Reference-Dependent Risk Attitudes

By Botond Kőszegi and Matthew Rabin*

We use Kőszegi and Rabin’s (2006) model of reference-dependent utility, and an

extension of it that applies to decisions with delayed consequences, to study pref-
erences over monetary risk. Because our theory equates the reference point with
recent probabilistic beliefs about outcomes, it predicts specific ways in which the
environment influences attitudes toward modest-scale risk. It replicates “classical”
prospect theory—including the prediction of distaste for insuring losses—when
exposure to risk is a surprise, but implies first-order risk aversion when a risk,
and the possibility of insuring it, are anticipated. A prior expectation to take on
risk decreases aversion to both the anticipated and additional risk. For large-scale
risk, the model allows for standard “consumption utility” to dominate reference-
dependent “gain-loss utility,” generating nearly identical risk aversion across situ-
ations. (JEL D81)

Daniel Kahneman and Amos Tversky’s (1979) model, but they also generate mutually inconsis-
prospect theory, and the literature building from tent predictions that to our knowledge have not
it, provide theories of risk attitudes based on a been formally reconciled. Under a status quo
few regularities. Most importantly, evaluation specification, loss aversion predicts the substan-
of an outcome is influenced by how it compares tial dislike of modest-scale risks involving both
to a reference point, with people exhibiting both gains and losses that has been widely observed,
a significantly greater aversion to losses than and diminishing sensitivity predicts the risk lov-
appreciation of gains, and a diminishing sen- ingness in high-probability losses found by many
sitivity to changes in an outcome as it moves researchers in the laboratory.1 Under specifica-
farther from the reference point. In addition, tions based on the lagged status quo—such as
people weight the probability of a prospect non- in Richard H. Thaler and Eric J. Johnson (1990)
linearly, overweighting small probabilities and and Francisco Gomes (2005)—diminishing sen-
underweighting high probabilities. sitivity predicts the willingness to take unfavor-
The implications of prospect theory have able risks to regain the previous status quo. This
been studied with several different specifica- “disposition effect,” which has been observed
tions of the reference point, including the status by Terrance Odean (1998) for small investors
quo, lagged status quo, and the mean of the cho- and by David Genesove and Christopher Mayer
sen lottery. These various approaches explain (2001) for homeowners, is inconsistent with the
many risk attitudes that are inconsistent with the substantial risk aversion predicted by a status
classical diminishing-marginal-utility-of-wealth quo model for gambles involving both gains

* Kőszegi: Department of Economics, University of

1
California, Berkeley, 549 Evans Hall, 3880, Berkeley, CA As has been pointed out by many researchers and formal-
94720 (e-mail: [email protected]); Rabin: Depart- ized by Rabin (2000), nontrivial modest-scale risk aversion
ment of Economics, University of California, Berkeley, is calibrationally inconsistent with the classical diminish-
549 Evans Hall, 3880, Berkeley, CA 94720 (e-mail: rabin@ ing-marginal-utility-of-wealth model. For instance, a
econ.berkeley.edu). We are grateful to Paige Skiba and person with $1 million in lifetime wealth who has CRRA
Justin Sydnor for research assistance, and to Erik Eyster, utility and who rejects a “fifty-fifty lose $500 or gain $550”
Vince Crawford, and three anonymous referees for detailed gamble would also turn down an equal-probability bet of
suggestions. We also thank seminar participants at Harvard losing $4,000 or gaining $100,000,000,000,000. While
University, London School of Economics, Stockholm nobody would turn down the large-scale bet, most people
School of Economics, and UC Berkeley for comments. would turn down the smaller gamble. Nicholas Barberis,
Rabin thanks the National Science Foundation for financial Ming Huang, and Thaler (2006), for example, find that the
support under grant SES-0518758, and Kőszegi thanks the majority of MBA students, financial advisors, and even
Central European University for its hospitality while some very rich investors (with median financial wealth over $10
of this work was completed. million) reject the $500–$550 bet.
1047
1048 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2007

