhaps by insisting that the
actual breaking is quite fragile: Had the window broken a few moments later, that
would have been a numerically dierent breaking. Then, arguably, the breaking that
actually occurs counterfactually depends on Suzys throw, but not on Billys.
Many people nd this view compelling until they realize that tightening the indi-
viduation conditions for events in this way results in a landslide of causal connec-
tions, too many of them unwanted. (It should also be remembered that variations
on late preemption yield examples immune from this treatment.) Your sore throat
occurred at precisely 4 p.m., but if you had not stopped by your oce to pick up your
bag before you commenced your walk (and thus your discussion), you would have
had a sore throat at 3:59 p.m. Under the fragility account, that sore throat would be
a numerically distinct event from the sore throat at 4 p.m. Hence, standard regularity
and counterfactual accounts of causation will count picking up your bag from your
oce a cause of the sore throat that actually occurred.
Although defenders of fragility accounts recognize that many more events will
count as causes and eects than untutored intuition will be comfortable with, they
believe the cost is worth the benet. David Coady, in Preempting Preemption
(chapter 13), argues that a counterfactual analysis of causation employing fragile
events will solve a number of problems, and along the way he gives an able defense
of the fragility strategy.17
Notice that the advantages that accrue to this strategy do not obviously extend to
an account that, while recognizing the existence of a highly fragile window-breaking,
also recognizes less fragile (and perfectly coincident) breakings, all diering in their
essences. On such an account, there is a window-breaking that necessarily occurs just
when it does; perhaps that breaking counterfactually depends on Suzys throw alone.
But there is also a breaking that could have occurred as much as a few moments
later. What caused it? Its not even intuitively clear what the answer is, but at any
rate it would be an embarrassment for an account of causation if it held that nothing
caused this event (e.g., because no event could be found on which it depends, etc.).
What reason could there be for admitting such uncaused events into ones ontology?
In fact, the reason standardly given is that we need such events as eects (and
causes). More precisely, perfectly coincident events diering only in their essences are
supposed to help us draw certain causal distinctions. For example, suppose Billys
throw makes a slight dierence to the way in which the window shatters, perhaps by
way of the gravitational pull his rock exerts on Suzys. We would like to say: Billys
throw doesnt cause the breaking, but it does help cause that breaking to happen
in just the way it does. Or, returning to the example considered toward the end of
last section, Billys action does not cause the reworks explosion, but it does cause it
to be the color it is. Distinguishing a breaking that necessarily happens in just the
way it does provides us with an event to be an eect of Billys throw; distinguishing
an explosion that is essentially red provides us with an event to be an eect of Billys
earlier action (and, by the way, to be a cause of the subsequent triggering of the
photodetector).
Of course, this reason for drawing purely modal distinctions between events pro-
vides no ingredients for solving the problem posed by late preemption; the theorist
needs to turn elsewhere for such a solution. There is a worse problem. Consider a
simple situation, in which Suzy, all alone, throws a rock at a window, breaking it.
The following sentence is trueand determinately so: Suzys throw is a cause of the
breaking. But to what do the expressions Suzys throw and the breaking refer?
There are far too many eligible candidates for it to be at all plausible that their
reference is determinate. So what? Why not just recognize a (large) number of ad-
missible assignments of referents to each expression? Well, because doing so leaves it
obscure how our sentence gets to have a determinate truth value. Suppose we assign
a comparatively fragile event to Suzys throw and a comparatively robust one to
the breaking; then our eect will in all likelihood not depend on our cause. Nor is
this trouble just for a counterfactual analysis: reverse matters, assigning a fragile
event to the breaking and a robust one to Suzys throw, and a typical regularity
account will have trouble making the latter out to be a cause of the former. The
upshot is that on standard accounts of causation, there will be admissible assignments
of referents that make our sentence false. And so even if there are admissible assign-
ments that make it true, the sentence will not come outas it shoulddeterminately
true.18
Why not opt for a more parsimonious account of eventsholding, for example,
that events are individuated by the spatiotemporal regions they occupy? But such a
seemingly attractive theory itself has unwelcome consequences. If we had a metal
sphere that as it rotated also heated up, we would have to identify the event of the
rotating of the sphere with the event of the heating up of the spherethey are the
same event because they occur in the same spatiotemporal region. But our earlier
reasons for wanting occasionally to distinguish coincident events remain in full force,
for it seems easy to imagine a situation in which the spheres rotation has dierent
eects from its heating.
A dierent approach to causal relata bypasses at least some of these problems by
holding that the causal relata are not events after all. (Or at least, the causal relation
can be said to hold between events only in virtue of holding between something else.)
Various alternative accounts of the causal relata have been put forward, the most
plausible of which take them to be property instances, facts, states of aairs, or
propositions.
Those who hold that the causal relata are property instances often argue that
causation obtains between instantiations of universals or between tropes. A dierent
but related view holds that causation obtains between states of aairs, where states of
aairs consist in particulars having properties. Such views can still accept that events
are causal relata in a derivative sense, since events may well be constructed from
property instances or states of aairs.
In Aspect Causation, chapter 8, Paul argues that we should take the causal
relata to be property instances (aspects), as this can help resolve some outstanding
problems with transitivity and allows us to bypass the need to provide individuation
conditions for events. If the causal relata are property instances, then we can con-
struct a reductive analysis of causation independently of an analysis of eventsa
result that, given the lack of consensus over what counts as a satisfactory theory of
events, philosophers of causation should welcome.
There is a quite dierent motivation for abandoning or at least qualifying the
default position that events are the primary causal relata. It stems from the need to
accommodate causation involving omissions: prevention, causation by omission, and
(arguably) causation by double prevention. (See the contributions by Lewis, Beebee,
Menzies, Mellor, and Armstrong.) Focus just on causation by omission, as when
Billys failure to show up for their lunch date causes Suzy to become disappointed.
The possibility of this kind of causation creates an immediate diculty for those who
would hold that causation must relate events. On the face of it absences (omissions)
are not particulars at all, and so not events; what are we to make, then, of causation
involving them? We seem to nd ourselves with a causal relation that is missing
a relatum. D. H. Mellor, in his For Facts as Causes and Eects (chapter 12),
responds by defending the view that most causes and eects are not particulars, but
rather facts. Mellor takes facts to be states of aairs expressed by true sentences,
statements or propositions.19 An advantage of Mellors position is that it allows him
to handle cases involving negative causation quite naturally. It is an unhappy phi-
losopher who is forced to believe in negative events, but Mellor by contrast can quite
serenely grant that there are negative facts.
Lewiss response, in Causation as Inuence (chapter 3), and especially Void
and Object (chapter 10), is to insist that when there are causal relata, they are
invariably eventsbut to go on to assert that in cases of causation by omission or
prevention, one of the relata is missing: So I have to say that when an absence is a
cause or an eect, there is strictly speaking nothing at all that is a cause or eect.
Sometimes causation is not a relation, because a relation needs relata and sometimes
the causal relata go missing (Causation as Inuence, p. 100). Nevertheless, he
claims, in all such cases there will be an appropriate pattern of counterfactual
dependencenot between two events, but, say, between an event and the fact that
no event of some specied type occurs. (Thus, Suzys sadness depends counter-
factually on the fact that no event of the Billy-showing-up type occurs.) Why this
relation of dependence does not qualify the foregoing fact as a cause is, to our minds,
left a bit unclear.
Beebee responds to problems involving absences by rejecting causation by omis-
sion. For Beebee, causation always relates particulars; hence, there can be no causa-
tion involving absences. She argues that commonsense intuitions about causation by
omission make trouble both for those who defend causation by omission and those
who reject it. For example, she must deny that Billys absence causes Suzys disap-
pointment. But those who disagree typically hold that mere counterfactual depen-
dence suces for causation by omission; consequently, they must hold that the failure
of Bill Gates to show up with a fat check for Suzy also causes her disappointment
(since if he had shown up, she would not have become disappointed). Such parity
leads Beebee to seek other desiderata on an account of causation in order to settle
whether absences can cause; she thinks that the attractiveness of the view that causa-
tion invariably relates events tells against this position. Finally, Beebee accounts for
the central role that causation by omission can play in our ordinary causal claims by
arguing that common sense fails to discriminate between causation and explanation.
Menzies takes a new approach to the problem of how our intuitions about omis-
sions seem to generate conicting results, arguing that we can develop an account
of causation based on relevant dierence-making that distinguishes between absences
as causes and absences as mere conditions. (The distinction between causes and
conditions is relative to a eld of normal conditions generated by a causal model.)
Menziess account is constructed so as to avoid the sorts of spurious causes that
Beebee cites and that make trouble for a nondiscriminating counterfactual account
of causation by omission.
Recognizing the special challenges that causation involving absences presents
suggests that it might be helpful to distinguish kinds of causation. There is garden-
variety causation, as when Suzys throw causes the window to break. There is cau-
sation by omission, as when Billys failure to stop her likewise causes the breaking.
There is prevention, as when (this time) Billy stops the rock midight, preserving the
window. There is double prevention, as when Sally knocks Billy aside before he can
block the rock, thus contributing to the windows breaking. And perhaps there are
still more exotic varieties of omission-involving causation.
It may well be that the dierent kinds of causation require dierent explications.
(Both Hall, in Two Concepts of Causation, chapter 9, and Armstrong, in Going
through the Open Door Again, chapter 19, argue as much.) If so, then it comes as
no surprise that uniform analyses of causation are vulnerable to counterexamples
and inadequate for complicated examples where dierent kinds of causation are
combined. Here it is well to recall some of the problems involving transitivity dis-
cussed in the previous section, where omissions played a central role in the structure
of the case (the inert network is a good example). Hall in particular argues that such
cases threaten transitivity only if we hold that transitivity must apply to the kinds of
causationin particular, causation by double preventionfeatured in them.
The problem of explaining causation by and of omission, and hence by extension
explaining what it is to prevent events from occurring or to allow events to occur, is
an area of continuing research. The importance of the problem is highlighted by the
special roles that preventing and allowing seem to play in related areas of inquiry, for
example, accounts of our ascriptions of moral and legal responsibility.
8 Summaries of the Contributions
Here we provide capsule summaries of the contributions that follow.
Chapter 2: Jonathan Schaer, Trumping Preemption
Schaer develops a novel kind of example that poses trouble for most existing
accounts of causation. So-called trumping preemption occurs when two (or more)
processes, each sucient to bring about some eect, go to completion. But unlike
cases of symmetric overdetermination, there is an asymmetry: One of the processes is
such that variation in it would have been followed by corresponding variation in the
eect, whereas the same is not true of the other. For short, the rst process trumps
the second. As an example, Schaer imagines a sergeant and a major simultaneously
shouting orders to their troops. In cases of conict, the troops obey the major;
but this time the order is the same. Schaer argues, rst, that in cases like this the
trumping event (e.g., the majors shout) is the sole cause of the eect (the troops
response), and, second, that existing counterfactual accounts of causation cannot
accommodate this fact.
Chapter 3: David Lewis, Causation as Inuence
Lewis replaces the old idea that causation is, at heart, counterfactual dependence
between events, with the thesis that causation is, at heart, counterfactual covariation
between events. He rst denes a relation of inuencewhich is, roughly, the
relation an event c bears to an event e just in case the manner of occurrence of e
depends counterfactually in a suitably systematic way on the manner of occurrence
of c. Lewis then takes causation to be the ancestral of inuence. As an illustration,
Lewiss solution to the problem of late preemptionas exhibited, say, in the example
of Suzy, Billy, and the broken windowis that Suzys throw exerts much more in-
uence on the shattering of the window than does Billys: counterfactually varying
Suzys throw in any of a number of ways will result in corresponding changes in the
shattering, whereas the same is not true of Billys throwor at least, not nearly to
the same extent. For that reason, Suzys throw, and hers alone, counts as a cause of
the shattering. En route to developing this account, Lewis treats a number of central
methodological and metaphysical issues, discussing transitivity, causation involving
omissions, the varieties of preemption, and the reasons for pursuing an analysis of
causation at all.
Chapter 4: John Collins, Preemptive Prevention
You catch a ball headed for a window. Common sense credits you with preventing
the window from breaking, even though I leapt behind you ready to take the catch if
you missed. Your catch was an act of preemptive prevention. My leap was a depen-
dence preventer; it prevented the windows not breaking from depending on your
catch. Yet if a brick wall had stood between you and the window, we are now reluc-
tant to say that your catch caused the window to remain unbroken. Collins suggests
that our ambivalence here is governed by how far-fetched it is to assume the absence
or failure of the dependence preventer. This suggests the possibility of an account of
causation in terms of would-be dependence. The analysis falls to Schaers examples
of trumping preemption. Collins goes on to argue that cases of masking and
nkish trumping pose new problems for a counterfactual analysis of dispositions.
Chapter 5: Stephen Yablo, Advertisement for a Sketch of an Outline of a
Prototheory of Causation
Yablo argues that we should return to the original, spartan idea that causation
is counterfactual dependencebut understand that dependence to be something he
calls de facto dependence. The idea is, roughly, that e de facto depends on c just
in case, had c not happenedand had suitably chosen other factors been held
xedthen e would not have happened. As an illustration, Yablo would say that the
shattering of the window de facto depends on Suzys throw because, had Suzy not
thrownand had it still been the case that Billys rock never struck the window
then the window would not have broken. Obviously, the success of this proposal
depends heavily on being able to pick out, in a principled way, which factors should
be held xed; much of Yablos paper is devoted to this question.
Chapter 6: Peter Menzies, Dierence-making in Context
Menzies argues that Lewiss counterfactual analysis of causation is insensitive to the
various context-relative ways in which commonsense distinguishes between causes
and background conditions. Menzies develops the conception of a cause as some-
thing that makes a dierence with respect to an assumed background or causal eld
by giving a detailed analysis of an account of dierence-making in terms of context-
sensitive counterfactuals.
Chapter 7: Ned Hall, Causation and the Price of Transitivity
Various purported counterexamples to the transitivity of causation have recently
appeared in the literature. Hall argues that these fall into two types. On the one
hand, there are examples that only seem to serve as counterexamples to transitivity:
Their intuitive status as such can, in fact, be resisted on a principled basis. On the
other hand, there are examples with much greater intuitive force, but that function as
counterexamples to transitivity only if one endorses the controversial thesis that
counterfactual dependence suces for causation. Hall suggests that it is better to
abandon the controversial thesis than to give up transitivity.
Chapter 8: L. A. Paul, Aspect Causation
Paul shows that reductive analyses of causation that take events to be the causal
relata face counterintuitive consequences with respect to the transitivity of causation,
and she proposes an analysis of causation where aspects, or property instances, are
the causal relata. Aspect causation combines elements of regularity theories with
inuence (patterns of counterfactual covariation) and takes property instances to be
the causal relata. Paul argues that changing the causal relata to property instances
resolves some outstanding problems with transitivity and allows the development of
a theory of causation that is not held hostage to a theory of event individuation.
Chapter 9: Ned Hall, Two Concepts of Causation
Hall argues that philosophers seeking to analyze causation should distinguish two
fundamentally dierent concepts of causation: dependence, which can to a rst ap-
proximation simply be identied with counterfactual dependence, and production,
which requires a very dierent analysis more in the spirit of Mackies original INUS
conditions account. The case for the distinction turns on examples that reveal sys-
tematic tensions between two sets of theses that might be thought to characterize
causation. On the one hand, causation seems to be a transitive, intrinsic, local rela-
tion (local in the sense that there is no causation at a temporal distance). On the
other hand, counterfactual dependence seems to suce for causation, and omissions
seem capable of standing in causal relations (to ordinary events and to each other).
Hall argues that it is impossible to provide an account of causation that does justice
to both sets of theses, but that the tensions can be resolved if we take the rst set to
characterize production and the second to characterize dependence.
Chapter 10: David Lewis, Void and Object
Lewis argues that the possibility of causation by omission requires an analysis that
can handle cases of causation without causal relations. (For an instance of causation
to be an instance of a causal relation, there must be relata for the relation. But an
absence cannot be a relatum, for an absence is nothing at all.) A counterfactual
analysis of causation can provide an analysis of causation that includes an analysis of
causation by omission, because it can dene causation without a causal relation:
Causation is (somehow) a matter of counterfactual dependence of events (or ab-
sences) on other events (or absences). Lewis also argues that the project of dening a
functional analysis of familiar instances of the causal relation (bi) to serve as a
basis for dening an (unrestricted) conceptual analysis of causation fails.
Chapter 11: Helen Beebee, Causing and Nothingness
Against Lewis, Beebee defends the view that there is no such thing as causation by
absence. We do often speak as though absences are causes. Thus: If Smith fails to
water her oce plants, and the plants die, we might say that her failure to water
them caused them to die. Yet common sense balks at the judgment that, for example,
Queen Elizabeths failure to water Smiths plants was also a cause of their death. The
problem is that commonsense assertions and denials of causation by absence seem to
be guided by normative considerations that have no place in a metaphysical account
of causation. Beebee suggests that the key to solving this problem lies in the distinc-
tion between causation and causal explanation. We can admit, she argues, that there
are causal explanations that invoke facts about absences without allowing that
absences are causes. Absences and omissions can causally explain, without being the
causes of what they explain.
Chapter 12: D. H. Mellor, For Facts as Causes and Eects
Mellor argues that facts, rather than particulars such as events, are causes and
eects. For Mellor, singular causal claims are best rendered in the form E because
C, where C and E are sentences and because is a sentential connective; E be-
cause C is interpreted as equivalent to the fact that C causes the fact that E.
Mellor argues that there is no reason to take causation to be a relation, much less a
relation between events, and that this perspective enables negative facts to be causes
and eects. For Mellor, facts are superior to events as causes and eects because fact
causation can handle causation of or by negative facts much more easily than event
causation can handle causation of or by omission.
Chapter 13: David Coady, Preempting Preemption
Coady defends what he calls the naive counterfactual analysis of causation, that c
causes e if and only if, if c had not occurred, e would not have occurred. This anal-
ysis is widely believed to be refuted by the possibility of preemption, which is a spe-
cies of redundant causation. Through a careful consideration of the dierent kinds of
events related in paradigmatic examples, Coady argues that there are in fact no cases
of preemption: All such cases are either cases of symmetrical overdetermination or
are not cases of redundant causation after all.
Chapter 14: Cei Maslen, Causes, Contrasts, and the Nontransitivity of Causation
Maslen develops an account according to which events cause other events only rela-
tive to contrast situations. For Maslen, causation in the world is an objective, three-
place relation between causes, contrasts, and eects, and the truth and meaning of
causal statements depend on the context in which they occur (for it is this context
that helps determine the relevant contrast situation). Maslen also argues that her
contrast theory of causation makes the intransitivity of the causal relation intuitively
plausible, and she gives compelling analyses of several counterexamples to the tran-
sitivity of (binary) causation.
Chapter 15: Igal Kvart, Causation: Probabilistic and Counterfactual Analyses
Kvart presents a probabilistic analysis of (token) causation. He argues that an
account of causal relevance must play a crucial role in any such analysis. Cases of
preemption will then be handled by the fact that the preempted cause is causally
irrelevant to the eect. Causal irrelevance is established by identifying a neutraliz-
ing event that screens o the eect in a stable way. Kvart argues that the resulting
account can deal with cases of probability-lowering causes. In the nal sections of the
essay it is argued that since any adequate account of counterfactuals must be based
on causal notions, the counterfactual analysis of causation will turn out to be circu-
lar. Philosophers have been too quick to abandon the probabilistic analysis of cau-
sation, Kvart suggests.
Chapter 16: Murali Ramachandran, A Counterfactual Analysis of Indeterministic
Causation
Ramachandran develops a counterfactual analysis of causation intended to work
under the assumption of indeterminism. He argues that three core problems for
Lewiss probabilistic theory and his latest causation as inuence theory (given in
this volume) are: (i) it (mistakenly) takes causation to be transitive; (ii) it precludes
the possibility of backward causation; and (iii) it cannot handle certain cases of
overdetermination. Ramachandran tackles these problems by way of three special
notions: (a) a notion of time-invariant overall world chance of an events occurring
already present in the notion of background chance; (b) a notion of accumulative
(counterfactual) chance-raising that discriminates between causal and noncausal
chance-raising chains; and (c) a notion of potential chance-raisers (as in c would
have raised the chance of e but for the occurrence of certain other events).
Chapter 17: Christopher Hitchcock, Do All and Only Causes Raise the Probabilities
of Eects?
According to probabilistic theories of causation, causes raise the probabilities of their
eects. Critics of these theories have tended to concentrate on putative examples of
non-probability-raising causes. Here Hitchcock focusses attention on the possibility
of probability-raising noncauses. These cases turn on fundamental questions about
the nature of indeterministic causation. Probabilistic theories usually presuppose that
indeterministic causes bring about their eects only via their contributions to a total
probability pool, there being no causal facts of the matter extending beyond these
probabilistic contributions. We can construct cases in which it is stipulated that there
are such further facts, but then there is no reason to think that the probabilities in
question are irreducible. The problem, Hitchcock suggests, is that the only cases in
which we do have reason to think that the probabilities are irreducible are cases
involving microphysics, where our ordinary causal intuitions are no longer reliable.
Chapter 18: Tim Maudlin, Causation, Counterfactuals, and the Third Factor
Maudlin argues that the key to understanding much of our thought about causation
lies in distinguishing laws of inertialaws that, roughly, specify how things behave if
left alonefrom laws of deviationlaws that specify in what conditions, and in
what ways, behavior will deviate from the inertial behavior. The primary notion of
cause, according to Maudlin, is the notion of an event that gures in some law of
deviation. He suggests further that the content of this distinction between lawsand
indeed whether such a distinction is even possiblewill depend signicantly on the
level of description of the phenomena under consideration.
Chapter 19: D. M. Armstrong, Going through the Open Door Again: Counterfactual
versus Singularist Theories of Causation
Armstrong criticizes the counterfactual analysis put forward by David Lewis and
defends a singularist theory. Armstrongs singularist theory holds that causation is a
two-term relationa universalholding between cause and eect (where causes and
eects are states of aairs). For Armstrong, causation is conceptually primitive. He
cites several reasons why we can take ourselves to have observational access to sin-
gular causal facts: ordinary language, experiences such as the perception of pressure
on ones body, our awareness of the operation of our will, and psychological experi-
mental evidence. Armstrong also discusses the treatments he favors for several issues
concerning causation, including causation by omission, action at a distance, proba-
bilistic causation, and the connection between causation and laws.
Notes
1. Let us be clear from the outset that by facts here we mean nonlinguistic items.
2. But see Woodward (2003) for a sophisticated development and defense of such an approach. Woodward
grants the circularity, but argues persuasively that the resulting account of causation is nevertheless highly
informative.
3. Albert (2000) also discusses the problem.
4. There is one other lesson from physics that is not so easily avoided, however. Here is an illustration:
Suppose you have before you a glass of hot water. What would happen were you to put an ice cube in it?
We would like to answer: The ice cube would melt. But statistical mechanics, together with the time-
reversibility of the fundamental dynamical laws that appear to govern our world, tells us that there would
be some extremely tiny (but nonzero) probability that the ice cube would grow, while the water heated up.
(Time-reversibility tells us that such an occurrence is nomologically possible, statistical mechanics that it is
extremely unlikely.) If so, then the answer we would like to give is falsesince it might be that the ice cube
would not melt. As far as we know, there is no decent way to construct truth conditions for the counter-
factual that avoids this result. But the counterfactual analyst need not be greatly concerned: She need
merely grant that her account of causation must be built on conditionals of the form if c had not oc-
curred, then the probability that e occurred would have been extremely close to zero. We will ignore
this complication henceforth.
5. See Maudlin (forthcoming) for discussion of how to extend this recipe to counterfactuals with more
complicated antecedents.
6. Observe that if backward causation is possible, and if causation is transitive, then simultaneous and
indeed self-causation are possible: e.g., Bill/Sues collision with itself causes that very same collision. Still,
there is a distinct kind of simultaneous causation whose possibility cannot be secured in this way, where
the causal relationship between c and its contemporaneous eect e is direct, involving the action of no
causal intermediates. We intended the example of the particle in the force eld to illustrate this sort of
simultaneous causation.
7. Here are the technical details: Let f T be a function that gives the particles position for each time T
(specied relative to some inertial frame of reference). We assume that the function is continuous, but not
that it is everywhere dierentiable. But suppose further that for some open interval extending forward
from time T t, the second derivative f
00
T is well dened and continuous; then the particles instanta-
neous acceleration at t, taken from the future, can be dened as the limit as T approaches t from
above of f
00
T. (For a well-behaved particle, this will equal the limit from below, and we can speak
simply of the acceleration at t, simpliciter.)
8. Typically, but not always: Ducasse (1926), for example, takes causation to be essentially nothing more
than spatiotemporal contiguity plus temporal priority. (That is, c causes e just in case c and e are spatio-
temporally contiguous, and c precedes e.) We view this as a kind of limit case of physical connection
accounts.
9. See also Davidson (1967). Note that Davidson does not treat the word cause in the way one might
expectnamely, by providing it with an explicit denition that would allow for the deduction of the target
sentence from sentences that did not themselves contain the word cause. Rather, he includes the word
cause in the statement of the laws themselves, a move that is both bizarre (given that the paradigm
examples of laws provided by fundamental physics never make use of this notion) and that inexplicably
ruins the chances for providing a reductive analysis of causation.
10. Given that Lewis quite explicitly takes himself to be providing a conceptual analysis of causation, it is
therefore mysterious to us why he continues thus: It would be still better, however, if theory itself went
indecisive about hard cases. If an analysis says that the answer for some hard case depends on under-
described details, or on the resolution of some sort of vagueness, that would explain nicely why common
sense comes out indecisive. Lewis nowhere explains why he thinks it would merely be better, as
opposed to crucial, to explain why common sense comes out indecisive when it does.
11. Note that this cannot be the whole story, for ordinary cases of preemption such as that exhibited in
gure 1.2 likewise exhibit this structure.
12. Suggested by Hartry Field, who learned it from Ellery Eells.
13. The case appears in McDermott (1995a).
14. See Lewis (1986d). For a closely related essentialist theory, see Yablo (1992a,b, 2000).
15. See for example Kim (1973a, 1980). According to Kim, each event corresponds to a triple of a con-
stitutive time, particular, and property. It is not entirely clear what it means for such elements to be con-
stitutive. But Kim explicitly distances himself from the view that they are constitutive in the sense that it is
essential to the events occurrence that the given particular has the given property at the given time.
16. Toward one extreme, but not all the way to the end: One could add, for example, that any dierence in
causal origins would make for a dierent event.
17. Lombard (1986) also holds that events are fragile with respect to time, and so could not have occurred
at any time other than the time they actually did.
18. See Bennett (1988) for an expert statement of this problem.
19. Others have also defended the idea that the causal relata are facts; e.g., Jonathan Bennett gives a well-
argued defense of the view that causes and eects are facts in his (1988).
2Trumping Preemption
Jonathan Schaer
Extant counterfactual accounts of causation (CACs) still cannot handle preemptive
causation. I describe a new variety of preemption, defend its possibility, and use it to
show the inadequacy of extant CACs.
Imagine that it is a law of magic that the rst spell cast on a given day match the
enchantment that midnight. Suppose that at noon Merlin casts a spell (the rst that
day) to turn the prince into a frog, that at 6:00 p.m. Morgana casts a spell (the only
other that day) to turn the prince into a frog, and that at midnight the prince
becomes a frog. Clearly, Merlins spell (the rst that day) is a cause of the princes
becoming a frog and Morganas is not, because the laws say that the rst spells are
the consequential ones. Nevertheless, there is no counterfactual dependence of the
princes becoming a frog on Merlins spell, because Morganas spell is a dependency-
breaking backup. Further, there is neither a failure of intermediary events along
the Morgana process (we may dramatize this by stipulating that spells work directly,
without any intermediaries), nor any would-be dierence in time or manner of the
eect absent Merlins spell, and thus nothing remains by which extant CACs might
distinguish Merlins spell from Morganas in causal status.
1 Trumping Preemption
In order to establish the possibility and relevance of trumping scenarios such as
the wizards case, I rebut the following imagined objections: that the causal judg-
ments evoked are unclear, that the laws involved are question-begging, and that
the case invoked is unrealistic, either in the sense that it is empirically implausible,
given what we know about our world, or in the sense that it is pretheoretically im-
plausible, given what we habitually assume about our world. I conclude that deliv-
ering the right verdict in trumping cases is an adequacy condition on an account of
causation.
Objection: It is not intuitively clear that Merlins spell is a cause and Morganas
not of the princes becoming a frog. The judgment of an otherwise successful theory
(assuming CACs are otherwise successful) ought to overrule the judgment of unclear
intuition. The wizards case, and trumping preemptions in general, should be left (to
use David Armstrongs phrase) as spoils to the victor.
Originally published in a special issue of the Journal of Philosophy (April 2000): 165181. Minor revisions
have been made for consistency.
Reply: To my mind, the judgment that Merlins spell is a cause and Morganas not
is intuitively clear (and so, to my mind, any account that cannot accommodate this
judgment just is not an account of causation). But no matter. To the philosopher who
does not share my intuitions, or does not attribute as much clarity to these intuitions,
I reply that any account of causation which hopes to respect the central connotations
of the causal concept had better agree that Merlins spell is a cause and Morganas
not. To treat trumping preemptions as spoils to the victor is to risk spoiling the point
of having a causal concept.
The concept of causation lies in a web of concepts including (most centrally) those
of law, explanation, counterfactual implication, agency, and evidence. To a crudest
approximation, (1) causes and their eects are subsumed under the laws; (2) causes
explain their eects; (3) causes, were they dierent, would counterfactually imply
dierent eects; (4) causes are means to their eects; (5) causes provide evidence both
to and from their eects. Some or all of these relations may need to be rened or
even ultimately abandoned. I submit, however, that any pair of events that satises
all these relations deserves to be considered clearly causally related, while any pair of
events that satises none deserves to be considered clearly not causally related.1
Consider the relation between causes and laws. Since the law is that the rst spell
cast on a given day match the enchantment that midnight, and since Merlins spell
is rst that day, his spell satises the antecedent conditions of the magic law, and so
calling his spell a cause of the enchantment that midnight respects the subsumption
of causal structure by nomic structure. Morganas spell, in contrast, does not satisfy
the antecedent conditions of the magic law, and so her spell has no relevant lawful
consequences. Thus, calling the spell a cause of the enchantment posits causation not
covered by law.
Now, consider the relation between causes and explanations. Intuitively, if you
want to know why the prince became a frog that midnight, Merlins spell is part of a
good answer. The deductive account of explanation vindicates this intuition:
(1) At noon, Merlin cast a spell to turn the prince into a frog.
(2) No other spells were cast that day prior to or contemporaneous with Merlins.
(3) It is a law that the rst spell cast on a given day match the enchantment that
midnight.
Therefore,
(4) The prince became a frog at midnight.2
Thus, treating Merlins spell as a cause respects the connection between causes
and explainers. Morganas spell, in contrast, intuitively explains nothing about the
princes fate, and the deductive account vindicates this intuition:
60 J. Schaer
(1) At 6:00 p.m., Morgana cast a spell to turn the prince into a frog.
(2) . . .
(3) It is a law that the rst spell cast on a given day match the enchantment that
midnight.
Therefore,
(4) The prince became a frog at midnight.
There is no way to ll in (2) so as to get a deductively valid argument without
rendering (1) superuous, since the argument will be deductively valid (in a way that
makes essential use of the laws in (3)) only when a rst spell is specied, at which
point Morganas spell need not be.3 Thus, calling Morganas spell a cause invents
causes that explain nothing.
Next, consider the relation between causes and counterfactual variance. Had
Merlins spell been dierent, such as a spell to turn the prince into a goat, then the
princes fate would have been dierent: He would have become a goat.4 Thus, the
princes fate counterfactually varies with Merlins magic, just as the barometer read-
ing counterfactually varies with the atmospheric pressure, in that in both cases the
noncompossible family of propositions representing dierences in the former coun-
terfactually depends on the noncompossible family of propositions representing dif-
ferences in the lattera relation that David Lewis5 points out is typical of such
intuitively causal processes as measurement, perception, and control. Not so for
Morganas spell: Had she cast a spell to turn the prince into a goat, to protect his
humanity, or even to make the sky the color of gold, come midnight all that would
happen would be the princes becoming a frog (because we are holding xed Merlins
prior spell as well as the laws). To call her spell a cause would be to countenance
causes whose alleged eects were counterfactually oblivious to them.
Next, as to the relation between causes and agency, it is widely agreed that If
an eect is an end, its causes are means to it.6 This is usually spelled out decision-
theoretically: if E is a desired end and C a prospective means, C will be eective if
and only if the probability of E given C is greater than the probability of E given
sC. One way to spell this idea out in more detail is to use agent probabilities (on
which actions are assumed free in having no causal antecedents), so that C will be
eective if and only if p
A
EjC > p
A
E j sC.7 On this test, Merlins spell is an
eective means for (E) the princes becoming a frog, since p
A
(EjMerlin-casts-his-
spell) 1, while p
A
(EjMerlin-does-not-cast-his-spell) < 1 (Morgana is assumed free,
so she might not cast her spell). Morganas spell, on the other hand, is ineective,
since p
A
(EjMorgana-does-not-cast-her-spell) 1 as Merlin has already acted to de-
termine the princes fate. Thus, treating Merlins spell as a cause and Morganas as a
Trumping Preemption 61
noncause of the princes becoming a frog respects the agential connotations of causes
as means.
Finally, consideration of the relation between causes and evidence will converge
on the same conclusion, namely, that Merlins spell is a cause of the princes meta-
morphosis and Morganas not. Intuitively, causes provide (epistemically valuable)
evidence both to and from their eects: If I see someone re a bullet at Joness heart,
I am licensed to infer that Jones will die; and, conversely, if I see Jones supine with a
bullet-shaped entry wound left of his sternum, I am licensed to infer that he was shot.
There is typically a relation of scale in both directions: The more details I know
about the cause, the more details I have license to predict concerning the eect; and
the more details I know about the eect, the more details I have license to retrodict
concerning the cause.
Suppose that all I know about Merlins spell is that it is the rst one cast that day;
then all I have license to predict is that there will be some enchantment at midnight.
If I learn more detail, such as that it targets the prince, then I can predict the result-
ing enchantment in more detail, namely, that it will concern the prince. And so on.
This relation of scale also holds in reverse. If I know that the prince became a frog,
I may infer that a spell of the form Presto, prince to frog! was rst that day. And if I
then learn that Merlins spell was rst that day, I can come to know its intimate details,
though I never witnessed it. Merlins magic ts the princes fate as the gun its bullet.
Morganas spell, once again, fails the test of causes. If I know that hers is not
the rst that day, I can infer nothing about what will happen at midnight from the
details of her performance. If I do not know that hers is not the rst that day, any
inferences I make based on her performance will be, if true, merely accidentally so
(and so not licensed).8 In reverse, if I know what enchantment has befallen the prince
that midnight, I have no license for inferring the details of later spells cast that day,
and any inferences I make concerning Morganas spell out of ignorance as to the fact
that Merlins came earlier will be at best accidentally true.
In summary, Merlins spell is nomically antecedent for, explanatory of, counter-
factually variant with, a means to, and evidence both to and from, the princes
becoming a frog at midnight, while Morganas spell is none of these. So every indi-
vidual measure of causeworthiness converges on the result that Merlins spell is a
cause of the princes becoming a frog, whereas Morganas spell is not.
Objection: The arguments that Merlins spell is a cause and Morganas not of the
princes metamorphosis all turn on the stipulated law that the rst spells cast match
midnight enchantments. But such stipulation begs the question against those (such as
Lewis, the leading proponent of CACs) who hold that laws supervene on occurrent
facts. Why not stipulate that last spells cast match midnight enchantments, or that at
62 J. Schaer
least they do when rst and last spells agree, since in any case the result is the same:
the prince becomes a frog at midnight? If these stipulations represent genuine possi-
bilities, then the supervenience of laws on facts is lost.
Reply: Even if the wizards scenario were incompatible with the supervenience of
laws on facts, this would still show (surprisingly) that the success of CACs is tied to
the supervenience of laws. Those who would deny supervenience in the manner of
Fred Dretske, Michael Tooley, and Armstrong9 (who hold that laws are contingent
necessitation relations between universals) could not avail themselves of this objec-
tion, since on their view there may well be a contingent necessitation relation be-
tween the universals rst spell cast on day x and midnight enchantment on day
x; those who would deny supervenience in the manner of John Carroll and Tim
Maudlin10 (who hold that laws are primitive) could not avail themselves of this
objection either, since on their view there is no barrier at all to the existence of a
primitive lawful relation between the rst spells and midnight enchantments. But no
matter. To the philosopher who is willing to tie the success of CACs to the super-
venience of laws on facts, I reply that the existence of laws favoring the rst spells
even when rst and later spells agree (as per the wizards) is fully compatible with the
supervenience of laws on facts, and is actually entailed by leading supervenientist
accounts of laws, given certain facts in the worlds history.
Suppose laws supervene on facts, and suppose for the sake of deniteness that (1)
laws supervene as per the MillRamseyLewis11 (MRL) account: L is a law of na-
ture if and only if L is a theorem of the axiomatization of the facts that best balances
simplicity and strength. Now (2) let the world history contain decisive competi-
tions between spells, where rst and later spells disagree, and where the rst spell
cast always wins (that is, you try rst to turn the queen into a frog, I try later to turn
her into a goat, and she becomes a frog). Finally (3), let the laws summarizing the
nonmagical remainder of the world history be independent of the magical laws (say,
the world is otherwise Newtonian). This suces to x the laws as per the wizards
case: When rst and later spells agree, it will be simpler (and just as strong) to regard
these cases as extensions of the same overall pattern of rst-spell dominance found in
(2), rather than as new examples of a second pattern of later-spell dominance or rst-
and-later-spell equality; this gain in simplicity will not compromise the simplicity/
strength of the other laws since they are assumed independent by (3), and so this
local and not-compromised-elsewhere gain in simplicity without loss in strength will
by (1) entail laws favoring the rst spells even when rst and later spells agree.
Somewhat more precisely, consider how the laws of a world in which the rst
spells dominate when rst and later spells disagree might cover cases in which rst
and later spells agree:
(a) Asymmetric in allowing only the rst spells to satisfy their antecedents (as per the
wizards): for all spells x and days y, if x is rst on y, then the x-matching enchant-
ment obtains that midnight.
(b) Symmetric in allowing both rst and later spells to satisfy their antecedents:
for all spells x and days y, if x is rst on y or there is a spell z which is rst on y and
which agrees with x, then the x-matching enchantment obtains that midnight.
(c) Asymmetric in allowing only later spells to satisfy their antecedents: for all
spells x and days y, if (x is rst on y and there is no spell z which is later on y and
which agrees with x) or (there is a spell z which is rst on y and which agrees with
x and there is no z
0
which is both later than x and which agrees with x), then the
x-matching enchantment obtains that midnight.
(d) Silent in these cases: for all spells x and days y, if x is rst on y and there is no
spell z which is later on y and which agrees with x, then the x-matching enchantment
obtains that midnight.
Now, it is apparent that (a) is simpler than (b), (c), or (d): (a) subsumes all cases
under one pattern/clause, that of rst-spell dominance, while (b)(d) require multiple
clauses to accommodate multiple patterns;12 it is especially revealing that (d), which
is silent on rst-later agreements, is actually more complicated than (a)to get the
laws not to say anything about rst-spell dominance, you need to add a silencing
clause. Further, it is obvious that (a) is as strong as (b) or (c) and stronger than (d):
(a)(c) have identical consequences as to which events are entailed, and how infor-
matively described each event is; whereas (d) is as strong but for cases of rst-later
agreement under which no events are entailed. So, on the MRL account, given that
laws are the theorems of the maximally simple and strong deductive system (I have
been supposing that the laws of magic are independent of the other laws), it follows
that the laws of magic say that the rst spells entail matching enchantments.
Generalizing away from the MRL account, the above result shows that any
account of laws which rules simplest laws at least possible (which I consider an ade-
quacy condition on an account of lawhood, given the role of simplicity in scientic
practice) must deliver at least some worlds with laws as per the wizards. So even
someone who thinks laws are contingent necessitation relations between universals in
the manner of Dretske, Tooley, and Armstrong, or who thinks laws are primitive in
the manner of Carroll and Maudlin, should accept that there are at least some worlds
with laws as per the wizards (they will presumably add that there are other worlds,
factually identical to the wizards with laws as in (b), (c), and (d), and maybe with no
laws at all). Thus, the laws invoked should be unproblematic on reection.13
64 J. Schaer
Objection: Causation is an empirical concept, and an account that mishandles
fairy-tale cases like the wizards will be no less empirically adequate. Even if the wiz-
ards case is a conceptual possibility, it involves magic and action-at-a-distance, and
so is too far-fetched to worry about.
Reply: I think accounts of causation need to be not just empirically but con-
ceptually adequate. Without prior conceptual analysis, how could we tell whether an
empirical investigation had hit its mark? But no matter. To the philosopher who
scorns such merely conceptual possibilities as those involving magic and action-at-
a-distance and demands empirically plausible scenarios, I reply that there exist em-
pirically plausible trumping scenarios.
Imagine a world whose laws of physics are very much like our own, but for rec-
ognizing types of elds, black, gray, and white, which do not superpose.
Whenever a particle is subject to just one of these elds (assume for simplicity that no
further forces are present), the particle will accelerate along a curved trajectory (g.
2.1). Whenever a particle is subject to multiple elds, the particle will accelerate as if
subject only to the darkest eld, for example (g. 2.2).14 Thus, the laws of physics at
this world would include a color-eld law, most simply formulated as: The intensity
of the darkest elds on a given object the color force on that object. Although this
world does not have the physics we now believe our world to have, it is relatively
easy to imagine actual physicists discovering further types of elds, performing the
experiments represented above, and supplementing our laws of physics in accord.
Source (black, gray, or white)
Figure 2.1
Black source White source
Figure 2.2
Moreover, this world is free of magic, action-at-a-distance, and other types of far-
fetched causal connections. It is in this sense that it is empirically plausible.
Now, if a particle is subject to both a black and a white eld that pull in the same
direction with the same magnitude, it is still only the black eld that causes the
resultant acceleration (g. 2.3). Here, we have an empirically plausible trumping
scenario: replace Merlin with the black source, Morgana with the white source, the
princes fate with the particles acceleration, spells with elds, and laws favoring
earlier spells with laws favoring darker eldsthe arguments from cause-worthiness
all carry over.
The elds case shows that all that is needed for trumping are three events, C1, C2,
and E, where E would be produced by C1 alone or by C2 alone (redundant causa-
tion), but where the laws render C1 the exclusive factor in the entailment of E
(trumping preemption). This is why the causal exotica of the wizards case is inessen-
tial. Nothing empirically implausible remains.
Objection: Causation is a folk concept, and an account of causation which mis-
handles fairy-tale cases like the wizards and faux-physical cases like the elds will be
no less useful in our everyday dealings with people, rocks, and the like. Even if the
wizards case is a conceptual possibility and the elds case an empirical possibility,
both involve behaviors so unlike those typical of local middle-sized dry objects that
both cases are too removed from our habitual assumptions about the world for us to
expect our concept of causation to be adequate to them.
Reply: I think the objector is overly pessimistic about the viability of our concepts
in unfamiliar scenarios. The intuitive clarity of a judgment in a given case strikes me
as strong evidence for the viability of the judged concept in that case (I hope to have
already convinced the reader of the intuitive clarity of causal judgments in trumping
cases). But no matter. To the philosopher skeptical of judgments in unfamiliar sce-
Black source
White source
Figure 2.3
66 J. Schaer
narios who requires scenarios that accord with the world we habitually take ourselves
to inhabit, I reply that there exist pretheoretically plausible trumping scenarios.
Imagine that in a world that could well be our own, the major and the sergeant
stand before the corporal, both shout Charge! at the same time, and the corporal
decides to charge.15 Orders from higher-ranking soldiers trump those of lower rank.
I hope you agree that the majors order, and not the sergeants, causes the corporals
decision to charge (for the same reasons as in the wizards and elds cases: Ranking
orders are like the rst spells and darkest elds). The soldiers case shows that every-
day scenarios can stand in the pattern constitutive of trumping preemption. Nothing
pretheoretically unfamiliar is needed.
The sophisticated philosopher might presume that the soldiers (unlike the wizards
and the elds) must be a case of standard preemption, by presuming that, as the
majors and sergeants orders engage the corporals mental mechanisms en route to
his decision module, there must be some (at least unconscious) intermediate lter
through which the majors order is passed but the sergeants order blocked. But those
who would reject the wizards and elds cases on the grounds that causation is a folk
concept are not entitled to such sophisticated presumptions. Perhaps the corporal
decides to charge via some such unconscious intermediate lter (as in standard pre-
emption), or perhaps he decides to charge because his training has simply made his
decision module exclusively sensitive to ranking orders (as in trumping preemption).
The folk do not know how the mind is wired here, and from the point of view of
attributing causation, they do not care.
In conclusion, trumping preemptions are intuitively clear, unproblematic in their
assumptions, and fully realistic in being both empirically and pretheoretically plau-
sible. Pending further objections, I take the above to show that delivering the right
verdict in trumping cases is an adequacy condition on an account of causation.
2 Counterfactual Accounts of Causation
CACs, developed especially by Lewis, in their simplest form claim that: C causes E if
and only if (i) C and E are actual, distinct events, and (ii) if C had not occurred, E
would not have occurred.
Preemptions show that simple counterfactual dependence is not necessary for cau-
sation, because, in evaluating the counterfactual implications of Cs not occurring,
we hold the presence of the backup xed, and so still get E. In response, defenders of
CACs have rened the counterfactual-dependence relation. The leading extant rene-
ments (Lewiss ancestral dependence, fragility, and quasi-dependence; Peter Menziess
continuous processes; Michael McDermotts minimal-counterfactual suciency;
Murali Ramachandrans minimal-dependence sets)16 all rely on failed intermediaries
along the backup process or on would-be dierences in the eect absent the main
process to distinguish preempting causes from preempted backups. For this reason,
none has the resources to distinguish trumping causes from trumped backups.
Ancestral dependence, quasi-dependence, continuous processes, and minimal-
dependence sets all rely on failed intermediaries along the backup process to
distinguish preempting causes from preempted backups. According to ancestral de-
pendence (suggested by Lewis to transitivize simple counterfactual dependence and
handle preemption cases in one swoop), C may cause E even though E does not
counterfactually depend on C, as long as there is an event D (or some nite set of Ds)
such that E counterfactually depends on D and D on C. Thus, where there are two
gunmen C and C
0
but only C shoots, the death E does not counterfactually depend
on either C or C
0
, but does presumably depend on the bullet-ying D, which in turn
depends on C. Ancestral dependence works only if there are failed intermediaries on
the backup process prior to the eect (early), since these failures are what create
counterfactual dependence of the eect on some post-failure intermediary D along
the main process, by spoiling the dependency-breaking redundancy. For this reason,
ancestral dependence miscounts the trumping cause as a noncause. There is no
intermediary along the trumping process such that the eect depends on it, either
because the causation is direct (no intermediaries), as in the wizards, or because the
trumped process is fully intact, as with the white eld that propagates all the way to
the particle, and with the sergeants order that remains with the corporal throughout
the decision process, breaking the dependency of the eect on all intermediaries
along the entire trumping process.
According to quasi-dependence (tentatively adopted by Lewis at the end of post-
script E to Causation), C may cause E even though E does not counterfactually
depend on C, as long as C and E are part of a process that is intrinsically such as
typically to induce counterfactual dependence (this relation of counterfactual or
quasi-dependence is then transitivized by taking the ancestral). Thus, where there are
two gunmen C and C
0
but only C shoots, the C to E process will still be intrinsically
such as typically to induce counterfactual dependence, because the presence of a
backup gunman is presumably atypical. What about the C
0
to E process? The only
reason I can see why this process (understood, for example, as the sequence of events
hC
0
armed and malevolent, E deadi) is intrinsically not such as typically to induce
counterfactual dependence is that there are failed requisite intermediaries along
this process, such as a failure of C
0
to pull the trigger.17 For this reason, quasi-
dependence miscounts the trumped backup as a cause. If Merlins spell is to count
68 J. Schaer
as a cause because the great majority of intrinsic duplicates of the sequence hMerlin
casts prince to frog, prince turns to frogi exhibit counterfactual dependency, then
Morganas spell must count as a cause as well, since the great majority of intrinsic
duplicates of hMorgana casts prince to frog, prince turns to frogi will exhibit just as
much dependency, since in both cases we abstract away from the atypical presence of
a rival spell caster; likewise for the eld-process and soldier-process duplicates.
Trumping cases show that the intuition that motivates quasi-dependencenamely,
that causation is determined by the intrinsic character of events (plus the laws)is
hasty, since the laws may well be sensitive to extrinsic characteristics, such as which
spell is earlier.
The continuous-process renement (suggested by Menzies as a solution to
making it the hard way preemption cases) requires that C and E be connected by
a temporally continuous chain of counterfactual dependencies. Here, failed inter-
mediaries function as discontinuities in the dependency chains from the backup.
Without these discontinuities, the backup will spoil the nal dependency of eect on
the immediately prior event along the causal process. Thus, because the trumped
backup runs continuously in the elds case, the nal dependency of the particles
acceleration on the black source is spoiled by the continued propagation of the white
eld, and so the continuity requirement miscounts the trumping cause as a noncause.
Worse, requiring temporal continuity rules out by brute stipulation the intuitive
possibility of action-at-a-distance and so could not count Merlins spell as a cause
even if Morgana had not cast a spell at all.18
According to minimal-dependence sets (devised by Ramachandran), preempter
C counts as a cause of E because the C-process involves only actual events, while
backup C
0
is not a cause because the C
0
-process includes failed events. First, Ram-
achandran denes an M-set for E as a nonempty set of events S such that:
(i) E is not a member of S, (ii) if no members of S had occurred, E would not have
occurred, and (iii) no proper subset of S meets (ii).
He then shows that M-sets will contain exactly one event, actual or not, from each
process, preempted or not, leading to E, and concludes that C causes E if and only if:
(iv) C and E are actual events, (v) C is in an M-set for E, and (vi) there are no two
M-sets for E which dier only in having one or more nonactual events in place of C.
The failed-intermediaries idea is implemented in (vi), since given that M-sets contain
exactly one event from each process, C will be replaceable by a nonactual event (and
thus not a cause) only if there is a nonactual event on its process. Ramachandran is
fully explicit in his reliance on the failed-intermediaries idea:
It seems true in all genuine cases of causal pre-emption . . . that the pre-empted processes do
not run their full course. . . . For any pre-empted cause x, of an event, y, there will be at least
one possible event . . . which fails to occur in the actual circumstances but which would have to
occur in order for x to be a genuine cause of y. . . . All genuine causes, on the other hand, do
seem to run their full course; indeed, they presumably count as genuine precisely because they
do so. (1997a, p. 273)
Thus, minimal-dependence sets miscounts the trumped backup as a cause, because it
runs its full course. In the wizards case, {Merlins spell, Morganas spell} will be an
M-set for the princes metamorphosis, but there will be no nonactual event substi-
tutable for Morganas spell, since there is no nonactual event along the process from
Morganas spell to the eect; likewise for the elds and soldiers cases.
Thus, ancestral dependence, quasi-dependence, continuous processes, and mini-
mal-dependence sets cannot distinguish trumper from trumped. It is important to see
that this failure is due, not to the mere details of each renement, but to the under-
lying assumption they share (call this the cutting assumption) that backup processes
are such because they have failed intermediaries. Trumping preemptions show that
the root idea of these renements is inadequate.
Both fragility and minimal-counterfactual suciency add the strictest standards of
event individuation to the failed-intermediaries idea. According to fragility (proposed
though tentatively rejected by Lewis on grounds of spurious causation), the above
ancestral-dependency analysis is to be implemented on the understanding that events
have their times and manners of occurrence essentially.19 Fragility-based approaches
are intended to handle late preemption cases in which the failed intermediaries on the
backup process occur at or after the time of the eect (since it then follows, assuming
no simultaneous/backward causation, that had the backup been left to produce the
eect, the eect would have been delayed) and are no help at all with trumping. In
the wizards case, for instance, the princes metamorphosis, absent Merlins spell,
would still take place at exactly midnight, and in precisely the same manner; the
analogous point applies to the elds and soldiers cases. Thus, the fragility renement
miscounts the trumping cause as a noncause.
According to minimal-counterfactual suciency (due to McDermott 1995a), the
root causal notion is of counterfactual suciency between strictly individuated
events, where a set of actual events C is counterfactually sucient for E if and only
if:
There is no set of actual events D, where E is not a member of D, such that (if the
members of C had all occurred and the members of D had all not occurred, then E
would not have occurred).20
Now, C is minimally counterfactually sucient for E if and only if:
70 J. Schaer
(i) C is counterfactually sucient for E, (ii) no proper subset of C is
counterfactually sucient for E, and (iii) there is no actual event D (distinct from E
and the members of C) such that for some proper subset C
0
of C, C
0
WfDg is
counterfactually sucient for E, and for some proper subset C
00
of C, C
00
WfsDg is
counterfactually sucient for E.
McDermott equates direct causation with minimal-counterfactual suciency, and
claims that, a part of a minimal sucient condition is always a cause (1995a,
p. 535). He denies the converse, and so supplements his account by dening processes
via chains of minimally counterfactually sucient conditions, relativizing causes to
processes, and distinguishing between realizing the real as opposed to nominal es-
sence of an event. I need not spell out these (considerable) complications here, since
the trumped backup disproves the claim that a part of a minimally counterfactually
sucient condition is always a cause, and since these complications merely counte-
nance more causes. Consider the set {Morganas spell}. It is a sucient condition,
according to the strictly individuated counterfactual-based denition, since its oc-
currence by itself would still lead to the princes becoming a frog exactly as and when
he actually does. It is clearly minimal, since (i) it is sucient, (ii) its only proper
subset, f g, is not sucient for the princes becoming a frog, and (iii) there is no
actual event D such that both f g WfDg as well as f g WfsDg are sucient for the
princes becoming a frog (in particular Merlins spell fails the latter). But Morganas
spell, despite being part of a minimally counterfactually sucient condition, is not a
cause of the princes becoming a frog.21
Thus fragility and minimal counterfactual suciency cannot distinguish trumper
from trumped. This failure is due to the underlying assumption they share (call this
the precision assumption), namely that backup processes are such that had they run
to completion the eect would have been somewhat dierent in time or manner.
Trumping preemptions show that the root idea of these renements is inadequate.
In conclusion, extant CACs still cannot handle preemptive causation. Trumping
preemptions show that the failed intermediary and would-be dierences strategies
that extant CACs use are inadequate as general solutions to the preemption prob-
lem.22 Of course there may yet be some new strategy for CACs which will prove
adequate, but at this point the prospects look dark.
Notes
1. Compare D. H. Mellor (1995, pp. 5866), who argues for the viability of indeterministic causation as
tting the connotations of causation.
2. The main extant accounts of explanation are (a) deductive, (b) causal, and (c) pragmatic accounts.
While (b) and (c) are not helpful here, because with regard to (b) the cause of the princes metamorphosis
is currently at issue, and with regard to (c) these accounts yield (by design) no objective fact, account (a)
proves helpful. I take it that most will concede that, whether or not deductive accounts are ultimately via-
ble, they at least provide reasonable codications of our intuitions in most cases.
3. There are tricks around this, such as by lling in (2) with some sort of disjunction, such as: either Mor-
ganas spell was rst that day, or it was not but what was rst agreed with Morganas in the enchantment it
called for. But these are just the tricks which a deductive account of explanation must rule out as implau-
sible anyway, and so in using the deductive account to justify our intuitions, we may assume (though we
may not even be able to say how) such tricks are excluded.
4. This follows on such leading accounts of counterfactual implication as David Lewiss in his (1973b).
According to Lewis, we evaluate the implications of the counterfactual supposition that Merlin had cast a
spell to turn the prince into a goat by postulating a minimal divergence miracle on which Merlin says
Presto, prince to goat! instead of Presto, prince to frog! while holding xed the course of history up to
Merlins conjuring (so holding xed the absence of any prior or contemporaneous spells that day) as well
as the laws (so presumably keeping the magic laws intact) to the extent compossible with the miracle; all
this results in Merlins casting the rst spell that day to turn the prince into a goat, where the rst spells
cast match midnight enchantments, and so all this results in the princes becoming a goat that midnight.
5. In his 1973a. Since I shall show that Lewiss counterfactual account of causation fails to accommodate
this variational type of counterfactual dependence in trumping cases, one way to read this result is that
Lewiss account of causation as counterfactual dependence between event occurrences fails to capture his
own intuitions about the general relation between counterfactual variation and causation.
6. See Mellor (1988), p. 230.
7. The main versions of decision theory are (a) evidential, and (b) causal (turning on whether the
probability of E given C is interpreted as (a) pEC=pC, or (b) p(C will cause E)). While (b) is not
helpful since whether C causes E is currently at issue, (a) proves helpful, and since (a) and (b) agree on
non-Newcomb cases, (a) can be used neutrally in the case at hand. (See Harper and Skyrms, eds., 1988.)
The point of introducing agent probabilities is to free evidentialism from the spurious correlations of
Newcomb cases. For further discussion, see Menzies and Price (1993).
8. In a sense, Morganas spell is evidence for the princes becoming a frog, since given certain (plausible)
prior probability distributions, one may assign a greater subjective probability to the princes becoming a
frog given Morganas spell. But this is not a sense of evidence relevant here, since in this sense, given
certain prior probability distributions, anything may count as evidence for anything (in particular the
increased probability that the conditionalizer may assign prince-to-frog given Morganas spell is exactly
parallel to the increased probability assignable victims death given backup shooters presence; Morganas
spell carries exactly as much and as little evidence of the eect as any preempted process). The sense of
evidence relevant here, which typically holds between cause and eect, which holds between Merlins
spell and the princes becoming a frog, and fails between Morganas spell and the princes becoming a frog,
is evidence with epistemic value: that which may confer knowledge.
9. See Dretske (1977a); Tooley (1987); Armstrong (1983).
10. See Carroll (1994); and Maudlin (forthcoming).
11. See Lewis (1973b); also Earman (1984).
12. We could, of course, be more precise, or employ dierent vocabulary, but no added precision or shift
of (natural) vocabulary should alter the greater simplicity of (a) as compared to (b)(d). Of course, su-
ciently gruesome predicates would alter this verdict, but the MRL account of lawhood, which I am
presently assuming, is viable only to the extent that our similarity judgments, based in part on our vocab-
ulary, have objective bite; so for purposes of showing that the wizards laws beg no questions against
supervenientists like Lewis, we may assume what the MRL advocate must.
13. Perhaps some will balk at the laws invoked, not because of worries about supervenience, but because
of disbelief in the possibility of laws projecting extrinsic properties, such as being rst cast on a given day.
I see no grounds for such disbelief (moreover none of the leading accounts of lawhood provides grounds
for such disbelief ). Further, I take worlds such as the wizards world, in which there is perfect covariance
between events with the extrinsic property of being-the-rst-spell-cast-on-day-x and events of midnight
72 J. Schaer
enchantments, and in which there may well be no such covariance between events merely-intrinsically-
described and midnight enchantments, to show exactly why such disbelief is unwarranted.
14. This does not mean that white elds in general are always trumped: there may be times when no black
elds exist, or the black elds might have nite spatial extent, or the black elds might trump only insofar
as they exceed a certain threshold. Nor does it mean that this particular white eld is completely epi-
phenomenal: it might trump on other types of particles, or trigger a white ag in the distance, or have any
number of other eects as long as they are not manifest on this particle. More interestingly, there might be
a functional law on which the ratio of darker eld to lighter eld inuences cycles through the epochs, with
the relevant experiments done in a 1:0 epoch. This raises the empirically open possibility that our world is
such a world (with respect to, say, types of electromagnetic force): our epoch just happens to be 1:1, so we
do not notice.
15. This particular case is a variant of one due to Bas van Fraassen, though the discovery of this type of
case is Ned Halls. A variant of one of Halls cases illustrates just how common these cases are: the Bradys
are deciding where to vacation, and Cindy and Bobby both plead Hawaii! Since Ma and Pa always do
what Cindy wants, the Bradys decide on Hawaii. Trumping is everywhere.
16. These accounts are oered, respectively, in Lewis (1973a) and (1986b); Menzies (1989a); McDermott
(1995a); Ramachandran (1997a).
17. Jonardon Ganeri, Paul Noordhof, and Ramachandran come to the same understanding of quasi-
dependence in their (1996), p. 220.
18. Menzies (1996) has recently moved to a theoretical denition of causation on which causation is that
intrinsic relation between actual, distinct events which typically induces counterfactual dependence be-
tween them. This denition is still troubled by trumping, since it takes as platitudinous that causation is an
intrinsic-to-its-pairs relation between events. Take dierent-day duplicates of the pair hMerlins spell,
princes metamorphosisi. Whether a given duplicate is causal will depend, inter alia, on whether there
was a Morgana spell at 9:00 a.m. Thus, trumping refutes Menziess central assertion that, The distinc-
tive mark of our intuitive concept of causation . . . is that it takes causal relations to be determined by the
natural properties of the relata and the natural relations holding between them, taken in isolation from
everything else happening in the world (ibid., p. 100).
19. We may distinguish temporal fragility from fragility of manner, and fragility of eects from fragility of
events generally. All Lewis needs (for his reply to late preemption) is the temporal fragility of eects. In
fact, all that is needed is the stipulation that the eect not be delayed, as L. A. Paul has shown (1998b).
Nothing here will turn on the details of how the fragility proposal is implemented.
20. This is a questionable use of counterfactuals. It is dicult enough to say what would happen on the
counterfactual supposition that one event not occur, and almost completely indeterminate what would
happen had most of the rest of the world other than C and E not occurred: Spacetime might simply col-
lapse. And the strict individuation of events which McDermott adds only makes things worse, because it
tightens the standards for saying that E still occurs.
21. Adding the rest of McDermotts machinery: the set P of all events involved in the process is {Morga-
nas spell, the prince becoming a frog}, the set Pi of intermediate events is f g, and the set Pc of primary
causal factors is {Morganas spell}. Now, it is clear that Morganas spell is counterfactually sucient rel-
ative to P, and minimally so, and so vis-a`-vis realizing the nominal essence of E, which is just the princes
becoming a frog. The added machinery will not help with the elds or soldiers, either. In the elds case (the
analogous points will hold for the soldiers), the process is something like {the insertion of the white source,
the propagation of the white eld, the white eld hitting the particle, the particle accelerating along a
curved trajectory}. The set Pc is {the insertion of the white source}. Now, all the intermediaries in the
process are held xed, and so the insertion of the white source now counts as counterfactually sucient
relative to P, and minimally so, and so vis-a`-vis the nominalization the particles accelerating along a
curved trajectory, and thus the presence of the white source is miscounted as a cause.
22. One solution to trumping is to analyze causation via nomic subsumption rather than counterfactual
dependence, since Merlins spell is subsumed under the magic law while Morganas is not (likewise for
the elds, and likewise presumably for the soldiers given that the corporals decision module works as
described). But this is a whole other story.
3Causation as Inuence
David Lewis
1 Lecture I
My second paper in my rst philosophy course defended a counterfactual analysis of
causation. Ive been at it, o and on, ever since. But its obvious that the simplest
counterfactual analysis breaks down in cases of redundant causation, wherefore we
need extra bells and whistles. Ive changed my mind once more about how those bells
and whistles ought to work.
This paper mostly presents the latest lessons Ive learned from my students. Under
the customs of the natural sciences, it should have been a joint paper, the coauthors
being (in alphabetical order) John Collins, Ned Hall, myself, L. A. Paul, and
Jonathan Schaer. But under the customs of philosophy, a paper is expected to be
not only a report of discoveries, but also a manifesto; and, happily, the ve of us
have by no means agreed upon a common party line. So, while Im much more than
usually indebted to the work of Collins, Hall, Paul, and Schaer, they cannot be held
responsible for the position I have reached.
1.1 Why Seek a Counterfactual Analysis?
The best reason to persist in trying to make a counterfactual analysis of causation
work is that the diculties that confront rival approaches seem even more daunting.
It is not a foregone conclusion that causation requires analysis at all. Is there,
perhaps, an unanalyzable relation of singular causation, which we know by percep-
tual acquaintance, and which we are therefore in a position to refer to and think
about? It might indeed turn out that this relation can be identied with some relation
already familiar to us either from physics or from metaphysical speculation; but if so,
that identication would be a physical or metaphysical hypothesis, not a matter for a
priori conceptual analysis.1
Hume, of course, taught that we never perceive causation, but only repeated suc-
cession. But its famously dicult to draw the line between whats true according
to perceptual experience all by itself and whats true according to a system of beliefs
shaped partly by perceptual experience and partly by previous beliefs. The boot
comes forward and touches the ball, and straightway the ball ies o through the
goalposts. Do I see that the one thing causes the other? Or do I infer it from what I
do see, together with my background knowledge about the ways of the world? I dont
Originally published in a special issue of the Journal of Philosophy (April 2000). Minor revisions have been
made for consistency.
know, and I dont know how to nd out. So Im in no position to deny that in such a
case Im perceptually acquainted with an instance of a causal relation; and thereby
acquainted as well with the relation that is instantiated.2
Im acquainted with a causal relationnot with the causal relation. Causal rela-
tions are many and various, and no amount of watching the footy will acquaint me
with all the causal relations there are, let alone all the causal relations there might
have been. And yet I seem to have picked up a general concept of causation, appli-
cable to all dierent kinds of causation, and applicable even to kinds of causation
never found in our own world. Thats the real problem, even if I concede pace Hume
that I sometimes perceive causation.
If ever I perceive causation, I perceive it when I watch the footy; or, to take the
customary example, if I watch the motions of billiard balls. But the causal mecha-
nism whereby a dinner too low in carbohydrate causes low blood sugar is utterly
dierent. The causal mechanism whereby our former congressman helped cause his
own defeat by literally singing the praises of Kenneth Starr is dierent again. And
so on, and so forth; not quite ad innitum if we limit ourselves to actuality. But we
should not limit ourselves to actuality, given that we can perfectly well understand
fantasies, or theologies, in which causation works by magical mechanisms entirely
alien to the world of our acquaintance. We are not perceptually acquainted with each
and every one of all these dierent actual and possible causal relations. If there is a
single causal relation, either it is a far from natural relation, a gruesomely disjunctive
miscellany, and so not the sort of relation we can become acquainted with by being
acquainted with some few of its disjuncts; or else the many disjuncts have something
in common. I think conceptual analysis is required to reveal what it is that all the
actual and possible varieties of causation have in common.
A parallel objection applies to the Canberra plan for causation. The plan is that
we rst elicit, Meno-fashion, the folk theory of causation that we all implicitly hold.
That done, we can dene causation as whatever it is that comes closest, and close
enough, to occupying the role specied by our folk-theoretical platitudes. We require
conceptual analysis of the job description, so to speak; but not of the actual occupant
of the role specied thereby.3,4 We leave it open that the role may be occupied by
dierent relations in dierent possible worlds, thereby explaining how our concept of
causation applies to causation by possible mechanisms alien to actuality. But the
problem of the many diverse actual causal mechanisms, or more generally of many
diverse mechanisms coexisting in any one world, is still with us. If causation is, or
might be, wildly disjunctive, we need to know what unies the disjunction. For one
thing the folk platitudes tell us is that causation is one thing, common to the many
causal mechanisms.
76 D. Lewis
The problem becomes especially acute when we remember to cover not only cau-
sation of positive events by positive events, but also causation by absences, causation
of absences, and causation via absences as intermediate steps. The most fundamental
problem is that absences are unsuitable relata for any sort of causal relation, by rea-
son of their nonexistence. This is everyones problem. It is not to be dodged by say-
ing that causation involving absences is really causation*, a dierent thing from
genuine causationcall it what you will, it still needs to be part of the story. It is
my problem too, and I shall return to it; but in the meantime, let the missing relata
objection join the miscellany objection as reasons to think that acquaintance with
the causal relation, or characterization of it as the occupant of a role, are not
workable rivals to a conceptual analysis of causation.
If we are convinced of that, one rival to the counterfactual analysis remains
standing. That is the analysis that says, roughly, that a cause is a member of a set
of conditions jointly sucient, given the laws of nature, for the eect (or perhaps for
a certain objective probability thereof ). (See White 1965, pp. 56104; and Mackie
1965.) This deductive-nomological analysis is descended from Humes constant con-
junction theory, just as our counterfactual analysis is descended from Humes o-
hand remark that if the rst object had not been, the second never had existed.5
However, we dont want to count C as a cause of E just because C belongs to
some sucient set or other for E: A sucient set remains sucient if we add irrele-
vant rubbish, and C might be exactly that. C should belong to a minimal sucient
set, and that is not easy to dene. It wont work just to say that no condition in the
set may be deleted without rendering the remainder insucient: that can be circum-
vented by mingling relevant and irrelevant information in such a way that each item
in the set contains some of each. I might suggest appealing to a counterfactual: we
want our sucient set to consist of items without which the eect would not have
occurred. (That doesnt work as it stands, but its at least a step in the right direc-
tion.) But now we have departed from the deductive-nomological analysis in the
direction of a counterfactual analysis.
Another diculty is that it can perfectly well happen that an eect is a member
of a minimal jointly sucient set for its cause; or that one eect of a common cause
is a member of a minimal jointly sucient set for another. The falling barometer
supposedly causes the low pressure; or the falling barometer supposedly causes the
storm. Even if we were willing to declare a priori that no cause ever precedes its ef-
fect, that would be no solution. The falling barometer does precede the storm. I know
of no solution to these familiar diculties within the connes of a purely deductive-
nomological analysis of causation.
Causation as Inuence 77
The nal rivals to a satisfactory counterfactual analysis of causation are the
unsatisfactory counterfactual analyses. Long, long ago, I thought it would suce to
say that event C is a cause of event E i E depends counterfactually on C; i, if C
had not occurred, E would not have occurred. But this turns out to need qualica-
tions before we have even a sucient condition for causation.
First. We need the right kind of relata. C and E must be distinct eventsand dis-
tinct not only in the sense of nonidentity but also in the sense of nonoverlap and
nonimplication. It wont do to say that my speaking this sentence causes my speak-
ing this sentence; or that my speaking the whole of it causes my speaking the rst
half of it, or vice versa; or that my speaking it causes my speaking it loudly, or vice
versa. Nor should C and E be specied in an overly extrinsic way; it wont do to say
that events a third of a century ago caused me to speak this sentence in the place
where once I was a student. (Though those events did cause my speaking simpliciter.)
See Kim (1973b) and Lewis (1986d).
Second. We need the right kind of counterfactual conditionals. Why cant we say,
given the laws connecting barometer readings and air pressure, that if the barometer
hadnt fallen, that would have been because the pressure wasnt low? Why cant we
then conclude that if the barometer hadnt fallen, there wouldnt have been a storm?
Yet if we say such things, why doesnt our counterfactual analysis fail in just the
same way that the deductive-nomological analysis did?I agree that were within
our linguistic rights to assert these backtracking, or back-and-then-forward, coun-
terfactuals. But they are out of place in the context of establishing causal connec-
tions. Here the much-bemoaned exibility of counterfactual conditionals is our
friend. When we imagine Caesar in command in Korea, we have a choice: We can
hold xed Caesars military knowledge, or we can hold xed the weaponry of the
Korean war. Likewise when we imagine the barometer not falling, we have a choice:
we can hold xed the previous history, or we can hold xed the lawful connections
between that history and what the barometer does. For purposes of analyzing cau-
sation, our policy in all such cases must be to prefer the rst choice to the second. If
need be, we hold history xed even at the price of a miracle (see my 1979a).
Now I think our oversimple counterfactual analysis succeeds in characterizing one
kind of causation. But other kinds are omitted. We have a sucient, but not a nec-
essary, condition for causation.
For one thing, we usually think that causation is transitive: if C causes D, which
in turn causes E, it follows that C causes E. That is why we can establish causal
connections by tracing causal chains. But we have no guarantee that the relation of
counterfactual dependence will be invariably transitive. (Shortly we shall see how its
transitivity can fail.) So we need to provide for causation not only by direct depen-
78 D. Lewis
dence, but by chains of stepwise dependence. We can do so by dening causation as
the ancestral of dependence (see my 1986b).
But that still does not suce to capture all cases of causation. We have at least
three items of unnished business. Probabilistic causation, preemptive causation, and
causation by, or of, absences are not yet fully covered. Here I shall mostly be dis-
cussing the second and third topics.
1.2 Probabilistic Causation
I have little to say about probabilistic causation. Not because I dont believe in it:
More likely than not, our world is so thoroughly indeterministic that most or all of
the causation that actually takes place is probabilistic. Whether our world is gov-
erned by indeterministic laws is settled neither by the Moorean fact that we make
free choices nor by a priori principles of sucient reason. Rather, it is a contingent
question of theoretical physics. If the best explanation of quantum phenomena
requires spontaneous collapses of the wave function, then we should believe in wide-
spread indeterminism. If Bohmian mechanics is a better explanation, we should be-
lieve that our world is deterministic after all. Either way, there is plenty of causation
in the world. Those who believe in widespread indeterminism still ascribe causal
connections. It would be preposterous to deny that the connections they ascribe
deserve the name they are given. Therefore chancy events can be caused. They can be
caused even when their causes do not make them inevitable, and do not even make
them highly probable. They can be caused even when they have some slight chance
of occurring spontaneously.
The probabilistic counterpart of the simplest sort of counterfactual dependence is,
roughly, probability-raising. C occurs, E has a certain chance (objective single-case
probability) of occurring, and as it happens E does occur; but without C, Es chance
would have been less. Likewise, the more complicated patterns of counterfactual de-
pendence that I shall be discussing later also come in probabilistic versions.
I used to think (substantial) probability-raising could simply take the place of
all-or-nothing counterfactual dependence in an analysis of causation (see my 1986b,
pp. 175180). But there is a problem: Not all probability-raising counts. One terror-
ist places an unreliable bomba genuinely indeterministic deviceon Flight 13;
another terrorist places an unreliable bomb on Flight 17. As it happens, the bomb on
Flight 13 goes o and the bomb on Flight 17 doesnt. The Age runs a headline:
Airline bomb disaster. The headline would have been just the same if it had been
the bomb on Flight 17 that went o, or if it had been both. So the bomb on Flight
17 raised the probability of the headline, but certainly didnt cause it. We want to
say that the raising that counts is the raising of the probability of the causal chain of
events and absences whereby the eect was actually caused. Raising the probability of
some unactualized alternative causal chain leading to the same eect doesnt count.
But it would be circular to say it that way within an analysis of causation. I hope
there is some noncircular way to say much the same thing, but I have none to oer.6
That said, let us set aside the probabilistic case. The proper treatment of causation
in a deterministic world will give us diculties enough. Those same diculties would
reappear for causation under indeterminism, and I hope the same solutions would
apply.
1.3 Preemption Revisited
It sometimes happens that two separate potential causes for a certain eect are both
present; and either one by itself would have been followed by the eect (or at least by
a raised probability thereof ); and so the eect depends on neither. Call any such sit-
uation a case of redundant causation. (For short: redundancy.) Some cases of redun-
dancy are symmetrical: Both candidates have an equal claim to be called causes of
the eect. Nothing, either obvious or hidden, breaks the tie between them. It may be
unclear whether we ought to say that each is a cause or whether we ought to say that
neither is a cause (in which case we can still say that the combination of the two is a
cause). But anyway it is out of the question to say that one is a cause and the other
isnt. Because its unclear what we want to say, these symmetrical cases are not good
test cases for proposed analyses of causation. Set them aside.
Other cases are asymmetrical. Its very clear what we want to say: One of the two
potential causes did cause the eect, the other one didnt. Call the one that did the
causing a preempting cause of the eect. Call the other one a preempted alternative,
or backup.
When our opinions are clear, its incumbent on an analysis of causation to get
them right. This turns out to be a severe test. The simplest sort of deductive-
nomological analysis unks: The preempted alternative is a member of a minimal
jointly sucient set for the eect, yet it is not a cause. The simplest sort of counter-
factual analysis likewise unks: The preempting cause is not a condition without
which the eect would have been absent, yet it is a cause. Both these attempts fail
because they treat the preempting cause and its preempted alternative alike, whereas
we know very well that one is a cause and the other is not. A correct analysis will
need to discern the source of the dierence.
1.4 Trumping
I used to think that all cases of preemption were cases of cutting: cases in which, rst,
there is a completed causal chain (often, but not necessarily, spatiotemporally con-
80 D. Lewis
tinuous) running from the preempting cause all the way to the eect; but, second,
something cuts short the potential alternative causal chain that would, in the absence
of the preempting cause, have run from the preempted alternative to the eect.7
Some think so still, but I have learned better.8
The Sergeant and the Major are shouting orders at the soldiers. The soldiers know
that in case of conict, they must obey the superior ocer. But, as it happens, there
is no conict. Sergeant and Major simultaneously shout Advance!; the soldiers
hear them both; the soldiers advance. Their advancing is redundantly caused: If the
Sergeant had shouted Advance! and the Major had been silent, or if the Major had
shouted Advance! and the Sergeant had been silent, the soldiers would still have
advanced. But the redundancy is asymmetrical: Since the soldiers obey the superior
ocer, they advance because the Major orders them to, not because the Sergeant
does. The Major preempts the Sergeant in causing them to advance. The Major
trumps the Sergeant.
We can speculate that this might be a case of cutting. Maybe when a soldier hears
the Major giving orders, this places a block somewhere in his brain, so that the signal
coming from the Sergeant gets stopped before it gets as far as it would have done if
the Major had been silent and the Sergeant had been obeyed. Maybe so. Or maybe
not. We dont know one way or the other. It is epistemically possible, and hence it is
possible simpliciter, that this is a case of preemption without cutting.
If we forsake everyday examples, we become free to settle by stipulation that we
have no cutting. We can stipulate, for instance, that the causal process in question
works by action at a distance. Nothing goes missing when the process is preempted,
because ex hypothesi there are no intermediate events to go missing. Here is such an
example. Suppose the laws of magic state that what will happen at midnight must
match the rst spell cast on the previous day. The rst spell of the day, as it happens,
is Merlins prince-to-frog spell in the morning. Morgana casts another prince-to-frog
spell in the evening. At midnight the prince turns into a frog. Either spell would have
done the job, had it been the only spell of the day; but Merlins spell was rst, so it
was his spell that caused the transmogrication. Merlins spell trumped Morganas.
Merlins spell was a preempting cause, Morganas was the preempted backup. But
we stipulate also that the causal process from spell to transmogrication has no
intermediate steps.9
1.5 Commonplace Preemption
Trumping shows that preemption does not require the cutting of a causal chain.
Nevertheless, the most familiar variety of preemption does work by cutting. The
causal chain from the preempting cause gets in rst: it runs to completion, and
the eect happens, while the chain from the preempted alternative is still on its way.
The preempted chain is cut. The eect itself is what prevents its nal steps.
Billy and Suzy throw rocks at a bottle. Suzy throws rst, or maybe she throws
harder. Her rock arrives rst. The bottle shatters. When Billys rock gets to where the
bottle used to be, there is nothing there but ying shards of glass. Without Suzys
throw, the impact of Billys rock on the intact bottle would have been one of the nal
steps in the causal chain from Billys throw to the shattering of the bottle. But,
thanks to Suzys preempting throw, that impact never happens.
I used to call such cases as this late preemption. (In hindsight, late cutting
would have been a better name.) I meant to contrast them with early preemption:
easy cases in which we have, if not direct counterfactual dependence of the eect
itself on the preempting cause, at least stepwise dependence. The eect depends on
some intermediate event, which in turn depends upon the preempting cause. (Or we
may have stepwise dependence through a longer chain of intermediates.) These are
cases in which dependence is intransitive, but we get the right answer by dening
causation as the ancestral of dependence.
There is a small industry devoted to solving the preemption problem under the
presupposed premise that preemption always works by cutting (see, e.g., Ram-
achandran 1997a). However well such solutions may (or may not) work in the cases
they were made for, they are not general solutions because they do not deal with
trumping. We may have to rest content with a patchwork of solutions, dierent ones
for dierent cases, but let us hope for something more ambitious.
1.6 Quasi-dependence Rejected
I used to think that late preemption (and maybe early preemption as well) could
be handled by appealing to the intuitive idea that causation is an intrinsic relation
between events10 (except insofar as being subject to such-and-such laws of nature is
an extrinsic matter, as I believe it to be). Take another case, actual or possible, which
is intrinsically just like the case of Suzy throwing her rock at the bottle (and which
occurs under the same laws), but in which Billy and his rock are entirely absent. In
this comparison case, we have a causal chain from Suzys throw to the shattering
which does exhibit counterfactual dependence and which is an intrinsic duplicate of
the actual chain from Suzys throw with Billy present. (Near enough. Doubtless the
presence of Billy and his rock makes some tiny dierence to the gravitational force
on Suzys rock, and therefore some negligible dierence to that rocks trajectory.) I
thought: If being a causal chain is an intrinsic matter, then both or neither of the two
chains that are intrinsic duplicates (and occur under the same laws) must be causal;
but the comparison chain, which exhibits dependence, surely is a causal chain; so the
82 D. Lewis
actual chain, even though thanks to Billy it does not exhibit dependence, must be a
causal chain too. I said that the actual chain exhibited quasi-dependence: it qualied
as causal by courtesy, in virtue of its intrinsic resemblance to the causal chain in the
comparison case.
Quasi-dependence was a bad idea, for ve reasons.
First. Imagine that Suzys and Billys rock-throwing takes place in a world with
laws just a little dierent from what we take to be the laws of our actual world: laws
under which ying objects sometimes make little random jumps. Imagine also that
Suzys rock is an intrinsic duplicate of Billys. Now consider the chain of events
consisting of Billys throw, the ight of Billys rock up to but not including the time
when it reaches the place where the bottle used to be, plus the impact of Suzys
duplicate rock on the bottle and the shattering of the bottle. Compare this chain with
another chain of events in which Suzy is absent, Billy throws, his rock takes a little
jump just before impact, it hits the bottle, and the bottle shatters. The original chain
and the comparison chain are intrinsic duplicates (or near enough) under the same
laws. The comparison chain exhibits counterfactual dependence. But now were
forced to conclude that the shattering quasi-depends on Billys throw! (As well as on
Suzys.) And thats the wrong answer: just as in the original case, Billys throw is not
a cause of the shattering, but rather a preempted alternative.11
Second. The intrinsic character of causation is, at best, a parochial feature of
our own possible world. It does not apply, for instance, to an occasionalist world
in which God is a third party to all causal relationships whatever between natural
events. And yet occasionalism certainly seems to be a genuine possibility. So if we
aim at conceptual analysis, not just a contingent characterization of the causal con-
nections that are found in this world of ours, we cannot assume a priori that causa-
tion is an intrinsic matter.
Third. Quasi-dependence gives the wrong answer in cases of trumping preemption.
The trumped causal chain runs to completion; therefore it is an intrinsic duplicate
(near enough) of an untrumped causal chain in a comparison case (under the same
laws) which exhibits counterfactual dependence. This reinforces our previous con-
clusion that quasi-dependence fails in some other possible worlds, for instance the
world in which Merlins spell trumps Morganas. But worse, it may mean that the
intrinsic character of causation is an overhasty generalization even about the causa-
tion that happens in our own world. It may be, for all we know, that our case of the
soldiers obeying the Major is a trumping case that actually happens.
Fourth. There is another kind of causal connection to which the intuition that
causation is an intrinsic matter does not apply. This is double prevention: a cause
prevents something which, had it not been prevented, would have prevented the
eect. The collision between billiard balls 1 and 2 prevents ball 1 from continuing on
its way and hitting ball 3 (g. 3.1). The collision of 1 and 3, had it occurred, would
have prevented the subsequent collision of balls 3 and 4. But since in fact the colli-
sion of 1 and 3 was prevented, the collision of 3 and 4 was unprevented. That is how
the collision of 1 and 2 causes the collision of 3 and 4. Its a straightforward case of
counterfactual dependence: Without the collision of 1 and 2, the collision of 3 and 4
would not have occurred. But notice that this counterfactual dependence is an ex-
trinsic matter. Had there been some other obstruction that would have stopped ball 1
from hitting ball 3, the collision of 3 and 4 would not have depended on the collision
of 1 and 2. So even in this very ordinary thisworldly case, the causal connection is
extrinsic.
Two more examples. Michael McDermotts: A crazed American President is
about to launch a nuclear attack on Russia; that attack would have provoked a coun-
terstrike, which would have prevented Joe Blow from eating breakfast the following
day. Luckily, the Presidents assistant intervenes to stop the attack. Joe Blows break-
fast depends counterfactually upon that intervention. But the dependence is an ex-
trinsic matter: Had Russia been uninhabited or unarmed, there would have been no
such dependence (McDermott 1995a).
Ned Halls: Billy, the pilot of the escort ghter, shoots down the interceptor that
would otherwise have shot down the bomber. Therefore the successful bombing of
the target depends counterfactually on Billys action. But again the dependence is
extrinsic. If the interceptor had been about to receive a radio order to return to base
3
1
2
4
Figure 3.1
84 D. Lewis
without attacking the bomber, then the successful bombing would not have depended
on Billys action.12
Fifth. Besides the overhasty intuition of the intrinsic character of causation, there
is another presupposition of the method of quasi-dependence that also breaks down
in cases of double prevention. That is the presupposition that we have a chain of
events that runs from the preempting cause to the eect. We need this chain of events
so that we can say what chain of events in the comparison case is its intrinsic dupli-
cate. But when we have causation by double prevention, there is often no continuous
chain of events running from cause to eect. Between the collision of balls 1 and 2
and the collision of balls 3 and 4, or between the intervention by the Presidents
assistant and Joe Blows breakfast, or between the shooting down of the interceptor
and the bombing of the target, nothing much happens. What matters, of course, is
what doesnt happen. Sometimes maybe we can assign denite locations to the pre-
vented intermediates, and thereby locate a chain of events and absences. Sometimes
not. If a preempting cause happens to work by double preventionand, once we
watch for them, cases of double prevention turn out to be very commonand if we
cannot assign any denite location to the relevant absences, there is no saying what
the intrinsic character of the comparison chain is required to match.
Put another way, the method of quasi-dependence breaks down when we have
causation at a distance; and causation at a distance, rather than being the far-fetched
possibility we might have thought it was, turns out to be a feature of commonplace
cases of double prevention. What is far-fetchedthough it may nevertheless turn out
to be the truth about the collapse of spatially spread-out wave functionsis action
at a distance; and that is only one variety of causation at a distance.13,14 If, for
instance, one body exerted a force on a distant body without any eld or any particle
going from one to the other, that would be action at a distance. Our billiard-table
example of double prevention, however, exhibits quite a dierent kind of causation
at a distance.
1.7 Fragility Corrected
There is an obvious solution to cases of late preemption. Doubtless you have been
waiting impatiently for it. Without Suzys preempting rock, the bottle would still
have shattered, thanks to Billys preempted rock. But this would have been a dier-
ent shattering. It would, for instance, have happened a little later. The eect that
actually occurred did depend on Suzys throw. It did not likewise depend on Billys.
Sometimes this solution is just right, and nothing more need be said. Suppose it were
alleged that since we are all mortal, there is no such thing as a cause of death. With-
out the hanging that allegedly caused Ned Kellys death, for instance, he would
sooner or later have died anyway. Yes. But he would have died a dierent death. The
event which actually was Kellys death would never have occurred.
The case of Suzys preempting throw is dierent, however. Its not just that with-
out it the bottle would have shattered somehow, sooner or later. Without it, the
bottle would have shattered at very nearly the same time that it actually did shatter,
in very nearly the same way that it actually did. Yet were usually quite happy to say
that an event might have been slightly delayed, and that it might have diered
somewhat in this or that one of its contingent aspects. I recently postponed a seminar
talk from October to December, doubtless making quite a lot of dierence to the
course of the discussion. But I postponed it instead of canceling it because I wanted
that very event to take place.
So if we say that the shattering of the bottle was caused by Suzys throw, because
without it that very shattering would not have occurred, we are evoking uncom-
monly stringent conditions of occurrence for that event. We are thinking that it
would take only a very slight dierence to destroy that event altogether, and put a
dierent substitute event in its place. We are supposing the shattering to be modally
fragile. This is not something we would normally suppose. We have no business rst
saying as usual that the very same event might have been signicantly delayed and
changed, and then turning around and saying that it is caused by an event without
which it would have been ever so slightly delayed and changed, and then saying that
this is because it takes only a very slight delay or change to turn it into a dierent
event altogether.
How much delay or change (or hastening) do we think it takes to replace an event
by an altogether dierent event, and not just by a dierent version of the same event?
An urgent question, if we want to analyze causation in terms of the dependence of
whether one event occurs on whether another event occurs. Yet once we attend to the
question, we surely see that it has no determinate answer. We just havent made up
our minds; and if we speak in a way that presupposes sometimes one answer and
sometimes another, we are entirely within our linguistic rights. This is itself a big
problem for a counterfactual analysis of causation, quite apart from the problem of
preemption.15
At least, it is a problem so long as we focus on whetherwhether counter-
factual dependence. But there are other kinds of dependence. There is, for instance,
when-on-whether dependence: When one event occurs depends counterfactually on
whether another event occurs. And that is only the beginning. But even this begin-
ning is enough to rehabilitate the obvious solution to late preemption, at least in very
many commonplace cases. Let us by all means agree that Suzys throw caused the
shattering of the bottle because, without her throw, the shattering would have been
86 D. Lewis
slightly delayed. But let us not go on to say that if it had been slightly delayed, that
would have turned it into a dierent event altogether. Let us rather say that Suzys
throw caused the shattering of the bottle in virtue of when-on-whether dependence.
When the shattering occurred depended on Suzys throw. Without Suzys throw, it
would not have occurred exactly when it actually did occur.
L. A. Paul has proposed an emended analysis of causal dependence: event E
depends causally on a distinct actual event C if and only if if C had not occurred,
then E would not have occurred at all or would have occurred later than the time that
it actually did occur (Paul 1998b).16 (Causation itself is the ancestral: C causes E i
there is a chain of such dependencies running from C to E.) This proposal does not
abandon the strategy of fragility, but corrects it. Instead of supposing that the event
itself is fragilewhich would y in the face of much of our ordinary talkwe in-
stead take a tailor-made fragile proposition about that event and its time. The nega-
tion of that fragile proposition is the consequent of our causal counterfactual. Now
we get the right answer to commonplace cases of late preemption. Suzys throw has-
tens the shattering, Billys doesnt. So Suzys throw causes the shattering, Billys
doesnt.
If we stopped here, we would be building into our analysis an asymmetry between
hasteners and delayers. We would be saying that an event without which the same
eect would have happened later is a cause, whereas an event without which the
same eect would have happened earlier is not.17 For that reason, among others, we
should not stop here. We should admit delayers as causes, even when the delayed
event is the very same event that would otherwise have happened earlieror at least,
to acknowledge our indecision about such questions, not clearly not the same event.
Were often ambivalent about the status of delayers. Perhaps that is because a
delayer often works by double prevention. It causes a later version of the event by
preventing an earlier version, which, had it happened, would have prevented the later
version. Then if we ask whether the delayer prevented the event or caused it, and we
overlook the possibility that it might have done both, we have to say prevented
(see Mackie 1992). To restore symmetry between hastening and delaying, we need
only replace the words or would have occurred later than the time that it actually
did occur by the words or would have occurred at a time dierent from the time
that it actually did occur. I favor this further emendation. (As does Paul.) But I
think we should go further still. Whats so special about time? When we thought that
without the actual causes of his death, Ned Kelly would have died a dierent death,
we were thinking not just that he would have died at a dierent time, but also that he
would have died in a dierent manner. According to the uncorrected strategy of fra-
gility, which supposes that events have very stringent conditions of occurrence, a
dierence either of time or of manner would suce to turn the eect into a numeri-
cally dierent event. And if, imitating Pauls correction, we relocate the fragility not
in the event itself but rather in a tailor-made proposition about that event, that will
be a proposition about whether and when and how the eect occurs. We could fur-
ther emend our analysis to require dependence of whether and when and how upon
whether: Without C, E would not have occurred at all, or would have occurred at a
time dierent from the time that it actually did occur, or would have occurred in a
manner dierent from the manner in which it actually did occur. (And we could
redene causation as the ancestral of this new kind of dependence.)
This formulation still distinguishes the case that event E occurs dierently from
the case that E does not occur at all. The distinction has been made to not matter,
but were still presupposing that there is a distinction. If were as indecisive about
such questions as I think we are, it would be better to avoid that presupposition.
Let an alteration of event E be either a very fragile version of E or else a very
fragile alternative event which may be similar to E, but is numerically dierent from
E. One alteration of E is the very fragile version that actually occurs: the unaltered
alteration, so to speak. The rest are unactualized. If you think E is itself very fragile,
you will think that all its unactualized alterations are alternatives, numerically dif-
ferent from E itself. If you think E is not at all fragile, you will think that all its
alterations are dierent versions of one and the same event. Or you might think that
some are alternatives and others are versions. Or you might refuse to have any
opinion one way or the other, and that is the policy I favor. Now we may restate our
current analysis of causal dependence. We can return to whetherwhether counter-
factual dependence, but with alterations of the eect put in place of the event itself:
Without C, the alteration of E which actually did occur would not have occurred.
However indecisive we may be about how fragile an event itself is, its actual alter-
ation is by denition fragile.
Now we say that Suzys throw caused the shattering of the bottle and Billys pre-
empted throw did not because, without Suzys throw, the alteration of the shattering
which actually did occur would not have occurred, and a dierent alteration would
have occurred instead. And here we are considering not only the slight delay before
Billys rock arrived but also any dierences to the shattering that might have been
made because Billys rock diers from Suzys in its mass, its shape, its velocity, its
spin, and its aim point.18
1.8 Spurious Causation
We have dealt with one objection against the fragility strategy: that it conicts with
what we normally think about the conditions of occurrence of events. But there is a
88 D. Lewis
second objection, and it applies as much to the corrected strategy as to the strategy in
its original form. All manner of irrelevant things that we would not ordinarily count
among the causes of the eect can be expected to make some slight dierence to its
time and manner. I once gave this example: If poison enters the bloodstream more
slowly when taken on a full stomach, then the victims death, taken to be fragilewe
might better say, the actual alteration of the victims deathdepends not only on the
poison but also on his dinner.19 If we heed still smaller dierences, almost everything
that precedes an event will be counted among its causes. By the law of universal
gravitation, a distant planet makes some minute dierence to the trajectory of Suzys
rock, thereby making a tiny dierence to the shattering of the bottle. So by adopting
the fragility strategy, whether in corrected or uncorrected form, we open the gate to a
ood of spurious causes.
Among the spurious causes that should have been deemed irrelevant is Billys
rock, the preempted alternative. For one thing, it too exerts a minute gravitational
force on Suzys rock. We wanted to say that (the actual alteration of ) the shattering
depended on Suzys throw and not on Billys, but that turns out to be not quite true.
Wellthese dierences made by spurious causes are negligible, so surely we are
entitled to neglect them? Just as its right to say that a box contains nothing when,
strictly speaking, it contains a little dust, so likewise we are within our linguistic
rights to say that Billys throw made no dierence to the shattering when, strictly
speaking, its gravitational eects made an imperceptibly minute dierence. And if for
some strange reason we chose to attend to these negligible dierences, would we not
then put ourselves in an unusual context where it is right, not wrong, to count all the
things that make negligible dierences as joint causes of the eect?
That would be a sucient reply, I think, but for the fact that sometimes the dif-
ference made by a preempting cause is also minute. Imagine that Suzys throw pre-
cedes Billys by only a very short time; and that the masses, shapes, velocities, spins,
and aim points of the two rocks also dier very little. Then without Suzys throw we
might have had a dierence equal to, or even less than, some of the dierences made
by causes we want to dismiss as spurious.
But even so, and even if Billys rock makes a minute dierence to the shattering by
way of its gravitational eects on Suzys rock, yet Suzys throw may make much
more of a dierence to the eect than Billys. The alteration that would have oc-
curred without Suzys throw, though not very dierent from the actual alteration,
may dier from it in time and manner much more than the alteration that would
have occurred without Billys. Though the dierence made by Billy and the dier-
ence made by Suzy may both count as small by absolute standards, yet the dierence
made by Billy may be small also in comparison to the dierence made by Suzy. That
would be enough to break the symmetry between Suzy and Billy, and to account for
our judgment that Suzys throw and not Billys causes the shattering. We speak of
the asymmetry as if it were all-or-nothing, when really it is a big dierence of degree,
but surely such linguistic laxity is as commonplace as it is blameless.
If, on the other hand, Billys throw does somehow make roughly as much dier-
ence to the eect as Suzys, that is a good reason to judge that Billys throw is not
after all a mere preempted alternative. Rather it is a joint cause of the shattering. So
in this case too we get the right answer.
2 Lecture II
2.1 Alterations of the Cause
Because were so indecisive about the distinction between alterations that are dier-
ent versions of the very same event and alterations that are dierent but similar
events, we ought to make sure that this distinction bears no weight in our analyses.
So far, were obeying that maxim only one-sidedly. The distinction doesnt matter
when applied to the eect, but it still matters when applied to the cause. What it
means to suppose counterfactually that C does not occur depends on where we
draw the line between C not occurring at all and C occurring dierently in time and
manner.
That makes a problem. What is the closest way to actuality for C not to occur?
It is for C to be replaced by a very similar event, one which is almost but not quite C,
one that is just barely over the border that divides versions of C itself from its nearest
alternatives. But if C is taken to be fairly fragile, then if C had not occurred and
almost-C had occurred instead, very likely the eects of almost-C would have been
much the same as the actual eects of C. So our causal counterfactual will not mean
what we thought it meant, and it may well not have the truth value we thought
it had.20 When asked to suppose counterfactually that C does not occur, we dont
really look for the very closest possible world where Cs conditions of occurrence are
not quite satised. Rather, we imagine that C is completely and cleanly excised from
history, leaving behind no fragment or approximation of itself. One repair would be
to rewrite our counterfactual analysis, or add a gloss on its interpretation, in order to
make this explicit (Lewis 1986b, p. 211).
But there is another remedy. We could look at a range of alterations of C, not just
one. As on the side of eects, we need not ever say which of these are versions of C
and which if any are alternatives to C. These alterations may include some in which
C is completely excised, but we need not require this. They may include some which
90 D. Lewis
are almost but not quite C, but nothing is to be said that restricts us to the closest
possible alterations. Then we look at the pattern of counterfactual dependence of
alterations of the eect upon alterations of the cause. Where C and E are distinct
actual events, let us say that C inuences E i there is a substantial range C
1
; C
2
; . . .
of dierent not-too-distant alterations of C (including the actual alteration of C) and
there is a range E
1
; E
2
; . . . of alterations of E, at least some of which dier, such
that if C
1
had occurred, E
1
would have occurred, and if C
2
had occurred, E
2
would
have occurred, and so on. Thus we have a pattern of counterfactual dependence of
whether, when, and how on whether, when, and how. (As before, causation is the
ancestral: C causes E i there is a chain of stepwise inuence from C to E.) Think
of inuence this way. First, you come upon a complicated machine, and you want
to nd out which bits are connected to which others. So you wiggle rst one bit and
then another, and each time you see what else wiggles. Next, you come upon a
complicated arrangement of events in space and time. You cant wiggle an event: it is
where it is in space and time, theres nothing you can do about that. But if you had
an oracle to tell you which counterfactuals were true, you could in a sense wiggle
the events; its just that you have dierent counterfactual situations rather than dif-
ferent successive actual locations. But again, seeing what else wiggles when you
wiggle one or another event tells you which ones are causally connected to which.
A process capable of transmitting a mark, in the sense of Reichenbach and
Salmon, is a good example of inuence (Reichenbach 1928, sections 21 and 43;
Salmon 1994). We have some sort of process extending along a continuous spatio-
temporal path. We can mark the process at one stage, and that mark will persist at
later stages. Or rathersince it is irrelevant whether there is actually anything
around that can make a markif the process were somehow marked at one stage,
that mark would persist at later stages. That is, we have patterns of inuence whereby
alterations of later stages depend counterfactually on alterations of earlier stages.21
The process capable of transmitting a mark might, for instance, be a ow of energy,
matter, momentum, or some other conserved quantity: if there were a little more or
less of the quantity at an early stage, there would be correspondingly more or less of
it at later stages (Fair 1979; Dowe 1992).
But transmission of a mark is only one special case of a pattern of inuence. In
general, we do not require that the alterations of E resemble the alterations of C that
map onto them. Nor do we require that suciently similar alterations of C map onto
similar alterations of E. Nor do we require a process along a spatiotemporally con-
tinuous path; we could have inuence of C upon E even if these were two separated
events with nothing relevant between them. And we do not require a manymany
mapping; the simplest sort of whetherwhether dependence, with only two dierent
alterations of E, still qualies as one sort of pattern of inuence. Recall Halls
example of causation by double prevention: The shooting down of the interceptor
causes the destruction of the target by preventing the shooting down of the bomber.
The shooting down of the interceptor does not much resemble the destruction of the
target; there was no continuous process linking cause and eect; and alterations of
the cause would in some cases have prevented the eect and in some cases not, but in
no case would they have made a (more than negligible) dierence to the eect with-
out preventing it altogether.22
Inuence admits of degree in a rough and multidimensional way. How many dif-
ferent C
i
s are there? How distant are the rest of them from the actual alteration of C,
and from one another? How much do the E
i
s dier from one another: How many
dierent ones are there, and when two of them do dier, how distant (on average, or
at maximum) are they? Plainly there are many ways in which something can be more
of a cause of some eect than something else is, even if it is not an all-or-nothing
dierence of inuence versus no inuence.
Now we are in a better position than before to say that Suzys throw is much more
of a cause of the bottles shattering than Billys. Even if the throws are so much alike
that removing Suzys throw altogether would make little dierence to the shattering,
its still true that altering Suzys throw while holding Billys xed would make a lot
of dierence to the shattering, whereas altering Billys throw while holding Suzys
xed would not. Take an alteration in which Suzys rock is heavier, or she throws a
little sooner, or she aims at the neck of the bottle instead of the side. The shattering
changes correspondingly. Make just the same alterations to Billys preempted throw,
and the shattering is (near enough) unchanged.23
(Although Billys throw does not inuence the shattering, Billys not throwing
before the time of Suzys throw does. This is a typical example of a delaying cause.
The doctors who treated Ned Kellys wounds lest he cheat the hangman by dying
prematurely were the hangmans accomplices: joint causes, along with the judge and
the hangman, of the death Ned actually died. Likewise Billys earlier nonthrow and
Suzys throw were joint causes of the shattering that actually occurred.)
Thanks to this latest emendation of the counterfactual analysis, cases of trumping
are covered along with commonplace preemption. Sergeant and Major both shout
Advance! The soldiers advance. Altering the Majors command while holding the
Sergeants xed, the soldiers response would have been correspondingly altered.
If the Major had said Take cover! they would have taken cover, if he had said
Retreat! they would have retreated, and so on. Altering the Sergeants command
while holding the Majors xed, on the other hand, would have made (near enough)
no dierence at all. If we look only at the whetherwhether dependence of the sol-
92 D. Lewis
diers response on the actual commands of the two ocers, we miss exactly the sort
of counterfactual dependence that breaks the symmetry between the two.24
Likewise for the two wizards. If Merlins rst spell of the day had been not prince-
to-frog, but rather king-to-kangaroo, the transmogrication at midnight would
have been correspondingly altered. Whereas if Morganas trumped spell had been,
say, queen-to-goanna (holding xed Merlins earlier spell and the absence of any
still earlier spell) what happened at midnight would have been exactly the same as it
actually was: The prince would have turned into a frog, and that would have been
all.
(Simon Keller has objected as follows. Suppose the soldiers are not perfectly
obedient, and they know that the Sergeant is better placed than the Major to spot
approaching danger. The Sergeant and the Major both shout Retreat! The soldiers
infer from the Sergeants order that they are in danger, so they retreat. The Ser-
geants order causes the retreat. Yet if the Sergeants order had been anything else,
they would not have inferred danger, so they would have obeyed the Major.Reply:
the soldiers think This is one of those exceptional times when its best to obey the
Sergeant. There is a range of alterations of the Sergeants order, namely the range
of alterations in which this thought is held xed, for which we would have corre-
sponding alterations of the soldiers response. True, if the Sergeants order had been
dierent, this thought would not have been there. But even when its true that if P, it
would not have been that Q, we can still entertain the counterfactual supposition that
P and Q.25 And we have not restricted ourselves to the alterations that are closest to
actuality.)
2.2 Transitivity of Causation
Causation, I previously said, is the ancestral of causal dependence. Event C causes
event E i there is a chain of dependencies running from C to E. That part of
my analysis has remained untouched, even as my denition of causal dependence
evolved from simple whetherwhether dependence between events to a pattern of
inuence. Is it still necessary to take the ancestral? Or does our improved denition
of causal dependence as a pattern of inuence allow us just to identify causation with
dependence?No. Inuence is not invariably transitive. If we want to ensure that
causation is invariably transitive, we still have to take an ancestral.
You might think that intransitivities of inuence could arise from intransitivities of
the counterfactual conditional itself. We know that it can be true that if P, it would
be that Q, and true also that if Q, it would be that R, yet false that if P, it would be
that R (see my 1973b, pp. 3233; Stalnaker 1968). But that is not the problem.
Though counterfactual transitivity itself is fallacious, a closely related inference
pattern is valid: from the premise that if P, it would be that Q, and the premise that if
both P and Q, it would be that R, it does follow that if P, it would be that R (Lewis
1973b, p. 35). Let the counterfactual from C
i
to D
i
be part of a pattern of inuence
of C on D; let the counterfactual from D
i
to E
i
be part of a pattern of inuence of D
on E; then it would seem that if both C
i
and D
i
, it would be that E
i
; so we do indeed
have the counterfactual from C
i
to E
i
, and likewise for the other counterfactuals that
constitute a pattern of inuence of C on E.
The real problem with transitivity is that a pattern of inuence need not map all
the not-too-distant alterations of C onto dierent alterations of D, or all the not-too-
distant alterations of D onto dierent alterations of E. Transitivity of inuence can
fail because of a mismatch between the two patterns of inuence.
In gure 3.2 I picture three possible patterns of inuence of C on E. The rst is nice
and simple: it maps several alterations of C oneone onto alterations of E. But less
nice patterns will still qualify. Let the actual alteration be at the center, and imagine
that distance from the center somehow measures closeness to actuality. (Theres no
need to make this distinction of inner and outer precise. Its only point is to make the
cases easier to picture.) We might have a pattern of inuence that maps the outer
alterations of C oneone onto dierent alterations of E, but funnels all the inner
alterations alike onto a single point (second picture). Or we might have a pattern that
maps the inner alterations of C oneone onto dierent alterations of E, but funnels
all the outer alterations alike onto a single point (third picture).
Now suppose C inuences D by a pattern that funnels all the inner alterations
onto a single point, while D inuences E by a pattern that funnels all the outer
alterations onto a single point (leftmost picture in g. 3.3); or vice versa (middle pic-
ture). Or we might have more complicated cases (rightmost picture). In each case, the
two patterns of inuence that take us from C to D to E are mismatched: The values
of the rst pattern do not coincide with the arguments of the second. So C inuences
D and D inuences E, but C does not inuence E. If we nevertheless want to say
C E C E C E
Figure 3.2
94 D. Lewis
that C causes E, we have to take the ancestral and say that causation outruns direct
inuence.
How might such a case arise? Here is a famous example (from Frankfurt 1969;
see also Heinlein 1951). The Neuroscientist knows exactly how she wants Jones to
behave. She hopes that Jones, left to himself, will behave just as she wants him to. By
reading his brain, she can predict what he will do if left to himself. She reads that he
will do what she wants, so she does nothing further. But if instead she had read that
he would stray from the desired path, she would have taken control. She would have
made him a puppet, manipulating his brain and nervous system directly so as to
produce the desired behavior. The initial state of Joness brain is a preempting cause
of his behavior; the idle Neuroscientist is a preempted backup. The moral of the
story is that preemptive causation, without dependence, suces to confer ownership
and responsibility for ones actions.
Let C be Joness initial brain state; let E be the desired behavior. Consider a time
after the Neuroscientist has read Joness brain, but before she would have seized
control if the reading had been dierent. Let D combine Joness brain state at that
time with the Neuroscientists decision not to intervene. C inuences D. D in turn
inuences E, since at the time of D its too late for the Neuroscientist to intervene. So
we have a two-step chain of inuence from C to D to E. But C does not inuence E:
any alteration of Joness initial brain state would have led to the same behavior in
the end, one way or the other.
The actual alteration of C is the one (assume it to be unique) that leads to exactly
the desired behavior. The actual alteration of E consists of the desired behavior; the
other alterations of E consist of dierent behavior. The actual alteration of D is the
one that leads to the desired behavior, and that includes the Neuroscientists decision
not to intervene. The inner alterations of D are those that would not lead to the
desired behavior, but that include the Neuroscientists decision to intervene in one
or another way. The outer alterations of D are those that would not lead to the
desired behavior, but that nevertheless include the Neuroscientists decision not
to intervene.26 These are arguments of the pattern of inuence from D to E, and
C D E C D E C D E
Figure 3.3
without them it would not be a pattern of inuence at all. But they are not among
the values of the pattern from C to D. The pattern of inuence of C on D maps the
actual alteration of C onto the actual alteration of D, and all other alterations of C
onto inner alterations of D. The pattern of inuence of D upon E maps all the inner
alterations of D onto the actual alteration of E, and the outer alterations of D onto
dierent alterations of E. Feeding the rst pattern into the second, we get a pattern
which maps all alterations of C onto the actual alteration of E. Thus the patterns
are mismatched in the way shown in the rightmost picture in gure 3.3. Transitivity
of inuence fails.
This is an easy case of early preemptionjust the sort of case that my strategy of
taking the ancestral was originally made for. If wed tried to make do without the
ancestral, and get by with inuence alone, it would remain unsolvedprovided that
we insist, as of course we should, that, with no intervention at all by the Neuro-
scientist, Joness initial brain state is indeed a cause of his behavior.
2.3 Transitivity Defended
Some will say that by making causation invariably transitive, our strategy of taking
the ancestral makes more trouble than it cures. It collides with a ock of alleged
counterexamples to transitivity of causation. Thus Ive incurred an obligation to deal
with these examples.
The alleged counterexamples have a common structure, as follows. Imagine a
conict between Black and Red. (It may be a conict between human adversaries, or
between nations, or between gods striving for one or another outcome, or just be-
tween those forces of nature that conduce to one outcome versus those that conduce
to another.) Black makes a move that, if not countered, would have advanced his
cause. Red responds with an eective countermove, which gives Red the victory.
Blacks move causes Reds countermove, Reds countermove causes Reds victory.
But does Blacks move cause Reds victory? Sometimes it seems not.
One of the best known of these BlackRed counterexamples comes from Jonathan
Bennett (1987). Forest re: Let Black be those forces of nature that want the forest
to survive; let Red be those forces of nature that want it to burn. Black protects the
forest from the May lightning by raining all over it in April. Red dries the forest o
again before more lightning comes. The forest burns in June. The April rain caused
there to be an unburnt forest in June, which in turn caused the June re. If causation
is invariably transitive, we must conclude that the rain caused the re.
Two more come from Michael McDermott (1995a). Shock C: Black is Cs friend,
Red is Cs foe. C will be shocked i the two switches are thrown alike. Black, seeing
that Reds switch is initially thrown to the left, throws his switch to the right. Red,
96 D. Lewis
seeing this, responds by throwing his switch also to the right. C is shocked. Blacks
throwing his switch caused Red to throw his switch, which in turn caused C to be
shocked. Thus Blacks attempt to protect C is thwarted. If causation is invariably
transitive, Blacks failed attempt to prevent the shock is actually among the causes of
the shock.
Dog-bite: Red wants to cause an explosion; Black (nature) wants him not to.
Blacks move: a dog bites o right-handed Reds right forenger. Reds counter-
move: with diculty, he uses his left hand to set o the bomb. The bomb explodes.
The dog-bite caused Red to set o the bomb with his left hand, which in turn caused
the explosion. If causation is invariably transitive, the dog-bite was a cause of the
explosion.
Another comes from Hartry Field (unpublished lecture). The bomb outside the
door: Black wants Red dead, so he leaves a bomb outside Reds door. Red nds it
and snus out the fuse. Red survives. Placing the bomb caused Red to snu out the
fuse, which in turn caused Reds survival. If causation is invariably transitive, placing
the bomb was a cause of Reds survival.
Three more examples come from Ned Hall (Two Concepts of Causation). The
deadly double dose: Black endangers Billy by giving him half of the deadly double
dose on Monday. Red counters by withholding the second half on Tuesday. Billy
survives. Mondays dose caused Tuesdays withholding, which in turn caused
Billys survival. If causation is invariably transitive, Mondays dose was a cause of
Billys survival.
The alarm clock: The ringing of the alarm clock summons the Black champion
forth into battle, where he is slain by the Red forces. Without him, Blacks cause
is lost. Red wins. The ringing clock caused the champion to be slain, which in turn
caused Reds victory. If causation is invariably transitive, the ringing clock was a
cause of Reds victory.
The inert network (g. 3.4): Red wants neuron F to re, Black wants it not to.
Since F is extraneously stimulated, it will re unless it is somehow inhibited. Blacks
move: Fire C, which has a stimulatory connection to D, which in turn has a stim-
ulatory connection to E, which in turn has an inhibitory connection to F. Reds
countermove (made in advance): Provide another stimulatory connection from C to
B, which in turn has an inhibitory connection to E. So E doesnt re, F is uninhib-
ited, and F does re. The neural network consisting of C; D; B, and E is inert, so far
as F is concerned; theres no way it could have prevented F from ring. Yet the ring
of C caused the ring of B; which in turn caused the nonring of E, which in turn
caused the ring of F. If causation (including causation by double prevention) is
invariably transitive, then the ring of C was a cause of the ring of F.
My last example is suggested by a familiar saying, Damned if you do, damned if
you dont: Black tries to do what God has commanded, but the Red Devil interferes
so that he messes it up. There aint no justice: God accepts no excuses. So Black
is damned. Blacks failed attempt at pious obedience caused the Devil to interfere,
which in turn caused Black to be damned. If causation is invariably transitive,
Blacks pious conduct caused him to be damned. In all these cases, there are two
causal paths the world could follow, both leading to victory for Red. The two paths
dont quite converge: Victory may come in one way or in another, it may come
sooner or it may come later, but Red wins in the end. Blacks thwarted attempt to
prevent Reds victory is the switch that steers the world onto one path rather than the
other. That is to say, it is because of Blacks move that Reds victory is caused one
way rather than the other. That means, I submit, that in each of these cases, Blacks
move does indeed cause Reds victory. Transitivity succeeds.
That is my considered opinion, but I do admit to feeling some ambivalence. Inso-
far as I can conjure up any inclination to accept the counterexamples, I think my
inclination has three sources, all of them misguided.27
First. In many of these cases Reds victory would have come sooner, or more
directly, without Blacks move. Blacks move prevents Reds victory as well as caus-
ing it: It causes one version but it prevents another. If we thought we had to choose,
we would wrongly infer that since it is a preventer it cannot be a cause. (Weve
already noted this ambivalence in the case of delaying causes generally.)
Second. Moves such as Blacks are in general conducive to victory for Black, not
for Red. If we mix up questions of what is generally conducive to what with ques-
tions of what caused what in this particular case, we may think it just a bit of good
common sense to say that Blacks moves advance Blacks cause, not Reds.28
F
D E
C B
Stimulatory Connection
Inhibitory Connection
Figure 3.4
98 D. Lewis
Third. We note that Blacks move didnt matter; Red would have won all the more
easily without it. The eect doesnt depend on the cause. The idea that causation
requires whetherwhether dependence may retain some grip on us. But if you ever
accept preemptive causation, you must have learned to resist that grip. Why yield to
it now? Its true that Blacks move didnt matter. But thats because the choice Black
faced (whether he knew it or not) was whether to have his defeat caused in one way
or in another; and, either way, Blacks defeat is caused.
In rejecting the counterexamples, and accepting that Blacks move is a cause of
Reds victory, I think I am doing what historians do. They trace causal chains, and,
without more ado, they conclude that what comes at the end of the chain was caused
by what went before. If they did not, they could say little about historical causation;
because, over intervals of any length, historical counterfactuals become so very
speculative that nothing much can be known about the dependence of any event on
its causal ancestors. And every historian knows that actions often have unintended
and unwanted consequences. It would be perfectly ordinary for a move by Black to
backre disastrously.
Ive assumed so far that the BlackRed examples are genuine test cases: We really
do have an event C that causes an event D that in turn causes an event E. But unless
the examples are carefully formulated, perhaps with the aid of somewhat articial
stipulations, that may not be so. It may rather be that C causes D
1
and D
2
causes E;
and D
1
and D
2
are dierent, even though perhaps we may refer to them by the same
name. If so, the example is not a test case, and if it turns out (contrary to my opin-
ion) that C does not cause E, that is no problem for the thesis that causation is
invariably transitive.
D
1
and D
2
might, for instance, be two dierent aspects of the same event: D-qua-
event-of-kind-A and D-qua-event-of-kind-B (see Paul, Aspect Causation, chapter 8
in this volume). Or D
1
and D
2
might be D taken with two dierent contrasts: D-
rather-than-X and D-rather-than-Y (see Maslen, The Context-Dependence of Cau-
sation, chapter 14 in this volume; Hitchcock 1996b). The contrast might be supplied
tacitly by contextual clues, or it might be explicit. I think the aspect proposal and the
contrast proposal dont dier much: The aspect D-qua-event-of-kind-A pretty much
amounts to the contrasted event D-rather-than-a-version-of-D-that-is-not-of-kind-A.
Id suggest that aspects and contrasts alike are best understood as constraints on the
range of relevant alterations.
2.4 Causation by Absences
Alterations, I said, are very fragile events. That was not quite right: Some of them
are absences. Absences can be causes, as when an absence of food causes hunger.
Absences can be eects, as when a vaccination prevents one from catching a disease.
And absences can be among the unactualized alterations of a cause or eect that
gure in a pattern of inuence.
Absences are not events. They are not anything: Where an absence is, there is
nothing relevant there at all.29 Absences are bogus entities. Yet the proposition that
an absence occurs is not bogus. It is a perfectly good negative existential proposition.
And it is by way of just such propositions, and only by way of such propositions,
that absences enter into patterns of counterfactual dependence. Therefore it is safe to
say with the vulgar that there are such entities as absences, even though we know
better. If there is no more beer in the fridge, it is a ction that the beer has been
replaced by something else, something called an absence of beer. We can say that
theres an absence of beer, sure enough; and its part of the ction that this proposi-
tion is made true by the existence of the absence. But the sober truth is rather that
this proposition is true because the proposition that there is some beer is false. That
said, I also insist that the ction is harmless, and we are within our linguistic rights to
indulge in it. Accordingly, I shall carry on make-believedly quantifying over absences
without apology.
(Should we conclude, then, that when we say that absences are causes, really it
is true negative existential propositions that do the causing?No; in other cases we
distinguish between the cause itself and the true proposition that describes it. For
instance, we distinguish the explosion from the proposition that an explosion oc-
curred at so-and-so place and time. The explosion caused the damage; the proposition
is a necessary being, abstract in one sense of that multifariously ambiguous term,
and doesnt cause anything. On absences, as also on the aspects of events, I have
met the friends of fact causation more than halfway; but I refuse to concede that
factstrue propositionsare literally causes.30 So I have to say that when an ab-
sence is a cause or an eect, there is strictly speaking nothing at all that is a cause or
eect. Sometimes causation is not a relation, because a relation needs relata and
sometimes the causal relata go missing [see my Void and Object, chapter 10 in this
volume]. But often, when one genuine event causes another, there are relata, and a
causal relation that holds between them. So if we ignore all causal judgments except
those framed by putting a because between clauses that express propositions, we
overlook part of our topic.)
One reason for an aversion to causation by absences is that if there is any of it at
all, there is a lot of itfar more of it than we would normally want to mention. At
this very moment, we are being kept alive by an absence of nerve gas in the air we
are breathing. The foe of causation by absences owes us an explanation of why we
sometimes do say that an absence caused something. The friend of causation by
100 D. Lewis
absences owes us an explanation of why we sometimes refuse to say that an absence
caused something, even when we have just the right pattern of dependence.31 I think
the friend is much better able to pay his debt than the foe is to pay his. There are ever
so many reasons why it might be inappropriate to say something true. It might be
irrelevant to the conversation, it might convey a false hint, it might be known already
to all concerned, and so on (Grice 1975).
Of course, such reasons for refusing to say whats true are not conned to causa-
tion by absences. Counterfactual analysis of causation?Yeah, yeah, my birth is a
cause of my death! said the scoer. His birth is indeed a cause of his death; but its
understandable that we seldom want to say so. The counterfactual dependence of his
death on his birth is just too obvious to be worth mentioning.
(In case youre tempted to agree with the scoer, consider this comparison of
cases. In actuality there are no gods, or anyway none who pay any heed to the lives
of mere mortals. You are born, and after a while you die. In the unactualized com-
parison case, the gods take a keen interest in human aairs. It has been foretold that
the event of your death, if it occurs, will somehow have a momentous impact on the
heavenly balance of power. It will advance the cause of Hermes, it will be a catas-
trophe for Apollo. Therefore Apollo orders one of his underlings, well ahead of time,
to see to it that this disastrous event never occurs. The underling isnt sure that just
changing the time and manner of your death would suce to avert the catastrophe;
and so decides to prevent your death altogether by preventing your birth. But the
underling bungles the job: you are born, you die, and its just as catastrophic for
Apollo as had been foretold. When the hapless underling is had up on charges of
negligence, surely it would be entirely appropriate for Apollo to complain that
your birth caused your death. And if its appropriate to say, presumably it must be
true. But now we may suppose that, so far as earthly aairs go, actuality and our
unactualized comparison case are alike in every detail. After all, the underling didnt
manage to do anything. We may also suppose that, so far as earthly aairs go, the
two cases are subject to exactly the same laws of nature. So, if you agree with the
scoer that your birth didnt cause your death in actuality, you must think that idle
heavenly dierences can make a dierence to what causes what here below! That is
hard to believe. To be sure, we earlier dismissed the thesis of the intrinsic character of
causation as an overhasty generalization. But here, all we need is that earthly causal
relations supervene on the intrinsic and nomological character of all things earthly.)
As I mentioned previously, Jaegwon Kim has drawn our attention to several
causes of noncausal counterfactual dependence. I said in reply that counterfactual
dependence is causal when it is dependence between entirely distinct events, neither
identical nor overlapping; and that events (or at least, those of them that are causal
relata) must be predominantly intrinsic (see Kim 1973b and my 1986d). Xanthippes
becoming a widow is a particular having of an extrinsic property; so it is not an event
at all (or anyway, it is not a causal relatum), unless it is taken to be identical to,
rather than distinct from, the event of Socrates death.
When we say that absences as well as events can be causes and eects, do Kims
problems reappear? I think not. First, it is hard to see how an absence could be
essentially a having of an extrinsic property. Second, it is safe to say that absences
and genuine events are always distinct from one another. And third, we can say when
two absences are distinct from one another: namely, when the corresponding nega-
tive existential propositions are logically independent.
It doesnt make sense for two distinct absences to dier slightly in detail. When we
have an absence, theres nothing (relevant) there at all, and thats that. So when an
absence is caused, we would expect a pattern of inuence that exhibits funneling to
an unusual degree. We can imagine a device that works in an extraordinarily precise
all-or-nothing fashion; or a Neuroscientist, or some other marvelous being, able to
exert extraordinarily precise and complete control; or we can just imagine a perfectly
ordinary case of prevention. If we then follow that with the funneling that comes
from the presence of a preempted backup, we may well end up with a mismatch be-
tween patterns of inuence in which transitivity of inuence fails. Small wonder,
then, that cases of preemptive preventionpreemptive causing of an absenceand
preemptive double prevention have appeared along with the BlackRed examples in
the debate over transitivity of causation. I say again that at worst we have causation
without direct inuence. I trace a chain; I take the ancestral; I say that when a pre-
empted preventer causes an absence which in turn causes some further event or
absence, then the preempted preventer is a cause of that further event or absence.
Part of what makes preemptive prevention hard, however, is doubt about whether
the absence really does cause anything further. Here is an example, due to Michael
McDermott.32 The elder catches the ball; he causes its absence just beyond his
hand. But a little further along its path there is a walla high, broad, thick, sturdy
wall. Further along still is a window. Does the elder cause the window to remain
unbroken? Does he thereby cause the owner of the window to remain cheerful?
We are ambivalent. We can think: Yesthe elder and the wall between them
prevented the window from being broken; but the wall had nothing to do with it,
since the ball never reached the wall; so it must have been the elder. Or instead we
can think: Nothe wall kept the window safe regardless of what the elder did or
didnt do.
A treatment of the case ought to respect our ambivalence. Rather than endorsing
the Yes or the No, it ought to show how we are within our linguistic rights in
102 D. Lewis
giving either answer. The indeterminacy of our naive judgments is best explained by
invoking some indeterminacy in our analysis. We are in a position to do this.
We have C, the catch. We have D, the absence of the ball from the place just
beyond the elders hand. We have E, the absence of the impact of the ball on the
window, or the nonbreaking of the window, or the continued good cheer of the
owner. Certainly we have a pattern of inuence of C on D. Whether we have inu-
ence of D on E is doubtful. There are alterations of D in which not only is the ball
present beyond the elders hand, but also it is on a trajectory that would take it over
the high wall and down again, or it is moving with energy enough to break through
the wall, and so on. Some of these alterations of D would indeed have led to alter-
ations of E. But are they relevant, not-too-distant, alterations? We may be in a
mood to think so, or we may be in a mood to think not. If we are in a mood to think
them relevant, we should conclude that D causes E, and by transitivity C also causes
E. That is the mood we are in when we are swayed by the thought that the elder
and the wall between them prevented the window from breaking. Whereas if we are
in a mood to think them not relevant, we should conclude that neither D nor C
causes E, and so the question of transitivity from C to D to E does not arise. That is
the mood we are in when we are swayed by the thought that the window was safe
regardless. But if anyone says that D causes E but C doesnt, and concludes that
transitivity fails, he is not stably in one mood or the other.
The Yale shadow puzzle is similar. Two opaque objects are between the sun and
the ground, in such a way that either one without the other would cast exactly the
same shadow. (There might be more than two; and they might even be many slices of
a single thick object.) The upper one is illuminated and stops the sunlight; the lower
one is unilluminated. Does the upper one cast a shadow on the ground? We can
think: Yesbetween them, the two cast a shadow, but the lower one stops no light
because no light ever reaches it, so the upper one must have done the job. Or we can
think: Nothanks to the lower one, the ground would have been shadowed regard-
less of whether the upper one was there or not.33 Again our ambivalence ought to
be respected. We can explain it as before. Consider the absence of light just beyond
the upper object; some far-fetched alterations of this absence would result in light
getting through or around the second object, but we may well be of two minds about
whether those alterations are too distant from actuality to be considered.
Yet another example of preemptive prevention comes from Ned Hall. The bomber
is protected by two escort ghters, piloted respectively by Billy and Hillary. When
the enemy interceptor arrives, Billy shoots it down; but had Billy failed, Hillary
would have succeeded. In either case, the shooting down of the interceptor prevents
the shooting down of the bomber, which, had it happened, would have prevented the
subsequent bombing of the target (Hall, Two Concepts of Causation, chapter 9 in
this volume). What Hall says about this case parallels the thoughts that favored say-
ing that the elder prevented the window from breaking, or that the upper object cast
the shadow on the ground: If Billys action was a cause of the bombing . . . where
Hillary was absent, then so too in this second case, which merely adds an alternative
that plays no active role. Halls view is defensible, provided he is in a mood not
to ignore those far-fetched alterations in which the interceptor succeeds in evading
both Billy and Hillary. But if so, then it is misleading (though literally true) for him
to deny, as he does, that the bombing depends on Billys action. Ignoring the far-
fetched alterations, its false that without Billys action the bombing would have
occurred anyway; whats true is that it might or might not have occurred. Saying that
it would have occurred is equally defensiblebut that calls for a dierent mood, one
in which those far-fetched alterations are ignored.
Notes
1. See, inter alia, D. M. Armstrong, Going through the Open Door Again, in this volume.
2. Or perhaps I feel a pressure on my own body; and perhaps it is analytic that a pressure involves a force,
and that a force is inter alia something apt for causing (or preventing) motion. Then it seems that Im
causally acquainted, if not with causation itself, at least with something conceptually linked to causation.
See Armstrong, The Open Door, this volume; and his (1962), p. 23, and (1997), pp. 211216.
3. The Canberra plan is derived, on one side, from Carnaps ideas about analyticity in a theoretical lan-
guage; and on the other side, from one version of functionalism about mental states. See inter alia Carnap
(1963), pp. 958966; my (1966); and Armstrong (1968). The Canberra plan has been applied to causation
in Tooley (1987), and in Menzies (1996). I discuss Menziess treatment in Void and Object, in this vol-
ume, saying that at best Menziess approach will succeed in dening one central kind of causation.
4. We could subsume the perceptual acquaintance strategy under the Canberra plan. We could take the job
description that species the role occupied by causation to consist almost entirely of platitudes about how
we are perceptually acquainted with causation.
5. An Enquiry Concerning Human Understanding, section VII.
6. For discussion of the problem, see inter alia Menzies (1989a); Woodward (1990); and Schaer (2000a).
Schaers version of the problem resists some of the strategies that might solve other versions.
7. Two points of terminology. Some say overdetermination to cover all sorts of redundancy; I limit it to
the symmetrical cases. Some say preemption to cover only those asymmetrical cases that do involve
cutting; I apply it to all asymmetrical cases.
8. For my former view, see the treatment of preemption in my (1986b), pp. 193212. For a recent claim
that all preemption involves cutting, see Ramachandran (1997a), p. 273: . . . in all genuine causes of
causal pre-emption, . . . the pre-empted processes do not run their full course. . . . All genuine causes, on the
other hand, do seem to run their full course; indeed, they presumably count as genuine precisely because
they do.
9. This example, and the discovery of trumping, are due to Jonathan Schaer. See his Trumping Pre-
emption, in this volume. The case of the soldiers is due to Bas van Fraassen.
10. See my (1986b), pp. 205207. For a similar proposal, see Menzies (1996) and (1999).
11. Here I am pretty much following Paul (1998a).
104 D. Lewis
12. Hall (1994); Hall, Two Concepts of Causation, chapter 9 in this volume.
13. For a misguided conation of causation at a distance with action at a distance, and consequent dis-
missal of causation at a distance as far-fetched, see my (1986b), p. 202.
14. Perhaps action at a distance is production at a distance, where production is one of the varieties of
causation distinguished by Ned Hall in Two Concepts of Causation. Or perhaps it is process-linkage
at a distance, where process linkage is explained as in Schaer (2001).
15. It is a problem that is seldom noted. However, see Bennett (1988), passim.
16. Actually is right, strictly speaking, only if the causal connection in question is set in the actual world.
More generally, E depends causally on C in world W i C and E occur in W and its true in W that
without C, E would not have occurred or would have occurred later than it did in W. But we need not
speak so strictly.
17. For advocacy of just such an asymmetry fact, see Bennett (1987); for reconsideration and rejection
of it, see Bennett (1988), pp. 6972.
18. Here Ive adopted a suggestion made by D. H. Rice at the Oxford Philosophical Society in 1984: If C
1
and C
2
are redundant causes of E, and E would have occurred more or less just as it did if C
2
had not
occurred, but would not have occurred more or less just as it did if C
1
had not occurred, then C
1
is a cause
simpliciter of E and C
2
is not.
19. Lewis (1986b), pp. 198199. Here I am indebted to Ken Kress.
20. See Bennett (1987), pp. 369370. (Bennetts point here is independent of his defense of a hastener-
delayer asymmetry elsewhere in that article.)
21. Salmon abandoned his mark transmission account of causal processes after Nancy Cartwright and
Philip Kitcher convinced him that it would need to be formulated in terms of counterfactuals; see Salmon
(1994).
22. However, some cases of causation by double prevention do exhibit a manymany pattern of inuence.
Suppose a whole squadron of bombers, with ghter escort, are on their way to destroy an extended target
by carpet-bombing; a squadron of interceptors arrives to attack the bombers. In the ensuing dogght, some
of the interceptors are shot down. The remaining interceptors shoot down some of the bombers. The
remaining bombers proceed to their assigned targets. Which parts of the target area get hit depends on
which bombers get through. Thus various alterations of the dogght would lead to various alterations of
the destruction of the target area.
23. Unless you alter Billys throw so much that his rock arrives rst, making Billy the preempting cause. In
a context in which were comparing Billys throw and Suzys, such alterations should be set aside as too
distant. I hope that the vagueness of the analysis at this point matches the vagueness of the analysandum,
and so need not be regretted.
24. Here I am indebted to Ned Hall.
25. See my (1973b), pp. 419, on counterfactuals as variably strict conditionals.
26. But we couldnt have had one of those alterations of D; because it would have to have been produced
by a prior brain state that would have led the Neuroscientist to intervene.True, but so what? We can still
entertain them as counterfactual suppositions, and they can still constitute part of the pattern of inuence
from D to E.
27. In the case of the deadly double dose, an inclination to accept the counterexample may have a fourth
source as well. Hall says that half of the double dose will cure Billys nonfatal illness. So when we are told
(truly, I take it) that Mondays dose causes Billy to survive, were apt to hear a hint that it does so by
curing his nonfatal illness. We too easily mistake the falsity of whats hinted for the falsity of whats actu-
ally said.
28. Compare Lombard (1990), p. 197.
29. Where an absence of spacetime itself is, there is nothing whatever there at all, relevant or otherwise. See
my Void and Object, chapter 10 in this volume, on voids as absences of spacetime, and on obstacles to
the reication of absences.
30. Pace Bennett (1988) and Mellor (1995), pp. 156162. Both Bennett and Mellor are willing to say that
one fact causes another, where a fact either is or corresponds to a true proposition. Mellor does indeed
deny that the two facts stand in a causal relation, where by relation he means a genuine universal exist-
ing in the world. (I deny that too.) But I am unappeased: even if causation is a relation only in some
lightweight, unserious sense, still it shouldnt be said to relate propositions.
31. Helen Beebee, in Causing and Nothingness, chapter 11 in this volume, states just this dilemma, but
chooses the wrong horn of it.
32. The example comes from McDermott (1995a). It is further discussed in Collins, Preemptive Preven-
tion, chapter 4 in this volume. The suggestion that our wavering intuitions are governed by how far-
fetched we nd the possibility of the ball getting past the wall comes from Collins; but I have transplanted
it from Collinss theory of causation as would-be dependence to my theory of causation as inuence.
33. The puzzle was much discussed at Yale circa 1968; see Todes and Daniels (1975). It reappears in
Sorensen (1999).
106 D. Lewis
4Preemptive Prevention
John Collins
As the ball ew toward us I leapt to my left to catch it. But it was you, reacting more
rapidly than I, who caught the ball just in front of the point at which my hand was
poised. Fortunate for us that you took the catch. The ball was headed on a course
that, unimpeded, would have taken it through the glass window of a nearby building.
Your catch prevented the window from being broken.
Or did it? Had you not made the catch, I would have caught the ball instead. My
leaping to catch the ball made your catch redundant. Given my presence, the ball
was never going to hit the window.
The example is a variant of one discussed by Michael McDermott. McDermott
writes:
Suppose I reach out and catch a passing cricket ball. The next thing along in the balls direc-
tion of motion was a solid brick wall. Beyond that was a window. Did my action prevent the
ball hitting the window? (Did it cause the ball to not hit the window?) Nearly everyones initial
intuition is, No, because it wouldnt have hit the window irrespective of whether you had
acted or not. To this I say, If the wall had not been there, and I had not acted, the ball
would have hit the window. So between usme and the wallwe prevented the ball from
hitting the window. Which one of us prevented the ball from hitting the windowme or the
wall (or both together)? And nearly everyone then retracts his initial intuition and says,
Well, it must have been your action that did itthe wall clearly contributed nothing.
(McDermott 1995a, p. 525)
These two cases, mine and McDermotts, are certainly very similar. And what
McDermott says about his example seems to me to be absolutely the right thing to
say about mine. When the redundancy of your catch is raised, when it is pointed out
that, given my presence, the window was never in any real danger of being broken,
I reply (with McDermott) as follows. If neither of us had reached for the ball, then
the ball would have hit the window. So between usyou and mewe prevented the
ball from hitting the window. Which one of us prevented the ball from hitting the
windowyou or I (or both of us together)? Well, clearly it must have been you, for it
was you and not I who took the catch. I contributed nothing.
Preemptive prevention is puzzling. My responses to these two cases dier, despite
the fact that they share the same basic structure. I am happier to agree with what
McDermott says about my version of the example than I am to concur with that
same account of his own version of the story. When I stand between your catch and
the window, I am happy to say that you prevented the window from breaking. It is
far less clear to me that that is the right thing to say when it is not just me but a solid
brick wall that stands as the second line of defense. When it comes to McDermotts
version of the story, I nd myself wanting to say the following things. (1) You pre-
vented the ball from hitting the wall. (2) The wall would have prevented the ball
from breaking the window. (3) Your catch prevented the wall from preventing the
ball from breaking the window. (4) Likewise, the presence of the wall prevented your
catch from preventing the ball from breaking the window. I am very reluctant to say,
in this case, that your catch prevented the ball from breaking the window. Given that
the wall was there, the window was never in any danger of being broken. The pres-
ence of the wall really does seem to make your catch irrelevant. At least, that is how
it seems to me.
Most, but not all, of the people I have asked shared the sense of ambivalence I
experience when contemplating these two cases. Many, but not all, of them agreed
with my judgments about the two examples. I think that these two cases of preemp-
tive prevention pose a genuine puzzle. But perhaps those not convinced that the
presence of the wall makes the catch irrelevant might be persuaded by an even more
extreme variation on the same theme. Consider a third version of the preemptive
prevention story. There is only one potential elder this time, no brick wall, and no
window. Suppose that at the moment you take the catch, the instantaneous velocity
of the ball is a vector directed at a point one hundred million miles from the Earth at
which Halleys comet will be located during its next but one swing through the solar
system. Suppose that the magnitude of the instantaneous velocity is exactly right for
a collision between the ball and the comet at that distant point at that future date (on
the supposition that is, that the Earth and its gravitational eld are absent). Did your
catch prevent the ball from hitting Halleys comet? No.
1 Counterfactual Dependence
Nonoccurrences of events, as well as occurrences, may act as causes and may be pro-
duced as eects. Prevention is a matter of causing the nonoccurrence of an event. To
prevent the window from being broken is to cause the nonoccurrence of a window-
breaking.
There are two ways we might proceed here. We might extend our ontology to in-
clude events of nonoccurrence, or we might continue simply to speak of the non-
occurrence of events. The rst approach leads to some diculties in specifying when
108 J. Collins
two events of nonoccurrence are one and the same event. The second approach re-
quires us to admit that causal relata may be other than events. Since I am dubious
about the ontological status of events of nonoccurrence, I propose to proceed as
follows.
To any possible event e there corresponds the proposition E that the event e occurs.
If e is any possible event, then the propositions E and @E will both be referred to as
propositions of occurrence. Note that I am using the term in such a way as to include
among the propositions of occurrence what might instead have been called proposi-
tions of nonoccurrence; no problem, as long as you understand what I mean. When-
ever C and E are propositions of occurrence, E would be true if C were true, and E
would not have been true had C not been true, we say that E is counterfactually depen-
dent on C. A sequence of true propositions of occurrence is said to be a chain of
counterfactual dependence when each proposition in the sequence is counterfactually
dependent on the proposition that precedes it. According to the simple counterfactual
analysis of causation, a causal chain is just a chain of counterfactual dependence. One
event (or absence) is a cause of another event (or absence) if and only if the proposi-
tion of occurrence for the former is linked to the proposition of occurrence for the
latter by some causal chain. (See Lewis 1973a.)
Suppose that C, D, and E are true propositions of occurrence. We have redundant
causation just in case: (i) E would not have been true had neither C nor D been true;
(ii) E would still have been true if exactly one of the propositions C and D had been
true and the other false. In the particular case where E is a proposition stating that a
particular possible event did not occur, we have a case of redundant prevention.
Preemption is a special kind of redundancy. Following Lewis, let us mark the dis-
tinction in the following way. Suppose that events c and d are redundant causes or
preventers of e. If it is clear, intuitively, that one of the two events c and d caused
or prevented e while the other waited in reserve, then we shall speak of preemption.
Those situations, on the other hand, in which c and d have an equal claim to having
caused or prevented e, will be referred to as cases of overdetermination. The distinc-
tion is a vague one to be sure, yet there are clear cases of preemptive causation and
prevention. The example with which the essay began is one such case.
2 Quasi-dependence
The phenomenon of preemption stands in the way of a simple counterfactual
analysis of causation. When it became clear that not all cases of preemption could
be handled by the simple counterfactual analysis, Lewis introduced the notion of
Preemptive Prevention 109
quasi-dependence. The idea was that whether or not a process is causal ought to de-
pend only on the intrinsic nature of the process itself, and on the laws of nature. If a
duplicate of some process is located in the same world, or in a world with the same
laws of nature, call it an isonomic duplicate. Lewiss suggestion was that any iso-
nomic duplicate of a causal process is a causal process. Call an isonomic duplicate of
a counterfactually dependent chain a quasi-dependent chain. Then, according to the
quasi-dependence analysis of causation, a causal chain is either a chain of counter-
factual dependence or a quasi-dependent chain. (See Lewis 1986b, pp. 193212.)
But the quasi-dependence strategy doesnt work either. There are clear counter-
examples to the claim that every isonomic duplicate of a causal process is a causal
process, that is, to the claim that causation is an intrinsic matter. One kind of coun-
terexample involves cases of causation by double prevention: A cause produces an
eect by preventing an event, which, had it not been prevented, would have prevented
the eect (for details of the examples see McDermott 1995a and Hall 2000). But
another kind of counterexample is provided by the phenomenon of preemptive pre-
vention. Your catch prevents the window from breaking when it preempts my catch,
but not when it merely prevents a collision with a brick wall from preventing the
window breaking. Yet the only dierence between these two cases has to do with
features extrinsic to the simple process involving your catch. The process that
includes the balls ight, your catch, and the windows not breaking is causal in the
case where my hand was poised behind yours to take the catch, but it is not causal in
the case where a brick wall is there instead of me. (If you dont agree with me about
this pair of examples, consider the Halleys comet case instead.) We are assigning a
dierent causal status to processes that are isonomic duplicates. So whether or not a
process is causal is not a matter intrinsic to it.
3 Dependence Prevention
Your catch prevented the window from breaking and also prevented me from catch-
ing the ball when I leapt. But my leap did some preventing as well. It was my leap
that prevented the windows not breaking from being counterfactually dependent on
your catch. My leap was a dependence preventer. I think that the reason we say your
catch prevented the window from breaking, despite the absence of counterfactual
dependence, is that it was only my happening to leap when I did that prevented the
not-breaking from depending on the catch. Sometimes, when e is not counterfactually
dependent on c, we count c as a cause of e when there would be dependence were it
not for some dependence preventer.
110 J. Collins
Of course this is not always so. We might recast the denition of redundancy as
follows: c and d redundantly cause e if and only if (i) c prevents e from being coun-
terfactually dependent on d; and (ii) d prevents e from being counterfactually de-
pendent on c. The redundant causes feature symmetrically in this denition. This
symmetry creates a problem. For just as my leap prevented the not-breaking from
depending on your catch, so did your catch prevent the not-breaking from depending
on my leap. In cases of preemption not only does the preempted backup prevent the
eect from depending on the preempting cause; the preempting cause also prevents
the eect from depending on the preempted backup. Yet we judge the causal roles of
preempter and preempted dierently.
Might we account for this by distinguishing varieties of dependence prevention?
I once hoped so. For there are two ways of preventing counterfactual dependence,
and in many cases the distinction between these two kinds of dependence prevention
matches the contours of our causal judgments. Consider some actual chain of coun-
terfactually dependent events. The rst way in which this chain of counterfactual
dependence might have been prevented is if something had prevented one or more
events in the chain from occurring. If you dont have a chain of actually occurring
events, then you dont have a counterfactually dependent chain of actually occur-
ring events. But an actual chain of events might have been prevented from being a
chain of counterfactual dependence without any event in that chain itself being pre-
vented. The second way of preventing dependence leaves the chain of events xed
and simply removes the counterfactual dependence. Call an event that prevents de-
pendence in this second way a pure dependence preventer.
We may use this distinction to break the symmetry in our example of preemptive
prevention. My leap and your catch were both dependence preventers, but only my
leap was a pure dependence preventer. My action prevented the windows not break-
ing from being counterfactually dependent upon your reaching out to catch the ball,
and it did so without preventing any occurrence (or nonoccurrence) in the sequence
initiated by your reaching out. Your catch also prevented the windows not breaking
from being counterfactually dependent upon my leap, but it did so only by prevent-
ing the occurrence of a key event in the chain that would otherwise have followed
from my action. In particular: your catch prevented me from catching the ball. This
suggests that:
c is a cause of e i there is a chain of counterfactual dependence linking c to e, or
there would be such a chain were it not for some pure dependence preventer.
This needs adjustment. To simplify things in the preceding paragraphs I wrote
as though all of our causal relataall of the elements of our causal chainshad
to be actually occurring events. But as we have already noted, nonoccurrences can
also cause and be caused. A more accurate expression of the above idea would be
couched in terms of propositions of occurrence. Thus:
A causal chain is a chain of true propositions of occurrence that is a chain of
counterfactual dependence or would be were it not for the truth of some proposition
of occurrence.
Then nally we might say that one event c (or the absence of the event c) is a cause
of another event e (or the absence of e) if and only if there is a causal chain linking
the proposition that c occurs (or doesnt occur) with the proposition that e occurs (or
doesnt occur). Think of this as the would-be dependence analysis of causation.1
The would-be dependence analysis handles well many of the hard cases of preemp-
tion that plagued the quasi-dependence theory. It has another nice feature as well.
It allows a solution to our puzzle about preemptive prevention. And the solution
it allows is an attractive one, for it not only enables us to distinguish cases like
mine from McDermotts; it also explains our feeling of ambivalence about such
examples.
I say that the dierence between McDermotts example and mine is this: the coun-
terfactual assumption of the absence of the pure dependence preventer in his story is
more far-fetched than the corresponding assumption of absence in mine. It does not
require much of a stretch to suppose that I simply get my timing slightly wrong, so
that when I leap, I do so at not quite the right moment to be ready to take the catch.
It is more far-fetched, on the other hand, to suppose that the brick wall be absent, or
that the ball would miraculously pass straight though it.
So let us adjust the would-be dependence analysis accordingly:
A causal chain is a chain of true propositions of occurrence that is a chain of
counterfactual dependence or would be were some true proposition of occurrence
false in some not too far-fetched way.
This analysis is vague, since it is a vague matter which alternatives to actuality
should count as being too far-fetched to be taken into account. But this vagueness
is a strength of the analysis, for it properly matches the vagueness of the analy-
sandum. Ambivalent intuitions about examples like McDermotts stem from uncer-
tainty about which suppositions count as being too far-fetched. In the right frame
of mind I might be persuaded to suppose away a solid brick wall. Some competent
speakers are happy to do so, McDermott among them. But fetch mere possibility
from far enough, and none will acquiesce in supposing the dependence preventer
absent. The Earth, its atmosphere, and its gravitational eld prevent the balls non-
112 J. Collins
collision with Halleys comet from depending on your catch. Far-fetched indeed the
supposition that you catch the ball in the Earths absence.
4 Trumping and Inuence
But not every chain that would be dependent were it not for some pure dependence
preventer is a causal chain. That is one of the lessons to be learned from Jonathan
Schaers discovery of trumping preemption (Schaer 2000b). So trumping preemp-
tion is bad news not only for the quasi-dependence theory, but for the would-be de-
pendence account as well. I now think of the latter as the would-have-been theory; I
no longer defend it.
In one of Schaers examples the laws of magic are such that what happens at
midnight is determined by the rst spell cast the previous day. At noon Merlin casts
the rst spell of the day: a spell to turn the prince into a frog. At six that evening
Morgana casts the second (and last) spell of the day: a spell to turn the prince into
a frog. At midnight the prince turns into a frog. The transguration is caused by
Merlin and not Morgana, yet it would be dependent on Morganas spell were it not
for Merlins trumping preemption, and Merlins preempting spell is a pure depen-
dence preventer if we allow that magical spells act without causal intermediaries.
It is tempting to dismiss such fantastic examples. More realistic cases have been
oered, and of these it might be claimed that they owe what plausibility they have
to a confusion of causes with normsthe only sense in which a majors order clearly
trumps a sergeants order rather than preempting it in some other way is a norma-
tive sense. But I think it would be a mistake to ignore the possibility of trumping
preemption on such grounds. We cannot rule out a priori the possibility that some
actual cases of preemption work that way. So trumping preemption must be dealt
with. I shall stick to discussing the fanciful examples here, since they are the cases in
which (by stipulation) the causal structure of the situation is most clear.
David Lewis has responded to the examples of trumping preemption by giving
up on the idea of providing a counterfactual analysis of c is a cause of e. The new
account provides in its place a counterfactual theory of causal inuence. The idea is,
roughly, this. Our original notion of the counterfactual dependence of one event on
another was a notion of whetherwhether dependence; e is dependent on c in this
sense just in case whether or not e occurs depends on whether or not c occurs. But
there are further varieties of dependence. Lewiss suggestion is that we think of
degree of causal inuence as being a matter of the degree to which whether, when,
and how one thing happens depends on whether, when, and how something else
happens (Lewis 2000). Thus the asymmetry in the MerlinMorgana example is
revealed by the fact that Merlin, but not Morgana, has a causal inuence on what
happens at midnight. Had Merlin cast a dierent spell at noon, something dierent
would have happened at midnight, but what happened at midnight was in no way
dependent on whether, when, or how Morgana acted later.
This new theory of causal inuence amounts to a change of topic. Lewis is not
oering a new answer to the old question: what is it for this event to be a cause of
that event? He is rather, oering an answer to a quite new question: what is it for
this event to have a causal inuence on that event? Certainly one event can have a
causal inuence on another without being among its causes. An example is provided
by the story of the Poison and the Pudding (Lewis 1986b, pp. 198199). If a poi-
son kills its victim more slowly and painfully when taken on a full stomach, then the
victims eating pudding before he drinks the poisoned potion has a causal inuence
on his death, since the time and manner of the death depend counterfactually on the
eating of the pudding. Yet the eating of the pudding is not a cause of his death.
The relation of cause and eect does not reduce in any obvious way to degree of
causal inuence. Dependence of the whetherwhether kind appears to play a special
role in our judgments of what is a cause of what. As it stands, the new theory lacks
the resources to identify an events causes from among those things that merely had
some causal inuence on it.
Furthermore, as Jacob Rosen has pointed out to me, there are cases of trumping
preemption that appear to be counterexamples to the necessity of Lewiss account
of causal inuence (Rosen, personal communication, November 1999). Changing
Schaers story slightly, let us suppose that the laws of magic distinguish the powers
of a wizard from those of a witch. As a wizard, Merlins options are very limited in-
deed. There is only one spell he may cast (standard prince-to-frog) and only one time
of day he may cast it (noon). No such limitations apply to Morgana. Let the rest of
the story remain the same. The princes transguration is now in no way dependent
on whether, or when, or how Merlin acts. So according to Lewiss account, Merlins
spell has no causal inuence on what happens at midnight. Yet I would claim, with
Rosen, that it is still Merlins spell, and not Morganas, that causes the prince to turn
into a frog.
5 Masking and Finkish Trumping
Time now to turn our attention to another failed counterfactual analysis. To say that
a glass window is fragile, to say, that is, that the window is disposed to shatter when
114 J. Collins
struck, would seem to be equivalent to saying that if the window were struck, it
would shatter. However, as C. B. Martin noticed long ago, this simple counterfactual
analysis of dispositions is false. For dispositions, like other properties, may be gained
or lost, and the conditions under which a disposition would be lost might be exactly
the dispositions conditions of manifestation. Call such dispositions nkish (Martin
1994).2
Further problems for the simple counterfactual analysis are presented by disposi-
tions that mask other dispositions (see Johnston 1993). To say that I mean addition
when I use the word plus is to ascribe certain dispositions to me; for example, the
disposition to give the sum of two numbers x and y when asked What is x plus y?
But of course it is not true for all values of x and y that were I asked What is x plus
y? I would respond by giving their sum. That is because the disposition that con-
stitutes my meaning addition by plus is masked by various other dispositions of
mine: my disposition to make mistakes, my disposition to decline to answer annoying
questions, my disposition to get old and die. . . .
I say: a masked disposition may fail to manifest itself even though it remains
present and its conditions of manifestation are satised. Some disagree. For example,
David Lewis writes:
. . . the rst problem we face in analysing any particular dispositional concept . . . is the prob-
lem of specifying the stimulus and the response correctly.
We might ohand dene a poison as a substance that is disposed to cause death if ingested.
But that is rough: the specications both of the response and the stimulus stand in need of
various corrections. To take just one of the latter corrections: we should really say if ingested
without its antidote. (Lewis 1999, pp. 144145)
Lewis may well be right about this particular case; it is probably false to say that
the disposition of a poison to kill is masked by its antidote. But I deny the general
claim that Lewiss example is intended to support. For to accept that claim would
be to rob ourselves of one of the most crucial conceptual resources provided by the
notion of a disposition. Some of our most important dispositional concepts are con-
cepts of dispositions that are components of other dispositions. These component
dispositions may have conditions of manifestation that fail to correspond to the an-
tecedent of any appropriate counterfactual conditional. That is how, for example,
we make sense of the dispositions of belief, desire, and meaning as component dis-
positions of an agents total behavioral disposition (Collins 1999). And that is how
we may reply to a Kripkensteinian skepticism about the notion of meaning (Kripke
1982). Kripkes rejection of a dispositional solution to the skeptical paradox presup-
poses the counterfactual analysis of dispositions.
David Lewis has oered a reformed counterfactual analysis of dispositions (Lewis
1999). The reformed theory relies on the thesis that a disposition must be absent
whenever the conditions for manifestation are satised and yet the manifestation fails
to occur. But if masking works the way I claim it does, that thesis is false.
Prior, Pargetter, and Jackson argue that every disposition must have a causal basis
(Prior, Pargetter, and Jackson 1982). Lewis agrees with them. A causal basis for the
windows disposition to shatter when struck is said to be some intrinsic property of
the window that would cause it to shatter were it struck. But if this counterfactual
analysis of the notion of a causal basis is correct, then examples of nkish trumping
establish that not every disposition has a causal basis. Consider an example that in-
volves preemption or mimicking rather than masking. Suppose that the window is
fragile, but that a playful sorcerer watches and waits, determined that if ever the
window is struck he will, quick as a wink, cast a spell to shatter the window. Sup-
pose that the spell would work not by changing any of the intrinsic properties of the
window, nor by cutting any causal chain initiated by the striking and any intrinsic
property of the window, but simply by the trumping preemption of whatever ordi-
nary causal process would lead an unenchanted duplicate of the window to shatter
when struck. Then the window would shatter were it struck, but its shattering would
not be caused by any intrinsic property. So, if the counterfactual analysis of the
notion of a causal basis is correct, the window has the disposition to shatter when
struck, but this disposition has no causal basis.
According to a (partially) reformed counterfactual analysis of what it is for the
window to be fragile:
The window is disposed to shatter when struck if and only if the window has an intrinsic
property B such that, if the window were struck and yet retained B, then the striking and the
windows having property B would jointly cause the window to shatter. (Lewis 1999, pp.
148149)3
If this is correct, then the window in the nkish trumping example is not disposed
to shatter when struck, despite the fact that it is an isonomic duplicate of a window
that is so disposed. Defenders of the reformed counterfactual analysis must deny that
dispositions are an intrinsic matter.
6 Conclusion
When one event prevents a second from preventing a third, this may be either cau-
sation by double prevention, or a case of preemptive prevention. Cases of preemp-
tive prevention deserve to be added to the standard set of hard cases against which
116 J. Collins
theories of causation are tested. The puzzle about preemptive prevention with which
this paper began reveals a new and interesting element of vagueness in our concept of
cause. I believe that this puzzle is best resolved by thinking in terms of dependence
prevention. When the occurrence of one event would have been dependent on
another were it not for some dependence preventer, we are often inclined to take this
as cause and eect, but provided only that it is not too far-fetched to suppose the
dependence preventer absent. Our ambivalent reactions to these cases are due to the
vagueness of too far-fetched.
The possibility of trumping preemption stands in the way of turning this would-be
dependence analysis into a general account of causation. And trumping cases, if
admitted, also lead to new problems for the counterfactual analysis of dispositions.
Discussions of causation and dispositions have, to a great extent, developed inde-
pendently of one another in the philosophical literature. It is time to bring these two
debates together, and time to consider alternatives to the counterfactual theory in
both cases. Our concepts of disposition, cause, and counterfactual dependence are
closely linked, and of these three the notion of a disposition appears to be most basic.
I favor reversing the order of analysis and developing a dispositional account of both
causation and the counterfactual conditional. But this is not the occasion to tell that
story.
Acknowledgments
Earlier versions of this paper were read to the Princeton Philosophy Colloquium on
March 28, 1997 and to my Fall 1999 Metaphysics Seminar at Columbia. Thanks are
due to those audiences, and in particular to Cian Dorr, Ned Hall, David Lewis, Cei
Maslen, Laurie Paul, Jacob Rosen, and Jonathan Schaer.
Notes
1. The material in this section was developed independently in late 1996 and presented in an earlier version
of this paper read to the Princeton Philosophy Colloquium in March 1997. In discussion following that
paper Cian Dorr drew my attention to a version of the would-be dependence analysis presented in
Ganeri, Noordhof, and Ramachandran (1996) and (1998).
2. It should be noted that Martin rst came up with the idea of nkish dispositions in 1957.
3. I am simplifying matters by omitting two features of Lewiss nal proposal, but those features are not
relevant to the point being made here.
5Advertisement for a Sketch of an Outline of a Prototheory of Causation
Stephen Yablo
1 Platos Distinction
A couple of thousand years before Hume made the remark that inspired the counter-
factual theory of causation, Plato said something that bears on the principal prob-
lems for that theory. The idea will seem at rst utterly familiar and of no possible
help to anyone, so please bear with me. What Plato said, or had Socrates say, is that
a distinction needs to be drawn between the cause and that without which the cause
would not be a cause (Phaedo, 98e).
This sounds like the distinction between causes and enabling conditions: condi-
tions that dont produce the eect themselves but create a context in which some-
thing else can do so; conditions in whose absence the something else would not have
been eective. And, indeed, that is what Plato seems to have had in mind. Crito
oers Socrates a chance to escape from prison. Socrates refuses and sends Crito on
his way. The cause of his refusal is his judgment that one should abide by the deci-
sion of a legally constituted court. But it is facts about Socrates body that allow the
judgment to be ecacious: if he had not had this apparatus of bones and sinews
and the rest, he could not follow up on his judgment, but it remains true that it is his
judgment on the question that really determines whether he will sit or run (Taylor
1956, pp. 200201).
Socrates bones and sinews are factors such that if you imagine them away, the
cause (Socrates judgment) ceases to be enough for the eect. Are there conditions
such that the cause ceases to be required for the eect, if you imagine them away?
There seem to be. Consider an example of Hartry Fields.
BOMB: Billy puts a bomb under Suzys chair; later, Suzy notices the bomb and
ees the room; later still, Suzy has a medical checkup (it was already arranged) and
receives from her doctor a glowing report.
Field intends this as a counterexample to transitivity, and so it is. The bomb is a
cause of the eeing is a cause of the glowing report; the bomb is not a cause of
the glowing report. But it is also an example of Platos distinction. Were it not for the
bombs presence, the glowing report would not have hinged on Suzys leaving the
room. The bomb does not help Suzys leaving to suce for the glowing report; rather
it makes Suzys action important, required, indispensable.
Apparently there are two kinds of factors without which the cause would not be a
cause. On the one hand we have enablers: facts G such that Oc & sG LsOe. On
the other we have what might be called ennoblers: facts G such that sOc & G L
sOe. Enablers make a dynamic contribution. They help to bring the eect about.
What an ennobler contributes is just a raising of status. Suzys removing herself from
the room is elevated from something that just happens to something that had to
happen, if Suzy was later going to be healthy.
Plato thinks that factors without which the cause would not be a cause are one
thing, causes another. He presumably, then, would say that enablers and ennoblers
are not to be regarded as causes.
About enablers, at least, it is not clear we should go along with him. If G is an
enabler, then it is a fact in whose absence the eect would not have occurred. And,
although there is some dispute about this, most say that that is good enough for
being a cause. Enablers are full-edged causes; it is just that they are pragmatically
counterindicated in some way. Platos distinction in its cause/enabler form can easily
be rejected. Lets suppose to keep things simple that it is rejected.
Consider now ennoblers. An ennobler contributes by closing o potential routes to
e, that is, all the routes not running through c. This if anything hurts es chances. So
there is no question of confusing an ennobler with a cause. Platos distinction in its
(unintended) cause/ennobler form is real and important. That a potential cause is
disarmed may be a factor in the manner of es occurrence, but it is not a factor in its
occurring as such.
2 Preemptive Causes
The counterfactual theory as handed down from Hume says that if the rst did not,
then the second had not been. For this to be plausible, we need to add that the rst
and second both occur, that the rst is distinct from the second, and so on. But let us
imagine that all of that is somehow taken care of, because our concerns lie elsewhere.
Let the simple counterfactual theory be just what Hume says:
(CF1) c causes e i e depends on c, that is, sOc LsOe
The simple theory cannot be right, because it ignores the possibility of backup
causes that would spring into action if the real cause failed. This is what Lewis used
to call the asymmetric overdetermination problem, and now calls preemption. An
example is
DEFLECT: Hit and Miss both roll bowling balls down the lane. Hits heavier ball deects
Misss lighter ball en route to the pin. Hits throw caused the pin to fall. But there is no de-
pendence since if it had not occurred, the eect would still have happened due to a chain of
events initiated by Misss throw. (Yablo 1986, p. 143)
120 S. Yablo
The problem here is obvious enough that one should probably date the counter-
factual theory, as opposed to the immediately withdrawn counterfactual hunch, to the
moment when it was rst clearly addressed, in Lewiss 1973 paper Causation.
Here is what Lewis says. Notice something about the chain of events initiated by
Misss throw. The Miss-chain was cut o before the Hit-chain had a chance to reach
the pin. It is true that the eect does not depend on the earlier part of Hits chain.
It does, however, depend on the part occurring later, after Misss chain is dead and
buried. And the part after the cuto point depends in turn on Hits throw. So the
eect depends on something that depends on Hits throwwhich suggests that
instead of (CF1) our analysis should be
(CF2) c causes e i there are d
i
such that sOc LsOd
1
& . . . & sOd
n
L
sOe
What (CF2) says is that causation need not be direct; it can be indirect, involving
dependency chains. If the ancestral of a relation R is written R
, it says that causa-

