Risk Aversion and Expected Utility Theory: Coherence For Small-And Large-Stakes Gambles
Risk Aversion and Expected Utility Theory: Coherence For Small-And Large-Stakes Gambles
Risk Aversion and Expected Utility Theory: Coherence For Small-And Large-Stakes Gambles
There is a sizable literature reporting the conclusion that expected utility theory cannot provide a
coherent positive theory of risk averse behavior. Recently-published articles have sharpened the
criticism with calibrations based on assumed concavity and additivity of initial wealth to income.
We demonstrate that the negative conclusions in this literature are valid for only one expected
utility model not, as claimed, for expected utility theory. The conclusions are not valid for the
expected utility model commonly used in bidding theory nor for a more general expected utility
model that we develop by extending the Arrow-Pratt characterization of agents comparative risk
aversion. Clarification of the distinction between expected utility theory and expected utility
models also makes it clear that loss aversion is consistent with expected utility theory.
Key Words: expected utility theory, risk aversion, loss aversion
JEL Classification Number: D81
1. Introduction
A central objective in developing an empirical science is coherence in the application of theory to
data from different sources. Thus, Rabin (2000a, 2000b) and Rabin and Thaler (2001) address an
important question concerning coherence of the application of concave expected utility theory to
explain risk-averse behavior for both small-stakes gambles used in laboratory experiments and
large-stakes gambles observed in everyday life. Although Rabin (2000b) singles out the work of
experimental economists for special criticism, his calibration theorem focuses ones attention
on the central issue in a sizable literature that questions the usefulness of expected utility theory
for modeling risk-averse behavior.
1
This criticism has been quite influential, in part because strongly-worded conclusions
have been repeated so many times. Thus, Rabin (2000a) tells readers that: Within the expected-
utility framework, the concavity of the utility-of-wealth function is not only sufficient to explain
risk aversion it is also necessary: Diminishing marginal utility of wealth is the sole explanation
for risk aversion (p. 202) Rabin also states that: The inability of expected utility theory to
provide a plausible account of risk aversion over modest stakes has been illustrated in writing in a
2
variety of different contexts (pp. 202-203) He continues by stating: Within the expected-
utility framework, for any concave utility function, even very little risk aversion over modest
stakes implies an absurd degree of risk aversion over large stakes. (p. 203)
Rabin (2000b) proves a calibration theorem and corollary for a concave expected-utility
model that he interprets as bringing the following to ones attention: this theory implies that
people are approximately risk neutral when stakes are smallWhile not broadly appreciated, the
inability of expected-utility theory to provide a plausible account of risk aversion over modest
stakes has become oral tradition among some subsets of researchers (p. 1281) In addition to
proof of the theorem, Rabin discusses at some length its purported implications for empirical
methods in economics. He is particularly critical of the use by experimental economists of
concave expected-utility theory to explain systematic, one-sided deviations from the predictions
of risk-neutral models because of his belief that expected-utility theory is manifestly not close
to the right explanation of risk attitudes over modest stakes (p. 1282) and if we think that
subjects in experiments are risk averse, then we know they are not expected-utility maximizers.
(p. 1286) He states that the right explanation is loss aversion, which is said to be inconsistent
with expected-utility theory: loss aversion, is a departure from expected-utility theory that
provides a direct explanation for modest scale risk aversion. (p. 1288)
The conclusions about expected-utility theory are stated most forcefully by Rabin and
Thaler (2001), as follows: We are arguing that when people decline gambles with positive
expected value for modest stakes, they are violating expected utility theory. (p. 227) They
conclude their paper as follows: it is time for economists to recognize that expected utility
theory is an ex-hypothesis, so that we can concentrate our energies on the important task of
developing better descriptive models of choice under uncertainty. (p. 230) This conclusion is
immediately preceded by the following: What should expected utility theory be replaced with?
We think it is clear that loss aversion and the tendency to isolate each risky choice must both be
key components of a good descriptive theory of risk attitudes. (p. 230)
3
These criticisms of expected utility theory, and the related discussions of loss aversion,
involve some issues for which there is a marked absence of clarity in the literature on decision
theory. In order to help clarify the fundamental differences between expected utility theory and
alternative decision theories, we draw attention to the distinction between expected utility theory
and expected utility models. We address the coherence question for both of the commonly-used
expected utility models, the expected utility of terminal wealth model and the expected utility of
income model, and for a more general model that we develop, the expected utility of income and
initial wealth model.
