Starmer-Developments in Non-Expected Utility Theory

American Economic Association
Developments in Non-Expected Utility Theory: The Hunt for a Descriptive Theory of

Choice under Risk
Author(s): Chris Starmer
Source: Journal of Economic Literature, Vol. 38, No. 2 (Jun., 2000), pp. 332-382
Published by: American Economic Association
Stable URL: https://www.jstor.org/stable/2565292
Accessed: 03-09-2019 08:52 UTC
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms
American Economic Association is collaborating with JSTOR to digitize, preserve and extend
access to Journal of Economic Literature
This content downloaded from 137.120.215.164 on Tue, 03 Sep 2019 08:52:00 UTC
All use subject to https://about.jstor.org/terms
Journal of Economic Literature
Vol. XXXVIII (June 2000) pp. 332-382
Developments in Non-Expected Utility

Theory: The Hunt for a Descriptive
Theory of Choice under Rislk
CHRIS STARMER1
1. Introduction stimulated developments in non-EU is

surely of widespread concern: put bluntly,
OW MANY THEORIES of decision the standard theory did not fit the facts.
making under risk and uncertainty As the standard theory of individual
can you think of? Readers of this article decision making, and as a core component
will no doubt be familiar with Expected of game theory, EUT constitutes a key
Utility Theory (EUT), the standard the- building block of a vast range of eco-
ory of individual choice in economics. nomic theory. It should be no surprise,
Many, I expect, will know of a few alter- therefore, that developing a better un-
natives to this model. But how many, I derstanding of the determinants of indi-
wonder, will be aware that these so- vidual choice behavior seemed a natural
called non-expected utility models now research priority to many theorists.
number well into double figures? An Around two decades of quite intensive
enormous amount of theoretical effort research on the topic has generated a
has been devoted towards developing al- great deal of theoretical innovation plus
ternatives to EUT, and this has run a much richer body of evidence against
hand-in-hand with an ongoing experi- which models can be judged. There can
mental program aimed at testing those be few areas in economics that could
theories. The good and proper division claim to have sustained such a rich in-
of labor suggests that a relatively small teraction between theory and evidence
group of specialists will be fully aware of in an ongoing effort to develop theories
the details of this literature. At the same in closer conformity with the facts.
time, the implications of developments Considered together, the accumulated
in this field are of more than passing in- theory and evidence present an oppor-
terest to the general economist, since what tunity to reflect on what has been
1 University of East Anglia, Norwich. I owe achieved. Perhaps the most obvious
thanks to Colin Camerer, Robin Cubitt, Mark question to address to this literature is
Machina, John Quiggin, Uzi Segal, Robert Sug- this: has it generated, or does it show
den, Peter Wakker, and George Wu, plus an
anonymous referee for extremely helpful com- the prospect of generating, a serious
ments on and discussions around this paper. I am contender for replacing EUT, at least
also grateful for support from the Economic and for certain purposes? If the question
Social Research Council of the UK (Risk and Hu-
man Behaviour Research Programme: Award No. seems disarmingly straightforward, pro-
L211252053). viding a clear-cut answer will not be.
332
Starmer: Developments in Non-Expected Utility Theory 333
Identifying a "best theory" naturally re- experimental testing, (2) how new theo-
quires judgements about the relative ries may help us to explain a range of
importance of predictive accuracy, sim- phenomena "in the field," and (3)
plicity, tractability, and so on. Such whether non-expected utility theory of-
judgments are complicated by the fact fers a viable alternative to EUT for
that the evidence, much of which de- everyday theoretical use. In the penulti-
rives from the experimental paradigm, mate Section 6, I discuss two emerging
is open to different interpretations. lines of enquiry which I see as particu-
In what follows, my aim will be to set larly exciting paths for future research.
out what I take to have been key A final section offers some concluding
theoretical developments in the area, to reflections.
review the related evidence and draw
conclusions about the current state of 2. Where It Began
play and the prospects for the future. In Although the primary purpose of this
doing so, rather than simply to present paper is to review alternatives to EUT,
an exhaustive list of models, my aim will that theory provides the natural point of
be to identify and discuss different departure, since most of the theories I
modeling strategies picking specific will be discussing can be understood as
models as illustrations. I also intend to generalizations of this base theory.2
narrow my sights in two significant re- EUT was first proposed by Daniel Ber-
spects. First, my focus will be on de- noulli (1738) in response to an apparent
scriptive as opposed to normative is- puzzle surrounding what price a reason-
sues. Second, I will concentrate on the able person should be prepared to pay
problem of modeling choices under risk to enter a gamble. It was the conven-
as opposed to the more general cate- tional wisdom at the time that it would
gory of uncertainty (the distinction is be reasonable to pay anything up to the
explained in the next section). Clearing expected value of a gamble, but Ber-
the ground in this way will, I hope, noulli presents this counterexample. A
sharpen the focus on one central re- coin is flipped repeatedly until a head is
search problem which continues to mo- produced; if you enter the game, you
tivate much of the research in this receive a payoff of, say, $2n where n is
arena: the endeavor to develop a "satis- the number of the throw producing the
factory" account of actual decision be- first head. This is the so-called St. Pe-
havior in situations of risk. It will be a tersburg game. It is easy to see that its
personal view, but one which I hope expected monetary payoff is infinite, yet
will help the interested nonspecialist Bernoulli believed most people would
find a trail through this expansive and only be prepared to pay a relatively
quite detailed literature. small amount to enter it, and he took this
The paper is organized as follows. intuition as evidence that the "value" of
Sections 2 and 3 set the scene with dis- a gamble to an individual is not, in gen-
cussions of the standard theory and the eral, equal to its expected monetary
evidence that prompted theorists to value. He proposed a theory in which
look for alternatives. Section 4 provides individuals place subjective values, or
the core overview of non-expected util- "utilities," on monetary outcomes and
ity theories. Section 5 seeks to evaluate
2 I shall not dwell on this account of EUT. For
what has been achieved so far, and in
those interested in further discussion an excellent
three subsections I discuss (1) how new starting place is Paul Schoemaker's (1982) review
theories have fared in a second phase of in this journal.
334 Journal of Economic Literature, Vol. XXXVIII (June 2000)
the value of a gamble is the expectation ternative models of choice, it will be

of these utilities. While Bernoulli's the- useful to present one set of axioms from
ory-the first statement of EUT- which EUT can be derived. In the ap-
solved the St. Petersburg puzzle, it did proach I adopt, at least to begin with,
not find much favor with modern preferences are defined over prospects
economists until the 1950s. This is where a prospect is to be understood as
partly explained by the fact that, in the a list of consequences with associated
form presented by Bernoulli, the theory probabilities. I will assume throughout
presupposes the existence of a cardinal that all consequences and probabilities
utility scale; an assumption that did not are known to the agent, and hence, in
sit well with the drive towards ordinaliza- choosing among prospects, the agent
tion during the first half of the twentieth can be said to confront a situation of
century. risk (in contrast to situations of uncer-
Interest in the theory was revived tainty in which at least some of the out-
when John von Neumann and Oskar comes or probabilities are unknown). I
Morgenstern (1947) showed that the ex- will use lowercase letters in bold (e.g.
pected utility hypothesis could be de- q, r, s) to represent prospects, and the
rived from a set of apparently appealing letter p to represent probabilities (take
axioms on preference. Since then, nu- it that p always lies in the interval
merous alternative axiomatizations have [0,1]). A given prospect may contain
been developed, some of which seem other prospects as consequences, but
highly appealing, some might even say assuming that such compound prospects
compelling, from a normative point of can be reduced to simple prospects fol-
view (see for example Peter Hammond lowing the conventional rules of prob-
1988).3 To the extent that its axioms can ability, any prospect q can be repre-
be justified as sound principles of ra- sented by a probability distribution q=
tional choice to which any reasonable (pi, . , pn) over a fixed set of pure
person would subscribe, they provide consequences X = (xi, . . . , Xn) where
grounds for interpreting EUT norma- pi is the probability of xi, pi ? 0 for all i,
tively (as a model of how people ought and zipi = 1. Hence, the elements of X
to choose) and prescriptively (as a prac- are to be understood as an exhaustive
tical aid to choice). My concern, how- and mutually exclusive list of possible
ever, is with how people actually choose, consequences which may follow from a
whether or not such choices conform particular course of action. While this
with a priori notions of rationality. Con- notation allows a prospect to be written
sequently, I will not be delayed by simply as vector of probabilities (as q
questions about whether particular axi- above) it will sometimes be useful to be
oms can or cannot be defended as explicit about the consequences too
sound principles of rational choice, and (e.g. by writing q = (xl, pi; . . .; X, pn))
I will start from the presumption that Given these preliminaries, the ex-
evidence relating to actual behavior pected utility hypothesis can be derived
should not be discounted purely on the from three axioms: ordering, continuity,
basis that it falls foul of conventional and independence. The ordering axiom
axioms of choice. requires both completeness and transi-
For the purpose of understanding al- tivity. Completeness entails that for all
q, r: either q > r or r > q or both where
3Such arguments, whilst widely accepted, are
nevertheless controversial. See, for example, Paul > represents the relation "is (weakly)
Anand (1992) and Sugden (1991). preferred to." Transitivity requires that
for all q, r, s: if q > r and r > s, then economics, so economists naturally

q > s. Continuity requires that for all pros- need a theory of individual decision
pects q, r, s where q > r and r > s: making under risk. EUT has much to
there exists some p such that (q, p; s, recommend itself in this capacity. The
1 - p) - r, where - represents the rela- theory has a degree of intuitive appeal.
tion of indifference and (q, p; s, 1 -p) It seems almost trivially obvious that
represents a (compound) prospect which any satisfactory theory of decision mak-
results in q with probability p; s with ing under risk will necessarily take ac-
probability 1 - p. Together the axioms of count of both the consequences of
ordering and continuity imply that pref- choices and their associated prob-
erences over prospects can be repre- abilities. These are, by definition, the
sented by a function V(.) which assigns dimensions relevant in the domain of
a real-valued index to each prospect. risk. EUT provides one very simple way
The function V(.) is a representation of combining probabilities and conse-
of preference in the sense that V(q) ? quences into a single "measure of
V(r) X q > r: that is, an individual will value" which has a number of appealing
choose the prospect q over the prospect properties. One such property is mono-
r if, and only if, the value assigned to q tonicity, which can be defined as fol-
by V(.) is no less than that assigned to r. lows. Let xi, . . . , xn be consequences
To assume the existence of some such ordered from worst (xl) to best (xv). We
preference function has seemed, to may say that one prospect q = (pql, . .
many economists, the natural starting pqn) first-order stochastically dominates
point for any economic theory of another prospect r = (pri, . . ., pmn) if
choice; it amounts to assuming that for all i=1,. . n:
agents have well-defined preferences, n n
while imposing minimal restriction on
the precise form of those preferences.
.Pqj 2 Y.Pj (2)
j=i j=i
For those who endorse such an ap-
proach, the natural questions center with a strict inequality for at least one i.
around what further restrictions can be Monotonicity is the property that sto-
placed on V(.). The independence ax- chastically dominating prospects are
iom of EUT places quite strong restric- preferred to prospects which they domi-
tions on the precise form of prefer- nate, and it is widely held that any satis-
ences: it is this axiom which gives the factory theory-descriptive or norma-
standard theory most of its empirical tive-should embody monotonicity. I will
content (and it is the axiom which most have more to say about this later.
alternatives to EUT will relax). Inde- The shape of the utility function also
pendence requires that for all prospects has a simple behavioral interpretation
q, r, s: if q > r then (q, p; s, 1 - p) > whereby concavity (convexity) of u(.)
(r, p; s, 1 - p), for all p. If all three axi- implies risk averse (prone) behavior; an
oms hold, preferences can be repre- agent with a concave utility function
sented by: will always prefer a certain amount x to
any risky prospect with expected value
V(q) = lipi . u(xi)(1
equal to x. Modeling risk preferences in
where q is any prospect, and u(.) is a this way does collapse some potentially
"utility" function defined on the set of distinct concepts into a single function:
consequences. any attitude to chance (e.g., like or
The concept of risk is pervasive in dislike of taking risks) and any attitude
towards consequences (e.g. a diminish- phenomena, first discovered by Maurice

ing marginal utility of money) must all Allais (1953), have played a particularly
be captured by the utility function. That important role in stimulating and shap-
need not imply any weakness of the the- ing theoretical developments in non-EU
ory. Indeed it is precisely the simplicity theory. These are the so-called common
and economy of EUT that has made it consequence effects and common ratio
such a powerful and tractable modeling effects. The first sighting of such effects
tool. My concern, however, is with the came in the form of the following pair
descriptive merits of the theory and, of hypothetical choice problems. In the
from this point of view, a crucial ques- first you have to imagine choosing be-
tion is whether EUT provides a suffi- tween the two prospects: si = ($1M,1)
ciently accurate representation of actual or ri = ($5M, 0.1; $1M, 0.89; 0, 0.01).
choice behavior. The evidence from a The first option gives one million U.S.
large number of empirical tests has dollars for sure; the second gives five
raised some real doubts on this score. million with a probability of 0.1; one
million with a probability of 0.89, other-
3. Descriptive Limitations of Expected wise nothing.5 What would you choose?
Utility Theory-The Early Evidence Now consider a second problem where
you have to choose between the two
Empirical studies dating from the prospects: S2 = ($1M, 0.11; 0, 0.89) or r2
early 1950s have revealed a variety of = ($5M, 0.1; 0, 0.9). What would you do
patterns in choice behavior that appear if you really faced this choice?
inconsistent with EUT. I shall not at- Allais believed that EUT was not an
tempt a full-blown review of this evi- adequate characterization of individual
dence.4 Instead, I discuss one or two ex- risk preferences and he designed these
amples to illustrate the general nature problems as a counterexample. As we
of this evidence, and offer a discussion shall shortly see, a person with expected
of its role in stimulating the develop- utility preferences would either choose
ment of new theories. With hindsight, it both "s" options, or choose both "r" op-
seems that violations of EUT fall under tions across this pair of problems. Allais
two broad headings: those which have expected that people faced with these
possible explanations in terms of some choices might opt for si in the first
"conventional" theory of preferences problem, lured by the certainty of be-
and those which apparently do not. The coming a millionaire, and select r2 in
former category consists primarily of a the second choice where the odds of
series of observed violations of the in- winning seem very similar, but the
dependence axiom of EUT; the latter of prizes very different. Evidence quickly
evidence that seems to challenge the as- emerged that many people did respond
sumption that choices derive from well- to these problems as Allais had pre-
defined preferences. Let us begin with dicted. This is the famous "Allais para-
the former. dox" and it is one example of the more
There is now a large body of evidence general common consequence effect.
indicating that actual choice behavior Most examples of the common conse-
may systematically violate the indepen- quence effect have involved choices be-
dence axiom. Two examples of such tween pairs of prospects of the following
4 Those interested in more thorough reviews are

recommended to consult Schoemaker (1982) and, 5 In Allais' original examples, consequences
more recently, Colin Camerer (1995). were French Francs.
form: s* = (y, p; c, 1 -p) and r* = (q, p; s = (y, p; 0, 1 -p) and r** = (x, 2p; 0,
c, l-p), where q=(x, 2; 0, 1-X) and 1 - p) where x > y. Assume that the
O <X< 1.6 The payoffs c, x and y are ratio of "winning" probabilities (X) is
nonnegative (usually monetary) conse- constant, then for pairs of prospects of
quences such that x > y. Notice that this structure, EUT implies that prefer-
both prospects s* and r* give outcome ences should not depend on the value
c with probability 1 - p: this is the of p,9 yet numerous studies10 reveal a
"common consequence" and it is an ob- tendency for individuals to switch their
vious implication of the independence choice from s** to r** as p falls.
axiom of EUT that choices between s* It would, of course, be unrealistic to
and r* should be independent of the expect any theory of human behavior to
value of C.7 Numerous studies, however, predict accurately one hundred percent
have found that choices between pros- of the time. Perhaps the most one could
pects with this basic structure are sys- reasonably expect is that departures
tematically influenced by the value of c. from such a theory be equally probable
More specifically, a variety of experi- in each direction. These phenomena,
mental studies8 reveal a tendency for however, involve systematic (i.e., pre-
individuals to choose s* when c = y, and dictable) directions in majority choice.
r* when c = 0. As evidence against the independence
A closely related phenomenon, also axiom accumulated, it seemed natural
discovered by Allais, is the so-called to wonder whether assorted violations
common ratio effect. Suppose you had of it might be revealing some underly-
to make a choice between $3000 for ing feature of preferences that, if prop-
sure, or entering a gamble with an 80 erly understood, could form the basis of
percent chance of getting $4000 (other- a unified explanation. Consequently, a
wise nothing). What would you choose? wave of theories designed to explain the
Now think about what you would do if evidence began to emerge at the end of
you had to choose either a 25 percent the 1970s. Most of these theories have
chance of gaining $3000 or a 20 percent the following features in common: (i)
chance of gaining $4000. A good deal of preferences are represented by some
evidence suggests that many people function V(.) defined over individual
would opt for the certainty of $3000 in prospects; (ii) the function satisfies or-
the first choice and opt for the 20 per- dering and continuity; and (iii) while
cent chance of $4000 in the second. V(.) is designed to permit observed vio-
Such a pattern of choice, however, is in- lations of the independence axiom, the
consistent with EUT and would consti- principle of monotonicity is retained. I
tute one example of the common ratio will call theories with these properties
effect. More generally, this phenome-
non is observed in choices among pairs To see why, consider any pair of options (s**
of problems with the following form: r**) where p = pi, then define a further pair of
options (s2*, r**) identical except having a lower
value of p = p2. Since there must be some oc, (1 >
6 It will be convenient to use a scaling factor X oc> O), such that P2 = cxpl, we can write s2* =
at several points in the paper, so to avoid repeti- (s**, cc; O, 1 - oc) and e* = (r*, o; 0, 1- c). It then
tion, assume 0 < X < 1 throughout. follows directly from independence that choices
7 The original Allais problems are recovered between such pairs of prospects should not
from this generalization setting x = $5M; y = $1M, depend on the value of p.
p = 0.11 and k=10/11. 10 Examples include Loomes and Sugden
8 Examples include Herbert Moskowitz (1974), (1987), Starmer and Sugden (1989), and Raymond
Paul Slovic and Amos Tversky (1974), Kenneth Battalio, John Kagel, and Komain Jiranyakul
MacCrimmon and Stig Larsson (1979). (1990).
conventional theories. The general spirit the so-called preference reversal phe-
of the approach is to seek "well be- nomenon first observed by psychologists
haved" theories of preference consis- Sarah Lichtenstein and Paul Slovic
tent with observed violations of inde- (1971) and Harold Lindman (1971). It
pendence; I call this general approach presents a puzzle for economics be-
the conventional strategy. cause, viewed from the standard theo-
There is evidence to suggest that fail- retical perspective, both tasks consti-
ures of EUT may run deeper than viola- tute ways of asking essentially the same
tions of independence. Two assump- question, that is, "which of these two
tions implicit in any conventional theory prospects do you prefer?" In these ex-
are procedure invariance (preferences periments, however, the ordering re-
over prospects are independent of the vealed appears to depend upon the
method used to elicit them) and de- elicitation procedure.
scription invariance (preferences over One explanation for preference rever-
prospects are purely a function of the sal suggests that choice and valuation
probability distributions of conse- tasks may invoke different mental pro-
quences implied by prospects and do cesses which in turn generate different
not depend on how those given distri- orderings of a given pair of prospects
butions are described). While these as- (see Slovic 1995). Consequently, the
sumptions probably seem natural to most rankings observed in choice and valu-
economists-so natural that they are ation tasks cannot be explained with
rarely even discussed when stating for- reference to a single preference order-
mal theories-there is ample evidence ing. An alternative interpretation ex-
that, in practice, both assumptions fail. plains preference reversal as a failure of
One well-known phenomenon, often transitivity (see Graham Loomes and
interpreted as a failure of procedure in- Robert Sugden 1983): assuming that the
variance, is preference reversal. The valuation task reveals true monetary
classic preference reversal experiment valuations, (i.e., M($) - $; M(P) - P),
requires individuals to carry out two preference reversal implies P > $ -
distinct tasks (usually separated by M($) > M(P) - P; which involves a vio-
some other intervening tasks). The first lation of transitivity (assuming that
task requires the subject to choose be- more money is preferred to less). Al-
tween two prospects: one prospect though attempts have been made to ex-
(often called the $-bet) offers a small plain the evidence in ways which pre-
chance of winning a "good" prize; the serve conventional assumptions-see
other (the "P-bet") offers a larger for example Charles Holt (1986); Edi
chance of winning a smaller prize. The Karni and Zvi Safra (1987); Uzi Segal
second task requires the subject to as- (1988)-the weight of evidence sug-
sign monetary values-usually mini- gests that failures of transitivity and
mum selling prices denoted M($) and procedure invariance both contribute to
M(P)-to the two prospects. Repeated the phenomenon (Loomes, Starmer,
studiesll have revealed a tendency for and Sugden 1989; Tversky, Slovic, and
individuals to chose the P-bet (i.e., re- Daniel Kahneman 1990).
veal P > $) while placing a higher value There is also widespread evidence
on the $-bet (i.e., M($) > M(P)). This is that very minor changes in the presen-
tation or "framing" of prospects can
11 Reviews of this evidence are contained in
Tversky and Richard Thaler (1990), Daniel Haus- have dramatic impacts upon the choices
man (1992), and Timo Tammi (1997). of decision makers: such effects are
failures of description invariance. Here that choices can, in general, be repre-