and losses relative to the current status quo. And m 1 w 2 , corresponds to the classical notion
under specifications based on the chosen lottery’s of outcome-based utility. Gain-loss utility,
certainty equivalent—such as in the disappoint- m 1 m 1 w 2 2 m 1 r 2 2 , depends on the difference
ment-aversion models of David E. Bell (1985), between the consumption utility of the outcome
Graham Loomes and Robert Sugden (1986), and of the reference level, with the shape of m
and Faruk Gul (1991)—loss aversion implies corresponding to the loss aversion and dimin-
substantial aversion to any risk. This strong risk ishing sensitivity of prospect theory. Some of
aversion, which is apparent in consumers’ choice our results are established by assuming only
of low insurance deductibles and purchase of what is commonly taken to be the stronger of
extremely expensive extended warranties and these two forces, loss aversion.
automobile service contracts, is inconsistent with We assume that the reference point relative
the risk lovingness in losses found in the lab and to which a person evaluates an outcome is her
in the case of the disposition effect. recent beliefs about that outcome. An employee
This paper uses the model from Kőszegi and who had expected a $50,000 salary will assess a
Rabin (2006) and an extension to study monetary salary of $40,000 as a loss, and a taxpayer who
risk, unifying the seemingly different risk atti- had expected to pay $30,000 in taxes will treat
tudes noted above as manifestations of the same a $20,000 tax bill as a gain. Because a person
preferences in different domains, and making may be uncertain about outcomes, our theory
novel predictions about behavior in situations allows for the reference point to be a distribu-
not studied in the related literature. Our model tion G 1 # 2 , with a wealth outcome w then evalu-
(a) combines the reference-dependent “gain-loss ated with “mixed feelings” as the average of the
utility” with standard “consumption utility”; different assessments u 1 w 0 r 2 generated by the r
(b) bases the reference point to which outcomes possible under G 1 # 2 . For simplicity, we abstract
are compared on endogenously determined from nonlinear decision weights: given a (sto-
lagged beliefs; and, to incorporate probabilistic chastic or deterministic) reference point, a sto-
beliefs, (c) allows for stochastic reference points. chastic wealth outcome is evaluated according
Because of feature (a), our theory is consistent to its expected reference-dependent utility.
with aversion to all large-scale risk as predicted Our model of how utility depends on beliefs
by classical expected-utility theory. Because of could be combined with any theory of how these
feature (b), it predicts both risk lovingness in beliefs are formed. As an imperfect but at the
response to surprise modest-scale losses, and— same time disciplined and largely realistic first
since anticipated premium payments do not pass, we assume that a person correctly pre-
generate sensations of loss while bad outcomes dicts her probabilistic environment and her own
in uncertain situations do—first-order risk aver- behavior in that environment, so that her beliefs
sion when a risk and the possibility to insure it fully reflect the true probability distribution of
are expected. Hence, our theory matches both outcomes. We begin in Section II by considering
status quo prospect theory and disappointment “surprise” (low-probability) decisions, modelled
aversion in domains where these models have in extreme form as situations where expectations
been applied, and more generally provides com- are given exogenously to the actual choice set.
parative-statics predictions on the extent of risk To illustrate implications for modest-scale risk,
taking as a function of the environment. Because where consumption utility is approximately lin-
of features (b) and (c), our theory predicts that ear, consider a person’s decision on whether to
the prior expectation of risk, even if it can now pay $55 to insure a 50 percent chance of hav-
be avoided, decreases risk aversion. Unlike all ing to pay $100. If she had expected to retain
the theories above, therefore, it predicts less the status quo of $0, our model makes the same
risk aversion when deciding whether to remove prediction as prospect theory: because of dimin-
expected risk than when deciding whether to ishing sensitivity, she does not wish to insure the
take on that risk. risk. If she had expected to pay $55 for insur-
For a wealth level w and reference wealth ance, however, paying that amount generates no
level r, Section I specifies a person’s utility gain or loss, while taking the gamble exposes
as u 1w|r2 ; m 1w 2 1 m 1m 1w 2 2 m 1r2 2 . The her to a fifty-fifty chance of losing $45 or gain-
reference-independent “consumption utility,” ing $55. With a conventional estimate of two-
VOL. 97 NO. 4 KŐSZEGI AND RABIN: REFERENCE-DEPENDENT RISK ATTITUDES 1049

