Choice under Uncertainty

Bibhas Saha
Durham University Business School

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Uncertainty
Value of your stock or investment may go up or go down. This is an uncertain
event, over which you have no control. Economic agents need to make a

decision before the uncertainty is resolved. We need a theory to explain this

decision making.

The most commonly used theory is called Expected utility theory due to von
Neumann and Morgenstern.
Two alternative theories are prospect theory and regret theory.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Expected utility approach
Suppose there are n states of the world, leading to n possible outcomes or
consequences of an event as (x1 , x2 , ..., x
P n ), which occur with probabilities
(p1, p2, ..., pn), pi = 1.

An event can be thought of as a lottery or a gamble. Let the agent’s utility

associated with each of these outcomes be u(xi ). If the consumer’s utility over
this event is
n
X
Eu = p1u(x1) + p2u(x2) + ... + pnu(xn) = piu(xi),
i=1

then we say that the agent is having a von Neumann-Morgenstern utility

function.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Attitude to risk
A key element of this approach is to recognise the agent’s attitude to risk. She
could be risk averse, risk-lover or risk neutral. The agent is defined to be
risk averse, iff
n
X
u(p1x1 + ... + pnxn) ≥ piu(xi), or u(Ex) ≥ Eu.
i=1

Alternatively, the agent is defined to be risk lover, iff

n
X
u(p1x1 + ... + pnxn) ≤ piu(xi), or u(Ex) ≤ Eu,
i=1

and risk neutral iff

n
X
u(p1x1 + ... + pnxn) = piu(xi), or u(Ex) = Eu.
i=1

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Risk aversion
Clearly, risk aversion implies (and is implied by) concavity of the utility
function u(.). Similarly, risk loving implies (and is implied by) convex u(.)
function and risk neutrality implies linear u(.) function.

In Figure 1 there are two graphs showing risk aversion and risk loving.
Suppose there is a lottery promising $5 with probability 1/2 and $15 with
probability 1/2. Expected prize is $10. In panel a the expected utility of the
lottery is less than the utility of the expected prize; in panel b it is reversed.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Risk aversion and Risk loving

u u u
u(15) u(15)
u(10)
Eu
u(5) Eu

u(10)
u(5)

£5 £10 £15 £5 £10 £15 Prize

Prize

Fig 1: (a) Risk aversion, (b) Risk loving

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Measures of risk aversion
Intuitively, the more concave the expected utility function, the more risk averse
the consumer. A well-known measure is Arrow-Pratt measure of risk
aversion:
u00(x)
r(x) = − 0 .
u (x)
This is also known as the measure of absolute risk aversion.

There is a related measure called relative risk aversion:

xu00(x)
ρ(x) = − 0 .
u (x)

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
Often we use utility functions that display constant absolute risk aversion
(CARA) or constant relative risk aversion (CRRA)

Example of CARA utility function: u(x) = 1r [1 − e−rx ].

Example of CRRA utility function: u(x) = xa , (a < 1)

(Verify yourself by taking first and second derivatives and using the ARA and
RRA definitions.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
Options
A risk averse individual has three options when facing uncertainty over an asset
(or any decision):

• Do nothing and live with risk (default option)

• Sell the asset (risk avoidance)

• Seek insurance against potential loss (buying insurance from market, or other
organisations/society)

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Pay risk premium to avoid risk
A risk averse agent would strictly prefer to have a certain payment
(appropriately set) over a lottery.

Suppose a lottery promises x1 with probability p and x2 with probability

(1 − p).
The expected income of the lottery is px1 + (1 − p)x2 . Expected utility is
pu(x1) + (1 − p)u(x2).

Define ‘Certainty Equivalent’: u(CE) = Eu(x)

See Fig. 2.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Risk premium
u
u(px1+(1-p)x2)

pu(x1)+(1-p)v(x2)

x
x1 CE px1+(1-p)x2 x2

Fig. 2: Risk Aversion

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Risk premium
This expected utility can also be achieved at a certainty amount CE . So if the
agent is given CE (or more), then she should be willing to give up the lottery.

risk premium = expected payoff - certainty equivalent=

px1 + (1 − p)x2 − CE .
Generally, risk premium increases with the variance of a gamble and the
degree of risk aversion of the ’gambler’

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
Relation between risk aversion and risk pre-
mium
There is a clear connection between risk premium and absolute risk aversion.
To see this,
Let µ be the expected value of a gamble: E(x) = µ (mean) and σ 2 be the
expected value of (x − µ)2 (variance). Then we can approximate u(CE) in
two different ways. First, we can write (by Taylor’s approximation)

u(CE) ≈ u(µ) + u0(µ)(CE − µ).