tion is dependence
.
The diagnosis implicit in (CF2) is that preemption arises because we had forgotten
about causal chains. I want to suggest an alternative Platonic diagnosis. Preemp-
tion happens because to take away a cause c is, sometimes, to take away more. It is
to take away the reason it is a cause. It is to take away factors that, although not
themselves causal, contribute to cs causal status by putting e in need of c. e can
hardly be expected to follow c out of existence, if the reasons for its depending on c
disappear rst.
So, look again at DEFLECT. Quoting a former self: if in fact Misss ball never
reaches the pin, then that is an important part of the circumstances. Relative to cir-
cumstances including the fact that Misss ball never makes it, what Hit did was nec-
essary for the pins toppling. If in those circumstances Hit hadnt rolled his ball down
the alley, the pin would have remained standing (Yablo 1986, p. 159). That Misss
ball never touches the pin is a fact that puts the eect in need of Hits throw. It is a
fact in virtue of which Hits throw is a cause. The trouble is that it is a fact put in
place by the throw itself, hence one that nkishly disappears when the relation is
counterfactually tested.
3 Holding Fixed
The diagnosis suggests a repair. If preemption is a matter of something nkishly giv-
ing way, the obvious thought is: Dont let it give way; hold the grounds of the causal
Outline of a Prototheory of Causation 121
connection xed. The test of causation in these cases is not whether e fails if c does,
but whether e fails if c fails with the right things held xed.
By dependence modulo G I will mean dependence with G held xed. This event
depends modulo G on that one i had that one failed to occur in G-type circum-
stances, this one would have failed to occur as well. Letting L
G
stand for depen-
dence modulo G, the suggestion is that
(CF3) c causes e i: for some appropriate G, sOc L
G
sOe.
Actually, of course, this is only an analysis-schema. An analysis would require a
clear, non-causation-presupposing statement of what makes for an appropriate G.
What does make for an appropriate G? Certainly G should ennoble c. But all we
have said about ennoblers is that they are conditions G such that e depends holding-
G-xed on c. As you might guess, and as will be discussed below, this purely formal
requirement can be met by logical trickery almost whatever c and e may be.
Thus where the standard counterfactual theory undergeneratesthe events that
depend on c (or depend
on c) are not all the events it causesthe present theory

has, or is in danger of having, the opposite problem. It may well be that e depends on
each of its causes modulo a suitably chosen G. But this is true also of events that do
not cause e.
4 Triviality/Polarity
There are actually two worries here, one building on the other. The rst is a worry
about trivialization: Everything depends on everything modulo a silly enough G. This
is illustrated by
JUMP: Suppose that e is Bob Beamons jumping 29
0
2 1/2
00
at the 1968 Olympics in
Mexico City. And let c be the burning out of a meteor many light years away. It is
not hard to nd facts modulo which the jump depends on the burnout. First choose
an event on which the jump depends pure and simplesay, Beamons tying his
shoes. Holding it xed that the tying occurs only if the burnout does, without the
burnout Beamon does not make the jump.
Unless some sort of restriction is put on admissible Gs, the requirement of depen-
dence modulo G is a trivial one satised by any pair of events you like. One could
stipulate that G should not be too cockamamie or too ad hoc or too cooked up for
the occasion. But that is hopeless. It is not just that too ad hoc is so vague. Sup-
pose that a standard of naturalness is somehow agreed on. No matter how high the
standard is set, there will be G-dependence without causation.
122 S. Yablo
Consider again BOMB. Certainly the doctors glowing report does not depend
simpliciter on Billys planting the bomb. But it does depend on it modulo the fact
that Suzys chair explodes. Holding the explosion xed, Suzy would not have been
healthy unless she had moved away, which she would not have had she not noticed
the bomb, which would not have been there to notice had it not been put there by
Billy. It seems on the face of it insane to credit Suzys good health to the bomb; she is
healthy despite the bomb, not because of it. And yet her health depends on the bomb
modulo a natural fact. Call that the polarity problem.
5 Stockholm Syndrome
Preemptive causes make themselves indispensable. They create the conditions given
which the eect would not have occurred without them. But there is more than one
way of doing that. The normal way is to produce the eect yourself, thereby pre-
venting other would-be causes from doing the job instead. The eect needs c modulo
the fact G that other avenues to the eect are closed o.
But if you look at our basic conditionthe condition of dependence modulo G
you can see that it supports an almost opposite scenario. E was going to happen
anyway, when c comes along to threaten it: to put its existence in jeopardy. Of
course, putting the eect in jeopardy is not all c does, or it would not even resemble a
cause. It also rescues e from the jeopardy. C threatens e with one hand, and saves it
with the other. The eect needs c to counter the threat G that c itself has launched.
Should e be grateful to c for blocking with one hand a threat it launches with the
other? Of course not. There is a word for that kind of inappropriate gratitude. You
might remember it if I quote from a Web site on the topic: In the summer of 1973,
four hostages were taken in a botched bank robbery at Kreditbanken in Stockholm,
Sweden. At the end of their captivity, six days later, they actively resisted rescue.
They refused to testify against their captors, raised money for their legal defense, and
according to some reports one of the hostages eventually became engaged to one of
her jailed captors. Stockholm Syndrome is the gratitude hostages feel toward cap-
tors who help them with problems brought on by the captivity. To give the present
sort of c causal credit would be the metaphysical equivalent of Stockholm Syndrome.
It is important to be clear about what is being rejected here. There is nothing
wrong with gratitude for actions taken against a threat that has already been
launched: not even if the action is taken by the one who launched it. If your kid-
napper takes pity on you and gets you a Mars bar, there is no requirement of inging
it back in his face. But suppose your kidnapper says, I appreciate that you are
grateful to me for various particular acts of mercy. But still I am hurt. Where is the
thanks I get for the action that occasioned these mercies, that is, the kidnapping?
That remark you should ing back in his face. I draw the following moral:
Dependence modulo G does not make for causation if (i) G is a threat to e that,
although (ii) countered by c, was also (iii) launched by c.
With this in mind, lets go back to the BOMB example. There is nothing wrong
with thanking the bomb for tipping you o to its presence, given that it has already
been planted. But what we are talking about here is gratitude toward the planting
itself. BOMB, then, is an example of Stockholm Syndrome. What may be less obvi-
ous is that JUMP is an example as well.
That there will be no shoe-tying unless the meteor burns out makes e vulnerable
from a new and unexpected direction; events that would cancel the burnout are now
being put in a position to cancel the jump too. (Thats (i).) There would been no such
fact as shoe-tying only if burnout, had the meteor not in fact burned out. (Thats (iii).)
It is the burnout, nally, that stops this fact from carrying out its threat against the
eect. (Thats (ii).)
6 Articial Needs
If e depends on c holding G xed, let us say that G puts e in need of c. Why is this not
enough for causation? The answer is that some needs are trumped up or articial.
This shows up in the fact that among es other needs are some that would, but for
c, have been all its needs. Or, to look at it from the point of view of the fallback
scenariothe closest scenario where c does not occurc is able to meet a need only
by making the eect needier than it had to be, indeed, needier than it would have
been had c failed to occur.
Say that e is Beamons big jump, and c is the burning out of that meteor. What are
the eects needs in the fallback scenario? What would the jump have depended on
had the meteor not burned out? It would have depended on Beamons tying his
shoes; on various earlier jumps whereby he won a place in the nals; on Mexicos bid
for the 1968 Olympics; and so on. Bringing in the burnout does not diminish these
needs one iota. Everything that had to happen before, still has to happen with the
meteor burning out. This is why the burnouts role is articial. It is strictly additional
to events that meet the eects needs all by themselves in its absence.
Now lets try to make this a teeny bit precise. History lets suppose has a branch-
ing time structure. There is the trajectory actually taken through logical space, and
the various branchings-o corresponding to other ways things could have devel-
oped. One branch in particular corresponds to the way things would have developed
124 S. Yablo
if c had not occurred. By the fallback scenario lets mean what happens after the
branching-o point on that alternative branch. By the actual scenario lets mean
what actually does happen after the branching-o point. The eects fallback needs
are the events it depends on in the fallback scenario, that is, the events it would
have depended on, had c not occurred.1 Recalling our subscripting conventions from
above, this can be written
FAN fx j sOx L
F
sOeg,
where F is short for sOc.2 The eects actual needs (for a given choice of G) are
the events it depends on modulo G in the actual scenario. This can be written
GAN fx j sOx L
G
sOeg.3
The need for c is articial i GAN covers FAN with c to spare; or, taking the per-
spective of the fallback scenario, FAN is identical to a subset of GAN not including
c. (I will write this FAN GAN
.) The point either way is that c speaks to a need