to-one loss aversion, she strongly dislikes this When a person makes a committed decision
gamble and buys the insurance. Yet if a person long before outcomes occur, she affects the ref-
had been expecting risk to start with, paying $0 erence point by her choice. For these situations,
instead of $100 can decrease expected losses, we introduce in Section IV the idea of a “choice-
and paying $100 might just decrease expected acclimating personal equilibrium” (CPE), defined
gains, so the gamble is less aversive. When the as a decision that maximizes expected utility
ex ante expected risk is the gamble itself, this given that it determines both the reference lottery
decreased risk aversion can be interpreted as and the outcome lottery. Except that we specify
an endowment effect for risk. When the ex ante the reference point as a lottery’s full distribution
expected uncertainty is very large, $100 cannot rather than its certainty equivalent, this concept
much change the extent to which money is eval- is similar to the disappointment-aversion models
uated as a loss rather than a gain, so the person of Bell (1985), Loomes and Sugden (1986), and
is close to risk neutral. Gul (1991). Like PPE, CPE predicts that the deci-
In Sections III and IV, we study attitudes to sion maker strongly prefers to insure expected
anticipated risks. We identify two implications of risks. But there is also an important difference.
our model: a person is more risk averse when she In some situations, a person would be better off
anticipates a risk and the possibility to insure it with the reduced uncertainty of expecting to buy
than when she does not—always displaying first- expensive insurance than not expecting to do so,
order risk aversion—and among such decisions but always choosing to avoid the expense of insur-
regarding anticipated risk, she is more risk averse ance at the last minute if not committed to buy-
when she can commit to insure ahead of time. ing it. Hence, the distaste for the risk manifests
When a decision is made shortly before the itself in behavior when the insurance decision is
outcomes resulting from it occur, at that moment made up front, as in CPE, but not when the deci-
the reference point is fixed by past expectations, sion is made later, as in PPE. In such situations,
so that the decision maker maximizes expected environments where CPE is appropriate generate
utility taking the reference point as given. Being greater risk aversion and higher expected utility
fully rational, therefore, she can expect behavior than environments where PPE applies.
only if she is willing to follow it through, given The sensitivity of behavior to the economic
a reference point determined by the expectation environment described above applies only to
to do so. Formalizing this idea, in Section III modest-scale choices, where risk attitudes are
we import from Kőszegi and Rabin (2006) the necessarily dominated by the gain-loss compo-
concept of an “unacclimating personal equi- nent of preferences. In Section V, we investigate
librium” (UPE), defined as behavior where the attitudes toward large-scale risk, where con-
stochastic outcome generated by utility-maxi- sumption utility cannot be assumed to be linear.
mizing choices conditional on expectations We show that under reasonable conditions, the
coincides with expectations. Positing that a reference point has only a minor impact on how
person can make any plans she knows she will a person evaluates very large gambles. A person
follow through, our analysis assumes that she is therefore prone to exhibit risk aversion reflect-
chooses her favorite UPE, the “preferred per- ing diminishing marginal utility of wealth inde-
sonal equilibrium” (PPE). pendently of the environment.
Applying PPE, we predict a very strong taste Beyond helping to explain in a unified frame-
for planning and executing the purchase of small- work seemingly contradictory behavior, we
scale insurance. The reason is a formalization hope the endogenous specification of the refer-
and elaboration of some previous researchers’ ence point helps make our model readily por-
(e.g., Kahneman and Tversky 1984) psychologi- table to many settings. To facilitate applications,
cal intuition that money given up in regular bud- in Appendix A we present an array of risk-
geted purchases is not a loss. In our model, a bad characterization concepts and results. In Section
outcome of an uncertain lottery is evaluated as VI, we conclude the paper by discussing some
a loss, but a fully expected premium payment is of the shortcomings of our model, emphasizing
not evaluated as a loss. Loss aversion therefore especially its failure to capture important ways
generates first-order risk aversion toward all that reference-dependent risk attitudes reflect
insurable risks. failures of full rationality.
1050 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2007