is one famous example due to Tversky sented by any well behaved preference
and Kahneman (1981) in which two function. If that is right, they lie out-
groups of subjects-call them groups I and side the explanatory scope of the con-
II-were presented with the following ventional strategy. Some might even be
cover story: tempted to say they lie outside the
scope of economic theory altogether.
"Imagine that the U.S. is preparing for the
outbreak of an unusual Asian disease, which That stronger claim, however, is contro-
is expected to kill 600 people. Two alternative versial, and I will not be content to put
programs to combat the disease have been away such challenging evidence so
proposed. Assume that the exact scientific es-
swiftly. For present purposes, suffice it
timate of the consequences of the programs
to make two observations. First, whether
are as follows:"
or not we have adequate economic theo-
Each group then faced a choice between ries of such phenomenon, the "Asian
two policy options. disease" example is clearly suggestive
that framing effects have a bearing on
Options presented to group I:
issues of genuine economic relevance.
"If program A is adopted, 200 people will be
Second, there are at least some theories
saved.
of choice that predict phenomena like
If program B is adopted, there is a 1/3
preference reversal and framing effects,
probability that 600 people will be saved,
and some of these models have been
and a 2/3 probability that no people will be
saved. widely discussed in the economics lit-
erature. Although most of these theories-
Options presented to group II: or at least the ones I will discuss-draw on
"If program C is adopted, 400 people will ideas about preference to explain choices,
die.
they do so in unorthodox ways, and many
draw on concepts more familiar to psy-
If program D is adopted, there is a 1/3
probability that nobody will die, and a 2/3 chologists than economists. The one fea-
probability that 600 people will die." ture common to this otherwise heterodox
bunch of theories is that none of them
The two pairs of options are stochasti- can be reduced to, or expressed purely
cally equivalent. The only difference is in terms of, a single preference function
that the group I description presents the V(.) defined over individual prospects. I
information in terms of lives saved while will call such models nonconventional
the information presented to group II is theories. These theories step into what
in terms of lives lost. Tversky and has been relatively uncharted water for
Kahneman found a very striking differ- the economics profession. One of the
ence in responses to these two presenta- aims of this piece is to reflect on the
tions: 72 percent of subjects preferred relative merits of the conventional and
option A to option B while only 22 per- nonconventional approaches.
cent of subjects preferred C to D. Simi-
lar patterns of response were found 4. Non-Expected Utility Theories
amongst groups of undergraduate stu-
4.1 The Conventional Strategy
dents, university faculty, and practicing
physicians. One way to approach this literature is
Failures of procedure invariance and by asking a question that motivated a num-
description invariance appear, on the ber of theories: what properties would a
face of it, to challenge the very idea conventional theory of preference need
x3
P3t
S2
Si ~~~~~~~2 ~~~xi s
x 0.01 0.89 0.9 1r1
[xl PI 0, x2 =$1m,X3=$5m]igurel.ExpetedUtiityIndfferencCurveSis
[xI = 0, x2 = $'lm X3 = $5m] Figure lb. Expected Utility Indifference Cu'rves
Figure la. The Allais Paradox Problems in a Probability
Triangle ($5M) increasing from bottom to top.
Hence si, which results in the interme-
to have in order to explain the known
diate consequence of $1M for sure, is
violations of independence? In order to
located at the bottom left corner of the
pursue that question, it will be helpful
triangle; S2 and r2, which each assign
to introduce an expositional device
positive probability to only two of the
known as the probability triangle dia-
three possible consequences, are lo-
gram,12 and this will also prove useful as
cated on the triangle boundaries; while
a vehicle for comparing the predictions
ri, which assigns positive probability to
of alternative theories.
all three consequences, lies on the inte-
Consider the class of prospects de-
rior of the triangle. Two lines have been
fined over three outcomes xi, X2, X3 such
drawn in the triangle joining the pairs
that x1 < X2 < X3. Since any such pros-
of prospects involved in the two
pects can be described as a vector of
choices. It is easy to establish that these
probabilities (pi, l-PI-p3, p3) we can two lines are parallel.
also locate them, graphically, in two-di-
Given ordering plus continuity, pref-
mensional probability space. Figure la
erences over prospects in any given tri-
is a probability triangle that does this
angle can be represented by a set of
for the four prospects {sl,rl,s2,r2} from indifference curves, hence, every con-
the original Allais paradox problems. By
ventional theory implies the existence of
convention, the horizontal axis mea-
a set of indifference curves in this space
sures the probability of the worst conse-
though the precise form of indifference
quence ($0) increasing from left to
curves varies between them.
right; the vertical axis measures the
The addition of the independence ax-
probability of the best consequence
iom of EUT restricts the set of indiffer-
12 Although the probability triangle had ence ap- curves to being upward sloping (left
peared in the literature many years before (see to right) linear and parallel. One such
Jacob Marschak 1950) Mark Machina's use of it in set of indifference curves is illustrated
the 1980s (see below) popularized it to the extent
that some have called this diagram the "Machina in Figure lb (preferences are increasing
triangle." moving northwest). Independence is a
Best
Consequence
Indifference curve for the

Xre isk averse agent X
r~~~~~~~~~~~
s Worst p <1
Consequence
Figure Ic. Different Degrees of Risk Aversion in EUT sI S2
Figure 2. Common Ratio Prospects
strong restriction which leaves only one

feature of the indifference curves unde- could have a less steep slope (in which
termined, that is, their slope. In EUT, case ri > si and r2 > S2). Finally, the
the slope of the indifference curves re- slope of indifference curves could cor-
flects attitude to risk and may vary be- respond exactly with that of the lines
tween individuals: the more risk averse joining pairs of prospects, in which case
the individual, the steeper the slope of ri - si and r2 S S2. But as noted above,
their indifference curves. To see why, people often violate EUT, revealing si
look at Figure Ic and consider two in- > ri in the left-hand problem, r2 > S2
dividuals: person 1 has indifference in the right-hand problem. Relative to
curves with the slope of the dashed line the predictions of EUT, in choosing r2
(hence s - r); person 2 has indifference over S2 these people are being more risk
curves with the slope of the solid line seeking than they should be, given their
(hence s - r'). Person 2 can be seen to choice of si over ri.
be the more risk averse in the sense that,A similar tendency is apparent in the
as we move northwest along the hypote- common ratio effect. A pair of common
nuse, relative to person 1, we must give ratio problems is illustrated in Figure 2.
her a higher chance of winning the best The pair of prospects {s**, r*}, near the
outcome in the riskier prospect in order left edge of the triangle, correspond
to generate indifference with the safe with the common ratio problems where
prospect s. p = 1. As p falls, we generate pairs of
In relation to the Allais paradox prob- prospects like {s**, r2*1, located on paral-
lems in Figure lb, for a given individ- lel lines further to the right in the trian-
ual, EUT allows three possibilities. In- gle. Assuming expected utility prefer-
difference curves could have a steeper ences, an individual must either prefer
slope than the lines connecting pros- the "safer option" in both choices or the
pects, in which case si > ri and S2 > r2. "riskier option" in both choices, yet
This is the case represented in Figure many people choose s** over r** and
lb. Alternatively, indifference curves r** over s2*. This is the common ratio
effect and, as in the common conse-

quence effect, relative to the predic-
tions of EUT, there is an "inconsis-
tency" in the risk attitudes revealed
r
across their choices.
Viewed in the context of the triangle,
this inconsistency is suggestive of a sys-
I
tematic pattern: relative to the predic-
tions of EUT, choices between pros-
pects located in the bottom right-hand I~~~~~~~~~~
corner appear more risk prone than q
should be expected given preferences

revealed for choices located leftwards
and/or upwards in the triangle. Any
conventional theory seeking to explain
these standard violations of EUT will
therefore need at least one quite spe- m s
cific property: indifference curves de- Figure 3. Indiffer
termining preferences over pairs of Utility
prospects located near the right-hand
corner of a given triangle-like, say {s**,expected utility preference functions
r2*1 will need to be relatively flat (re- V(.), so that U(.;q) can be thought of as
flecting more risk-prone behavior), the "local utility function" of V(.) about
compared with indifference curves de- q. For example, the property "concavity
termining choices over pairs of pros- of U( .;q) at every q" is equivalent to
pects, like Is", rt*1, near to the left- global risk aversion of V(.).
hand edge of the triangle. All of the Given the existence of phenomena
proposed conventional alternatives to like the common ratio and common
EUT are able to generate this property, consequence effects, Machina hypothe-
though they do so in a variety of ways. sized that the local utility functions
4.1.1 The "Fanning-out" Hypothesis
U( .;q) become more concave as we
move from (first order) stochastically
Having observed this apparent con- dominated to stochastically dominating
nection between different violations of distributions. Loosely speaking, this es-
independence, Mark Machina (1982) sentially empirical assumption (which
proposed an analytical extension of EUT Machina calls "Hypothesis II") implies
(termed "generalized expected utility a tendency for agents to become more
analysis"), along with a specific hy- risk averse as the prospects they face
pothesis on the shape of non-expected get better; in the context of the trian-
utility indifference curves. Analytically, gle, it means that indifference curves
he noted that under expected utility, become steeper, or "fan out" as we
where V(q) = X U(xi). pi, the utility move val- northwest. Figure 3 illustrates the
ues U(xi) = 8V(q)/6pi are the probability general pattern of indifference curves
derivatives of V(.). He then showed that implied by Hypothesis II. Notice that
standard expected utility results (e.g., they are drawn as wavy lines: general-
risk aversion concavity of U(.)) also ized expected utility theory requires in-
hold for the probability derivatives difference curves to be smooth but does
U(xi;q) = 6V(q)/6pi of smooth non- not imply that they must be linear
(though they may be). It is very easy to

see that this fanning-out property gen-
erates implications consistent with the
common consequence and common ra-
tio effects. Since indifference curves
are relatively steeply sloped in the
neighborhood of prospect m, m lies on
a higher indifference curve than q or r.
Flatter indifference curves in the bot-
tom right-hand corner of the triangle
are such that t lies on a higher indiffer-
ence curve than s. Hence, for an indi-
vidual whose indifference curves fan out
we can construct prospects over which we Figure 4. Weighted Utility Theory with Indifference
will observe a common consequence effect Curves Fanning Out
(e.g. m > q and t > s) and a common
ratio effect (e.g. m > r and t > s). gram, excepting the special case of EUT,
A whole family of models have this this axiom has the effect of requiring
fanning-out property ak, within this there to be some point at which all indif-
family, one important subset consists of ference curves cross. The location of this
those models that restrict indifference point, which could lie inside or outside
curves to be linear. One example is Soo of the triangle boundary, depends upon
Hong Chew and MacCrimmon's (1979) the specifications of the functions u(.),
weighted utility theory in which prefer- and g(.). Transitivity can be preserved by
ences over prospects are represented by making the point from which curves ra-
the function: diate lie outside the boundary of the tri-
V(q) = [-pi . g(xi) . u(xi)] / [-pi . g(xi)] (3), angle and, in order to explain the com-
mon ratio and common consequence
where u(.) and g(.) are two different effects, the origin of indifference curves
functions assigning non-zero weights to must lie somewhere to the southwest of
all -consequences. The model incorpo- the triangle, as in Figure 4. Having re-
rates EUT as the special case in which stricted the model in this way,'3 we can
the weights assigned by g(.) are identical then understand it as a special case of
for every consequence. Weighted utility Machina's theory (including Hypothesis
has been axiomatized by, among others, II) in which indifference curves are
Chew and MacCrimmon (1979), Chew constrained to be linear.
(1983), and Peter Fishburn (1983), and It is not obvious to me that weak in-
different variants are discussed in Fish- dependence has much, if any, intuitive
burn (1988). Essentially these axiomati- appeal, and the main rationale for as-
zations involve a weakened form of the suming it in weighted utility theory is
independence axiom which constrains in- presumably that it results in a simple
difference curves to be linear without re- mathematical function capable of gen-
quiring them to be parallel. One version erating fanning out and hence explain-
of weak independence is this: if q > r ing the early violations of EUT. Other
then for each pq there exists a corre- models with very similar properties
sponding pr such that (q, pq; s, 1 - pq) >
(r, pr; s, 1 - pr) for all s. If we think in 13 Chew and MacCrimmon (1979b) explain the
terms of preferences in the triangle dia- conditions necessary to generate this property.
have been based on psychologically

grounded hypotheses. One example is
the theory of disappointment developed
by David Bell (1985) and Loomes and
Sugden (1986). While this theory lacks
axiomatic foundations, it has a more ob-
(q,p;r, l-p)
vious intuitive interpretation. In the
r
version presented by Loomes and Sug-
den, preferences over prospects can be
represented by the function:
V(q) = lipi[u(xi) + D(u(xi) - U)] (4), q
where u(xi) is interpreted as a measure