that is piled arbitrarily on top of what would, in cs absence, have been all the needs.
7 Counterparts
Imagine that Beamon as a child was initially attracted to chess rather than long
jump. He was waiting to sign up for chess club when someone threw a rock at him. It
was because of the rock incident that he wound up in track. And now here is the in-
teresting part: the rock came from the burnt-out meteor. If not for the burnout, it
would not have been that rock-throwing the eect depended on but a related one (in
which the bully hurled a dierent rock). Since the rock-throwing that e would have
needed is a dierent event from the one it does need, it would seem that GAN does
not cover FAN at all, let alone with c to spare.
I answer that dierent event does not have to mean dierent need. One event can
meet the same need as another, as that need manifests itself in their respective sce-
narios.4 Articiality is still a matter of the eects actual needs subsuming its fallback
needs, so long as we understand this in the following way. Suppose that x is an event
needed in the fallback scenario; one nds in GAN, not perhaps that very event, but
an event meeting the same need (henceforth, a counterpart of x). This complicates
things a little, but not much. Where earlier we required FAN to be identical to a
subset of GAN not including c, now we ask only that it coincide with such a subset,
where sets coincide i their members are counterparts.5 (I will write this FANA
GAN
.)
When do events speak in their respective scenarios to the same need? The idea is
this. Needs that e would have had in cs absence can be paired o with actual needs
in ways that preserve salient features of the case: energy expended, distance traveled,
time taken, place in the larger structure of needs. One wants to preserve as many of
these features as possible, while nding matches for the largest number of needs. One
asks: How much of the fallback structure is embeddable in the actual one? What is
the maximal isomorphic embedding? Events speak to the same need if they are linked
by this embedding.
8 De facto Dependence
It is not enough for causation that a G can be found that puts e in need of c. Causes
must meet real needs, and the need met by c might be trumped up or articial. A fact
G makes the need for c articial i it assigns other needs that would, but for c, have
been all of es needs.
An issue I have nessed until now is how the rst Gthe one that puts e in need of
clines up with the second onethe one that makes the need articial. Suppose we
say of the second G that it enfeebles c, as we said of the rst that it ennobles it.
Does it suce for causation that an ennobler G exists that is not itself an enfeebler?
No, for there is almost always a G like that, namely, the material bicondi-
tional e occurs $ c occurs. That this ennobles c should be clear. That it does not
enfeeble c can be seen as follows. (1) GAN is limited to events x on which c
counterfactually depends. (If sOx LOc,6 then Oc $ Oe LsOx LOc; so by
the export-import law, sOx & Oc $ Oe LOc; so sOx & Oc $ Oe LOe
i sOx & Oc $ Oe & Oc LOe; so sOx & Oc $ Oe LOe; so x is not in
GAN.) (2) FAN is almost certainly not limited to events on which c counterfactually
depends. That c fails to depend on x has no tendency at all to suggest that e would
not have depended on x in cs absence. (1) and (2) make it unlikely that GAN in-
cludes FAN, or hence that G enfeebles c.
Where does this leave us? It is not enough for causation that an ennobler G can be
found that is not itself an enfeebler. It is, I suggest, enough that an ennobler can be
found such that no comparably natural enfeeblers exist. And so I propose a denition
(DD) One event de facto depends on another i some G putting the rst in need of
the second is more natural than any H that makes the need articial,
and I make the following claim
(CF3) c is a cause of e i e de facto depends on c.
126 S. Yablo
It is understood that c and e both occur, that they are suitably distinct, and that
various unnamed other conditions are met; I have in mind the same sorts of extra
conditions as the counterfactual theorist uses. Sometimes (CF3) will be written (DF)
to emphasize that it relies on a new type of dependence, albeit one dened in terms of
counterfactual dependence.
You might have expected me to say that c is a cause i e either depends counter-
factually on c or, failing that, de facto depends on it. That formulation is ne but it
is equivalent to what I did say, for de facto dependence has ordinary counterfactual
dependence as a special case. If e counterfactually depends on c, then it depends on
c modulo the null condition. The null condition is our ennobler, and what needs to
be shown is that there are no comparably natural enfeeblers. But there cannot be
enfeeblers at all, for enfeeblers presuppose fallback needsevents that e depends on
in cs absenceand e does not even occur in cs absence.
9 Triviality and Polarity
One worry we had is that even if c and e are completely unrelated, still e is put in
need of c by the fact that k occurs only if c occurs, where k is an event on which the
eect counterfactually depends.
I say that although this is true, the victory is short-lived, because the very fact of
unrelatedness means that it will be easy to nd an H making the need articial.
Usually we can let H be the null condition. That is, e counterfactually depends out-
right (holding nothing xed) on events that would have been enough in cs absence.
This is just what we would expect if c has, causally speaking, nothing to do with e.
Beamons jump depends on all the same things if the burnout occurs as it would have
depended on absent the burnout.
The need for c is articial i it is over and above what would, but for c, have been
all the needs. An equivalent and perhaps clearer way of putting it is that c must
either meet a fallback needwhich it does if for some f in FAN, c meets the same
need as f or cancel onewhich it does if for some f in FAN, no actual event
meets the same need as f . The need for c is articial i c fails to address any fallback
needs, meaning that it neither meets any fallback needs nor cancels any.
I take it as given that Billys planting of the bomb does not meet any fallback
needs. The question is whether it cancels any. Suppose that Suzy needs to stay
hydrated, or she becomes very sick. She has set her Palm Pilot to remind her at noon
to act on this need. The fallback scenario has her sitting quietly in her chair at noon.
She has a drink of water, water being the one hydrous stu available in the room.
The actual scenario has Suzy catching her breath on the sidewalk when her Palm
Pilot beeps. She eats some Italian ice, that being the one hydrous stu available on
the sidewalk. Any isomorphism worth its salt is going to associate these two events.
The drinking and eating are counterparts; they speak to the same need. One imagines
that the same can be done for all of the eects fallback needs. Anything the glowing
report needed absent the bomb, it still needs. The reason Billys action is not a cause
is that it fails to address any fallback needs.
Suppose that I am wrong about that. Suppose the eects fallback needs are not all
preserved into the actual situation; or suppose they are all preserved but one maps to
the planting of the bomb. Then, I claim, the planting starts to look like a cause.
Case 1: There is an f in FAN such that Billys action meets the same need as f .
Suzy needs exercise or she becomes very sick. She has set her Palm Pilot to remind
her to exercise at 11:45. As things turn out, she doesnt hear the beeping because
she has just spotted a bomb under her chair. Running from the bomb gives her the
needed exercise and so saves her health. If that is how it goes, then Billys planting
the bomb meets the same need as would have been met by Suzys setting her Palm
Pilot. And now we are inclined to reason as follows. Billys planting the bomb meets
the need for an exercise-reminder; the need was not articial because it would have
been there bomb or not; so there is no objection to treating what Billy did as a cause.
Case 2: There is an f in FAN such that no actual event meets the same need as f .
Billys planting the bomb does not in fact meet the same need as Suzys setting her
Palm Pilot. The Palm Pilot, if she had heard it, would have led Suzy to do push-ups,
thus exercising her muscles. The bomb leads her instead to run, thus exercising her
heart and lungs. These are entirely dierent forms of exercise. Either one of them
would have stopped Suzy from getting sick, but the similarity ends there. Now we are
inclined to reason as follows. The eect originally had need of muscle exercise, that
being the only kind of exercise possible in the room. It is relieved of that need by
Billys planting of the bomb; for Suzy now runs, thus exercising her heart and lungs.
So there is no objection to treating what Billy did as a cause. (Analogy: You have a
at tire and need a jack to get back on the road. I can help you either by meeting that
need, or by relieving you of it. I do the rst if I provide you a jack. I do the second if
I bend over and lift the car myself.)
10 Preemption
I say that eects really do depend on their preemptive causes. There is no counter-
factual dependence, because the causality rests on a fact G; and had c not occurred,
128 S. Yablo
that fact would not have obtained. But we can restore the dependence by holding
G xed. I dont know how to argue for this except by going through a bunch of
examples.
Recall DEFLECT. Certainly the eect is put in need of Hits throw by the fact G
that Misss ball never gets close to the pin. It might be thought, though, that the need
was articial.
The eects fallback needs are (lets say) for Misss throw, her balls rolling down
the aisle, and her balls hitting the pin. These needs would seem to recur in the
actual situation as needs for Hits throw, his balls rolling down the aisle, his balls
hitting the pin. If that is how things line up, then Hits throw meets the same need
as was met in the fallback scenario by Misss throw; and so the need it meets is not
articial.
Suppose on the other hand that the fallback needs are held not to recur in the
actual situation. Then articiality is averted through the canceling of needs rather
than the meeting of them. These are intuitive considerations but they suggest that a
fact that makes the need for Hits throw articial will not be easy to nd. I do not
doubt that you could construct one by brute force, but a brute force H will not be as
natural as our existing G, the fact that Misss ball never gets close.
A tradition has arisen of treating early and late preemption as very dierent af-
fairs. But this is for theoretical reasons to do with Lewiss ancestral maneuver, which
works for early preemption but not late; intuitively the two sorts of preemption seem
much on a par. The de facto theory agrees with intuition here. Consider
DIRECT: Hit and Miss both roll balls down the lane. The balls do not come into
contact. Hits ball knocks the pin into the gutter. A moment later, Misss ball
reaches the spot where the pin formerly stood.
Once again, it is part of the circumstances that Misss ball never gets close to the
pin. That no other ball gets close puts the eect in need of Hits throw. It is true that
some H might expose the need as articial. But such an H would have to be con-
structed by brute force. There is no more reason to expect a natural enfeebler in this
case than in the previous one.7
11 Overdetermination
Overdetermination occurs when an eect e depends on two events taken together
without depending on either taken alone; and (what distinguishes it from preemp-
tion) neither can lay claim to being more of a cause than the other. Consider
TOGETHER: Knock and Smack roll their balls at the same time; the balls hit the
pin together and it falls over; either ball alone would have been enough.
It is not hard to nd suitable Gs. The eect depends on Knocks throw, holding
xed the fact G
k
that Smacks ball does not hit the pin unaccompanied, that is, unless
another ball also hits. And it depends on Smacks throw, holding xed the fact G
s
that Knocks ball does not hit the pin unaccompanied.
It is not hard to nd suitable Hs either; indeed we have already found them. G
k
makes the need for Smacks throw articial, and G
s
does the same for Knocks. To
see why, suppose that Knock had not thrown. The eect would have depended on
Smacks throw, the forward motion of his ball, and the like. These events are still
needed in the actual situation, if we hold xed the fact G
s
that Knocks ball does not
hit alone.
Assuming that these are the most natural cause-makers and -breakers to be had,
does the de facto theory call Knocks throw (e.g.) a cause? Is the eect put in need of
it by a fact more natural than any fact making the need articial?
That depends. One reading of more natural is strictly more natural. If that is
what is meant, then neither throw is a cause; each prima facie connection is broken
by a fact exactly as natural as the one that established it. But the phrase could also
be taken weakly, to mean at least as natural as. If, as claimed, the makers and
breakers are the same, then the weak reading makes both throws out to be causes.
True, each occurs under conditions given which the eect takes no notice of it; but
then each also occurs under conditions no less natural given which the eect needs it.
Ties go to the runner on the weak reading, so we have two bona de causes. Our
uncertainty about overdeterminers reects indecision about what to mean by more.
(This is intended less as an explanation of the uncertainty than a rational reconstruc-
tion of it.)
12 Asymmetry
Suppose that c aects not whether e occurs but only when it occurs. Could that be
enough to make c a cause? An example is given by Jonathan Bennett.
RAINDELAY: There was heavy rain in April and electrical storms in the
following two months; and in June the lightning took hold and started a forest re.
If it hadnt been for the heavy rain in April, the forest would have caught re in
May (Bennett 1987, p. 373).
Bennett says that no theory should persuade us that delaying a forests burning
for a month (or indeed a minute) is causing a forest re. . . . And then he points out
130 S. Yablo
something interesting. Although you cannot cause a re by delaying somethings
burning, you can cause a re by hastening somethings burning (ibid.). So, consider
LIGHTNING: There are no rains in April. The re happens in May owing to May
lightning, rather than in June owing to the lightning that strikes then. The lightning
is a cause of the re even though the re would still have occurred without it. That
the time of occurrence would have been later rather than earlier seems to make all
the dierence.
Bennetts examples raise two problems for standard counterfactual accounts. One
is that they cannot explain the asymmetry, that is, why hasteners seem more like
causes than delayers. Also, though, they have trouble explaining why there should
be causation here at all. I assume with Bennett that hasteners bring it about that the
very same event occurs earlier than it would have. If in fact the re would still have
occurred without the lightning, how can the lightning be regarded as a cause?
The form of that question ought to seem pretty familiar. It is the standard pre-
emption question: How can c be a cause, when the eect would have occurred with-
out it thanks to c
0
waiting in the wings? The answer is the same as always: It is a
cause because the eect depends on it modulo a certain fairly natural fact, and
nothing that natural exposes the dependence as fraudulent. It is a part of the cir-
cumstances that the woods do not catch re in June (or later). Holding that xed,
without the May lightning there would not have been a re. The May lightning
causes the re because the re depends on it, holding xed that May is its last
opportunity.
But there is an obvious objection. The eect also fails to occur before a certain
time, and this would seem to obliterate the intended asymmetry. Holding xed the
lack of a re before June, if not for Aprils rain there would not have been a re at
all. June was the window of possibility, and it was the rain that kept the forest going
until that window opened.
The dierence between rain and lightning is not that the rst meets no need;
rather, it has to do with the kind of need involved. Suppose the rain had not fallen,
so that the forest burned in May. Then the things that were done to preserve it from
May until June would not have been required. (The loggers wouldnt have had to go
on strike, the rangers wouldnt have had to apply the ame retardant, and so on.)
That the rain introduces new needs would not be a problem if it addressed some old
ones. But it doesnt. The things that would have been needed for the May re, had
the rain not fallen, continue to be needed as conditions of the June re. (A landslide
late in April threatens to bury the forest under rubble; the June re needs it to change
course just as much as the May re would have.)
Now we see why the rain makes a bad cause. It piles on new needs without can-
celing any old ones. The lightning, by contrast, cancels a whole month of old needs.
The pattern here is typical of the genre. Just by their denition, hasteners are liable to
speak to fallback needs; they reduce the time period over which the eect is in jeop-
ardy and so cancel any needs pertaining to the period that is chopped o. Just by
their denition, delayers often bring about a situation in which the eect needs more
than it would have had the delaying event not occurred. The eect is in jeopardy for
longer and has needs pertaining to the extra time. This is why hasteners tend to be
causes and delayers tend not to be.
13 The Hastener Theory
I have treated hastening as a special case of preemption. One might try the reverse,
assimilating preempters to hasteners (Paul 1998b). A cause is an event in whose ab-
sence e would not have occurred, or at any rate would not have occurred as early as
it did. If we count never occurring as the limiting case of delay, then the claim is that
causes are hasteners, that is, events in whose absence the eect would have been
delayed. One problem for this view is that hasteners are not always causes. Here is an
example due to Hugh Rice (1999, p. 160):
REFLEXES: Slow Joe and Quick-Draw McGraw are shooting at Billy the Kid. Joe
res rst, but since his gun res slower-moving bullets, it is not too late for
McGraw (if he res) to cause the death. And so it happens. McGraw (blest . . .
with super fast reexes) was aware of Joes ring and as a result (wishing to have
the glory of killing Billy for himself ) red a little earlier than he would otherwise
have done. . . . It seems that McGraws ring was a cause of e, but that Joes ring
was not (ibid.). Both shots hasten the death. So both count on the hastener theory
as causes. Intuitively, however, it is McGraws shot that kills Billy.
What does the de facto theory say about this? It is not hard to nd a G modulo
which the death depends on McGraws shot. As the situation in fact develops, Joes
bullet never comes into contact with Billy (it passes untouched through the hole left
by McGraws bullet). Holding that xed, Billys death would not have occurred were
it not for McGraw. This same G also enfeebles Joes shot. Had Joe not red, Billys
death would have depended on McGraws shot, the motion of his bullet, and so
on. Those are its fallback needs. The deaths actual needs are the events on which
it depends holding xed that Joes bullet never made contact. Prima facie it would
seem that the deaths fallback needs are all preserved into the actual scenario: Any-
132 S. Yablo
thing the eect depended on absent Joes shot, it continues to depend on given that
Joes bullet doesnt hit anything.
I said that hasteners tend to reduce needs pertaining to the time period over which
the eect is no longer in jeopardy. That assumes, however, that the counterpart re-
lation puts a lot of emphasis on temporal as opposed to other factors. Oftentimes
other factors will seem just as important, or more important. Suppose that by kicking
a bowling ball already en route to the pin, I get it to arrive more quickly. Ordi-
narily my kick would count as a cause. This time, though, the main threat to the
balls forward motion is from equally spaced gates that open and shut according to a
complicated pattern. The eect occurs only if the ball makes it through each of the
gates. Then we might feel that the eects needs are better conceptualized in terms of
number of gates than number of seconds. To the extent that kicking the ball leaves
its chances with the gates unchanged, the need it meets comes to seem articial.
Certainly the kick seems like less of a cause when it is stipulated that the obstacles
are distributed spatially rather than temporally.
I said that delayers often bring about a situation in which the eect needs strictly
more than it would have, had the delaying event not occurred. The eect is in jeop-
ardy for longer and has needs pertaining to the extra time. But again, this is only a
trend, not a strict rule. Sometimes by putting an eect o for a bit we can cut down
on other, more important needs. Consider a variant of REFLEXES: McGraw is
standing further from Billy than Slow Joe. When Joe sees that McGraw has red, he
res his slower bullet on a trajectory that has it deecting McGraws bullet o to the
side before reaching Billy. Joes ring makes the eect happen later than it would
have, but it is still a cause. Counterparthood is judged in respect not of time but of
dependency relations; Joes ring meets the need that McGraws would have met, or,
on an alternative accounting, it cancels it. It is Joes shot that kills Billy, despite the
fact that Billy lives a little longer because of it.
14 Trumping Preemption
A second recent response to the preemption problem focuses on events causally in-
termediate between c and e. It exploits the fact that, in all the usual cases, e would
have depended on events other than those actual intermediaries had c failed to occur
(Ganeri, Noordhof, and Ramachandran 1998). A third focuses on the manner in
which the eect occurs, if caused by something other than c. There is nothing in the
nature of preemption, though, that requires intermediate events, or that the eects
characteristics should vary according to its cause.
SPELL: Imagine that it is a law of magic that the rst spell cast on a given day [matches] the
enchantment that midnight. Suppose that at noon Merlin casts a spell (the rst that day) to
turn the prince into a frog, that at 6:00 p.m. Morgana casts a spell (the only other that day) to
turn the prince into a frog, and that at midnight the prince becomes a frog. Clearly, Merlins
spell . . . is a cause of the princes becoming a frog and Morganas is not, because the laws
say that the rst spells are the consequential ones. Nevertheless, there is no counterfactual
dependence of the princes becoming a frog on Merlins spell, because Morganas spell is a
dependency-breaking backup. Further, there is neither a failure of intermediary events along
the Morgana process (we may dramatize this by stipulating that spells work directly, without
any intermediaries), nor any would-be dierence in time or manner of the eect absent Mer-
lins spell. . . . Thus nothing remains by which extent [counterfactual accounts of causation]
might distinguish Merlins spell from Morganas in causal status. (Schaer, Trumping Pre-
emption, chapter 2 in this volume, p. 59)
What does our sketch of a prototheory say about this case? First, we should look
for a G such that the eect depends modulo G on Merlins spell. How about the fact
that no one casts a spell before Merlin does? Holding that xed, there would have
been no transformation had Merlin not cast his spell. Perhaps a no less natural H
can be found that enfeebles Merlins spell; I have not been able to think of one. It is
perhaps enough to show that, unlike the other approaches mentioned, the de facto
dependence account is not at an absolute loss here.
15 Switching
A switch is an event that changes the route taken to the eect. It may not be ob-
vious how switching so described goes beyond standard preemption, but consider an
example.
YANK: A trolley is bearing down on a stalled automobile. The car lies 110 yards
ahead on the trackor rather tracks, for just ahead the track splits into two 100-
yard subtracks that reconverge ten yards short of the car. Which subtrack the
trolley takes is controlled by the position of a switch. With the switch in its present
position, the trolley will reach the car via subtrack U (for unoccupied). But Suzy
gives the switch a yank so that the trolley is diverted to subtrack O (for occupied).
It takes subtrack O to the reconvergence point and then crashes into the car.
Certainly the crash does not counterfactually depend on the yank; had Suzy left
the switch alone, the trolley would have taken subtrack U to the car, and the crash
would have occurred as ever. Thus the simple counterfactual theory (CF1) does not
classify the yank as a cause. The ancestralized theory (CF2) sees things dierently;
134 S. Yablo
the eect depends on events that depend on the yankthe trolleys regaining the
main line from track O, for instanceso what Suzy did was a cause. (The verdict
does not change if track O was mined; Suzy was hoping to get the trolley blown up,
and would have succeeded had not the bomb squad arrived.)
What does the de facto theory say? There is no trouble nding a G such that the
crash depends modulo G on the yank. Holding xed that subtrack U is untraveled,
had the switch not been pulled there would have been no way forward; the trolley
would, lets assume, have derailed. The worry is that some comparably natural H
makes the need articial. And, indeed, the null fact makes it articial. Here in the
actual scenario, the eect has need of 100 one-yard motions down track O. Had the
yank not occurred, its needs would have been for 100 one-yard motions down track
U. Because the yank lies apart from what might as well have been all the eects
needs, the de facto theory does not call it a cause.
The de facto theory lets the yank be a cause i it either meets a fallback need or
cancels one. As the case was rst stated, it does neither, but suppose we tweak it a
little. Suppose that O is shorter, or that U was disconnected when Suzy pulled the
switch. Then there are needs the eect would have had that the yank does away with,
and so the role it plays is not entirely articial. Alternatively, suppose the switch
operates not by rearranging the tracks, but by physically grabbing hold of the train
and forcing it away from U and down O. Then the yank does meet a fallback need,
the one that would in its absence have been met by the trains continued momentum.
The door is thus open to the yanks being classied as a cause.
This is a good a place to acknowledge that although technically, everything that e
would have depended on counts as a fallback need,8 in practice not all such needs are
taken equally seriously. Suppose that track U has been disconnected for years, and
heroic eorts are required to x it. That it makes those eorts unnecessary earns the
yank causal credit. But what if the track is constantly reversing itself; it is part of Us
design to connect when it senses an approaching trolley and disconnect when the
trolley is safely past. Then the need that gets canceled may be considered too slight to
protect the yank from charges of articiality. I have no criterion to oer of when a
fallback need is suciently serious that c can escape articiality by canceling it. But
two relevant questions are these: were the eect to fail, what are the chances of its
failing for lack of x? And how counterfactually remote are the scenarios where x is
the culprit? A fallback need may not count for much if it is the last thing one would
think of as the reason why e would fail.
Some have said that an event that makes minor changes in the process leading
to e is not its cause, whereas an event that makes major changes is one. Our theory
agrees, if minor changes are changes whereby all the same needs have to be met.
Consider in this connection an example of Ned Halls, in Causation and the Price of
Transitivity (chapter 7 in this volume, p. 191):
THE KISS: One day, [Billy and Suzy] meet for coee. Instead of greeting Billy with her usual
formal handshake, however, Suzy embraces him and kisses him passionately, confessing that
she is in love with him. Billy is thrilledfor he has long been secretly in love with Suzy, as
well. Much later, as he is giddily walking home, he whistles a certain tune. What would have
happened had she not kissed him? Well, they would have had their usual pleasant coee to-
gether, and afterward Billy would have taken care of various errands, and it just so happens
that in one of the stores he would have visited, he would have heard that very tune, and it
would have stuck in his head, and consequently he would have whistled it on his way home. . . .
But even though there is the failure of counterfactual dependence typical of switching cases (if
Suzy hadnt kissed Billy, he still would have whistled), there is of course no question whatso-
ever that as things stand, the kiss is among the causes of the whistling.
That seems right: The kiss is among the causes of the whistling. But the example
is not really typical of switching cases, or at least, it is missing features present in
pure cases like YANK. The eects fallback needs (its needs absent the kiss) are
heavily weighted toward the period after Billy leaves the coee shop. They include,
for instance, Billys deciding to drop into that particular store, the stores staying
open until he arrives, the playing of that particular tune, and so on. It is because
Suzys kiss relieves the eect of this heavy burden of late-afternoon needs that we are
ready to accept it as a cause.
Notes
1. If it seems odd to think of events as needs, remember that need can mean thing that is needed. (The
dogsled was piled high with our winter needs.) Needs in the ordinary sense do not exist in our system.
Their work is done by events considered under a soon to be introduced counterpart relation, the relation of
meeting-the-same-need-as.
2. According to the export-import law for counterfactuals, A LB LC is equivalent to A & B LC.
This implies that sOx & sOc LsOe, the membership condition for FAN, is equivalent to sOc L
sOx LsOe, which says that e would have depended on x had c not occurred. I assume that the law is
close enough to correct for our purposes, or at least that the indicated consequence is close enough to
correct.
3. FAN and GAN are to be understood as limited to events occurring after the point at which the actual
world and the nearest c-less world begin to diverge.
4. Also, same event does not have to mean same need. An event that meets one need here might meet
another there, or it might meet no need at all.
5. I will be taking counterparthood to be symmetric and oneone. But there might be reasons for relaxing
these requirements. Take rst symmetry. There might be an x in FAN whose closest actual correspondent
meets, not the same need as x, but a bigger need: one with the need met by x as a part. This closest
actual correspondent ought to qualify as a counterpart of x. So, the argument goes, counterparts should be
events meeting at least the same need, which makes counterparthood asymmetric. There is a similar worry
136 S. Yablo
about the oneone requirement. It might take a pair of events to meet the need x meets all by itself in the
fallback scenario; or vice versa. I propose to ignore these complexities.
6. I get from ssOx LsOc to sOx LOc by conditional excluded middle (CEM). CEM is generally
controversial, but it seems in the present context harmless; we are not trying to show that Oc $ Oe is
bound to ennoble c without enfeebling it, but just that this is the likely outcome.
7. What if we change the example so that Misss ball does hit the pin, after it has been knocked down?
Then G should be this: Misss ball never gets close to the pin when it is in an upright position, i.e., when it
is in a condition to be toppled. Holding xed that Misss ball never approaches the pin at any relevant
time, it remains the case that without Hits throw, the pin would not have been knocked over.
8. Remember that attention is limited to events occurring after the branch-point: the point at which the
nearest c-less world begins to depart from actuality.
6Dierence-making in Context
Peter Menzies
1 Introduction
Several approaches to the conceptual analysis of causation are guided by the idea
that a cause is something that makes a dierence to its eects. These approaches seek
to elucidate the concept of causation by explicating the concept of a dierence-maker
in terms of better-understood concepts. There is no better example of such an ap-
proach than David Lewiss analysis of causation, in which he seeks to explain the
concept of a dierence-maker in counterfactual terms. Lewis introduced his coun-
terfactual theory of causation with these words: We think of a cause as something
that makes a dierence, and the dierence it makes must be a dierence from what
would have happened without it. Had it been absent, its eectssome of them, at
least, and usually allwould have been absent as well (Lewis 1986a, pp. 160161).
According to Lewis, a cause c makes a dierence to an eect e in the sense that if the
cause c had not occurred, the eect e would not have occurred either. As we shall see
in section 2, Lewiss theory says there is more to the concept of causation than this
counterfactual condition.
Lewis is on the right track, I think, in saying that we think of a cause as something
that makes a dierence and that this thought is best explicated in terms of counter-
factual concepts. However, I shall argue that the particular way in which Lewis spells
out the concept of a cause as dierence-maker is unsatisfactory. Lewiss articulation
of this concept is distorted by a specic metaphysical assumption: specically, that
causation is an absolute relation, speciable independently of any contextual factors.
The distortion induced by this assumption is reected in the undiscriminating
manner in which his theory generates countless causes for any given eect. However,
commonsense judgment is much more discriminating about causes than is Lewiss
theory. Accordingly, I claim that Lewiss analysis faces the problem of proigate
causes, and I outline some specic problem cases in section 3. In the following section
I argue that Lewiss most recent formulation of his counterfactual analysis (Lewis
2000) faces the same problem of proigate causes, and I also argue that an initially
promising solution to the problem that appeals to pragmatics does not succeed.
The key to solving the problem of proigate causation, I argue, is to give up the
metaphysical assumption that causation is an absolute relation, speciable inde-
pendently of context. In sections 5 and 6 I attempt to analyze the concept of a cause
as a dierence-maker in a way that integrates a certain contextual parameter into the
relevant truth conditions. The analysis employs counterfactual concepts, but ones
that are sensitive to context. I use this account in section 7 to explain the problem
cases of proigate causation, cited in section 3.
I need to note two restrictions that I intend to impose on my discussion. The rst
restriction is that I shall consider only cases of deterministic causation. I shall ignore
cases of probabilistic causation, not because they do not exist, but because they do
not raise any special issues in connection with the problem of proigate causes.
The second restriction is that I shall consider only cases of nonredundant causa-
tion. Any counterfactual rendering of the idea of a cause as a dierence-maker must
address some tricky questions in dealing with redundant causationboth symmetri-
cal cases involving overdetermination by two or more genuine causes and asymmet-
rical cases involving preemption by two or more potential causes only one of which
is genuine. Such examples raise serious questions about the viability of purely coun-
terfactual analyses of causation: They seem to show that a cause need not make a
counterfactual dierence to its eect owing to the presence of alternative causes wait-
ing in the wings. It would take us a long way from our present concerns to determine
whether these examples really show this.1
My aim in this essay is quite limited. I intend merely to explicate the idea of one
conditions making a dierence to another. I shall claim that the condition expressing
this idea is a necessary, though not sucient, condition for causation. To bring the
necessary condition up to suciency further concepts have to be added: Specically,
I would argue, the concept of a process linking cause with eect. How such a concept
is to be added to the dierence-making condition, and whether the condition, so sup-
plemented, is adequate to deal with cases of redundant as well as nonredundant
causation, are questions to be pursued on another occasion. For my immediate pur-
poses, it will suce to have a necessary condition for causation, with which to rule
out the proigate causes generated by Lewiss theory.
2 Lewiss 1973 Theory of Causation
Lewis has presented two counterfactual theories of causation: the original (1973a)
theory and a later (2000) renement of the theory.2 In this section I shall discuss his
conceptual analysis of causation in the context of the earlier (1973a) version of the
theory, deferring consideration of the later version of the theory until section 4.
The way Lewis frames his conceptual analysis is inuenced by a number of meta-
physical assumptions about the causal relation. One of these assumptions, which I
will not contest here, is that causation relates dated, localized events (Lewis 1986a,
pp. 241269). He means to include events, in the ordinary sense, that involve changes:
explosions, battles, conversations, falls, deaths, and so on. But he also means to in-
140 P. Menzies
clude events in a broader sense that do not involve changes: a moving objects con-
tinuing to move, the retention of a trace, the presence of copper in a sample, and so on.
Another metaphysical assumption Lewis makes about causation is that it is an
absolute relationabsolute in the particular sense that it is not relative to any con-
textual parameter and so does not vary in nature from one context to another. It is
this assumption that I wish to contest here. The assumption is not an explicit feature
of his analysis, but is rather a consequence of the way he denes the central concept
of causal dependence:
(1) Where c and e are distinct actual events, e causally depends on c if and only if
e counterfactually depends on c: That is, (i) if c were to occur, e would occur; and
(ii) if c were not to occur, e would not occur.
Lewis actually works with a simpler denition than this. He imposes a centering
principle on the similarity relation governing counterfactuals to the eect that no
world is as similar to the actual world as the actual world is to itself. This principle
implies that the counterfactual above, If c were to occur, e would occur, is auto-
matically true in virtue of the fact that c and e are actual events. So his simpler
working denition is:
(2) Where c and e are distinct actual events, e causally depends on c if and only if
e would not occur if c were not to occur.
The absolute character of causal dependence does not follow from this denition
alone. It is possible, after all, to argue that the counterfactual constructions that de-
ne causal dependence are to be understood in a context-dependent way. This move
is far from implausible in view of the notorious sensitivity of counterfactual con-
structions to contextual factors. But, in fact, Lewis argues that the counterfactuals
that dene causal dependence are to be read as nonbacktracking counterfactuals; and
he species the similarity relation that governs them in terms of a unique, context-
invariant set of weights and priorities for comparing respects of similarity (Lewis
1979a). The absolute, invariant character of the concept of causal dependence stems
ultimately from the absolute, invariant character of the similarity relation for non-
backtracking counterfactuals.
A nal metaphysical assumption that Lewis makes about causation is that it is a
transitive relation. Causal dependence, as dened in terms of counterfactual depen-
dence, is not transitive. To ensure the transitivity of causation, Lewis denes causa-
tion in terms of the ancestral of causal dependence:
(3) Where c and e are distinct actual events, c is a cause of e if and only if there is
a chain of stepwise causal dependences between c and e.
Dierence-making in Context 141
By dening causation in terms of the ancestral of causal dependence, Lewis is also
able to deal with some examples of preemptionthe so-called examples of early pre-
emption. (See Lewis 1986a, pp. 193212.) Though I believe this assumption is false
and can be shown to be unnecessary for the treatment of preemption examples, I
shall not dispute it here.
My present target, to repeat, is the second assumption, to the eect that causation
is an absolute relation; or more precisely, the assumption that the truth conditions
for causal claims can be specied without reference to any contextual factors. Lewis
acknowledges that our causal talk is often selective, focusing on some salient causes
while backgrounding other less salient causes. However, this selectivity is to be ex-
plained by pragmatic principles of conversational exchange, which leave the objec-
tive truth conditions of causal claims untouched.
On Lewiss view, the causal history of an event is a vast, complicated relational
structure: the relata in the structure are events and the relation that structures them is
causal dependence. Out of this vast structure, the human mind may focus selectively
on some part and call it the cause of the given event. Indeed, dierent minds pur-
suing dierent inquiries may focus on dierent parts of this structure. But the prin-
ciples of invidious selection, as Lewis calls them, by which fragments of a causal
history are selected for attention, operate on an already fully determinate causal his-
tory. The principles of selection are independent of the relational structure itself. In
this connection he writes:
The multiplicity of causes and the complexity of causal histories are obscured when we speak,
as we sometimes do, of the cause of something. That suggests that there is only one. . . . If
someone says that the bald tyre was the cause of the crash, another says that the drivers
drunkenness was the cause, and still another says that the cause was the bad upbringing that
made him reckless, I do not think that any of them disagree with me when I say that the causal
history includes all three. They disagree only about which part of the causal history is most
salient for the purposes of some particular enquiry. They may be looking for the most re-
markable part, the most remediable or blameworthy part, the least obvious of the discoverable
parts. . . . Some parts will be salient in some contexts, others in others. Some will not be salient
in any likely context, but they belong to the causal history all the same: the availability of
petrol, the birth of the drivers paternal grandmother, the building of the fatal road, the posi-
tion and velocity of the car a split second before the impact. (Lewis 1986a, pp. 215216)
3 The Problem of Proigate Causation
How satisfactorily does this theory capture the idea of a cause as a dierence-maker?
A fair answer would have to be: Not very well or Well enough, but only with a lot
of auxiliary assumptions about pragmatics. The main problem I wish to highlight
142 P. Menzies
here is the proigate manner in which the theory generates causes for any given eect.
Below I list a number of examples that illustrate this defect of the theory. They are
familiar examples for the most part, but it is useful to have a catalog of them before us.
According to Lewiss theory, any event but for which the eect would not have
occurred is one of its causes. But, as is widely recognized, this generates some absurd
results.
Example 1, The Lung Cancer A person develops lung cancer as a result of years of
smoking. It is true that if he had not smoked he would not have developed cancer. It
is also true that he would not have developed lung cancer if he had not possessed
lungs, or even if he had not been born. But it is absurd to think his possession of
lungs or even his birth caused his lung cancer.
Common sense draws a crucial distinction between causes and background con-
ditions. It ranks the persons possession of lungs and his birth as background con-
ditions, so disqualifying them from being dierence-makers for the eect. Several
philosophers of causation have stressed the importance of this commonsense distinc-
tion in connection with the view of causes as dierence-makers. J. L. Mackie, for
example, says that what we call a cause is what makes the dierence in relation to
some assumed background or causal eld (1974, p. 35).3
Perhaps the most extensive and penetrating investigation of the distinction be-
tween causes and conditions is that of H. L. A. Hart and A. M. Honore in their
seminal work Causation in the Law (1985). They argue that the distinction is relative
to context in two ways. One form of relativity might be called relativity to the context
of occurrence.4 They contrast our causal judgments about the following situations.
Example 2, The Presence of Oxygen If a building is destroyed by re, it may be true
that the re would not have taken hold but for the oxygen in the air, the presence of
combustible material, and the dryness of the building. But these are mere conditions
of the re. On the other hand, if a re breaks out in a laboratory or in a factory,
where special precautions are taken to exclude oxygen during the experiment or
manufacturing process, it would not be absurd to cite the presence of oxygen as a
cause of the re. In both situations it may be true that the re would not have
occurred if oxygen had not been present. (Modied from Hart and Honore 1985,
pp. 3536.)
The second form of context-relativity might be called relativity to the context of
inquiry. With this form, rather than two situations eliciting dierent judgments about
causes and conditions, as in the example above, it is one and the same situation that
elicits dierent judgments depending on the type of inquiry being undertaken. Here
are some of Hart and Honores examples.
Example 3, The Famine and the Ulcerated Stomach The cause of a great famine in
India may be identied by an Indian peasant as the drought, but the World Food
Authority may identify the Indian governments failure to build up reserves as the
cause and the drought as a mere condition. Someone who prepares meals for a
person suering from an ulcerated condition of the stomach might identify eating
parsnips as the cause of his indigestion, but a doctor might identify the ulcerated
condition of the stomach as the cause and the meal as a mere condition. (Modied
from Hart and Honore 1985, pp. 3536.)
Lewiss theory is insensitive to the dierent context-relative ways in which com-
mon sense draws the distinction between causes and conditions. His theory treats
mere conditions as causes because they are factors without which the eect would
not have taken place.5
Hart and Honore argue convincingly, in my view, that the suggestions made by
various philosophers for drawing the commonsense distinction between causes and
conditions are unsatisfactory. They reject J. S. Mills suggestion that the distinction is
the epistemically based distinction between causal factors revealed by investigation
and causal factors known before investigation. They point out that we would count a
dropped cigarette as the cause of a re even when we learn from science, what we
may not have initially known, that the presence of oxygen is among the conditions
required for its occurrence. Hart and Honore also reject R. G. Collingwoods sug-
gestion that the distinction is the practically based distinction between factors con-
trollable by the investigator and factors not so controllable. They argue that the
discovery of the cause of cancer would still be the discovery of the cause, even if it
were useless for the cure of the disease; and that drought is the cause of the failure of
crops and so of famine, and lightning the cause of a re even for those who can do
nothing to prevent them (Hart and Honore 1985, pp. 3437).
Hart and Honore also argue that it is wrong to identify the conditions as the
ordinary course of nature unaected by human intervention. They observe that the
commonsense distinction is often an artifact of human habit, custom, or convention.
Because nature can be harmful unless we intervene, we have developed customary
techniques, procedures, and routines to counteract such harm. These become a sec-
ond nature. For example, the eect of a drought is regularly neutralized by gov-
ernment precautions in conserving water; disease is neutralized by inoculation; rain
by the use of umbrellas. When such procedures are established, the cause of some
harm is often identied as an omission or failure on the part of some agent to carry
out the neutralizing procedures, as the example of the famine illustrates (Hart and
Honore 1985, pp. 3738).
The fact that omissions, absences, and failures are recognized as causes and eects
poses a prima facie diculty for Lewiss theory, which requires causation to link
144 P. Menzies
events. Lewis concedes that an absence is a bogus kind of entity that cannot be
counted as an event. Nonetheless, in partial solution to this diculty, he argues that
the proposition that an absence occurs is not bogus; and since such propositions can
enter into counterfactual dependences, we can talk in a derivative way about causa-
tion by absences.
Lewiss extension of causal relata to include absences exacerbates the problems
already noted. With the inclusion of absences as possible causes of a given eect, the
blurring of the distinction between causes and conditions by Lewiss theory generates
even more counterintuitive results. The next examples are illustrative of the di-
culties Lewiss theory encounters with absences and other nonoccurrences.
Example 4, The Absence of Nerve Gas I am writing this essay at my computer. If,
however, there were nerve gas in the air, or I were attacked with amethrowers, or
struck by a meteor shower, I would not be writing the essay. But it is counterintuitive
to say that the absence of nerve gas, amethrower attack, and meteor strike are
causes of my writing the essay.
Example 5, The Multiple Omissions A healthy plant requires regular watering dur-
ing sweltering hot weather. A gardener whose job it is to water the plant fails to do
so and the plant dies. But for the gardeners omission, the plant would not have died.
On the other hand, if anyone else had watered the plant, it would not have died. But
it seems to absurd to say that the omission of everyone else to water the plant was a
cause of its death.
Common sense draws a distinction between these negative occurrences, ranking
some of them as causes and others as mere conditions. In the rst example, it rates all
the absences as conditions, but in the second example it distinguishes the gardeners
omission from the other omissions, making it a cause. In contrast, Lewiss theory
treats all these nonoccurrences equally as causes.
In summary, the examples cited in this section all point to the counterintuitive
causal judgments licensed by Lewiss theory. They illustrate that his theory, or at
least its truth-conditional component, is too proigate in its attribution of causes
because it does not respect the context-relative way in which common sense distin-
guishes between causes and conditions.
4 A Possible Defense in Terms of Contrastive Explanation
We have so far been considering the original (1973a) version of Lewiss counter-
factual theory. Does the more recent (2000) version of the theory fare any better with
these problematic examples?
The more recent version of the theory also employs a counterfactual rendering
of the idea that a cause makes a dierence to its eect. But the counterfactuals it
employs do not simply state dependences of whether one event occurs on whether
another event occurs. The counterfactuals state dependences of whether, when, and
how one event occurs on whether, when, and how another event occurs. A key idea in
the formulation of these counterfactuals is that of an alteration of an event. This is an
actualized or unactualized event that occurs at a slightly dierent time or in a slightly
dierent manner from the given event. An alteration is, by denition, a very fragile
event that could not occur at a dierent time or in a dierent manner without being
a dierent event. Lewis stipulates that one alteration of an event is the very fragile
version that actually occurs.
The central notion of the new version of the theory is that of inuence:
(4) If c and e are distinct events, c inuences e if and only if there is a substantial
range c
1
; c
2
; . . . of dierent not-too-distant alterations of c (including the actual
alteration of c) and there is a range e
1
; e
2
; . . . of alterations of e, at least some of
which dier, such that if c
1
had occurred, e
1
would have occurred, and if c
2
had
occurred, e
2
would have occurred, and so on.
Where one event inuences another, there is a pattern of counterfactual depen-
dence of whether, when, and how on whether, when, and how. As before, the notion
of causation is dened as an ancestral relation.
(5) c causes e if and only if there is a chain of stepwise inuence from c to e.
This theory is designed to handle cases of redundant causation. For example, con-
sider a standard case of late preemption. Billy and Suzy throw rocks at a bottle. Suzy
throws rst so that her rock arrives rst and shatters the bottle. However, without
Suzys throw, Billys throw would certainly have shattered the glass. Suzys throw is
the preempting cause of the shattered bottle, Billys throw the preempted potential
cause. The revamped theory explains why we take Suzys throw, and not Billys
throw, to be the cause of the shattering of the bottle. If we take an alteration in
which Suzys throw is slightly dierenther rock is lighter or she throws sooner
(while holding Billys throw xed), we nd that the shattering is dierent too. But if
we make the same alterations to Billys throw (while holding Suzys xed), we nd
that the shattering is unchanged (Lewis 2000, p. 191).
The important question for our purposes is whether this new version of the theory
helps to deal with any of the problematic examples above. I cannot see that it does.
The theory might try to explain the distinction between causes and conditions by
showing that eects are sensitive to alterations in causes in a way that they are not
146 P. Menzies
sensitive to alterations in conditions. But this does not seem to be the case. In the
lung cancer example, for instance, altering the mans possession of lungs will change
the way his lung cancer develops as surely as altering his habits of smoking. Eects
are sensitive to alterations in conditions as much as to alterations in causes. So the
theory does not do anything to explain the distinction between causes and conditions.
Indeed, Lewis concedes as much himself. He says that the new version of the
theory, as much as the old version, appears to generate too many causes. Any pres-
ence or absence linked by a pattern of inuence, or a chain of such patterns, will
count as a cause, although this seems to go against how we think and speak of
causes. However, he oers a defense of his theory in terms of Grices (1975) prag-
matic theory of conversational implicature. It is literally true that any presence or
absence linked to an event in the way distinguished by his theory is a cause of that
event, but it is not always conversationally appropriate to mention it as a cause. He
writes: There are ever so many reasons why it might be inappropriate to say some-
thing true. It might be irrelevant to the conversation, it might convey a false hint, it
might be known already to all concerned . . . (this volume, p. 101).
Lewis belongs to a long tradition of philosophers who have tried to isolate objec-
tive truth conditions for causal statements from pragmatic considerations of context.
J. S. Mill famously claimed that the only objective sense of cause is that of the total
cause of some eect. He dismissed the context-dependent way in which we ordi-
narily talk of some partial conditions as causes and others as conditions.6 Of course,
Lewiss position is slightly more subtle than Mills. Whereas Mill dismissed our or-
dinary talk as unsystematic and muddled, Lewis gestures at the outlines of a possible
explanation in the form of Grices maxims of conversational exchange.
However, Lewis provides scant detail of the way Grices maxims are meant to
apply to particular examples. Which maxims are relevant? How are they to be em-
ployed? There is, moreover, a question whether Grices principles are especially well
suited to explaining the specic causal judgments in question. Grices maxims of
conversation are very general principles of rationality applied to information ex-
change. Yet the principles that lie behind our judgments about the examples of the
last section seem to be particular to causal judgments. As general principles of ratio-
nal information exchange, Grices maxims miss out on these particular causation-
specic principles.
What are these causation-specic principles? Several philosophers investigating the
pragmatics of causal explanation have stressed the importance of using contrastive
why-questions to analyze the interest-relativity of causal explanation. (See, in par-
ticular, Garnkel 1981 and van Fraassen 1980, chapter 5.) They have argued that
seeing causal explanations as answers to contrastive why-questions aords a way of
understanding how context lters out from the vast causal history of an event those
causes that are salient for certain explanatory concerns. For example, an enormous
range of causal factors can be cited in explanation of a particular eclipse of the
moon. Nonetheless, we can restrict attention to certain kinds of causal factors by
seeking explanations of specic contrasts: why did the eclipse occur rather than not
occur? Why was it partial rather than complete? Why did it last two hours rather
than some other interval of time? Dierent contexts can be seen as implicitly request-
ing explanations of these dierent contrasts. This work on contrastive explanation
suggests a strategy for explaining our causal judgments about the examples of the
last section: Preserve Lewiss counterfactual analysis of causation, but add to it an
account of contrastive explanation that can explain the context-sensitivity of ordi-
nary causal discourse.7
There are several accounts of contrastive explanation available. Which should we
use? Lewis has developed one account, which, as it happens, is tailormade for our
purposes, though it must be noted that Lewis himself does not envisage the applica-
tions to which we shall put it. He writes in his paper Causal Explanation:
One way to indicate what sort of explanatory information is wanted is through the use of
contrastive why-questions. Sometimes there is an explicit rather than. . . . Then what is
wanted is information about the causal history of the explanandum event, not including in-
formation that would also have applied to the causal histories of alternative events, of the sorts
indicated, if one of them had taken place instead. In other words, information is requested
about the dierence between the actualised causal history of the explanandum and the unac-
tualised causal histories of its unactualised alternatives. Why did I visit Melbourne in 1979,
rather than Oxford or Uppsala or Wellington? Because Monash invited me. That is part of the
causal history of my visiting Melbourne; and if I had gone to one of the other places instead,
presumably that would not have been part of the causal history of my going there. It would
have been wrong to answer: Because I like going to places with good friends, good philosophy,
cool weather, and plenty of trains. That liking is also part of the causal history of my visiting
Melbourne, but it would equally have been part of the causal history of my visiting any of the
other places, had I done so. (Lewis 1986a, pp. 229230)
On this account, we explain why an event e rather than an event e
occurred by
giving information about the actual causal history of e that dierentiates it from the
counterfactual causal history of e
.
Can the strategy of conjoining this account with the counterfactual analysis help
to explain our intuitive judgments about the examples of section 3? Let us be clear
about what the strategy is in the rst place. It involves two auxiliary assumptions.
The rst is the assumption that every ordinary causal statement can be seen as a re-
sponse to an implicit contrastive why-question. The second is the assumption that
our ordinary talk about causes is to be explained in terms of the way contrastive
148 P. Menzies
why-questions selectively lter out from the objective causes, delivered by the coun-
terfactual analysis, those relevant in particular contexts.
Let us consider how well this strategy works by seeing how it applies to the ex-
amples of section 3. It must be said that it works surprisingly well with some of them.
For example, it explains reasonably well why we do not consider it appropriate to
say about the lung cancer example that the mans birth and his possession of lungs
were causes of his lung cancer. It is natural to assume that such causal statements
would be attempts to answer the why-question Why did the man get lung cancer
rather than not? However, the mans birth and possession of lungs fail to be objec-
tive causes that are present in the actual causal history of his lung cancer but absent
from the counterfactual causal history of his not getting lung cancer. If the man had
not developed lung cancer, it would still plausibly be part of his causal history that he
was born and possessed lungs.
The strategy also provides a convincing explanation of the relativity of our causal
judgments about causes and conditions to the context of inquiry. For instance, in the
ulcerated stomach example, the meal-preparer and the doctor make dierent judg-
ments about causes and conditions because they address dierent contrastive why-
questions. The meal-preparer is addressing the question Why did the person get
indigestion on this occasion rather than some other? The persons ulcerated stomach
is a condition that is present in both the actual history and the counterfactual his-
tories and so is disqualied from counting as a factor that dierentiates between
them. On the other hand, the doctor is addressing the question Why does this person
rather than other people get indigestion? Here what the person ate is a factor com-
mon to his causal history and the causal histories of other people, while his ulcerated
stomach condition is a factor that dierentiates them.
Without doubt, these explanations of our commonsense judgments have a ring of
plausibility to them. Nonetheless, I think they cannot be the complete story, as the
principles they rely on have some major gaps or inadequacies.
First, Lewiss account of contrastive explanation relies on backtracking counter-
factuals. We have to be able to work out whether some objective cause would be
present or absent from the history that would have had to occur if some alternative
to the actual eect had occurred. But the principles that guide the reasoning behind
such backtracking counterfactuals have not been formulated. For example, to get the
right results in the lung cancer example we have to infer that that the person would
have been born and possessed lungs even if he had not developed lung cancer. And
to get the correct answer in the absence of nerve gas example we have to infer that
if I were not writing at my computer, it would not be because nerve gas had been
intruded into my oce, or because I had been attacked by amethrowers, or been
struck by a meteor. But why are these inferences alone reasonable? Backtracking
reasoning, unguided by any principles or unconstrained in any way, could equally
well lead to the opposite conclusions. Clearly, this strategy, if it is to give the correct
verdicts about the examples, must articulate some fairly detailed principles regard-
ing the appropriate kind of backtracking reasoning. Until these principles have been
articulated, the strategy is incomplete.
Second, one of the central assumptions of the strategythat every causal state-
ment must be understood in the context of an implicit contrastive why-questionis
too strong. This assumption may hold for some cases, but it is dubious whether
it holds for all. This becomes clear if we allow, as I think we should, for cases of
probabilistic causation, in which the cause brings about the eect but does so with a
chance of less than one. It may be true, for instance, that bombarding a radioactive
particle causes it to decay, but the bombardment is not something that dierentiates
the actual causal history leading to decay from the counterfactual history leading to
nondecay. The atom may fail to decay in the counterfactual history not because the
atom is not bombarded, but because the bombardment does not, as a matter of pure
chance, lead to decay.
Third, Lewiss account of contrastive explanation does not capture an important
feature of contrastive explanations. This point is clearer where contrasts between
compatible alternatives, rather than incompatible alternatives, are being explained.
An example of a contrast between compatible alternatives is Carl Hempels (1965)
much discussed example of syphilis and paresis. Paresis is a late developmental stage
of the disease syphilis, but, as it happens, few people with syphilis contract paresis.
Nonetheless, we can still explain why Jones, rather than Smith, contracted paresis
by saying that only Jones had syphilis. But this cannot be the right explanation on
Lewiss account: syphilis does not dierentiate between the actual case in which
Jones gets paresis and the counterfactual history in which Smith gets paresis, since
the only way in which Smith could get paresis is by rst developing syphilis. Such
examples highlight a feature of contrastive explanations not captured in Lewiss ac-
count. Sometimes the correct contrastive explanation compares actual with actual,
rather than actual with counterfactual. In the example under consideration, it cites
an actual feature that dierentiates Smith and Jonesa feature present in Joness
case but absent from Smiths case.
Finally, and most important, the two-part strategy is unsatisfactory from an ex-
planatory point of view. It unnecessarily duplicates the use of the idea of a cause as
something that makes a dierence: rst in the analysis of objective cause as some-
thing that makes a counterfactual dierence; and then again in the contrastive ex-
planation account of the context-sensitive cause as something that dierentiates
150 P. Menzies
actual from counterfactual histories. These uses of the idea are clearly independent;
neither is derived from the other. Yet it would surely be a surprising fact, requiring
elaborate explanation, if our framework for conceptualizing causation used in two
dierent but crucial ways the very same idea of dierence-making. It would be much
more likely that our conceptual framework was developed on the basis of a single
fundamental application of this idea.
For these reasons, then, the two-part strategy is not as promising as it rst ap-
peared. What is required is a unitary account of causes as dierence-makers that
explains the success of this strategy while avoiding its failures. In my view, if we are
to develop such an account, we must draw a distinction between two kinds of theories
of the context-sensitivity of causal discourse. Add-on context-sensitive theories, like
Lewiss, apply pragmatic principles such as Grices maxims or principles about con-
trastive explanation to independently determined truth conditions. In contrast, inte-
grated context-sensitive theories make the context-sensitivity intrinsic to the truth
conditions of causal claims by making the truth conditions relative to certain con-
textual parameters. I shall recommend adopting an integrated rather an add-on ac-
count of causal claims.8
5 Causal Models
As we have seen, Lewiss view that causation relates events is confounded by the fact
that common sense also allows absences and omissions as causes and eects. The
diculty shows that we need to be inclusive about the relata of causation. To be as
inclusive as possible, I shall talk of factors as causes and eects. Factors are meant to
include anything that common sense dignies as causes and eectsevents, states of
aairs, absences, omissions, and other nonoccurrences.9 I shall reserve the uppercase
variables C, D, E, and so on for factors.
Any theory of causes as dierence-makers must make a connection with Mills
method of dierence for detecting causes and testing causal claims. A crucial part of
the method is a dierence observation between a positive instance in which some
eect E is present and a negative instance in which E is absent. If some condition C
is present in the positive instance and absent in the negative instance, it is, at least,
part of what makes the dierence to E. Mackie (1974, pp. 7172) points out that
there are two forms the classical dierence observation can take. One form is the
before-and-after experiment in which some change C is introduced, either naturally
or by deliberate human action, into an otherwise apparently static situation. The
state of aairs just after the introduction is the positive instance and the state of
aairs just before it is the negative instance. If the introduction is followed, without
any further intervention, by some change C, then we reason that C is part of what
made the dierence to E. The other form the classical dierence observation can
take is the standard controlled experiment, where what happens in the experimental
case is compared with what happens in a deliberately controlled case that is made to
match the experimental case in all ways thought likely to be relevant other than C,
whose eects are under investigation.
Mackie points out that dierent conceptual analyses of causes as dierence-makers
are modeled on the two forms of the classical dierence observation. For example,
C. J. Ducasses (1968) theory of causation is clearly modeled on the before-and-after
observation. It states that the cause of a particular change E is the particular change
C that alone occurred in the close environment of E immediately before it. However,
I agree with Mackie that this analysis is inadequate as an account of causation, as
it fails to distinguish between causal and noncausal sequences of events. Consider
Mackies pair of contrasting sequences (1974, p. 29). In one sequence, a chestnut
is stationary on a at stone. A person swings a hammer down so that it strikes the
chestnut directly from above and the chestnut is attened. In the other sequence, a
chestnut is stationary on a hot sheet of iron. A person swings a hammer down so that
it strikes the chestnut directly from above. At the instant the hammer touches it, the
chestnut explodes with a loud pop and its fragments are scattered around. Couched
as it is in terms of actual changes, Ducasses theory is hard pressed to deliver the
correct verdict that the hammer blow is a cause in the rst sequence but not in the
second.
Mackie argues that these examples show that the relevant contrast in the dierence
observation is not the before-and-after contrast, but the experimental-and-control
contrast (1974, chapter 2). We judge that the hammer blow is the cause of the eect
in the rst sequence because if we were to intervene in the course of events to pre-
vent the hammer from striking the chestnut, the attening would not occur; and we
judge that the hammer blow is not the cause of the eect in the second sequence be-
cause if we were to intervene to prevent the hammer striking the chestnut, the ex-
plosion would still occur. Mackie argues that the conceptual analysis based on the
experimental-and-control form of the dierence observation must appeal to modal
notions, in particular to conditionals. More specically, he argues that the concep-
tual analysis of cause as dierence-maker must appeal to two conditionals, one fac-
tual and the other counterfactual:
(6) Where C and E are distinct factors, C makes a dierence to E if and only if E
would occur if C were to occur and E would not occur if C were not to occur.
152 P. Menzies
This analysis captures the idea involved in the experimental-and-control contrast
precisely because one conditional represents what happens in the experimental case
and the other what happens in the control case.
I nd much of what Mackie says about the experimental-and-control contrast idea
to be illuminating. However, his discussion of this idea is marred by confusions
about the conditionals that are supposed to capture this contrast. Especially confus-
ing is his metalinguistic account of conditionals, according to which they do not have
truth conditions. Nonetheless, I am going to take, as a starting point for my discus-
sion, Mackies claim that the experimental-and-control form of the dierence obser-
vation is the relevant analogical basis for a conceptual analysis of dierence-making.
I shall also take, as a starting point for my discussion, the thesis that this contrast can
be spelled out in terms of a pair of conditionals, one representing what happens in
the experimental case and the other representing what happens in the control case.
(Unfortunately, we shall have to wait until the next section to see the full justication
for these assumptions.) However, I shall reject Mackies confusing account of con-
ditionals, in favor of a more orthodox truth-conditional account in terms of possible
worlds. Under such an account, the central idea of dierence-making can be spelled
out in the following schematic terms.
(7) Where C and E are distinct factors, C makes a dierence to E if and only if
every most similar C-world is an E-world and every most similar @C-world is a
@E-world.
This formulation neatly captures the idea of a cause as a dierence-maker: Where
two relevantly similar possible worlds dier with respect to C they also dier with
respect to E, and vice versa. A condition that just happens to covary with another in
their actual instances will not modally covary in the way required to count as making
the dierence.
I should state at the outset that, while using the standard possible worlds frame-
work for understanding conditionals, I understand the possible worlds in a slightly
unconventional way. The possible worlds I shall employ are miniworlds rather than
alternative large-scale universes: they are alternative courses of development of typi-
cally small-scale systems. They are best understood as being similar to the trajec-
tories in the state space posited by a scientic theory to describe the behavior of
systems of a certain kind. A theory may seek to describe the behavior of a certain
kind of system in terms of a set of state variables fS
1
; . . . ; S
n
g. The accompanying
state space will be an n-dimensional space and the trajectories in this space will be
temporally ordered sequences of states in this space. So while I use the traditional
term possible world, it should always be kept in mind that I understand it typically
in the miniworld sense, where the miniworlds are analogous to trajectories in a
state space for a typically small-scale system of a certain kind.
The all-important question to be answered about the possible worlds formulation
of the dierence-making idea is: which worlds count as relevantly similar to the actual
world? As we have seen, Lewis thinks that, for each causal claim about an event that
makes a dierence to another, the corresponding counterfactuals are to be read in
terms of a unique kind of similarity relation. In this respect, my position diers from
Lewiss, in that I think that the relevant similarity relations are context dependent,
with causal statements in dierent contexts requiring dierent similarity relations.
Causal statements must be understood, I shall argue, as relative to a certain contex-
tual parameter; and depending on the way the parameter is set, an appropriate kind
of similarity is determined for a given causal statement.
The contextual parameter in question reects the fact that our causal thinking is
steeped in abstraction. It is a platitudebut one worth repeatingthat the world is
exceedingly complex in its causal structure. Within any spatiotemporal region, there
are many levels of causation, and within each level many cross-cutting and inter-
secting causal processes. To determine the structure of these processes, we are nec-
essarily forced, by the nitude of our minds, to focus selectively on some aspects of
what is going on and to ignore others or place them in the background. The causal
schemas by which we interpret the world are irremediably permeated by abstractions
that enable this selective focusing. There seem to be several forms of abstraction that
underlie our causal thinking.
One form of abstraction underlying our thinking about the causal structures of
a concrete situation involves the identication within the situation of a particular set
of objects as forming a system of a certain kind. A particular system may consist of
a great many objects or very few, of very large objects or very small ones. Astron-
omers and cosmologists investigate vast systemssolar systems, galaxies, or the
whole cosmos. The systems investigated by biologists and economistseconomies,
markets, species, populations, and so onare smaller, but still large by human stan-
dards. On the other hand, the systems investigated by particle physicists are small
by any standard. It is not always easy to determine which objects belong to a partic-
ular system. This is not just because of our epistemic limitations, but because the
spatiotemporal boundaries of the system are indeterminate. How many astronomical
bodies are in the Milky Way galaxy? How many organisms belong to a population
of marsh frogs? It is dicult to answer these questions because the spatiotemporal
boundaries of these systems are not perfectly determinate. Nonetheless, the indeter-
minate localization of systems does not stop scientists from conceptualizing causal
structures in terms of them.
154 P. Menzies
The form of conceptual abstraction under consideration involves not just the iden-
tication of a particular set of objects, but the identication of this set of objects as
constituting a system of a certain kind. But what is a system? A simple answer to this
question is that a particular system is a set of objects that have certain properties and
relations. But not any old properties and relations are relevant to the identication of
a system. For example, a set of astronomical bodies can be individuated as a partic-
ular planetary system by way of each astronomical bodys relation to other bodies in
the system, but not by way of their relations to objects outside the system; a partic-
ular population of marsh frogs may be individuated in terms of the frogs relational
property of living in a particular marsh, but not in terms of extraneous relational
properties involving far-distant objects. In short, a system is a set of constituent
objects that is internally organized in a distinctive fashion; and the properties and
relations that congure the objects into a system must be intrinsic to the set of con-
stituent objects.
The concept of intrinsic properties and relations has been much discussed. How-
ever, the signicant concept under consideration here is not the concept of properties
and relations that are intrinsic tout court, but those that are intrinsic to a set of ob-
jects. It will suce for our purposes to explain the intuitive idea behind these con-
cepts, rather than to present a full analysis of them, which turns out to be slightly
tricky. Modifying an idea of Jaegwon Kims (1982) concerning the simple concepts, I
shall say that:
(8) A property F is extrinsic to a set of objects if and only if, necessarily, one of its
members has F only if some contingent object wholly distinct from the set exists.
For example, the extrinsic properties of a set of astronomical bodies would include
being observed by some human and being a certain distance from the Earth (assum-
ing the Earth is not in the set).10
The concept of a property intrinsic to a set of objects is dened in converse
fashion:
(9) A property F is intrinsic to a set of objects if and only if, possibly, one of its
members has F although no contingent object wholly distinct from the set exists.
For example, the intrinsic properties of a set of astronomical bodies would include
the mass and shape of the individual astronomical bodies. But the intrinsic properties
of the set need not be all intrinsic properties simpliciter. For example, the property of
being gravitationally attracted to another body that is also a member of the set is an
intrinsic property of the set, though it is not an intrinsic property simpliciter.11
There are, literally, uncountably many particular systems, but very few of them
are of any interest to us. For the most part, we are interested in the kinds of systems
that evolve in lawful ways. For example, certain systems of astronomical bodies and
certain systems of biological organisms have intrinsic properties and relations that
change over time in regular ways described by certain laws. Identifying a kind of
system involves identifying the intrinsic properties and relations that are shared by
particular systems and that conform to certain laws. The state variables employed in
a scientic theory correspond to the intrinsic properties and relations that constitute
a kind of system. In Newtonian mechanics, for instance, the state variables used to
describe the behavior of mechanical bodies are the properties of mass, position, and
momentum. The following denition captures these ideas:
(10) A kind of system K is a set of particular systems sharing the same intrinsic
properties and relations (state variables) whose evolution over time conforms to
certain laws.
By denition, the state variables that determine a given kind of system are intrinsic
properties and relations of the particular systems belonging to the kind. More gen-
erally, a kind of system supervenes on a set of intrinsic properties and relations in the
sense that any two particular systems with the same intrinsic properties and relations
must both belong, or both fail to belong, to a given kind of system.
I have said that a certain contextual parameter determines the similarity relation
relevant to working out whether some condition makes a dierence to another in a
given concrete situation. I propose that one element of this contextual parameter is a
kind of system. It is, I claim, an automatic and inevitable feature of the way in which
we conceptualize the causal relations of a concrete situation that we see the concrete
situation as an instance of a certain kind of system.
The other element in the similarity-determining contextual parameter is the set of
laws governing the kind of system under consideration. This element of the contex-
tual parameter reects a further type of abstraction involved in our causal thinking,
for almost invariably the laws governing the kinds of system of interest to us are
ceteris paribus laws. Such laws state that the relevant systems evolve along certain
trajectories provided nothing interferes. For example, the law of gravity in New-
tonian mechanics states that, provided there is no other interfering force, the force
exerted by one object on another varies directly as the product of their masses and
inversely as the square of the distance between them. The law of natural selection
states that, provided there is no force besides that of selection at work, if organisms
possessing a heritable trait F are tter than organisms with an alternative heritable
156 P. Menzies
trait F
0
, then the proportion of organisms in the population having F will increase.
Georey Joseph (1980) suggests that such laws would be better called ceteris absen-
tibus laws, as they usually describe the evolution of the relevant systems under the
assumption that all interfering factors or forces are absent. Such an assumption is,
often enough, an idealization, because most kinds of systems are subject to interfer-
ing inuences in addition to the causal inuences described by their relevant laws.
Idealization is central to our causal thinking, as is evident from the ubiquity of
ceteris paribus laws. Still, many philosophers have thought that ceteris paribus laws
are disreputable in some way. For example, it is sometimes objected that ceteris
paribus laws are vacuous because the ceteris paribus condition cannot be specied
nontrivially. (See, for instance, Fodor 1991.) The law that ceteris paribus all Fs are
Gs, the objection runs, is really just the vacuous law that all Fs, unless they are not
Gs, are Gs. This objection has no cogency at all, in my view. The law of gravity that
tells us how, in the absence of other causal inuences, gravity exerts a force on ob-
jects is far from trivial. It makes a substantive claim about the world because the
concept of an interfering causal inuence can be explicated informatively. Without
being overly precise, we can explicate the concept in the following terms:
(11) A factor I is an interfering factor in the evolution of a system of kind K in
conformity with the laws L if and only if:
(i) I instantiates an intrinsic property or relation in a particular system of kind K;
(ii) I is caused by some factor instantiating a property or relation extrinsic to the
system of kind K; and
(iii) the laws governing the causation of I by the extrinsic factor are distinct from
the laws L.
Condition (i) simply states that the interfering factor is an intrinsic feature of the
system in question. But condition (ii) says that this factor must have a causal source
extrinsic to the system. Condition (iii) says that the causation of the factor can be
explained independently of the laws governing the system in question. The paradigm
example of an interfering factor is the result of an intervention by a human agent
in the workings of the system. For example, the gravitational force of the Earth on a
simple pendulum can be counteracted by a simple human intervention in the swing of
the pendulum. While human interventions are not the only kinds of interfering fac-
tors, they form the analogical basis for our thinking about interfering forces. They
constitute the most familiar type of situation in which an external force, operating
according to its own distinctive laws, can intervene or intrude into the workings of a
system.
Given this explication, we can see that the hypothesis that the ceteris paribus laws
of Newtonian mechanics hold true of some system, say, the system of planets orbit-
ing around the Sun, involves a substantial claim about the world. The hypothesis
commits one not only to making certain predictions about the orbits of the planets,
but also to explaining prediction failures in terms of the external interfering forces
whose causal explanation is, in some sense, independent of the system and laws
under consideration. Ceteris paribus laws are, to use the words of Pietroski and Rey
(1995), like checks written on the banks of independent explanations, their sub-
stance and warrant deriving from the substance and warrant of those explanations. It
may be questionable on some occasions whether the check can be cashed, but that
hardly demonstrates the general inadequacy of the institution of bank checks.
Another common objection to ceteris paribus laws is that, even if ceteris paribus
clauses can be specied nontrivially, they cannot be specied determinately. (See, for
instance, Schier 1991.) This is because it is impossible to specify in advance all the
interfering factors whose absence is required to enable a given system to evolve in
accordance with given laws. Without doubt, there is truth in this claim. But it is a
mistake to think this somehow impugns the determinacy of a ceteris paribus law. It
is a mistake to say that the statement There is only one person in the room
alternatively There is one person in the room and no one elsehas no determinate
sense because one cannot specify in advance every person whose absence is required
to verify the negative existential. This mistake rests on a confusion about what the
determinacy of negative existentials requires. The objection to the determinacy of
ceteris paribus laws rests on exactly the same confusion.
To capture the fact that our causal thinking is permeated by the two kinds of ab-
straction identied above, I shall say our causal judgments about a concrete situation
must be understood as relative to a causal model of the situation. I represent a causal
model of a situation as an ordered pair hK; Li, where the rst element K is the kind
of system in terms of which we conceptualize the situation, and the second element
L is the set of laws, typically ceteris paribus laws, governing the evolution over time
of that kind of system. In using the term causal model, I hope to highlight the
continuity between commonsense causal thinking and the causal theorizing of the
natural and social sciences. Several philosophers of sciencenotably, Nancy Cart-
wright (1983, 1999), Ronald Giere (1988), Fred Suppe (1979, 1989), and Bas van
Fraassen (1980, 1989)have emphasized that theorizing in these sciences often pro-
ceeds by way of idealized causal models in which ceteris paribus laws play a central,
indispensable role. I would claim that these features of scientic practice have their
roots in everyday causal reasoning.
158 P. Menzies
No doubt the claim that causal judgments about a concrete situation are to be
understood as relative to a causal model of the situation will strike many as confused
and erroneous. So let me try to forestall some misunderstandings of this claim.
First, I am not claiming that causation is mind-dependent in some idealist sense. I
am simply explicating the scientic commonplace that the causal structure of any
particular situation can be modeled in several ways. I interpret this commonplace as
meaning that a given situation can be viewed as instantiating dierent kinds of sys-
tems obeying dierent laws. In the analysis to follow, the claim that the dierence-
making relation is relative to a causal model M hK; Li should be understood in
terms of a conditional construction of the following form: if the given situation in-
stantiates a kind of system K governed by laws L, then C makes a dierence to E if
and only if . . . , where what replaces the dots will state a perfectly objective condition
about the world. If the given situation satises the antecedent of this conditional, it is
a completely mind-independent matter whether some factor in the situation makes a
dierence to another.
Second, I am not endorsing a crude relativism to the eect that any causal model
of a situation is as good as any other, or more specically, any kind of system is just
as natural as any other for determining causal relations. There are natural kinds,
in my view, but it is the job of metaphysics and science rather than conceptual anal-
ysis to investigate what they are. However these investigations turn out, a plausible
metaphysics is likely to allow that any particular spatiotemporal region instantiates
several kinds of systems. Perhaps an extremely austere physicalism committed to the
existence of a unied eld theory would assert that every situation is to be modeled
in terms of a unique physical kind of system subject to the unied eld equations.
However, any less austere metaphysics is likely to conclude that several, perhaps
imperfectly natural kinds of systems may be instantiated in a given spatiotemporal
region. In this case, a conceptual analysis should be able to make sense of the alter-
native causal judgments about these dierent kinds of systems.
Finally, I am not saying that a causal model must be specied in terms of known
kinds of systems and known laws. My discussion has been inuenced by philosophers
of science who argue that scientic theories are best understood as abstract models.
Of practical necessity, they discuss known scientic theories in terms of known kinds
of systems and known laws. But I do not wish to conne causal models to what is
actually known. Common sense and scientic practice accept a realism according to
which we may be ignorant of the intrinsic properties and relations that constitute a
kind of system, and we may yet have to discover all the laws governing a kind of
system. Indeed, our causal judgments may presuppose a causal model that can be
specied imperfectly only in terms of an incompletely known kind of system and set
of laws. But the analysis to follow can proceed satisfactorily in terms of an objectied
causal model along these lines: if the given concrete situation is an instance of this
imperfectly known kind K obeying the imperfectly known laws L, then a dierence
making claim is true if and only if . . . .
6 The Similarity Relation and Dierence-making
How exactly are causal judgments about a concrete situation relative to a causal
model? The relativity of causal judgments to a model consists, I shall argue, in the
fact that the model determines the respects of similarity used in evaluating whether a
putative cause makes a dierence to an eect. I will try to explain the way a model
determines these respects of similarity in several stages.
Let us suppose that we are considering the structure of causal relations in a par-
ticular system of kind K in a certain interval of time (t
o
t
n
). The following denition
captures the way in which a causal model determines the fundamental respects of
similarity that are relevant to determining whether one condition makes a dierence
to another.
(12) A model hK; Li of an actual system of kind K generates a sphere of normal
worlds that consists of all and only worlds w such that:
(i) w contains a counterpart to the actual system and this counterpart has exactly
the same K-determining intrinsic properties and relations as the actual system at
time t
o
;
(ii) w does not contain any interfering factors (with respect to the kind K and laws
L) during the interval (t
o
t
n
);
(iii) w evolves in accordance with the laws L during the interval (t
o
t
n
).
For each sphere of normal worlds, there is a conjunctive proposition that is true of
all and only the worlds in the sphere. I shall label this conjunctive proposition F
M
,
and, taking over terminology reintroduced by Mackie (1974, p. 35) say that it speci-
es a eld of normal conditions (generated by the model M).
Each of the worlds in the sphere of normal worlds generated by the model M
exemplies a course of evolution that is normal, in a certain sense, for a system of
the kind K evolving in accordance with the laws L.12 The conditions imposed on the
these worlds represent default settings of the variablesthe initial conditions, the
laws, and the absence of interferersthat can inuence the way the system evolves
through time. If we are investigating the causal relations in a system of a certain
160 P. Menzies
kind, as it evolves through a given interval of time, it is reasonable to assume that
the initial conditions of the system are the kind-determining intrinsic properties and
relations that the system possesses at the beginning of the interval; that the system
evolves in accordance with the laws governing the kind of system in question; and
that none of the factors that can interfere with the lawful evolution of the system is
present. These are default settings in the sense that the assumption that they obtain
constitutes a reasonable starting point for our causal investigations, an assumption
that we relinquish only when forced to do so. This is not to say that we are always
aware of what these default settings are. As mentioned above, there is no reason to
think that we will always have complete knowledge of all the initial conditions of
a given system, or all the laws governing systems of that kind, or all the possible
interferers that can hinder the given systems lawful evolution. Nonetheless, we move
from the assumption that the actual system will evolve in accordance with these de-
fault settings, whatever their precise details, only when we have good reason to think
it must deviate from them.
The normal worlds generated by a model are those that form the background to
any consideration of whether some factor makes a dierence to another. These nor-
mal worlds may hold xed, as part of the eld of conditions, intrinsic properties and
relations of the system that are causally relevant to the eects displayed by the sys-
tem. A special case is that in which the system does not contain any such causally
relevant factors. In Newtonian mechanics, a system subject to no forces at all is
such a special case. The zero-force law of Newtonian mechanicsthe rst law of
motiontells us that such a system will remain at rest or travel at a constant veloc-
ity. Similarly, a special case in population genetics is a population subject to no evo-
lutionary forces. The evolution of gene frequencies in such a population is described
by the zero-force HardyWeinberg law. In contrast to these special cases, the typical
case is one in which the initial conditions of the system already entail that the system
is subject to certain forces. The very description of a Newtonian system consisting
of two particles with certain masses and a certain distance apart will entail that it
is subject to gravitational forces. And the very description of a population whose
members have certain properties entailing dierential tnesses will ensure that the
population is subject to the force of natural selection.13 Even when a system already
possesses an array of causal forces, it makes sense to ask about the causal signi-
cance of additional causal forces. The condition for dierence-making provides us
with a test of the causal signicance of these extra factors.
The sphere of normal worlds generated by a model is tied, in some sense, to the
actual world. Worlds earn their membership in the sphere by virtue of their resem-
blance to the way the actual system under consideration would evolve in conformity
with actual laws. Nonetheless, it is important to note that the actual world need
not itself belong to the sphere of normal worlds, for these worlds represent how the
actual system would evolve in conformity with the laws in the absence of any inter-
ferers. In many cases, therefore, these worlds are ideal ones. The actual world, as we
know, may be far from ideal in that the evolution of the actual system may be sub-
ject to many interfering forces. The presence of any interfering factor disqualies the
actual world from belonging to the sphere of normal worlds. It follows from this that
the centering principle that Lewis (1973b, pp. 2631) imposes on the similarity rela-
tion for counterfactuals fails to hold here, both in its strong and its weak forms. Its
strong form states that there is no world as similar to the actual world as the actual
world itself, so disallowing ties for most similar world. Its weak form, which allows
for ties for most similar world, states that there is no world more similar to the actual
world than the actual world itself. That the actual world need not belong to the
sphere of normal worlds generated by a model means we must abandon the centering
principle in both its forms.
So far, we have attended to the question of which worlds count as the normal
worlds generated by a model. But denition (7) of a dierence-making factor C re-
quires a specication of the most similar C-worlds and the most similar @C-worlds.
It is best to spell out the denition of the most-similar C-worlds by considering
subcases.
One subcase we need not consider is that in which both the conditions C and @C
are consistent with the eld of conditions F
M
. This case cannot arise because it is self-
contradictory. For both C and @C to hold consistently with the eld of conditions,
C would have to hold in some of the normal worlds and @C would have to hold
in other normal worlds. But since the laws governing these worlds are deterministic,
these worlds would have to dier either with respect to their initial conditions, or
with respect to the presence of interfering factors. In either case, the worlds could not
satisfy all the conditions (12(i)(iii)) required for membership in the sphere of normal
worlds.
The rst subcase we need to considerI shall call it subcase Iis the case in
which the eld of conditions F
M
implies the putative dierence-making factor C. In
this kind of case, the initial conditions of the system and the laws, in the absence of
interfering factors, imply that the factor C holds in the system. As an illustration,
consider a slight modication of Mackies example: A specially designed machine
swings a hammer so that it strikes a chestnut directly from above. Suppose we are
considering whether the hammers striking the chestnut (C) makes a dierence to the
attening of the chestnut (E), where the hammer strike is an outcome of the lawful
evolution of the relevant system from its initial conditions.
162 P. Menzies
In this kind of case, it is simple to specify which worlds are to count as the most
similar C-worlds. They are simply the C-worlds that belong to the sphere of normal
worlds, that is, those C-worlds that hold xed the eld of conditions F
M
. A compli-
cation arises, however, when it comes to specifying which worlds are to count as the
most similar @C-worlds. Clearly, @C is not consistent with F
M
, and so the normal
worlds included in F
M
are not eligible to be the most similar @C-worlds.
To work out which are the most similar @C-worlds in these circumstances, we
need to specify the worlds that dier from the normal worlds no more than is neces-
sary to allow for the realization of @C. In other words, we must nd the minimal
revision of the eld of conditions F
M
that is consistent with @C. There are three ele-
ments that determine F
M
: the initial conditions of the system, the laws governing the
system, and the absence of interfering factors. We can get a set of most similar @C-
worlds by systematically revising each of these elements to allow for the realization
of the counterfactual @C. And each of the resulting revisions counts, in some sense,
as a minimal @C-inducing revision of F
M
. Indeed, each of these revisions generates a
similarity relation that corresponds to a certain style of counterfactual reasoning.
For example, we could revise the eld of conditions F
M
to allow for @C by revis-
ing the laws that govern the system in question while holding xed the initial con-
ditions and the absence of interferers. Evidently, this kind of revision is required to
entertain counterlegals such as If force were given by mass times velocity, then. . . .
However, this kind of revision is not relevant in the present context, in which we are
treating counterfactual antecedents that concern particular matters of fact. Another
possibility, more relevant in the present context, is to revise F
M
by altering the initial
conditions while holding xed the laws and the absence of interferers. This type of
revision corresponds to the style of counterfactual reasoning by which we infer how
the past conditions must have been dierent in order for some counterfactual ante-
cedent to be true. This kind of backtracking reasoning lies behind a counterfactual
such as If the hammer had not struck the chestnut, then the operating machine
would have had a malfunction of some kind. However, the one thing we know from
counterfactual analyses of causation is that the required similarity relation must not
allow for backtracking reasoning of this kind, on pain of generating countless in-
stances of spurious causation.14
The only option left is to revise F
M
by allowing for the presence of an interfering
factor that would realize the counterfactual antecedent @C, while holding xed the
initial conditions and laws of the system. In other words, the most similar @C-worlds
are like the worlds stipulated in (12) above, in that they preserve the initial conditions
of the actual system (condition (i)) and the laws governing the system (condition
(iii)), but they dier from these worlds in that they allow for an interfering factor that
realizes the counterfactual antecedent @C (not condition (ii)). As discussed above,
the paradigm of such an interfering factor is an external human intervention in a
system. In small-scale systems open to human manipulation, the kind of interference
that would realize a counterfactual antecedent is to be understood in terms of a
human intervention.
For example, in the modied Mackie example, the most similar worlds that make
it true that the hammer does not strike the chestnut are easily imagined: they are
simply worlds in which the relevant machine runs on from its initial conditions in
conformity with the relevant laws, but at some point a human agent intervenes to
prevent the hammer from striking the chestnut. With large-scale systems not open
to human manipulation, the interference that realizes the counterfactual antecedent
can be understood in terms of a miracle that interrupts the lawful evolution of the
system.15 But even here, I would argue, the analogy with human intervention guides
the way we think in these cases about the miraculous realization of the counter-
factual antecedents.
It is useful at this point to be able to specify which worlds are to count as the
most similar C-worlds for any antecedent C, whether or not it is entailed by the eld
of conditions F
M
. To be able to do this, we need an ordering of spheres of worlds
in terms of their similarity to the normal worlds in F
M
. (Compare Lewis 1973b, pp.
1316.) If the ordering is to carry information about similarity to the normal worlds,
it must satisfy certain conditions.
(13) Let fS
0
; . . . ; S
n
g be an ordered set of spheres of worlds. This set is centered on
the sphere of normal worlds S
0
F
M
if and only if S
0
is included in every other
sphere. The set is nested if and only if for any spheres S
i
and S
j
in the set, either S
i
is included in S
j
or S
j
is included in S
i
.
When the ordered set of spheres is centered and nested in this sense, it can convey
information about the similarity of worlds to the normal worlds. A particular sphere
around the sphere of normal worlds will contain just those worlds that resemble the
normal worlds to a certain degree. The dierent spheres will correspond to dierent
degrees of similarity to the normal worlds. The smaller the sphere, the more similar
to the normal worlds will be a world falling within it. In other words, if one world
falls within a sphere and another world lies outside that sphere, the rst world will
resemble the normal worlds more closely than the second.
This purely formal specication of the ordering of spheres answers some questions
of logic. However, if it is to be applied to particular examples, it must be made more
specic with a detailed description of the respects of similarity to the normal worlds
that receive signicant weighting in the interpretation of conditionals. A complete
164 P. Menzies
description of these weightings would require an extensive discussion. However, it
will suce for our treatment of the particular examples of this essay to note one im-
portant principle that seems to govern our intuitive judgments about this matter.
(14) Weightings of similarity principle: In determining the respects of similarity to
the normal worlds generated by a model M, it is of rst importance to preserve the
initial conditions and the laws of the relevant kind of system; and it is of second
importance to preserve the absence of interfering factors.
One obvious implication of this principle is that it allows us to read certain coun-
terfactuals in the characteristic nonbacktracking manner. It permits a counterfactual
antecedent to be realized in a world by an external intervention in the relevant sys-
tem if the laws and initial conditions of the system are preserved in that world. The
principle has another implication that will be relevant to our discussion. We obvi-
ously entertain counterfactuals whose antecedents concern changes in the initial con-
ditions of a system. It is perfectly intelligible to say about the modied Mackie
example, for instance, If the initial conditions of the hammer-striking machine had
been dierent, then the situation would have evolved dierently. But the principle
at hand tells us that we have to go farther out from the normal worlds to nd worlds
that permit this counterfactual antecedent than we have to go to nd worlds that
permit the counterfactual antecedent of If the machines hammer had not struck the
chestnut, it would not have been attened. Both antecedents can be realized in a
world that permits an external intervention in the system. But a world that realizes
the rst antecedent will, of necessity, involve a change in the initial conditions of the
system, whereas a world that realizes the second antecedent will not. The weightings
principle implies that the rst world must be less similar to the normal worlds than
the second world.
With this ordering in hand, we can dene the most similar C-worlds in a perfectly
general way that covers the case in which C is entailed by the eld of conditions F
M
and the case in which it is not entailed.
(15) The most similar C-worlds generated by a model M are the C-worlds that
belong to the smallest C-permitting sphere in the ordering of spheres governed by
the weightings principle.
This is perfectly general also in that it covers not just subcase I, in which C is im-
plied by the eld of conditions F
M
, but also the yet-to-be-considered subcase II, in
which it is @C rather than C that is implied by F
M
. In this second subcase, the most
similar @C-worlds are simply the @C-worlds belonging to the sphere of normal
worlds. However, to nd the most similar C-worlds in this subcase, we have to go
out from the sphere of normal worlds to nd worlds that allow for the realization of
C by intervention or miracle.
We are nally in a position to explicate the idea of one factor making a dierence
to another in a way that acknowledges the relativity to models.
(16) C makes a dierence to E in an actual situation relative to the model M of
the situation if and only if every most similar C-world generated by the model is an
E-world and every most similar @C-world generated by the model is a @E-world.
Of course, these truth conditions bear an unsurprising resemblance to the standard
truth conditions for counterfactuals. If the truth conditions for counterfactuals are
relativized in a way to match those given above, then the condition can be reformu-
lated to yield the following one:
(17) C makes a dierence to E in an actual situation relative to the model M if
and only if C L
M
E and @C L
M
@E.
Here the subscript M on the counterfactual operator signies that the operator is
dened with respect to the ordering of spheres generated by the model M.
This counterfactual construction is similar to the notion of counterfactual depen-
dence that plays the central role in Lewiss counterfactual analysis. Indeed, it will be
useful to be able to take over this terminology. But the way I will use the term is
dierent from the way Lewis uses it in two respects. First, the counterfactuals that
dene counterfactual dependence do not, for the reasons given above, conform to the
centering principle that Lewis imposes on counterfactuals. On the other hand, I be-
lieve that they conform to the limit assumption to the eect that there is a smallest
sphere of antecedent-permitting worlds for any entertainable antecedent. Lewis con-
siders this an optional principle for counterfactuals. But, in fact, it applies automati-
cally to the counterfactuals I have dened, since it follows from the way in which a
model generates an ordering of spheres of worlds centered on the normal worlds.
The other way in which the present denition of counterfactual dependence diers
from Lewiss is the obvious one bearing on the relativity to a model. The notion of
counterfactual dependence, as I will use it, inherits the relativity to a model of the
counterfactuals that dene it. The truth conditions of counterfactuals in my theory
are dened over the most similar antecedent-worlds generated by a model. Lewiss
notion involves no such relativity, for he assumes that there is only one kind of
system to considerthe whole universeand so his worlds are maximal worlds. He
also assumes that worlds are governed by exceptionless laws without ceteris paribus
conditions, and so he makes no use of the notion of an interferer in a system, which
166 P. Menzies
I believe is required to explain the content of ceteris paribus conditions. There are
further dierences between the accounts, but they follow from these.
Let me conclude this section by connecting up our recent discussion with the ear-
lier discussion of the idea that motivates analyses of causation in terms of making a
dierence. As we have seen, Mackie argues for a conceptual analysis of a cause as
what makes the dierence in relation to some assumed background or causal eld.
This idea is best understood, he argues, in terms of the experimental-and-control
contrast, rather than the before-and-after contrast. The former contrast can be cap-
tured by a pair of conditionals, with one conditional corresponding to the experi-
mental case and the other conditional corresponding to the control case.
We can now see how to make sense of Mackies claims. Let us suppose that a
conditional is an experimental conditional if we do not have to leave the sphere of
normal worlds to nd the most similar antecedent-worlds. In other words, the ante-
cedent would be realized by allowing the system in question to evolve lawfully with-
out interference. On the other hand, let us suppose that a conditional is a control
conditional if we do have to leave the sphere of normal worlds to nd the most sim-
ilar antecedent-worlds. Or, in other words, the antecedent would be realized only by
an intervention in the lawful evolution of the system in question from its initial con-
ditions. Given this terminology, we can see that one of the conditionals that dene
the dierence-making condition (17) will be an experimental conditional and the
other a control conditional.
It is important to realize, though, that the two conditionals in (17) do not always
line up in the same way with the experimental and control cases. It depends on
whether we are considering subcase I or subcase II. In subcase I, the experimental
conditional is C L
M
E and the control conditional is @C L
M
@E. An example in
which both these conditionals are true is represented in gure 6.1. In this gure, the
concentric circles represent the spheres of worlds, with the smallest sphere F
M
repre-
senting the sphere of normal worlds generated by a model M. The symbol @
denotes the actual world.
In subcase II, C L
M
E is the control conditional and @C L
M
@E the experi-
mental conditional. An example in which both these conditionals are true is repre-
sented in gure 6.2. (Notice that in both subcases C L
M
E is a factual conditional in
the sense that C and E both actually hold, while @C L
M
@E is a genuine counter-
factual conditional since @C and @E actually fail to hold. Hence, the experimental
control dichotomy crosscuts the factualcounterfactual dichotomy.)
Finally, I wish to make a connection with another classic discussion of causation.
In their work on causation in the law, Hart and Honore claim that the concept of
Figure 6.1
Figure 6.2
168 P. Menzies
a cause as making a dierence has its home in a certain paradigm situation. They
write:
Human action in the simple cases, where we produce some desired eect by the manipulation
of an object in our environment, is an interference in the natural course of events which makes
a dierence in the way these develop. In an almost literal sense, such an interference by human
action is an intervention or intrusion of one kind of thing upon a distinct kind of thing. Com-
mon experience teaches us that, left to themselves, the things we manipulate, since they have
a nature or characteristic way of behaving, would persist in states or exhibit changes dier-
ent from those which we have learnt to bring about in them by our manipulation. The notion,
that a cause is essentially something which interferes with or intervenes in the course of events
which would normally take place, is central to the commonsense concept of cause, and at least
as essential as the notions of invariable or constant sequence so much stressed by Mill and
Hume. Analogies with the interference by human beings with the natural course of events in
part control, even in cases where there is literally no human intervention, what is identied as
the cause of some occurrence; the cause, though not a literal intervention, is a dierence to the
normal course which accounts for the dierence in the outcome. (Hart and Honore 1985,
p. 29)
Again it is possible to see the point of Hart and Honores remarks in the light of
the framework developed above. They are, in eect, considering, as a paradigm case,
the kind of causal situation that falls under subcase II. The normal course of events
for a system of some kind, free from any external interference, makes @E true. If it
turns out that E actually holds, then an explanation is required in terms of some
factor that makes a dierence to this normal course of events. It will be some factor
C such that both the experimental conditional @C L
M
@E and the control
conditional C L
M
E are true. By the denition of these conditionals, C will count as
an interfering factor in the normal course of development of the system, the kind of
interfering factor that is paradigmatically exemplied by an external intervention in
the system by an agent. My only reservation about the quotation from Hart and
Honore is that it focuses attention on just one of the two important subcases of the
idea of making a dierence, ignoring the important subcase I.
7 The Phenomena Explained
Let us return to the various puzzle cases that were cited as problems for Lewiss ac-
count of dierence-making in section 3. As we saw in that section, Lewiss account
blurs the commonsense distinction between causes and conditions. Let us see whether
the present account of dierence-making can explain this distinction.
Most philosophical discussions of the distinction between causes and conditions
take it for granted that the distinction is best elucidated in a specic explanatory
setting. (See, for example, Hart and Honore 1985, pp. 3244; and Mackie 1974, pp.
3437.) The setting is one in which some unexpected factor E stands in need of ex-
planation, specically in terms of something that dierentiates it from the normal
situation in which @E obtains. This assumption makes sense in terms of the present
framework. It simply amounts to the supposition that the causesconditions distinc-
tion is best understood in terms of examples falling under subcase II. In examples of
this kind, the eld of normal conditions F
M
generated by a model entails @E, so that
when E unexpectedly obtains, it requires explanation in terms of a dierence-making
factor. Let us proceed on the assumption of this explanatory setting, though always
keeping in mind that this is just one of two possible subcases.
It is interesting to note in this connection a certain abstract implication of the logic
of dierence-making. It follows from the description of the assumed explanatory
setting as exemplifying subcase II that the factor identied as making the dierence
to E must be identied as an interfering factor in the system. The logic of the situa-
tion implies that the factor C that makes the dierence to E cannot intersect with the
eld of normal conditions F
M
, so that the smallest C-permitting sphere of worlds
includes worlds in which C is realized by way of an external interference or inter-
vention in the system. (Figure 6.2 represents the situation of the explanatory setting.)
This implication explains the way in which commonsense and scientic explanations
of abnormal occurrences in systems often describe the dierence-making factor as an
interferer or intervention in the system. Such factors are seen as intrusions into the
system that account for the deviation from the normal course of events.16
In terms of the explanatory setting we are assuming, it is easy to identify the con-
ditions for a causal judgment made relative to a model M. They are simply the con-
ditions that belong to the eld of normal conditions F
M
. Typically speaking, they will
be conditions relating to the initial state of the system in question, conditions relating
to the absence of interferers, and any conditions that follow from these and the laws
governing the system. Notice that these conditions are not restricted to ones obtain-
ing contemporaneously with the dierence-making factor. To modify an example
from Hart and Honore (1985, p. 39): if lightning starts a re in the grass, and shortly
after, a normal gentle breeze gets up and the re spreads to a forest, then the light-
ning caused the forest re and the breeze was a mere condition of it. Even though the
breeze was subsequent to the lightning, it can be identied as a condition so long as it
arises lawfully from the initial conditions of the relevant system and is not an inter-
ference in the system.
It follows from this identication of the conditions that no condition can be cited
as a dierence-maker for the unexpected or abnormal event E in the explanatory
setting under consideration. A dierence-maker C must be an actually obtaining fac-
170 P. Menzies
tor such that all the most-similar C-worlds are E-worlds and all the most-similar
@C-worlds are @E-worlds. Figure 6.2, which illustrates this explanatory setting,
shows that no actual condition entailed by F
M
can meet this requirement. In partic-
ular, any actual condition X entailed by F
M
will be such that all the most similar X-
worlds are @E worlds, contrary to what is required. Consequently, the identication
of a condition as an actual factor entailed by F
M
implies that conditions cannot be
dierence-making causes. This is exactly as it should be.
Let us make the discussion concrete by reconsidering the cases cited in section 3,
starting with the lung cancer case (example 1). The essential step in treating this ex-
ample is the identication of the causal model that guides our intuitions about it. It is
natural to specify the relevant model here as involving a person living according to
the laws of normal healthy functioning. If the person gets lung cancer, then it is rea-
sonable to look for some factor that makes the dierence to this eect with respect to
the normal worlds generated by this modelsome factor such as his smoking, for
instance. Yet, even when the model is specied in these broad terms, we can see that
the eld of normal conditions generated by this model will hold xed, as initial con-
ditions, the fact that the person has been born and has lungs. Hence, these conditions
cannot be cited as a dierence-making factors.
The situation is represented diagrammatically in gure 6.3. The eld of normal
conditions F
M
generated by the relevant model M entails that the person in question
does not get lung cancer (LC). However, as things actually turn out, the person
develops lung cancer, so that an explanation is required in terms of a dierence-
maker. The gure shows that persons smoking (S) is such a dierence-maker, since
Figure 6.3
both of the conditionals S L
M
LC and @S L
M
@LC are true. However, the gure
also shows that the persons birth (B) and his possession of lungs (L) are mere con-
ditions in this example, as they are entailed by the eld of normal conditions. As
such, they cannot qualify as dierence-makers for the persons lung cancer. (In par-
ticular, it turns out that the wrong conditionals B L
M
@LC and L L
M
@LC hold,
so that B and L cannot be make the appropriate counterfactual dierence to LC.)
The gure also shows that there are outer spheres of worlds, quite dissimilar to the
normal worlds of F
M
, that permit the absence of these conditions. In these outer
spheres, it is true that @B L
M
@LC and @L L
M
@LC, as one would expect to be
the case.
Of course, the causal claims about the person can be interpreted in terms of a dif-
ferent causal model. By changing the example, we can make a dierent causal model
more salient. Consider an example given by Lewis, in which we are to suppose that
there are gods who take a keen interest in human aairs:
It has been foretold that the event of your death, if it occurs, will somehow have a momentous
impact on the heavenly balance of power. It will advance the cause of Hermes, it will be a ca-
tastrophe for Apollo. Therefore Apollo orders one of his underlings, well ahead of time, to see
to it that this disastrous event never occurs. The underling isnt sure that just changing the time
and manner of your death would suce to avert the catastrophe; and so decides to prevent
your death altogether by preventing your birth. But the underling bungles the job: You are
born, you die, and its just as catastrophic for Apollo as had been foretold. When the hapless
underling is had up for charges of negligence, surely it would be entirely appropriate for
Apollo to complain that your birth caused your death. (Lewis, Causation as Inuence,
chap. 3 of this volume, p. 101)
It would, indeed, be appropriate for Apollo to make this causal claim. But notice
how the causal model has changed, with a dierent kind of system and dierent laws
of functioning involved. Indeed, this example is better seen as exemplifying subcase I
rather than II, as the eld of normal conditions F
M
generated by the relevant model
entails the factor to be explained, namely, that you will die (D). This is represented
diagrammatically in gure 6.4. Nonetheless, your birth (B) does make a dierence to
your death in view of the fact that Apollos underling could have intervened in the
system (I) to prevent your birth and so your death. Here the two counterfactuals
required for your birth to be a dierence-maker for your death, that is, B L
M
D and
@B L
M
@D, both hold true with respect to the sphere of normal worlds generated by
the relevant model. But it does require a radical change of causal model to get this
result.
The present framework provides a ready explanation of the two forms of contex-
tual relativity underlying the commonsense causesconditions distinction. It is easy
172 P. Menzies
to see how the explanation should work for the presence of oxygen case (example 2).
Here our readiness to rank the presence of oxygen as a condition in one situation and
a cause in another simply reects the fact that the two situations involve dierent
kinds of systems with dierent initial conditions. In the rst situation, in which a re
takes hold of a building, the presence of oxygen is an initial condition, held xed in
the eld of normal conditions, whereas in the second situation, in which oxygen is
excluded from a delicate experimental or manufactory setup, it is not an initial con-
dition, and so it is eligible to be a dierence-maker for the eect.
The explanation of the relativity of the causesconditions distinction to the context
of inquiry is equally straightforward. In the ulcerated stomach case (example 3), the
causal claims made by the dierent inquirers are explained by the fact that they are
employing dierent models. For example, the person who prepares meals for the
patient implicitly employs a model that focuses on the person with the ulcerated
stomach as a xed initial condition. On the other hand, the doctor tacitly employs a
model that focuses on the person as a normally functioning human without ulcerated
stomach as a xed initial condition. These inquirers both seek factors that make a
dierence to the patients indigestion, but they do so with respect to the dierent
spheres of normal worlds generated by their dierent models.
Another virtue of the present framework is that it allows us to discriminate be-
tween absences as causes and absences as mere conditions. In allowing absences and
other nonoccurrences as causes, we need not open the oodgate to a host of spurious
causes. For instance, the natural causal model for interpreting the multiple omissions
case (example 5) ranks the gardeners omission as the cause of the plants death,
Figure 6.4
while backgrounding other peoples omissions as mere conditions. This causal model
is one that takes, as its system for investigation, a healthy plant functioning under
a regime of regular watering by the gardener. Accordingly, this example can be seen
to exemplify subcase II, in which the eld of normal conditions F
M
entails that the
gardener waters the plant (GW) and the plant survives (PS). The situation is rep-
resented in gure 6.5. When the plant fails to survive, an explanation in terms of
a dierence-maker is required. The gardeners actual omission can act as such a
dierence-maker, since the two appropriate counterfactuals hold true, GW L
M
PS
and @GW L
M
@PS. (Notice that in this case what explains the abnormal occur-
rence of the plants death is actually a nonoccurrence. It seems that common sense
sometimes regards a nonoccurrence as an interfering factor that perturbs the normal
course of development of some kinds of systems.) However, the omission by every-
one else (@SW) to water the plant is disqualied from acting as a dierence-maker,
as this omission is held xed in the eld of normal conditions. (However, an outer
sphere allows it to be the case that someone else waters the plant and in this outer
sphere it holds true that SW L
M
PS.)
A similar explanation can be given of our causal judgments about the absence of
nerve gas case (example 4), when it is interpreted as exemplifying subcase II. As
a matter of fact, this is a rather strained interpretation, as one has to imagine a eld
of normal conditions generated by an appropriate model that entails that I am not
writing at my computer. However, if we do imagine this, then given that I am so
writing, it is reasonable to ask for some explanation in terms of a dierence-maker.
Figure 6.5
174 P. Menzies
But this dierence-maker cannot be supplied by the absence of nerve gas, the absence
of amethrower attack, or absence of meteor strikefor these absences are held
xed in the eld of normal conditions, in view of the fact that the intrusion of nerve
gas, or amethrower attack, or meteor strike would count as an interference in the
system.
It is much easier to understand this example as exemplifying subcase I; that is,
the eld of normal conditions generated by the relevant model entails the actually
obtaining factormy writing at my computer. However, a striking fact emerges
when one construes the example in this way. The various absences mentioned above
can each count as a dierence-maker for my writing. For instance, the intrusion of
nerve gas into the situation in which I am writing at my computer is naturally re-
garded as an interference, whose absence should be held xed in the eld of normal
conditions. However, the rules for a models generating spheres of worlds (in par-
ticular, the weightings principle) permit the presence of the nerve gas in an outer
sphere of worlds. The consequence of this is that the two counterfactuals required for
the absence of nerve gas (@NG) to count as a dierence-maker for my writing (W)
can both hold. Figure 6.6 represents the situation in which the two counterfactuals
NG L
M
@W and @NG L
M
W hold. The same line of reasoning shows that the ab-
sence of any factor that could be regarded as an interferer can, in the right circum-
stances, act as a dierence-maker for my writing at my computer, when the example
is construed as exemplifying subcase I rather than II. I conjecture that it is not
absurd to judge in these circumstances that a sustaining cause of my writing at my
Figure 6.6
computer is the collective absence of all interfering factors, including the absence of
nerve gas, the absence of amethrower attack, and the absence of meteor strike.
In the discussion above, I have argued that the commonsense distinction between
causes and conditions makes sense only relative to a eld of normal conditions gen-
erated by a causal model. Given such a model, the conditions of some eect can be
explained as those factors belonging to the eld, and the causes as those factors that
make the dierence to the eect relative to this eld. It is worth comparing this
characterization of the distinction with an alternative one that has become popular.
On this alternative characterization, the distinction is an entirely pragmatic one to
be cashed out in terms of contrastive explanation. We have already seen the out-
lines of this kind of approach in section 4. The main idea is that commonplace causal
judgments are implicit answers to contrastive why-questions of the form Why
does E
1
, rather than E
2
; . . . ; E
n
, obtain? where the members of the contrast class
fE
1
; . . . ; E
n
g may or may not be mutually compatible. On this approach, a cause is
an actual objective cause that dierentiates E
1
from the other members of the
contrast class; and the conditions are those factors that are common to all possible
situations that could realize a member of the contrast class. This approach also
makes the distinction a context-sensitive one because dierent contexts may contrast
E
1
with dierent sets of alternatives, so aecting which factors count as causes and
conditions.
There are, to be sure, similarities between these characterizations of the causes
conditions distinction. One obvious similarity is that they both characterize the dis-
tinction in terms of a contextually generated space of possibilities. In the present
framework, it is the space of normal worlds generated by a model; in the alternative
framework, it is the space of contrasting alternatives. Still, there are, in my view,
some serious shortcomings to this alternative characterization, some of which have
been touched on earlier.
One of these is that the characterization must operate with an independently
motivated notion of an objective cause. The factor that dierentiates E
1
from the
other members of the contrast class cannot be a causally irrelevant factor: it must
be an objective cause present in the E
1
situation, but not in the others. But this
requires an explication of what an objective cause is. The explication cannot, on
pain of unnecessary duplication, appeal to the idea of a dierence-maker. Another
shortcoming of the characterization is that it leaves radically underspecied what
conditions are common to the realizations of the dierent members of the contrast
class. A specication of these commonalities requires a description of a similarity
relation between the possible situations realizing the various members of the contrast
class. It is totally unclear what this similarity relation involves. Lewis attempts, as we
176 P. Menzies
have seen, to specify a similarity relation by appeal to backtracking reasoning, but
this attempt fails to deliver determinate verdicts in many cases.
For these reasons, I believe, the alternative characterization of the causescondi-
tions distinction in terms of contrastive explanation cannot bear the explanatory
weight that many have placed on it. Why then, it may be asked, does the account of
the phenomena in terms of contrastive explanation, sketched in section 4, work as
well as it does? There are several reasons, I would suggest. First, the treatment of
causal judgments as answers to contrastive why-questions puts the emphasis in the
right place, namely, on the context-relativity of these judgments. Second, the speci-
cation of the contrast class, embodied in a contrastive why-question, carries infor-
mation about the kind of system that is being investigated and its laws of normal
functioning. In other words, we can read o from a class of contrasting alternatives
information about the real contextual determinant of our causal judgmentsthe
underlying causal model. But the contrast class is, at best, an indirect source of this
information.
So, I oppose the popular strategy of explaining the commonsense view of causes
as dierence-makers pragmatically in terms of contrastive explanation. Rather, I
recommend the reverse procedure of explaining contrastive explanation in terms of
the present independently motivated account of dierence-making. Let me outline in
broad detail how such an explanation should work. Suppose the contrastive why-
question Why does E
1
, rather than E
2
; . . . ; E
n
, obtain? has been posed, where the
members of this contrast class are all actual or possible factors of systems of kind K
operating according to laws L. Then, a satisfactory answer to this question should
cite some actual factor C that makes a dierence to E
1
relative to the model M
hK; Li. Where the members of the contrast class are incompatible outcomes in the
same system, it follows from the denition of C as a dierence-maker for E
1
that we
have an automatic explanation why none of the alternative possible outcomes could
have occurred. Similarly, where the members of the contrast class are compatible
outcomes of dierent systems of kind K, it follows from the denition of C as a
dierence-maker for E
1
that we have an automatic explanation why none of the
alternative outcomes actually occurred (for, given that these have the same initial
conditions and conform to the same laws, these systems would have to have E
1
if
they had the factor C).
This account of contrastive explanation overcomes the diculties facing Lewiss
account that we encountered in section 4. For example, it does not leave indetermi-
nate what the various members of the contrast class have in common. It species
these commonalities precisely in terms of the eld of normal conditions generated by
the relevant causal model. Again, it handles the examples such as Hempels, where
a contrastive explanation is required of two compatible alternatives. We explain
why Jones, rather than Smith, got paresis, by citing the dierence-making factor of
syphilis that actually applies to Jones, but not to Smith. Most important, this ac-
count does not involve an unnecessary duplication of the idea of a dierence-maker.
It follows from the denition of dierence-making that, if we have factor C that
makes a dierence to E
1
, then we have a contrastive explanation of why E
1
rather
than E
2
; . . . ; E
n
occurred. This factor does its work of dierentiating the contrasting
alternatives precisely because it is, by hypothesis, a dierence-maker for E
1
.
8 Conclusion
One of the aims of this essay has been to explore the conception of a cause as
what makes the dierence in relation to some assumed background or causal eld
(Mackie 1974, p. 71). I have tried to explicate this conception by giving an account
of dierence-making in terms of context-sensitive counterfactuals. This account
explains the way in which we distinguish causes from background conditions in terms
of the way in which an implicit contextual parameter of a causal model generates
a similarity ordering among possible worlds. It is important to elucidate this dimen-
sion of context-sensitivity in our causal judgments not only to get the conceptual
analysis of causation right, but also to avoid philosophical puzzles that arise from
conceptions of causation that are too simplistic. For example, the puzzle in the phi-
losophy of mind about mental causationthe puzzle of how mental states can play
a role in the causation of behavior independent of the role played by the physical
states on which they supervenearises because philosophers overlook the way in
which causal models implicitly guide our judgments about causation, or so I argue
(in Menzies 2002).
Notes
1. I have considered some of the problems faced by the counterfactual approach to causation in connec-
tion with preemption cases in Menzies (1996). There I argue that cases of preemption show that a purely
counterfactual analysis of causation will not work: at some point we must make an appeal to a concept of
causation as an intrinsic relation or process in order to deal with them. (See also Menzies 1999.) I show
how to marry counterfactual intuitions about causation with intuitions about intrinsic processes by way of
a RamseyCarnapLewis treatment of causation as a theoretical relation. On this treatment, a counter-
factual dependence is a defeasible marker of causation: when the appropriate conditions are satised, it
picks out the process that counts as the causal relation. Such a treatment of causation as a theoretical re-
lation obviously can be framed around the counterfactual explication of the idea of a cause as a dierence-
maker recommended in this chapter.
178 P. Menzies
2. Actually, Lewis has presented three theories of causation, the third being the quasi-dependence theory
that he tentatively sketched in postscript E to the paper Causation (Lewis 1986b).
3. Mackie credits the term eld of causal conditions to his teacher John Anderson, who used it to resolve
diculties in Mills account of causation in Anderson (1938).
4. I take the terms for the dierent kinds of context relativity from Gorovitz (1965).
5. It might be argued that these examples demonstrate only that the construction c is the cause of e
is context sensitive; and that the counterfactual theory is best understood as an account of the context-
insensitive construction c is a cause of e. Thus it might be argued that even the World Food Authority
would admit that the drought was a cause of the famine and that the doctor would allow that eating
parsnips was a cause of the indigestion. While this defense on the basis of common usage seems faintly
acceptable with some examples, it fails in other cases. Even with a liberal understanding of the words, it
seems a stretch to say that a persons birth and possession of lungs were among the causes of his lung
cancer. It seems that even the expression a cause displays some degree of context sensitivity. For dis-
cussion and elaboration of this point see Unger (1977).
6. Mill wrote: Nothing can better show the absence of any scientic ground for the distinction between
the cause of a phenomenon and its conditions, than the capricious way in which we select from among the
conditions that which we choose to denominate the cause. See Mill (1961), p. 214.
7. The clearest exponent of this strategy of adding on a theory of contrastive explanation to a theory of
objective causation is Peter Lipton: see his (1990) and (1991), chapters 35. Lipton does not, however,
endorse Lewiss counterfactual theory as the right theory of objective causation. Others who have
emphasized the importance of contrastive explanation for understanding ordinary causal discourse include
Gorovitz (1965), Mackie (1974), Dretske (1973), Woodward (1984), and Hitchcock (1996b).
8. It is worth observing that the idea of an integrated account of context-sensitivity is not altogether foreign
to Lewiss style of counterfactual theory. In the revamped (2000) version of the theory, context enters the
theory in an important but inconspicuous way. The notion of a not-too-distant alteration of the cause
introduces an important contextual element into the truth conditions of causal statements. A not-too-
distant alteration of the cause is an alteration that is relevantly similar to the cause by the standards de-
termined by the context. The approach I shall advocate is similar in building context-sensitivity into the
truth conditions, but it will draw on contextually determined standards of similarity for counterfactuals
rather than events.
9. I develop a fuller account of factors in Menzies (1989b), though that paper uses the term situations
rather than factors.
10. A problem infects these denitions that is parallel to the problem Lewis pointed out for Kims deni-
tion of the simple concepts. Modifying some concepts Lewis introduced, let us say that a system S is
accompanied if and only if it coexists with some contingent object wholly distinct from it, and lonely if and
only if it does not so coexist. The denitions I have presented amount to saying that the extrinsic properties
of a system are those implied by the accompaniment of the system and the intrinsic properties of a system
are those compatible with its loneliness. The problem is that loneliness of a system is intuitively an extrinsic
property of the system (since it can dier between duplicates of the system), but it counts as an intrinsic
property by the denition (since it is compatible with itself ). One possible remedy to this problem may be
to adapt to our purposes the renement of Kims original idea to be found in Langton and Lewis (1998).
This renement is supposed to circumvent the defect Lewis detected in Kims original idea.
11. The concepts of relations that are extrinsic or intrinsic to a set of objects can be dened in a similar
manner. A relation R is extrinsic to a set of objects if and only if, necessarily, the relation holds between
two members of the set only if some other contingent object wholly distinct from the set exists. Conversely,
a relation R is intrinsic to the set if and only if, possibly, the relation holds between two constituents of the
system although no contingent object wholly distinct from the set exists.
12. In using the terms sphere of normal worlds and eld of normal conditions, I am not invoking the
ordinary notion of normal. Rather, the given denitions stipulate the intended sense in which I use the
terms, though I hope this sense bears some relation to the ordinary notion of normal.
13. For an illuminating discussion of the role of zero-force laws, such as the rst law of motion and the
HardyWeinberg law, in default assumptions about the evolution of systems, see Elliott Sobers discussion
(1984, chap. 1).
14. See, for example, Lewiss discussion of the perils of allowing backtracking counterfactuals in the
counterfactual analysis of causation in his postscripts to Counterfactual Dependence and Times Arrow
and Causation in his (1986a).
15. For an account of the similarity relation for nonbacktracking counterfactuals that appeals to the
miraculous realization of antecedents, see Lewis (1979a).
16. For discussions of episodes in the history of science that highlight the signicance of causes as intru-
sions in the normal course of events, see Toulmin (1961), chapters 34; and Sober (1980).
180 P. Menzies
7Causation and the Price of Transitivity
Ned Hall
1 Introduction and Preliminaries
That causation is, necessarily, a transitive relation on events seems to many a bed-
rock datum, one of the few indisputable a priori insights we have into the workings
of the concept. Of course, state this Transitivity thesis so boldly and it will likely
come under dispute; one can reasonably worry that the appearance of a priori insight
glosses over a lack of cleverness on our part. Might not some ingenious counter-
example to Transitivity lurk nearby, waiting for a philosopher acute enough to
spot it?
Yes, comes the recent reply. In the last several years a number of philosophers,
myself included, have exhibited examples that appear to undermine Transitivity (al-
though, as well see, the target of my own examples was, and remains, quite dier-
ent). Ill review a representative sample in the next section. First I need to sketch the
main argument of the paper, and take care of some preliminaries.
I claim that the examples do have something to teach us about the metaphysics
of causation, but that it is emphatically not that Transitivity fails. Close inspection
reveals that they pose no threat to that thesis. But theyor rather, certain of them,
constructed with sucient caredo show that Transitivity conicts with the follow-
ing thesis:
Dependence Necessarily, when wholly distinct events c and e occur, and e counter-
factually depends on c (read: if c hadnt happened, e wouldnt have), then c is thereby
a cause of e.
Dependence should not be confused with the transparently false claim that coun-
terfactual dependence is necessary for causation; it is rather the claim that such de-
pendence is sucientat least, when the events in question are wholly distinct, so
that the relation of dependence does not hold merely in virtue of their mereological
or logical relationships.1 Nor is Dependence the claim that causation can be given a
counterfactual analysis, though most (if not all) counterfactual analyses endorse it.
One might hold that no analysis of causation is possible, and yet endorse Dependence.
Or again, one might hold that it is counterfactuals that demand an analysis in terms
of causation, and yet maintain Dependence. Whats more, Dependence surely has a
great deal of intuitive plausibility, at least when we are careful to read the counter-
factual in a suitable nonbacktracking sense, one that rules out such judgments as
that if c had not occurred, then it would have to have been the case that those earlier
events that were causally sucient for it did not occur.2 Given the plausibility of
Dependence, its conict with Transitivity is all the more striking.
In the end, Ill suggest that were we forced to choose, we should reject Dependence
in favor of Transitivity. But Ill also suggest that we arent forced to choose, since we
can perfectly well recognize dierent kinds of causal relation in which events can
stand, so that an obvious resolution of the conict between Dependence and Tran-
sitivity presents itself: The kind of causal relation characterized by the rst thesis is
not the same asindeed, fundamentally diers fromthe kind of causal relation
characterized by the second. Still, a manifest asymmetry distinguishes these causal
relations, in that the kind of causal relation for which Transitivity holds is clearly the
more central of the two. So the price of transitivitya price well worth payingis
to give up on the claim that there is any deep connection between counterfactual de-
pendence and (the central kind of ) causation.
Here is the plan for getting to these conclusions: First Ill run through some of
the alleged counterexamples to Transitivity, and highlight their apparent common
structure (sec. 2); examples with this structure I call short-circuits. An immediate
problem will emerge for my claim that Dependence is the culprit, and not Tran-
sitivity; for while that claim neatly deects some of the counterexamples (call these
the Easy ones), it does not help deect the others (call these the Hard ones).
But the Easy and Hard cases share the short-circuit structure, and whats worse,
it seems that it is in virtue of this structure that they are (or at least appear to be)
counterexamples to Transitivity. So targeting Dependence seems quite mistaken.3
Not so, but it takes some work to see why. I start by sharpening the challenge the
Hard cases pose to Transitivity by distinguishing three dierent problems for that
thesis, arguing that they arent deep problems, and observing that their solutions help
not at all in dealing with the Hard cases, at least if those cases are formulated care-
fully enough (sec. 3). The main argument follows: In section 4 Ill introduce a kind
of causal structure that I call a switch, in which one event c interacts with a causal
process in such a way as to redirect the causal route by which that process brings
about a given event e. This structure diers in obvious ways from the apparent
common structure of the alleged counterexamples. It also might seem to pose trouble
of its own for Transitivity, for reasons that I explain. Ill argue that the trouble is
illusory, and that a number of considerations tell in favor of counting c a cause of
e, when c is a switch with respect to e. That argument sets the stage for section 5,
where I revisit the alleged counterexamples to Transitivity, and show that what dis-
tinguishes the Easy cases from the Hard ones is exactly that the Hard ones, in addi-
tion to possessing the short-circuit structure, also possess a (much less noticeable)
182 N. Hall
switching structure. So the intuitive verdictthat they provide us with cases in which
c is a cause of d, and d of e, but not c of eis mistaken: c is a cause of e, in virtue
of being a switch with respect to e. Since the Easy cases do not exhibit the switch-
ing structure, no such maneuver helps neutralize their threat to Transitivity. But
the original, obvious optiontreat them as counterexamples only to Dependence
remains available, and should be taken.
I close, in section 6, by highlighting two dierent lessons we learn from the exam-
ples. The rst is metaphysical: What the Easy cases show us is that Transitivity and
Dependence conict. Of course, that verdict leaves it open which one must go; Ill
sketch a number of reasons to hold onto Transitivity at the cost of Dependence. The
second is methodological: Briey, the Hard cases remind us that the business of
mining for intuitions about causation is risky. It is all too easy to mistake fools gold
for the real thing, and one must therefore be careful to subject ones hypothetical
cases to careful analysis before trying to buy fancy conclusions with them. The Easy
cases stand up under such analysis; the Hard cases dont.
2 The Counterexamples and Their Apparent Common Structure
Here is an example from Michael McDermott (1995a): A man plans to detonate a
bomb. The day before, his dog bites o his right forenger, so when he goes to press
the button he uses his left forenger instead. Since he is right-handedand so would
have used his right forengerthe dogbite causes his pressing of the button with his
left forenger. This event in turn causes the subsequent explosion. But, intuitively,
the dogbite does not cause the explosion.
Here is another example, adapted from one devised by Hartry Field: An assassin
plants a bomb under my desk (philosophy is a dangerous business, you know). I nd
it, and safely remove it. His planting the bomb causes my nding it, which in turn
causes my continued survival. But, intuitively, his planting it does not cause my con-
tinued survival.
Next, one from Igal Kvart (1991b): A mans nger is severed in a factory accident.
He is rushed to the hospital, where an expert surgeon reattaches the nger, doing
such a splendid job that a year later, it functions as well as if the accident had never
happened. The accident causes the surgery, which in turn causes the nger to be in a
healthy state a year later. But, intuitively, the accident does not cause the nger to be
in a healthy state a year later.
Finally, a refreshingly nonviolent example of my own: Billy and Suzy are friends.
Suzy is mischievous, and Billy is forever trying to keep her out of trouble. Billy sees
Suzy about to throw a water balloon at her neighbors dog. He runs to try to stop
Causation and the Price of Transitivity 183
her, but trips over a tree root and so fails. Suzy, totally oblivious to him, throws the
water balloon at the dog. It yelps. She gets in trouble. Billys running toward Suzy
causes him to trip, which in turn causes the dog to yelp (by Dependence: If he hadnt
tripped, he would have stopped her from throwing and so the dog wouldnt have
yelped). But, intuitively, his running toward her does not cause the dogs yelp.
This last example is an Easy caseEasy because it is so obvious how to respond
to it in a way that safeguards Transitivity. After all, the only sense in which Billys
trip causes the dogs yelp is that it prevents somethingBilly continuing to run
toward Suzy, reaching her in time to stop her from throwing the balloon, and so
onwhich, had it happened, would have prevented the yelp. But no causal process
connects the trip to the yelp (and not because the connection is an example of magi-
cal action-at-a-distance). So we are well within our rights to deny that, in this and
similar cases of double-prevention, the sort of dependence the yelp has on the trip
is causal dependenceand if so, the example poses no threat to Transitivity. Or,
more cautiously, we might allow that the trip is, in some sense, a cause of the yelp
but that it is not this rather nonstandard sense of cause we have in mind when
we assert Transitivity, but rather the ordinary, garden-variety sense.4 Still, no threat
to Transitivity. The casualty, rather, is Dependence: It is either false, or must be taken
to be true only of a sort of nonstandard, second-class kind of causation.
But this response falls limp when it comes to the three other cases. Take Kvarts,
for example: Its not merely that the healthy state of the nger counterfactually de-
pends on the surgery; no, a perfectly ordinary causal process also connects the two.
Likewise, a perfectly ordinary causal process connects the accident to the surgery. So
how can we blame Dependence, and not Transitivity, for the counterintuitive result?
Worse, it might seem that Ive misdiagnosed my own Easy case. For observe that
all four examples have the following salient causal structure in common: An event
c occurs, beginning (or combining with other events to begin) some process that
threatens to prevent some later event e from occurring (call this process Threat).
But, as a sort of side-eect, c also causes some event d that counteracts the threat
(call this event Savior). So c is a cause of Savior, and Saviorby virtue of
counteracting Threatis a cause of e. Butor so it seems to many philosophers
c is not thereby a cause of e, and so Transitivity fails. That diagnosis seems to apply
equally to all four examples. So shouldnt we favor it over a diagnosis that only ts
the last of the examples?
No. The correct picture is a more complicated one, according to which the last
example deserves exactly the diagnosis I gave it, and the rst three Hard cases de-
serve a quite dierent diagnosis. And the reason is that the Hard cases have a quite
dierent causal structure from the last, Easy case.
184 N. Hall
Bringing out this dierent structure will take some care. As a preliminary, I will
contrast the foregoing alleged counterexamples to Transitivity with three very dif-
ferent kinds of counterexample. We will see that these other counterexamples are
relatively benignin that they leave ample room for a mildly qualied version of
Transitivitybut that they are also quite unlike the apparently more virulent strain
exhibited above.
3 Contrast Class: Benign Counterexamples to Transitivity
Take any event efor example, my recent pressing of the letter t on my keyboard.
Begin to trace its causes. Its immediate causes are relatively few, and near to hand:
some signals in neurons in my brain, arm, and hand, the prior presence of that bit of
the keyboard, and so on. As we trace back further, though, the causes become more
numerous, scattered, and miscellaneous: They will likely include my birth, as well as
the births of all those who were instrumental in the design and manufacture of this
keyboard; if so, the number of more remotely ancestral births that count as causes of
my simple act of typing becomes truly staggering. And we havent even considered
those parts of es causal history that concern more than its merely human aspect. Go
back far enough, and it may be that every event (or at least: every event that takes
place at that stage of es backward light cone) is a cause of e.
One might balk at calling all such events causes of e. Surely they are not as much
causes of e as are its most proximate causes. Fair enough. The simplest way to ac-
commodate this intuition is to say that causation comes in degrees, so that the more
proximate a cause is to a given eect, the greater the degree to which it is a cause of
that eect. Perhaps, then, what happens as we travel backward down es causal his-
tory is that the degree of causal inuence of the events we nd gradually diminishes;
perhaps it eventually diminishes to the point where we must say that we no longer
have causes of e at all. Such a picture of causation is harmless enough, though it
raises the question of how to measure the attenuation of causal inuence along a
causal chain. Assuming that question settled, the needed modication to Transitivity
is obvious. Ideally, it will have the following form: When c is a cause of d to degree
x, and d is a cause of e to degree y, then c is a cause of e to degree f x; ywhere
the function f returns values lower than, but close to, the lesser of x and y. Assum-
ing some low threshold value a such that for b < a, is a cause of e to degree b no
longer implies is a cause of e, Transitivity will still hold for choices of c; d, and e
where the values x and y lie well above the threshold. Or so we can hope the story
would go.
At any rate, this issue, however interesting, has no bearing on our topic, as the
examples discussed in the last section purported to exhibit failures of transitivity at
early (backwardly speaking) stages of an events causal historytoo early to pin the
blame on the sort of gradual attenuation of causal inuence under discussion here.
Two other kinds of counterexample to Transitivity focus on the possibility of dis-
connect between the parts of d that c causes and the parts of d that cause e. It might
be that c counts as a cause of d by virtue of causing one part of d, but that d counts
as a cause of e by virtue of the causal action of a dierent part. If so, it need not
follow that c is a cause of e.
What distinguishes the two kinds of counterexample is the notion of part at issue.
Begin with the ordinary, mereological notion. A dance performance, for example,
has as distinct parts the performances of each dancer. Let d be such a performance,
and suppose that it has parts d
1
and d
2
, consisting of the performances of Billy and
Suzy, respectively. Among the causes of d
1
is, let us suppose, Billys joining the dance
troupe several months earlier. And among the eects of d
2
is, let us suppose, the ap-
pearance of a favorable review in the next mornings paper (for Suzy is the star of the
show). We might count Billys joining as a cause of the dance, in virtue of its role in
causing one part of the dance (namely, Billys). And we might count the dance as a
cause of the favorable review, in virtue of the eect of one part of it (Suzys brilliant
performance). But we would notor at least, not therebycount Billys joining as
a cause of the favorable review. (E.g., suppose hes a lousy dancer.) If so, we need
to modify Transitivity once again, presumably by introducing a distinction between
primary and derivative senses of cause, so that (for example) it is only in the de-
rivative sense that Billys joining the dance troupe is a cause of the performance, and
the performance of the favorable review, whereas it is only in the primary sense that
causation is transitive. Again, interesting but irrelevant, for none of our counter-
examples takes advantage of this loophole.
Finally, L. A. Paul has highlighted a thirdand arguably much more signicant
way in which Transitivity must be qualied. (See her Aspect Causation, chapter
8 in this volume.) Put abstractly, we might have a situation in which c causes d to
have a certain property or (to use Pauls terminology) aspect A, and d in turn causes
e, in virtue of some other aspect B, but intuitively c is not a cause of e. Arguably,
McDermotts counterexample provides a case in point: The dog-bite does not
cause the button-pushing per se, but rather causes it to have a certain aspectthat
is, causes it to be a button-pushing with the left hand. Moreover, the button-pushing
causes the explosion, but simply in virtue of being a button-pushing; that it is a
button-pushing with the left hand is causally irrelevant to the explosion. Paul argues
that in cases like this, Transitivityproperly construedsimply doesnt apply: In
186 N. Hall
order for it to apply, we must not merely have a single intermediate event to serve as
a link between c and e, but must have a single intermediate event-cum-aspect.
I think that is a very important point, and I highly recommend Pauls expert
treatment of it. But observe that there is no hope of applying it uniformly to cases
with the short-circuiting structure. Consider for example Kvarts case: We cannot
single out some aspect of the surgery, and argue that (i) the injury causes the surgery
to have that aspect, but (ii) it is only in virtue of some other aspect that the surgery
causes the nger to be healthy. Or again, a simple modication of McDermotts
example renders it immune to Pauls treatment: Suppose that after the dog-bite, the
man does not push the button himself but orders an underling to do so. The relevant
intuitions do not change: the dog-bite causes the order, and the order causes the ex-
plosion, but the dog-bite does not cause the explosion. The only way I can see to
apply Pauls observations is by way of a rather strained insistence that there is one
eventcall it a making the button be depressedwhich the dog-bite causes to
have the aspect being an order, and which otherwise would have had the aspect
being a button-pushing. That resolution is unattractive enough to warrant the
search for an alternative. So while I certainly think that Pauls observation is very
useful in a wide range of casesindeed, Ill put it to use myself, in the next section
I dont think it provides the means for an adequate response to the challenge raised
by the Hard cases. In fact, I think the right response has a quite dierent shape:
Transitivity does apply to those cases, and yields the correct conclusionintuition
notwithstandingthat c is a cause of e. The next section provides the crucial under-
pinnings for this response.
4 Switches
A distinctive relationship that an event c can bear to an event e is that of helping to
determine the causal route by which e is produced. In such cases, I call c a switch
with respect to e. Here is an example.
The Engineer An engineer is standing by a switch in the railroad tracks. A train
approaches in the distance. She ips the switch, so that the train travels down the
right-hand track, instead of the left. Since the tracks reconverge up ahead, the train
arrives at its destination all the same; let us further suppose that the time and manner
of its arrival are exactly as they would have been, had she not ipped the switch.
Let c be the engineers action and e the trains arrival. Pick an event d that is part
of the trains journey down the right-hand track. Clearly, c is a cause of d and d of e;
but is c a cause of e? Is her ipping the switch a cause of the trains arrival? Yes, it is,
though the opposing reaction surely tempts. After all, goes this reaction, isnt it
clear that the switching event makes no dierence to whether the train arrives, but
merely determines the route by which it arrives?
To some extent we can accommodate this reaction: Yes, of course the switching
makes no dierence to the arrival, in the sense that had it not occurred, the train
would have arrived all the same. And yes, of course the switching helps determine
the causal route to the arrival. But it goes too far to conclude that the switching is
not itself among the causes of the arrival. A number of considerations reinforces this
conclusion; since the issue is crucial to the proper diagnosis of the Hard cases, it will
pay to review these considerations with some care.
The Explanation Test Ask for a complete explanation of the trains arrivalHow
did it get there? What processes led to its arrival?and a complete answer must
include the switching event. But, plausibly, this complete answer will simply consist
in a description of the arrivals causal history.
The Subtraction Test Remember that the train travels down the right-hand track.
Certain rather boring events extraneous to this causal process are the persistence
through time of the various bits of left-hand track; these are events that would have
combined with the presence of the train to cause its arrival, had it traveled down the
left-hand track. Suppose some of these events away; that is, subtract them (you
neednt go so far as to replace them by pure void; air and dirt will do). Subtract
enough of them, in fact, that the left-hand track ends up having a sizable gap in it.
Plausibly, this subtraction should not alter the facts about the arrivals causes, since
the events subtracted are quite separate from and extrinsic to the events that consti-
tute the causal processes leading up to the arrival.5 But in the altered situation, the
engineers action quite obviously helps bring about the trains arrival; after all, she
steers the train onto the only track that will get it to its destination. So, given that
subtracting extraneous events should not alter the causal history of the arrival, we
should also say that in the original case her switching is part of this causal history.
More generally, when a switch alters the causal pathway of some process M, there
will typically be other processes that M would have interacted with to bring about the
eect e, but thatthanks to the switchit does not interact with. Since the events
that go into these other processes are, as things stand, causally idle features of the
environment, it seems plausible that we ought to be able to remove them from the
environment without altering the causal status of the switch c with respect to the ef-
fect e. But in the scenarios where they are absent, it becomes perfectly clear that c is a
cause of e.
188 N. Hall
The Variation Test The example had a little intuition-pump built into it, given that
we stipulated that if the engineer had not thrown the switch the train would have
arrived at exactly the same time and in exactly the same manner. Suppose we had
told the story slightly dierently, so that if the engineer had not thrown the switch,
the time and/or manner of the arrival would have been dierentsay, the train
would have arrived an hour later. This alteration reverses, I think, the intuitive ver-
dict (such as it was) that the switching is not a cause of the arrival. But that verdict
should not be so sensitive to the details of the casewhich in turn suggests that the
no-dierence-whatsoever intuition-pump was misleading us.
Pauls Observation If the details of the trains counterfactual journey down the left-
hand track are suciently similar to the details of its actual journey down the right-
hand trackand its certainly natural to interpret the story as if this were truethen
the right thing to say might be that, at a suitably abstract (i.e., nondetailed) level of
description, the two journeys consist of exactly the same events. To borrow Pauls
terminology, what the switching does is make it the case that these events have cer-
tain aspects: Specically, the switching makes it the case that they are travelings-
down-the-right-hand-track rather than travelings-down-the-left-hand-track. On such
an analysis, no two-step chain of the right sort connects the switching to the ar-
rival, and so Transitivity simply does not apply. Perhaps our tendency to think of
the actual and counterfactual journeys as really the same journey, dierently realized,
lies behind whatever reluctance we might feel to calling the switching a cause of the
arrival.
As a test, hold the details of the arrival xed, but alter the extraneous events so
drastically that the way the train gets to its destination, in the counterfactual situa-
tion in which it travels down the left-hand track, is completely dierent from the way
it in fact gets to its destination: It stops after a short while, gets taken apart, shipped
piecemeal to a point near its destination, reassembled, and all this in such a way as
to guarantee that nothing distinguishes its counterfactual from its actual arrival.
Then Pauls observation doesnt apply, since there is no event that can happen as a
traveling-down-the-right-hand-track, and as a shipping-of-disconnected-parts. But,
by the same token, there is much less temptation to deny that the switching is a cause
of the arrival.
Salience Notice, rst, that among the (highly nonsalient) causes of the trains ar-
rival is the presence of a certain section S of the track down which the train travels,
during the time that the train is in fact on S. And among the causes of the presence of
S at this time is the presence of S a day earlier. So we should say: Among the causes
of the arrival is the presence of S a day earlier. Thats certainly odd, but its odd in
the familiar way that writers on causation have grown accustomed to: Whatever the
exact nature of the causal concept we are trying to analyze, it surely picks out a very
permissive relation, one that doesnt distinguish between events that we would nor-
mally single out as causes and events we would normally ignore because their causal
relationship to the eect in question is too boring or obvious to be worth mentioning,
or is easily hidden from view as part of the background conditions. The presence
of S a day earlier is like that.
But the relationship that the switching event bears to the arrival is just like the re-
lationship that the earlier presence of S bears to the arrival: The relationship that the
switching event bears to the setting of the switch as the train passes over it is just like
the relationship that the earlier presence of S bears to the later presence of S, and the
relationship the setting of the switch as the train passes over it bears to the arrival is
just like the relationship the later presence of S bears to the arrival.
Switching per se Is Not Causing The last paragraph helps bring out something
rather important, which is that the switching bears two distinct causal relationships
to the arrival. The rst is the highly nonsalient one just outlined: The switching event
causes the setting of the switch, which interacts with the passing train in the same
way as the mere presence of a piece of track interacts with the passing train, etc. The
second is the narratively more vivid relation: The engineers action makes a dier-
ence to the causal route by which the arrival happens. But it is only in virtue of the
rst relation that the switching is a cause of the arrival. To see this, consider how the
two relations can come apart, as in the following example: Billy and Suzy are about
to throw rocks at a bottle. Suzy, the quicker of the two, is winding up when suddenly
her muscle cramps, stopping her from throwing. So Billys rock gets there rst. Had
there been no cramp, Suzys rock would have struck the bottle before Billys, break-
ing itin which case Billys rock would have struck only empty air. Either way, the
bottle breaks; what the cramp does is to determine the causal route to the breaking
(i.e., whether it happens via Suzys throw, or via Billys). But the cramp is manifestly
not a cause of the breaking. What this shows is that switching per se is not causing;
that is, it does not follow from the fact that an event c determines the causal route to
an event e that c is among es causes.
But recall that the kind of switch I am focusing ona kind of which the engineers
action is an instanceis an event c that interacts with a causal process in such a way
as to redirect the causal route by which that very process brings about a given event
e. I claim that in cases of this kind of switchingcall them cases of interactive
switchingc is a cause of e. But that will be so in virtue of cs interaction with the
causal process that results in e, and not merely because c is a switch with respect to e.
190 N. Hall
And this fact, in turn, helps account for our reluctance (such as it is) to count the
switching as a cause of the arrival: We naturally focus on the narratively most salient
relationship that the rst event bears to the secondnamely, that of helping to de-
termine the causal routeand, recognizing that this relationship does not suce for
causation, conclude that there is none to be had.
An obvious test of this hypothesis is to construct a case where the causal role is the
vivid one, and the switching role is not. Here is such a case:
The Kiss Billy and Suzy have grown up. One day, they meet for coee. Instead of
greeting Billy with her usual formal handshake, however, Suzy embraces him and
kisses him passionately, confessing that she is in love with him. Billy is thrilledfor
he has long been secretly in love with Suzy, as well. Much later, as he is giddily
walking home, he whistles a certain tune. What would have happened had she not
kissed him? Well, they would have had their usual pleasant coee together, and af-
terward Billy would have taken care of various errands, and it just so happens that in
one of the stores he would have visited, he would have heard that very tune, and it
would have stuck in his head, and consequently he would have whistled it on his way
home, and just to give the case the right shape let us stipulate that matters are
rigged in such a way that the time and manner of the whistling would not have dif-
fered at all. So the kiss is an event that interacts with a certain processcall this
process Billys dayin such a way as to redirect the causal route by which that
process bring about a certain eect (the whistling). But even though there is the fail-
ure of counterfactual dependence typical of switching cases (if Suzy hadnt kissed
Billy, he still would have whistled), there is of course no question whatsoever that as
things stand, the kiss is among the causes of the whistling.
Those who still balk at calling the engineers action a cause of the trains arrival
should answer the following question: What is the relevant dierence between that
case and this one? None, so far as I can seethat is, none relevant to the question
whether the switching event c is a cause of the eect e. But there are clear and sig-
nicant dierences relevant to an explanation of why our intuitions about The
Engineer are less rm and settled than our intuitions about The Kiss. For in The
Engineer, it is the switching role of the engineers action that stands out, and not her
causal contribution to the arrival; and switching per se is not causing. But in The
Kiss, the switching role is quite obscuredin fact, in telling the story one must
forcibly draw attention to itwhereas the causal contribution of Suzys kiss couldnt
be more obvious.
To wrap up: The points canvased in this section (i) provide us with ample positive
reason to count interactive switches as causes; and (ii) provide us with several plausible
explanations of the source of any contrary intuitions. I conclude that the price of
accepting interactive switches as causes is quite small, certainly small enough to
be worth paying. Especially so since, as were about to see, doing so provides us the
means to disarm the Hard counterexamples to Transitivity.
5 Two Kinds of Short-circuits
Recall the apparent common structure of the counterexamples, common to both
Easy and Hard cases: One event c helps to initiate a Threat, which, if not stopped,
will prevent e from occurring. But c also causes a Savior event that in turn counter-
acts Threat. So c causes Savior, and Saviorby counteracting Threatcauses e. But
c does not thereby cause e. Or so the stories go.
Now, however, we are in a position to see that there are two very dierent ways
in which Threat might be counteracted. Ill bring them out by means of examples,
building on the case discussed in the last section.
Drunken Bad Guys The Bad Guys want to stop the Good Guys train from getting
to its destination. Knowing that the switch is currently set to send the train down
the left-hand track, the Bad Guy Leader sends a demolition team to blow up a sec-
tion of that track. En route to its mission, the team stops in a pub. One pint leads to
another, and hours later the team members have all passed out. The Good Guys,
completely oblivious to what has been going on, send the train down the left-hand
track. It arrives at its destination.
Clever Good Guys The Bad Guys want to stop the train from getting to its desti-
nation. Knowing that the switch is currently set to send the train down the left-hand
track, the Bad Guy Leader sends a demolition team to blow up a section of that
track. This time, its a crack team, not so easily distracted. However, the Good Guys
have gotten wind of the Bad Guys plans, and send word to the engineer to ip the
switch. She does so. The Bad Guy demolition team blows up a section of the left-
hand track, but to no avail: Thanks to the engineers action, the train travels down
the right-hand track. It arrives at its destination.
Let c be the Bad Guy Leaders sending the team on its mission. Let e be the trains
arrival. Let d
1
be the event, in Drunken Bad Guys, of the teams stopping in the pub.
Let d
2
be the event, in Clever Good Guys, of the engineers ipping the switch.
Events d
1
and d
2
are, in each case, the Savior events. (Of course we could have
chosen other events to play this role; nothing hinges on the particular choice made.)
Clearly, c, in each case, causes them. But it turns out to be entirely too quick to as-
192 N. Hall
sume that they have the same causal status with respect to e, for there is a world of
dierence between the ways that each event counteracts Threat.
In Drunken Bad Guys, what the Savior event does is to cut short the Threat pro-
cess. Notice, furthermore, that the way that Threat is cut short requires no causal
connection whatsoever between the Savior event (the teams stopping in the pub) and
the nal eect (the trains arrival). The intuition that cthe Bad Guy Leaders
sending out the teamis not a cause of ethe trains arrivalis quite correct; but
that fact poses no threat to Transitivity, for the intermediate event d
1
is likewise not a
cause of e. Dependence may say otherwisebut so much the worse for that thesis.
The way the Threat is counteracted in Clever Good Guys, by contrast, is entirely
dierent. The engineers ipping of the switch does not cut short Threat at all; on the
contrary, that process goes to completion, resulting in the destruction of a portion of
the left-hand track. Rather, what the Savior event does is to switch the causal process
leading to the trains arrival from a pathway that is vulnerable to Threat to one which
is immune to Threat. Quite obviously, such a method for counteracting Threat must
involve a switch, of the sort examined in the last section. Therefore, given the con-
clusions of that section, this way of counteracting Threat necessarily establishes a
causal connection between the Savior event and the nal eect (assuming, as I will
throughout this discussion, that the switch is an interactive switch).
What of the event c, the Bad Guy Leaders sending out the team? That too, in
Clever Good Guys, is a cause of the arrival, by virtue of being a cause of the engi-
neers ipping the switch. But isnt that counterintuitive? Counterintuitive enough
that we should reject the tacit appeal to Transitivity made in the foregoing by virtue
of clause? Well, yes, its counterintuitivebut there is excellent reason to think that
intuition is being misled, here. For what intuition naturally focuses upon is the nar-
ratively most vivid role that c plays in the story. And this is exactly the same short-
circuiting role that c plays in Drunken Bad Guys: It initiates a Threat to e, and also
causes that Threat to be counteracted. Intuition is perfectly right to insist that no
event c can be a cause of an event e merely by virtue of (i) doing something that
threatens to prevent e; and (ii) doing something that counteracts that very threat.
Indeed, that is the lesson of Drunken Bad Guys, for in that story the only causal re-
lationship c has to e is that of threatening to prevent e and simultaneously causing
that threat to be counteracted. Intuition stridently objects to calling c a cause of e, in
that case; and whats more, no argument can be found to persuade her to change her
mind. So she should be heeded.
But in Clever Good Guys, such an argument is available. For what we naturally
fail to noticegiven the narrative prominence of the short-circuiting roleis that c
bears another causal relation to e: namely, c begins a process that ultimately interacts
with the processes leading up to e in such a way as to alter the causal route they fol-
low to e. All the reasons for counting switches as causes apply here with equal force.
So we must patiently explain to intuition that she is being distracted by irrelevant
details of the case, and should reverse her verdict.
One way to see that these detailsspecically, the fact that c is a short-circuit with
respect to ereally are irrelevant is to compare Clever Good Guys to the following
case:
Mischievous Bad Guys This time, the Bad Guy Leader has no intention of stopping
the train from getting to its destination. But he knows the Good Guys are paranoid,
and likes to play tricks on them. Just for fun, he sends out the demolition teamnot
with orders to blow up the left-hand track, but with orders to go down to the pub
and enjoy themselves. Still, he knows full well that once the Good Guy Spies have
reported that the team has been sent out, the Good Guys will panic and order the
engineer to ip the switch. Sure enough, they do. The engineer ips the switch. The
train travels down the right-hand track. It arrives at its destination.
There is no short-circuit here, because there is no Threat to be counteracted. So in
that sense, Mischievous Bad Guys diers dramatically from Clever Good Guys. But
in the important sense, the two cases are exactly the same: The relevant relationship
the Bad Guy Leaders action bears to the arrival diers not at all. Whats more, there
is no question (not, at least, once we absorb the lessons of the last section) that in
Mischievous Bad Guys, the Leaders action is a cause of the arrivala verdict we
should therefore carry over to Clever Good Guys. The fact that in this latter case, the
Leaders action is also a short-circuit with respect to the arrival matters not one whit;
rather, it serves merely as a distraction, making the case dicult to judge, and o-
hand intuitions about it quite suspect.
This diagnosis neatly extends to the Hard cases introduced in section 2. Consider
Kvarts (Ill leave the others as an exercise). Does the injury cause the nger to be
healthy? Intuitively no, but on inspection yes: For the accident that befalls the man
redirects a certain process (namely, the process consisting, roughly, of him and his
movements) onto a causal pathway dierent from the one it would have followed,
and in fact both pathways (actual and counterfactual) issue in the same result (a
healthy nger). So the accident is a switch with respect to the healthy nger. Observe,
furthermore, how easily we can apply the various considerations canvased in the last
section to reinforce this conclusion. Ill consider just the variation test: Alter events in
the environment suciently to make it the case that if the accident had not occurred,
the ngers state of health would have been worse (say, because the man would have
suered some other, much more serious accident). Then we would not hesitate to
count the accident as one of the causes of the ngers later state of health.6 Again,
194 N. Hall
what this test brings out is the (somewhat hidden) switching structure of the example;
that it has the short-circuiting structure as well is merely a distraction.
Here, then, is the point we have come to: (i) When c is a switch with respect to e
(of the interactive kind), then c is a cause of e, although for various reasons this fact
may not be immediately obvious. (ii) What distinguishes the Hard cases from the
Easy cases is exactly that, whereas both exhibit events c and e such that c is a short-
circuit with respect to e, in Hard cases it is also true that c is a switch with respect
to e. (iii) When we render the intuitive verdict about Hard cases that c is not a cause
of e, we wrongly focus solely on the fact that c is a short-circuit with respect to e.
Observing (correctly enough) that c is not thereby a cause of e, we wrongly conclude
that it is not a cause of e, simpliciter. But c is also a switch with respect to e, so that
conclusion is mistaken. The mistake is natural enough, though, given that the short-
circuiting role is by far the more vivid one. (iv) Since, however, no such diagnosis is
available for the Easy cases, the intuitive verdict that, in such cases, c is not a cause
of e stands. But in such cases we can always nd an intermediate event d such that c
is clearly a cause of d, and e counterfactually depends on d, but does so in the odd
double-preventing kind of way. Hence such cases exhibit an intractable conict be-
tween Dependence and Transitivity. (v) Since, in such cases, there are independent
grounds for doubting that d is a cause of e (most notably: no causal process connects
them), the appropriate response is to deny Dependence.
Points (iv) and (v) deserve more discussion; Ill reserve discussion of (v) for the
next section. Now, given that I have been at such pains to nd grounds for resisting
the intuitive verdict about the Hard cases, one might reasonably wonder whether,
with enough ingenuity, we might nd similar grounds for resisting the intuitive ver-
dict about the Easy cases. I say no: The intuitive verdict about those cases is correct,
and that is why they have something useful to teach us about causation (namely, that
counterfactual dependence does not suce for it). While I cannot hope to establish
this claim conclusively, I can at least show that not one of the strategies that helped
us with the Hard cases helps in the slightest with the Easy cases. Lets review them,
using Drunken Bad Guys as our canonical example of an Easy case:
The Explanation Test Does the fact that the Bad Guy Leader sent out the demoli-
tion team help explain the trains arrival? Of course not. The closest this fact can get
to appearing in an explanation of the arrival is to appear in a very specic sort of
explanation-request: Namely, we can ask, Why did the train arrive at its destina-
tion, given that the Bad Guy Leader sent out a demolition team to stop it from doing
so? The obvious and correct reply is to cite the relevant double-preventer: The de-
molition team got side-tracked by the pub; this event prevented it from doing some-
thing (blow up the track) which in turn would have prevented the trains arrival.
Keeping in mind that the given that clause might well be tacitly assumed in the
context of the explanation-request, it seems plausible that double-preventers of the
sort exhibited by this example can be explanatory only relative to this specic sort of
explanation-request. But in the context of such a request, it would be silly to add the
content of the given that clause as an extra bit of explanatory information. No: On
the assumption that the given that clause cites a genuine threat (which we can take
to be a presupposition of the why-question), the question is answered completely by
citing those events that counteracted the threat, and explaining how they did so.
The Subtraction Test Intuitively, the Bad Guy Leader doesnt help cause the trains
arrival by sending out the demolition team. Can we reverse this intuitive verdict by
subtracting events from the environment of the processes that issue in the arrival?
At the very least, can a selective subtraction make it the case that the arrival
counterfactually depends on the Leaders action? Well, lets review the events that
are available for subtraction. On the one hand, there are the events that make up
the trains journey down the tracks, and the rather less noticeable events that consist
in the persistence of the appropriate bits of track. No doubt there are other even less
noticeable events that contribute to the arrival: the presence of sucient oxygen to
allow the trains engines to burn fuel, and so on. We neednt canvass them all, for its
clear that the most that can happen if we subtract some of these events is that the
arrival wont happen. And that is, of course, not a situation in which (i) the arrival
happens; and (ii) its occurrence counterfactually depends on the Leaders action.
Where else to look? Well, there are the events that make up the rest of the story:
the Leaders action itself, the teams journey, the drunken revelry in the pub, and so
on. Subtract some of these, and the most that will happen is that we get a situation in
which the team succeeds in stopping the train (e.g., if we remove the pub itself ). No
help there, either. So the example fails the subtraction test.
The Variation Test It likewise fails this test. Remember that the object here is to
selectively alter events so that the manner of the arrival at least depends on the
Leaders action. Keeping in mind the dierent sorts of events that are available, it
becomes clear at once that the only eect such alterations could possibly have is to
make the Leaders action prevent the arrival (e.g., alter the potency of the beer in the
pub suciently so that the team doesnt become drunk, and therefore continues on
its mission). So the example fails the variation test, as well.
It was clear all along that it was bound to fail both the subtraction test and the
variation test. For what those tests bring out, when they succeed, is that (i) there is an
alternative causal pathway to the given eect e; (ii) the eect of the alleged cause c is
to switch the processes leading to e away from this alternative and onto a dierent
196 N. Hall
path; and therefore (iii) altering the character of the path not choseneither in an
extreme way (the subtraction test), or in a modest way (the variation test)will yield
a situation in which e, or the manner of es occurrence, does depend on c. In this new
situation, if c had not happened, then the processes aimed at e would have followed
the alternative (and now altered) causal pathway, and therefore either would have
missed their mark (the subtraction test), or would have produced e in a slightly dif-
ferent manner (the variation test). But in the Drunken Bad Guys caseas, indeed, in
all Easy casesthe alleged cause simply does not act as a switch. So of course the
subtraction and variation tests fail, for such cases.
Pauls Observation Could it be that Transitivity simply doesnt apply, because for
any candidate intermediate event d, we will nd that while in a loose sense, c (the
Leaders action) causes d and d causes e (the trains arrival), strictly speaking what
happens is that c causes d to have a certain aspect, but it is only in virtue of some
other aspect that d causes e? Lets nd out, by taking our intermediate d to be the
teams decision to stop in the pub. Now, Dependence (and common sense) yields the
verdict that the Leaders action is a cause of d; likewise, Dependence (but not, this
time, common sense) yields the verdict that d is a cause of the arrival. But as far as
I can tell, aspect intuitions are silent: It is not that we can x on some way that
d happens, and say with condence that what the Leaders action really does is
to cause d to happen in this way; likewise for the relationship d bears to e. So this
strategy is not available: The defender of Dependence cannot say (not on these
grounds, anyway) that Transitivity does not apply, and so the Easy cases exhibit no
conict between that thesis and Dependence.
Salience Return to the case of The Engineer. In that case, it can be agreed on all
sides that the engineers action causes an eventthe setting of the switch as the train
passes over itthat interacts with the processes that lead to the arrival; the question
was whether this interaction is of the right sort to qualify the engineers action as one
of the arrivals causes. And we were able to answer that question in the armative
by observing that the setting of the switch bears the same sort of relationship to the
arrival as other events that clearly are causes of the arrival, albeit highly nonsalient
ones (e.g., the presence of the bits of track that the train passes over). But if we try
to apply this strategy to the case at hand, arguing that the Bad Guy Leaders action
is simply a highly nonsalient cause of the arrival, we cannot even get it o the
groundfor the Leaders action causes no event that interacts at all with the pro-
cesses that lead up to the arrival, let alone interacts with them in the right sort of way.
To put the point rather mildly, we appear to have run out of options for defend-
ing the claim that the Leaders action is too a cause of the arrival, and intuition be
damned. I conclude that we should side with intuition in this, as in all other Easy
cases. The conict between Dependence and Transitivity is unavoidable.
6 Metaphysical and Methodological Lessons
Which to give up? Pose the question abstractly, and it may be hard to decide. But the
details of the examples that exhibit the conict in fact make the question easy: For
what is so striking about those examples is that the intermediate event d causes the
eect e only in the sense that it prevents something which would have prevented e.
Thus, Billys trip stops him from preventing Suzys throw; the demolition teams de-
cision to stop in the pub winds up preventing them from blowing up the track, which
would in turn have prevented the trains arrival; and so forth. No causal process
connects d with e; d does not interact with other events to help bring about e; and so
on. Of all of the characteristics we expect from the causal relationship, the only one
exhibited by the relationship between d and e is that e counterfactually depends on d.
And even that characteristic is entirely optional, as witness standard cases of pre-
emption, in which one event causes a second, even though an alternative, preempted
process would have brought about the second, had the rst not done so. The striking
conict with Transitivity thus conrms a suspicion that should have been there from
outset: namely, that double-prevention is not causation. If we had independent reason
to think that Transitivity fails, then we might hold on to Dependence, even in the face
of the examples. But once the Hard cases have been defused, we have no such inde-
pendent reason.
We do have independent reasons to doubt Dependence, on the other hand
reasons that go beyond (or, perhaps, simply properly articulate) the doubts about
that thesis that naturally arise when we focus on cases of double-prevention. For
example, Dependence can be shown to conict both with a prohibition on action-at-
a-distance, and with the claim that the causal structure of a process is determined by
its intrinsic, noncausal character, together with the laws that govern it. But since that
story is longand I have told it elsewhere (in Two Concepts of Causation, this
volume)I wont go into detail here. Given our focus on Transitivity, however, it is
appropriate to point out two new sorts of trouble that arise if we extend Dependence
in a natural way to cover cases of prevention, and of causation by omission. More
exactly, suppose we endorse the following theses:
(1) Event c causes the omission of an event of type E if it is the case that c occurs,
no event of type E occurs, and if c had not occurred, an event of type E would have
occurred.
198 N. Hall
(2) The omission of an event of type C causes event e if it is the case that no event
of type C occurs, e occurs, and if an event of type C had occurred, then e would
not have occurred.
Thus, (1) is a natural application of Dependence to causation of omission, and (2) is a
natural application of Dependence to causation by omission.
Then, if Transitivity holds, we get wonderful results, as in the following three
cases:
Preemptive Strike One day, Suzy suddenly and viciously attacks Billyso viciously
that he is quite incapacitated, as a result. Asked to explain herself, she states that she
was merely preventing him from attacking her. But, her interrogators exclaim,
you know full well that he had no intention of attacking youthe two of you were
friends! Yes, she replies, he had no intention of attacking me then. But he cer-
tainly does nowso its a good thing I incapacitated him.
Unfortunately, the combination of Dependence (extended as in (1)) and Tran-
sitivity endorses this reasoning: For her attack causes his subsequent incapacitation;
and, were he not incapacitated, he would attack her, whereby his incapacitation
causes his failure to attack her. By Transitivity, her attack did indeed cause him to
fail to attack her.
No Cure Billy is sick, and needs a certain drug to cure his disease. But cure requires
two doses, whereas only one is available. (Suppose further that one, by itself, has no
eect on the disease.) In a gesture of futility, Dr. Jones gives Billy the one dose on
Monday. On Tuesday, Dr. Smith comes in, intending to do the same thing, but of
course failing to. Billy remains sick. Whats worse, Dr. Jones gets blamed for Billys
continued ill health, on the following grounds: By giving Billy the dose on Monday,
he caused Dr. Smiths failure to give Billy the dose on Tuesday (for if Jones hadnt
given the dose, Smith would have). But Smiths failure to give the dose on Tuesday in
turn caused Billy to remain ill, since if Smith had given Billy the dose, thenwith
one dose already in his systemBilly would have been cured. By Transitivity, Jones
made Billy remain sick.
Collision Avoidance The engineer ips the switch, sending the train down the right-
hand track; recall that the tracks reconverge up ahead. She promptly puts in for
merit pay, on the grounds that she has prevented a collision. After all, she causes the
absence of a train on the left-hand track (for if she hadnt ipped the switch, the train
would have been there), which in turn causes there to be no collision (for if there had
been a train on the left-hand track, there would have been a collision at the point of
reconvergence). So her action caused the omission of a collision.
Now, it is certainly true that in all of these cases, there is a great temptation to
deny the relevant counterfactuals. For example, we all surely want to insist that
if there had been a train on the left-hand track, that would have been because the
engineer did not ip the switch, and so there (still) would have been no collision. But
however sensible such backtracking reasoning is, it is simply not available to the
loyal defender of Dependence, since that thesis requires a scrupulously nonback-
tracking reading of the counterfactual. So the loyal defender must, on pain of deny-
ing Transitivity, embrace these remarkable results. All the more reason to shift
allegiance elsewhere.
Moreover, the cost of denying Dependence should not be overestimated. We cer-
tainly need not abandon the truism that where there is counterfactual dependence
between distinct events, there is typically causation. Nor, I think, need we deny that
such dependence is a kind of causal relationso long as we are clear that it is not the
kind of causal relation we have in mind when we think of processes, interaction, and
Transitivity. For obvious reasons, we might call this latter kind of causal relation
production; our thesis should then be that production and counterfactual dependence
are both causal relations, but that only production obeys Transitivity.7 To this claim
we should add that counterfactual dependence is not the central kind of causal rela-
tion. As evidence, observe that there is a clear asymmetry in our intuitive reactions to
cases of counterfactual dependence without production, on the one hand, and cases
of production without counterfactual dependence, on the other. As an example of the
latter, suppose Billy and Suzy both throw perfectly aimed rocks at a window. Suzys
gets there rst, breaking it. Intuition unhesitatingly picks out her throw, and not
Billys, as a cause of the breaking. But when we have a case of dependence without
productionas in Drunken Bad Guys, for exampleintuition is more equivocal; it
makes sense, for example, to ask whether, by entering the pub, the demolition team
really helps cause the trains arrival. Why such a question makes sense is perfectly
clear: By stressing the word really (or functional equivalents such as strictly
speaking, etc.), one invokes a context where the central kind of causation is salient;
and entering the pub is not a cause of the arrival, in this sense.
Let us now turn from the metaphysical to the methodological: What lessons does
our discussion have to teach us about how we ought to conduct a philosophical in-
vestigation of causation? Very simply put, the lesson is this: Treat intuitions about
cases with care. It will not do simply to construct hypothetical cases, consult our o-
hand intuitions about them, and without further ado use the results to draw sweeping
(or less-than-sweeping, for that matter) conclusions about the nature of causation.
For such an approach oers no insight into the source of our intuitions, and under-
standing that is crucial if we are to know what, if anything, these intuitions have
200 N. Hall
to teach us. To gain such insight, there is no substitute for close inspection of the
casesclose enough that we can be condent that we have adequately discerned
their salient structure.
Ignoring this point can lead to either of two opposing errors, both of which are to
be found in the literature. The rst consists in excessive gullibility, a willingness to
take intuitions about cases at face value, without interrogating their credentials. Thus
McDermott, when discussing his dog-bite case as well as others, fails to conduct any-
thing like a detailed analysis of his cases, and instead rests content with reporting
the results of various informal polls (of naive subjects, no less) about themas if
these results could establish anything more than the unsurprising result that most of
us nd it counterintuitive to count the dog-bite as a cause of the explosion. Of course
thats counterintuitive, and of course such counterintuitive results count as prima
facie reason to doubt Transitivitybut prima facie reasons can wither upon scrutiny,
and these ones do. McDermott, unfortunately, seems not to see the need for such
scrutiny.
Lewis, in his discussion of this issue, makes a rather more interesting mistake.
Notice rst the way he characterizes the alleged counterexamples to Transitivity
(Causation as Inuence, chapter 3 in this volume, p. 96):
The alleged counterexamples have a common structure, as follows. Imagine a conict between
Black and Red. (It may be a conict between human adversaries, or between nations, or be-
tween gods striving for one or another outcome, or just between those forces of nature that
conduce to one outcome versus those that conduce to another.) Black makes a move which, if
not countered, would have advanced his cause. Red responds with an eective countermove,
which gives Red the victory. Blacks move causes Reds countermove, Reds countermove
causes Reds victory. But does Blacks move cause Reds victory? Sometimes it seems not.
Now, one can argue about whether Lewis has adequately captured the common
structure of the counterexamples. (What is Reds countermove in Drunken Bad
Guys? Putting a pub in the path of the team?) Never mind: What is important to
notice is that Lewis completely overlooks the crucial distinction between counter-
examples that do and counterexamples that dont involve switches. And this leads
him to a curious and quite fallacious dissolution of them. After running through
various of the examplesincluding ones that involve no switcheshe writes (p. 98):
In all these cases, there are two causal paths the world could follow, both leading to victory for
Red. Blacks thwarted attempt to prevent Reds victory is the switch that steers the world onto
one path rather than the other. That is to say, it is because of Blacks move that Reds victory
is caused one way rather than the other. That means, I submit, that in each of these cases,
Blacks move does indeed cause Reds victory. Transitivity succeeds.
Patently, this argument applies only to those casesmy Hard casesthat in-
volve a switch. By contrast, its a confusion to think that in Drunken Bad Guys, for
example, the Leaders decision to send out the team causes the trains arrival to come
about one way rather than another: no, the processes leading up to the arrival are
exactly as they would have been, had the Leader not made the decision.8 Moral: If
youre going to start the job of analyzing cases, then nish it.
Have I nished it? Yes, as far as I can tell. That is, on careful reection I cant
discern any further structure to the various examples that matters, or that distin-
guishes some of them from others. But of course such a conclusion is by its nature
tentative. Supposing it correct, we can now measure the price of Transitivity, at least
in the current market. Put the point this way: We knew all along that counterfactual
dependence is not transitive; that is what we learn in our rst course on the semantics
of this conditional. And that should have at least raised the suspicion that there was
something overly optimistic in the view that counterfactual dependence bears a deep
connection to causation, in the way that the Dependence thesis brings out. The price
of Transitivity is that we must nally face up to this suspicion, recognize that it was
well foundedand therefore look elsewhere for the tools with which to analyze our
central concept of causation.
Notes
1. For valuable discussion of this point, see Lewis (1986d).
2. See for example Lewis (1973a).
3. Autobiographical note: It was precisely because of this problem for my thesis that Transitivity conicts
with Dependence that I became interested in these counterexamples in the rst place. For it had seemed
clear enough to me that there were casesmy Easy casesthat exhibited this conict; but then I found
in the literature other cases, posed not as counterexamples to Dependence but as counterexamples to
Transitivity, and I did not see how my thesis bore on them. Had I simply misdiagnosed my own cases,
construing them as trouble for Dependence when in fact that principle was quite irrelevant, and they
werelike their cousins in the literaturecounterexamples to Transitivity, pure and simple? No, I hadnt.
But it took me a while to see why not.
4. This is the response I favor. For reasons, together with much more detailed discussion of other double-
prevention cases, see my Two Concepts of Causation (chap. 9 in this volume). For discussion of related
issues, see my (2002a) and (2002b).
5. For detailed discussion and defense of the intrinsicness thesis being relied on here, see my (2002b).
6. A slightly dierent consideration pointed out to me by Tim Maudlin also applies: Alter the surgery
enough so that the nger ends up in a much better state of health than it would have, without the accident.
Here, too, we would not hesitate to count the accident as a cause of the ngers state of health. The rele-
vant methodological point still applies: Prima facie, the correct judgments about the causal status of the
accident should not be so sensitive to these ne details.
7. Certain other principles of interest distinguish the two varieties of causation as well; see my Two Con-
cepts of Causation, chapter 9 in this volume.
202 N. Hall
8. Perhaps Lewis thinks that the dierence consists in this: Given the Leaders decision, the trains arrival is
caused in part by the demolition teams stopping in the pub; without that decision, the arrival would not
have this event as one of its causes. But the intuition that Lewis is leaning on, as applied to this case, is that
the Leaders decision is the switch that steers the world onto one path rather than the other. That intu-
ition is obviously out of place, in this example: We do not have here a choice between two dierent
paths. Then again, perhaps Lewis means to be appealing to a much stronger principle to the eect that if
c; d, and e are such that (i) d is a cause of e; and (ii) if c had not occurred, d would not have been a cause of
e; then it follows that (iii) c is also a cause of e. But that principle is not available to him. To see why, ring
a small change on the Drunken Bad Guys case. This time, the Leader sends the team out without telling
them their mission; they are simply ordered to go to a certain point, and wait for written instructions to
arrive. In fact, they never make it to the rendezvous, thanks to the pub. But if they hadnt stopped in the
pub, they would have reached the rendezvous, shortly thereafter received instructions to blow up the track,
etc. Then let c be the sending of the instructions to the rendezvous, d be the teams stopping in the pub,
and e be the trains arrival. Lewis wants to endorse Dependence, so he will say that d is a cause of e
(without, presumably, drawing a distinction between this kind of causation and the ordinary kind). But
everyone must agree that if c had not occurredif no instructions had been sentthen d would not have
been a cause of e; for in such a circumstance, the team simply poses no threat whatsoever to the trains
arrival. So by the principle under consideration, c is a cause of e: The sending of the instructions helps
cause the train to arrive. Which, of course, is silly.
8Aspect Causation
L. A. Paul
While skiing, Suzy falls and breaks her right wrist. The next day, she writes a phi-
losophy paper. Her right wrist is broken, so she writes her paper using her left hand.
(Assume, as seems plausible, that she isnt dexterous enough to write it any other
way, e.g., with her right foot.) She writes the paper, sends it o to a journal, and it is
subsequently published. Is Suzys accident a cause of the publication of the paper?1
Of course not. Below, I will show that none of the major contenders for a theory
of events coupled with a theory of causation succeeds against examples like that of
Suzys accident, and that the reason for this derives from an underlying tension
between our beliefs about events and our goals for theories of causation. I will then
argue that property instances should be taken, in the rst instance, as the causal
relata, and propose an analysis of causation that I call aspect causation.
Aspect causation combines elements of regularity theories and David Lewiss new
inuence theory of causation2 with property instances. Combining lawful entailment
with inuence handles problems involving redundant causation for regularity and
counterfactual based theories. Changing the causal relata to property instances re-
solves transitivity problems and allows us to develop an account of events without
holding our theory of causation hostage to our theory of event individuation.
1 Event Causation: Problems
First, we need some background and terminology: actual events can be dened ac-
cording to a continuum from ne grained to coarse grained. Speaking loosely, the less
ne grained a theory of individuation species events to be, the fewer events the
theory implies. There is a parallel continuum from fragile to robust for the individu-
ation of events with respect to modal contexts, needed for counterfactual theories
of causation. The less fragile the event, the weaker the requirements events in other
possible worlds must meet to be numerically the same as the event that actually
occurred.
One of two approaches is usually used to analyze causation: a regularity approach
or a counterfactual approach. Roughly, versions of regularity or covering-law anal-
yses for events state that an event c causes an event e i the occurrence of c, together
with the right conditions and right regularities or laws, is sucient for the occurrence
of e. It is a consequence that causation is transitive: If c is a cause of d, and d is a
cause of e, then c is a cause of e. Counterfactual analyses in their simplest form hold
that for any two actual, distinct events c and e, e depends causally on c i e depends
counterfactually on c, that is, had c not occurred then e would not have occurred.
Causation is usually distinguished from causal dependence in order to ensure tran-
sitivity: c is a cause of e i there is a chain of causal dependencies running from c to
e. Both accounts require transitivity to avoid important counterexamples,3 and both
require an acceptable theory of eventhood in order to provide an acceptable (re-
ductive) analysis of causation. In order to provide an underpinning for analyses of
causation based on events as causal relata, several characterizations of the identity
conditions for events have been put forward.
Fragile Events Under a fragility view, one must hold that when I pick up my cup
of coee, had I picked up the coee a millisecond later, had the coee swirled in a
slightly dierent fashion, or had it been a fraction of a degree hotter, the event of
my picking up the coee cup wouldnecessarilyhave been a dierent event. Since
anything that brings about a new event counts as a cause of it, the view gives us
a plethora of spurious causes. If an explosion on the sun makes the summer day on
which I drink my coee a fraction of a degree hotter, thus keeping my coee ever so
slightly warmer than it would have been had the explosion not occurred, that explo-
sion counts as a cause of my drinking my coee.
If events are fragile, then Suzys skiing accident is among the causes of her writing
her paper with her left hand, and her writing of the paper is a cause of it being pub-
lished. For fragility theorists, writing the paper with her left hand is necessarily a
dierent event than writing the paper in some other way. So the skiing accident is
a cause of the event that actually occurred: the event e of the papers-being-written-
with-the-left-hand, since the skiing accident is part of a sucient condition for e, and
if the skiing accident had not occurred, e would not have occurred. Given that the
skiing accident occurred, if the writing of the paper occurs at all, it occurs with the
left hand, since after the accident she cant write it any other way. Under a counter-
factual analysis that ensures transitivity and treats events as fragile, if the event of the
writing of the paper with the left hand had not occurred, then the publication would
not have occurred. Such a fragility theorist is committed to the view that Suzys
skiing accident is a cause of her papers being published.
The implausibility of such conclusions increases with every link in the causal chain,
and confronts advocates of fragility with a multitude of counterintuitive cases. Per-
haps the fragile event theorist can provide some sort of an account that can make
one-step cases such as the suns explosion causing my drinking of my coee seem less
counterintuitive. But the implausibility of holding, for example, that the skiing acci-
206 L. A. Paul
dent is among the causes of Suzy publishing her paper, and the ease with which such
two-step (or multi-step) examples can be constructed shows us that explaining one-
step cases is not good enough.
Fine-grained Events Using ne-grained events as part of a theory of causation gives
us many of the same undesirable results as the use of fragile events, coupled with a
need to accept an implausible ontology. Jaegwon Kims view is perhaps the most
well known: he combines a regularity approach with a very ne-grained theory of
events. For Kim, events are identied by their constitutive triples, which comprise an
individual, a property exemplied by that individual, and the time when the property
is exemplied. Although an event may exemplify many properties, it is the constitu-
tive triple that, in a sense, denes the event.4 For Kim, we have two numerically dif-
ferent events when they dier in their constitutive triples: That is, when they dier in
their constitutive individuals, properties, or times.
When cat C. Louise sneezes loudly, according to Kim at least two events occur
in that region of spacetime: C. Louises sneezing and C. Louises sneezing loudly.
Moreover, when she sneezes, she sneezes in the kitchen, on a summer morning while
dislodging a ea; so we also have the simultaneously occurring events of a sneezing
on a summer morning, a sneezing in the kitchen, and a sneezing while dislodging a
ea. We can continue our modication of the constitutive property indenitely, so
we have an innite number of events occurring in the region of C. Louises sneezing.
For Kim, these events are all dierent but not necessarily distinct: In some sense
(although he does not develop the notion), he thinks the many events that occur
when C. Louise sneezes are all included in her sneezing.5 The claim is supposed to
be similar to the claim that C. Louise is made up of many dierent C. Louises, one
for each hair or molecule that C. Louise might lose but still remain C. Louise. C.
Louise is made up of many spatiotemporal parts, and many of these parts are less
maximal than all of C. Louise, but these proper parts are enough to count as C.
Louise themselves.
But the attempt to run an analogy between included events and spatiotemporal
parts fails.6 The dierence between events that are qualitatively richer than other
events (such as the dierence between a sneezing in the kitchen, and a sneezing at
10 a.m.) is not purely analogous to the dierence between spatiotemporal parts. If
Kim were claiming that the event of C. Louises sneezing in the kitchen and sneezing
at 10 a.m. occupied slightly dierent spatiotemporal regions, then we could see how
one event could overlap another, and how part of one event could include part of
another event. But both of these events are said to occupy the very same spatio-
temporal region. The only way to make sense of the parts analogy is to take some
Aspect Causation 207
events as logical parts of other events, giving logical overlap, not spatiotemporal
overlap. But then the idea that these are dierent events (which are supposed to be
particulars, or things that occupy regions of spacetime) rather than just dierent
properties is still just as counterintuitive.
For Kim, when Suzy writes her paper with her left hand, multiple events occur:
Suzys writing of her paper, Suzys writing of her paper with her left hand, and so on.
All of the occurrent events that include Suzys writing of her paper with her left
hand (such as Suzys writing her paper with her left hand by writing on a keyboard,
Suzys writing her paper with her left hand by pecking with one nger on a keyboard,
and so on, if this was how Suzy actually wrote her paper) are eects of the accident.
And all of these eects are among the causes of the publication of the paper, for they
all include Suzys writing of the paper, and so they are sucient under the laws to
cause the publication of the paper.
David Lewis argues for a counterfactual approach combined with ne-grained
events, and denies that events must be fragile.7 For Lewis, actual events have both
strong and weak essences, so that when an event that is C. Louises sneezing loudly
occurs, a second event (with a weaker essence) also occursC. Louises sneezing.
Events, as under Kims view, are greatly multiplied.
For Lewis, the skiing accident is among the causes of the event of Suzys writ-
ing her paper that is essentially a writing of her paper and is essentially with her left
hand, as well as a cause of all the other events that occurred essentially with the left
hand (i.e., Suzys writing of the paper that is essentially a writing of the paper and
essentially with the left hand, but only accidentally a writing using a keyboard, etc.).
If the accident had not occurred, none of these events, as essentially specied, would
have occurred. And (holding that the appropriate closest possible world in which we
evaluate whether or not the publication of the paper occurs is a world in which no
event that is essentially Suzys writing of her paper occurs) each of the events that
involve Suzys writing of her paper that is essentially a writing of her paper, which
includes many events caused by the accident, is a cause of the publication of the
paper. So Lewis must also accept the conclusion that the skiing accident is a cause of
the publication.
The problems for the Kimian and Lewisian views highlight a side eect of their
multiplication of entities: many unexpected instances of what we might call a kind of
redundant causation. For Kim, each event that includes Suzys writing of her paper
causes the event that has as its constitutive property the publication of the paper.
(Of course, there are many other publications of the paper caused at the same time as
the one we are focusing on, but well ignore those events for simplicitys sake.) For
Lewis, each event that involves Suzys writing of her paper essentially causes a ver-
208 L. A. Paul
sion of a publication of her paper that is essentially a publication. Multiplying events
in either the Kimian or the Lewisian way gives counterintuitive results in examples
like that of the accident, results in a new and implausible kind of redundant causa-
tion, and, as I shall argue below, is not necessary. The point is not that the size of
the ontology of events is unacceptable, but rather that ne-grained theories of events
cannot be motivated by the claim that they improve the results given by their com-
panion theories of causation.8
Coarse-grained Events Donald Davidson holds that the correct theory of causa-
tion involves a regularity approach, and that events are individuated by the regions
of spacetime they occupy.9 This view has the happy consequence that C. Louises
sneezing is the same event as the event of her sneezing loudly, and the event of her
sneezing in the kitchen, and so on. Such a view takes events to be coarse grained.
Defenders of coarse-grained events seem to have a natural solution to cases like
the skiing accident causing the publication. They can hold that Suzys writing the
paper with her left hand and Suzys writing the paper are one and the same event,
and use this claim to break the link between the skiing accident and the writing of the
paper. A natural way to justify the claim would be to point out that that the actual
writing of the paper and the actual writing of the paper with the left hand occupy the
same spatiotemporal region.
Those who hold a version of the regularity analysis could argue that the skiing
accident is not a part of a sucient condition for the writing of the paper, so it is
not a cause of it. Counterfactual theorists could adopt a moderately robust view of
events together with the coarse-grained approach and argue that the writing of the
paper would have been the very same event even if it had been written with the right
hand, so if the skiing accident had not occurred, the writing of the paper would still
have occurred. With either approach, since Suzys skiing accident was not a cause of
the existence of the paper, it was not a cause of her publishing the paper.
Something about this seems rightit seems to capture part of what we mean when
we deny that the accident was a cause of the publication of the paper. But it cant
be quite right, because the response can be used to generate problems. For suppose
there was more to tell about the case of the skiing accident. Suzys left hand, unused
to writing, begins to cramp severely after nishing the paper. She visits a doctor, and
spends large amounts of money on prescription drugs in order to dull the pain. It
seems right to say that the writing of the paper with Suzys left hand is among the
causes of the pain in her left hand, and her visit to the doctor and expenditure of
large amounts of money on prescription drugs is caused by the pain. Moreover, it
seems right to say that the skiing accident to Suzys right hand is among the causes of
her writing the paper with her left hand and so a cause of the pain in her left hand.
But if we are coarse-grained theorists, we have already decided in order to resolve
earlier problems that the event of Suzys writing the paper is the same event as her
writing the paper with her left hand, and that the skiing accident did not cause the
event of the writing of the paper. But if the skiing accident did not cause the event of
the writing of the paper when we determine what caused the publication, then neither
did it cause the event of the writing of the paper when we determine what caused the
costly prescription. The link in the causal chain between the skiing accident and the
writing of the paper has already beenby stipulationbroken. But we do want to
say that the skiing accident was a cause of the cramping hand and the costly pre-
scription, and moreover, we want to say this because we think that the accident, by
causing the writing of the paper to occur in a particular wayit was a writing with
the left handwas a cause of the subsequent cramping of the left hand (and thus of
the costly prescription).
So we shouldnt try to preserve causal intuitions by breaking the link between the
skiing accident and the writing of the paper. Further, common sense would have it
that the writing of the paper (however it was written) is a cause of the publication of
the paper, so its not a good idea for the coarse-grained theorist to try to deny the
second link in the chain. I suspect that the problem is perfectly general: For any
theory of events that individuates coarsely (i.e., for theories that do not individuate
events as nely as can be with respect to properties), we can design puzzles like the
one with the cramping hand.10
This leaves advocates of event causation in a predicament, since whatever standard
of individuation for events they impose, there exist clear problem cases. Distinguish-
ing between events coarsely respects many of our causal judgments and common-
sense views about individuation. But by holding that some changes (no matter how
minor) in events do not result in a dierent event, we lose needed exibility when just
such a minor change aects the causal story that we want to tell. On the other hand,
theories of events that are as ne grained as can be build every detail of the event
into the causal story that we tell. Such precise specication gives us many cases of
spurious causationunless we reject the transitivity of event causationwhich is a
steep price to pay for those of us who value our commonsense views about causation.
2 Property Instances
But there is a clear, intuitive solution to the case of the accident. To see it, set aside
the idea that events are the only kinds of causes and eects, and think in terms of
property instances with respect to events or individuals. With this in mind, consider
our example. It seems right to say that because Suzys accident involved Suzys right
210 L. A. Paul
hand being broken, it caused the writing of the paper to be a writing with the left
hand. But the left-handedness of the writing had nothing to do with the papers
being accepted for publication: As Ive told the story, all that caused the paper to be
accepted for publication was that the paper was written in the rst place, that (pre-
sumably) it was a good paper, and so on. So the accidents property of being an
accident to Suzys right hand did not cause the paper to have the property of being
accepted for publication, but merely to have the property of being written with the
left hand. This becomes clear when we think of causes and eects as instances of
properties rather than as events. Its not that the case shows us causation isnt tran-
sitive, but rather that this is a case where the question of transitivity does not arise,
since the property caused by the accident and the property that causes the publica-
tion are not the same property. Transitivity just doesnt apply.
The follow-up example, where the skiing accident is a cause of the cramping in my
left hand, does involve transitivity: The skiing accidents property of being an acci-
dent to Suzys right hand caused the writing of the paper to be a writing with the left
hand, and the writing of the paper being a writing with the left hand caused the left
hand to cramp. This solution conforms to our intuitive understanding of what hap-
pened, and as a result generates the correct answer. Examples like these give us a
strong reason to accept the view that property instances rather than events simpliciter
are the causal relata. We get further exibility by allowing the property instances to
involve individuals as well as events.
There is a related reason to take property instances as the causal relata. As the
example of Suzys accident helps to bring out, there is an essential tension between
the goal of developing an adequate, acceptable theory of event individuation and the
goal of developing an adequate theory of causation. It is not uncommon to adopt
an extremely ne grained or fragile theory of events in order to get better results for
ones theory of event causation (this is what usually motivates advocates of ne-
grained views), but the example of the accident shows us that such maneuvers bring
trouble elsewhere.
By separating the two accounts, we can be free to develop our best theory of event
individuation (and our best theory of individuation for individuals) apart from our
theory of the causal relata and the causal relation. Even if the Suzy examples dont
convince, the history of problems developing an acceptable theory of events as a
companion to a theory of causation, together with the lack of parsimony (both of
causal claims and events) of views like Kims and Lewiss, should be enough to mo-
tivate a switch to property instances.
There is one issue that needs immediate explication: how does my view that the
causal relata are property instances dier from Kims theory of events as property
exemplications? The obvious dierence between the views is that my account shows
us how to keep transitivity while explaining cases like Suzys accident. But the root of
the dierence is less obvious: It is based on the fact that Kims account requires that
causation relate events, dened using constitutive triples, not property exemplica-
tions simpliciter.
To see how this is important, recall that Kims theory of events is metaphysically
substantial: The property instances referred to by the triples really are events. It is
not the view that we should dene the causal relata and then call whatever we end
up with events. Because it is a denition of events, not just any property that can
be exemplied counts as an event: Properties that modify other properties are not
themselves eventspresumably because modiers are even less acceptable, intui-
tively, as real events than other properties are.
For Kim, our example of the property of being performed with the left hand would
not count as an event itself but rather (as it should) as a modier of the property of
being a writing of a paper. It is hard to see a way to make sense of the idea that the
property of being performed with the left hand could count as any sort of event. Kim
handles modiers by combining the modier with an appropriate constitutive prop-
erty to create a new constitutive property, which in turn denes a new event. So for
Kim, the property of being performed with the left hand is not an event in its own
right: The writing of the paper counts as an event, and the writing of the paper with
the left hand is another event. In this way, Kim hopes to be able to make room for
the ways in which we seem to want to take account of the importance of properties,
including modiers, in cases of causation, yet retain the idea that causation is a
relation between events. Unfortunately, by creating new events for properties that
are merely modiers, Kim opens the door to counterexamples like that of Suzys
accident.
Kim could revise his view and claim that instead of creating a new event that is the
writing of the paper with the left hand, we say instead that the event of the writing of
the paper (dened by the triple of Suzy, the property of being a writing of a paper,
and the time at which the writing occurred) has the property of being performed with
the left hand.11 But while this seems sensible, it wont help Kim, since in order to get
the right result in the case of Suzys accident, it is the property of the event, being
performed with the left hand, that has to count as an eect (and as a cause, with re-
spect to the cramping). Since Kim holds that causation is a relation between events,
the modifying property of being performed with the left hand cant be a cause or
eect unless it somehow counts as an event in its own right.
By switching to property instances as causes and eects we sidestep the problems
that Kim, Lewis, Davidson and others with related views face when developing
212 L. A. Paul
theories about causation. Of course, we must still rely on an adequate theory of
property individuation, but any of these theories of events and event causation must
rely on a theory of properties as well. Removing a problematic level of analysis and
solving dicult problems with transitivity should be sucient motivation to make
the switch.
I will use aspect to refer to a property instance: an aspect is a particulars (a
particular event or individual) having a property. Aspects are things that correspond
one to one with thing-property pairs such that the property is had by the thing; so
aspects are in an important sense part of the spatiotemporal world. Dened as such,
aspects correspond to tropes (if there are any tropes), but the denition of property
is intended to be exible: Whether property instances involve exemplications of
universals, sets of particulars, states of aairs or tropes I need not say nor choose
between.12 A few more details: Aspects involving conjunctive properties can be
causes or eects i each conjunct is. The properties instanced must be suitably natu-
ral, so aspects that involve gruesome or disjunctive properties are not eligible to serve
as causes or eects. The question of how extrinsic properties can be and still be
paired with particulars to serve as causal relata needs further investigation.13
Taking aspects as the causal relata does not exclude events as causes and eects,
for it may be that when an aspect causes another aspect or group of aspects, the as-
pect or aspects that are caused are sucient to imply or constitute an event. In some
cases, if an aspect causes enough properties to be instantiated, or perhaps enough
essential properties to be instantiated, we may say that the aspect causes an event in
virtue of causing the particular to have those properties.14 If aspect c causes Billy to
have the property of dying in a particular way at a particular time, this aspect might
be sucient for the instantiation of an event, namely, the death of Billy. If so, then
aspect c is a cause of the event of the death of Billy, in virtue of being a cause of
aspects sucient to constitute the event. Depending on your standards for partic-
ulars, the aspects sucient to constitute a particular may be many or few, common
or rare. The power of the aspects account is that it will deliver the right causal judg-
ment consistent with the standard of particulars adopted.
Counterfactual and regularity accounts based on aspects could rely on appropriate
versions of the following denitions:
1. Counterfactual dependence of aspects: For any two distinct, actual events or in-
dividuals c and e, and logically distinct properties p and q: aspect e
q
(es having q) is
counterfactually dependent on aspect c
p
(cs having p) i, had c occurred (existed)
without p, then e would not have occurred (existed) with q. (I do not specify whether
c without p is c or is numerically dierent from c.)15
2. Lawful entailment of aspects: For any two distinct, actual events or individuals c
and e, and any two logically distinct properties p and q, aspect c
p
(cs having p)
lawfully entails aspect e
q
(es having q) i cs exemplication of p is subsumed by the
antecedent of the right law or laws that entail a consequent subsuming es exempli-
cation of q. This is a souped-up kind of suciency, one that excludes properties had
by particulars not linked to the eect via a law of nature from being counted as part
of a lawfully sucient condition for an eect.16
Under a regularity account where lawful entailment implies causation, being a
breaking of the right hand is a cause of the left-handedness of the writing. But it is
the papers property of being a good paper (rather than the property of being written
with the left hand) that lawfully entails the papers being published, so the accidents
being a breaking of the right hand is not a cause of the publication.
Under a counterfactual analysis where dependence implies causation, the left-
handedness of the writing depends on the breaking of the right hand. However, the
publication of the paper does not depend on the left-handedness of the writing: It
depends on the property of being a good paper. Our transitivity puzzle is dissolved.
3 Reductive Analysis: Problems
Although the simple versions of the counterfactual and regularity accounts can be
used to solve the transitivity puzzles we have been discussing, they cannot serve as
theories of causation. Recent work on causation shows that problems with pre-
emptionwhere potential causes c and b both occur, and each in the absence of the
other can cause the eect e, yet it is intuitively clear that c is a cause of e and b is
notrequire a more sophisticated treatment of the causal relation.
Two kinds of preemption are responsible for the worst of the problems: late pre-
emption and trumping. In cases of late preemption, the central issue involves the fact
that the preempted cause b would have caused the very same eect e, but slightly
later than the preempting cause c caused it. In these cases, the preempted causal
chain is prevented by the occurrence of the eect itself, before the preempted cause
can cause it.
C. Louise crouches, aiming for an unfortunate y. Possum also crouches, aiming
for the same y. C. Louise pounces, and catches the y. She then eats it. Possum,
though agile, is heavier than C. Louise and so pounces more slowly, and the y is
eaten by the time he arrives. If C. Louise hadnt eaten the y, Possum would have
eaten it in the very same way, but just a few moments later. Counterfactual and reg-
ularity theories of event causation seem to have serious problems with examples like
214 L. A. Paul
these: The (event that is the robust) catching of the y does not depend counter-
factually on C. Louises crouching, and conversely, both C. Louises crouching and
Possums crouching seem to be lawfully sucient for the catching.
I have argued elsewhere that eects in cases of late preemption do depend on their
causes, for when the eect occurs depends on whether the preempting cause oc-
curred, but when the eect occurs does not depend on whether the preempted cause
occurs.17 We can make a similar claim in terms of lawful entailment: The preempt-
ing cause, but not the preempted cause, lawfully entails the eects occurring when
it did.
Trumping examples give a new twist to the problem. In these examples, preempted
cause b would have caused the very same eect e, but for the fact that a law species
that if c occurs, c causes e and b does not. In these cases, b could have brought about
the eect at the very same time, and with the very same properties, as c did.18
C. Louise crouches, aiming for another y. Possum also crouches, aiming for the
same y. C. Louise jumps. Possum, who has been practicing, jumps a moment later,
but his (newly acquired) agility makes him able to catch the y at the same time as C.
Louise. Unfortunately for Possum, there is a little-known law that states that ies,
when pounced on by multiple cats, are captured by the cat who jumps rst. Since C.
Louise jumps before Possum, she gets the y. If C. Louise hadnt jumped, Possum
would have captured the y in the very same way and at the very same time. C.
Louises pounce, albeit through no intrinsic feline merit, trumps Possums.
Trumping cases where the eect would have occurred at the very same time and in
the very same way if it had been caused by the preempted cause are of concern only
if we want our analysis to be able to handle action at a distance.19 It seems to me
that trumping is simply a new variant of early preemption, and cases not involv-
ing action at a distance can be solved using stepwise dependence.20 Nevertheless, the
cases are quite interesting and I prefer an analysis that can handle them to one that
cannot.
Unsurprisingly, counterfactual and regularity theories have problems with trump-
ing: The catching of the y does not depend on C. Louises crouching, and con-
versely, C. Louises crouching and Possums crouching each lawfully entail the
catching. Since there is no dierence in when or whether the eect occurs if the pre-
empting cause does not occur, the solution for late preemption cannot be straight-
forwardly applied to trumping cases.
Lewis sees that restricting dependence to when or whether an eect occurs if the
cause occurs is too limiting21 and proposes that we dene causation in terms of a
pattern of dependencies between events. The relevant events are alterations of the
events for which the causal relation is being evaluated. These alterations are either
the actual c and e (actual refers to the world of the example), or very fragile ver-
sions of, or alternatives to, c and e. The alterations are used to help us represent dif-
ferent ways c and e could have occurred (or dierent ways events that are very much
like c and e could have occurred).
Under Lewiss account, an event c inuences an event e i there is a substantial
range c
1
; c
2
; c
3
. . . and e
1
; e
2
; e
3
. . . of not-too-distant alterations of c and e such that
if c
1
had occurred, then e
1
would have occurred, and if c
2
had occurred, e
2
would
have occurred, and so on. In this way we check to see if whether, when, and how e
occurs depends on whether, when, and how c occurs. If there is a suciently large
range of direct dependencies between alterations of e and c, then c inuences e, and
so c causes e. To preserve transitivity, as in the original counterfactual analysis,
Lewis takes the ancestral: c causes e i there is a chain of stepwise inuence from c
to e.
The new analysis handles preemption problems elegantly: The preempting cause, if
changed, would have caused changes in the eect. If C. Louise had batted the y in-
stead of catching it, or if she had caught the y dierently, the eect at the end of the
corresponding causal chain would have been dierent: a batting instead of a catch-
ing, or a catching that occurred with more or less enthusiasm. Not so for Possum:
Whether hed bat or how hed catch would make no dierence to the eect.
But the approach has two serious defects. First, the weakening of the dependence
requirement to include more ways in which the eect can depend on the cause allows
spurious causation. In the preemption cases, it seems right that if Possum had
pounced earlier or (in the late preemption case) with more agility, the eect would
have occurred earlier or in a dierent way. But if the eect depends on Possums acts,
then by the analysis above Possums act counts as a cause. This is clearly an unde-
sirable consequence.
Lewis recognizes this problem and attempts to minimize it by arguing that in most
cases alterations in C. Louises act make more of a dierence than changes in Pos-
sums act, and further that alterations in Possums act, in the context of comparing
this act to C. Louises, are much more distant than alterations of C. Louises act.
Dierences in degree or distance correspond to dierences in inuence and justify
calling C. Louises act, but not Possums, the cause. Lewis argues further that if it
turned out after taking degree and distance into account that there was not much
dierence between the inuence of Possums act and C. Louises act, we would be
justied in calling Possums a cause as well.
The view is defensible with respect to cases of trumping where both potential
causes could bring about the very same eect. But in cases of late preemption we are
not justied in calling Possums preempted act a cause no matter how much inuence
216 L. A. Paul
he has. The fact remains that Possums act does not lawfully entail the catching of
the y when it actually occurred, and intuitively, lawful entailment is a necessary
condition for causation. (Or at least it is for worlds like our own.) The problem with
inuence can be put more generally: Consider two events a and b that we would
normally take to be causally unrelated. Take a to be my body temperature and b to
be the white pages of the manuscript strewn across my desk. If my body temperature
were altered so that I radiated sucient heat, the white paper would turn brown and
curl at the edges. But surely the temperature of my body is not a cause of the white-
ness of the paper: It does not lawfully entail the whiteness. Inuence alone is not su-
cient for causation.
The second problem with the account involves transitivity. To preserve transitivity
(and to help solve some particularly worrying cases of early preemption), Lewis takes
the ancestral of the inuence relation to be causation. However, this move commits
him to counterintuitive results with respect to a ock of (supposed) counterexamples
to transitivity.22
Many of the cases have the same general form: Some series of events (call this
event pathway A), initiated by event a, starts to occur. If all the threatened events of
A occurred, the series would culminate in the causing of event c. However, before the
series of events that make up event pathway A have all occurred, event b causes event
c via a dierent chain of events (call this event pathway B) connected to but dierent
from A. For example, a train rushes toward Jesse James, who is tied to the tracks. If
the train continues on its track, it will run over and kill James (event pathway A). A
few minutes before the train runs over Jesse, his brother Frank ips a switch that
causes the train to veer left onto a dierent track (event pathway B). Unluckily for
Jesse, this track converges to the original track just before the spot where he is tied,
and the train runs him over anyway. We assume that the trains diversion to the left-
hand track did not delay the train or change the event of Jesses death in any way.
(We might assume that the original track meandered a bit before converging with the
left track, so that each track was exactly the same length.)
Now, intuitively, we want to deny that Franks ipping of the switch was a cause
of the train running over and killing James. But under the inuence accountas
under accounts of event causation generallywe cannot. The event of Frank ip-
ping the switch inuences the event of the trains being on the left hand track and the
event of the trains being on the left hand track inuences the convergence at the
point just before it ran over Jesse, since there are alterations of the event of the train
being on the left-hand track just before it converges (namely, the alteration where
the event is completely excised from history) that would result in the convergence
not occurring. Since the event of the trains running on the track through the
convergence point toward Jesse inuences his death, the event of the ipping of the
switch is a cause of Jesses death.
The case is related to our skiing accident case above, and Lewiss new account fails
to handle it for the same reason that his earlier account fails it: A reliance on events
as the causal relata allows too much information into the causal claim, and when this
extra information is combined with transitivity, spurious causal results are easy to
generate.23
4 Aspect Causation
Both lawful entailment and inuence go some way toward capturing the content of
the causal relation. So why not combine the two? For it is because of a lack of lawful
entailment that the inuence account errs in counting certain events as causes, and it
is because a inuences eect e and b does not in cases where a trumps b that a regu-
larity account errs in counting b as a cause. Each analysis alone is too permissive, but
combined they can give us a simple, strong, and elegant analysis of causation. In the
rst part of this paper, I argued that property instances, not events, are causes and
eects: Problems with transitivity help to make this clear. Accordingly, I propose the
following analysis of causation, based on the denitions (1) and (2) given in section 2
and Lewiss denition of inuence given in section 3:
Aspect Causation: For any two aspects c
p
and e
q
:
(i) if c
p
lawfully entails e
q
, and
(ii) if c
p
inuences e
q
,
then e
q
is directly caused by c
p
. Taking the ancestral of direct causation in order to
give us causation, c
p
is a cause of e
q
i e
q
is directly caused by c
p
or there is a chain of
direct causation running from c
p
to e
q
.
The idea is this: Take any aspect c
p
that lawfully entails an eect e
q
; for each such
aspect c
p
, check to see that the eect exhibits dependence on the cause by check-
ing for an appropriate pattern of dependence of aspects on aspects. Let alterations
p
1
; p
2
; . . . ; and q
1
; q
2
; . . . of property instances p and q be property instances that
might be similar to but numerically dierent from p or q, and check to see if c
p1
had
occurred, then e
q1
would have occurred, and so on.24 If and only if the appropriate
pattern of dependence exists, c
p
is a cause of e
q
. By including inuence in my ac-
count, I allow for a certain amount of vagueness, but for far less than in the original
inuence theory. The account tightens up the inuence theory in two major ways: It
prevents anything from counting as a cause if it does not lawfully entail the eect,
218 L. A. Paul
and it prevents illicit information from being included in causal claims by taking
aspects rather than events (in the rst instance) as the causal relata. It seems correct
to say that both inuence and lawful entailment capture part of the nature of causa-
tion, and my hope is that the two combined suce for a simple and strong analysis
of the (deterministic) causal relation in the actual world.25
By requiring causes to lawfully entail their eects, we eliminate the major prob-
lems with spurious causation that Lewis faced. In our example of late preemption, C.
Louises having the property of pouncing in some particular way at some particular
time counts as a cause of the ys having the property of being caught in a particular
way at a particular time. The properties of C. Louises act, being a pouncing in such and
such a way and at such and such a time lawfully entail the ys having the property of
being caught when and how it actually was, and if the properties of C. Louises act,
being a crouching in such and such a way and at such and such a time, were changed,
the y would not have been caught when and how it actually was. We can even say that
since the properties of C. Louises act caused the catching of the y in a particular
way at a particular time, the event of the ys being caught was caused simpliciter.
But properties of Possums act, being a pouncing with such and such a momentum
and starting at such and such a time, do not lawfully entail the eect as it actually
occurred. (Possum could not have pounced the way he did, when he did, and brought
about the eect when it occurred the way it did.) Alterations of properties of his
act that would aect the time of the y-catching and thus give spurious dependence
are not relevant to our evaluation of the situation, since these properties were what
entailed his arriving too late in the rst place. Likewise, my bodys having a temper-
ature of (about) 98.6
F does not lawfully entail the color of the paper on my desk,

