Decision Making Under

Uncertainty
Decision Making Under Uncertainty
Expected utility theory attempts to explain how economic agents choose
among
different risky prospects.

For historical reasons, Economists refer to a risky prospect as a lottery

or a gamble.
These terms, however, are only generic names for what could be a real-
world risky prospect that may have nothing to do with pure gambling,

such as investing in stock market, making a capital investment with

uncertain future revenue stream
or deciding which career to enter.
Formally, a lottery is a random variable L whose outcomes WI, W2, .... Wn are
economically meaningful events that occur, respectively, with known
probabilities P1, P2, …Pn

The outcomes may be anything that individual decision maker may value,
though we will often assume, without loss of generality, that the outcomes Wi
refer to different levels of wealth.

To simplify the 'exposition, we will take the set of possible outcomes WI, W2, ...
, Wn as given and fixed and indicate a lottery L simply by the probabilities with
which these outcomes occur.

Specifically, we write
..
to indicate a lottery in which agent realizes wealth Wi with probability Pi for
i = 1,2, ... , n.

Example: Suppose that obtaining an MS from OSU will result in a $100

thousand annual salary with probability 0.6 and a $50 thousand annual salary
with probability 0.4; further suppose that not obtaining an MS will result in a
$50 thousand annual salary with probability 0.5 and a $10 thousand annual
salary with probability 0.5. With outcomes consisting of annual salaries of

$100. $50, and $10 thousand, the former (MS) lottery may be denoted

L1 = (0.6,0.4,0.0)

and the latter (non MS) lottery may be denoted

L2 = (0.0,0.5,0.5).
A compound lottery is a lottery whose possible outcomes are simple lotteries. If
L1 and L2 are lotteries and pai 1 and pai2 are a pair of probabilities, we write

to denote the compound lottery in which lottery

is awarded with probability pai1

 and lottery

is awarded with probability π2. We assume agents view compound lotteries as the
equivalent simple lotteries implied by iterating probabilities.

Thus, the compound lottery

is equivalent to the simple lottery

Where

Pi= π1p1i+ π2p2i i=1,2,…n

Example 1 The outcome of obtaining an MS degree from Purdue University is also
random. With probability 0.4, a Purdue MS will be equivalent to an OSU MS; with
probability 0.6, a Purdue MS will equivalent to no MS at all. Thus, continuing the
preceding example, pursuing an MS from Purdue is the compound lottery

L = 0.4L1 + 0.6L2,

which is equivalent to the simple lottery

L = 0.4(0.6,0.4,0.0) + 0.6(0.0, 0.5, 0.5) = (0.24,0.46,0.30)

That is, obtaining an MS degree from Purdue will result in a $100 thousand annual
salary with probability 0.24 .. a $50 thousand annual salary with probability 0.46, and a
$10 thousand annual salary with probability 0.30
As in static consumer theory, we will assume that an agent has a well­defined
preference ordering ≥ over the set of all lotteries, and write LI ≥L2 to indicate
that the agent prefers lottery L1 to lottery L2 or is indifferent between the two.

If L1 ≥ L2 and L2 ≥ L1, we write L1~L2 to indicate the agent is indifferent

between the two lotteries

We will make certain assumptions regarding an agent's preferences over

lotteries. In particular, if L1, L2 and L are lotteries, we will assume:

• Completeness: Either L1 ≥ L2, L2 ≥ L1 or both

• Transitivity: If L1 ≥ L and L ≥L2, then L1 ≥ L2

•Continuity: If L1 ≥L ≥L2 then for some probability π

L= πL1 +(1-π)L2

•Independence: If L1 ~ L2, then, for any probability π

πL1+ (1- π)L ~ πL2 +(1- π)L
The independence axiom states that if the agent is Indifferent between
lotteries L1 and L2. then he is indifferent between any two compound lotteries
in which L1 is replaced by L2, and vice versa.

The independence axiom is a controversial assumption, but, together with

the other three assumptions, yields a very strong and practical result

In particular. if all four axioms hold, there exists a real-valued function U

defined on the set of all outcomes w such that if L1 = (P11, P12, ... ,P1n) and
L2 = (P21, P22 ... ,P2n) are lotteries, then

We call the function u a von Neuman-Morgenstern utility of wealth function

The existence of a von Neuman-Morgenstern utility of wealth function is a very
convenient result. It allows us to compare lotteries by performing relatively
simple computations.