of "basic" utility (that is, the utility of xi,
considered in isolation from the. other Figure 5a. Probability Mixtures of Prospects q, r
consequences of q) and U is a measure
of the "prior expectation" of the utility A series of other models with linear
from the prospect. The model assumes indifference curves including implicit
that if the outcome of a prospect is expected utility (Eddie Dekel 1986) and
worse than expected (i.e., if u(xi) < U) a implicit weighted utility (Chew 1989)
sense of disappointment will be gener- allow fanning out, but also permit more
ated. On the other hand, an outcome complex patterns. For example, Faruk
better than expected will stimulate "ela- Gul (1991) and William Neilson (1992)
tion." With D(.) = 0, the model reduces present models based on implicit ex-
to EUT. This additional function, how- pected utility which generate a mixture
ever, is intended to capture a particular of fanning-in and fanning-out within a
intuition about human psychology: that given triangle.14 The crucial axiom in
people dislike disappointment and so act these models is a weakened form of in-
to avoid it. More specifically, this is cap- dependence called betweenness: if q > r,
tured by assuming that agents are "disap- then q > (q, p; r, (1-p)) > r for all
pointment averse" (D(h) is concave for p < 1. It is this assumption that imposes
h < 0) and "elation prone" (D(.) is convex linearity on indifference curves and,
for h > 0). The theory then implies a ten- conversely, it is implied by any model
dency for indifference curves to fan out that assumes linear indifference curves.
in the triangle. The theory of disappoint- Behaviorally, betweenness implies that
ment has close affinity with earlier mod- any probability mixture of two lotteries
els based on moments of utility. In EUT, will be ranked between them in terms
the value of a prospect is the (probability of preference and, given continuity, an
weighted) mean of utility. Allais (1979) individual will be indifferent to random-
proposed a model in which V(.) may also ization among equally valued prospects.
depend on the second moment of utility, To understand the connection between
that is, the variance of utility about the these behavioral and geometric proper-
mean. Hagen (1979) extended this idea ties, look at Figure 5a and consider an
to include the third moment of utility, or individual offered a compound gamble
skewness. Sugden(1986) shows that prop-
erties of D(.) imposed in disappointment 14 These models were proposed in response to
later evidence (see Section 5) which suggests be-
theory can be interpreted as restrictions havior is more complex than pure fanning-out
on Hagen's general model of moments. theories imply.
~~~~~ \
Figure 5c. Quasi-concave Preferences. Preference for
Randomization
Figure 5b. Quasi-convex Preferences. Aversion to
Randomization
preferences are strictly quasi-concave,
giving a p chance of prospect q and a indifference curves are convex, as in
1 - p chance of r. Geometrically, the Figure 5c, hence by similar reasoning
simple prospect induced by this com- individuals prefer to randomize among
pound gamble must lie along the straight equally valued prospects. Some signifi-
line joining q and r (for any 0 < p < 1). cant theoretical results in economics ex-
For an individual with linear indifferencetend to a non-expected utility world if
curves, it follows that for any q r, the agents' preferences satisfy betweenness
indifference curve through q and r co- (see Section 5.3 below).
incides with the set of simple prospects Various models have been proposed
induced by (q, p; r, 1 -p). Hence, with that do not impose betweenness. Chew,
linear indifference curves, the individ- Larry Epstein, and Segal (1991) pro-
ual indifferent between q and r is also pose quadratic utility theory which re-
indifferent to randomization between lies on a weakened form of betweenness
them. Once betweenness is relaxed, this called mixture symmetry: if q - r then
indifference to randomization no longer (q, p; r, (1 -p)) - (q, (1 -p); r, p). In
holds and two important cases can be this model, indifference curves may
distinguished: quasi-convex preferences switch from concave to convex (or vice
and quasi-concave preferences. A pref- versa) as we -move across the triangle.
erence function is strictly quasi-convex Joao Becker and Rakesh Sarin (1987)
if for every q ? r, V(q, p; r, (1 - p)) < propose a model with even weaker re-
max[V(q) ,V(r)] for all p. When prefer- strictions. Their lottery-dependent util-
ences are quasi-convex, indifference ity assumes only ordering, continuity,
curves are concave, as in Figure 5b, and and monotonicity. The basic model is
consequently the individual will be conventional theory for minimalists as,
averse to randomization among equally without further restriction, it has virtu-
valued prospects (notice that prospects ally no empirical content. The authors
r and s in Figure 5b lie on a higher in- discuss a particular "exponential form"
difference curve than probability mix- which implies fanning-out.
tures of the two prospects which lie An important subset of the between-
along the dashed line). Conversely, when ness non-conforming theories has an
additional feature absent from the mod- such effects could be captured in mod-
els discussed so far. To this point we els incorporating decision weights. A
have considered a variety of conven- number of such theories can be under-
tional theories, each of which generates stood as variants of the following func-
the property of fanning-out. Although tional form where the wi terms represent
.they achieve it in different ways, there decision weights:
is one structural similarity between these
theories: each operates by assigning V(q) = liwi . u(xi). (5)
subjective weights-or utilities-to con- I will call this the decision weighted form.
sequences; the value assigned to any Theories of this type were first discussed
given prospect is then determined by by Ward Edwards (1955, 1962). In its
some function that combines these utili- most basic form, consequences are treated
ties with objective probabilities. An- in the way that probabilities are handled
other variant of the conventional strat- in the standard theory and enter "raw"
egy involves the use of probability with u(xi) =xi for all i. Edwards called
transformation functions which con- this subjective expected value, and in the
vert objective probabilities into subjec- version presented by Jagdish Handa
tive decision weights. An important (1977) the decision weight attached to
feature of these models is that, except- each outcome is determined by a prob-
ing special cases, betweenness does not ability weighting function tc(pi) which
hold. transforms the individual probabilities of
4.1.2 Theories with Decision Weights each consequence directly into weights.
As in most theories that incorporate prob-
There is evidence for the view that
ability weights, *t.) is assumed to be in-
individuals have subjective attitudes to creasing with sT(1) = 1 and sT(O) = 0, and I
probabilities which are distinct from at- will retain these assumptions from now on.
titudes to consequences. For instance, The subjective expected value form has not
according to Nick Pidgeon et al. (1992), been widely used, but theories that allow
when people are asked to make judge- nonlinear transformations of both prob-
ments about the likelihood of death oc- abilities and consequences have received
curring from different causes, they tend much more attention. In the simplest vari-
to underestimate the number of deaths ant of this latter type of model, individuals
from relatively frequent causes, while are assumed to maximize the function:
overestimating deaths due to relatively
infrequent causes. Similarly, apparent V(q) = In(pi) . u(xi). (6)
biases in the subjective odds revealed in I will call this form simple decision
studies of racetrack betting have been weighted utility. 15 Both this and subjec-
explained as bettors being either oversen- tive expected value, because they trans-
sitive to the chances of winning on long form the probabilities of individual con-
shots (Mukhtar Ali 1977; Richard sequences directly into weights (i.e., wi =
Thaler and William Ziemba 1988), or lt(pi)), have the property that V(q) will
oversensitive to the chances of losing on not generally satisfy monotonicity. To
favorites (Bruno Jullien and Bernard see this, suppose for the sake of example
Salanie 1997). These effects might be that n(.) is convex, then apart from the
revealing misperception of objective
15This form has sometimes been called subjec-
probabilities or a tendency for individu-
tive expected utility, but this label is now more
als to subjectively weight objective commonly used to refer to Leonard Savages'
probabilities. Either way, in principle, (1954) formulation of EUT.
extremes of the probability scale, st(p) + dependent model as "the most natural and
t(l -p) < 1 and there will be some ? > O useful modification of the classical ex-
such that gambles of the form (x, p; x + E,pected utility formula" and, as testament
1-p) will be rejected in favor of (x,1) to this, it has certainly proved to be one
even though they stochastically dominate of the most popular among economists.
the sure option. A similar argument ap- In this type of model the weight attached
plies for any departure from linearity, to any consequence of a prospect de-
and the only way to ensure general pends not only on the true probability
monotonicity in this type of theory is to of that consequence but also on its rank-
set decision weights equal to objective ing relative to the other outcomes of the
prospect.
probabilities (i.e., Wi = It(pi) = pi for all i) With consequences indexed as
in which case the theory reduces to before such that xi is worst and xn best,
EUT. This property was first noted by we can state rank-dependent expected
Fishburn (1978) and since then has been utility theory as the hypothesis that
widely viewed as a fatal objection to agents maximize the decision weighted
models that attach decision weights to form with weights for i = 1,..., n - 1
the raw probabilities of individual conse- given by:
quences. For example, Machina (1983, p.
Wi=7t(pi+ * +pn)-(pi+1+ .** +pn)
97) argues that any such theory will be,
and Wi = 7(pi) for i = n.
"in the author's view at least, unaccept-
able as a descriptive or analytical model In this model there is a meaningful dis-
of behavior." The point seems to have tinction between decision weights (w)
been generally accepted and, while many and probability weights (t). Richard
theorists have wished to retain the idea Gonzalez and George Wu (1999, p. 135)
that probabilities may be subjectively suggest an interpretation of the prob-
weighted, the thrust of work in this ability weighting function as reflecting
stream of the literature over the last two the underlying "psychophysics of risk,"
decades has been towards variants of the that is, the way that individuals subjec-
decision weighted form that satisfy tively "distort" objective probabilities;
monotonicity. the decision weight then determines how
There are two distinct strands to this the probability weights enter the value
contemporary literature: one conven- function V(.). Notice that t(pi +. . . +
tional, the other distinctly nonconven- pn) is a subjective weight attached to the
tional. The nonconventional route is that probability of getting a consequence of xi
taken by Kahneman and Tversky (1979) or better, and 7t(pi +1 + . . . + pn) is a
in prospect theory, but that model takes weight attached to the probability of get-
us outside the bounds of conventional ting a consequence better than xi, hence
theory and so I postpone further discus- in this theory *t.) is a transformation on
sion of it until the next section. Theo- cumulative probabilities. This procedure
rists following the conventional route for assigning weights ensures that V(.) is
have proposed decision weighting models monotonic. It also has the appealing
with more sophisticated probability trans- property that, in contrast to the simple
formations designed to ensure mono- decision weighting models which assign
tonicity of V(.). One of the best-known the same decision weight to any conse-
models of this type is rank-dependent quence with probability p, the weight at-
expected utility theory, which was first tached to a consequence may vary ac-
proposed by John Quiggin (1982). cording to how "good" or "bad" it is. So
Machina (1994) describes the rank- in principle this would allow for, say,
extreme outcomes to receive particularly

1 _________________ _--
high (or low) weights. A less appealing
//
feature of the model is that a small
change in the value of some outcome of
a prospect can have a dramatic effect on
its decision weight if the change affects
the rank order of the consequence; but a
change in the value of an outcome, no
matter how large the change, can have ,' /zp)
no affect on the decision weight if it
0~~~~~~~~~~~~~
does not alter its rank.
The predictions of the rank-depen-
dent model depend crucially on the
form of t(.). If t(.) is convex, this gen-
erates a set of concave indifference
curves (implying aversion to randomiza-
tion) which are parallel at the hypote-
nuse but fan out as we move left to
right across the triangle and fan in (i.e., Figure 6. An (inverted) S-shaped Probability Weighting
Function
become less steep) as we move verti-
cally upwards. Aside from the hypote-
nuse parallelism which holds for any cave u(.) will be universally risk averse;
sT(.) (see Camerer 1989), the reverse and an ag ent with a convex utility func-
pattern of indifference curves (i.e., con- tion can be risk averse if they are suffi-
vex curves, horizontal fanning in, and ciently pessimistic (See Chew, Karni,
vertical fanning out) is generated with a and Safra 1987; Alain Chateauneuf and
concave K(.). Mich'elle Cohen 1994).
Curvature of K(.) in the rank-depen- Although rank-dependent theory does
dent model has been interpreted as re- not imply generalized fanning out, the
flecting "optimism" and/or "pessimism" early evidence of EUT violation can be
with respect to probabilities (see Quig- explained either by assuming a simple
gin 1982; Menahem Yaari 1987; Enrico convex n(.) or by more complex specifi-
Diecidue and Peter Wakker 1999). Con- cations. One possibility is the function
sider, for example, the prospect q = (xi, displayed in Figure 6 which has Tc(p) .= p
0.5; X2, 0.5). Assigning weights to the for a unique value of p = p *; it is con-
consequences of q according to the cave below p * and convex above it,
rank-dependent method above gives hence "low" probabilities (below p * ) are
WI = 1 - 0(O.5) and W2 = st(O.5). With overweighted. Quiggin (1982) proposes
*(.) convex, t(0.5) < 0.5, hence the this form with p * = 0.5. He is drawn to
weight attached to the lower ranking this partly because it explains the early
consequence, xI, will be higher than the violations of EUT and partly because it
weight attached to the larger conse- has the appealing property that 50-50
quence. This overweighting of the bets will be undistorted by probability
lower-ranked consequences relative to weighting. While there is little empiri-
higher-ranked consequences can be in- cal support for the crossover at p = 0.5,
terpreted as a form of pessimism. Pessi- research over a period of fifty years,
mism also has a close connection to risk from Malcolm Preston and Phillip
aversion: a pessimistic agent with a con- Baratta (1948) to Drazen Prelec (1998),
lends support to the hypothesis of an two prospects q, r have a "common tail"

(inverted) s-shaped decision-weighting such that for some j, pqi = pri for all i
function (see Section 5.1.1). A useful from j to n. Ordinal independence re-
discussion of the theoretical properties quires that preferences between q and
necessary and sufficient for an s-shaped r be unaffected by the substitution of
weighting function can be found in this common tail, in both prospects,
Tversky and Wakker (1995). with any other common tail. This axiom
Axiomatizations of rank-dependent is necessary for any rank-dependent
expected utility have been presented model. The contribution of Chew and
by, among others, Segal (1990), Wakker Epstein constructs a theoretical bridge
(1994), Mohammed Abdellaoui (1999), between the rank-dependent models
and Yaari (1987), who examines the spe- and the betweenness-conforming theo-
cial case of the model with linear utility ries (i.e., those with linear indifference
(this is essentially a rank-dependent re- curves discussed above) by presenting a
formulation of Handa's proposal with general model which contains each class
U(xi)=xi). Wakker, Ido Erev, and Elke as a special case (see also the "correc-
Weber (1994) provide a useful discus- tion and comment" by Chew, Epstein,
sion of the axiomatic foundations of rank- and Wakker 1993).
dependent expected utility in which A further extension to the rank-de-
they demonstrate the essential differ- pendent model discussed by Starmer and
ence between EUT and rank-dependent Sugden (1989), Luce and Fishburn
expected utility is that the latter theory (1991), and Tversky and Kahneman
relies on a weakened form of indepen- (1992) involves a distinction between
dence called co-monotonic indepen- consequences that are "gains" and those
dence. It is an implication of the stan- that are "losses." This approach draws
dard independence axiom that if two on Kahneman and Tversky's earlier work
prospects q and r have a common out- on prospect theory. It is to this model
come x, which occurs with probability ,that we now turn, and in doing so we
p, in each prospect, then substituting x cross the boundary into nonconventional
for some other outcome y in both pros- territory.
pects will not affect the preference or-
4.2 Nonconventional Theories
der of q and r. The same may not be
true in the rank-dependent model, how- 4.2.1 The Procedural Approach and
ever, because such substitutions may af- Reference Dependence
fect the rankings of consequences and
hence the decision weights. Co-mono- Each of the theories we have consid-
tonic independence asserts that prefer- ered so far models choice as preference
ences between prospects will be unaf- maximization and assumes that agents
fected by substitution of common behave as if optimizing some underly-
consequences so long as these substitu- ing preference function. The "as if" is
tions have no effect on the rank order significant here: the conventional ap-
of the outcomes in either prospect. proach, interpreted descriptively, seeks
Various generalizations of the rank- to predict which choices are made and
dependent model have been proposed typically, there is no presupposition that
(e.g. Segal 1989, 1993; Chew and Ep- the model corresponds with any of the
stein 1989; Jerry Green and Jullien mental activities actually involved in mak-
1988). In Green and Jullien, the crucial ing choices. While this underlying meth-
axiom is ordinal independence. Suppose odology dominates economic theory,
another approach more common in the Johnson argue that selection of choice
psychology literature seeks to model the procedures is "adaptive and intelligent"
processes that lead to choice. I will call (p. 14), and though decisions may not
such theories procedural theories. A be optimal in the conventional sense,
common feature of such theories is to the selection of decision rule does in-
assume that agents draw on decision volve optimization but with unusual
heuristics or rules of one kind or an- constraints (e.g. information processing
other when making their choices. The capacity) and/or objectives (e.g. the
problem is then to identify the set of choice of strategy might be influenced
decision heuristics the agent may draw by considerations such as a desire to be
on, and to specify the conditions under able to justify a choice to a third party).
which particular rules will be followed. Indeed, as John Conlisk (1996, p. 672)
In such theories, it is common for prob- points out, "bounded rationality is not a
lem context to be an important determi- departure from economic reasoning,
nant of choice-rule selection. For in- but a needed extension of it."
stance, there may be a tendency to While models of bounded rationality
choose the rule that is easiest to apply have been applied with some success
in the given context, and ease of appli- elsewhere in economics-see Conlisk's
cation may depend on how a problem is (1996) review in this journal-full-
presented. Consequently, it seems natu- blown procedural models of decision
ral to expect phenomena like framing under risk, like that of Payne, Bettman,
effects within this framework. and Johnson, have not received much
One recent and quite general proce- attention from the economics profes-
dural model has been developed by sion. Nevertheless, there has been a de-
John Payne, James Bettman, and Eric gree of cross-fertilization, and some
Johnson (1993). They assume that theories involving a procedural element
agents have at their disposal a range of have appeared in the economics litera-
possible choice heuristics that might be ture. Examples include the models pro-
applied to a given decision task. These posed by Kahneman and Tversky (1979),
include expected utility calculations, Ariel Rubinstein (1988), and Marc Lavoie
satisficing rules, lexicographic choice (1992).
rules, and so on. 16 In their adaptive The most widely discussed of these is
model the decision maker "decides how Kahneman and Tversky's (1979) pros-
to decide," trading off the desire to pect theory. In this theory, choice is
make a "good" decision against the cog- modeled as a two-phase process. In the
nitive effort involved in applying differ- first phase, prospects are "edited" using
ent rules in a given context. Here, as in a variety of decision heuristics; in the
other procedural models, the agent is second, choices among edited prospects
conceived of as boundedly rational, an are determined by a preference func-
agent with limited computational ability tion which, for a restrictive class of
and, perhaps, imperfectly defined ob- prospects,17 can be represented by the
jectives, attempting to cope with an
17The original version of prospect theory does
often complex decision environment.
not provide a general preference representation over
Yet, boundedly rational does not equate prospects. Strictly speaking, it only applies to pros-
with dumb. Payne, Bettman, and pects of the form (xl, pl; x2, p2; 0, (1 - pl - p2)).
The function assumed in prospect theory coin-
16 For a discussion of satisficing rules see Her- cides with the function defined dere in the case of
bert Simon (1955) and for an example of a lexi- "regular prospects" where either pl + P2 < 1, or
cographic procedure see Tversky (1969). xl?O?x2,orxl <?<X2.
Value of Outcome
steeper in the domain of losses. In their
later paper Tversky and Kahneman
(1992) interpret these restrictions as
implications of two more general prop-
erties of perception and judgement:
diminishing sensitivity and loss aver-
Losses (x <0) Gains (x > 0) sion. Diminishing sensitivity holds that
the psychological impact of a marginal
change will decrease as we move fur-
ther away from a reference point. So,
for example, relative to the status quo,
the difference between a gain of $10
and $20 will seem larger than the dif-
ference between gains of $110 and
$120. More generally, the assumption
Figure 7. The Valuation of Outcomes in of diminishing
Prospect Theorysensitivity applied to the
outcome domain entails diminishing

simple decision-weighted utility form marginal utility for gains (i.e., u"(x) < 0
defined in 6 above. Two features of this for x ? 0) and diminishing marginal
theory distinguish it clearly from any of disutility for losses (i.e., u"(x) ? 0 for x <
the theories we have discussed so far. 0). So property (i) of the utility function
First and most obvious is the editing is a direct implication of diminishing
phase, but a second distinguishing fea- sensitivity. Loss aversion is the princi-
ture is that, in prospect theory, out- ple that "losses loom larger than corre-
comes are interpreted as gains and sponding gains" (Tversky and Kahne-
losses relative to a reference point. For man 1992, p. 303). They justify this
present purposes we may think of the second feature of the function partly by
reference point as status quo wealth. an appeal to intuition and partly by ap-
The motivation for handling conse- pealing to empirical evidence (e.g. the
quences in this way is that it allows fact that most people find symmetric
gains and losses to be evaluated quite bets of the form (x, 0.5; -x, 0.5) "dis-