We explain that the calibration theorem has no implication for the expected utility of
income (EUI) model because of its fixed reference point of zero income. The expected utility of
income and initial wealth (EUI&IW) model includes initial wealth as a reference point but the
calibration theorem has no implication for this model because initial wealth is not additive to
income in the utility function for this model. In contrast, if there is empirical support for the
pattern of small-stakes risk aversion hypothesized in Rabin (2000a, 2000b) and Rabin and Thaler
(2001), then the calibration theorem has implications for applicability of the expected utility of
terminal wealth (EUTW) model.
Since modeling risk aversion with the EUTW model is called into question by Rabins
calibration theorem, and since the calibration theorem has no implication for the EUI&IW model,
it is important to ascertain whether central analytical results in the economics of uncertainty that
have been derived with the EUTW model have analogues that can be derived with the EUI&IW
model. We begin the study of this question by extending the Arrow (1971) and Pratt (1964)
characterization of comparative risk aversion to the EUI&IW model. Subsequently, we use the
EUI&IW model to derive an analogue of Arrows (1971) classic two-asset portfolio allocation
proposition.
Clarification of the distinction between expected utility theory and specific expected
utility models also helps to clarify the implications of loss aversion (Kahneman and Tversky,
4
1979) for decision theory. We explain that loss aversion is consistent with some expected utility
models, hence it is consistent with expected utility theory. Thus, if a researcher observes loss-
averse behavior, he should not, ipso facto, be motivated to reject expected utility theory in favor
of some alternative such as prospect theory (Kahneman and Tversky, 1979), although loss
aversion is inconsistent with the expected utility of terminal wealth model.
We mention a small part of the literature that reports data showing risk-averse behavior in
laboratory experiments. The common features of these experiments are: (1) consistent one-sided
deviations of subjects choices from the predictions of risk neutrality, in the direction consistent
with risk aversion; (2) no relevance of loss aversion; and (3) researchers use of expected utility
models for which the calibration theorem has no implication. Clarification of these issues is
important to decision theory and to proper review for journals of papers reporting empirical
applications of expected utility theory.
2
2. Expected Utility Theory vs. Expected Utility Models
We define expected utility theory as the theory of decision-making under risk based on a set of
axioms for a preference ordering that includes the independence axiom or an alternative that
implies that the (expected) utility function that represents the ordering is linear in probabilities.
Linearity in probabilities implies that indifference curves in the Machina (1982) triangle diagram
and the probability simplex are parallel straight lines. Thus, we include within expected utility
theory any model of decision-making under risk that has parallel straight-line indifference curves
in the Machina triangle diagram.
Expected utility theory and alternative decision theories are concerned with the properties
of preference orderings of probability distributions of prizes. The identity of the prizes depends
on the decision context that is modeled and on assumptions made by the theorist. We shall
confine our discussion to decision contexts in which the prizes are amounts of money. Within this
context, we identify two expected utility models that are commonly used, the expected utility of
5
terminal wealth model and the expected utility of income model. We also discuss a third model,
the expected utility of income and initial wealth model. The distinctions among the models are in
the assumed identity of the prizes. All of the models have parallel straight-line indifference
curves in the Machina triangle, hence are included within expected utility theory.
2.1 The EUTW Model
The expected utility of terminal wealth (EUTW) model is based on the assumption that the prizes
are amounts of terminal wealth. This model was used in the seminal work of Arrow (1971) and
Pratt (1964) in which they developed the measures for comparing agents risk attitudes. This
model is commonly used in theoretical and empirical papers on various topics.
It will help explain some essential distinctions between models to briefly review the
familiar triangle-diagram representation of indifference curves for simple gambles (Machina,
1987). Consider three amounts of monetary gains (or amounts of income),
i
y , i = 1,2,3 such that
3 2 1
y y y < < . Assume that the
i
y occur with probabilities,
i
p , i = 1,2,3 such that
1
3 2 1
= + + p p p . If w is the decision-makers initial wealth, then the expected utility function
for the EUTW model is written as
(1)
=
+ =
3
1
) (
i
i i W
y w u p U .
Since
3 1 2
1 p p p = , indifference curves for expected utility function (1) can be represented
in the triangle diagram in Figure 1. Indifference curves are parallel straight lines because
expected utility function (1) is linear in probabilities. Higher (respectively lower) Arrow-Pratt
absolute risk aversion implies steeper (respectively flatter) slope of the indifference curves. Thus
if expected utility function (1) exhibits non-constant absolute risk aversion then the slopes of the
indifference curves in Figure 1 depend on the amount of initial wealth, w. In the special cases in
6
which expected utility function (1) exhibits constant absolute risk aversion or the agent is risk
neutral, the slope of the indifference curves in Figure 1 is invariant to changes in w.
