The Geography of Inter-State Resource Wars

Francesco Caselliy, Massimo Morellizand Dominic Rohnerx

July 23, 2014

Abstract
We establish a theoretical as well as empirical framework to assess the role of
resource endowments and their geographic location for inter-State conict. The main
predictions of the theory are that conict is more likely when at least one country
has natural resources; when the resources in the resource-endowed country are closer
to the border; and, in the case where both countries have natural resources, when the
resources are located asymmetrically vis-a-vis the border. We test these predictions
on a novel dataset featuring oileld distances from bilateral borders. The empirical

We wish to thank Johannes Boehm, Cyrus Farsian, Patrick Luescher and Wenjie Wu for excellent
research assistance. Helpful comments from Robert Barro, Luis Corchon, Tom Cunningham, Oeindrila
Dube, Joan Maria Esteban, Erik Gartzke, Michael Greenstone, Sebnem Kalemli-Ozcan, Hannes Mueller,
Peter Neary, Nathan Nunn, Elias Papaioannou, Costantino Pischedda, Giovanni Prarolo, Jack Snyder,
Silvana Tenreyro, Mathias Thoenig, Andrew Wood, Pierre Yared, Fabrizio Zilibotti, three anonymous
referees, and conference and seminar participants in Barcelona, Bocconi, Copenhagen, East Anglia, Harvard, Lausanne, LSE, Lucerne, Manchester, Munich, NBER Political Economy Programme, NBER Income
Distribution and Macroeconomics Programme, Oxford, Princeton, Sciences Po Paris, SED, St. Gallen,
ThReD, York, and Zurich are gratefully acknowledged.
y
London School of Economics, BREAD, CEP, CEPR, CFM, and NBER. Email: f.caselli@lse.ac.uk.
Acknowledges nancial support from the Leverhulme Trust.
z
Columbia University, THRED and NBER. Email: mm3331@columbia.edu. Financial support of the
Program for Economic Research at Columbia University is gratefully acknowledged.
x
University of Lausanne. Email: dominic.rohner@unil.ch. Financial support from the Swiss National
Science Foundation (SNF grant no. 100014-122636) is gratefully acknowledged.

analysis shows that the presence and location of oil are signicant and quantitatively
important predictors of inter-State conicts after WW2.

Introduction

Natural riches have often been identied as triggers for inter-state war in the public debate
and in the historical literature.1 The contemporary consciousness is well aware, of course,
of the alleged role of natural resources in the Iran-Iraq war, Iraqs invasion of Kuwait, and
the Falklands war. At the moment of writing, militarized tensions involving territorial
claims over areas known, or thought, to be mineral-rich exist in the South China Sea, the
East China Sea, the border between Sudan and South-Sudan, and other locations. But
the historical and political science literatures have identied a potential role for natural
resources in dozens of cases of wars and (often militarized) border disputes, such as those
between Bolivia and Peru (Chaco War, oil, though subsequently not found), Nigeria and
Cameroon (Bakassi peninsula, oil), Ecuador and Peru (Cordillera del Condor, oil and other
minerals), Argentina and Uruguay (Rio de la Plata, minerals), Algeria and Morocco (Western Sahara, phosphate and possibly oil), Argentina and Chile (Beagle Channel, sheries
and oil), China and Vietnam (Paracel Islands, oil), Bolivia, Chile, and Peru (War of the
Pacic, minerals and sea access).2
However, beyond individual case studies there is only very limited systematic formal
and empirical analysis of the causal role of resources in inter-state conict, and of the

E.g. Bakeless, 1921; Wright, 1942; Westing, 1986; Klare 2002; Kaldor et al., 2007; De Soysa et al.,

2011; and Acemoglu et al., 2012.

2
References for these conicts include: Price (2005) for Nigeria-Cameroon, Franco (1997) for Ecuador
and Peru, Kocs (1995), for Argentina and Uruguay and Algeria and Morocco, BBC (2011) for Algeria and Morocco, Anderson (1999) for China and Vietnam, Carter Center (2010) for the War of the
Pacic.

Other examples of (militarized) border disputes over areas (thought to be) rich in oil and

other resources include Guyana-Suriname, Nicaragua-Honduras, Guinea-Gabon, Chad-Libya, BangladeshMyanmar, Oman-Saudi Arabia, Algeria-Tunisia, Eritrea-Yemen, Guyana-Venezuela, Congo-Gabon, Equatorial Guinea-Gabon, Greece-Turkey, Colombia-Venezuela, Southern and Northern Sudan (see Mandel,
1980; McLaughlin Mitchell and Prins, 1999; Carter Center, 2010).

underlying mechanisms. This paper aims to begin to ll this gap.

The key idea of the paper is to relate the likelihood of conict between two countries to
the geographical location of natural-resource deposits vis-a-vis the two countriesbilateral
border. The reasoning is simple: reaching, seizing, and holding on to areas belonging
to another country is progressively more di cult and costly the further away these areas
are from the border. The further an advancing army has to go, the more opportunities
the defender has to stop the advance, the longer and more stretched the supply lines
become, the greater the likelihood that the local population will be hostile, etc. Therefore,
if countries do indeed engage in military confrontations in order to seize each others
mineral reserves, as hypothesized in the case-study literature, they should be relatively
more tempted when these reserves are located near the border. Accordingly, we ask whether
countries are more likely to nd themselves in conict with countries with mineral deposits
near the border than with neighbors with minerals far away from the border.
As a preliminary check on the plausibility of this, Figure 1 presents a simple scatterplot
which suggests that the geographic location of oil deposits could be related to cross-country
conict. Each point in the graph is a pair of contiguous countries. On the vertical axis we
plot the fraction of years that the pair has been in conict since World War II, while on
the horizontal axis we measure the (time average of) the distance to the bilateral border
of the closest oil eld. (Clearly only country pairs where at least one country has oil elds
are included).3 The graph clearly shows that country pairs with oil near the border appear
to engage in conict more often than country pairs with oil far away from the border [the
correlation coe cient is -.11 (p-value: 0.01)].
The crude correlation in Figure 1 could of course be driven by unobserved heterogeneity
and omitted variables. For example, it could be that some countries that have oil near the
border just happen to be more belligerent, so that country-pairs including such countries
spuriously ght more often. Hence, the rest of the paper engages in a more careful, modelbased empirical investigation that controls for omitted factors, including country xed

Note that for visual convenience we have trimmed both axes, removing the 1% outliers with highest

levels on the axes. The data in the gure is described in detail in Section 3.1.

60
% years hostility
20
40
0

500

1000
1500
Min. oil distance (in km)

2000

2500

Figure 1: Oil distance from the border and bilateral conict

eects, and is sensitive to the issue of border endogeneity.

To see the benet of focusing on the geographical location of resource deposits, contrast
our approach with the (simpler) strategy of asking whether countries are more likely to
nd themselves in conict with neighbors who have natural resources than with neighbors
that are resource-less. There are two shortcomings of this strategy. First, it tells us little
about the mechanism by which resource abundance aects conict. For example, it could
just be that resource-abundant countries can buy more weapons. Second, the potential for
spurious correlation between being resource-rich and other characteristics that may make a
country (or a region) more likely to be involved in conict is non-trivial. For both reasons,
while we do look at the eects of resource abundance per se, we think it is crucial to focus
most of the analysis on the geographic distribution of resource deposits.
To the best of our knowledge, there is no theoretical model that places conict (whether
over resources or otherwise motivated) inside a geographical setting. Given the prominence
of the concept of territorial war, this omission may seem surprising. Hence, we begin the
paper by developing a simple but novel two-country model with a well-dened geography,
where each country controls some portion of this geography, so there is a meaningful notion
of a border, and where the two countries can engage in conict to alter the location of
the border. This provides a simple formalization of territorial war (which could have

applications well beyond the present focus on resource wars).

We use our model of territorial war to generate testable implications on the mapping
from the geographical distribution of natural resources to the likelihood of conict. We
assume that each of the two countries may or may not have a resource deposit (henceforth
oil, for short). The one(s) that have oil have the oil at a particular distance from the initial
bilateral border. If a war leads one of the two countries to capture a portion of territory
that includes an oil eld, the control over the oil eld shifts as well.
We obtain rich testable implications which go well beyond the simple intuition with
which we have opened this Introduction. The model belongs to a much more general class of
models of conict where one players gain (gross of the cost of engaging in conict) equals
the other players loss. We remark that in such games, under very general conditions,
the likelihood of conict is increasing in the asymmetry of payos. Increases in payo
asymmetry make the player which is expected to win more aggressive, and the one that
is expected to lose less aggressive. Since one party can initiate conict unilaterally, the
former eect tends to dominate.
Hence, the presence and geographic distribution of natural-resource deposits increases
conict if it increases payo asymmetry. Compared to the situation where neither country
has oil, the appearance of oil in one country clearly increases payo asymmetry: the
heightened incentive of the resource-less country to seek conict to capture the others oil
tends to dominate the reduced conict incentive of the resource-rich country (which fears
losing the oil). Similarly, ceteris paribus, payo asymmetry increases with the proximity
of the oil to the border: as the oil moves towards the border the incentive of the oil-less
country to ght increases more than the incentive for the oil-rich one is reduced. When
both countries have oil, conict is less likely than when only one does, but more likely
than when there is no oil at all. More importantly, conditional on both countries having
oil, the key geographic determinant of conict is the oil eldsasymmetric location: the
more asymmetrically distributed the oil elds are vis-a-vis the border the more likely it is
that two oil-rich countries will enter into conict. The overall message is that asymmetries
in endowments and location of natural resources translate into asymmetries in payos and
are thus potentially important determinants of territorial conict.
While our theory applies to any type of resource endowment, our empirical work fo-

cuses on oil, for which we were able to nd detailed location information (and which is the
resource most commonly conjectured to trigger conict). We test the models predictions
using a novel dataset which, for each country pair with a common border (or whose coastlines are relatively near each other), records the minimum distance of oil wells in each of
the two countries from the international border (from the other countrys coastline), as
well as episodes of conict between the countries in the pair over the period since World
War II.
We nd that indeed having oil in one or both countries of a country pair increases the
average dispute risk relative to the baseline scenario of no oil. However, this eect depends
almost entirely on the geographical location of the oil. When only one country has oil,
and this oil is very near the border, the probability of conict is more than three times as
large as when neither country has oil. In contrast, when the oil is very far from the border,
the probability of conict is not signicantly higher than in pairs with no oil. Similarly,
when both countries have oil, the probability of conict increases very markedly with the
asymmetry in the two countriesoil locations relative to the border.
Our results are robust to concerns with endogeneity of the location of the border, because they hold when focusing on subsamples of country pairs where the oil was discovered
only after the border was set; in subsamples where the border looks snaky, and hence
likely to follow physical markers such as mountain ridges and rivers; and in subsamples
where the distance of the oil is measured as distance to a coastline rather than to a land
border. They are also robust to controlling for a large host of country and country-pair
characteristics often thought to aect the likelihood of conict. Since country xed effects are included, they are also robust to unobservable factors that may make individual
countries more prone to engage in conict.
Most theoretical work on war onset in political science and economics takes the belligerentsmotives as given. The objective is rather either to study the determinants of ghting
eort (Hirshleifer, 1991, Skaperdas, 1992), or to identify impediments to bargaining to
prevent costly ghting (Bueno de Mesquita and Lalman, 1992, Fearon, 1995, 1996, 1997,

Powell, 1996, 2006, Jackson and Morelli, 2007, Bevi and Corchn, 2010).4 Our approach
is complementary: we assume that bargaining solutions are not feasible (for any of the
reasons already identied in the literature), and study how the presence and location of
natural resources aect the motives for war.
The paper is thus closer to other contributions that have focused on factors that enhance the incentives to engage in (inter-state) conict. On this, the literature so far has
emphasized the role of trade (e.g., Polachek, 1980; Skaperdas and Syropoulos, 2001; Martin et al., 2008; Rohner et al., 2013), domestic institutions (e.g., Maoz and Russett, 1993;
Conconi et al., 2012), development (e.g., Gartzke, 2007; Gartzke and Rohner, 2011), and
stocks of weapons (Chassang and Padr i Miquel, 2010). Natural resources have received
surprisingly little systematic attention in terms of formal modelling or systematic empirical
investigations. Acemoglu et al. (2012) build a dynamic theory of trade and war between
a resource rich and a resource poor country, but their focus is on the interaction between
extraction decisions and conict, and they do not look at geography. De Soysa et al. (2011)
cast doubt on the view that oil-rich countries are targeted by oil-poor ones, by pointing
out that oil-rich countries are often protected by (oil-importing) superpowers.5
Unlike in the case of cross-country conict, there is a lively theoretical and empirical
literature, nicely summarized in van der Ploeg (2011), on the role of natural resources
in civil conict. The upshot of this literature is that natural-resource deposits are often
implicated in civil and ethnic conict.6 Our paper complements this work by investigating

These authors variously highlight imperfect information, commitment problems, and agency problems

as potential sources of bargaining failure. See also Jackson and Morelli (2010) for an updated survey.
5
De Soysa et al. (2011) also nd that oil-rich countries are more likely to initiate bilateral conict against
oil-poor ones. Colgan (2010) shows that such results may be driven by spurious correlation between being
oil rich and having a revolutionary government. In Appendix B we look at a similar directed dyads
approach and nd that, in our sample, oil-rich countries are relatively less prone to be (classied as)
revisionist, attacker, or initiator of conict, and that their propensity to attack is decreasing in their oil
proximity to the border. This dierence in results could be due to dierences in sample (we only look at
contiguous country pairs), or methods (we include a full set of country and time xed eects and various
additional controls).
6
The vast majority of the civil-conict literature focuses on total resource endowments at the country
level (see, e.g. Michaels and Lei, 2011, Ross, 2012, van der Ploeg and Rohner, 2012, and Cotet and Tsui,

whether the same is true for international conict.7

The remainder of the paper is organized as follows. Section 2 presents a simple model
of inter-state conict. Section 3 carries out the empirical analysis, and Section 4 concludes.

The Model

2.1

Preliminary Remarks: Asymmetric Payos and Conict

Many conict scenarios can be crudely captured by the following static, two-player game:
Player B

Player A

Action 0

Action 0

Action 1

0; 0

x + cA ; x + cB

Action 1 x + cA ; x + cB

x + cA ; x + cB

where x,cA ,cB are real numbers. Action 0 is a peace action that, if played by both
parties, maintains the status quo, here normalized to (0; 0). Action 1 is a conict

2013, for recent examples and further references), but recently a few contributions have begun exploiting
within-country distributional information. For example, Dube and Vargas (2013) and Harari and La
Ferrara (2012) nd that localities producing oil are more prone to civil violence; Esteban et al. (2012) nd
that groups whose ethnic homelands have larger endowments of oil are more prone to being victimized;
Morelli and Rohner (2014) nd that inter-group conict is more likely when total resources are more
concentrated in one of the ethnic groups homelands. None of these studies make use of information on
the distance of natural resource deposits from country/region/ethnic homeland borders. Needless to say,
there are also several studies examining the eect of the spatial distribution of resources and geographic
features on outcomes other than conict (e.g. Caselli and Michaels, 2013, for corruption; Alesina et al.,
2012, for development).
7
Much as in the literature linking resources to domestic conict, our results imply that the net gain
from resource discoveries may be well below the gross market value of the discovered reserves. Aside from
the risk of losing the oil to its neighbors, countries have to factor in the economic cost of ghting to protect
it. Based on their review of the literature Bozzoli et al. (2010) conclude that mass conict causes GDP
growth to fall by between 1 and 3 percentage points. Using our preferred specication for the probability
of conict, these values imply that a country which nds oil right at its border (with a country that has
no oil) should expect to lose between 1 and 3 percent of GDP to war every 9 years or so.

action, such as initiating a war. The parameter x ( x) is the expected (gross) payo of
the conict to player A (B). If x > 0 player A is the expected winner. For example, x
could represent the capture of a strategic location or a mineral resource deposit currently
located in country B, weighted by the probability that A succeeds at capturing it. Finally,
c is a country-specic cost (or benet if positive) of undertaking the conict action.8
The condition for observing peace, dened as neither player playing the conict action,
is that
cB

cA :

Hence, if conict is usually costly (i.e. most of the time cB < 0, cA < 0), we will typically
see conict unless x is relatively small in absolute value. The absolute value jxj is a measure
of payo asymmetry: it captures both the extent of the expected gains of one player, and
the extent of the expected losses of the other. Hence, in conict games we expect to observe
conict when payos are asymmetric.
In real world situations the prize from conict jxj is often persistent over time. For
example a strategic location often retains its value over years or decades. Yet conict
among two players is only observed some of the time. To capture this pattern, we can
assume that the cost of the conict action, ci , is a random variable. The idea is that there
are good times and bad times to ght. For example, the perceived cost of conict
may be particularly low during an economic boom, or if the opponent is going through a
period of political upheaval.
While it is natural to think of ci as being negative most of the time, we can also
imagine situations where ci is occasionally positive, reecting the fact that sometimes
countries have very compelling ideological or political reasons to ght wars. For example,
governments facing a collapse in domestic support have been known to take their countries
to war to shore up their position by riding nationalist sentiments. In other cases they have
felt compelled (or at least justied) to take action to protect the interests of co-ethnic
minorities living on the other side of the border. Hence, it makes sense to assume that c

Needless to say, the applicability of this framework goes well beyond international (or civil) conicts.

It extends to, e.g., price wars over market share, industrial disputes, divorce, and many others.

is a random variable which takes values on the real line.

Suppose then that h : R ! R+ is the probability density function of ci , i = A; B, and
H is the corresponding cumulative distribution function. Then the probability of observing
peace is H(x)H( x). How does this probability change with changes in payo asymmetry?
By inspection, we have immediately the following
Remark 1. The probability of peace is nonincreasing in jxj if and only h(jxj)=H(jxj)
h( jxj)=H( jxj).
In other words, increases in payo asymmetry always increase conict if the Inverse
MillsRatio of the cost of conictdistribution evaluated at a positive value of this cost
is less than the Inverse Mills Ratio evaluated at the symmetric negative value. Clearly
H(jxj) > H( jxj) so we would expect this condition to hold much of the time. Indeed,
it holds in all cases where h is symmetric and single peaked around a negative mode, or
whenever H is log-concave. The vast majority of commonly-used distributions dened on
R are either symmetric or log-concave (or both).
The intuition for this condition is straightforward. H( jxj) is the probability that
the conicts prospective winner chooses the peace action, and h( jxj)=H( jxj) is the
percentage decrease in this probability when payo asymmetry jxj increases. Similarly,
h(jxj)=H(jxj) is the percentage increase in the likelihood that the prospective loser will
play the peace action. The Inverse Mills ratio condition simply states that the former
exceeds the latter. Now, because H( jxj) < H(jxj), it will typically be the case that
the proportional increase in the bellicosity of the expected winner exceeds the proportional
increase in the dovishness of the loser, causing an increase in conict. A simple way to
think about this is that the winner is the player responsible for most conicts, so what
happens to this players incentive to engage in conict matters more than what happens
to the other players incentives.
In the next subsection we set up a simple model of conict over natural resources which
ts squarely in this general setup. We will see how the existence and spatial distribution
of natural resources aects payo asymmetry, and hence the likelihood of conict.