differently. This capacity, it turns out, tinctly unattractive"). Loss aversion is
has some quite interesting implications. modeled by imposing u'(x) < u'(-x).
In prospect theory outcomes are The evaluation of risky prospects in-
evaluated via a utility function'8 with volves a probability weighting function
the shape of that in Figure 7. It is and, in the original version of prospect
kinked at the reference point (i.e., theory, Kahneman and Tversky proposed
status quo, x = 0) and notice two further a weighting function that underweights
properties: (i) it is concave for gains "large" and overweights "small" prob-
and convex for losses, and (ii) it is abilities. The endpoints are such that
18 Kahneman and Tversky explicitly avoid using t(1) = 1 and t(0) = 0, but the function is
the term "utility" to describe this function, prefer- not defined for probabilities close to
ring instead the term "value function." I suspect
zero and one; unusual things may hap-
they had in mind a conception of value indepen-
dent of risk and wished to distance themselves pen in these regions-for example,
from the notion of utility in EUT where utilities "very small" probabilities might be ig-
may partly reflect attitudes to chance. Here I re-
nored. It is worth noting that in a later
vert to utility terminology, but with a timely re-
minder that the appropriate interpretation of "util- version of prospect theory (see cumula-
ity" varies between theories. tive prospect theory below), Kahneman
352 Journal of Economic Literature, Vol. XXXVIII (tune 2000)
and Tversky adopt the widely used in- negative outcomes may display risk
verted-s weighting function. This is seeking if all outcomes are changed
partly because that specification fits to losses of the same absolute magni-
their data well, and no doubt partly to tude. Kahneman and Tversky report
resolve the ambiguity about what hap- evidence for this kind of effect from an
pens at the end points in the original experiment involving choices among pros-
version, but there is also an underlying pects of the form S5 = (X, p; 0, 1 - p) and
theoretical rationale. The principle of r5s=(y, 2p; 0, I- kp). For given abso-
diminishing sensitivity, which deter- lute values of x and y the majority of
mines some of the important charac- subjects revealed S5 > r5 when y > x > 0
teristics of the utility function, can alsoand r > S5swhen y <x < 0.
provide a psychological rationale for an The "Asian disease" example dis-
(inverted) s-shaped probability weight- cussed at the end of Section 3 is consis-
ing function: a function with the prop- tent with the reflection effect. In that
erty of diminishing sensitivity will be example, the choice between prospects
steepest close to a reference point, was affected by the description of op-
hence on the assumption that the end tions. When outcomes were framed as
points of the probability scale constitute lives saved, the majority of choosers
natural reference points, diminishing were attracted to a sure gain of 200 out
sensitivity implies a probability weight- of 600 lives; when framed as losses the
ing function that is steep near zero and majority rejected the sure loss of 400
one but relatively flat around the mid- out of 600 deaths, preferring instead to
dle. The inverted-s has precisely these take the risk. The effect observed there
properties. Hence, if diminishing sensi- can be interpreted as a reflection effect
tivity is a general principle of percep- with risk aversion in relation to gains
tion, it provides a common psychologi- and risk seeking for losses. Before we
cal underpinning for properties of both could think this an explanation of the
the utility function and the probability Asian disease problem, however, we
weighting function. need an account of how consequences
Kahneman and Tversky (1979) argue are interpreted. From an objective
that their theory is able to capture a standpoint, two hundred lives saved out
wide range of observed behavior toward of six hundred is the same thing as four
risk, including standard violations of the hundred lives lost, hence a full explana-
independence axiom (e.g. the common tion would require a theory of how
ratio and common consequence ef- framing affects whether an outcome is
fects), and a variety of field data, plus interpreted as a gain or a loss. Kahne-
an extensive range of data generated man and Tversky go some way towards
from their own experiments. The theory this in their discussion of editing.
also has some unusual properties, one Prospect theory assumes that prior to
of which is the so-called reflection ef- the second stage of evaluation, individu-
fect. The fact that concavity of the util- als will edit prospects using a variety of
ity function in the domain of gains is heuristics. One of the major editing op-
mirrored by convexity in the domain of erations involves the coding of outcomes
losses means behavior towards risk can as gains and losses relative to a refer-
be likewise mirrored across the two do- ence point. Kahneman and Tversky argue
mains. For instance, a given individual that the reference point will typically be
who displays risk aversion in a choice the current asset position, but they al-
among particular prospects with non- low the possibility that "the location of
the reference point, and the consequent whether a particular rule is applied
coding of outcomes as gains or losses, depends upon whether or not it is sali-
can be affected by the formulation of ent. Although they have no formal the-
the offered prospects, and by the expec- ory of salience they do present evidence
tations of the decision maker" (p. 274). that editing is context dependent. One
Notice that this possibility of differen- example shows that cancellation is used
tial coding under the two problem de- in some cases where it is salient and not
scriptions is a necessary step in explain- in others (see their discussion of the
ing responses to the Asian disease "isolation effect," p. 271).
problem. While some economists might One further rule-I will call it the
be tempted to think that questions dominance heuristic-has the effect of
about how reference points are deter- eliminating stochastically dominated
mined sound more like psychological options from the choice set prior to
than economic issues, recent research is evaluation. The addition of the domi-
showing that understanding the role of nance heuristic does not, however, re-
references points may be an important move all possibility of monotonicity vio-
step in explaining real economic behav- lation. Kahneman and Tversky assume
ior in the field (see, for example, Chip that individuals scan the set of options
Heath, Steven Huddart, and Mark Lang and delete dominated prospects if they
1998). are detected. This ensures the deletion
Several of the other editing routines of "transparently" dominated options, but
in prospect theory are essentially rules leaves open the possibility that some
for simplifying prospects and transform- dominated options survive application
ing them into a form that can be more of the routine. Since the preference
easily handled in the second phase. One function is not generally monotonic,
such operation is the rule of combina- such options may ultimately be chosen.
tion which simplifies prospects by com- This strategy for imposing mono-
bining the probabilities associated with tonicity has the further, perhaps sur-
identical outcomes. For example, a pros- prising, implication that choices may be
pect described as (xl, pi; Xl, p2; X3, p3; . . non-transitive.
. ) If n(.) is nonlinear, then
may be evaluated as the simplified pros- prospect theory implies that there will
pect (xl, (pl+ p2); X3, p3; . . . ). Notice be some q and r where q stochastically
that these two prospects are not, in gen- dominates r such that V(r) > V(q).19 So
eral, equivalent if sT(.) is nonlinear. De- long as this dominance is transparent,
cision makers may also simplify pros- the dominance heuristic ensures that
pects by rounding probabilities and/or there will be no direct violation of
outcomes. Further operations apply to monotonicity and r will not be chosen
sets of prospects. The operation of can- over q. In general, however, it should
cellation involves the elimination of ele- be possible to find some other prospect
ments common to the prospects under s, such that V(r) > V(s) > V(q). If there
consideration. Hence a choice between
prospects q' = (x, p; q, 1 - p) and r' = (x, 19 To see how nonlinearity of t(.) can generate
p; r, 1 - p) may be evaluated as a choice violations of monotonicity, consider a simple case
where q=(x, 1) and r=(x-?, p; x, 1 -p). Sup-
between q and r. Although cancellation pose ? > 0 hence q dominates r: If Tc(.) is concave,
is effectively an application of the inde- probabilities are overweighted, and the dominated
option r is preferred for some ?. Now suppose ? <
pendence axiom of EUT, the editing
0, hence r dominates q: if tc(.) is convex, prob-
phase does not imply that choices will abilities are underweighted, and the dominated
generally satisfy independence, since option q is preferred for some ?.
is no relation of dominance between s placed by zeros. Tversky and Kahne-

and either of q or r, then pairwise man's model, cumulative prospect the-
choice among these three gambles will ory, is more general in that it allows the
generate a systematic cycle of choice in decision weighting function to be dif-
which q >c r and r > c s and s >c q ferent for the positive and negative
where >c is the relation "is chosen components. The development of these
over." Quiggin (1982, p. 327) calls this so-called sign- and rank-dependent
an "undesirable result." Quiggin's reac- models demonstrates that important as-
tion would not be untypical of econo- pects of prospect theory can be cap-
mists more generally, most of whom tured within a formal model that is es-
have taken both transitivity and mono- sentially conventional, without the need
tonicity to be fundamental principles to invoke an editing phase.
which any satisfactory theory should In these later models, the procedural
embody. On the other hand, several element central to prospect theory has
economists, Quiggin included, have disappeared.20 No doubt the abandon-
thought aspects of prospect theory ap- ment of editing does leave some things
pealing and have sought to build the unexplained. For instance, framing ef-
relevant features into models more in fects do suggest that choices are context
keeping with conventional theoretical dependent in complex yet subtle ways,
desiderata. For example, part of Quig- and the procedural approach seems to
gin's motivation in developing rank- provide the more natural arena in which
dependent expected utility theory was to model this. On the other hand, intro-
to establish that a central feature of ducing elements of bounded rationality
prospect theory-nonlinear decision does considerably complicate the theo-
weights-can be built into a preference retical structure of models in ways that
function without sacrificing monotonic- render them less compatible with the
ity. By constructing decision weights rest of economic theory. For example,
cumulatively, we obtain a (transitive) working with a set of decision rules
preference function that is monotonic seems clumsy, relative to the neatness
without the need for an additional edit- and tractability of optimizing a single
ing routine. Papers by Starmer and Sug- function; unlike conventional models,
den (1989), Luce and Fishburn (1991), procedural models often exhibit a degree
and Tversky and Kahneman (1992) of indeterminacy.21
show that the rank-dependent form can Might such arguments provide suffi-
be extended to capture another key cient grounds for defending a general
element of prospect theory: valuing theoretical presumption that agents be-
outcomes relative to reference points. have "as if' fully rational? Conlisk
In Starmer and Sugden's model, any (1996) reviews a series of methodologi-
prospect q is valued by the function cal arguments which might be used to
V(q) V+(q) + V-(q) where V+(q) is the make such a case against incorporating
rank-dependent expected utility of a
transformed prospect q+; this is equiva- 20 Although Tversky and Kahneman do mention
that editing may be important, their 1992 model
lent to q excepting that any outcomes of
has no formal editing p ase and their references
q that are losses are replaced by zeros. to it are virtually asides.
Similarly, V-(q) is obtained by applying 21 For instance, in prospect theor ,the outcome
of editing can depend on factors t at are under-
the standard rank-dependent form to a
determined by the theory, such as the order in
transformed prospect q-; in this case, which operations are applied (see M. K. Steven-
any outcomes that are gains are re- son, J. R. Busmeyer, and J. C. Naylor 1991).
Starmer: Developments in Non-Expected Utilit, Theor, 355
ideas of bounded rationality into eco- sons occur between the consequences
nomics. He concludes that it is hard to of alternative choice options.
make any convincing case against. If Since the theory has to allow com-
that's correct, and I for one am per- parisons between choice options, it can-
suaded, then the question to ask is not be a conventional theory that assigns
whether departures from conventional values independently to individual pros-
models are of sufficient concern, from pects. Loomes and Sugden propose a
an empirical point of view, to justify the theory of pairwise choice in which pref-
theoretical costs involved. We shall be erences are defined over pairs of acts,
examining this issue shortly (see Sec- where an act maps from states of the
tion 5). First, however, we consider an world to consequences.22 Let Ai and Aj
alternative avenue of departure from be two potential acts that result in out-
the conventional approach. comes xis and Xjs, respectively, in state
of the world S. The utility of conse-
4.2.2 Non-Transitive Preference Theory
quence xis is given by a function
As we have seen, many have taken the M(xis,xjs) which is increasing in its first
view that the standard independence ax- argument and decreasing in its second.
iom of EUT can be sacrificed for the sake This function allows the utility from
of explaining the data. Transitivity, how- having xi, to be suppressed by "regret"
ever, may be another matter. It might be when xis < xjs, or enhanced by "rejoic-
tempting to think that transitivity is so ing" when xis > xjs. The individual then
fundamental to our ideas about prefer- seeks to maximize Is ps .M(xis,xjs) where
ence that to give it up is to depart from ps is the probability of state S. Regret
theories of preference altogether. Can theory reduces to EUT in the special
we speak of people maximizing anything case where M(xis,xjs) = u(xis)
if they don't have transitive preferences? Although preferences are defined
It turns out that the answer is yes. over acts, the theory can be applied to
There is at least one well-known the- choices between prospects given some
ory of choice based on a model of non- assumption about how outcomes are
transitive preference. The theory I have correlated between them. One interest-
in mind was proposed simultaneously by ing case is when consequences are un-
Bell (1982), Fishburn (1982), and correlated between prospects; that is,
Loomes and Sugden (1982). I will begin when prospects are statistically inde-
by discussing a version of this theory pendent. In a choice between a pair of
presented by Loomes and Sugden (1987). such prospects q and r, if q is chosen,
Loomes and Sugden call their theory re- the probability of getting xi and missing
gret theory, and its central premise is out on xj is given by pqiprj where pqi is
closely akin to the psychological intui- the probability of consequence xi in q
tion at the heart of the theory of disap- and prj the probability of xj in prospect
pointment. In that theory, it is assumed r. Preferences between q and r are
that an individual compares the out- then determined by the expression:
comes within a given prospect giving rise
to the possibility of disappointment when >= r jpqipiJ(xi, Xj) 0 (7)
the outcome of a gamble compares un-
favorably with what they might have
22As a theory of pairwise choice, regret theory
had. Regret theory allows comparisons
has limited applicability, but ways of generalizing
between consequences to affect choice, the theory have been suggested by Sugden (1993)
but in this case, the relevant compari- and Quiggin (1994).
skew-symmetric bilinear utility or SSB).

where NJ(xis,xs) _ M(xi,,xj,) - M(xj,,xi,).
Fishburn's
The function Nf(.,.) is skew symmetric by model is identical with re-
gret theory
construction, hence Nj(x,y) _-N(y,x) and for statistically independent
f(x,x) =_ 0 for all x,y. prospects, and we can think of regret
If prospects are statistically indepen- theory as a generalization of SSB which
dent, the addition of a further assump- extends it to non-independent pros-
tion which Loomes and Sugden call pects: in this realm, regret aversion has
regret aversion23 implies that indiffer- some very interesting implications.
ence curves will fan out in the prob- Consider three stochastically equiva-
ability triangle. Regret aversion re- lent actions Al, A2 and A3, each of which
quires that for any three consequences gives each of the consequences x > y >
x .> y > z, (x, z) > v(X, y) + v(y, z).- z in one of three equally probable states
The interpretation of the assumption is of the world si, S2 and s3. Any con-
that large differences between what you ventional theory entails a property of
get from a chosen action and what you equivalence, that is, indifference be-
might have gotten from an alternative tween stochastically equivalent options,
give rise to disproportionately large hence, for any such theory, Ai - A2 -
regrets; so people prefer greater cer- A3. In regret theory, however, it matters
tainty in the distribution of regret. Un- how consequences are assigned to states,
der these conditions regret theory is and for particular assignments, regret
equivalent to Chew and MacCrimmon's theory implies a strict preference between
weighted utility theory, and so indiffer- stochastically equivalent acts, violating
ence curves in the probability triangle equivalence. For example, suppose that
will have the pattern described in Figure the three acts involved the following
4 above (see Sugden 1986 for a simple assignment of consequences to states:
demonstration of this). Consequently,
SI S2 S3
regret theory is able to explain the stan-
A1 z y x
dard violations of the independence A2 x z y
axiom for statistically independent A3 y x z
prospects.24
If we consider the class of all statisti- If we consider preferences between the
cally independent prospects-not just first two acts, regret theory implies:
those with up to three pure conse-
quences-weighted utility theory is a A1 A2 X [tV(z,x) + if(y,z) + ox,y)] - (8)
special case of regret theory. Specifi-
cally, the representation in expression 7 Using the skew symmetry of (.,.), the
is obtained from Chew and MacCrim- term in square brackets is equal to
mon's axiom set by relaxing transitivity. [V(x,y) + j(y,z) - xj(x,z)]. Assuming re-
This is the route by which Fishburn gret aversion, this will be negative,
(1982) arrived at this model (he calls it hence regret theory implies a strict pref-
erence A2 > Al. It is easy to see that the
23 In their early discussions of regret theory, same reasoning applied to the other two
Loomes and Sugden called this assumption 'con- possible pairwise comparisons implies A3
vexity. "
24 Some instances of the common consequence > A2 and Al > A3. Hence, regret theory
effect have involved statistically non-independent also implies a cycle of preference of the
options, and these cases are not consistent with form: A2 > Ai, A3> A2, Ai > A3. Now
regret theory (unless we assume agents treat op-
tions as if they are independent even when they consider adding some small positive
are not). amount e to one consequence of action
Starmer: Developments in Non-Expected Utilitu Theoru 357
Al. The resulting action, call it A*, sto- theory can be interpreted as a form of
chastically dominates each of the original preference reversal.
actions. But since regret theory implies So, regret theory offers the tantalizing
A2 > Ai we should expect A2 > A* for at opportunity of explaining violations of
least some ?> 0. Hence regret theory independence and preference reversal
also implies violations of monotonicity. within a theory of preference maximiza-
Relative to the conventional approach tion. Of course, since observation of
then, preferences in regret theory are preference reversal pre-dates the devel-
not at all well-behaved: they satisfy nei- opment of regret theory, that phenom-
ther monotonicity nor transitivity and enon offers only weak support for the
the theory allows strict preferences be- unconventional predictive content of re-
tween stochastically equivalent acts. gret theory. More recent research has
While such properties may seem pecu- aimed at testing some novel predictions
liar to the eye of the conventional of regret theory, and some of the results
economist, from the descriptive angle, from this line of research are discussed in
the crucial question is whether such im- the next section.
plications of the theory are borne out
by actual behavior. Shortly after propos- 5. Evaluating Alternatives to Expected
ing regret theory, Loomes and Sugden
Utility Theory
(1983) argued that at least one might
be. Consider the following three acts la- 5.1 The Recent Experimental Evidence
beled $, P and M with monetary conse-
quences x > y > m > 0 defined (for the sake Starting in the mid-1980s, a number
of simplicity) over three equiprobable of researchers turned their attention to-
states: wards testing non-expected utility theo-
ries. The majority of this work involved
SI S2 S3 experimental testing, some of it de-
$ x 0 0
signed to compare the predictive abili-
P y y 0
M m m m ties of competing theories; some de-
signed to test novel implications of
The actions labeled $ and P have the particular theories; and some designed
structure of typical $- and P-bets: they are to test the descriptive validity of par-
binary gambles where $ has the higher ticular axioms. A very large volume of
prize, and P the higher probability of "win-work has emerged in this arena, provid-
ning"; the third act gives payoff m for sure.ing a much richer evidential base
Loomes and Sugden show that, given re- against which theories can be judged.
gret aversion, pairwise choices over acts The purpose of this section is to discuss
with this structure may be cyclical, and if what has been learned in this second
a cycle occurs it will be in a specific di- phase of the hunt, and its organization
rection with P > $, M > P and $ > M. reflects the dichotomy between conven-
Now recall that in a standard experiment, tional and nonconventional models
subjects reveal P > $ in a straight choice adopted in the theoretical discussion.
between options but place a higher value The first part discusses evidence that
on $ relative to P in separate valuation bears directly on the choice between
tasks. If we interpret choices from {$, MI
conventional models; the second part
addresses evidence with a bearing on
and {P, Ml as analogues of valuation tasks
the relative merits of conventional and
asking "is $ (or P) worth more or less than
m" then the cycle predicted by regret nonconventional approaches.
5.1.1 Choosing among Conventional up along the left-hand edge of prob-

Models ability triangles. Such behavior would
be consistent with a tendency for indif-
As we have seen, conventional theo- ference curves to fan in. These facts
ries all imply the existence of indiffer- mitigate in favor of theories like disap-
ence curves in the probability triangle, pointment aversion, implicit utility,
and certain of their key properties can quadratic utility, and models with de-
be expressed in terms of characteristics cision weights, all of which allow a
of the indifference maps they generate. mixture of fanning-in and fanning-out.
For instance, Machina's theory implies A second general lesson in the data
generalized fanning-out, while other seems to be don't impose betweenness.
theories imply a mixture of fanning-in There is considerable evidence-a good
and fanning-out. A large number of ex- part of it is reviewed in Camerer and
perimental studies have explicitly exam- Teck-Hua Ho (1994)-that choices are
ined individual behavior in choices inconsistent with the assumption of lin-
among prospects in probability trian- ear indifference curves. Together these
gles. The data generated from these two requirements narrow the field con-
