Expected Utility Theory

Mark Dean

GR5211 - Microeconomic Analysis 1

 Introduction

Up until now, we have thought of subjects choosing between

objects
Used cars
Hamburgers
Monetary amounts
However, often the outcome of the choices that we make are
not known
You are deciding whether or not to buy a share in AIG
You are deciding whether or not to put your student loan on
black at the roulette table
You are deciding whether or not to buy a house that straddles
the San Andreas fault line
In each case you understand what it is that you are choosing
between, but you don’t know the outcome of that choice
In fact, many things can happen, you just don’t know which
one
 Risk vs Uncertainty

We are going to di¤erentiate between two di¤erent ways in

which the future may not be know
Horse races
Roulette wheels
What is the di¤erence?
 Risk vs Uncertainty

When playing a roulette wheel the probabilities are known

Everyone agrees on the likelihood of black
So we (the researcher) can treat this as something we can
observe
Probabilities are objective
This is a situation of risk
 Risk vs Uncertainty

When betting on a horse race the probabilities are unknown

Di¤erent people may apply di¤erent probabilities to a horse
winning
We cannot directly observe a person’s beliefs
Probabilities are subjective
This is a situation of uncertainty (or ambiguity)
 Choices Under Risk

So, how should you make choices under risk?

Let’s consider the following (very boring) fairground game
You ‡ip a coin
If it comes down heads you get $10
If it comes down tails you get $0
What is the maximum amount x that you would pay in order
to play this game?
 Approach 1: Expected Value

You have the following two options

1 Not play the game and get $0 for sure
2 Play the game and get $x with probability 50% and $10 x
with probability 50%
Approach 1: Expected value
The expected amount that you would earn from playing the
game is
0.5( x ) + 0.5(10 x )
This is bigger than 0 if

0.5( x ) + 0.5(10 x) 0
5 x

Should pay at most $5 to play the game

 The St. Petersburg Paradox

This was basically the accepted approach until Daniel

Bernoulli suggested the following modi…cation of the game
Flip a coin
If it comes down heads you get $2
If tails, ‡ip again
If that coin comes down heads you get $4
If tails, ‡ip again
If that comes down heads, you get $8
Otherwise ‡ip again
and so on
How much would you pay to play this game?
 The St. Petersburg Paradox

The expected value is

1 1 1 1
$2 + $4 + $8 + $16 + ...
2 4 8 16
= $1 + $1 + $1 + $1 + ......
= ∞

So you should pay an in…nite amount of money to play this

game
Which is why this is the St. Petersburg paradox
 The St. Petersburg Paradox

So what is going wrong here?

Consider the following example:

Example
Say a pauper …nds a magic lottery ticket, that has a 50% chance
of $1 million and a 50% chance of nothing. A rich person o¤ers to
buy the ticket o¤ him for $499,999 for sure. According to our
‘expected value’method’, the pauper should refuse the rich
person’s o¤er!
 The St. Petersburg Paradox

It seems ridiculous (and irrational) that the pauper would

reject the o¤er
Why?
Because the di¤erence in life outcomes between $0 and
$499,999 is massive
Get to eat, buy clothes, etc
Whereas the di¤erence between $499,999 and $1,000,000 is
relatively small
A third pair of silk pyjamas
Thus, by keeping the lottery, the pauper risks losing an awful
lot ($0 vs $499,999) against gaining relatively little ($499,999
vs $1,000,000)
 Marginal Utility

Bernoulli argued that people should be maximizing expected

utility not expected value
u (x ) is the expected utility of an amount x
Moreover, marginal utility should be decreasing
The value of an additional dollar gets lower the more money
you have
For example

u ($0) = 0
u ($499, 999) = 10
u ($1, 000, 000) = 16
 Marginal Utility

Under this scheme, the pauper should choose the rich person’s
o¤er as long as
1 1
u ($1, 000, 000) + u ($0) < u ($499, 999)
2 2
Using the numbers on the previous slide, LHS=8, RHS=10
Pauper should accept the rich persons o¤er
Bernoulli suggested u (x ) = ln(x )
Also explains the St. Petersberg paradox
Using this utility function, should pay about $64 to play the
game
 Risk Aversion