10

2.2

Territorial Conict

The world has a linear geography, with space ordered continuously from

1 to +1: In

this world there are two countries, A and B. Country A initially controls the [ 1; 0]
region of the world, while country B controls [0; +1]: In other words the initial border is
normalized to be the origin. Each country has a resource point (say an oil eld) somewhere
in the region that it controls. Hence, the geographic coordinates of the two resource points
are two points on the real line, one negative and one positive. We call these points GA and
GB , respectively. These resource points generate resource ows RA and RB , respectively.
For simplicity the Rs can take only two values, RA ; RB 2 f0; Rg, where R > 0. Without
further loss of generality we normalize R to be equal to 1.9
The two countries play a game with two possible outcomes: war and peace. If a conict
has occurred, there is a new post-conict boundary, Z. Intuitively, if Z > 0 country A has
won the war and occupied a segment Z of country B. If Z < 0 country B has won. The
implicit assumption here is that in a war the winner will appropriate a contiguous region
that begins at the initial border.
We make the following assumptions on the distribution of Z:
Assumption 1 Z is a continuous random variable with domain R, density f , and cumulative distribution function F .
In sum, the innovation of the model is to see war as a random draw of a new border
between two countries: this makes the model suitable for the study of territorial wars. Note
that the distribution f need not be symmetric, much less symmetric around 0. The position
in space of the distribution will depend on the relative strength of the two countries. If
most of the mass point is over the positive real numbers, then a potential war is expected
to result in territorial gains for country A (the more so the more to the rightis the mass
of the distribution), so country A can be said to be stronger. Needless to say, since Z is
dened on R, it is possible for the (expected) weaker country to win.
We assume that each countrys objective function is linearly increasing in the value of

In Section 2.4.2 we allow for arbitrary values of RA and RB .

11

the natural resources located in the territory it controls (at the end of the game). This
means that, ceteris paribus, a country would like to maximize the number of oil elds it
controls.
Besides the oil, there is an additional cost or benet from conict, ci , i = A; B, which
is a catch-all term for all the other considerations that aect a countrys decision to go to
war. As in the previous subsection, we assume that ci , i = A; B is a continuous random
variable dened on R, with density h, cumulative distribution function H, and satisfying
the Inverse Mills Ratio condition h(jcj)=H(jcj) < h( jcj)=H( jcj). This implies that
increases in payo asymmetry increase the likelihood of conict.10
This discussion results in the following payo functions. If the outcome is peace, the
payos are simply RA for country A and RB for country B, as by denition there is no
border change (and hence also no change in property rights over the oil elds). If there
has been a war, the payos are:
UAC = RA I(Z > GA ) + RB I(Z > GB ) + cA ;
UBC = RA I(Z < GA ) + RB I(Z < GB ) + cB ;
where UiC is the payo for country i after a conict, and I( ) is the indicator function.
The rst two terms in each payo function are the oil elds controlled after the war. For
example, country A has hung on its eld if the new border is to the right of it, and
similarly it has conquered Bs oil if the new border is to the right of it. The last term
represents the non-territorial costs or benets from war. Note that implicitly (and for
simplicity) we assume that countries are risk neutral.11;12

10

We discuss relaxing the assumption that the two countries draw the cs from the same distribution in

footnote 18 below.
11
Our payo functions implicitly assume that the value of the oil elds is the same in case of war or
without. It would be trivial to allow for some losses in the value of the oil in case of conict. For example
we could assume that conquered oil only delivers R to the conqueror, with

2 (0; 1]. The qualitative

predictions would be unchanged.

12
In order to use our framework to study other aspects of territorial war, it will typically make sense
to assume that Z enters directly into the payo functions, reecting that countries may care about their
territorial size per se (which in our model is equivalent to the measure of the real line it controls). For

12

The timing and actions of the model are as follows. First, each country i draws a cost
of conict ci , i = A; B. Then each country decides whether or not to declare war, and does
so to maximize expected payos. If at least one country declares war, war ensues. In case
of war, nature draws the new boundary, Z. Then payos are collected.

2.3

Analysis

This game is readily seen to have the structure discussed in Section 2.1. Both countries
prefer peace (conditional on their draw of c) if E(UAC )

RA and E(UBC )

RB (where

the expectation is taken after observing ci ). Given assumption 1 these conditions can be
rewritten as
cB

RA F (GA ) + RB [1

F (GB )]

Hence, the probability of peace is H(x)H( x), where x =

(1)

cA :
RA F (GA ) + RB [1

F (GB )].

Changes in RA , RB , GA , and GB increase the likelihood of conict if they increase jxj.

The expression for x clearly conveys the basic trade-o countries face in deciding
whether to initiate a conict (over and above the trade-os that are already subsumed
in the ci terms): conict is an opportunity to seize the other countrys oil, but also brings
the risk of losing ones own. Crucially, the probabilities of these two events depend on the
location of the oil elds. Consider the decision by country A [second inequality in (1)].
If its own oil is very far from the border (GA , and hence F (GA ), is small) then country
A is relatively unlikely to lose its oil, which makes country A in turn less likely to choose
peace. Similarly, if country Bs oil is nearer the border (GB small, so 1

F (GB ) large),

the prospects of capturing Bs oil improve, and A once again is less likely to opt for peace.
Remark 2. The case where RA = 0 ( RB = 0) is isomorphic to the case where
GA !

1 ( GB ! 1).

For the purposes of evaluating the likelihood of peace, it makes no dierence if one

example, controlling more territory provides more agricultural land to exploit, or more people to tax.
Indeed in a previous version of this paper we added the term +Z ( Z) to the expression for UAC (UBC ).
However this addition complicates the statements of our results, so we have dropped these terms in the
current version to focus on the mechanism we are interested in.

13

country does not have oil, or if its oil is located innitely far from the border. This
observation, which follows directly from inspection of equations (1), simplies slightly the
presentation of the results, as it implies that the cases where only one or neither country
have oil are limiting cases of the case where both countries have oil. In particular, we can
denote the probability of peace as P (GA ; GB ), i.e. simply as a function of the location of
the oil elds. With some slight abuse of notation, we then denote the probability of peace
when only country A has oil (no country has oil) as P (GA ; 1) (P ( 1; 1)).
Proposition 1
(i) P (GA ; 1)

P ( 1; 1);

(ii) @P (GA ; 1)=@GA

(iii) P (GA ; GB )
(iv) P (GA ; 1)

0;

P ( 1; 1);
P (GA ; GB ) if and only if 1

(v) @P (GA ; GB )=@GA

0 if and only if 1

F (GB )
F (GA )

2F (GA );
F (GB )

0.

The proposition, which follows nearly directly from the Inverse MillsRatio condition,
enumerates ve testable implications about how the presence and location of oil aects the
likelihood of conict among two countries. Parts (i), (iii), and (iv) compare the likelihood
of conict when neither, only one, or both countries have oil. Parts (ii) and (v) look at
how the likelihood varies with the location of the oil. In the rest of the section we discuss
what these predictions say and how they come about within the logic of the model.
Part (i) of the proposition establishes that conict is more likely when one country
has oil than when neither country does. Recall that a discovery of oil in one country has
opposite eects on each countrys incentives to go to war. The country which found the
oil becomes less likely to wish to get into a conict because it has more to lose, while the
other country has an additional potential prize from going to war. The proposition says
that (as long as the Inverse MillsRatio of the distribution of c is well behaved) the latter
eect systematically dominates, so the likelihood of conict goes up. The reason is that
the oil discovery in country A creates a payo asymmetry.
Part (ii) says that when oil is only in one country the probability of conict increases
when oil moves closer to the border. The reason is that the movement of the oil towards
the border increases the likelihood that country B (A) will capture (lose) the oil, thus
exacerbating payo asymmetry.
14

Part (iii) tells us that two countries both having oil are more likely to experience
a conict than two countries both not having oil. Oil always makes one country more
aggressive, because with oil payos will always be more asymmetric than without oil, and
this is enough to trigger more conicts. In this sense under our assumptions the mere
presence of oil is always a threat to peace.
Part (iv) compares the situation when both countries have oil to the situation when
only country A has oil. It says that the discovery of oil in the second country will typically
defuse tensions. The intuition of course is that when the second country nds oil payo
asymmetry declines. The country that would typically have been responsible for most
conict becomes less aggressive, as it becomes concerned with the possibility of losing its
newly-found oil. Country A does become more aggressive, but this is typically insu cient
to create a more belligerent atmosphere. The exception is when the oil in country A was
initially much further away from the border than the new oil discovered in country B
which is the meaning of the conditioning statement in part (iv). In this case, the new
discovery in country B can actually increase payo asymmetry: from mildly asymmetric in
favor of B to strongly asymmetric in favour of A. Unconditionally, however, i.e. without
knowledge of the locations of the two countries oil elds, we expect pairs where both
countries have oil to engage in less conict than pairs where only one does.
Finally, part (v) looks at the marginal eect of moving oil towards the border in one
country, while leaving the other countrys oil location unchanged. To better understand
this part, it is useful to look at the following special case.
Corollary
If f is symmetric around 0, then @P (GA ; GB )=@GA

0 if and only if jGA j

GB :

In other words, when both countries have oil, changes in distance that increase the
asymmetry of oil locations tend to increase the asymmetry of payos. Consider starting
from a situation of perfect symmetry, or

GA = GB . When f is symmetric around 0,

the incentive to ght for the other countrys oil exactly cancels out with the deterrent
eect from fear of losing ones oil (see equations (1)). However, as soon as we break this
symmetry, say by moving As oil towards the border, country B becomes an expected
winner. The conditioning statement in the proposition generalizes this intuition, as F (GA )

15

will tend to be larger than 1

F (GB ) when GA is closer to border than GB :13

The empirical part of the paper tests predictions (i)-(v).

2.4
2.4.1

Discussion and Extensions

Conict and border changes

The key modelling choice we have made is to think of international wars as potentially
border-changing events. The long (and very incomplete) list of examples of territorial
wars and militarized border disputes in the Introduction supports this assumption. The
International Relations literature provides further systematic evidence. Kocs (1995) has
found that between 1945 and 1987 86% of all full-blown international wars were between
neighboring states, and that in 72% of wars between contiguous states unresolved disputes
over territory in the border area have been crucial drivers. The unstable nature of borders
is well recognized. According to Anderson (1999) about a quarter of land borders and some
two-thirds of maritime borders are unstable or need to be settled. Tir et al. (1998) identify,
following restrictive criteria, 817 territorial changes between 1816 and 1996, many of which
are the result of international conicts. According to Tir et al. (1998) and Tir (2003) 27%
of all territorial changes between 1816 and 1996 involve full-blown military conict, and
47% of territorial transfers involve some level of violence. Weede (1973: 87) concludes that
"the history of war and peace is largely identical with the history of territorial changes as
results of war."
The data described in the next section also supports the existing evidence. In our
panel of country pairs 0.4% of all observations feature border changes (corresponding to
90 cases of border change). Yet, conditional on the two countries being in conict with
each other, the incidence of border changes goes up to 7.4%. In other words the probability
of a border change increases 19-fold in case of war.14 In Appendix C we show that conict

13

However in the case where f is not symmetric jGA j

GB is not su cient for movements away from

symmetry to generate more conict. The prediction could be overturned if the country whose oil is moving
towards the border is much stronger militarily (i.e. if f is very skewed in its favor).
14
Conversely, while only 6% of observed country pairs are in conict, 30% of country pairs experiencing

16

remains a signicant predictor of border changes after controlling for time and country
xed eects. Indeed we go further and show that the presence and location of oil elds has
some predictive power for border changes, despite the very infrequent occurrence of such
changes.
Having said that, it is also important to stress that the model emphatically does not
predict that all conicts will be associated with border changes. All of our results and
calculations allow for the distribution of Z to have a mass point at 0. Indeed, a signicant
mass point at 0 appears likely in light of the gures above.
It is also important to point out that, strictly speaking, the distribution function f need
not be the true distribution of post-conict border locations Z. f is the distribution used
by the decision-makers in the two countries, but this need not be the rational-expectations
distribution. Anecdotal observation suggests that overoptimism is often a factor in war
and peace decisions, so our guess is that the objective numbers cited above are probably
lower bounds on the probabilities assigned by leaders to their chances of moving the border
in case of war. For example, it seems likely that Saddam Hussein overstated his chances
of permanently shifting the borders of Iraq with Iran (rst) and Kuwait (later).
2.4.2

Allowing for Variation in R

With our assumption that R 2 f0; 1g we have normalized all non-zero oil endowments.
It is trivial to relax this assumption to look at the eects of changes in RA and RB . In
particular, as implied by our Remark above, an increase in RA has identical qualitative
eects of a movement of As oil towards the border, while an increase in RB is akin to a
move of Bs oil towards the border. Our propositions can therefore readily be reinterpreted
in terms of changes in quantities.
Unfortunately, testing these predictions would require data on oil eld-level endowments
that we have no access to. Potentially, predictions for changes in the Rs might be tested
using variation in oil prices, as an oil price increase is an equiproportional increase in
both RA and RB . For example, for the case where only one country has oil, our theory

a border change are in conict.

17

would predict that increases in oil prices tend to lead to an increase in the likelihood of
conict. However, ample anecdotal evidence suggests that short-term oil prices are very
responsive to conicts involving oil-producing countries, so it would be very di cult to
sort out a credible causal path from oil prices to conict.15 Another issue is that what
matters for war should be the long-term oil price: it is not clear that current oil prices are
good forecasts of long-run ones.
2.4.3

Endogenous F

Oil as a source of military strength In our model the discovery of oil in one country
tends to make this country less aggressive, as it fears losing the oil, and the other more
aggressive, as it wishes to capture it. We may call this a greed eect. However, the
discovery of oil may also provide the discoverer with nancial resources that allow it to
build stronger military capabilities. If oil rich countries are militarily stronger, they might
also be more aggressive as the odds of victory go up. Their neighbors may also be more
easily deterred. Hence, there is a potential strength eect that goes in the opposite
direction to the greedeect.16
However, while the fact of having oil may have some ambiguous implications through
the opposing strengthand greedeects, the geographical location of the oil should only

15

Even interacting oil distance from the border with the World oil price would be di cult to interpret,

as market participantsassessment of the disruption caused by a war to oil supplies might depend on the
distance of the oil from the border. In particular, when the oil is close to the border the ghting is more
likely to disrupt oil production and shipment.
16
The strength eect could easily be added to our model by making F a decreasing function of RA
and an increasing function of RB . However, this would not be enough to fully bring out the ambiguity
discussed in the text. For example, it is easy to see that parts (i)-(iii) of our Proposition would still go
through exactly unchanged. Hence, it would still be the case that, e.g., discovery of oil in one country
unambiguously leads to greater likelihood of conict. This is because in our model the only territorial
benet of conict is oil merely being stronger does not make country A more aggressive. In footnote 12
we have alluded to a previous version of the model where countries have territorial aspirations over and
above the control of oil (i.e. Z enters the payo function). In that model the strength eect is present
and the empirical predictions are correspondingly a bit more ambiguous.

18

matter through greed. Oil will increase resources for ghting irrespective of its location, but
the risk of losing it will be more severe if the oil is near the border. Hence, our predictions
concerning the eect of oil location on conict which are the focus for our most distinctive
empirical results should be unaected by the strength argument. As mentioned in the
Introduction, this is one key reason to focus on the geographic distribution of the oil in
the empirical work.
In any case, while we dont model the strengtheect explicitly, in our empirical work
we are able to fully control for it, by including various measures of each countrys aggregate
oil endowments.17
Other sources of asymmetric strength Having oil endowments is just one reason
why one country may be militarily stronger than another. For example, a larger country,
a richer country, or a more ethnically homogeneous country could also be expected to
be stronger. The same mechanisms that may lead the strength eect to qualify the
predictions of the greed eect are thus involved in thinking about these other reasons
for military asymmetry, and lead to similar qualications. For example, if it is the militarily
stronger country which nds oil, it is no longer necessarily the case that payos become
more asymmetric.
Endogenous arming decisions In the discussion so far, we have assumed that the two
countries take their relative strength, represented by the function F , as given. As we show
in the Online Appendix, a similar line of reasoning applies if each country can make military
investments to improve its odds of success in case of conict. Consider, for example, an
increase in GA when country B does not have oil. In the baseline model, this has only a
direct positive eect on country Bs chances of capturing the oil, and unambiguously leads
to more payo asymmetry and hence more conict. In the extended model, however, the

17

Note that while we do not have data on oil-eld-level oil endowments, we do have data on country-level

endowments. The former would be required to test the comparative statics of the model with respect to
RA and RB , i.e. the eect of endowments through the greed eect. The latter are su cient to test for
the strengtheect, which depends only on aggregate endowments, and not on their spatial distribution.

19

shift of the oil towards the origin can cause the two countries to change their armaments.
If country A responds by arming much more than B, it is conceivable that this indirect
eect will dominate over the direct eect, resulting in a decline of country Bs prospects
of capturing country As oil. This may reduce payo asymmetry, and hence, unlike in
the baseline model, an increase in GA no longer unambiguously increases conict. Having
said this, the scenario where the likelihood of conict declines seems very implausible.
Specically, it is not at all clear why A should respond to the increase in GA with a greater
arming eort than B, much less that the disparity in response should be so large as to
more than negate the direct eect.18

Empirical Implementation

3.1

Data and Empirical Strategy

3.1.1

Sample

We work with a panel dataset, where an observation corresponds to a country pair in a

given year, e.g. Sudan-Chad in 1990. Country pairs are included if they satisfy Stinnett
et al. (2002)s direct contiguitycriterion: the two countries must either share a land (or
river) border, or be separated by no more than 400 miles of water. There are 606 pairs of
countries satisfying this criterion.19 The dataset covers the years 1946-2008.
All variables are described in detail in Appendix A, which also contains Table 6 with
summary descriptive statistics. Here we focus on the key dependent variable and the
independent variables of interest.