I. Reference-Dependent Utility reference lottery. For example, if the reference

lottery is a gamble between $0 and $100, an out-
In this section we present the one-dimen- come of $50 evokes a mixture of two feelings, a
sional version of the utility function in Kőszegi gain relative to $0 and a loss relative to $100.3
and Rabin (2006). As is standard for models of When w is drawn according to the probability
risky choice outside of full-fledged life-cycle measure F, utility is given by
consumption models in macroeconomics, our
theory takes as a primitive the choice set of
(2) U 1FZG2 5 u 1w|r) dG 1r2 dF 1w2.
gambles a person focuses on, in isolation from 22
other risks and choices she faces.
For a riskless wealth outcome w [ R and For simplicity and contrary to Kahneman and
riskless reference level of wealth r [ R, utility is Tversky (1979) and its extensions, we assume
given by u 1w|r2 ; m 1w2 1 m 1m 1w2 2 m 1r2 2.2 that preferences are linear in probabilities. This
The term m 1 w 2 is intrinsic “consumption util- means that our model will get some predic-
ity” usually assumed relevant in economics, tions—e.g. regarding insurance of low-prob-
and the term m 1 m 1 w 2 2 m 1 r 2 2 is the reference- ability losses—wrong.
dependent gain-loss utility. Our model assumes Our utility function is closely related to that
that how a person feels about gaining or los- of Sugden (2003). In his model, outcome lotter-
ing relative to a reference point depends on ies are compared to reference lotteries state by
the changes in consumption utility associated state, capturing a form of state-contingent dis-
with such gains or losses. This separation and appointment missing from our theory.4 Another
interdependence of economic and psychologi- alternative to our formulation, pursued by Bell
cal payoffs is analogous to assumptions made (1985), Loomes and Sugden (1986), Gul (1991),
previously by Bell (1985), Loomes and Sugden and Jonathan Shalev (2000), is to collapse the
(1986), and Veronika Köbberling and Peter P. reference lottery into some type of certainty
Wakker (2005). equivalent.5 With such a specification, two
To accommodate our assumption below that
the reference point is beliefs about outcomes, we
allow for the reference point to be a probability 3
More than saying a person separately compares an
measure G over R: outcome to all components of the reference lottery, our
formulation of m[ below implies that losses relative to a
stochastic reference point count more than gains, so that the
(1) U 1wZG2 5 u 1wZr2 dG 1r2.
2 $50 above yields negative gain-loss utility. An alternative
specification is one where the relief of avoiding the $0 out-
This formulation captures the notion that the come outweighs the disappointment of not getting the $100
evaluation of a wealth outcome is based on com- outcome. This alternative seems difficult to reconcile with
loss aversion relative to riskless reference points. It would
paring it to all possibilities in the support of the also seem to imply that people will seek to endow them-
selves with risks because the chance for a pleasant relief
from getting better outcomes outweighs the potential disap-
2
Like many models, this paper assumes preferences are pointment from getting bad outcomes.
4
over monetary wealth rather than consumption. Strictly While this state-contingency seems in some cases to be
speaking, this means our model corresponds to a single- more realistic than our approach, and we do not know the
period setting. While we have not verified how our results extent to which the two models can be reconciled within
extend to a multiperiod consumption model, the shortcut a broader framework, our prediction of a state-indepen-
of using wealth may be appropriate even in that case, both dent disappointment when receiving worse-than-expected
because people often experience sensations of gain and outcomes seems pervasively realistic, and is missing from
loss directly from wealth changes, and because wealth is Sugden’s approach. In addition, Sugden’s model (unlike
a summary statistic for consumption and hence may gen- ours) makes the implausible prediction that a person
erate similar gain-loss sensations. As a person’s wealth becomes very risk loving when given the option to replace
increases, for example, her anticipation of increased con- the reference lottery with a riskless amount.
5
sumption throughout the future is likely to generate a sense There is some suggestive evidence of mixed feelings
of gain similar to that presumed in our model. A developing when there are multiple counterfactuals relative to which
body of research on dynamic models of reference-depen- outcomes can be evaluated. For instance, Jeff T. Larsen et al.
dent utility, such as Rebecca Stone (2005) and Kőszegi and (2004) find that when a subject receives $5 from a lottery
Rabin (2007), may help extend and explore the robustness that could have paid $5 or $9, she has both positive and
of the results we develop in this paper. negative emotions—presumably from winning $5 and not
VOL. 97 NO. 4 KŐSZEGI AND RABIN: REFERENCE-DEPENDENT RISK ATTITUDES 1051

reference lotteries that have the same certainty assume that a person’s reference point is the
equivalent generate the same risk preferences. rational expectations about the relevant out-
This is inconsistent with our theory’s prediction come she held between the time she first focused
that a person is more inclined to accept a risk if on the outcome and shortly before it occurs. For
she had been expecting risk, a prediction that example, if an employee had been expecting a
seems broadly correct based on the little avail- salary of $100,000, she would assess a salary
able evidence. But as we discuss below, the main of $90,000, not as a large gain relative to her
difference between all these models and ours is status quo wealth, but as a loss relative to her
in the specification of the reference point. expectations of wealth.6 As we explain in detail
We assume m satisfies the following properties: in Kőszegi and Rabin (2006), our primary moti-
vation for equating the reference point with
A0. m 1 x 2 is continuous for all x, twice dif- expectations is that this assumption helps unify
ferentiable for x 2 0, and m 1 0 2 5 0. and reconcile existing intuitions and discus-
sions. We assume rational expectations both
A1. m 1 x 2 is strictly increasing. to maintain modelling discipline (much like
in other rational-expectations theories), and
A2. If y . x $ 0, then m 1y 2 1 m 12y 2 , m 1x 2 because we feel in most situations people have
1 m 12x 2. some ability to predict their own environment
and behavior. Unfortunately, relatively little evi-
A3. ms 1 x 2 # 0 for x . 0 and ms 1 x 2 $ 0 for dence on the determinants of reference points
x , 0. currently exists. Some existing evidence does,
however, provide empirical support for expec-
A4. m9– 102/m91 102 K l . 1, where m91 102 K tations as a component of the reference point.
lim xS0 m91|x|2 and m9– 102 K lim xS0 m912|x|2. In analyzing play in the huge-stakes game show
“Deal or No Deal,” for example, Thierry Post et
Properties A0–A4, first stated by David al. (forthcoming) find evidence that past expec-
Bowman, Deborah Minehart, and Rabin (1999), tations affect behavior. In the game, a contestant
correspond to Kahneman and Tversky’s (1979) “owns” a suitcase with a randomly determined
explicit or implicit assumptions about their prize. Gradually, the contestant learns informa-
“value function” defined on w 2 r. Loss aver- tion about the prize in her bag (by opening other
sion is captured by A2 for large stakes and A4 bags and learning what is not in her bag). At
for small stakes, and diminishing sensitivity is each stage, a “bank” offers a riskless amount of
captured by A3. While the inequalities in A3 are money to replace the amount in the bag. A con-
most realistically considered strict, to charac- testant’s acceptance or rejection of the offer is
terize the implications of loss aversion without an indication of her risk aversion. A key finding
diminishing sensitivity as a force on behavior, is that contestants reject better offers when they
we define a subcase of A3: have received bad news in the last few rounds,
suggesting that they are less risk averse in these
A3r. For all x 2 0, ms 1 x 2 5 0. contingencies.7