Notice that if people

Maximize expected utility
Have decreasing marginal utility (i.e. utility is concave)
They will be risk averse
Will always reject a lottery in favor of receiving its expected
value for sure
Risk Aversion
 Expected Utility

Expected Utility Theory is the workhorse model of choice

under risk
Unfortunately, it is another model which has something
unobservable
The utility of every possible outcome of a lottery
So we have to …gure out how to test it
We have already gone through this process for the model of
’standard’(i.e. not expected) utility maximization
Is this enough for expected utility maximization?
 Data

In order to answer this question we need to state what our

data is
We are going to take as our primitve preferences
Not choices
But we know how to go from choices to preferences, yes?
But preferences over what?
In the beginning we had preferences over ‘objects’
For temptation and self control we used ’menus’
Now ‘lotteries’!
 Lotteries

What is a lottery?
Like any lottery ticket, it gives you a probability of winning a
number of prizes
Let’s imagine there are four possible prizes
a(pple), b(anana), c(elery), d(ragonfruit)
Then a lottery is just a probability distribution over those
prizes 0 1
0.15
B 0.35 C
B C
@ 0.5 A
0
This is a lottery that gives 15% chance of winning a, 35%
chance of winning b, 50% of winning c and 0% chance of
winning d
 Lotteries

More generally, a lottery is any

0 1
pa
B pb C
p=B C
@ pc A
pd

Such that
px 0
∑ x px = 1

De…nition
Let X be some …nite prize space, The set ∆(X ) of lotteries on X is
the set of all functions p : X ! [0, 1] such that

∑ p (x ) = 1
x 2X
 Expected Utility

We say that preferences have an expected utility

representation if we can
Find utilities on prizes
i.e. u (a), u (b ), u (c ), u (d )
Such that
p q if and only if

pa u (a) + pb u (b ) + pc u (c ) + pd u (d )
> qa u (a) + qb u (b ) + qc u (c ) + qd u (d )

i.e ∑x px u (x ) ∑x qx u (x )
 Expected Utility

De…nition
A preference relation on lotteries on some …nite prize space X
have an expected utility representation if there exists a function
u : X ! R such that

p q if and only if
∑ p (x )u (x ) ∑ q (x )u (x )
x 2X x 2X

Notice that preferences are on ∆(X ) but utility numbers are

on X
 Expected Utility

What needs to be true about preferences for us to be able to

…nd an expected utility representation?
Hint: you know a partial answer to this
An expected utility representation is still a utility
representation
So preferences must be
Complete
Transitive
Re‡exive
 Expected Utility

Unsurprisingly, this is not enough

We need two further axioms
1 The Independence Axiom
2 The Archimedian Axiom
 The Independence Axiom

Question: Think of two di¤erent lotteries, p and q. Just for

concreteness, let’s say that p is a 25% chance of
winning the apple and a 75% chance of winning the
banana, while q is a 75% chance of winning the
apple and a 25% chance of winning the banana. Say
you prefer the lottery p to the lottery q. Now I o¤er
you the following choice between option 1 and 2
1 I ‡ip a coin. If it comes up heads, then you get
p. Otherwise you get the lottery that gives you
the celery for sure
2 I ‡ip a coin. If it comes up heads, you get q.
Otherwise you get the lottery that gives you the
celery for sure
Which do you prefer?
 The Independence Axiom

The independence axiom says that if you must prefer p to q

you must prefer option 1 to option 2
If I prefer p to q, I must prefer a mixture of p with another
lottery to q with another lottery
The Independence Axiom Say a consumer prefers lottery p to
lottery q. Then, for any other lottery r and number
0 < α 1 they must prefer

αp + (1 α )r

to
αq + (1 α )r

Notice that, while the independence axiom may seem intutive,

that is dependent on the setting
Maybe you prefer ice cream to gravy, but you don’t prefer ice
cream mixed with steak to gravy mixed with steak
 The Archimedean Axiom