18

Much the same style of argument applies if we relax the assumption that cA and cB are drawn from

identical distributions. For example, a natural extension would be to assume that cA (cB ) is positively
(negatively) related to the mean of the distribution of Z, i.e. that the country that expects the largest
territorial gains also expects the largest non-territorial ones (or to pay a less devastating non-territorial
cost for the conict). For example, if the oil is found (or moves towards the border) in the country with
higher mean c i.e. the country responsible for most conicts the deterrence eect on the country with
oil may dominate over the greed eect experienced by the country without oil.
19
Approximately 60% of the country pairs in the sample are separated by a land or river border.

20

3.1.2

Dependent Variables

Our main dependent variable is a measure of inter-state disputes, from the "Dyadic Militarized Interstate Disputes" (MID) data set of Maoz (2005). The MID data is the most
widely used data on interstate hostilities.20 Compared to alternative (and less widely used)
data sets such as the UCDP/PRIO Armed Conict Dataset (Uppsala Conict Data Program, 2011) it has the advantage of not only including the very rare full-blown wars
between states, but also smaller scale conicts, and to provide a relatively precise scale of
conict intensity.
In Maoz (2005) interstate disputes are reported on a 0-5 scale. The highest value, 5, is
reserved for sustained combat, involving organized armed forces, resulting in a minimum
of 1,000 battle-related combatant fatalities within a twelve month period.This extremely
violent form of confrontation, which we will refer to as War, is rare: only 0.4% of our
observations meet this criterion. The next highest value, 4, is for Blockade, Occupation
of territory, Seizure, Attack, Clash, Declaration of war, or Use of Chemical, Biological, or
Radioactive weapons.While still very violent, this type of confrontation, which is labelled
Use of Force,is much more frequent, occurring in as many as 5.2% of our observations.
Accordingly, we construct our main dependent variable, which we call Hostilityby combining all episodes of War and Use of Force.21 We also present robustness checks using
War only,22 including disputes receiving a value of 3 in Maoz (2005) - Hostility+,23 and
even specications relating the intensity of conict to the presence and geographic location

20

MID data, as well as the Stinnett et al.s contiguity variable, are accessible through the Correlates

of War project. Related papers in economics using this data include for example Martin et al. (2008),
Besley and Persson (2009), Glick and Taylor (2010), Baliga et al. (2011) and Conconi et al. (2012).
21
It is standard practice in the empirical literature on international conict to aggregate over more than
one of the Maoz (2005) categories. For example, Martin et al. (2008) and Conconi et al. (2012) code a
country pair to experience conict when hostility levels 3, 4 or 5 are reached.
22
The dataset from Maoz (2005) only runs until 2001. As alternative data on full-blown wars is readily
available, when we check the results using "War" we update this variable using the UCDP/PRIO Armed
Conict Dataset (Uppsala Conict Data Program, 2011).
23
Disputes receive a mark of 3 when they meet the criterion of "Display use of force", which is reserved
for "Show of force, Alert, Nuclear alert, Mobilization, Fortify border, Border violation".

21

of oil.
An alternative approach is to investigate data which identies the aggressor in a bilateral conict (as in Colgan (2010) and De Soysa et al. (2011)). However, in many cases,
identication of the aggressor requires subjective and possibly unreliable judgments. Furthermore, if a country perceives a potential threat, it may choose to attack rst, and it is
not clear that data focusing on the direction of attack are always able to account for such
preemptive strikes.24 We submit that our approach based on distance of the oil from the
border oers a more robust strategy. Having said this, in Appendix B we use additional
data from Maoz (2005) to look at how the presence and the distance of the oil from the
border dierentially aect the likelihood that the oil rich or the oil poor country is classied
as "revisionist", "attacker" or "initiator of conict".
3.1.3

Explanatory Variables of Interest

Our main independent variables are one-period lagged measures of the presence and distance of oil elds in each country in the pair from the bilateral border or from the other
countrys coastline. To construct these we have combined two sources. The rst source is
the CShapes dataset of Weidmann et al. (2010), which contains historically accurate georeferenced borders for every country and year. The dataset accounts for border changes
over time, both the ones originating from state creation and split-ups, and those arising
from border adjustments. Their border adjustment information is based on Tir et al.
(1998).
The second source is a time varying and geo-referenced dataset on the location of oil
and gas elds from Lujala et al. (2007, PETRODATA). It includes the geographic coordinates of hydrocarbon reserves and is specically designed for being used with geographic
information systems (GIS). In total, PETRODATA consists of 884 records for onshore and
378 records for oshore elds in 114 countries. Note that PETRODATA includes all oil
and gas elds known to exist, including those not yet under production, which is clearly

24

See, e.g., Gaubatz (1991), Gowa (1999), Potter (2007), and Conconi et al. (2012), for more detailed

versions of these and other criticisms of the direct dyad approach.

22

appropriate given that incentives to appropriate will likely be similar for operating and
not-yet-operating elds.25
Using Geographical Information System (GIS) software, we merge these two data sets
so that we can pinpoint each oil eld position vis-a-vis a countrys borders as well as visa-vis the coastline of neighboring countries. Then, for each country pair and for each oil
eld belonging to one of the two countries, we measure the oil elds minimum distance to
the other countrys land border, as well as the minimum distance to the coastline of the
other country. The oil elds distance to the other country is then the minimum of these
two.26 The minimum oil distance from the other country is the minimum across all oil
eldsminimum distances.
On the basis of these data, we have constructed the following ve explanatory variables.
"One" is a dummy variable taking the value of 1 when only one country in the pair has
oil. Similarly, "Both" takes a value of 1 if both countries of the pair have oil. The omitted
baseline category hence is the case where none of the countries in the pair has oil. "One
x Dist" is the product of the One dummy with the distance of the oil from the border.
Similarly, "Both x MinDist" is the product of the Bothdummy and the minimum of the
distances of the oil from the border in the two countries. Analogously, "Both x MaxDist"
captures the distance from the border in the country whose oil is further from the border.
Note that an increase in "Both x MinDist" (holding "Both x MaxDist" constant) is a
movement towards symmetry, while an increase in "Both x MaxDist" (holding "Both x
MinDist" constant) is a movement away from symmetry.27

25

The main data sources of PETRODATA include World Petroleum Assessment by U.S. Geological

Survey (USGS, 2000), Digital database on Giant Fields of the World by Earth Sciences and Resources
Institute at the University of South Carolina (ESRI-USC, 1996), and World Energy Atlas by Petroleum
Economist (Petroleum Economist, 2003).
26
Needless to say in many cases there is no land border and in many others there is no coastline, so in
these cases the distance variable is just the distance from the coastline (border).
27
The attentive reader will have noticed that, in constructing our key dependent variables, we have
taken the min operator three times: rst, for each oil eld in a country, between its distance to the other
countrys border and the other countrys coastline (distance of oil eld to other country); second, for each
country, among all its oil eldsdistances to the other country (minimum oil distance to other country);

23

In our main specications, all the distance variables are normalized to lie between 0
and 1, to reduce their range, and constructed so that there are diminishing marginal
costs from geographical distance. In particular, the functional form is 1

, where

d is the crude geographical distance in hundreds of Kms. The idea for the diminishing
costs is that conquering the rst Km in the enemys territory may be a more momentous
decision than conquering the 601st Km when one has already captured the rst 600. In our
benchmark specication we set

= 1, which is equivalent to assuming that diminishing

marginal costs set in fast, consistent with our intuition. We present robustness checks for
alternative functional forms (including unscaled distance).
3.1.4

Control Variables

In all regressions we control for the average and the absolute dierence of land areas in the
pair. We also present specications which further include the average and absolute dierence of GDP per capita, the average and absolute dierence of population, the average and
absolute dierence of ghting capabilities, the average and absolute dierence of democracy scores, the number of consecutive years the two countries have been at peace before
the current period, the volume of bilateral trade (scaled by the sum of GDPs), a measure
of genetic distance between the populations of the two countries, a dummy for membership in the same defensive alliance, a dummy for historical inclusion in the same country,
kingdom or empire, a dummy for having been in a colonial relationship, two dummies for
civil war incidence in one or both of the countries in the pair, and two dummies for OPEC
membership of one or both countries in the pair.28 Finally, in important robustness checks

and, third, between the two countriesminimum oil distances (MinDist). MaxDist is the max between the
two countriesminimum oil distances.
28
Population and GDP could aect the likelihood of conict in myriad ways, e.g. through relative
military strength; ghting capabilities directly aect the chances of success, so clearly enter the calculation
of whether to engage in conict; democracy scores are included to account for the democratic peace
phenomenon (Maoz and Russett, 1993); joint membership in alliances or in OPEC, or historical roots in
the same kingdom or empire, may oer countries venues to facilitate the peaceful resolution of conicts;
previous history of conict is meant to absorb unobserved persistent factors leading to conict between
the two countries; recent history of domestic civil wars captures one factor that may weaken one country

24

we discuss later, we further include several measures of the amounts of oil production and
reserves in the two countries. Again, all variables are discussed in detail in Appendix A.
3.1.5

Specication and Methods

Our benchmark specication is a linear-probability model that takes the form

HOSTILITYd;t+1 =

Onedt + (One x Dist)dt

+ Bothdt + (Both x MinDist)dt + ! (Both x MaxDist)dt

+X0dt + udt ;
where d indexes country pairs, t indexes time, and X is the vector of afore-mentioned
controls. We consider alternative functional forms (including probit and logit) in robustness
checks.
Crucially, our preferred specication for the error term udt includes a full set of country dummies as well as a full set of time dummies. This implies that the key source of
identication for, say, the eect of One is the relative propensity of a given country to
experience conict with its oil-rich neighbors and with its oil-poor neighbors. We will nd a
positive estimate of

when the same country has more conicted relations with its oil-rich

neighbors. The identication of the other coe cients is driven by similar within-country
comparisons, e.g. is a given country more likely to nd itself in conict with a neighbor
whose oil is near the border than with one whose oil is far away from the border.
The inclusion of country xed eects serves to limit the inuence of unobservable determinants of a countrys proneness to engage in conict that may be spuriously correlated
with its having oil (or with its having many neighbors having oil). Clearly the theory also
predicts that countries with oil (or with oil-rich neighbors) should engage in conict more
frequently, a prediction that is not allowed to inuence the results when xed eects are
included. If we could be condent that our control variables fully absorbed all the other
determinants of bilateral conict, specications that omit xed eects might be preferable

and tempt the other to take advantage; bilateral trade has been found to matter for bilateral conict by
Martin et al. (2008), and so has genetic distance [Spolaore and Wacziarg (2013)].

25

(and we report them). But it is unlikely that the controls fully account for all the determinants of conict that may be spuriously correlated with having oil (or neighbors with
oil). This is particularly true in the present context: the dyadic specication means that
all controls must be dened at the country-pair level, not at the country level. Hence, it
is likely that the controls are quite ineective at accounting for country-level covariates of
oil endowments and conict outcomes.
For the same reason, the country xed eects are insu cient to absorb the inuence
of factors that aect the likelihood of conict at the country-pair level, leading us to
prefer specications that include controls. The downside of including the controls is, of
course, that some of them may be endogenous to bilateral conict. This is why we present
specications with and without controls - except for the controls based on surface area
which seem clearly exogenous.
Ideally one would use country-pair xed eects, and base all inference on time-series
variation in the variables of interest. Identication of the oil-related coe cients would
then be driven by (i) oil discoveries, which switch the dummies Oneand Bothfrom 0
to 1, and potentially change the distance measures (if the newly-discovered eld is closer
to the border than all the pre-existing elds); and (ii) changes in borders. Unfortunately
there are too few (relevant) oil discoveries and border changes in our dataset to provide
su cient power for identication.29
In all regressions reported in then main body of the paper the standard errors (which
are always displayed in parenthesis below the coe cient) are clustered at the country-pair
level. In the Online Appendix we further report standard errors based on methods that
attempt to implement clustering at the country level. The Online Appendix also reports
results that control for region-year xed eects, and also presents ndings for a simple

29

For each country pair and each of the ve variables of interest we have calculated the number of

changes during the sample period. Across the ve variables, the fraction of country pairs experiencing no
change whatsoever varies between 75 and 85%, and the fraction experiencing no more than one change
is between 95 and 100%. Unsurprisingly, this time variation is too small to yield sharp estimations once
country-pair xed eects are included. With the full set of controls, the estimated coe cients (standard
errors) are

=0.018 (0.029),

=-0.019 (0.032),

=0.049 (0.044),

26

=0.021 (0.042), and ! =-0.078 (0.046).

cross-section of country pairs.

3.2

Results

Table 1 presents the baseline regressions for the main dependent variable, Hostility. In the
rst four columns we use all oil elds to construct our measures of oil endowments and
distance, while in columns 5-8 we only use oshore oil, and in columns 9-12 only onshore
oil. In column 1 we show the coe cients on our variables of interest only after controlling
for annual time dummies and average and absolute dierence in land areas. In column 2
(3) we add all the other controls (country xed eects), and in column 4 we present our
preferred specication with both controls and xed eects. The estimates are reasonably
stable across the four specications, though statistical signicance tends to improve as we
add country xed eects and the further controls.
When we include the full set of country xed eects and controls (column 4), both the
presence and geographic location of the oil are statistically signicant predictors of bilateral
conict. As predicted by the model a country pair with one or both countries having oil is
signicantly more prone to inter-state disputes than a pair with no oil whatsoever (which is
the omitted category). More importantly, when only one country has oil, the likelihood of
conict signicantly drops when the oil is further away from the border. Similarly, when
both countries have oil, the likelihood of conict is decreasing in the distance from the
border of the oil that is closest to the border - a movement towards symmetry. The only
prediction of the model for which the support is weak concerns the distance of the furthest
oil eld: while the sign of the coe cient is positive, as predicted, it is not statistically
signicant.30
Quantitatively, the eect of geographic location is very sizeable. Figure 2 shows the

30

Generally speaking, in most specications we have run the two versions with xed eects (with and

without controls for country-pair characteristics) generate quite similar results. However, the xed eects
are important. For example, in the specications with controls but without country xed eects the
coe cients on the key distance variables drop by one third to one half, and lose signicance in about
50% of cases. This suggests that there are unobservable country-level determinants of conict that are
spuriously positively correlated with oil distance.

27

probability of conict implied by the regression coe cients in Column 4 as a function of

the oils distance from the border (when all the controls are set at their average values). As
already noted, the average risk of conict in our sample is 5.7 percent. This drops to 3.1
percent for country pairs in which neither country has oil. In contrast, when one country
in the pair has oil, and this oil is right at the border (Distance = 0), the probability of
conict is about 3.5 times as large: 10.8 percent. But this greater likelihood of conict
is very sensitive to distance. Indeed when the oil is located at the maximum theoretical
value for our distance measure (Distance = 1) the likelihood of conict is similar to the
likelihood when neither country has oil.31 The last two bars in the gure look at the case
where both countries have oil. In the rst instance, asymmetry is maximal: one country has
oil right at the border (MinDist=0), the other at the maximum distance (MaxDist=1). The
likelihood of conict is over 2.5 times as large as in the case where neither country has oil,
or 8 percent. In the second instance, we look at a case of perfect symmetry: both countries
have oil at a distance that is one half of the maximum distance (MinDist=MaxDist=0.5).
The likelihood of conict is a much more modest 3.4 percent.32
The remaining columns in the table investigate whether the results are driven particularly by oshore or onshore oil. In particular, in columns 5-8 we construct our variables of
interest using exclusively information on oshore oil elds (so, e.g., if in a country pair all
the oil is onshore the pair is treated as a no oil pair), and then repeat the four specications with no controls, only country-pair controls, only country dummies, and all xed
eects and controls. In columns 9-12 we do the same for onshore oil. It turns out that
the coe cients on the variables of interest and patterns of signicance are quite similar for
oshore and onshore oil (and thus to the baseline case). Hence, if the mechanism driving
the results is the one implied by our theory, it seems that having another countrys oil near
ones coastline is as temptingas having it near ones border.

31

Hence, our models formal isomorphism between the cases of no oil and innitely distant oil seems

to hold empirically.
32
In constructing Figure 2 we have used the coe cient on the interaction term between the oil in two
countriesdummy and the maximum distance variable even though it is statistically insignicant. Because
it is a very small number, however, using 0 instead has only a minor eect on the quantities in the table.

28

29

(2)
0.029
(0.027)
-0.044
(0.027)
0.028
(0.020)
-0.044
(0.035)
-0.014
(0.036)
All
No
Yes
11303
0.090

(3)
0.049*
(0.027)
-0.073***
(0.026)
0.034
(0.029)
-0.105***
(0.030)
0.016
(0.030)
All
Yes
No
19962
0.145

(4)
0.077**
(0.030)
-0.086***
(0.027)
0.045*
(0.027)
-0.089***
(0.029)
0.004
(0.029)
All
Yes
Yes
11303
0.158

(5)
0.087
(0.054)
-0.107*
(0.056)
0.023
(0.030)
-0.088*
(0.047)
0.048
(0.065)
Offshore
No
No
19962
0.020

Dependent variable: Hostility

(6)
(7)
(8)
0.081**
0.143*** 0.124***
(0.040)
(0.048)
(0.035)
-0.087** -0.138*** -0.118***
(0.043)
(0.048)
(0.034)
0.048*
0.110***
0.073**
(0.027)
(0.035)
(0.030)
-0.028
-0.107**
-0.066**
(0.035)
(0.051)
(0.032)
-0.048
0.012
-0.010
(0.050)
(0.065)
(0.044)
Offshore
Offshore
Offshore
No
Yes
Yes
Yes
No
Yes
11303
19962
11303
0.091
0.145
0.156
(9)
0.048
(0.042)
-0.079*
(0.044)
0.009
(0.024)
-0.102***
(0.038)
0.059
(0.043)
Onshore
No
No
19962
0.021

(10)
0.063*
(0.038)
-0.072*
(0.039)
0.024
(0.020)
-0.096**
(0.043)
0.042
(0.045)
Onshore
No
Yes
11303
0.092

(11)
0.058*
(0.033)
-0.103***
(0.033)
0.020
(0.032)
-0.122***
(0.032)
0.047
(0.034)
Onshore
Yes
No
19962
0.146

(12)
0.132***
(0.044)
-0.137***
(0.041)
0.047
(0.031)
-0.125***
(0.036)
0.043
(0.038)
Onshore
Yes
Yes
11303
0.160

Table 1: Baseline results for Hostility

Note: The unit of observation is a country pair in a given year. The sample covers all contiguous country pairs and the years 1946-2001. Method: OLS with robust
standard errors clustered at the country-pair level. Significance levels *** p<0.01, ** p<0.05, * p<0.1. All explanatory variables are taken as first lag. All specifications
control for the average and the absolute difference of land areas in the pair, intercept and annual time dummies. Additional controls are: The average and absolute
difference of GDP per capita, the average and absolute difference of population, the average and absolute difference of fighting capabilities, the average and absolute
difference of democracy scores, dummy for one country having civil war, dummy for both countries having civil war, bilateral trade / GDP, dummy for one country being
OPEC member, dummy for both countries being OPEC member, genetic distance between the populations of the two countries, dummy for membership in the same
defensive alliance, dummy for historical inclusion in the same country, kingdom or empire, dummy for having been in a colonial relationship, and years since the last
hostility in the country pair.