In particular, we need only compute the "expected utility" provided by the two
lotteries. The agent always prefers the lottery that provides the greatest
expected utility
Example 2: Suppose an agent is presented with two lotteries

LI = (0.2,0.4,0.1, 0.2, 0.1)

And

L2 = (0.1, 0.5, 0.2, 0.1, 0.1)

over five possible wealth levels w = 10, 20, 30, 40,and 50.

. Further suppose that the agent's preferences satisfy the von Neuman-
Morgenstern axioms and that he ascribes utilities u=1.0, 1.4, 1.7, 2.0,
and 2.2 to the five wealth levels, respectively. The agent's expected
utility with lottery 1 is:

EU1 =0.2 ·1.0 +0.4 ·1.4 +0.1 ·1.7 +0.2 ·2.0 +0.1 ·2.2 =1.55

and with lottery 2 is

EU 2 =0.1 .1.0 +0.5 ·1.4 +0.2 ·1.7 +0.1 ·2.0 +0.1 ·2.2 =1.56.
Thus, the agent prefers lottery 2 over lottery 1.
Suppose for the sake of argument, that an agent faces the prospect of two
possible Wealth outcomes, w1 and w2, with probabilities P1 and P2,
respectively.

Denote the agent's expected utility bv u bar = P1U(w1) + P2u(w2)

Agent’s expected wealth by w bar= PI w1 + P2w2.

We consider three distinct possibilities for the curvature of u: u is convex, u is

linear, and u is concave
Figure 1: Convex Utility of Wealth function; ubar>uwbar, risk loving
Prefers risky prospects over receiving wealth with certainty
Figure 2.: Linear utility of wealth function ; ubar=uwbar, risk neutral
Figure 3; Concave utility of wealth function; uwbar>ubar; risk averse
Prefers expected wealth with certainty over risky prospects
Jensen's inequality generalises the risk aversion result, above result to
more general risky prospects with a continuum of outcomes:

If u is concave and defined for all possible values of a continuous random

variable w
Then
U(E(w)≥Eu(w)

Example: Demand for Insurance

Suppose a risk-averse agent has a current wealth level W, but faces a

possible loss L with probability p.

Also suppose that the agent may purchase any amount of coverage K
against the loss at a premium rate γ

That is, the agent can pay a premium γK ( for a contract that pays her an
indemnity K in the event of a loss, but nothing otherwise.

How much coverage should the agent purchase?

With insurance, the agent ends up with wealth W - L + K - γ K with probability p
and wealth W- γK with probability 1- p.

The agent achieves her most preferred position by purchasing coverage K that
maximizes her expected utility of wealth

Eu(w) = g(K) = (1 - p) u(W-yK) + pu (W- L + K -yK)

This is achieved by setting derivative of expected utility with respect to K to 0:

g'(K) = (-y)(1- p) u'(W - yK) + (1 - y) pu‘ (W - L + K - yK) = 0

y /p. u'(W - y K) = 1- y/1-p. u’ (W-L+K- yK)

(Since the agent is risk-averse, u is concave, and so is expected utility as a
function of the coverage level K; thus, the first-order condition is both
necessary and sufficient for a maximum.)
Now suppose that insurance is actuarially fair. That is, the expected
indemnity pK equals total premiums yK, so that p = y. Then,

u'(W - yK) = u'(W -L + K - yK).

w-γK=w-L+K-γK
And thus
K=L

That is, if the insurance is actuarially fair, a risk-averse agent will purchase
full coverage, completely eliminating any uncertainty.
It can also be shown that if insurance is not actuarially fair,

that is, the insurer sets a premium, that exceeds the loss probability p in order
to make a profit,
then a risk-averse agent will purchase less than full coverage and retain some
of the risk.

Measuring Risk A version

One way to measure an agent's aversion to a specific risk is by the amount of
money the agent would be willing to pay to eliminate it completely.