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
Relation between risk aversion and risk pre-
mium
Again use Taylor expansion to write Eu as
1
Eu(x) ≈ E{u(µ) + u0(µ)(x − µ) + u00(µ)(x − µ)2}.
2

Recall by the definition of certainty equivalence u(CE) = Eu(x).

Setting two Taylor expansions being approximately equal to each other,

1
u(µ) + u0(µ)(CE − µ) ≈ E{u(µ) + u0(µ)(x − µ) + u00(µ)(x − µ)2}
2
1
u(µ) + u0(µ)(CE − µ) ≈ Eu(µ) + u0(µ)E(x − µ) + u00(µ)E{(x − µ)2}
2

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
Relation between risk aversion and risk pre-
mium
1
u(µ) + u0(µ)(CE − µ) ≈ u(µ) + 0 + u00(µ)σ 2,
2
00
u (µ) 1
µ − CE ≈ − 0 σ 2 = r(µ)σ 2.
2u (µ) 2
The left hand side expression represents risk premium, which will be approx-
imately equal to half of the variance multiplied by the degree of absolute risk
aversion.

We can rearrange the terms and write:

1
CE = µ + r(µ)σ 2.
2

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
Normal distribution gambles and risk premium
Suppose x ∈ (−∞, +∞) and x follows normal distribution with (µ, σ 2 ) as
mean and variance respectively. Note σ 2 is constant for a normal distributrion.

Then we can write :

00
u (µ) 2
µ − CE = − 2u0 (µ) σ

Substituting the Arrow-Pratt measure of absolute risk aversion r(µ) =

−u00(µ)/u0(µ) we can write

µ − CE = 12 r(µ)σ 2.

Risk premium will increase with the degree of risk aversion r(.) and variance
of the gamble.
If the degree of risk aversion is decreasing in wealth (i.e. r0 (µ) < 0), the risk
premium will also be declining in wealth.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 CARA utility function and risk premium
In the case of constant variance gambles (for example, under normal
distribution) and agent’s constant absolute risk aversion (CARA) utility
function it can be shown that the agent selects from a set of gambles the one
that offers the highest certainty equivalent.

Suppose a gamble’s returns are normally distributed with mean µ and variance
σ 2 (σ 2 being constant) and the agent has the CARA utility function
u(x) = 1r [1 − e−rx].

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 CARA utility function and risk premium
Since Eu(x) = u(CE) ( i.e. the expected utility of a gamble is equal to the
utility of the certainty equivalent of the gamble), we an write

1h −r(CE)
i
Eu(x) = u(CE) = 1 − e ,
r
1h −r(µ− 2r σ 2 )
i
= 1−e .
r

Then the risk averse agent will pick the gamble that offers highest CE ,
because at that gamble the expected utility becomes maximum. This is also
known as the mean-variance preference, widely used in finance and agency
theories.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Application: Buying insurance
Suppose a risk averse consumer has wealth W which may lose its value by L
with probability p (bad state). The consumer can buy insurance with a cover D
for which she has to pay premium γD, 0 < γ < 1.

Her expected utility from buying the insurance is

Eu = pu(W − L + D − γD) + (1 − p)u(W − γD).

Question is: What is the optimal D?

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Application: Buying insurance
The first order condition is:

pu0(W − L + D(1 − γ))(1 − γ) − (1 − p)γu0(W − γD) = 0,

pu0(W − L + (1 − γ)D)
or = γ.
pu0(W − L + D(1 − γ)) + (1 − p)u0(W − γD)

The optimal cover should be such that the ratio of the expected marginal utility
in the bad state to the overall expected marginal utility is equal to the rate of
premium γ .

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Application: Buying insurance
Now, determine γ . Assume: the insurance market is competitive and firms
earn zero expected profit. That is, γD = pD or equivalently

Eπ = −p(1 − γ)D + (1 − p)(γD) = 0, or γ = p.

This is called the actuarially fair insurance premium. When this rate is
substituted into the consumer’s optimal cover equation, we get

u0(W − L + (1 − γ)D) = u0(W − γD) or D = L.

The consumer will demand ‘full coverage’ so that her

marginal utilities are equalized across all states of nature.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Insurance:Graph
Let WG denote the consumer’s wealth in the good state and WB be the same in
bad state.
If she does not buy any insurance, then she has a probabilistic endowment of
(W, W − L).

If she buys a cover D0 , then her future wealth position changes to

((W − γD0), (w − L + D0(1 − γ)).

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Insurance: Graph

WB Eu

Slope =
-(1-p)/p WG=WB

W-L+D’(1-γ)

W-γL Optimal choice

Endowment point
W-L

W-D’γ W-γL W WG

Fig. 3: Insurance choice

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Insurance: Explaining the graph
The straight line is like a budget line.
It captures all the possible future wealth positions (that may result from
insurance) such that the insurance premium rate is actuarially fair.