Type Oil
Country FE
Add. Controls
Observations
R-squared

Both x MaxDist

Both x MinDist

Both

One x Dist

One

(1)
0.034
(0.032)
-0.050
(0.035)
0.022
(0.021)
-0.077**
(0.035)
0.026
(0.040)
All
No
No
19962
0.019

Figure 2: Quantitative Eects

30

31

Dep. var.: War

(2)
(3)
0.003
0.008
(0.005)
(0.010)
-0.002
-0.010
(0.005)
(0.008)
0.010**
0.009
(0.005)
(0.009)
-0.008
-0.008*
(0.008)
(0.005)
-0.003
-0.005
(0.008)
(0.006)
OLS
OLS
No
Yes
Yes
No
11303
23768
0.033
0.073
n/a
n/a
(4)
0.016**
(0.007)
-0.010**
(0.005)
0.020**
(0.009)
-0.007
(0.006)
-0.006
(0.008)
OLS
Yes
Yes
11303
0.103
n/a

(5)
0.036
(0.033)
-0.059*
(0.035)
0.033
(0.024)
-0.092**
(0.041)
0.025
(0.047)
OLS
No
No
19962
0.024
n/a

Dep. var.: Hostility+

(6)
(7)
0.026
0.051*
(0.027)
(0.030)
-0.046*
-0.083***
(0.027)
(0.027)
0.035
0.037
(0.026)
(0.033)
-0.051
-0.131***
(0.039)
(0.035)
-0.018
0.021
(0.041)
(0.035)
OLS
OLS
No
Yes
Yes
No
11303
19962
0.105
0.155
n/a
n/a
(8)
0.082***
(0.032)
-0.094***
(0.028)
0.055*
(0.033)
-0.111***
(0.033)
0.005
(0.031)
OLS
Yes
Yes
11303
0.179
n/a

(9)
0.480
(0.382)
-0.804*
(0.429)
0.384
(0.278)
-1.050***
(0.404)
0.254
(0.408)
Poisson
No
No
19962
n/a
-16589.39

Dep. var.: Dispute intensity

(10)
(11)
0.454
0.530
(0.332)
(0.337)
-0.793** -1.022***
(0.321)
(0.292)
0.263
0.478
(0.272)
(0.363)
-1.039*** -1.724***
(0.345)
(0.395)
0.158
0.143
(0.348)
(0.348)
Poisson
Poisson
No
Yes
Yes
No
11303
19962
n/a
n/a
-6058.98 -12419.69

(12)
1.768**
(0.789)
-2.134***
(0.624)
0.503
(0.506)
-1.575***
(0.438)
0.241
(0.441)
Poisson
Yes
Yes
11303
n/a
-5217.07

Table 2: Baseline results for alternative dependent variables

Note: The unit of observation is a country pair in a given year. The sample covers all contiguous country pairs and the years 1946-2008. Robust standard errors clustered at the
country-pair level. Significance levels *** p<0.01, ** p<0.05, * p<0.1. The oil variables are constructed using all oil fields (onshore and offshore). All explanatory variables are taken
as first lag. All specifications control for the average and the absolute difference of land areas in the pair, intercept and annual time dummies. Additional controls are: The average
and absolute difference of GDP per capita, the average and absolute difference of population, the average and absolute difference of fighting capabilities, the average and absolute
difference of democracy scores, dummy for one country having civil war, dummy for both countries having civil war, bilateral trade / GDP, dummy for one country being OPEC
member, dummy for both countries being OPEC member, genetic distance between the populations of the two countries, dummy for membership in the same defensive alliance,
dummy for historical inclusion in the same country, kingdom or empire, dummy for having been in a colonial relationship, and years since the last hostility in the country pair.

Method
Country FE
Add. Controls
Observations
R-squared
Log pseudolikelihood

Both x MaxDist

Both x MinDist

Both

One x Dist

One

(1)
0.005
(0.008)
-0.007
(0.008)
0.004
(0.006)
-0.003
(0.005)
-0.004
(0.007)
OLS
No
No
23768
0.005
n/a

3.2.1

Robustness

Alternative dependent variables Table 2 presents results from our benchmark specications, but using alternative measures of conict. Given the similarity between the
oshore and onshore results, and to save space, we focus on exercises that treat oshore
and onshore oil equally, as in the rst four columns of Table 1. In columns (1)-(4) we use
the most stringent denition of conict, namely "War." Because of the very infrequent occurrence of "War" (sample mean 0.004), these regressions have much less statistical power
than those using Hostility, and some of our variables of interest accordingly lose statistical signicance. Nevertheless, perhaps surprisingly, the coe cients on the One and Two
dummies, as well as the coe cient on Distance, remain signicant. Quantitatively, the
coe cients are smaller than when using Hostility, though the impact of distance on War
is still economically very sizable.33
In a similar spirit, columns (5)-(8) present results using a denition of conict broader
than Hostility, namely including conicts classied as having intensity 3 in the Maoz data
set. The results are very similar to the ones using our baseline Hostility measure, with
the coe cients of most of our key variables being sizeable and highly signicant for our
preferred specication of column (8).
Finally, in columns (9)-(12) we further exploit Moazs ner-grained classication of
conicts on a 0-5 scale, by running Poisson Maximum Likelihood specications for conict
intensity. We nd that our oil-distance variables exert an economically and statistically
signicant eect on the intensity of conict, much as they do on conict occurrence.
In Appendix B we estimate further regressions in the same spirit to predict whether
country A is classied as more "revisionist", "attacker" or "initiator of conict" than
country B. We nd robust evidence that oil-rich countries are less likely to be classied
in any of the above categories, and that this eect becomes stronger as the oil gets closer
to the border. There is also fairly strong evidence that countries are more likely to be

33

For example, while in an average country pair the risk of war is 0.4% per year, this risk goes up to

2.4% which is 6 times higher in the most dangerous conguration where only one country in the pair
has oil and this is located right at the border.

32

classied as revisionist towards neighbors that have oil, the more so the closer the oil is to
the border.
Alternative distance scales, functional forms, subsamples To further assess the
robustness of our results, Table 3 presents variants of our preferred specication of column
4 of Table 1. In column 1 we re-scale the distance of oil elds from the border using a
plain natural log function, while in column 2 we use raw oil distance. The results are very
similar to the ones of the benchmark regression. In columns 3, 4 and 5 we replace our
linear probability model with, respectively, logit, probit and rare events logit (ReLogit)
estimators.34 The results are again very similar to our benchmark.35 To further reduce
unobserved heterogeneity, in column 6 we restrict the sample to country pairs where one
or both of the countries have oil (hence dropping all country pairs without any oil). The
results of the benchmark continue to hold in this restricted sample. In column 7 we show
that our results are robust to dropping country pairs including Israel, a country that has
been involved in frequent conict in an oil-rich region of the world (but not necessarily
because of oil). Finally, column 8 shows that our results are also robust to dropping country
pairs with oil elds that straddle the border (i.e. for which MinDist=0). This indicates that
the ndings are not driven by hostilities arising from di culties in managing common-pool
resources.

34

The rare events logit (ReLogit) estimator is from Tomz et al. (2003), and adjusts the estimation for

the fact that the dependent variable takes much more often a value of 0 than of 1. The ReLogit estimator
is not designed for the inclusion of xed eects and for robust standard errors. Hence, we remove all xed
eects and use standard errors without the robust option, but still cluster at the country-pair level.
35
Note that the sample size drops in columns 3 and 4 with the logit, resp. probit estimators as countries
with no variation in the dependent variable (i.e. countries being in all periods in peace with all their
neighbors) drop from the sample when country xed eects are included.

33

34

(1)
0.050
(0.034)
-0.004*
(0.002)
0.049
(0.031)
-0.005***
(0.002)
-0.000
(0.002)

(2)
0.048***
(0.018)
-0.008***
(0.002)
0.047*
(0.024)
-0.011**
(0.006)
-0.005*
(0.003)

(3)
2.603***
(0.992)
-3.063***
(0.829)
0.772
(0.623)
-2.050***
(0.497)
0.312
(0.500)
-0.090***
(0.028)
-0.032
(0.029)
-0.092***
(0.030)
0.002
(0.030)

(6)

(7)
0.066**
(0.028)
-0.077***
(0.026)
0.043*
(0.026)
-0.050**
(0.023)
-0.026
(0.026)

(8)
0.077**
(0.030)
-0.086***
(0.027)
0.046*
(0.027)
-0.088***
(0.029)
0.002
(0.029)

Table 3: Robustness with respect to Estimator and Sample

Only pairs
Sample
All
All
All
All
All
with oil
w/o Israel
w/o Dist. 0
Method
OLS
OLS
Logit
Probit
ReLogit
OLS
OLS
OLS
Country FE and TE
Yes
Yes
Yes
Yes
No
Yes
Yes
Yes
Scale distances
Nat.log.
in 100 km
Standard
Standard
Standard
Standard
Standard
Standard
Observations
11303
11303
8840
8840
11303
9839
11158
11294
R-squared
0.154
0.161
0.319
0.310
0.232
0.171
0.155
0.158
Note: The unit of observation is a country pair in a given year. The sample covers all contiguous country pairs and the years 1946-2001. In all columns
except (5) robust standard errors clustered at the country pair level; in column (5) robust non-clustered standard errors (ReLogit does not allow for
clustering). Significance levels *** p<0.01, ** p<0.05, * p<0.1. The oil variables are constructed using all oil fields (onshore and offshore). All
explanatory variables are taken as first lag. All specifications control for intercept, the average and the absolute difference of land areas in the pair,
the average and absolute difference of GDP per capita, the average and absolute difference of population, the average and absolute difference of
fighting capabilities, the average and absolute difference of democracy scores, dummy for one country having civil war, dummy for both countries
having civil war, bilateral trade / GDP, dummy for one country being OPEC member, dummy for both countries being OPEC member, genetic distance
between the populations of the two countries, dummy for membership in the same defensive alliance, dummy for historical inclusion in the same
country, kingdom or empire, dummy for having been in a colonial relationship, and years since the last hostility in the country pair. All columns, with
the exception of (5), also include country fixed effects and annual time dummies (ReLogit does not allow for FE and TE).

Both x MaxDist

Both x MinDist

Both

One x Dist

One

Dependent variable: Hostility

(4)
(5)
1.089**
0.886**
(0.448)
(0.450)
-1.362***
-1.121**
(0.386)
(0.460)
0.285
0.359
(0.291)
(0.311)
-1.084***
-1.341***
(0.244)
(0.485)
0.175
0.241
(0.258)
(0.474)

Oil endowments In Section 2.4.3 we noted that countries with oil might experience
more frequent conict simply because oil revenues confer resources that can be spent on
weaponry and other military capabilities. Our regressions already control for GDP and
a measure of military capability, so in principle this eect should indirectly already have
been absorbed by these variables. Having said this, in order to make sure that our distance
variables do not spuriously correlate with oil endowments, in Table 4 we perform further
robustness checks with respect to the overall quantitative endowments of oil in the two
countries in each pair.36 Specically, we control (in turn) for the mean and dierence of:
oil output (column 1), estimated oil reserves (2), and oil output as a share of GDP (3).
Further, we control for oil output in the country with oil closest to the border, and in the
country with oil further from the border (column 4). In column 5 we perform the same
analysis as in column 4, but this time with oil reserves.
We can make three broad observations from the results of Table 4. First, and most
important, the results relating to distance of the oil from the border are very robust, both
in magnitude and in statistical signicance, to controlling for the overall oil endowments.
Second, the One and Two dummies are less systematically signicant, especially the latter.
However, we have veried that this eect is due to a drop in the sample size. Third, the oil
output/endowment variables are only rarely statistically signicant predictors of conict,
possibly because their inuence is already captured by the controls for GDP and for military
capabilities.
3.2.2

Endogenous Borders

In interpreting our regressions so far we have implicitly assumed that borders are located
randomly in space - or at least without consideration for the presence and location of the
oil. There may be reasons to query this identifying assumption, as the process by which
borders come about may be aected by the spatial distribution of oil elds. Indeed in our
own model the ex-post border is certainly endogenous to the oils location, since countries

36

Recall that we do not have oil eld-level information on endowments, so we cannot test the models

predictions with respect to oil eld size.

35

One
One x Dist
Both
Both x MinDist
Both x MaxDist
Oil Production (mean)
Oil Production (difference)

(1)
0.082**
(0.036)
-0.101***
(0.034)
0.042
(0.032)
-0.070***
(0.025)
-0.013
(0.028)
-0.008**
(0.004)
0.002
(0.001)

Dependent variable: Hostility

(2)
(3)
(4)
0.061
0.078**
0.084**
(0.039)
(0.036)
(0.037)
-0.082**
-0.099***
-0.104***
(0.034)
(0.034)
(0.035)
0.013
0.039
0.036
(0.042)
(0.032)
(0.032)
-0.072***
-0.072***
-0.070***
(0.025)
(0.024)
(0.024)
-0.009
-0.010
-0.014
(0.028)
(0.027)
(0.027)

Oil Reserves (mean)

(5)
0.061
(0.039)
-0.083**
(0.034)
0.016
(0.042)
-0.072***
(0.026)
-0.012
(0.029)

0.062
(0.123)
-0.041
(0.031)

Oil Reserves (difference)

Oil Production / GDP (mean)

-0.188***
(0.063)
0.063
(0.043)

Oil Production / GDP (difference)

Oil Production (further)

0.000
(0.001)
-0.001
(0.001)

Oil Production (closer)

Oil Reserves (further)

0.020
(0.035)
Oil Reserves (closer)
0.018
(0.034)
Observations
9580
7089
9240
9580
7089
R-squared
0.167
0.181
0.161
0.166
0.181
Note: The unit of observation is a country pair in a given year. The sample covers all contiguous country pairs and the
years 1946-2001. Method: OLS with robust standard errors clustered at the country-pair level. Significance levels ***
p<0.01, ** p<0.05, * p<0.1. The oil variables are constructed using all oil fields (onshore and offshore). All explanatory
variables are taken as first lag. All specifications control for intercept, annual time dummies, country fixed effects for each
country of the dyad, the average and the absolute difference of land areas in the pair, the average and absolute
difference of GDP per capita, the average and absolute difference of population, the average and absolute difference of
fighting capabilities, the average and absolute difference of democracy scores, dummy for one country having civil war,
dummy for both countries having civil war, bilateral trade / GDP, dummy for one country being OPEC member, dummy
for both countries being OPEC member, genetic distance between the populations of the two countries, dummy for
membership in the same defensive alliance, dummy for historical inclusion in the same country, kingdom or empire,
dummy for having been in a colonial relationship, and years since the last hostility in the country pair.

Table 4: Robustness with respect to oil quantities

36

enter into (potentially) border-changing conict with a view of capturing each others oil.
But even ex-ante borders, i.e. borders drawn before countries have made conict-peace
decisions, could have been inuenced by the location of oil. For example, a country with
more bargaining power might have insisted on deviating somewhat from naturalborders
in order to insure oil elds remained on its side. Or, colonial powers might have chosen
to draw post-colonial borders so as to make sure that oil elds are located in the country
more likely to be friendly to its interests - or perhaps so as to divide the oil elds between
the two countries in order to diversify the risk of disruption arising from turbulence in any
one country.
In order to address these concerns, we follow three distinct strategies. The rst strategy
is to focus on observations were we know that the border predates the discovery of oil. The
second strategy is to focus on observations in which the border has the physical appearance
of a natural border. The third strategy is to focus on observations in which the distance
variables are distances of the oil from a coastline, which are necessarily exogenous.
We begin with borders that were drawn/set before the oil was discovered. If the parties
do not know the oil is there, they cannot be inuenced by its presence when drawing the
border or ghting over territory. We implement two versions of this idea in columns (1)(2) and, respectively, (3)-(4) of Table 5. In columns (1)-(2) we drop from our sample all
observations featuring a border that has changed subsequently to the rst oil discovery in
either country in the pair. More specically, we use information from Lujala et al. (2007)
to identify the date at which oil was rst discovered in either country in the pair, and
we use information from Tir et al. (1998) to identify all dates at which borders changed
between the two countries. We then drop from the analysis all observations dated after
the rst border change following the rst oil discovery.37 The results of columns (1)-(2)

37

Hence, if toil is the date at which oil was rst discovered in either country, and t1 , t2 , t3 , ... are the

ordered dates of border changes (i.e. ti > ti

1 ),

we (i) dene ~{ such that t~{

< toil and t~{

toil , and

(ii) drop all observations dated t > t~{ . Note that if oil was discovered in a country pair before 1946, and
the border experienced one or more changes between the date of discovery and 1946, the country pair is
dropped entirely from the analysis. Also note that we do not observe border changes before 1816, so t1
is the rst border change after 1816. However oil was a nearly valueless commodity before 1816 so any

37

show that our key ndings are statistically and economically robust to dropping borders
that changed after oil discoveries.38
The exercise in columns (1)-(2) is suitable to remove concerns with ex-post endogeneity,
i.e. with border changes in response to oil discoveries. However, it is still potentially
vulnerable to ex-ante endogeneity, i.e. with the position of the oil aecting the drawing of
the original borders. To address ex-ante endogeneity, in columns (3)-(4) we further drop all
country pairs which rst came to share a border (for example when one or both countries
rst came into existence) after oil was rst discovered in either of them.39 Again, despite
the substantial drop in sample size, our headline results on minimum distance turn out to
be robust.
Our second strategy to assess the threat to identication posed by endogenous borders
is to drop country pairs whose borders look articial. This strategy, inspired by recent
work by Alesina et al. (2011), consists of building, for each bilateral land border, a measure
of the deviation of the actual border from a relatively smooth arc (see Appendix A for a
detailed description). We name this variable border snakiness. The idea is that the
smoother the border (the less snaky it is), the more likely it is to have been designed
articially, while the more snakyit is, the more likely it is to follow natural geographical
features like mountain ridges or rivers. Based on this reasoning, in columns (5)-(6) we re-

border change before that date cannot conceivably have been motivated by oil.
38
The removal of borders which changed after oil discoveries, as well as the other strategies examined
in Table 5, involves a signicant drop in sample size. Our full set of controls induces further losses due
to missing values. For these reasons, we include in the table specications with no controls alongside our
benchmark specications with controls.
39
Following on the same notation, denote now t0 the date at which the border between two countries
rst came in existence. We now drop all the same observations as in columns (1) and (2) and, in addition,
all those satisfying toil

t0 . As before, however, all pairs where the border was drawn before 1816 (which

is the start date of the Correlates of War data on state creation) are kept in the analysis, on the ground
that oil could not have inuenced these borders even if its presence was known at the time. To nd out the
earliest establishment of current borders for all pairs, we have used data from Strang (1991), Correlates
of War (2010), CIA (2012) and Encyclopedia Britannica (2012). Note that we use the date of the rst
drawing of the currently active borders, even if this date is earlier than independence, e.g. when borders
were already drawn in colonial times.