Suppose an agent faces uncertain wealth w. The certainty equivalent level of

wealth w* is the level of wealth the agent would accept with certainty in
exchange for his uncertain wealth. That is,

Eu(w) = u(w*)
We know by Jensen's inequality that if the agent is risk-averse,

u(Ew bar) > Eu(w bar) = u(w*)

which implies

E w bar> w *

We define the risk premium to be the difference between the certainty

equivalent wealth and the expected uncertain wealth

Π= E w bar-w*

The risk premium measures the agent's willingness to pay to eliminate all
risk. It is positive if the agent is risk-averse.
Figure 4: Certainty Equivalent Income

Consider Figure 4, which illustrates the case of an agent who faces an

uncertain wealth prospect
In this figure, the wealth level w* provides the same expected utility as his
uncertain wealth prospect.

Thus, the agent is indifferent between receiving w* with certainty and facing his
uncertain wealth prospect. As such, w* is the agent's certainty-equivalent
wealth and

π = w bar – w*

is his risk premium.

Example: Mr. Smith is a salesman who works on commission.

In a good year, he earns $80,000 and, in a bad year, he earns $60,000.

Good years and bad years are equally probable, so that Smith's expected
annual income is $70,000.

Smith's employer offers him the opportunity to continue in his current position,
but at a fixed salary that is to be negotiated.

As he enters the negotiations with his employer, Smith decides he would

accept $66,000, but not a penny less.

It follows that Smith's certainty-equivalent income is $66,000, implying that he

assesses the cost of risk associated with working on commission at $ 4,000 =
$70,000-$66,000.
Example: An investor with utility of wealth function u(w) = √w and initial wealth
100 is offered a risky asset with two equally likely payoffs, -10 and 10. The risk
premium π placed on the risky asset by the investor satisfies

Eu(w bar) = u(Ew bar - π)

that is,
0.5 √ 90+ 0.5 √110 = √100 - π
which can easily be solved with a calculator for π = 0.2506
Two widely used measures of risk aversion are, respectively, the
Arrow­Pratt measures of absolute and relative risk aversion:

A(w) = -u"(w)/ u’ (w)

R(w) = -wu"(w)/ u’(w)

The two measures are independent of the utility function used to represent
agent preferences;

the latter, but not the former, is also independent of the units used to measure
wealth.

The sign of the absolute and relative risk aversion measures indicate the
agent's attitude toward risk. Since we assume that u'(w) > 0, both measures will
be positive if the agent is risk-averse
both measures will be zero if the agent is risk-neutral; and both measures will
be negative if the agent is risk-loving

Both measures of risk aversion are intimately related to the risk premium.

Suppose an agent with current wealth w is presented with a small pure risk
€. That is, he faces an uncertain wealth

W bar=w+€

where € is a zero-mean random variable with standard deviation σ and

coefficient of variation v = σ /w relative to current wealth.

Then an approximate expression for the risk premium π, the amount of

money the agent is willing to pay to entirely eliminate the pure risk € , is

π ≈ ½ A(w) σ sq.

That is, the risk premium is directly proportional to the agent's absolute risk
aversion and to the absolute magnitude of the risk, as measured by its
variance.
 Similarly

That is, the risk premium, as a proportion of current wealth, is directly

proportional to the agent's relative risk aversion and the relative
magnitude of the risk, as measured by the square of the coefficient of
variation.

It is generally presumed that the wealthier an agent is, the less he is

willing to pay to eliminate a given risk.

In other words, for a given risk, the risk premium should decline with
wealth. It can be shown that, for any risk, the risk premium decreases
with wealth

if, and only if, the coefficient of risk aversion is globally decreasing in
wealth, that is, if and only if, A'(w) < 0 for all w.

An agent whose utility of wealth function satisfies this condition is said to

exhibit decreasing absolute risk aversion or DARA preferences
Example : Suppose an agent faces terminal wealth levels of 4 and 6 with equal
probability, so that his wealth has an expectation of 5, a variance of 1, and a
coefficient of variation 1/5. Further suppose that the agent possesses a utility of
wealth function u(w) = -w-2. Then the agent's certainty-equivalent income w' is
characterized by :

-w*-2=-.5.4-2-.5.6-2=-.02257

So that w*=4.707

And his risk premium is

Π=w bar-w*=5-4.707=.293

Use Arrow-Pratt approximation and compute pai?