and so it cannot be counted as a cause no matter what the overall pattern of depen-
dence looks like.
For those who resist analyses involving counterfactuals for fear of unwanted
vagueness, my account should help to calm those worries as well. These worries can
arise when determining the truth values of counterfactuals, for the selection function
might specify more than one closest possible world, allowing conicting truth values
for the relevant counterfactuals. Unless we adopt a selection function that species
only one closest world, ties in some cases are possible, and restricting causation to
cases where there are no ties is ad hoc.26 If we merely emend a counterfactual-based
account of causation to say that a pattern of counterfactual dependence under some
precisication of the similarity relation is sucient for causation, then our counter-
factual constraint is too permissive, allowing causation whenever there is vagueness.
On the other hand, requiring dependence under all precisications is clearly too re-
strictive, preventing causation whenever there is vagueness.
But if we require lawful entailment for causation, then we can rely on the more
permissive version of counterfactual dependence, yet eliminate most of the vagueness
worries. Although we might have not have a pattern of dependence under all pre-
cisications, intuitively, if we have dependence under one or more precisication of
the similarity relation together with lawful entailment in the actual world we can say
there is causation without being unduly permissive. There are additional benets to
combining lawful entailment with dependence, for lawful entailment brings in natu-
ral restrictions with respect to disjunctive, overly gruesome, and logically necessary
properties.27 Including counterfactuals (and restrictions on backtracking counter-
factuals) with an account of lawful suciency, on the other hand, helps us address
problems for regularity accounts with distinguishing causes from eects and eects of
a common cause.28
The second major change is the use of aspects rather than events simpliciter as
causes and eects. Return to our case of late preemption with C. Louise and Possum.
It seems right to say that Possums mass lawfully entails some of the properties of C.
Louises eating of the y (those properties caused by the minute gravitational forces
his mass exerts on C. Louise and the y), and that if Possums mass were changed
these properties would also change. So Possums mass counts as a cause of some of
the properties of the eating. Is this a problem for our account?29
I think rather that it is an advantage. In some contexts we might need to ask about
what caused these minute gravitational eects, and then we would need to be able to
cite Possums mass as among the causes. Note that, unlike Lewis, I need not hold
that Possums mass is a cause of the eating of the y; rather, his mass is merely
among the causes of the minute gravitational eects. Lewis simply denies that we
should attend to these negligible dierences: But in a context where we need to ex-
plain these dierences such a denial is inexplicable. We need aspects to have an ade-
quately precise account of the causal story.
Turning from the problems with simple spurious causation to those with transi-
tivity, moving to aspects solves the so-called transitivity problem where the ipping
of the switch has to count as a cause of Jesses death. When we eliminate the extra,
unwanted information brought in by the events, we see that transitivity is not the
culprit: Events as the causal relata are responsible for the unwanted consequences.
As the story was told, when the switch is ipped, the only property (or properties)
it causes are the trains being on the left-hand track, and any other properties to do
solely with the trains being on the left (e.g., the trains having the property of passing
the scenery on the left more closely than if it had traveled on the right). The ipping
lawfully entails the trains traveling on the left, and the ipping of the switch inu-
ences the traveling on the left. The other properties of the trains travel (which are
220 L. A. Paul
relevant to Jesses death) while it travels on the left are caused by properties of the
trains prior motion, mass, and so on, not by the ipping of the switch: The ip-
ping does not lawfully entail them, and they are not inuenced by the ipping. (If the
ipping had occurred dierently or not at all, all of these properties would have
occurred just as they did in the original story.)
Because the only (relevant) eect of the ipping is the trains traveling on the left,
the ipping does not cause any of the properties of Jesses death. The trains merely
having the property of traveling on the left does not lawfully entail the death of
James. Nor does the property of traveling on the left inuence properties of Jamess
death, since even if the train had traveled on the right, Jamess death would not have
been aected.30
Note that, under the aspects account, if we were to tell the train story dierently,
the account would reect the right changes in the results. Telling the story very dif-
ferently might muddle our intuitions enough to make it unclear whether the ipping
should count as a cause of properties of the death. In such a case, an answer either
way, or no clear answer at all, is acceptable. Telling the story slightly dierently, so
that the train is somehow slowed or changed by taking the left track and the death
happens later or is more violent than if the train had taken the right track, makes the
ipping of the switch count among the causes of those properties (e.g., the time or
manner of death)just as it should.31
5 Conclusion
By recognizing that property instances are the causal relata, we can conform to com-
mon sense; we can develop, unimpeded, the account of the individuation of events
(and individuals) that best conforms to our antecedent beliefs; and we can allow our
causal relata to be at least as ne grained as facts.
By combining this view with lawful entailment and inuence, we arrive at a well-
developed theory of property causation (perhaps most useful in the context of cur-
rent discussions of mental causation), we solve many of the most pressing problems
with extant analyses of causation, and we have an account of causation that ts well
with actual scientic practice. By this last claim, I mean that when doing science, we
identify the potential causes eligible to bring about an eect (those that lawfully en-
tail the eect), and then alter the potential causes and observe the eect for changes.
When doing metaphysics rather than empirical science, we evaluate close possible
worlds to see if alterations of the potential cause (one that lawfully entails the eect)
result in alterations of the eect. If we have the right sort of dependence, we have
causation. If not, not. The account I propose is modeled after the ways we rely on to
learn about causation in experimental contexts, so it is no surprise that it does such a
good job of handling cases where we have clear causal intuitions and a full descrip-
tion of the situation. We thus clear away a number of problems that have hampered
the development of the analysis of the causal relation.
Acknowledgments
I received helpful comments from many, but I am especially indebted to Ned Hall,
Chris Hitchcock, David Lewis, Jonathan Schaer, Stephen Yablo, and an audience
at the 1998 Australasian Association for Philosophy.
Notes
1. The example is modeled on a nice example used in Michael McDermott (1995a). McDermott uses his
example to create trouble for a particular kind of counterfactual account that takes events to be coarse
grained but does not discuss its broader implications.
2. Lewis (2000).
3. See David Lewis (1973a) and (1986b).
4. Kim (1973a) and (1993), pp. 3352.
5. Kim (1993), pp. 4546.
6. Bennett (1988), p. 83.
7. Lewis (1986d).
8. Stephen Yablos work goes a great way toward handling many of the problems I discuss in this paper,
but also relies on extremely strong essentialist claims and the ontology these claims imply. I prefer a more
parsimonious account (if one can be had). Yablo develops his account in his (1992a, 1992b, 2000).
9. Davidson (1969, 1970) originally tried to individuate events by their causes and eects, but such a
strategy requires a prior analysis of causation. Quine (1985) argued that the strategy didnt succeed, and
Davidson (1985) agreed. After writing this paper, I learned that Douglas Ehring (1997) and Daniel Haus-
man (1998) raise similar transitivity worries for Davidson.
10. There is one more move that the coarse grained theorist might try: Following Davidson (1967), he
might say that The accident caused the publication of the paper is true, but not causally explanatory. In
other words, the claim would be that, strictly speaking, the accident caused the publication even though
our singular causal statement doesnt mention any causally relevant properties of the accident as part of
the description it gives of the cause (the accident). This seems as though it would allow the coarse-grained
theorist to accept the rst link of the causal chain (as well as subsequent links). But this response is inade-
quate, for the skeptic can return with the point that no matter how the accident is described, the event of
the accident has no properties that are causally relevant to the eect (the publication). Davidsons argu-
ment requires that the event that counts as the cause have some properties under some description that are
causally relevant to the eect, even if these are unobvious (e.g., such as the properties of Suzys birth that
are causally relevant to Suzys writing her paper).
11. Kim (1993), pp. 4445.
12. Williams (1953) and Campbell (1990) suggest that tropes should be the causal relata. Douglas Ehring
(1997) argues for persisting tropes as causal relata, and Hausman (1998) argues that tropes aid causal
explanation.
222 L. A. Paul
13. For the purpose of using aspects as part of a theory of causation, I leave the controversial question
of how we dene particulars aside by stipulating that the aspects dont have to be part of the structure
we take to be the particular. Just as C. Louises sneezing isnt part of the structure that is C. Louise (and
unlike the way C. Louises left front paw is part of the structure that is C. Louise), the aspects that are the
causal relata need not be the structure we take to dene them.
14. D. H. Mellor (1995, chap. 12) has a very nice discussion of (what he calls) the dierence between
causing and aecting particulars.
15. In my account of dependence, I am assuming a standard ordering relation on closeness of worlds based
on Lewiss similarity ordering, in which propositions about property instances are eligible to be included
among the relevant conditions for the evaluation of counterfactuals.
16. Viz., Kims (1973a) account of subsumption. My account requires an adequate specication of the
causal laws and the properties they refer to, but does not require allegiance to a particular theory of
laws: one could rely on the DretskeArmstrongTooley view of nomic necessity, some sort of primitive-
ness claim, or the MillRamseyLewis view that the right regularities are those that belong to the set that
does the best job of systematizing and organizing physical information. The denition might change
slightly depending on the theory of properties adopted: For simplicity, Ive written the denition here for
aspects taken as tropes.
17. Paul (1998b).
18. Schaer (2000b).
19. I do not claim that there can be no such thing as causation by action at a distance. Rather, once we
eliminate the need for chains of events in our cases of causation, our intuitions change enough to motivate
a change in the conditions necessary for causation: Causation by action at a distance is dierent enough
from ordinary causation to need an analysis of its own. (So cases that combine ordinary causation with
causation by action at a distance will require a combined analysis, and its no surprise if cases involving
action at a distance are not handled by an analysis for contiguous causation.) I suspect that the analysis of
causation by action at a distance will be a version of the lawful entailment analysis for aspects. Hall, Two
Concepts of Causation, chapter 9 in this volume, argues that causation by action at a distance is a kind of
causation that diers from the ordinary kind, and he develops an analysis based on minimally sucient
sets. Similarly, there are problems for causation by omission.
20. See Lewiss (1973a) and (1986b) discussions of early and late preemption, and his stepwise solution to
early preemption. In these papers Lewis lumps cases involving action at a distance in with late preemption
(where the eect is delayed): I think this is a miscategorization.
21. Lewis (2000).
22. Ibid. Lewis refers to these cases as Black-Red examples. Also see Hall (2000) for some particularly
ingenious cases. There are more variations of these kinds of cases than I can deal with here, including well-
known probabilistic versions. The earliest discussions of the (deterministic) cases that Ive found are in van
Inwagen (1987), Rowe (1993), and Fischer and Ravizza (1993b).
23. Lewiss inuence account also fails to handle the skiing example: The skiing accident inuences the
writing of the paper and the writing of the paper inuences the publication.
24. Following Lewiss account of alterations of events, I take no rigid stand on much we can alter a prop-
erty for our account of inuence: It is a matter of degree and context. I also leave vague how we decide
which alterations are closer to the original p and which are more remote.
25. I hope, with Lewis, that a suitably modied version will be able to handle indeterministic causation.
26. Adopting Stalnakers (1980) account would allow us to select only one closest world. This does not
solve our problem but merely shifts the worry, since depending on the selection function we adopt we
could get dierent values for the counterfactual. In other words, truth or falsity of the counterfactual
would be relative to the selection function. On Stalnakers account we can supervaluate, so a counter-
factual conditional is true (false) i it is true (false) for every selection function, but the conditional is
undened in the case of vagueness.
27. For those who nd inuence unacceptable or who are not concerned about trumping cases, I advocate
a view that combines simple counterfactual dependence and lawful entailment (versions of (1) and (2) in
sec. 2) together with aspects.
28. Lewis (1973a) shows how counterfactual based accounts can address these problems.
29. Lewis discusses this problem in Lewis (2000), as does Jonathan Schaer in Schaer (2001).
30. Here I am relying on traveling on the right as the relevant property alteration to evaluate inuence.
31. My solution in this section helps to resolve problems with explaining why Frank is not morally
responsible for the death of James because he ipped the switch: Merely ipping the switch didnt make
him causally responsible for the death, so neither did it make him morally responsible.
224 L. A. Paul
9Two Concepts of Causation
Ned Hall
1 Introduction
Causation, understood as a relation between events, comes in at least two basic and
fundamentally dierent varieties. One of these, which I call dependence, is simply
that: counterfactual dependence between wholly distinct events. In this sense, event c
is a cause of (distinct) event e just in case e depends on c; that is, just in case, had c
not occurred, e would not have occurred. The second variety is rather more dicult
to characterize, but we evoke it when we say of an event c that it helps to generate
or bring about or produce another event e, and for that reason I call it production.
Here I will articulate, defend, and begin to explore the consequences of this distinc-
tion between dependence and production. A synopsis:
After taking care of some preliminaries (sec. 2), I will argue for the distinction in a
slightly devious manner, by starting with a broad-strokes critique of counterfactual
analyses of causation (sec. 3). The reason for this approach is plain: Since I end up
endorsing the simplest kind of counterfactual analysisalbeit only as an analysis of
one kind of event-causationit makes sense to pay some attention to the prospects
for this and kindred analyses, and to examine why there is no hope of turning them
into analyses of a univocal concept of event-causation. Specically, my critique will
aim to show that the best attempts to shore up counterfactual analyses in the face of
well-known and stubborn counterexamples (involving certain kinds of overdetermi-
nation) rely on three general theses about causation:
Transitivity: If event c is a cause of d, and d is a cause of e, then c is a cause of e.
Locality: Causes are connected to their eects via spatiotemporally continuous se-
quences of causal intermediates.
Intrinsicness: The causal structure of a process is determined by its intrinsic, non-
causal character (together with the laws).
These thesesparticularly the second and thirdwill require more discussion and
elaboration, which will come in due time. For now, contrast them with the thesis that
lies at the heart of all counterfactual analyses of causation:
Dependence: Counterfactual dependence between wholly distinct events is sucient
for causation.
The simplest counterfactual analysis adds that dependence is necessary for causa-
tion. As a general analysis of causation, it fails for well-known reasons, which we will
review shortly. Consequently, very few counterfactual analysts endorse this necessary
condition (but see David Coadys Preempting Preemption, chapter 13 in this vol-
ume). But to my knowledge, all endorse the sucient condition codied in the thesis
of Dependence. Indeed, it is probably safe to say that Dependence is the cornerstone
of every counterfactual analysis.
What is the trouble? Simply this: A hitherto ignored class of examples involving
what I call double-prevention reveals deep and intractable tensions between the
theses of Transitivity, Locality, and Intrinsicness, on the one hand, and Dependence,
on the other (sec. 4).
In section 5, Ill add to my case by arguing that exactly parallel tensions divide the
rst three theses from the thesis of
Omissions: Omissionsfailures of events to occurcan both cause and be caused.
This thesis will also need further elaboration and discussion.
One immediate result is that counterfactual analyses are doomed to failure (unless,
as I think, they are understood to be targeted narrowly at just one kind of event-
causation): for they need the rst three theses if they are to cope with the well-known
counterexamples involving overdetermination, but they cannot abide these theses if
they are to cope with the counterexamples involving double-prevention (or, for that
matter, if they admit omissions as causes and eects).
Although important, this result is eclipsed by a more signicant lesson that I will
develop in section 6. For the ve theses I have mentioned are, I claim, all true. Given
the deep and intractable tensions between them, that can only be because they charac-
terize distinct concepts of causation. Events can stand in one kind of causal relation
dependencefor the explication of which the counterfactual analysis is perfectly
suited (and for which omissions can be perfectly suitable relata). And they can stand
in an entirely dierent kind of causal relationproductionwhich requires an en-
tirely dierent kind of analysis (and for which omissions are not suitable relata). De-
pendence and Omissions are true of the rst of these causal relations; Transitivity,
Locality, and Intrinsicness are true of the second. Ill close section 6 by defending this
claim against some of the most obvious objections.
How are production and dependence to be analyzed? Dependence, I think, is easy;
it is counterfactual dependence, nothing more nor less (with, perhaps, the proviso
that counterfactual dependence itself can come in dierent varieties; see sec. 7 for
brief discussion). Production is trickier, and in section 7 Ill oer a speculative pro-
posal about its analysis, conned to the special case of deterministic laws that permit
no action at a temporal distance or backward causation. But Ill say at once that I
226 N. Hall
am much more condent of the propriety of the distinction than I am of this partic-
ular gloss on the production half of it.
Ill close, in section 8, by suggesting some ways in which the distinction between
production and dependence might be put to work, and by highlighting what I think
are the most important bits of unnished business.
2 Preliminaries, and a Brief Methodological Sermon
There are, in the literature, more than a dozen versions of a counterfactual analysis
of causation that I am aware of. To attack them all, in detail, would require (to
borrow an apt term from Tim Maudlin) a kind of philosophical trench warfare
that only deeply committed partisans could nd engaging. Ill confess to a taste for
trench warfare, but I wont indulge it here. Instead, I will follow a dierent strat-
egy, focusing my critique on the simplest counterfactual analysis, according to which
causation is counterfactual dependence between wholly distinct events. It will be far
more illuminating to explore the most basic problems for this analysisalong with
the clearest and most plausible strategies for confronting themthan it would be
to wind through the convolutions built into the multitude of more sophisticated
variants.
In order to develop this critique as constructively as possible, we must avoid vari-
ous methodological pitfalls. For that reason, it will be important to characterize, if
only in a rough way, the causal relation that is the target of the counterfactual anal-
ysis. I take the analysis to concern the concept of causation as a transitive, egalitarian
relation between localized, datable events. Lets look at the parts of this character-
ization in turn.
Begin with the relata. In understanding them to be events, I am taking sides on
an issue that has seen much recent controversy.1 I grant that there may be senses in
which noneventsfacts, properties, maybe even thingscan cause and be caused;
certainly we speak of event types as doing so, as when we say that lightning causes
res. All the same, I assume that there is a clear and central sense of causethe
one at issue herein which causes and eects are always events. (In sec. 6, Ill qual-
ify this assumption slightly, suggesting that dependence, at least, can admit more
kinds of relata.)
I will, furthermore, follow common practice by stretching ordinary usage of the
term event to cover such things as, for example, the presence, at the appropriate
time, of the oxygen and dry timber that combine with the lightning bolt to produce
the forest re. I will also take it for granted that we can adequately discern when
Two Concepts of Causation 227
two events fail to be wholly distinctthat is, when they stand in some sort of
logical or mereological relationship that renders them unsuited to stand in causal
relationshipsand that we can tell when a description is too disjunctive or extrinsic
to succeed in picking out an event. Without such assumptions, it is far too easy to
make a hash of the simple analysis, and the analyses that build on it, by way
of alleged counterexamples to the claim that counterfactual dependence is sucient
for causation.2
Of course, I do not at all mean to suggest that it is an easy matter to provide an
adequate philosophical account of events that meets these criteria. I certainly wont
try to provide any such account here. What I will do is avoid choosing examples
where any of the controversies surrounding the nature of events makes a dierence.3
Turn next to the characterization of the relation. Transitivity is straightforward
enough: If event a is a cause of event b, and b a cause of c, then a is thereby a cause
of c. What I mean by egalitarian can best be made clear by contrast with our usual
practice. When delineating the causes of some given event, we typically make what
are, from the present perspective, invidious distinctions, ignoring perfectly good
causes because they are not suciently salient. We say that the lightning bolt caused
the forest re, failing to mention the contribution of the oxygen in the air, or the
presence of a sucient quantity of ammable material. But in the egalitarian sense of
cause, a complete inventory of the res causes must include the presence of oxy-
gen and of dry wood. (Note that transitivity helps make for an egalitarian relation:
Events causally remote from a given event will typically not be salientbut will still
be among its causes, for all that.)
Now for a brief methodological sermon: If you want to make trouble for an anal-
ysis of causationbut want to do so on the cheapthen its convenient to ignore the
egalitarian character of the analysandum. Get your audience to do the same, and you
can proceed to elicit judgments that will appear to undermine the analysis, but which
are in fact irrelevant to it. Suppose that my favorite analysis counts the big bang as
among the causes of todays snowfall (a likely result, given transitivity). How easy it
is to refute me, by observing that if asked what caused the snowfall (better still: what
was the cause of it), we would never cite the big bang! Of course, the right response
to this refutation is obvious: It conates the transitive, egalitarian sense of cause
with a much more restrictive sense (no doubt greatly infected with pragmatics) that
places heavy weight on salience.
A simple mistake, it would seem. But the same sort of mistake shows up, in more
subtle forms, in examples drawn from the literature. It will be helpful to work through
a few illustrative casesones that show, incidentally, how even rst-rate authors can
sometimes go astray.
228 N. Hall
First, Bennett (1987, pp. 222223; italics in the original), who is here concerned
with Lombards thesis that an events time is essential to it:
Take a case where this is true:
There was heavy rain in April and electrical storms in the following two months; and in June
the lightning took hold and started a forest re. If it hadnt been for the heavy rain in April,
the forest would have caught re in May.
Add Lombards thesis to that, and you get
If the April rain hadnt occurred the forest re wouldnt have occurred.
Interpret that in terms of the counterfactual analysis and you get
The April rains caused the forest re.
That is unacceptable. A good enough theory of events and of causation might give us reason
to accept some things that seem intuitively to be false, but no theory should persuade us that
delaying a forests burning for a month (or indeed for a minute) is causing a forest re.
Lombard agrees that Bennetts result is unacceptable. It is a bit of good common
sense that heavy rains can put out res, they dont start them; it is false to say that
the rains caused the re (Lombard 1990, p. 197; italics in the original).
Lombard discusses a second example that shows that the essentiality of an events
time is not at issue (ibid., pp. 197198):
Suppose that Jones lives in a very dangerous neighborhood, and that one evening Smith at-
tempts to stab him to death. Jones is saved because of the action of Brown who frightens
Smith o. However, a year later, Jones is shot to death by the persistent Smith. So, if Browns
action had not occurred, Joness death due to the shooting would not have occurred, since he
would have died of stab wounds a year earlier. But, I nd it intuitively quite unacceptable to
suppose that Browns action was a cause of Joness dying as a result of gunshot a year later.
Finally, Lewis discusses a similar example (Lewis 1986d, p. 250):
It is one thing to postpone an event, another to cancel it. A cause without which it would have
occurred later, or sooner, is not a cause without which it would not have occurred at all. Who
would dare be a doctor, if the hypothesis under consideration [that an events time is essential
to it] were right? You might manage to keep your patient alive until 4:12, when otherwise he
would have died at 4:08. You would then have caused his death. For his death was, in fact, his
death at 4:12. If that time is essential, his death is an event that would not have occurred had
he died at 4:08, as he would have done without your action. That will not do.
If these examples are meant to provide rock-solid data on which the counter-
factual analysis (and perhaps others) founders, then they uniformly failfor in each
case, we can nd independently plausible premises that entail the allegedly unaccept-
able consequences. Of course that doesnt show that the consequences are true. But
it does show that we make a serious methodological mistake if we treat those of our
intuitions that run counter to them as nonnegotiable data.
First we must disentangle irrelevant but confusing issues. It is probably right that
an events time is not in every case essential to it; but ( pace Lewis) that doesnt help
in any of the three cases. This is more or less obvious in the rst two cases (the June
re is not the same as the re that would have happened in May; the death by
shooting is not the same as the death by stabbing that would have happened a year
earlier). So consider Lewiss case. Supposedly, it will not do to assert that the
doctors action is among the causes of the patients death. But what does this have
to do with the proximity of the actual time of death to the time at which the patient
would have died? Suppose you manage to keep your patient alive until June of 2004,
when otherwise he would have died in June of 2003. Would you then have caused his
death, since without your action the death he in fact died would not have occurred?
It is no less (and no more) unacceptable to say yes in this case than it is to say
yes in Lewiss case. But if, following Lewis, we conclude that the actual death is
the same as the death that would have occurred a year earlier, then we are taking
the denial of the essentiality of times to a ridiculous extreme. Such a denial, however
warranted, does not give the counterfactual analyst the means to respond eectively
even to Lewiss problem.
The analyst can, however, draw on our brief methodological sermon to point to
two sorts of judgments about causation that the three examples implicitly trade on
but illegitimately, since these judgments concern types of causation that are not at
issue. We can all agree that heavy rains can put out res, they dont start them,
just as we can agree that smoking causes lung cancer, but regular exercise doesnt.
So what? The intuitions called on here do not concern the concept of causation as a
transitive, egalitarian relation between events, but rather some other concept of cau-
sation as an inegalitarian relation between event-types. (Never mind that starting a
re is not the only way to be one of its causes!) Similarly, we can all agree that it is
the lightning that causes the forest re, and nothing elseincluding the heavy rains.
Again, so what? Here we seem to have in mind a restricted, inegalitarian concept of
event-causation according to which events that are to count as causes must be par-
ticularly salient in some respect; but judgments involving this concept matter not at
all to the counterfactual analysis, since it concerns the weaker and more inclusive
transitive, egalitarian concept.
Unfortunately, Bennett, Lombard, and Lewis have all muddied the question of
whether the counterfactual analysis is adequate by choosing examples where intu-
230 N. Hall
itions of the two types just discussed are particularly strong and seemingly salient: Its
not the rainfall that causes the June re, but rather the lightning; moreover, its just
good common sense that heavy rains dont cause res!4 Of course, while recog-
nizing these points you might still judge these cases to have some intuitive force as
counterexamples. Fair enough; they do. But a more careful examination shows how
hasty it would be to take any such intuitions as decisive. Ill make the case against
Bennett, after which it will be clear enough how to proceed against Lombard and
Lewis that we can leave those cases aside.
The idea is to nd an event intermediate between the cause and its alleged eect
that is clearly a cause of the second, and at least plausibly an eect of the rst. So
observe that among the causes of the June re is not just the lightning but also the
very presence of the forest, lled with ammable material. The presence of the forest
in the hours before the lightning strikes is itself an event, or perhaps a collection of
events. This event is a cause of the June re (albeit not a salient cause). What are its
causes? A typical counterfactual analysis will claim that one of its causes is the April
rainfall, since without the rainfall the forest would have been destroyed in May. But
we can argue for the plausibility of this claim independently, by noting that the fol-
lowing judgments seem, intuitively, to be correct: It is in part because of the April
rains that the forest is present in June; any complete causal explanation of the forests
presence must cite the role of the April rains in preventing its destruction; the April
rains are at least in part responsible for the presence of the forest in June.5
One could deny the truth of these judgments, or deny that they show that the April
rainfall is a cause of the forests presence in June, or deny that causation is transitive
in the way that is needed to complete the inference to the claim that the April rainfall
is among the causes of the June forest re. But unless one can nd some grounds for
supporting such denialsgrounds independent of the mere intuitive implausibility
of the claim in questionthen this implausibility will fail to provide a particularly
compelling reason for giving up the counterfactual analysis. (Exactly parallel points
apply to the other two examples.) Happily, I think the conclusions drawn in section
6 clear up what is going on in the rainfall case (and the other cases), precisely by
showing that counterfactual analyses utterly fail to capture one important sense of
causeproductionand that in this sense the April rains are not among the
causes of the June re. But it will take some work to get there, and along the way
we must not be distracted by the temptations of such bogus refutations as those
we have just examined. Intuitions about cases must be heeded, to be sure. But not
blindly.
Onward. It will help to have a means of representing simple causal structures; ac-
cordingly, I will adopt the neuron diagrams used by Lewis (1986b).
The diagram in gure 9.1 depicts a pattern of neuron rings. Dark circles represent
ring neurons, while arrows represent stimulatory connections between neurons. The
order of events is left to right: In gure 9.1 neuron a res, sending a stimulatory sig-
nal to neuron b, causing it to re; bs ring in turn sends a stimulatory signal to
neuron c, causing it to re. We will also need a way to represent prevention of one
event by another. So let us add inhibitory connections to the neuron diagrams, rep-
resented by a line ending in a solid dot, as shown in gure 9.2. In the left-hand dia-
gram, neurons a and c re simultaneously; cs ring causes d to re, which in turn
causes e to re. However, thanks to the inhibitory signal from c, as ring does not
cause b to re; bs failure to re is represented by leaving its circle unshaded. The
right-hand diagram shows what would have happened if c had not red.
Calling these diagrams neuron diagrams is merely picturesque; what is impor-
tant about them is that they can provide, in a readily digestible form, partial repre-
sentations of many causal structures.
3 The Simple Counterfactual Analysis, and Two Kinds of Overdetermination
3.1 The Simple Analysis
Both for simplicity and to avoid needless controversies, I will focus only on the coun-
terfactual analysis as it applies to worlds with deterministic laws that permit neither
backwards causation nor action at a temporal distance (although the lessons of the
a b c
Figure 9.1
c d
e
a b
c fires
c d
e
a b
if c hadnt fired
Figure 9.2
232 N. Hall
paper apply much more generally, as far as I can see). I will also leave aside cases, if
such there be, where a cause is simultaneous with one of its eects.
Here and throughout I will denote events by lower-case italicized letters a, b,
c, and so forth; the proposition that an event e occurs by Oe; and the counter-
factual or subjunctive conditional by L (read: were it the case that . . . then it
would be the case that . . .).6 The simple analysis is as follows:
Event c is a cause of event e i
(i) c and e are wholly distinct events;
(ii) Oc; Oe, and sOc LsOe are all true.
An immediate problem arises, whose solution requires the counterfactual condi-
tional to be understood in a rather specic way. In gure 9.1, it is certainly correct to
say that if a hadnt red, c wouldnt have; but it may also be correct to say that if c
hadnt red, a wouldnt have. (If you dont like the sound of that, try the happier
locution: If c hadnt red, it would have to have been that a didnt re.) If so, the
analysis wrongly says that c is a cause of a. (Note the harmless but convenient am-
biguity: Letters sometimes refer to events, sometimes to neurons.)
Two responses seem possible. We could augment the analysis by adding some
third condition to guarantee the asymmetry of causation (for example: causes must
precede their eects). Or we could deny the truth of the oending counterfactual,
appealing to an account of the conditional that secured the falsehood of all such
backtrackers (to use Lewiss apt term: see his 1979a). Swain (1978), for example,
opts for the rst alternative, Lewis (1979a and 1986a) (and most other counterfactual
analysts) for the second.7
The rst response doesnt work, partly for reasons that have been well explored
and that I wont rehearse in detail here (e.g., merely adding the requirement that
causes precede their eects wont help if, say, c and e are joint eects of some event
a, with c occurring before e; for we could still reason that if c hadnt happened, it
would have to have been that a didnt happen, and therefore that e didnt happen). A
dierent, often unnoticed reason for rejecting the rst response deserves some dis-
cussion, however: The problem is that this response implicitly supposes that back-
trackers threaten only the suciency of the above analysis. If that were true, it would
make sense to add further conditions, so as to make the analysis less liberal. But
backtrackers also undermine its necessity, as gure 9.2 shows.
In gure 9.2, d is, clearly, a cause of e. But if, in evaluating counterfactuals with
the antecedent d does not occur, we proceed by making minimal alterations to
the past events that led to d, then we will reach a counterfactual situation in which c
does not occur, but a still doesthat is, a counterfactual situation in which e occurs.
That is, if we allow as true the backtracker sOd LsOc, then the right-hand dia-
gram also describes what would have happened if d hadnt red, and so the condi-
tional sOd LsOe is false. Then how can it be that d turns out to be a cause of e?
Adding extra conditions to (i) and (ii) provides no answer.8 (Nor will it help to lib-
eralize the analysis in the standard way, by taking causation to be the ancestral of
counterfactual dependence. For the problem that threatens the connection between
e and d will equally threaten the connection between e and any event that mediates
between d and e.)
In short, reading the counterfactual in a backtracking manner destroys the depen-
dence of e on d. Thats not only trouble for the simple analysis: Its just wrong, since
it manifestly is the case that if d hadnt red, e wouldnt have. Or, more cautiously,
there manifestly is an acceptable reading of the counterfactual conditional accord-
ing to which this is true. I will henceforth take it for granted that both the simple
analysis and its more elaborate kin employ such a nonbacktracking reading of the
conditional.
It is not obvious how to provide a general semantics for the counterfactual that
will secure this reading. Fortunately, for the purposes of this paper we need only
come up with a rule for evaluating counterfactuals of the form sOc LsOe, where c
and e both occur, and c precedes e. Following Maudlin (2000), I suggest the follow-
ing: Letting t be the time of occurrence of the given event c, we evaluate the condi-
tional sOc LsOe by altering the state of the world at t just enough to make the
antecedent true, evolving that state forward in time in accordance with the (actual)
laws, and seeing whether the consequent comes out true.9 So, in gure 9.2, if d hadnt
red, circumstances contemporaneous with its ring such as the nonring of b would
have been unchanged, and so e would not have red.
3.2 Early Preemption
The simple analysis may appear quite able to stave o challenges to its suciency.
But obvious problems beset its claim to necessity. Consider a case of ordinary pre-
emption, as in gure 9.2. The ring of e is overdetermined by the simultaneous rings
of a and c. But not in a way that leaves us at all uncertain as to what causes what:
Without question, c (and not a) is a cause of e, even though if c hadnt occurred, e
would have occurred anyway, thanks to an alternative process, beginning with a,
that c preempts.
There is an obvious strategy for handling this kind of case. First, we liberalize our
analysis, by taking causation to be the ancestral of counterfactual dependence: c is a
cause of e i there are events d
1
; . . . ; d
n
such that d
1
counterfactually depends on c, d
2
234 N. Hall
depends on d
1
; . . . ; and e depends on d
n
. Next, we look for an event d (or sequence
of events) intermediate between the preventing cause c and the eect e, such that e
depends on d and likewise d on c. The strategy works handily in the case before us
(provided, once again, that we are careful not to interpret the counterfactual in a
backtracking sense, according to which, had d not red, it would have to have
been the case that c didnt re, and so b would have red, and so e would still have
red).
Observe how natural this embellishment to the simple analysis isand observe
that it gives a central role to the Transitivity thesis.
3.3 Late Preemption
Other, quite ordinary cases of overdetermination require a dierent treatment. Con-
sider a case of so-called late preemption, as in gure 9.3. Neurons a and c re simul-
taneously, so that e res at the same time as b; the inhibitory signal from e therefore
prevents d from ring. If c hadnt red, e still would have; for in that case d would
not have been prevented from ring and so would have stimulated e to re. Likewise
for every event in the causal chain leading from c to e: If that event had not occurred,
e would nevertheless have red. So the strategy of nding suitable intermediates
breaks down; for it to succeed, e would have to depend on at least one event in the
chain leading back to c, and it does not.
Here is another example, with a slightly dierent structure; it illustrates how ab-
solutely mundane these cases are. Suzy and Billy, expert rock-throwers, are engaged
in a competition to see who can shatter a target bottle rst. They both pick up rocks
and throw them at the bottle, but Suzy throws hers a split second before Billy. Con-
sequently Suzys rock gets there rst, shattering the bottle. Since both throws are
perfectly accurate, Billys would have shattered the bottle if Suzys had not occurred,
so the shattering is overdetermined.
Suzys throw is a cause of the shattering, but Billys is not. Indeed, every one of the
events that constitute the trajectory of Suzys rock on its way to the bottle is a cause
of the shattering. But the shattering depends on none of these events, since had any
of them not occurred the bottle would have shattered anyway, thanks to Billys ex-
pert throw. So the transitivity strategy fails.
Three alternative strategies for dealing with this kind of case suggest themselves.
The rst rests on the observation that Suzys throw makes a dierence to the time
and manner of the shattering, whereas Billys does not. The second rests on the
observation that Suzys throw is connected to the shattering by a spatiotemporally
continuous chain of causal intermediates, whereas Billys is not. And the third rests
on the observation that there is a sequence of events connecting Suzys throw to the
c