38

estimate our baseline specications only on the subset of country pairs with above median
snakiness. Once again despite the massive loss of sample size, the key results appear robust
(except for the coe cient on Both x MinDist,which loses signicance in the specication
with the full set of controls).
Our third and nal strategy to assuage concerns with endogeneity builds on the fact
that coastlines, as opposed to land borders, are (mostly) exogenous to human activity.
Recall that our sample contains both country pairs that share a land border and country
pairs that do not share a land border but are separated by less than 400 miles of water. In
the latter case, by construction, all our oil distance variables are distances of oil elds to the
other countrys coastline. Because both the oil location, and the position of the coastline
are natural phenomena, it is di cult to think of plausible mechanisms that would lead
these distances to respond to incentives by the two countries in the pair. Accordingly, in
columns (7)-(8) we re-estimate our main specications on the subsample of pairs that do
not share a land border.40 Even with this most restrictive criterion for inclusion in the
sample we nd that our headline results largely hold (except for the last column where the
reduction in sample size is most extreme and where our key variables now narrowly miss
the 10% signicance threshold).

40

Note that by construction the subset for the results in columns (7) and (8) is a strict subsample of the

corresponding samples in the other columns of this table. This is because we have treated coastlines as
pre-existing any oil discovery (so all country pairs without a land border are retained in columns (1)-(4))
and because we have treated all bodies of water separating countries (other than rivers) as natural,and
hence assigned maximum snakiness to country pairs that do not share a land border (so all country pairs
without a land border are included in columns (5)-(6)). Recall that only about 40% of country pairs do
not share a land border. It may also be appropriate to note that in this subsample about 50% of the pairs
have the closest oil onshore.

39

40

No border changes after oil

discovery
No
Yes
16504
9482
0.151
0.149

(2)
0.087**
(0.034)
-0.103***
(0.032)
0.036
(0.026)
-0.055**
(0.025)
-0.025
(0.030)

(8)
0.191
(0.117)
-0.195
(0.122)
0.055
(0.038)
-0.023
(0.039)
-0.044
(0.046)
Only country pairs without
land border
No
Yes
8168
4341
0.172
0.148

(7)
0.072**
(0.031)
-0.061**
(0.025)
0.154***
(0.056)
-0.157**
(0.073)
0.016
(0.081)

Table 5: Controlling for potentially endogenous or articial borders

Note: The unit of observation is a country pair in a given year. The sample covers all contiguous country pairs and the years 1946-2001. Method: OLS with
robust standard errors clustered at the country-pair level. Significance levels *** p<0.01, ** p<0.05, * p<0.1. The oil variables are constructed using all oil
fields (onshore and offshore). All explanatory variables are taken as first lag. All specifications control for intercept, annual time dummies, country fixed
effects for each country of the dyad, and the average and the absolute difference of land areas in the pair. Additional controls are: The average and
absolute difference of GDP per capita, the average and absolute difference of population, the average and absolute difference of fighting capabilities, the
average and absolute difference of democracy scores, dummy for one country having civil war, dummy for both countries having civil war, bilateral trade /
GDP, dummy for one country being OPEC member, dummy for both countries being OPEC member, genetic distance between the populations of the two
countries, dummy for membership in the same defensive alliance, dummy for historical inclusion in the same country, kingdom or empire, dummy for
having been in a colonial relationship, and years since the last hostility in the country pair.

Sample
Additional controls
Observations
R-squared

Both x MaxDist

Both x MinDist

Both

One x Dist

One

(1)
0.052*
(0.029)
-0.079***
(0.027)
0.045*
(0.026)
-0.100***
(0.030)
-0.011
(0.038)

Dependent variable: Hostility

(3)
(4)
(5)
(6)
0.019
0.042
0.130***
0.081**
(0.027)
(0.026)
(0.037)
(0.032)
-0.042*
-0.061**
-0.127***
-0.084***
(0.024)
(0.024)
(0.035)
(0.029)
0.007
0.012
0.164***
0.045
(0.023)
(0.027)
(0.052)
(0.034)
-0.083***
-0.061***
-0.135**
-0.038
(0.028)
(0.023)
(0.062)
(0.042)
0.026
0.009
-0.018
-0.028
(0.030)
(0.026)
(0.071)
(0.051)
No border changes after oil
discovery, historical borders
older than oil discovery or
Removed 50% with least
1816
"snaky" border
No
Yes
No
Yes
11771
7266
9907
5399
0.231
0.148
0.187
0.203

Conclusions

In this paper we have studied the eect of natural resource endowments, as well as their
geographic distribution, on the risk of inter-state conict. We have built a simple model
that predicts the risk of inter-state disputes to be largest in the presence of natural resource
asymmetry. The most dangerous situations are the ones where only one country of the
pair has oil, and this oil is close to the border. When both countries have oil, conict risk
is maximal when the location of oil elds is maximally asymmetric.
We have tested these predictions empirically with a novel geo-referenced dataset designed to capture these geographical asymmetries. Controlling for a battery of determinants of bilateral conict, as well as country xed eects and annual time dummies, we
nd large quantitative eects from asymmetric oil location. For example, country pairs
where only one country has oil near the border are as much as three to four times more
likely to engage in conict than country pairs with no oil, or where the oil is very far from
the border, or when both countries have oil near the border. These results are robust to
several strategies to deal with the potential endogeneity of bilateral borders.
While our theoretical model is novel and has the advantage of simplicity, it also has
several limitations. The theoretical framework is static, and is thus unable to capture a
host of interesting dynamic eects, particularly as regards the endogenous evolution of
borders and, hence, country size and location [e.g. Alesina and Spolaore (2003)]. This
is a priority for future work.41 Empirically, the priority is to complement our data on
oil eld location with data on oil eld size and reserves. In addition, our theory applies
equally, and our empirical methods could be usefully applied to, mineral natural resources
other than oil. Finally, one could enrich the geographic dimension of both theory and
empirics. For example, our analysis is silent on the location of oil elds vis-a-vis the
country capital, but recent work suggests a weakening of political and institutional links
away from the capital [Campante et al. (2013), Campante and Do (2014), Michalopoulos

41

In the Online Appendix we take a rst step towards a dynamic model, via a dynamic extension

with discrete geography. The qualitative predictions of the static model are robust, but even this simple
extension promises to generate additional interesting predictions that we plan to pursue in future work.

41

and Papaioannou (2014)], which might have implications for the propensity of peripheral
areas to be targets of foreign military action.

References
[1] Acemoglu, Daron, Michael Golosov, Aleh Tsyvinski, and Pierre Yared, 2012, "A Dynamic Theory of Resource Wars", Quarterly Journal of Economics 127, 283-331.
[2] Alesina, Alberto, William Easterly and Janina Matuszeski, 2011, "Articial States",
Journal of the European Economic Association 9, 246-77.
[3] Alesina, Alberto, Michalopoulos, Stelios and Elias Papaioannou, 2012, "Ethnic Inequality," NBER WP 18512.
[4] Alesina, Alberto, and Enrico Spolaore, 2003, The Size of Nations, MIT Press.
[5] Anderson, Ewan, 1999, "Geopolitics: International Boundaries as Fighting Places",
Journal of Strategic Studies 22, 125-136.
[6] Bakeless, John Edwin, 1921, The economic causes of modern war: a study of the
period: 1878-1918, New York: Moat, Yard and company.
[7] Baliga, Sandeep, David Lucca, and Tomas Sjstrm, 2011, "Domestic Political Survival and International Conict: Is Democracy Good for Peace?", Review of Economic
Studies 78, 458-486.
[8] Barbieri, Katherine, and Omar Keshk, 2012, Correlates of War Project Trade Data
Set Codebook, Version 3.0, http://correlatesofwar.org.
[9] BBC, 2011, "Western Sahara Prole", http://www.bbc.co.uk/news/world-africa14115273 (9 December 2011).
[10] Bevi, Carmen, and Luis Corchn, 2010, "Peace Agreements without Commitment".
Games and Economic Behavior 68, 469487.
[11] Besley, Timothy, and Torsten Persson, 2009, "The Origins of State Capacity: Property
Rights, Taxation, and Politics", American Economic Review 99, 1218-1244.
42

[12] Bozzoli, Carlos, Tilman Brueck, and Simon Sottsas, 2010, "A Survey of the Global
Economic Costs of Conict", Defence and Peace Economics 21, 165-176.
[13] British Petroleum, 2009, "BP Statistical Review of World Energy June 2009", dataset,
http://www.bp.com/statisticalreview.
[14] Bueno de Mesquita, Bruce, and David Lalman, 1992, War and Reason, New Haven:
Yale University Press.
[15] Campante, Filipe, and Quoc-Anh Do, 2014, "Isolated Capital Cities, Accountability,
and Corruption: Evidence from US States," American Economic Review, forthcoming.
[16] Campante, Filipe, Bernardo Guimaraes, and Quoc-Anh Do, 2013, "Isolated Capital
Cities and Misgovernance: Theory and Evidence," unpublished, Harvard University.
[17] Carter Center, 2010, "Approaches to Solving Territorial Conicts: Sources, Situations,
Scenarios, and Suggestions", mimeo, Carter Center.
[18] Caselli, Francesco, and Guy Michaels, 2013, "Do Oil Windfalls Improve Living Standards? Evidence from Brazil," American Economic Journal: Applied Economics 5,
208-238.
[19] Chassang, Sylvain, and Gerard Padr i Miquel, 2010, "Conict and Deterrence under
Strategic Risk," Quarterly Journal of Economics 125, 1821-1858.
[20] CIA, 2012, "World Factbook", dataset, https://www.cia.gov/library/publications/theworld-factbook/.
[21] Colgan, Je, 2010, "Oil and Revolutionary Governments: Fuel for International Conict", International Organization 64, 661-694.
[22] Conconi, Paola, Nicolas Sahuguet, Maurizio Zanardi, 2012, "Democratic Peace and
Electoral Accountability", Journal of the European Economic Association, forthcoming.
[23] Correlates

of

War,

2010,

"Correlates

http://www.correlatesofwar.org/.

43

of

War

Project",

dataset,

[24] Cotet, Anca M., and Kevin K. Tsui, 2013, "Oil and Conict: What Does the Cross
Country Evidence Really Show?," American Economic Journal: Macroeconomics 5,
49-80.
[25] De Soysa, Indra, Erik Gartzke, and Tove Grete Lie, 2011, "Oil, Blood, and Strategy: How Petroleum Inuences Interstate Conict", mimeo, Norwegian University of
Science and Technology, Trondheim, and UCSD.
[26] Dube, Oendrila, and Juan Vargas, 2013, "Commodity Price Shocks and Civil Conict:
Evidence from Colombia", Review of Economics Studies 80, 13841421.
[27] Encyclopedia Britannica, 2012, "Encyclopedia Britannica - Academic Edition", encyclopedia, http://www.britannica.com/.
[28] ESRI-USC, 1996, "Giant elds of the world", digital database, technical report 95-472.
Columbia, SC: Earth Sciences and Resource Institute, University of South Carolina.
[29] Esteban, Joan, Massimo Morelli, and Dominic Rohner, 2012, "Strategic Mass
Killings", mimeo, IAE, Columbia University and University of Zurich.
[30] Fearon, James, 1995, "Rationalist Explanations for War", International Organization
49, 379-414.
[31] Fearon, James, 1996, "Bargaining over Objects that Inuence Future Bargaining
Power", unpublished, University of Chicago.
[32] Fearon, James, 1997, "Signaling Foreign Policy Interests: Tying Hands versus Sinking
Costs", Journal of Conict Resolution 41, 68-90.
[33] Franco, Jerey, 1997, "Peru Ecuador Border Dispute", ICE Case Studies 5.
[34] Gartzke, Erik, 2007, "The Capitalist Peace", American Journal of Political Science
51, 166-91.
[35] Gartzke, Erik, and Dominic Rohner, 2011, "The Political Economy of Imperialism,
Decolonization, and Development", British Journal of Political Science 41, 525-556.
[36] Gaubatz, Kurt Taylor, 1991, "Election Cycles and War", Journal of Conict Resolution 35, 212-244.
44

[37] Gowa, Joanne, 1999. Ballots and Bullets: The Elusive Democratic Peace. Princeton:
Princeton University Press.
[38] Harari, Mariaavia and Eliana La Ferrara, 2012, "Conict, Climate and Cells: A
disaggregated analysis", mimeo, Bocconi University.
[39] Heston, Alan, Robert Summers, and Bettina Aten, 2009, Penn World Table Version
6.3, Center for International Comparisons of Production, Income and Prices at the
University of Pennsylvania.
[40] Hirshleifer, Jack, 1991, "The Technology of Conict as an Economic Activity," American Economic Review 81, 130-134.
[41] Jackson, Matthew O. and Massimo Morelli, 2007, "Political Bias and War", American
Economic Review 97, 1353-73.
[42] Jackson, Matthew O. and Massimo Morelli, 2010, "The Reasons of Wars - An Updated Survey", in Chris Coyne (ed.), Handbook on the Political Economy of War,
Cheltenham UK: Edward Elgar Publishing.
[43] Kaldor, Mary, Terry Karl, and Yahia Said, 2007, Oil Wars, London: Pluto.
[44] Klare, Michael, 2002, Resource Wars: The New Landscape of Global Conict, New
York: Holt.
[45] Kocs, Stephen, 1995, "Territorial Disputes and Interstate War, 1945-1987", Journal
of Politics 57, 159-175.
[46] Lujala, Pivi, Jan Ketil Rod and Nadja Thieme, 2007, "Fighting over Oil: Introducing
a New Dataset", Conict Management and Peace Science 24, 239-56.
[47] Mandel, Robert, 1980, "Roots of the Modern Interstate Border Dispute", Journal of
Conict Resolution 24, 427-454.
[48] Maoz,

Zeev,

2005,

"Dyadic

MID

Dataset

(version

2.0)",

http://psfaculty.ucdavis.edu/zmaoz/dyadmid.html.
[49] Maoz, Zeev and Bruce Russett, 1993, "Normative and Structural Causes of Democratic Peace, 1946-1986", American Political Science Review 87, 624-38.
45

[50] Martin, Philippe, Thierry Mayer, and Mathias Thoenig, 2008, "Make Trade not
War?", Review of Economic Studies 75, 865-900.
[51] Mayer, Thierry, and Soledad Zignago, 2011, "Notes on CEPIIs distances
measures (GeoDist)", CEPII Working Paper 2011-25, data available under:
http://www.cepii.fr/anglaisgraph/bdd/distances.htm#sthash.yi5CHEGw.dpuf.
[52] McLaughlin Mitchell, Sara, and Brandon Prins, 1999, "Beyond Territorial Contiguity:
Issues at Stake in Democratic Militarized Interstate Disputes", International Studies
Quarterly 43, 169-183.
[53] Michalopoulos, Stelios, and Elias Papaioannou, 2014, "National Institutions and Subnational Development in Africa," Quarterly Journal of Economics 129, 151-213.
[54] Michaels, Guy, and Yu-Hsiang Lei, 2011, "Do Giant Oileld Discoveries Fuel Internal
Armed Conicts?", mimeo, LSE.
[55] Morelli, Massimo, and Dominic Rohner, 2014, "Resource Concentration and Civil
Wars", mimeo, Columbia University and University of Lausanne.
[56] OPEC,

2012,

"OPEC

member

countries",

website,

http://www.opec.org/opec_web/en/about_us/25.htm.
[57] Petroleum Economist, 2003, World Energy Atlas, London: Petroleum Economist.
[58] Ploeg, Frederick van der, 2011, "Natural Resources: Curse or Blessing?", Journal of
Economic Literature 49, 366420.
[59] Ploeg, Frederick van der, and Dominic Rohner, 2012, "War and Natural Resource
Exploitation," European Economic Review 56, 1714-1729.
[60] Polachek, Solomon, 1980, "Conict and Trade", Journal of Conict Resolution 24,
55-78.
[61] Polity IV, 2009, "Political Regime Characteristics and Transitions, 1800-2007",
Dataset, www.systemicpeace.org/polity4.
[62] Potter, Philip, 2007, "Does Experience Matter? American Presidential Experience,
Age, and International Conict", Journal of Conict Resolution 51, 351-378.
46

[63] Powell, Robert, 1996, "Uncertainty, Shifting Power, and Appeasement", The American Political Science Review 90, 749-764.
[64] Powell, Robert, 2006, "War as a Commitment Problem", International Organization
60, 169-203.
[65] Price, Felicia, 2005, "The Bakassi Peninsula: The Border Dispute between Nigeria
and Cameroon", ICE Case Studies 163.
[66] Rohner, Dominic, Mathias Thoenig, and Fabrizio Zilibotti, 2013, "War Signals: A
Theory of Trade, Trust and Conict", Review of Economic Studies 80, 1114-1147.
[67] Ross, Michael, 2012, The Oil Curse: How Petroleum Wealth Shapes the Development
of Nations, Princeton: Princeton University Press.
[68] Skaperdas, Stergios, 1992, "Cooperation, Conict, and Power in the Absence of Property Rights", American Economic Review 82, 720-39.
[69] Skaperdas, Stergios, and Constantinos Syropoulos, 2001, "Guns, Butter, and Openness: On the Relationship between Security and Trade", American Economic Review
91, 353-357.
[70] Spolaore, Enrico, and Romain Wacziarg, 2009, "The Diusion of Development", Quarterly Journal of Economics 124, 469-529.
[71] Spolaore, Enrico, and Romain Wacziarg, 2013, "War and Relatedness," unpublished,
UCLA.
[72] Stinnett, Douglas, Jaroslav Tir, Philip Schafer, Paul Diehl, and Charles Gochman,
2002, "The Correlates of War Project Direct Contiguity Data, Version 3", Conict
Management and Peace Science 19, 58-66.
[73] Strang, David, 1991, "Global Patterns of Decolonization, 1500-1987", International
Studies Quarterly 35, 429-454.
[74] Tir, Jaroslav, 2003, "Averting Armed International Conicts Through State-to-State
Territorial Transfers", Journal of Politics, 65, 1235-1257.