When we apply A39 below, we will param-

eterize m as m91 102 5 h and m9– 102 5 lh . h, so 6
Our theory posits that preferences depend on lagged
that h can be interpreted as the weight attached expectations, rather than on expectations contemporaneous
to gain-loss utility. with the time of consumption. This does not assume that
beliefs are slow to adjust to new information or that people
To determine behavior, the utility function are unaware of the choices they have just made—but that
introduced above needs to be combined with preferences do not instantaneously change when beliefs do.
a theory of reference-point determination. As When somebody finds out five minutes ahead of time that
a disciplined and largely realistic first pass, we she will for sure not receive a long-expected $100, she pre-
sumably immediately adjusts her expectations to the new
situation, but five minutes later she will still assess not get-
ting the money as a loss.
7
winning $9, respectively. Losing $5 when it is also possible For further examples of evidence of expectations-
to lose $9 evokes similar mixed feelings. based counterfactuals affecting reactions to outcomes,
1052 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2007

Many researchers have noted over the years II. Risk Attitudes in Surprise Situations
that the reference point may to some extent be
influenced by expectations. But most previous In the next three sections, we investigate the
formal models either equate the reference point decision maker’s attitudes toward modest-scale
with the status quo, or leave it unspecified, and risk, such as $100 or $1,000, where consump-
none explicitly equates it to recent beliefs about tion utility can be taken to be approximately
outcomes. To our knowledge, the disappoint- linear—and where we therefore derive formal
ment-aversion models by Bell (1985), Loomes results under the assumption that m 1 w 2 5 w.9 In
and Sugden (1986), and Gul (1991), come closest Section V, we return to an exploration of large-
to saying that the reference point is recent expec- scale risk, where risk preferences can be sub-
tations.8 But because these models assume the stantially influenced by diminishing marginal
reference point is formed after choice, they treat utility of wealth. We organize our results on
surprise situations differently from our theory, modest-scale risk into three sections accord-
predicting in particular first-order risk aversion ing to the expectational environment the deci-
for surprise losses. Our model can be thought sion maker faces, considering in turn “surprise,”
of in part as unifying prospect theory and dis- PPE, and CPE situations. But a number of
appointment-aversion theory in one framework, themes link the three sections. Propositions 1,
while also proposing a solution concept (PPE) 2, 5, and 6 identify common ways in which the
for situations where choices are anticipated but decision maker becomes less risk averse if she
not committed to in advance. We are not aware had been expecting, or is now facing, more risk.
of any model that attempts such a unification. Proposition 3 shows that a person is first-order
Although our model is in the tradition of most risk averse when she anticipates a risk as well
of economics in that it posits a utility function as the possibility to insure it. And Propositions
with certain properties, it differs from much 4, 7, and 8 show that the more a risk can be pre-
of the foundational literature on choice under pared for, the greater is the risk aversion dis-
uncertainty in that it does not derive the utility played in behavior: the person’s risk aversion is
function from axioms capturing those properties. greater when she expects an insurance decision
It also differs from most economic theories in than when she does not, and even greater when
making explicit how utility depends on a mental she can commit early to purchasing insurance.
state—beliefs—that is not directly observable This section begins the analysis with risk-tak-
in choice behavior. Obviously, neither of these ing behavior when the reference point is fixed,
features means that our model does not have considering both deterministic and stochastic
observable or falsifiable implications. Indeed, in reference points. The analysis is the limiting
Appendix B we show how to extract the full util- case of UPE/PPE behavior when the decision
ity function 1m 1 # 2 and m 1 # 2 2 from behavior in maker finds herself in an ex ante low-probabil-
a limited set of decision problems, and because ity situation, so that she has fixed expectations
our model provides a complete mapping from formed essentially independently of the relevant
decision problems to possible choices, this com- choice set.
pletely ties down predictions for all other deci-
sion problems. And the many propositions in the
paper derive general restrictions that our model
implies for observed behavior. 9
If m[ were not linear (but remained differentiable),
some of our results would survive unchanged, and the
others could be modified by restating them in terms of
see Victoria Husted Medvec, Scott F. Madey, and Thomas expected consumption utilities instead of expected values.
Gilovich (1995), Barbara Mellers, Alan Schwartz, and Ilana But because this would complicate our statements, and
Ritov (1999), and Hans C. Breiter et al. (2001). because consumption utility is so close to linear for mod-
8
In addition, the notion of “loss-aversion equilibrium” est stakes, we assume m[ is linear. To give a sense of the
that Shalev (2000) proposed for multiplayer games can calibrational appropriateness of this approximation, note
be interpreted as saying that each player’s reference point that even for a person who has a low $1 million in lifetime
is expectations. As we discuss below, rewriting Shalev’s wealth and a very high consumption-utility coefficient of
model using our specification of stochastic reference points relative risk aversion of 10, winning or losing $1,000 (a dif-
and applying it to individual decision making corresponds ference of $2,000 in wealth) changes marginal consump-
to the UPE solution concept. tion utility by only 1.8 percent.
VOL. 97 NO. 4 KŐSZEGI AND RABIN: REFERENCE-DEPENDENT RISK ATTITUDES 1053