The other axiom we need is more techincal

It basically says that no lottery is in…nitely good or in…nitely
bad
The Archimedean Axiom For all lotteries p, q and r such that
p q r , there must exist an a and b in (0, 1) such
that

ap + (1 a )r q bp + (1 b )r
 The Expected Utility Theorem

It turns out that these two axioms, when added to the

‘standard’ones, are necessary and su¢ cient for an expected
utility representation

Theorem
Let X be a …nite set of prizes , ∆(X ) be the set of lotteries on X .
Let be a binary relation on ∆(X ). Then is complete,
re‡exive, transitive and satis…es the Independence and
Archimedean axioms if and only if there exists a u : X ! R such
that, for any p, q 2 ∆(X ),

p q
if and only if ∑ px u ( x ) ∑ qx u ( x )
x 2X x 2X
 The Expected Utility Theorem

Proof?
Do you want us to go through the proof?
Oh, alright then
Actually, Necessity is easy
You will do it for homework
Su¢ ciency is harder
Will sketch it here
 Proof

Key to the proof is the following lemma

Lemma If is complete, re‡exive, transitive and satis…es the
Independence and Archimedean axioms then
1 p q and 0 α<β 1 implies

βp + (1 β )q αp + (1 α )q

2 p q r and p r implies that there exists a

unique α such that

q α p + (1 α )r
 The Expected Utility Theorem

Step 1
Find the best prize - in other words the prize such that getting
that prize for sure is preferred to all other lotteries. Give that
prize utility 1 (for convenience, let’s say that a is the best prize)
Step 2
Find the worst prize - in other words the prize such that all
lotteries are preferred to getting that prize for sure. Give that
prize utility 0 (for convenience, let’s say that d is the worse
prize)
Step 3
Show that, if a > b, then

aδa + (1 a ) δd bδa + (1 b ) δd

where δx is the lottery that gives prize x for sure (this is

intuitively obvious, but needs to be proved from the
independence axiom)
 The Expected Utility Theorem
Step 4
For other prizes (e.g. b), …nd the probability λ such that the
consumer is indi¤erent between getting apples with probability
λ and dragonfruit with probability (1 λ), and bananas for
sure. Let u (b ) = λ. i.e.
0 1 0 1 0 1
0 1 0
B 1 C B 0 C B 0 C
B C B C B C
@ 0 A u (b ) @ 0 A + (1 u (b )) @ 0 A
0 0 1
(for us to know such a λ exists requires the Archimedean
axiom)
Step 5
Do the same for c, so
0 1 0 1 0 1
0 1 0
B 0 C B 0 C B 0 C
B C B C + (1 u (c )) B C
@ 1 A u (c ) @ 0 A @ 0 A
0 0 1
 The Expected Utility Theorem

So now we have found utility numbers for every prize

All we have to do is show that p q if and only if
∑ x 2 X px u ( x ) ∑ x 2 X qx u ( x )
Let’s do a simple example
0 1 0 1
0 0
B 0.25 C B 0.75 C
p=B C B
@ 0.75 A , q = @ 0.25 A
C

0 0
 The Expected Utility Theorem

First, notice that

0 1 0 1 0 1
0 0 0
B 0.25 C B 1 C B 0 C
p=B C B
@ 0.75 A = 0.25 @
C + 0.75 B
A @
C
0 1 A
0 0 0

But
 The Expected Utility Theorem

But
0 1 0 1 0 1
0 1 0
B 1 C B 0 C B 0 C
B C u (b ) B C u (b )) B C
@ 0 A @ 0 A + (1 @ 0 A
0 0 1

and
0 1 0 1 0 1
0 1 0
B 0 C B 0 C B 0 C
B C u (c ) B C u (c )) B C
@ 1 A @ 0 A + (1 @ 0 A
0 0 1
 The Expected Utility Theorem

0 0 1 0 11
1 0
B B 0 C B 0 CC
p 0.25 B B C B CC
@u (b ) @ 0 A + (1 u (b )) @ 0 AA
0 1
0 0 1 0 11
1 0
B B 0 C B 0 CC
+0.75 B B C B CC
@u (c ) @ 0 A + (1 u (c )) @ 0 AA
0 1
 The Expected Utility Theorem