"triangle experiments" provide a van- siderably: if we want a theory of mixed
tage point from which we can ask the fanning with nonlinear indifference
following question: suppose one were curves, of the theories reviewed above
attempting to construct a conventional the only contenders are quadratic utility,
theory now, with the aim of accounting lottery-dependent utility, and models
for the evidence currently available, are with decision weights.
there any obvious properties one should A third widely observed finding argu-
seek to build in? ably nudges the decision weighting
Although the evidence is both rich models into the lead: behavior on the
and complex, a number of stylized facts interior of the probability triangle tends
apply across a range of studies. In my to conform more closely to the implica-
view, three observations seem particu- tions of EUT than behavior at the bor-
larly robust. First, if you want a theory ders. Although significant off-border
consistent with the available data don't violations are observed in at least some
impose generalized fanning-out. Evi- experiments (see for example Wu and
dence from a wide range of studies re- Gonzalez 1996) several studies, including
veals behavior inconsistent with linear those of Conlisk (1989), Camerer (1992),
parallel indifference curves, but the David Harless (1992), and Garry Gigliotti
patterns actually observed are more and Barry Sopher (1993), suggest that
complex than generalized fanning-out. violations of EUT are concentrated in
For example, while numerous studies comparisons between options involving
reproduce behavior consistent with Al- prospects on or near to the borders of
lais paradox violations of EUT in choice triangles. It is important to note that
pairs moving left to right along the bot- this observation is unlikely to rescue
tom edge of the probability triangle, an- EUT for practical purposes. A natural
other finding replicated across a range interpretation of the "border effect" is
of studies, including Camerer (1989), that individuals are particularly sensi-
Chew and William Waller (1986), Bat- tive to changes in the likelihood of out-
talio, Kagel, and Jiranyakul (1990), and comes with "extreme" probabilities (i.e.,
Starmer (1992), is a tendency for behav- moving off the border of the triangle we
ior to become less risk averse moving introduce a low probability event; in the
vicinity of each corner, some outcome is for the probability weighting function.
near certain). It is very easy to think of Lattimore, Baker, and Witte (1992) use
important choice scenarios involving a probability weighting function of the
real prospects with "extreme" prob- form:
abilities; for example, individual deci-
sions about participation in national or =(pi) =Pi / [| aP + jP
state lotteries or collective decisions
about nuclear power generation involve for i, k= 1,2, ... n,ki and c, >0 (n
high magnitude outcomes (winning the is the number of outcomes as usual).
lottery, suffering the effects of a radia- This captures a number of other pro-
tion leak) occurring with very small posed forms (e.g. those of Uday Kar-
probabilities. Consequently, there are markar 1978 and Quiggin 1982) as spe-
good reasons to model sensitivity to "ex- cial cases. With cc= 13= 1, t(pi)= pi,
treme" probabilities. One obvious way hence we get EUT. More generally, the
to do it is via decision weights.25 parameter I^ controls the inflection point
In summary, if one is looking to orga- and 1B < 1 generates the inverted-s with
nize the data from the large number of the consequent over-weighting of "small"
triangle experiments, then the decision- probabilities below the inflection point,
weighting models are probably the best and underweighting above it. With ca < 1,
bet. Moreover, there is a striking de- *t(.) is "sub-certain" in the sense that
gree of convergence across studies re- the sum of weights (Xi 7c(pi)) will be less
garding the functional form to use; for than unity. Lattimore, Baker, and Witte
best predictions the key ingredient (1992, p. 381) describe this as "'prospect
seems to be an inverted s-shaped pessimism' in the sense that the value of
weighting function. Empirical support the prospect is reduced vis-a-vis certain
for this specification comes from a wide outcomes." In their empirical estimates,
range of studies including Pamela Latti- they find that allowing nonlinear deci-
more, Joanna Baker, and Ann Witte sion weights offers significant improve-
(1992); Tversky and Kahneman (1992); ment in predictive power over EUT
Camerer and Ho (1994); Abdellaoui (which is the best model for only about
(1998); and Gonzalez and Wu (1999), 20 percent of their subjects). The best-
all of which fit the decision-weighting fitting weighting function is generally
model to experimental data. Collec- the inverted-s exhibiting greater sensitiv-
tively, these studies show that models ity to high and low probabilities relative
with s-shaped probability transforma- to mid-range probabilities. They also re-
tions offer significant predictive im- port differences between the best-fitting
provement over EUT and outperform weighting functions for gains and losses
other rivals. Most of the studies in this (for example "pessimism" is more pro-
vein, at least those conducted in recent nounced for losses) though the interpre-
times, employ the rank dependent tation of these differences is potentially
transformation method, though differ- confounded by the fact that, in their
ent mathematical forms have been used study, gains are measured in units of
money while losses are measured in units
25 Another theoretical possibility suggested by
Neilson (1992) is to allow the utility function de-
of time.
fined over outcomes to depend on the number of Single parameter weighting functions
outcomes: this generates different behavior on and have been proposed by Tversky and
off the border, But experimental tests of the model
(see Stephen Humphrey 1998) have not been sup-
Kahneman (1992) and Prelec (1998).
portive. Tversky and Kahneman suggest the
form t(p)=p6/[(p8+ (1-p)6)l16]. This rameter model may be useful for explain-
generates the inverted-s for 0< < 1, ing differences between individuals. For
and reducing 8 lowers the crossover other purposes, however, parsimony
point while accentuating the curvature favors the one-parameter versions.
of the function. Their empirical analysis Conventional theory can claim a suc-
supports the s-shaped weighting function cess here: a one-parameter extension to
and also reveals systematic differences EUT can offer significantly improved
in behavior for gains and losses; specifi- predictive power for a large body of
cally, indifference curves in the best- data generated mainly from triangle ex-
fitting models for losses resemble those periments. If we want to predict behav-
for gains flipped around a 45-degree line. ior over simple choices like this we
This supports the case for a model that know a lot about how to improve on
distinguishes between gains and losses EUT. Reflection over a broader range
(i.e. a model with a reference point) of experimental evidence, however, sug-
though virtually no work is done by the gests that we are still a long way from a
weighting function here; essentially, satisfactory general account of behavior
the same probability-weighting function under risk.
works well for both gains and losses.
5.1.2 A Case for the Unconventional
Prelec proposes the function i(p)=
exp(-(-ln p)a). With 0 < oc < 1, this gen- I now turn to a discussion of labora-
erates the inverted-s with a fixed inflec- tory evidence which, in my view, pro-
tion point at p = Ile = 0.37. Visually, ox is vides a substantive challenge to certain
the slope of n(.) at the inflection point, key assumptions that underpin conven-
and as ox approaches unity, n(.) becomes tional preference theories. Again, the
approximately linear; as it approaches evidence I cite does not constitute a
zero, n(.) approximates a step function. thorough review; instead I draw on ex-
Prelec argues that a crossover in the vi- amples of phenomena that seem both
cinity of l/e is consistent with the data challenging and well established.
observed across a range of studies. A i. Violations of monotonicity: It might
novel feature of Prelec's contribution is be tempting to think that violations of
to provide an axiomatization for this form,monotonicity must be rare for two rea-
and he also discusses a two-parameter sons based on casual empiricism: indi-
generalization. The two-parameter ver- viduals are not stupid; and, if they were,
sion is similar in spirit to the "linear inwe would see market institutions trad-
log odds form" discussed by Gonzalez ing on that stupidity (e.g. casinos com-
and Wu (1999) in that it allows the cur- peting with each other by advertising
vature and elevation of the weighting worse odds of winning than their ri-
function to be manipulated (more or vals!). Experimental evidence, however,
less) independently. In the latter form, supports two stylized facts about mono-
probability weights are given by: tonicity. First, very few people will
choose a stochastically dominated op-
ir(pi) = 87/[8iY + (1 - poy]
tion from a choice set when it is trans-
The parameter 8 primarily controls parently obvious that the option is
the absolute value of c(.) by altering the dominated. Second, choices are not gen-
elevation of the function, relative to the erally monotonic and systematic viola-
45-degree line, while y primarily con- tions of monotonicity can be generated
trols curvature. Gonzalez and Wu's data in contexts where the relation of domi-
suggests that the flexibility of a two-pa- nance is opaque (i.e. not obvious to the
chooser). One sharp illustration of this the violation is opaque, presumably be-
is provided by the following example cause they do not have generally mono-
due to Tversky and Kahneman (1986): tonic preferences. Further examples of
monotonicity violation can be found in
Consider the following pair of lotteries, described by Loomes, Starmer, and Sugden (1992);
the percentage of marbles of different colors in each box Michael Birnbaum and Laura Thompson
and the amount of money you win or lose depending on
the color of a randomly drawn marble. Which lottery
(1996); Birnbaum and Juan Navarette
do you prefer? (1998); and J. W. Leland (1998).
ii. Event-splitting effects: It is well
Option A: known in marketing circles that by "un-
90% white 6% red 1% green 1% blue 2% yellow
packing" positive attributes of a good
$0 win $45 win $30 lose $15 lose $15
into multiple sub-attributes one can
Option B: make a good seem more desirable. For
90% white 6% red 1% green 1% blue 2% yellow example, instead of just describing a car
$0 win $45 win $45 lose $10 lose $15 as having "good performance," you can
make it seem more attractive by subdi-
It is very easy to see that option B domi-
viding performance into acceleration,
nates option A since, for every color, the
cornering, braking, and so on. It turns
prize for option B is always at least as
out that the attractiveness of risky op-
good as the prize for option A and in
tions can be influenced by "unpacking"
some cases it is better. Kahneman and
probabilities in an analogous way. Con-
Tversky presented this problem to 88
sider acts defined over a set of mutually
subjects and found that all of them chose
exclusive and exhaustive states of the
B. Now consider this slightly modified
world S = {sj: i = 1, . . . , n}. Define
version of the above problems:
four events Ei, E2, E3 and E4 such that
Option C: Ei consists of the set of states {si: i = 1, . . ..
90% white 6% red 1% green 3% yellow k - 11; E2 consists of the remaining
$0 win $45 win $30 lose $15 states {si: i = k, . . . , n}. Events E3 and
E4 partition E2 into two distinct subsets
Option D:
90% white 7% red 1% green 2% yellow
of states such that E3 = {si: i = k, . . . k
$0 win $45 lose $10 lose $15 +j1 and E4 = {si: i = k +j + 1, . . . , n}.
Now consider two particular acts A and
Options C and D are stochastically B where A gives consequence x condi-
equivalent to A and B respectively; the tional on Ei and consequence y condi-
only difference is a minor change in the tional on E2; B gives x conditional on
presentation which "simplifies" the op- Ei and y conditional on each of events
tions by assigning each prize to a single E3 or E4. The only difference between
color. This framing of the options, how- A and B is that A is described as result-
ever, also makes it more difficult to de- ing in outcome y for a single event
tect the dominance of D over C. In fact, whereas act B is described as resulting
Kahneman and Tversky found that a main y for two distinct events. In most
jority of subjects (58 percent) chose the theories this difference is irrelevant and
dominated option C. This finding- the two acts are simply regarded as two
which is consistent with the original two- alternative descriptions of the same
phase version of prospect theory-sup- prospect q = (x, p; y, I -p) where p is
ports the view that although people do the probability of Ei.
not purposefully choose to violate mono- There is clear evidence that such re-
tonicity, they might do so in cases where descriptions do matter, however. More
specifically, studies by Starmer and in objective data or subjective impres-
Sugden (1993) and Humphrey (1995) sion. The theory explicitly allows for
have shown that when an event that judged probability to be enhanced by
gives some outcome y is split into two "unpacking" of events in line with the
sub-events, there is a tendency for that event-splitting effect. Prospective refer-
consequence to carry more weight even ence theory proposed by W. Kip Viscusi
though its total objective probability is (1989) provides another possible ac-
unchanged. This is the event-splitting count of event splitting; this works by
effect. One implication of it is that indi- assuming that individuals may distrust,
viduals are more likely to choose par- and hence revise, objective probability
ticular options if events containing rela- information. Another possible explana-
tively attractive consequences are tion is provided by simple decision-
subdivided into two (or perhaps more) weighted utility: take the formulation in
events. Similar effects have been found (6) above, then assume a sub-additive
by Martin Weber, Franz Eisenftihr, and weighting function such that 7t(pi) +
Detlof vonWinterfeldt (1988) in the 1t(p2) > 1C(pl + p2) for at least some pi,
context of multi-attribute choice. This p2. A theory of this form would be con-
suggests the possibility, albeit specula- sistent with event splitting and also the
tive, that event splitting may be a con- violations of monotonicity observed by
sequence of some more general prop- Loomes, Starmer, and Sugden (1992).26
erty of judgement which extends Historically, of course, theories of this
beyond the domain of risk. The exis- type with such a simple construction of
tence of an event-splitting effect has decision weights have been rejected in
some potentially important implications favor of more sophisticated theories like
at the level of policy. Take, for example, rank dependent models precisely because
the risks of skin cancer from exposure they predict such behavior.
to ultraviolet radiation (UVR). There iii. Violations of transitivity: There
are a variety of cancers that may result is well-established evidence-Tversky
from UVR. If event-splitting effects (1969) produced some of the earliest-
generalize to this context, then the per- that cyclical choice is a robust and rea-
ception of the risks from UVR may be sonably general phenomenon. This vio-
affected by whether the risks are de- lates a central principle of most
scribed collectively as a single risk of economic theory. Some economists
"skin cancer" with a given probability, might be tempted to argue that eco-
or alternatively described as a series of nomics can explain non-transitive be-
risks of different cancers, each occur- havior under risk using regret theory:
ring with a smaller probability. The evi- regret theory is, after all, a model of
dence from the laboratory suggests that preference maximization (i.e. in the tra-
the risks would appear worse under the dition of economic theory) and it allows
second, disaggregated description. non-transitive behavior. The truth is,
One possible explanation for the however, that even regret theory cannot
event-splitting effect is the support actually explain what we are learning
theory of Tversky and Koehler (1994). about intransitivities in risky choice.
This theory distinguishes between 26 It has turned out that a large part of the ex-
events and "hypotheses" which are de- perimental evidence which seemed to support
novel predictions of regret theory is more likely
scriptions of events. "Support" is a mea-
due to event-splitting effects. This possibility was
sure of the strength of evidence in favor first suggested by Step hen Davies during a semi-
of a hypothesis, and it may be grounded nar at the University of East Anglia.
While it is true that experimental just at odds with most available theo-
studies have found particular forms of ries. Arguably, it strikes deeper since it
intransitivity specifically predicted by might be read as suggesting a flaw in
regret theory (see Loomes, Starmer, the conventional modeling strategy. In
and Sugden 1989, 1991),27 later experi- seeking models of actual behavior, theo-
ments by Starmer and Sugden (1998) rists following the conventional strategy
confirm the robustness of the phenome- have sought theories built upon consis-
non, but also allow us to reject a variety tency principles like transitivity and
of possible explanations for the occur- monotonicity. Not only has this been
rence of the cycle, including the expla- the standard approach, but those rela-
nation offered by regret theory. The tively rare theories that have not con-
bottom line is that economists do not formed with these principles have been
have a theory of non-transitive behavior widely criticized as unacceptable or im-
that is consistent with the available evi- plausible. For example, the original ver-
dence, though some of the evidence is sion of prospect theory might explain at
suggestive of the kind of theory that least some of the above evidence, but
would be needed. Here's an example. these explanations would all rely on ele-
Earlier in this paper, I noted that, be- ments of prospect theory-like proce-
cause the original version of prospect dural rules, or "unsophisticated" prob-
theory combines simple nonlinear deci- ability weighting functions-which have
sion weights with a dominance heuris- been criticized, ignored, or abandoned.
tic, the theory implies violations of tran- But, like it or not, it seems that theories
sitivity. Some have thought that a of well-behaved preferences in the con-
limitation of the theory, but such judge- ventional mould will not provide gen-
ments may have been premature (and eral descriptive models consistent with
confusing normative and descriptive is- the experimental evidence. Conse-
sues). Starmer (1999a) reports an ex- quently, I would argue that if we genu-
periment that tests for the specific form inely seek descriptive models capable of
of intransitivity implied by prospect explaining the patterns observed in labo-
theory and finds it. I would suggest ratory behavior, our conventional theo-
there is a general lesson here that runs retical desiderata may need rethinking:
beyond simply observing a new form of in particular, there should be no prior
intransitive behavior: don't judge the supposition that the best models will be
predictions of descriptive theories using ones based on principles of rational
normative principles of choice; judge choice, no matter how appealing those
them against empirical evidence. may seem from a normative point of view.
No theory currently available in the
economics literature can successfully 5.2 Evidencefrom the Field
organize the observations in i-iii above.
One might argue it is still early days I have heard some economists argue
and that the way forward is to develop a that they would take more notice of
further generation of theories in the non-EU models if they could be shown
light of the accumulating data. It is cases where they help to explain real-
surely important to acknowledge, how- world phenomena of practical interest
ever, that this type of evidence is not to economics. It is a fair point, but pro-
ponents of non-expected utility theory
27 For further evidence and discussion see
Loomes and Taylor (1992) and Tversky, Slovic, can muster some strong responses. Let
and Kahneman (1990). me illustrate by way of a couple of
examples. It is well-known that EUT agents will be approximately risk neu-

does a poor job explaining some of the tral for small risks (since the utility
things economists are traditionally very function will be almost linear). This
interested in, like insurance behavior theoretical property is at odds with peo-
and the demand for assets: in both cases ples' actual risk attitudes as revealed
non-expected utility models may offer a through their reactions to probabilistic
better understanding of the determinants insurance and so on: people demand a
of some real market behavior, though much greater reduction in premium
in the second case a nonconventional than the actuarially fair adjustment for
approach might hold the solution. accepting a small positive risk of claim
The standard theory of insurance nonpayment.
based on EUT has some implications Segal and Spivak go on to show that
that have long been regarded as highly the counter-intuitive implications of
implausible. For example, a risk-averse EUT carry through to non-expected
expected utility maximizer will not buy utility theories which have similar
full insurance in the presence of posi- smoothness properties. This captures a
tive marginal loading (see J. Mossin large number of alternatives to EUT
1968). This implication, Karl Borch and, in fact, only a single type of theory.
(1974) suggests, is "against all observa- escapes their net: the decision weight-
tion." More recently, Wakker, Thaler, ing models. It is easy to see why models
and Tversky (1997) have made a similar with probability transformations do not
point in relation to "probabilistic insur- imply approximate risk neutrality for
ance." Think of probabilistic insurance small risks since risk averse behavior
as a policy with some fixed probability q can be generated by nonlinear prob-
that a claim will not be paid in the event ability weighting even where the utility
of an insured loss. Wakker, Thaler, and function is linear. So, for example, aver-
Tversky show that an expected utility sion to probabilistic insurance is easily
maximizer willing to pay a premium c explained by overweighting of the small
for full insurance against some risk probability of non-payment. As such,
should be willing to pay a premium ap- decision weighting models stand out as
proximately equal to the actuarially ad- leading contenders to explain aspects of
justed premium (1 - q).c for probabilistic insurance behavior which it has long
insurance. Survey evidence, however, been known standard theory cannot
shows that people are extremely averse handle. There is growing evidence that
to probabilistic insurance and their will- probability weighting may be an impor-
ingness to pay for it is much less than tant ingredient in explaining a variety of
standard theory allows. field data relating to gambling and in-
If expected utility can't explain insur- surance behavior, and several examples
ance behavior, can non-expected utility are discussed by Camerer (forthcoming).
theory do any better? Part of the an- Another field-phenomenon that has
swer is provided by Segal and Avia Spi- perplexed economists is the size and
vak (1990), who show that a number of persistence of the excess return on
implications of EUT for insurance and stocks over fixed income securities. This
asset demand which are widely recog- is the so-called equity premium puzzle
nized to be counter intuitive have a and it is the economics equivalent of
common origin. They arise because, the crop circle: we have seen it in the
with any smooth (i.e. differentiable) field, but we have real trouble explain-
utility function, EUT implies that ing how it got there. Since the return
on stocks is more variable, standard the- like the sign and rank dependent theo-
ory is consistent with some difference ries, the other ingredient in this expla-
in the long-run rates of return, but since nation of the equity premium-i.e.,
Rajnish Mehra and Edward Prescott myopia-belongs in another tradition.
(1985) it has been recognized that the This is essentially a bounded rationality