2.2 The EUI Model
The expected utility of income (EUI) model is based on the assumption that the prizes are
amounts of income (or, equivalently, changes in wealth or gains and loses). The EUI model is
commonly used in theoretical modeling, most notably in the theory of auctions. Thus, most
bidding models are based on the expected utility axioms, the assumption that the prizes are
amounts of income, and the assumption of Bayesian-Nash equilibrium.
3
Considering the same three-outcome lottery as above, the expected utility function for the
EUI model is written as
(2)
=
=
3
1
) (
i
i i I
y v p U .
Obviously, indifference curves for this model are parallel straight lines whose slope is
independent of initial wealth.
2.3 The EUI&IW Model
We here discuss a more general model, the expected utility of income and initial wealth
(EUI&IW) model. Assume that the prizes are ordered pairs of amounts of initial wealth and
income, ) , (
i
y w , i = 1,2,3. Then the expected utility function can be written as
(3)
=
=
3
1
) , (
i
i i IW
y w p U .
Indifference curves for this model are parallel straight lines; therefore it is an expected utility
model. Because it appears that the properties of the EUI&IW model have not previously been
developed, we extend the Arrow-Pratt theory of comparative risk aversion to this model, as
follows.
7
Let F denote the probability distribution function for the random amount of income, Y.
The expected utility function for agent , = j can be written as
(4) )) , ( ( y w E U
j
F
j
IW
=
for either discrete or continuous distributions of random income. The Arrow-Pratt measure of
absolute risk aversion for this model is
(5)
) , (
) , (
) , (
2
22
y w
y w
y w A
j
j
j
= .
The risk premium,
j
is defined by
(6) )) , ( ( )) , ( ) ( , ( y w E F w y E w
j
F
j
F
j
= .
Assume that the function,
j
is strictly increasing in income, y ; then there exists a y-inverse
function,
j
, defined by
(7) )) , ( , ( y w w y
j j
= .
Define the function
(8) )). , ( , ( ) , ( u w w u w g
=
The Arrow-Pratt measures of comparative risk attitudes for agents and are as given
in the following proposition, which states that: (a) the absolute risk aversion measure for agent
is greater than the absolute risk aversion measure for agent , if and only if, (b) the risk
premium for agent is greater than the risk premium for agent , if and only if, (c) the
(Bernoulli) utility function for agent is a strictly increasing and strictly concave transformation
of the utility function of agent of the form given by definition (8) above and statement (c) in
the proposition. A version of this proposition that applies to the EUI model is derived simply by
specifying that
j
, and hence all related functions, are constant functions of w. A proof of the
proposition is contained in the appendix.
8
Proposition 1. If
a
and
A
g
g =
(
=
Statement (a.1) implies that ), , ( , 0 ) , (
2
u w u w g > because 0 ) , (
2
> y w
and 0 ) , (
2
> y w
,
) , ( y w . Statement (a.3) implies that ), , ( , 0 ) , (
22
u w u w g < if and only if
). , ( ), , ( ) , ( y w y w A y w A >
We next show that (c) (b). One has
(a.4)
)). , ( ( ))) , ( , ( , ( ( ))) , ( , ( (
))) , ( ( , ( )))) , ( ( , ( , (
y w E y w w w E y w w g E
y w E w g y w E w w
F F
= = >
=
Therefore
(a.5)
). , ( ) ( ))) , ( ( , (
)))) , ( ( , ( , ( , ( ))) , ( ( , ( ) , ( ) (
F w y E y w E w
y w E w w w y w E w F w y E
F F
F F F
= >
= =
Therefore (c) (b).
We next show that (b) (c). One has
(a.6)
)). , ( ( )))) , ( ( , ( , ( )) , ( ) ( , (
)) , ( ) ( , ( )))) , ( ( , ( , (
y w E y w E w w F w y E w
F w y E w y w E w w
F F F
F F
= = >
=
Therefore
32
(a.7)
))). , ( , ( ( )))) , ( , ( , ( ( )) , ( (
)))) , ( ( , ( , ( ))) , ( ( , (
y w w g E y w w w E y w E
y w E w w y w E w g
F F F
F F
= = >
=
Therefore, g is strictly concave in ). , ( , 0 ) , ( :
22
u w u w g u <
Proof of Proposition 2.c.