The agent's utility of wealth function exhibits constant relative risk
aversion of 3, since

R(w) = -wu"(w) /u'(w) = (-) -6w-3 /2w-3 =3

Thus, from the Arrow-Pratt approximation above,

Π ~ 0.5· 5 . 3/25 = 0.300
which is clearly a reasonable approximation for the exact figure
0.293
Paradoxes of Expected Utility

Expected utility theory has been criticized for not adequately explaining all
observed behavior under uncertainty.

For example, many people who gamble also purchase insurance.

This phenomenon is typically explained away by incorporating an

entertainment value from gambling

and noting that people will gamble small amounts but will purchase insurance
against big potential losses.

Two other, more formal puzzles that are frequently cited in the economics
literature are known as the Allais and Ellsberg paradoxes
Allais Paradox

Consider four lotteries that yield winnings of 0, 1 million, and 5 million

dollars. respectively according to the following probabilities:

L1= (0.00.1.00,0.00)
L2 = (0.01, 0.89, 0.10)
L3 = (0.89,0.11, 0.00)
L4 = (0.90, 0.00, 0.10)
In actual experiments, most people prefer lottery 1 over lottery 2 and
prefer lottery 4 over lottery 3. However, these choices violate the
axioms of expected utility theory.
To prove this, suppose an agent's preferences satisfy the
axioms of expected utility and that he ascribes utilities uo, u1,
and u5 to wealth of 0, 1 million, and 5 million dollars,
respectively. Then

L1 >L2 => Eu(L1) > Eu(L2)

0.00uo + 1.00u1+ O.OOu5 > O.Oluo + 0.89u1 + 0.10u5
0.89uo + 0.11u1 + 0.00u5 > 0.90uo + O.OOUI + O.10u5
=> Eu(L3 ) > Eu(L4)
L3 >L4

That is, if an agent’s preferences satisfy the axioms of

expected utility and he prefers lottery 1 over lottery 2, he
should prefer lottery 3 over lottery 4.
Ellsberg Paradox

Suppose Hat X has 50 red balls and 50 black balls and Hat Y has 100 red and
black balls, but in unknown proportions.

Given a choice between the two hypothetical gambles

• XR = Get $100 if a red ball drawn from hat X, $0 otherwise

• YR = Get $100 if a red ball drawn from hat Y, $0 otherwise

most people choose X R

Given a choice between the two hypothetical gambles

X B = Get $100 if a black ball drawn from hat X, $0 otherwise

• Yo = Get $100 if a black ball drawn from hat Y, $0 otherwise

most people choose XB

However, these two choices are inconsistent according to expected utility theory.
Let p denote the unknown proportion of red balls I hat Y.
Then
XR>YR =>. Eu(XR) > Fu(YR) ==:- 0.5u(O) +- 05u(100) > (1 - p)u(0) +- pu(100)
=>-0.5u(0) – 0.5u(100) > -pu(0) -+ (p - l)u(100)
pu(0) +- (1 - p)u(100) > O.5u(O) +- 0.5u(100)
EU(YB):> Eu(XB)
YB >XB

Thus, regardless of the value of p, if an agent's preferences satisfy the axioms of

expected utility and he chooses XR over YR, then he should choose YB over XB.
How does one explain this paradox? 'Most people are distrustful.

If offered a choice between the first pair of gambles, in which a red ball earns a
reward, an individual would suspect that the lottery manager has placed very
few red balls, if any, in hat Y;

that is, he would assume that p is small, making hat X the intelligent choice.

If offered a choice between the second pair of gambles, in which a black ball
earns a reward, an individual would suspect that the lottery manager has
placed very few black balls, if any, in hat Y
that is, he would assume that p is large, making hat X the intelligent choice,

In other words, the individual’s assessment of the number of black and red balls in hat Y
is contingent on how the gambles are presented

If the individual were convinced that the proportion of black and red balls in hat Y was
Determined by a fair process that was not biased in favour of either black or red balls,
then the individual would conclude that all four gambles offer a 50% chance of paying
$100, making him/her indifferent between them.

Which would be consistent with expected utility theory