observed disparity implies implausibly assumption, and while the one-year
high degrees of risk aversion in stan- time horizon has a nice ring of plausi-
dard models of asset pricing. One possibility to it, it sits much more naturally
ble explanation for (part of) the equity alongside procedural theories like the
premium has been suggested by Ep- original version of prospect theory.
stein and Stanley Zin (1990). They show Bounded rationality assumptions seem
that a recursive utility model using rank to be providing the missing links neces-
dependent preferences predicts an eq- sary to explain an increasing range of
uity premium, though only about one economic phenomena (see Camerer
third of the size that is usually ob- 1998 for a recent review of applications
served. A full-and in my view much in individual decision making).
more convincing-account has been It also seems likely that the concept
suggested by Shlomo Benartzi and of loss aversion will become increas-
Thaler (1995), who show that the level ingly important in economics. Evidence
of equity premium is consistent with for the existence of loss aversion in both
prospect theory, with the added as- risky and riskless environments now
sumption that agents are myopic (i.e., seems overwhelming-a few of many
they assess expected returns over possible references are B. J. McNeil et
"short" time horizons). The crucial ele- al. (1982); Jack Knetsch and J. A. Sin-
ment of prospect theory for this expla- den (1984); William Samuelson and
nation is loss aversion. In the short run, Richard Zeckhauser (1988); Knetsch
there is a significant chance that the re- (1989); George Loewenstein and Daniel
turn to stocks is negative, so if, as loss Adler (1995); Kaisa Herne (1998)-and
aversion implies, investors are particu- loss aversion may well explain other
larly sensitive to these possible negative puzzles in field data such as the dispar-
returns, that would explain the equity ity between measurements of willing-
premium for myopic investors. But just ness to pay and willingness to accept
how loss averse and how myopic do (see Ian Bateman et al. 1997) plus a va-
agents have to be for this explanation to riety of other examples relating to con-
work? Benartzi and Thaler show that, sumption and labor supply decisions
assuming people are roughly twice as discussed in Camerer (forthcoming). Yet
sensitive to small losses as to corre- more examples relating to the evalu-
sponding gains (which is broadly in line ation of opportunity costs, sunk costs,
with experimental data relating to loss and search behavior are discussed in the
aversion), the observed equity premium much earlier paper by Thaler (1980).
is consistent with the hypothesis that in- These examples show that there are
vestments are evaluated annually. This important field phenomena that non-
is a very simple, and to my mind, intui- expected utility models may be neces-
tively appealing account of another im- sary to explain. Of course no theory is
portant field phenomenon which has perfect, so might it be that these are
defied explanation in standard theory. exceptional cases and that EUT is still a
Notice that while loss aversion can be reasonable approximation for a wide
accommodated in conventional models range of field behavior? I suspect this is
little more than wishful thinking. In a take-up of new models. It is fair to say
new paper, Matthew Rabin (forthcom- that giving up EUT raises some deep
ing) presents a "calibration theorem" theoretical questions in fundamental ar-
which shows that expected utility theory eas of economics like game theory and
has some grossly implausible implica- the analysis of dynamic choices. For
tions. The central result is that an ex- example, Nash equilibrium may fail
pected utility maximizer who displays to exist with non-EU preferences, and
risk aversion in cases where outcomes choices may b^e dynamically inconsis-
are modest will display ludicrously high tent. There are difficult problems to ad-
degrees of risk aversion over large dress here, but in mind of that perhaps
stakes and the result holds for any con- we should expect progress to occur
cave utility function. Rabin argues that slowly and not take the gentle pace as
this general property of EUT is at odds evidence of intractability.
with intuition and observation in both It is worth noting that many standard
the lab and the field. For example, results and techniques are robust to
there is a huge amount of evidence some relaxations of the independence
which shows risk aversion over small axiom: that is, economic theory as we
stake laboratory choices, and explaining know it does not simply implode when
this using EUT implies pathologically non-expected utility preferences are al-
risk averse behavior over larger out- lowed. Although EUT has been a cen-
comes. Rabin's argument should jolt us tral building block in core areas of eco-
out of wishful thinking, since it suggests nomics, many tools and results that
that EUT is implausible as a general have been developed assuming it actu-
account of behavior under risk. ally require weaker assumptions. This
was an important message of Machina's
5.3 Theoretical Applications
(1982) analysis which goes through even
Since my concern is with theories as though his empirical hypothesis (that
descriptive models, I have placed em- indifference curves fan out) was not
phasis on assessing the predictive power supported by the data. His generalized
of alternative theories. But while pre- expected utility analysis enables us to
diction is important, it is not every- extend theoretical results and insights
thing. Other important questions sur- derived from EUT to a non-expected
round the theoretical usefulness of utility framework. For example, so long
alternatives to EUT. The standard the- as preferences have the necessary
ory is, without doubt, a potent simplifi- smoothness properties, we can charac-
cation which can be easily applied in a terize risk aversion, stochastic domi-
range of theoretical contexts, and its nance preference, and comparative risk
use is pervasive. While a good deal of aversion in terms of properties of local
effort has been devoted to developing utility functions. Hence, much of our
alternatives to EUT, by comparison, the understanding of these aspects of risk
use of such models in theoretical work preference remains intact for a wide
outside of the specialist literature has class of non-expected utility models
been limited. Does this suggest that al- (see Machina 1987, 1989). Moreover,
ternative models are too complex or in- our understanding of risk aversion has
tractable to be useful in a broader been substantially refined by discus-
theoretical context? In general I think sions of the concept in a non-expected
the answer is no and that other factors utility framework. For instance, the
most likely explain the relatively slow rank dependent approach allows us to
decompose risk aversion into elements quasi-convex, then they will be unwill-
deriving from, respectively, attitudes to ing to randomize in the way required
consequences and attitudes to chance for equilibrium. Those interested only
(Wakker 1994); and as we have seen, in normative analysis might brush this
such models also provide accounts of aside, but since we know that between-
observed risk behavior inconsistent with ness is not supported empirically, and
received notions of risk preference. violations of it often go in the direction
In several areas of applied theory, pa- of quasi-convexity (see Camerer and Ho
pers have emerged showing that well- 1994) a natural question for the de-
known results usually derived assuming scriptively minded economist is: how do
EUT do not rely on independence. For we analyze strategic behavior when
example, in the theory of auctions be- preferences are quasi-convex? Vincent
tweenness is sufficient to guarantee Crawford (1990) provides an answer
value revealing behavior in ascending showing that a generalization of Nash,
bid auctions (Karni and Safra 1989a); in the "equilibrium in beliefs," which coin-
a model of search discussed by Karni cides with the standard concept for
and Safra (1990), qualitatively similar quasi-concave preferences, also exists
stopping rules characterize the optimal when preferences are quasi-convex.
behavior of EU maximizers and non-EU Crawford's analysis makes an important
maximizers so long as the preferences theoretical step in showing how stan-
of the latter are quasi-convex. Such cor- dard game theoretic tools can be ex-
respondences, while interesting, are of tended to handle players with non-EU
course special cases. In general, the be- preferences. Other related discussions
havior of EU maximizers and non-EU of non-expected utility theory in the
maximizers does not coincide, so given context of games can be found in Karni
the well-documented predictive failure and Safra (1989a,b) and Dekel, Safra,
of EUT, there is surely a good case for and Segal (1991).
seeking to develop general tools of eco- It is true that giving up EUT has dra-
nomic analysis capable of handling non- matic implications in some areas of the-
EU preferences. Such tools would then ory, and one pertinent example is the
help us to understand what implications area of dynamic choice. If EUT does
failures of EUT have for a wider class not hold, then sequential choices may
of economic phenomena. While that be dynamically inconsistent. To appre-
sounds like a major undertaking, some ciate the significance of this, consider a
significant progress has already been sequential choice problem represented
made, and one example is in the area of by a standard decision tree. An agent who
game theory. is dynamically inconsistent may identify
It is well-known that if players' pref- an optimal path viewed from the initial
erences do not satisfy the independence choice node, but then be unwilling to
axiom of EUT, Nash equilibrium may take actions that form part of that opti-
fail to exist. Independence is not neces- mal path at choice nodes further down
sary for existence: quasi-concavity or the tree. Wakker (1999) suggests an
betweenness is enough. The problem analogy between dynamic inconsistency
case for standard game theory is quasi- and schizophrenia: the dynamically in-
convexity, and it is easy to understand consistent agent has something akin to a
the intuition behind this. In games split personality, with different aspects
where the only equilibria are mixed of the person revealing themselves in
strategies, if players' preferences are different parts of the tree. Although
368 Journal of Economic Literature, Vol. XXXVIII (Tune 2000)
some might regard this as a "problem" butions are Machina (1989b), Karni and
with non-expected utility models, I Safra (1989b, 1990), McClennen (1990)
think that conclusion could be mislead- and Segal (1990, 1997). Since the mod-
ing for two reasons, one theoretical, the els proposed by these authors give up
other empirical. different principles (for example, Ma-
From the theoretical point of view it china and McClennen relax consequen-
is important to note that relaxation of tialism; Segal (1990) relaxes the reduc-
independence does not necessarily im- tion of compound lotteries axiom) an
ply dynamic inconsistency. Machina obvious question to ask is: which princi-
(1989b) has shown that agents with ple or principles of dynamic choice are
non-expected utility preferences can be actually implicated when the indepen-
dynamically consistent if we are pre- dence axiom of EUT is violated? As yet,
pared to sacrifice the assumption of relatively little work has addressed this
consequentialism. An implication of issue directly, though a recent experi-
consequentialism in standard decision- mental investigation by Cubitt, Starmer,
tree analysis is that agents are entirely and Sugden (1998a) suggests a surpris-
forward looking: at any given decision ing answer. In this experiment, common
node, the consequentialist decision ratio type violations of independence
maker ignores any part of the tree that appear to be due to the failure of a
cannot be reached moving forward from principle which we call timing indepen-
that node. In contrast, Machina argues dence. The answer is surprising since
that risks born in the past may be rele- timing independence is implicit in most
vant to current decisions and he pro- proposed models of dynamic decision
vides some telling examples of where making: the only exception I know of is
that could be the case. As such he de- Karni and Safra's (1989b, 1990) model
fends the notion of a dynamically con- of behaviorally consistent choice.
sistent non-EU agent by rejecting I bring this section to a close with a
consequentialism. brief smorgasbord of applications. Non-
It has only recently been properly unexpected utility models have been used
derstood that axioms of EUT, including across a reasonably diverse range of
the independence axiom, follow from theoretical applications. Here are some
assuming certain principles of dynamic examples using conventional ap-
choice (see Hammond 1988; Edward proaches. Epstein and Zin (1989, 1991)
McClennen 1990; Robin Cubitt 1996). use non-expected utility models in the
This provides a new form of normative context of intertemporal consumption
defence for EUT. On the other hand, and asset demand. Their approach al-
since we know that independence fails lows the separation of parameters (i.e.,
empirically, at least one of the dynamic for risk aversion, intertemporal substi-
choice principles that jointly imply it tution, and preference over the timing
must be failing too. It follows that if we resolution of uncertainty) which are
want to predict the behavior of real confounded in the conventional ex-
agents in dynamic contexts we will need pected utility approach. Epstein (1995)
models of dynamic decision making that discusses a range of applications of
relax the suspect dynamic choice princi- other non-expected utility models in
ple(s) implicit in EUT. Several papers macroeconomics, finance, and game
have investigated models of dynamic theory. Epstein and Segal (1992) derive
decision making that relax standard as- a social welfare function based on qua-
sumptions. Among the important contri- dratic utility theory. Segal (1988b), Karni
(1995), and Machina (1995) consider utilizes the notion of regret. A con-
the implications of non-expected utility nected literature suggests that consid-
for insurance. Chew (1985) and Neilson erations of regret may be important par-
(1992) apply the model of implicit util- ticularly in the context of health-related
ity to, respectively, demand revelation decisions. For example, Weinstein
in an auction context and asset demand. (1986) considers whether regret ought
The most popular non-expected util- to play a role in clinical decisions; Rene
ity models in applied work have been Richard (1994) presents survey evidence
those based on the rank dependent indicating that anticipation of post-deci-
form. Early applications include Yaari's sional emotions, particularly regret,
(1987) analysis of portfolio selection; may be important factors in decisions
Chew, Karni, and Safra (1987) on risk relating to contraception. Richard ar-
aversion; Segal, Spivak, and Ziera (1988) gues that it may be possible to influ-
on savings and risk aversion; Epstein ence individual behavior through in-
and Zin (1990) on asset pricing in a re- formation campaigns that highlight
cursive utility framework; plus Quig- potential regrets, and he discusses the
gin's (1991a,b; 1993) discussions of implications for public policy in relation
comparative statics, optimal lottery de- to, among other things, the spread of
sign, portfolio selection, and informa- sexually transmitted diseases such as
tion revelation. More recently, the model AIDS. Richard Smith (1996) proposes a
has been applied to a diverse range of modified version of regret to account
topics, including work by Chateauneuf for the valuation of health states. Not
and Cohen (1994) and Cohen (1995) on all of these applications involve formal
risk aversion; Eide (1995) on the effects economic models. They are none the
of punishment in a model of crime; a less interesting. Indeed, there is a sense
series of papers investigating utility elici- in which the formal models are lagging
tation procedures (Wakker and Daniel behind the less formal discussions.
Deneffe 1996; Han Bleichrodt and There is considerable evidence that
Quiggin 1997; and Bleichrodt, Jose Luis considerations of regret do influence
Pinto, and Wakker 1999); plus an analy- ordinary people in important decisions.
sis of dynamic decision making which At the same time, it seems clear that
allows a psychological role for anxiety in our existing models do not provide good
relation to unresolved lotteries (Wu formalization of such processes.
1999). It has to be said that, overall, the vol-
Attempts have also been made to ap- ume of work applying non-expected
ply nonconventional theories. The ap- utility models looks quite small given
plications of prospect theory discussed how long some of the theories have
in Section 5.2 (above), Thaler (1980) been available. I think things may be
and Camerer (forthcoming), are rele- changing and that we will see increasing
vant here. Also Jeffrey Rachlinski use of models based on the rank depen-
(1990, 1994, 1996) has applied prospect dent form. Until recently, the sheer va-
theory to the economics of litigation riety of competing models probably
and civil negotiation. Applications of re- counted against their use. Too many al-
gret theory include Cubitt and Sugden's ternatives were on offer with no obvious
(1998) model of preference evolution (I way to discriminate between them (bear
discuss this below). Milton Weinstein in mind that many of these theories
and Robert Quinn (1983) propose a were proposed to explain the same, rela-
model of post-decisional "blame" which tively small, set of choice anomalies).
But now that much more evidence has these models are essentially conventional,
accumulated, it seems clear that there and since their use seems to be expand-
are quantitatively important phenomena ing, general claims to the effect that they
that should not be ignored in general are intractable or not useful in econom-
economic analysis. One of these is ics more broadly seem unconvincing.
surely the phenomenon of nonlinear Perhaps there is a case for thinking that
probability weighting. The rank depen- the position we should now aim for is
dent model is likely to become more one in which models like cumulative
widely used precisely because it cap- prospect theory become the default in
tures this robust empirical phenomenon applied economics with EUT used as a
in a model which is quite amenable to convenient special case, but only when
application within the framework of we can be confident that loss aversion
conventional economic analysis. and probability weighting are insignifi-
Loss aversion is another empirically cant. While that position may be some
important concept, and I sense that way off, my prediction is that the use
economists are becoming more inter- of models incorporating probability
ested in studying the implications of as- weights and loss aversion will grow rap-
suming loss-averse preferences for a idly, and my normative judgement is
range of economic issues. Tversky and that, if it doesn't, it ought to.
Kahneman (1991) present a model-
6. Phase III: New Directions in the
based on prospect theory-that applies
Theory of Choice under Risk
the ideas of reference dependence and
loss aversion in riskless choice, and at- I have argued that there is an estab-
tempts are currently under way to exam- lished case for taking non-expected util-
ine the implications of rank dependent ity models seriously and, further, taking
preferences for fundamental theoretical unconventional approaches to modeling
issues in economics. For example, Munro seriously. My arguments have, in several
(1998) examines the implications for places, relied heavily upon experimental
welfare economics of assuming reference evidence to make the case. This strategy,
dependent preferences; Munro and however, begs an important question,
Sugden (1998) examine the conditions which is: are the choices of badly be-
necessary for general equilibrium in an haved experimental subjects particu-
economy where agents have reference larly, if at all, relevant to economic en-
dependent preferences. quiry? The experimental paradigm is
Sign and rank dependent models- relatively new to economics and is still
like cumulative prospect theory-cap- viewed with suspicion by some mem-
ture both of these empirically important bers of the discipline. Since most of the
phenomena in a theoretically compact data driving developments in this area
way. And, while not all of the empirical have come from experimental investiga-
evidence fits this approach, it does pro- tions, could it be that choice anomalies
vide an account consistent with some of are revealing defects of the experimen-
the most robust stylized facts from a tal method (in relation to economics) as
range of experimental studies.28 Since opposed to flaws in conventional eco-
nomic theory? I have argued elsewhere
28 For those interested in where rank-dependent (Starmer 1999b) that it would be hard
models fail, aside from the cases discussed in Section
for economists to argue for a blunt re-
5.1.2 which count against all conventional theo-
ries, see Wakker, Erin, and Weber (1994), Wu (1994) jection of experimental data per se. That
and Bimbaum, Jamie Patton, and Melissa Lott (1999). said, there are grounds for meaningful
debate about the theoretical and em- to learn. Of course, many important de-
pirical significance of laboratory-ob- cisions are taken rarely and afford lim-
served choice anomalies. Some econo- ited opportunity for repetition, change,
mists have challenged the empirical learning, and so on. Nevertheless, it is
significance of laboratory anomalies for interesting to ask whether the "anoma-
economic investigation by suggesting lies" observed in laboratory behavior
that they may not generalize to eco- disappear in market contexts or where
nomically meaningful contexts. Since appropriate conditions for learning exist.
laboratory experiments are usually de- Charles Plott (1996) argues that, in
signed to control precisely those vari- the presence of sufficiently strong in-
ables that economic theories identify as centives, laboratory anomalies tend to
important, the arguments along these disappear when subjects are allowed to
lines that merit serious consideration, adjust their behavior on the basis of ex-
in my view, are those that offer some perience gained through repetition. As
account of why behavior in laboratory Plott puts it: "Behavior seems to go
contexts may not extend to contexts of through stages of rationality that begin
general concern. I shall focus on two with a type of myopia when faced with
lines of argument of particular interest unfamiliar tasks. With incentives and
here as they are beginning to open up practice, which take the form of repeated
new and exciting theoretical accounts of decisions in the experimental work (but
individual choice behavior. might include play, banter, discussions
with others, stages of commitment, etc.),
6.1 The Evolution of Preference
the myopia gives way to what appears to
Most of the data we have been dis- be a stage of more considered choices
cussing has come from individual choice that reflect stable attitudes or prefer-
experiments where subjects undertake a ence" (1996, p. 248). Plott calls this the
series of one-off tasks. Typically both discovered preference hypothesis. The
the tasks and the environment will be argument is essentially empirical: he
unfamiliar to subjects. Moreover, al- draws on a range of experimental evi-
though most experiments involve real dence indicating a tendency for anoma-
-usually monetary-incentives, the most lous behavior to converge on the predic-
common reward mechanism is the ran- tions of economic theory when choices
dom lottery incentive system. In experi- are repeated in market-like settings.
ments with this design, subjects are re- The collected evidence is, without doubt,
warded according to their response to one impressive, but the vast majority of it
task which is randomly selected at the relates to experiments in which prefer-
end of the experiment.29 This provides ences of the experimental subjects have
little or no opportunity for subjects to been "induced" or controlled.30 The
revise their behavior in the light of feed- purpose of such experiments is usually
back on the consequences of prior choices. to investigate whether particular market