0 1
1
B 0 C
= (0.25u (b ) + 0.75u (c )) B C
@ 0 A+
0
0 1
0
B 0 C
(1 0.25u (b ) 0.75u (c )) B C
@ 0 A
1
 The Expected Utility Theorem

So p is indi¤erent to a lottery that puts probability

(0.25u (b ) + 0.75u (c ))

on the best prize (and the remainder on the worst prize)

But this is just the expected utility of p
Similarly q is ind¤erent to a lottery that puts

(0.75u (b ) + 0.25u (c ))

on the best prize

But this is just the expected utility of q
 The Expected Utility Theorem

So p will be preferred to q if the expected utility of p is higher

than the expected utility of q
This is because this means that p is indi¤erent to a lottery
which puts a higher weight on the best prize than does q
QED (ish)
 Expected Utility Numbers

Remember that when we talked about ’standard’utility

theory, the numbers themselves didn’t mean very much
Only the order mattered
So, for example

u (a ) = 1 v (a ) = 1
u (b ) = 2 v (b ) = 4
u (c ) = 3 v (c ) = 9
u (d ) = 4 v (c ) = 16

Would represent the same preferences

 Expected Utility Numbers

Is the same true here?

No!
According to the …rst preferences
1 1
u (a ) + u (c ) = 2 = u (b )
2 2
and so
1 1
a+ c b
2 2
But according to the second set of utilities
1 1
v (a ) + v (c ) = 5 > v (b )
2 2
and so
1 1
a+ c b
2 2
 Expected Utility Numbers

So we have to take utility numbers more seriously here

Magnitudes matter
How much more seriously?

Theorem
Let be a set of preferences on ∆(X ) and u : X ! R form an
expected utility representation of . Then v : X ! R also forms
an expected utility representation of if and only if

v (x ) = au (x ) + b 8 x 2 X

for some a 2 R++ , b 2 R

Proof.
Homework
 Risk Aversion

We motivated the move from expected value maximization to

expected utility maximization on the basis of risk aversion
Does EU imply risk aversion?
No!
Consider someone who has u (x ) = x
They will be risk neutral
Consider someone who has u (x ) = x 2
They will be risk loving
So risk attitude has something to do with the shape of the
utility function
 Risk Aversion
For this section we will think about lotteries with monetary
prizes
Let δx be the lottery that gives prize x for sure and E (p ) be
the expected value of a lottery p

De…nition
We say that a decision maker is risk averse if, for every lottery p

δ E (p ) p

We say they are risk neutral if

δ E (p ) p

We say they are risk loving if

δ E (p ) p
 Risk Aversion

We can say the same thing a di¤erent way

De…nition
The certainty equivalence of a lottery p is the amount c such
that
δc p
The risk premium is
E (p ) c
 Risk Aversion

Lemma
For a decision maker whose preferences are strictly monotonic in
money
1 They are risk averse if and only if for any p the risk premium
is weakly positive
2 They are risk neurtal if and only if for any p the risk premium
is zero
3 They are risk loving if and only if for any p the risk premium
is weakly negative
 Risk Aversion and Utility Curvature
We have made the claim that there is a link between risk
aversion and the curvature of the utility function
 Risk Aversion and Utility Curvature

We can make this statement tight

Theorem
An expected utility maximizer
1 Is risk averse if and only if u is concave
2 Is risk neutral if and only if u is linear
3 Is risk loving if and only if u is convex

Proof.
Comes straight from Jensen’s inequality: for a random variable x
and a concave function u

E (u (x )) u (E (x ))
 Measuring Risk Aversion

We might want a way of measuring risk aversion from the

utility function
Intuitively, the more ‘curvy’the utility function, the more risk
averse
How do we measure curvature?
The second derivative u 00 (x )!
Is this a good measure?
No, because we can change the utility function in such a way
that we don’t change the underlying preferences, and change
u 00 (x )
 The Arrow Pratt Measure

One way round this problem is to use the Arrow-Pratt

measure of absolute risk aversion
u 00 (x )
A(x ) =
u 0 (x )

This measure has some nice properties

1 If two utility functions represent the same preferences then
they have the same A for every x
2 It measures risk aversion in the sense that the following two
statements are equivalent
The utility function u has a higher Arrow Pratt measure than
utility function v for every x
Utility function u gives a higher risk premium than utility
function v for every p
 The Arrow Pratt Measure

Why is it called a measure of absolute risk aversion?