That is, it represents the set

{((W − γD), (W − L + D(1 − γ)) : (1 − p)Dγ − pD(1 − γ) = 0}.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Insurance: Explaining the graph
Note that WG = W − γD; so we can write γD = W − WG ,
and similarly from WB = W − L + D(1 − γ) we can write
D(1 − γ) = WB − W + L.

Substitute these expressions into the zero profit equation:

(1 − p)Dγ − pD(1 − γ) = 0. This is then rewritten as,
W − pL − [pWB + (1 − p)WG] = 0.
This line has a slope dWB
dWG
= − 1−p
p
.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Insurance: Explaining the graph
Given this budget line, the consumer maximizes her expected utility, and we
know that the optimal point would be such that D = L.

The result of complete insurance depends on two crucial assumptions: (i) the
premium rate is actuarially fair (i.e. γ = p) and (ii) the probability of the bad
state (p) is beyond the control of the consumer.

The first assumption is violated if the insurance market is not competitive, or if

the insurance market suffers from asymmetric information.

The second assumption is violated when the agent’s action alters the risk: ex:
car insurance, or medical insurance.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Axioms of Expected utility theory (optional)
A lottery (or a gamble) will represent a probability distribution over these n
consequences.

Lottery L1 may be denoted as, L1 : p1 ◦ x1 ⊕ p2 ◦ x2 ⊕ ... ⊕ pn ◦ xn .

Similarly, define another lottery L01 as L1 : p01 ◦ x1 ⊕ p02 ◦ x2 ⊕ ... ⊕ p0n ◦ xn .

We can indicate a lottery by a point on the probability simplex.

Lotteries L1 and L01 are also called prospects (but need to be written
differently).

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Compound lottery (optional)
The lotteries L1 and L01 are called simple lotteries. A lottery L001 is called a
compound lottery if L001 : α ◦ L1 ⊕ (1 − α) ◦ L01 .

Then we can write L001 as a reduced lottery:

L001 : [αp1 + (1 − α)p01] ◦ x1 ⊕ ... ⊕ [αpn + (1 − α)p0n] ◦ xn.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Consumer’s preference over the lottery space
(optional)
We assume that the consumer has a well-defined preference over the lottery
space.

To start with, she does not care whether a lottery is a simple lottery or a
compound lottery. She cares only about the net probabilities.

Suppose
L001 : [αp1 + (1 − α)p01] ◦ x1 ⊕ ... ⊕ [αpn + (1 − α)p0n] ◦ xn.
and if L2 : q1 ◦ x1 ⊕ q2 ◦ x2 ⊕ ... ⊕ qn ◦ xn and if qi = αpi + (1 − α)p0i for
i = 1, 2, ..., n,
then she is indifferent between L2 and L001 .

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Axioms of preference (optional)
1. Completeness: For any pair of simple lotteries (L1 , L2 ) the consumer
prefers L1 to L2 , or L2 to L1 or both.
That is, either L1  L2 or L2  L1 or both.

That is, the consumer can rank every possible lottery.

2. Transitivity: If L1  L2 and L2  L3 , then L1  L3 .

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Axioms of preference (optional)
Continuity: The preference relation  on the space of simple lotteries is
continuous if for any L1 , L2 and L3 , the following two sets are closed:

{α ∈ [0, 1] : α ◦ L1 ⊕ (1 − α) ◦ L2  L3}
and

{α ∈ [0, 1] : L3  α ◦ L1 ⊕ (1 − α)L2} .
In words, A small change in the probabilities does not change the nature of the
ordering of two lotteries. If a compound lottery of L1 and L2 is preferred to
L3, then any other compound lottery sufficiently close to this lottery will also
be preferred to L3 .

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Axioms of preference (optional)
Independence axiom: The preference relation  satisfies the independence
axiom if for all (L1 , L2 , L3 ) and α ∈ (0, 1) we have L1  L2 if and only if

α ◦ L1 ⊕ (1 − α) ◦ L3  α ◦ L2 ⊕ (1 − α) ◦ L3.

Ranking of two lotteries must be preseved with respect to a third lottery.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 The expected utility theorem
Suppose the preference relation  satisfies axioms 1-4. Then we can assign a
number ui to each outcome x1 , x2 , ..., xn such that for any two lotteries
0 0 0 0 0
L = (p1, p2, ..., pn) and
Pn L = (p 1 , p2 ,
Pn 0 ..., p n ) , L  L if and only if
i=1 pi ui ≥ i=1 pi ui .
Pn Pn
In other words, U (L) = i=1 pi ui and U (L0 ) = i=1 p0i ui , and
U (L) ≥ U (L0) implies L  L0.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Comment on the independence axiom
Axioms of completeness, transitivity and continuity ensure that the expected
utility function exists.