f
i
r
e
s
i
f
c

h
a
d
n
t

f
i
r
e
d
c
e
a
b
d
c
e
a
b
d
F
i
g
u
r
e
9
.
3
236 N. Hall
shattering that has the right sort of intrinsic character to count as a causal sequence,
whereas no such sequence connects Billys throw to the shattering. Let us consider
these strategies in turn.
There are various ways to implement the rst strategy. For example, we could
deny that the eect that does the preventing is numerically the same as the eect that
would have occurred via the alternative process. (E.g., in gure 9.3, the ring of e
that would have occurred, had c not red, is not the same event as the ring that
actually occurs.) If so, then our two examples do exhibit the needed pattern of coun-
terfactual dependence, since the eect that actually occurred would not have oc-
curred without its cause (although a very similar event would have occurred in its
place). Alternatively, we could remain silent about the individuation of events, and
simply employ a slightly dierent counterfactual in the analysissay, by counting
c a cause of e if and only if, had c not occurred, e would not have occurred at the
time it actually did (Paul 1998b). Lewis (Causation as Inuence, chapter 3 in this
volume) argues that we should count c a cause of e if there is a suitable pattern of
counterfactual dependence between various dierent ways c or something like it
might have occurred and correspondingly dierent ways in which e or something like
it might have occurred. (Lewis proposes taking causation itself to be the ancestral of
this relation.)
These approaches are uniformly nonstarters. Never mind the well-known problems
(e.g., that noncauses can easily make a dierence to the time and manner of an
events occurrencea gust of wind that alters the course of Suzys rock ever so
slightly, for example). What seems to have gone unnoticed is that it is not at all es-
sential to examples of late preemption that the genuine cause make any dierence to
the time or manner of the eect. As Steve Yablo pointed out to me, its easy enough
to construct cases in which c is clearly a cause of e, but in which neither c nor any
event causally intermediate between it and e makes the slightest dierence to the way
e occurs. Yablo observes that we can simply alter the story of Billy and Suzy. This
time, Billy throws a Smart Rock, equipped with an onboard computer, exquisitely
designed sensors, a lightning-fast propulsion systemand instructions to make sure
that the bottle shatters in exactly the way it does, at exactly the time it does. In fact,
the Smart Rock doesnt need to intervene, since Suzys throw is just right. But had
it been any dierentindeed, had her rocks trajectory diered in the slightest, at
any pointthe Smart Rock would have swooped in to make sure the job was done
properly. Sure, the example is bizarre. But not in a way that matters in the slightest
to the evaluation of the causal status of Suzys throw: Smart Rock notwithstanding,
her throw is still a cause of the shatteringeven though neither it nor any event that
mediates between it and the shattering makes a dierence to the time or manner of
that shattering.
I wont consider these approaches further. It will be far more instructive for us to
focus on the two alternative strategies.
Suzys throw is spatiotemporally connected to the shattering in the right way, but
Billys is not. So perhaps we should add the Locality thesis as a constraint on the
analysis: Causes have to be connected to their eects via spatiotemporally continu-
ous sequences of causal intermediates. Now, on the face of it this is a step in entirely
the wrong direction, since it makes the analysans more stringent. But if we simulta-
neously liberalize the analysis in other respects, this strategy might work. For ex-
ample, we might say that c is a cause of e just in case there is a spatiotemporally
continuous sequence of events connecting c with e and a (possibly empty) set S of
events contemporaneous with c such that each later event in the sequence (includ-
ing e) depends on each earlier eventor at least would have, had the events in S not
occurred. That will distinguish Suzys throw as a cause, and Billys as a noncause.
Of course, since action at a distance is surely possible, and so Locality at best a
highly interesting contingent truth, this amended counterfactual analysis lacks gen-
erality. But it is patently general enough to be of value. At any rate, it is not so im-
portant for our purposes whether this strategy, or some variant, can handle all cases
of late preemption. What is important is that it is a plausible and natural strategy to
pursueand it gives a central role to the Locality thesis.
Lewis has proposed a third, dierent strategy. He begins with the intuition that the
causal structure of a process is intrinsic to it (given the laws). As he puts it (1986b,
p. 205):
Suppose we have processescourses of events, which may or may not be causally connected
going on in two distinct spatiotemporal regions, regions of the same or of dierent possible
worlds. Disregarding the surroundings of the two regions, and disregarding any irrelevant
events that may be occurring in either region without being part of the process in question,
what goes on in the two regions is exactly alike. Suppose further that the laws of nature that
govern the two regions are exactly the same. Then can it be that we have a causal process in
one of the regions but not the other? It seems not. Intuitively, whether the process going on in
a region is causal depends only on the intrinsic character of the process itself, and on the rele-
vant laws. The surroundings, and even other events in the region, are irrelevant.
In cases of late preemption, the process connecting cause to eect does not exhibit
the right pattern of dependencebut only because of accidental features of its sur-
roundings. The process that begins with Suzys throw and ends with a shattered
bottle does not exhibit the right pattern of dependence (thanks to Billys throw), but
it is intrinsically just like other possible processes that do (namely, processes taking
238 N. Hall
place in surroundings that lack Billy, or a counterpart of him). Lewis suggests, in
eect, that for that reason Suzys throw should count as a cause.
Clearly, Lewis is trying to parlay something like the Intrinsicness thesis into an
amended counterfactual analysis, one adequate to handle cases of late preemption.
Now, I think there are serious problems with the details of Lewiss own approach
(spelled out in the passage following that just quoted), but since that way lies trench
warfare, I wont go into them. I do, however, want to take issue with his statement
of the Intrinsicness thesis, which is too vague to be of real use. What, after all, is a
process or course of events? If it is just any old sequence of events, then what he
says is obviously false: We might have a sequence consisting of the lighting of a fuse,
and an explosionbut whether the one is a cause of the other is not determined by
the intrinsic character of this two-event process, since it obviously matters whether
this fuse was connected to that exploding bomb.
I will simply give what I think is the right statement of the Intrinsicness thesis, one
that eschews undened talk of processes.10 Suppose an event e occurs at some
time t
0
. Then consider the structure of events that consists of e, together with all of
its causes back to some arbitrary earlier time t. That structure has a certain intrinsic
character, determined by the way the constituent events happen, together with their
spatiotemporal relations to one another. It also has a certain causal character: In
particular, each of the constituent events is a cause of e (except e itself, of course).
Then the Intrinsicness thesis states that any possible structure of events that exists in
a world with the same laws, and that has the same intrinsic character as our given
structure, also duplicates this aspect of its causal characterthat is, each duplicate of
one of es causes is itself a cause of the e-duplicate.11
Three observations: First, same intrinsic character can be read in a very strict
sense, according to which the two structures of events must be perfect duplicates.
Read this way, I think the Intrinsicness thesis is close to incontrovertible. But it can
also be read in a less strict sense, according to which the two structures must be, in
some sense, suciently similar in their intrinsic characters. Read this way, the thesis
is stronger but still highly plausible. Consider again the case of Billy and Suzy, and
compare the situation in which Billy throws his rock with the situation in which he
doesnt. Clearly, there is a strong intuition that the causal features of the sequence
of events beginning with Suzys throw and ending in the shattering should be the
same in each case, precisely because Billys throw is extrinsic to this sequence. But
it is too much to hope for that the corresponding sequences, in each situation, be
perfect duplicates; after all, the gravitational eects of Billys rock, in the situation
where he throws, will make minute dierences to the exact trajectory of Suzys rock,
and so on. So if it is the Intrinsicness thesis that gives voice to our conviction that,
from the standpoint of Suzys throw, the two situations must be treated alike, then
we should read the same intrinsic character clause in that thesis in the less strin-
gent way.
Doing so quite obviously leaves us with the burden of explaining what near-but-
not-quite-perfect duplication of intrinsic character consists in. I wont try to unload
that burden here. It will emerge that for my main purposes, that doesnt matter, since
in order to use the Intrinsicness thesis to argue that dependence and production are
two distinct kinds of causation, I can read same intrinsic character in the more
stringent sense. (Alas, we will also see that my own preferred analysis of production
will require the less stringent reading. For extensive discussion of these and other
issues involving Intrinsicness, see my 2002b.)
The second observation to make about the Intrinsicness thesis is that it is some-
what limited in scope: it does not apply, in general, to situations in which there is
causation at a temporal distance, or to situations in which there is backward causa-
tion. Roughly, the problem is that the relevant structure of events must be complete
in a certain respect, consisting in a complete set of joint causes of the given eect e,
together with all those events that mediate between these causes and e. I wont go
into the reasons why it must exhibit this kind of completeness (but see my 2002b).
But consider a case where the eect takes place at one oclock, and we have collected
together all of its causes that occur at noon, as well as those that occur between noon
and one. If there is action at a temporal distance, then some of the other causes with
which the noon causes combine to bring about the eect might have occurred before
noon, in which case our structure wont be suciently complete. If there is backward
causation, then some of the events that mediate between the noon causes and the
eect might occur outside the given interval, in which case our structure wont be
suciently complete. Either way, there is trouble. It is partly in order to nesse this
trouble that I have limited my focus by ignoring both backward causation and cau-
sation at a temporal distance.
The third observation to make about the Intrinsicness thesis is that we must
assumeon pain of rendering the thesis trivially falsethat the structure of events
against which we compare a given structure includes no omissions. Let the structure
S consist of e, together with all of its causes back to some arbitrary earlier time t. And
let the structure S
0
simply consist of S, together with some arbitrary omission that
occurs at some point in the relevant interval. Plausibly, this omission will contrib-
ute nothing to the intrinsic character of S
0
for it simply consists in the failure of
some type of genuine event to occur. So S
0
will perfectly match S. If we apply the
Intrinsicness thesis uncritically, we immediately get the absurd result that the added
omissionwhatever it iscounts as a cause of e. Now, it was already fairly clear
240 N. Hall
that whatever the guiding intuition is behind the Intrinsicness thesis, it does not con-
cern omissions. This result conrms the suspicion. So the nal clause of the Intrin-
sicness thesis should read: . . . any possible structure of genuine events (not including
any omissions) that exists in a world with the same laws, and that has the same
intrinsic character as our given structure, also duplicates. . . . (It doesnt follow that
Sthe structure picked out as consisting of e, together with all of its causes back
to some earlier time tmust include no omissions. Well take up the question of
whether it can in sec. 5, below.)
Perhaps the counterfactual analyst can use the Intrinsicness thesis to handle the
problem of Billy and Suzy. After all, in the alternative circumstances in which Billys
throw is absent, it seems correct to say that the causal history of the shattering (back
to the time of Suzys throw) consists exactly of those events on which it depends.
Whats more, this structure matches a structure that takes place in the actual circum-
stances, where Billys throw confounds the counterfactual relations; Suzys throw,
being a part of this structure, will therefore count as a cause of the shattering, thanks
to the Intrinsicness thesis. To be sure, this is no more than a suggestion of a revised
analysis. But again, what is important is that it is a plausible and natural suggestion
to pursueand it gives a central role to the Intrinsicness thesis.
4 Double Prevention
And now for something completely dierent: a kind of example that spells trou-
ble for the suciency of the simple analysis, by showing that the cornerstone thesis
of Dependence runs headlong into conict with each of Transitivity, Locality, and
Intrinsicness.
4.1 Example
Suzy and Billy have grown up, just in time to get involved in World War III. Suzy is
piloting a bomber on a mission to blow up an enemy target, and Billy is piloting
a ghter as her lone escort. Along comes an enemy ghter plane, piloted by
Enemy. Sharp-eyed Billy spots Enemy, zooms in, pulls the trigger, and Enemys
plane goes down in ames. Suzys mission is undisturbed, and the bombing takes
place as planned. If Billy hadnt pulled the trigger, Enemy would have eluded him
and shot down Suzy, and the bombing would not have happened.
This is a case of what I call double prevention: one event (Billys pulling the
trigger) prevents another (Enemys shooting down Suzy), which had it occurred
would have prevented yet another (the bombing). The salient causal structure is
depicted in gure 9.4. Neurons a; b, and c all re simultaneously. The ring of c
prevents e from ring; if e had red, it would have caused f to re, which in turn
would have prevented g from ring. Thus, if c had not red, g would not have. So c
is a cause of g: Billys pulling the trigger is a cause of the bombing.
This consequence of the counterfactual analysis might seem natural enough. After
all, wouldnt we give Billy part of the credit for the success of the mission? Isnt
Billys action part of the explanation for that success? And so on. On the other hand,
it might seem quite unnaturalfor the scue between Billy and Enemy takes place,
let us suppose, hundreds of miles away from Suzy, in such a way that not only is she
completely oblivious to it, but it has absolutely no eect on her whatsoever. Here she
is, in one region, ying her plane on the way to her bombing mission. Here Billy and
Enemy are, in an entirely separate region, acting out their fateful drama. Intuitively,
it seems entirely unexceptionable to claim that the events in the second region have
no causal connection to the events in the rstfor isnt it plain that no physical
connection unites them?
So far, it might seem that we have a stalemate: two contrary intuitions about the
case, with no way to decide between them. (Indeed, my informal polling suggests that
intuitive judgments vary quite a lot.) Not so: Both the judgment that we have a case
of causation here, and the thesis of Dependence that endorses this judgment, run into
trouble with each of the theses of Locality, Intrinsicness, and Transitivity.
4.2 Problems with Locality
We all know what action at a distance is: We have a case of it if we have a cause, at
least one of whose eects is not connected to it via any spatiotemporally continuous
causal chain.12 I take it that action at a distance is possible, but that its manifestation
in a world is nevertheless a highly nontrivial fact about that world. Yet if Billys
action counts as a cause of the bombing, then the quite ordinary and mundane rela-
a g
b e f
c
c fires
a g
b e f
c
if c hadnt fired
Figure 9.4
242 N. Hall
tionship it bears to the bombing also counts as a case of action at a distance. Worse:
It counts as a case of action at a temporal distancesomething that one might rea-
sonably argue is not possible, and at any rate something for which one will search the
history of physics in vain for precedent. Is this all it takes to achieve such non-
locality? (And to think that philosophers have been fussing over Bells inequalities!)
If so, we would be hard pressed to describe laws that didnt permit action at a (tem-
poral) distance. For example, even the classical laws that describe perfectly elastic
collisions would have to be judged nonlocal, since they permit situations in which
one collision prevents a second, which, had it happened, would have prevented a
thirdso that we have dependence of the third collision on the rst, but no con-
necting sequence of causal intermediates. In short, it appears that while Dependence
doesnt quite contradict Locality, it renders it satisable only by the most trivial laws
(e.g., laws that say that nothing ever changes). Thats wrong: The distinction between
laws that do and laws that dont permit action at a distance is interesting; to assimi-
late it to the all-but-vacuous distinction between laws that do and laws that dont
permit double prevention is a mistake.
A remarkably frequent but entirely unsatisfactory response is the following: Billys
action is connected to the bombing via a spatiotemporally continuous causal chain
its just that this chain consists, in part, of omissions (namely, the various failures of
Enemy to do what he would have done, had Billy not red). Now, its not just that
such reliance on causation by omission is desperate on its face. Its that even if we
grant that these omissions exist and are located where the events omitted would have
occurred (a nontrivial supposition: right now I am at home, and hence fail to be in
my oce; is this omission located there or here?), it doesnt help. For there is no
reason to believe that the region of spacetime these omissions occupy intersects the
region of spacetime that Suzy and her bomber actually occupy; to hold otherwise
is just to mistake this region with the region she would have occupied, had Billy
not red. We can agree that had Billy not red, then the Enemy-region would
have intersected the Suzy-region; but if, say, Suzy would have swerved under those
circumstances, then its just false to suppose that this counterfactual Enemy-region
( the actual omission-of-Enemy-region) intersects the actual Suzy-region.
Of course, the debate can take various twists and turns from here: There are fur-
ther stratagems one might resort to in an eort to interpolate a sequence of omissions
between Billy and the bombing; alternatively, one might deny that causation without
a connecting sequence of causal intermediates really is sucient for action at a dis-
tance. It wont prot us to pursue these twists and turns (but see my 2000b); suce it
to say that the stratagems fail, and the prospects for a replacement for the sucient
condition seem hopeless.
4.3 Problems with Intrinsicness
Lets rst recall what the Intrinsicness thesis says, in its careful formulation: Suppose
an event e occurs at some time t
0
. Consider the structure of events S that consists of
e, together with all of its causes back to some arbitrary earlier time t. Then any pos-
sible structure of events that exists in a world with the same laws, and that has the
same intrinsic character as S, also has the same causal character, at least with respect
to the causal generation of e.
For the purposes of this section, we can read has the same intrinsic character
as as perfectly duplicateswe wont need to compare structures of events that
exhibit near-but-not-quite-perfect match of intrinsic character.
Now for some more detail. When Billy shot him down, Enemy was waiting for his
home basehundreds of miles awayto radio him instructions. At that moment,
Enemy had no particular intention of going after Suzy; he was just minding his own
business. Still, if Billy hadnt pulled the trigger, then Enemy would have eluded him,
and moments later would have received instructions to shoot down the nearest suit-
able target (Suzy, as it happens). He would then have done so. But Billy does shoot
him down, so he never receives the instructions. In fact, the home base doesnt even
bother to send them, since it has been monitoring Enemys transmissions and knows
that he has been shot down.
Focus on the causal history of the bombing, back to the time of Billys action.
There is, of course, the process consisting of Suzy ying her plane, and so on (and,
less conspicuously, the process consisting in the persistence of the target). If Depen-
dence is true, then the causal history must also include Billys action and its immedi-
ate eects: the bullets ying out of his gun, their impact with Enemys fuselage, the
subsequent explosion. (Perhaps we should also throw in some omissions: the failure
of Enemy to do what he would have done, had he somehow eluded Billy. It makes
no dierence, since their contribution to the intrinsic character of the resulting causal
history is nil.) Let this structure of events be S.
Two problems now emerge. In the rst place, the intrinsic character of S fails to
determine, together with the laws, that there are no other factors that would (i) stop
Enemy, if Billy somehow failed to; (ii) do so in a way that would reverse the intuitive
verdict (such as it is) that Billys action is a cause of the bombing. Suppose, for in-
stance, that we change the example by adding a bomb under Enemys seat, which
would have gone o seconds after the time at which Billy red. And suppose that
within this changed example, we can nd a duplicate of Sin which case the speci-
cation of the intrinsic character of S must leave out the presence of the bomb. That
shows (what was, perhaps, apparent already) that the dependence of the bombing on
Billys action is a fact extrinsic to S. If we decide that in this changed example, Billys
244 N. Hall
action is not a cause of the bombing (since, thanks to the bomb under Enemys seat,
he in fact poses no threat to Suzy), then we must either give up the Intrinsicness
thesis, or grant that the causal history of the bombing (back to the time of Billys
action) wasnt described completely by S. Neither option is attractive. Let us call this
the problem of the extrinsic absence of disabling factors (disabling in the sense that if
they were present, there would be no dependence of the bombing on Billys action).
Much more serious is the problem of the extrinsic presence of enabling factors
(enabling in the sense that if they were absent, there would be no dependence of the
bombing on Billys action). For consider a third case, exactly like the rst except
in the following critical respect: The home base has no intentions of sending Enemy
orders to shoot anyone down. In fact, if Billy hadnt pulled the trigger, then the
instructions from the home base would have been for Enemy to return immediately.
So Enemy poses no threat whatsoever to Suzy. Hence Billys action is not a cause of
the bombing. Yet the structure of events S is duplicated exactly in this scenario. So if
the Intrinsicness thesis is right, then that causal history S must not in fact have been
complete; we must have mistakenly excluded some events for which the third scenario
contains no duplicates. Presumably, these events will be the ones that constitute the
monitoring of Enemy by his home base, together with the intentions of his superiors
to order him to shoot down the nearest appropriate target.
But now we are forced to say that these events count as causes of the bombing.
That is ridiculous. It is not that they have no connection to the bombing, its just that
their connection is much more oblique: All we can say is that if they hadnt hap-
pened, then the bombing would not have depended on Billys action. And notice,
nally, that it is exactly the inclusion of Billys action as part of the causal history
S that is the culprit: Once we include it, we must also include (on pain of denying
Intrinsicness) all those events whose occurrence is required to secure the counter-
factual dependence of the bombing on this action.
To see this problem more vividly, compare the events depicted in gures 9.5 and
9.6. Here, f is a stubborn neuron, needing two stimulatory signals in order to re.
Neuron h, in gure 9.5, res shortly after the time at which neurons a; b, and c all re
(so I have abused the usual left-to-right conventions slightly). In the left-hand dia-
gram of gure 9.5, g depends on c, but in gure 9.6 it does not; indeed, it would be
quite ridiculous to claim, about the left-hand diagram of gure 9.6, that c was in any
sense a cause of g.
But now consider the causal history of g, in the left-hand diagram of gure 9.5,
and suppose thatin keeping with Dependencewe count c as part of this causal
history. Then it would seem that this causal history is duplicated exactly in the
left-hand diagram of gure 9.6in which case either Intrinsicness is false, or c in
gure 9.6 is, after all, a cause of g. The only way out of this dilemma is to deny
Dependenceor else to insist, against all good sense, that the causal history of g, in
gure 9.5, also includes the ring of h (which is not duplicated in gure 9.6). But of
course it does not: In gure 9.5, the ring of h is necessary, in order for g to depend
on c; but that does not make it one of gs causes.
4.4 Problems with Transitivity
A more striking problem appears when we focus on the transitivity of causation. I
begin by adding yet more detail to the example.
Early in the morning on the day of the bombing, Enemys alarm clock goes o. A
good thing, too: If it hadnt, he never would have woken up in time to go on his pa-
trolling mission. Indeed, if his alarm clock hadnt gone o, Enemy would have been
nowhere near the scene at which he was shot down. It follows that if Enemys alarm
c fires if c hadnt fired
a g
b e f
c h
a g
b e f
c h
Figure 9.5
a g
b e f
c h
a g
b e f
c h
Figure 9.6
246 N. Hall
clock hadnt gone o, then Billy would not have pulled the trigger. But it is also true
that if Billy hadnt pulled the trigger, then the bombing would never have taken
place. By transitivity, this ringing is one of the causes of the bombing.
Figure 9.7 helps to reinforce the absurdity of this conclusion. Neuron e can never
re. If c does not re, then e wont get stimulated by d, whereas if c does re, then the
stimulation from d will be blocked by the inhibitory signal from b. So e poses no
threat whatsoever to the ring of f . The little four-neuron network that culminates
in e is, from the standpoint of f , totally inert.
Clearly, cs ring cannot be a cause of f s ring. At most, we might character-
ize cs ring as something which threatens to prevent f s ring, by way of the c-d-e
connectionwith the threat blocked by the c-b-e connection. Yet if both Depen-
dence and Transitivity are correct, then cs ring is a cause of f s ring. For if c
hadnt red, then b would not have red. Likewise, if b had not red, then f would
not have red (recall here that backtracking is forbidden: We cannot say that if b
had not red, then it would have been that c didnt re, and so f would have red
all the same). Since f depends on b, and b depends on c, it follows from Depen-
dence and Transitivity that cs ring is a cause of f s ring. That consequence is
unacceptable.
Certain examples with this structure border on the comic. Billy spies Suzy about to
throw a rock at a window. He rushes to stop her, knowing that as usual hes going
to take the blame for her act of vandalism. Unfortunately for him, he trips over a
tree-root, and Suzy, quite oblivious to his presence, goes ahead and breaks the win-
dow. If he hadnt tripped, he would have stopped herso the breaking depends on
the tripping. But if he hadnt set out to stop her, he wouldnt have trippedso, by
a f
d e
b c
c fires
a f
d e
b c
if c hadnt fired
Figure 9.7
the combination of Transitivity and Dependence, he has helped cause the breaking
after all, merely by setting out to stop it! That conclusion is, of course, just silly.13
Conclusion: If the thesis of Dependence is true, then each of Locality, Intrinsicness,
and Transitivity is false. More precisely, if Dependence is true at a world, and the
events in that world exhibit a causal structure rich enough to provide even one case
of double prevention like each of the ones we have been examining, then each of
Locality, Intrinsicness, and Transitivity is false at that world. In the next section, well
see that an exactly parallel conclusion can be drawn with respect to the thesis that
omissions can be causes and eects.
5 Omissions
The thesis of Omissions brings in its wake a number of dicult questions of ontol-
ogy: Does it imply a commitment to a peculiar kind of event whose occurrence
conditions essentially involve the failure of some ordinary type of event to occur?
Does it make sense to speak of the failure of c to occur, where c is supposed
to refer to some ordinary event? (For perhaps such singular reference to nonactual
events is impossible; alternatively, perhaps it is possible, but the circumstances in
which we want to cite some omission as a cause or eect typically underdetermine
which ordinary event is omitted.) Do omissions have locations in space and time?
If so, what determines these locations? (Recall the remarks in sec. 4.2: Right now
I am at home and hence fail to be in my oce; is this omission located there or here?)
And so on. I am simply going to gloss over all of these issues and assume that a
counterfactual supposition of the form omission o does not occur is equivalent to
the supposition that some ordinary event of a given type C does occur (at, perhaps, a
specic place and time)where the type in question will be xed, somehow, by the
specication of o (or perhaps by context, or perhaps by both). At any rate, however
justied complaints about the ontological status of omissions might be, they are em-
phatically not what is at issue, as were about to see.
In what follows, Ill make the case that examples of causation by omission rou-
tinely violate each of Locality and Intrinsicness. The techniques I employ can be
adapted so straightforwardly to make the same points about prevention (i.e., causa-
tion of omission) that we can safely leave those cases aside. Displaying the conict
between Omissions and Transitivity will require a case in which we treat an omission
as an eect of one event and as a cause of another.
Finally, I am also going to gloss over the remarkably tricky question of when, ex-
actly, we have a case of causation by or of omissiona question to which the thesis
of Omissions only gives the vague answer, sometimes. For example, is it enough to
248 N. Hall
have causation of e by the failure of an event of type C to occur for e to counter-
factually depend on this omission? Or must further constraints be satised? If notif
dependence is all that is requiredwe get such unwelcome results as that my act of
typing has among its causes a quite astonishing multitude of omissions: the failure of
a meteorite to strike our house moments ago, the failure of the President to walk in
and interrupt me, and so on. If, on the other hand, we insist that mere dependence
is not enough for causation by omission, then we face the unenviable task of trying
to characterize the further constraints. Im going to sidestep these issues by picking
cases that are uncontroversial examples of causation by omissionuncontroversial,
that is, on the assumption that there are any such cases.
5.1 Problems with Locality
We can draw on the story of Suzy, Billy, and Enemy to show that, even if we waive
worries about whether omissions have determinate locations, Locality fails for typi-
cal cases of causation by omission. Focus on a time t at which Enemy would have
been approaching Suzy to shoot her down, had he not been shot down himself. Had
Enemy not been absent, Suzys mission would have failed; so the bombing depends
on, as we might put it, the omission of Enemys attack. More than this: The omis-
sion of Enemys attack is among the causes of the bombingat least, if there is to
be causation by omission at all, this case should certainly be an example. But once
again, it appears that the connection between this omission and the bombing must
also qualify as a case of action at a distance, for no spatiotemporally continuous se-
quence of causal intermediates connects the two events. As before, the problem is not
with nding a suitable location for the omission; it is rather that nothing guarantees
that the sequence of omissions that proceeds from it (Enemys failure to approach,
pull the trigger, etc.) will intersect Suzys actual ight. We can grant that the region
of spacetime in which these omissions take place intersects the region she would
have occupied, had Enemy not been absent. But to suppose that this region is the
same as the region she actually occupies is to commit the same mistake as before.
5.2 Problems with Intrinsicness
Whatever omissions are, they are notably lacking in intrinsic character. We already
saw that for this reason, the Intrinsicness thesis needed to be phrased rather carefully:
When we have picked out an event e and a structure of events S comprising e and all
causes of e back to some earlier time, it is to be understood that any structure against
which we compare S is composed solely of genuine events, not omissions. (On the
other hand, no harm comes of letting S include omissions, at least on the assumption
that they contribute nothing to its intrinsic character.) Still, it is for all that consistent
to hold that Intrinsicness applies to causation by omission, as follows: Suppose that
e occurs at time t
0
, and that S consists of e and all causes of e back to some earlier
time t. Suppose further that we count the omission o as one of es causes, and that o
occurs (in whatever sense is appropriate for omissions) in the interval between t
and t
0
. Then if structure S
0
intrinsically matches S, there must be some omission o
0
corresponding to o that causes the event e
0
in S
0
that corresponds to e in S (never
mind that o
0
is not part of S
0
). In short, we might think that causation of an event
by omission supervenes on the intrinsic character of that events positive causal
history.
This conjecture is false. To show why, Ill argue that both of the problems we saw
in section 4.3the problem of the extrinsic lack of disabling factors and the problem
of the extrinsic presence of enabling factorsrecur in this context. A simple neuron
diagram will serve to illustrate each. In gure 9.8, d is a dull neuron that needs two
stimulatory signals in order to re. Thus, d fails to re even though stimulated by c;
still, since c res, es ring depends on the failure of b to re (at, say, time t, which we
will take to be the time of as ring). Note nally that if b had red and f had as well
(at t), then e would have red all the same.
Let us suppose, in keeping with the Omissions thesis, that the failure of b to re at t
is among the causes of es ring. Let S consist of e, together with all of its (positive)
causes back to time t. Then if Intrinsicness applies to causation by omission in the
way we have suggested, any nomologically possible structure that duplicates S will
a
d
e
b
c
a
d
e
b
c
f f
if b had fired b doesnt fire
Figure 9.8
250 N. Hall
exhibit the same causal relationships: In particular, there will be an omission that
duplicates bs failure to re and that will be a cause of the event that duplicates
es ring. Shown in gure 9.9 is one such possible structure, embedded in slightly
dierent surroundings. And another is shown in gure 9.10, again in dierent
surroundings.
The problem is that in each case, bs failure to re is no longer a cause of es ring,
contra the requirements of our conjecture about how Intrinsicness covers causation
by omission. In gure 9.9, the ring of f renders bs failure to re quite irrelevant to
whether e res, showing that when bs failure to re is a cause of es ring, this is
owing in part to the extrinsic absence of disabling factors. Likewise, in gure 9.10, cs
failure to re renders the behavior of b irrelevant, showing that when bs failure to
a
d
e
b
c
f
Figure 9.9
a
d
e
b
c
f
Figure 9.10
re is a cause of es ring, this is owing in part to the extrinsic presence of enabling
factors. So the leading idea behind the Intrinsicness thesisthat it is the intrinsic
character of some events causal history that (together with the laws) makes it the
case that this is its causal historycomes directly into conict with the Omissions
thesis.
5.3 Problems with Transitivity
As before, more striking problems emerge when we combine the theses of Omissions
and Transitivity. To see how easy it is to concoct an absurdity from these two ingre-
dients, consider the following variant on our story: This time, Enemys superiors on
the ground had no intention of going after Suzyuntil, that is, Billy shoots Enemy
down. Outraged by this unprovoked act of aggression, they send out an all-points-
bulletin, instructing any available ghter to go after Suzy (a much more valuable
target than Billy). Alas, Enemy was the only ghter in the area. Had he somehow
been present at the time of the broadcast, he would have received it and promptly
targeted and shot down Suzy; his absence is thereby a cause of the bombing. But, of
course, his absence is itself caused by Billys action. So by Transitivity, we get the
result that Billys action is a cause of the bombing. Lest the details of the case be
distracting, lets be clear: All Billy does is to provoke a threat to the bombing; luckily
for him, the very action that provokes the threat also manages to counteract it. Note
the similarity to our earlier counterexample to Transitivity: Enemys action (taking
o in the morning) both causes a threat to the bombing (by putting Enemy within
striking range of Suzy) and counteracts that threat (by likewise putting Enemy within
Billys striking range).
Conclusion: If the thesis of Omissions is true, then each of Locality, Intrinsicness,
and Transitivity is false. More precisely, if Omissions is true at a world, and the
events in that world exhibit a causal structure rich enough to provide cases of the
kinds we have just considered, then each of Locality, Intrinsicness, and Transitivity is
false at that world.
6 Diagnosis: Two Concepts of Causation
Here are two opposed reactions one might have to the discussion so far:
Counterfactual dependence is not causation. In the rst place, its not (as everyone
recognizes) necessary for causation. In the second place, the best attempts to tart it
up in such a way as to yield a full-blown analysis of causation rely on the three theses
of Locality, Intrinsicness, and Transitivityand the lesson of double prevention (a
lesson also supported by considering the causal status of omissions) is that these
252 N. Hall
theses contradict the claim that dependence is sucient for causation. The theses are
too important; this latter claim must be given up. But give up Dependence, and
youve torn the heart out of counterfactual analyses of causation.
Nonsense; counterfactual dependence is too causation. Here we have two wholly
distinct events; moreover, if the rst had not happened, then the second would not
have happened. So we can saynotice how smoothly the words glide o the tongue
that it is in part because the rst happened that the second happened, that the rst
event is partly responsible for the second event, that the occurrence of the rst event
helps to explain why the second event happened, and so on. Nor do we reverse these
verdicts when we discover that the dependence arises by way of double prevention;
that seems quite irrelevant. All of these locutions are causal locutions, and their ap-
propriateness can, quite clearly, be justied by the claim that the second event de-
pends counterfactually on the rst event. So how could this relation fail to be causal?
To be sure, its another question whether we can use it to construct a full-blown
analysis of causation, but at the very least we have the result that counterfactual de-
pendence (between wholly distinct events) is sucient for causationwhich is just to
say that Dependence is true.
The claims of both of the foregoing paragraphs are correct, but not by making a
contradiction true: Rather, what is meant by causation in each case is dierent.
Counterfactual dependence is causation in one sense: But in that sense of cause,
Transitivity, Locality, and Intrinsicness are all false. Still, they are not false simpliciter;
for there is a dierent concept of causationthe one I call productionthat ren-
ders them true. Thus, what we have in the standard cases of overdetermination we
reviewed in section 3 are not merely counterexamples to some hopeless attempt at
an analysis of causation, but cases that reveal one way the concepts of dependence
and production can come apart: These cases uniformly exhibit production without
dependence. What we have in the cases of double prevention and causation by omis-
sion we examined in sections 4 and 5 are not merely more nails in the con of the
counterfactual analysis, but cases that reveal the other way the two causal concepts
can come apart: These cases uniformly exhibit dependence without production. Sim-
ilarly, we can now diagnose the intuitions Bennett is pumping in his April rains/June
forest re case. For while there is a sense in which the rains do cause the rethe re
clearly depends on the rainsthere is an equally good sense in which they dont
the rains do not help to produce the re. That is because (surprise!) we have here a
case of double prevention: The rains prevent an event (re in May) that, had it oc-
curred, would have prevented the June res (by destroying the ammable material).
The principal virtues of my claim are thus clear: It allows us to maintain each of
the ve theses. It provides us with a natural and compelling diagnosis of the most
important problem cases for analyses of causation. And it should come as no sur-
prise that the distinction between production and dependence has gone unnoticed,
for typically the two relations coincide (more exactly, I think, production typically
coincides with the ancestral of dependence; more on this in sec. 7.4, below).
An additional virtue of the position, perhaps less obvious than the foregoing
ones, concerns the ontological status of omissions. Those who endorse the Omissions
thesis might worry that they are thereby committed to the existence of a special sort
of eventas if the truth of the failure of an event of type C to occur caused e to
occur required the existence of something that answered to the description, failure
of an event of type C to occur. But if the only sense in which omissions can cause
and be caused is that they can enter into relations of counterfactual dependence, then
this worry is quite misplaced. For talk of causation by and of omissions turns out
to be nothing more than a way of talking about claims of the form, if an event of
type C had occurred, then . . . and if . . . , then an event of type C would have oc-
curred. Manifestly, neither locution carries an ontological commitment to a strange
sort of negative event. So, if I am right, anxieties about whether we can nd a
place for omissions in the causal order rest on a basic confusion about what it means
to attribute causal status to omissions.
This observation connects to a broader point, which is that dependence, under-
stood as a relation between events, is unduly restrictive. Quite generally there can
be counterfactual dependence between facts (true propositions), where these can be
positive, negative, disjunctive, or whateverand where only rarely can we
shoehorn the facts so related into the form, such-and-such an event occurred.
When we canwhen we can say that the fact that e occurred depends on the fact
that c occurredthen we can go ahead and call this a kind of event-causation. But to
see it as anything but a special case of a causal relation with a much broader domain
would be, I think, a mistake.14
We can bring my thesis into still sharper focus by considering some of the more
obvious objections to it. It seems wise to begin by directly confronting what many
will see as the most damning objectionwhich is simply that it posits two concepts
of event-causation. This might strike some as an extravagantly high price to pay:
After all, when possible we should be conservative, and conservatism argues for
taking our concept of event-causation to be univocal. At the very least, shouldnt
we view the bifurcation of our concept of event-causation as a serious cost of my
proposal?
No, we should notand not because we shouldnt be conservative. Its rather that
this objection mistakes a perfectly sensible methodological maxim with a reason to
254 N. Hall
believe. The methodological maxim goes: When trying to come up with an analysis
of a concept, start out by operating under the assumption that the concept is uni-
vocal. I think thats sound. But it doesnt at all follow that it is somehow antecedently
more probable that the concept in question is univocallet alone so probable that any
analysis that says otherwise pays a high price. In the face of the right sorts of rea-
sons to prefer a nonunivocal analysis, we should give up our operative assumption
and we shouldnt expect those reasons to have to carry an extra-heavy burden of
proof because of the intrinsic plausibility of the hypothesis of univocality.
To think otherwise manifests a basic confusion. Its rather as if I had lost my keys
somewhere in this room; I have no idea where. They might be over there, where its
dark and a lot of debris obscures things; or they might be over here, where its sunny
and uncluttered. It makes exceedingly good sense for me to start by looking in the
sunny and uncluttered part of the roomto act as if I believed my keys were there.
But that is not because I do believe they are there, or even because I consider it more
likely than the alternative (as if the hypothesis that life is easy has some intrinsic
plausibility to it!). Its rather that if my keys are in the uncluttered area, then I will
soon nd themand if they are not, I will quickly nd that out as well.
In the same way, when we go to analyze some concept of philosophical interest, it
makes exceedingly good sense to start by looking for a univocal analysis. For even if
we are wrong, and some hidden ambiguity lurks in our ordinary applications of the
concept, the very problems we will encounter in trying to come up with a univocal
analysis will (if we are careful and attentive) be diagnostic of this ambiguity. (The
critique of the counterfactual analysis carried out in secs. 35 was partly designed to
be a case in point.) But it is foolishness to mistake this advice for a reason to believe
that the concept is univocal. Indeed, if I consider the hypothesis that our concept of
event-causation is univocal, I see no reason whatsoever to judge it to be highly proba-
ble, antecedently to any investigation. And after sucient investigationin particu-
lar, after basic principles governing our application of cause have been shown to
come into conictI think its plausibility is just about nil.
A more subtle objection is the following: What I have really shown is not that
there are two concepts of causation, but rather that there are two kinds of causation,
two dierent ways in which one event can be a cause of another. That may well be
right; certainly, I was happy to begin this paper by announcing that event-causation
comes in two varieties. I do not know how to judge the matter, because I am not
suciently clear on what underlies this distinction between concepts and kinds. Com-
pare a nice example borrowed from Tim Maudlin: There are at least three dierent
ways of being a mother. We might call them DNA-mother, womb-mother, and
nurturing mother. Does that mean we have three dierent concepts of motheran
ambiguity largely unnoticed only because those we call mothers are typically all
three? I dont know. At any rate, in the case at hand it doesnt matter in the slightest.
I am quite content to agree that I have (merely) shown that there are two kinds of
causationas long as those who insist on this rendering of my thesis agree that the
two kinds answer to very dierent criteria and consequently require very dierent
analyses. That claim alone is enough to show how unwise it would be, when attempt-
ing to provide a philosophical account of event-causation, merely to forge blindly
ahead, trying to come up with an analysis that can successfully run the gauntlet of
known problem cases. If I am right, any such single analysis is doomed to failure.
A third, more congenial objection begins by granting the distinction between pro-
duction and dependence, but denying that dependence deserves to be counted a kind
of causation at all. Now, I think there is something right about this objection, in that
production does seem, in some sense, to be the more central causal notion. As
evidence, consider that when presented with a paradigm case of production without
dependenceas in, say, the story of Suzy, Billy, and the broken bottlewe unhesi-
tatingly classify the producer as a cause; whereas when presented with a paradigm
case of dependence without productionas in, say, the story of Suzy, Billy, and
Enemyour intuitions (well, those of some of us, anyway) about whether a genuine
causal relation is manifested are shakier. Fair enough. But I think it goes too far to
deny that counterfactual dependence between wholly distinct events is not a kind of
causal relation. Partly this is because dependence plays the appropriate sort of roles
in, for example, explanation and decision. (See sec. 8, below, for more discussion of
this point.) And partly it is because I do not see how to accommodate causation
of and by omissions (as we should) as a species of production; counterfactual de-
pendence seems the only appropriate causal relation for such negative events to
stand in.
This last point brings up a fourth possible objection, which is that in claiming that
there are two kinds of causation, each characterized by a dierent subset of the ve
theses, I have overstepped my bounds. After all, even if the arguments of sections 4
and 5 succeed, all they establish is, roughly, (i) that Dependence contradicts each of
Locality, Intrinsicness, and Transitivity; and (ii) that Omissions likewise contradicts
each of Locality, Intrinsicness, and Transitivity. It obviously doesnt follow that De-
pendence and Omissions should be bundled together and taken to characterize one
kind of causation, nor that Locality, Intrinsicness, and Transitivity should be bundled
together and taken to characterize another. Perhaps the ambiguity in our ordinary
causal talk is more multifarious and messy than this claim allows.
256 N. Hall
Dead right. And even though I think that further investigation could unearth more
positive reasons for dividing the ve theses into the two groups I have chosen, I do
not have such reasons to oer here. For what its worth, I do have a strong hunch
that, as noted above, there couldnt be anything more to causation of and by omis-
sions than counterfactual dependence; hence the pairing of Omissions with Depen-
dence. And in the next section Ill propose an analysis of production that gives
central roles to both Intrinsicness and Transitivity, as well as to a slightly weak-
ened version of Locality. But thats hardly enough to warrant conviction. Rather,
whats wanted are more probing arguments as to why our ordinary notion of event-
causation should fracture cleanly along the lines I have drawn. Lacking such argu-
ments, I will fall back on the methodological maxim discussed above: Given that we
can no longer take it as a working hypothesis that the concept of causation is uni-
vocal, let us nevertheless adopt the most conservative working hypothesis available
to us. Since we have yet to nd any reason to think that Dependence conicts with
Omissions, or that conceptual tensions threaten the happy union of Locality, Intrin-
sicness, and Transitivity, let us assumeagain, as a working hypothesisthat the
rst two theses characterize one causal notion, the last three another.
And let us now consider how the two causal notions are to be analyzed.
7 The Two Concepts Analyzed
Part of the taskanalyzing dependenceis easy: It is simply counterfactual depen-
dence between distinct events. More cautiously, we might want to admit another
kind of counterfactual dependence as well. Perhaps counterfactual covariation
manifested when the time and manner of one events occurrence systematically coun-
terfactually depend on the time and manner of anothersshould count as a kind
of causation as well, to be classied as a close relative of dependence. (Its clearly
possible to have dependence without covariation, as in typical cases of double pre-
vention; Schaer [Trumping Preemption, chapter 2 in this volume] provides com-
pelling examples of covariation without dependence, as well.) No matter; given that
these counterfactual locutions are themselves well understood, our work here is
basically already done for us. (But see sec. 8.2 for some tentative reservations.)
Production is harder. In this section I will put forth my own proposalspeculative,
and, as well see, somewhat limited in scopefor a reductive analysis of this relation.
I will set it out in two parts. The rst, less speculative part outlines a certain strategy
for developing an analysis, which I call the blueprint strategy. The second, more
speculative part describes my (currently) preferred way of implementing this strategy.
7.1 The Blueprint Strategy
Suppose we have an analysis that succeedswhen circumstances are nicein sin-
gling out a portion of the causal history of some target event e, where this is under-
stood to be the history of es producers. (When circumstances are not nice, let the
analysis fall silent.) It might be a simple counterfactual analysis: When circumstances
are nice (when there is no double prevention, overdetermination, etc.), the causal
history of e back to some earlier time t consists of all those events occurring in that
interval on which e depends. Or it might be a Mackie-style analysis: The causal his-
tory consists of all those events (again, occurring in that interval) that are necessary
parts of some sucient condition for e. Or it might be some other kind of analysis.
Thenprovided we can say with enough precision what it takes for circumstances to
be nicewe can use the Intrinsicness and Transitivity theses to extend the reach of
this analysis, as follows:
First, suppose we examine some events c and e, and nd that our analysis is silent
as to whether c is a cause of e. Still, we nd that c and e belong to a structure of
events S such that (i) S intrinsically matches some other structure of events S
0
(occur-
ring in a world with the same laws as the world of S); and (ii) our analysis counts S
0
as a segment of the causal history of e
0
(where e
0
is the event in S
0
that corresponds
to e in S). That is, our analysis counts S
0
as a rich enough set of causes of e
0
for the
Intrinsicness thesis to apply. It follows that S has the same causal structure as S
0
(at
least, with respect to the target event e), and hence that c is a cause of e.
For convenience, let us say that when the conjunction of our analysis with the
Intrinsicness thesis counts c as a cause of e, c is a proximate cause of e. Then,
second, we parlay proximate causation into causation simpliciter by means of the
Transitivity thesis: Causation is simply the ancestral of proximate causation. In short,
we use our original analysis to nd a set of blueprints for causal structures, which we
can then use to map out (if we are lucky) the causal structure of any set of events, in
any circumstances, by means of the Intrinsicness and Transitivity theses.
This strategy has the virtue of factoring the analysis of production into two parts:
the analysis that produces the blueprints, and the extension of any such anal-
ysis into a full analysis of production by means of the Intrinsicness and Transitivity
theses. Still, two potential diculties deserve mention. First, recall that the Intrin-
sicness thesis as Ive stated it presupposes that there is neither action at a temporal
distance nor backward causationso without a more general statement of the Intrin-
sicness thesis, the full analysis of production will necessarily be limited in scope. Sec-
ond, recall that for the purposes to which I put the Intrinsicness thesis above (to reveal
conicts with Dependence and Omissions), the same intrinsic character clause in
that thesis could be understood in a relatively clear and uncontroversial sense, namely
258 N. Hall
as requiring that the two structures of events in question be perfect duplicates. That
is, to make trouble for Dependence and Omissions we only needed to assume,
roughly, that two event-structures that perfectly match one another in intrinsic
respects likewise match in causal respects. But the blueprint strategy aords us no such
luxury.
To see why, consider our old standby example of Billy, Suzy, and the broken bot-
tle. Suppose that our unadorned analysis (whatever it turns out to be) falls silent
about whether Suzys throw is a cause of the breakingand this, thanks to the con-
founding presence of Billys throw. And suppose that the counterfactual situation
in which Billys throw is absent is one whose causal structure our analysis succeeds
in capturingin particular, the counterpart to Suzys throw, in that situation, is
counted a cause of the counterpart to the breaking. Victory!for surely we can say
that when Billys throw is present, Suzys still counts as a cause, because it belongs
to a structure of events (the throw, the ight of the rock, etc.) that matches an ap-
propriate blueprint structurenamely, the structure found in the counterfactual
situation where Billys throw is absent. Dont we have here a vindication of the
blueprint strategy?
Yes, but only if the notion of matching is more liberal and, regrettably, vague
than the restrictive, relatively precise notion of perfect match. For the two sequences
of eventsthe one beginning with Suzys throw, in the case where Billy also throws,
and the one beginning with her throw, in the case where he doesntwill not match
perfectly: For example, tiny gravitational eects from Billys rock will guarantee that
the trajectories of Suzys rock, in each case, are not quite the same. So we are left
with the unnished business of saying what imperfect match consists in, and of spe-
cifying how imperfect it can be, consistent with the requirements of the blueprint
strategy. While I do not think these diculties undermine the blueprint strategy, I
wont try to resolve them here. (But see my 2002b for a detailed proposal.)
7.2 Implementing the Strategy (First Pass)
Now for my own story about what makes for nice circumstances, and how an
analysis should proceed under the assumption that they obtain. As usual, I will as-
sume determinism, but I will also assume that there is no action at a temporal
distance, nor backward causation (not merely because I wish to slot the following
analysis into the blueprint strategy, but also because I do not yet know how to make
the analysis itself work, without these assumptions). First, some terminology.
Suppose that at time t, the members of some set S of events all occur, and that e
occurs at some later time t
0
. I will say that S is sucient for e just in case the fact that
e occurs follows from
(i) the laws, together with
(ii) the premise that all the members of S occur at t, together with
(iii) the premise that no other events occur at t.
The entailment here is metaphysical, not narrowly logical. I will say that S is
minimally sucient for e just in case S is sucient for e, but no proper subset of S is.
(We might want to add a premise to the eect that relevant background conditions
obtain. I prefer to treat any such conditions as encoded as members of S.) Do not
be distracted by the fact that in typical situations, (iii) will be false (though (i) and (ii)
will of course be true); that is quite irrelevant to the purposes to which we will put the
notions of suciency and minimal suciency.
Finally, the quantier in (iii) must be understood as ranging over only genuine
events, and not omissions, else this premise is inconsistent. Suppose, for example,
that our set S does not include a kiss at a certain location l; the no other events
occur requirement will therefore entail that no kiss occurs at t at l. Then consider
the omission o, which consists in the failure of a kiss to take place at t at l, and sup-
pose that o is also not a member of S. To require, in addition, that this event not
occur, is just to require that a kiss does take place at t at l. Quite obviously, we cant
add that requirement consistently.
It is a more or less obvious consequence of the foregoing restriction that omissions
cannot be producers. With modest assumptions about the laws, we can also prove
that no omission can be producedthat is, there is no omission o and event c such
that c helps to produce o. For in order for some omission to be produced, there must
be at least one example of an omission o and a set of events S such that S is mini-
mally sucient for o. Now for the assumption: The laws of evolution are such that
the unique state of the world in which nothing at all happensno event occurs
remains unchanged, evolving always into itself. If the laws are like this, then S
cannot possibly be minimally sucient for o, simply because a proper subset of S
namely, the empty setwill be sucient for o.
Roughly, we can say that where the members of S occur at t, S is sucient for
later event e just in case, had only the events in S occurred at t, e would still have
occurred; S is minimally sucient if the same is not true for any proper subset of S.
(Since I have employed a counterfactual locution here, one might want to call the
resulting analysis a counterfactual analysis. Call it what you willjust dont confuse
it with those analyses that take Dependence as their starting point.)
It seems that the problems that confound the usual attempts to analyze causation
all have to do with stu going on in the environment of the genuine causal process,
stu that ruins what would otherwise be the neat nomological relationships between
260 N. Hall
the constituents of that process and the given eect. An attractive and simple idea is
that if, at a time, there is a unique minimally sucient set for our target eect e, then
such environmental noise must be absentso that circumstances are niceand
we can take it that this unique minimally sucient set contains all and only the pro-
ducers of e that occur at that time. If so, then one way of implementing the blueprint
strategy becomes obvious: Suppose that e occurs at t
0
, and that t is an earlier time
such that at each time between t and t
0
, there is a unique minimally sucient set for
e; then the segment of es causal history back to time t consists of all and only the
events in these sets.
It will turn out that this simple idea wont work without a signicant adjustment.
But rst some good news.
One embarrassment for a Mackie-style analysis is the so-called problem of com-
mon eects (gure 9.11): d and e are both eects of c; hence from the fact that d
occurs (at t, say), together with the laws, together with an appropriate specication
of the circumstances, it follows that e occurs (at t
0
, say). (The reasoning is something
like the following: Given the circumstances, d could occur only if c caused it; but in
that case e must have occurred as well.)
Our provisional analysis has no such problem, since d cannot be part of a mini-
mally sucient set for e. To see why, let us rst simplify matters by folding a speci-
cation of the background conditions (i.e., the existence of and connections between
the various neurons) into the denition of suciency (nothing will hinge on this).
Then there are, at the relevant time t, only two events occurring: d and b. The set
fbg is of course sucient for e (hence minimally sucient, since it contains but one
member), since it follows from the claim that b alone occurs at time t, together with
the laws, together with the description of the neuron network, that e occurs. Put
more simply, had b alone occurred at time t, e would (still) have occurred. But the set
fdg is not sucient for e, since, clearly, in a circumstance in which d alone res, e
does not re. So fbg is the unique minimally sucient set for e. So b is a (producing)
cause of e.
c
d
b e
Figure 9.11
Objection: It is nomologically impossible for b to occur alonefor if it occurs,
then c must have red, and so d must re as well.
The best reply is to insist that once we get clear on what the relevant dynamical
laws are, we will see that they do allow as a possibility an instantaneous state of
the world in which the four neurons are connected as shown, but b alone is ring.
(The idea is that, in general, dynamical laws place only relatively weak constraints on
which instantaneous states are possible.) Since that reply requires too long a digres-
sion to develop here, let me fall back on a simpler one: If the objection has any force,
then it spells just as much trouble for the notion that there is a coherent nonback-
tracking reading of the counterfactual conditional. For consider the conditional if d
had not red, e would have red anyway. True, yes? But if true, it is true only on a
nonbacktracking reading. And that reading requires us to make sense of a possibility
in which (i) at the relevant time, the instantaneous state of the world is one in which
the four neurons are connected as shown, but b alone is ring; (ii) the actual dynam-
ical laws govern the evolution of that state from the given time onward. That is ex-
actly the possibility that my account requires.
So our analysis does not count d as a cause of e. Nor will it, when we add in the
Intrinsicness and Transitivity theses: For d cannot inherit causal status from any blue-
print; in order to do so the blueprint would have to contain a copy of d, a copy of e,
and other events contemporaneous with the d-duplicate and with which it formed a
minimally sucient set. These other events, moreover, would have to be duplicated
in the events of gure 9.11. In short, the duplicate would have to be the result of
subtracting events from gure 9.11 in such a way that d remained, and belonged to
a minimally sucient set for e. But there is manifestly no way to perform such a
subtraction. And it is equally clear that no sequence of blueprints can connect up d
with e.
Next, consider ordinary preemption (see g. 9.2). At the time of occurrence of a
and c, there are two minimally sucient sets for e: fcg and fag. (For ease of exposi-
tion, here and in the rest of this section Ill suppress mention of those events whose
occurrence consists in the presence of the relevant stimulatory and inhibitory con-
nections.) So circumstances are not nice. Still, there is no problem with getting c to
come out as a cause of e, for there is an obvious blueprint contained within the cir-
cumstances that would have occurred, if a had not red. But no such blueprint con-
necting a with e can be found in the circumstances that would have obtained, had
c not red; for in those circumstances, the causal history of e will include the ring
of b, and so there will be no match between this causal history and any part of the
actual structure of events.
262 N. Hall
Next, late preemption: Happily, such cases receive exactly the same diagnosis as
the case of ordinary preemption, and so need no special treatment.
Finally, double-prevention (see g. 9.4). Observe rst that there is, at the time of
occurrence of a; b, and c, a unique minimally sucient set for g, namely, fag. So if c
is to qualify as a cause of g, we must nd a blueprint, or sequence of blueprints, that
will connect up c with g. But that is evidently impossible. For such a blueprint would
have to describe a causal history for g that is dierent from the one that actually
obtains; otherwise, this causal history would contain a duplicate of a, in which case
the duplicate of c could not be part of a minimally sucient set for the duplicate of g.
But it is apparent that there is no sequence of events connecting c with g that could
serve as such an alternate causal history.
7.3 Implementing the Strategy (Final Pass)
So far, so good. Unfortunately, two diculties scotch the key idea that when there is
a unique minimally sucient set for e at a time, then its elements are all and only the
producers of e (at that time).
First, there are producers that belong to no minimally sucient set (gure 9.12). In
the diagram, a and c are clearly the producers of e; yet the unique minimally su-
cient set for e contains just c. (Remember that e, here, is a stubborn neuron, requiring
two stimulatory signals to re.)
Second, there are nonproducers that belong to unique minimally sucient sets
(gure 9.13). In the diagram, a is clearly not a producer of e; yet it is included in
fa; cg, the unique minimally sucient set for e. (Notice that this is a special case of
d
e
b c
a d
e
b c
a
a fires if a hadnt fired
Figure 9.12
double prevention, one that eludes the treatment we gave of the standard sort of case
exhibited by gure 9.4.)
I suggest these problems arise because we have overlooked an important constraint
on the internal structure of causal histories. Suppose that e occurs at time t
2
, and that
we have identied the set of its producers at both time t
0
(call this set S
0
) and time t
1
(call this set S
1
) t
0
< t
1
< t
2
. Then it had better be the case that when we trace the
causal histories of the elements of S
1
back to t
0
,
(i) we nd no events outside of S
0
for otherwise transitivity of production would
have been violated; and
(ii) we do nd all the events inside S
0
for otherwise we would have to say that one
of these events helped produce e, but not by way of any of the t
1
-intermediates.
Return to the diagrams. In gure 9.12, our analysis tells us (among other things)
that d is a producer of e and that a is a producer of dbut fails to deliver the con-
sequence that a is a producer of e. That is an example of a failure to meet constraint
(i). In gure 9.13, our analysis identies d as the sole intermediate producer of e, and
identies c as its only producerthus misdescribing a as a producer of e that some-
how fails to act by way of any intermediates. That is an example of a failure to meet
constraint (ii).
The way to x these problem is to make our nice circumstances analysis still
more restrictive, by building into it the two foregoing constraints. We begin as be-
fore, by supposing that e occurs at t
0
, and that t is an earlier time such that at each
time between t and t
0
, there is a unique minimally sucient set for e. But now we add
d e
b c
a
d e
b c
a
a fires if a hadnt fired
Figure 9.13
264 N. Hall
the requirement that whenever t
0
and t
1
are two such times t
0
< t
1
and S
0
and S
1
the corresponding minimally sucient sets, then
(i) for each element of S
1
, there is at t
0
a unique minimally sucient set for it; and
(ii) the union of these minimally sucient sets is S
0
.
This added requirement gives expression to the idea that when we, as it were, identify
the producers of e directly, by appeal to their nomological relationship to e, we must
get the same result as when we identify them by tracing back through intermediate
producers.
Were now in a position to state the analysis of production:
Given some event e occurring at time t
0
and given some earlier time t, we will
say that e has a pure causal history back to time t just in case there is, at every time
between t and t
0
, a unique minimally sucient set for e, and the collection of these
sets meets the two foregoing constraints. We will call the structure consisting of the
members of these sets the pure causal history of e, back to time t.
We will say that c is a proximate cause of e just in case c and e belong to some
structure of events S for which there is at least one nomologically possible structure
S
0
such that (i) S
0
intrinsically matches S; and (ii) S
0
consists of an e-duplicate, to-
gether with a pure causal history of this e-duplicate back to some earlier time. (In
easy cases, S will itself be the needed duplicate structure.)
Production, nally, is dened as the ancestral of proximate causation.
I do not mean to pretend for an instant that the analysis I have oered stands
in no need of detailed elaboration or defense. Of course it does. But that task can be
left for another occasion, since my aim here is quite modest: I mean only to make it
at least somewhat plausible that a reductive analysis of production can be had
thereby blocking the objection that once we distinguish production and dependence
and relegate counterfactual analyses to the role of explicating only the latter concept,
then we will be stuck with production as an unanalyzable causal primitive. That
would indeed be unfortunate. But I think there is little reason for such a pessimistic
assessment.
7.4 The Coincidence of Production and Dependence
Let me close this section by considering more closely why production and depen-
dence so often coincide. First, suppose that c, which occurs at t, is a producer of e. In
typical casesthat is, if the environment does not conspire in such a way as to ruin
the ordinarily neat nomological relationships between c and ec will belong to a
unique minimally sucient set for e. Let S be this set. Other events occur at t; let us
collect these together into the set T. Then consider what happens in a counterfactual
situation where c does not occur (keeping in mind that there may be more than one
such situation): (i) The other events in S occur. (ii) The events in T occur. (iii) Pos-
sibly, in place of c, some other event c
0
occurs.15
Let S
0
S fcg. Thenmodulo a small assumptionS
0
WT cannot be su-
cient for e. For suppose it is. Then given (the small assumption) that not every subset
of S
0
WT that is sucient for e contains a proper subset that is also sucient for e
(given, that is, that the sucient subsets of S
0
WT are not innitely nested), S
0
WT
will contain a minimally sucient set for e. But then S will not be the unique such
minimally sucient set, contra the hypothesis.
So it will not follow from the premise that all the events in S
0
WT occur, together
with the premise that no other events occur, together with the laws, that e occurs.
And that means that in a counterfactual situation in which c does not occur, and in
which no event takes its place, e does not occur. That is one way it could turn out that
e depends counterfactually on c.
More likely, though, some event c
0
will take the place of c. Furthermore, if this
event conspires in the right way with the other events in S
0
WT, then e will occur all
the same. Suppose, for example, that in the actual situation, Suzy throws a rock,
breaking a window. Billy is absent this time, so we can assume that her throw is part
of a unique minimally sucient set for the breaking. But suppose further that if she
hadnt thrown, her hand would (or simply might) have fallen by her side, brushing
against a switch, ipping it and thereby activating a catapult that would have hurled
a brick at the window, breaking it. Then that is a situation in which the counter-
factual alternative to c (or: one of the alternatives) conspires with other events to
bring about e. But notice that it takes some work to rig an example so that it has
this feature; typically, when c is part of a unique minimally sucient set for e, it will
be the case that if c hadnt occurred, then whatever event replaced it would not so
conspirewhich is to say that it will be the case that e depends on c.
Finally, even in the case of Suzy, the rock, and the stand-by catapult, we will have
stepwise dependence of the windows breaking on her throw: Picking an event d that
forms part of the rocks ight, we will have dependence of the breaking on d and of d
on the throw. We could tinker further to destroy one of these dependencies, but of
course that is not enough: We would need to tinker enough to block every two-step
chain of dependenciesand every three-step chain, and every four-step chain, and so
on. It is possible to do thiseven while guaranteeing the existence, at each stage, of
a unique minimally sucient set for the windows breakingbut only at the cost of
making the example even more atypical.16 And that shows that if I am right about
the correct analysis of the central kind of causation, then it is no great surprise that
266 N. Hall
the simple idea that causation should be understood as the ancestral of counter-
factual dependence worked as well as it did.
8 Applications, Open Questions, and Unnished Business
8.1 Applications
In the last three decades or so, causation seems to have become something of a philo-
sophical workhorse: philosophers have oered causal accounts of knowledge, per-
ception, mental content, action, explanation, persistence through time, and decision
making, to name a few. It wont be possible to discuss any of these topics in de-
tail, but I will focus briey on the latter three as a way of beginning to explore the
broader consequences of the distinction between dependence and production.
Before doing so, however, let us just observe that even the most cursory inspection
of the philosophical roles causation plays vindicates one of the three central argu-
ments for the distinction, which is that Transitivity and Dependence conict. Recall
one of the examples used to display the conict: Billy spies Suzy, and runs toward
her in an eort to stop her from throwing a rock at a window; en route he trips, and
as a consequence doesnt reach her in time to stop her. The window breaks. If Billy
hadnt tripped, the window would not have broken (because he would have stopped
Suzy). If Billy hadnt run toward Suzy, he wouldnt have tripped. Suppose we con-
clude, via a confused appeal to Dependence and Transitivity, that Billys running
toward Suzy was one of the causes of the windows breaking. Still, we will have
to admit that for a cause that is so proximate to its eect, it is quite strange: It
is not something we would cite as part of an explanation of the breaking, we would
not hold Billy at all responsible for the breaking on account of having helped to
cause it in this way, and so on. In short, consider any of the typical roles that
causation plays in other arenas, and you will nd that the sort of relation Billys
action bears to the breaking quite obviously plays none of those roles. That should
add to the conviction (if such addition were needed) that this relation is not one of
causation.
That is not to say that we cannot describe this relation in causal terms, since of
course we can: Billy does something that both (i) initiates a process that threatens to
interrupt the window-breaking process, and (ii) causes an event that interrupts this
potential interrupter. So the relation of Billys action to the window-breaking has a
perfectly denite causal structure. But that does not make it a kind of causation.
If we look more closely at some causally infused concepts, I think we can nd
more direct manifestation of the dierence between production and dependence, even
in the kind of brief and selective treatment I am about to oer.
Begin with persistence. On one well-known view, what it is for an object to persist
from time t
1
to time t
2
is for it to have temporal parts at t
1
; t
2
, and the intervening
times such that earlier ones are appropriately connected to later ones. What I want to
focus on is not the controversial ontology of temporal parts, but the nature of the
connectionwhich, on typical formulations, has got to be partly causal. The ques-
tion is: Could the causal relation involved in this connection be one of mere depen-
dence, without production?
Good test cases are not easy to come by, mainly because we already know that
for an enduring object of any complexity, the causal component of the connec-
tion between its earlier and later stages has to be understood as much more restric-
tive than either dependence or production, and the restrictions are not easy to spell
out.17
Lets strive for as simple a case as possiblesay, one involving the persistence
through time of an electron (assuming for the moment a naive conception of the
electron as a classical point-particle). Suppose we have two electron-stages, located at
t
1
and t
2
t
1
< t
2
. Plausibly, it is a necessary condition for the stages to be stages of
the same persisting electron that the rst be a cause of the second. But I think this
necessary condition cannot possibly be met if the second merely depends on the
rst. Suppose, for example, that the presence of the rst results in the prevention of
something that would itself have prevented the presence of the second. If we know
that that is the only causal connection that obtains between the two stages, then we
know enough to conclude that they cannot be stages of one and the same electron.
On the other hand, if we know that the rst stage helps to produce the second stage,
then while we may not yet know enough to conclude that they belong to the same
electron, it does not matter whether we learn that the second stage fails to depend on
the rst (as it might, because of some backup process that would have led to an
electron in the same place at the same time).
Thus, while the matter certainly merits further investigation, we can conclude that
with respect to persistence through time of a simple object like an electron, produc-
tion is the important causal notion (or at least: one of the basic ingredients in this
causal notion), whereas dependence is irrelevant. Persistence of more complex objects
seems unlikely to dier in this respect.
Consider next an arena in which the relative importance of the two kinds of cau-
sation is reversed: causal decision theory. When you face a range of options, and
causal decision theory says (very roughly) that the rationally preferable one is the one
most likely to have as a causal consequence the best (by your lights) outcome, the
notion of causal consequence at work is clearly that of dependence, and not pro-
268 N. Hall
duction. Or rather, it is a natural generalization of dependence, where we allow that
more than just events can be suitable relata: facts, say, or states of aairs.
Our standard stories of double prevention already illustrate the irrelevance of pro-
duction to decision making. There is, we can suppose, nothing whatsoever that Billy
can do to help produce the bombing, but that doesnt matter in the slightest: Whether
there is a bombing clearly depends on the action he takes, and it is his beliefs about
this dependence and its detailed structure that will guide his decisions, insofar as he is
rational.
Here is another kind of example that makes vivid the irrelevance of mere produc-
tion (once you see the trick you can generate endless variations): You want your
team to win; that is the only thing in the world that matters to you. You know that
if you do nothingjust stay where you are, sitting on the sidelinesyour fellow
teammates will, with certainty, achieve victory. On the other hand, if you insist on
playing, the team will probably lose (you are not, alas, very good). Still, if you play
and the team manages to win, then your own actions will have helped to produce
that win (you have your good days, and with luck this might be one of them). If what
matters in decision is what your dierent courses of action would be likely to pro-
duce, then you should playfor only then does your action have a chance of helping
to produce something desirable. That you should clearly not play helps show that the
productive upshot of your actions is not what matters. (Dont say: But if you sit on
the sidelines, then you help produce the victory by not playing. That is to confuse
the kind of causation that omissions can enter into with production, and we have
already seen ample reason why such confusion should be avoided.)
There is a needed qualication that by now might be obvious. For there is an ob-
vious way that production can matter, quite a lot, to decision: namely, if the out-
comes to which the agent attaches value are themselves partly characterized in terms
of what produces what. To modify our example, suppose that what matters to you
is not that your team win, per se, but that you help bring about your teams victory.
If so, then it will be perfectly rational (though selsh) for you to insist on playing.
Again, if something awful is going to happen no matter what you do, it may yet
matter quite a lot to you whether it happens partly as the productive upshot of your
behavior.
So production can matter to decision, after all. But seen as an objection to my
thesis this is just an equivocation, since that thesis concerns the kind of causal rela-
tion that connects action to outcome, and not the taxonomy of outcomes themselves.
Even when you choose to avoid a certain course of action because it would result
in your having helped produce the evil deed, the sense of because is clearly that of
dependence: What matters is how the possible outcomes (evil deed you helped pro-
duce vs. evil deed in which you had no hand) depend on your action.
Let us nally consider an arenacausal explanationin which both production
and dependence play a role; what well nd is that these roles are interestingly
dierent.
Recall, once more, the story of Billys failed attempt to stop Suzy from breaking
the window. In seeking an explanation of the windows shattering, we might, on the
one hand, ask what brought that event about, what led up to it. To questions of this
sort, it would be strange, to put it mildly, to cite Billys trip; rather, whats wanted is
information about the producers of the shattering. On the other hand, we can ask
why the shattering occurred, given that Billy started running toward Suzy. When we
ask a question of this formWhy did e occur, given that c occurred?we are
obviously presupposing that c set in motion some process or processes that would
have prevented e, had circumstances been dierentand so we want to know about
events whose occurrence kept circumstances from being dierent. (Sometimes we
make the presupposition quite explicit: Why did the window break, given that Billy
set out to stop Suzy?) With respect to our story, the obviously correct answer is that
Billy tripped (or that he wasnt watching where he was going, or that hes clumsy and
so prone to tripping, etc.). It wont do to cite an arbitrary producer or e, unless it also
happens to play the role of stopping cs occurrence from preventing es. Try it: Why
did the window shatter, given that Billy started running toward Suzy (with the in-
tention of stopping her)? Because Suzy threw a rock at the window. Highly mis-
leading, to say the least.
The catch-all explanation-request that philosophers often focus onWhy did
the window break?obscures the dierence between these two more rened ways
of requesting an explanation. Indeed, in the right contexts, the question Why did
the window break? might be appropriately answered either by Billy tripped or
by Suzy threw. That shows that both dependence and production have causal-
explanatory roles to play. But it doesnt showindeed it hides the factthat these
roles exhibit interesting and striking dierences. There is, after all, a world of dier-
ence between asking, of some event, what led up to it, and asking why it occurred,
given that something else was poised to prevent itnever mind that each question
could, in the right context, be conveyed by Why did it happen?
That concludes my necessarily brief discussion of the implications that the distinc-
tion between production and dependence might have for other areas of philosophy.
I hope it has been detailed enough to make further such inquiry seem worthwhile. I
will close now with a brief look at a few ways in which the picture of causation that
270 N. Hall
emerges from the foregoing treatment turns out to be more complicated than one
might have thought.
8.2 Unnished Business
First, there are certain kinds of cases that we have some inclination to call cases of
causation, but that also elude classication in terms of production or dependence.
Here is an example, a slight variation on the story of Billy, Suzy, and Enemy: This
time, there is a second ghter plane escorting Suzy. Billy shoots down Enemy exactly
as before, but if he hadnt, the second escort would have. Figure 9.14 captures the
salient causal structure.
It is no longer true that if Billy hadnt pulled the trigger, then the bombing
wouldnt have happened. In gure 9.14, it is no longer true that if c hadnt red, then
g wouldnt have. Nevertheless, given that Billys action is partly responsible for the
success of the bombing in the rst case, where the second escort was absent, then
surely there is some inclination to grant him such responsibility in this second case,
which merely adds an alternative that plays no active role. In the diagram, the su-
peruous preventive chain from d should not, it seems, change cs status as a kind of
cause of g.
Notice that our judgment that Billy is partly responsible for the success of the
bombing is quite sensitive to the nature of the backup preventer. For example, sup-
pose that Suzy is protected not by a second escort, but by a Shield of Invulnerability
that encloses her bomber, making it impervious to all attacks. We have the same
relations of counterfactual dependence: If Billy had not red, the bombing would still
have been a success, but if Billy had not red and the Shield had not been present,
a g
b e f
c
a g
b e f
c d d
Figure 9.14
the bombing would not have been a success, and so on. But only with great strain
can we get ourselves to say that, in this case, Billy is partly responsible for the success
of the bombing.
The issue here is really exactly the same as one that has received some discussion
in the literature, concerning the nature of prevention and how best to analyze it.
Thus, McDermott (1995a) asks us to consider a case in which a catcher catches a ball
ying toward a window; the window, however, is protected by a high, thick brick
wall, and so of course would not have broken even if the catcher had missed her
catch. Does the catcher prevent the window from being broken? It seems not. But if
we replace the wall with a second catcherone who would have caught the ball if
the rst catcher had missedthen judgments tend to be reversed. All I have done
with the case of Billy, Suzy, and the backup escort is to insert such a case of pre-
empted prevention into the story. In short, the tricky problem of how to understand
the exact nature of preempted prevention generates, as a kind of side eect, a prob-
lem for how to understand certain kinds of double prevention (namely, where the
rst preventer has a preempted backup). (For an insightful treatment of the problem
of preempted prevention, see Collins 2000, reprinted as chapter 4 in this volume.)
Here is one possible explanation for what is going on in these cases:18 When we
judge that Billy is partly responsible for the success of the bombing, that is because
we are treating the bombing as counterfactually dependent on his actionnot, ad-
mittedly, on the ordinary way of understanding the counterfactual, but on a slightly
dierent way that holds xed slightly dierent facts about the scenario. That is, in
the actual scenario the backup escort does not, let us suppose, re on Enemy. If we
hold that fact xed when evaluating the counterfactual, then we get the result that if
Billy had not red, then (holding xed the fact that the backup escort does not re)
the bombing would not have succeeded. Moreover, this reading of the counterfactual
seems permissible (although not obligatory). On the other hand, a parallel reading of
the counterfactual in the version of the story where Suzy is protected by a Shield of
Invulnerability seems strained: What extra fact is to be held xed? Presumably, we
have to hold xed the fact that the Shield does not repel Enemys missilesbut it is
not at all clear how to construct (without overly gratuitous deviations from actuality)
a counterfactual situation in which this is the case, but Billy does not re.
I am not at all sure about the prospects for this proposal, and hence I take it that
solving the problem of preempted prevention is a piece of unnished business that
aects my account of causation, by way of complicating the dependence half of it.
A dierent issue arises from cases that look like production until they are exam-
ined up close. I will consider just one such case, and then tentatively oer a lesson I
think we should draw from it. The boxed neuron in the diagram in gure 9.15 func-
272 N. Hall
tions as an AND gate: When and only when input neurons c and d both re, it
res, causing e to re. A paradigm case of the production of one event by two other
events, it would seem. But is it? Note that I did not tell you what it was for the boxed
neuron to re, what such ring consists in. Figure 9.16 oers a closer look at the in-
ner workings of the box. Here it appears that d is not a producer after all. We could
leave the matter there, with the fairly obvious observation that our judgments about
the causal structure of this setup are, of course, going to be defeasible, given further
information about the detailed workings of the setup. But I think a dierent and
somewhat more interesting lesson is called for. Notice rst that it would be inapt to
say that d is not a producer of the ring of the box; it is rather that within the chain of
events that constitutes the ring of the box, the incoming signal from d plays the role
of double-preventer. It does not, for example, help produce the ring of the right-
most minineuron inside the box.
Now, that observation might seem to be of little import. For example, it doesnt
seem to bear at all on the question of what the relationship between the ring of d
and the ring of e is. Is that production or dependence? But in fact I think the ob-
servation is relevant. For I think it is correct to say that when c and d both re, the
AND gate
e
c
d
Figure 9.15
AND gate, close-up
c
d
e
Figure 9.16
ring of d helps produce the ring of the AND neuron, but I also think it is cor-
rect to say that the ring of d does not help to produce the ring of the right-most
constituent neuron in the box. So far that leaves us some room to maneuver with
respect to the ring of e: We could, for example, adopt the view that if d helps pro-
duce any event that is itself a producer of e, then by transitivity d helps produce the
ring of e. Then d comes out as a producer of e, since it helps produce the ring of
the AND neuron, which in turns helps produce the ring of e. Alternatively, we
could adopt a much more stringent standard, and say roughly that in order for d to
be a producer of e, every event causally intermediate between the two must be pro-
duced by d and producer of e. Since the ring of the right-most neuron in the box is
such a causally intermediate eventand since d does not help produce itd will not
come out as a producer of e.
What I think we should say is that depending on context, either answer can be
correct. More specically, I think that causal judgments are tacitly relative to the
level of description we adopt when giving an account of the relevant chain of events
(and that this choice of level of description will be a feature of the context in which
we are making our causal judgments). In giving an account of the events in gure
9.15, for example, we can adopt a level of description that includes such categories
as ring of the AND neuron; that is, we will provide some such description as
neurons c and d both re, each sending a stimulatory signal to the boxed neuron,
which then res, emitting a stimulatory signal that reaches e, which then res.
Alternatively, we can adopt a level of description that speaks not of the ring of the
AND neuron, but rather of the various rings of the neurons within it and of the
stimulatory and inhibitory connections between them. Relative to the rst choice of
level of description, d comes out as a producer of e; relative to the second, it does
not.19
Obviously, another major piece of unnished business is to spell out the relevant
notion of levels of description, and to explain exactly how such levels nd their
way into the contexts in which we make our causal judgments. Here Ill content my-
self with responding to one objection. Some may view this introduction of context-
sensitivity into the account as a cost; but as far as I can see such an attitude manifests
the same confusion we saw earlier of a sound methodological precept with an un-
sound a priori conviction about the workings of our causal concepts. Just as it makes
good methodological sense to begin an investigation of our concept of causation with
the working hypothesis that it is univocal, so too it makes good sense to adopt as
an initial working hypothesis the view that it is not context sensitive. But theres
no sense whatsoever in maintaining such hypotheses when our investigations have
revealed complexities too serious for them to accommodate. If I am right, the view
274 N. Hall
we are pushed to is that our thinking about causation recognizes two basic and fun-
damentally dierent varieties of causal relation, and that which relation is in play
in any given situation isor least candepend on contextually specied features of
how we are conceptualizing that situation. That this is a view that quite obviously
needs detailed argument and defense should make it seem unattractive only to those
with an excessive devotion to the curious notion that philosophical life should be
easy.
Notes
1. See for example Mellor (1997) and Bennett (1988), as well as Mellor, For Facts as Causes and Eects,
chapter 12 in this volume.
2. See Kim (1973b); see also the discussion in Collins, Hall, and Paul, Counterfactuals and Causation:
History, Problems, and Prospects (chapter 1 in this volume) and Lewis, Causation as Inuence (chap-
ter 3 in this volume).
3. For excellent discussions of the issues involved in providing a full-blown philosophical account of
events, see Lewis (1986c) and Bennett (1988).
4. Notice also that Bennett and Lewis both use the locution A causes B rather than the weaker A is a
cause of Bthus illegitimately suggesting that a particularly salient causal connection is being asserted.
That wont do; after all, doesnt it also sound wrong to say, e.g., that the forests presence caused the re?
But it doesnt sound so bad to say that it was a cause of, or among the causes of, the re.
5. Notice that the distracting intuitions evoked by Bennetts example are silent here: There is no good bit
of common sense analogous to Lombards observation that heavy rains . . . dont start [res]; further-
more, no event stands out as a particularly salient cause of the forests presence (although in the right
context, the April rains just might).
6. For standard treatments of the counterfactual, see, e.g., Stalnaker (1968), Lewis (1973a), and Lewis
(1973b). Whether these standard treatments are adequate to the needs of the counterfactual analysis is a
question we will take up shortly.
7. Note that Lewiss account of the counterfactual conditional does not rule out backtrackers in principle,
but only when the world exhibits an appropriate sort of (contingent) global asymmetry; in this way he
hopes to leave room for the possibility of backward causation.
8. Swain (1978, see especially pp. 1314) has overlooked this point. He considers a case with exactly the
structure of that described by gure 9.2, yet fails to notice that his views on counterfactuals deny him the
resources needed to secure the link between c and e.
9. The usual caveats apply: There might be more than one such minimal alteration to the state at t, or there
might be an innite sequence of minimal alterations, each more minimal than its predecessor. Either way,
the proper x is to take the conditional to be true just in case there is some alteration A such that the
consequent comes out true for every choice of alteration A
0
that is at least as minimal as A. Note also that
we would need to amend this rule, if we wanted our analysis to accommodate backward causation and
action at a temporal distance.
10. But see Hall (2002b) for detailed argument and discussion.
11. The word structure is intentionally ambiguous: We could take the structure to be the mereological
fusion of all the events, or we could take it to be a set-theoretic construction out of them (most simply, just
the set of them). It doesnt matter, as long as were clear on what duplication of event-structures amounts to.
12. For various reasons, not worth the long digression their spelling out would require, I do not think we
can add an only if to the if to get i. See my (2002b) for discussion.
13. There are, in the literature, various other apparent counterexamples to Transitivity that cannot be
handled merely by denying Dependence (see, e.g., McDermott 1995a). But on my view they are only ap-
parent; see my Causation and the Price of Transitivity (chapter 7 in this volume) for detailed discussion.
14. I do not think that whenever we have counterfactual dependence between two facts, the statement
asserting this dependence should be construed as causal. It depends on why the dependence holds. For
example, If it hadnt been that P, it wouldnt have been that Q could be true because Q entails P, in
which case we shouldnt view this sentence as expressing a causal truth. Alas, I dont think we can hope
to circumscribe the causal dependence claims merely by demanding that the facts in question be logically
independent; for what of counternomics, such as If gravity had obeyed an inverse-cube law, then the
motion of the planets would not have obeyed Keplers laws? While it seems intuitively clear when we
have a relation of dependence that holds for the right sorts of reasons to count as causal and when we
dont, I will leave the project of elucidating these reasons for another occasion.
15. Consider, for example, Suzys throw, in the simple situation where Billys doesnt also throw: Had this
throw not occurred, it is not that nothing would have happened in its place; rather, her arm would have
(say) simply remained at her side. Thats an event, even if not one ordinarily so called.
16. For example, imagine that there is some exquisitely sensitive alarm system, which will be triggered if
the rock deviates at all from its actual path; triggering it in turn initiates some other process that will break
the window. It is possible to set up such an example so that at each moment t of the rocks ight, there is a
unique minimally sucient set for the breaking (a set that will include the part of the ight that is occur-
ring at t); but plausibly the breaking will depend on not one of the events making up the ight. Its worth
contrasting this sort of failure of dependence with the more usual cases of preemption: In those cases, there
is in addition to the process that brings about the eect some other, rival process that somehow gets cut
o. Here, by contrast, there is no such rival process; there only would be, if the events in the main process
were dierent. We might therefore call these cases of overdetermination by a preempted merely potential
alternative. They pose special problems for counterfactual analyses, since many of the strategies for dealing
with cases of overdetermination by preempted actual alternatives fail to apply.
17. You step into the machine, and as a consequence (i) your body dissolves, and (ii) the machine transmits
a signal to a second, distant machine. It receives the signal and as a consequence produces a body that is
an exact replica of yours, as it was when you stepped into the rst machine. Is it you? Even those of us who
believe that teletransportation is possible cant say yes without knowing more of the causal details. For
example, it might be that the second machine had been ready for a long time to produce a bodyone that
just by chance happened to be exactly like yoursand that the signal from the rst machine merely acted
as a catalyst. Then the two person-stages are stages of dierent people, even though the rst is a cause, in
both senses, of the second.
18. Inspired by a proposal of Steve Yablos (personal communication).
19. Although I will not pursue the matter here, it is worth noting that it is quite easy to accommodate this
relativity to level of description within the analysis of production oered in section 7.
276 N. Hall
10Void and Object
David Lewis
1 The Deadly Void
The void is deadly. If you were cast into a void, it would cause you to die in just a
few minutes. It would suck the air from your lungs. It would boil your blood. It
would drain the warmth from your body. And it would inate enclosures in your
body until they burst.1
What Ive said is literally true, yet it may be misleading. When the void sucks
away the air, it does not exert an attractive force on the air. It is not like a magnet
sucking up iron lings. Rather, the air molecules collide and exert repulsive forces
on one another; these forces constitute a pressure that, if unresisted, causes the air to
expand and disperse; the void exerts no force to resist the pressure; and that is why
the air departs from the lungs.
Likewise, when the void boils the blood, there is no ow of energy from the void
into the blood. It isnt like a stove boiling a kettle of water. The blood is already
warm enough to boil, if its vapor pressure is unresisted; the void exerts no counter-
pressure; and so the boiling goes unprevented.
Likewise, when the void drains your warmth, what happens is that your thermal
energy, left to itself, tends to dissipate; and the void provides no inux of energy to
replace the departing heat.
And when the void inates enclosures, again what happens is that the enclosed
uids exert pressure and the void exerts no counterpressure. So nothing prevents the
outward pressure from doing damage.
In short, you are kept alive by forces and ows of energy that come from the ob-
jects that surround you. If, instead of objects, you were surrounded by a void, these
life-sustaining forces and ows would cease. Without them, you would soon die.
That is how the void causes death. It is deadly not because it exerts forces and sup-
plies energy, but because it doesnt.
2 Void and Vacuum
The deadly eects of a void would be just like those of a commonplace vacuum.
Nevertheless, I distinguish the two. The more we learn about the vacuum, the more
we nd out that it is full of causally active objects: force elds, photons, and vir-
tual particles. Spacetime itself, if curved, can serve as a repository of energy. And
perhaps that is not the end. The void, on the other hand, is entirely empty. Thus, if
there is a vacuum within these four walls, there may be quite a lot of objects between
the walls that are capable of exerting forces and supplying energy. Whereas if there is
a void within these walls, then (even though the walls are some distance apart) there
is nothing at all between the walls. What?Not even any spacetime? Not even any
at, causally inert spacetime?No, not even any spacetime. Nothing at all.
The void is what we used to think the vacuum was. It is what the vacuum is ac-
cording to a relational theory of spacetime, according to which particles are sur-
rounded by nothing at all, and are separated not in virtue of substantival spacetime
between them but rather by direct external distance relations from one particle to
another.
Whether or not any such relational theory is true, I take it to be, in some good
sense, a genuine possibility for the world. Nothing rules it out a priori. It is a
broadly logical or conceptual possibility. But conceptual possibility is governed
by a combinatorial principle that says that we can generate new possibilities by
patching together (copies of ) parts of other possibilities.2 So if a relationist world
is possible, and a world full of substantival spacetime is likewise possible, then by
patching together parts of these two worlds, we get a world that consists of sub-
stantival spacetime interrupted by occasional voids. The walls and the spacetime
within them are distinct existences; ergo it is possible for either one to exist without
the other. If the walls exist without the spacetime (and without any other objects be-
tween the walls) then there is a void between the walls.
A void is conceptually possible. But probably it is impossible in another sense: It
violates the laws of nature. Nothing you or nature can do will make a void, but only
a vacuum. Our horric counterfactuals about what would happen if you were cast
into a void are contrary not only to fact but to law.
Yet they are none the worse for that. It is no mistake to conate the deadly void
and the deadly vacuum in our counterfactual reasoning, because the dierence be-
tween the two makes very little dierence to the lethal eects. If you were cast into
the vacuum of outer space, high-energy photons would be the least of your problems.
It is the absences that would do you in. And when it comes to absences, the void is
like the vacuum, only more so. Therefore, although presumably we may not enter-
tain the supposition that you are cast into a void under laws that are exactly those of
the actual world, we know very well which features of actuality to hold xed in sup-
posing there to be a void.
3 Menzies on Causation as Intrinsic
Peter Menzies has identied a source of dissatisfaction with our most popular ap-
proaches to the analysis of causation.3 We would prefer, he says, to think of the
278 D. Lewis
causal relation as an intrinsic relation between cause and eect: a relation that is
instantiated by a pair of events just in virtue of the (natural) properties and relations
of that pair itself, and so supervenes just on the (natural) properties and relations of
the pair; a relation that is independent of the (natural) properties and relations of all
things that are entirely distinct from that pair; and hence a relation that would be
instantiated equally by any other pair of events that shared exactly those (natural)
properties and relations. It would be instantiated equally by any such duplicate pair
regardless of whether the duplicate pair was actual or merely possible; regardless of
whether the duplicate pair was all alone in its universe or whether it was accom-
panied by contingent objects distinct from itself; and regardless of what pattern was
constituted by the accompanying objects, if there were any.4
Two events c and e stand in a relation of constant conjunction i, throughout the
universe, there are many events of the same kind as c, and every one of them is
accompanied by an event of the same kind as e. An analysis that identies cau-
sation with constant conjunction is subject to many diculties and admits of many
improvements, but one diculty remains: constant conjunction is not an intrinsic
relation, and neither is any of its improved descendants.
Two events c and e stand in a relation of counterfactual dependence i, if c had
not occurred, e would not have occurred either.5 A counterfactual analysis of causa-
tion is again in need of improvement. We need to consider a pattern of counter-
factuals that goes well beyond the simple counterfactual dependence of e on c.6
Whatever this pattern may be, it obtains in virtue of whatever makes single counter-
factuals true or false. And in this, a large part is played by the laws of nature. (That
is so even when we entertain counterlegal suppositions. We may have to bend the
laws, but we do not give them away altogether.) And laws, whatever else they may
be, are at least exceptionless regularities throughout the universe (or some large part
of the universe). So counterfactual dependence is not an intrinsic relation, and nei-
ther is any of its improved descendants.
If we thought that causation ought to be analyzed as an intrinsic relation between
events, then constant-conjunction analyses and counterfactual analyses would both
be in trouble. The trouble would go deep. It would persist no matter how well we
succeeded in the game of counterexamples and improvements.
4 Menzies on Causal Functionalism
What sort of analysis of causation would portray the causal relation as intrinsic?
Menzies oers a plausible recipe: he applies the Canberra plan to causation.7 We
start with a platitudinous folk theory of the causal relation. This theory says that the
Void and Object 279
causal relation does so-and-so and such-and-such. Thereby it species a functional
role that the causal relation is said to occupy. We stipulate that the causal relation is
the relation, if there is one, that does in fact occupy the specied role.
(At least, that is the ideal case. But if there is no unique relation that perfectly
occupies the role, we might resort to semantic satiscing. We might grant that an
imperfect occupant of the role is an imperfect, but good enough, deserver of the
name causal relation, if it comes near enough and there is no better candidate to
be had. Or if several candidates were tied for best, so that they deserved the name
equally, we might grant them equal claim by treating the name as ambiguous or in-
determinate in reference.)8
Menzies lists three crucial platitudes. First, the causal relation is a relation be-
tween wholly distinct events. Second, it is a relation that is intrinsic to the event-pairs
that instantiate it. Third, it is typically, though perhaps not invariably, associated
with a probabilistic version of counterfactual dependence. In most cases in which it
relates an event c to an event e, though perhaps not in a minority of atypical cases,
the actual objective single-case probability of event e is substantially higher than it
would have been if c had not occurred. (For instance, in the most extreme case, the
actual probability of e is exactly 1, without even an innitesimal chance of not-e;
whereas without c, the probability of e would have been exactly 0, without even an
innitesimal chance of e, so that e would denitely not have occurred.)
By building the intrinsic character of the causal relation into the denitive func-
tional role as a crucial platitude, Menzies makes sure that the occupant of the role, if
there is one (and if it is a perfect occupant) must be an intrinsic relation. Does this
strategy satisfy Menziess desideratum?
Not entirely. In this case, as in others, the functionalist strategy creates an ambi-
guity. We have the functional role; we have the relation that is the actual occupant of
that role. But presumably it is a contingent matter what relation occupies the role. In
a dierent possible world, perhaps with dierent laws of nature, some dierent rela-
tion may occupy the role. Now let c and e be two events that both occur in some
possible world w. Shall we say that c causes e i c stands to e in the relation that
occupies the role in actuality? Or i c stands to e in the relation that occupies the
role in world w? If we make the former decision, then indeed (at least if the actual
occupation of the role is perfect) causation is an intrinsic relation. But if we make the
latter decision, then causationby which I mean the relation such that, for any
events c and e in any world w, c bears it to e in w i c causes e in wwill be a dis-
junctive aair. It will not be the same as any of the various role-occupants in the
various worlds.9 So even if all these role-occupants are intrinsic relations, it will not
follow that causation itself is an intrinsic relation.
280 D. Lewis
The folk well might have left this subtle ambiguity unresolved. Indeed, they might
never have noticed it. After all, they are mostly interested in causation as it takes
place here in our actual world (or in worlds similar enough to ours that they could be
expected to have the same role-occupant). What mostly matters is that the actual
occupant of the role is an intrinsic relation. So I think that if Menziess strategy is
otherwise successful, then it satises his intuitive desideratum quite well enough. But
we cannot say unequivocally that he has analyzed causation as an intrinsic relation.
5 Causal Functionalism and the Void
Menziess strategy is not otherwise successful. What purports to be a general anal-
ysis of causation turns out to apply only to one kind of causation. Causation by the
void, and causation by absences more generally, have no place in Menziess account.
Omission is omitted.
Menziess topic is the causal relation. A relation requires relata. The void aords
no causal relata: Theres nothing there at all, so theres nothing for events to happen
to, so the void is devoid of events. And even if we allow causal relata to belong to
other categories, still there would be none of them in the voidbecause theres
nothing at all in the void. A vacuum, or an almost-void that still contains at, inert
spacetime, is almost as bad. It has parts; and it may contain events in some tolerant
sense of that word. But it contains nothing that could plausibly be said to bear a
causal relation to the death of a victim. The victim dies, as I said before, not because
of what is there, but because of what isnt. Indeed, whenever any eect is caused by
an absence of anything, we have the problem of the missing relatum. (And likewise
whenever anything causes an absence.) A void, being the absence of any objects at
all, is just the most extreme case of an absence.
Faced with the problem of the missing relatum, we have four possible lines of
response.
(1) We could deny, in the face of compelling examples to the contrary, that ab-
sences ever cause anything. We could deny, for instance, that the void is deadly.
(Likewise, we could deny that anything ever causes an absence. In other words, we
could deny that there is any such thing as prevention.) Simply to state this response is
to complete the reductio against it.
(2) We could reify absences nonreductively. A void, so we might say, is a sui gen-
eris entity, but it is none the worse for that. It is eligible to serve as a causal rela-
tum. It springs up automatically and necessarily whenever, and only whenever, all
else goes away; it is conceptually impossible not to have a void between the walls and
not to have anything else there either. So much the worse, says the reier, for the
Void and Object 281
combinatorial principle, which claims that existential statements about distinct things
are independent.10
(3) We could reify absences reductively. We could identify absences with compar-
atively uncontroversial objects that, as others would say, are somehow associated
with those absences. For instance, we could identify a hole with the hole-lining that,
as wed normally say, immediately surrounds the hole. (Strange to say, some holes
are made of cheese and some of limestone! Strange to say, no holes are exactly where
we would have thought they were!) Or we could identify an absence with a bit of
unoccupied spacetime, if we were not such uncompromising combinatorialists as to
countenance an absence of spacetime itself. One way or another, we can cook up
ersatz absences to serve as relata of the causal relationthough surely they will seem
to be the wrong relata, since we dont really think of these ersatz absences as hav-
ing the same eects (or causes) as the absences they stand in for.11 We might, for
instance, imitate the identication of holes and hole-linings on a grander scale. Take
the most inclusive void of all; and take the mereological fusion of all objects of what-
ever kind. On the principle of identifying hole with hole-lining, and void with sur-
rounding objects, we might identify this greatest void with the greatest object.12
(4) The best response is to concede that a void is nothing at all, and that a lesser
absence is nothing relevant at all and therefore cannot furnish causal relata. Yet
absences can be causes and eects. So I insist, contra Menzies, that causation cannot
always be the bearing of a causal relation. No theory of the causal relation, neither
Menziess theory nor any other, can be the whole story of causation.
The intrinsic character of causation is not our present problem. I do indeed fear
that the intrinsic character of causation is more a hasty generalization than an a
priori desideratum.13 But even if we struck the intrinsic character of causation o
our list of folk platitudes, wed still be trying to characterize the causal relation, so
wed still be in trouble. Any relation needs relata, whether it is intrinsic or not. So the
problem of missing relata hits any relational analysis of causation.14
But does any analysis escape the problem of missing relata?Yes; a counterfactual
analysis escapes. We do not have to reify the void in order to ask what would have
happened if the void had not been there. The void causes death to one who is cast
into it because if, instead, he had been surrounded by suitable objects, he would not
have died. (Here we must assume that if the victim had not been surrounded by the
void, he would instead have been surrounded by the life-sustaining objects that nor-
mally surround usnot by liquid nitrogen, or clouds of nerve gas, or a hail of bul-
lets.) Likewise for lesser absences. If the cause is an absence, then to suppose away
the cause counterfactually is not to attend to some remarkable entity and suppose
282 D. Lewis
that it does not exist. Rather, we need only suppose that some unremarkable entity
does exist. Absences are spooky things, and wed do best not to take them seriously.
But absences of absences are no problem.
Note well that in defending a counterfactual analysis, I am not claiming that
all causation consists in a relation of counterfactual dependence between (distinct)
events. That theory would not escape the problem of missing relata. A relation of
counterfactual dependence is still a relation, a relation still needs relata, and absences
still fail to provide the needed relata. The counterfactual analysis escapes the prob-
lem because, when the relata go missing, it can do without any causal relation at all.
6 Menziess Analysis Retargeted
So far, I have been arguing that Menzies is not entirely right. But in fact I think that
a large part of what he says is right, can be separated from the parts that are wrong,
and can be accepted even by one who favors a counterfactual analysis. That will be
our business for the rest of this essay. Menzies has not given us a fully general anal-
ysis of causation, but he has given us something. I think he has given us the right
analysis of the wrong analysandum. Let us introduce a name for that which Men-
ziess functionalist analysis succeeds in analyzing. Let us call it bi. That word enjoys
some unocial currency among those who conceive of causation much as Menzies
does; so let it be our word for the kind of causation that ts their conception, even if
we accept as we should that this is not the only kind of causation.
A theory built around Menziess three crucial platitudes species a functional
role for a relation: an intrinsic relation between distinct events that is typically, but
perhaps not invariably, associated with a probabilistic version of counterfactual de-
pendence. Bi is dened to be the occupant of this functional role, if such there be.
There is the actual occupant of the bi-role, unless we are badly wrong about the
ways of our world. Other possible worlds, some of them, might have dierent rela-
tions occupying the bi-role. In case of imperfect or nonunique occupation of the
bi-role, we might resort to semantic satiscing in the ways already considered.
What sort of relation might bi be? We can echo much of what Menzies says,
overlooking that he says it not about bi but about causation.15 Bithe actual
occupantmight, or it might not, be some relation well known to physics. It
might, for instance, be force. (More precisely, it might be the relation of exerting a
force upon.) Or, taking up David Fairs suggestion about the actual nature of
causation,16 it might be a relation of transfer of energy or momentum. It might
be a Humean-supervenient relation. Or it might be a relation posited by some
Void and Object 283
anti-Humean metaphysic of nomological necessity.17 It might be a perfectly natural
relation, or it might be more or less disjunctive. Myself, Id like to think that the
actual occupant of the bi-role is Humean-supervenient, physical, and at least fairly
natural; but nothing else I shall say here is premised on that hope.
Bi might have been one of the relations posited by bygone physical theories that
have turned out to be false in our actual world, but are true in various other worlds.
Some worlds obey Aristotelian dynamics; the occupant of the bi-role in such a world
might be a kind of force, if we may call it that, which is proportional to velocity
rather than acceleration. Or the occupant of the bi-role might have been a relation
of transfer of impetus, where impetus is rather like inertia except that it fades away
spontaneously, like the heat in a red-hot poker.18 Or bi might have been some
otherworldly physical relation unlike anything that has ever crossed our minds.
Or bi might have been something even stranger. Consider a possible world where
occasionalism is true: God is a third party to every causal relationship between events
in nature. Then the best available candidate to occupy the bi-role would be an im-
perfect occupant of the role and an imperfect deserver of the name. It would be
a relation between events that did not require accompaniment by anything else in
nature but did require accompaniment by God; so it would not be, strictly speaking,
an intrinsic relation. Yet we could reasonably judge that this imperfect candidate
deserved the name quite well enough.
7 Varieties of Causation
Causation by absence is not an instance of bi. Nevertheless it can be described in
terms of bi. If you were cast into the deadly void, the absences that would kill
you would be absences of bi; because, if instead you were surrounded by suitable
objects, events involving those objects would stand in the bi-relation to the events
that would constitute your continuing life. Equivalently: Events involving those ob-
jects would prevent the event of your death by standing in the bi-relation to events
incompatible with it.
Beginning with bi itself, we can dene several varieties of causation.
(1) Event c directly causes event e i c stands to e in the relation that occupies the
bi-role. For short: i c bis e.
(2) The absence of any event of kind C directly causes event e i, had there been
an event c of kind C, c would or might have bied some event d incompatible with
event e.
284 D. Lewis
(3) Event c directly causes the absence of any event of kind E i c bis some event
d incompatible with any event of kind E.
(4) The absence of any event of kind C directly causes the absence of any event of
kind E i, had there been an event c of kind C, c would or might have bied some
event e of kind E.19
But there are also cases of indirect causation: An event (or absence) c causes an
event (or absence) that then causes another event (or absence) e. Or c may cause both
an event and an absence, which then jointly cause e. Or there may be a chain with
three steps, or four, or any number. Cases of indirect causation need not reduce to
direct causation. Take, for example, a case of causation by double prevention: Event
c causes the absence of any event of kind D, which absence in turn causes event e.20
It does not follow, and it may be false, that c bis e. Even if bi itself is intrinsic, the
causal relation of c to e in cases of double prevention is sometimes extrinsic.21
So the functional analysis of bi aords a basis for dening many varieties of
causation. All the varieties there could possibly be? We have no assurance of that.
Maybe some possible worlds have no occupant at all of the bi-role, not even an
imperfect occupant; and maybe in some such worlds the actual occupant of the bi-
role also is nowhere to be found. And when we depart counterfactually from such
a biess world, taking care to avoid gratuitous dierences from our starting point,
presumably the counterfactual situations we reach will be equally biess. Yet might
there not be some sort of causation in a biess world? Maybe all the causal relations
of events in such a world are thoroughly extrinsic, far more so than in the case of the
occasionalist world we imagined before. The intuition of the intrinsic character of
causation may indeed be right for one basic variety of causation in the actual world,
but it is by no means given a priori.
8 Ambiguity? Disjunction?
What shall we conclude from this proliferation of dierent varieties of causation?
Has it turned out that we have not one concept of causation but many, so that many
dierent analyses are required to capture the many dierent senses of the word
cause? Or is it rather that our one concept of causation is a radically disjunctive
concept, so that a correct analysis must consist in a long list of alternative cases?
Were in trouble either way.
If causation in a biess world is possible, that is a problem equally for both
hypotheses. On the many-concepts hypothesis, the problem is that one concept of
Void and Object 285
causation (at least) will never be reached by our chain of denitions starting with the
functional denition of bi. On the disjunctive-concept hypothesis, the problem is
that one disjunct (at least) will never be reached. To complete an inventory of senses,
or to complete the disjunctive analysis of the single sense, it seems that we must nd
some dierent starting point.
Another problem for the many-concept hypothesis is that it requires distinctions in
our thinking that sometimes we do not make, need not make, and are in no position
to make. If one event directly causes another, for instance, that is causation in one
sense; whereas if one event causes another indirectly, in a case of double prevention
(or in some still more indirect case) that is causation in a dierent sense. But when we
neither know nor care whether the causation we have in mind is direct or indirect,
what concept of causation are we employing then?
(Example: The frightened passenger pulls the cord, knowing that this will cause the
train to stop. Being moderately well informed, he knows that pulling the cord opens
a valve connecting a reservoir to the outside air. The changed pressure in the reser-
voir changes the balance of forces on the brake shoes, thereby applying the brakes
and stopping the train. But the passenger doesnt know whether this train is tted
with air brakes or vacuum brakes. If air brakes, then the air in the reservoir is nor-
mally above atmospheric pressure; so opening the valve lowers the pressure and
removes a force, and so the stopping of the train is a case of double prevention. If
vacuum brakes, then the air in the reservoir is normally below atmospheric pressure;
so opening the valve raises the pressure and applies a force, and so the stopping of
the train is a case of direct causation. But so long as he can cause the train to stop,
its all the same to the passenger what kind of causation it is.)
The disjunctive-concept hypothesis now seems better. But it faces an urgent, if not
absolutely compulsory, question. Why do we disjoin exactly these disjuncts? Why is
the disjunction of just this long list of alternatives anything more than a miscellane-
ous gerrymander? What makes it a natural kind?
It is as if we came upon some people who had a peculiar taxonomy for birds. They
group together a kind that includes swans, but not ducks or geese; eagles and hawks,
but not vultures; magpies and crows, but not ravens or currawongs or mudlarks, and
indeed no other birds at all. We would be entitled to ask why just these birds are
included, and it would not be good enough just to say that all classes are equally
classes, and that these people happen to have picked out this class.
The many-concepts hypothesis and the disjunctive-concept hypothesis are both un-
satisfactory. Yet if we analyze causation by starting with the functional analysis of
bi, and going on to dene other varieties of causation one by one in terms of bi,
that is the choice we come to.
286 D. Lewis
9 Conclusion
I think we are aiming our answers at the wrong question. Menzies went wrong
when he took the functional denition of bi to be the whole of a conceptual analy-
sis of causation. We still go wrong if we take it to be even a rst step toward con-
ceptual analysis. We should look elsewhere for a conceptual analysis. And we should
look elsewhere for a question to which the invocation of bi aords a satisfactory
answer.
What is causation? As a matter of analytic necessity, across all possible worlds,
what is the unied necessary and sucient condition for causation?It is somehow
a matter of counterfactual dependence of events (or absences) on other events (or
absences).
What is causation? As a matter of contingent fact, what is the feature of this
world, and of other possible worlds suciently like it, on which the truth values of
causal ascriptions supervene?It is bi: the pattern of relatedness of events to one
another by the relation that is the actual occupant of the bi-role. Bi is literally the
basic kind of causation, in this world anyway: the basis on which other varieties of
causation supervene.
Two dierent answers to two dierent questions. They are not in competition. I
conjecture that both are right.
If bi is oered not as conceptual analysis but as a basis for supervenience, then it
matters little if the varieties of causation, when described in terms of bi, are many
and diverse. Unifying the miscellany is a job for conceptual analysis. And if bi is
oered as a supervenience basis for causation as it takes place here in our world, then
the possibility that some other variety of causation takes place in biess worlds re-
mote from actuality is no cause for alarm.
Let me say more fully what I have in mind. Doubtless all will agree that the visual
qualities of dot-matrix picturessay, the quality of looking clutteredsupervene on
the arrangement of light and dark pixels, at least if we restrict our attention to black
and white pictures with maximum contrast. This means that no two possible pictures,
at least no two that both fall within our restricted class, dier in their visual qualities
without also diering in their arrangement of pixels. If one looks cluttered and the
other doesnt, they cannot be alike pixel for pixel. Likewise, the thesis under consid-
eration says that no two possible worlds, or at least no two that fall within a certain
restricted class, dier in respect of the truth values of causal ascriptions without also
diering in their bi-relations. At least one bi-related pair of events in one world
must fail to correspond to any bi-related pair in the other world, and that is what
makes the causal dierence.
Void and Object 287
The narrower is the restricted class of worlds within which all causation is said to
supervene on bi, the weaker and safer and less interesting our thesis will be. How
shall we strike the balance between safety and interest? Since we are especially inter-
ested in causation as it takes place in our actual world, the actual world had better
fall within the restricted class. But our supervenience thesis, if restricted to one single
world, would be utterly trivial; and besides, our interest extends at least to worlds
that only narrowly escape being actual. I propose that the restricted class should
consist of exactly those worlds that satisfy two conditions. (1) They are worlds where
the relation that occupies the bi-role is the same relation that occupies that role in
actuality. And (2) they are worlds where the laws of nature are the same as the actual
laws of nature.
Condition (2) is not motivated just by caution. Remember our starting point: when
the void, or some lesser absence, causes an eect, that is because of the absence of
what it takes to prevent that eect (nothing counteracts the vapor pressure that
would cause warm blood to boil, and so on). At any rate, that is how causation by
absences works under the actual laws of nature, or so we think. But might it not be
otherwise under dierent laws? Suppose there were a fundamental law that said that
a certain spell would turn a prince into a frog i that prince was within a mile of the
edge of a void. Under such a law, a void could cause a transmogrication in a way
that had nothing to do with absence of bi. There could be such a lawwhy not?
But it is irrelevant to capturing our opinions about how causation works in actuality.
So we unabashedly bar the monster: We stipulate that a world with such a law falls
outside the range to which our actuality-centered supervenience thesis is meant to
apply. And for good measure, but perhaps with needless caution, we likewise stipu-
late that all worlds that depart from the laws of actuality fall outside the range.
It is because our supervenience thesis is restricted, and because the restriction
makes reference to actuality, that our thesis is contingent. Had we started from some
dierent possible world, and restricted according to the same two conditions, we
would have restricted the thesis to a dierent class of worlds. The thesis restricted to
worlds that satisfy our conditions relative to this world may be true, but the thesis
restricted to worlds that satisfy the same conditions relative to some other world may
be false. Or, if we are unlucky, vice versa.
We saw how causation by absences, at least as it takes place in our actual world,
could be dened piecemeal using various counterfactuals about bi. So if all causa-
tion is to supervene on bi, these counterfactuals about bi must supervene on bi.
That is plausible enough. For when we depart counterfactually from a given world,
we make no gratuitous changes. Except insofar as the supposition we are making
requires dierences, the character of the given world carries over into the counter-
288 D. Lewis
factual situation. In particular, the laws governing bi tend to carry over.22 Return,
one last time, to the victim cast into the void (or into something as much like a void
as the actual laws allow). If instead he were surrounded by suitable objects, those
objects would conform to the laws that actually govern bi. If not, that would be a
gratuitous dierence between the counterfactual situation and the actual world.
Notes
1. Here I follow the lead of Martin (1996): it seems that the void has . . . terrible causal powers (p. 62).
Martin later says that voids are causally relevant but not causally operative (p. 64), but I do not know
what he means by this.
2. On combinatorial principles see my (1986e), pp. 8692; and Armstrong (1989).
3. Menzies (1996) and (1999). Menzies might better have suggested not that causation is an intrinsic rela-
tion, but rather that being a causal chain is an intrinsic property. For present purposes, we may leave this
correction unmentioned.
4. Here I have combined denitions of intrinsic relations taken from my (1983b), n. 16; and from Langton
and Lewis (1998). These denitions dier, but can be expected to pick out the same class of relations.
5. See my (1973a) and (1986b).
6. See my Causation as Inuence, chapter 3 in this volume.
7. Menzies (1996). The Canberra plan is modeled on analytic functionalism in the philosophy of mind, and
on Carnaps proposal for dening analyticity in a theoretical language. See Carnap (1963).
8. See Bedard (1993).
9. It will be what I called a diagonalized sense in my (1970).
10. Casati and Varzi (1994) defend nonreductive reication of holes: Holes are immaterial bodies that
depend for their existence on the arrangement of matter. Martin, in his (1996), is probably best classed as a
nonreductive reier, despite his emphatic warnings that absences are not things. (His conicting suggestion
that they are localized states of the universe would seem to be retracted by his denial that they are nat-
ural properties of things [p. 58]. The universe would seem to be a thing, states would seem to be properties,
and properties of local emptiness would seem to be not unnatural.)
11. Lewis and Lewis (1970) is, for the most part, a dialogue between a reductive and a nonreductive reier.
Neither one notices the option of paraphrasing hole-statements in terms of quantication over ersatz holes
without also claiming that holes are identical to ersatz holes. Frank Jackson calls attention to that option
in his (1977), p. 132.
12. Or, since the greatest object has the property of totalityof being all there iswe might identify the
greatest void with the having by that object of the property of totality. This ersatz void is as wrong as it
can be in its eects (and causes): It causes what objects cause, not what the void unassisted by objects
causes. And it is as wrong as it can be in its location, being exactly where the void isnt. It violates a com-
binatorial principle, since it cannot possibly coexist with any extra object. For that very reason it is of
doubtful ontological status, being a having of a merely extrinsic property.
D. M. Armstrong uses havings of totality as truthmakers for negative existential statements, inter alia in
his (1997), chapter 13. But he never asks these totality states of aairs to serve as ersatz absences.
13. See Causation as Inuence, chapter 3 in this volume.
14. How about an analysis in terms of a relation between propositions, where, in the case of causation by
or of an absence, one of the related propositions is a negative existential?Not problematic; but not what
Id call a relational analysis of causation. We dont want to say that a cause- or eect-describing proposi-
tion is itself the cause or eect.
Void and Object 289
15. Menzies (1999).
16. Fair (1979).
17. It might for instance be the relation of instantiating a law, where a law is taken to be at once a state of
aairs involving universals and a universal in its own right. See Heathcote and Armstrong (1991).
18. See Buttereld (1957), pp. 813.
19. The denitions of cases (2)(4) resemble, but dier from, the denitions given in Fair (1979), p. 247,
and also the somewhat dierent denitions given in the nal section of Dowe (1999). Note that these
counterfactual denitions are not the same as the counterfactuals about events and absences that would
appear in a counterfactual analysis of causation.
20. See Ned Hall, Two Concepts of Causation, chapter 9 in this volume; and McDermott (1995a).
21. As is noted in McDermott (1995a) in connection with his example of Nixon, Haig, and Joe Blows
breakfast. See also the billiard-ball example in Causation as Inuence, p. 84 of this volume.
22. Unless violating a law yields the least departure overall. See Lewis (1979a).
290 D. Lewis
11Causing and Nothingness
Helen Beebee
1 Introduction
According to the way we normally speak about the world, absences can be causes.
We might say, for instance, that Joness failure to close the re doors was a cause of
the raging re that destroyed the building, or that Smiths failure to water her oce
plants was a cause of their death, or that lack of rain was a cause of the bush re.
So it seems that an adequate theory of causation is going to have to rule that there is
such a thing as causation by absence.
I dont believe there is any such thing as causation by absence. In this essay Im
going to try to defend that claim against the objection just raised, partly by arguing
that there are features of commonsense assertions and denials of causation by ab-
sence to which no theory of the metaphysics of causation ought to be doing justice,
and partly by making a positive claim about the role absences play in our explana-
tory practices, a claim that, I think, allows me to rebut the objection that the repu-
diation of causation by absence ies in the face of too much of our commonsense
understanding of causation to be taken seriously.
The reason I deny that theres any such thing as causation by absence is that I
want to uphold the view that causation is a relation between events. To be rather
more picturesque, I subscribe to what Helen Steward has recently dubbed the net-
work model of causation: The complete causal history of the universe can be rep-
resented by a sort of vast and mind-bogglingly complex neuron diagram of the
kind commonly found in discussions of David Lewis, where the nodes represent
events and the arrows between them represent causal relations.1 Or, to put it another
way, the causal history of any event is, as Lewis puts it, a relational structure
(1986c, p. 216).
I think its fair to say that the network model is the dominant model of causation
in contemporary metaphysicslargely owing to the inuence of Lewis and David-
son.2 The network model lies behind most of the theories of causation currently on
oer; it also lies behind a vast amount of the literature on the philosophy of mind.
If there is causation by absence, then the network model cant be rightor at
least, it cant be right assuming there are no such things as negative events. And I
assume in what follows that there are no negative eventswhich is to say, more or
less, there are no events whose essence is the absence of a property or particular. If
Joness failure to close the re door is not an event, and if this failure was a cause
of the re, then the full causal history of the re is not exhausted by the network of
events and causal relations between them, for there will be no event of Joness fail-
ure, and hence no causal relation between his failure and the re. The network model
cannot accommodate the factif it is a factthat Joness failure caused the re, and
hence cannot be the whole causal truth about reality.
How should we solve the problem? One solutionthe one I favoris to hang
onto the network model of causation and deny that there is any causation by ab-
sence. On this view, Smiths failure to water her oce plants was not a cause of their
death. This seems to be a deeply unpopular solutionand surprisingly so, given the
prevalence of the network model in contemporary metaphysics. Heres what Lewis
has to say about it in Void and Object (chapter 10, this volume): We could deny,
in the face of compelling evidence to the contrary, that absences ever cause anything.
We could deny, for instance, that the void is deadly. . . . Simply to state this response
is to complete the reductio against it.3
The other obvious solution is to hang onto causation by absence and scrap the
network model. This is the solution for which Hugh Mellor has long been arguing,4
and has been suggested more recently by Lewis in Void and Object. According to
Mellors view, causation is not a relationor at least, the most basic kind of causa-
tion is not relational. Rather, causal facts have the form E because C, where C
and E are facts and because is a sentential connective. Since facts are not partic-
ulars, on Mellors view causation is not a relation. And since facts can be facts about
absences, causation by absence causes no special problems on his view. In much the
same vein, Lewis proposes that causation be analyzed in terms of counterfactual de-
pendence not between events, but between propositionspropositions that need not
assert that some event or other occurs.
A third solutionone that is suggested but not condoned by Lewis in Void and
Objectis to take the relational kind of causation as the most basic kind of causa-
tion, and dene causation by absence in terms of relational causation.5 According to
this kind of solution, the network model captures all the basic causal facts, but there
are other causal facts whose obtaining supervenes on the actual pattern of causal
relations together with some extra counterfactuals.
I shall call the rst view the relationist view, since it upholds the thesis that causa-
tion is a relation between events. I call the other two views nonrelationist views, since
they both uphold the thesis that causation is not always a relation. The central claim
of this essay is that, if our aim is to do as much justice as possible to commonsense
intuitions about causation by absence, then it doesnt much matter whether we up-
hold relationism or nonrelationism.
This might seem like a surprising claim, since relationism, unlike nonrelationism,
is forced to deny a whole host of causal claims that common sense holds true. As I
292 H. Beebee
said at the beginning, undoubtedly many of us think that Joness failure to close the
re doors was a cause of the re, or that the lack of rain was a cause of the bush re,
and so on. Fair point: The relationist view really does deny all of that. But this is to
tell only half the story. For common sense also makes a lot of negative causal claims
about absences. Take Brown, who lives on the other side of the city and has no con-
nection whatever with either Smith or Jones. Commonsense intuition has it, I think,
that neither Browns failure to close the re doors nor his failure to water Smiths
plants were causes of the re or of the plants death. The relationist view, of course,
gets these cases right: Browns omissions are not causes, because no absencesand
therefore no omissionsare. But nonrelationism gets these negative causal judgments
wrong. According to the relationist, the commonsense claim that Joness omission
caused the re is false; according to the nonrelationist, the commonsense claim that
Browns omission did not cause the re is false.
So when it comes to doing justice to commonsense intuitions about causation by
absence, neither view fares very well. Luckily, I dont think this matters very much.
Briey, the reason it doesnt much matter is this. As Ill argue in sections 2 and 3,
commonsense intuitions about which absences are causes and which arent are highly
dependent on judgments that it would be highly implausible to suppose correspond
to any real worldly dierence at the level of the metaphysics of causation. For in-
stance, sometimes common sense judges the moral status of an absence to be relevant
to its causal status. But no philosopher working within the tradition Im concerned
with here thinks that the truth conditions for causal claims contain a moral element.
It follows that whatever we think about whether or not causation is a relation, were
going to have to concede that common sense is just wrong when it takes, say, moral
dierences to determine causal dierences. There is no genuine causal dierence be-
tween those cases that common sense judges to be cases of causation by absence and
those that it judges not to be cases of causation by absence. Hence both the rela-
tionist and the nonrelationist must agree, and with good reason, that commonsense
judgments about causation by absence are often mistakenthough of course they
disagree about which commonsense judgments are mistaken.
In sections 4 and 5, I show how absences can gure in causal explanations even
though they do not cause anything. So although it is false to say that Smiths failure
to water her oce plants caused their death, it is nevertheless true to say that the
plants died because Smith failed to water them. Here I appeal to Lewiss analysis of
causal explanation, according to which to explain an event is to provide information
about its causal history: information that need not be restricted merely to citing the
events causes. I claim that common sense judges some absences to be causes because
it fails to distinguish between causation and causal explanation, and (in section 6)
Causing and Nothingness 293
that the sorts of distinctions discussed in sections 2 and 3 that ground commonsense
assertions and denials of causation by absence are best seen as distinctions between
explanatorily salient and nonsalient absences. The moral is that commonsense intui-
tions about causation by absence are no more damaging to the relationist than they
are to the nonrelationist; hence, those intuitions provide no good reason for aban-
doning the network model of causation.
2 Causation by Absence and Common Sense
In what follows, Im going to assume that absences that are causesif there are
anyare at least necessary conditions of their eects. I dare say this is a false as-
sumption: Maybe (again, if there is any causation by absence at all) there are cases
where, had the absent event occurred, the eect in question would still have had
some chance of coming about. I ignore this possibility for the sake of simplicity.
With the assumption that absences are at least necessary conditions of what they
cause in place, we might try the following denition of causation by absence:
(I) The nonoccurrence of an event of type A caused event b if and only if, had an
A-type event occurred, b would not have occurred.
This denition is dierent from both Lewiss and Dowes,6 and may well be inad-
equate for reasons other than the one discussed below. However, since both Lewiss
and Dowes denitions can easily be seen to fall prey to the objection too, it will
make things easier if we make do with the simpler denition (I).
Not surprisingly, (I) gets the right answer in cases that common sense judges to
be genuine cases of causation by absence. Here are some examples.7 Flora normally
gratuitously waters her neighbors orchids. But she stops watering them, and they
die. Common sense judges that Floras omission was a cause of the orchids death,
since had she watered them, as she usually does, the orchids would not have died.
Second example: Zs dog is bitten by an insect, contracting an eye disease as a result,
which Z ignores. The dog loses its sight. Intuitively, Zs negligence caused the dogs
blindness; and again, the denition gets this right: Had Z not ignored the eye disease,
the dog would not have lost its sight. Third example: An old tree decays and falls to
the ground, crushing some rare wild owers. The park ranger failed to inspect the
tree, thereby making himself a cause of the owers death. Again, (I) gets this right:
Had the park ranger inspected the tree (and thus had it carefully removed rather than
allowing it to fall), the owers would have survived. Fourth example: A geranium
survives the London winter because there is no frost. Intuitively the absence of frost
was a cause of its survival; and, once again, this judgment is in accordance with (I).
294 H. Beebee
So far so good. The trouble is that our denition is too inclusive: It renders all
sorts of absences as causes that common sense does not recognize as causes. Com-
monsense intuition identies Floras failure to water the orchids as a cause of their
death, but not the failure of the other neighbors, and certainly not the failure of
people on the other side of the world who neither know nor care about the orchids
even though its perfectly true of each of them that had they watered the orchids,
they would not have died. Similarly, while common sense judges the dog owners or
the park rangers omission to be a cause of the dogs blindness or the crushing of
the owers, it does not judge the omissions of others as causes of those events, even
though the relevant counterfactuals are true too. And similarly for the geranium
example: We would ordinarily judge that the absence of a hungry geranium-eating
koala was no part of the causal history of the geraniums survival, even though had
such a koala been present the geranium would not have survived.
Why is this? What grounds common senses discrimination between absences that
are and are not causes of an event? Hart and Honore, in their book Causation in the
Lawa classic text for those interested in commonsense causal judgmentsclaim
that what makes us single out one omission as a cause but not another is abnormal-
ity. We regard Floras failure to water her neighbors orchids as a cause of their
death because Floras failure is abnormal: She normally does water them. On the
other hand, there is nothing abnormal in the failure of the other neighbors, or of
other people in other countries, to water the orchids; hence, those failures are not
causes of the orchids death.8
The abnormality criterion seems to work for the orchid case, and it seems plausible
to suppose that it also works for cases like the geranium case: cases that are not
omissions by human agents. The absence of frost in winter in London is quite un-
usual, and hence according to Hart and Honores criterion may count as a cause of
the geraniums survival. The absence of hungry geranium-eating koalas, on the other
hand, is perfectly normal, and hence does not qualify as a cause of the geraniums
survival.
Stapleton claims, howeverand I think shes rightthat there are many cases in
which the abnormality criterion fails to explain our commonsense causal judgments.
In the case of Zs dog, for instance, she says that we would regard Zs conduct as
a cause because against the backdrop of Zs ownership of the dog we expect Z to
have acted, not because of expectations generated by past conduct but for moral
reasons (1994, p. 122). And in the park ranger case, she says, we single out the park
ranger and nobody else because she, and nobody else, has a legal duty to inspect the
tree. The point here is that we take these causal judgments to be correct even if the
relevant omissions are not in any way abnormal: The judgments stand even if Z is
generally very bad at looking after his dog, or if it is the park rangers rst day on
the job.
Stapleton also claims that whether an omission is to count as a cause can depend
on epistemic features. Suppose, for instance, that a certain drug in fact has harmful
side eects, but that this risk is unforeseeable: We could not reasonably expect the
drug manufacturer to have known about it. In such a case, she says, we should not
say it was the defendant-manufacturers failure to warn of the unforeseeable risk
which caused the injuryclearly a silly idea which no common sense version of cau-
sation could accommodate (ibid., p. 124). On the other hand, if the manufacturer
did know of the risk, or if we could reasonably have expected them to have found out
about the risk, then we would say that their failure to warn consumers of the risk was
a cause of the side eects.
So it seems that common sense singles out an absence as a cause when, and only
when, it stands out in some wayeither from what normally happens, or from some
norm the absence of which (generally an omission) counts as a violation. The norm
might be moral, legal, or epistemic, as Stapletons examples illustrate; but other sorts
of norm may well play a similar role. Owens failure to get to the ball after Beckhams
cross into the penalty area counts as a cause of Englands defeat, but Seamans fail-
ure to do so does noteven though, had Seaman (the goalkeeper) somehow managed
to get to the ball, he would undoubtedly have scored and England would have won.
Why? Because Owen is an attacker and Seaman is the goalkeeper; its an attackers
job, and not the goalkeepers, to score goals from crosses into the penalty area.
What all this points toward is a denition of causation by absence that goes some-
thing like this:
(II) The absence of an A-type event caused b if and only if
(i) b counterfactually depends on the absence: Had an A-type event occurred, b
would not have occurred; and
(ii) the absence of an A-type event is either abnormal or violates some moral, legal,
epistemic, or other norm.
This denition, I think, does justice to as many commonsense causal judgments
about absences as anyone could wish for. The trouble is that although it works qua
linguistic analysis of the ordinary concept of causation by absence, it doesnt look
like the sort of analysis we ought to be giving of the metaphysics of causation by
absence.
Take the violation-of-norms part of the denition. If we take the denition to give
the truth conditions of causation by absence claims, then causal facts about absences
296 H. Beebee
depend in part on normative facts: facts about whether a moral or epistemic or other
norm has been violated. But nobody within the tradition of the metaphysics of
causation that Im concerned with here thinks that causal facts depend on human-
dependent norms.
Even if taking causal facts to depend on normative facts werent in itself such a
bad thing, it would in any case make the truth of causal claims turn out to be a
relative matter. For instance, you and I might dier in our epistemic standards, so
that I count the side eects of a drug as foreseeable and you dont; hence I count the
drug manufacturers failure to warn consumers as a cause of their side eects but you
do not. Either there is some absolute epistemic standard that marks out what some-
one can reasonably be expected to foreseewhich seems wildly implausibleor the
causal status of the failure is relative to dierent standards.
A similar point applies to the abnormality part of the denition. For one thing,
how often something has to happen in order to count as normaland hence for
its absence to count as abnormalis always going to be rather uid. And for an-
other, the same absence might count as normal relative to one kind of regularity and
abnormal relative to another. Suppose, for instance, that Flora is in the habit of
doing gratuitous repetitive favors for her neighbors, like watering their gardens or
washing their cars, but that she invariably stops doing them after a month or so.
Now is Floras failure to water her neighbors orchids after thirty days of watering
them normal or abnormal? There doesnt seem to be any principled way of answer-
ing that question, so it looks as though the causal status of her omission is simply
indeterminate.
It follows that if we want facts about causation to be reasonably determinate and
not relative to extraneous facts about whether a putative cause happens to count
as abnormal, immoral, illegal, or whatever, then any account of the metaphysics of
causation by absence is going to have to be pretty revisionary: No adequate account
of what in the world makes causal claims about absences true or false is going to be
able to condone all, or even most, of the verdicts given by commonsense usage.
3 A Way Out for the Nonrelationist?
The denition (II) of causation by absence I proposed in the previous section had
two conditions on when the absence of an A-type event is to count as a cause of
b: the counterfactual conditionhad an A-type event occurred, b would not have
occurredand the condition that the absence be either abnormal or the violation of
some norm. I claimed that no respectable theory of the truth conditions of causal
claims ought to respect the second of these conditions; which leaves us with the rst,
counterfactual condition, that is, with (I). As weve already seen, an analysis of cau-
sation by absence that consisted solely in this condition would be far too inclusive.
It would render all of us causally responsible for indenitely many events happening
right now, all over the worldand doubtless beyond. One of the causes of Maites
drinking her coee in Mexico City just now was my failure to shoot her before she
had a chance to put the kettle on. One of the causes of your reading these words right
now is the absence of a lion from the room. And so on.
But, in principle at least, the nonrelationist who thinks there really is such a thing
as causation by absence is not required to think that theres that much causation by
absence, for she might be able to invent a more restrictive denition that will do a lot
better than what weve got so far. That is to say, maybe our original denition (I) can
be supplemented by some other clause that rules out the spurious cases of causation
by absence like the one I just gave, without ruling out the allegedly nonspurious cases
we want to keep, but which doesnt appeal to normative features of absences.
Well, on behalf of my nonrelationist opponent, I generously oer the following
denition. Its the best I can dobut not, as well see, good enough:
(III) The absence of an A-type event caused b if and only if:
(i) if an A-type event had occurred, b would not have occurred; and
(ii) an A-type event occurs at a world that is reasonably close to the actual world.
The denition discounts vast numbers of absences from being causes on the
grounds that worlds where the absent event occurs are very distant worlds. For in-
stance, my failure to shoot Maite no longer counts as a cause of her drinking her
coee because, Im happy to say, a world where I do shoot her is a very distant world
indeed.
One might even go so far as to claim that the denition gets all the cases right
(though Ill argue in a moment that it doesnt). We might try to claim, for instance,
that what makes Floras failure to water her neighbors orchids a cause of their death
but not the failure of her other neighbors is not, as we rst thought, the fact that
Floras failure, unlike those of the other neighbors, is abnormal; rather, the dierence
is that Flora is more strongly disposed to water the orchids than are her neighbors
as evidenced by the dierences in her and their past behavior. So a world where
Flora waters the orchids is closer to our own than is a world when another neighbor
does so. Zs failure to attend to his dogs eye infection, as opposed to our failure,
counts as a cause of its blindness because Z was in a position to do something about
it and we were not: Hence a world where Z takes his dog to the vet is reasonably
close to our own, and one where one of us takes the dog to the vet is not.
298 H. Beebee
The thought here, then, is that abnormality and violation of norms are not part
of the truth conditions of causation by absence claims. Rather, those considerations
inform our judgments about how reasonable it isthat is to say, epistemically
reasonableto expect the absent event to happen; and thus how reasonable it is
to suppose that the event in question happens at a reasonably nearby world. For
example, we generally assume that dog owners are disposed to behave in a way con-
ducive to the welfare of their pets, whereas people in general are not particularly
disposed to go out of their way to behave in a way conducive to the welfare of other
peoples pets. And we can construe this assumption as an assumption about closeness
of worlds: If Z is strongly disposed to keep a close eye on his dogs welfare and I
am not, then a world where Z manifests this strong disposition and takes the dog
to the vet is plausibly closer than a world where I, who have no disposition whatever
to do so, take Zs dog to the vet for him. Hence Zs omission is a cause of the dogs
blindness but mine is not.
Things seem to be looking up for the nonrelationist; but unfortunately I dont
think the situation is as rosy as Ive made it look. I have two objections to the sug-
gested denition. First, I think that commonsense intuition would still discriminate
between, say, Zs omission and mine even if we knew perfectly well that Z is a terri-
ble pet owner and consistently fails to look after his dog properly. Even if we think
that Z has no disposition whatever to take the dog to the vet, and hence that a world
where he does so is just as distant as a world where someone else does it for him, we
still think of Zs omission and nobody elses as a cause of the dogs blindness.
Similarly, we still judge a drug companys failure to warn consumers of foreseeable
side eects as a cause of their illness even if we know that the company is extremely
disreputable and rarely bothers to carry out the proper tests. I take it that common
sense simply doesnt endorse the view that if youre negligent enoughif your dis-
position to behave in accordance with norms is weak enoughyour negligence lit-
erally wont have any eects.
These observations support my initial claim that the norms are doing the work by
themselves, as it were, rather than merely informing our judgments about closeness
of worlds. Still, Im basing this conclusion on some disputable claims about com-
monsense intuition that you may not share, so this isnt a decisive objection.
The second objection, I think, is more telling: I can see no sensible way of speci-
fying what reasonably close amounts to. Its one thing to judge relative similarity
of worldsto judge that, say, a world where Flora waters her neighbors orchids is
closer than any world where someone else waters them. But its quite another thing
to try and impose a metric on this ordering, so that any A-world within a certain
distance from the actual world is close enough for the absence of A to count as a
cause, whereas any A-world outside that distance is too far away for the absence of A
to count as a cause.
Even if there were a way of specifying what reasonably close amounts to, there
are two further problems. One is that any choice of maximum distance is going to be
entirely arbitrary; and the question of which absences are causes and which are not
ought not to be decided by a piece of arbitrary stipulation. The other problem is that
whereas relative similarity of worlds might explain why one absence and not another
counts as a cause in any particular caseZs taking his dog to the vet happens in a
close enough world, say, but not my taking itI see no reason to suppose that the
same standards will apply across all cases. What counts as close enough for causa-
tion in one case might not count as close enough for causation in another.
All of these problems for the proposed denition (III), I think, have the same
source: There just isnt any objective feature that some absences have and others lack
in virtue of which some absences are causes and others are not. So any denition of
causation by absence that seeks to provide a principled distinction between absences
that are and are not causes is bound to fail: No such denition will succeed in carving
nature at its joints. If this is right, then the nonrelationist is going to have to concede
that there just is no principled reason to regard Floras failure to water the orchids as
a cause of their death but not to regard my failure to shoot Maite as a cause of her
coee-drinking. And this, I suggest, hardly puts nonrelationism ahead of relationism
when it comes to doing justice to commonsense intuitions.
In fact, there is another possible way out of the problem for the nonrelationist, and
that is to deny that the alleged distinction in our commonsense talk and thought be-
tween absences that are and are not causes really exists. I have been claiming that,
according to commonsense intuition, Floras failure to water her neighbors orchids
was a cause of their death, but the failure of other peoplepeople who neither know
nor care about the orchidswas not. But one might object that this is not really the
position endorsed by common sense. Rather, in ordinary circumstances we fail to
identify the failures of other people as causes of the orchids death, in the sense that it
does not occur to us to mention those failures part in the causal history of the death
of the orchids (since in ordinary conversational circumstances it would be inappro-
priate for us to do so). But to fail to mention that those other failures are causes is
not to hold that they are not causes.
No doubt there is some truth in this suggestion. Perhaps, if asked directly whether
the failures of other people to water the orchids was a cause of their death, at least
some people would say yespresumably because they would appreciate the similar-
ity between Floras failure, which they do count as a cause, and the failures of those
other people.
300 H. Beebee
I see no reason to suppose, however, that the point generalizes. The number of
possible events, or combinations of events, that are such that, had they occurred, the
orchids would not have died, is absolutely enormous. The plants would have sur-
vived, for example, if Floras neighbor had installed a sprinkler system that then got
activated accidentally while she was away, or if her roof had started leaking at a
point just above the orchids during a rainstorm, or if a cow had somehow entered her
house and icked a nearby glass of water onto the orchids with its tail, or. I do not
think that most people would happily accept that the failure of each of these events
to occur was equally a cause of the orchids death. Of course, this is an empirical
claim about what people are ordinarily inclined to judge. But I have not come across
any evidence to suggest that the claim is false.
4 Causation and Causal Explanation
I shall return to the commonsense distinction between absences that are and are
not causes in section 6. In this section and the next, however, I set this issue aside in
order to focus on another distinction: the distinction between causation and causal
explanation. In this section I defend the view that not all causal explanations are
reports of causation: The explanans of a causal explanation need not stand to the
explanandum as cause to eect. And in section 5 I defend the view that causal
explanations that involve facts about absences can be seen as explanations of just this
kind: We do not need absences as causes in order for facts about absences to be the
explananta of causal explanations.
In Causal Relations (1967), Davidson argues for a distinction between causa-
tion and causal explanation by concentrating on the logical form of causal state-
ments. He argues, using a version of the Slingshot example, that causation is a
relation between events rather than factsin other words, that the canonical form of
causal statements is c caused e, where c and e are events and caused is a two-place
relation, rather than E because C, where C and E are facts and because is a
sentential connective. Davidson reserves the E because C locution for causal ex-
planation. For Davidson, then, no causal explanations are themselves causal claims,
since they simply do not have the right logical form.
In opposition to Davidson, Mellor has long maintained that all causal explana-
tions are in fact causal claims. For Mellor, facts are the most basic kind of causal
relata, and the canonical form of causal statements is E because C.9 According
to Mellor, causation sometimes also relates eventsbut only sometimes. When its
true that the match lit because I struck it, its also true that the striking caused the
lighting; but when its true that, say, Kim has no children because she used contra-
ception, there is no corresponding true statement of the form c caused e.10
If absences are to gure in causal explanations without doing any causing, there
must be a distinction between causation and causal explanation: Some causal expla-
nations cannot be reports of causation. But this latter claim needs to be defended
against the following objection: How, one might ask, can a causal explanation be
genuinely causal if the explanans doesnt stand to the explanandum as cause to
eect? Or, as Mellor puts it, how can facts explain other facts causally without
causing them? (1995, p. 130). Well, here I want to appeal to Lewiss theory of ex-
planation. For Lewis, to explain an event is to provide some information about its
causal history (1986b, p. 217); and the causal history of an event, he says, is a
relational structure (ibid., p. 216). Not surprisingly, then, Lewiss account of expla-
nation is tailor-made to t the network model of causation. For present purposes, the
most important feature of Lewiss account of explanation is that it does not amount
to the view that every explanation involves picking out a cause (or some causes) of
an event; the way in which causal facts enter into an explanation can be more com-
plex than that. One can give information about an events causal history in all sorts
of other waysby saying, for instance, that certain events or kinds of event do not
gure in its causal history, or by saying that an event of such-and-such kind occurred,
rather than that some particular event occurred. The moral here, then, is that some-
thing can be the explanans of a causal explanation without itself being a cause of the
event cited in the explanandum.
For example, suppose that Lee Harvey Oswald really did shoot JFK. Then the
following three sentences are all true:
(1) Oswalds shot caused JFKs death.
(2) JFK died because Oswald shot him.
(3) JFK died because somebody shot him.
On Mellors view, all three are causal truths. On Lewiss view (as expressed in
Lewis 1973a, 1986b,c), only the rst is, strictly speaking, a causal truth; the second
and third are causal explanations rather than reports of causation. Its pretty obvious
what makes (2) a causal explanation: Each of the explanans and the explanandum
asserts that a particular event (Oswalds shot and JFKs death respectively) occurred,
and those two events are in fact causally related. The case of (3) is a little more
complicated: Here, the explanans does not assert that a particular event occurred, for
there is no event essentially describable as someones shooting JFK. Such an event
would be disjunctive: It would be the event of Lee Harvey Oswalds shooting or the
302 H. Beebee
man on the grassy knolls shooting or Jackie Kennedys shooting or . . . and so on.
And there is no more such an event than there is an event of my-birthday-party-or-
your-morning-bath.11 Rather, the explanans of (3) asserts that there was some event
or other that was a shooting of JFK by someone. Hence the explanans of (3) does
not stand to the explanandum as cause to eect. Nonetheless, (3) still counts as a
true causal explanation because, although it doesnt tell us which event caused JFKs
death, it tells us something about the deaths causal history, namely, that it included
a shooting by someone.
The notion of providing information about an events causal history is further
expanded by Jackson and Pettits account of program and process explanations.12
This account is designed to show how multiply realizable properties (like functional
and dispositional properties) can gure in true causal explanations without being
causally ecacious. For example, suppose that to be fragile is to be disposed to
break when dropped, and that a glass is dropped and duly breaks. We recognize that
it is the glasss molecular structure that causes it to break, and not its fragility; but we
do not want to conclude that the fragility cannot gure in a worthwhile explanation
of why it broke. (Likewise, we do not want the view that mental states are functional
states to preclude the possibility of explaining, say, actions in terms of beliefs and
desires.)
Jackson and Pettits solution to the problem is to distinguish between process
and program explanation. Process explanation is explanation in terms of actual
underlying physical processes: an explanation of why the glass broke in terms of
molecular structure, say, or an explanation of why I went to the shop in terms of
my neural processes. Program explanations, on the other hand, tell us not what the
actual underlying processes were, but rather that those processes satisfy a particu-
lar functional or dispositional or higher-order description. The presence of fragility
programs for the presence of a causally ecacious molecular property. Although
it is the actually present causally ecacious property, and not the programming
property, that gures in the causal history of the breakage, the program explanation
still tells us something about that causal historynamely, that it involves some e-
cacious property or other for which fragility programs.
As Jackson and Pettit say:
The process story tells us about how the history actually went: say that such and such partic-
ular decaying atoms were responsible for the radiation. A program account tells us about how
that history might have been . . . telling us for example that in any relevantly similar situation,
as in the original situation itself, the fact that some atoms are decaying means that there will be
a property realizedthat involving the decay of such and such particular atomswhich is
sucient in the circumstances to produce radiation. In the actual world it was this, that and
the other atom which decayed and led to the radiation but in possible worlds where their place
is taken by other atoms, the radiation still occurs.13
If we adopt Lewiss analysis of causal explanation, then, we can distinguish be-
tween causation and causal explanation and still be able to say why causal explana-
tions count as causal explanations; for such explanations can give information about
the causal history of the event to be explained even though the explanans does not
stand to the explanandum as cause to eect.
5 How Can Causally Inert Absences Explain Anything?
In the last section, I defended the view that facts can causally explain without being
causes of what they explain. In this section, I argue that facts about absences, omis-
sions, and failures do just that. The because locution isor at least sometimes
isan explanatory locution; moreover, as we have seen, because claims can be
truecan reveal information about causal historywithout the explanans standing
to the explanandum as cause to eect. So we can repudiate the claim that absences
can be causes and perfectly well grant that there are true causal explanations whose
explananta concern absences.
First, though, a point about the logical form of causation by absence claims. Much
of the time in our everyday talk, we speak as if absences, omissions, and failures are
things. We say that the void is deadly just as we would say that the Chrysler Building
is very tall: The sentence has the same subject-predicate form, and we count it as
true, even though there could be no object in the world picked out by the denite
description the void. Similarly, we say that Floras failure to water the orchids
caused their death just as we might, in dierent circumstances, say that Floras
throwing them on the re, or cutting them up, caused their death. Our everyday
causation-by-absence claims often have the logical form of a relational sentence
and we count them as trueeven though one of the singular terms anking the re-
lation does not refer to anything. Absences, omissions, and failures get assimilated to
the familiar ontological category of events even though they are not events.
So what? Well, suppose for a moment that there really is causation by absence, as
the nonrelationist claims. Even so, one ought to think that expressing causation by
absence facts using the c caused e locution is at best highly misleading. For such
sentences have a relational form, yet whatever it is that makes them true, it is not the
obtaining of any relation. The discussion in section 4 suggests an obvious paraphrase
of causation-by-absence claims into the because locution: A more ontologically
perspicuous way of saying that Floras failure to water the orchids caused their death
304 H. Beebee
is to say that the orchids diedor that their death occurredbecause Flora failed to
water them.
The relationist, on the other hand, cannot claim that the orchids died because
Flora failed to water them is a paraphrase of Floras failure caused the orchids
death, for, although we can perhaps paraphrase away the relational form of the
latter sentence, we cannot paraphrase away the fact that it is a causal claim; and
according to the relationist, absences do not cause anything. So if we wantas I
wantto hold that the sentence the orchids died because Flora failed to water
them is explanatory but not causal, it cannot be a paraphrase of the relational
sentence.
Still, I agree with the nonrelationist (as I have portrayed her) on this much: We
would do better to use the because locution than the caused locution when we
are talking about absences. The nonrelationist should think so because the because
locution is less misleading, whereas I think so because the because locution will
allow us to say something true, whereas the caused locution will not.
I say that common sense is just mistaken when it asserts that an absence or an
omission caused some event. Its not an especially bad mistake. Often we move be-
tween the E because C and c caused e locutions without going wrong: It doesnt
much matter whether I say the match lit because I struck it or instead my striking
the match caused it to light; or whether I say the crash occurred because there was
an avalanche or instead the avalanche caused the crash. Often causal explana-
tions go hand in hand with causal relations between events. Often, but not always.
When I say Floras failure to water the plants caused their death instead of the
death occurred because Flora failed to water them, I say something false instead of
something true. Its not surprising, nor a matter for particular concern, that we make
this error in our everyday talk; no serious harm is done. But I see no reason why,
from a philosophical perspective, we should not rule the move out of order.
It remains to be shown, of course, that because claims involving absences really
can give information about the causal history of the event to be explained without
the absence being a cause of that event. What sort of information does such an ex-
planation give us about the event whose occurrence we want to explain? Well, when
I say that the orchids death occurred because Flora failed to water them, you learn
something minimal about the deaths causal history: that it did not include an event
of Floras watering the orchids. But you also learn something about the causal struc-
ture of nearby worlds where Flora didnt fail to water the orchidsnamely, that the
causal processes that ensued at those worlds did not cause (or perhaps might not
have caused) the orchids death. And if you doubted the veracity of my explanation,
I could ll in more details about how the causal history of the world would have
gone had Flora watered the plants: Water would have been taken up by the roots,
sustained the cell structure, and so on. None of this, of course, is information about
what causal processes there were in the actual world; it is information about what
causal processes there are in the closest world(s) where the actually absent event
occurs. As with program explanation, the information provided about causal history
is modal information; and it is information that is explanatorily relevant to the
orchids death.
Similarly for Lewiss void. Suppose that as punishment for her negligence, Floras
neighbor casts her into the deadly void. Floras blood boils, the air is sucked from
her lungs, and so on. If, as I claim, there is no causation by absence, then the void
causes none of these unfortunate events. Strictly speaking, the void is not deadlyif
deadliness is the capacity to cause death. But Floras death is not thereby rendered
uncaused: There are plenty of positive events going on in her body that do cause
her death. Nor, on the account oered here, is the void explanatorily irrelevant to
Floras death. Its perfectly true to say that Floras blood boiled because there were
no forces presentforces that ordinarily keep us aliveto stop it. When we cite the
void in our explanation of Floras death, we describe how Floras causal history
would have gone had she not been cast into the void. We do not say what actually
caused her death; rather we point out that the sorts of events that would have caused
her to remain alive did not occur.
I see no reason, then, why explanations invoking facts about absences should not
be seen as genuine causal explanations, even though absences do not cause anything.
Of course, I have not shown that this relationist account of the role of absences in
causal explanations is any better than a nonrelationist account like Mellors, accord-
ing to which facts about absences causally explain in virtue of their being causes. I
merely hope to have shown that the relationist about causation is not committed to
denying that absences have a legitimate role to play in our explanatory talk.
6 Commonsense Judgments and the Pragmatics of Explanation
Its time to return to the issues discussed in the rst three sections of the essay. There
I argued that commonsense assertions and denials of causation by absence depend on
considerationsnormative features, for examplethat have no place in an account
of the metaphysics of causation. Since, from the perspective of metaphysics, we can
draw no relevant distinction between, for example, Floras failure to water the plants
and your failure to do so, or between Zs failure to attend to his dogs eye infection
and Zs neighbors cousins best friends failure to do so, the defender of causation by
306 H. Beebee
absence must concede that there is much more causation by absence about than we
ordinarily think there is. The denier of causation by absence, on the other hand, must
concede that there is much less causation by absence than we ordinarily think there
is: There is none at all, in fact.
As a denier of causation by absence, what do I say about the commonsense dis-
tinction between Floras failure to water the orchids caused their death (true) and
your failure to water the orchids caused their death (false)? I say, of course, that
both are false. But the corresponding explanatory claims (the orchids died because
Flora failed to water them; the orchids died because you failed to water them) are
both true. However, they do not, in most contexts, count as equally adequate ex-
planations. When we explain why the orchids died, we must, if our explanation is to
count as adequate as well as true, be sensitive to why the explanation was requested
in the rst place. In the context where my interlocutor is requesting information that
is relevant to the issue of whom to blame for the death of the orchids, it would be
highly misleading of me to say that they died because you didnt water them. The
truth of my utterance doesnt depend on the moral question of who is to blame; but
the adequacy of my explanation does, in this context, so depend. Similarly, its true
to say that I attended the seminar because I wasnt attacked by a hungry polar bear;
its just very hard to imagine a context within which someone who asked me why I
attended would be satised with that explanation.
Like the distinction between events and absences, the distinction between a true
explanationa true because statementand an adequate explanation is one that
common sense has a tendency to ignore. If you judge that the orchids died because
Flora didnt water them but not because you didnt water them yourself, or that I
attended the seminar because I didnt have anything better to do and not because I
was not attacked by a hungry polar bear, you mistake lack of explanatory salience
for falsity.
Of course, the believer in causation by absence can tell the same story. I do not
claim that my account is any more plausible than that of someone who thinks there
is causation by absence; I merely hope to have persuaded you that it is no less plau-
sible, and hence that it is no objection to the network model of causation that it
entails that there is no causation by absence.
The causal history of the world is a mass of causal processes: Events linked by a
vast and complex web of causal relations. In order that the causal history of the
world should look the way it does look, rather than some other way, there must have
been no extra events impinging on itfor those extra events would have had eects
that would have changed the causal history of the world in various ways. If Godzilla
had impinged upon the causal history of the world, that causal history would have
gone very dierently. We might even, if circumstances demanded it, want to explain
happenings in the world by citing Godzillas absence (though its hard to imagine
that we should ever want to do so). But I see no need to think of Godzillas lack of
impingement as a kind of causation.
Notes
1. See Steward (1997), chapter 7.
2. See for instance Lewis (1973a, 1986b), and Davidson (1967).
3. Lewis, chapter 10 in this volume, p. 281.
4. See Mellor (1987) and (1995, chapter 11). Steward (1997) also believes that the network model should be
abandoned.
5. See Lewis, chapter 10 in this volume, pp. 284285. Another proposal along similar lines can be found in
Dowe (1999).
6. See Lewis, chapter 10 in this volume, p. 284 and Dowe (1999).
7. The rst three examples are taken from Stapleton (1994), chapter 6, section 2.
8. See Hart and Honore (1985), p. 38.
9. Of course, relata isnt really the right word, since Mellor agrees with Davidson that because does
not pick out a causal relation.
10. See Mellor (1995), pp. 162165.
11. See Lewis (1986d), p. 267.
12. See, for example, Jackson and Pettit (1990).
13. Jackson and Pettit (1990), p. 117.
308 H. Beebee
12For Facts as Causes and Eects
D. H. Mellor
1 Introduction
Philosophers of causation, including those who deny that there is any, need to say
what they take it to be; just as atheists need to say what they disbelieve in. So, to
avoid begging the question for or against gods or causation, we must start not with
the debatable extensions but with the intensions of these concepts. That is, we must
say what we think the existence of gods or of causation entails. Only then can we say
whether and where they exist and what follows from that. In short, to reverse an
overrated adage, we must start not with the usethe unreective applicationof
these terms by believers but with the point of that use, that is, with their meaning.
But while semantics may have to have the rst word here, it can hardly have the
last, at least not in the case of causation. Only if what we must mean is demonstrably
necessary or impossible, as in the case of gods it might be, can that fact settle the
question. With causation, the question is more complex, since dierent people have
taken causation to mean too many vague, contentious, and conicting things, not
all of which we can have.1 We must therefore be prepared to nd that whatever in
the world, if anything, deserves to be called causation fails to live up to some of
our ideas of it. Take the pre-Humean idea that we can know a priori what causes
what. Now that we know that we cannot know this (with trivial exceptions, like the
cause of e, if any, causes e), most of us have dropped the idea rather than conclude
that there is no causation.
Similarly with other erstwhile connotations of causation, such as determinism. By
this I mean the idea that the presence and/or absence of causes somehow compels
that of their eects: in other words, that causes must in some strong sense be su-
cient and/or necessary for what they cause. This idea too has been challenged, by
convincing cases of seemingly indeterministic causation ranging from medicine (peo-
ples smoking causing them to get cancer) to microphysics (the triggering of atomic
explosions). But what makes these cases convincing, that is, what makes us want to
keep them in causations extension, is that they do meet other connotations of cau-
sation. For despite the indeterminism, these apparent causes still precede, explain,
give grounds for predicting, and provide means to, their apparent eects (Mellor
1995, chs. 57). That is why we want to call them causes.
In short, the conict of determinism with causations apparent extension arises
because it is not required by other, more important connotations. However, it is at
least consistent with them, unlike the transitivity that Lewis (1973a, in 1986a, p. 167)
and others have foisted on causation, by making it the ancestral of an otherwise
credible but nontransitive relation R. This contradicts most of causations other con-
notations, as such notorious causal chains as those from losing a nail to losing a
kingdom (e.g., Lowe 1980), or from a buttery wings apping to a tornado (Stewart
1989, p. 141), show. For while each member of such chains explains, gives grounds
for predicting, and is a means to the next one, their rst and last members stand
in none of these relations, which is why no one really thinks the rst causes the last.
So here too, but with far less excuse, a failure to t causations apparent exten-
sion arises from forgetting its principal connotations, that is, what we think follows
from saying that one thing causes another. Still, at least in this case the question is
simpleshould we identify causation with the nontransitive R or its trivially transi-
tive ancestral?even if the answer usually given is wrong.
Determinism is of course a harder and tastier nut to crack than transitivity, be-
ing related in more complex and interesting ways to causations other connotations.
Here, however, having tried to crack it in my (1995), I merely note that cracking such
nuts takes more than semantics, that is, more than specifying a relation and calling it
causation. We must also see if a real relation exists that both meets enough of the
specication to count as causation and has a credible extension, that is, links most if
not all of what we take to be causes and eects.2 And to do that we need not only a
semantics but a metaphysics, and in particular an ontology, to tell us rst what rela-
tions there are that might serve our turn, and then how much of our turn they can
serve.
It is of course no news that theories of causation need an ontology as well as a
semantics, Davidsons well-known theory being an obvious case in point. For on the
one hand in his (1967) he argues on semantic grounds that causation must relate
particulars, in the sense of entities that rst-order quantiers range over, of the kinds
that he calls events.3 And on the other he argues independently in his (1969)
that there must be particulars of these kinds, as well as of the less contentious kinds
exemplied by people, plants, and planets.
That Davidsons semantics for causation needs an ontology that includes events
is obvious, since without them he would have far too few particulars to provide
all singular causes and eects. Suppose, for example, you do something because you
decide to. Then if there are events, your decision (one particular) can cause your
action (another particular). But if there are not, and the only relevant particular is
you, the causation here must link not particulars but facts, namely the fact of your
deciding to do something and the fact of your doing it. To yield an extensionally
credible theory of causation, Davidsons semantics needs events.
310 D. H. Mellor
Similarly for those who argue, as I do (1995, chs. 911), that most causes and
eects are facts. We need a credible ontology of facts, defensible against among other
things the so-called slingshot arguments of Davidson (1967) and others, which pur-
port to show that if causation links any facts it links them all, which we know it does
not do. Here too, having argued the matter in my (1995), I shall simply assume that
these and other objections to facts can be met, and that neither facts nor events can
be ruled out independently of the theories of causation that invoke them. How, then,
are we to decide between these theories?
The answer, as in boxing and science, is that what cannot be settled by a knockout
must be settled on points. What matters is which theory of causation does best over-
all, when rated not only for its semantics and ontology but also, and mainly, for how
well it explains its subject matter, namely the connotations and apparent extension of
causation. That is the question here, as it is for other philosophical theories, and as
it is for scientic ones; and here, as there, the eventual answer may well alter some of
our initial semantic and ontological assumptions.
Compare, for example, an imaginary history of scientic theories of sh, that is,
theories ofto start withmiddle-sized self-propelled organisms living under water.
Theories of sh can of course not alter this ontology or semantics too much, on pain
of changing the subject: No theory of galaxies will ever be a theory of sh. Even so,
our current theory of sh has changed our initial piscatory assumptions quite a lot.
On it, many self-propelled underwater organisms are not sh: Some because they are
not animals; others, like whales, because they are mammals, that is, they work dif-
ferently and hence are only fools sh.4
Similarly, I say, with philosophical theories, including theories of causation. Some
apparent cases of causation (such as the phenomena of nonlocality in quantum phys-
ics) may, like whales, need excluding on theoretical grounds from what we end up
calling causation. New distinctions, such as that between causing and aecting (see
sec. 6), may also need drawing in our theory, just as shellsh may need distinguishing
from other underwater animals.
This view of scientic theories is familiar enough in the philosophy of science.
But it may still need selling to some philosophers of philosophy, for whom semantic
analysis remains the be-all and end-all of the subject. I should say therefore that
denying that this is what distinguishes philosophy from science is not to try and re-
duce the former to the latter. On the contrary, it is to try and restore to philosophy
the serious ontological theorizing that an unwarranted subservience to science and
semantics has inhibited for too long, and which is only now reviving as metaphysi-
cians shake o their scientic and semantic shackles.
For Facts as Causes and Eects 311
2 Facts and Particulars
How, then, do facts and particulars compare as singular causes and eects? Partic-
ulars may indeed have a semantic head start: Even to me the spark caused the re
sounds more natural, or at least more causal, than there was a re because there
was a spark. But that of course is not the end of the matter. It certainly does not
rule out causal truths of the form
(1) E because C,
where C and E are sentences and because is a sentential connective, as
opposed to
(2) c causes e,
where c and e are particulars and causes is a two-place predicate. (1) may still, as I
shall argue, win on points. But how?
First, to give (1) a chance, we must exclude its noncausal instances, such as those
used to give noncausal explanations. This we can do by at, by restricting it to in-
stances that are equivalent to the fact that C causes the fact that E. This restriction
begs no relevant questions, because it relies only on the uncontentious assumption
that E because C entails C and E, just as c causes e entails the existence of
c and e. For then, on the weak reading of fact given by the principle that, for all
sentences, statements or propositions P,
(3) P is true i it is a fact that P,
it follows that the facts that C and that E exist i C and E are true. And this
being so, we can simplify what follows by reading C and E not only as sen-
tences but also as shorthand for the fact that C and the fact that E, a usage in
which we can then rewrite (1) as
(1
0
) C causes E,
where C and E are facts in the weak sense given by (3). It is in this sense that (1
0
)
and hence (1) represent causes and eects as facts, just as (2) represents them as
particulars.
The contentious assumptions here are not semantic but ontological. Do the facts
that (1) and (1
0
) require, and the particulars that (2) requires, really exist? More pre-
cisely, do most if not all of the facts and particulars entailed by seemingly true
instances of (1), (1
0
), and (2) exist?5 This is not an easy or uncontentious thing to
show. Some philosophers, for example, still reject the Davidsonian events needed for
312 D. H. Mellor
my doing something because I decide to (an instance of (1)) to yield an instance of
(2), namely my decision causes my action. Here, however, if only to keep (2) in the
race, I shall take for granted the existence of the particular events, such as decisions,
actions, sparks, and res, that true instances of (2) need.
What of the facts that true instances of (1) and (1
0
) need? To these there are
several objections, notably the slingshot mentioned above, an argument whose va-
lidity means we can reject it only by rejecting one of its assumptions. These assump-
tions are that we cannot falsify a true E because C by replacing either
(i) C or E with a logical equivalent, or
(ii) a term referring to a particular with a coreferring term,
that is, with another term referring to the same particular. Of these two assumptions
the one I reject is the transparency assumption, (ii). I reject it because it implies, for
example, that if Tony Blair is the Prime Minister because he won the election is
true, so is Tony Blair is Tony Blair because he won the election, which is absurd.
For if Tony Blair is the Prime Minister, Tony Blair and the Prime Minister refer
to the same person. So for Tony Blair is Tony Blair because he won the election to
fail to follow, as it must, (ii) must fail in this case, no doubt because the fact that
Tony Blair is Tony Blair diers from the fact that he is Prime Minister.
But what makes these facts dier? What, in general, individuates the facts that I
say are causes or eects? My answer is the one that Davidson gave for events in his
(1969), namely that any events d and d
0
are identical i they have all the same causes
and eects, that is, i replacing d by d
0
in any c causes d or d causes e
would never change its truth value. Similarly, I say in my (1995, ch. 9.3), for facts
that are causes or eects. Any such facts D and D
0
are identical i they have all the
same causes and eects, that is, if replacing the sentence D by D
0
in any causal
D because C or E because D would never change its truth value.6
From this the falsity of (ii) follows at once. For as Tony Blairs being Tony Blair
does not have all the same causes and eects as his being the Prime Minister, Tony
Blair is the Prime Minister because he won the election can be, as it is, opaque, that
is, not transparent. So some instances of E because C are opaque.7 But then why
should they not be?
3 The Relata of Causation
The stock answer to this question relies on the seemingly innocuous assumption that
causation is a relation. It relies on this assumption because, for any (e.g., two-term)
relation R, we must be able to say transparently that it relates any entities a and b.
For if what R relates are a and b themselvesas opposed to aspects of or facts about
themthen to say so we must need only to refer to a and b and say that R relates
them: How we refer to a and b must be irrelevant to the truth of what we say. This
is why any simple relational statement of the form aRb must be transparent for a
and b. Yet, as we have just seen, some instances of E because C are opaque. Does
this not show that causation does not relate facts, and hence that causes and eects
are not facts but particulars?
No; but to see why not we must rst look not at (1) but at (1
0
), C causes E. This
is transparent for the facts C and E: Replacing the referring term C or E in (1
0
)
by any other term for the same fact will never change its truth value. That follows at
once from my causal criterion of identity for facts. For if replacing C by C
0
(or
E by E
0
) did change the truth value of C causes E, that criterion would
automatically make C and C
0
(or E and E
0
) dierent facts, so that C and C
0
(or
E and E
0
) would not refer to the same fact. But then this change in the truth
value of C causes E would not show it to be opaque.
All that the opacity in some E because C shows is that, within a true sentence
C or E, substituting coreferring terms for a particular may make that sentence
correspond to a dierent fact: as replacing the Prime Minister by Tony Blair
does in Tony Blair is the Prime Minister. But as this induces no opacity in sen-
tences of the form C causes E, it does not show that causation cannot relate facts.
Nor of course have we shown that causation cannot relate particulars: As causal
relata, both facts and particulars are still in the ring. But facts still seem to win on
points. For if there can be a re because there is a spark, there can also fail to be a
re because there is no spark. Similarly, if I can act because I decide to, I can also fail
to act because I do not decide to. In other words, just as E because C can be true
if E and C are true, so @E because @C can be true if E and C are
false. This poses no problem for facts, because the weak sense of fact given by our
principle (3) allows there to be negative facts.
For particulars, however, these cases do pose a problem. What particulars does
causation relate when there not being a spark causes there not to be a re? They
cannot be negative onesa nonspark and a nonreas there are demonstrably no
such entities. For suppose there is a long spark and a hot re: This entails that there is
a spark and a re, since something that is both a spark and long must be a spark, just
as something that is both a re and hot must be a re. In short, these entailments are,
as Davidson (1969) says, just cases of conjunctions entailing their conjuncts. But with
negative particulars, the entailments go the other way: If there is no spark, it follows
that there is no long spark, and no short one either. But this cannot be because a
314 D. H. Mellor
nonspark exists and is both long and short, since nothing can be that; any more than
a nonre can be (as it would have to be) both hot and cold, to make its existence
entail (as it would have to) that of both a hot nonre and a cold nonre.
How, then, without negative particulars, can causation relate particulars when the
cause or eect is that there is no particular of some kinda spark, a re, a decision,
an action? Here advocates of particular causes and eects face a dilemma, since they
must either deny that there is causation in these cases or nd some positive partic-
ulars for causation to relate. But the former restricts causations extension too much,
by ruling out too many obvious cases; and the latter is often made impossible by the
lack of suitable particulars. For positive particulars in most of these cases are clearly
either irrelevant (one need not decide not to act in order not to decide to act), in-
scrutable (what particulars does the nonexistence of sparks and res entail that can
make the former cause the latter?) or nonexistent, as in what Lewis (following Mar-
tin 1996) starts his second paper in this volume by calling the deadly void, that
would cause you to die in just a few minutes. It would suck the air from your lungs.
It would boil your blood. . . . Yet in all these cases, most of which lack any particu-
lar that is obviously capable of being the cause or eect, there is always an obviously
capable fact: namely, the fact that there is no particular of a suitable kind.
4 Negative Causes
Semantically, then, facts in my weak sense can provide causes and eects in far more
apparent cases of causation than particulars can, because they can be negative. But
semantics, I have argued, is not enough: we also need a credible ontology, which we
shall now see that negative facts may not provide. So as King Lear says to his silent
daughter Cordelia: How? Nothing can come of nothing. Speak again. And so,
being more willing than Cordelia to heave my heart into my mouth, I shall.
There is of course an innocuous reading of Lears maxim, namely that only actual
causes can have actual eects. This I accommodated in section 2 by requiring C
causes E to entail the existence of the facts C and E. However, this reading is not
strong enough to make the present point, since in my weak sense of fact all it
means is that E because C entails C and E. But when, as here, C or E is
a negative existential sentence, all this means is that no particular satises some de-
scription. That is the kind of absence that many philosophers nd it hard to credit
with any ecacy.
Yet why should absences not have eects? Perhaps it is because causes need to
come in kinds to enable laws of nature to x the kinds of eects they have: making
forces cause accelerations, sparks res, decisions actions, and so on. But then how
can nothing, being of no specic kind, have specic kinds of eects? The answer is of
course that absences can come in as many kinds as presences can. A lack of force
causes a lack of acceleration; an absence of sparks causes an absence of re; indeci-
sion causes inaction; and so on. These are all well-dened kinds of absence, with
well-dened kinds of eects; and the presence of nothing is merely the conjunction of
all such absences. Why then should we deny that absences in general, and a void in
particular, can have eects of specic kinds?
I can think of three sources of this denial. First, there is the contingency of most
and perhaps all causation, which may make the ecacy of even negative causes con-
tingent on the existence of something else. That I grant: but then this something
else may also be a negative fact, as when the absence of one force will cause an
object not to accelerate only in the absence of other forces. So although some nega-
tive facts will certainly lack their normal eects in Lewiss void, it does not follow
that they will have no eects, still less that the void itself will have none.
Next, there is the idea of causal ecacy as an intrinsic property of causes, in some
sense of intrinsic that would stop absences having such properties. But what can
this sense be? It cannot be the usual one, of failing to entail the existence of other
things. For in that sense the absence of a force, a spark, or a decision is as intrinsic
as its presence; and so therefore may its ecacy be. And yet in this sense, of course,
ecacy can never be intrinsic, since being a cause always means having, that is,
entailing the existence of, a distinct eect. So either way negative causes are no worse
o than positive ones.
This is why few if any philosophers now think that having eects is a property of
causes, as opposed to the relation between causes and eects discussed in section 3.
That relation may indeed depend on its relata havingor, if its relata are facts,
containingthe properties that make them instantiate the laws that x the kinds of
eects that given kinds of causes have; and I agree with most philosophers that cau-
sation does depend in this way on laws and hence on the properties of causes. But as
laws can and often do include negationsas in the law that bodies acted on by no
forces will not acceleratethis still gives us no reason to deny that negative facts can
be causes or eects.
What does give us a reason to deny this is the very idea of causation as a relation
in the sense of note 2, the reason being that real relations need real relata, which
negative facts cannot be. Let us see why not. The objection here is not that the
opacity of E because C stops it reporting a relation, since it does not, as we saw in
section 2: C causes E is as transparent for C and E as c causes e is for c and e.
The real objection to negative causes and eects is, as I hinted at the start of this
section, not semantic but ontological.
316 D. H. Mellor
To see the objection, let us look again at the principle I used in section 2 to gen-
erate all the facts I needed to supply my factual causes and eects, namely that for all
P, including negative existential ones,
(3) P is true i it is a fact that P.
Together with the uncontentious principle that, for all P,
P is true i P,
(3) entails that, for all P,
P i it is a fact that P.
This reading of fact thus makes it trivially true that, for example,
murder is wrong i it is a fact that murder is wrong;
Jim will probably win tonight i it is a fact that Jim will probably win tonight; and
quarks have spin i it is a fact that quarks have spin.
Yet it cannot follow from this that, for murder to be wrong, Jim to probably win
tonight, and quarks to have spin, the world must contain objective values, proba-
bilities, future-tensed facts, and theoretical entities. Theories of value, probability,
time, and the nature of scientic theory that deny this are not so easily refuted.
The sense of fact given by (3) is thus far too weak to show that causation relates
facts, since (3) does not tell us what in the world, if anything, by making P true,
makes P a fact in this weak sense.8 So in particular, (3) does not show that a negative
existential P is made true by the existence of something which a causal (or any
other) relation could link to anything else. And indeed it is obvious that it is not,
since by denition what makes any such P true is that no particular of some kind
exists. But then facts can no more supply all the relata that causation seems to need
than particulars can: Negative facts cannot have eects by being related to them. So
what does causation relate?
5 Nonrelational Causation
To answer this question we need to apply Ramseys heuristic maxim, that in such
stalemates the truth lies not in one of the two disputed views but in some third possi-
bility which has not yet been thought of, which we can only discover by rejecting
something assumed as obvious by both the disputants (Ramsey 1925, pp. 1112).
Here the assumption I propose to reject is that causation is a relation. It is not, and
the idea that it islike the idea that it is transitiveis a mere formal prejudice,
unwarranted by any of its substantive connotations.
To see this, we must note rst that, for causation to be a relation, statements of it,
like
(1
0
) C causes E, and
(2) c causes e,
must do more than meet the semantic criteria for relational statements by
(a) entailing C and E, and c and e, and
(b) being transparent for them;
they must also be made true by a relation holding between C and E or c and e. It is
this ontological assumption that I say the lack of relata in many apparent cases of
causation should make us reject, to prevent an implausible restriction on causations
extension.
To show how this can work, I start with my innocuous reading of King Lears
maxim that nothing can come of nothing, namely that only actual causes can have
actual eects. This, as we saw in section 2, requires E because C to entail C and
E, and hence the facts C and E, and c causes e to entail c and e. So for C and c
to be causes, and for E and e to be eects, these entities must exist, whether or not
they are the relata of any relation. Of course if causation does relate C and E, or c
and e, this will indeed entail that the entities it relates exist. But if they must exist in
any case to satisfy Lears maxim, that entailment is superuous. In other words, the
fact that causes and eects meet the merely semantic existential criterion (a) is no
reason to think that causation really is a relation, since they must meet this criterion
anyway.
Nor should we be impressed by the fact that causes and eects meet the transpar-
ency criterion (b). For as we saw in section 2, the transparency for C and E of C
causes E, and for c and e of c causes e, will follow from a causal criterion of
identity for C and E, and for c and e, whether causation is a relation or not.
If these criteria are neutral, others, notably criteria for what factual properties and
relations there are, imply that causation is not a relation.9 Take Shoemakers (1980)
criterion, that the factual properties that exist are those that combine to x the causal
powers of particularsas when having the properties of being steel and of being
sharp-edged combine to give a knife the power to cut. By that criterion causation
itself will obviously not be a property in this world or any other. Nor will causation
meet my (1997) criterion, that the factual properties and relations that exist are those
318 D. H. Mellor
that occur in laws of nature, since in my view no such law includes causation itself as
a property or relation.
Nor do substantive theories of what makes C causes E or c causes e true
support the idea of causation as a relation, as we shall now see. There are of course
many such theories, but for present purposes we may divide all the serious ones into
just two kinds: those that require causes and eects to instantiate laws of nature, and
those that require causation to imply something about an eects prospects without
its cause. Let us take these in turn.
Some theories make laws, and hence general causation (such as smoking causing
cancer), a relation between the properties and relations involved (Armstrong 1983).
But this, even if true (which I doubt), does not make the singular causation that
concerns us relational. For suppose it is a law that all Fs are, in certain circum-
stances, followed by Gs.10 How does this help to make (the fact that there is) an F
cause (there to be) a G? The answer, on almost any view, is that it does so by making
a certain non-truth-functional conditional true (or assertible),11 namely that, in any
relevant circumstances, if there were an F, it would be followed by a G. But as this
conditional entails neither its antecedent nor its consequent, it does not entail that
any such particulars exist. So whatever makes it true need notand clearly does not
entail the existence of the F and the G (or of the facts that there is an F and is a G)
that are thereby shown to be cause and eect. In short, far from law-based theories
of causation requiring it to be a relation, they imply that it is not.
Similarly for theories that invoke an eects prospects without its cause: These too
require causation only to entail conditionals, for example, that if in the circumstances
there were no F, there would also be no G.12 But as this does not entail the falsity
of either its antecedent or its consequent, its truthmaker also need notand again
clearly does notentail the existence of either the cause or the eect, and so cannot
be a relation between them.
So far so good for the idea that causes and eects need not be relata and can
therefore include negative facts. But not yet good enough, for even if causation itself
is not a relation, it may still require causes and eects to be related in space and time.
It may, for example, require that causes precede their eects, be contiguous to any
immediate eects, and be linked to others by dense sequences of intermediate causes
and eects. Does this not require causes and eects to be the relata, if not of causal
then at least of temporal and spatial relations, which we have seen that negative facts
cannot be?
To see why it does not require this, consider rst the temporal analogue of (1),
namely E after C, as in there was a re after there was a spark, I did it after
I decided to do it, and so on. There are also spatial analogues, as in there was a
re above where there was a spark and I did it where I decided to do it. In
short, spatial and temporal relations can relate facts as easily as particulars. But then
what about negative facts: How can the nonexistence of a spark or a re, or my not
deciding or not doing something, be the relata of such relations?
The answer is that they need not be. For all these negative facts have locations,
just as positive ones do: There is always a time and place at which there is no spark,
re, decision, action, and so on. So even if these are negative facts about things or
events, they are positive facts about times and places. But then whatever spacetime
relations causation entails need not relate these causes and eects but only their
spatiotemporal locations.13 Causes and eects themselves need not be relata at all.
6 Particular Causes and Eects
By denying that causation is a relation, and thus enabling negative facts to be causes
and eects, we can match causations apparent extension far better than any theory
that limits causes and eects to particulars. But on the other hand, we need not limit
causes and eects to facts. For since our causes and eects need not be truthmaking
facts, they need not be facts at all. As well as the nontruthmaking facts C and E
entailed by true instances of E because C, our causes and eects can include all the
particulars referred to in true instances of c causes e.
But then, since many truths of these forms come in pairs that clearly stand or fall
togetheras I acted because I decided to does with my decision caused my
actionwe must ask which comes rst: the factual causes and eects or the partic-
ular ones? In my (1995, ch. 11.3), I show how facts can come rst, by instances of E
because C entailing all true instances of c causes e, as follows: c causes e only if,
in a suitably restricted region of spacetime, for some F and G,
(i) there is a G because there is an F,
(ii) c is the F, and
(iii) e is the G.14
This is what makes facts the primary singular causes and eects, since what makes
particulars causes or eects is that facts about them are, not that a causal relation
holds between them.
Can the derivation also go the other way, from particulars to facts? Yes, but only
up to a point, and then only by making particulars too fact-like. Take Lewiss
(1973a) analysis of c causes e, using conditionals like if c did not occur (i.e.,
exist), e would not. This does yield some instances of E because C, but only of
320 D. H. Mellor
the form e exists because c does, which ts far too few cases, and not only because
it cannot cope with negative causes and eects. It also cannot cope with the many
cases in which c does not cause e, but only aects it, by causing an e to be G that
would exist even if it were not G:15 as when an injection aects a dentists drilling of
a tooth, by making it painless, without causing it, that is, without causing there to be
a drilling. Similarly, dierent facts about c may aect e in dierent ways, as when an
unwelcome content P of my speech (c) causes you (e) to believe P while its quiet
tone causes you to react calmly to that news.
All these cases yield natural instances of e is G because c is F, true for some
F and G and false for others. It is far harder to t them into the Procrustean form
of e exists because c does. That requires distinguishing the event of a drilling from
the event of its being painless, the event of my saying something from the event of
my saying it quietly, and so on. But this multiplies particulars beyond all necessity
and sense, destroying in particular Davidsons (1969) obviously correct explanation,
endorsed in section 3 above, of why, for example, a painless drilling must be a drill-
ing and a quiet speech a speech. The plain fact is that these so-called events are not
particulars at all but facts: the fact that there is a drilling, and the fact that it is
painless; that I say something, and that I say it quietly; and so on. To call these
entities events just to preserve the claim that causation relates particulars is to
evacuate that claim of almost all its content: a sure sign in philosophy, as in science,
of a degenerating research program.
7 The Positive Facts of Causation
I conclude that causation is not only not a relation, but that most causes and eects
are not particulars but facts, in the nontruthmaking sense of fact given by the
principle that P is true i it is a fact that P. But how then is causation embodied?
What must our world contain, besides whatever is needed to make C and E
true, to make E because C true?
We saw in section 5 that most if not all theories make causation entail one or two
conditionals, about an eects prospects with and without its cause. For E be-
cause C we may for present purposes write these conditionals as C LE and
@C L@E, where L is a connective that we may assume has something like
Lewiss (1973b) possible-world semantics. This is of course contentious, but that does
not matter here. Whatever their semantics, as these conditionals are contingent, and
incomplete truth-functions of C and E, at least one of them will need something
in this world besides C and E to make it true. And whatever that is will be what
embodies causation. All the other apparent consequences of Cs causing E are either
also causal (if causation is dense) or spatiotemporal (Cs preceding or being contig-
uous to E) or not ontological at all: for example, that C explains E, gives grounds
for predicting E, and provides a means to E. It is the truthmakers for conditionals
like C LE and @C L@E that add causation to a world of spatiotemporally
ordered facts.
And what those truthmakers are we already know from section 5. They are par-
ticulars having those properties discussed in my (1997) or Shoemaker (1980), namely
those that make particulars instantiate laws, or combine to give particulars their
causal powers. In either case, the way they do this is by being truthmakers for con-
ditionals: as when any thing of mass m, if acted on by any net force f , would ac-
celerate in the direction of f at f /m.16 And similarly for temperatures, pressures,
the strengths of elds across spacetime, and so on. It is particular things, events, and
spacetime points having properties like these that are the positive facts that embody
the worlds causation. Hence the causal limitations of the void. The problem is not
that a void cannot contain causes and eects: It can, since its negative facts can make
just as good causes and eects as positive facts. The problem is that it lacks the pos-
itive facts, the instances of properties, that such negative facts need in order to make
them causes and eects. It is our properties, not the voids, that make it deadly to us.
Notes
1. By causation here I mean singular causation, as in Freds smoking causing him to get cancer, rather
than general causation, as in smokings causing cancer.
2. By real relations (hereafter relations) I mean the relational counterparts of what Lewis (1983b) calls
natural properties, and I and Alex Oliver (1997) just call properties: namely, those that entail an
objective resemblance between the particulars that share them. Relations then are simply properties of
ordered pairs, triples, etc. of particulars, i.e., respects in which the ordered n-tuples that share them objec-
tively resemble each other. Whether these properties and relations are universals, resemblance classes of
particulars or tropes, or something else again, is immaterial for present purposes.
3. For clarity I too shall only use events to mean certain kinds of particulars. The common habit of
calling causes and eects events, whatever they may be, only causes confusion when, as here, we are
trying to say what causes and eects are.
4. Compare: Animals with the appearance of cats but reptilic internal structure . . . would not be cats; but
fools cats (Kripke 1972, p. 321).
5. I.e., exist in the past, present, or future of the actual world. In what follows, those who think that only
the present exists should read my exists as did, do, or will exist, whereas those who think that other
possible worlds exist should read it as is actual.
6. Note that these criteria, being only of actual and not of counterfactual identity, do not require events or
facts that have causes or eects to have them necessarily. Just because d and d
0
or D and D
0
are identical i
they have all the same actual causes and eects, it does not follow that the very same d (i.e., d
0
) or D (i.e.,
D
0
) could not have had any dierent causes and eects, and often it clearly could.
7. C and E need not state identities, or be necessary truths, for E because C to be opaque, as many
cases of mental causation show. For example, seeing Jim win his race may cause you to believe that he
322 D. H. Mellor
does so without causing you to believe that the youngest runner (which he is) does so. For other opaque
instances of E because C, see my (1995, ch. 12.45).
8. By a truthmaker for P I mean something whose existence entails P: see, e.g., Restall (1996). Dis-
putes about which P needs truthmakers need not concern us here, except that I take it for granted that if
P does, @P does not, since all it takes to make @P true is that the truthmaker for P does not
exist. (This distinction is not syntactic or semantic: However P and its negation may be represented in
thought or language, I take the positive onePto be the one that has a truthmaker.) This is why I
deny that negative existential instances of C and E in E because C have truthmakers that could be
the relata of a causal relation.
9. By factual properties and relations I mean nonevaluative and nonidentity ones whose instances are
not provable a priori, as opposed to such apparent properties and relations as being good, better than, or
identical to Fred, and such properties and relations of numbers as being prime or less than. If causation is a
relation at all, it is certainly factual.
10. By Fs and Gs I mean particulars satisfying the predicates F and G, whether or not F or G is
a property by my or Shoemakers criteria. In particular, to cover indeterministic laws, G may mean
having a certain chance of satisfying another predicate H.
11. This is just to cater for those who think these conditionals lack truth values; it makes no odds to the
ensuing argument.
12. Or, in indeterministic cases, that the chance of a G would be such-and-such, e.g., less than it would
have been with an F; again, it makes no odds to the argument.
13. For more details, see my (1998), chs. 8.6 and 10.4.
14. (i)(iii) are necessary, not sucient, since they only entail that c causes or aects e: causing it if being G
is essential to e, aecting it if not. See my (1995), ch. 12.
15. See note 14.
16. Provided f does not alter m, a proviso needed to enable so-called nkish properties to provide truth-
makers for conditionals: See Lewis (1997) and Mellor (2000).
13Preempting Preemption
David Coady
1 The Naive Counterfactual Analysis of Causation
Since David Lewis rst published Causation it has been generally assumed that
the most straightforward counterfactual analysis of causation (henceforth, the naive
analysis) cannot succeed, because of a class of counterexamples that, following
Lewiss nomenclature, have come to be called cases of preemption. Consequently,
there has been a debate about how to respond to this alleged fact.1 I argue that this
debate is premature, since the naive analysis has not been refuted or been shown to
conict with any compelling intuitions.
The naive analysis, which is widely believed to be refuted by the existence, or pos-
sible existence, of cases of preemption, claims that a singular causal proposition of
the form c caused e (where c and e are distinct events) is true i a counterfactual of
the form If c had not occurred, then e would not have occurred is true.2
Restricting the scope of the analysis to distinct events is necessary, if we are to
exclude certain noncausal relations. Although there is no standard denition of
distinctness in the literature, it seems clear that it is at least partly a mereologi-
cal notion: Two events are not distinct, if they have a common part. For example, it
is probably true that without World War I, none of the battles of that war would
have occurred, but this does not, I think, mean that World War I caused those
battles.
Counterfactual dependence can also be due to a logical rather than a causal (or
mereological) relation between events. It seems to be the case, for example, that
when one slams a door, one also closes it: Without the closing, the slamming would
not occur. This does not mean, however, that the closing causes the slamming (see
Kim 1973b, p. 571). Although there seems to be some sense in which these events are
not distinct, it is not obvious what that sense is. We could say that they are iden-
tical and therefore not distinct, but this seems hard to reconcile with the intuition
that they may dier causally.3 We could say that they have a common part, but it is
not clear that we are entitled to do so.4 It remains an important task to articulate a
satisfactory sense of distinctness; the standard objections to the naive analysis,
however, have nothing to do with this diculty.
While there is widespread agreement that the naive counterfactual analysis is a non-
starter, defenders of some kind of counterfactual analysis (as well as those who reject
the counterfactual approach altogether) are divided among themselves over whether
the naive analysis fails in both directions, that is, the truth of a causal proposition is
neither necessary nor sucient for the truth of its corresponding counterfactual, or
just one direction, that is, the truth of a causal proposition is sucient, but not
necessary, for the truth of its corresponding counterfactual.
Maudlin and Horwich are members of the former group. Their views about coun-
terfactuals entail that there are circumstances in which it would be correct to say that
if c had not occurred, neither would e, but wrong to say that c caused e, even though
c and e are distinct events.5 This could be the case if c is an eect of e rather than
a cause of it (Horwich 1989, pp. 169170; Maudlin 1994, pp. 128129). Maudlin
argues that it could also be the case if c and e are eects of a common cause (1994,
p. 129).6 The issue is not strictly whether backtracking or eects-of-a-common-cause
counterfactuals can be true. It would not be credible to suppose that they are never
true (or at least assertible). Rather, the issue is whether they are true in the same
sense as causal counterfactuals are.7
Defending the naive analysis does not necessarily mean defending Lewiss posi-
tion that causal counterfactuals are standard, whereas the others are nonstandard; it
would be enough to establish that the causal ones are dierent from the other two,
and that this dierence can be explicated without appeal to causal notions. Although
I believe this can be done, it is beyond the scope of the current essay, in which my
ambition is restricted to persuading you that the analysis does not fail in the other
direction.
David Lewis and many of those inuenced by his work agree with me that the
right kind of counterfactual proposition entails its corresponding causal proposition,
but they claim that there are realistic thought experiments about redundant causation
that show that the entailment doesnt go in the other direction. Without exception,
the extensive philosophical literature that has been generated from Lewiss discussion
of these cases accepts his claim that they provide counterexamples to the naive anal-
ysis. I disagree.
Lewis denes redundant causation in Postscript E of Causation:
Suppose we have two events c
1
and c
2
, and another event e distinct from both of them; and in
actuality all three occur; and if either one of c
1
and c
2
had occurred without the other, then
also e would have occurred; but if neither c
1
nor c
2
had occurred, then e would not have
occurred. Then I shall say that c
1
and c
2
are redundant causes of e. (Lewis 1986a, p. 193)8
As a defender of the naive counterfactual analysis, I must either deny that there are
redundant causes, or deny that redundant causes are ever genuine causes. I will do
the latter; that is, I will insist that in circumstances like those described by Lewis,
neither c
1
nor c
2
causes e, since e does not depend counterfactually on either of them.
I do, however, count the combination of c
1
and c
2
as a cause of e.9
326 D. Coady
The Lewiseans will have no problem with this approach for some cases. Lewis
claims there are two kinds of redundant causation; preemption and symmetrical
overdetermination.10 In a case of symmetrical overdetermination, c
1
and c
2
have an
equal claim to being causes of e. In such cases it may be unclear whether we should
say that c
1
and c
2
are each causes of e or that neither is, but it is at least clear that
there is no reason to say that one is whereas the other is not. Because it is unclear
what to say about such cases, Lewis correctly considers them to be poor test cases for
analyses of causation. This is not true of cases of preemption, however:
In a case of preemption, the redundant causes are not on a par. It seems clear that one of
them, the preempting cause, does the causing; while the other, the preempted alternative, waits
in reserve. The alternative is not a cause; though it could and would have been one, if it had
not been preempted. (Lewis 1986a, p. 199)
A preempting cause, then, is an event that is both a genuine cause, and a redundant
cause, of another eventa possibility that is inconsistent with the naive analyis.
There are cases in which it does seem clear that a redundant cause is a genuine
cause, but I believe this appearance is illusory and can be explained away. I will
argue that each apparent case of preemption should be redescribed in one of two
ways, with the correct way depending on context. In some contexts, I will argue, the
alleged preemption is in fact symmetrical overdetermination, in which case neither of
the redundant causes is a genuine cause. In other contexts, the alleged preempting
cause is in fact a straightforward (i.e., nonredundant) cause.
The context-dependence of some examples results from an ambiguity about the
identity of the eect (i.e., the e of the analysis). In other examples, it results
from an ambiguity about the identity of the cause (i.e., the c of the analysis).
2 Ambiguity about the Identity of the Eect
Lewis has recently oered the following example of what he calls commonplace
preemption: Billy and Suzy throw rocks at a bottle. Suzy throws rst or maybe
she throws harder. Her rock arrives rst. The bottle shatters. Billys rock passes in-
nocuously through the place where the bottle has just been (Lewis 2000, p. 184).11
Allegedly this is a counterexample to the naive analysis, because the bottle shattering
is caused by Suzys throw, despite not being counterfactually dependent on it.
2.1 A Common Suggestion
Lewis anticipates one objection to this description of the example, namely that
the shattering that would have occurred without Suzys throw would have been a
Preempting Preemption 327
dierent event from the one that did occur (Lewis 2000, p. 185; see also Lewis 1986a,
p. 197).12 This may seem plausible, since the counterfactual shattering clearly would
have occurred dierently from the actual one; specically, the counterfactual shat-
tering would have occurred later than the actual shattering.
2.2 The Standard Reply
Lewis responds to this suggestion by saying that it presupposes extremely stringent
criteria for the transworld identity of events: In the terminology popularized by
Lewis, it takes events to be modally fragile. An event is fragile, in this sense, if, or to
the extent that, it could not have occurred dierently.13 Lewis oers two arguments
against extreme fragility. First, it seems evident that we can engage in counterfactual
suppositions in which an actual event occurs very dierently; so, it seems, we ordi-
narily suppose that events are reasonably robust.14 Second, when combined with
counterfactual analyses of causation, extreme fragility would imply the falsehood
that events are caused by anything that even slightly aects them.15 Common sense
seems clear that there is a distinction between aecting an event and causing it, and
supposing events to be too modally fragile would rob us of this distinction.
2.3 A New Suggestion
I think the appropriate response to both of these arguments is to concede that not all
events are fragile and, specically, that there is an event that is a legitimate candidate
for being called the shattering that is suciently robust that, had Suzy not thrown,
it would still have occurred. I think there is another event, however, that is an
equally legitimate candidate for being called the shattering, which is suciently
fragile that it would not have occurred without Suzys throw. The fragile shatter-
ing is caused by Suzys throw, and is counterfactually dependent on it. The robust
shattering is not caused by Suzys throw, and is not counterfactually dependent on
it: Rather, the robust shattering is caused by the combination of Suzys and Billys
throws, and is counterfactually dependent on this combination. On neither inter-
pretation do we have a counterexample to the naive counterfactual analysis of
causation.16
Objection 1: There is obviously only one shattering in this example.
Reply: Lewiss own arguments against extreme fragility suggest that there are both
fragile and robust shatterings. Furthermore, a plurality of shatterings is not really a
problem, if we can also defend a sense in which there is only one. I will do so.
Although it is true, as Lewis points out in his rst argument against extreme fra-
gility, that we can make counterfactual suppositions according to which an actually
occurring event would have occurred at a quite dierent time, or in a quite dierent
328 D. Coady
manner, we can also often (and perhaps always) make a qualitatively identical sup-
position according to which it would have been a dierent event that occurs at that
time and in that manner. Lewis gives the following example of this phenomenon:
You can say: the performance should have been postponed until the singer was over his
laryngitis; then it would have been better. You can just as well say, and mean nothing dier-
ent: the performance should have been cancelled, and another, which would have been better,
scheduled to replace it. Theres no right answer to the question how fragile the performance
is. (Lewis 1986a, p. 197)
Or rather, there is more than one right answer, because there is a fragile perfor-
mance, which could not have occurred very dierently, and there is a robust perfor-
mance, which could. Both are permissible referents of the performance.
These performances are logically related; the fragile performance implies the ro-
bust performance.17 Elsewhere, Lewis has explicitly argued that some events imply
other events, and that a natural language description may refer to the implying event
in some contexts, and to the implied event in others (Lewis 1986d, p. 255). I will
follow Michael McDermott in calling this position the proigate theory of events
(McDermott 1995a, pp. 531532).18 Lewiss argument for it can be understood as a
natural development of his second argument against extreme fragility.19 That argu-
ment invited us to draw conclusions about the fragility of events from our considered
beliefs about what caused them. As Lewis recognizes, however, our considered causal
beliefs about what we are inclined pretheoretically to call a single event may vary,
depending, as he puts it, on how fragile we take the event to be (Lewis 1986a, p. 197;
2000, p. 186).20 Sometimes, this will depend on how detailed the expression used to
denote the event is.21 Sometimes the same natural language expression can be used
to refer, in dierent contexts, to either fragile or robust versions of the event.22
In either case, so long as we are committed to the idea that causation is a relation
between events, we need events that imply other events, in order to accommodate the
kind of ne-grained causal judgments we actually make. Without them we would
not be able to distinguish the question What caused the bottle to shatter? from the
question What caused it to shatter at the particular time it did?23 Surely these are
dierent questions; surely an adequate theory of causation should be able to distin-
guish them.
An adequate theory of causation, in this respect, is like an adequate theory of
probability, and, for a closely related reason, standard probability theory is also
committed to what seems pretheoretically to be too many events. I throw a single die,
for example, and get a two; I also get an even number. To speak of the outcome of
this probability experiment (without specifying a sample space) is to speak ambigu-
ously.24 One may be referring to the relatively fragile event of getting a two, or the
relatively robust one of getting an even number, even though we are strongly inclined
to say that these are the same event. If the proigate theory of events is committed to
invidious double-counting, so is standard probability theory.
It may be replied that this shows only that standard probability theory is wrong
to treat probabilities as attributes of events, and that a great deal of philosophical
literature is equally wrong to treat causation as a relation between events. In fact, it
is partly our ability to understand causal propositions of diering degrees and kinds
of specicity that has led some philosophers to deny that causation, at least in its
most fundamental sense, is a relation between events.25 It has been suggested that
it is instead a relation between aspects of events, or between events-in-contrast-to-
alternatives, or between facts (see respectively Paul 2000; Hitchcock 1996b; Bennett
1988).
I suggest that we should say instead that what is, in one sense, a single event, may,
in another sense, be many; there is one for each of the shifting judgments about its
causes that we want to make.26 It is a familiar kind of paradox: Freds house taken
as including the garage, and taken as not including the garage, have equal claim to
be his house. The claim had better be good enough, else he has no house. So Fred
has two houses. No! (Lewis 1999, p. 180).27 In fact, Lewis allows that there is some
sense in which Fred has two houses, but he also recognizes the need to articulate a
sense in which he has just one. To do this he borrows van Fraassens strategy for
dealing with unmade semantic decisions: We may say a statement is super-true i it is
true on all interpretations; it is super-false i it is false on all interpretations; it has a
super-truth-value-gap i it is true on some interpretations and false on others (van
Fraassen, 1966).
Super-truth occupies the role in a less than fully interpreted language that truth
simpliciter occupies in a fully interpreted language: It is that at which honest speakers
aim. So we are entitled to say that Fred has only one house, because it is super-true
that Fred has only one house, that is, it is true on each interpretation of house.
Similarly, I can agree with common sense that there is only one bottle shattering,
since it is super-true, that is, true on each interpretation of the shattering, that there
is only one of them.
Even if you are unconvinced by van Fraassens approach, I do not see any reason
to suppose that the proigate theory of events gives us any problems we didnt
already have. A new example of an old problem is not a new problem.
Objection 2: This position does not do justice to the intuition that Suzys throw
caused the shattering.
Reply: I think it does, because not only does it grant that there is a sense in which
this intuition is accurate, it is also consistent with the position that anyone who as-
330 D. Coady
serts that Suzys throw caused the shattering is asserting something true. In defense
of this claim I can do no better than quote Lewis on a closely related issue:
There is a rule of accommodation: what you say makes itself true, if at all possible, by creating
a context that selects the relevant features so as to make it true. Say that France is hexagonal,
and you thereby set the standards of precision low, and you speak the truth; say that France is
not hexagonal . . . and you set the standards high, and again you speak the truth. In parallel
fashion, I suggest that those philosophers who preach that origins are essential are absolutely
rightin the context of their own preaching. They make themselves right: their preaching
constitutes a context in which de re modality is governed by a way of representing . . . that
requires match of origins. (Lewis 1986e, pp. 251252)
In parallel fashion, I suggest that the philosophers who say that Suzys throw caused
the shattering are rightin the context created by what they say. Thus I think I do
justice to this intuition, despite there being a sense in which it is mistaken.
My position requires there to be contexts in which Suzys throw would not count
as a cause of the shattering. This seems to be true of contexts in which we have no
detailed interest in when the bottle shatters. I submit that in such contexts we should
deny that Suzys throw caused it, since it would have occurred anyway.
Objection 3: Suppose Billys throw had some slight inuence on the time or man-
ner of the shattering, thus causing a very fragile shattering. My account would then
imply, contrary to intuition, that there is a sense in which Billys throw is a cause of
the shattering.
Reply: I do not think this is a problem. There is no danger of losing the obvious
causal asymmetry between Billys throw and Suzys throw. Both may both be causes
of the shattering according to very stringent standards of fragility; neither will be
causes according to very robust standards of fragility; but according to ordinary
standards of fragility, Suzys throw will be a cause and Billys will not. I think this
does justice to the intuition that Billys throw is, and Suzys throw is not, a cause of
the shattering simpliciter.
Objection 4: It is not an essential feature of this example that there be any time
dierence between the shattering that occurred with Suzys throw and the shattering
that would have occurred without Suzys throw. Suppose, for example, that if not for
Suzys throw, Billy would have thrown harder or earlier than he actually does, in
which case the bottle would have shattered at the same time it actually did.28
Reply: Just as we can distinguish the question What causes the bottle to shatter?
from the question What causes the bottle to shatter at the particular time it does?
we can also distinguish both these questions from the question What causes the
bottle to shatter in the particular way it does? I submit that we can defend the
intuition that Suzys throw causes the bottle shattering in these circumstances, if
Suzys throw is a correct answer to the third question, even if it is not a correct
answer to the rst two.
Objection 5: Suppose the bottle was valuable, and that later in the day its owner
was disappointed to discover that it had shattered. The intuition that Suzys throw
caused the disappointment is, if not as strong as the intuition that it caused the shat-
tering, then almost as strong. There is no need, however, to suppose that there was
any dierence (temporal or otherwise) between the disappointment that occurred
with Suzys throw and the one that would have occurred without it.
Reply: I think there are two reasonable responses to this challenge. The rst tries
to respect the intuition that Suzys throw is, and Billys throw is not, a cause of the
disappointment. The second tries to explain the intuition away. Which response is
best will depend on context.
To respect the intuition that Suzys, but not Billys, throw is a cause of the disap-
pointment, I must identify the actual disappointment with the one that would have
occurred without Billys throw, but not with the one that would have occurred with-
out Suzys throw. I can justify doing so by appealing to the fact that the former
pair of events (actual disappointment, disappointment without Billys throw) have a
common history that the latter pair (actual disappointment, disappointment without
Suzys throw) lack.
Many Lewiseans will reject this approach because of the extrinsic identity condi-
tions of the postulated disappointment. Lewis and Maudlin have argued that such
extrinsic events should be treated with suspicion, because they appear to stand in
relations of noncausal counterfactual dependence (Lewis 1986d, pp. 263264; Maud-
lin 1994, p. 128). If Oedipus had not killed Laius, Jocasta would not have been wid-
owed; but does that mean that the killing caused the widowing? We do not want
to waste our time on such scheinprobleme as exactly when Jocasta became a widow
(instantaneously in some preferred reference frame? When she intersects the light
cone of the murder?) . . . If this is superluminal causation, it is not the sort to be of
any concern (Maudlin 1994, p. 128). Maudlin suggests that the way to avoid such
scheinprobleme is to restrict the analysis of causation to local physical events. We
need not be persuaded by this reasoning. We could instead say that the killing does
not cause the widowing because they are not distinct events, rather than because the
latter is not local.
Suppose you remain suspicious of extrinsic events, or are convinced that the dis-
appointment that appears to be caused by Suzys throw is not one of them. Another
approach to this kind of example is suggested by Lewis himself in a passage in which
he considers whether it is plausible to eliminate redundant causation altogether by
332 D. Coady
taking events to be extremely fragile (i.e., the common suggestion). He rejects one
possible objection:
Suppose we did follow this strategy wherever we could. Wouldnt we still have residual cases
of redundancy, in which it makes absolutely no dierence to the eect whether both of the re-
dundant causes occur or only one? Maybe so; but probably those residual cases would be mere
possibilities, far-fetched and contrary to the ways of this world. Then we could happily leave
them as spoils to the victor. For we could plausibly suggest that commonsense is misled: its
habits of thought are formed by a world where every little thing that happens spreads its little
traces far and wide, and nothing that happens thereafter is quite the same as it would have
been after a dierent past. (Lewis 1986a, pp. 197198)
If this were right, then we could assume that the absence of Suzys throw would have
made some dierence to the shattering, even though no such dierence is explicitly
mentioned in the story.
If Suzys throw were to make no dierence at all to the shattering, I think we are
entitled to dismiss any inclination we have to think of it as still being a cause as mis-
taken, this mistake being the product of a persistent (though in this case erroneous)
assumption that the shattering would have occurred somewhat dierently, if not
for it.
3 Ambiguity about the Identity of the Cause
The strategy I have outlined so far will not be applicable to every putative example
of preemption in which it appears that the absence of the so-called preempting cause
would have made little or no dierence to the time or manner of the eect. Lewis
attributes the following example of trumping preemption to van Fraassen:
The Sergeant and the Major are shouting orders at the soldiers. The soldiers know that in case
of conict, they must obey the superior ocer. But, as it happens, there is no conict. Sergeant
and Major simultaneously shout Advance!; the soldiers hear them both; the soldiers ad-
vance. Their advancing is redundantly caused: if the Sergeant had shouted Advance! and the
Major had been silent, or if the Major had shouted Advance! and the Sergeant had been
silent, the soldiers would still have advanced. But the redundancy is asymmetrical: since the
soldiers obey the superior ocer, they advance because the Major orders them to, not because
the Sergeant does. The Major preempts the Sergeant in causing them to advance. (Lewis
2000, p. 183)29
Rather than appealing to the fragility of the eect to respect the intuition that the
Majors order is, and the Sergeants order is not, a cause, I will appeal to the fragility
of the cause, that is, the fragility of the Majors order.
The proigate theory of events allows us to distinguish two orders shouted by the
Major; the more robust one is essentially an order, but only accidentally an order-to-
advance (it would still have occurred had the Major shouted a dierent order); the
more fragile one is essentially an order-to-advance. The robust one is not a cause of
the soldiers advance, since to suppose it not to have occurred is to suppose that the
Major shouted no order at all, in which case the Sergeants order would have been
eective.
But, I claim, the fragile version of the Majors order does cause the soldiers ad-
vance: If the fragile version of the Majors order had not occurred, the soldiers would
not have advanced. This seems to follow from familiar possible worlds semantics for
counterfactuals, according to which counterfactuals are true i the nearest possible
worlds in which the antecedent is true are also worlds in which the consequent is
true. The nearest worlds in which the Majors fragile order does not occur are not
worlds in which he shouts no order, but worlds in which he shouts a dierent order.
In such worlds the soldiers would not have advanced, since they would still have
obeyed the Major. This gives us a sense in which the Majors order caused the sol-
diers advance. By contrast, even an extremely fragile version of the Sergeants order
did not cause the soldiers advance, since that advance would have been unchanged
(or close enough) by him either shouting a dierent order or none at all.
Yet again my position is inspired by Lewis himself. He once argued that a robust
version of a greeting, but not a fragile version of the same greeting, may cause a
reply, because the nearest worlds in which the fragile one does not occur are worlds
in which the robust one still does (Lewis 1986d, p. 255).30 Lewis no longer seems to
accept this kind of reasoning. He now claims that to suppose a cause not to have
occurred is to suppose it to be completely and cleanly excised from history, leav-
ing behind no fragment or approximation of itself (Lewis 2000, p. 190). The naive
analysis will have diculty reconciling Lewiss new position with the intuition, which
I mentioned earlier, that the closing of a door may prevent someone from entering a
room whereas the slamming of the door (which implies the closing) does not. Lewiss
earlier position and mine accommodate the variability of our considered judgments
about the eects of what seems pretheoretically to be a single event.31
Objection 1: Worlds in which the Major shouts no order will, all else being equal,
be closer to the world of this example than worlds in which he shouts a dierent
order. This is because the closest worlds in which the Major shouts no order are, like
the world of the example, worlds in which the soldiers advance; whereas the closest
worlds in which the Major shouts a dierent order will, unlike the world of the ex-
ample, be worlds in which the soldiers do not advance.
334 D. Coady
Reply: The objection assumes, and I agree, that closeness is something like simi-
larity. Similarity, after the time of the antecedent, however, should play little or no
role in our calculations. Otherwise we would have to say that if Nixon had pressed
the button, nuclear war would have been avoided (see Lewis 1986a, pp. 4348). I am
inclined to agree with Jonathan Bennett that similarity at the time of the antecedent
is the only relevant similarity (Bennett 1984, pp. 7274). Surely, all else being equal,
a world in which the Major shouts just about any order is, at that time, more like a
world in which he shouts an order to advance than to a world in which he shouts no
order at all.
Objection 2: Suppose the Major very much wanted the soldiers to advance, but
hesitated to order them to do so, because he does not like giving orders.32 Arguably,
in this situation, the nearest possible worlds in which a fragile version of his order
does not occur are ones in which he gives no order, since it appears that the change
required for there to be a dierent order would be greater than the change required
for there to be no order.
Reply: In this situation the relevant version of the Majors orderthe one that
does the causingwould be not only essentially an order to advance, but also one
that was essentially given by someone who does not like giving orders. If it had not
occurred, a suciently fragile version of the soldiers advance would not have
occurred either.
4 Conclusion
Lewis recognizes that questions about the fragility of what seems pretheoretically
to be a single event need not have a determinate answer. This feature of his thought
has been largely ignored in the philosophical literature inuenced by his work on
causation.
One of the many places in which it plays a role is his discussion of what he calls a
class of cases distinguished by doubt as to whether they exhibit redundancy at all
(Lewis 1986a, p. 195). He asks us to suppose, for example, that neurons c1 and c2 are
set up in such a way that, if red, they will stimulate neuron e to re, and that neuron
e will re more vigorously if it is stimulated by both neurons, than if it is stimulated
by just one. In fact, both c1 and c2 re, doubly stimulating e, which res vigorously:
Is this vigorous ring of e a dierent event from the feeble ring that would have occurred if
either one of c1 and c2 had red alone? Then we have joint causation, in which the eect
depends counterfactually on each of the causes, and there is no redundancy. Or is it that
numerically the same ring would have occurred, despite a dierence in manner, with single
stimulation? (Lewis 1986a, p. 196)
Lewiss answer is that it is hard to say. He adds that the diculty cannot be
blamed on underdescription of the details (ibid.). He does not blame the diculty
on our epistemic faculties either; we do not have to resign ourselves to not knowing
whether this is a case of redundant causation. Rather, he suggests that there may be
no right answer (ibid., 197). Perhaps a better way to put his point, however, is to
say that there is no wrong answer. I think Lewis would agree with me, that we can
say that this is a case of redundant causation or we can say that it is a case of joint
causation without redundancy, preferably on dierent occasions, and either way we
will be right. As Lewis has pointed out, by saying something the truth of which
depends on context, we can often create a context that makes it true.
If I am right, putative cases of preemption are also cases distinguished by doubt as
to whether they exhibit redundancy. In the example under consideration the indeter-
minacy is between symmetrical overdetermination and joint causation, whereas the
relevant indeterminacy in putative cases of preemption is between symmetrical over-
determination and causation by the so-called preempting cause.33
It is a tribute to the profound inuence of David Lewiss work on the study of
metaphysics in general, and causation in particular, that there is such widespread
acceptance of the position that the most straightforward counterfactual analysis of
causation breaks down in cases of preemption. The real explanation, however, for
what I have argued is a mistake, are the ways in which putative cases of preemp-
tion are representedways that make it natural to conate dierent (though not
distinct) events. I have tried to show how natural language can conate logically
related events. If I am right, the neuron diagrams popularized by Lewis do so as
well.
I do not claim to have proven that my position is correct. If anyone continues to
believe that one or more cases of preemption are genuine, that is, that they really do
represent a situation in which an eect does not depend counterfactually on one of its
causes, then I do not know how to convince them. But I hope I have at least under-
mined the popular notion that it is obvious that the naive counterfactual analysis of
causation is refuted by cases in which a causal proposition is true, even though its
corresponding counterfactual is not.
Acknowledgments
I thank Hartry Field, without whom this article would have been very dierent in
time and manner (i.e., later and worse), and David Lewis, without whom not even a
very robust version of it could have occurred at all.
336 D. Coady
Notes
1. Participants in this debate, other than Lewis, include Tim Maudlin, Paul Horwich, Jonathan Bennett,
David Armstrong, Martin Bunzl, Douglas Ehring, Michael McDermott, Richard Miller, Murali Ram-
achandran, L. A. Paul, Paul Noordhof, Alex Byrne, Ned Hall, Jonathan Schaer, Bruce LeCatt, William
Goosens, Judea Pearl, and others. See the references for those who have published on this subject.
2. This is historically related to one of Humes analyses of causation. The form of the analyzing counter-
factual used by him, however, was if the rst object had not been, the second never had existed (An
Enquiry Concerning Human Understanding, section VII). I doubt that this dierence is philosophically sig-
nicant. The problems with his regularity analysis are another story about another body of literature. I
think general causal propositions, like Smoking causes lung cancer may also be analyzed in counter-
factual terms, but that too is another story about another body of literature.
3. Suppose the slamming, but not the closing, is caused by bad temper; and that the closing, but not the
slamming, causes someone to be unable to enter. (See Goldman 1970, p. 3.)
4. In fact, it is an implication of Lewiss theory of events that the closing and the slamming may well have
a common part, in virtue of the latter literally being a part of the former. Unfortunately Lewiss theory of
events has ontological commitments (i.e., unactual entities) that many, myself included, do not accept. (See
Lewis 1986d.)
5. Strictly speaking, Lewis (or at least a temporal part of him) would not disagree. He would claim, how-
ever, that such counterfactuals are true only on a nonstandard interpretation that is irrelevant to the anal-
ysis of causation. (See Lewis 1986c.)
6. I assume Horwich would agree.
7. For the sake of clarity I am interpreting backtracking counterfactual more narrowly than does Lewis,
from whom I take the terminology. A backtracker, for me, is a counterfactual that asserts that if an event
had not occurred, then neither would one of its causes; an eects-of-a-common-cause counterfactual is one
that asserts that if one of a pair of events with a common cause had not occurred, then neither would the
other; a causal counterfactual is one that asserts that if an event had not occurred, then neither would one
of its eects.
8. Lewis takes this to be redundant causation in its simplest form. He passes over other kinds of redundant
causation. One example of another alleged kind of redundant causation is probabilistic redundant cau-
sation. I think of putative cases of this phenomenon as situations in which there is a certain probability of
redundant causation simpliciter, rather than as situations exhibiting a special type of probabilistic redun-
dant causation. This will not be relevant to my discussion.
9. The mode of composition could be mereological or set-theoretical.
10. It is convenient to think of preemption as asymmetrical redundancy, and to think of symmetrical over-
determination as symmetrical redundancy. My position is that all redundancy is symmetrical.
11. According to Lewiss original taxonomy of preemption, this would be late preemption, since the pre-
empted process is cut o by the eect itself. Lewis no longer thinks that preemption requires the preempted
process to be cut (whether by the eect or otherwise); it is clear, however, that most, if not all, realistic
examples of it do involve cutting. This issue is not relevant to my approach.
12. Lewis calls this a common suggestion in Causation. His response to it remains unchanged in
Causation as Inuence.
13. I will say that an event is robust if, or to the extent that, it is not fragile. The distinction between fra-
gility and robustness can also be drawn in terms of essences; a fragile event has a rich essence, whereas a
robust event has an impoverished essence.
14. Of particular relevance to the example under consideration is our ability to engage in counterfactual
suppositions in which an actually occurring event is delayed.
15. Lewis attributes this argument against extreme fragility to Ken Kress. It has been extremely popu-
lar. Peter Menzies has argued against extreme fragility by appealing to the undesirability of treating an
explosion at the center of the sun, which produces a neutron, which in turn passes through a persons body
as she dies, as a cause of her death (Menzies 1989a, 649). Jonathan Bennett has cited in support of the
same conclusion the unacceptability of supposing that Platos wiping of Socrates brow just before Socrates
died was a cause of his death (Bennett 1988, p. 65).
16. It should become increasingly clear how my approach is intended to apply to other putative cases of
preemption in the literature.
17. For the sake of simplicity, I interpret the concept of implication between events more broadly than does
Lewis, from whom I take the terminology. I say that event e implies event f i, the proposition that e
occurs implies the proposition that f occurs; in other words, every possible world in which e occurs is a
world in which f occurs. The dierence between our concepts of implication between events is not relevant
to my argument.
18. McDermott does not explicitly dene proigacy, but I think my denition conforms closely to his
usage. McDermott does not approve of proigacy. He claims that Lewis seems to nd the proigate
theory of events an embarrassment. I do not believe this is true; but I do think that Lewiss discussion of
preemption fails to do justice to its implications.
19. I endorse the following argument, whether or not it is Lewiss.
20. Lewis and I disagree, however, about the extent of this indeterminacy. He claims, for example,
that Ned Kelly would have died a dierent death, if not for being hanged (Lewis 2000, p. 185). I think
it all depends on context. One can imagine Kelly consoling himself for his premature death with
the thought that it would have happened eventually anyway, thus taking his death to be extremely
robust.
21. Kims example of a slamming that is also a closing illustrates this phenomenon.
22. Lewis has argued that Johns tension may cause a relatively fragile greeting that is essentially too loud,
and not cause a relatively robust version of the same event that is only accidentally too loud. Furthermore,
he suggests that the expression Johns Hello may refer to the fragile version in some contexts, and the
robust one in others (Lewis 1986d, p. 255).
23. There do seem to be contexts in which they would be the same question. In particular, the former
question could be an abbreviation of the latter. I think, however, that the distinction Im speaking of is
clear, and of a familiar kind.
24. Just as, on my account, to speak of the bottle shattering, without contextual guidance, is to speak
ambiguously.
25. Another reason is that unchanging states, as well as what we would ordinarily think of as events seem,
at least sometimes, to cause and be caused. A third reason is that the failure of an event to occur can, at
least sometimes, appear to cause and be caused. I think unchanges and omissions should both be treated as
kinds of events, but that is yet another story about yet another body of literature.
26. There may also be some that we do not want to make for pragmatic reasons.
27. I use so many examples from Lewiss own work because I like them, but also because I want to per-
suade him that he should adopt my approach. So far I have been unsuccessful.
28. This has now become a case of early preemption, in which the preempted process is cut o by a branch
process diverging from the process that connects the preempting cause with the eect. This contrasts with
late preemption in which the preempted process is cut o by the eect itself. In this case the branch process
would be the ow of information by which Billy becomes aware that Suzy will throw. This causes him to
throw later or less forcefully than he otherwise would have.
29. Jonathan Schaer is responsible for the term trumping. Like Lewis, he has discussed this example
as a possible case of it (Schaer 2000b). This would qualify as a case of trumping, in Lewiss and
Schaers sense, only if no cutting (early or late) takes place in it. The alleged possibility of this is not rel-
evant to my approach. I think Schaers other putative example of trumping preemption should get similar
treatment.
338 D. Coady
30. Jonathan Bennett seems to be reasoning in the same way when he says that, on the counterfactual
analysis, an extremely fragile death of Socrates would not have caused Plato to grieve (Bennett 1988,
p. 65).
31. This is similar to the way in which, as we have seen, the proigate theory of events, when combined
with the naive analysis, accommodates the variability of our considered judgments about the causes of
what seems pretheoretically to be a single event.
32. Hartry Field has pointed out this possibility to me.
33. There may also be a legitimate (if nonstandard) sense in which putative cases of preemption exhibit
joint causation.
14Causes, Contrasts, and the Nontransitivity of Causation
Cei Maslen
1 Introduction
Whether one event causes another depends on the contrast situation with which the
alleged cause is compared. Occasionally this is made explicit. For example, a tooth-
paste company claims that regular brushing with their product will cause teeth to
become up to two shades whiter than brushing with another leading brand. Here,
the comparison not only helps to specify a range of shades of white, but also to
specify a contrast situation. Regular brushing with their product is not compared
with, say, irregular brushing with their product or not brushing at all, but with reg-
ular brushing with another leading brand.
More often the contrast situation is not made explicit, but is clear from the con-
text. Hence, in general, the truth and meaning of causal statements depend on the
context in which they occur. In section 2, I give a more complete formulation of
this claim, illustrate it with an example, and compare it with similar and super-
cially similar views. The theory is incomplete without a description of how contrast
events are xed by the context, which I supply in section 3. In section 4, I discuss
the context-dependence of counterfactuals. Of course, a major motivation for the
theory is the extent to which it can avoid obstacles that have defeated other theories
of causation, problems such as the nontransitivity of causation, preemption, causa-
tion by absences, and causation under indeterminism. In section 5, I explain how a
contrastive counterfactual account solves the rst of these problems: analyzing the
nontransitivity of causation.1
2 A Contrastive Counterfactual Account
Causal statements are systematically dependent on context.2 The meaning and truth
conditions of causal statements are dependent on contrast events that are seldom
explicitly stated, but are xed by conversational context and charity of interpreta-
tion. Occasionally confusion about contrast events leads to misunderstandings and
indeterminacy of meaning and truth value. This may be expected when causal state-
ments are taken out of context, for example, in some philosophical discussions.
This isnt to make causation a subjective matter. The causal structure of the world
is an objective, mind-independent three-place relation in the world between causes,
contrasts, and eects. (Compare this with discovering that motion is relative to frame
of reference. This is not to discover that motion a subjective matter.)
It would perhaps be ideal to study the properties of causal statements (e.g., non-
transitivity and context-dependence) without appealing to a specic formal analysis
of our concept of causation. All formal analyses of causation are controversial, and
complicated. The claims that causation is context dependent and that it is nontran-
sitive are partly independent of any specic analysis. However, in practice it is im-
possible to have detailed discussions of these aspects without settling on one analysis.
The few philosophers who discuss the context-dependence view can be classied into
all the major schools of thought on causation. Hitchcock3 incorporates it into a
probability-raising account. Field4 discusses probabilistic and nonprobabilistic ver-
sions of a regularity or law-based view. Holland5 presents a counterfactual account. I
support a counterfactual account also, and I concentrate on singular causal claims.
2.1 Account
1
For distinct events c, c
and e, c is a cause of e relative to contrast event c
i c and
e actually happened and if c had not happened, but contrast event c
had happened
instead, then e would not have happened.
This account takes events as the fundamental causal relata. Either Kims or Lewiss
denitions of events would serve this purpose.6 My hope is that causal sentences with
other kinds of relata (physical objects, processes, facts, properties, or event aspects)
can be reexpressed in terms of event causation, but I will not argue for this claim
here.7 I discuss counterfactuals briey in section 4 below. I think that our intuitive
understanding of this grammatical form is strong, so I do not commit myself to an
analysis here.8
The only restrictions I place on the contrast event is that it be compossible with the
absence of the cause and distinct from the absence of the cause.9 I also require that
the cause, eect, and contrast event be distinct events.10 This may seem hopelessly
liberal. What is to stop someone from claiming that brushing my teeth this morning,
in contrast to being hit by a meteorite, was a cause of my good humor, or that the
price of eggs being low, in contrast to the open re having a safety guard, was a
cause of the childs burn? The inanity of these examples arises from inappropriate
contrast events. Describing appropriateness of contrast is a dicult task, and I have
little more to say about it at this stage than that events are usually contrasted with
events that occur at a similar time, and that might have replaced them.11
Lets assume that any complex of events (the event that occurs just in case the
constituent events occur) is itself an event. It will be useful to dene a contrast situa-
tion as the complex of a contrast event and the event in which the absence of the
cause consisted.12 With this terminology, account
1
is:
342 C. Maslen
Event c is a cause of event e relative to a contrast situation i had the contrast
situation happened, then e would not have happened.
There may also be explicit and implicit alternatives to the eect. For example, a
friends opinions were a cause of my renting the video Annie Hall rather than the
video Mighty Aphrodite. Account
1
can be generalized in the following way to allow
for contrasts to the eect and for a range of implicit contrasts. (However, I will
mostly continue to work with the simpler account
1
.)
2.2 Account
2
Event c, relative to contrast situations fc
g, is a cause of event e, relative to