47

[75] Tir, Jaroslav, Philip Schafer, Paul F. Diehl and Gary Goertz, 1998, "Territorial
Changes, 1816-1996: Procedures and Data", Conict Management and Peace Science 16, 89-97.
[76] Tomz, Michael, Gary King, and Langche Zeng, 2003, "ReLogit: Rare Events Logistic
Regression", Journal of Statistical Software 8, 127.
[77] Uppsala Conict Data Program, 2011, "UCDP/PRIO Armed Conict Dataset
v.4-2011, 1946 2010", dataset, http://www.pcr.uu.se/research/ucdp/datasets
/ucdp_prio_armed_conict_dataset/.
[78] USGS, 2000, "U.S. Geological Survey world petroleum assessment 2000 Description
and results", U.S. Geological Survey World Energy Assessment Team. USGS Digital
Data Series DDS-60. Multi Disc Set. Version 1.1. http://pubs.usgs.gov/dds/dds-060/.
[79] Weede, Erich, 1973, "Nation-Environment Relations as Determinants of Hostilities
Among Nations", Peace Science Society (International) Papers 20, 67-90.
[80] Weidmann, Nils, Doreen Kuse, and Kristian Skrede Gleditsch, 2010, "The Geography
of the International System: The CShapes Dataset", International Interactions 36,
86-106.
[81] Westing, Arthur, 1986, Global Resources and International Conict: Environmental
Factors in Strategic Policy and Action, Oxford: Oxford University Press.
[82] World

Bank,

2009,

"World

Development

Indicators",

dataset,

http://go.worldbank.org/U0FSM7AQ40.
[83] Wright, Quincy, 1942, A Study of War, Chicago: University of Chicago Press.

Appendix A: Data
This appendix describes the variables used in section 3, and provides summary descriptive
statistics in Table 6.

48

4.1

Dependent Variables

The dependent variables, "Hostility", "War", "Hostility+" and "Dispute intensity" have
been explained in detail in subsection 3.1. We now explain the dependent variables used
in Appendices B and C, respectively.
Revisionist (Attacker) [Initiator]: We use the variables revstata and revstatb
(sidea and sideb) [rolea and roleb] from Maoz (2005) which take values 1 if state A and,
respectively, state B are deemed to seek to change the borders (be on the side of the
initiator) [be the conicts initiator], and 0 otherwise. Note that it is possible that in a
country pair either both, one or neither of the countries are revisionist (attacker) [initiator].
We construct our dependent variable for Appendix B as the dierence between the dummy
for the rst-listed country in the pair (country A) and the dummy for the second-listed
country (country B). This variable can be interpreted as a measure of relative aggressiveness
of country A. This allows us to run a specication quite similar in spirit to the benchmark
model we used for the other dependent variables. In particular, we estimate the impact
on the relative aggressiveness of A of: oil in country A, distance of country As oil to the
border, oil in country B, and distance of country Bs oil to the border.
"Territorial Change": Our dependent variable for Appendix C is a dummy taking a
value of 1 if there has been a territorial change in a given pair year. From Tir et al. (1998),
version 4.01 obtained from http://www.correlatesofwar.org/.

4.2

Explanatory Variables

The explanatory variables One, Both, Dist, MinDist, and MaxDist have also been described
in the detail in the main text. The others are as follows.
"Land area": In 1000 Square kilometers. From World Bank (2009).
"GDP per Capita": Real Gross Domestic Product per Capita (in 1000), Current Price
National Accounts at PPPs. From Heston et al. (2009).
"Population": In Millions. From Heston et al. (2009).
"Capabilities": Capability scores from Correlates of War (2010).
"Polity Score": Democracy scores ranging from -10 (strongly autocratic) to +10 (strongly
democratic). From Polity IV (2009).
"Number of years since the last hostility, resp. war between the countries in the pair":
49

Authorscalculations, based on the "hostility", "war", resp. "hostility (broad denition)"

variables.
"Bilateral trade /GDP": Sum of total bilateral trade between the two countries of the
pair divided by the sum of their total GDPs. Bilateral trade data from Barbieri and Keshk
(2012), GDP data from Heston et al. (2009).
"Genetic distance between the populations of the two countries": Genetic distance between current plurality groups (variable "fst_distance_dominant"), from Spolaore and
Wacziarg (2009).
"Defensive pact": Dummy taking a value of 1 if the countries of the pair are together
in a defense pact, and 0 otherwise. From Correlates of War (2010).
"Historical inclusion in the same country, kingdom or empire": Dummy variable taking
value 1 if countries were or are the same country (variable "smctry"), from Mayer and
Zignago (2011).
"Having been in a colonial relationship": Dummy variable taking value 1 for pairs that
were ever in colonial relationship (variable "colony"), from Mayer and Zignago (2011).
"CW1": Dummy with value of 1 if there is a civil war in one country of the pair, and
0 otherwise. Constructed using data from Uppsala Conict Data Program (2011).
"CW2": Dummy with value of 1 if there is a civil war in both countries of the pair,
and 0 otherwise. Constructed using data from Uppsala Conict Data Program (2011).
"OPEC1": Dummy with value of 1 if one country in the pair is an OPEC member,
and 0 otherwise. From OPEC (2012).
"OPEC2": Dummy with value of 1 if both countries in the pair are OPEC members,
and 0 otherwise. From OPEC (2012).
"Oil production": In 10 million tones (mean = 3). From British Petroleum (2009).
"Oil reserves": In 100 billion barrels. From British Petroleum (2009).
"Oil production/GDP": Total value of current oil production / GDP. Production quantities and prices from British Petroleum (2009), corresponding GDP in current prices from
World Bank (2009).
"Border snakiness": Authorscalculations. Using the geo-referenced shapes of bilateral
country borders from Weidmann et al. (2010), we compute an index of bilateral border
snakiness, using the following formula: "Border snakiness" = "Actual bilateral border

50

length" / (0.5 * "Convex hull below the bilateral border" + 0.5 * "Convex hull above the
bilateral border"). This measure takes a value of 1 when the border is a straight line, while
its value increases when the border becomes more winding, resp. snaky.

Appendix B: Directed Dyads

The results of the regressions with directed dyads are displayed in Table 7.

Appendix C: Border Changes

The results of the regressions with border changes are displayed in Tables 8 and 9.

51

Variable
Hostility
War
Hostility+ (Int. 3, 4 and 5)
Hostility scale (cont.)
State A revisionist
State A attacker
State A initiator
Border change
One
One x Dist
Both
Both x MinDist
Both x MaxDist
Land area (mean)
Land area (diff)
GDP p.c. (mean)
GDP p.c. (diff)
Pop. (mean)
Pop. (diff)
Capabilities (mean)
Capabilities (diff)
Democracy (mean)
Democracy (diff)
Bilat. Trade / GDP
Genetic distance
Defensive pact
Ever same country
Colonial relation
Civil war 1
Civil war 2
OPEC 1
OPEC 2
Oil prod. (mean)
Oil prod. (diff)
Oil res. (mean)
Oil res. (diff)
Oil/GDP (mean)
Oil/GDP (diff)
Border snakiness

Obs.
20564
24387
20564
20564
19965
19965
19965
24387
24387
24387
24387
24387
24387
24366
24366
18075
18075
20418
20418
20489
20489
20055
20055
17201
23566
19948
22738
22738
24387
24387
24387
24387
19547
19547
15606
15606
18600
18600
24387

Mean
0.057
0.004
0.072
0.284
0.011
0.006
0.008
0.004
0.349
0.285
0.512
0.253
0.332
1330.520
1928.770
6.099
3.941
31.991
45.064
0.013
0.019
0.086
5.499
0.003
374.012
0.389
0.156
0.062
0.263
0.041
0.134
0.025
2.850
4.674
0.088
0.139
0.061
0.077
1.929

Std. Dev.
0.233
0.066
0.259
1.014
0.216
0.263
0.256
0.061
0.477
0.424
0.500
0.382
0.431
2277.200
3909.240
7.219
5.982
63.520
109.178
0.024
0.040
6.508
6.063
0.007
563.105
0.488
0.363
0.242
0.440
0.197
0.340
0.157
5.975
9.930
0.252
0.396
0.125
0.149
0.744

Table 6: Summary Statistics

52

Min.
0
0
0
0
-1
-1
-1
0
0
0
0
0
0
0.300
0.010
0.108
0
0.053
0.003
0
0
-10
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1

Max.
1
1
1
5
1
1
1
1
1
1
1
1
1
13365.100
17055.200
80.983
99.604
682.280
1129.500
0.200
0.362
10
20
0.121
2908
1
1
1
1
1
1
1
52.175
56.950
2.014
2.643
1.125
1.213
2.757

53

-0.022
(0.014)
0.033**
(0.016)
0.031*
(0.017)
-0.035**
(0.017)
No
No
19962
0.006

(2)
(3)
State A revisionist
-0.012
-0.044**
(0.010)
(0.019)
0.022
0.032**
(0.015)
(0.016)
0.003
-0.004
(0.022)
(0.014)
-0.018
-0.025*
(0.021)
(0.014)
No
Yes
Yes
No
11303
19962
0.022
0.058
-0.034**
(0.017)
0.045***
(0.015)
0.000
(0.016)
-0.022*
(0.013)
Yes
Yes
11303
0.089

(4)
-0.006
(0.015)
0.027*
(0.015)
0.008
(0.018)
-0.014
(0.017)
No
No
19962
0.005

(5)

(6)
(7)
State A attacker
-0.018
-0.022
(0.014)
(0.017)
0.018
0.033*
(0.015)
(0.017)
-0.010
-0.007
(0.017)
(0.020)
-0.005
-0.021
(0.016)
(0.017)
No
Yes
Yes
No
11303
19962
0.014
0.046
-0.041***
(0.015)
0.040***
(0.013)
-0.010
(0.018)
-0.023
(0.016)
Yes
Yes
11303
0.048

(8)
-0.006
(0.015)
0.029*
(0.016)
0.004
(0.018)
-0.011
(0.018)
No
No
19962
0.006

(9)

(10)
(11)
State A initiator
-0.016
-0.024
(0.015)
(0.016)
0.019
0.037**
(0.015)
(0.016)
-0.012
-0.014
(0.017)
(0.019)
-0.001
-0.017
(0.016)
(0.017)
No
Yes
Yes
No
11303
19962
0.017
0.048

-0.041***
(0.015)
0.041***
(0.013)
-0.017
(0.018)
-0.016
(0.017)
Yes
Yes
11303
0.054

(12)

Table 7: Regressions with Directed Dyads

Note: The unit of observation is a country pair in a given year. The sample covers all contiguous country pairs and the years 1946-2001. Method: OLS with robust standard
errors clustered at the country-pair level. Significance levels *** p<0.01, ** p<0.05, * p<0.1. The oil variables are constructed using all oil fields (onshore and offshore). All
explanatory variables are taken as first lag. The dependent variable in the columns 1-4 is the dummy for country A being revisionist minus the dummy for country B being
revisionist (hence the dependent variable takes values of -1, 0, and 1). The construction of the dependent variable is analogous for columns 5-8 and 9-12 with being
attacker, resp. initiator instead of revisionist as underlying variable. All specifications control for intercept, land areas of both countries and annual time dummies.
Additional controls for each country in the pair are: Population, GDP per capita, democracy score, capabilities, dummy for having a civil war, dummy for being OPEC
member, and all the controls at the country pair level, which are, years since the last hostility in the country pair, bilateral trade / GDP, genetic distance between the
populations of the two countries, dummy for membership in the same defensive alliance, dummy for historical inclusion in the same country, kingdom or empire, and
dummy for having been in a colonial relationship.

Country FE
Add. Controls
Observations
R-squared

Oil B x MinDist B

Oil B

Oil A x MinDist A

Oil A

(1)

(1)

(2)
(3)
(4)
Dependent variable: Border Change
Hostility
0.018***
0.015***
(0.006)
(0.004)
War
0.070***
0.064***
(0.022)
(0.020)
Country FE
No
No
Yes
Yes
Observations
20564
24387
20564
24387
R-squared
0.013
0.014
0.033
0.031
Note: The unit of observation is a country pair in a given year. The sample covers all
contiguous country pairs and the years 1946-2008. Method: OLS with robust standard
errors clustered at the country-pair level. Significance levels *** p<0.01, ** p<0.05, *
p<0.1. All specifications control for intercept and annual time dummies.

Table 8: Conict and Border Changes

54

55

(2)
0.001
(0.002)
-0.001
(0.001)
0.006**
(0.003)
-0.002
(0.005)
-0.005
(0.004)
All
No
Yes
11303
0.011

(3)
0.004
(0.003)
-0.003
(0.002)
0.013***
(0.004)
-0.001
(0.002)
-0.009***
(0.003)
All
Yes
No
23768
0.027

(4)
0.005*
(0.003)
-0.004
(0.003)
0.008*
(0.004)
-0.002
(0.003)
-0.005
(0.005)
All
Yes
Yes
11303
0.035

Dependent variable: Border change

(5)
(6)
(7)
(8)
0.014*
0.014*
0.017**
0.020*
(0.007)
(0.009)
(0.007)
(0.011)
-0.016**
-0.016*
-0.019**
-0.021**
(0.008)
(0.009)
(0.007)
(0.009)
0.006
0.001
0.011**
0.004
(0.004)
(0.002)
(0.005)
(0.005)
0.002
0.000
0.000
-0.002
(0.002)
(0.002)
(0.003)
(0.002)
-0.008
-0.002
-0.010*
0.000
(0.006)
(0.004)
(0.006)
(0.003)
Offshore
Offshore
Offshore
Offshore
No
No
Yes
Yes
No
Yes
No
Yes
23768
11303
23768
11303
0.012
0.012
0.027
0.037
(9)
0.002
(0.002)
-0.003
(0.002)
0.005**
(0.002)
-0.003
(0.004)
-0.002
(0.004)
Onshore
No
No
23768
0.011

(10)
0.001
(0.002)
-0.001
(0.002)
0.006**
(0.003)
-0.004
(0.006)
-0.003
(0.004)
Onshore
No
Yes
11303
0.011

(11)
0.004
(0.003)
-0.003
(0.003)
0.011**
(0.005)
-0.004
(0.003)
-0.006*
(0.003)
Onshore
Yes
No
23768
0.026

(12)
0.005
(0.004)
-0.004
(0.004)
0.008
(0.006)
-0.002
(0.004)
-0.006
(0.005)
Onshore
Yes
Yes
11303
0.035

Table 9: Oil Location and Border Changes

Note: The unit of observation is a country pair in a given year. The sample covers all contiguous country pairs and the years 1946-2008. Method: OLS with robust standard
errors clustered at the country-pair level. Significance levels *** p<0.01, ** p<0.05, * p<0.1. All explanatory variables are taken as first lag. All specifications control for the
average and the absolute difference of land areas in the pair, intercept and annual time dummies. Additional controls are: The average and absolute difference of GDP
per capita, the average and absolute difference of population, the average and absolute difference of fighting capabilities, the average and absolute difference of
democracy scores, dummy for one country having civil war, dummy for both countries having civil war, bilateral trade / GDP, dummy for one country being OPEC
member, dummy for both countries being OPEC member, genetic distance between the populations of the two countries, dummy for membership in the same defensive
alliance, dummy for historical inclusion in the same country, kingdom or empire, dummy for having been in a colonial relationship, and years since the last border change
in the country pair.

Type Oil
Country FE
Add. Controls
Observations
R-squared

Both x MaxDist

Both x MinDist

Both

One x Dist

One

(1)
0.002
(0.002)
-0.002
(0.002)
0.006***
(0.002)
0.000
(0.003)
-0.006*
(0.003)
All
No
No
23768
0.011

Online Appendix for

The Geography of Inter-State Resource Wars
Francesco Caselli, Massimo Morelli, and Dominic Rohner
July 24, 2014

Endogenous Armament

The model is identical to the model in the paper, except that we introduce a new stage
at the very beginning. In this new rst stage, each country decides how much to invest in
armaments. Armament spending by country i is denoted by Wi . It carries a cost (Wi ).
The two countriesspending in armaments aect the distribution of Z.
Assumption 1.1 Z is a continuous random variable with domain R, density f (Z; WA ; WB ),
and cumulative distribution function F (Z; WA ; WB ).
Following the choices of WA and WB , the game proceeds exactly as in the baseline
model. Hence, if the outcome is peace, the payos are RA
RB

(WA ) for country A, and

(WB ) for country B. If there has been a war, the payos are:
UAC = RA I(Z > GA ) + RB I(Z > GB ) + cA

(WA );

UBC = RA I(Z < GA ) + RB I(Z < GB ) + cB

(WB ):

where I is an indicator function.

In the second stage of the game the players take WA and WB as given. Hence, following
the identical logic of the Proposition in the main text, we see that the probability of peace
is
P (GA ; GB ; RA ; RB ; WA ; WB ) = H( x)H(x)
x = RB [1

F (GB ; WA ; WB )]
1

RA F (GA ; WA ; WB );

and it is decreasing in the absolute value of x.

In the rst stage, each country i chooses Wi . Country Bs problem is
8
9
< [1 P (G ; G ; R ; R ; W ; W )] [R F (G ; W ; W ) + R F (G ; W ; W )] =
A
B
A
B
A
B
A
A
A
B
B
B
A
B
max
;
WB :
;
+P (G ; G ; R ; R ; W ; W )R
(W )
A

which can be rewritten as

max f [1

H( x)H(x)]x + RB

WB

(WB )g :

From this, we see that even in the two-stage game with endogenous arming eort our
Remark 2 in the baseline model still holds: the case RA = 0 is isomorphic to the case
GA !