Some have argued that such "raw" be- forms generate the equilibria predicted
havior may have little in common with by economic theory when we know what
the behavior of agents in economic en- the subjects' preferences are. For our
vironments where there is opportunity purposes, however, the more relevant
question is whether agents who lack,
29 For a discussion of this random lottery design,
and evidence on its validity, see Starmer and Sug-
den (1991); Jane Beattie and Loomes (1997); and 30 For an account of the induced preference
Cubitt, Starmer, and Sugden (1998b). methodology see Smith (1976).
say, expected utility preferences might market is introduced; it occurs because

evolve behavior more consistent with the chosen market mechanism (a
them through market participation. sealed-bid auction) happens to select a
There has been relatively little empiri- price in the middle of the distribution
cal investigation of how preferences of bids. An experiment conducted by
themselves might be affected by expo- John Bone, John Hey, and John Suck-
sure to market mechanisms. Plott dis- ling (1999) suggests that repetition and
cusses only one study with a direct group discussion increases Allais-type
bearing on this. He cites experimental violations of EUT. Thus, there is a good
work by James Cox and David Grether case for thinking that patterns of behav-
(1996) as support for the conclusion ior change in some environments in-
that "the classical preference reversal volving markets and/or repetition, but
can be seen as a product of inexperience as yet there is no sound empirical basis
and lack of motivation, and it goes away for asserting a general tendency towards
with experience of a market setting" expected utility preferences under "mar-
(Plott 1996, p. 231). Although there is ket conditions." The evidence is at best
some support for that claim in Cox and mixed. Is there a theoretical basis for as-
Grether, their results are complicated suming that behavior may evolve towards
and do not provide the basis for a gen- expected utility (or anything else)?
eral assertion that anomalies like pref- For decisions that are rare and/or ir-
erence reversal tend to disappear in any reversible (e.g. childbearing, marriage,
economically relevant market context. job-taking, decisions relating to health
There is some evidence, albeit rela- and education) the scope for learning
tively limited, relating to whether other must be limited. Repeated choices that
violations of expected utility persist in generate feedback (e.g. stock market
environments that might allow subjects' decisions, some consumption decisions)
behavior to evolve. For example, Mik- are the more obvious candidates. As yet
hail Myagkov and Plott (1997) find evi- economics lacks any well-developed
dence consistent with prospect theory theoretical account of learning, the im-
(e.g. implications of reference depen- pact of incentives, and so on. Moreover,
dence and diminishing sensitivity) in an at least some of the work that does exist
experimental market environment with shows there can be no presumption that
repeated decisions. While the authors learning will always generate behavior
identify some tendency for behavior to that converges on full-blown rationality
change in a direction consistent with ex- (see for example Timermann 1994).
pected utility preferences, the data do There is, of course, a considerable and
not reveal general convergence on the growing volume of research on evolu-
predictions of the standard model. Even tionary models in economics-a good
where violations of EUT do fall in mar- part of this literature is reviewed by
ket contexts, preference discovery may Richard Nelson (1995)-though so far
not be the right interpretation of the there has been relatively little attempt
effect. For example, Dorla Evans (1997) to examine whether theories of prefer-
finds that using a market mechanism to ence may be based on evolutionary
elicit valuations for gambles, rather foundations. Exceptions are Karni and
than an individual pricing task, leads to David Schmeidler (1986) who argue
a marked reduction in betweenness vio- that the expected utility hypothesis may
lations. But the effect is not due to any be derived from a principle of self-pres-
preferences having changed when the ervation, plus recent papers by Tilman
Borgers and Rajiv Sarin (1996, 1997) typical strategy has been to test
and Cubitt and Sugden (1998). These whether observed violations of a theory
later contributions which examine mod- can be explained as "random error" and
els in which preferences evolve under theories are rejected only when the de-
the pressure of some selection mecha- parture seems systematic (i.e., non-ran-
nism reach more skeptical conclusions. dom). For example, in an experimental
Although the selection mechanisms test for a common ratio effect in prob-
considered by the two sets of authors lems like those of Figure 2, an individ-
are quite different-in Borgers and Sarin, ual subject could violate expected util-
the selection mechanism is reinforce- ity in either of two ways: they could
ment learning, in Cubitt and Sugden it choose s1** then r2** (as predicted by
is imitation-a common conclusion fanning-out theories) or they could
emerges: expected utility preferences choose ri** then s2**. It would be com-
evolve only under restrictive assump- mon for researchers to adopt the null
tions. Moreover, Cubitt and Sugden ar- hypothesis that individuals choose ac-
gue that imitation does not even imply cording to expected utility plus random
pressure to select for transitive prefer- error and to further assume that ran-
ences over prospects. While there re- dom error implies the two violating pat-
mains much to be done before we could terns are equally likely. Hence, on this
claim to have an adequate under- procedure, simply observing some peo-
standing of the evolution of behavior in ple choosing consistently with fanning
real market environments, already it out is not enough to reject the null; the
seems clear that it would be rash to as-. typical test examines whether violation
sume a general tendency for individual consistent with the common ratio effect
behavior in relation to risk to converge occurs significantly more frequently
with conventional economic assump- than its mirror image. Similarly, in test-
tions simply by virtue of it taking place ing for the cycle of choice predicted by
within some real market institution. regret theory, some researchers have
operationalized a null hypothesis of
6.2 Theories of Stochastic Preference
EUT plus error by assuming that the
In the 1990s a new direction in the probability of observing the predicted
hunt for a descriptive theory has choice cycle equals that of observing a
emerged with the publication of a num- cycle in the opposite direction. While
ber of papers investigating models of such assumptions lacked any substan-
stochastic preference. This literature is tive theoretical underpinnings, in the
generating new ways of modeling choice absence of any generally accepted the-
behavior and new methods for testing ory of randomness, researchers had lit-
existing theories. It also provides a rea- tle choice but to base empirical tests
son to re-examine some of the earlier on ad hoc assumptions about an error
conclusions drawn from experimental generating process.
evidence. A series of recent papers by Hey and
The models we have discussed so far Chris Orme (1994), Harless and Camerer
are essentially deterministic and, if in- (1994), and Loomes and Sugden (1995)
terpreted literally each could be re- has opened the way to a more general
jected by a single contrary observation. approach by suggesting different ways
This seems too strong a test, and in of modeling stochastic preferences. Each
most empirical work researchers have approach provides a general framework
interpreted theories stochastically. The for developing stochastic versions of
alternative (deterministic) "core" theo- core theory, but the parameters of the
ries of preference such as expected util- core theory to be applied to any given
ity theory or some non-expected utility choice are determined by a random pro-
theory, though the interpretation of the cess. So, if the core theory is expected
source of randomness differs between utility, the random draw determines
the models. In the Hey and Orme the individual's degree of risk aversion,
model, the choice between a pair of independently, for each choice. In this
prospects q and r is determined by the case, the stochastic element is inherent
sign of: in preferences as opposed to random
deviation about true preferences.
HO = [V(q) - V(r)] + e
It is well-known that the data from
where V(.) is the preference function of choice experiments show a great deal of
a deterministic core theory. If c = 0, variability. For instance, a common
preferences are determined purely by the finding is that individuals confronted
core theory, and if HO is positive (nega- with the same pairwise choice problem
tive) then q (r) is chosen. Hey and Orme twice within a given experiment fre-
model the stochastic component by as- quently give different responses on the
suming that e is a normal variate with a two occasions.31 Stochastic choice is
mean of zero. Consequently, the sign of more convincing than indifference as an
HO can be reversed by the draw of a account for such intrinsic variability,
large enough e with the "right" sign. This but explicitly introducing randomness
is essentially the Fechner model discussed into models of choice also provides al-
earlier by Gordon Becker, Morris DeGroot, ternative possible explanations for at
and Jacob Marchak (1963). Hey and Orme least some of the violations of EUT that
interpret the randomness in this model motivated non-expected utility models.
as some kind of calculation error. Notice For instance, assuming expected utility
that the larger the difference in values as the core theory, Loomes and Sugden
assigned to the prospects by the core show that all three error models allow
theory, the less likely it is that true pref- the possibility of systematic cycles of
erences will be overturned by the error choice. Thus the introduction of a sto-
term. Harless and Camerer's approach is chastic component might allow the ex-
different. They assume that any decision planation of intransitive choice, without
reveals true preferences (as defined by giving up the assumption of transitive
the core theory) with probability 1 - e, preferences in the core theory. They
but there is some constant probability e, also show that a model combining EUT
that the individual chooses at random. with the HO specification can generate
The error generating mechanism here is behavior consistent with fanning out (or
akin to the trembling hand idea, familiar fanning in) like the common ratio ef-
in game theory. In these first two models fect. Given this, it seems natural to ask
the stochastic element reflects deviations just how much of the known data could
from "true" preferences resulting from be explained by a stochastic version of
miscalculations, slips or trembles, and so EUT.
on. Loomes and Sugden consider a ran- An analysis of experimental choice
dom preference model-also discussed by 31 A variety of studies including Starmer and
Becker, deGroot, and Marchak-which Sugden (1989), Camerer (1989), Hey and Orme
has a different interpretation: for any (1994), and T. Parker Ballinger and Nathaniel Wil-
cox (1997) find that between one-quarter and one-
given choice, the individual acts on pref- third of subjects "switch" preferences on repeated
erences satisfying the restrictions of the questions.
data conducted by Harless and Camerer sions may be sensitive. A parallel study
(1994) sheds some light on this. They by Hey and Orme (which appears back-
proposed their "tremble" theory as part to-back with Harless and Camerer in
of an econometric methodology de- the 1994 volume of Econometrica) con-
signed to compare the predictive power cludes with a more positive assessment
of EUT and a variety of alternatives of EUT. Their analysis, which uses the
against experimental choice data while HO error specification, allows the best
allowing for randomness. Their study, model to be determined separately for
which examined data from tens of thou- each individual (Harless and Camerer
sands of choices from over twenty prior fit models to aggregate data). Hey and
studies, provides a highly comprehen- Orme find that EUT works as well as
sive snap-shot of the evidence. The ap- any other model for a substantial minor-
proach they take allows them to assess ity of their subjects (almost 40 per-
the relative performance of alternative cent). Even so, for the majority of their
theories given different trade-offs be- subjects at least one non-expected util-
tween parsimony (the number of pat- ity model outperforms the standard the-
terns the theory allows) and predictive ory and the rank dependent model is
accuracy. With most emphasis placed on a leading contender (along with the
predictive accuracy (and least on parsi- quadratic utility model).
mony) Harless and Camerer find that Once we think in terms of alternative
the "best" theory will be one that allows models of error, the problem of theory
some mixed fanning in and fanning out; selection becomes more complex since
with the premium on parsimony, simple the number of available models is now
models like expected utility, or expected the product of the available core theo-'
value, win. Interestingly, however, when ries and error specifications. Harless
choices involve prospects with "mixed and Camerer, and Hey and Orme, each
support" (i.e. involve some choices on compare models for a given error speci-
triangle boundaries), they find that there fication. A number of researchers have
is no trade-off between parsimony and now begun to address the problem of
fit which picks expected utility as the choosing between error specifications.
best theory. In general they conclude Loomes and Sugden (1998) and Enrica
that "The pairwise-choice studies sug- Carbone (1997) compare the predictive
gest that violations of EU are robust performance of EUT under each of the
enough that modeling of aggregate eco- three error specifications. A tentative find-
nomic choice behavior based on alterna- ing from this research seems to be that
tives to EU is well worth exploring" the trembling hand model performs
(p. 1287). In addition, they find that relatively poorly. However, Loomes,
there is room for improvement in the Moffat, and Sugden (1998) investigate
predictive power of new theories: "For the possibility of two more sophisti-
every theory there is systematic variation cated error theories that extend either
in excluded patterns which could in the Fechner model or the random pref-
principle be explained by a more refined erence model to allow the possibility of
theory" (p. 1284). trembles. They compare two core theo-
Although this study is based on an ries-expected utility and rank-depen-
impressive data base, it is important to dent expected utility-and find that the
recognize that their analysis involves data supports rank-dependent expected
specific assumptions about error, parsi- utility combined with a random prefer-
mony, and so on, to which the conclu- ence error term plus trembles. They
report another finding too, which gives He thought a central theoretical prob-
pause for thought. Their data derives lem remained to be solved before we
from an experiment in which subjects could claim to have developed a reason-
were confronted with repeated choices. ably general descriptive model of choice.
They find two significant trends in their In my view they were both right.
data: (i) randomness due to trembles There is substantial evidence that
tended to decay with experience; and EUT is likely to be descriptively mis-
(ii) deviations from EUT decayed with leading in at least some important
experience. Is this evidence of individu- contexts and, given the accumulating
als discovering expected utility prefer- evidence supporting, in particular,
ences? I think the truth is that we don't probability weighting and loss aversion,
yet know. There is growing evidence we have at least some well-grounded
that there are some dynamic processes hypotheses about important factors gen-
at work in repeated choice contexts. For erating departures from the standard
example, Hey and Orme (1994), Ballin- theory. From a practical point of view
ger and Wilcox (1997), and Loomes and the rank-dependent models (including
Sugden (1998) all report systematic the sign-dependent variants) provide a
variation in risk aversion across re- convenient way of modeling these well
peated choices. In my view, an examina- established influences on choice, and
tion of those processes constitutes an- there seems good reason to push for-
other extremely interesting avenue for ward the task of examining what impli-
future research. cations such models have in general
economic contexts.
7. Concluding Thoughts At the same time, there seems to be a
good case for pushing at the limits and
In 1994 I attended a conference on
perhaps stepping beyond the bounda-
the' foundations of utility and risk32
ries of conventional theorizing. In the
which concluded with a roundtable dis-
past, some have justified the conven-
cussion where members of a panel, in-
tional approach on the grounds that its
cluding Mark Machina and Robert Sug-
assumptions were supported by the evi-
den, were asked to offer their personal
dence. It is hard to do that in any con-
reflections on the future of non-
vincing way now, as it seems quite plain
expected utility. Machina argued that
that real behavior refuses to be con-
while much good work had been done
fined by the limits of conventional theo-
in developing theoretical alternatives to
rizing. We are discovering that a wide
EUT, the field of applications was rela-
variety of behavior, in both the lab and
tively underdeveloped. Perhaps the
the field, cannot be adequately ex-
time had come, he suggested, to spend
plained within the conventional frame-
more time considering the implications
work. While existing unconventional
of these alternatives for a wider collec-
theories have their own limitations,
tion of economic problems. Sugden, on
models like regret and prospect theory
the other hand, suggested that the gath-
(the original version) illustrate the pos-
ering evidence shows actual choice be-
sibility of working outside the conven-
havior to be more complex than our
tional framework. Insights derived from
models, and he voiced skepticism about
these theories have proved useful in un-
the ability of conventional theorizing to
derstanding real behavior even if the
provide an account of such complexity.
formal theories have not been widely
32 FUR VII, Oslo. applied. Moreover, the investigation of
these theories has led to the discovery hope is that research on these topics
of new empirical phenomena, including, will continue to be driven by an effort
for example, the discovery of the cycli- to confront our theories with the evi-
cal choices predicted by prospect the- dence, however challenging the data
ory and the event-splitting effect. This mavb hP.
must surely count as progress too. REFERENCES
Let me offer a final thought about
Abdellaoui, Mohammed. 1998. "Parameter-Free
the significance of the evidence that has Eliciting of Utilities and Probability Weighting
been driving this research program. Functions," working paper, GRID, ENS,
Cachan, France.
Perhaps "anomalies" in choice behavior
. 1999. "A Genuine Rank-Dependent Gen-
will turn out to be insignificant for a eralization of von Neumann-Morgenstern Ex-
broad range of economic problems. I pected Utility Theorem," working paper,
GRID, ENS, Cachan, France.
have to confess that could be true. On
Ali, Mukhtar M. 1977. "Probability and Utility Es-
the other hand, theories of choice lie at timates for Racetrack Bettors," J. Polit. Econ.
the very heart of economics, and the 85, pp. 803-15.
Allais, Maurice. 1953. "Le Comportement de
data we have suggests that choice be-
l'Homme Rationnel devant le Risque: Critique
havior displays complex patterns in des Postulats et Axiomes de l'Ecole Ameri-
even very simple contexts. Given that caine," Econometrica, 21, pp. 503-46.
. 1979. "The Foundations of a Positive The-
datum line, is there a good reason for
ory of Choice Involving Risk and a Criticism of
thinking that behavior in a more com- the Postulates and Axioms of the American
plex reality will better conform to our School," in Expected Utility Hypotheses and the
Allais Paradox. M. Allais and O. Hagen, eds.
simple models? It would surely be
Dordrecht: Reidel.
genuinely surprising if that were true, Anand, Paul. 1993. Foundations of Rational
and a major intellectual achievement if Choice under Risk. Oxford U. Press and Claren-
it could be demonstrated. Had it not don.
Ballinger, T. Parker and Nathaniel T. Wilcox.
been for the program of research inves- 1997. "Decisions, Error and Heterogeneity,"
tigating failures of EUT and alternative Econ. J. 107:443, pp. 1090-105.
Bateman, Ian; Alistair Munro, Bruce Rhodes,
models of choice, we may not even have
Chris Starmer, and Robert Sugden. 1997. "A
asked such questions. Before this work Test of the Theory of Reference Dependent
became well known, most economists Preferences," Quart. J. Econ. 112:2, pp. 479-
505.
probably took it for granted that EUT
Battalio, Raymond C.; John. H. Kagel, and
was the right model of individual Komain Jiranyakul. 1990. "Testing between Al-
choice. Even if it turns out that EUT ternative Models of Choice under Uncertainty:
Some Initial Results," J. Risk Uncertainty, 3,
can be supported using arguments
pp. 25-50.
based on learning or evolution, we will Beattie, Jane and Graham Loomes. 1997. "The
have learned something new and impor- Impact of Incentives upon Risky Choice Experi-
ments," J. Risk Uncertainty, 14, pp. 149-62.
tant, that is, why and in what circum-
Becker, Gordon M.; Morris H. DeGroot, and Ja-
stances EUT applies. Currently, one can cob Marschak. 1963. "Stochastic Models of
point to some evidence that behavior Choice Behavior," Behavioral Sci. 8, pp. 41-55.
Becker, Joao L. and Rakesh K. Sarin. 1987. "Gam-
becomes "better-behaved" by exposure
ble Dependent Utility," Manage. Sci. 33, pp.
to markets, but there is relatively little 1367-82.
evidence for this and it is mixed. It Bell, David. 1982. "Regret in Decision Making un-
der Uncertainty," Operations Res. 20, pp. 961-
seems to me that some of the most in-
81.
teresting current research relates to -. 1985. "Disappointment in Decision Mak-
how behavior evolves over time and ing under Uncertainty," Operations Res. 33, pp.
1-27.
across different institutional settings.
Benartzi, Shlomo and Richard H. Thaler. 1995.
As yet we have only the vaguest appre- "Myopic Loss Aversion and the Equity Pre-
ciation of the mechanisms involved. My mium Puzzle," Quart. J. Econ. 110, pp. 73-92.
Bernoulli, Daniel. 1954. "Exposition of a New tic Preference Theory Using Experimental
Theory on the Measurement of Risk," (original Data," Econ. Letters, 57:3, pp. 305-11.
1738) Econometrica, 22, pp. 23-36. Chateauneuf, Alain and Michele Cohen. 1994.
Birnbaum, Michael H. anc Juan B. Navarrete. "Risk Seeking with Diminishing Marginal Util-
1998. "Testing Descriptive Utility Theories: ity in a Non-Expected Utility Model," J. Risk
Violations of Stochastic Dominance and Cumu- Uncertainty, 9, pp. 77-91.
lative Independence," J. Risk Uncertainty, 17, Chew, Soo Hong. 1983. "A Generalization of the
pp. 49-78. Quasilinear Mean with Applications to the Mea-
Birnbaum, Michael H. and Laura A. Thompson. surement of Income Inequality and Decision
1996. "Violations of Monotonicity in Choices Theory Resolving the Allais Paradox," Econo-
between Gambles and Certain Cash," Amer. J. metrica, 51, pp. 1065- 92.
Psych. 109, pp. 501-23. 1985. "Implicit-Weighted Utility Theories,
Birnbaum, Mi6hael H.; Jamie N. Patton, and M-Estimators, and Non-Demand Revelation of
Melissa K. Lott. 1999. "Evidence against Rank- Second-Price Auctions for an Uncertain Auc-
Dependent Utility Theories: Tests of Cumula- tioned Object," working paper 15, Johns Hop-
tive Independence, Interval Independence, kins U.
Stochastic Dominance, and Transitivity," Org. . 1989. "Axiomatic Utility Theories with the
Behavior Human Decision Processes, 77:1, pp. Betweenness Property," Annals Operational
44-83. Res. 19, pp. 273-98.
Bleichrodt, Han and John Quiggin. 1997. "Charac- Chew, Soo Hong and Larry G. Epstein. 1989. "A
terising QALYs under a General Rank-Depen- Unifying Approach to Axiomatic Non-expected
dent Utility Model," J. Risk Uncertainty, 15:2, Utility Theories," J. Econ. Theory, 49, pp. 207-