contrast situations fe
g, i had any events from fc
g occurred, then an event from

fe
g would have occurred.13

2.3 Illustrations
Here are two examples to illustrate the context-dependence of causation. First, sup-
pose I have recipes for three cookies: hazelnut cookies, walnut cookies, and pecan
cookies. I decide to make hazelnut cookies, and then I oer one to Stuart. Unfortu-
nately Stuart has a nut allergy; he is allergic to all nuts. He eats the cookie and
becomes ill. Was making the hazelnut cookies one of the causes of his illness?
Well, if I hadnt made the hazelnut cookies I would have made the pecan cookies,
because I have only three cookie recipes and I have no walnuts in the house. So in
one sense, making the hazelnut cookies was not a cause of his illness: Relative to
making pecan cookies, making hazelnut cookies was not a cause of his illness. How-
ever, relative to the alternative of making no cookies at all, making the hazelnut
cookies was a cause of his illness. Taken out of context, there is no correct answer
to the question of whether making hazelnut cookies was a cause of Stuarts illness.
Out of context, the only valid claims we can make are relative claims. However, the
conversational context and unspoken assumptions can make some alternatives
more salient than others. Our awareness that I have only nut cookie recipes makes
the alternatives of baking pecan cookies and walnut cookies salient alternatives
for us, and in the context of this paper the causal claim making hazelnut cookies
was a cause of his illness is naturally interpreted as meaning making hazelnut
cookies, relative to making pecan cookies or walnut cookies, was a cause of his
illness.
The second example, of an electric circuit, comes from Daniel Hausman, though
he uses it for a dierent purpose:
Causes, Contrasts, and Nontransitivity 343
The weak circuit and the strong circuit power a solenoid switch, which closes the bulb
circuit. If only the weak circuit is closed, the current through the solenoid is 4 amperes. If
only the strong circuit is closed, the current through the solenoid is 12 amperes. If both are
closed, about 16 amperes ow through the solenoid. It takes 6 amperes to activate the solenoid
switch, close the bulb circuit, and turn on the light bulb. Whether the weak circuit is closed
or not aects how much current is owing through the solenoid, but it has no inuence on
whether the light goes on.14
Suppose that on this occasion, I close both the strong and the weak circuit, a cur-
rent of 16 amperes ows through the coil, the solenoid switch closes, and the bulb
lights. Was the presence of a current of 16 amperes through the coil a cause of the
bulb lighting? The answer to this question is relative to the contrast event.
16 amperes owing through the coil, rather than 12 amperes owing through the
coil, was not a cause of the bulbs lighting.
16 amperes owing through the coil, rather than 4 amperes owing through the coil,
was a cause of the bulbs lighting.
16 amperes owing through the coil, rather than 0 amperes owing through the coil,
was a cause of the bulbs lighting.
The fact that there are cases like these, where the context-dependence is clear, gives
excellent support to the context-dependence view.
2.4 Other Views
At this point, it will be helpful to compare account
2
with similar and supercially
similar views.
Figure 14.1
344 C. Maslen
I am not proposing a revisionary account of causation. The realization that causa-
tion is context dependent and that this can lead to misunderstandings might prompt
us to propose a revisionary account of causation. We could suggest that in the future,
to avoid misunderstandings in precise or important uses of causal claims, we should
always state alternative events together with our causal claims. We could even fol-
low Bertrand Russell in suggesting that we abandon the concept of causation alto-
gether.15 Instead we would use only precise counterfactuals with our assumptions
spelled out. But the state of our causal concept does not warrant such an extreme
reaction.
The context-dependence that I have described in the concept of a cause is addi-
tional to the generally accepted context-dependence of the cause or the decisive
cause. I am concerned here only with what it takes to be a cause or one of the
causes of a given eect. The context-dependence of the concept of the cause is
more obvious. It might be suspected that there is only one source of context depen-
dence here and that is from the concept of the cause. It could be argued that we
confuse a cause with the cause when we read the examples, and this is why we
nd them convincing. However, this simply isnt borne out by our intuitions. We do
have an additional contextual element here.
The context-dependence of causation is distinct from the widely accepted view that
explanation is context dependent. Van Fraassen is the major defender of the latter
view: If, as Im inclined to agree, counterfactual language is proper to explana-
tion, we should conclude that explanation harbors a signicant degree of context-
dependence.16 He argues that the context-dependence of explanation takes the form
of determining both the salience of explanatory factors and also the contrast class.
However, those who claim that explanation is dependent on context assume that
causation is independent of context. For example, it seems to be implied by van
Fraassens claims that the propositions of the causal net are scientic propositions
and scientic propositions are not context-dependent in any essential way17 that
causal statements are (essentially) independent of context. I disagree; both explana-
tion and causation are context dependent.
Account
2
is similar in many ways to Lewiss recent causation as inuence view
and to event-feature views of causation, for example, Pauls aspect causation view.
I do not have space to discuss all the similarities and dierences here. However,
note that all three views imply that causal statements are strongly context dependent.
Lewis talks of a substantial range of not-too-distant alterations of the cause; what
constitutes a substantial range and what constitute not-too-distant alterations of the
cause presumably dier with context. Event-feature views require the context to de-
termine event features from event nominalizations.
3 Contrast Events and Context
The context of a statement is the circumstances in which it occurs. The truth of a
statement may depend on features of the context as well as on matters of fact. De-
pendence on matters of fact is contingency; dependence on features of context is
context-dependence. Following the two-stage scheme of Stalnaker, we may say that
an interpreted sentence together with a context determines a proposition, and a prop-
osition together with a possible world determines a truth value.18 Hence, which prop-
osition is expressed by an interpreted sentence may depend on the context in which it
occurs. A classic example of a context-dependent sentence is I went to the store; its
interpretation depends on the identity of the speaker, owing to the indexical I.19
In general, a large variety of contextual features may be required to interpret a
sentence on a particular occasion, for example, the intentions of the speaker, the
knowledge, beliefs, expectations, or interests of the speaker and his audience, other
speech acts that have been performed in the same context, the time of utterance, the
eects of the utterance, the truth-value of the proposition expressed, the seman-
tic relations between the proposition expressed and some others involved in some
way.20 Features of context such as speaker and time of utterance are almost always
obvious and readily observable. The set of contrast events is a theoretical feature of
context; hence, we owe a description of how contrast events connect with more ob-
vious features of context. Otherwise, when disputes arise it might seem as though the
metaphysician is magically summoning the set of contrast events or choosing the set
to t her case.
Suppose that the contrast event is what the speaker has in mind for replacing the
cause. That is, it is what the speaker has in mind to be dierent in an imagined
counterfactual situation where the cause is removed. Two objections to this immedi-
ately arise. Firstly, unphilosophical speakers may not have anything of the right
sort in mind when uttering causal statements. When asked what they had in mind,
many speakers may admit that they hadnt thought about it. However, usually these
speakers have no problem responding with a contrast event when prompted. These
responses give the intended interpretation of the original causal statement, in some
sense of intended. Second, as we are not mind readers, if the have in mind pic-
ture were the whole story then communication would be very haphazard.
The rest of the story is that which contrast set the speaker has in mind, if not
communicated directly, may be clear to the audience by the previous course of the
conversation, acknowledged assumptions, limitations, plans, and presuppositions.
General pragmatic principles play a large part in this. The set of contrast events is
what Lewis calls a component of conversational score.21 We have a tendency to
346 C. Maslen
interpret utterances generously or charitably, as true or probable, relevant, useful,
and informative. The set of contrast events is xed and developed through the course
of the conversation. Sometimes the set is left unsettled or vague until a dispute arises.
In a few cases this vagueness even leads to ambiguity in the causal statement.
Often there are physical limitations on the ways in which the cause could have
been omitted. This gives us a default set of contrast events. Consider the following
example. My opponent and I are both very competitive. We each would have been
happy to win, but we are both unhappy when we reach a draw at chess.
Reaching a draw at chess was a cause of us both being unhappy.
Presuming that we nish the chess game, there are only two ways in which the event
of reaching a draw could have been omitted: by my winning the game, or by my
opponent winning the game. Hence, there are two natural interpretations of the
causal statement: Reaching a draw at chess, in contrast to my winning the game,
was a cause of us both being unhappy; and Reaching a draw at chess, in contrast
to your winning the game, was a cause of us both being unhappy.
Either of these alternative outcomes would have left one of us happy. Hence,
reaching a draw can truly be called a cause of us both being unhappy. The natural
set of contrasts here are the events in which the absence of the cause would have
consisted. Given our assumption that we did actually nish the chess game, the ab-
sence could only have consisted in my winning the game, or my opponents winning
the game.
In some cases, there is only one probable way in which the cause would have been
omitted, and this is the natural or default contrast. Suppose that my opponent is a
much better chess player than I. If we hadnt reached a draw, he would almost cer-
tainly have won. Given our assumptions, the contrast situation here is naturally lim-
ited to one alternative. The following causal claim is probably true in that context.
Reaching a draw on the chess game was a cause of my opponent being unhappy.
The default contrast is always overruled by what the speaker has in mind or in-
tends as the contrast. For instance, suppose I mistakenly believe that I am the better
player. I could have in mind a contrast with the case in which I had won the game
and deny the above causal claim. (If my opponent does not realize I have this mis-
taken belief, then this will probably lead to misunderstanding.) Even if I recognize
that my opponent would have won the game if we hadnt drawn, I could make it
clear that Im contrasting with a wider set, and then deny the causal statement above.
I could contrast with the set {playing the chess game and my winning, playing the
chess game and his winning, not playing the chess game at all}.22
4 The Context-dependence of Counterfactuals
The counterfactual analysis of causation is one of the most popular approaches to
analyzing causation. Hence, it is strange that counterfactuals are widely acknowl-
edged to be context dependent, while causation is not widely acknowledged to be so.
The context-dependence of counterfactuals has been observed and discussed in
most major works on counterfactuals.23 Here are some classic examples exhibiting
context-dependence:
If Caesar had been in command in Korea he would have used the atom bomb.
If Caesar had been in command in Korea he would have used catapults. (Example
due to Quine 1960, p. 222)
If this were gold it would be malleable.
If this were gold then some gold things would not be malleable. (Example due to
Chisholm 1955, p. 103)
If New York City were in Georgia then New York City would be in the South.
If New York City were in Georgia then Georgia would not be entirely in the South.
(Example due to Goodman 1947, p. 121)
All of these counterfactuals are context dependent. Consider just the rst pair. Each
statement can be reasonably asserted in the same situation. (We presuppose that
Caesar was ruthless, ambitious, and indierent to higher authority.) Yet the rst
statement implies that the second statement is false. (We presuppose that it is possi-
ble that Caesar was in command in Korea and it is not possible that he uses both
catapults and atom bombs.) The second statement cannot be both true and false, so
it must express at least two dierent propositions depending on factors other than the
background facts. It is context dependent.
Acknowledging the context-dependence of counterfactuals can help us to under-
stand the context-dependence of causation. I do not wish to commit myself to one
analysis of counterfactuals here. However, let me mention how two successful and
sophisticated analyses of counterfactuals account for their context-dependence. One
formulation of Lewiss analysis of counterfactuals is as follows: A counterfactual If
it were that A, then it would be that C is (non-vacuously) true if and only if some
(accessible) world where both A and C are true is more similar to our actual world,
overall, than is any world where A is true but C is false.24 On this account, the
context-dependence of counterfactuals arises because our judgments of overall simi-
348 C. Maslen
larity of possible worlds depend on context. The delineating parameter for the
vagueness of counterfactuals is the comparative similarity relation itself: the system
of spheres, comparative similarity system, selection function, or whatever other en-
tity we use to carry information about the comparative similarity of worlds.25
The tacit premise view of counterfactuals is presented by Chisholm and also by
Tichy.26 On this view, a counterfactual is true i its antecedent together with its tacit
premises logically entail its consequent, and the tacit premises are true. The tacit
premises are simply those assumptions that have been presupposed in the conversa-
tion or those that the speaker has in mind on the occasion of utterance. The context-
dependence is obvious on this view.
It is interesting that Lewis observes in his original paper on causation that The
vagueness of similarity does infect causation, and no correct analysis can deny it.27
His original theory already accounts for some context-dependence, admittedly in a
subtle way.
5 The Nontransitivity of Causation
Is causation transitive? That is, is it true for all events a, b, and c that if a is a cause of
b and b is a cause of c then a is a cause of c? Some causal chains are clearly transitive.
Suppose that the lightning is a cause of the burning of the house, and the burning of
the house is a cause of the roasting of the pig. (Suppose that the pig was trapped in
the house.) Then surely the lightning is also a cause of the roasting of the pig. But is
this true for all events a, b, and c?
In the past, the transitivity of causation was commonly assumed in the literature
without argument.28 But more recently a host of ingenious examples have been pre-
sented as counterexamples to the transitivity of causation.29 Before discussing these
examples, lets briey consider one argument for the transitivity of causation.
The only argument that I have found in the literature for the claim that causation
is transitive comes from Ned Hall. Hall argues that rejecting transitivity seems intui-
tively wrong: it certainly goes against one of the ways in which we commonly justify
causal claims. That is, we often claim that one event is a cause of another precisely
because it causes an intermediate, which then causes another intermediate, . . . which
then causes the eect in question. Are we to believe that any such justication is
fundamentally misguided?30
That we do sometimes justify causal claims this way is an important considera-
tion. I agree with Hall that it is common practice to refer to intermediates in a causal
chain in justifying causal claims. This seems to apply across many dierent elds of
application of the singular causal concept: history, science, law, and ethics. I dont
think that this practice is fundamentally misguided, but it may be a rule of thumb
that should be supplemented with restrictive guidelines. If we are to reject tran-
sitivity, we have a pressing need for a general rule for distinguishing the cases in
which transitivity holds from the cases in which it fails, and an explanation of why
transitivity sometimes fails. Hall agrees with me here. He issues a challenge: Cau-
sation not transitive? Then explain under what circumstances we are right to follow
our common practice of justifying the claim that c causes e by pointing to causal
intermediates.31 I take up this challenge in the next section after presenting the
counterexamples.
The alleged counterexamples to transitivity are diverse; I will describe three di-
cult and representative cases: bomb, birthday, and purple re. The rst comes
from Field.32 Suppose that I place a bomb outside your door and light the fuse.
Fortunately your friend nds it and defuses it before it explodes. The following three
statements seem to be true, and thus seem to show that causation is nontransitive.
(1a) My placing the bomb outside the door is a cause of your friends nding it.
(1b) Your friends nding the bomb is a cause of your survival.
(1c) My placing the bomb outside the door is not a cause of your survival.
On a more cheerful note, suppose that I intend to buy you a birthday present, but
when the time comes I forget. Fortunately, you remind me and I buy you a birthday
present after all. The following three statements seem to be true, and thus seem to
show that causation is nontransitive.
(2a) My forgetting your birthday is a cause of your reminding me.
(2b) Your reminding me is a cause of my buying you a birthday present.
(2c) My forgetting your birthday is not a cause of my buying you a birthday
present.
Finally, consider Ehrings purple re example.33 Davidson puts some potassium
salts into a hot re. Because potassium compounds give a purple ame when heated,
the ame changes to a purple color but otherwise stays the same. Next, the heat of
the re causes some ammable material to ignite. Soon the whole place is ablaze,
and Elvis, sleeping upstairs, dies of smoke inhalation. The following three statements
seem to be true and again seem to show that causation is nontransitive.
(3a) Davidsons putting potassium salts in the replace is a cause of the purple
re.
350 C. Maslen
(3b) The purple re is a cause of Elviss death.
(3c) Davidsons putting potassium salts in the replace is not a cause of Elviss
death.
6 A Contrast Analysis of the Counterexamples
Transitivity is dened only for binary relations, but, on the context-dependence view,
causation is not a binary relation. (It is either a three-place relation between a cause,
a context, and an eect, or a four-place relation between a cause, two sets of con-
trast events, and an eect, depending on how you count it.) We will discuss a related
property, the variable-context transitivity of the three-place causal relation. The
three-place causal relation has variable-context transitivity just in case for all events
e
1
, e
2
, e
3
, and for all contexts c
1
, c
2
, c
3
, if e
1
causes e
2
in context c
1
, and e
2
causes e
3
in context c
2
, then e
1
causes e
3
in context c
3
. That is, the causal relation has variable-
context transitivity just in case it appears transitive no matter how you change the
context.
Lets analyze the bomb example. The example can be interpreted in many ways
depending on the contrast events assumed in statements (1a), (1b), and (1c). Two
natural contrasts with the cause in (1a) are the contrast with my placing nothing
outside the door and the contrast with my carefully concealing the bomb outside the
door. My placing the bomb outside the door, in contrast to my placing nothing out-
side the door, is a cause of your friends nding the bomb, because if I had placed
nothing outside the door your friend wouldnt have found the bomb. My placing the
bomb outside the door, in contrast to my carefully concealing the bomb outside the
door, is a cause of your friends nding the bomb, because if I had carefully con-
cealed the bomb outside the door your friend would not have found it (let us say).
Some interpretations of the example do not yield counterexamples to transitivity.
Statement (4c) of the following is false: It seems plausible that my placing the bomb
outside the door, in contrast to my carefully concealing the bomb outside the door, is
a cause of your survival. So the following causal chain is transitive.
(4a) My placing the bomb outside the door (vs. carefully concealing it) is a cause
of your friends nding it (vs. overlooking it).
(4b) Your friends nding the bomb (vs. overlooking it) is a cause of your survival
(vs. death).
(4c) My placing the bomb outside the door (vs. carefully concealing it) is not a
cause of your survival (vs. death).
However, with other natural contrasts, we do have a nontransitive causal chain.
For example, all three of the following statements seem true.
(5a) My placing the bomb outside the door (vs. placing nothing outside the door)
is a cause of your friends nding the bomb (vs. nding nothing).
(5b) Your friends nding the bomb (vs. nding nothing) is a cause of your
survival (vs. death).
(5c) My placing the bomb outside the door (vs. placing nothing outside the door)
is not a cause of your survival (vs. death).
We can develop a sucient condition for a causal chain to be transitive from a
special case of inference by transitivity of counterfactuals:
(T) j Lw, w Lj, j Lc 8 w Lc.34
Consider a general causal chain:
(6a) a, with contrast situation c
1
is a cause of b, with contrast situation d
1
.
(6b) b, with contrast situation c
2
, is a cause of e, with contrast situation d
2
.
(6c) a, with contrast situation c
3
, is a cause of e, with contrast situation d
3
.
Suppose that (C1) c
1
c
3
, d
1
c
2
, and d
2
d
3
(in other words, events a, b, and
e have the same contrasts throughout) and that (C2) if c
2
had occurred then c
1
would have to have occurred (a backtracking counterfactual).35 Given (T), and
account
2
, these conditions together, (C1) and (C2), are sucient for a causal chain
to be transitive.
This sucient condition can help us to understand what is going on in example (5)
to yield nontransitivity. Example (5) passes (C1), but fails (C2). It is not true that had
your friend found nothing outside the door then I would have to have placed nothing
outside the door. Had your friend found nothing outside the door, it might have been
because I had carefully concealed the bomb so that she overlooked it. There is a
sense in which the contrast situations in the example are incompatible with each
other, and this incompatibility leads to nontransitivity.
Lets return briey to the other counterexamples. Here is one natural interpreta-
tion of the birthday example:
(7a) My forgetting your birthday (vs. my remembering your birthday) is a cause
of your reminding me (vs. your forgetting to remind me, or our both forgetting
your birthday).
352 C. Maslen
(7b) Your reminding me (vs. your forgetting to remind me, or our both forgetting
your birthday) is a cause of my buying you a birthday present (vs. buying you
nothing).
(7c) My forgetting your birthday (vs. my remembering your birthday) is not a
cause of my buying you a birthday present (vs. buying you nothing).
Notice that condition (C2) fails. It is not true that if you had forgotten to remind me
about your birthday then I would have to have remembered by myself. If you had
forgotten to remind me about your birthday then I might have forgotten too. Also,
it is not true that if we had both forgotten your birthday then I would have to have
remembered your birthday. On the contrary, if we had both forgotten your birthday
then I would not have remembered your birthday.
Here is the example with some other contrast situations. Statement (8b) is false: If
I had remembered your birthday by myself then I would have bought you something
(let us say). So the example exhibits transitivity. Furthermore, conditions (C1) and
(C2) are satised. The same events always have the same contrasts throughout, and
if I had remembered your birthday by myself then I would have remembered your
birthday.
(8a) My forgetting your birthday (vs. my remembering your birthday) is a cause
of your reminding me (vs. my remembering your birthday by myself ).
(8b) Your reminding me (vs. my remembering your birthday by myself ) is a cause
of my buying you a birthday present (vs. buying nothing).
(8c) My forgetting your birthday (vs. my remembering your birthday) is not a
cause of my buying you a birthday present (vs. buying nothing).
Finally, here is the purple re example with some natural contrasts spelled out:
(9a) Davidsons putting potassium salts in the replace (vs. Davidsons putting
nothing in the replace) is a cause of the purple re (vs. a yellow re).
(9b) The purple re (vs. no re) is a cause of Elviss death (vs. Elviss survival).
(9c) Davidsons putting potassium salts in the replace (vs. Davidsons putting
nothing in the replace) is not a cause of Elviss death (vs. Elviss survival).
This case is nontransitive because it fails both conditions (C1) and (C2). The event
of the purple re is contrasted with a yellow re in (9a) but contrasted with no re in
(9b). Furthermore, it is not true that if there had been no re then Davidson would
have to have put nothing in the replace. If there had been no re, he might have
decided to put potassium salts in the replace anyway.
7 A Fine-grained Event Analysis of the Counterexamples
Hausman describes how allowing for ne-grained events enables us to explain exam-
ples of this sort without rejecting the transitivity of causation.36 The example is not
of the right form to be a counterexample to transitivity, because of the equivocation
about which event is being referred to by the phrase the purple re. In terms of
Kimian events, the purple re could either designate the triple [the re, being pur-
ple, time] or the triple [the re, being a re, time]. In terms of Lewisian events, the
phrase could either designate a strong event, the purple re, which is essentially
purple, or designate a weak event, the re, which is only accidentally purple. It is
the strong event of the purple re (or the triple [the purple re, being purple, time])
that is caused by Davidsons action, and it is the weak event of the re (or the triple
[the purple re, being a re, time]) that is a cause of Elviss death.
The ne-grained event analysis of the purple re example is similar to the analy-
sis in terms of implicit contrasts that I gave above. Whereas Hausman locates
context-dependence in the reference of the phrase the purple re, I locate the
context-dependence in the interpretation of the whole sentence, Davidsons putting
potassium salts in the replace is a cause of the purple re. However, if the phrase
the purple re vs. a yellow re designates a strong event of the purple re and the
phrase the purple re vs. no re designates a weak event of a purple re, which
they plausibly do, then it can be shown that the two analyses of this example are
equivalent.
However, I do not see how the ne-grained event analysis can explain in a similar
fashion the nontransitivity of the bomb example. Perhaps we can locate an equivo-
cation in the event referred to by the phrase your friends nding the bomb by
looking at the contrast analysis of the example. Suppose that there are two dierent
events referred to by the phrases your friends nding the bomb, in contrast to your
friends overlooking the bomb and your friends nding the bomb, in contrast to
there being no bomb and your friend nding nothing, and that the phrase your
friends nding the bomb could designate either event. Perhaps we could dene
event
1
as an event that occurs in worlds in which I place a bomb outside the door
and your friend nds it, and that does not occur in worlds in which I place a bomb
outside the door and your friend overlooks it or in worlds in which I do not place a
bomb outside the door. And we could dene event
2
as an event that occurs in worlds
354 C. Maslen
in which either I place a bomb outside your door and your friend nds it or I place
a bomb outside your door and your friend overlooks it, and that does not occur
in worlds in which I do not place a bomb outside the door. Then, after some work,
we could show that the example involves an equivocation rather than a failure of
transitivity.
But how could the phrase your friends nding the bomb designate event
2
?
(Surely overlooking the bomb is not just another way of nding the bomb!) To ana-
lyze the example in this way we have to build conditions external to the event into
the event identity conditions for the event. The contrast analysis of this example is
more plausible.
Acknowledgments
Thanks to David Lewis, Ram Neta, and Nick White for helpful comments on drafts
of this paper.
Notes
1. In my dissertation (Maslen 2000) I argue that the contrastive counterfactual account can also solve
the problems of preemption, causation by absences, and causation under the indeterminism of modern
physics.
2. This context-dependence is additional to any context-dependence arising from indexicals in the cause or
eect phrases of the statement.
3. Hitchcock (1996a).
4. Fall 1997 seminar on causation at New York University.
5. Holland (1986).
6. Kim (1973a, p. 222) denes an event as a concrete object (or n-tuple of objects) exemplifying a property
(or n-adic relation) at a time. According to Lewis (1986d), for the purposes of causation an event may be
dened as a predominantly intrinsic, nondisjunctive property of spacetime regions.
7. For a detailed discussion of the varieties of causal relata and one way of reducing causal sentences
involving other relata to those involving Kim-style events see Menzies (1989b).
8. However, at the least a distinction between backtracking and standard counterfactuals must be
drawn, in order to rule out spurious cases of backward causation. For example, see Lewis (1979a), p. 34.
In normal conditions, backtracking counterfactuals are those saying that the past would be dierent if the
present were somehow dierent. It must be shown that backtracking counterfactuals can always be iden-
tied syntactically.
9. I do not discuss the status of absences in this essay. However, if, for example, the contrast event were
itself the absence of the eect, the causal statement would clearly be trivially true.
10. See Lewis (1986d) for a sophisticated treatment of the distinctness of events. If cause and eect are not
required to be distinct events, then this denition would mistakenly judge various relations of noncausal
determination to be causation. Furthermore, it would entail that every event is self-caused. However,
requiring cause and eect to be distinct events rules out self-causation by atan unwelcome result.
(Thank you to an anonymous referee for pointing this out.) Lewis neatly sidesteps this problem in his
original account, by requiring only events that are directly causally dependent to be distinct (see postscript
F of Lewis 1986b). He thus allows for the logical possibility of the closed causal loops that appear in some
time-travel stories. I cannot see how to expand account
1
to yield the same results without assuming
transitivity.
However, notice that account
1
does not require cause and eect to be distinct; rather, account
1
is
restricted to cases where cause and eect are distinct. It is simply silent about self-causation and other cases
of causation between nondistinct events.
11. Lewis solves a similar problem in his (2000), where alterations play a similar role to that of contrast
events in account
2
. Lewis limits his account to not-too-distant alterations of the cause, where an altera-
tion of event E [is] . . . either a very fragile version of E or else a very fragile alternative event that is similar
to E, but numerically dierent from E (p. 188).
12. If the contrast event is itself the event in which the absence of the cause consisted, then the contrast
situation is the same as the corresponding contrast event.
13. This claim is ambiguous. I mean: Event c, relative to a set of contrast events fc
g, is a cause of event
e, relative to a set of contrast events fe
g, i for every c
i
in fc
g there is an e
i
in fe
g such that if c
i
had
happened then e
i
would have happened. Often there is only one natural contrast to the cause c
, and one
natural contrast to the eect e
. In that case account

3
simplies to the following: Event c, relative to
contrast events c
, is a cause of event e, relative to contrast event e
, i if c
had happened then e
would
have happened.
14. Hausman (1992). A solenoid is a coil of wire and a solenoid switch is a relay; when current passes
through the coil a magnetic eld is induced in the coil, and the switch closes. There are also two manual
switches in the circuit. A circuit like this is common in many household appliances, for example the three-
way switch found on some standing lamps.
15. The word cause is so inextricably bound up with misleading associations as to make its complete
extrusion from the philosophical vocabulary desirable. Instead we should talk of functional dependence,
according to Russell (1953).
16. Van Fraassen (1980), p. 118.
17. Van Fraassen (1980). Van Fraassen then uses this to reject the counterfactual theory of causation on
the grounds that counterfactuals are context dependent and causation is not.
18. Stalnaker (1972), p. 385. However, see Lewis (1980), for a discussion of how a one-level scheme diers
only supercially from Stalnakers two-level scheme, and for a broader construal of context-dependence
encompassing contingency.
19. Note that there is further context-dependence in this sentence associated with the past tense of the
sentence, the use of the denite article the and the use of the verb went (see Fillmore 1971). In total,
the truth and meaning of the sentence depend on many features of the context including the identity of the
speaker, the location of the speaker, the location of the audience, the time of utterance, and previous
statements in the conversation.
20. Stalnaker (1972), p. 384.
21. Lewis (1979b).
22. Van Fraassen (1980), pp. 126146, treats the question of how a contrast class is obtained from the
context in the case of explanation and why-questions, and some of this also applies to the case of causa-
tion. This is very relevant to the discussion in this section. However, note that van Fraassen is analyzing
explanation, not causation, and that his examples are of contrasts with the eect.
23. For example, Chisholm (1946), Goodman (1979), Lewis (1973b), and Stalnaker (1972).
24. Lewis (1979a), p. 41.
25. Lewis (1973b), p. 93.
26. Chisholm (1946); Tichy (1984).
27. Lewis (1986a), p. 163.
356 C. Maslen
28. For example, Lewis (1986a), p. 167, simply asserts [c]ausation must always be transitive, and pro-
ceeds to add this requirement to his analysis of causation. Further examples of those writing around the
same time who simply assume the transitivity of singular causation are Ehring (1987), Rosenberg (1992),
and Hitchcock (1995a).
29. As far as I can see, these date back to Jig-chuen Lee (1988), but similar examples have been presented
independently, and it took some time for their existence to be widely recognized.
30. The passage quoted appeared in an earlier draft version of Halls essay, chapter 9 in this volume.
31. Ibid.
32. In conversation.
33. Ehring (1987), p. 323. A version of this example and a discussion also occurs in Ehring (1997), p. 76.
34. See Lewis (1973b), p. 33.
35. If an analysis of backtracking counterfactuals is not available, the following condition is also sucient
for transitivity together with (C1): (c
1
& c
2
) Ld
3
.
36. Hausman (1992). Also see Ehring (1997), p. 76. Ehring accounts for examples of this sort in a parallel
way by employing an event-feature theory of causation.
15Causation: Probabilistic and Counterfactual Analyses
Igal Kvart
In the bulk of this essay (sections 113) I oer a probabilistic analysis of token cau-
sation. The probabilities employed are chances. In the analysis of cause, the main
task is identifying the right notion of probability increase; but causal relevance is a
crucial, delicate, and widely overlooked ingredient in the analysis of causation (and
of counterfactuals). In a chancy world, causal irrelevance is secured either through
probabilistic irrelevance or through the presence of a so-called neutralizer, which es-
sentially screens o C from A in a stable way and is such that A is not a cause of it.
Despite appearances to the contrary, the account is not circular. The core diagnosis
of early preemption, late preemption, double preemption, and overlapping cases is
that the preempted cause is causally irrelevant to the eect (and thus is not a cause of
it). The proper analysis of such cases consists in establishing this causal irrelevance
by specifying the pertinent neutralizer.
Various workers in the eld have in the past become convinced that a probabilistic
analysis of cause is not viable and that other resources such as processes must be
employed. The major problem that beset probabilistic analyses was the problem of
causes that lower the probability of their eects. In the analysis of cause in section 1,
I show how this diculty can be overcome. Since the requisite notion of causal rele-
vance is also analyzed probabilistically, one upshot of this essay is that probabilistic
analysis of (token) causation has been abandoned prematurely.
In the remainder of the essay (sections 1415) I derive a counterfactual analysis
of cause from my account of counterfactuals and the thesis that cause amounts to
some positive causal relevance. Although the account is extensionally adequate and
(I hope) illuminating, it turns out to be circular. (This part may be read indepen-
dently of the rst part.)
1 Cause
I assume a chancy world, where the chance of C is conditional on some prior1 world-
state or history (and possibly other events), that is, having the form: P(C/W
t
), where
W
t
is the world history up to (or a world-state just prior to) t (t being earlier than
the beginning of t
C
, which is the interval to which C pertains).2 It is crucial for the
proper understanding of the account presented here, and in particular regarding par-
ticular examples, that the notion of probability employed throughout be construed as
that of chance.
I have argued that, for A to be a token cause of C,3 A must raise the chance of
C ex post facto, that is, while taking into account the interval between A and C.
Traditionally, probabilistic analyses of causation have had to face the diculty of
how to reconcile the intuitive idea that causes raise the probability of their eects
with certain cases where causes seem to lower them. On the token level, this problem
arises in particular when only the history of the world prior to the antecedent is taken
into account.4 Put in terms of chances, the natural way of expressing probability in-
crease is:
PC=A:W
A
> PC=@A:W
A
1
(called ab initio probability increase; in short: aipi ). For instance, suppose I drop the
chalk (A), and the chalk then falls on the oor (C) (absent any complications). A,
intuitively, is a cause of C, and ab initio probability increase obtains.
However, in (1), only the history prior to A, that is, W
A
, is taken into account. Ex
post facto probability increase, on the other hand, must take into account the inter-
mediate history as wellthat between A and C. The condition for being a cause,
then, is not ab initio probability increase, which is not sucient for being a cause,5
but rather a condition of ex post facto probability increase (and a particular form of
it at that), which I shall spell out now.
Ex post facto probability increase can be illustrated in cases of ab initio probabil-
ity decrease (i.e., (1) with < instead of >; in short: aipd ). Ex post facto
probability increase despite ab initio probability decrease requires that there be
an actual intermediate event E that yields probability increase when held xed,
that is, when added to both sides of the ab initio probability decrease condition.
Thus:
PC=A:E:W
A
> PC=@A:E:W
A
: 2
Call an event such as E an increaser.6 To illustrate:
Example 1 The Comeback Team had been weak for quite a while, with poor
chances of improving during the next season. Consequently, there were very high
odds of its losing. Nevertheless, x bet $Y, a signicant portion of her nancial worth,
on its winning (A). Later, but before the beginning of the games, the Comeback
Team was acquired by a new wealthy owner, an event that had been quite unlikely
earlier. The new owner then acquired a few rst-rate players. Consequently, the
teams performance was the best in the season (E), x won her bet, and C occurred: x
improved her nancial position.7
360 I. Kvart
As of t
A
, A yielded a probability decrease of C, that is, ab initio probability de-
crease. But given E, A yielded a higher chance of C.8 Hence E is an increaser for A
and C. And indeed, intuitively, A was surely a cause of C.
A note of caution: The term increaser should not lead one to think that an
increaser increases the chance of C when added to the condition in PC=A:W
A
. The
import of an increaser E as such does not involve the characterization of the rela-
tion between a conditional probability once with E added to the condition and once
without it. Rather, its the relation of two conditional probabilities, both with E in the
condition, one with A, the other with @A. For instance, consider a student x who
took an exam:
A: x gave a bad answer to question b.
Yet:
E: x answered question d correctly.
C: x received a high grade.9
Of course A yields ab initio probability decrease for C. Yet E is not an increaser,
since:
PC=A:E:W
A
< PC=@A:E:W
A
:
So E does not raise the probability of C when held xed in the condition on both
sides of the ab initio probability decrease condition. Yet E raises the probability of C
given A (and W
A
) when considered by itself, that is:
PC=A:E:W
A
> PC=A:W
A
:
Thus, E is not an increaser (for A and C).
However, ab initio probability increase for A and C need not yield that A is a
cause of C, since there might be a decreaser for A and C, that is, an intermediate E
fullling (2) with < instead of >, which undermines the indication of ex post
facto probability increase of the ab initio probability increase. Hence ab initio prob-
ability increase is not a sucient condition for being a cause. That raising the prob-
ability is not a sucient condition for being a cause, absent cutting of causal routes,
has not been properly noticed in the literature.10 Call an intermediate event that is
either an increaser or a decreaser a reverser.
An analogous problem may plague the presence of an increaser: An increaser
might have a further decreaser for it, that is, an intermediate event11 F fullling
Probabilistic and Counterfactual Analyses 361
PC=A:E:F:W
A
< PC=@A:E:F:W
A
: 3
Condition (3) undermines the indication of ex post facto probability increase yielded
by the increaser E. The possibility of a decreaser for an increaser shows that the in-
dication of ex post facto probability increase yielded by an increaser depends on the
choice of that increaser, and that a choice of other intermediate events that are
held xed need not yield probability increase. For A to be a cause of C, it must have
an increaser E without a further decreaser for it (such as F in (3)). Call such an
increaser stable (or strict).12 The probability increase indicated by a stable increaser
is stable since it is not reversed when other intermediate events are taken into
account.
In addition to being extensionally adequate, it is also intuitively plausible that
the notion of cause is one of probability increase that is both ex post facto and
stable. Compare the assertibility conditions of a causal claim from a retrospective
perspective (from some suciently later time) without limitations of knowledge of
facts or of chances. The role of an assertion of ab initio probability increase,
when restricted to knowledge of facts that pertain up to t
A
13 (rather than the objec-
tive notion used above, which is relativized to W
A
) is useful for prediction when
one is at t
A
and is concerned to assess whether A will bring about Cthis is ab
initio assessment. But the notion of As raising the probability of C that is needed
for the notion of As being a cause of C must be robust enough so that the proba-
bility increase involved proves stable across the intermediate history. Stable ex
post facto probability increase is thus the requisite constraint for the notion of
cause.
For the sake of a uniform terminology, consider the case of ab initio probability
increase a case of a null increaser14 (as in the above example of the chalk that fell to
the oor); similarly, consider the case of ab initio probability decrease a case of a null
decreaser. A null increaser may be stable (i.e., if there is no decreaser). In such a case,
its presence constitutes a sucient condition for being a cause.15
In analogy to a stable increaser, we have the notion of a stable decreaser, namely,
a decreaser for which there is no further increaser. The presence of a stable increaser
yields positive causal relevance, and the presence of a stable decreaser yields negative
causal relevance. The copresence of a stable increaser and a stable decreaser yields
mixed causal relevance.
We saw that for A to be a cause of C there must be a strict increaser for A and
C. We can thus reformulate this conclusion by stating that As being a cause of C
amounts to there being some positive causal relevance for A and C.16
362 I. Kvart
2 Hitchcocks Randomizer and Doctors Dispositions
Against probabilistic accounts of causation, and in favor of contrastive accounts,
Hitchcock raised an example of the following kind:
A: the doctor prescribed to the patient x 100 ml of the medicine;
and indeed:17
C: the patient recovered.
Accordingly, the medicine was indeed helpful at the level of 100 ml. 200 ml would
have been very likely to be much more eective.
Was A a cause of C? I say, intuitively, yes. Hitchcock says no, at least not without
contrastive qualication. On his view, whether A comes out a cause of C on a prob-
abilistic analysis depends on the doctors dispositions. Assume, in variation 1, that
the doctor above was disposed to prescribe 200 ml (he acted in fact contrary to his
disposition). Variation 2 is just like variation 1 except that the doctor was disposed to
prescribe 0 ml, yet here too he acted contrary to his own disposition. That is, in the
two variations the doctor was on dierent dispositional curves between the values of
0 ml and 200 ml even though in fact he acted in the same way (prescribed 100 ml),
and the patient recovered. Whether there is probability increase or decrease depends
on the dispositional curve of the doctor in question and thus varies between the
two variations, since the most likely event in the nonactual case, namely, given @A,
would be very dierent for dierent curves. But, Hitchcock says, the outcome of
whether A was a cause of C must not depend on these dispositions. Hence proba-
bilistic accounts fail and a contrastive account is called for, in which cause in ana-
lyzed as a three-term relation between A, C, and a contrastive alternative. In sum,
Hitchcock correctly observes that the result of whether A was a cause of C must not
come out dierently in the two variations. But since, on his view, probabilistic ac-
counts of causation yield just this result, he concludes that the fault lies with the
notion of cause being construed as a two-place relation.
However, Hitchcocks reasoning relies on a simplistic conception of probabilistic
causation. In view of the account of the last section, A here had a mixed causal im-
pact on C: a positive causal impact in view of the doctors not having avoided giving
the patient this medicine (prominent in the case of the second variation), and a neg-
ative causal impact in view of his failing to give the patient the higher dose (promi-
nent in the case of the rst variation). Thus, consider:
E: the dosage that the patient took didnt exceed 100 ml
E is a strict increaser, since, given E and the fact that the doctor prescribed 100 ml
(A), the probability of the patients taking a 100 ml dosage is high, and thus the prob-
ability of C (the patient was cured) is higher than it is given E and @A (i.e., its not
being the case that the doctor prescribed 100 ml).18 On the other hand,
F: the dosage the patient took was not less than 100 ml
is a stable decreaser, since, given F and A, again, the probability that the patient took
100 ml is high, and thus the probability of C (the patient was cured) is lower than it
is given F and @A. Hence in both variations there is a stable increaser and a stable
decreaser, which is characteristic of mixed causal relevance cases. To be a cause, re-
call, is to have some positive causal relevance, and obviously there is some positive
causal relevance in all mixed causal relevance cases, including the above. Hence, on
the above analysis, A comes out a cause of C in both variations, as it should.
This is the outcome of the analysis regardless of the disposition of the doctor, and
Hitchcocks main concern is to secure just such independence. The crucial point here,
again, is to recognize the importance of stable reversers and to avoid construing
probability increase in the simplistic form of ab initio probability increase. Once
stable reversers are taken into account, the outcome of the above probabilistic ac-
count (regarding whether A was a cause of C) does not vary with the doctors dis-
positions. This ts not only Hitchcocks intuitions but also his robustness principle,
to the eect that whether A was a cause of C in the two variations ought to depend
on the physiological condition of the patient and the chemical properties of the pill
but be independent of the doctors dispositions. Hitchcock considers a randomizer
that determines a pattern of such dispositions, and he maintains that whether A
was a cause of C should be independent of the randomizers selected pattern of dis-
positions. Hitchcocks intuitions and principle are, therefore, respected on the above
analysis. For Hitchcock, this example motivates favoring a contrastive account of
cause in view of the inadequacy of the simplistic form of the probabilistic account
of causation. But once an adequate probabilistic account is adopted, the counter-
intuitive outcome is blocked, and Hitchcocks example no longer provides a motiva-
tion for ruling out the ordinary construal of causation as a two-place relation and
opting, instead, for a contrastive conception.
A much briefer, and ultimately more satisfactory, response to Hitchcock is that A
0
is a cause of C on the analysis presented here:
A
0
: the doctor prescribed the medicine to the patient.
There is ab initio probability increase of A
0
to C, with no reversers. Hence A
0
is a
cause of C (given causal relevance). A, however, is an informational expansion of A
0
,
364 I. Kvart
and hence it is also a cause of C, in view of the principle of Cause Preservation under
Informational Expansion.19
Yet there remains the issue of the causal impact of the aspect at a dose of 100 ml
in A on C. The analysis of whether this aspect of A was a cause of C requires an
analysis of aspect causation,20 though its outcome would not change the conclusion
that A was a cause of C.
3 Causal Relevance Neutralizers
The existence of a stable increaser is a necessary but not a sucient condition for
somethings being a cause. Causal relevance (of A to C) is also a necessary condi-
tion for As being a cause of C. So if A is not causally relevant to C, A is not a cause
of C even if there is a stable increaser for A and C. This happens if there is, in ad-
dition to the stable increaser, an intermediate event that neutralizes the would-be
causal relevance of A to C. So, the presence of a stable increaser does not suce for
causal relevance, and thus does not suce for being a cause. Consider the following
example:
Example 2 x was pursued by two enemies who wanted to kill him. They discovered
that he was about to be at a particular meeting place in an area covered with heavy
snow at a particular time. Enemy 1 arrived with his attack dog, discovered that there
was a cave with a very small entrance right next to the meeting place, and hid there
with his dog. Enemy 2, with his gun ready, found another place to hide, overlooking
that meeting place. At the designated time, x arrived at the meeting place, and indeed:
A: enemy 1 released his dog at t
1
.
However, owing to the heavy mass of snow that covered the slope above the
entrance of the cave:
E: an avalanche completely sealed the entrance to the cave at t
1
dt.
This happened right after the dog was released, and before it had a chance to charge
forward. Both enemy 1 and his dog were thus trapped in the blocked cave. However,
on observing the arrival of x:
B: enemy 2 shot at x
and indeed hit his mark:
C: x was injured.
Intuitively, A was not a cause of C, but B was. However, A had ab initio proba-
bility increase to C, that is:
PC=A:W
A
> PC=@A:W
A
;
and no decreaser,21 so A had a null stable increaser to C. But A was nonetheless not
a cause of C, since A ended up being causally irrelevant to C.
For there to be causal relevance, there must be probabilistic relevance of A to C. A
is probabilistically relevant to C if there is either an increaser or a decreaser for A and
C. Increasers and decreasers may be null or not. The presence of a null increaser or a
null decreaser for A and C is tantamount to:
PC=A:W
A
0PC=@A:W
A
4
A non-null increaser or decreaser F for A and C yields:
PC=A:F:W
A
0PC=@A:F:W
A
5
Call such an F a dierentiator (for A and C). In the case of the ab initio inequality
of (4), consider an empty intermediate event a null dierentiator (for A and C). Thus,
a null dierentiator is either a null increaser or a null decreaser, and a dierentiator
(null or not) is either an increaser or a decreaser (null or not). Thus, A is probabil-
istically relevant to C just in case A makes a probabilistic dierence to C, either
directly (as in (4)) or via an intermediate event that is held xed (as in (5)). A is
therefore probabilistically relevant to C just in case there is a dierentiator (null or
not) for A and C.
The absence of a null dierentiator amounts to (4) not holding, that is, to the
presence of probabilistic equality:
Equiprobability: PC=A:W
A
PC=@A:W
A
Without probabilistic relevance there is no causal relevance; and yet A was prob-
abilistically relevant to C in the last example. What accounts, then, for the causal
irrelevance there? It was the presence of an intermediate event that neutralized the
would-be causal relevance of A to C. In that example it was E. Call such an event a
causal relevance neutralizer. So if A is probabilistically relevant to C, A is causally
relevant to C just in case there is no causal relevance neutralizer for them. That is,
if there is a causal relevance neutralizer for them, A is not causally relevant to C,
despite probabilistic relevance. Our problem, then, is how to characterize causal rel-
evance neutralizers. (We can call them for short just neutralizers.)
A would-be causal chain from A to C may be neutralized if it is cut o; but this
is not the only way. Such a chain may be diverted or simply dissipate. Further, even
366 I. Kvart
if the would-be causal chain from A to C was cut o, the event that did the cutting
need not secure that there was no causal relevance.22 The role of securing that there
was no causal relevance of A to C is that of a neutralizer.
4 Candidates for Causal Relevance Neutralizers
A neutralizer E, which secures causal irrelevance of A to C (despite probabilistic rel-
evance), must screen o A from C; that is, it must fulll:
PC=A:E:W
A
PC=@A:E:W
A
6
An event E satisfying (6) is a screener for A and C (or, alternatively, a blocker
for them).23 But a neutralizer for A and C must also screen A o from C in a stable,
that is, unreversed way: There must not be any intermediate event F that undoes this
screening-o. In other words, there must not be an intermediate24 event F such that:
PC=A:E:F:W
A
0PC=@A:E:F:W
A
7
As an illustration, consider example 3.
Example 3 Two pipes lead from a faucet to a pool. The faucet can be in only one of
two positions. The faucets being in any one of the two positions allows the water to
ow exclusively through one pipe or exclusively through the other. I switched the
faucet in the direction of the right-hand pipe (A). That pipe allowed for a much
better water ow than the other pipe. And indeed, the pool lled up in due course
(C). Hence there is ab initio probability increase for A and C (and thus there is a null
dierentiator for them). But after I switched the faucet, a stray bullet hit the right-
hand pipe at a certain spot before the water reached it. Consequently, there was
a large hole in the right-hand pipe, and both pipes became equal in terms of their
water-ow conductivity (E). E is thus a screener for A and C.
The mere presence of a screener need not yield stable ex post facto probabilistic
equality, since a screener may have a dierentiator for it. If E is a screener, fullling
(6), for which there is no F, fullling (7), then the probabilistic equality of (6) is in-
deed stable. Call such an event E that satises (6) for which there is no such F sat-
isfying (7) a stable screener for A and C (or, alternatively, a stable blocker). A stable
screener, then, yields a stable probabilistic equality. A neutralizer for A and C must
be a stable screener.25
If E is a stable screener (for A and C), so would be conjunctive expansions of E.26
To focus on the intrinsic features of neutralizers, well conne our attention to lean
stable screeners, that is, stable screeners without extra information that plays no role
in their being stable screeners.27 Call lean stable screeners candidates for causal rele-
vance neutralizers (or for short: candidates for neutralizers).
5 The Analysis of Neutralizers
So far we have specied what role a neutralizer is taken to play. But we must specify
now what a neutralizer is. My thesis is the following:
Thesis: E is a neutralizer for A and C just in case E is a lean stable screener for A
and C of which A is not a cause.
This thesis need not strike us as being intuitively compelling; but, I hold, it is cor-
rect28 (although space doesnt allow to spell out the full motivation here).29 It also
threatens circularity, since it employs the notion of cause, but this threat of circular-
ity will evaporate, as we will see below.
First, let us apply the above analysis to example 2 above, with the dog and the
avalanche. Surely, intuitively, A (enemy 1 released his dog at t
1
) ended up being
causally irrelevant to C (x was injured), even though there is a null dierentiator for
A and C. Yet E (i.e., an avalanche completely blocked the entrance to the cave at
t
1
dt) was a stable screener for A and C, and intuitively A was not a cause of E
(and indeed, in terms of our analysis, there was also no dierentiator for A and E).
Yet B (i.e., enemy 2 shot at x) has a null stable increaser for C, and thus there is
a dierentiator for them, and there is no causal relevance neutralizer. B, therefore,
intuitively as well as by the above analysis, ended up being a cause of C, whereas A
wasnt, since it ended up being causally irrelevant to C.
Now to the threat of circularity. Suppose that there is no dierentiator for A and
C (null or not). Then A is causally irrelevant to C, since probabilistic relevance is a
necessary condition for causal relevance.
So suppose there is no equiprobability for A and C, or else there is equiprobability
but there is a dierentiator. Then A is probabilistically relevant to C, and so it is
causally relevant to it unless there is a neutralizer for A and C.
So consider a candidate for neutralizer E
1
for A and C: a lean, stable screener for
A and C. It now remains to be established whether A is a cause of E
1
in order to
determine whether E
1
is a neutralizer for A and C. If there is no stable increaser for A
and E
1
, A is not a cause of E
1
; E
1
is therefore a neutralizer for A and C, and A is
causally irrelevant to C.
So suppose there is a stable increaser for A and E
1
. Then A is causally relevant to,
and a cause of, E
1
(and hence E
1
is not a candidate for a neutralizer for A and C)
unless there is a neutralizer for A and E
1
. So consider a candidate for neutralizer E
2
368 I. Kvart
for A and E
1
. If there is no stable increaser for A and E
2
, A is not a cause of E
2
;
hence E
2
is a neutralizer for A and E
1
; so A is causally irrelevant to E
1
and thus not
a cause of it; hence E
1
is a neutralizer for A and C, and A is thus causally irrelevant
to C.
Otherwise, if there is a stable increaser for A and E
2
, we must look for a candidate
neutralizer for A and E
2
. And so on. This procedure may terminate if there is an E
n
with no stable increaser for A and E
n
; or else there may be an innite regress.
If a chain of this sort terminates, it terminates in the establishment of some E
n
as a
neutralizer (for A and E
n1
); hence A is, in this case, causally irrelevant to C. Thus, if
such a chain terminates, there is a conclusive positive result of causal irrelevance; call
such an E
n
an endpoint neutralizer. But what happens if no such chain terminates?
Recall that this is a prima facie case of causal relevance of A to C, since there is a
dierentiator for them. Yet there is no neutralizer for A and C, since no such chain
terminates. Hence there is no neutralizer to overrule the prima facie causal relevance
of A to C. So A is therefore causally relevant to C.30
Thus, there may be innite regress, but there is no circularity, since the innite
regress case is a case of causal relevance.
The above presentations of the analyses of cause and of causal relevance summa-
rize the more detailed analyses, presented elsewhere with more detailed arguments
for their adequacy.31 The reader who needs to be further convinced by the adequacy
of these analyses is advised to look at the more detailed presentations.32
6 Illustrating Causal Relevance through Innite Regress
Consider example 4:
A: x red at y at t
0
.
C: y was shot at t
1
.
A, in this case (which is a simple, straightforward case with no complications),33 is
intuitively a cause of C; hence it is causally relevant to C. Surely there is ab initio
probability increase of A to C; hence there is a null dierentiator. So in checking for
causal relevance of A to C, by our analysis, we look for a neutralizer candidate for A
and C. And indeed, consider:
E
1
: xs bullet was at t
1
at point p
1
in midair, with momentum m
1
, spin s
1
, etc.
E
1
does screen o A from C and is in fact a lean, stable screener, and thus it is a
neutralizer candidate. But intuitively, A is surely a cause of E
1
. And indeed, there is
ab initio probability increase of A and E
1
, and hence there is a null stable increaser.
So E
1
is not a neutralizer for A and C unless there is a neutralizer E
2
for A and E
1
.
And indeed, consider E
2
, which is just like E
1
only with t
2
, p
2
, m
2
, s
2
, and so on in-
stead of t
1
, p
1
, m
1
, s
1
, and so on. E
2
is a neutralizer candidate for A and E
1
just as
E
1
was a neutralizer candidate for A and C. But again, A is intuitively a cause of E
2
.
And indeed, A has a null stable increaser to E
2
. Hence E
2
is not a neutralizer for A
and E
1
unless there is a neutralizer E
3
for A and E
2
; and so on.
There is thus an innite series of such E
i
s34 where the corresponding t
i
s converge
to some temporal point between A and C. That there is only innite regress of this
sort (i.e., there not being a suitable terminating chain) establishes that A is causally
relevant to C, since it establishes that there is no neutralizer for A and C (in the pres-
ence of a dierentiator).
7 Lewiss Two Bombs and the Newspaper
Consider the following example:35
A: terrorist 1 places a bomb on ight 13 at t
1
.
B: terrorist 2 places a bomb on ight 17 at t
2
.
Both bombs are unreliable. The bomb of terrorist 1 goes o, and:
C: the newspaper prints a headline at t
3
: Airline bomb disaster.
The bomb of terrorist 2 fails; thus:
E: no bomb went o on ight 17,
but the headline would have been just the same had it gone o (with or without the
other bomb). The bomb on ight 17 raises the probability of the eect, yet it is not a
cause of C. Clearly, A is intuitively a cause of C, and B is not.
Indeed, both A and B bear ab initio probability increase to C, with no reversers.
The dierence between them is that A is causally relevant to C but B is not. E is a
stable screener for B and C,36 and B is not a cause of E, since there is ab initio
probability decrease of B to E and no reverser. So E is a neutralizer for B and C, and
B comes out as causally irrelevant to C and thus not as a cause of it. (Hence B didnt
cause C, since being a cause of C is a necessary condition for causing C.) On the
other hand, there is no neutralizer for A and C; hence A is a cause of C. The outcome
then is intuitively the right one, and the analysis yields it in a noncircular way.
370 I. Kvart
8 Late Preemption
I will now illustrate the analysis by applying it to the case of late preemption37 that
posed a severe problem for Lewiss analysis of causation.
Suzy threw a stone at the window at t
1
(A
1
), and Billy threw a stone at the window
at t
2
(A
2
).38 Suzys stone hit the window rst (it doesnt matter which one of the two
threw the stone rst), and the window shattered at t
3
(C). Intuitively, A
1
was a cause
of C, and A
2
was not.
However, both A
1
and A
2
yield ab initio probability increase for C, and neither
has a reverser. A
2
, however, should come out as causally irrelevant to C. And indeed,
consider:
G: Billys stone did not hit the window (until t
3
e).
(t
3
is the upper edge of the interval t
3
.) G is a stable screener for A
2
and C, and there
is ab initio probability decrease for A and G (and there is no increaser). Hence G is a
neutralizer for A
2
and C, and so A
2
is causally relevant to C.
Another neutralizer for A
2
and C is:
E: Billys stone was at t
3
e at a distance not smaller than d from the window, not
traveling toward it at velocity greater than v.
(d and v are the actual distance and velocity of Billys stone at t
3
e.)
E is a screener for A
2
and C,39 since, given E, B makes no dierence to the win-
dows shattering at t
3
, and E is a stable increaser. A
2
is not a cause of E since A
2
doesnt increase the chance of E (and there is no stable increaser).40 Hence E is a
neutralizer for A
2
and C, and A
2
is not a cause of C. A
1
, on the other hand, has no
neutralizer for C, and so it is thus a cause of C.
By way of testing probabilistic accounts of cause (including the account above),
Christopher Hitchcock oered a counterexample of late preemption that is structur-
ally similar. In it, bullet 1 shattered a fragile vase by hitting it, though barely so,
while bullet 2, shot at around the same time, missed the vase by an extremely tiny
distance at the very same time. A (bullet 1 was emitted) was a cause of C (the vase
shattered at t), but B (bullet 2 was emitted) was not. Yet, Hitchcock claims, B in-
creased the probability of C all the way through.41
However, A comes out as a cause of C and B does not, since G
, a counterpart of
the above G, is a neutralizer for B and C (though not for A and C):
G
: bullet 2 did not hit the vase (until t
e).
G
screens o B from C and lowers the chance of C (without an increaser). So G
is
a neutralizer for B and C, and B doesnt come out as a cause of C.
Yet E
, a counterpart of the above E, is also a neutralizer for B and C (though not

for A and C):
E
: bullet 2 was at t
e at a distance not smaller than ds from the window, not

traveling toward it at velocity greater than v.
(ds and v are the actual distance and velocity of bullet 2 at t
e.)42
Again, E
is a stable screener for B and C, and B is not a cause of E
for the same

considerations regarding E above. Hence E
is a neutralizer for B and C, so B does

not come out as a cause of C. There is ab initio probability increase for A and C,
with no reverser and no neutralizer (including E
); hence A comes out as a cause of

C, which is the right result.43
9 Early Preemption
Consider example 5, a neuron-type case (of the sort introduced by Lewis). Suppose
both neurons c
1
and c
2
re. Neuron c
1
has a single, direct connection to neuron e.
Neuron c
2
also has a single connection to e, but one that passes through neuron
d. These connections transmit stimulatory signals. There is also a connection from
c
1
to d that transmits an inhibitory signal. Assume that the inhibitory and stim-
ulatory signals respectively yield high probability that the inhibited neuron does
not re and that the stimulated neuron res.44 Finally, assume that the inhibitory
signal from neuron c
1
reaches neuron d before the stimulatory signal from neuron
c
2
. So:
A: neuron c
1
red;
B: neuron c
2
red;
C: neuron e red;
D: neuron d didnt re.
Intuitively, A, but not B, was a cause of C.45 There is ab initio probability increase
for A and C and for B and C, and there are no reversers. There is no neutralizer
for A and C; but D is a lean stable screener for B and C, and B is not a cause of D
(there is ab initio probability decrease for B and D, no reverser, and hence no stable
increaser). So D is a neutralizer for B and C, and hence A is a cause of C whereas B
is not.
372 I. Kvart
One can further assume, without changing the signicant features of the case, that
the route from c
1
to e was suciently less conducive to es ring than the route from
c
2
. In this case, A lowered the probability of C (ab initio), while still being a cause
of it.
Lewiss way of handling early preemption relied on the presence of an intermediate
event between A and C (later than D), in the form of an intermediate neuron on the
path from c
1
to e, which allowed his counterfactual analysis to resort to stepwise
causation, essentially relying on such an intermediate event. Note, however, that the
above analysis of the example resorts to no such intermediate event. Although
a suitable intermediate event need not involve an additional neuron,46 Lewiss han-
dling of the example requires the presence of some suitable intermediate events,
whereas the above analysis does not. Thus, a ctional example that involves action
at a distance or discrete time, with no intermediate events between A and C (at least
after D), seems immune to Lewiss standard counterfactual treatment of early pre-
emption, but not to the above analysis.
10 The Doctor and the Convict
The convict was executed at 11:00 p.m. as scheduled. He was, however, terminally ill
and on the verge of dying at any moment.
A: the doctor made great eorts throughout the evening to keep him alive until
11:00 p.m.
C: the convict died at 11:00 p.m.47
A, a delayer, is intuitively a cause of C, even though A is not a cause of C
1
:
C
1
: the convict died that evening.
A has ab initio probability increase to C, since it raised the chance that the convict
was alive until 11:00 p.m., and there is no neutralizer for A and C. Yet there is no ab
initio probability increase for A and C
1
: There is ab initio probability decrease, and
there is no reverser, so A doesnt come out as a cause of C
1
.48
11 Early Preemption with Dierent Time Specications
Early preemption cases allow for variations that yield dierent results depending
on how the temporal parameter is specied. Consider, for instance, the following
example.
Example 6
A: Oswald red at Kennedy at t
1
.
C: Kennedy was hit at t
3
.
(t
3
is instant-like.)
Contrary to the Warren commission, assume that there was a standby assassin x,
highly procient and well positioned. However, x didnt shoot, since Oswald did.
Consider:
G: x didnt shoot at t
3
T
(for a suitably very short T).49 Intuitively, G was not a cause of C: G does not have
any positive causal impact on C. Indeed, there is ab initio probability decrease of G
and C, and there is no reverser.50
Consider, on the other hand:
B: x came to the knoll.
Intuitively, B was not a cause of C.51 G
0
, however, is a stable screener for B and C:
G
0
: x didnt shoot during t
3
T
0
, t
3
e
(for a suitably short T
0
).52 But B ended up not being a cause of G
0
, since there is ab
initio probability decrease for B and G
0
, and there is no reverser. Hence G
0
is a neu-
tralizer for B and C, and B comes out as causally irrelevant to Cthe right result.
But now consider a candidate other than B and G:
D: x didnt shoot at t
2
(t
2
is considerably prior to t
1
). Intuitively, the fact that x didnt shoot at t
2
made it
feasible for Kennedy to be hit at t
3
(rather than at an earlier time), and thus was a
cause of it. D is thus an omission that is a cause of C.53
On our analysis, there is ab initio probability increase of D and C, no reverser, and
no neutralizer. (D is of course a cause of neutralizer candidates such as there was no
bullet at the air at t
2
dt.)54 Such outcomes reect the use of narrow event individ-
uation that provides for exibility regarding the choice of the temporal parameter.55
12 Double Preemption
Consider example 7 (due to Ned Hall):
A: Suzy set out on a bombing mission.
374 I. Kvart
Yet:
D: an enemy interceptor was launched to abort Suzys mission.
However, his own mission was aborted, since:
E: Billys plane intercepted the enemys interceptor.
Consequently, Suzys mission was successful:
C: Suzys bomb hit the target.
A and E were, intuitively, causes of C. Surely there is ab initio probability increase
for A and C and for E and C, and no reversers. Moreover, there is no neutralizer for
A and C and for E and C. In particular, F:
F: Suzys bomb was launched as planned
is not a neutralizer for either event, since, although it is a lean stable screener for
both vis-a`-vis C, both A and E were intuitively causes of F (and indeed, there is ab
initio probability increase, no reverser, and no neutralizer for both).
D, however, was intuitively not a cause of C. It had ab initio probability de-
crease to C with no reverser and thus with no stable increaser. Thus, D surely had
no positive causal relevance to C. Yet D was not purely negatively causally rel-
evant to C, as one may think in view of the presence of a stable decreaser and the
absence of any stable increaser. Rather, D was causally irrelevant to C. This is so
because there was a neutralizer for D and C. It is not E, since D was a cause of E
(there is a null stable increaser, and no neutralizer). But F is a lean stable screener for
D and C, and D was not a cause of F (there is ab initio probability decrease with no
reverser, and so with no stable increaser). Hence D ended up being causally irrele-
vant to C.
This aspect of the case illustrates the following phenomenon. In general, a neu-
tralizer E (for A and C) is causally relevant to C: if it secures the abortion of a
would-be cause of C, it has negative causal relevance to C; if it secures the abor-
tion of a would-be negative cause of C (a negative cause is an event which
bears purely negative causal relevance to C), it has positive causal relevance to C
(and thus is a cause of C). It fails to be either, and thus to be causally relevant
to C as well, in case there is yet another neutralizer for it and C. That is, there may
in general be a later neutralizer for A and C, which is also a neutralizer for the
earlier neutralizer E and C, thus rendering the rst neutralizer causally irrelevant
to C.
13 Omissions
Example 8 1, 2, and 3 are billiard balls. 2 was hit (B), and in turn it hit 3 (C). In the
meantime, x had an opportunity to hit 1 as it started its course in a certain direction
d (during a brief interval T). x didnt hit 1 in that direction during T (A). @A would
yield a high chance of 1s colliding with 2 on its way, diverting it from its course and
preventing 3 from being hit. Hence there is ab initio probability increase of A and C,
and intuitively A (which is an omission) is a cause of C. Indeed, there is no reverser,
and hence there is a null stable increaser.
There are stable screeners for A and C; yet A is a cause of them. For instance:
F: 2 is at s
0
at t
0
(on its way to hit 3, after T and close to but before t
C
; s
0
is outside the range of di-
rection d for ball 1). There is ab initio probability increase of A and F; there are no
reversers; hence there is a stable increaser, and, intuitively, A was a cause of F (there
is no neutralizerit is an innite regress case).
Or, alternatively, consider:
E: 1 is stationary at s during T
1
.
T
1
is a suitable, large enough interval before t
C
;56 s is where 1 was all along. There is
ab initio probability increase of A and E, and there is no reverser; hence there is a
stable increaser, and no neutralizer. Hence A comes out as a cause of C by the anal-
ysis, which is the right result. Omissions, on widespread intuitions, can be causes, and
the above analysis allows for this phenomenon.
Concluding Remark about Type Causation An account of token causation in terms
of chances of the sort presented here would, if successful, open the way to analyzing
type (generic) causation in terms of token causation. A main ingredient of type cau-
sation, I suggest, would be that type causation requires sucient presence of token
causes; that is, there should be enough57 token events of the cause type that are
token causes of token events of the eect type. (The notion of generic cause seems to
allow for relativity to a reference class, e.g., a country, an era, or circumstances.)
So much for the probabilistic analysis of cause. We now move to a counterfactual
analysis of cause (section 15).
14 Counterfactuals
Lewis oered a counterfactual analysis of causation for an indeterministic world in
terms of chance-consequent counterfactuals, which I discussed elsewhere.58,59 The
376 I. Kvart
probabilistic analysis of causation oered above was designed for an indeterministic
world, with the probabilities being objective chances.60 At the end of section 1 we
saw that the informal thrust underlying this probabilistic analysis is that A is a cause
of C i A has some positive causal relevance to C. But the details of this probabilistic
analysis need not concern us regarding the counterfactual analysis of cause in the
next section.
The notion of some positive causal relevance stems from my analysis of counter-
factuals,61 some main points of which I will summarize here. I argued that the core
counterfactuals of the form @A >@C are counterfactuals with false, factual ante-
cedents @A compatible with their prior history W
A
(called n-d counterfactuals, for
natural divergence).62 My analysis for these counterfactuals has been that the coun-
terfactual is true i the consequent is inferable on the basis of the antecedent and the
so-called implicit premises (through the laws of nature).63 The implicit premises for
such counterfactuals consist of the following three sets: rst, the history prior to A,
that is, W
A
. Second, the set of true statements describing events belonging to the in-
termediate course (i.e., the actual course of events pertaining to the interval between
t
A
and t
C
64) that are causally independent of A. Third, the set of true statements
describing events belonging to the intermediate course to which A is purely nega-
tively causally relevant.65 (For an example illustrating this analysis, see the stock
market example below, this section.)
As an example of purely negative causal relevance, consider: z struggled to prevent
x from ring at y (A), though unsuccessfullyx shot y (E) nonetheless. The fact A
that z struggled to prevent x from ring was purely negatively causally relevant to
Ethe fact that x shot y.
E is causally independent of A just in case A is causally irrelevant to E.66 A has
some positive causal relevance to E just in case A is causally relevant to E but not
purely negatively causally relevant. The notions of causal relevance and purely neg-
ative causal relevance, which are crucial to the above analysis of counterfactuals,
thus provide for a denition of the notion of some positive causal relevance, which is,
as I argued above, tantamount to the notion of cause. A has purely negative causal
relevance to C just in case A is causally relevant to C but has no positive causal
relevance to C. Thus, analyses of causal relevance and of some positive causal rele-
vance, which have been provided in this paper, yield an analysis of purely negative
causal relevance.
If A is causally irrelevant to E, A is not a cause of E.67 According to the above
conception of cause, if A is purely negatively causally relevant to E, then again A is
not a cause of C. If A is causally relevant to E without being purely negatively caus-
ally relevant to E, then A has some positive causal relevance to E and thus is a cause
of E. Consequently, the above analysis of counterfactuals @A >@C68 and of semi-
factuals @A > C (i.e., counterfactuals with true consequents) can be presented in
terms of the notion of cause as follows:
@A >@C (or @A > C) is true just in case @C (or C) is inferable on the basis of
@A, W
A
and the set of true statements describing events in the intermediate course
of which A is not a cause.
In view of the above, a counterfactual analysis of cause follows, as we shall see.69
In A Theory of Counterfactuals I divided the set of semifactuals exclusively and ex-
haustively into three sets: First, irrel-semifactuals, that is, semifactuals @A > C such
that A is causally irrelevant to C. From the above analysis it follows that all such
counterfactuals are true, since C is preserved among the implicit premises in virtue of
its causal independence from A. Second, pn-semifactuals ( pn for purely negative),
which are the semifactuals @A > C such that A has purely negative causal rele-
vance to C.70 All these semifactuals are also true, since C qualies as an implicit
premise. Third, the remaining set of semifactuals, called cts (for con-type-sems, ab-
breviating counterfactual-type-semifactuals), since they may be true or false, just like
regular counterfactuals.
As another example of irrel-semifactual (as well as a further illustration of the
above analysis), consider the following.71 x contemplates selling certain stocks. After
deliberating, x sells his stock (A) on Monday. On Tuesday morning, the stock market
skyrockets (E). Consider someone saying on Tuesday evening: Had x not sold
his stock, he would have been able to retire. On a plausible way of eshing out
this example, this counterfactual is intuitively true.72 On the above inferential
schema, the counterfactual is true only if event Ethe stock market skyrockets on
Tuesday morningis retained among the true statements from which the counter-
factual consequent is to be inferred (through the laws). And indeed, the actual A is
causally irrelevant to E. @A and its prior history W
A
do not suce to yield the con-
sequent @C (Cx is not able to retire) through the laws. A dierent intermediate
history W
A; C
that does not include events such as E need not uphold this counter-
factual.
As an example of a pn-semifactual, consider the above example of purely negative
causal relevance where z struggled to prevent x from ring at y (A), though unsuc-
cessfully. Had z not struggled to prevent x from ring at y, a fortiori x would have
shot y. This semifactual is true. The fact A that z struggled to prevent x from ring at
y was purely negatively causally relevant to Ethe fact that x shot y. @A > E here
is thus a pn-semifactual and is therefore true.
378 I. Kvart
15 The Counterfactual Analysis of Cause
We can now deduce a counterfactual analysis of cause from this analysis of counter-
factuals and the above conception of cause. If @A > C is an irrel-semifactual or a pn-
semifactual, then A is not a cause of C, since the requirement of some positive causal
relevance is violated. Now consider any other semifactual @A > C (which, as such, is
a cts). A is causally relevant to C and without purely negative causal relevance to C.
Hence A has some positive causal relevance to C and thus is a cause of C.
This last conclusion covers any semifactual not belonging to the above two sets
(i.e., of irrel-semifactuals and pn-semifactuals) regardless of whether the semifactual
is true or false. Hence, if @A > C is a cts, A is a cause of C. If @A > C is a false
semifactual, then its not an irrel-semifactual or a pn-semifactual (since the latter are
always true); so @A > C is a cts and A is a cause of C. In particular, if the counter-
factual @A >@C is true, then the complementary semifactual @A > C is false, and
hence A is a cause of C. The venerable intuition that if the counterfactual @A >@C
is true, then A is a cause of C, thus logically follows from the above analyses of
counterfactuals and the above conception of cause.
This completes the counterfactual analysis of cause. A is a cause of C73 i the
semifactual @A > C is neither an irrel-semifactual nor a pn-semifactual.
We can summarize these ndings by providing a breakdown of the conditions for
As being a cause of C and for As not being a cause of C in counterfactual terms of
the above sort, thereby providing a counterfactual analysis of cause. This summary
can be expressed by the chart in gure 15.1 (where A c C abbreviates A is a cause
of C and sem abbreviates semifactual).
Hume considered the truth of @A >@C as necessary and sucient for As being a
cause of C. In that Hume was wrong: The truth of @A >@C is indeed a sucient
condition (as shown aboveHume was right about that), but it falls considerably
short of being a necessary condition for As being a cause of C. Lewiss analysis
required causation to be transitive.74 However, transitivity cuts across the various
categories in the above analysis and, as I have argued elsewhere,75 causation is not
transitive. In addition to the cases where the nonsemifactual counterfactual @A >
@C is true, in which A is a cause of C, there are two other important types of cases,
characterizable counterfactually, where A is a cause of C as well: the set of cases
in which the counterfactual @A >@C as well as the complementary semifactual
@A > C are both false; and the set of cases with true cts. These three counterfactual
categories together constitute the scope of cases where A is a cause of C.76
Although it is extensionally adequate, this analysis of causation is circular, since
the above analysis of counterfactuals is based on the ingredients of the notion of
cause (i.e., if not on the notion of cause itself, then on the notions of causal relevance
and purely negative causal relevance). An analysis of cause in terms of counter-
factuals, which themselves are analyzed in terms of cause (or suciently closely re-
lated notions) is circular. I have argued at length that any adequate analysis of n-d
counterfactuals must respect the constraint that all intermediate events causally in-
dependent of the negation of the antecedent, A (or such that A is purely negatively
causally relevant to them), be preserved among the implicit premises.77 Hence an
adequate analysis of counterfactuals should be based on causal notions, and thus no
noncircular counterfactual analysis of cause is feasible.
Acknowledgments
A much longer version of this essay was written in summer 1997. The account of the
rst 6 sections was presented as a series of two lectures in UNAM, Mexico City, in
early August 1997, and I thank the audience, and particularly Dorothy Edgington,
for their reaction. It was also presented in California State University at Sacramento
and at the University of Washington in October 1997, and at the University of
Miami in April 1998, and I thank the audience, particularly Marc Lange, Alan
Goldman, and Risto Hilpinen, for their reactions. Section 2 was added in August
1998, after a discussion with Ned Hall. Sections 2 and 10 and were added in October
1998. Section 7 was added in the middle of March 1999. I also want to thank Alex
Byrne and Christopher Hitchcock for discussions of some of these themes.
Figure 15.1
380 I. Kvart
Notes
1. I assume a nonrelativistic framework. However, since in general the order of temporal priority is
assumed here, this assumption can be construed as having the eect that the notion of cause is frame
dependent.
2. I thus assume an indeterministic world.
3. In this essay I deal only with token causes and token causal relevance. I employ narrow event individu-
ation, Kim-style (which is called for if token events are to have probabilities), appropriately extended. For
further details, see my (2001b), note 21.
4. The problem arose both in cases of generic causation as well as in cases of token causation. Recourse
to the purported transitivity of causation partly addresses this issue, but only regarding the necessary con-
dition for being a cause, where transitivity might be taken to play a role. However, as various people have
argued (myself included), causation is not transitive. For my arguments against transitivity, see my
(1991b), pp. 125160, section 5, where I argued that owing to diagonalization causal relevance is not
transitive. But causal relevance is a necessary condition for being a cause, which gives rise to the non-
transitivity of cause. And indeed, the notion of cause is vulnerable to diagonalization as the notion of
causal relevance is. Cause is thus not transitive as well. For a dierent counterexample against the tran-
sitivity of causation, see my (1997), section 5, and my (2001a), section 1. Michael McDermott (1995a)
oered the dog-bite counterexample, which, however, I argued, is unsuccessful (see my 2001a, section 1,
and my forthcoming b, section 12). Other successful counterexamples against cause transitivity have been
oered by Ned Hall and Hartry Field (see Lewis 2000, p. 194; Hall, chapters 7 and 9 in this volume).
The problem of causes that lower the probability of their eects, which has often been handled by re-
course to transitivity, is here handled by the requirement of a stable increaser (see below). The corre-
sponding problem of noncauses despite probability increase is handled on my account, in one kind of case,
through the requirement of a stable increaser (and not just ab initio probability increasesee below), and
in another kind of case, through the requirement of causal relevance, which involves another kind of in-
termediate eventsneutralizers (see below). (Peter Menzies noted the failure of raising the probability of
the eect as a sucient condition for being a cause in view of causal chains that are cut o. See Menzies
1989a; see also my 1991a.)
5. For detailed arguments, see my (1991a), which shows that causal relevance depends strongly on the
intermediate course. Cause of course requires causal relevance. See also my (1997), section 8.
6. For a detailed analysis of this notion of cause, understood in terms of ex post facto probability increase
and employing the notion of an increaser, see my (1997), sections 612.
It is desirable to compare the advantages and disadvantages of the approach summarized here with
alternative approaches, e.g., the counterfactual approach (for an indeterministic world, for instance a` la
Lewis) or with the process approach (e.g., a` la Salmon or Dowe). Limitations of space for this essay pre-
clude such a detailed comparison. In my (1997) I have undertaken a fairly detailed critical comparison,
insofar as the notion of cause is concerned, to Lewiss (1986b), Salmons, and Dowes analyses (see sections
1, 2, 3, 4, 15, 16, 17). Suce it to say, with respect to the process approach, that the process account of
cause has a hard time handling instances of causation that are not processlike, such as cases of omissions,
absences, preventions, and double preventions. Attempts to extend the approach by relying on counter-
factuals (e.g., Dowes 2000) to make it possible to address such cases, as well as Salmons original resort
to counterfactuals in characterizing causal processes (in Salmon 1984, chapter 7), lead to circularity when
considered in terms of my analysis of counterfactuals in terms of causal ingredients (see my 1997) and un-
dermine causation as a unied phenomenon once based both on processes as well as counterfactuals. Fur-
ther, noncircular analyses of exactly what in general a process and in particular a causal process is are hard
to come by. Salmons (1984) attempt regarding causal processes relied on counterfactuals, and its exten-
sional adequacy has been criticized (see, among others, my critique in my 1997, sections 3, 4, 15, 17, and
further references there). Still further, attempts to dene causal processes without reliance on counter-
factuals, such as Dowes, seem to be limited to processes characterized in physical terms and to be inca-
pable of moving up to higher levels involving human actions and mental states, thereby limiting the
applicability of the approach. (See also the discussion of Dowes position in my 1997, notes 22, 31, 34.)
7. Assume that xs nancial state at t
A
can be summarized by Z. Strictly speaking, C should be read as: xs
nancial position was signicantly better than Z.
8. This notion, strictly speaking, doesnt carry causal implications and is strictly probabilistic: All it means
is that, when E is held xed, the probability of C with A is higher than with@A.
9. Assume that x answered the questions in their alphabetical order.
10. Menzies noted such insuciency in cases when causal routes are cut o (note 3 above).
11. In discussing intermediate events, I will consider only actual events.
12. I have used the notion of a strict increaser in my earlier writings, but the notion of a stable increaser
seems more suitable.
13. t
A
is the time interval to which A pertains.
14. Thus, one may consider an empty intermediate event as the null increaser in this case.
15. Given causal relevance (see below). I oered this account of stable increasers as a solution to the
problem (for probabilistic accounts of causation) of causes that lower the probability of their eects, a
well-known problem, as well to the problem of events that raise the probability of later events that are
nonetheless not their eects, in my (1994). When I wrote that paper I was not yet entirely sure whether to
commit myself fully to the thesis that a case with some positive causal impact, or a case with a stable
(strict) increaser (given causal relevance), amounts to a case of causation, and I thus proceeded to use,
instead, the stable increasers analysis as a stepping stone to the more complex notion of overall positive
causal impact.
16. Given, of course, causal relevancehere and above regarding causal relevance. For more on mixed
causal relevance, see my (2001a), section 3.
17. Hitchcock (1993) and (1996a). For an analogous example, see Hitchcock (1996b). I assume that the
chances that the patient x would follow his doctors orders were high (and x did indeed follow them). The
way the patient recovered was as expected in both variations.
18. In the@A-case, it is plausible that there is a higher chance that the patient did not act on his disposition
to fulll his doctors orders, and thus the chances that he did not take the medicine at all seems higher than
in the A-case.
19. See my (2001a), note 54. The recourse to A
0
and to the principle of Cause Preservation under Infor-
mational Expansion might be indispensable if a radical enough version of the case is articulated. Notice,
however, that here I do not address explanation. In particular, I do not address contrastive explanation.
Regarding the rst doctor, I do not address the issue of whether the patient recovered because the doctor
prescribed 100 ml. I exclusively address the issue of whether the fact that he prescribed 100 ml was a cause
of the fact that the patient recovered, and this, I submit, is true.
20. See my (forthcoming b).
21. I assume here that the avalanche didnt aect the chance that the second enemy hit x. If, contrary to
this assumption, the avalanche decreased it, A may well have had purely negative causal relevance to C,
yet still without having been a cause of C. For how purely negative causal relevance implies no cause, see
my (1997), sections 6, 7 and 13; my (2001b), section 3; and my (2001a), section 4. For a more detailed
account of this conception of causal relevance, see my (2001b).
22. For further details, see my (2003), sections 7 and 8. Here I abbreviate a causal-relevance neutralizer
as a neutralizer, rather than as a crn, as I did in earlier writings.
23. I use these two terms interchangeably.
24. I.e., its occurrence time must be between A and C.
25. In general (as will be clear once we present the full analysis of a neutralizer), if there is one neutralizer,
there are many. Various informational expansions of a neutralizer also fulll the role of a neutralizer. In
securing causal irrelevance, we need to secure the presence of just one event fullling the role of a neutral-
izer. If a certain event E in fact secures the neutralization of the causal relevance of A to C, but this is not
revealed ab initio, that is, from our usual perspective of (the lower limit of ) t
A
, i.e., given W
A
alone, then
382 I. Kvart
some informational expansion of E secures it ab initio as well if the informational expansion adds further
information that secures that various would-be causal threads were neutralized. Therefore, if causal irrele-
vance is attested to by an event E that in fact secures the neutralization of the would-be causal relevance of
A to C, but is not a stable screener, it will also be attested to by a neutralizer, which is a stable screener
(e.g., by some conjunctive expansion of E). Hence, without loss of generality, we can conne our attention
in securing the condition for a neutralizer to stable screeners.
26. A conjunctive expansion of a stable screener E that yields a stable screener of it is an event E.F, for
some intermediate F.
27. We need not be terribly precise heresee note 24: All we need is to characterize an event that secures
the neutralization of would-be causal relevance. If there is one, there are many, and for establishing causal
irrelevance it doesnt matter which one we pick. Hence we may characterize neutralizers in a way that
brings out the fact that they serve that role and yet do not include ingredients that are immaterial to that
role.
28. Exception: Causal relevance and the presence of a stable increaser are necessary, but not entirely su-
cient, for an events being a cause. It must further be ascertained, when checking for causes, that there is no
purely negative causal relevance despite the presence of a stable increaser. This can happen if the route of
the would-be positive causal relevance (indicated by a stable increaser) is neutralized by a positive relevance
neutralizer. This case can be analyzed with the notions used here to analyze causal relevance, but I will not
discuss it further. See my (forthcoming c). I ignore this complication below.
29. For a motivation for this thesis, see my (2001b), esp. section 5. Stable screeners can be exhaustive
channelers (see there) or neutralizers. Exhaustive channelers typically are such that A is a cause of them
(though not always, since causal relevance transitivity fails, as I have argued in my 1991b, section 5). Thus,
a promising way of excluding exhaustive channelers from the array of stable screeners while retaining
neutralizers is by requiring that A not be a cause of the stable screener.
30. One might say that the presence of an endpoint neutralizer is a conclusive defeater for the prima facie
causal relevance presented by the presence of a dierentiator.
31. For a more detailed presentation of the above analysis of cause, see my (1997), sections 612. For a
more detailed presentation of the above analysis of causal relevance, see my (2001b).
32. See the references in the previous note.
33. Assume x had only one bullet.
34. In fact, many.
35. The following example was given by David Lewis in his (2000), as an example of probabilistic causa-
tion for which he has no analysis.
36. The outcome would have been dierent if the bomb on ight 17 had been discovered and the paper had
printed the headline: Airline bomb disaster; another bomb discovered. But in the story as specied
above, the headline is as specied in C.
37. The classication and the terminology of early and late preemption are due to David Lewis.
38. The example is Ned Halls. In the version of late preemption discussed here, the time of the eect, t
C
(which is t
3
) is a small interval covering just the time of the actual shattering. An analogous treatment
applies when the temporal quantier in C is construed as existential, with t
C
specifying a reference time
rather than an occurrence time, and thus with the eect being construed as nonfragile; see my (2003), sec-
tion 9.
39. The distance d could not be traversed at velocity v within the time interval e under the circumstances.
To take an extra precaution to secure that E is a screener, one can add to E a clause to the eect that the
circumstances surrounding the trajectory of Billys stone during t
E
; t
3
e=10 were as could be expected as
of t
2
, given B. (Of course, if they were dierent, one might need to have a dierent clause regarding such
facts, e.g., that there was no sudden intermediate thrust accelerating Billys rock.) A requisite neutralizer
need not stretch all the way to the upper end of the t
2
; t
3
interval, as is evidenced in e in E. And indeed,
since the temporal interval for E involved e, the upper limit of t
E
was earlier than that of t
C
. Yet one may
opt for a neutralizer for A
2
and C that does not employ such an e; but then one needs to heed the con-
straint on such neutralizers (applying when the upper limit of the occurrence time of the neutralizer coin-
cides with that of C). See my (2003), sections 7, 8.
40. On a natural and plausible reading of the example, A
1
and A
2
are causally independent of each other.
Surely the example is entirely dierent if one assumes that A
2
was a cause of A
1
(with A
2
being earlier).
(Note: The two throws may be epiphenomena of a common cause. But such cases do not pose a problem
for a chance analysis of cause such as the one presented above since whatever is indicated by the earlier
A
i
that might tempt you to think it increases the probability of A
j
is already indicated by W
Ai
[since the
common cause is recorded there], and hence there is no ab initio probability increase.)
41. Christopher Hitchcock, Do All and Only Causes Raise Probabilities of Eects? (chapter 17 in this
volume). t is a time interval consisting of the actual occurrence time of C.
42. Under the circumstances, the distance d could not be traversed at velocity v within the time interval e.
One may take here too an extra precaution of the sort suggested in note 38.
43. Again, on a natural and plausible reading of the example, A and B are causally independent of each
other.
44. That is, given that no opposing signal arrived rst. Later signals dont matter.
45. One can set the pertinent times to suit the example. One can assume that, without an inhibitory signal
from c
1
to d, the signal from c
2
is likely to reach e before (or at least no later than) the signal from c
1
.
46. It can be the passage of the stimulatory signal through an intermediate point.
47. A was in particular a cause of the timing of C, i.e., that it was at 11:00 p.m. that the convict died. This
is a version of Lewiss Kelly example. A similar example is in Michael McDermott (1995a). I have used an
example of a cause that is a delayer (though I didnt call it so) to a similar eect in the case of the wife and
her husbands pills in my (1994), section IV. For an analysis of causes of timing and of delayers, see my
(forthcoming b).
48. In this kind of case there may be a dierence between adherents of narrow and broad event individu-
ation. Adherents of broad individuation may feel ambiguous about whether the doctors action was a
cause of the murderers death: in an obvious sense it was not, yet in another sense, it was a cause of the
timing of the death (since it was a delayer). Narrow individuation, as used here, makes it possible to bring
out the dierence vis-a`-vis the timing factor in an explicit and nonambiguous way. For further elaboration
on the time-and-manner directedness of being a cause (which is also alluded to briey in the next section)
and the way it is brought out in the above probabilistic treatment, see my (forthcoming b).
49. I.e., take T that maximizes the chance of xs hitting Kennedy at t
3
. I take it that, under a plausible
construal of the example, ab initio probability decrease is built into the case. In a variation structured dif-
ferently, in which G raises the chance of C, G may well be a cause of C.
50. G is causally relevant to C, since there is a dierentiator and no neutralizer: Candidates such as xs
bullet is in his gun at t
3
t, for small enough t, are not neutralizers since, intuitively, G is a cause of them.
51. Assume that Oswald was not aware that B (and that x did not in fact make a dierence to the ecacy
of Oswalds ring in any other way).
52. I.e., take T
0
that allows for a nonnegligible probability of Kennedys being shot at t
3
by x given that x
shot at t
3
T
0
.
53. Thus, D is a delayer. In particular, D was a cause of the timing of C; see my (forthcoming b).
54. If one takes D
0
, which is D but with an extended temporal interval T instead of an instant-like t
2
, and
similarly if one takes such an A
0
instead of A and such a C
0
instead of C, then, whether D
0
is a cause of C
0
depends on various features of the case, unspecied so far, that have to do with whether there is ab initio
probability increase for D
0
and C
0
(such as: Who is a better shooterx or Oswald? How much better? To
what extent would a shot by one spoil the chances of the other to aim well and be in a position to shoot?).
That is, the example is then underspecied regarding the presence of a stable increaser.
55. E.g., as between fragile and nonfragile temporal specications, as reected in the dierence between at
a temporal point vs. during (where the latter is construed as an existential temporal quantier). This
384 I. Kvart
dierence in turn reects (especially in cases of nonomissions) the dierence between occurrence times and
reference times; see the previous note and the previous section. For an elaborate treatment of this exibil-
ity, see my (forthcoming b).
56. It may overlap with T.
57. That is, more than a relatively low threshold. The threshold seems to be numeric rather than propor-
tional, e.g., relative to the set of events that are tokens of the cause type, or relative to the set of events that
are tokens of the cause type for which a token of the eect type is also present.
58. This section and the next, which oer a counterfactual analysis of cause, are quite condensed. For a
more detailed and less condensed presentation, see my (2001a).
59. See my (1997), esp. sections 1, 2, 5, 15, and 16.
60. For my rst attempt to analyze causal notions by chance conditions, see chapter 4 of my (1975), which
became chapter 4 of my (1986) (currently available from Ridgeview). (But note endnote 22 in the Errata
there.) The formal analysis of counterfactuals and, in particular, the version presented there of the chance
analysis of related causal notions, appeared in my (1979). The analysis of causal relevance and of purely
positive causal relevance presented in my (1986) was superseded by the accounts I have oered later
(including the one in this essay).
61. Kvart (1986), especially chapter 2. For a summary of the analysis of n-d counterfactuals (see below),
see also my (1992), and especially my (2001a), section 5. This analysis of counterfactuals is conducted
under the assumption of indeterminism.
62. See my (1986), p. 29. What this terminology reects is that alternative courses of events in which@A is
true are pertinent to this kind of counterfactuals only if they diverge from the actual course right before t
A
.
For present purposes, we can assume standard temporal order, i.e., that t
A
< t
C
.
63. Regarding inferability, see my (2001a), note 19.
64. All the way to the upper end of t
C
.
65. This notion of causal independence was analyzed probabilistically in the rst part of this essay (since
causal independence is the converse relation of causal irrelevance). Purely negative causal relevance
amounts to no positive causal relevance and some negative causal relevance. Some positive causal rele-
vance and some negative causal relevance amount to the presence of a stable increaser and a stable
decreaser, respectively, given causal relevance (but note the qualication in note 27).
66. In various discussions of mental causation, causal relevance is considered to be some variant of positive
causal relevance. I take this to be a misnomer, and this is emphatically not the sense of causal relevance I
employ here and elsewhere: negative causal relevance is also causal relevance.
67. This is true apart from any particular analysis. But it follows from the above conception of cause, since
some positive causal relevance implies causal relevance.
68. Of the n-d type.
69. The analysis here is of As being a cause of C, where A and C are true, factual statements.
70. See above, the end of section 1. Purely negative causal relevance amounts to causal relevance without
positive causal relevance. In my (1986) I talked about pp-semifactuals, since I focused there on@A rather
than on A.
71. This example is a variant of example 8 in my (1986), chapter 2. For convenience, I allow A, B, E, etc.
to vary through facts (states, events, etc.) as well as through the sentences that specify them.
72. I assume that the stock, which was of substantial value, was sold in a private transaction (which
remained condential at least until Tuesday evening) to an individual who, in turn, was not at all interested
in further selling the stock and in fact did not sell it (at least until Tuesday evening). The sharp increase of
xs stock investment would allow for his retirement. xs stock was invested in an index fund that reects the
overall performance of the stock market.
73. A and C are true; A is temporally prior to C.
74. In Lewis (1973a) and (1986b). Lewis needs transitivity to handle cases of early preemption. Lewiss
analysis in his (2000) retains his adherence to cause transitivity. For a discussion of Lewiss theory in his
(2000), see my (2001c).
75. See my (1997), section 5, and my (2001a), section 1. Other eective examples against cause transitivity
were oered by Ned Hall (chapter 9 in this volume) and Hartry Field (unpublished).
76. For further details, see my (2000a), section 7 and 8.
77. See, in addition to the previous section, my (1986), ch. 2, sec. V, esp. V.6 and V.7 (pp. 3840). This
holds not just for n-d counterfactuals, but for l-p counterfactuals as well; ibid., ch. 9, sec. I.2 (pp.
236237).
386 I. Kvart
16A Counterfactual Analysis of Indeterministic Causation
Murali Ramachandran
Attempts to analyze indeterministic causation in terms of counterfactual conditionals
have generally met fairly persuasive refutations.1 An account is developed in this
essay that avoids many of the problems encountered so far. Section 1 outlines David
Lewiss counterfactual approach and ve problem-cases for his (1986a) probabilistic
account of causationthese problems motivate the present project. In section 2, the
idea of background chance leads to a notion of time-invariant overall world chance;
the thesis that overall-world-chance-raising is sucient for causation forms the basis
of the new analysis. In section 3 a simple version that handles three of the ve prob-
lems noted in section 1 is considered. Section 4 concerns the issue of transitivity;
the idea of accumulative chance-raising is used to discriminate between causal and non-
causal chance-raising chains. On the resulting account, causation is not transitive (but
it is not intransitive either). Section 5 introduces the idea of potential chance-raisers
as a means of handling cases of overdetermination by nonactual events. The account
arrived at in section 5 is intended to work under the assumption that no event is de-
termined. Section 6 proposes a simple modication in order to accommodate deter-
ministic causation as well. Section 7 concludes the essay with a couple of caveats.
1 Lewiss Approach
1.1 Deterministic Causation
Lewis (1973a) analyzes deterministic causation in terms of counterfactual dependence,
this in terms of counterfactual conditionals, and these in terms of possible worlds plus
degree of similarity between these worlds and actuality. In short: For any actual,
distinct events C and E, C causes E if
(1) E (counterfactually) depends on C, that is: if
(1a) If C were to occur, then E would occur; and
(1b) If C were not to occur, then E would not occur, that is: if
(1a*) The nearest (i.e., most similar) C-worlds are E-worlds; and
(1b*) The nearest not-C-worlds are not-E-worlds.
C causes E i there is a series of events [C, X
1
; . . . ; X
n
, E] such that each event in the
chain depends on the former one. Here are some of the salient denitions and fea-
tures of the account:
The nearest A-world is a world that diers just enough from actuality so as to ren-
der A true. (So, (1) and (1*) are trivially true if C and E are actual events, for the
nearest C-world will in that case be the actual world itself, which is also an E-world.)
Dependence is asymmetric: If A depends on B, B cannot also depend on A. (This