Proposition 2 of Kőszegi and Rabin (2006) it makes a set of novel predictions regarding the
shows that when the decision maker expects effect of prior uncertainty on behavior, identi-
to keep the status quo, her behavior is identical fying senses in which the expectation of risk
to that predicted by prospect theory modified decreases aversion to the expected as well as
to equate decision weights with probabilities. additional risks. To state these results, we use
Hence, our model is consistent with much of H 1 Hr to denote the distribution of the sum of
the evidence motivating status quo prospect independent draws from the distributions H and
theory. It can also be used to interpret the H9. (Thus, 1H 1 H92 1z2 5 1 H 1z 2 s 2 dH91s 2.)
“disposition effect” found by Odean (1998) When it creates no confusion, a real number will
for stocks and Genesove and Mayer (2001) for denote both a deterministic wealth level and the
houses—whereby people appear disproportion- lottery that assigns probability 1 to that amount
ately reluctant to sell an asset for less than they of wealth. Proposition 1 says that under A3r, a
paid.10 The intuition commonly invoked for the person is no more willing to accept a given lot-
disposition effect, formalized by Gomes (2005) tery if it is added to a riskless reference point
in a version of prospect theory based on the than if it is added to a lottery and/or evaluated
lagged status quo, is that because the purchase relative to a risky reference point.
price operates as a reference point for the sell-
ing price, people will be risk-seeking in waiting PROPOSITION 1: Suppose m 1 # 2 is linear and
for a price to recover before selling.11 While our m 1 # 2 satisfies A39. For any lotteries F, G, and
model cannot explain all aspects of the disposi- H and constant w, if U 1 w 1 F 0 w 2 $ U 1 w 0 w 2 ,
tion effect, by basing the reference point on the then U 1 H 1 F 0 G 2 $ U 1 H 0 G 2 .
expected resale price, it does help to determine
when and how the effect is likely to be observed. Since m 1 # 2 is linear and A39 is satisfied, a
Because home and stock owners usually expect small change in an outcome is evaluated solely
to make money, they will be risk loving when according to the previously expected probabili-
these investments unexpectedly lose money. But ties of getting higher and lower outcomes, and
when a person foresees a good chance of losing not according to the distance from those higher
some money—for example, when investing in a and lower outcomes. Now when F is added to a
highly risky stock—she will be less willing to riskless reference point w, positive outcomes of
take chances to break even. And if she expects F are assessed as pure gains, and negative out-
an investment—such as a house in a booming comes of F are assessed as pure losses. But when
market or inventory a merchant expects to resell F is instead added to a lottery H and is evaluated
at a large margin—to make a large positive relative to a lottery G, positive outcomes of F
return, she may even be reluctant to sell at prices partially eliminate losses suffered from H rela-
insufficiently above the purchase price. tive to G, and are hence evaluated more favor-
While our model only replicates or qualifies ably than pure gains; and negative outcomes of
classical prospect theory in the settings above, F in part merely eliminate gains from H relative
to G, and are hence evaluated less unfavorably
10
than pure losses. For both these reasons, the
The disposition effect is closely related to the “break- decision maker is more willing to accept F.
even effect” coined by Thaler and Johnson (1990) for
monetary gambles. They predicted that following losses, An important implication of Proposition 1—
gambles that offer a chance to break even become espe- obtained by setting H 5 w and G 5 F —is a type
cially attractive. of endowment effect for risk: a person is less risk
11
Barberis and Wei Xiong (2006) show that prospect averse in eliminating a risk she expected to face
theory based on the lagged status quo does not necessarily than in taking on the same risk if she did not
imply an increased risk lovingness after losses. Intuitively,
for risks a person takes on voluntarily, losses are typically expect it. This prediction of our model contrasts
smaller than gains. Hence, a person is typically closer to with previous theories of reference-dependent
the reference point after a loss than after a gain, and due utility with which we are familiar. While little
to loss aversion this can mean greater aversion to substan- evidence on the issue seems to be available,
tial amounts of risk. In the Odean (1998) and Genesove and
Mayer (2001) studies, however, individuals can take more choice experiments by Jack L. Knetsch and J.
incremental risk—waiting more or less exactly until the A. Sinden (1984) and evidence on hypothetical
price returns to the reference point. choices in Michael H. Birnbaum et al. (1992) do
1054 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2007