To see this, let’s think of a function for which A(x ) is constant
ax
u (x ) = 1 e

Note u 0 (x ) = ae ax and u 00 (x ) = a2 e ax so A(x ) = a

This is a constant absolute risk aversion (CARA) utility
function
 The Arrow Pratt Measure
Claim: for CARA utility functions, adding a constant amount
to each lottery doesn’t change risk attitues
i.e if δx p then δx +z is preferred to a lottery p 0 which adds
an amount z to each prize in p
To see this note that
u (x ) ∑ p (y )u (y )
y

1 e ax
∑ p (y ) 1 e ay
y

) 1 e ax
1 ∑ p (y )e ay
y

e az
e ax
e az
e az
∑ p (y )e ay
e az
y

) 1 e a (x +z )
∑ p (y ) 1 e a (y +z )
y

) u (x + z ) ∑ p (y )u (y + z )
y
 Relative Risk Aversion

Is this a sensible property?

Maybe not
Means that you should have the same attitude to a gamble
between winning $100 or losing $75 whether you are a student
earning $20,000 a year or a professor earning millions!
Perhaps a more useful measure is relative risk aversion

xu 00 (x )
R (x ) = xA(x ) =
u 0 (x )
 Relative Risk Aversion

An example of a Constant Relative Risk Aversion measure is

x1 ρ 1
u (x ) =
1 ρ

Note that u 0 (x ) = x ρ, u 00 (x ) = ρx ρ 1 and so R (x ) = ρ

CRRA utility functions have the property that proportional
changes in prizes don’t a¤ect risk attitudes
i.e if δx p then δαx is preferred to a lottery p 0 which
multiplies each prize in p by α > 0
 Relative Risk Aversion

To see this note that

u (x ) ∑ p (y )u (y )
y

x1 ρ 1 ∑y p (y )y 1 ρ 1
)
1 ρ 1 ρ
) x 1 ρ
∑ p (y )y 1 ρ
y

) α 1 ρ 1 ρ
x ∑ p (y ) α1 ρ 1 ρ
y
y
1
(αx )1 ρ 1 ∑y p (y ) (αy )
ρ
1
)
1 ρ 1 ρ
u (αx ) ∑ p 0 (y )u (y )
y
 Are People Expected Utility Maximizers?

Because of the work we have done above, we know what the

‘behavioral signature’is of EU
The independence axiom
Essentially this is picking up on the fact that EU demands
preferences to be linear in probabilities
Does this hold in experimental data?
 The Common Ratio E¤ect

What would you choose?

Many people choose C1 and D2
The Common Ratio E¤ect
 The Common Ratio E¤ect

This is a violation of the independence axiom

Why?
Because

D1 = 0.25C 1 + 0.75R
D2 = 0.25C 2 + 0.75R

where R is the lottery which pays 0 for sure

Thus independence means that

C1 C 2 ) D1 D2
 The Common Consequence E¤ect

What would you choose?

Many people choose A1 and B2
The Common Consequence E¤ect
 Explanations

What do you think is going on?

Many alternative models have been proposed in the literature
Disappointment: Gul, Faruk, 1991. "A Theory of
Disappointment Aversion,"
Salience: Pedro Bordalo & Nicola Gennaioli & Andrei Shleifer,
2012. "Salience Theory of Choice Under Risk,"
One of the most widespread and straightforward is probability
weighting
 Probability Weighting

Maybe the problem that the Allais paradox highlights is that

people do not ’believe’the probabilities that are told to them
For example they treat a 1% probability of winning $0 as if it
is more likely than that
‘I am unlucky, so the bad outcome is more likely to happen to
me’
The di¤erence between 0% and 1% seems bigger than the
di¤erence between 89% and 90%
This is the idea behind the probability weighting model.
 Simple Probability Weighting Model