enables Lewis to resolve the problems of eects, epiphenomena, and simple cases of
preemptive causationsee Lewis 1986a, pp. 170172.)
Causation is a relation between events; events are particularsthey have specic

spatiotemporal locations; but it is important that where and when an event occurs
are generally not essential to its occurrence: It might have occurred at a (slightly)
dierent time and place.
The analysis identies what are loosely speaking causes; it does not identify
dominant causes, or the cause.
Singular (brute, one-o) causation is permitted: There may be causation without

covering laws (or regularities).
Causation at a distance is permitted: Causes and their eects need not be linked
together by a series of spatiotemporally contiguous events.
Simultaneous and backward causation is permitted: Causes need not precede their
eects.
The transitivity of causation is assumed: If A causes B, and B causes C, then A

causes C.
1.2 Indeterministic Causation
The world is indeterministic if there are actual events that might have failed to occur
without violation of any actual laws. A natural way of capturing this idea in terms of
chance is to say that the world is indeterministic if there is at least one actual event,
E, such that the chance of Es occurring, given that its (actual) causes have occurred,
is less than 1. Indeterminism denies the suciency of causes for their eects, not their
(nearby) necessity; in other words, an event, C, may not guarantee the occurrence of
an eect, E, but this is not to deny that if C had not occurred, then E would not have
occurred. So, one may wonder why Lewiss original theory needs to be modied at
all. The problem arises when E has a background chance of occurring, as in the fol-
lowing example:
The nucleus of an atom in a radioactive substance is bombarded by a subatomic
particle (event C) and the atom decays (event E). Event C greatly increased the
chance of event E occurring, but E had a nonzero background chance of occurring
388 M. Ramachandran
regardless; hence, if C had not occurred, E might still have occurred. (See Mellor
1995, pp. 5253, for an informed outline.)
Lewis wants C to count as a cause of E in such cases, where E has a greater chance
of occurring because of Cs occurrence. So, he incorporates counterfactual chances
into his account. The new account is given by the following denitions. For any dis-
tinct events C and E,
(Df1) C raises the chance of E i the chance of Es occurring would be much
higher if C were to occur than if C were not to occur.
(Df2) A series of events [X
1
; . . . ; X
n
] forms a chance-raising chain (CR-chain) i
each event in the chain raises the chance of the subsequent one. (X
1
is then said to
chainwise-raise the chance of X
n
.)
(Df3) C causes E i C chainwise-raises the chance of E.
Two features of the notion of chance at play here are signicant. First, Lewis takes
the chance of Es occurring, ch(E), to vary over time; so, for example, Es occurrence
might have been very unlikely ten minutes ago, but quite likely two minutes ago. So,
in the above denition of chance-raising, ch(E) is to be assessed immediately after the
time of Cs actual occurrence (t
C
). The second signicant feature is that for any
time t after the period over which E might have occurred, the chance at t of Es
occurring is 1 if E did occur, or 0 if E failed to occur.
1.3 Problems for Lewiss Theory
I will outline ve problems for Lewiss theorythese motivate the new account to be
developed in this essay.
Problem 1. Nonoccurrence of (required) intermediary event (g. 16.1) Suppose C, A,
B, and E are neurons linked by very reliable stimulatory axons (signied by the for-
ward arrows). C and A re, B fails to re, and E res spontaneously, unstimulated
(its background chance of occurring is fullled, as it were). Now, at time t
C
, just
C A B E
Figure 16.1
Nonoccurrence of intermediary events.
A Counterfactual Analysis of Indeterministic Causation 389
after Cs actual ring (event C
), the chance of Es ring (E
) is higher than it would

have been at that time if C hadnt red: Thus, C
raises the chance of E
. So, by
Lewiss theory, C
is a cause of E
, which is plainly incorrect, since the C
pro-
cess was cut short.
Problem 2. Timely chance-raising is necessary Suppose C raises the chance of Es
occurring, but that E occurs spontaneously outside the range of Cs inuence; that is,
E occurs before or after C could have caused it. But, by Lewiss theory C nevertheless
comes out a cause of E. The simple solution to this problem we shall be adopting,
roughly, is to make it a necessary condition of Cs causing E that C chainwise-raises
the chance, not just of Es occurring, but also of its occurring when it did.2
Problem 3. Chancy backward causation is precluded by the denition Lewis
thought it a virtue of his account of deterministic causation that it did not rule out
the possibility of backward causation, that is, the possibility of an eect preceding its
cause. However, the present theory rules out direct indeterministic backward causa-
tion, because it is not possible for an event, C, to be a chance-raiser of an earlier
event, E, that had a chance of occurring regardless. Heres why: ch(E) is to be as-
sessed at time t
C
, immediately after Cs occurrence; in the actual world (the nearest
C-world), ch(E) at t
C
will be 1, since E has already occurred by then; if C had not
occurred, E might still have occurred when it did (by hypothesis), so ch(E) at t
C
might still have been 1.
Problem 4. Causation is transitive by the denition On Lewiss account causation
is transitive by stipulationfor, if there is a CR-chain between A and B and between
B and C, there will be one between A and C. But there appear to be counterexamples
to transitivity. Here is a variation on Michael McDermotts (1995a) dog-bite exam-
ple. Singh was due to detonate a bomb but is involved in an accident (event C) that
prevents this. Patel detonates the bomb instead (event D) and the bomb duly ex-
plodes (event E). C raises the chance of D, and D raises the chance of E; thus, C
causes D and D causes E; but it seems plainly wrong to hold that C (the accident) is
a cause of E (the explosion). So, causation is not transitive.
However, this is not to say that causation is intransitive. Suppose Patel has a sti
drink (event F) after pressing the button to calm his nerves. C is a cause of D and D
is a cause of F; and in this case, it seems right to say C is a cause of F. So, to solve
the transitivity problem, we need a way to discriminate between the two sorts of
cases.3
Problem 5. Overdetermination by nonactual events (g. 16.2) Take the diagram in
gure 16.2 to be a neuron diagram as in gure 16.1. Suppose the CE-axon is un-
390 M. Ramachandran
reliable and the DE-axon is very reliable; C res (event C
), D fails to re (D
fails
to occur), and E res (E
). Clearly, C
is a cause of E
; but C
does not raise the

chance of E
; indeed, if C had not red, D might have red, in which case the chance
of E
would have been higher than it actually is.

The account to be developed in this paper is intended to handle all of the above
problems.4
2 Overall World Chance
In the decay example E was said to have a small, nonzero background chance ( b
say) of occurring. I take this to mean that there are nearby worlds with the same laws
where C (the bombardment) does not occur and where E has chance b of occurring
regardless of whether E does occur there or not. In the E-worlds among these worlds,
it would seem quite proper to say that E occurred despite having only b chance of
occurring. On this notion of chance, the chance at a time t of Es occurring is not 1 if
E has already occurred by t; likewise, it is not 0 if t is after the time E might have
occurred and E has failed to occur. The chance, in this sense, of Es occurring is
constant: It is the chance E had of occurring in the world in question.
One may think of such chances as being determined by the laws of the world tak-
ing all the events of that world into account. If the laws of the world are determin-
istic, these chances will be 0 or 1; if the laws are not all deterministic, at least some of
these chances will be between 0 and 1.
We should not think of chance in this sense as simply being the background chance
of an events occurring. The world considered in the decay example is a world in
D
C E
Figure 16.2
Overdetermination by nonactual event.
which both C and E occur. It would seem perfectly proper to say of this world that E
had a much higher than b chance, a, say, of occurring. What chance it had of
occurring does not become 1 after it occurs; it had a chance, period. And it would
seem correct to say, since we are assuming indeterminism, that there is a nearby
world where E had the same chance, a, of occurring, but simply failed to occur.
I shall call chance in this sense overall world chance (or simply owchance), and say
that owch(E) is a at w when the owchance of Es occurring in world w is a. I propose
to analyze causation in terms of owchance-raising, dened as follows:
(Df4) For any distinct events C and E, C raises owch(E) i owch(E) would be
much higher if C were to occur than if C were not to occur.
One of the problems we noted for Lewiss theory was that an event C may raise
ch(E) without raising the chance of Es occurring when it did. The same presumably
goes for owchance-raising. In such cases, it does not seem correct to count C a cause
of E. So, our analysis needs to take the owchance of Es occurring when it did into
account.
For any event E and time t, let owch(E; t) stand for the owchance of Es oc-
curring at time t.5 Where E is an actual event and t
E
is the time of Es actual occur-
rence, owch(E; t
E
) is the owchance of Es occurring when it did.
(Df5) For any event C and actual event E, C raises owch(E; t
E
) i owch(E; t
E
)
would be much higher if C were to occur than if C were not to occur.
The core idea behind the new account of causation then is that raising both
owch(E) and owch(E; t
E
) is sucient for being a cause of E.
3 Causation by Way of Owchance-raising: A Simple Analysis
In developing the theory, I shall make the simplifying assumption that no event is
fully determined, that is, there is no event E such that owch(E) 1 in the actual
world.
The rst point to note is that it is not necessary for causes to be owchance-raisers
of their eects. For example, C will not raise the owchance of E if there is a crucial
intermediary event D that does not counterfactually depend on C. Consider gure
16.3. In this scenario, neurons C, D, and E re (events C
, D
, and E
); suppose D
has a small background chance of ring even if unstimulated. Now, if C
had not
occurred, D
might have occurred spontaneously, in which case owch(E
) would be
the same as, not lower than, its actual value. So C
does not raise the owchance of

392 M. Ramachandran
E
. Yet, if other conditions are metif, say, the events occur at appropriate times
we might well want to count C
a cause of E
. So, being an owchance-raiser is not

necessary for causation.
This result compels us to accept that what are only chainwise owchance-raisers
of an event may be causes of that event, which suggests the following simple analysis.
(Df6) A series of events [X
1
; . . . ; X
n
] forms an owchance-raising chain (an owchain)
i each event in the chain raises the owchance of the next event.
(Df7) A series of actual events [X
1
; . . . ; X
n
] forms a temporal owchance-raising
chain (a towchain) i each event in the chain raises the owchance of the next events
occurring when it did.
Analysis 1: For any actual, distinct events C and E, C causes E i there is a series of
actual events [C, X
1
; . . . ; X
n
, E] that forms an owchain and a towchain.
The analysis resolves three of the problems for Lewiss theory mentioned in section
1.3.
Problem 1: No causation if the process is incomplete Consider gure 16.1 again;
because B fails to re, there is no appropriate owchain between C
and E
. For ex-
ample, owch(E
) would not be lower than it is if A
had not occurred: B
s failure to
occur means that A
had no inuence on E
in the rst place.

Problem 2: Causation requires timely owchance-raising This is explicit in the
specication of the analysis.
Problem 3: Backward causation is not ruled out The analysis does not preclude the
possibility of an event C being a cause of an earlier event E. The analysis does not
consider the chance at a certain time of Es occurring (the owchances in question are
constant over time in any world). Of course, this does not explain how backward
owchance-raising is possible or render it more plausible; but the important point is
that, unlike Lewiss approach, this analysis allows that backward owchance-raising
and hence backward causation are not explicitly ruled out by the very denition of
causation.
C D E
Figure 16.3
Nontransitivity of owchance-raising.
Another result of note is that inhibitors of inhibitors of an event E do not generally
come out as causes of E.
In gure 16.4, C lowers the owchance of D (shown as a line ending in a dot) and of
Ds occurring when it did, and D fails to occur; D, had it occurred, would have
lowered the owchance of E and of Es occurring when it did. If C had not occurred,
D might still have failed to occur, in which case owch(E) and owch(E; t
E
) would not
be any lower. So, C does not come out a cause of E. This contrasts with the deter-
ministic case, where preventers of preventers do come out as causes by Lewiss ap-
proach: If C had not occurred, D would have occurred and prevented Es occurrence;
so, C comes out a cause of E because E straightforwardly depends on C.
Theorists are divided on the question of whether inhibitors of inhibitors are causes,
and intuitions seem to vary from case to case; so, I do not think the above result
presents a decisive point for or against the present account.6 But the account must be
rejected neverthelessat any rate, modiedbecause it does not resolve the tran-
sitivity problem or the problem of overdetermination by nonactual events (problems
4 and 5 in sec. 1.3).
4 The Transitivity Problem and Accumulative Owchance-raising
Let us reconsider the transitivity problem. Our problem was that we wanted the ac-
cident to come out a cause of Patels taking the sti drink, but not a cause of the
explosion, even though there are owchains linking the accident to both. But, as gure
16.5 reveals, there is a signicant dierence between the two chains. If neither the
accident nor Patels pressing of the button had occurred, then the owchance of the
explosions occurring might not have been lower than it isfor, Singh might have
pressed the button instead (an event inhibited if not prevented by the accident). By
contrast, if neither the accident nor Patels pressing of the button had occurred, then
the owchance of Patel taking the sti drink would have been less than it is. This dif-
ference between the owchains suggests a simple modication to analysis 1 to cope
with the transitivity problem.
C D E
Figure 16.4
Inhibition of an inhibitor.
394 M. Ramachandran
(Df8) An owchain of actual events, [X
1
; . . . ; X
n
], is an accumulative owchain i the
owchance of any event in the chain would have been lower than it actually is if none
of the former events in the chain had occurred. (X
1
is then an accumulative
owchance-raiser of X
n
.)
Analysis 2: For any distinct actual events C and E, C causes E i there is a series
of actual events [C, X
1
; . . . ; X
n
, E] that forms a towchain and an accumulative
owchain.7
In our example, the series of events [the accident, Patel pressing the button, the
explosion] is an owchain but not an accumulative owchain; so, the accident does not
come out a cause of the explosion. But, [the accident, Patel pressing the button, Patel
taking the sti drink] is an accumulative owchain, so we get the desired verdict that
the accident is a cause of Patels drinking the hard stu. Let us turn our attention
now to the remaining problem, that of overdetermination.
5 Overdetermination and Potential Owchance-raising
Consider gure 16.2 again. Intuitively, C
causes E
; but it fails come out as such by

analysis 2, for if C
had not occurred, then D
might have occurred, in which case

owch(E
) and owch(E
; t
E
) would have been higher than they actually are. The key
to resolving this problem lies in the fact that the actual values of owch(E
) and
owch(E
; t
E
) are higher than they would have been if neither C
nor D
were to oc-
cur. C
raises these owchances relative to the nonoccurrence of D*; in other words, it

would have raised them but for the possible occurrence of D
. The idea that causes

need only raise owchances relative to the nonoccurrence of other events forms the
basis of the successor to analysis 2.8
accident
Patel pressing the button
Singh pressing the button
explosion
Patels stiff drink
Figure 16.5
Failure of transitivity.
The following denitions, cousins of Df4Df8, simply incorporate the idea of rel-
ative owchance-raising (C and E are distinct actual events and S is a [possibly empty]
set of possible events):
(Df4*) C S-raises owch(E) i C raises owch(E) relative to the nonoccurrence of the
events in S; that is, i owch(E) would be much higher if C were to occur without
the events in S than if neither C nor the events in S were to occur.
(Df5*) C S-raises owch(E; t
E
) i C raises owch(E; t
E
) relative to the nonoccurrence
of the events in S.
(Df6*) A series of events [X
1
; . . . ; X
n
] forms a S-owchain i each event in the chain
raises the owchance of the next event relative to the nonoccurrence of the events in
S. (X
1
X
n
are said to form a potential owchain if there is such a S.)
(Df7*) A series of actual events [X
1
; . . . ; X
n
] forms a S-towchain i each event in
the chain raises the owchance of the next events occurring when it did relative to
the nonoccurrence of the events in S.
(Df8*) A S-owchain of events, [X
1
; . . . ; X
n
], is an accumulative S-owchain i the
owchance of any event in the chain would have been lower than it actually is if
none of the former events in the chain nor any of the events in S had occurred.
Now, causation is dened thus:
Analysis 3: For any distinct actual events C and E, C causes E i there are (possibly
empty) sets of possible events, S and W, and a series of actual events [C, X
1
; . . . ; X
n
,
E] that forms an accumulative S-owchain and an W-towchain.
(The denition allows that in some cases there may be no single set of events S
such that the events in a causal process form both a S-owchain and a S-towchain. Of
course, S and W will be one and the same set on many occasions.)
Simple result 1 If C S-raises owch(E) and C W-raises owch(E; t
E
), where C and E
are distinct actual events and S and W are sets of possible events, then C causes E.
Proof Suppose the antecedent of the above conditional is met. Then, it follows
trivially by denition that [C, E] is a S-owchain and a W-towchain. Now, [C, E] is an
accumulative S-owchain if it is true that the chance of E would be lower than it
actually is if none of the former events in the chain (i.e. C) nor any of the events in S
were to occur. But this condition is already met if, as has been assumed, C and E are
actual events and C S-raises owch(E). Hence, C comes out a cause of E.
396 M. Ramachandran
Simple result 2 If C raises owch(E) and owch(E; t
E
), where C and E are distinct
actual events, then C causes E.
Proof Take S W q, the empty set. Then C S-raises owch(E) and C W-raises
owch(E; t
E
). From simple result 1 we then get that C causes E.
Let us check whether we get the right verdict in our overdetermination example
(gure 16.2). Let S W fD
g. C
S-raises owch(E
): As noted earlier, owch(E
)
would be higher if C
were to occur without D
than if neither C
nor D
were to
occur. The same goes for C
, which likewise W-raises owch(E
; t
E
). It now follows
from simple result 1 that C is a cause of E, as required.
Analysis 3 handles all the problem-cases we have considered so far. But, the sigma-
chance-raising approach, as we might call it, ushers in a fresh problem that calls for a
qualication to the condition specied in the analysis.
Problem 6 Suppose a stone is thrown at a window (event A); someone in the build-
ing, unaware of event A, opens the window slightly (event C)(imagine here a win-
dow that opens outward rather than upward). The stone hits the window, and the
window shatters (event E); however, E would have occurred even if the window
hadnt been opened: The stone would simply have struck and shattered it a bit later.
Intuitively, C (the opening of the window) is not a cause of E (the windows shatter-
ing), but the present analysis delivers the opposite verdict.9 Heres why:
Step 1 Let S W the set of possible but nonactual movements of the stone be-
tween the open-window position and the closed-window position that are prevented
by the actual impact.
Step 2 C S-raises owch(E): owch(E) would be much higher if C were to occur
without any of the events in S than it would be if neither C nor these events were to
occur. If the window is not opened and none of the events in S occurs, then the stone
does not even reach the window.
Step 3 C W-raises owch(E; t
E
).If the window is not opened and none of the
events in W occurs, then there is (close to) zero chance of the window shattering when
it did.
Step 4 By simple result 1, steps 2 and 3 dictate that C causes E, as claimed (but not
as desired).
My strategy for tackling this problem rests on the following observations. There is
a possible owchain from A (the stone-throwing) to E (the window shattering) that
involves (a) the actual movements of the stone up until the impact with the open
window and (b) certain nonactual events that are in S, namely, certain possible-
but-unrealized movements of the stone between the open-window position and the
closed-window position that were prevented from occurring by actual window-
shattering. What is special about this possible owchain is that if none of its actual
events were to occur, then C (the window-opening) would not have been a S-owch(E)-
raiser at all. Cs being a S-owch(E)-raiser depends on the occurrence of these events;
if these events do not occur, the stone does not reach even the open-window position
hence, the owchance of the window shattering will be close to zero whether the
window is open or closed.
So, I suggest that an event, C, should count as a cause of another event, E, if C
and E meet the condition specied by analysis 3 but only so long as they are not in a
situation structurally similar to the above example. Formally:
Analysis 4: For any distinct actual events C and E, C causes E i
(1) there are sets of possible events, S and W, and a series of actual events [C,
X
1
; . . . ; X
n
, E] that forms an accumulative S-owchain and an W-towchain; and
(2) there is no potential owchain [Y
1
; . . . ; Y
j
, Z
1
; . . . ; Z
k
, E] where (2
1
)(2
3
) all
hold:
(2
1
) the Y-events are actual;
(2
2
) Z
1
is non-actual and in S; and
(2
3
) if none of the Y-events were to occur, then C might have occurred but would
not have been a S-raiser of owch(E).
This is the nal analysis I put forward for indeterministic causation. I believe it
handles all the problem-cases that have been mentioned in this essay. So far, we have
been working under the assumption of full-scale indeterminism, the view that no
event is fully determined. The question arises: Can the proposed approach be devel-
oped so as to accommodate deterministic causation?
6 A Counterfactual Analysis of Causation (Deterministic or Not)
At present, the only problem I see emerging from the relaxing of the assumption of
full-scale indeterminism is of the following sort.
Problem 7 Read gure 16.6 as a neuron diagram, and suppose that any neuron has
zero chance of ring if unstimulated and a chance of 1 of ring if stimulated without
398 M. Ramachandran
interference. In this case, A res (A
occurs) and prevents F from ring, thereby

cutting the CE process. Intuitively, A
is a cause of E
, but C
is not. Analysis 4,
however, delivers the verdict that both A
and C
are causes of E
. Heres how C
comes out a cause:

Step 1 Let S W fA
g.
Step 2 C S-raises owch(E
): If C
had occurred without A
, F
and G
would have
occurred, in which case owch(E
) would have been 1; and if neither C
nor A
had
occurred, then E
would have had zero owchance of occurring.

Step 3 C W-raises owch(E; t
E
) (reasoning as above).
Step 4 Hence, by simple result 1 in section 5, C
comes out a cause of E
, as
claimed (but not as desired).
By similar reasoning, letting S W fG
g, say, we get the correct verdict that A
causes E
. To resolve this problem, then, we must nd a signicant dierence be-

tween any S (or W) set one may use to get C
to come out a cause and any S (or W)

set one may use to get A
to come out a cause.

Here is such a dierence. In the case of C
, we can add a nonactual event to the

given S setcall the resulting S#such that C
does not S#-raise owch(E
). For
example, we had S fA
g; add G
to this, so S# fA
; G
g; now, if C
were to
occur without any of the events in S#namely, A
and G
then E
would have 0
owchance of occurring; hence, C
does not S#-raise owch(E
).
By contrast, in the case of A
, there is no nonactual event we can add to a corre-

sponding S set to get a parallel result. The reason for this disparity between A
and
C F G
A B D
E
Figure 16.6
Deterministic preemption.
C
is simple: the A
process is actually completeevery event in it actually

occurswhereas the C
process is incompleteit does involve nonactual events.

This suggests a simple solution to problem 7. All we need modify in the account
we currently have are the S-owchance raising denitions, Df4* and Df5* in section 5;
we simply replace them with:
(Defn4y) C S-raises owch(E) i for any superset of S, S#, containing just
nonactual events in addition to the members of S, C raises owch(E) relative to the
nonoccurrence of the events in S#.
(Defn5y) C S-raises owch(E; t
E
) i for any superset of S, S#, containing just
nonactual events in addition to the members of S, C raises owch(E; t
E
) relative to
the nonoccurrence of the events in S#.
These minor modications are, I hazard, sucient to handle the problems deter-
minism throws up. We need not, the reader will be pleased to note, modify analysis
4s specications at all.
One interesting result is that counterfactual dependence is no longer sucient for
causation. Consider the inhibitor-of-an-inhibitor from section 3 (gure 16.4) again.
We noted that in the deterministic version of that problem, we have counterfactual
dependence: E depends on C because it prevents a preventer, D, of E. By Lewiss
accounts, C is therefore a cause of E. But not by our account. Let S be any set of
events such that C S-raises owch(E)(consider S q, the empty set, for example);
adding the nonactual event D to S will yield a set S# where C does not S#-raise
owch(E). For, owch(E) would be no higher if C were to occur without D than if
neither C nor D occurred. This result is fortuitous, because it means that inhibitors of
inhibitors are treated alike whether indeterminism is assumed or not.
7 Summing Up
The analysis we have arrived at is, so far as I know, the only counterfactual account
on the market that promises to handle the problems mentioned in the essay. Of
course, everything hinges on the coherence of overall world chance, or at any rate, on
the existence of a suitable notion of overall chance that does not vary over time. The
few remarks I have made in support will strike some readers as quite inadequate. It
may turn out that no such notion is tenable, or, what I think is more likely, that the
most feasible candidates are ones that do preclude the possibility of backward cau-
sation. But I hope to have at least shown how the mentioned problems may be
resolved with a suitable, albeit pending, notion of time-invariant chance.
400 M. Ramachandran
One problem I have left well alone: Schaers (2000b) problem of trumping pre-
emption.10 I cannot at present see how any counterfactual account will be able to
handle all the varieties of trumping preemption. But I do not think one should reject
the counterfactual approach on the basis of just this; if, overall, the approach sits
better with our intuitive verdicts on the whole range of test-cases than do other ap-
proaches, one should perhaps question our intuitive verdicts in this special case.11
Acknowledgments
The earliest incarnation of this essay was presented at a workshop on Chance and
Cause jointly organized by the University of Tasmania and the London School of
Economics in September 1999. Other versions have been presented at the University
of Sheeld, the University of Sussex, and the London School of Economics. I thank
the audiences involved for the valuable discussions. Thanks in particular to Stephen
Barker, Helen Beebee, Ron Chrisley, Cian Dorr, Phil Dowe, Jonardon Ganeri, Paul
Noordhof, and David Papineau. Rob Charleston deserves special thanks for exten-
sive comments on late drafts, as does Ned Hall for doing the diagrams for this essay.
Notes
1. Some examples: Lewis (1986b) has been refuted by Menzies (1989a); Menzies (1989a) has been rejected,
if not refuted, by Menzies himself (1996); Noordhof (1998) persuades me that my M-set analysis (1997a,
1998)which I claimed handled indeterministic causation without invoking chances at allignores inde-
terminism proper; and, to my mind at any rate, Noordhof s own (1999) account has decisive counter-
examples (see my 2000; Noordhof 2000).
2. This problem is mentioned in Byrne and Hall (1998, p. 44, fn. 4), Paul (1998a), and Noordhof (1998).
Lewis (2000) now takes it to be sucient for C being a cause of E that Es time of occurrence chainwise
depends on C; the trouble with this proposal is that mere hasteners and delayers of an event come out as
causes of that event. I nd this counterintuitive. The solution proposed here ( pace Ramachandran 1998
and Noordhof 1999) does not have this consequence. See Paul (1998b) for an alternative line.
3. Paul (2000) despairs of nding a solution to this problem on the counterfactual approach and proposes a
radically dierent account that, e.g., makes explicit appeal to laws and takes the causal relata to be fea-
tures of events ( property instances) rather than the events simpliciter. As we shall see, however, the prob-
lem can be resolved with counterfactuals alone.
4. For the record, Lewis (2000), his latest, causation as inuence approach, is vulnerable to all of these
problems, save the problem about timely chance-raisingbut even here, the problem is avoided only at
the cost of ushering in new problems (see note 2 above).
5. Some remarks on owch(E; t). First, owch(E; t) will typically vary with t: so, e.g., one might hold that E
has a smaller owchance of occurring now than in 2 minutes time. Second, since we are taking events to be
particulars, for any event E, there will be times t
1
and t
2
such that owch(E; t) 0 for any time t before t
1
or after t
2
intuitively, t
1
and t
2
mark the time interval within which E might have occurred, or, rather,
had a positive owchance of occurring. Third, for any time t, owch(E; t) will be less than owch(E)unless t
is the only time at which E has a positive owchance of occurring, in which case these chances will be the
same. (I expect owch(E) can be dened in terms of owch(E; t), taking the latter to be a function of time; I
leave the formulation to the experts.) Finally, I have taken t to be an instant for the sake of simplicity;
obviously, an event will have a duration, so it is wrong to talk of its occurring at a particular instant t. But,
so far as I can see, no major diculties arise if we take the t in owch(E; t) to stand for a particular
time interval rather than an instant.
6. I am actually inclined towards the view that inhibitors of inhibitors are causes (see my 1998); but Hall
(2000), for example, is inclined to the opposite view.
7. The reader may wish to test his or her grasp of the denitions by checking that this condition is trivially
met if C raises owch(E) and owch(E; t
E
).
8. This idea is used in the analysis of deterministic causation by Ganeri, Noordhof, and Ramachandran
(1996), (1998) and is utilized for the indeterministic case by Noordhof (1999).
9. I initially presented this as a problem for Noordhof s (1999) account but owned that it might be an in-
escapable problem for the sigma-dependence approach generally (see my 2000, pp. 312313, including fn.
4 on p. 313). The solution oered in this essay gets the approach o the hook.
10. Many of the articles in this collection discuss the problem, so I will not rehearse it.
11. For example, I have mild but persistent misgivings about the fact that the formulations of trump-
ing preemption examples beg the question in favor of the covering-law model of causation. Maybe
reductionist-minded counterfactual-theorists should just deny the coherence of the examples.
402 M. Ramachandran
17Do All and Only Causes Raise the Probabilities of Eects?
Christopher Hitchcock
1 Introduction
According to probabilistic theories of causation, causes raise the probabilities of their
eects. Opponents of probabilistic theories of causation oer putative counterexam-
ples. A moments reection should lead us to expect such counterexamples to be of
two types: (1) causes that appear not to raise the probabilities of their eects; and (2)
events that appear to raise the probabilities of other events, without causing those
events. Almost all of the cases that have been discussed in the literature have been of
the rst sort; these can be eectively handled using resources that are already avail-
able. Counterexamples of the second sort, which have been largely ignored, still pose
a threat. I will explore some options for avoiding these counterexamples, and more
important, argue that they raise fundamental questions about the nature of indeter-
ministic causationquestions that transcend issues about the correctness of any par-
ticular philosophical theory.
2 Probabilistic Causation
The motivation for a probabilistic theory of causation lies in the desire to have a
theory that encompasses causal relationships that are inherently indeterministic or
chancy. The central idea behind probabilistic theories of causation is that causes
raise the probabilities of their eects. Perhaps the simplest and most natural way to
formulate this idea is in terms of conditional probability. C raises the probability of
E, and hence C causes E, exactly if PE=C > PE=@C.
This proposal is subject to a well-known problem. Coughing raises the probability
of lung cancer: Coughers are more likely to be smokers than are noncoughers, and
since smokers are more likely to contract lung cancer than nonsmokers, coughers will
also be more likely to contract lung cancer than noncoughers. Coughing does not,
however, cause lung cancerthe correlation between coughing and lung cancer is
spurious. Spurious correlations can give rise to counterexamples of both sorts de-
scribed in the introductory section.
The standard solution to this problem is to require that causes raise the proba-
bilities of their eects ceteris paribus. There are at least two well-developed ap-
proaches based on this idea. According to one (see, e.g., Cartwright 1979; Skyrms
1980; Humphreys 1989; Eells 1991), C causes E i PE=C&B > PE=@C&B, for a
variety of background conditions B. By conditioning on background conditions, we
hold xed the presence or absence of confounding factors.
Lewis (1986) analyzes the concept of ceteris paribus probability-raising dierently.
According to his theory, C causes E if the probability that E would occur in the
actual world, at the time when C occurred, is higher than the probability that E
would occur at the corresponding time in any of the closest possible worlds in which
C does not occur.1 In this theory single-case probabilities in dierent possible worlds,
not conditional probabilities, are compared. For example, suppose that Barney
smokes, coughs, and develops lung cancer. Barneys coughing is a cause of his lung
cancer if he would have been less likely to contract lung cancer had he not coughed.
But the relevant counterfactual is false: The closest possible world in which Barney
does not cough is one in which he nonetheless smokes. In such a world, the proba-
bility of his developing lung cancer remains unchanged. In what follows, I will not
distinguish between the dierent methods of articulating the basic probability-raising
idea unless necessary.
My focus will be on what I call actual causation: the relation that holds between
two particular events when one does, in fact, cause the other. This is to be distin-
guished from the relation that holds when one event has (or is of a type that has) a
tendency to cause another.2 As we shall see, the most troubling sort of counter-
example to probabilistic theories of causation threatens only theories of actual cau-
sation. Even if the counterexamples cannot be circumvented, probabilistic theories
of causal tendencies remain attractive. Unlike some authors (e.g., Sober 1985; Eells
1991), I nd it unpalatable that these are simply two independent causal relations:
Surely whether one event did in fact cause another has a great deal to do with
whether the one (was of a type that) had a tendency to cause the other (see Hitchcock
1995a for further discussion). This leaves us with good reason to continue pursuing
probabilistic theories of causation even in the face of the diculties discussed below.
3 Making It the Hard Way
Many authors have argued that there are examples in which a cause lowers the
probability of its eect, even when one holds other factors xed. Perhaps the most
famous example, due originally to Deborah Rosen, is presented in Suppes (1970).
(See also Salmon 1984, chapter 7.) A golfer lines up to drive her ball, but her swing
is o and she badly slices the ball, sending it on a trajectory well to the right of the
hole. Her slice decreases the probability that it will land in the cup for a hole-in-one.
By chance, however, the ball bounces o a tree trunk at just the right angle to send it
on a trajectory back toward the cup. As it happened, her slice did cause the ball to go
into the cup, even though the slice lowered the probability of this outcome.
404 C. Hitchcock
This sort of example has received considerable attention. I will present only the
solution that I favor. Probabilistic theories of causation require that we compare the
probability of the eect in the presence of the cause, and also in its absence. But a
cause may be absent in many ways. For example, our golng heroine might have hit
the ball squarely, hooked the ball to her left, or barely nicked the ball (causing it to
move hardly at all); her swing might have been so bad that she missed the ball com-
pletely, or she might have refrained from swinging. These are all ways in which she
might not have sliced the ball. The intuition appealed to in the previous paragraph,
that the slice lowered the probability that the ball would land in the cup, suggests
that the relevant alternative is the one where she hits the ball squarely. But why
should this be so? Perhaps the relevant alternative is the one in which the golfer
refrains from swinging altogether; relative to such an alternative, the slice actually
increases the probability of a hole-in-one. My view (developed in Hitchcock 1993,
1996a,b) is that there is no objectively correct alternative for purposes of probability
comparison. Rather, causal claims are contrastive in nature; they are true or false
relative to a specic alternative. Thus, the golfers slice caused the hole-in-one, rela-
tive to the alternative in which she abstains from swinging, but not relative to the
alternative in which she hits it squarely. In the latter case, we say that the ball landed
in the cup despite the badly sliced shot.
There are a number of advantages to this approach. First, it avoids the need to
select a privileged alternative from among the many alternatives to the candidate
cause. Second, it can accommodate the seemingly incompatible judgments that the
golfers slice caused the hole-in-one and the ball landed in the cup despite the
badly sliced shot are both true causal claims. We can accept both claims because we
can hear them (perhaps with the help of a pragmatic rule of accommodation) as
being made relative to dierent alternatives. Third, the relativity of causal claims to
specic alternatives can account for a curious linguistic feature of causal claims. To
borrow an example from Dretske (1977b), suppose that Susan has stolen a bicycle,
resulting in her arrest. We might agree that Susans stealing the bicycle caused her to
be arrested, while denying that Susans stealing the bicycle caused her to be arrested.
The placement of stress in a causal claim can aect its truth value. According to
the proposed view, the dierent placement of stress points to dierent alternatives.
Susans stealing the bicycle caused her to be arrested relative to alternatives in which
she acquired a bicycle by some other means, but not relative to alternatives in which
she stole some other item.
I conclude that putative counterexamples involving probability-lowering causes
can be eliminated using a renement to probabilistic theories of causation that is
motivated on a number of independent grounds.
Do All and Only Causes Raise the Probabilities of Eects? 405
4 Three Scenarios
Peter Menzies (1989a, 1996) describes a case involving causal preemption in which
a probability-raising event is not a cause. He takes such cases to be isolated and
atypical:3 It is striking how a probabilistic theory of causation, appropriately for-
mulated, follows the contours of our intuitive conception of causation with amazing
accuracywith the exception of the problem cases to do with pre-emption and over-
determination (Menzies 1996, p. 100).
In fact, however, preemption is in no way essential to Menziess example.
James Woodward (1990) describes a case with the following structure. Suppose that
C
1
and C
2
both increase the probability of E, and would do so regardless of whether
the other is present or not. On some particular occasion, C
1
, C
2
, and E occur.
Although both C
1
and C
2
are potential causes of E, could it not be the case that in
fact only C
2
causes E on this particular occasion? That is, couldnt this be one of
those cases where C
1
fails to cause E, not because it is preempted, but simply be-
cause its operation is chancy? If so, then C
1
raises the probability of E without
causing it.4
Woodward oers the following more specic example in a review of Humphreys
(1989):
Scenario 1: Suppose we know . . . that each of two dierent carcinogenic materials C
1
and C
2
[e.g., gyromitrin and diazonium metabolite] can cause [stomach tumors] in mice (E). Suppose
also that the operation of each of these causes is non-deterministic but governed by a stable
probability distribution: each increases the probability of cancer, although to some value
strictly less than one. Suppose also that there is no evidence for any interaction eect between
C
1
and C
2
when both are present. Now suppose that a particular mouse is exposed to . . . C
1
and C
2
and develops cancer E. It follows on Humphreys account that since both C
1
and C
2
increase the possibility of cancer, both cause or have causally contributed to cancer. But why
should we believe this? How do we know that the cancer was not instead caused by C
1
alone
or C
2
alone? We know that when . . . C
1
occur[s] in isolation (without . . . C
2
) . . . , it is perfectly
possible for [it] to increase the probability . . . of E and yet . . . fail to cause . . . E. Here proba-
bility increase is not sucient for actual . . . causation. How do we know that the envisioned
case is not also one of these cases, in which C
1
fails to cause E . . . and E is instead caused by
C
2
[alone]? (Woodward 1994, p. 366; with minor changes in notation. Woodward notes that
the example is originally due to Humphreys 1989, p. 15)
The potentially troublesome implicationthat both carcinogens caused the stom-
ach tumoris not unique to Humphreyss theory, but it is shared by all of the
probabilistic approaches to singular causation mentioned above. Note that the as-
sumption of indeterminism is essential to the example. If, given the relevant back-
406 C. Hitchcock
ground conditions, the presence of diazonium metabolite increases the probability
of stomach tumor to 1, then the presence of gyromitrin cannot further raise the
probability.
Is Woodwards scenario coherent? Could it be that only one of the carcinogens
actually causes the stomach tumor? Consider two potential analogues.
Scenario 2 Maria and Lucinda are neighbors. Every week, each of them buys a
one-dollar ticket for the state lottery, which pays one million dollars as the grand
prize. Due to the diminishing marginal utility of money, Lucinda and Maria decide
that they would each rather have two chances at half a million dollars than one
chance at a million, and so they decide to pool their resources. Each of them buys a
ticket as before, but under the agreement that the prize money will be shared if either
of them wins. In fact, Marias ticket wins, and the neighbors split the million-dollar
prize.
Scenario 3 Desmond and James decide to do something similar. They walk to the
store together, slap their dollars down simultaneously, and purchase two tickets, each
registered in both names. In fact, one of those tickets wins, and the two men share a
million dollars.
In each scenario, how might we address the question: whose dollar purchased the
winning ticket? Although the two women shared the money as agreed, and although
each ticket increased the chance that the prize would be won, there is still a fact of
the matter as to whose dollar had been converted to one million: It was Marias. In
scenario 3, however, there is no such fact. Although only one of the tickets actually
won, there is no fact of the matter about whose dollar had actually purchased this
ticket.
Probabilistic theories of causation presuppose that indeterministic causation works
on analogy with the third scenario. Various causes increase the probability of an ef-
fect by contributing to a probability pool. Once the probability of the eect is de-
termined, the dice are cast, and the event either occurs or it does not. The individual
causes make no additional contribution to the outcome; they bring it about (or not)
only via their contribution to the probability pool. David Lewis and Paul Hum-
phreys both defend the analogy between scenario 1 and scenario 3. Humphreys
(1989) discusses scenario 1 explicitly:
In the example at hand, the situation with both carcinogens present is dierent from the situ-
ation with only onethe chance is higher than with either alone because both chemicals have
contributed to the value of that chance. And that is all there is. To think otherwise is to con-
ceive of the example in terms of a deterministic image where the tumor was entirely caused
by the rst chemical and the second chemical was thereby irrelevant. But that is not the situa-
tion. The second chemical is not irrelevant on this (or any other) occasion for it contributes to
the chance on this occasion, as does the rst chemical, and after they have done this, nothing
else causal happens. It is . . . a matter of sheer chance whether the tumor occurs or not. (pp.
3637)
Lewis rejects an analogous counterexample: [T]he objection presupposes that the
case must be of one kind or the other: either E denitely would have occurred with-
out C
2
, or it denitely would not have occurred without C
2
. But I reject the presup-
position that there are two dierent ways the world could be, giving us one denite
counterfactual or the other . . . (Lewis 1986, p. 180, with minor modications in
notation). Both writers articulate a picture of probabilistic causation in which causes
bring about their eects only via their contributions to a probability pool.
At issue is not merely the correctness of the probabilistic approach to causation
although that is important in itselfbut how we are to conceive of indeterministic
causation in general. This has implications that extend beyond metaphysics into risk
analysis and tort law. Suppose, for example, that a pregnant woman is exposed to a
particular teratogen, which increases the probability that her child will possess a
particular kind of birth defect (which may also occur in the absence of the teratogen).
Assume, moreover, that these are objective probabilities and do not merely reect
our ignorance about background conditions. If the birth defect occurs, is there a re-
sidual issue of whether it was caused by the teratogen, and hence of whether the
woman may be entitled to receive payment for damages? American tort law answers
in the armative, but it is plausible that this answer is founded on a denial of the
assumption that the underlying probabilities are objective and not merely epistemic.
If the relationship between exposure to the teratogen and the formation of birth
defects is indeterministic, however, current jurisprudence may be deeply misguided
on this issue. This critique of existing tort law is carried out in some detail by Para-
scondola (1996), although he assumes without argument that scenario 1 assimilates
to scenario 3.
Although the intuitions of the opposing camps are clear, the case seems inconclu-
sive. In particular, our immediate reaction to scenario 1 is to request more informa-
tion. By what mechanism do the tumors form? How was the mouse exposed to the
carcinogens? Which molecules of which substance made contact with which tissue
cells? Which cells are the tumorous ones? And so on. In the absence of such further
information, we have no clear intuitions about whether this case assimilates to sce-
nario 2 or scenario 3. It will thus be helpful to examine some further examples that
elicit clearer intuitions.
408 C. Hitchcock
5 Three More Scenarios
Consider three more examples in which two potential causes are present and an eect
occurs.
Scenario 4 Suppose that a source emits polarized photons where the angle of po-
larization depends on the physical conguration of the source. The emitted photons
then strike a polarizer that is aligned so that horizontally polarized photons are
transmitted. Initially, the source is congured so as to emit vertically polarized pho-
tons; any such photon will be absorbed by the polarizer. Now suppose that Juan and
Jennifer simultaneously push the source, each pushing it with a force sucient to
rotate the apparatus through 30
. The source emits a photon, which now has a

probability of sin
2
(60
) 0.75 of being transmitted by the polarizer, and the photon

is in fact transmitted.
In this scenario, each push increased the probability that the photon would be
transmitted. It seems clearly wrongheaded to ask whether the transmission was really
due to Juans push rather than Jennifers push. Rather, this case seems to t the
model of probabilistic causality described by Humphreys and Lewis: Each push
increased the probability of the eect (the transmission of the photon), and then it
was simply a matter of chance. There are no causal facts of the matter extending
beyond the probabilistic contribution made by each. This is a very simple case in
which our intuitions are not clouded by tacit deterministic assumptions. If this case
is typical of genuine indeterministic causation, then the sort of case envisaged by
Woodward would indeed be incoherent.
Scenario 5 also supports this conclusion:
Scenario 5 A series of identical ropes are tied to a wall, and it is determined that
when either Gene or Pat pulls on a rope as hard as possible for ten seconds, there is a
fty percent chance that the rope will break. When both pull on a rope for ten sec-
onds, there is a seventy-ve percent chance that it will break. On a given occasion,
both pull and the rope does break. (Woodward, personal communication.)
Once again, our intuition is that it would be wrongheaded to say that on this
occasion the breaking of the rope is due to the pull of Gene rather than Pat or
vice versa. This example is less esoteric than the previous one, although we currently
have no principled reason to think that in a scenario such as this, the probabilities
would be irreducible and have this structure. This is a common problem in dis-
cussions of indeterministic causation: The only examples in which we have prin-
cipled reasons for thinking that there are irreducible probabilities having a certain
structure involve microphysics, where our normal causal intuitions begin to lose their
moorings.
Scenario 6 spells trouble.
Scenario 6 Suppose that two gunmen are shooting at a Ming vase. Each one has a
fty percent chance of hitting the vase, and each one shoots independently, so the
probability that the vase shatters is 0.75. (For simplicity, we will ignore the possibil-
ity that the vase might survive a bullet hit.) As it happens, the rst gunmans shot hits
the vase, but the second gunman misses.
In this example, the second gunman shot at the vase, his shooting increased the
probability that the vase would shatter (from 0.5 to 0.75), and the vase did in fact
shatter. Nonetheless, it seems clear that we should not say that the second gunmans
shot caused the vase to shatter. So here we have an apparent counterexample to
probabilistic theories of causation: The second shot increased the probability that
the vase would shatter, but it did not cause the vase to shatter. This example has the
same structure as Woodwards, but here it is clear that the eect is due to one of the
probability-raising factors, and not to the other. The two gunshots do not simply
contribute to the probability of the vases shattering, after which nothing else causal
happens. Something else causal does happen: The bullet red from the rst gun
strikes the vase, while the bullet from the second gun misses. Note that this example
threatens only probabilistic accounts of actual causation. This example would not
lead us to withdraw the causal tendency claim that shooting at vases causes them to
shatter. Likewise, even if Woodwards intuitions about scenario 1 are correct, that
would not lead us to deny that gyromitrin or diazonium metabolite tend to cause
stomach tumors.
In the next three sections, I will discuss prospects for replying to this counterex-
ample. The results will be seen to be inconclusive: although no line of response suf-
fers a knockout blow, none is wholly persuasive, either. In particular, many of the
responses to be considered must postulate further empirical facts, beyond those pre-
sented above. While we do not know for sure that those facts do not obtain, we
readily judge that the second shooter did not cause the vase to shatter without asking
whether such facts obtain. Moreover, the facts adduced are suciently esoteric that it
is implausible that our judgment about this case tacitly presupposes them.
6 Causal Processes
The shattering of the vase can be traced back to the ring of the rst shooters gun
via the bullet that was red from his gun, whereas there is no analogous process
410 C. Hitchcock
connecting the second shooter to the target. It is natural to point to this asymmetry
in judging that the ring of one of the guns, but not the other, was a cause of the
vases shattering.
Spatiotemporal causal processes, such as the speeding bullets of scenario 6, play a
central role in Salmons (1984) theory of causation (see also Dowe 1992; Salmon
1994). It is not clear, however, how the appeal to causal processes can resolve our
problem. There are processes connecting the second shooter to the vase: photons and
sound waves, for example. Intuitively, these are not processes of the right sort. Un-
fortunately, causal process theories such as Salmons do not have the resources to
formulate the distinction between causal processes of the right sort and those of
the wrong sort; at the very least, such a distinction will presuppose something like
the probabilistic theory of causation that is at issue. Thus, for example, if we try to
get around this problem by saying that there must be a connecting process that
transmits a sucient quantity of energy to shatter the vase, we must then address the
question of what constitutes a sucient amount of energy. The answer to this ques-
tion will almost certainly have to be formulated in terms of counterfactuals or ceteris
paribus conditional probabilities.5
In any event, there are reasons that mitigate against requiring that causation al-
ways involve connection via continuous processes. Suppose that a gunman res at a
human target, but a secret service agent jumps in front of the gunmans bullet,
allowing herself to be hit by it. The agents brave action caused the target to remain
alive, even though there are no processes connecting the agent and the target (except
for irrelevant ones such as sound waves and photons).6 A probabilistic theory of
causation thus cannot be saved by adding the simple requirement that causes be
connected to their eects by continuous processes.
7 Precise Specication of Events
Consider the following scenario:
Scenario 7 Barney smokes, and he also spends a lot of time in the sun. These two
proclivities are not connected; for example, Barney is not forced to go outside in
order to smoke. Barneys smoking increases the probability that he will get lung
cancer. By increasing his probability of getting lung cancer, Barneys smoking
increases the overall probability that he will suer from some form of cancer, and
analogously for his exposure to the sun. In fact, Barney develops skin cancer. A for-
tiori, Barney develops cancer of some form or other.
Since Barneys smoking increased the probability that he would develop cancer of
some sort, and Barney did indeed develop cancer, probabilistic theories of causation
would seem to rule that his smoking caused him to develop cancer. But this seems
clearly wrong: Barneys cancer is skin cancer, not lung cancer, so it was not caused
by his smoking. Although both smoking and sun exposure increased the probability
that he would develop cancer, only Barneys exposure to the sun actually caused him
to get cancer.
Few philosophers would be troubled by this prima facie counterexample to prob-
abilistic theories of causation. Our rst reaction is that such theories yield the
wrong verdict only because we have described the eectBarneys getting cancer
imprecisely. There is a more detailed level of descriptionBarneys contracting skin
cancerwhere probabilistic theories render the correct verdict: Barneys smoking did
not raise the probability of skin cancer, and hence it did not cause Barneys skin
cancer.
Let us formulate the general principle that seems to be at work here:
(PSE) Suppose that on a particular occasion, events occur that instantiate types C
and E. Even if, relative to the relevant background conditions, C raises the
probability of E, if there are more precise specications of the events in question, C
0
and E
0
, such that C
0
does not raise the probability of E
0
, then we should not say
that C causes E.
How, precisely, one ought to formulate this principle of the precise specication of
events will depend on ones theory of eventsI will leave it to the reader to refor-
mulate (PSE) in accord with her favorite theory of events.
Perhaps appeal to (PSE) will enable us to avoid the counterexample of the second
shooter. Although the ring of the second shot increased the probability that the vase
would shatter, there may exist a more precise specication of the outcome (say,
involving the shattering into very small pieces of a certain region of the vase directly
facing the rst shooter) such that the second shot did not increase its probability.7
But although (PSE) seems to provide a natural diagnosis of scenario 7, it does not
seem to get at our reasons for saying that the rst gunmans shot caused the vase to
shatter, while the second gunmans shot did not. Moving from the description Bar-
neys cancer to the more specic Barneys skin cancer seems wholly natural,
whereas the move from the vases shattering to the more detailed level of descrip-
tion seems ad hoc.
Moreover, the application of (PSE) in this case involves an empirical presupposi-
tion: that there is some level of description such that the second gunmans shot made
412 C. Hitchcock
no dierence whatsoever to the probability that the vase would shatter in just that
way. Perhaps the second gunmans shot did increase the probability, ever so slightly,
that the vase would shatter in precisely the manner that it did. If so, (PSE) will be of
no avail. If (PSE) succeeds at all, then, it will succeed because of empirical facts we
have no (current) warrant for accepting, and which appear to play no role in our
straightforward causal judgments about scenario 6.
8 Ebb and Flow in the Probability Pool
There is one regard in which scenarios 4 and 6 are signicantly dierent. In scenario
4, both Juan and Jennifer push the polarizer, yielding a 0.75 probability that the
photon will be transmitted. Nothing further happens to aect the probability of
transmission once the position of the polarizer is set. This ts nicely with the proba-
bility pool conception of indeterministic causation: The two pushes determine the
probability of transmission, and nothing else causal happenstransmission follows
by sheer chance. Time is not literally discrete in this scenario, but it might as well be.
It is implausible that scenario 6 works in this way: We do not imagine that the two
guns re, xing the probability of shattering at 0.75 until the very moment of shat-
tering. Rather, we imagine that the actual trajectory followed by the bullets depends
on random gusts of wind, uctuations of air pressure, or perhaps even on indeter-
ministic uctuations intrinsic to the bullets themselves. The initial 0.5 probability
that each shot will hit the target is an average over all of these possibilities. As the
bullets speed through the air, the probability that each one will hit the vase does not
remain frozen at 0.5, but changes in response to indeterministic events that occur
after the guns were red. The probability pool does not remain calm, but rises and
falls as probability ows in and out.
In cases such as that described in scenario 6, there may be a vast number of inter-
mediate events whose chance occurrence will aect the probability of the outcome.
Perhaps a probabilistic theory of causation that is sensitive to the ebb and ow of
the probability of shattering after the two shots are red could yield the right ver-
dict about scenario 6. Such theories have been oered by Kvart (1997, inter alia),
Menzies (1989a, but repudiated in 1996), and Eells (1991, chapter 6). I will not
describe each of these theories in detail, but will rather focus on the features of the
probability pool that they might exploit in dealing with the case of the two gunmen.
Suppose that there is some time t before the shattering of the vase, such that the
trajectory of the bullet from second gun at t gives it no chance of hitting the vase:
Although the bullet from the second gun emerged from the barrel with a probability
of 0.5 of hitting the vase, by time t, random changes in the bullets trajectory have
put it on a path that will denitely steer clear of the vase. Then, the probability at t
that the vase will shatter is exactly the same as it would have been had the second
gunman never red. Under this supposition, the accounts of Menzies (1989a) and
Eells (1991) correctly rule that second gunmans shot did not cause the vase to shat-
ter. I will not defend these claims in detail here, but will instead focus on a simpler
proposal suggested by Hall (personal communication). Hall recommends a minor
modication of Lewiss counterfactual theory of causation: Instead of requiring
that causes raise the probability of their eects at the time the cause occurs, we
must require that causes raise the probabilities of their eects shortly before the time
at which the eect occurs. In the case at hand, the probability at time t that the
vase will shatter would have been no dierent if the second gunman had never
red; hence Halls proposal rules that the second gunmans shot is not a cause of the
shattering.
Another possibility emerges if instead we focus on the trajectory of the rst bullet.
Suppose now that there is some time t at which the bullet red from the rst gun is
determined to hit the vase. If we hold xed the trajectory of this bullet at time t, then
the ring of the second gun no longer makes a dierence to the probability of shat-
tering: The probability is 1 either way. This move is not ad hoc: The trajectory of the
rst bullet at time t is a cause of shattering, and so should (according to many prob-
abilistic theories of causation) be held xed when evaluating the probabilistic rele-
vance of the second gunmans shot. If the trajectory of the rst bullet at time t is held
xed, the second gunmans shot no longer raises the probability of shattering; hence,
a suciently rened probabilistic theory of causation correctly rules that the second
shot is not a cause of shattering.
While these approaches hold out some promise for handling the case of the second
shooter, each of them requires assumptions about the probabilistic details that were
not included in the original description of the scenario. If the probability that the rst
bullet will miss the vase or that the second bullet will hit remains positive right up
until the time of shattering, the proposed resolutions will not succeed.8 Although the
probabilistic assumptions discussed above are not altogether implausible, our igno-
rance about such matters of probabilistic detail surely outstrips our ignorance of the
causal relations in scenario 6. Our intuition that the rst shot caused the shattering
and the second did not does not depend on such probabilistic details. It is particu-
larly troubling that the proposals discussed in sections 7 and 8 presuppose that cer-
tain probabilities take on values of 0 and 1. One of the motivations for developing
probabilistic theories of causation was the thought that the world, for all we know,
414 C. Hitchcock
might be indeterministic. It might, for all we know, be thoroughly indeterministic;
perhaps everything is possible, no matter how improbable. Surely we should want
our probabilistic theories of causation to apply in such a world as well.
9 The Last Scenario
I present for consideration a nal scenario.
Scenario 8 A box contains two carbon-14 atoms, and a very sensitive detector is
placed near the box. The presence of each atom increases the probability that a decay
event will be detected during some given time interval. In fact, a decay event is de-
tected, and one of the atoms in the box is now a nitrogen atom.
Intuitively it seems that it was the presence of one atom (the one that is now a ni-
trogen atom), and not the other, that caused the detection event.9 Yet in this sce-
nario, none of the maneuvers canvased in sections 6 through 8 seems possible. Is this,
at long last, a clean counterexample to probabilistic theories of causation?
We must proceed with caution here, for we are within the realm of the notorious
measurement problem of quantum mechanics. There are many interpretations of
what is going on in this example, no one of them universally accepted. As it turns
out, those interpretations that lend themselves to the natural conclusion that it is the
presence of one atom and not the other that caused the detection event tend not to be
the ones that close the door on the various maneuvers described above. Suppose we
understand the decay event in terms of a realist, or hidden-variable interpretation
(such as that of Bohm 1952). In this case, one of the atoms decays at some time prior
to the detection event, and a decay particle travels to the detector. There is a spatio-
temporally contiguous process (the decay particle) connecting the atom to the detec-
tor, and there is a more detailed specication of the detection event available (one
including hidden variables).
Suppose, by contrast, that we understand this scenario in terms of a more tra-
ditional Copenhagen-style interpretation. The atoms enter into superpositions of
decayed and undecayed states, thus bringing about a certain probability that a mea-
surement will reveal that a decay event has occurred. When the measurement is per-
formed, the superposition collapses and a decay is in fact detected. On this picture,
the detection of the decay particle is not a distinct event from the decay event itself.
As an indeterministic result of that decay/detection event, one of the atoms is now a
nitrogen atom. The decay of that atom is better thought of as a part of the mea-
surement interaction than as a cause of it.
This is not an exhaustive partition of possible interpretations of scenario 8, of
course; but it should suce to convince us that this is the wrong place to look for a
trouble-free counterexample to probabilistic theories of causation.
10 Conclusion
Note that our question is not one of epistemologyhow do we know in cases like
scenario 6 which potential cause is an actual cause? We do knowotherwise the
counterexamples would have no force. In scenario 6, we know that the second gun-
shot did not cause the vase to shatter because the two events were not connected by
an appropriate type of spatiotemporal process. In scenario 7, we know that Barneys
smoking did not cause his cancer because Barney developed skin cancer, and that is
not the sort of cancer that smoking causes. In scenario 8 (assuming some realistic
interpretation of atomic decay), we know that the presence of one atom but not the
other caused the detection event, because atoms cause detection events by decaying,
and atoms that decay are transformed from an atom of one type to another. In each
case, there is a marker that distinguishes the genuine cause from the spurious proba-
bility-raiser.10 But as metaphysicians, we are interested in providing a general theory
of causation. The markers described above are heterogeneous in nature: in scenario
6, the marker resides in events intermediate between the putative cause and its puta-
tive eect; in scenario 7, it is the nature of the eect itself that marks the actual cause;
in scenario 8, a separate eect (the transformation of the decaying atom) marks
the actual cause of the detection event. For a theory of causation to exploit these
markers, something more must be said about what they have in common in virtue of
which they are causal markers. It will not do to simply say that the actual cause is the
one that is marked as such.
It might be tempting to conclude that probabilistic theories of actual causation
must ultimately fail. But there are still reasons for resisting this conclusion. The suc-
cess of probabilistic theories of causal tendencies, together with the presumed con-
nection between causal tendencies and actual causation, should lead us to resist the
conclusion that actual causation has nothing to do with probability-raising. Even if
we accept the example of the two gunmen as a direct counterexample to probabilistic
theories of actual causation, scenarios 4 and 5 suggest that there are cases in which
the probability pool model articulated by Lewis and Humphreys is appropriate. It
will still be necessary, then, to map out the boundaries of this form of indeterministic
causation, where probabilistic theories of causation directly apply. The counter-
example of the second shooter, therefore, is not oered as a call to abandon prob-
416 C. Hitchcock
abilistic theories of actual causation, but rather as a call for more work to be done.
Whatever the result, our picture of causation in an indeterministic world can only
come into sharper focus.
Acknowledgments
For helpful comments and suggestions, thanks go to Joseph Berkovitz, Phil Dowe,
Ellery Eells, Dan Hausman, Paul Humphreys, Igal Kvart, Laurie Paul, Jon Schaer,
and Jim Woodward.
Notes
1. The condition described is only sucient, according to Lewis. Since we will be primarily concerned with
counterexamples to the suciency of probability-raising for causation, I will not present the more com-
plex, weakened requirement that Lewis takes to be both necessary and sucient.
2. Note, however, that not all of the probabilistic theories of causation mentioned above are intended by
their authors as theories of actual causation.
3. Menzies does also mention problems due to randomly occurring events (1996, p. 100, n. 24), but he
does not state clearly what sort of problems he has in mind.
4. Woodward notes that a similar counterexample is possible without the presence of C
2
, if there is some
probability that an event of type E will occur spontaneously.
5. For a more detailed critique of causal process theories of causation, see my (1995b).
6. A similar example is described in Otte (1986); see also my (1998) and Hall, Two Concepts of Causa-
tion (chapter 9 in this volume).
7. Note that this proposal diers slightly from a similar proposal rejected by Lewis (1986b, pp. 204205).
That proposal was to move to a more precise specication of events in order admit the existence of cer-
tain causal relations otherwise denied by the theory; (PSE) aims to reject the existence of causal relations
otherwise admitted by the theory.
8. If the probabilities remain positive but approach zero in the limit as the actual time of shattering nears,
then suitable reformulations of the proposals discussed above may succeed. I will not explore this possi-
bility further here.
9. Schaer (2000a) presents an example with a similar structure.
10. This line of thought has been pressed on me independently by Joseph Berkovitz and Jonathan Schaer.
See also Hausman (1998), section 13.2.
18Causation, Counterfactuals, and the Third Factor
Tim Maudlin
In his classic paper Causation (1973a), David Lewis urged that regularity accounts
of causation, having continually failed to resolve the counterexamples and problem-
atic cases with which they were faced, be abandoned in favor of an alternative strat-
egy: the analysis of causation in terms of counterfactuals. More than a quarter of a
century later, the time has come to evaluate the health and prospects of Lewiss pro-
gram in its turn.
No one would now deny that there is some deep conceptual connection between
causation and counterfactuals. As Lewis points out, Hume himself (incorrectly) para-
phrased one of his analyses of causation using a counterfactual locution (Hume 1902
[1748], sec. VII):
. . . or in other words, where if the rst had not been, the second had never existed.
In honor of Hume, let us call the counterfactual If C had not occurred, E would not
have occurred, when both C and E did actually occur, the Hume counterfactual.
When we think we know a cause of some event, we typically assent to the cor-
responding Hume counterfactual. Furthermore, our interest in causes often has a
practical aspect: We want to know the causes of events so that we can either prevent
or foster similar sorts of events at other times. Causal claims are therefore deeply
implicated with the sorts of future subjunctives used in practical deliberation: If we
should do X (which we might or might not, for all we now know) then the result
would be Y. The future subjunctive is a close cousin to the counterfactual condi-
tional, since accepting a future subjunctive commits one to accepting the corre-
sponding Hume counterfactual in the event that the antecedent does not come about.
But Humes dictum hardly survives a few minutes contemplation before counter-
examples crowd into view. The sort of counterfactual dependency Hume cites is not
necessary for causation: Perhaps the eect still would have occurred despite the ab-
sence of the cause since another cause would have stepped in to bring it about. The
dependency is also not uncontroversially sucient for causation. If John Kennedy
had not been assassinated on November 22, 1963, he would have still been president
in December 1963. But surely too, if Kennedy had still been president in December
1963, he would not have been assassinated in November of that year: When asked to
consider the latter counterfactual we do not imagine Kennedy killed in November
and resurrected in December. So the counterfactual dependencies go both ways and
the causal arrow only one.
These sorts of problems have been thoroughly chewed over in the literature, and
no detailed review of the strategies used to respond to them is needed here. The
problem of backup (potential) causes can perhaps be solved by using counterfactual
dependency to dene direct causation and then taking the ancestral to dene causa-
tion, or by making the eect qua eect very fragile in some way and arguing that the
backup could not have produced it at exactly the right time or right place or with all
of the right properties, or by appealing to a matching strategy, in which the cause
that is operative when the backup is present is identied by matching it to an other-
wise identical case where the backup is absent (and the Hume counterfactual holds).
Attempts to x the counterfactual analysis have become ever more subtle, and com-
plicated, and convoluted (as perhaps some other essays in this volume will attest). I
have neither the space not the desire to address them all individually.
I do not want to quarrel with these sorts of solutions in detail because I want to
argue that the attempt to analyze causation in terms of counterfactuals of this sort is
wrongheaded in a way that no amount of ne-tuning can x. Causation is not to be
analyzed in terms of counterfactual dependency at all, no matter how many equants
and epicycles are appended to the original rough draft.1
If causation is not to be analyzed in terms of counterfactual dependency, how are
we to explain the systematic connections between judgments about causes and judg-
ments about counterfactuals? Such connections may be secured analytically if cau-
sation can be dened in terms of counterfactuals, but I am denying that this can be
done. Connections would also obtain if counterfactuals could be reduced analytically
to causal claims, but that is an even less appetizing project than the one we are re-
jecting. The only other possibility is that a third factor is involved, some compo-
nent of the truth conditions of counterfactual claims that is also a component of the
truth conditions of causal claims. This third factor would provide the analogue of a
common cause explanation for the systematic connections between causal claims
and counterfactuals: Neither underpins the other, but the third factor underpins them
both.
The prospects for a third factor explanation obviously depend on the identi-
cation of the missing factor. I think it can be identied: What links causation and
counterfactuals by guring in the truth conditions for both is natural law. Laws play
one role in determining which counterfactuals are true, and another role in securing
causal connections. The necessary connexion that Hume sought at the heart of
causation is nomic necessity.2
1 Knowledge of Causation without Knowledge of any Hume Counterfactual
Let us consider what one must know, in at least one case, in order to know what
caused an event. Suppose you know that the laws that govern a world are the laws of
420 T. Maudlin
Newtonian mechanics. And suppose you also know that forces in this world are all
extremely short range: forces exist only between particles that come within an ang-
strom of each other. And suppose particle P is at rest (in an inertial frame) at t
0
and
moving at t
1
, and that in the period between t
0
and t
1
only one particle, particle Q,
came within an angstrom of P. Then, I claim, we know with complete certainty what
caused P to start moving: It was the collision with Q.
Thus: Given that we know the laws and we know some very circumscribed set of
particular facts, we know the cause of Ps motion. But do we know what would have
happened if Q had not collided with P (i.e., if Q had not approached within one
angstrom of P)? We do not. Suppose, for example, that the vicinity of our particles is
chock-a-block full of monitoring equipment, which tracks the exact trajectory of Q,
and jammed with particle-launching devices loaded with particles just like Q and
designed to launch these particles so as to collide with P just as Q did if Q should
deviate in any way from its path. There are (for all I have told you) innumerable
such devices, monitoring the path at arbitrarily close intervals, all prepared to step in
should Q fail to collide with P. I hereby warn you of the presence of these contrap-
tions, unspecied in number and construction, and ask you now whether we know
what would have happened if Q had not collided with P.
We do not have enough information to evaluate this counterfactual, both because
the way Q fails to collide has not been specied and because the exact construction
and disposition of the monitoring devices has not been indicated. Perhaps for many
sorts of deviation, some other particle, just like Q, would have collided with P at just
the same time and in just the same way. Perhaps this is even true for all possible
deviations, or all the sorts of deviation we would consider relevant. So perhaps we
could be convinced that P would have moved at exactly the same time and in just the
same way even if Q had not collided with it. Still, none of this has anything at all to do
with the fact that the collision with Q is what caused P to move. The existence of the
monitoring devices and potential backup particles is simply irrelevant to the claim
that the collision with Q was the cause. In fact, once we know the laws we dont even
care what would have happened if Q had not collided with P: Perhaps P would not
have moved, or perhaps it would have because something else collided with it. The
information we have (viz., the laws of nature in this world) allows us to identify the
cause without knowing that any Hume counterfactual is true.
The counterfactual analyst can respond in several ways. One would be to insist
that we know the causes in this case because we know that some Hume counter-
factual is true, although (owing to the lack of information) we dont know which
one. No matter how many backup systems there are, there must be some point on
the trajectory of Q such that had it miraculously swerved at that point, P would not
Causation, Counterfactuals, and the Third Factor 421
have moved since none of the backups would have had a chance to re. But since I
have not specied the number or location of the backups, how does one know this is
true? What if there is an innite sequence of backups, progressively faster and faster
reacting, monitoring at ever closer and closer intervals?
Lewis (1986b) addresses this sort of criticism, claiming that we need not pay much
heed to our intuitions about such recondite circumstances. Discussing exactly this
sort of problem (as raised by William Goosens) and a case involving action at a dis-
tance, he responds: I do not worry about either of these far-fetched cases. They both
go against what we take to be the ways of this world; they violate the presuppositions
of our habits of thought; it would be no surprise if our common-sense judgments
about them went astrayspoils to the victor! (1986b, p. 203). Lewiss strategy here
deserves some comment. The rules of the game in this sort of analytic project are
relatively clear: Any proposed analysis is tested against particular cases, usually
imaginary, for which we have strong intuitions. The accuracy with which the judg-
ments of the analysis match the deliverances of intuition then constitutes a measure
of the adequacy of the analysis. Unfortunately, it is often the case that the question
of how the intuitions are arrived at is left to the side: Getting the analysis to deliver
up the right results, by hook or by crook, is all that matters, and this, in turn, en-
courages ever more baroque constructions. But if we care about intuitions at all, we
ought to care about the underlying mechanism that generates them, and in this case
it seems plausible that a simple argument accounts for our judgment, an argument
entirely unaected by the existence of backup devices, no matter how numerous. The
case does dier from those of our actual experience, and in such a way that the en-
visioned situation will surely never occur. But the way that the example diers from
more familiar ones makes no dierence whatsoever to the reasoning that allows us to
identify the cause: So long as none of the backup devices res, their number and
complexity is perfectly irrelevant. We have such strong intuitions about the case even
though it is far-fetched because the respect in which it is far-fetched makes no dier-
ence to the method by which the cause is identied. The example shows that causes
are not identied via the Hume counterfactual.
What one would like to say, of course, is that the following counterfactual it true:
If Q had not collided with P and none of the back-up devices had red, then P would
not have moved. And this is correct. So we want to hold xed, in evaluating the
counterfactual, the nonring of potential alternative causes. But in order to know
what to hold xed, we already have to know quite a lot about the (actual and coun-
terfactual) causal structure, since we have to identify the possible-but-not-actual-
alternative-causes to hold xed. But now the project looks pretty hopeless as a way
422 T. Maudlin
of using counterfactuals to get a handle on causation: We already have to bring in
causal judgments when determining which counterfactual to consider.
So if we do not know that the collision with Q caused P to move by knowing what
would have happened if Q had not collided with P, how do we know it? The argu-
ment is quite simple and straightforward, but it is worth laying out. Since the laws of
this world are by hypothesis those of Newtonian physics, we know that particle P,
which is at rest at t
0
, will remain at rest unless some force is put on it. And since the
forces in this world are all short range, we know that no force will be put on P unless
some particle approaches within one angstrom of it. And since P does start to move,
we know that some particle did approach within one angstrom of it, which is why it
started moving. And since Q was the only particle to approach closely enough, we
know that it was the collision with Q that caused P to move. End of story.
There are some counterfactuals implicit in this reasoning. If nothing had put a
force on P it would not have accelerated. That, of course, follows from Newtons
rst law of motion. And if we knew that if Q had not put a force on P nothing else
would have, then we would have the Hume counterfactual. But we do not know this.
So as it stands we can only infer: If Q had not collided with P, either P would not
have started moving or something else would have collided with it. This is not the
Hume counterfactual; hence the headaches for the counterfactual analysis.
Various sorts of sophistication of the original simple counterfactual analysis might
succeed in getting the case right, once enough detail is given. For example, if Q had
not collided with P, but some other particle had instead, then it does seem plausible
that P would not have moved o in quite the same way as it did. This seems plausi-
ble even though we cannot begin to specify how Ps motion would have diered. It
also seems plausible that there is some moment in Qs trajectory late enough that,
had Q miraculously disappeared, it would be too late for a backup to reeven
though we have no clue about how late that point must be. So maybe these sorts of
analysis can get the right result. But the curious thing is that even if they do get the
right result, it is obviously not on the basis of that result that we judge Q to be the
cause. Rather, it is because we already judge Q to be the cause that we think that, with
enough detail, we would get the right result. This is obvious since our judgment that
Q is the cause is secure even before the details are lled in. And the advantage of
the little argument rehearsed above is that it explains how we know that Q was the
cause, given the laws of motion and the fact that only Q collided with P, without
having to know any more details about the situation. So even if it does turn out that
had Q not collided, P would have at least moved dierently, that is irrelevant (given
the laws) to identifying the cause.
Since it is facts about the laws that help us identify the cause in this case, and since
laws are obviously deeply implicated in the evaluation of counterfactuals, I suggest
that we stop trying to analyze causation directly in terms of counterfactuals and
consider anew how laws play a role in determining causes. At one end of the spec-
trum we have seen how the laws of Newtonian mechanics can enable one to iden-
tify the cause in the example discussed above, even if one does not know any Hume
counterfactual. Let us now look at an example from the other end of the spectrum.
2 Knowledge of Counterfactuals without Knowledge of Causes
If an analysis of causation in terms of counterfactuals were correct, then, unless the
analysis itself contains some vague concepts, xing determinate truth values for all
counterfactuals should x the truth values of all causal claims. Of course, in many
circumstances counterfactuals themselves may not have classical truth values: The
antecedent or consequent of the conditional may be too vague. But supposing that
the truth values of all relevant counterfactuals are sharp, then one would suppose
that an analysis of causation in terms of counterfactuals should render the causes
sharp. But we can imagine situations in which all relevant counterfactuals are deter-
minate but causal claims are not, because it is unclear exactly what the laws are. The
example is rather articial, but illustrative.
Recall the rules of John Conways Game of Life. Life is played on a square grid,
using discrete moments of time. At any moment, each square in the grid is either
empty or occupied. At any given moment of time, whether a square is occupied or
not depends on the how that square and the eight immediately adjacent squares were
occupied at the previous moment. For example, if four or more of the adjacent
squares are occupied at one instant then the given square will be empty at the next
instant, and if one or none of the adjacent squares are occupied then at the next
moment the square will be empty. Conways rules cover all possibilities, so the state
of the grid evolves deterministically through time.
Now imagine a world in which space and time are discrete, and in which space
consists in a rectangular grid of points that persist through time. And imagine that
there is only one sort of matter in this world, so that every discrete part of space is, at
any moment, either empty or occupied by the matter. And imagine that the patterns
of occupation evolve in denite deterministic patterns similar to those in Conways
game, but somewhat less orderly. As in Conways game, one can predict with cer-
tainty whether a given point will be occupied or not at an instant if one knows only
the previous pattern of occupation of that point and the eight adjacent points. But
424 T. Maudlin
unlike Conways game, the rules of evolution cannot be distilled into a few simple
rules.
In principle, the rules of evolution must specify for each of the 512 possible pat-
terns of occupation of a 3-by-3 grid whether the central square is or is not occupied
at the next instant.3 Imagine picking the rules in a random way: For each of the 512
possible input patterns, ip a coin. If the coin lands heads, the central square will be
occupied the next instant, if tails it will be empty. One now has a complete set of
rules that will determine how patterns evolve through time. Now imagine a physical
world of innite extent and with a completely variegated state of occupation whose
evolution in time always conforms to these rules.
(It is a key point that we are now imagining a world whose evolution everywhere
conforms to these rules, but we are reserving judgment about whether we ought to
say that the evolution is generated by these rules.)
If we have chosen the rules randomly, then it is overwhelmingly likely that there
will be patterns of occupation that dier only in a single location, but which both
yield the same result. For example, suppose it turns out that both patterns A and B in
gure 18.1 are always succeeded by the central square being occupied. The question I
want to ask is this: In such a world, is the bit of matter in the bottom central location
of pattern A a cause (or an essential part of a cause) of the central square being oc-
cupied the next instant or not?
In the course of this inquiry, I will take a few points to be uncontroversial. First, I
take it that the sort of physical world I have described at least could be a world that
follows laws of the evolution of the distribution of matter. On anything like the Mill
RamseyLewis account of laws, I take it that if the world is extensive and variegated
Figure 18.1
enough, then there must be laws: After all, of all possible distributions of matter
throughout space and time, only a set of measure zero will conform to these rules
through all time. It is therefore extremely informative to know that these rules are
always obeyed. I further assume that having granted that the world is law governed,
we grant that the truth values of all counterfactuals concerning the evolution of dis-
tributions of matter are determined. Since the rules are deterministic, we can use
them to determine how any distribution of matter would have evolved, even if that
distribution never actually occurs in the world we are imagining. So at the funda-
mental physical level, there is no dispute about counterfactuals: Given any complete
pattern at a time, we know how it would have evolved had it been dierent in some
precisely specied way.
My central claim is this: Even though there is no dispute, at the fundamental level,
about counterfactuals, there can still be a dispute about causation, and further, this
dispute about causation arises as a consequence of a parallel dispute about laws. The
issue here is not how the case ought to be judged, or what our intuitions are about
the case, or whether a particular theory gets the intuitions right. The issue is rather
that unproblematic knowledge of all counterfactuals in this case does not seem to
settle the question of causal structure, and furthermore, unproblematic knowledge of
counterfactuals does not settle what the laws are in this same case. This again sug-
gests that it is the laws, not the counterfactuals per se, that underwrite causal claims.
There are two lines of argument that can be addressed to the question of whether
the bit of matter in the bottom central location of pattern A is a cause of the suc-
ceeding occupation or not. The rst line makes straightforward appeal to the Hume
counterfactual, and denies that the matter in that location is a cause. It is beyond
dispute in this case that the Hume counterfactual is false: Even if the bottom central
location had not been occupied (i.e., even if what was pattern A had been instead
pattern B), still the central square in the succeeding instant would have been oc-
cupied. In this sense, it is beyond dispute that the presence or absence of matter in
that location makes no dierence to how the distribution of matter will evolve. Fur-
thermore (this argument continues) although sometimes the Hume counterfactual
can fail to be true for genuine causes and eects, that is only because causation is the
ancestral of direct causation, and in direct causation the Hume counterfactual is al-
ways true. But since space and time are discrete here, and there are no intervening
moments of time or intervening places in space, if there is any causation it must be
direct causation. There is simply no place for there to be a chain of direct causes be-
tween pattern A and the succeeding occupation of the central square. So since there
is no indirect causation (no room for it) and there is no direct causation (the Hume
counterfactual is false) there is no causation at all linking the matter in the bottom
426 T. Maudlin
central location to the matter occupying the central square in the next instant. Hence
the case against causation.
Now the case for causation. In arguing in favor of the matter in the bottom center
being a cause, one would rst point out that in general the laws of evolution in this
world seem to link complete 3-by-3 patterns of occupation to the succeeding state of
the central square. If the state of the central bottom location were never relevant to
the evolution, then there would be little dispute about its causal ecaciousness. But
suppose that patterns C and D in gure 18.2 are such that C is always followed by an
occupied central square and D by an empty one. In this case, the Hume counter-
factual holds, and the matter in the bottom central square of pattern C would be
among the causes of the occupation of the central square in the next instant. The idea
is that the complete set of 512 transition rules is so miscellaneous that there is no
signicantly more compact way to convey it than by specifying each of the 512 rules
separately. It does not, for example, make things signicantly simpler to try to
bundle patterns A and B together into one rule (which refers to only the pattern of
occupation of the eight locations other than the bottom central one) since one would
have to indicate, either explicitly or implicitly, that in this case the state of the bot-
tom central location need not be specied, although in other cases it must be.
So the advocate of causation argues as follows. The overall structure of the evolu-
tion in this world suggests that the laws of nature connect complete 3-by-3 patterns
to the succeeding state of the central square. (After all, that is how we generated the
rules: by ipping a coin anew for each of the 512 patterns.) Since there is no inter-
esting generic pattern to the rules, the most uniform thing to do is to regard the fun-
damental laws of this world as all rules that take complete 3-by-3 patterns as input
and give either an occupied or empty central square as output. From this point of
Figure 18.2
view, it is just a coincidence that pattern A and pattern B, which dier only in the
state of the bottom central square, both lead to occupied central squares. (After all, if
we generate the rules by ipping a coin, it is just a coincidence that they do.) As far
as the laws of nature are concerned, pattern A and pattern B have nothing in com-
mon at all; they are, as it were, each atomic and distinct. From this point of view,
then, the state of the central bottom square is of vital importance, for it determines
whether the transition falls under the law governing pattern A or the law governing
pattern B, where these laws are regarded as fundamentally distinct. If an A pattern
gives rise to an occupied central square, then although it is true that the square would
still have been occupied had the bottom central location been empty, that would
have been a case of an alternative and distinct cause of the same eect. The B pattern
is a sort of backup mechanism, on this view, which would produce the same eect
as the A pattern had the A pattern been dierent in a specied way. But whether the
transition from the B pattern counts as alternative mechanism or as an instance of
the same mechanism as the transition from the A pattern depends on what the laws
are that govern the situation. Both transitions could be instances of the operation of
the same law, or instances of two distinct laws, and our causal judgments depend on
which we take to be the case.
We saw above that one route to save the counterfactual analysis was to add
something to the antecedent: that all potential backup causes be held xed as non-
ring if in fact they did not re. We saw that this spells trouble for trying to use the
counterfactuals to analyze causation. Now we are seeing that what counts as an
alternative or backup cause may depend on the nature of the laws governing the
situation, so it is not puzzling that analyses that leave the laws aside will run into
trouble.
I do not wish to adjudicate between the arguments pro and con the causal eec-
tiveness of the bottom central location in this case; I rather simply want to acknowl-
edge that there can be a dispute, a reasonable dispute, in a case like this where the
counterfactuals are all beyond question. I do think, as a matter of fact, that the pro
causation argument would eventually win out, if no interesting generic metarules
could be found that simplied the presentation of the 512 dierent transitions. But
even if one thinks that the con side would win, it is hard to deny that each side has a
point, and that the point ultimately turns on how one is to conceive of the laws of
nature in such a world.
We can easily imagine the 512 dierent transitions being such that they can be
encapsulated by very simple rules, such as the Conway rules mentioned above. And
we can imagine intermediate cases, where instead of giving all 512 transitions sepa-
428 T. Maudlin
rately, one can give a few very complicated, generic rules. And we can imagine the
case discussed above, where no interesting simple metarules exist. There is a slippery
slope between these cases. I take it that on Lewiss account of laws, there will be
points along this slope where it will be indeterminate what the laws are: points where
there will be reasonable disagreement about whether the fewer complicated generic
rules are simpler overall than the 512 separate transition rules. And where the laws
are in dispute, it may turn out (as it does above) that the causes are in dispute, all
while the truth values of the counterfactuals remain unquestioned.
So our pair of examples displays a nice symmetry. In the Newtonian case, we
showed that we can know the causes without knowing any Hume counterfactuals if
only we know the laws. And in our modied Game of Life, we have shown that we
can know all the counterfactuals without being sure about the causes if there is a
dispute about the laws. Between the two of them, they make a strong case for the
idea that it is the laws rather than the counterfactuals that determine the causes.
3 The Role of Laws
If we are to explain the ubiquitous connections between causal claims and counter-
factuals by appeal to the role of laws in the truth conditions of each, we must at least
sketch how those truth conditions go. In the case of counterfactuals the basic story is
extremely obvious and requires little comment. If I postulate that a world is governed
by the laws of Newtonian mechanics, or the laws of the Game of Life, then certain
conditionals are easy to evaluate: those that postulate a particular complete physical
state at a given time as the antecedent and some other particular or generic physical
state at a later time as the consequent. Given the state specied in the antecedent, the
laws then generate all later states, and the sort of state specied by the consequent
either occurs or does not. That is, one uses models of the laws as possible worlds
and then employs the usual possible worlds semantics for the truth conditions of the
conditional.
If the antecedent is not stated directly in the physical vocabulary, or is vague or
generic, or is only a partial specication of the state of the world, then various sorts
of lling in must be employed to yield a set of states that satisfy the antecedents,
which then generate a set of models, and again things proceed in the usual way.
There is much to be amplied in this sketch, and emendations needed for proba-
bilistic laws, and so on, but the role of laws in the whole process is evident enough.
Let us omit any more detail about counterfactuals, then, and turn to the much more
interesting topic of causation.
I do not think that there is any uniform way that laws enter into the truth con-
ditions for causal claims, as there is a fairly uniform way they enter into the truth
conditions for counterfactuals. Rather, I think that laws of a very particular form,
wonderfully illustrated by Newtons laws of motion, support a certain method of
evaluating of causal claims, whereas a more generic and somewhat less intuitive use
must be found for other sorts of laws. As we will see, the laws of the Life world lack
all of the interesting characteristics of Newtons laws, so we will again use them for
illustrative purposes.
Lets start, then, by returning to our colliding particles. Why are we so condent in
identifying the collision with particle Q as the cause of particle P being set in motion?
Recall the exact form of Newtons laws, the form that makes them so useful as aids
for tracking causes in a case like this. The rst law, the law of inertia, states that a
body at rest will remain at rest and a body in motion will continue in motion at a
uniform speed in a straight line, unless some force is put on it. The rst law species
inertial motion, that is, how the motion of an object will progress if nothing acts on
it. The second law then species how the state of motion of an object will change if a
force is put on it: It will change in the direction of the force, and proportionally to
the force, and inversely proportionally to the mass of the object. Note that the sec-
ond law is in a way parasitic on the rst: The rst species what is to count as a state
of motion (uniform motion in a straight line), and the second how, and in what cir-
cumstances, the state changes.
The structure of Newtons laws is particularly suited to identifying causes. There is
a sense, I think, in which the continuation of inertial motion in a Newtonian universe
is not caused. If a body is at rest at one time, and nothing acts on it (i.e., no force acts
on it), then it sounds odd to ask what causes it to remain at rest. It sounds odd to say
that the bodys own inertial mass causes it to remain at rest, since there is no force
that the mass is resisting, and the inertial mass is just a measure of a bodys resistance
to force. And it sounds odd to say that the law of inertia itself causes the body to
remain at rest, since it seems similar to a category mistake to ascribe causality to
those very laws. Of course, the body remains at rest because no force acts and be-
cause the inertial state of motion of a body at rest is to remain at rest, but it certainly
sounds odd to cite the absence of forces as a cause of remaining at rest.
Or at least, if there is any cause of a body at rest remaining at rest in a Newtonian
universe it is a sort of second-class cause: The rst-class Newtonian causes are forces
(or the sources of forces), and what they cause, the rst-class form of a Newtonian
eect, is a change or deviation from an inertial state of motion. There is no doubt
that for Newton, once the rst law is in place, one can ask what causes the Earth to
orbit the Sun (rather than travel at constant speed in a straight line), and that the
430 T. Maudlin
cause in this case is the gravitational force produced on the Earth by the Sun. It is
this sort of conceptual structure that allows us to so readily identify the cause of the
motion in particle P above: Since the inertial state of P is for it to remain at rest, its
change into a state of motion requires a cause, that is, a force, and the only force is
provided by Q. Without the law of inertia, it would not be clear that the sudden on-
set of motion in P required any cause at all: Perhaps particles sometimes just spon-
taneously start moving. Or perhaps the inertial state of motion of a particle could be
a jerky motion: The particle moves at a given velocity for a while, then stops for a
while, then resumes at the same velocity, so the onset of motion in P is just the con-
tinuation of its natural state of motion.
Let us denominate laws quasi-Newtonian if they have this form: There are, on the
one hand, inertial laws that describe how some entities behave when nothing acts on
them, and then there are laws of deviation that specify in what conditions, and in
what ways, the behavior will deviate from the inertial behavior. When one conceives
of a situation as governed by quasi-Newtonian laws, then typically the primary no-
tion of an eect will be the deviation of the behavior of an object from its inertial
behavior, and the primary notion of a cause will be whatever sort of thing is men-
tioned in the laws of deviation.
Laws, of course, need not be quasi-Newtonian. At a fundamental level, the laws
of the Game of Life are not. In that game, if a cell is occupied at one moment
then it has no natural or innate tendency to remain occupiedor to become
unoccupiedat the next moment. The patterns evolve in an orderly way, but the
behavior of the overall pattern cannot be analyzed into, on the one hand, inertial
states of the parts and, on the other, interactions that cause deviations from those
states. There is no division of the rules of evolution into the inertial rules and the
rules for change of inertial state.
That does not, of course, mean that the Life world is without causes. There are
causes and eects, but the way we conceive of them, at the fundamental level, is quite
dierent from the way we think of causes in a quasi-Newtonian setting. If a particu-
lar cell in the Life world (or, more obviously, the modied Life world) is occupied at
a moment, then the cause of that is the complete 3-by-3 pattern of occupation cen-
tered on that cell the instant before, and similarly if a cell is empty. And the cause of
that 3-by-3 pattern a moment before is the complete 5-by-5 pattern of occupation
centered on that cell two moments before, and so on. As we go further back in time,
the reach of the ancestral causes grows outward like a ziggurat, the Life-world ana-
logue of the past light-cone of an event in a relativistic universe. At a fundamental
level, the only proper causes in the Life world are these complete patterns at a
time, whichin conjunction with the lawsgenerate the successive patterns that
culminate in the given cell being occupied or empty at the given moment. (In the
context of the discussion I am assuming that we agree that only a full 3-by-3 pattern
is a cause of the succeeding state of the central square.)
This notion of causation will be available in any world governed by deterministic
laws, whether quasi-Newtonian or not, and so the philosopher hankering after the
most widely applicable concept of causation will likely be drawn to it. It is obviously
an instance of the INUS concept of causation, for example. But I think that we fall
back on this notion of causation only when circumstances demand it: Our natural
desire is to think of the world in quasi-Newtonian terms, in terms of inertial behavior
(or natural behavior) and deviations from inertial behavior: in terms of, to use
a concept from mathematical physics, perturbation theory. Largely, this is because
it is much easier to think in these terms, to make approximate predictions on the
basis of scanty data, and so on. And often circumstances allow us to think in quasi-
Newtonian terms even when the underlying laws are not quasi-Newtonian, or to
think in macrolevel quasi-Newtonian terms quite dierent from the laws that obtain
at the fundamental level. Indeed, the search for quasi-Newtonian laws does much to
explain the aims of the special sciences.
4 Causation and Macrotaxonomy in the Special Sciences
I am a realist about laws: I think that there are laws, and that their existence is not a
function of any human practices. I am also a primitivist about laws: I do not think
that what laws there are is determined by any other, distinctly speciable set of facts,
and that in particular it is not determined by the total physical state of the universe.
And I am a physicalist about laws: The only objective primitive laws I believe in
are the laws of physics. Speaking picturesquely, all God did was to x the laws and
the initial state of the universe, and the rest of the state of the universe has evolved
(either deterministically or stochastically) from that. Once the total physical state of
the universe and the laws of physics are xed, every other fact, such as may be,
supervenes. In particular, having set the laws of physics and the physical state, God
did not, and could not have, added any further laws of chemistry or biology or psy-
chology or economics.
We do not, however, understand the vast majority of what we do understand by
reecting on the laws of physics. For example, there is much that I understand about
how the computer I am now using works, and precious little of that derives from
detailed knowledge of the physics of the machine. Rather, I understand its operation
by thinking about it in terms of some lawlike generalizations, generalizations that
432 T. Maudlin
resemble laws at least insofar as being regarded as supporting counterfactual claims
and being conrmed by positive instances.
In this sense, it is a lawlike generalization about the computer that when it is
turned on and the word processing program is running and there is document open,
pressing a key on the keyboard will be followed by the corresponding letter appear-
ing at the point where the cursor is, and the cursor will move over one space to the
right (unless it is the end of a line, in which case it moves all the way to the left on
the next line, unless it is at the bottom of the window, etc.). This generalization,
which could be made more precise and extensive at the cost of much tedium, is taken
to support counterfactuals: If I had hit the letter z on the keyboard instead of
s just before the last colon, the word that would have appeared would have been
counterfactualz.
No doubt it is by means of such generalizations that we understand how to use the
computer, predict how it will function, explain it to others, and so on. And no doubt
this generalization, albeit lawlike in certain respects, is not in any metaphysically in-
teresting sense a law. If I should hit a key and the corresponding letter did not ap-
pear, then it is not that any law would be broken: Rather, the computer would be
broken (or misprogrammed, or crashed, etc.). And no doubt if the generalization is
correct and the counterfactuals it implies are true, that is ultimately because of the
physical structure of the machine operating in accord with the laws of physics. The
lawlikeness of the macrogeneralizations, insofar as they are lawlike, is parasitic on
the laws of physics in a way that the laws of physics are not parasitic on anything.
The point is that this is how the special sciences work: They seek to impose a tax-
onomy on the physical structure of the world (the concept keyboard, for example,
is not and cannot be reduced to the vocabulary of physics) in such a way that the
objects as categorized by the taxonomy fairly reliably obey some lawlike general-
izations that can be stated in terms of the taxonomy. Generalizations about how
computers, or cumulus clouds, or volcanoes, or free markets behave are evidently of
this sort.
Talk about carving nature at the joints is just shorthand for nding a macro-
taxonomy such that there are reasonably reliable and informative and extensive
lawlike generalizations that can be stated in terms of the taxonomy, and the more
reliable and informative and extensive, the closer we have come to the joints.
Again, I claim nothing particularly novel, or astounding, about this observation.
But if the foregoing analysis is correct, we are now in a position to add something
new. We have already seen that certain forms of laws, namely, quasi-Newtonian laws,
allow the identication of causes to be particularly simple and straightforward. So
insofar as the special sciences seek to use causal locutions, it will be a further desid-
eratum that the lawlike generalizations posited by the sciences be quasi-Newtonian.
The special sciences, and plain common sense as well, will seek to carve up the
physical world into parts that can, fairly reliably, be described as having inertial
states (or inertial motions) that can be expected to obtain unless some speciable sort
of interference or interaction occurs. Or at least, those special sciences that manage
to employ taxonomies with quasi-Newtonian lawlike generalizations can be expected
to support particularly robust judgments about causes.
A few examples. We obviously understand computers in quasi-Newtonian terms:
Depending on the program being run, there is an inertial state or inertial motion, and
that motion can be expected to continue unless some input comes from the keyboard
or mouse. The input is then an interfering factor, whose inuence on the inertial state
is specied by some laws (the program). We similarly understand much of human
biology in quasi-Newtonian terms. The inertial state of a living body is, in our usual
conception of things, to remain living: That is why coroners are supposed to nd
a cause of death to put on a death certicate. We all know, of course, that this
macrogeneralization is only a consequence of the clever construction of the body and
a lot of constant behind-the-scenes activity: By all rights, we should rather demand
multiple causes of life for every day we survive. Nonetheless, the human body, in
normal conditions, is suciently resilient that the expectation of survival from day to
day is a reliable one, and the existence of the right sort of unusual circumstance im-
mediately preceding death is typical. We do sometimes say that people just die of old
age (which is obviously not a normal sort of cause) when no salient cause of death
exists, and our acceptance of this locution illustrates our awareness that the quasi-
Newtonian generalization In the normal course of things (in the absence of forces)
living human bodies remain alive is not really a law at all.
Most critically for our present purposes, we typically think of the operation of
neurons in quasi-Newtonian form: The inertial state of a neuron is not to re, and it
departs from that state only as a result of impinging forces, namely, electrochemi-
cal signals coming from other neurons. These forces come in dierent strengths
and can be either excitatory or inhibitory, and there is a lawlike formula that de-
scribes how the neuron will depart from its inertial state depending on the input to it.
Of course, I do not mean to defend this conception as a bit of real neurophysiology,
but that is how we lay people (and particularly we philosophers) tend to think of
neurons. So conceived, it is often very easy to identify the cause of the ring of a
neuron. If a neuron res and only one excitatory impulse came into it, then that im-
pulse was the cause of ring. End of story. Counterfactuals about what would have
happened had that excitatory impulse not come in (whether, for example, some other
434 T. Maudlin
impulse would have come in) are simply irrelevant. Details of the larger neural net in
which these two are embedded are simply irrelevant. The only thing that would not
be irrelevant would be the discovery that the quasi-Newtonian generalization is false,
for example, because neurons sometimes spontaneously re with no input.
The widespread use of Lewisian neuron diagrams in discussions of the nature of
causation is, from this point of view, both a beautiful conrmation and a deep puz-
zle. We like neuron diagrams because our intuitions about what is causing what are
strong and robust. Those who want to analyze causation in terms of counterfactuals
think that the diagrams are useful as tests of counterfactual analyses: The trick is to
nd some condition stated in terms of counterfactuals about ring patterns that picks
out all and only the causes. Those conditions then tend to get very complicated and
quickly embroil one in issues like backtracking, miracles, intermediate states, and so
on. But it is perfectly apparent that our strong and robust intuitions in this case are
not generated by fancy considerations of counterfactuals at all: They are generated
by the application of quasi-Newtonian laws to the situation, and the counterfactuals
be damned. So the puzzle is why it has not been apparent that the very diagrams
used to discuss counterfactual analyses have not been recognized as clear illustrations
of the wrongheadedness of the counterfactual approach.
The great advantage of the special sciences not being fundamental is the lati-
tude this provides for constructing macrotaxonomies well described by quasi-
Newtonian generalizations. For example, one can try to secure the reliability of a
quasi-Newtonian law of inertia by demanding that one nd an interfering factor if the
inertial state changes. In a case like nding the cause of death this can often be
done. There is a lot of biological activity all the time, and even absences and lacks
can count: One can die of starvation or suocation. There is also latitude in carving
the joints: One can shift boundaries so as to identify systems that obey more robust
generalizations. This sort of boundary-shifting helps explain our causal intuitions in
some cases that have been discussed in the literature.
Consider the following example by Michael McDermott:
Suppose I reach out and catch a passing cricket ball. The next thing along in the balls direc-
tion of motion was a solid brick wall. Beyond that was a window. Did my action prevent
the ball hitting the window? (Did it cause the ball not to hit the window?) Nearly every-
ones initial intuition is, No, because it wouldnt have hit the window irrespective of whether
you had acted or not. To this I say, If the wall had not been there, and I had not acted,
the ball would have hit the window. So between usme and the wallwe prevented the ball
from hitting the window. Which one of us prevented the ball hitting the windowme or the
wall (or both together)? And nearly everyone then retracts his initial intuition and says,
Well, it must have been your action that did itthe wall clearly contributed nothing.
(1995a, p. 525)
McDermotts argument is quite convincing, but a puzzle remains. Why was nearly
everyones initial reaction the wrong one? What were they thinking? Was it merely
the falsehood of the Hume counterfactual, If you had not caught the ball it would
have hit the window, that makes people judge that catching the ball is not a cause
of the ball failing to hit the window? Then why does one not make this error (or not
so easily) when the rst catcher is followed by a second, infallible catcher rather than
a brick wall?4 The Hume counterfactual also fails here (if you had not caught the
ball, the second catcher would have, and it still would not have hit the window), but
it seems clear in this case that you are the actual cause here, and the second catcher
merely an unutilized backup. The pair of examples makes trouble for any counter-
factual analysis, since the counterfactuals in each case are identical, substituting the
second catcher for the wall. So if we judge causes by means of counterfactuals, our
intuitions should swing the same way in both cases, but they dont.
Here is an account that makes sense of the data. In judging causes, we try to carve
up the situation into systems that can be assigned inertial behavior (behavior that can
be expected if nothing interferes) along with at least a partial specication of the
sorts of things that can disturb the inertial behavior, analogous to the Newtonian
forces that disturb inertial motion. Let us call the things that can disturb inertial be-
havior threats: They are objects or events that have the powerif they interact in
the right wayto deect the system from its inertial trajectory. We then think about
how the situation will evolve by expecting inertial behavior unless there is an inter-
action with a threat, in which case we see how the threat will change the behavior. A
threat can itself be threatened: Its inertial trajectory might have it interacting with a
target system, but that trajectory be deected by something that interferes with it.
This is what we mean by neutralizing a threat, or by preventing an event.
Now what counts as inertial behavior, and what counts as a threat, depends on
how we carve the situation up into systems. If we consider a normal window on its
own, its inertial behavior is to remain unbroken (something must cause it to shatter,
but nothing causes it not to if nothing interferes with it). The sorts of thing that count
as threats are objects with sucient mass and hardness and (relative) speed: A hurled
rock is a threat to a window but a lofted marshmallow is not. The cricket ball in the
example is obviously a threat to break the window (although the case just deals with
hitting the window, the same point holds). The cricket ball is furthermore in a state
of motion such that its (Newtonian) inertial trajectory has it hit the window, so for
the threat to be neutralized, something must act to deect the ball. This is clearly the
rst catcher in both cases, and the presence of the wall or second catcher never comes
into the analysis.
But let us now carve up the situation a bit dierently. Let us call the window plus
brick wall system a protected window. The inertial state of the window in a pro-
436 T. Maudlin
tected window system is to remain unbroken, and indeed to remain untouched. To
count as a threat to this state, an object must be able to penetrate (or otherwise cir-
cumvent) the wall, at least if it is approaching from the direction protected by the
wall. Thought of this way, the cricket ball is not a threat to the inertial behavior of
the window; only something like an artillery shell would be. So in the given situation,
there are no threats to the window at all, and a fortiori there are no threats that are
neutralized, and so no prevention: Nothing causes the protected window not to be hit
since that is its inertial state and nothing with the power to disturb the inertial state
threatened.
In sum, if we carve things up one way (window plus wall plus catcher plus ball)
we get systems governed by quasi-Newtonian generalizations that yield the judg-
ment that the catcher prevented the window being hit; if we carve them up another
way (protected window plus catcher plus ball) we get systems governed by quasi-
Newtonian generalizations that yield the judgment that the window was never
threatened with being hit, so nothing had to prevent it. And the interesting thing is
that both systematizations, together with the corresponding quasi-Newtonian general-
izations, yield the same counterfactuals. So no counterfactual analysis of causation
has the resources to account for the disparity of judgments: Carving up the world
dierently can give dierent (special science) laws, governing dierent systems, but it
should not give dierent truth values to counterfactuals.
If this analysis is correct, McDermott has played something of a trick on his test
subjects. Their original intuition that the catcher does not prevent the window being
hit is perfectly correct, relative to conceptualizing the situation as containing a pro-
tected window. McDermott then asks a series of questions that essentially require
that one reconceive the situation, carving it instead into a window, a ball, and a
compound wall plus catcher system. So thought of, the ball is clearly a threat to the
window, the threat needs to be neutralized (or deected) to prevent the window being
hit, and the wall plus catcher system does the deecting. The last step is to give the
credit to either the wall or the catcher or both, and here the catcher clearly wins. All
of this is perfectly legitimate: It is just no more legitimate that lumping the window
and wall together and judging that nothing prevented the hit.
The window plus two catchers example tends to conrm the analysis: Here the
natural tendency would be to regard the two catchers as equally autonomous sys-
tems. It would be odd to carve up the situation so that the window plus second
catcher is regarded as a single system (although one could tell stories that would
promote this). So our intuitions go dierently from the wall case, even though the
corresponding counterfactuals are the same.
The only remaining question is why typical subjects would have the tendency to
regard the situation as containing a protected window, rather than in terms of a
regular window, a threatening ball, and a pair of potential neutralizers. Lots of hy-
potheses come to mind, and the only way to test them would be to try many more
sorts of examples. Windows and walls are obviously more similar than catchers and
walls, since windows and walls are inert bits of architecture. It is simpler to keep
track of things using the protected window systematization, since there are fewer
threats and hence fewer threats that need to be tracked. The protected window
systematization also has the advantage of yielding (in this case, at least) causal judg-
ments that agree with the Hume counterfactual: The catcher does not prevent the
window from being hit, and the window would not have been hit even if the catcher
had not acted. (The last two considerations are held in common with the two
catcher case, where our intuitions go the other way, but in that case there is positive
pressure to group the catchers together.)
A thought experiment recommends itself: Imagine various ways for the wall to
metamorphose into a second catcher and track when ones causal intuitions ip. But
that is a task for another paper.
Alternative taxonomies need not be at the same degree of resolution, as in the
foregoing example. Ned Hall has argued (Two Concepts of Causation, chap. 9,
this vol.) that cases of double prevention (preventing an event that would have pre-
vented another event) are typically not regarded as causation (in the sense of pro-
duction). Jonathan Schaer (2000c) has then pointed out that guns re by double
prevention, but one is not tempted to conclude that pulling the trigger of a gun is not
a cause of the ring. But surely what guides out intuitions in this case are lawlike
macrogeneralizations in quasi-Newtonian form. The inertial law for guns is just: A
gun doesnt re if nothing acts on it (e.g., pulls the trigger, jars it, etc.). Anything that
regularly results in a gun ring (particularly pulling the trigger) counts as a cause
that changes the inertial state, no matter how the trick is done at the microlevel. This
judgment is not reversed even if one decides that the right thing to say at the microlevel
is that pulling the trigger does not produce the ring. If dierent taxonomies can allow
for dierent lawlike generalizations and hence dierent causal judgments, we have all
the makings for interminable philosophical disputes, since causal judgments can be
reversed simply by changing taxonomies, as McDermotts example illustrates.
5 Remote Causation
The foregoing account has been concerned with the analysis of what we may call
proximate or immediate causation: Where a situation is conceptualized as governed
by quasi-Newtonian laws, the laws will specify what counts as an inertial state
and therefore what counts as a deviation from an inertial state (a rst-class eect),
438 T. Maudlin
and also what sorts of things or events (causes) bring about such deviations. The
interfering factor is then the proximate cause of the deviation, as in the billiards
example.
This analysis does not, however, solve the problem of remote causation. We com-
monly identify events or actions as causes that are not the proximate causes of their
eects: The loss of the horseshoe nail ultimately causes the loss of the battle not di-
rectly but by a long chain of intermediates. The loss of the horseshoe nail is only the
proximate cause of the loss of the horseshoe.
Prima facie, the use of quasi-Newtonian laws appears to be of some help in this
regard. If the laws have the form of specifying inertial states and interfering factors,
then any concrete situation can be represented by what we may call an interaction
diagram, a graph depicting inertial motions of objects as straight lines and diver-
gences from inertial motion as always due to the interaction with some interfering
element. One might think, for example, of Feynman diagrams as examples of the
form. Note that if all interactions are local, or by contact, then interaction diagrams
will look like spacetime diagrams, and the lines will represent continuous processes in
spacetime, but this is not essential to the form: if there is action-at-a-distance, then
two lines can intersect on an interaction diagram even if the objects they represent
never come near each other in spacetime.
Interaction diagrams supply one simple method for identifying remote causes:
Trace the diagram backward from an event, and every node one comes to counts as a
cause. But this method will not, in many cases, accord with our intuitions. Being in
the interaction diagram for an event may well be a necessary condition for being a
cause, but it is unlikely to be sucient. Interaction diagrams will include events that
we consider not to be causes (in the usual sense) but failed attempts at prevention,
such as the futile course of chemotherapy that fails to stem the cancer. Interaction
diagrams also stretch back indenitely in time, to events too remote to be commonly
considered causes. They are a bit more parsimonious than the entire back light-cone
of an event, but not that much.
This sort of situation is, of course, the siren song for the analytic philosopher.
Perhaps we can identify commonsense remote causes by use of an interaction dia-
gram plus some further condition (the Hume counterfactual, for example). But the
avalanche of counterexamples to the many theories of causation that have been
oated ought to give us pause. Perhaps there is some reason that the analysis of re-
mote causation has proved so dicult.
So lets step back a moment and ask what purpose there is to identifying remote
causes in the sort of situation we are contemplating, that is, where we know the laws,
the immediate causes of events, and the complete interaction diagram. What more
could we want? What do we gain by trying to distinguish events that count as remote
causes from others that are not?
One thing we might have is a practical concern for prediction and/or control. We
might like to know how we could have prevented a certain event, or whether a simi-
lar event is likely to occur in future situations that are, in some specied respects,
similar. But in that case, all we would really care about is the Hume counterfactual
and that, as we know by now, is neither sucient nor necessary for causation. So
what other aim might we have?
In some circumstances, we are interested in remote causes because we wish to
assign responsibility for an event, for example, when oering praise, or assigning
blame, or distributing rewards, or meting out punishments. In these circumstances,
identifying a remote cause is often tantamount to establishing a responsible agent,
and many of our intuitions about remote causes have been tutored by our standards
of culpability. If this is correct, then we might make some progress by reecting on
such standards.
The game of basketball provides an example where rules for assigning credit as
a remote cause have been made reasonably explicit. Consider this case: A pass is
thrown down court. The center leaps and catches the ball, then passes it to Forward
A, who dunks the ball. Forward A gets credit for the eld goal, and the center gets an
assist: He counts as a remote cause of the points being scored. Note that these attri-
butions do not change even if it is also true that the pass was intended for Forward
B, who was standing behind the center, and that had the center not touched the ball,
Forward B would have caught the pass and scored, or would have passed it to A.
Even if the Hume counterfactual does not hold for the center (even had he not
caught the ball, the points would have been scored), the center counts, unproblem-
atically, as a cause.
Now consider an apparently analogous case. John enters a car dealership, unsure
whether he will buy a car. He is met by an Ocial Greeter, who directs him to
Salesman A. Salesman A makes the pitch and convinces John to buy the car. Had the
Greeter not intercepted him, John would have run into Salesman B, who also would
have convinced him to buy the car (or who, perhaps, would have directed him to
Salesman A). In this case, the Greeter can not, intuitively, claim credit for the sale:
He did not even remotely cause John to buy the car. He was a cause of the cause,
a cause of Johns hearing Salesman As pitch (at least in the scenario where the al-
ternative is a pitch from Salesman B), but we are not inclined to accept transitivity
here.
Doubtless there are perfectly good reasons for the dierence in practices for as-
signing credit in these cases: Typically, in a basketball game, had the assist not
440 T. Maudlin
occurred points would not have been scored, so one wants to recognize and encour-
age those who give assists. But typically, in the sort of situation described in the car
dealership, whether a sale is made does not depend on the actions of the Greeter. But
what is typical does not aect the individual case: We can make the counterfactual
structure of this particular pair of examples as analogous as we like (the schematic
neuron diagrams for the scenarios can be made identical) without changing our
views about who deserves credit and who does not. If so, then no generic account of
remote causation couched in terms of interaction diagrams or counterfactual struc-
ture will always yield intuitively acceptable results: In dierent contexts, our intu-
itions will pick out dierent nodes on the same diagram as remote causes. In these
cases, the denition of a remote cause is, as Hume would put it, drawn from cir-
cumstances foreign to the cause, that is, from statistical features of other cases that
are regarded as similar. As such, the standards violate the desideratum that causation
be intrinsic to the particular relations between cause and eect.
If standards for identifying remote causes vary from context to context, and, in
particular, if they depend on statistical generalities about types of situations rather
than just on the particular details of a single situtation, then the project of providing
an analysis of remote causation is a hopeless task. We might usefully try to artic-
ulate the standards in use in some particular context, but no generic account in terms
of interaction diagrams or counterfactual connections will accord with all of our
strongly felt intuitions.
6 The Metaphysics of Causation
If the foregoing analysis is correct, then (1) what causes what depends on the laws
that govern a situation; (2) judgments of causation are particularly easy if the laws
have quasi-Newtonian form; (3) everyday judgments (intuitions) about causation
are based not on beliefs about the only completely objective laws there are (viz.,
physical laws) but rather more or less precise and reliable and accurate lawlike gen-
eralizations; (4) the same situation can be brought under dierent sets of such gen-
eralizations by being conceptualized dierently, and those sets may yield dierent
causal judgments even though they agree on all the relevant counterfactuals.
To what extent, then, is causation itself objective or real? At the level of
everyday intuition, the freedom to dierently conceptualize a situation implies
that ones causal judgments may not be dictated by the complete physical situation
per se. Further, the lawlike generalizations appropriate to the conceptualization can
be criticized on objective grounds: They could be more or less accurate or clear or
reliable. If, for example, windows sometimes spontaneously shatter (as a result of,
say, quantum uctuations), then the reasoning that the ball made it shatter because
the ball was the only thing to hit it (and its inertial state is to remain unbroken unless
something hits it) is no longer completely trustworthy, the less so the more often
spontaneous shattering occurs. Lawlike generalizations are also supposed to support
counterfactuals, but those counterfactuals must ultimately be underwritten by physi-
cal law, so a close examination of the physics of any individual situation could
undercut the macrogeneralization applying to that case. The quest for greater scope,
precision, and reliability of generalizations tends to force one to more precise micro-
analysis, ultimately ending in the laws of physics, which brook no exceptions at all.
If the laws of physics turn out to be quasi-Newtonian, then there could be a fairly
rich objective causal structure at the fundamental level. But if, as seems more likely,
the laws of physics are not quasi-Newtonian, then there may be little more to say
about physical causation than that the entire back light-cone of an event (or even the
entire antecedent state of the world in some preferred frame) is the cause of the event,
that being the minimum information from which, together with the laws of physics,
the event can be predicted. Quasi-Newtonian structure allows one to dierentiate the
merely inertial part of the causal history of an event from the divergences from
inertial states that are paradigmatic eects, but without that structure it may be im-
possible to make a principled distinction within the complete nomically sucient
antecedent state. In that case, causation at the purely physical level would be rather
uninteresting: If all one can say is that each event is caused by the state on its com-
plete back light-cone, then there is little point in repeating it. None of this, of course,
is of much interest to physics per se, which can get along quite well with just the laws
and without any causal locutions.
There is much more that needs to be said about the role laws, or lawlike general-
izations, play in the truth conditions of causal claims. The proposal I have made is
more of a sketch than a theory, and it has all the resources of vagueness to help in
addressing counterexamples and hard cases. Perhaps there is no adequate way to
make the general picture precise. But the counterfactual approach to causation has
had a good long run, and it has not provided simple and convincing responses to the
problem cases it has faced. Perhaps it is time to try another approach.
Notes
1. Ned Hall has suggested that there are really two concepts of causation, one of which (dependence)
amounts to nothing more than the Hume counterfactual, whereas the other (production) cannot be so
442 T. Maudlin
analyzed (cf. his Two Concepts of Causation, chapter 9 in this volume). If one wishes, this essay can be
read as an account of production, although I suspect that at least many cases of dependence can be cov-
ered using these methods.
2. Natural law can be employed in analyses of causation in many ways. One interesting, and distinct,
approach is advocated by Schaer (2001) in his PROPs analysis: Schaer uses laws to identify the process
that actually brought an event about.
3. If one imposes some natural symmetry constraints, such as 90
rotational symmetry, then there will be