Let V(a) denote the expected utility of investing an amount, a in the risky asset with cumulative
distribution function, F for the random rate of return per dollar invested, R ; that is.
(a.8) )) , ( ( ) ( ar w E a V
F
= .
Assume that 0 ) ( > r E
F
and hence 0 > a . Let a(w) be the optimal choice of a as a function of
the initial wealth, w; this will satisfy the first-order condition as an identity:
(a.9) 0 ) ) , ( ( ) (
2
= r ar w E a V
F
.
Differentiating with respect to w gives
(a.10)
) ) , ( (
) ) , ( (
2
22
21
r ar w E
r ar w E
dw
da
F
F
= .
Since 0 ) , (
22
< ar w , one has
(a.11) ( ) ) ) , ( (
21
r ar w E sign
dw
da
sign
F
=
|
.
|
\
|
.
We show that if ) , ( ) , (
2 21
y w y w decreases in y then ( ) 0 ) ) , ( (
21
> r ar w E sign
F
, and
hence from (a.11) the amount, a invested in the risky asset increases. The other cases can be
derived by similar arguments.
Consider first the case where r > 0. Since ) , ( ) , (
2 21
y w y w decreases in y, one has
(a.12)
) 0 , (
) 0 , (
) , (
) , (
2
21
2
21
w
w
ar w
ar w
< ,
which can be rewritten as
33
(a.13) ) , (
) 0 , (
) 0 , (
) , (
2
2
21
21
ar w
w
w
ar w
> .
Multiplying both sides by r > 0, one has
(a.14) r ar w
w
w
r ar w ) , (
) 0 , (
) 0 , (
) , (
2
2
21
21
> .
A similar argument shows that the inequality in (a.14) is satisfied for 0 < r as well. Therefore,
taking expectations over all values of the random rate of return, and using (a.9), gives one
(a.15) 0 ) ) , ( (
) 0 , (
) 0 , (
) ) , ( (
2
2
21
21
= > r ar w E
w
w
r ar w E
F F
.
34
Table 1. If Averse to 50-50 Lose $100 / Gain $110, Will Reject or Accept 50-50 $Lose /$Gain Bets
$ Lose $ Gain
Table I Rejections Table II Rejections Eq. (10) Acceptances Eq. (11) Acceptances
400 550 550 655 690
600 990 990 1,000 1,040
800 2,090 2,090 1,350 1,385
1,000 718,190 1,695 1,730
2,000 12,210,880 3,425 3,460
4,000 60,528,930 6,890 6,925
6,000 180,000,000 10,355 10,390
8,000 510,000,000 13,820 13,855
10,000 1,300,000,000 17,285 17,320
20,000 160,000,000,000 34,605 34,640
35
Table 2. Ten Paired Lottery-Choice Decisions
Option A
Option B
Expected Payoff
Difference
1/10 of $40.00, 9/10 of $32.00
2/10 of $40.00, 8/10 of $32.00
3/10 of $40.00, 7/10 of $32.00
4/10 of $40.00, 6/10 of $32.00
5/10 of $40.00, 5/10 of $32.00
6/10 of $40.00, 4/10 of $32.00
7/10 of $40.00, 3/10 of $32.00
8/10 of $40.00, 2/10 of $32.00
9/10 of $40.00, 1/10 of $32.00
10/10 of $40.00, 0/10 of $32.00
1/10 of $77.00, 9/10 of $2.00
2/10 of $77.00, 8/10 of $2.00
3/10 of $77.00, 7/10 of $2.00
4/10 of $77.00, 6/10 of $2.00
5/10 of $77.00, 5/10 of $2.00
6/10 of $77.00, 4/10 of $2.00
7/10 of $77.00, 3/10 of $2.00
8/10 of $77.00, 2/10 of $2.00
9/10 of $77.00, 1/10 of $2.00
10/10 of $77.00, 0/10 of $2.00
$23.30
$16.60
$9.90
$3.20
-$3.50
-$10.20
-$16.90
-$23.60
-$30.30
-$37.00
36
Figure 1. Expected Utility Indifference Curves in the Triangle Diagram
increasing preference
1
1 0
) (
1 1
x prob P =
) (
3 3
x prob P =
37
Figure 2. Loss Aversion
38
Figure 3. Risk-Averse Choices from Pairs of Binary Lotteries
0
0,2
0,4
0,6
0,8
1
1,2
1 2 3 4 5 6 7 8 9 10
Row
P
r
o
p
o
r
t
i
o
n
o
f
O
p
t
i
o
n
A
C
h
o
i
c
e
s