Approach 1: Simple probability weighting

Let’s start with expected utility

U (p ) = ∑ p (x )u (x )
x 2X

And allow for probability weighting

V (p ) = ∑ π (p (x ))u (x )
x 2X

Where π is the probability weighting function

This can explain the Allais paradox
For example if π (0.01) = 0.05
 Simple Probability Weighting Model

However, the simple probability weighting model is not popular

For two reasons
1 It leads to violations of stochastic dominance
2 It doesn’t really capture the idea of ‘pessimism’
 Pessimism

Think back to the Allais paradox

0 1 0 1
0 0.01
@ 1 A @ 0.89 A
0 0.1

It seems as if the 1% probability of $0 is being overweighted

Is this just because it is a 1% probability?
Or is it because it is a 1% probability of the worst prize
If it is the latter, this is something that the simple probability
weighting model cannot capture
Weights are only based on probability
 Pessimism

Consider the following two examples

Example
Lottery p :49% chance of $10, 49% of winning $0, 2% chance of
winning $5

Example
Lottery p :49% chance of $10, 49% of winning $0, 2% chance of
losing $1000

Would you ‘weigh’the 2% probability the same in each case?

Arguably not
If you were pessimistic then you might think that 2% is ‘more
likely’in the latter case than in the former
Can’t be captured by the simple probability weighting model
 Rank Dependent Utility

Because of these two concerns, the simple probability

weighting model is rarely used
Instead people tend to use rank dependent utility
(sometimes also called cumulative probability weighting)
Probability weighting depends on
The probability of a prize
Its rank in the lottery - i.e. how many prizes are better or
worse than it
In practice this is done by applying weights cumulatively
Here comes the de…nition
It looks scary, but don’t panic!
 Rank Dependent Utility

De…nition
A decision maker’s preferences over ∆(X ) can be represented by
a rank dependant utility model if there exists a utility function
u : X ! R and a cumulative probability weighting function
ψ : [0, 1] ! [0, 1] such that ψ(0) = 0 and ψ(1) = 1, such that the
function U : ∆(X ) ! R represents , where U (p ) is constructed
in the following way:
1 The prizes of p are ranked x1 , x2 , . . . , xn such that
x1 x2 xn
2 U (p ) is determined as
! !!
n i i 1
U (p ) = ψ(p1 )u (x1 ) + ∑ ψ ∑ pj ψ ∑ pk u (xi )
i =2 j =1 k =1
 Rank Dependent Utility

Let’s go through an example: for prizes 10 > 5 > 0 let p be

equal to 0 1
0.1
@ 0.7 A
0.2
How do we apply RDU?
 Rank Dependent Utility

Well, …rst note that there are three prizes, so we can rewrite
the expression above as

U (p ) = ψ(p1 )u (x1 )
+ (ψ (p1 + p2 ) ψ (p1 )) u (x2 )
+ (ψ (p1 + p2 + p3 ) ψ (p1 + p2 )) u (x3 )

The weight attached to the best prize is the weight of p1

The weight attached to the second best prize is the weight on
the probability of
Getting something at least as good as the second prize
Minus the probability of getting something better than the
second prize
And so on
Notice that if ψ is the identity function this is just expected
utility
 Rank Dependent Utility

In this speci…c case

U (p ) = ψ(p1 )u (x1 )
+ (ψ (p1 + p2 ) ψ (p1 )) u (x2 )
+ (ψ (p1 + p2 + p3 ) ψ (p1 + p2 )) u (x3 )

Becomes

U (p ) = ψ(0.1)u (10)
+ (ψ (0.8) ψ (0.1)) u (5)
+ (ψ (1) ψ (0.8)) u (0)
 Introduction

In the …rst class we drew a distinction betweem

Circumstances of Risk (roulette wheels)
Circumstances of Uncertainty (horse races)
So far we have been talking about roulette wheels
Now horse races!
 Risk vs Uncertainty

Remember the key di¤erence between the two

Risk: Probabilities are observable
There are 38 slots on a roulette wheel
Someone who places a $10 bet on number 7 has a lottery with
pays out $350 with probability 1/38 and zero otherwise
(Yes, this is not a fair bet)
Uncertainty: Probabilities are not observable
Say there are 3 horses in a race
Someone who places a $10 bet on horse A does not necessarily
have a 1/3 chance of winning
Maybe their horse only has three legs?
 Subjective Expected Utility