1, and the case RB = 0 is isomorphic to GB ! 1. In other words, we can just

write WA as a function of GA and GB , and WB as a function of GA and GB . Then we

can dene a function (Z; GA ; GB )

F (Z; WA (GA ; GB ); WB (GA ; GB )). Accordingly, the

probability of peace is
P (GA ; GB ) = H( x)H(x)
x = RB

RB (GB ; GA ; GB )

RA (GA ; GA ; GB ):

As in the baseline model, we set RA ; RB 2 f0; 1g. We then obtain the following
predictions:
Proposition 1.1
(i) P (GA ; 1)

P ( 1; 1);

(ii) @P (GA ; 1)=@GA

1 (GA ; GA ; 1)

P ( 1; 1);

(iv) P (GA ; 1)

P (GA ; GB ) if and only if

1 (GA ; GA ; GB )

(GB ; GA ; GB ) > 0 and

where

(GA ; GA ; 1)

(GB ; GA ; GB )

(GA ; GA ; GB )

(GB ; GA ; GB )

or 1

(GA ; GA ; GB )

(GA ; GA ; 1);

(v) @P (GA ; GB )=@GA

0 and

2 (GA ; GA ; 1);

>

(iii) P (GA ; GB )

(GA ; GA ; GB )

of

0 if

1 (x; y; z)

0 if and only if either 1

[ 2 (GA ; GA ; GB ) +
1 (GA ; GA ; GB )

is the derivative of

<

2 (GB ; GA ; GB )]

[ 2 (GA ; GA ; GB ) +

with respect to x, and

with respect to y:

2 (GB ; GA ; GB )] ;
2 (x; y; z)

is the derivative

The proposition says that parts (i) and (iii) of the original proposition in the baseline
model are entirely robust to endogenous arming eort (or any other endogenization of
the distribution of contest success). This is not surprising. Even if the two countries can
respond to oil discoveries by arming there is still a country who (weakly) expects to do
better from war, and conict unambiguously goes up relative to the case where neither
country had oil.
Part (ii) is robust as long as

1 (GA ; GA ; 1)

0 so our original prediction fails only if

value. We can think of

>

2 (GA ; GA ; 1).

2 (GA ; GA ; 1)

Clearly

1 (GA ; GA ; 1)

is negative and large in absolute

as the directeect of moving As oil towards the border, which

is to make As oil within easier reach for country B. On the other hand,

2 (GA ; GA ; 1)

is the indirect eect that occurs as the two countries endogenously change their armaments. For

to be negative it must be the case that the relative military strength of

country A increases as country As oil moves towards the border. This may be the case,
for example, if country A responded to the movement of the oil towards the border by
ramping up military spending much more than country B. Clearly both countries may
have an incentive to increase military spending as country As oil moves towards the border, but there is no particular reason to presume that country A will do more - at any
rate not so much more that

becomes su ciently negative for the indirect eect from

endogenous arming to overcome the direct geographic eect. In the special case where the
two countries have the same marginal cost from arming we expect

= 0, and our original

proposition goes through unchanged.

The notion of direct and indirect eects helps interpreting the new parts (iv)
and (v) as well. To interpret part (iv) note that we can think of the quantity 1
(GB ; GA ; GB )

(GA ; GA ; GB ) as country As (expected) net territorial gain from con-

ict after the discovery of oil in country B, and of

(GA ; GA ; 1) as country As (ex-

pected) loss (equal to Bs expected gain) before Bs oil discovery. Then the proposition says that: (i) if country A continues to be the net territorial loser after the discovery of oil in country B [1

(GB ; GA ; GB )

(GA ; GA ; GB )

0], then the new discov-

ery has reduced conict unless Bs discovery of oil on its own soil has increased Bs
benets from conict compared to when only A had oil. Clearly the direct eect of
the discovery militates against this [this condition for less conict is always satised if

>

(GA ; GA ; GB ) = (GA ; GA ; 1)]. What is required is that B increases its military spending relative to A so much as to overturn the direct eect. Once again, it is not clear why
this should be the case. (ii) If instead country A becomes the net territorial winner after the new discovery [1

(GA ; GA ; GB ) > 0], then the discovery in the

(GB ; GA ; GB )

second country has reduced conict if Bs net benet from conict before the discovery
exceeds As benets after. As in the baseline model, we would expect this to be the case
in most scenarios, since in the two-oil scenario both countries have something to lose from
conict, while in the one-oil scenario the net territorial winner has no territorial losses to
fear.
This reasoning also allows us to easily interpret part (v). If 1
(GB ; GA ; GB )

(GA ; GA ; GB )

0 then A is the territorial loser, and the relevant condition for increased

conict from an increase in GA ,

says that the direct eect

1 (GA ; GA ; GB )

[ 2 (GA ; GA ; GB ) +

2 (GB ; GA ; GB )],

> 0 must exceed any indirect eect from endogenous arming.

In particular, to reverse the predictions of the baseline model the territorial loser must
respond to a movement of its oil towards the border by a disproportionate increase in
armaments (so that

is not only negative but also large in absolute number). Again, it

is hard to see why this should generally be the case. (The case were B is the territorial
loser is entirely symmetric, of course).

A Dynamic Model

In this section we show that the sharp predictions of our model summarized in Proposition
1 extend to a dynamic setting in which a war can be followed by other wars. Given that
in the model in the main text a war has a stochastic outcome, the dynamic model will
continue to have that feature, but we need to "discretize" the model in order to avoid
having to handle a continuum of realizations of the state variable, namely the border.
In particular, we now switch to a nitegeography which identies the world with the
segment [ Z; Z], Z > 0; and assume that the border can take only 5 values (locations),
namely

Z,

Z ; Z0 ; Z , and Z, with Z0 = 0, Z > 0.1 Given any distribution of oil on

This discretization is for convenience but it also has the avor of a realistic feature. Namely, in case

the space, we need to study the equilibrium strategies at the initial state of the world,
Z0 (the middle), fully anticipating the continuation war-or-peace behavior that would be
taken at all other nodes Zt =

Z or Zt = Z if reached as a consequence of a war. The

game has an innite number of periods.

In case of war let pZ denote the probability of a change of the state "one step to the
right" and qZ the probability of a shift "one step to the left", with 1

pZ

qZ being the

probability of remaining at border Z. We make the following two assumptions:

Assumption 2.1 (Equal strength) pZ = qZ = p 8z =

Z ; Z0 ; Z ; p

=q

= pZ =

qZ = 0.
The expression p

= q

= pZ = qZ = 0 means that once a country is fully

conquered it ceases to exist, implying that the border will never change anymore. The
present discounted value of being conquered (and ceasing to exist) is normalized to zero,
i.e. V A (Z =

Z) = V B (Z = Z) = 0.

Assumption 2.2 (Equal cost of war) At time 0 there is a realization of the net benet
parameter c, from a distribution H on the real line. This parameter remains constant for
the subsequent periods and it is the same for both countries.
Both the equal strength assumption and the equal c assumption can be removed, only
at the cost of more complex algebra. With these two assumptions we can focus exclusively
on the role of the geographic asymmetries.
Consistent with the basic model, assume that the instantaneous utility of a country
depends only on oil rents: uA (Zt ) = RA I(Zt > GA ) + RB I(Zt > GB ), where Gi , i = A; B
denotes the location of an oil well on the line, and I is an indicator function. It is symmetric
for B.
The game starts at time 0 at Z0 and then the two players are free to decide in every
subsequent period whether to accept the current state of the world (border) or try to change

one country succeeds in moving the border, the new border is often located along the next natural
barrier or ridge within the territory of the other country. For example, if one country manages to breach
a river border, the defender are likely to retreat to the next river and retrench there.

it with a war. We shall assume that in every period one of the two players is randomly
selected with equal probability to move rst. This assumption serves the purpose of ruling
out all weakly dominated equilibria.2
The timing under war is such that at the beginning of the war period the net benet
of war c (negative on average under most realistic assumptions on the distribution of c) is
received, and in the same period the border shifts or stays. The instantaneous utility of
the period is the one corresponding to the new border.
In terms of oil endowments, in order to mimic the situations discussed in the main text,
we consider the following possibilities:
Case No-No: RA = RB = 0 (no oil);
Case Far-No: Ri = 1, R

= 0, jGi j > Z (oil only in one country, far from the

= 0, jGi j < Z (oil only in one country, near the border);

border);
Case Close-No: Ri = 1, R

Case Far-Far: RA = RB = 1, jGA j > Z , jGB j > Z (oil in both countries, far from
the border);
Case Close-Close: RA = RB = 1, jGA j < Z , jGB j < Z (oil in both countries, near
the border);
Case Far-Close: RA = RB = 1, jGi j > Z , jG i j < Z (oil in both countries, one far
and one near the border);
where

i stands for B (A) when i stands for A (B).

Given the equal strength assumption and the equal cost of war assumption, the three
asymmetric cases can be described just in terms of endowment and distance: both case
Far-No and case Close-No are asymmetric in terms of endowment, but the latter case has
closer oil to the initial border; case Far-Close does not have endowment asymmetry but
only distance asymmetry.
Let us consider the Markov Perfect Equilibrium of this game, which allows players to

Notice that an alternative way of ruling out weakly dominated equilibria would be to focus on a

"trembling hand perfect" equilibrium renement, which would lead to identical results. Note also that all
results of the static model of the main text would be unchanged if we made there the analogous assumption
of sequential moves.

consider only payo relevant information from the history of play.This standard selection
gives prominence to the state variable "border" alone.
Proposition 2.1 There exists a unique Markov Perfect Equilibrium (MPE). In the MPE,
the probability of war in period 0 is:
1. Highest in case Close-No;
2. Lowest in cases No-No, Far-Far, and Close-Close;
Proof.
At the two absorbing one-country states the value for the unique surviving state is
V B ( Z) = V A (Z) = kR=(1

), where k denotes the number of oil wells in the whole

region, which can take values 0,1,2.3

Let us now describe the values at the other states. For peace in every period,
V i (Z) =

ui (Z)
; i = A; B; Z =
1

Z ; Z0 ; Z :

Besides these "value of peace" simple expressions, denote by W i (Z) the continuation
expected value of a war starting from a border Z.
Assume rst that in states Z0 and Z there will always be peace, and let us derive the
indierence condition in state
(uA ( Z ))=(1

Z between war and peace. We know that V A ( Z ) =

), and will set V A ( Z ) = W A ( Z ) to nd the threshold. Given that

W A( Z ) = c + p

uA (Z0 )
+ (1
1

2p) uA ( Z ) + V A ( Z ) ;

the threshold for indierence is

cA =
Assume now that at states

p
1

[2uA ( Z )

uA (Z0 )]:

Z and Z there will always be peace. Solving for the

indierence threshold at state Z0 yields:

Here we normalize the oil quantity per well to R, which is slightly more general than in the baseline

model, where it is normalized to 1.

cA
0 =
Assuming that at states

(2uA (Z0 )

uA ( Z )

uA (Z )):

Z and Z0 there will always be peace, solving for the indif-

ference threshold at state Z yields:

cA
+ =

p
1

(2uA (Z )

uA (Z0 )

kR):

The thresholds for country B are analogous.

Consider case No-No: In the absence of oil, the thresholds dened above are
A
B
B
cA = cA
= cB
0 = c+ = c
0 = c+ = 0:

For any c < 0, i.e., whenever the net benets of war are negative, peace is the unique
Markov Perfect Equilibrium. In fact, with c < 0, not only it is an equilibrium to stay
peaceful given the (consistent) assumptions at the adjacent nodes, but it remains a best
response in case 0 even when at an adjacent node there is war.
On the other hand, whenever c > 0, given that the expected value of eternal peace is 0,
while with positive c the expected value of seeking war is strictly positive for both players,
the unique MPE involves war at every node.
Consider now case Far-No. The thresholds become:
cA =

p
1

R;

A
B
B
cA
0 = c+ = c0 = c+ = 0;

cB =

p
1

R:

Clearly A would always have (weakly) lower incentives to start war than B. Hence, we
can focus on Bs incentives, which are the binding constraint for the existence of a Markov
p

Perfect Equilibrium with peace. Clearly, when c <

R, there will always exist a MPE

with peace (note that this is a higher cost of war with respect to the one needed in case
0), and no other MPE can exist.
Now consider c >

p
1

R. Assuming war in node

Z and peace in node Z , we can

derive Bs indierence condition in node Z0 . The value of a war in state Z0 is

W B (Z0 ) = c + p(uB ( Z ) + W B ( Z )) + p
8

uB (Z )
+ (1
1

2p)

uB (Z0 )
;
1

where
W B( Z ) = c + p

+p

uB (Z0 )
+ (1
1

2p)(uB ( Z ) + W B ( Z ));

c + p 1R + p u 1 (Z0 ) + (1 2p)uB ( Z )
W ( Z )=
:
1
(1 2p)
B

Imposing the indierence condition W B (Z0 ) = V B (Z0 ) = (uB (Z0 ))=(1

c=
Note that

p
(1 3p)

be war at state

), we obtain

p
p
R
(1 3p) 1

< 1 always holds. Hence, for

R<c<

p
p
(1 3p) 1

Z , but not at states Z0 and Z . For c slightly above

there will be a zone where there is war in states

R, there will
1

p
p
(1 3p) 1

Z and Z0 , but peace at state Z .

What matters are the war incentives at state Z0 , and hence we can conclude that in
case a war will occur i c >

p
p
(1 3p) 1

R.

Let us now analyze case Close-No, for which the peace in all periods thresholds derived
above become
p

cA =

1
p

cB =
p

When c <

Assume that at

R; cA
0 =
R ; cB
0 =

R; cA
+ = 0;

R; cB
+ = 0:

1
1

R, there will always be peace. Now suppose that c is a bit larger.

Z there will be war; at Z instead peace. Let us compute the condition

under which B is indierent between peace and war at Z0 .

Knowing the values at the boundaries, we also know that V B (Z0 ) = (uB (Z0 ))=(1
and V B (Z ) = (uB (Z ))=(1

). The value of war at Z0 is

W B (Z0 ) = c + p[uB ( Z ) + W B ( Z )] + p

uB (Z )
+ (1
1

2p)

uB (Z0 )
1

where
W B( Z ) = c + p

R
1

+p

uB (Z0 )
+ (1
1

2p)[uB ( Z ) + W B ( Z )]:

From the condition W B (Z0 ) = V B (Z0 ) = (uB (Z0 ))=(1

c=
Note that

1
1

(1 p) p
(1 3p) 1

R<

1
1

) it follows that

(1 p) p
R:
(1 3p) 1

p
p
(1 3p) 1

R, which implies that the ex ante probability

of war is higher when the oil is close to the border.

9

For consistency with the assumption of war at

ference threshold for A at
of war at

Z , we need to also derive the indif-

Z when there is war at node Z0 , and peace at Z . The value

Z is
W A ( Z ) = c + p[uA (Z0 ) + W A (Z0 )] + (1

uA ( Z )
1

2p)

where
W A (Z0 ) = c + p

uA ( Z )
uA (Z )
+p
+ (1
1
1

From W A ( Z ) = V A ( Z ) = (uA ( Z ))=(1

c=
This means that for c >

1
1

(1 p) p
(1 3p) 1

1
1

2p)[uA (Z0 ) + W A (Z0 )]:

), we obtain

(1 p) p
R:
(1 3p) 1
R at

Z there is war, as assumed for computing

Bs threshold at node Z0 .
1
1

In summary, in case Close-No war will occur at Z0 i c >

(1 p) p
(1 3p) 1

R.

Following similar steps, it is possible to show that case Far-Far and case Close-Close
are equivalent to case No-No.
Finally, consider case Far-Close, with B having oil close to the border. The peace in
all periods thresholds become
cA =
cB =

p
1
p
1

R; cA
0 =
R; cB
0 =

R; cA
+ =

1
p
1

R; cB
+ =

p
1
p
1

R;
R:

Clearly, B has always lower incentives to attack at Z0 than A given the assumptions
at the adjacent nodes. So we need to check the incentives of A to attack. Now compute
the indierence threshold of A at Z0 , assuming war at both adjacent states. The value of
war at Z0 is
W A (Z0 ) = c + p(uA ( Z ) + W A ( Z )) + p(uA (Z ) + W A (Z )) + (1

2p)(

where
W A( Z ) = c + p

W A (Z ) = c + p

uA (Z0 )
+ (1
1

2p)(uA ( Z ) + W A ( Z ));

2R
uA (Z0 )
+p
+ (1
1
1
10

2p)(uA (Z ) + W A (Z )):

uA (Z0 )
);
1

From W A (Z0 ) = V A (Z0 )) = (uA (Z0 ))=(1

c=

), we obtain
pR
:
(1 4p)

Again, we need to check for consistency that at

Z and Z B will indeed stage war,

given war at Z0 . Analogous computations yield that indeed B is indierent between war
and peace at

Z and Z in the presence of war at Z0 when

c=

1
(1 p)
(1 3p))(1

(1

pR:

Given that unambiguously

1
(1 p)
(1 3p))(1

(1

we conrm that indeed B stages war at

pR <

pR
;
(1 4p)

Z and Z , as assumed, when c >

1

Now compare the case Far-Close threshold c =

pR
(1 4p)

olds we found before:

pR
(1 4p)

>

1
1

(1 p) p
(1 3p) 1

pR
.
(1 4p)

pR
.
(1 4p)

In sum, in case Far-Close war occurs at Z0 i c >

1

for war at Z0 to the thresh-

R, which implies that war is more likely

in case Close-No than in case Far-Close.

In contrast, whether war is more likely in case Far-Close or case Far-No is ambiguous,
as one cannot sign

(1 )
(1 4p)

p
.
(1 3p)

End of Proof of Proposition 2.1

Note that the probability of war at time 0 in case Far-No can be higher or lower than
that in case Far-Close, depending on parameter values. The ambiguity is due to the fact
that case Far-No is endowment asymmetric, while case Far-Close is endowment symmetric
but has the close oil eect that is missing in case Far-No.
Beside proposition 2.1, which serves the purpose of a robustness check of the main
results of our static theoretical model, this dynamic extension allows to make some extra
predictions:
Corollary 2.1 Conditional on observing a war at time 0, the probability of another war
next period is higher if the rst period war is won by the attacker than when it is won by
the defender.

11

Given that the attacker is (in this equal strength and equal cost-benet benchmark)
always the country that is either oil less or at least has no oil close to border, while the
target country has oil, then if the attacker wins the other country will have roughly the
same incentives to attack back, whereas if the defender wins, then the border shifts away
from the oil, reducing the incentives to attack by the initial attacker.