The independence axiom imposes some order in the preference, namely the
shape of ’indifference curves’ on the probability space.

But the independence axiom is not very intuitive and is regarded controversial.
Unfortunately it is also central to the expected utility theory.

The following is an example where the independence axiom is violated (in

reality).

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Allais Paradox
Suppose there are three prizes: x1 = 0, x2 = $50, 000, x3 = 200, 000; of
course u(x3 ) > u(x2 ) > u(x1 )

and there are four lotteries over these prizes:

L1 = (0, 1, 0) and L01 = (0.01, 0.89, 0.10).,

L2 = (0.89, 0.11, 0) and L02 = (0.90, 0, 0.10).
Note that L01 = L1 + (0.01, −0.11, 0.10) and L02 = L2 + (0.01, −0.11, 0.10).

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 The Allais paradox
If u(L1 ) > u(L01 ) then by the expected utility theory we should also have
u(L2) > u(L02).
EU of L1 = u(x2), and EU of L01 = 0.01u(x1) + 0.89u(x2) + 0.10u(x3).
Suppose the EU of L1 > EU of L01 . That is,
u(50, 000) > 0.01u(0) + 0.89u(50, 000) + 0.10u(200, 000).
Add 0.89u(0) − 0.89u(50, 000) on both sides and get
0.89u(0) + 0.11u(50, 000) > 0.90u(0) + 0.10u(200, 000).
That is, u(L2 ) > u(L02 ).

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
The Allais paradox
Prediction of the expected utility theory: if L1  L01 , L2 must be preferred to
L02.
The reason for preferring L1 is clear because of the certainty payment of
$50, 000.
In the second choice the chance of getting nothing is very high whether it is
L2 or L02; but swapping the (near equal) probabilities between $50, 000 and
$200, 000 makes L02 more attractive than L2. Most people in reality would
choose L02 than L2 , in contradiction to the prediction of the expected utility
theory.

The unreasonable prediction of the expected utility theory comes from the in-
dependence axiom.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Reaction to the Allais paradox
First, Allais paradox is of limited significance because it involves out of the
ordinary payoffs and probabilities close to 0 and 1.

Second, introduce the idea of regret.

The regret of getting nothing from L01 against the assured prize of 50,000 in L1
justifies L1  L01 . On the other hand, with the choice between L2 and L02 no
such clear-cut regret potential exists. So L02  L2 is clearly admissible and
consistent.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Reaction to the Allais paradox
Third, introduce the idea of disappointment.

Suppose there are three outcomes: ‘a trip to Grand Canyon’, ‘watching a

movie about Grand Canyon’ and ‘staying home’.

Suppose there are two lotteries. The first lottery consists of ‘going to Grand
Canyon’ with 99.9% probability and ‘watching a movie about Grand Canyon’
with probability 0.1%. The second lottery consists ‘going to Grand Canyon’
with probability 99.9% and ‘staying home’ with probability 0.1%.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Reaction to the Allais paradox
The independence axiom forces one to prefer the first lottery over the second.
Yet, one might choose the second lottery. One may feel so disappointed after
not getting ‘a trip to Grand Canyon’ that one may prefer ‘staying home’ to
‘watching the movie about Grand Canyon’.

In the context of the Allais paradox the choice of L02 over L2 might be justified
if the consumer is likely to feel disappointed after choosing L2 , if $50, 000
comes up - disappointed for not choosing L02 .

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Non-expected utility theories for decision
making under uncertainty
Prospect theory: utility representations are obtained by assigning subjective
probabilities to the outcomes, instead of assigning the objective probabilities.

Consider the set of outcomes (x1 , x2 , ..., xn ) and the agent has the following
preference over the outcomes: x1 ≺ x2 ≺ ... ≺ xn . The agent’s utility over
the lottery L = (p1 , p2 , ..., pn ) takes the form:

U (L) = π(p1)u(x1) + π(p2)u(x2) + ... + π(pn)u(xn)

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit
 Prospect theory
Assignment of probabilities: π(pi ) is the subjective probability of the outcome
xi and this subjective probability depends on the objective probability pi in a
particular way. Typically,

π(pi) > pi, if pi < 1/2

= pi if pi = 1/2
< pi if pi > 1/2.
This is justified by saying that agents are less used to evaluating very low or
very high probability events for which they might feel anxious.

Such probability assignments are observed in betting markets, which lead to a

betting pattern called favourite-longshot bias.

•First •Prev •Next •Last •Go Back •Full Screen •Close •Quit