If we want to model situations of uncertainty, we cannot think

about preferences over lotteries
Because we don’t know the probabilities
We need a di¤erent set up
We are going to thing about acts
What is an act?
 States of the World

First we need to de…ne states of the world

We will do this with an example
Consider a race between three horses
A(rchibald)
B(yron)
C(umberbach)
What are the possible oucomes of this race?
Excluding ties
 States of the World

State Ordering
1 A, B ,C
2 A, C, B
3 B, A, C
4 B, C, A
5 C, A, B
6 C, B, A
 Acts

This is what we mean by the states of the world

An exclusive and exhaustive list of all the possible outcomes in
a scenario
An act is then an action which is de…ned by the oucome it
gives in each state of the world
Here are two examples
Act f : A $10 even money bet that Archibald will win
Act g : A $10 bet at odds of 2 to 1 that Cumberbach will win
 Acts

State Ordering Payo¤ Act f Payo¤ Act g

1 A, B ,C $10 -$10
2 A, C, B $10 -$10
3 B, A, C -$10 -$10
4 B, C, A -$10 -$10
5 C, A, B -$10 $20
6 C, B, A -$10 $20
 Subjective Expected Utility Theory

So, how would you choose between acts f and g ?

SEU assumes the following:

1 Figure out the probability you would associate with each state
of the world
2 Figure out the utility you would gain from each prize
3 Figure out the expected utility of each act according to those
probabilities and utilities
4 Choose the act with the highest utility
 Subjective Expected Utility Theory

So, in the above example

Utility from f :

[π (ABC ) + π (ACB )] u (10)

+ [π (BAC ) + π (BCA)] u ( 10)
+ [π (CBA) + π (CAB )] u ( 10)

where π is the probability of each act

Utility from g :

[π (ABC ) + π (ACB )] u ( 10)

+ [π (BAC ) + π (BCA)] u ( 10)
+ [π (CBA) + π (CAB )] u (20)
 Subjective Expected Utility Theory

Assuming utility is linear f is preferred to g if

[π (ABC ) + π (ACB )] 3
[π (CBA) + π (CAB )] 2

Or the probability of A winning is more than 3/2 times the

probability of C winning
 Subjective Expected Utility Theory

De…nition
Let X be a set of prizes, Ω be a (…nite) set of states of the world
and F be the resulting set of acts (i.e. F is the set of all functions
f : Ω ! X ). We say that preferences on the set of acts F has a
subjective expected utility representation if there exists a utility
function u : X ! R and probability function π : Ω ! [0, 1] such
that ∑ω 2Ω π (ω ) = 1 and

f g
, ∑ π (ω )u (f (ω )) ∑ π (ω )u (g (ω ))
ω 2Ω ω 2Ω
 Subjective Expected Utility Theory

Notes
Notice that we now have two things to recover: Utility and
preferences
Axioms beyond the scope of this course: has been done twice -
…rst by Savage1 and later (using a trick to make the process a
lot simpler) by Anscombe and Aumann2
Utility pinned down to positive a¢ ne transform
Probabilities are unique

1 Savage,
Leonard J. 1954. The Foundations of Statistics. New York, Wiley.
2 Anscombe, F. J.; Aumann, R. J. A De…nition of Subjective Probability.
The Annals of Mathematical Statistics 34 (1963), no. 1, .
 The Ellsberg Paradox

Unfortunately, while simple and intuitive, SEU theory has

some problems when it comes to describing behavior
These problems are most elegantly demostrated by the
Ellsberg paradox
This thought experiment has sparked a whole …eld of decision
theory
 The Ellsberg Paradox - A Reminder

Choice 1: The ’risky bag’

Fill a bag with 20 red and 20 black tokens
O¤er your subject the opportunity to place a $10 bet on the
color of their choice
Then elicit the amount x such that the subject is indi¤erent
between playing the gamble and receiving $x for sure.
Choice 2: The ‘ambiguous bag’
Repeat the above experiment, but provide the subject with no
information about the number of red and black tokens
Then elicit the amount y such that the subject is indi¤erent
between playing the gamble and receiving $y for sure.
 The Ellsberg Paradox

Typical …nding
x >> y
People much prefer to bet on the risky bag
This behavior cannot be explained by SEU?
Why?
 The Ellsberg Paradox

What is the utility of betting on the risky bag?