indicate that people tend to be less risk averse maker had been expecting risk or is already
when selling a lottery they are endowed with facing risk, she is already exposed to stochastic
than when buying the lottery. utility-decreasing sensations of loss, so taking
A second way in which expecting risk additional risk does not add so much exposure
decreases risk aversion is that a person is to losses.
approximately risk neutral in accepting a lottery To illustrate immediate and later points,
that is “small” relative to the reference lottery. Table 1 shows some of our model’s implications
in a parameterized example with consumption
PROPOSITION 2: Suppose m[ is linear. For utility m 1 w 2 5 10,000 ln 1 w 2 , and gain-loss util-
any lottery F with positive expected value: ity m 1 x 2 5 !x for x $ 0 and m 1 x 2 5 23!2x
for x # 0.12 We consider three fifty-fifty gambles
(i) There exist A, e . 0 such that if the lot- of different scales. Rows correspond to environ-
tery G satisfies PrG[r [ 1k 2 A, k 1 A2] , ments in which the gambles might be evaluated.
e for all constants k, then U 1H 1 F Z G2 . For each gamble, the first column identifies the
U 1H Z G2 . U 1H 2 F Z G2 for any lottery H. premium the decision maker is willing to pay
for insurance against the risk in each environ-
(ii) For any continuously distributed lottery ment, calculated as the difference between the
G, there is a t{ . 0 such that for any t [ gamble’s mean and its certainty equivalent. The
10, t{ 4 and any lottery H, U 1H 1 t ? FZG2 . second column gives the coefficient of relative
U 1HZG2 . U 1H 2 t ? FZG2. risk aversion r~ that would be inferred from the
gamble’s certainty equivalent if one assumed
The proposition identifies attitudes toward reference-independent CRRA utility.
a given lottery F with positive expected value, Gamble I is a small-stakes gamble that pays
showing ways in which the decision maker is risk either $1,000,100 or $999,900. The first panel
neutral when evaluating multiples of F. Part 1 of the table considers surprise situations with
says that if the reference lottery is sufficiently various (exogenous) expectations. The decision
widely distributed, the decision maker will take maker displays extreme risk lovingness if the
F and reject 2F. Part 2 says that fixing a con- gamble is a small loss relative to expectations,
tinuously distributed reference lottery, she will and extreme risk aversion if it is a small gain or
take a sufficiently small multiple of F and reject involves both a loss and a gain. Although loss
the same multiple of 2F. Intuitively, if the extra aversion is a stronger force than diminishing
lottery is small relative to the reference lottery sensitivity, the decision maker displays more
G, it is unlikely to turn a gain or loss relative risk aversion for gains than for mixed gambles
to G into the opposite, and there is little dimin- because in the latter case she evaluates the cer-
ishing sensitivity over its range. Hence, neither tainty equivalent as a loss.13 And as Propositions 1
loss aversion nor diminishing sensitivity plays a and 2 predict, the prior expectation of risk
major role in its evaluation. decreases risk aversion, with expected risk on
Part 2 of Proposition 2 is related to Proposi- the order of $1,000 or $10,000 already lowering
tion 1 in Barberis, Huang, and Thaler (2006)
showing that a loss-averse (in their terminology 12
Note that, unless A39 holds, our model is not invariant
“first-order risk-averse”) decision maker is only to affine transformations of m[, so the appropriate speci-
second-order risk averse if she faces background fication of m[ and m[ involves a substantive assumption
risk. In addition, Proposition 2 demonstrates about their relative scaling. This scaling amounts to an
that even if a person now faces no background assumption about the speed of diminishing sensitivity in
gain-loss utility. Indeed, the choice to specify our example
risk (H is riskless), the mere prior expectation of with m(w2 5 10,000ln(w2 is designed to get gain-loss utility
risk makes her second-order risk averse. to dominate for an appropriate range of small stakes, and
Crucially for a number of later results, Propo- consumption utility to dominate for an appropriate range
sitions 1 and 2 do not imply that a person is of large stakes.
13
unbothered by risk she expects or faces, even Indeed, while not shown in the table, the riskless
reference point that makes the person indifferent between
if it does make her less risk averse. In fact, the the reference point and the gamble—a situation where the
same force that decreases risk aversion also insurance premium is not evaluated as a loss—corresponds
decreases expected utility: when the decision to a much higher, $48 premium for insurance.
VOL. 97 NO. 4 KŐSZEGI AND RABIN: REFERENCE-DEPENDENT RISK ATTITUDES 1055