Clustering at the Country Level

In the main body of the paper the standard errors are clustered at the country-pair level,
which is standard practice in literatures focusing on dyadic relationships (e.g. the vast
literature on gravity models in international trade, as well as of course much of the empirical
literature on international conict). One criticism of this approach (which applies to nearly
all contributions in those literatures as well) is that it treats two dierent dyads featuring
the same country as independent observations. This suggests that it could be preferable
to have some sort of country-level clustering.
In a gravity-trade setting, Cameron, Gelbach, and Miller (2011) propose to use their
multi-clustering approach, where the rst country in each pair is assigned to one cluster
and the second country to a second cluster. Spolaore and Wacziarg (2009) do the same in
an application where the dyadic outcome is the dierence in development between the two
countries.
As can be seen in Table 1 below, our results are very robust to applying the Cameron
et al. multi-clustering approach, treating country A and country B as dierent clusters.
Nevertheless, this approach is actually quite problematic. Whenever one uses a dyadic
data set, there will always be some countries that necessarily have to appear as country A
in some observations, and as country B in others. Clustering at the country A, country
B level will therefore create (for many countries) two separate clusters, one when the
country appears as country A, and one when it appears as country B. For this reason,
the multi-clustering approach is really not suitable to dyadic settings when one wants to
cluster at the country level.
Fafchamps and Gubert (2007) propose an alternative in the context of individual-level
data in a village-network setting. Their proposal looks more promising but even in the
network literature we couldnt nd other signicant papers that use it as the primary
12

method to estimate standard errors.4 Those authors who have used it, have tended to nd
massive losses of signicance using Fafchamps-Gubert clustering. Needless to say that the
technique is unheard of in the international context, even though it is potentially applicable
to the entirety of the enormous trade literature on gravity equations, where the problem of
each country appearing in multiple dyads also applies (as well as in the literature on war,
of course). We therefore feel that the Fafchamps-Gubert method has not been adequately
road testedby the literature for us to make the big leap of adopting it as our benchmark
method. But we report results using this method in Table 2 of this appendix. Consistent
with the sparse references to results using this method in the network literature, we also
nd a massive increase in estimated standard errors. However, even with this method some
of our core results survive, or remain very close to the margin of signicance.

Results with Region x Year Fixed Eects

Results with xed eects for world regions interacted with the year xed eects are reported
in Table 3.

Results with Cross-Section of Country Pairs

Results obtained after collapsing the panel into a simple cross-section of country pairs,
where all variable values now correspond to the averages over the 1946-2001 period, are
reported in Tables 4 (OLS estimates) and 5 (Poisson regressions).
References
Cameron, Colin, Jonah Gelbach, and Douglas Miller, 2011, Robust Inference with
Multi-way Clustering,Journal of Business and Economic Statistics 29, 238-249.
Fafchamps, Marcel and Flore Gubert, 2007, The Formation of Risk-Sharing Net-

We have checked Google Scholar for the rst 100 papers which cite the Fafchamps-Gubert paper.

Of these, only 11 actually use their method for their preferred specication. Of these, only 4 are not
(co-)authored by Fafchamps.

13

works,Journal of Development Economics 83, 326-50.

Spolaore, Enrico and Romain Wacziarg, 2009, The Diusion of Development,Quarterly Journal of Economics 124, 469-529.

14

15

(1)
0.034
(0.032)
-0.050*
(0.031)
0.022
(0.025)
-0.077**
(0.036)
0.026
(0.042)
All
No
No
19962
0.019

(2)
0.029
(0.029)
-0.044
(0.028)
0.028
(0.023)
-0.044
(0.035)
-0.014
(0.039)
All
No
Yes
11303
0.090

(3)
0.049
(0.030)
-0.073***
(0.026)
0.034
(0.032)
-0.105***
(0.036)
0.016
(0.032)
All
Yes
No
19962
0.145

(4)
0.077**
(0.033)
-0.086***
(0.032)
0.045*
(0.026)
-0.089**
(0.038)
0.004
(0.036)
All
Yes
Yes
11303
0.158

(9)
0.048
(0.043)
-0.079**
(0.039)
0.009
(0.029)
-0.102***
(0.036)
0.059
(0.043)
Onshore
No
No
19962
0.021

(10)
0.063
(0.040)
-0.072*
(0.040)
0.024
(0.024)
-0.096**
(0.039)
0.042
(0.045)
Onshore
No
Yes
11303
0.092

(11)
0.058
(0.037)
-0.103***
(0.033)
0.020
(0.037)
-0.122***
(0.034)
0.047
(0.034)
Onshore
Yes
No
19962
0.146

(12)
0.132***
(0.050)
-0.137***
(0.048)
0.047
(0.031)
-0.125***
(0.041)
0.043
(0.042)
Onshore
Yes
Yes
11303
0.160

Table 1: Country clustering following Cameron et al.

Note: The unit of observation is a country pair in a given year. The sample covers all contiguous country pairs and the years 1946-2001. Method: OLS with robust standard
errors clustered at the levels of "first-country in the dyad" and "second-country in the dyad". Significance levels *** p<0.01, ** p<0.05, * p<0.1. All explanatory variables
are taken as first lag. All specifications control for the average and the absolute difference of land areas in the pair, intercept and annual time dummies. Additional controls
are: The average and absolute difference of GDP per capita, the average and absolute difference of population, the average and absolute difference of fighting capabilities,
the average and absolute difference of democracy scores, dummy for one country having civil war, dummy for both countries having civil war, bilateral trade / GDP,
dummy for one country being OPEC member, dummy for both countries being OPEC member, genetic distance between the populations of the two countries, dummy for
membership in the same defensive alliance, dummy for historical inclusion in the same country, kingdom or empire, dummy for having been in a colonial relationship, and
years since the last hostility in the country pair.

Type Oil
Country FE
Add. Controls
Observations
R-squared

Both x MaxDist

Both x MinDist

Both

One x Dist

One

Dependent variable: Hostility

(5)
(6)
(7)
(8)
0.087*
0.081**
0.143**
0.124***
(0.051)
(0.041)
(0.063)
(0.047)
-0.107*
-0.087**
-0.138**
-0.118**
(0.056)
(0.043)
(0.065)
(0.046)
0.023
0.048*
0.110***
0.073**
(0.032)
(0.028)
(0.042)
(0.032)
-0.088**
-0.028
-0.107**
-0.066*
(0.043)
(0.031)
(0.052)
(0.034)
0.048
-0.048
0.012
-0.010
(0.060)
(0.046)
(0.068)
(0.046)
Offshore
Offshore
Offshore
Offshore
No
No
Yes
Yes
No
Yes
No
Yes
19962
11303
19962
11303
0.020
0.091
0.145
0.156

16

(1)
0.034
(0.033)
-0.050
(0.034)
0.022
(0.029)
-0.077**
(0.034)
0.026
(0.046)
All
No
No
19962
0.076

(2)
0.029
(0.030)
-0.044
(0.028)
0.028
(0.024)
-0.044
(0.037)
-0.014
(0.042)
All
No
Yes
11303
0.131

(3)
0.049
(0.034)
-0.073**
(0.032)
0.034
(0.034)
-0.105***
(0.038)
0.016
(0.036)
All
Yes
No
19962
0.195

(4)
0.077*
(0.040)
-0.086**
(0.038)
0.045
(0.032)
-0.089**
(0.041)
0.004
(0.042)
All
Yes
Yes
11303
0.196

(9)
0.048
(0.046)
-0.079*
(0.044)
0.009
(0.032)
-0.102***
(0.035)
0.059
(0.047)
Onshore
No
No
19962
0.078

(10)
0.063
(0.041)
-0.072*
(0.040)
0.024
(0.026)
-0.096**
(0.043)
0.042
(0.050)
Onshore
No
Yes
11303
0.133

(11)
0.058
(0.042)
-0.103**
(0.040)
0.020
(0.038)
-0.122***
(0.039)
0.047
(0.039)
Onshore
Yes
No
19962
0.196

(12)
0.132**
(0.059)
-0.137**
(0.057)
0.047
(0.034)
-0.125***
(0.047)
0.043
(0.052)
Onshore
Yes
Yes
11303
0.198

Table 2: Country clustering following Fafchamps and Gubert

Note: The unit of observation is a country pair in a given year. The sample covers all contiguous country pairs and the years 1946-2001. Method: OLS with robust standard
errors clustered at the country level. Significance levels *** p<0.01, ** p<0.05, * p<0.1. All explanatory variables are taken as first lag. All specifications control for the
average and the absolute difference of land areas in the pair, intercept and annual time dummies. Additional controls are: The average and absolute difference of GDP per
capita, the average and absolute difference of population, the average and absolute difference of fighting capabilities, the average and absolute difference of democracy
scores, dummy for one country having civil war, dummy for both countries having civil war, bilateral trade / GDP, dummy for one country being OPEC member, dummy for
both countries being OPEC member, genetic distance between the populations of the two countries, dummy for membership in the same defensive alliance, dummy for
historical inclusion in the same country, kingdom or empire, dummy for having been in a colonial relationship, and years since the last hostility in the country pair.

Type Oil
Country FE
Add. Controls
Observations
R-squared

Both x MaxDist

Both x MinDist

Both

One x Dist

One

Dependent variable: Hostility

(5)
(6)
(7)
(8)
0.087
0.081*
0.143**
0.124**
(0.055)
(0.043)
(0.068)
(0.052)
-0.107*
-0.087*
-0.138**
-0.118**
(0.060)
(0.045)
(0.068)
(0.050)
0.023
0.048*
0.110**
0.073**
(0.032)
(0.027)
(0.044)
(0.034)
-0.088***
-0.028
-0.107**
-0.066*
(0.025)
(0.034)
(0.052)
(0.034)
0.048
-0.048
0.012
-0.010
(0.046)
(0.045)
(0.069)
(0.047)
Offshore
Offshore
Offshore
Offshore
No
No
Yes
Yes
No
Yes
No
Yes
19962
11303
19962
11303
0.077
0.132
0.195
0.193

17

(2)
0.025
(0.026)
-0.045*
(0.026)
0.019
(0.019)
-0.047
(0.031)
-0.011
(0.033)
All
No
Yes
11303
0.121

(3)
0.027
(0.028)
-0.071***
(0.027)
0.001
(0.030)
-0.105***
(0.032)
0.017
(0.031)
All
Yes
No
19962
0.164

(4)
0.065**
(0.028)
-0.092***
(0.026)
0.015
(0.025)
-0.085***
(0.029)
-0.000
(0.029)
All
Yes
Yes
11303
0.186

(5)
0.072
(0.048)
-0.107**
(0.050)
0.003
(0.036)
-0.088
(0.054)
0.040
(0.073)
Offshore
No
No
19962
0.060

Dependent variable: Hostility

(6)
(7)
(8)
0.083**
0.132*** 0.126***
(0.037)
(0.044)
(0.029)
-0.087** -0.130*** -0.107***
(0.040)
(0.044)
(0.028)
0.044*
0.094*** 0.085***
(0.026)
(0.034)
(0.028)
-0.038
-0.110**
-0.059*
(0.034)
(0.052)
(0.032)
-0.032
0.025
-0.006
(0.048)
(0.064)
(0.041)
Offshore
Offshore
Offshore
No
Yes
Yes
Yes
No
Yes
11303
19962
11303
0.122
0.162
0.182
(9)
0.036
(0.040)
-0.080**
(0.040)
-0.002
(0.026)
-0.116***
(0.035)
0.051
(0.039)
Onshore
No
No
19962
0.064

(10)
0.057
(0.036)
-0.071*
(0.037)
0.014
(0.020)
-0.098**
(0.040)
0.043
(0.043)
Onshore
No
Yes
11303
0.124

(11)
0.040
(0.035)
-0.100***
(0.034)
-0.004
(0.032)
-0.124***
(0.034)
0.048
(0.034)
Onshore
Yes
No
19962
0.165

(12)
0.130***
(0.043)
-0.148***
(0.040)
0.034
(0.031)
-0.122***
(0.035)
0.039
(0.037)
Onshore
Yes
Yes
11303
0.189

Table 3: Baseline results with region*year xed eects

Note: The unit of observation is a country pair in a given year. The sample covers all contiguous country pairs and the years 1946-2001. Method: OLS with robust
standard errors clustered at the country-pair level. Significance levels *** p<0.01, ** p<0.05, * p<0.1. All explanatory variables are taken as first lag. All specifications
control for the average and the absolute difference of land areas in the pair, intercept and region*year fixed effects. Additional controls are: The average and absolute
difference of GDP per capita, the average and absolute difference of population, the average and absolute difference of fighting capabilities, the average and absolute
difference of democracy scores, dummy for one country having civil war, dummy for both countries having civil war, bilateral trade / GDP, dummy for one country being
OPEC member, dummy for both countries being OPEC member, genetic distance between the populations of the two countries, dummy for membership in the same
defensive alliance, dummy for historical inclusion in the same country, kingdom or empire, dummy for having been in a colonial relationship, and years since the last
hostility in the country pair.

Type Oil
Country FE
Add. Controls
Observations
R-squared

Both x MaxDist

Both x MinDist

Both

One x Dist

One

(1)
0.020
(0.029)
-0.052
(0.032)
0.007
(0.022)
-0.084***
(0.032)
0.014
(0.036)
All
No
No
19962
0.061

18

Table 4: Baseline results for the simple cross-section of country pairs using OLS

Dependent variable: Share of years with hostility

(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
One
0.064
0.090
0.087
0.217
0.101**
0.099**
0.145*
0.259***
0.082
0.137
0.129
0.273
(0.052)
(0.073)
(0.080)
(0.139)
(0.051)
(0.047)
(0.080)
(0.097)
(0.069)
(0.092)
(0.102)
(0.168)
One x Dist
-0.081
-0.119
-0.113**
-0.169*
-0.124**
-0.117**
-0.182*** -0.196***
-0.104
-0.148
-0.145*
-0.209*
(0.054)
(0.074)
(0.057)
(0.091)
(0.053)
(0.049)
(0.060)
(0.063)
(0.072)
(0.096)
(0.075)
(0.113)
Both
0.022
0.015
0.036
0.174
0.003
0.037
0.021
0.206
0.012
0.020
0.085
0.206
(0.018)
(0.019)
(0.120)
(0.185)
(0.027)
(0.032)
(0.116)
(0.143)
(0.019)
(0.016)
(0.139)
(0.217)
Both x MinDist
-0.043
-0.059*
-0.092**
-0.114**
-0.065*
-0.065*
-0.058
-0.071*
-0.067*
-0.099**
-0.102**
-0.139***
(0.033)
(0.034)
(0.039)
(0.045)
(0.038)
(0.034)
(0.049)
(0.043)
(0.035)
(0.039)
(0.042)
(0.052)
Both x MaxDist
0.004
-0.001
0.006
0.020
0.046
-0.018
-0.040
-0.024
0.026
0.035
0.028
0.052
(0.036)
(0.034)
(0.040)
(0.046)
(0.049)
(0.047)
(0.062)
(0.053)
(0.038)
(0.039)
(0.044)
(0.053)
Type Oil
All
All
All
All
Offshore
Offshore
Offshore
Offshore
Onshore
Onshore
Onshore
Onshore
Country FE
No
No
Yes
Yes
No
No
Yes
Yes
No
No
Yes
Yes
Add. Controls
No
Yes
No
Yes
No
Yes
No
Yes
No
Yes
No
Yes
Observations
596
411
596
411
596
411
596
411
596
411
596
411
R-squared
0.046
0.282
0.548
0.611
0.051
0.275
0.545
0.600
0.051
0.296
0.549
0.614
Note: The unit of observation is a country pair. The sample covers all contiguous country pairs. Method: OLS with robust standard errors. Significance levels *** p<0.01, **
p<0.05, * p<0.1. All specifications control for the average and the absolute difference of land areas in the pair and intercept. Additional controls are: The average and
absolute difference of GDP per capita, the average and absolute difference of population, the average and absolute difference of fighting capabilities, the average and
absolute difference of democracy scores, dummy for one country having civil war, dummy for both countries having civil war, bilateral trade / GDP, dummy for one country
being OPEC member, dummy for both countries being OPEC member, genetic distance between the populations of the two countries, dummy for membership in the same
defensive alliance, dummy for historical inclusion in the same country, kingdom or empire, and dummy for having been in a colonial relationship.

19

Table 5: Baseline results for the simple cross-section of country pairs using Poisson

Dependent variable: Share of years with hostility

(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
One
1.046
1.233*
0.440
1.626
1.324***
1.294***
1.878***
1.800
1.192
1.768***
1.520
4.588**
(0.655)
(0.663)
(1.058)
(2.191)
(0.455)
(0.366)
(0.616)
(2.016)
(0.730)
(0.648)
(1.144)
(2.247)
One x Dist
-1.417**
-1.767*** -1.889*** -3.351*** -1.748*** -1.660*** -2.826*** -3.310***
-1.664**
-1.749*** -2.172*** -3.889***
(0.682)
(0.603)
(0.605)
(1.004)
(0.488)
(0.320)
(0.495)
(0.803)
(0.781)
(0.621)
(0.641)
(1.202)
Both
0.401
0.289
-0.658
-1.451
0.073
1.135**
1.178
-0.001
0.206
0.415
1.497
2.814
(0.325)
(0.335)
(1.646)
(3.856)
(0.420)
(0.535)
(1.152)
(3.663)
(0.323)
(0.295)
(1.865)
(3.817)
Both x MinDist
-0.751
-1.594*** -1.662*** -2.794***
-0.947**
-1.181**
-1.091
-2.024*
-1.139**
-2.230*** -1.653*** -3.707***
(0.503)
(0.436)
(0.518)
(0.751)
(0.450)
(0.501)
(0.766)
(1.177)
(0.480)
(0.466)
(0.501)
(0.646)
Both x MaxDist
0.119
0.373
-0.244
0.880
0.631
-0.870
-1.789*
-1.507
0.425
0.997**
0.113
2.180***
(0.495)
(0.415)
(0.498)
(0.787)
(0.604)
(0.622)
(0.930)
(1.086)
(0.491)
(0.448)
(0.500)
(0.687)
Type Oil
All
All
All
All
Offshore
Offshore
Offshore
Offshore
Onshore
Onshore
Onshore
Onshore
Country FE
No
No
Yes
Yes
No
No
Yes
Yes
No
No
Yes
Yes
Add. Controls
No
Yes
No
Yes
No
Yes
No
Yes
No
Yes
No
Yes
Observations
596
411
596
411
596
411
596
411
596
411
596
411
Log pseudolikelihood
-107.1
-61
-78.2
-50.5
-107
-61.4
-78.7
-50.8
-106.8
-60.8
-78.2
-50.5
Note: The unit of observation is a country pair. The sample covers all contiguous country pairs. Method: Poisson regression with robust standard errors. Significance levels *** p<0.01,
** p<0.05, * p<0.1. All specifications control for the average and the absolute difference of land areas in the pair and intercept. Additional controls are: The average and absolute
difference of GDP per capita, the average and absolute difference of population, the average and absolute difference of fighting capabilities, the average and absolute difference of
democracy scores, dummy for one country having civil war, dummy for both countries having civil war, bilateral trade / GDP, dummy for one country being OPEC member, dummy for
both countries being OPEC member, genetic distance between the populations of the two countries, dummy for membership in the same defensive alliance, dummy for historical
inclusion in the same country, kingdom or empire, and dummy for having been in a colonial relationship.