fewer than 512 distinct cases to deal with.
4. This example is discussed by John Collins in Preemptive Prevention, chapter 4 in this volume.
19
Going through the Open Door Again: Counterfactual versus
Singularist Theories of Causation
D. M. Armstrong
1 Introduction
This is a revised version of The Open Door (Armstrong 1999), a paper that criti-
cized David Lewiss counterfactual theory of causation, arguing instead for a singu-
larist theory. Things move quite fast in contemporary analytic philosophy, because
workers on a particular topic are now much more likely to know each others work
and to communicate and inuence each other. This is a great advance. Philosophical
problems are so dicult to resolve that they seem beyond the compass of individuals
working on their own. Ive already found reason to try to rework a number of the
things that I said in the original version.
2 Lewiss Theory
In a volume such as this, devoted to causation, we can summarize Lewiss theory
rather quickly.1 He is concerned with causation in particular cases, one token event
bringing about another, rather than (true) causal generalizations. Causation is the
ancestral of the relation of causal dependence, and causal dependence holds between
events if and only if the counterfactual if C had not occurred, then E would not
have occurred holds, with the counterfactual stipulated not to be a backtracker.
This counterfactual, Lewis asserts, is true if and only if there exists a possible world
where c fails to occur and e fails to occur, a world that is closer in overall similarity
to the actual world than any world in which c fails to occur but e does occur. In-
structions are given for ranking dierent worlds in terms of overall resemblance.
Holding previous history xed, keeping the laws of nature as constant as possible,
and minimizing the size of the miracles required to instantiate the antecedents of the
counterfactuals are the most important ways of maximizing such resemblance.
Lewiss talk of possible worlds here is to a degree misleading. It is important to
realize, as I did not originally realize, and I think many others have not realized, that
these counterfactuals are supposed to hold solely in virtue of features of the world in
which the causal relation holds. As I would put it, the truthmaker for causal truths is
to be found solely in the world in which the relation holds. (I think this follows
straight from the contingency of the causal relation, a contingency that Lewis does
not doubt.) In his theory of causation the possible worlds enter as mere calculational
devices. He has given me as an example the way that we might say with truth that a
Originally published in Preyer and Siebelt (2001). Minor revisions have been made for consistency.
person is a Montague rather than a Capulet, without being committed to the view
that these families are actual. The ctional families are used as no more than a sort
of calculational device.
An important doctrine that Lewis upholds is the thesis of Humean supervenience,
the thesis that causes and laws supervene on the particular, local, matters of fact. But
we should not let this encourage us to run his theory of causation together with
his theory of law. His theory of law is neo-Humean, the MillRamseyLewis view
that laws are the most systematic and coherent subset of the regularities that the
world contains. It is interesting to notice that given this theory of laws, including
causal laws, but also given irreducibly singular causation, something close to
Humean supervenience still holds: Causes have become particular, local, matters of
fact.2 It seems possible, also, to hold a Lewisian counterfactual theory of causality,
together with a non-Humean theory of laws. For instance, one might hold that laws
are contingent connections holding directly between universals. (If the connections
are held to be necessary, that might be dicult to t together with counterfactual
causality.)
3 Neuron Diagrams and Singular Causation
Lewis has examined at length many types of situation that constitute prima facie
objections to a counterfactual analysis, and I will shortly discuss three of these. But
rst a general remark. In the course of these discussions Lewis introduces his justly
celebrated neuron diagrams. These involve imaginary systems of neurons that are
hooked up according to certain often quite complex causal patterns of rings and
inhibitions of ring. They are then used to illustrate such situations as epipheno-
menal events, preemptive causation, causal redundancy, and so on.
When we look at these diagrams, we can immediately see that they are possible
causal patterns, in most cases empirically possible patterns (you could construct such
a circuit). Counterfactual theories of causation (and, it may be added, regularity
theories and probability-raising theories of causation) struggle with these diagrams.
Wittgenstein spoke of an open door that we had only to see and go through to escape
philosophical confusion. See Malcolm (1958), p. 44.3 The solution that I recommend
to the problems posed by the neuron diagrams is very simple. Where there is an
arrow in a diagram showing that one neuron brings it about that another neuron
res, or is rendered incapable of ring, take it that here there is a genuine two-term
relation of singular causation holding between cause and eect. Where there is no
such arrow, deny that there is any such relation. This is the open door.
446 D. M. Armstrong
4 Negative Causes and Eects
There is a deep predisposition to think that where cause and eect are not contiguous
in both space and time we can always trace an intermediate chain of causes that links
the cause to its distant eect. It is true that we do admit cases of the following sort.
An object is struck, but does not break. Later it is struck again, rather lightly, and
unexpectedly breaks. We understand very well the suggestion that the object broke
because the original striking caused a weakness that persisted and at least helped to
cause the later breaking. This notion that objects can continue to bear eects would
readily be admitted as a qualication of the doctrine of contiguity.
I will, however, record here, but not argue for the position, that I think that the
persistence is still a case of causation. It is however, immanent, that is, remaining
within causality, as opposed to transeunt, that is, interactive causality. (The termi-
nology is due to W. E. Johnson 1924, ch. 9.) This is because I accept the doctrine of
temporal parts for continuing objects and, further, think that the glue that holds the
parts together has to be more than spatiotemporal continuity and a certain continu-
ity of character. So, like Russell, I think of a continuing thing as a causal line, an
information-preserving causal line. However, perhaps unlike Russell, I think of the
line as a line of singular causes. If this is granted, the theory becomes much easier to
defend.
Some think that this contiguity may break down in certain cases. It has been
feared that gravitation would have to be taken as action at a distance. Now, how-
ever, gravitational waves are accepted. More recently, it has been wondered whether
the puzzling EPR phenomenon may force the abandonment of locality. But,
prima facie, we do not expect causes to jump gaps in either space or time.
But what about preventions and omissions? We regularly speak of preventions as
causes of what they prevent, eects that never occur. Omissions are things that are
not done but that have positive eects. What is more, preventions and omissions
seem susceptible to a Lewisian counterfactual analysis. The father grabs the child
running down the drive, thus preventing the child from being hit by the car ap-
proaching in the road. If he had not acted thus, the child would have been run over.
The grabbing is causal interaction in good standing. But there is no causal chain that
leads from the grabbing to the eect of the childs not being run over. Or again,
the father nods o and fails to grab the child. As a result the car hits the child. If he
had not nodded o, he would have grabbed the child. Child and car interact, but
no causal chain proceeds from the father. Preventions, something positive, bring
about the absence of the thing prevented (often desirable). Omissions are absences
Going through the Open Door Again 447
that result in something positive happening (often undesirable). But how can absen-
ces have a place in a causal chain? Nothing will come of nothing.
Let us consider the great causal net of the world,4 the total causal history of the
world, the great neuron diagram as it were, where cause and eect are at every point
positive realities (positive events, states of aairs, or whatever). What needs to be
added to this causal net to give us the preventions and the omissions, the causal
relations that involve absences? It seems that we need to add the laws of nature
(taken ontologically, not as statements). They are needed as truthmakers for the
counterfactual truths that are obviously involved in preventions and omissions. But
given the causal net and its laws, we do not need in addition preventions and omis-
sions. Preventions and omissions exist, of course. But they are not an ontological
addition.
Lewis, it seems, should not deny the existence of the causal net in my sense of the
phrase, the net that involves positive causation alone. (He would, of course, deny
that it is a net involving genuine singular causal connections.) But since the laws for
him are no more than the most systematic and pervasive regularities to be found in
the world, this causal net is no more than the series of instantiations of these regu-
larities, or if he admits fundamental noncausal laws, a subset of these instantiations
of regularities. So causation involving prevention, omission, or absence is no onto-
logical addition to his Humean world.
But what positive account are we to give of these negative eects and negative
causes? I believe that Phil Dowe has shown us the way here. His rst presentation of
his view is to be found in section 6 of Good Connections: Causation and Causal
Processes (Dowe 1999). The account is developed at much greater length in chapter
6 of his book Physical Causation (2000). Let us speak of the causation to be found in
the causal net as causal action. (It has been spoken of as bi.) Dowe argues that when
we have prevention, omission, and so on, then some causal action is always found. In
prevention, for instance, the actual preventing is always causal action. Flipping a
switch to turn a light o is causal action. We may call the current not reaching the
bulb an eect of this action. But if we are thinking solely in terms of causal action,
the only relevant eect is the positive happenings that break the connection of cur-
rent and bulb.5
What has to be added is, fairly obviously, a counterfactual. In the present case, the
counterfactual will perhaps be if the switch had not been ipped, then the current
would have continued to ow, and so kept bulb burning. But now follows the key
feature in Dowes account. These counterfactuals are not the familiar Lewis coun-
terfactuals that have as their aim giving an account or analysis of causality. We may
448 D. M. Armstrong
call them instead Dowe counterfactuals. Their special feature is that they make no
attempt to analyze causality. Causalitycausal actionappears as an unanalyzed
notion in these counterfactuals. In the case of prevention, the counterfactuals give us
in the consequent condition the causal consequences that would ow from the ab-
sence of the successful prevention.
These Dowe counterfactuals will have to have truthmakers, of course. Counter-
factuals cannot hang on ontological air. A key part of the truthmaker will be the
relevant laws of causal action, together presumably with the boundary conditions in
which the prevention takes place. But I will not go into further detail. Dowes latest
discussion of the topic, in chapter 6 of his book, nds a great deal of complex detail
to discuss in the various cases where the causality involved is not mere causal action.
But it seems to me that he has succeeded in showing that, ontologically, negative
causation is a sort of parasitic case. Henceforth, therefore, this essay is concerned
with causal action, positive causality, alone.
But before leaving the topic I call attention to one important point that Dowe
makes. Because causal action is left unanalyzed in his account of prevention and
omission, his account of these phenomena is available, pretty well promiscuously, to
other theories of causation. Regularity theories, probability-raising theories, coun-
terfactual theories, singularist theories, empirical theories such as transference of
energy or the conservation of certain quantities, can all avail themselves of Dowes
treatment of negative causation, and in particular the Dowe counterfactuals. (Coun-
terfactual theories would plug in, say, Lewisian counterfactuals in giving, in terms of
their lights, a deeper account of the Dowe counterfactuals.) The only exception is
agency theories, since prevention, at least, is a means to an end, and the end may be
something negative.
5 Probabilistic Causation
I will now consider Lewiss discussion of chancy causation in section B of the 1986
postscript to his 1973 paper Causation, the sort of thing that we perhaps get in
quantum phenomena. The following case gives him diculty:
. . . c occurs, e has some chance x of occurring, and as it happens e does occur; if c had not
occurred, e would still have had some chance y of occurring, but only a very slight chance
since y would have been very much less than x. We cannot quite say that without the cause,
the eect would not have occurred; but we can say that without the cause, the eect would
have been very much less probable than it actually was. . . . I think we should say that e de-
pends causally on c, and that c is a cause of e. (p. 176)
The obvious thing to say instead, it seems to me, is that e may or may not causally
depend on c, and that c may or may not be the cause of e. Suppose, to take the case
that seems to have the best chance of empirical instantiation, that besides c a poten-
tial cause c
1
is also present and that it is the latter that gives the smaller chance y of e
occurring. It seems to be a perfectly objective question, when e occurs, whether it is c
or c
1
that is the cause, although it is more likely to have been c. It might even be
possible to settle the matter empirically. If the counterfactual theory cannot deliver
this answer, that seems to be a weakness of that theory.
As he makes clear on page 180, Lewis would reject this reply. My style of answer
is considered there. He points out that in the sort of case just considered, two dier-
ent counterfactuals would compete in cases where c and c
1
were both present fol-
lowed by the occurrence of e. In one case, where c is the cause, we have if not c,
then not e, but it is false that if not c
1
, then not e. But with c
1
as the cause, the
truth values of these two conditionals are reversed. He says:
But I reject the presupposition that there are two dierent ways the world could be, giving us
one denite counterfactual or the other. That presupposition is a metaphysical burden quite
out of proportion to its intuitive appeal; what is more its intuitive appeal can be explained
away.
The presupposition is that there is some hidden feature that may or may not be present in
our actual world, and which if present would make true the [relevant] counterfactual. (1986b,
pp. 180181)
Lewis goes on at some length discussing candidates for the hidden feature. But
there is an obvious candidate, and not a mysterious one either if we are not blinded
by the rhetoric of Hume and the Humeans. It is simply causation in the single case.
Either c does this causing, or c
1
does. And some of us believe that we observe that
this relation holds in particular favorable cases. If so, singular causation is not even
hidden in every case.
A case that Lewis does not consider, but which brings out the strength of the sin-
gularist position here, is one where there are two possible probabilistic causes of just
one eect, and the chance of each possible cause being the actual cause is equal.
(Perhaps with a multiplication of the two chances giving the chance of an overdeter-
mination.) Suppose that there are two bombardments of an atom, with the same
chance of the atom emitting a particle, which the atom duly does.6 Does there not
seem to be an objective question, which of the two bombardments actually did the
job?
Notice that we can very well have probabilistic causation of a genuine chance
event (Lewis 1986b, p. 183) on this singularist view, although probabilistic causa-
tion is a rather unhappy and potentially misleading phrase. What we have rather, I
450 D. M. Armstrong
suggest, is the probability, less than strictly 1, of causing occurring. This causing is
the very same causing that we have with deterministic causing. Event c had a genuine
chance of causing event e. Perhaps it did. The causing will then be ordinary singular
causation. Event c
1
had a lesser chance of causing e. Perhaps c
1
was luckyyou
cant lose them all, as the folk sayand c
1
was the event that caused e. Ordi-
nary singular causing again. An open door, I suggest. It is a measure of the power
of a singularist theory that it sweeps aside problems about chancy causation so
easily.
I nd that a little mnemonic is useful: not probabilistic causation but the proba-
bility of causing. In cases where an event c has a less than strictly 1 chance of causing
e and does cause e then we can say that c res. If c fails to re, there is no causa-
tion at all, though there was a chance of it.7
A nal point. Lewis proposed his counterfactual theory as a conceptual analysis of
causation. I have been arguing that a singularist theory is preferable as conceptual
analysis, because it yields more intuitive results in dicult and disputed cases. (The
elimination of negative causation is an add-on that seems to simplify the ontology of
singular causation.) Lewis suggested to me that quantum physics may actually pro-
vide evidence against singularism. If so, as a scientic realist I ought to bow to the
scientic evidence. I am not sure how to take this. It does not seem to me that sin-
gular probabilistic causation would be a hidden variable in the sense abjured in
quantum theory. But I may be wrong here, or perhaps there are other strong reasons
in quantum theory to reject singularist causation. I can certainly conceive giving up
singularism under the weight of such evidence, but like St. Augustine, I say not now.
But even if the theory had to be abandoned, I would still claim that it does best jus-
tice to our concept of causation.
6 Causation and Counterfactuals
Suppose it is true that event c causes event e. We know that, because of the possibil-
ity of a backup cause, it is false that it is entailed that if c had not occurred, then e
would not have occurred. There also seems to be the bare possibility, pointed out to
me by Michael Tooley, that, without c, e would have occurred without any cause at
all. So what entailment can we nd? It is not at all clear that we can do anything
better than this: c causes e entails that: if not-c, then, in the absence of a backup
cause, or e coming to exist totally uncaused, then not-e.
It may well be, of course, both that a backup cause is absent, and that, likely
enough, totally uncaused events never occur. If so, then the stronger if not-c, then
not-e will be true. Because of this, it seems fair enough to say that the truth that c
causes e supports the proposition if not-c, then not-e. But without the qualifying
clauses, this support is not entailment.
It seems to me that this throws out an important challenge to those, such as Lewis,
who hold that the counterfactual analysis of causation is conceptually true. The onus
is on them to give us an entailment of c causes e that involves counterfactuals but
makes no mention of causation. I have no proof that the onus cannot be discharged,
but I do suspect that it cannot be.
7 Menzies on Singular Causation
But what is this causation that so many analytic philosophers nd so mysterious? In
what I think is a very important paper, Probabilistic Causation and the Pre-emption
Problem (1996), Peter Menzies suggests that such causation is a theoretical en-
tity. Look at the dierence a word or two makes! Lewis stigmatized singular cau-
sation as a hidden feature, a phrase well calculated to arouse our Humean reexes.
He also says that it is mysterious (1986b, p. 182). But the phrase theoretical en-
tity puts a new complexion on matters. And then Menzies goes on to suggest that
we should avail ourselves here of the treatment of theoretical entities to be found in
the work ofDavid Lewis. For some further development of Menziess position, see
his paper Intrinsic versus Extrinsic Conceptions of Causation (Menzies 1999).
Lewiss best-known application of his theory of theoretical entities is to our con-
cepts of mental states. His idea is to collect the platitudes of folk psychology that
concern the causal relations of mental states, sensory stimuli, and behavioral re-
sponses. Call these platitudes when conned to a particular sort of mental state, pain,
say, the causal role of pain. Pain is a theoretical entity, namely, that entity that plays
the causal role of pain. Here the notion of cause is treated as an undened primitive.
But, says Menzies, why not apply the same treatment to causation itself ? Collect the
platitudes of singular causation in the same sort of way. Causation is that intrinsic
relation between singular events for which the causal platitudes hold.
In previous work on causality Menzies had emphasized not so much counter-
factuals as the raising of probabilities: A cause raises the probability of its eect. It is
known, however, that there are cases that pose great diculty for theories of this
type. But we know that a cause does typically raise the probability of its eect.
Menzies therefore oers us the following Lewisian-style denition of the causal rela-
tion: It is the intrinsic relation that typically holds between two distinct events when one
increases the chance of the other event (Menzies 1996, p. 101). The word intrinsic
here is meant to capture the singularist insight. Menzies says, furthermore, that the
notion of increase of chance is to be understood in the counterfactual manner found
452 D. M. Armstrong
in Lewiss theory. So the counterfactuals that typically hold in the case where one
event causes another, but which may fail in cases of late preemption and other sorts
of unusual cases, can be part of the causal platitude.
All very ne, I think. For myself I would favor taking Menziess causal plati-
tude as xing the reference of the singular causal relation rather than analyzing the
concept. Menzies tells me that his treatment was meant as conceptual analysis, but
he does not want to bar a retreat to reference-xing. But I have three further points
to make about Menziess treatment, points to be developed in the succeeding sec-
tions. First, I think that we can supplement the epistemology involved in a quite
important way. Second, I think that Menziess platitude stands in need of some ex-
pansion. Third, and more speculatively, I have hopes that we can get behind the
Menzies denition to a deeper, more ontologically revealing, account of the nature of
singular causation.
8 The Epistemology of Singular Causation
I begin by noting that Lewiss causal role account of the mind makes the notion of
the mental a theoretical one. I have great sympathy with this idea, though I might
elsewhere niggle about details. But Lewis would not have denied that this theoretical
entity is also one to which we have some observational access in the rst-person case.
We have introspective access to some of our own mental goings-on. We do not have
introspective access to the intrinsic nature of these goings-on, certainly not if this in-
trinsic nature is as Lewis and I believe it is, namely, physical in nature and going on
inside our heads. But we do have introspective access to our own mental goings-on,
perhaps as goings-on that typically play a certain causal role.
This should embolden us to wonder whether we do not have observational access
to singular causation, at least in certain favorable cases. There is in fact a good deal
of evidence that opponents need to explain away.
(i) Ordinary Language Our ordinary language, as a number of philosophers have
emphasized, constantly involves reference to the perception of causality in the singu-
lar case. This is a little obscured in the reporting of what we perceive, because we so
often use verbs that make no explicit mention of the causality involved. G. E. M.
Anscombe (1971, p. 8) oers what she calls A small selection: scrape, push, wet,
carry, eat, burn, knock over, keep o, squash, make, (e.g. noises, paper boats), hurt.
It would seem that all theories that deny the existence of singular causation must as-
sert that such perceptual reports invariably involve (unselfconscious and immediate)
inference. The contrast will be with the perception of shapes, color-surfaces, motions,
smells, and so on. The contrast does not seem particularly plausible on the surface,
but it must be made if there is no genuinely singular causation.
(ii) Two Salient Cases Two sorts of case that philosophers have advanced as
involving noninferential perception of singular causation are our awareness of pres-
sure on our own body, and our introspective awareness of the (successful) operation
of our own will. I myself think that perception of pressure on our own body is the
nearest there is to a knock-down case for noninferential perception of causality,
though strangely enough there has been little discussion of it in the literature. (I in-
cline to suspect the fell inuence of the old representative or veil-of-perception theory
of perception.) I rst argued for this view in my (1962), though not then perceiving
that it was incompatible with Humean theories of causation, and more recently in
my (1997), section 14.6. An important defense of the same position is to be found in
Evan Fales (1990), chapter 1. He points out, very interestingly, that the phenomen-
ology of the perception of pressure matches the properties of a physical vector.
Many philosophers have pointed to the introspective awareness of the causality of
the will. Fales gives names (see his ch. 1, n. 12). I suppose the case is more salient in
the literature than that of pressure because few would dare to postulate a veil-of-
perception intervening between oneself and ones own mental operations. Hume is
aware of this objection to his position, and he argues against it in the appendix to the
Treatise and later at length in the Enquiry Concerning Human Understanding, sec-
tions 5253. But it is clear he had come to see that here, perhaps, was a place where
the idea of causality might be derived from an impression, but that he had this real-
ization only after he had given his own account of causation. The arguments he gives
require attention, as one would expect, but they are not especially impressive. See the
thorough and convincing examination of these arguments in Menzies (1998).
(iii) Experimental Work Finally, there is the experimental work of the psycholo-
gist A. Michotte of Louvain University (1963). He argues for a visual apprehension
of causality, and in chapter 13 for a tactile-kinaesthetic awareness.8 The visual case, I
think, though quite strong, is not quite so convincing as the perception of pressure,
or even awareness of the operation of our own will.
9 Some More Platitudes
It seems to me that Menzies has left out two things from his collection of platitudes
that (in my view) x the reference for the singular causal relation. The rst is what
we might call the same cause, same eect platitude. I mean by this that it is regu-
larly true, even allowing that there are many apparent exceptions in ordinary expe-
454 D. M. Armstrong
rience, that from the same sort of cause the same sort of event, or at least roughly the
same probability distribution of events, will follow. We might call it the regularity
platitude, and we can see it as the truth that the regularity theory of causation grasps
but misleadingly makes the whole truth about causation. Perhaps, however, addition
of this platitude is implicit in Menzies (1996) section 5, where he follows up his
Lewisian-style denition of causation by calling attention to: . . . the powerful intu-
ition that causal relations relate events in virtue of specic features of the events
themselves, in particular the properties they exemplify (1996, p. 105).
But the intuition as stated is not quite enough. We want the idea that not only
does causation depend on the properties of the thing doing the causing, but also that
from the same properties the same eects tend to ow.
A nal platitude that I think we should include is linked with the agency or ma-
nipulative theory of causality.9 Causes characteristically stand to their eects as
means, or empirically possible means stand to their ends. We see, then, that regular-
ity, counterfactual, probability-raising, and agency theories, though all should be
rejected, each do contribute their important platitude to xing the reference of the
causal relation.
10 The Deeper Nature of Causation
Finally, we come a deeper, if more controversial suggestion, which I owe to Adrian
Heathcote (Heathcote and Armstrong 1991), that an account of singular causation
can be given as the instantiation of a law.
Laws link properties. In the simplest case, the instantiation of a certain property
determines, or in the probabilistic case merely makes probable, that either that same
particular or some related particular instantiates a certain property. That, surely,
would explain the raising of the probability of the eect that is so regularly involved.
(It would also explain the counterfactuals that we associate with causation, and why
causes are means to ends.) A law (an entity in the world) provides the ground in na-
ture for an inference. The cause is like the antecedent. If you know the singular cause
and the law (which may be probabilistic only and/or may involve ceteris paribus
conditions of operation) then, in general though not invariably, you can infer a
raised probability of the presence of the eect.
Heathcote and I think of this identity thesis as an empirical identity claim.10 It will
not, however, be like the identication of the player of the mental causal role with
brain processes. The latter is a contingent identity claim, because it does not tell us
the purported essence of the mental. The brain processes are just those entities that
happen to play the mental role. But our claim is that instantiation of law gives us the
essence (or perhaps only part of the essenceexplanation to come shortly) of singu-
lar causation. A better model than mindbrain identity will therefore be the iden-
tity of heat with molecular motion or water with H
2
O molecules. It is a Kripkean
necessity.
I now indicate briey a consequence of combining the identication of singular
causes with instantiations of a law (or laws) with my view, argued at length else-
where,11 that laws are relations of universals. Suppose that laws link properties;
suppose that the ultimate properties, the properties that divide the world along its
ultimate joints, are universals; suppose that the ultimate laws, at least, are linkages of
universals. Such a linkage of universals will itself be a (complex) universal. It follows
that each instantiation of a fundamental law will be the instantiation of a universal.
But, famously, each instantiation of a universal is complete in itself, so the law will
be present completely in each instantiation. Hence singular causation will be a com-
pletely intrinsic relation. The causal structure of a process will be determined solely
by the intrinsic character of that process.
This result was unsought, but I think it is a welcome consequence of my theory of
laws.12 By contrast, any Hume-inspired theory of laws makes the lawlike nature of
an instantiation of the law an extrinsic property of the instantiation.
But there is a problem that my view of singular causation faces. I said three para-
graphs ago, in an aside, that instantiation of a law may be no more than part of the
essence of singular causation. My reason was this. Not all laws are causal laws, or at
least not all are obviously so. So it seems that all I am entitled to say is that singular
causation is instantiation of a certain sort of law. My identication is therefore in-
complete. And it would be a sad anticlimax to say that singular causation is instan-
tiation of a causal law.
One satisfactory solution to this problem from my point of view would be that all
the fundamental laws are, or can fairly be represented as being, causal laws. The no-
tion of the causal net has already been introduced. What will need to be said is that
the net exhausts reality and is causal through and through. The lines of the net pro-
ceed by immanent singular causation. Where lines intersect there is transeunt causa-
tion. It would be a natural corollary of this thesis that the fundamental laws are all
causal.
A nal word about the attractive empirical theories of Aronson, David Fair,
and Phil Dowe. (Among these theories, I lean to Dowes conserved quantity ap-
proach.) If the theory sketched above is on the right track, it would seem that these
theories are really giving us theses about the most general laws that govern all phys-
ical action. The theories, I think, require the notion of singular causal connection to
make them work, and so they cannot be the account of what that connection is.
456 D. M. Armstrong
Acknowledgments
A slightly earlier version of this paper was published in Reality and Humean Super-
venience: Essays on the Philosophy of David Lewis, eds. Gerhard Preyer and Frank
Siebelt, New York: Rowman and Littleeld, 2001, by kind permission of the editors
of the present volume.
I am indebted for comments to Phil Dowe, Michael McDermott, and Peter
Menzies. Particular thanks go to the late David Lewis.
Notes
1. Lewiss essay in this volume, Causation as Inuence, is not examined here. It is my impression, how-
ever, that the new points presented there do not change matters much from my point of view.
2. As Lewis pointed out to me, the denition he gives of Humean supervenience in his (1986a), pp. ixx.
explicitly restricts the fundamental external relations to the geometrical ones.
3. The full quotation runs: A person caught in a philosophical confusion is like a man in a room who
wants to get out but doesnt know how. He tries the window but it is too high. He tries the chimney but it
is too narrow. And if he would only turn around, he would see that the door had been open all the time! I
am indebted to Nicholas Smith for tracing the reference. Like all of us, Wittgenstein failed to go through a
few open doors himself.
4. The phrase causal net is inspired by C. B. Martins power net. See also his (1993), but without that
striking phrase.
5. Lewis in his last years was keen to emphasize that even where we ip the switch to turn the light on,
rather than o, there is still what he calls prevention of prevention. The gap in the circuit when the
switch is switched o prevents the current owing to the bulb. Switching on prevents this prevention.
But notice that here there is nothing that is not causal action. This prevention enables the current to ow
and light the bulb. To represent a situation of this sort in a neuron diagram, one would need a side chain
coming in and aecting a particular neuron in the main chain in such a manner that the main chain was
able to continue ring beyond that neuron.
6. For cases of this sort I am beholden to Michael Tooley. See my (1983), p. 133.
7. I may well have picked up this term res from the writings of Nancy Cartwright, who tells me that she
sometimes uses it in the same way. In her (1989), p. 108, she says that she is trying to model a very special
concept of probabilistic causality, a concept according to which the cause either contributes its entire in-
uence or it does not contribute at all. That is my idea also, although it does not seem so special to me.
8. Menzies (1998) also cites Michotte and iluminatingly links his work with Fodors modularity hypothesis.
9. See Gasking (1955) and (1996), and Mellor (1995), ch. 7.
10. Davidson (1995), replying to criticism from Elizabeth Anscombe, argues that it is possible to get from
causes to laws by conceptual argument alone. I am inclined to think that this does not work, but if it does,
then so much the better.
11. Most recently in my (1997), chs. 15 and 16.
12. In the case of a merely probabilistic law I contend that the law, as opposed to the antecedent condition
of the law, is instantiated only in the positive cases, only where the law res.
References
Albert, David. 2000. Time and Chance. Cambridge, Mass.: Harvard University Press.
Anderson, John. 1938. The Problem of Causality. Australasian Journal of Philosophy and Psychology 16:
127142.
Anscombe, G. E. M. 1971. Causality and Determination: An Inaugural Lecture. Cambridge: Cambridge
University Press.
Armstrong, D. M. 1962. Bodily Sensations. London: Routledge and Kegan Paul.
Armstrong, D. M. 1968. A Materialist Theory of the Mind. London: Routledge and Kegan Paul.
Armstrong, D. M. 1983. What Is a Law of Nature? New York: Cambridge University Press.
Armstrong, D. M. 1989. A Combinatorial Theory of Possibility. New York: Cambridge University Press.
Armstrong, D. M. 1997. A World of States of Aairs. New York: Cambridge University Press.
Armstrong, D. M. 1999. The Open Door. In Sankey (1999): 175185. A revised version, Going
through the Open Door Again: Counterfactual vs. Singularist Theories of Causation, appeared in Preyer
and Siebelt (2001) and is reprinted in this volume.
Bacon, John, Keith Campbell, and Lloyd Reinhardt (eds.). 1993. Ontology, Causality, and Mind: Essays in
Honour of D. M. Armstrong. New York: Cambridge University Press.
Barker, S. 1999. Counterfactuals, Probabilistic Counterfactuals and Causation. Mind 108: 427469.
Bedard, Katherine. 1993. Partial Denotations of Theoretical Terms. Nous 27: 499511.
Bell, J. S. 1964. On the EinsteinPodolskyRosen Paradox. Physics 1: 195200.
Bennett, Jonathan. 1984. Counterfactuals and Temporal Direction. Philosophical Review 93: 5791.
Bennett, Jonathan. 1987. Event Causation: the Counterfactual Analysis. Philosophical Perspectives 1:
367386. Reprinted in Sosa and Tooley (1993): 217233.
Bennett, Jonathan. 1988. Events and Their Names. Indianapolis, Ind.: Hackett.
Bogdan, R. J. (ed.). 1984. D. M. Armstrong. Boston: Reidel.
Bohm, D. 1952. A Suggested Interpretation of Quantum Theory in Terms of Hidden Variables, Parts I
and II. Physical Review 85: 166193.
Bunzl, Martin. 1980. Causal Preemption and Counterfactuals. Philosophical Studies 37: 115124.
Buttereld, H. 1957. The Origins of Modern Science, 13001800, second edition. New York: Macmillan.
Byrne, Alex, and Ned Hall. 1998. Against the PCA-analysis. Analysis 58: 3844.
Campbell, Keith. 1990. Abstract Particulars. Oxford: Blackwell.
Carnap, Rudolf. 1963. Replies and Systematic Expositions. In Schilpp (1963): 963966.
Carroll, John. 1994. Laws of Nature. New York: Cambridge University Press.
Cartwright, Nancy. 1979. Causal Laws and Eective Strategies. Nous 13: 419437.
Cartwright, Nancy. 1983. How the Laws of Physics Lie. Oxford: Clarendon Press.
Cartwright, Nancy. 1989. Natures Capacities and Their Measurement. Oxford: Clarendon Press.
Cartwright, Nancy. 1999. The Dappled World. Oxford: Oxford University Press.
Casati, Roberto, and Achille Varzi. 1994. Holes and Other Supercialities. Cambridge, Mass.: MIT Press.
Chisholm, Roderick M. 1946. The Contrary-to-Fact Conditional. Mind 55: 219307.
Chisholm, Roderick M. 1955. Law Statements and Counterfactual Inference. Analysis 15: 97105.
Collins, John. 1999. Indeterminacy and Intention. In Lewis Hahn (ed.), The Philosophy of Donald
Davidson, The Library of Living Philosophers, volume 27, 501528. Chicago, Ill.: Open Court.
Collins, John. 2000. Preemptive Prevention. Journal of Philosophy 97: 223234. Reprinted in this
volume.
Davidson, Donald. 1967. Causal Relations. Journal of Philosophy 64: 691703. Reprinted in Davidson
(1980): 149162.
Davidson, Donald. 1969. The Individuation of Events. In Nicholas Rescher (ed.), Essays in Honor of
Carl G. Hempel, 216234. Dordrecht: Reidel. Reprinted in Davidson (1980): 163180.
Davidson, Donald. 1970. Events as Particulars. Nous 4: 2532. Reprinted in Davidson (1980): 181
187.
Davidson, Donald. 1980. Essays on Actions and Events. New York: Oxford University Press.
Davidson, Donald. 1985. Reply to Quine on Events. In Lepore and MacLaughlin (1985): 172176.
Davidson, Donald. 1995. Law and Cause. Dialectica 49: 263279.
Davidson, Donald, and Gilbert Harman (eds.). 1972. Semantics of Natural Language. Dordrecht: Reidel.
Dowe, Phil. 1992. Wesley Salmons Process Theory of Causality and the Conserved Quantity Theory.
Philosophy of Science 59: 195216.
Dowe, Phil. 1996. Backwards Causation and the Direction of Causal Processes. Mind 105: 227248.
Dowe, Phil. 1999. Good Connections: Causation and Causal Processes. In Sankey (1999): 247263.
Dowe, Phil. 2000. Physical Causation. New York: Cambridge University Press.
Dretske, Fred. 1973. Contrastive Statements. The Philosophical Review 82: 411437.
Dretske, Fred. 1977a. Laws of Nature. Philosophy of Science 64: 248268.
Dretske, Fred. 1977b. Referring to Events. Midwest Studies in Philosophy 2: 9099.
Ducasse, C. J. 1926. On the Nature and the Observability of the Causal Relation. Journal of Philosophy
23: 5768.
Ducasse, C. J. 1968. Truth, Knowledge, and Causation. London: Routledge and Kegan Paul.
Earman, John. 1984. Laws of Nature: The Empiricist Challenge. In Bogdan (1984): 191223.
Edgington, D. 1995. On Conditionals. Mind 104: 235329.
Edgington, D. 1999. Counterfactuals and the Importance of Hindsight. Unpublished paper, delivered at
a London School of Economics workshop on Chance and Cause.
Eells, Ellery. 1991. Probabilistic Causality. Cambridge: Cambridge University Press.
Ehring, Douglas. 1987. Causal Relata. Synthese 73: 319328.
Ehring, Douglas. 1989. Preemption and Probabilistic Counterfactual Theory. Philosophical Studies 56:
307313.
Ehring, Douglas. 1990. Preemption, Direct Causation, and Identity. Synthese 85: 5570.
Ehring, Douglas. 1994. Preemption and Eells on Token Causation. Philosophical Studies 74: 3950.
Ehring, Douglas. 1997. Causation and Persistence. New York: Oxford University Press.
Elga, Adam. 2000. Statistical Mechanics and the Asymmetry of Counterfactual Dependence. Philosophy
of Science (suppl. vol. 68, PSA 2000): 313324.
Fair, David. 1979. Causation and the Flow of Energy. Erkenntnis 14: 219250.
Fales, Evan. 1990. Causation and Universals. London: Routledge.
Fetzer, James H. (ed.). 1988. Probability and Causality. Boston: Reidel.
Fillmore, Charles. 1971. How to Know Whether Youre Coming or Going. In Karl Hyldgard-Jensen
(ed.), Linguistik. 369379. Frankfurt: Athenaum Verlag.
Fine, Kit. 1975. Review of Lewis, Counterfactuals. Mind 84: 451458.
Fischer, John Martin, and Mark Ravizza (eds.). 1993a. Perspectives on Moral Responsibility. Ithaca, NY:
Cornell University Press.
Fischer, John Martin, and Mark Ravizza. 1993b. Responsibility for Consequences. In Fischer and
Ravizza (1993a): 322347.
Fodor, J. 1991. You Can Fool Some of the People All the Time, Other Things Being Equal: Hedged
Laws and Psychological Explanation. Mind 100: 1934.
460 References
Frankfurt, Harry. 1969. Alternate Possibilities and Moral Responsibility. Journal of Philosophy 66:
829839.
Ganeri, Jonardon, P. Noordhof, and M. Ramachandran. 1996. Counterfactuals and Preemptive Causa-
tion. Analysis 56: 219225.
Ganeri, Jonardon, P. Noordhof, and M. Ramachandran. 1998. For a (Revised) PCA-analysis. Analysis
58: 4547.
Garnkel, A. 1981. Forms of Explanation. New Haven, Conn.: Yale University Press.
Gasking, Douglas. 1955. Causation and Recipes. Mind 64: 479487. Reprinted in Oakley and ONeill
(1996): 106115.
Gasking, Douglas. 1996. Hypotheticals, Recipes, and Causation. In Oakley and ONeill (1996):
116131.
Giere, R. 1988. Explaining Science. Chicago, Ill.: University of Chicago Press.
Goldman, Alvin. 1970. A Theory of Human Action. Englewood Clis, N.J.: Prentice-Hall.
Goodman, Nelson. 1947. The Problem of Counterfactual Conditionals. Journal of Philosophy 44:
113128. Reprinted in Goodman (1979).
Goodman, Nelson. 1979. Fact, Fiction, and Forecast. Cambridge, Mass.: Harvard University Press.
Goosens, William K. 1979. Causal Chains and Counterfactuals. Journal of Philosophy 76: 489495.
Gorovitz, S. 1965. Causal Judgements and Causal Explanations. Journal of Philosophy 62: 695711.
Grice, H. P. 1975. Logic and Conversation. In P. Cole and J. L. Morgan (eds.), Syntax and Semantics,
volume 3. New York: Academic Press.
Hall, Ned. 1994. New Problems for an Analysis of Causation. Unpublished paper presented at the
annual conference of the Australasian Association of Philosophy.
Hall, Ned. 2000. Causation and the Price of Transitivity. Journal of Philosophy 97: 198222. Reprinted
in this volume.
Hall, Ned. 2002a. Non-locality on the Cheap? A New Problem for Counterfactual Analyses of Causa-
tion. Nous 36: 276294.
Hall, Ned. 2002b. The Intrinsic Character of Causation. In Dean Zimmerman (ed.), Oxford Studies in
Metaphysics, volume 1. Oxford: Oxford University Press, 2002.
Harper, William L., and B. Skyrms (eds.). 1988. Causation in Decision, Belief Change, and Statistics. Bos-
ton: Kluwer.
Hart, H. L. A., and A. M. Honore. 1985. Causation in the Law, second edition. Oxford: Clarendon Press.
Hausman, Daniel. 1992. Thresholds, Transitivity, Overdetermination, and Events. Analysis 52: 159
163.
Hausman, Daniel. 1998. Causal Asymmetries. New York: Cambridge University Press.
Heathcote, Adrian, and D. M. Armstrong. 1991. Causes and Laws. Nous 25: 6373.
Heinlein, Robert. 1951. The Puppet Masters. Garden City, N.Y.: Doubleday.
Hempel, C. 1965. Aspects of Scientic Explanation. New York: The Free Press.
Hitchcock, C. 1993. A Generalized Probabilistic theory of Causal Relevance. Synthese 97: 335364.
Hitchcock, C. 1995a. The Mishap at Reichenbach Fall: Singular vs. General Causation. Philosophical
Studies 78: 257291.
Hitchcock, C. 1995b. Discussion: Salmon on Explanatory Relevance. Philosophy of Science 62: 304
320.
Hitchcock, C. 1996a. Farewell to Binary Causation. Canadian Journal of Philosophy 26: 267282.
Hitchcock, C. 1996b. The Role of Contrast in Causal and Explanatory Claims. Synthese 107: 395
419.
References 461
Hitchcock, C. 1998. Causal Knowledge: That Great Guide of Human Life. Communication and Cogni-
tion 31: 271296.
Holland, Paul. 1986. Statistics and Causal Inference. Journal of the American Statistical Association 81:
945960.
Horwich, Paul. 1989. Asymmetries in Time. Cambridge, Mass.: MIT Press.
Hume, David. 1888 (1739). A Treatise of Human Nature, ed. L. A. Selby-Bigge. Oxford: Clarendon Press.
Hume, David. 1902 (1748). An Enquiry Concerning Human Understanding, ed. L. A. Selby-Bigge. Oxford:
Clarendon Press.
Humphreys, P. 1989. The Chances of Explanation: Causal Explanations in the Social, Medical, and Physi-
cal Sciences. Princeton, N.J.: Princeton University Press.
Jackson, Frank. 1977. Perception: A Representative Theory. Cambridge: Cambridge University Press.
Jackson, Frank, and Philip Pettit. 1990. Program Explanation: A General Perspective. Analysis 50: 107
117.
Johnson, W. E. 1924. Logic, vol. 3. Cambridge: Cambridge University Press.
Johnston, M. 1993. Objectivity Regured: Pragmatism without Vericationism, Appendix 2: Com-
plexities in the Notion of a Disposition. In J. Haldane and C. Wright (eds.), Reality, Representation, and
Projection, 119121. New York: Oxford University Press.
Joseph, G. 1980. The Many Sciences and the One World. Journal of Philosophy 77: 773790.
Kim, Jaegwon. 1971. Causes and Events: Mackie on Causation. Journal of Philosophy 68: 426441.
Kim, Jaegwon. 1973a. Causation, Nomic Subsumption, and the Concept of Event. Journal of Philoso-
phy 70: 217236.
Kim, Jaegwon. 1973b. Causes and Counterfactuals. Journal of Philosophy 70: 570572.
Kim, Jaegwon. 1980. Events as Property Exemplications. In M. Brand and D. Walton (eds.), Action
Theory, 159177. Dordrecht: Reidel. Reprinted in Kim (1993): 3352.
Kim, Jaegwon. 1982. Psychophysical Supervenience. Philosophical Studies 41: 5170.
Kim, Jaegwon. 1993. Supervenience and Mind: Selected Philosophical Essays. New York: Cambridge
University Press.
Knowles, Dudley (ed.). 1990. Explanation and Its Limits. Cambridge: Cambridge University Press.
Kripke, Saul. 1972. Naming and Necessity. In Davidson and Harman (1972): 253355.
Kripke, S. 1982. Wittgenstein on Rules and Private Language. Cambridge, Mass.: Harvard University
Press.
Kvart, Igal. 1975. Counterfactual Conditionals. Ph.D. dissertation, University of Pittsburgh.
Kvart, Igal. 1979. A Formal Semantics for Temporal Logic and Counterfactuals. Logique et Analyse 22:
3562.
Kvart, Igal. 1986. A Theory of Counterfactuals. Indianapolis, Ind.: Hackett.
Kvart, Igal. 1991a. Counterfactuals and Causal Relevance. Pacic Philosophical Quarterly 72: 314
337.
Kvart, Igal. 1991b. Transitivity and Preemption of Causal Impact. Philosophical Studies 64: 125160.
Kvart, Igal. 1992. Counterfactuals. Erkenntnis 36: 141.
Kvart, Igal. 1994. Overall Positive Causal Impact. Canadian Journal of Philosophy 26: 267282.
Kvart, Igal. 1997. Cause and Some Positive Causal Impact. In J. Tomberlin (ed.), Philosophical Per-
spectives 11: Mind, Causation, and World, 401432. Atascadero, Calif.: Ridgeview.
Kvart, Igal. 2001a. A Counterfactual Theory of Cause. Synthese 127: 389427.
Kvart, Igal. 2001b. Causal Relevance. In Bryson Brown (ed.), New Studies in Exact Philosophy: Logic,
Mathematics and Science, 5990. Oxford: Hermes Scientic Publications.
462 References
Kvart, Igal. 2001c. Counterexamples to Lewiss Causation as Inuence. Australasian Journal of Phi-
losophy 79: 411423.
Kvart, Igal. 2003. Probabilistic Cause, Edge Conditions, Late Preemption and Discrete Cases. In Phil
Dowe and Paul Noordhof (eds.), Cause and Chance: Causation in an Indeterministic World. London:
Routledge.
Kvart, Igal. Forthcoming a. Mental Causation and Probabilistic Causation. Pacic Philosophical
Quarterly.
Kvart, Igal. Forthcoming b. Cause: Time and Manner.
Kvart, Igal. Forthcoming c. Partial Causal Neutralizers.
Langton, Rae, and David Lewis. 1998. Dening Intrinsic. Philosophy and Phenomenological Research
58: 333345. Reprinted in Lewis (1999): 116132.
LeCatt, Bruce. 1982. Censored Vision. Australasian Journal of Philosophy 60: 158162.
Lee, Jig-chuen. 1988. The Nontransitivity of Causation. American Philosophical Quarterly 25: 8794.
LePore, Ernest, and Brian McLaughlin. 1985. Actions and Events: Perspectives on the Philosophy of Don-
ald Davidson. New York: Blackwell.
Lewis, David. 1966. An Argument for the Identity Theory. Journal of Philosophy 63: 1725. Reprinted
with additional material in Lewis (1983a): 99107.
Lewis, David. 1970. How to Dene Theoretical Terms. Journal of Philosophy 67: 427446. Reprinted in
Lewis (1983a): 7895.
Lewis, David. 1973a. Causation. Journal of Philosophy 70: 556567. Reprinted in Lewis (1986a): 159
172.
Lewis, David. 1973b. Counterfactuals. Cambridge: Harvard University Press.
Lewis, David. 1973c. Counterfactuals and Comparative Possibility. Journal of Philosophical Logic 2:
418446. Reprinted in Lewis (1986a): 331.
Lewis, David. 1979a. Counterfactual Dependence and Times Arrow. Nous 13: 455476. Reprinted with
postscripts in Lewis (1986a): 3266.
Lewis, David. 1979b. Scorekeeping in a Language Game. Journal of Philosophical Logic 8: 339359.
Reprinted in Lewis (1983a): 233249.
Lewis, David. 1980. Index, Context, and Content. In Stig Kanger and Sven O
hman (eds.), Philosophy

and Grammar. Dordrecht: Reidel. Reprinted in Lewis (1998): 2144.
Lewis, David. 1983a. Philosophical Papers, volume I. Oxford: Oxford University Press.
Lewis, David. 1983b. New Work for a Theory of Universals. Australasian Journal of Philosophy 61:
343377. Reprinted in Lewis (1999): 855.
Lewis, David. 1986a. Philosophical Papers, volume II. Oxford: Oxford University Press.
Lewis, David. 1986b. Postscripts to Causation. In Lewis (1986a): 172213.
Lewis, David. 1986c. Causal Explanation. In Lewis (1986a): 214240.
Lewis, David. 1986d. Events. In Lewis (1986a): 241269.
Lewis, David. 1986e. On the Plurality of Worlds. Oxford: Blackwell.
Lewis, David. 1993. Many, but Almost One. In Bacon, Campbell, and Reinhardt (1993): 2338.
Reprinted in Lewis (1999): 164182.
Lewis, David. 1997. Finkish Dispositions. Philosophical Quarterly 47: 143158. Reprinted in Lewis
(1999): 133151.
Lewis, David. 1998. Papers in Philosophical Logic. Cambridge: Cambridge University Press.
Lewis, David. 1999. Papers in Metaphysics and Epistemology. Cambridge: Cambridge University Press.
References 463
Lewis, David. 2000. Causation as Inuence. Journal of Philosophy 97: 182197. An expanded version of
the paper appears in this volume.
Lewis, David, and Stephanie Lewis. 1970. Holes. Australasian Journal of Philosophy 48: 206212.
Reprinted in Lewis (1983a): 39.
Lipton, P. 1990. Contrastive Explanation. In D. Knowles (ed.), Explanation and Its Limits. Cambridge:
Cambridge University Press, 1990.
Lipton, P. 1991. Inference to the Best Explanation. London: Routledge.
Lombard, Lawrence. 1986. Events: A Metaphysical Study. London: Routledge and Kegan Paul.
Lombard, Lawrence. 1990. Causes, Enablers, and the Counterfactual Analysis. Philosophical Studies 59:
195211.
Lowe, E. J. 1980. For Want of a Nail. Analysis 40: 5052.
Mackie, J. L. 1965. Causes and Conditions. American Philosophical Quarterly 2: 245264.
Mackie, J. L. 1974. The Cement of the Universe. Oxford: Clarendon Press.
Mackie, Penelope. 1992. Causing, Delaying, and Hastening: Do Rains Cause Fires? Mind 101: 483500.
Malcolm, Norman. 1958. Ludwig Wittgenstein: A Memoir. New York: Oxford University Press.
Martin, C. B. 1993. Power for Realists. In Bacon, Campbell, and Reinhardt (1993): 175186.
Martin, C. B. 1994. Dispositions and Conditionals. Philosophical Quarterly 44: 18.
Martin, C. B. 1996. How It Is: Entities, Absences, and Voids. Australasian Journal of Philosophy 74: 57
65.
Maslen, Cei. 2000. Causes, Eects, and Contrasts. Princeton Ph.D. dissertation. Ann Arbor, Mich.: Uni-
versity Microlms International.
Maudlin, Tim. 1994. Quantum Non-locality and Relativity. Oxford: Blackwell.
Maudlin, Tim. Forthcoming. A Modest Proposal Concerning Laws, Counterfactuals, and Explanation.
McDermott, Michael. 1995a. Redundant Causation. British Journal for the Philosophy of Science 46:
523544.
McDermott, Michael. 1995b. Lewis on Causal Dependence. Australasian Journal of Philosophy 73: 129
139.
McGrath, Sarah. 2002. Causation by Omission. Chapter 2 of Sarah McGrath, Causation in Metaphysics
and Moral Theory, Ph.D. dissertation, Massachusetts Institute of Technology.
Mellor, D. H. 1987. The Singularly Aecting Facts of Causation. In P. Pettit, R. Sylvan, and J. Norman
(eds.), Metaphysics and Morality. Oxford: Basil Blackwell.
Mellor, D. H. 1988. On Raising the Chances of Eects. In Fetzer (1988): 229240.
Mellor, D. H. 1995. The Facts of Causation. London: Routledge.
Mellor, D. H. 1997. Properties and Predicates. In Mellor and Oliver (1997): 255267.
Mellor, D. H. 1998. Real Time II. London: Routledge.
Mellor, D. H. 2000. The Semantics and Ontology of Dispositions. Mind 109: 757780.
Mellor, D. H., and A. Oliver (eds.). 1997. Properties. Oxford: Oxford University Press.
Menzies, Peter. 1989a. Probabilistic Causation and Causal Processes: A Critique of Lewis. Philosophy of
Science 56: 642663.
Menzies, Peter. 1989b. A Unied Theory of Causal Relata. Australasian Journal of Philosophy 67: 59
83.
Menzies, Peter. 1996. Probabilistic Causation and the Pre-Emption Problem. Mind 105: 85117.
Menzies, Peter. 1998. How Justied Are the Humean Doubts about Intrinsic Causal Links? Communi-
cation and Cognition 31: 339364.
464 References
Menzies, Peter. 1999. Intrinsic versus Extrinsic Conceptions of Causation. In Sankey (1999): 313329.
Menzies, Peter. 2002. The Causal Ecacy of Mental States. In J. Monnoyer (ed.), The Structure of the
World: The Renewal of Metaphysics in the Australian School. Paris: Vrin.
Menzies, Peter, and Huw Price. 1993. Causation as a Secondary Quality. British Journal for the Philos-
ophy of Science 44: 187205.
Mill, J. S. 1961. A System of Logic. London: Longmans.
Miller, Richard W. 1987. Fact and Method: Explanation, Conrmation, and Reality in the Natural and the
Social Sciences. Princeton, N.J.: Princeton University Press.
Michotte, A. 1963. The Perception of Causality. Translated by T. R. Miles and Elaine Miles. London:
Methuen.
Noordhof, P. 1998. Problems for the M-set Analysis of Causation. Mind 107: 457463.
Noordhof, P. 1999. Probabilistic Causation, Preemption and Counterfactuals. Mind 108: 95125.
Noordhof, P. 2000. Ramachandrans Four Counterexamples. Mind 109: 315324.
Oakley, I. T., and L. J. ONeill. 1996. Language, Logic, and Causation: Philosophical Writings of Douglas
Gasking. Melbourne: Melbourne University Press.
Otte, R. 1986. Reichenbach, Causation, and Explanation. In A. Fine and P. Machamer (eds.), PSA
1986, volume 1, 5965. East Lansing, Mich.: Philosophy of Science Association.
Parascondola, M. 1996. Evidence and Association: Epistemic Confusion in Toxic Tort Law. In
L. Darden (ed.), PSA 1996, Part I. Philosophy of Science 63: S168S176.
Paul, L. A. 1998a. Problems with Late Preemption. Analysis 58: 4853.
Paul, L. A. 1998b. Keeping Track of the Time: Emending the Counterfactual Analysis of Causation.
Analysis 58: 191198.
Paul, L. A. 2000. Aspect Causation. Journal of Philosophy 97: 235256. Reprinted in this volume.
Pearl, Judea. 2000. Causality: Models, Reasoning, and Inference. Cambridge: Cambridge University Press.
Pietroski, P., and G. Rey. 1995. When Other Things Arent Equal: Saving Ceteris Paribus Laws from
Vacuity. British Journal for the Philosophy of Science 46: 81110.
Preyer, Gerhard, and Frank Siebelt (eds.). 2001. Reality and Humean Supervenience: Essays on the Philos-
ophy of David Lewis. New York: Rowman and Littleeld.
Price, Huw. 1994. A Neglected Route to Realism about Quantum Mechanics. Mind 103: 303336.
Prior, Elizabeth, Robert Pargetter, and Frank Jackson. 1982. Three Theses about Dispositions. Ameri-
can Philosophical Quarterly 19: 251253.
Quine, W. V. O. 1960. Word and Object. Cambridge, Mass.: MIT Press.
Quine, W. V. O. 1985. Events and Reication. In LePore and McLaughlin (1985): 162176.
Ramachandran, Murali. 1997a. A Counterfactual Analysis of Causation. Mind 151: 263277.
Ramachandran, Murali. 1997b. Redundant Causation, Composite Events, and M-sets. Acta Analytica
16/17: 149158.
Ramachandran, Murali. 1998. The M-set Analysis of Causation: Objections and Responses. Mind 107:
465471.
Ramachandran, Murali. 2000. Noordhof on Probabilistic Causation. Mind 109: 309313.
Ramsey, F. P. 1925. Universals. In Ramsey (1990): 830.
Ramsey, F. P. 1990. Philosophical Papers. Ed. D. H. Mellor. Cambridge: Cambridge University Press.
Reichenbach, Hans. 1928. Philosophie der Raum-Zeit-Lehre. Berlin: Walter de Gruyter. English transla-
tion published as Reichenbach (1958).
Reichenbach, Hans. 1958. The Philosophy of Space and Time. New York: Dover.
References 465
Restall, G. 1996. Truthmakers, Entailment, and Necessity. Australasian Journal of Philosophy 74: 331
340.
Rice, Hugh. 1999. David Lewiss Awkward Cases of Redundant Causation. Analysis 59: 157164.
Rosenberg, Alex. 1992. Causation, Probability, and the Monarchy. American Philosophical Quarterly
29: 305318.
Rowe, William L. 1993. Causing and Being Responsible for What Is Inevitable. In Fischer and Ravizza
(1993a): 310321.
Russell, Bertrand. 1953. On the Notion of Cause. In Russell, Mysticism and Logic. London: Penguin.
Salmon, W. C. 1984. Scientic Explanation and the Causal Structure of the World. Princeton, N.J.:
Princeton University Press.
Salmon, Wesley. 1994. Causality without Counterfactuals. Philosophy of Science 61: 297312.
Sankey, Howard (ed). 1999. Causation and Laws of Nature. Dordrecht: Kluwer.
Schaer, Jonathan. 2000a. Overlappings: Probability-raising without Causation. Australasian Journal of
Philosophy 78: 4046.
Schaer, Jonathan. 2000b. Trumping Preemption. Journal of Philosophy 97: 165181. Reprinted in this
volume.
Schaer, Jonathan. 2001. Causes as Probability-raisers of Processes. Journal of Philosophy 98: 7592.
Schier, S. 1991. Ceteris Paribus Laws. Mind 100: 118.
Schilpp, P. A. (ed.). 1963. The Philosophy of Rudolf Carnap. La Salle, Ill.: Open Court.
Shoemaker, S. 1980. Causality and Properties. In van Inwagen (1980): 109135.
Shope, R. K. 1978. The Conditional Fallacy in Contemporary Philosophy. Journal of Philosophy 75:
397413.
Skyrms, B. 1980. Causal Necessity: A Pragmatic Investigation of the Necessity of Laws. New Haven,
Conn.: Yale University Press.
Sober, E. 1980. Evolution, Population Thinking, and Essentialism. Philosophy of Science 47: 350383.
Sober, E. 1984. The Nature of Selection. Cambridge, Mass.: MIT Press.
Sober, E. 1985. Two Concepts of Cause. In P. D. Asquith and P. Kitcher (eds.), PSA 1984, vol. II, 405
424. East Lansing, Mich.: Philosophy of Science Association.
Sorensen, Roy. 1999. Seeing Intersecting Eclipses. Journal of Philosophy 96: 2549.
Sosa, Ernest, and Michael Tooley (eds.). 1993. Causation. Oxford: Oxford University Press.
Stalnaker, Robert. 1968. A Theory of Conditionals. In Nicholas Rescher (ed.), Studies in Logical
Theory. Oxford: Blackwell.
Stalnaker, Robert. 1972. Pragmatics. In Davidson and Harman (1972): 380397.
Stalnaker, Robert. 1980. A Defense of Conditional Excluded Middle. In William L. Harper, Robert
Stalnaker, and Glenn Peirce (eds.), Ifs: Conditionals, Belief, Decision, Chance, and Time, 87104. Dor-
drecht: Reidel.
Stapleton, J. 1994. Product Liability. London: Butterworths.
Steward, H. 1997. The Ontology of Mind. Oxford: Clarendon Press.
Stewart, I. 1989. Does God Play Dice? Oxford: Blackwell.
Suppe, F. 1979. Introduction to The Structure of Scientic Theories. Urbana, Ill.: University of Illinois
Press.
Suppe, F. 1989. The Semantic Conception of Theories and Scientic Realism. Urbana, Ill.: University of
Illinois Press.
Suppes, P. 1970. A Probabilistic Theory of Causality. Amsterdam: North-Holland.
466 References
Swain, Marshall. 1978. A Counterfactual Analysis of Event Causation. Philosophical Studies 34: 119.
Taylor, A. E. 1956. Plato: The Man and His Work. Cleveland, Ohio: Meridian.
Tichy, Pavel. 1984. Subjunctive Conditionals: Two Parameters vs. Three. Philosophical Studies 45: 147
179.
Todes, Samuel, and Charles Daniels. 1975. Beyond the Doubt of a Shadow: A Phenomenological and
Linguistic Analysis of Shadows. In D. Ihde and R. M. Zaner, (eds.), Selected Studies in Phenomenology
and Existential Philosophy. The Hague: Nijho.
Tooley, Michael. 1987. Causation: A Realist Approach. New York: Oxford University Press.
Tooley, Michael. 1990. Causation: Reductionism versus Realism. Philosophy and Phenomenological
Research 50 (supplement): 215236.
Toulmin, S. 1961. Foresight and Understanding: An Inquiry into the Aims of Science. Indianapolis, Ind.:
Indiana University Press.
Unger, Peter. 1977. The Uniqueness of Causation. American Philosophical Quarterly 14: 177188.
van Fraassen, Bas. 1966. Singular Terms, Truth-Value Gaps, and Free Logic. Journal of Philosophy 63:
481495.
van Fraassen, Bas. 1980. The Scientic Image. Oxford: Clarendon Press.
van Fraassen, Bas. 1989. Laws and Symmetry. Oxford: Oxford University Press.
van Inwagen, Peter (ed.). 1980. Time and Cause. Dordrecht: Reidel.
van Inwagen, Peter. 1987. Ability and Responsibility. Philosophical Review 87: 201224.
White, Morton. 1965. Foundations of Historical Knowledge. New York: Harper and Row.
Williams, Donald. 1953. On the Elements of Being. Review of Metaphysics 7: 318, 171192.
Woodward, J. 1984. A Theory of Singular Causal Explanation. Erkenntnis 21: 231262.
Woodward, J. 1990. Supervenience and Singular Causal Statements. In Knowles (1990): 215216.
Woodward, J. 1994. Review of Humphreys The Chances of Explanation. British Journal for the Philos-
ophy of Science 45: 353374.
Woodward, J. 2003. Making Things Happen: A Theory of Causal Explanation. Oxford: Oxford University
Press.
Yablo, Steve. 1986. Things. Ph.D. dissertation, University of CaliforniaBerkeley. Ann Arbor, Mich.:
University Microlms International.
Yablo, Steve. 1992a. Cause and Essence. Synthese 93: 403449.
Yablo, Steve. 1992b. Mental Causation. Philosophical Review 101: 245280.
Yablo, Steve. 2000. The Seven Habits of Highly Eective Thinkers. In Bernard Elevitch (ed.), Proceed-
ings of the Twentieth World Congress of Philosophy, volume IX: Philosophy of Mind and Philosophy of
Psychology. Charlottesville, Virginia: Philosophy Documentation Center.
References 467
General Index
Absences. See also Facts, negative; Void
and abnormality criterion, 295
of absences, 283
as background conditions, 173175
and bi, 284285, 288, 448449
cannot be causes, 25, 3436, 49, 53, 281, 291308
and causal explanation, 291, 293294, 301, 304
306
and two concepts of causation, 253
as causes and eects, 99102, 108109, 112, 248
249, 341
as causes and eects on Lewiss original analysis,
19, 2425, 144145, 447
common sense and causation by, 294297, 300,
306307
and intrinsic character of causation, 240241, 244,
248252
location of, 243, 248249
moral norms and the causal status of, 293, 296
299, 306
and the nature of causal relata, 4849, 77, 100,
151, 281282, 314
and neuron diagrams, 232
and nonrelational view of causation, 100, 283,
292, 298300, 304
ontological status of, 49, 100, 145, 248, 254, 281
283, 304, 448
and probabilistic causation, 376
and production, 226, 256, 260
spurious cases of causation by, 298, 301
and transitivity of causation, 198199, 252
Accommodation, rule of, 405
Action at a distance
and causation at a distance, 85
and double prevention, 29, 184, 198, 242243
as a far-fetched case, 38, 422
excluded from analyses of dependence and
production, 226, 232, 258259
and intrinsic character of causation, 240
in interaction diagrams, 439
and Lewiss treatment of early preemption, 24, 33,
373
and Locality thesis, 238, 242243
and trumping preemption, 6566, 215
Actual causation, 404, 410, 416
Agency, 6061
Agency theory of causation, 449, 455
Albert, David, 8
Alterations of an event, 8896, 99, 103104, 146,
215221, 345
Ancestral dependence maneuver
and backtracking, 234
and causation as inuence, 51, 93, 9596, 216
217
and the fragility strategy, 88
in Lewiss original analysis, 22, 39, 79, 121, 141
142, 206, 267, 310, 420, 445
and trumping preemption, 6770
Armstrong, D. M., 13, 32, 48, 50, 59, 6364, 319
Aronson, J., 456
Aspect analysis of causation, 43, 5253, 186189,
197, 205222, 330, 345, 365
Australian rules football, 286
Background conditions, 52, 143144, 146, 169,
172178, 190, 403. See also Normal conditions
Backtracking
and aspect analysis of causation, 220
and contrastive causation, 352
and contrastive explanation, 149150, 177
and Dependence thesis, 181, 200
and Lewiss original analysis, 1719, 78, 141,
233235, 445
and naive counterfactual analysis, 326
and neuron diagrams, 435
and production, 262
and similarity of worlds, 59, 165
and spurious causation, 163
and transitivity of causation, 247
Backup systems, 421423, 428, 451
Backward causation, 612, 17, 38, 226, 232, 240,
258259, 388, 390, 393, 400
Barker, Stephen, 401
Beebee, Helen, 25, 3436, 4850, 53, 401
Bennett, Jonathan, 20, 96, 130131, 229231, 253,
330, 335
Berkovitz, Joseph, 417
Bi, 53, 283288, 448449
Blocker. See Screener
Blueprint strategy. See Production
Bohm, David, 415
Byrne, Alex, 380
Canberra plan. See Causal functionalism
Carroll, John, 6364
Cartwright, Nancy, 12, 158, 403
Causal
action (see Bi)
functionalism, 76, 279287, 452453
models, 158166, 170178
process theories (see Physical connection accounts
of causation)
relevance, 359380
relevance neutralizer (see Neutralizer)
tendencies, 404, 410, 416
Causation at a distance, 85, 240, 388. See also
Action at a distance
Centering principle, 141, 162, 164, 166
Ceteris paribus conditional probabilities, 411
Ceteris paribus laws. See Laws
Chance, 7, 359, 377. See also Probabilistic
causation; Probability
background, 387388, 391392
overall world (ow-), 55, 387, 391398
Chance-raising analyses of causation, 387401
Charleston, Rob, 401
Chisholm, Roderick, 3, 348349
Chrisley, Ron, 401
Coady, David, 46, 54, 226
Collingwood, R. G., 144
Collins, John, 26, 5152, 75, 115, 272
Common eects problem, 261
Commonsense causal judgments, 197, 295296,
299301, 305307
Comparative similarity. See Similarity
Conceptual analysis and synthesis, 3039, 159
Conserved quantity theory of causation. See
Physical connection accounts of causation
Context dependence
of causation, 5, 31, 139, 141142, 327, 331, 341
345, 348, 441
of causal judgments, 178, 274275, 341
of counterfactuals, 5, 78, 141, 178, 341, 348
of explanation, 345
of sentences in general, 346
Context of inquiry, 143
Context of occurrence, 143
Contrastive account of causation, 42, 54, 330, 341,
351, 363364, 405
Contrast situation, 341343, 346347, 352353
Control, 440
Control conditional, 167, 169
Conway, John H., 424425, 428
Cotenability, 4
Counterfactual covariation, 51, 61, 257. See also
Inuence, causation as
Counterfactual dependence
and absences, 254, 294, 296, 298
of alterations of the eect on alterations of the
cause, 91
and ancestral maneuver (see Ancestral maneuver)
of aspects, 213
asymmetry or symmetry of, 388, 419
and causal explanation, 270
and causation, the basic idea, 3
as one of two concepts of causation, 53, 200, 225
227, 257
and two concepts of causation, 287
dened, 1, 109, 279
and dierence-making, 166
dicult cases to classify, 271274
kinds of, other than whetherwhether, 8688, 92
93, 113, 146
and naive counterfactual analysis, 326, 336
noncausal, 101, 332
not intrinsic, 279
not necessary for causation, 67, 181, 419420
not transitive, 39, 202
in Lewiss original analysis, 1516, 2223, 141, 387
probabilistic, 280
without production, 253, 256, 268
and production often coincide, 265266
and simple counterfactual analysis, 206
sucient for causation (see Dependence thesis)
and switching, 191
and vagueness, 219220
Counterfactuals
causation, do not settle facts about, 424426
and chance, 7
context dependence of (see Context dependence of
counterfactuals)
dened, 2
Dowe, 448449
Hume, 419429, 436440
intransitivity of, 93
Kvart semantics for, 376378
metalinguistic theory, 153
natural divergence, 377
nonmonotonicity, 34
StalnakerLewis semantics for, 47, 90, 219, 234,
321, 348349
tacit premise view, 349
vagueness of, 219220, 349
Counterlegals, 163, 278279
Covering law analysis. See Deductive-nomological
analysis
Cutting, 8182
Davidson, Donald, 1617, 209, 212, 291, 301, 310
314, 321
Davidsonian account of causation, 1617
Decision making, 269
Decreaser. See Increaser
Deductive-nomological analysis of causation, 77
De facto dependence analysis of causation, 52,
126135
Default settings, 160
Delaying or hastening, 35, 8687, 89, 92, 130133,
229230, 373
Dependence (as opposed to production). See
Counterfactual dependence
Dependence modulo G, 122124, 132, 134
Dependence prevention, 51, 110112, 117
Dependence thesis. See also Counterfactual
dependence
and absences, 198199
conicts with Intrinsicness, 226, 241242, 244
246, 256, 258259
470 General Index
conicts with Locality, 241243, 256
conicts with Transitivity, 182184, 193, 195202,
241242, 246248, 256, 267
cornerstone of counterfactual analysis, 226, 253
paired with Omissions, 226, 256257
stated, 52, 181, 225
Determinism, 309310
Dierence-making, 52, 39140, 143, 151154, 161
162, 166167, 169178
Dierentiators, 366370
Dispositions, 114117
Dorr, Cian, 117, 401
Double preemption. See Double prevention
Double prevention
and absences, 85
no causal process in, 184
causal relation extrinsic in, 83, 85, 110, 285
vs. causation, 41, 198
contrasted with probabilistic cases, 394, 400
and counterfactual covariation, 257
and delaying, 87
and Dependence thesis, 226, 241242, 252253
in detailed workings, 273, 286, 438
examples of, 2829, 8485, 241242, 263264,
374375
excluded from blueprint strategy, 258, 263
and explanation, 195196, 253
laws that do/do not permit, 243
and persistence, 268
preemptive, 102
and preemptive prevention, 116, 272
and probability-raising analysis, 374375
Dowe, Phil, 13, 91, 294, 401, 411, 417, 448449,
456457
Dretske, Fred, 6364, 405
Ducasse, C. J., 152
Duplication, 259, 263
Edgington, Dorothy, 380
Eells, Ellery, 403404, 413414, 417
Eects, problem of, 388
Egalitarian concept of causation, 227230,
388
Ehring, Douglas, 13, 350
Elga, Adam, 7, 8
Enabling conditions, 119120
Enfeeblers, 126127, 132, 134
Ennoblers, 119120, 122, 126127
Entropy, 8
Epiphenomena, problem of, 388
Events
as causal relata, 44, 211, 140, 227, 310314, 329,
342, 388
distinctness of, 44, 325, 228
extrinsic, 332
individuation of, 21, 22, 4548 (see also Fragility)
Kims theory of, 212
no negative events, 49, 291, 314315
nonoccurrence of, 20, 21
precise specication of, 411412 (see also Fine-
grained event analysis)
proigate theory of, 329330, 334
Evidence, causes as, 62
Experimental-and-control contrast, 152153, 167
Experimental conditional, 167, 169
Explanation
and absences, 291
absences can gure in causal explanation without
being causes, 54, 293, 304305
not all causal explanations report causation, 34,
301302
contrastive, 147150, 176178
Davidson and Mellor on, 301
and double prevention, 195196, 253, 267
and egalitarian concept of causation, 231
pragmatics of, 306307
process and program explanation, 303, 306
roles of dependence and production in, 270
and switching, 188
Explanation test, 188, 195196
Extrinsic absence of disabling factors problem, 245,
250251
Extrinsic presence of enabling factors problem,
245, 250, 252
Extrinsic properties, 155157
Factors, 151
Facts
aspects are as ne-grained as, 221
as causal relata, 49, 54, 227, 254, 292, 301302,
306, 310316, 321, 330
are not causal relata, 100
as causal explainers, 304
individuation of, 313
negative facts, 49, 54, 254, 314316, 319320
negative facts cannot have eects by being related
to them, 317
negative facts are located, 320
Fair, David, 13, 91, 283, 456
Fales, Evan, 454
Fallback scenario, 125, 127129, 132, 135136
Far-fetched examples, 3739, 6567, 113, 422
Field, Hartry, 97, 119, 183, 336, 342, 350
Fine, Kit, 6
Fine-grained event analysis, 45, 205, 207211, 354
355, 411413
Finkish dispositions, 115, 121
General Index 471
Finkish trumping, 52, 116117
Fodor, Jerry, 157
Fragility
and alterations of an event, 88, 99, 146, 216
of events, 45, 205, 211, 328
and late preemption, 4547, 8587, 331333,
420
and proliferation of spurious causes, 46, 8890,
206207, 229230, 328330
of propositions, 8788
and trumping preemption, 7071, 333335
Frankfurt, Harry, 95
Future similarity objection, 67
Ganeri, Jonardon, 133, 401
Garnkel, Alan, 147
Giere, Ronald, 158
Goldman, Alan, 380
Goodman, Nelson, 3, 348
Goosens, William, 422
Grice, H. P., 147, 151
Hall, Ned, 75, 117, 222, 380, 401
and aspects, 44
and two concepts of causation, 3132, 50, 53
and double prevention, 2829, 84, 92, 110, 374,
438
and preemptive prevention, 103104
and switching, 136
and transitivity of causation, 41, 44, 52, 97, 349
350
Hart, H. L. A., 143144, 167170, 295
Hastening. See Delaying or hastening
Hausman, Daniel, 343, 354, 417
Heathcote, Adrian, 455
Heinlein, Robert, 95
Hempel, Carl, 150, 177178
Hilpinen, Risto, 380
Hitchcock, Christopher, 27, 5556, 222, 330, 342,
363364, 371, 380
Holes, 282
Holland, 342
Honore, A. M., 143144, 167170, 295
Horwich, Paul, 326
Hume, David, 1, 7577, 119120, 169, 379, 419,
454
Humean supervenience, 446
Humphreys, Paul, 403, 406409, 416417
Immanent causation, 447
Increasers, 360362, 364, 366376
Indeterminism. See Probabilistic causation
Indirect causation, 285
Inertial behavior, 436437
Inuence, causation as, 9196, 103, 113, 146, 205,
215221, 345
Inhibitors of inhibitors, 394, 400. See also Double
prevention
Interaction diagrams, 439, 441
Interfering factors, 157158, 160, 162167, 169
170, 175176
Intervention, 144, 157, 164167
Intrinsic character of causation. See also
Intrinsicness thesis
and absences, 240241, 249252
not given a priori, 285
and bi, 285
and the blueprint strategy for production, 258
259, 262
and causal functionalism, 279281
and two concepts of causation, 53, 253, 256257,
285
and causation at a distance, 85, 240
and double prevention, 29, 8385, 110, 226, 244
245, 252
and late preemption, 82, 237239
more or less strict senses of same intrinsic
character, 239240, 244, 258259
as an overhasty generalization, 85, 101, 282
and quasi-dependence, 8285, 110
and singular causation, 456
and switching, 188
Intrinsicness thesis
and blueprint strategy for production, 258259, 262
bundled with Locality and Transitivity, 253, 256
257
conicts with Dependence, 226, 241242, 244
246, 256, 258259
conicts with Omissions, 248252
rst stated, 225
more carefully formulated, 239241, 244
Intrinsic properties, 155157, 159, 161
Intuitions about cases, 200201, 242
INUS conditions analysis of causation, 53, 432
Jackson, Frank, 116, 303
Johnston, Mark, 115
Joseph, Georey, 157
Keller, Simon, 93
Kim, Jaegwon, 22, 78, 101102, 155, 207209,
211212, 325, 342, 354
Kinds (vs. concepts) of causation, 255256, 285
286
Kripke, Saul, 115
Kripkean necessity, 456
472 General Index
Kvart, Igal, 27, 3839, 55, 183184, 187, 194, 413,
417
Kvarts counterfactual analysis of causation, 379
380
Lange, Marc, 380
Lawful entailment, 218221
Lawlike generalizations, 432434
Laws of nature
and absences, 260, 316
and action at a temporal distance, 243
and bi, 288289
and causal models of systems, 158167, 170172,
177
causes subsumed under, 60
ceteris paribus, 157158, 166167
and deductive-nomological analysis of causation,
77
determining truth of causal claims, 421429, 441
determining truth of counterfactuals, 5, 279, 377,
445, 448
and Dowe counterfactuals, 449
DretskeTooleyArmstrong account of, 6364,
319, 446, 455
dynamical, 262
exceptionless, 166, 279
of inertia and deviation, 14, 56, 430431
and intrinsic character of causation, 110, 198, 239,
244, 252
in Lewiss original analysis, 16, 232
as link between counterfactuals and causation,
420442
MillRamseyLewis account of, 6264, 425, 429,
446
and nomological entailment accounts of
causation, 1415
and physical connection accounts of causation,
449, 456
physicalism about, 432
primitivism about, 6364, 432
probabilistic, 19, 2627, 38, 79, 388, 392
quasi-Newtonian, 431442
realism about, 432
reductionist accounts of, 12
and regularity analysis of causation, 205, 214, 218
singular causation as instantiation of, 455456
and systems, 156, 159
in Tooleys antireductionist argument, 34
and the void, 278
zero-force, 161
Lears maxim, 315, 318, 448
Lewis, David, 117, 222, 355, 457
and absences as causes, 19, 2425, 36, 77, 79, 99
104, 145, 281284, 288, 292294, 447448
and absences of absences, 283
and action at a distance, 85, 422
and alterations of an event, 8896, 99, 103104,
146, 215, 345
and ancestral maneuver, 2224, 68, 79, 88, 9396,
121, 129, 141142, 216217, 310, 373, 445
and asymmetry of counterfactual dependence,
388
and background conditions, 144, 146, 169
and backtracking, 58, 1719, 78, 141, 149, 176
177, 233, 445
and backward causation, 67, 388, 390
and bi, 53, 283289
and cases dicult to classify, 286
and causal functionalism, 76, 279287
and causal relata, 7778, 100, 102, 140, 145, 151,
281283
and causation at a distance, 85, 388
and causation without causal relata, 4849, 53,
282283, 292
and centering principle, 141, 162, 166
and conceptual analysis, 30, 33
and context dependence of causation, 141142,
144
and context dependence of counterfactuals, 78,
141
and context dependence of explanation, 172
and counterfactual dependence, 8688, 9193,
101, 113, 146, 166, 287, 279, 292
and counterfactual variance, 61
and counterlegals, 279
and delaying and hastening, 35, 8687, 89, 92,
229231, 329
and dispositions, 115116
and double prevention, 8385, 87, 102, 285286,
400
and egalitarian notion of cause, 388
and events, 101102, 140, 208, 211212, 328329,
342, 354, 388
and explanation, 100101, 142, 147151, 172,
177, 293, 302304
and extrinsic events, 332
and facts as causes, 100
and far-fetched examples, 422
and fragility, 70, 8588, 8990, 99, 146, 208, 229
231, 328329, 333335
and future similarity objection, 6
and holes, 282
and Humean supervenience, 446
and indeterministic causation, 376, 387
and indirect causation, 285
and inuence, 2, 39, 51, 55, 9196, 102103, 113
114, 146, 205, 215218, 237, 345
General Index 473
Lewis, David (cont.)
and intrinsic character of causation, 8285, 110,
238239, 279, 281, 285
and kinds vs. concepts of causation, 285286
and laws of nature, 12, 6263, 279, 288289, 425,
429, 446
and limit assumption, 166
and mark transmission, 91
and mereology, 282
and miracles, 711, 78, 445
and missing relatum problem, 281283
and natural kinds, 286
and neuron diagrams, 231232, 291, 336, 435, 446
and nomological entailment accounts, 15, 77, 80
original analysis of causation, 1, 1524, 140, 320
321, 371, 387, 419
and overdetermination, 120
and perceiving causation, 7576
and possibility, 278
and possible worlds, 445
and pragmatics, 36, 142, 147, 346
and preemption, 23, 7992, 96, 120121, 129, 142,
146, 217, 327, 336, 372373, 388
and probabilistic causation, 19, 24, 55, 79, 370,
376, 387393, 404, 407409, 414, 416, 449450
and problems of eects and epiphenomena, 388
and proigate causation, 139140, 143, 145, 147,
208209
and quasi-dependence, 68, 8285, 109110
and redundant causation, 75, 80, 146, 208209,
326327, 332336
and regularity analyses, 1, 419
and reication of absences, 282
and rule of accommodation, 331, 336
and semantic satiscing, 280, 283
and semantics for counterfactual conditional, 34,
2021, 90, 93, 164, 166, 321, 348, 445
and similarity, 411, 141, 154, 162, 164, 176177,
335, 349, 445
and simple counterfactual analysis of causation,
67, 7578, 80, 279, 325, 327, 336
and simultaneous causation, 388
and singular causation, 388, 446452
and spoils to the victor, 32
and spurious causation, 70, 88, 216, 219220
and supervaluation, 330
and supervenience, 283284, 287288
and temporal asymmetry, 78
and theoretical entities, 452453
and transitivity of causation, 4142, 68, 78, 93
94, 9699, 102, 141, 201, 216217, 309, 379, 388,
390
and trumping preemption, 8082, 9293, 113, 333
and the void, 277278, 281282, 289, 306, 315
316
Life, game of, 424432
Limit assumption, 166
Locality thesis
bundled with Intrinsicness and Transitivity, 257
conicts with Dependence, 226, 241243, 248
conicts with Omissions, 248249, 252
and late preemption, 238
stated, 225
Lombard, Lawrence, 229231
Lowe, E. J., 310
Mackie, John
and dierence-making, 143, 151153, 162167,
178
and experiment-and-control contrast, 152153,
167
and explanation, 170
and eld of normal conditions, 160, 167, 178
and INUS conditions account, 53
and regularity analysis of causation, 1, 15, 22, 77,
258, 261
Mackie, Penelope, 87
Macrotaxonomy, 432435
Malcolm, Norman, 446
Mark transmission. See Physical connection
accounts of causation
Martin, C. B., 115
Masking, 52, 114115
Maslen, Cei, 5, 42, 54, 117
Matching. See Duplication
Maudlin, Tim
and concepts vs. kinds, 255
and extrinsic events, 332
and laws of inertia and deviation, 14, 56
and laws as primitive, 6364
and preemptive prevention, 26, 37
and third-factor account of causation, 16, 56
and philosophical trench warfare, 227
and semantics for counterfactuals, 234, 326
McDermott, Michael, 457
and double prevention, 84, 110
and minimal counterfactual suciency, 68, 70
71
and preemptive prevention, 102, 107108, 112,
272, 435438
and proigate theory of events, 329
and transitivity of causation, 96, 183, 186187,
201, 390
McGrath, Sarah, 25
Mellor, Hugh, 10, 4849, 54, 292, 301302, 306,
389
Mental causation, 178, 221
474 General Index
Menzies, Peter, 457
and absences as causes, 4850
and background conditions, 52
and continuous processes, 6768
and dierence-making, 50, 52
and intrinsic character of causation, 278279
and perception of causation, 454
and probabilistic causation, 406, 413414
and singular causation, 452452, 454455
Mereology, 186, 207208, 282, 325
Methodology, 3039, 183, 200202, 227230, 254
257, 274
Michotte, A., 454
Mill, J. S., 144, 147, 151, 169
Mimicking, 116
Minimal-counterfactual suciency, 6871
Minimal-dependence sets, 6870
Minimal suciency, 260266
Miniworlds, 153154
Miracle, 511, 78, 166, 435, 445
Morgenbesser, Sidney, 7
Naive counterfactual analysis of causation
not refuted by standard preemption cases, 54,
325336
and preemption, 120
and production, 258
stated, 120, 225, 233
target of the analysis, 227228
Natural kinds, 159, 286
Neta, Ram, 355
Network model of causation, 291294, 302, 307
Neuron diagrams, 1720, 231232, 335336, 434
435, 441, 446, 448
Neutralizers, 55, 365375
Newtonian mechanics, 421430
Nomological entailment accounts of causation, 13
15, 77, 80
Noordhof, Paul, 133, 401
Normal conditions and worlds, 50, 160167, 170
177
Normative considerations and causation, 25, 54,
296297
Objectivity of causation, 441442
Occasionalism, 284
Omissions. See Absences
Omissions thesis
conicts with Intrinsicness, 248252, 256, 258
259
conicts with Locality, 248249, 252, 256
conicts with Transitivity, 248 252, 256
paired with Dependence, 226, 256257
stated, 226
Opacity. See transparency assumption
Ordinary language, 453
Overall world chance (ow-chance). See Chance
Overdetermination, 3233, 51, 109, 120, 129130,
258, 327, 336, 387
Papineau, David, 401
Parascondola, M., 408
Pargetter, Robert, 116
Paul, L. A., 75, 117, 417
and aspect causation, 4344, 5253, 186187,
189, 197, 330, 345
and delaying or hastening, 87, 132, 237
and transitivity, 4344, 5253, 186187
Perceiving causation, 56, 7576, 453454
Persistence through time, 268, 447
Pettit, Philip, 303
Physical connection accounts of causation, 1314,
91, 283284, 411, 449, 456
Physicalism, 159
Pietrowski, Paul, 158
Platitudes. See Causal functionalism
Plato, 119120
Polarity problem, 123
Pragmatic considerations, 3637, 147, 405
Precision assumption, 71
Prediction, 440
Preemption
and alterations of an event, 88
and ancestral dependence maneuver, 96, 121, 129,
146, 234
and aspect analysis of causation, 219220
and causation as inuence, 92, 9596, 216,
237
and contrastive causation, 341
and cutting, 8182
and deductive-nomological analysis, 7980
and de facto dependence, 128
and delaying or hastening, 131, 237
and dependence modulo G, 123
double (see Double prevention)
vs. double prevention, 198
early, 23, 8082, 96, 214, 234, 262263, 372373,
388
and fragility, 46, 8587, 328
late, 23, 82, 8587, 146, 214215, 219220, 235,
237239, 263, 371
and overdetermination, 109, 120, 327, 336
and probabilistic causation, 55, 359, 371374
and probability-raising without causation, 406
and production, 262263
General Index 475
Preemption (cont.)
and quasi-dependence, 83, 238239
simple counterfactual analysis not refuted by,
325336
and spurious causation, 8990, 219220
and transitivity of causation, 35
trumping (see Trumping preemption)
and would-be dependence analysis, 111112
Preemptive prevention, 26, 37, 5152, 102104,
107117, 272, 435438
Prevention. See Absences as causes and eects;
Double prevention
Prior, Elizabeth, 116
Probabilistic causation. See also Chance;
Probability
and causal processes, 410411
causal relevance account of, 359380
and contrastive causation, 341342, 405
and event specication, 411412
and far-fetched cases, 38
and Lewiss original analysis, 19, 24, 403404
as nomological entailment, 15
problems for, 2627, 55, 79, 403409
and quantum mechanics, 415
and singular causation, 449451
Probabilistic relevance, 366368
Probability
increase (or decrease), 359376
lowering causes, 55, 359360, 403405
pool, 55, 407408, 413, 416
raising noncauses, 27, 55, 7980, 370, 403, 406
409
Processes
and causation by absence, 307
continuous, 6870
and double prevention, 85, 184
and interaction diagrams, 439
and intrinsic character of causation, 238239, 456
and physical connection accounts of causation, 91
and preemption, 237
and probabilistic causation, 410411
and production, 200, 260261
and switching, 188, 191, 193198
Production
analyzed as ancestral of proximate causation,
265
ancestral of dependence, typically coincides with,
254, 265266
blueprint strategy for, 257265
causal explanatory role of, 270
vs. dependence, 53, 200, 225227, 231, 240, 267
without dependence, 253, 256, 269
dicult cases to classify, 271274, 438
irrelevant to decision making, 269
required for persistence, 268
transitivity of, 264
Proigate causation, 139, 142143, 185, 249
Properties, 213, 220
Property causation, 221
Property instances. See Aspect analysis of
causation
Proximate causation, 258, 265, 438
Pure causal history, 265
Quantum mechanics, 415
Quasi-dependence, 6769, 8285, 109110, 113
Quine, Willard V. O., 1, 5, 348
Ramachandran, Murali, 27, 55, 6870, 82, 133
Ramsey, Frank, 317
Reductionist accounts of causation, 1216, 19, 27,
30, 31, 34
Redundant causation
and causation as inuence, 146, 205
dened, 80, 109, 111, 326327
and ne-grained events, 208
and fragility, 332333, 335336
and simple counterfactual analysis, 75
Regularity analysis of causation, 1, 19, 205, 209,
214215, 218, 220, 342, 449, 455
Reichenbach, Hans, 91
Relata, causal
absences as, 77, 226
aspects as, 43, 48, 205, 211212, 219, 221
causation without, 49, 53, 282283
events as, 4448, 78, 102, 140, 151, 206, 218, 220,
227, 291
factors as, 151
facts as, 49, 54, 227, 254, 292, 301302, 306, 310
316, 321, 330
problem of missing, 100, 281283
Relational/nonrelational views of causation, 292
300, 304306, 313, 317321
Remote causation, 439440
Responsibility, 440
Reverser. See Increaser
Rey, Georges, 158
Rice, Hugh, 132
Risk analysis, 408
Robustness. See Fragility
Rosen, Deborah, 404
Rosen, Jacob, 114, 117
Rule of accommodation, 331, 336
Russell, Bertrand, 345, 447
Salience, 189191, 197, 230231, 345
Salmon, Wesley, 13, 91, 404, 411
Schaer, Jonathan, 75, 117, 222, 417
476 General Index
and double prevention, 438
and trumping preemption, 28, 34, 5051, 113
114, 134, 257, 401
Schier, Stephen, 158
Screener, 367376
Second-class causation, 430
Selection function, 219, 349
Semantic satiscing, 280
Semifactuals, 378379
Shoemaker, Sydney, 12, 318, 322
Short-circuit structure, 182, 187, 192195
Similarity
and causal models, 160, 162165, 178
and centering principle, 141, 162
context dependence of, 154, 156, 178, 300, 348
349
future, 67, 335
metric, 299
and StalnakerLewis semantics for counter-
factuals, 49
and vagueness, 220
Simple counterfactual analysis, 109, 206, 423. See
also Lewis, original counterfactual analysis;
Naive counterfactual analysis
Simultaneous causation, 910, 17, 388
Singular causation, 56, 75, 388, 446456
Skyrms, Brian, 403
Slingshot argument, 301, 311, 313
Sober, Elliot, 404
Special sciences, 433435
Spoils to the victor, 3233, 59, 422
Spurious causation, 70, 8889, 100101, 163, 173,
210, 216, 218220, 249
Spurious correlations, 403
Stalnaker, Robert, 4, 93, 346
States of aairs, 213
Statistical mechanics, 78
Steward, Helen, 291
Stockholm syndrome, 123124
Subjunctive conditional. See Counterfactual
conditional
Subtraction test, 188, 196197
Supervaluations, 330
Supervenience of laws on occurrent facts, 6263
Suppe, Fred, 158
Suppes, P., 404
Sustaining causation, 175
Swain, Marshall, 233
Switching, 40, 134136, 182183, 187202, 217,
220221
Systems, 154166, 172, 177
Temporal asymmetry, 512, 17, 233
Third factor, 420
Thomson, Judith, 42
Threats, 436438
Tichy, Pavel, 349
Time travel, 911. See also Backward causation
Tooley, Michael, 13, 27, 31, 34, 6364, 451
Tort law, 408
Transeunt causation, 447
Transfer (of energy or momentum etc.). See
Physical connection accounts of causation
Transitivity of causation
and absences as causes and eects, 198200, 252
and ancestral maneuver, 23, 39, 68, 9395, 141,
309, 234235
and aspect analysis, 4344, 53, 186189, 197,
205206, 210220
assumed by Lewis, 141, 388
assumption a mere prejudice, 318
as bedrock datum, 35, 39, 181
and two concepts of causation, 50, 182183, 200,
226, 252253, 256
and causation as inuence, 9396
contrast analysis of alleged counterexamples, 42,
54, 351354
alleged counterexamples to, 29, 4043, 52, 119,
182202
defended, 3940, 9699, 181, 349
denied, 341342, 379, 390
and double prevention, 29, 50, 183184, 198,
246248
and ne-grained event analysis, 210, 354355
problem for probabilistic analysis, 394
and production, 200, 226, 253, 256258, 264
and short-circuit structure, 182, 193197
structure of alleged counterexamples to, 182202
and switching, 187197
and tracing of causal chains, 78
Transitivity thesis. See also Transitivity of
causation
and blueprint strategy for production, 258, 262
bundled with Intrinsicness and Locality, 253, 256
257
conicts with Dependence, 182184, 193, 195202,
241242, 246248, 256, 267
conicts with Omissions, 248 252, 256
stated, 225
Transparency assumption, 313314, 318
Trench warfare, 227, 239
Tropes, 213
Trumping preemption, 5971
and causation as inuence, 9293, 113114
and covariation without dependence, 257
and de facto dependence account, 133134
examples of, 2728, 5052, 214216
nkish (see Finkish trumping)
General Index 477
Transparency assumption (cont.)
and fragility of events, 333335
and hidden mechanisms, 3
left well alone, 401
as preemption without cutting, 8083
Type causation, 376
van Fraassen, 147, 158, 330, 333, 345
Variation test, 189, 194, 196197
Void, 188, 277278, 281282, 292, 304, 306, 315
316, 322
White, Morton, 16, 77
White, Nick, 355
Wittgenstein, Ludwig, 446
Woodward, James, 406, 409410, 417
Would-be-dependence analysis of causation, 51,
112113
Yablo, Stephen, 42, 52, 222, 237
478 General Index
Index of Examples
Absences as causes (apparently genuine cases)
Beckham, Owen, and Seaman (Beebee), 296
billiard balls (Kvart), 376
bushre caused by lack of rain (Beebee), 291
dog-owners negligence (Stapleton), 294296,
298300, 306
drug company fails to warn of side eects
(Stapleton), 296297, 299
enemys failure to attack (Hall), 249
father fails to grab child (Armstrong), 447
re door not closed (Beebee), 291, 293
geranium survives frostless winter (Beebee), 294
295
hunger (Lewis), 99
indecision causes inaction (Mellor), 316
no force, no acceleration (Mellor), 316
no spark, no re (Mellor), 314316, 320
park rangers negligence (Stapleton), 294296
plants not watered, 53, 145, 173174, 291295,
297301, 304306
vaccination (Lewis), 100
void (Lewis), 277278, 281282, 288, 292, 304,
306, 315316, 322
Absences as causes (apparently spurious cases)
Bill Gates, 25, 49
amethrower attack, 145, 149, 175176
geranium-eating koala, 295
Godzilla, 307308
lion, 298
Maites coee, 298, 300
meteor strike, 145, 150, 175176, 249
nerve gas, 36, 100, 145, 149, 174176
president, 249
Queen Elizabeth, 54
Aspects. See also Transitivity, putative counter-
examples involving aspects or parts of events
Billys death (Paul), 213
Background conditions and causes
famine and ulcerated stomach (Hart and Honore),
144, 149, 173
lightning, breeze, and re (Hart and Honore),
170
lung cancer (Menzies), 143, 147, 149, 171
presence of oxygen (Hart and Honore), 3637,
143144, 173, 228
Causation (miscellaneous)
doomed window, 10
time-traveling billiard ball, 911
Ceteris paribus laws
HardyWeinberg law, 161
natural selection, 156
Newtons rst law, 161
Newtonian law of gravitation, 156
Context-dependence of counterfactuals
Caesar in Korea (Quine), 5, 78, 348
malleability of gold (Chisholm), 348
New York City in Georgia (Goodman), 348
Contrastive causation
chess game drawn (Maslen), 347
cookie recipe nut allergy (Maslen), 343
electric circuit (Hausman), 343344
price of eggs, no safety guard (Maslen), 342
stolen bicycle (Dretske), 405
toothbrushing and meteorite (Maslen), 342
toothpaste (Maslen), 341
video rental (Maslen), 343
Contrastive explanation
eclipse of the moon, 148
visit to Melbourne (Lewis), 148
syphilis and paresis (Hempel), 150, 177178
Counterfactuals (miscellaneous)
Billy and Suzys conversation, 20
coin toss you would have won (Morgenbesser), 7
egg uncooks and unbreaks (Elga), 8
Kennedy assassination (Maudlin), 419
Nixon and the nuclear button (Fine), 6, 335
stock market (Kvart), 378
switching on a light (Armstrong), 448
Deductive nomological analysis, diculties for
barometer and storm, 7778
neuron diagram, 261262
De facto dependence
Beamons jump and meteor (Yablo), 122, 124
125, 127
exercise (Yablo), 128
hydration (Yablo), 127128
Delaying and hastening
Brown delays Joness death (Lombard), 229230
doctor keeps patient alive until 4:12 (Lewis), 35,
229230
forest re (Bennett), 96, 130132, 229231, 253
kicked bowling ball (Yablo), 133
lightning (Yablo), 131132
Ned Kellys death (Lewis), 8587, 92, 338, 373
poison and pudding (Lewis), 89, 114
Slow Joe and Quick-Draw McGraw (Rice), 132
133
Double prevention. See also Transitivity, putative
counterexamples involving double prevention
air brakes (Lewis), 286
AND gate (Hall), 272274
billiard balls (Lewis), 8485, 290
Billy trips (Hall), 183184, 198, 248, 267, 270
Double prevention (cont.)
Drunken Bad Guys (Hall), 192203
escort ghter and interceptor, 84, 92, 241245,
249, 252, 256, 269, 374375
inhibitor of inhibitor (Ramachandran), 393, 400
neuron diagram, 2829, 242, 263264, 393, 400
Nixon, Haig, and Joe Blows breakfast
(McDermott), 8485, 290
Sally pushes Billy, 50
sitting on the sidelines (Hall), 269
squadrons of bombers and interceptors (Lewis),
105
trigger mechanism (Schaer), 438
Event extrinsically specied
Suzys email message, 22
Events that are not distinct
battles of World War I (Coady), 325
door closing and slamming (Kim), 2122, 325,
334
killing and widowing, 102, 332
shattering bottle, 2122
Experimental-and-control contrast
chestnut and hammer (Mackie), 152, 162165
Fine-grained events
cat sneeze (Paul), 207209
walking and talking, 45
Finkish trumping
sorcerer and window (Collins), 116
Fragility of events. See also Delaying and
hastening
death of Socrates (Bennett), 338339
explosion on sun, 206, 338
major and sergeant (van Fraassen), 333335
postponement or cancellation (Lewis), 86,
329
rocks thrown at bottle, 328333
sore throat, 46
vigorous or feeble ring (Lewis), 335336
Inuence without causation
body heat (Paul), 217, 219
injection and tooth drilling (Mellor), 321
Intervention
pendulum (Menzies), 157
Intrinsicness
Billy and his rock are absent, 82
neuron diagram, 250252
rocks make random jumps, 83
Laws and counterfactuals
computer word processor (Maudlin), 432434
Conways Game of Life (Maudlin), 424431
Masking
meaning addition by plus, 115
poison and antidote (Lewis), 115
Methodological points
lost keys (Hall), 255
motherhood (Maudlin), 255256
peculiar taxonomy for birds (Lewis), 286
theories of sh (Mellor), 311
Overdetermination, symmetrical
bowling balls (Yablo), 130
neuron ring (Lewis), 335336
rocks thrown at window, 3233
Physical connection accounts
Billy ushes when Suzy kisses him, 14
Preemption, early
backup gunman, 68
deected bowling ball (Yablo), 120121, 129
Kennedy, Oswald, and the grassy knoll (Kvart)
(probability), 374
neuron ring, 232, 234235, 262
neuron ringprobabilistic (Kvart), 372337
neuroscientist and Jones (Frankfurt), 9596
Suzy, rock, and catapult, 266
Preemption, late
cat pounces (Paul), 214217, 219220
bowling balls (Yablo), 129
neuron ring, 235236
smart rock (Yablo), 237
rocks thrown at bottle/window, 23, 37, 4647, 51,
8283, 8590, 92, 146, 200, 235241, 256, 259,
327333, 371
Preemption of merely potential alternative
exquisitely sensitive backup system (Hall,
Maudlin), 276, 421423
Preemptive prevention
backup escort ghter (Hall), 103104, 271272
elder, backup elder, and window (Collins), 26,
37, 51, 107108, 110112, 272, 436438
elder, and Halleys comet (Collins), 108, 110,
112113
elder, wall, and window (McDermott), 26, 51,
102103, 107108, 110112, 272, 435438
shield of invulnerability (Hall), 272
Yale shadow puzzle, 103
Probabilistic causation
airline bomb disaster (Lewis), 79, 370
attack dog and avalanche (Kvart), 365366, 368
bombarding an atom (Armstrong), 450
bullets precise momentum and spin (Kvart), 369
370
carbon-14 decay (Hitchcock), 415416
comeback team (Kvart), 360361
480 Index of Examples
doctor prescribes 100ml (Hitchcock), 363365
faucet and pipes, pool (Kvart), 367
zzled stimulatory signal, 2627
hole-in-one (Rosen), 404405
inhibitor of inhibitor (Ramachandran), 393, 400
lottery ticket (Hitchcock), 407
lung cancer and skin cancer (Hitchcock), 411
412, 416
photons and polarizer (Hitchcock), 409, 413, 416
radioactive decay, 150, 388
rope-pulling (Hitchcock), 409, 416
smoking, coughing, and cancer (Hitchcock), 403
404
spontaneous ring (Ramachandran), 389390
spontaneous shattering (Maudlin), 442
stone and outward opening window
(Ramachandran), 397398
student taking exam (Kvart), 361
teratogen, 408
two carcinogens (Woodward), 406408
vase and bullets (Hitchcock), 371372, 410416
Production and persistence
electron (Hall), 268
teletransportation (Hall), 276
Proigate causes
big bang causes todays snowfall (Hall), 228
birth causes my death (Morgenbesser), 35, 101
buttery eect, 310
car crash (Lewis), 142
cow, glass of water, plants (Beebee), 301
Hermes and Apollo (Lewis), 35, 101, 172
loss of a nail, 310, 439
pressing t key (Hall), 185
Switching, interactive. See also Transitivity,
putative counterexamples involving interactive
switching
basketball assist (Maudlin), 440441
kiss (Hall), 136, 191
Mischievous Bad Guys (Hall), 194
Switching, noninteractive
muscle cramp (Hall), 190
Transitivity
lightning and roast pig (Maslen), 349
sti drink (Ramachandran), 390, 394395
Transitivity, putative counterexamples involving
absences as causes and eects
collision avoidance (Hall), 199
damned if you do, damned if you dont (Lewis),
98
deadly double dose (Hall), 97, 105, 199
preemptive strike (Hall), 199
aspects or parts of events
dance performance (Hall), 186
dog-bite (McDermott), 97, 183, 186187, 201, 381
purple re (Ehring), 350351, 353354
red rework (Thomson), 4244
skiing accident (Paul), 43, 205214, 218
skiing accident and writing cramp (Paul), 209212
double prevention
alarm clock (Hall), 97, 246247
Billy trips (Hall), 183184, 198, 247248, 267, 270
bomb outside the door (Field), 4042, 97, 119,
123124, 127128, 183, 350352, 354355
Drunken Bad Guys (Hall), 192198, 200203
forest re (Bennett), 96, 130131, 229231, 253
inert network (Hall), 4142, 9798, 247
interactive switching
Clever Good Guys (Hall), 192194
forgetting your birthday (Maslen), 350, 352353
greeter and car salesmen (Maudlin), 440441
severed nger (Kvart), 183184, 187, 194, 202
shock C (McDermott), 9697
Singh and Patel (Ramachandran), 390, 394395
train tracks, 40, 42, 134136, 187191, 197, 217
218, 220221
Trumping
Brady Bunch vacation (Hall), 73
color-eld laws (Schaer), 6570, 73
jumping cats (Paul), 215217
major and sergeant (van Fraassen), 51, 6770, 73,
81, 83, 9293, 113, 333335
Merlins limited options (J. Rosen), 114
Merlin and Morgana (Schaer), 5973, 81, 83, 93,
113114, 134
neuron polarity, 2728
Index of Examples 481

Causation and Counterfactuals PDF

Uploaded by

Copyright:

Available Formats

Causation and Counterfactuals PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Causation and Counterfactuals PDF

Uploaded by

Copyright:

Available Formats

What are some of the main topics discussed in the document?

What are some of the main topics discussed in the document?

What philosophical issues related to causation and counterfactuals are examined?

What philosophical issues related to causation and counterfactuals are examined?

CAUSATION AND COUNTERFACTUALS

E D I T E D B Y JOHN COLLI NS, NED HALL, AND L. A. PAUL

. Maslen argues that in

, it says that causa-

on c) are not all the events it causesthe present theory

.) The point either way is that c speaks to a need

F does not lawfully entail the color of the paper on my desk,

and e, c is a cause of e relative to contrast event c

g, is a cause of event e, relative to

g, i had any events from fc

g occurred, then an event from

g would have occurred.13

. In that case account

, is a cause of event e, relative to contrast event e

had happened then e

: bullet 2 did not hit the vase (until t

screens o B from C and lowers the chance of C (without an increaser). So G

, a counterpart of the above E, is also a neutralizer for B and C (though not

e at a distance not smaller than ds from the window, not

is a stable screener for B and C, and B is not a cause of E

for the same

is a neutralizer for B and C, so B does

); hence A comes out as a cause of

Dependence is asymmetric: If A depends on B, B cannot also depend on A. (This

Causation is a relation between events; events are particularsthey have specic

Singular (brute, one-o) causation is permitted: There may be causation without

The transitivity of causation is assumed: If A causes B, and B causes C, then A

), the chance of Es ring (E

) is higher than it would

raises the chance of E

, which is plainly incorrect, since the C

does not raise the

would have been higher than it actually is.

might have occurred spontaneously, in which case owch(E

does not raise the owchance of

. So, being an owchance-raiser is not

) would not be lower than it is if A

had not occurred: B

in the rst place.

; but it fails come out as such by

had not occurred, then D

might have occurred, in which case

raises these owchances relative to the nonoccurrence of D*; in other words, it

. The idea that causes

): As noted earlier, owch(E

were to occur without D

, which likewise W-raises owch(E

occurs) and prevents F from ring, thereby

comes out a cause:

had occurred without A

) would have been 1; and if neither C

would have had zero owchance of occurring.

comes out a cause of E

g, say, we get the correct verdict that A

. To resolve this problem, then, we must nd a signicant dierence be-