pp. 151-65. 40.
Bleichrodt, Han; Jose Luis Pinto, and Peter P. Chew, Soo Hong; Larry G. Epstein, and Uzi Segal.
Wakker. 1999. "Using Descriptive Findings of 1991. "Mixture Symmetry and Quadratic Util-
Prospect Theory to Improve Prescriptive Appli- ity," Econometrica, 59, pp. 139-63.
cations of Expected Utility," plenary presenta- Chew, Soo Hong, Larry G. Epstein, and Peter P.
tion at SPUDM, Mannheim. Wakker. 1993. "A Unifying App roach to Axi-
Bone, John; John Hey, and John Suckling. 1999. omatic Non-Expected Utility Theories: Correc-
"Are Groups More (or Less) Consistent than tion and Comment," J. Econ. Theory, 59, pp.
Individuals?" J. Risk Uncertainty, 18, pp. 63- 183-88.
81. Chew, Soo Hong; Edi Karni, and Zvi Safra. 1987.
Borch, Karl H. 1974. The Mathematical Theory of "Risk Aversion in the Theory of Expected Util-
Insurance. Lexington, MA: .Lexington Books. ity with Rank-dependent Probabilities," J. Econ.
Borgers, Tilman and Rajiv Sarin. 1996. "Naive Re- Theory, 42, pp. 370-81.
inforcement Learning with Endogenous Aspira- Chew, Soo Hong and Kenneth MacCrimmon.
tions," mimeo, U. College London. 1979a. "Alpha-nu Choice Theory: a Generalisa-
-. 1997. "Learning through Reinforcement tion of Expected Utility Theory," working paper
and Replicator Dynamics," J. Econ. Theory, 77, 669, U. British Columbia.
pp. 1-14. . 1979b. "Alpha Utility Theory, Lottery
Camerer, Colin F. 1989. "An Experimental Test of Composition, and the Allais Paradox," Faculty
Several Generalised Utility Theories," J. Risk of Commerce and Business Admin. working pa-
Uncertainty, 2:1, pp. 61-104. per 686, U. British Columbia.
. 1992. "Recent Tests of Generalizations of Chew, Soo Hong and William S. Waller. 1986.
Expected Utility Theory," in Utility: Theories "Empirical Tests of Weighted Utility Theory,"
Measurement, and Applications. Ward Ed- J. Math. Psych. 30, pp. 55-72.
wards, ed. Norwell, MA: Kluwer, pp. 207-51. Cohen, Michele. 1995. "Risk-aversion Concepts in
. 1995. "Individual Decision Making," in Expected- and Non-expected Utility Models,"
Handbook of Experimental Economics. J. Kagel Geneva Papers Risk Insurance Theory, 20, pp.
and A. E. Roth, eds. Princeton: Princeton U. 73-91.
Press. Conlisk, John. 1989. "Three Variants on the Allais
. 1998. "Bounded Rationality in Individual Example," Amer. Econ. Rev. 79, pp. 392-407.
Decision Making," Experimental Econ. 1, pp. 1996. "Why Bounded Rationality," J.
163-83. Econ. Lit. 34:2, pp. 669-700.
. Forthcoming. "Prospect Theory in the Cox, James C. and David M. Grether. 1996. "The
Wild: Evidence from the Field," in Choices, Preference Reversal Phenomenon: Response
Values and Frames. Daniel Kahneman and Mode, Markets and Incentives," Econ. Theory,
Amos Tversky, eds. Cambridge: Russell Sage 7, pp. 381-405.
Foundation. Crawford, Vincent, P. 1990. "Equilibrium without
Camerer, Colin F. and Teck-Hua Ho. 1994. "Vio- Independence," J. Econ. Theory, 50, pp. 127-
lations of the Betweenness Axiom and Non- 54.
linearity in Probability," J. Risk Uncertainty, 8, Cubitt, Robin P. 1996. "Rational Dynamic Choice
pp. 167-96. and Expected Utility Theory," Oxford Econ. Pa-
Carbone, Enrica. 1997. "Investigation of Stochas- pers, 48, pp. 1-19.
Starmer: Develovments in Non-Exvected Utilitu Theoru 379
Cubitt, Robin P.; Chris Starmer, and Robert Sug- Gonzalez, Richard and George Wu. 1999. "On the
den. 1998a. "Dynamic Choice and the Common Shape of the Probability Weighting Function,"
Ratio Effect: An Experimental Investigation," Cognitive Psych. 38, pp. 129-66.
Econ. J. 108:450, pp. 1362-80. Green, Jerry and Bruno Jullien. 1988. "Ordinal In-
-. 1998b. "On the Validity of the Random dependence in Nonlinear Utility Theory," J.
Lottery Incentive System," Experimental Econ. Risk Uncertainty, 1, pp. 355-87.
1, pp. 115-31. Gul, Faruk. 1991. "A Theory of Disappointment in
Cubitt, Robin and Robert Sugden. 1998. "The Se- Decision Making under Uncertainty," Econo-
lection of Preferences through Imitation," Rev. metrica, 59, pp. 667-86.
Econ. Stud. 65, pp. 761-71. Hagen, Ole. 1979. "Towards a Positive Theory of
Dekel, Eddie. 1986. "An Axiomatic Charac- Preferences under Risk" in Expected Utility Hy-
terization of Preferences under Uncertainty: potheses and the Allais Paradox. M. Allais and
Weakening the Independence Axiom," J. Econ. 0. Hagen, eds. Dordrecht: Reidel.
Theory, 40, pp. 304-18. Handa, Jagdish. 1977. "Risk, Probability, and a
Dekel, Eddie.; Zvi Safra, and Uzi Segal. 1991. "Ex- New Theory of Cardinal Utility," J. Poult. Econ.
istence and Dynamic Consistency of Nash Equi- 85, pp. 97-122.
librium with Non-expected Utility Preferences," Hammond, Peter. 1988. "Consequentialist Foun-
J. Econ. Theory, 55, pp. 229-46. dations for Expected Utility," Theory and Deci-
Diecidue, Enrico and Peter P. Wakker. 1999. "On sion, 25, pp. 25-78.
the Intuition of Rank-dependent Expected Util- Harless, David W. 1992. "Predictions about Indif-
ity," paper presented at FUR IX, Marrakesh. ference Curves inside the Unit Triangle: a Test
Edwards, Ward. 1955. "The Prediction of Deci- of Variants of Expected Utility," J. Econ. Behav-
sions among Bets," J. Experimental Psych. 50, ior Org. 18, pp. 391-414.
pp. 201-14. Harless, David W. and Colin F. Camerer. 1994.
. 1962. "Subjective Probabilities Inferred "The Predictive Utility of Generalized Expected
from Decisions," Psych. Rev. 69, pp. 109-35. Utility Theories," Econometrica, 62, pp. 1251-
Eide, Erling 1995. "RDEU Models of Crime," 89.
Law and Economics working paper C-1, Uni- Hausman, Daniel. 1992. The Inexact and Separate
versitetet i Oslo. Science of Economics. Cambridge U. Press.
Epstein, Larry G. 1992. "Behavior under Risk: Re- Heath, Chip; Steven Huddart, and Mark Lang.
cent Developments in Theory and Applica- 1999. "Psychological Factors and Stock Option
tions," in Advances in Economic Theory: Sixth Exercise," Quart. J. Econ. 114:2, pp. 601-27.
World Congress, Vol. 2. Jean-Jacques Laffont, Herne, Kaisa. 1998. "Testing the Reference-De-
ed. Cambridge: Cambridge U. Press. pendent Model: An Experiment on Asymmetri-
Epstein, Larry G. and Uzi Segal. 1992. "Quadratic cally Dominated Reference Points," J. Behav-
Social Welfare Functions," J. Polit. Econ. 100, ioral Decision Making, 11, pp. 181-92.
pp. 691-712. Hey, John D. and Chris Orme. 1994. "Investigat-
Epstein, Larry G. and Stanley E. Zin. 1989. "Sub- ing Generalizations of Expected Utility Theory
stitution, Risk Aversion, and the Temporal Be- Using Experimental Data," Econometrica, 62,
havior of Consumption and Asset Returns: A pp. 1291-326.
Theoretical Framework," Econometrica, 57, pp. Holt, Charles A. 1986. "Preference Reversals and
937-69. the Independence Axiom," Amer. Econ. Rev.
1990. "'First-order' Risk Aversion and the 76, pp. 508-15.
Equity Premium Puzzle," J. Monetary Econ. 26, Humphrey, Stephen J. 1995. "Regret Aversion or
pp. 387-407. Event-Splitting Effects? More Evidence under
1991. "Substitution, Risk Aversion, and Risk and Uncertainty," J. Risk Uncertainty, 11,
the Temporal Behavior of Consumption and As- pp. 263-74.
set Returns: An Empirical Analysis," J. Polit. - . 1998. "More Mixed Results on Boundary
Econ. 99, pp. 263-86. Effects," Econ. Letters, 61, pp. 79-84.
Evans, Dorla A. 1997. "The Role of Markets in Jullien, Bruno and Bernard Salanie. 1997. "Esti-
Reducing Expected Utility Violations," J. Polit. mating Preferences under Risk: The Case of
Econ. 105, pp. 622-36. Racetrack Bettors," IDEI and GREMAQ Tou-
Fishburn, Peter C. 1978. "On Handa's 'New The- louse U. working paper.
ory of Cardinal Utility' and the Maximization of Kahneman, Daniel and Amos Tversky. 1979.
Expected Return," J. Polit. Econ. 86, pp. 321-24. "Prospect Theory: An Analysis of Decision un-
1982. "Nontransitive Measurable Utility," der Risk," Econometrica, 47:2, pp. 263-91.
J. Math. Psych. 26, pp. 31-67. Karmarkar, Uday S. 1978. "Subjectively Weighted
. 1983. "Transitive Measurable Utility," J. Utility: A Descriptive Extension of the Ex-
Econ. Theory, 31, pp. 293-317. pected Utility Model," Org. Behavior Human
. 1988. Nonlinear Preference and Utility Performance, 21, pp. 61-72.
Theory. Brighton: Wheatsheaf Books. Karni, Edi. 1995. "Non-expected Utility and the
Gigliotti, Garry and Barry Sopher. 1993. "A Test Robustness of the Classical Insurance Para-
of Generalized Expected Utility Theory," The- digm: Discussion," Geneva Papers Risk Insur-
ory and Decision, 35, pp. 75-106. ance Theory, 20, pp. 51-56.
Karni, Edi and Zvi Safra. 1987. "'Preference Re- . 1983. "A Rationale for Preference Rever-
versal' and the Observability of Preferences by sal," Amer. Econ. Rev. 73, pp. 428-32.
Experimental Methods," Econometrica, 55, pp. . 1986. "Disappointment and Dynamic
675-85. Consistency in Choice under Uncertainty," Rev.
. 1989a. "Dynamic Consistency, Revelations Econ. Stud. 53:2, pp. 271-82.
in Auctions and the Structure of Preferences," . 1987. "Some Implications of a More Gen-
Rev. Econ. Stud. 56, pp. 421-34. eral Form of Regret Theory," J. Econ. Theory,
. 1989b. "Ascending Bid Auctions with Be- 41:2, pp. 270-87.
haviorally Consistent Bidders," Annals Opera- . 1995. "Incorporating a Stochastic Element
tional Res. pp. 435-46. into Decision Theories," European Econ. Rev.
. 1990. "Behaviorally Consistent Optimal 39, pp. 641-48.
Stopping Rules," J. Econ. Theory, 51, pp. 391- -____. 1998. "Testing Different Stochastic Speci-
402. fications of Risky Choice," Economica, 65, pp.
Karni, Edi and David Schmeidler. 1986. "Self- 581-98.
Preservation as a Foundation of Rational Behav- Loomes, Graham and Caron Taylor. 1992. "Non-
ior Under Risk," J. Econ. Behavior Org. 7, pp. transitive Preferences over Gains and Losses,"
71-81. Econ. J. 102, pp. 357-65.
Knetsch, Jack L. 1989. "The Endowment Effect Luce, R. Duncan and Peter C. Fishburn. 1991.
and Evidence of Non-reversible Indifference "Rank and Sign-dependent Linear Utility Mod-
Curves," Amer. Econ. Rev. 79, pp. 1277-84. els for Finite First-Order Gambles" J. Risk Un-
Knetsch, Jack L. and J. A. Sinden. 1984. "Willing- certainty, 4, pp. 29-59.
ness to Pay and Compensation Demanded: Ex- MacCrimmon, Kenneth and Stig Larsson. 1979.
perimental Evidence of an Unexpected Dispar- "Utility Theory: Axioms versus Paradoxes," in
ity in Measures of Value," Quart. J. Econ. 99, Expected Utility Hypotheses and the Allais
pp. 507-21. Paradox. M. Allais and 0. Hagen, eds. Dor-
Lattimore, Pamela K.; Joanna R. Baker, and Ann drecht: Reidel
D. Witte. 1992. "The Influence of Probability Machina, Mark J. 1982. "'Expected Utility' Theory
on Risky Choice: A Parametric Examination," J. without the Independence Axiom," Econo-
Econ. Behavior Org. 17,pp. 377-400. metrica, 50, pp. 277-323.
Lavoie, Marc. 1992. Foundations of Post-Keynes- . 1983. "The Economic Theory of Individ-
ian Economic Analysis. Aldershot, UK: Edward ual Behavior toward Risk: Theory, Evidence,
Elgar. and New Directions," Technical Report 433,
Leland, J.W. 1998. "Similarity Judgements in Dept. Economics, Stanford U.
Choice under Uncertainty: A Reinterpretation . 1987. "Choice under Uncertainty: Prob-
of the Predictions of Regret Theory," Manage. lems Solved and Unsolved," J. Econ. Perspec-
Sci. 44, pp. 659-72. tives, 1, pp. 121-54.
Lichtenstein, Sarah and Paul Slovic. 1971. "Rever- . 1989a. "Comparative Statics and Non-Ex-
sals of Preference between Bids and Choices in pected Utility Preferences," J. Econ. Theory,
Gambling Decisions," J. Experimental Psych. 47, pp. 393-405.
89, pp. 46-55. . 1989b. "Dynamic Consistency and Non-
Lindman, Harold R. 1971. "Inconsistent Prefer- Expected Utility Models of Choice under Un-
ences among Gambles," J. Experimental Psych. certainty," J. Econ. Lit. 27, pp. 1622-68.
89, pp. 390-97. . 1994. "Review of 'Generalized Expected
Loewenstein, George F. and Daniel Adler. 1995. Utility Theory: The Rank-Dependent Model"' J.
"A Bias in the Prediction of Tastes," Econ. J. Econ. Lit. 32:3, pp. 1237-38.
105, pp. 929-37. . 1995. "Non-Expected Utility and the Ro-
Loomes, Graham; Peter Moffat, and Robert Sug- bustness of the Classical Insurance Paradigm,"
den. 1998. "A Microeconometric Test of Alter- Geneva Papers Risk Insurance Theory, 20, pp.
native Stochastic Theories of Risk Choice," 9-50.
ERC discussion paper 9806, U. East Anglia. Marschak, Jacob. 1950. "Rational Behavior, Uncer-
Loomes, Graham; Chris Starmer; and Robert Sug- tain Prospects, and Measurable Utility," Econo-
den. 1989. "Preference Reversal: Information- metrica, 18, pp. 111-41.
Processing Effect or Rational Non-transitive McClennen, Edward F. 1990. Rationality and Dy-
Choice?" Econ. J. 99:395, pp. 140-51. namic Choice: Foundational Explorations. Cam-
1991. "Observing VioLations of Transitivity bridge: Cambridge U. Press.
by Experimental Methods," Econometrica, 59:2, Mehra, Rajnish and Edward C. Prescott. 1985.
pp. 425-39. "The Equity Premium Puzzle," J. Monetary
. 1992. "Are Preferences Monotonic? Test- Econ. 15, pp. 145-61.
ing Some Predictions of Regret Theory," Eco- McNeil, B. J.; S. G. Pauker, H. C. Sox, and Amos
nomica, 59, pp. 17-33. Tversky. 1982. "On the Elicitation of Prefer-
Loomes, Graham and Robert Sugden. 1982. "Re- ences for Alternative Therapies," New England
gret Theory: An Alternative Theory of Rational J. Medicine, 306, pp. 1259-62.
Choice under Uncertainty," Econ. J. 92, pp. Moskowitz, Herbert. 1974. "Effects of Problem Pre-
805-24. sentation and Feedback on Rational Behavior in
Allais and Morlat-Type Problems," Decision . 1994. "Prospect Theory and the Econom-
Sciences, 5, pp. 225-42. ics of Litigation," PhD thesis, Stanford U.
Mossin, J. 1968. "Aspects of Rational Insurance . 1996. "Gains, Losses, and the Psychology
Purchasing," J. Polit. Econ. 76, pp. 553-68. of Litigation," South. Cal. Law Rev., 70, pp.
Munro, Alistair. 1998. "Loss Aversion and Welfare 113-85.
Economics," Economics Research Centre work- Richard, Rene. 1994. "Regret Is What You Get:
ing paper, U. East Anglia. The Impact of Anticipated Feelings and Emo-
Munro, Alistair and Robert Sugden. 1998. "A The- tions on Human Behaviour," PhD thesis, Psych.
ory of General Equilibrium with Reference De- Faculty, U. Amsterdam.
pendent Preferences," mimeo, U. East Anglia. Rubinstein, Ariel. 1988. "Similarity and Decision-
Myagkov, Mikhail and Charles R. Plot. 1997. "Ex- Making Under Risk," J. Econ. Theory, 46, pp.
change Economies and Loss Exposure: Experi- 145-53.
ments Exploring Prospect Theory and Competi- Samuelson, William F. and Richard Zeckhauser.
tive Equilibria in Market Environments," Amer. 1988. "Status Quo Bias in, Decision Making," J.
Econ. Rev. 87, pp. 801-28. Risk Uncertainty, 1, pp. 7-59.
Neilson, William S. 1992. "A Mixed Fan Hypothe- Savage, Leonard J. 1954. The Foundations of Sta-
sis and its Implications for Behavior toward tistics. NY: Wiley.
Risk," J. Econ. Behavior Org. 19, pp. 197-211. Schoemaker, Paul. 1982. "The Expected Utility
-____. 1979. "Some Mixed Results on Boundary Model: Its Variants, Purposes, Evidence and
Effects," Econ. Letters, 39, pp. 275-78. Limitations," J. Econ. Lit. 20:2, pp. 529-63.
Nelson, Richard R. 1995. "Recent Evolutionary Segal, Uzi. 1988a. "Does the Preference Reversal
Theorizing about Economic Change," J. Econ. Phenomenon Necessarily Contradict the Inde-
Lit. 33:1, pp. 48-90. pendence Axiom?" Amer. Econ. Rev. 78, pp.
Neumann, John von and Oskar Morgenstern. 233-36.
1947. The Theory of Games and Economic Be- . 1988b. "Probabilistic Insurance and An-
haviour. 2nd ed. Princeton: Princeton U. Press. ticipated Utility," J. Risk Insurance, 55, pp.
Payne, John W.; James R. Bettman, and Eric J. 287-97.
Johnson. 1993. The Adaptive Decision Maker. 1989. "Anticipated Utility: A Measure
Cambridge U. Press. Representation Approach," Annals Operational
Pidgeon, Nick. C. Hood, D. Jones, B. Turner and Res. 19, pp. 359-73.
R. Gibson. 1992. "Risk Perception," in Risk: . 1990. "Two-Stage Lotteries without the
Analysis, Perceptions and Management. Report Independence Axiom," Econometrica, 58, pp.
of a Royal Society Study Group. London: The 349-77.
Royal Society. . 1993. "The Measure Representation: A
Plott, Charles, R. 1996. "Rational Individual Be- Correction," J. Risk Uncertainty, 6, pp. 99-107.
havior in Markets and Social Choice Processes: . 1997. "Dynamic Consistency and Refer-
The Discovered Preference Hypothesis," in Ra- ence Points," J. Econ. Theory, 72, pp. 208-19.
tional Foundations of Economic Behavior. K. Segal, Uzi and Avia Spivak. 1990. "First Order ver-
Arrow, E. Colombatto, M. Perleman, and C. sus Second Order Risk Aversion," J. Econ. The-
Schmidt, eds. London: Macmillan and NY: St. ory, 51, pp. 111-125.
Martin's, pp. 225-50. Sega2, Uzi; Avia Spivak, and J. Ziera. 1988. "Pre-
Prelec, Drazen. 1998. "The Probability Weighting cautionary Savings and Risk Aversion: An An-
Function," Econometrica, 66, pp. 497-527. ticipated Utility Approach," Econ. Letters, 27,
Preston, Malcolm G. and Phillip Baratta. 1948. pp. 223-27.
"An Experimental Study of the Auction-Value Simon, Herbert, A. 1955. "A Behavioral Model of
of an Uncertain Outcome," Amer. J. Psych. 61, Rational Choice," Quart. J. Econ. 69, pp. 99-
pp. 183-93. 118.
Quiggin, John. 1982. "A Theory of Anticipated Slovic, Paul. 1995. "The Construction of Prefer-
Utility," J. Econ. Behavior Org. 3:4, pp. 323-43. ences," Amer. Psychologist, 50, pp. 364-71.
. 1991a. "Comparative Statics for Rank-De- Slovic, Paul and Amos Tversky. 1974. "Who Ac-
pendent Expected Utility Theory," J. Risk Un- cepts Savage's Axiom?" Behavioral Sci. 19, pp.
certainty, 4, pp. 339-50. 368-73.
. 1991b. "On the Optimal Design of Lotter- Smith, Richard D. 1996. "Is Regret Theory an Al-
ies," Economica, 58:229, pp. 1-16 ternative Basis for Estimating the Value of
_ . 1993. Generalized Expected Utility The- Health Care Interventions?" Health Policy, 37,
ory. Dordrecht: Kluwer. pp. 105-15.
. 1994. "Regret Theory with General Choice Smith, Vernon. 1976. "Experimental Economics:
Sets," J. Risk Uncertainty, 8:2, pp. 153-65. Induced Value Theory," Amer. Econ. Rev. Pro-
Rabin, Matthew. Forthcoming. "Risk Aversion and ceedings, pp. 274-79.
Expected-Utility Theory: A Calibration Theo- Starmer, Chris. 1992. "Testing New Theories of
rem," Econometrica. Choice under Uncertainty Using the Common
Rachlinski, Jeffrey J. 1990. "Prospect Theory and Consequence Effect," Rev. Econ. Stud. 59, pp.
Civil Negotiation," working pa er 17, Center on 813-30.
Conflict and Negotiation, Stanord U. 1. 999a. "Cycling with Rules of Thumb: An
Experimental Test for a New Form of Non- mulative Representation of Uncertainty," J.

Transitive Behaviour," Theory and Decision, 46, Risk Uncertainty 5:4, pp. 297-323.
pp. 141-58. Tversky, Amos and Derek J. Koehler. 1994. "Sup-
1999b. "Experiments in Economics: port Theory: A Nonextensional Representation
Should We Trust the Dismal Scientists in White of Subjective Probability," Psych. Rev. 101, pp.
Coats?," J. Econ. Method. 6, pp. 1-30. 547-67.
Starmer, Chris and Robert Sugden. 1989. "Viola- Tversky, Amos; Paul Slovic, and Daniel Kahne-
tions of the Independence Axiom in Common man. 1990. "The Causes of Preference Rever-
Ratio Problems: An Experimental Test of Some sal," Amer. Econ. Rev. 80:1, pp. 204-17.
Competing Hypotheses," Annals Operational Tversky, Amos and Richard H. Thaler. 1990.
Res. 19, pp. 79-102. "Preference Reversals," J. Econ. Perspectives, 4,
. 1991. "Does the Random-Lottery Incen- pp. 201-11.
tive System Elicit True Preferences?" Amer. Tversky, Amos and Peter P. Wakker. 1995. "Risk
Econ. Rev. 81, pp. 971-79. Attitudes and Decision Weights," Econo-
. 1993. "Testing for Juxtaposition and Event metrica, 63, pp. 1255-80.
Splitting Effects," J. Risk Uncertainty, 6, pp. Viscusi, W. Kip. 1989. "Prospective Reference
235-54. Theory: Toward an Explanation of the Para-
. 1998. "Testing Alternative Explanations of doxes," J. Risk Uncertainty, 2, pp. 235-64.
Cyclical Choices," Economica, 65:259, pp. 347- Wakker, Peter P. 1994. "Separating Marginal Util-
61. ity and Probabilistic Risk Aversion," Theory and
Stevenson, M. K.; J. R. Busmeyer, and J. C. Nay- Decision, 36, pp. 1-44.
lor. 1991. "Judgement and Decision-Making 1999. "Justifying Bayesianism by Dynamic
Theory," in New Handbook of Industrial Or- Decision Principles," plenary paper presented
ganizational Psychology. M. Dunnette and at FUR IX, Marrakesh.
L. N. Hough, eds. Palo Alto, CA: Consulting Wakker, Peter P. and Daniel Deneffe. 1996. "Elic-
Psychology Press. iting von Neumann-Morgenstern Utilities
Sugden, Robert. 1986. "New Developments in the When Probabilities Are Distorted or Un-
Theory of Choice Under Uncertainty," Bul. known," Manage. Sci. 42, pp. 1131-50.
Econ. Res. 38, pp. 1-24. Wakker, Peter P.; Ido Erev, and Elke U. Weber.
1991. "Rational Choice: A Survey of Contri- 1994. "Comonotonic Independence: The Criti-
butions from Economics and Philosophy," Econ. cal Test between Classical and Rank-Depen-
J. 101, pp. 751-85. dent Utility Theories," J. Risk Uncertainty, 9,
. 1993. "An Axiomatic Foundation for Re- pp. 195-230.
gret Theory," J. Econ. Theory, 60, pp. 159-80. aker, Peter P.; Richard H. Thaler, and Amos
Tammi, Timo. 1997. "Essays on the Rationality of Tversky. 1997. "Probabilistic Insurance," J. Risk
Experimentation in Economics: The Case of Uncertainty, 15, pp. 7-28.
Preference Reversal," PhD thesis, U. Joensuu, Weber, Martin; Franz Eisenftihr, and Detlof von
Finland. Winterfeldt. 1988. "The Effects of Splitting At-
Thaler, Richard H. 1980. "Toward a Positive The- tributes on Weights in Multi Attribute Utility
ory of Consumer Choice," J. Econ. Behavior Measurement," Manage. Sci. 34, p p. 431-45.
Org. 1, pp. 39-60. Weinstein, Milton, C. 1986. "Risky Choices in
Thaler, Richard H. and William T. Ziemba. 1988. Medical Decision Making: A Survey," Geneva
"Parimutuel Betting Markets: Racetracks and Papers Risk Insurance, 11, pp. 197-216.
Lotteries," J. Econ. Perspectives, 2, pp. 161-74. Weinstein, Milton C. and Robert J. Quinn. 1983.
Timmerman, Allan. 1994. "Can Agents Learn to "Psychological Considerations in Valuing Health
Form Rational Expectations? Some Results on Risk Reductions," Nat. Resources J. 23, pp.
Convergence and Stability of Learning in the 659-73.
UK Stock Market," Econ. J. 104, pp. 777-97. Wu, George. 1994. "An Empirical Test of Ordinal
Tversky, Amos. 1969. "Intransitivity of Prefer- Independence," J. Risk Uncertainty, 9, pp. 39-
ences," Psych. Rev. 76, pp. 31-48. 60.
Tversky, Amos and Danief. Kahneman. 1981. "The . 1999. "Anxiety and Decision Making with
Framing of Decisions and the Psychology of Delayed Resolution of Uncertainty," Theory
Choice," Science, 211, pp. 453-58. and Decision, 46, pp. 159-98.
1986. "Rational Choice and the Framing Wu, George and Richard Gonzalez. 1996. "Curva-
of Decisions," J. Bus. 59:4, pt. 2, pp. s251-78. ture of the Probability Weighting Function,"
. 1991. "Loss Aversion in Riskless Choice: A Management Sci. 42, pp. 1676-90.
Reference-Dependent Model," Quart. J. Econ. Yaari, Menahem E. 1987. "The Dual Theory of
106:4, pp. 1039-61. Choice Under Risk," Econometrica, 55, pp. 95-
1992. "Advances in Prospect Theory: Cu- 115.

Starmer-Developments in Non-Expected Utility Theory

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Starmer-Developments in Non-Expected Utility Theory

Uploaded by

Copyright:

Available Formats

American Economic Association

Developments in Non-Expected Utility Theory: The Hunt for a Descriptive Theory of

Developments in Non-Expected Utility

1. Introduction stimulated developments in non-EU is

the value of a gamble is the expectation ternative models of choice, it will be

for all q, r, s: if q > r and r > s, then economics, so economists naturally

towards consequences (e.g. a diminish- phenomena, first discovered by Maurice

4 Those interested in more thorough reviews are

failures of description invariance. Here that choices can, in general, be repre-

Indifference curve for the

Figure Ic. Different Degrees of Risk Aversion in EUT sI S2

Figure 2. Common Ratio Prospects

strong restriction which leaves only one

effect and, as in the common conse-

corner appear more risk prone than q

should be expected given preferences

(though they may be). It is very easy to

have been based on psychologically

where u(xi) is interpreted as a measure

extreme outcomes to receive particularly

lends support to the hypothesis of an two prospects q, r have a "common tail"

outcome domain entails diminishing

is no relation of dominance between s placed by zeros. Tversky and Kahne-

skew-symmetric bilinear utility or SSB).

5.1.1 Choosing among Conventional up along the left-hand edge of prob-

examples. It is well-known that EUT agents will be approximately risk neu-

say, expected utility preferences might market is introduced; it occurs because

Experimental Test for a New Form of Non- mulative Representation of Uncertainty," J.

You might also like