The probability of drawing a red ball is the same as the
probability of drawing a black ball at 0.5
So whichever act you choose to bet on, the utility of the
gamble is
0.5u ($10)
 The Ellsberg Paradox

What is the utility of betting on the ambiguous bag?

Here we need to apply SEU
What are the states of the world?
Red ball is drawn or black ball is drawn
What are the acts?
Bet on red or bet on black
 The Ellsberg Paradox

State r b
red 10 0
black 0 10

How do we calculate the utility of these two acts?

Need to decide how likely each state is
Assign probabilities π (r ) = 1 π (b )
Note that these do not have to be 50%
Maybe you think I like red chips!
 The Ellsberg Paradox

Utility of betting on the red outcome is therefore

π (r )u ($10)

Utility of betting on the black outcome is

π (b )u ($10) = (1 π (r ))u ($10)

Because you get to choose which color to bet on, the gamble
on the ambiguous urn is

max fπ (r )u ($10), (1 π (r ))u ($10)g

is equal to 0.5u ($10) if π (r ) = 0.5

otherwise is greater than 0.5u ($10)
should always (weakly) prefer to bet on the ambiguous urn
intuition: if you can choose what to bet on, 0.5 is the worst
probability
 The Ellsberg Paradox

61% of my last class exhibited the Ellsberg paradox

For more details see Halevy, Yoram. "Ellsberg revisited: An
experimental study." Econometrica 75.2 (2007): 503-536.
 Maxmin Expected Utility

So, as usual, we are left needing a new model to explain

behavior
There have been many such attempts since the Ellsberg
paradox was …rst described
We will focus on ’Maxmin Expected Utility’by Gilboa and
Schmeidler3

3 Gilboa,
Itzhak & Schmeidler, David, 1989. "Maxmin expected utility with
non-unique prior," Journal of Mathematical Economics, Elsevier, vol. 18(2),
pages 141-153, April.
 Maxmin Expected Utility

Maxmin expected utility has a very natural interpretation....

The world is out to get you!
Imagine that in the Ellsberg experiment was run by an evil and
sneaky experimenter
After you have chosen whether to bet on red or black, they will
increase your chances of losing
They will sneak some chips into the bag of the opposite color
to the one you bet on
So if you bet on red they will put black chips in and visa versa
 Maxmin Expected Utility

How should we think about this?

Rather than their being a single probability distribution, there
is a range of possible distributions
After you chose your act, you evaluate it using the worst of
these distributions
This is maxmin expected utility
you maximize the minimum utility that you can get across
di¤erent probability distributions
Has links to robust control theory in engineering
This is basically how you design aircraft
 Maxmin Expected Utility

De…nition
Let X be a set of prizes, Ω be a (…nite) set of states of the world
and F be the resulting set of acts (i.e. F is the set of all functions
f : Ω ! X ). We say that preferences on the set of acts F has a
Maxmin expected utility representation if there exists a utility
function u : X ! R and convex set of probability functions Π and

f g
, min
π 2Π
∑ π (ω )u (f (ω )) min
π 2Π
∑ π (ω )u (g (ω ))
ω 2Ω ω 2Ω
 Maxmin Expected Utility

Maxmin expected utility can explain the Ellsberg paradox

Assume that u (x ) = x
Assume that you think π (r ) is between 0.25 and 0.75
Utility of betting on the risky bag is 0.5u (x ) = 5
What is the utility of betting on red from the ambiguous bag?

min π (r )u ($10) = 0.25u ($10) = 2.5

π (r )2[0.25,0.75 ]

Similary, the utility from betting on black is

min (1 π (r )) u ($10) = 0.25u ($10) = 2.5

π (r )2[0.25,0.75 ]

Maximal utility from betting on the ambiguous bag is lower

than that from the risky bag