Table 1—Attitudes toward Three 50250 Gambles

Gamble I Gamble II Gamble III

$999,900/$1,000,100 $990,000/$1,010,000 $500,000/$1,500,000
WTP r~ WTP r~ WTP r~
Surprise with different (fixed) expectations
$500,000 $0.01 1.01 $50 1.01 $135,840 1.02
$990,000 0.02 3.4 329 6.5 132,477 0.99
$999,900 19.2 3,936 469 9.4 132,248 0.99
$1,000,000 9.17 1,844 432 8.6 132,246 0.99
$1,000,100 233.1 27,171 397 7.9 132,244 0.99
$1,010,000 2,0.01 25.6 2712 214 132,028 0.99
$1,500,000 ,0.01 0.97 49 0.97 128,429 0.96
U[$999,900, $1,000, 100] 3.66 733 433 8.7 132,246 0.99
U[$999,000, $1,001,000] 0.27 52 507 10.2 132,246 0.99
U[$990,000, $1, 010, 000] 0.02 3.5 275 5.5 132,248 0.99
PPE with different amounts of background risk
none $48.0 11,389 $757 15 $138,502 1.04
U[2$100, $100] 8.54 1,717 779 14 138,467 1.04
U[2$1,000, $1,000] 0.56 111 656 13 138,394 1.03
U[2$10,000, $10,000] 0.03 6.0 259 4.7 138,176 1.03
CPE with different amounts of background risk
none $71.0 23,348 $757 15 $138,502 1.04
U[2$100, $100] 28.5 6,041 779 14 138,467 1.04
U[2$1,000, $1,000] 1.84 368 656 13 138,394 1.03
U[2$10,000, $10,000] 0.07 14 362 7.2 138,176 1.03
Notes: The table assumes a consumption utility of m 1w2 5 10, 000 ln 1w2 and a gain-loss utility of m 1x 2 5 !x for x $ 0 and
m 1x 2 5 23 !2 x for x , 0. For each gamble and environment, we calculated the gamble’s certainty equivalent in that envi-
ronment. WTP is the difference between the gamble’s mean ($1,000,000 in each case) and the certainty equivalent. This is
the premium the decision maker is willing to pay for insurance. r~ is the coefficient of relative risk aversion that would be
inferred from the gamble’s certainty equivalent if one assumed reference-independent CRRA utility. U[x, y] refers to a dis-
crete uniform distribution between x and y, with atoms on multiples of $50.

the decision maker’s willingness to pay for is equivalent to the one-dimensional version of
insurance to essentially zero. the model in Kőszegi and Rabin (2006).
Suppose the decision maker has probabilis-
III. UPE and PPE Risk Attitudes tic beliefs over possible compact choice sets
described by {D1, 1 2 q; D2, q}, where choice
While the previous section considered risk sets D1, D2 ( D 1R2 occur with probabilities
attitudes in surprise situations, our primary 1 2 q and q, respectively. All of our results in
interest is in behavior when the decision maker the next two sections are for the case q 5 0, so
correctly predicts the choice set she faces. From that the decision maker knows the single choice
the psychological hypothesis that the reference set she will face. But we keep our definition a
point for evaluating outcomes is equal to lagged little more general because—as we explain
rational beliefs about those outcomes, we develop below—this allows us to capture the kinds of
two reduced-form models that differ in when surprise situations discussed in the previous sec-
the decision maker makes a committed choice. tion as limiting cases of low-probability choice
In this section, we analyze her behavior in one sets.
extreme possibility, when she anticipates the Since the person makes her decision shortly
decision she faces but cannot commit to a choice before the outcome resulting from it, at that time
until shortly before the outcome. Although our the beliefs determining the reference point are
model is ambiguous as to the interpretation of past and hence unchangeable. This means that
“shortly before,” insurance choices on short- she maximizes utility taking the reference point
term rentals such as cars or skis probably best as given, so that she can rationally expect to fol-
correspond to such a case. The resulting model low a plan of behavior only if she is willing to
1056 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2007

follow it given a reference point generated by the There can be multiple UPE in a given situa-
expectation to do so. tion—there can be multiple self-fulfilling expec-
tations—and generically different UPE yield
DEFINITION 1: A selection F1 [ D1, F2 [ D2 different expected utilities. But the person’s
is an unacclimating personal equilibrium expectations are based on her own plans on what
1UPE) if for each l [ 1, 2 and any F9l [ Dl, to choose once the time comes. It seems likely,
U 1Fl| 11 2 q 2F1 1 qF22 $ U 1F9l | 11 2 q 2F1 1 therefore, that she will choose the best plan she
qF22 .14 knows she will follow through on.

If the person expects to choose F1 and F2 DEFINITION 2: A selection F1 [ D1, F2 [ D2

from choice sets D1 and D2, respectively, then is a preferred personal equilibrium 1PPE) if it is
given her expectations over possible choice a UPE, and U 1 11 2 q 2F1 1 qF2| 112q 2 F1 1 q
sets, she expects the distribution of outcomes F22 $ U 1 11 2 q 2F91 1 q F92 Z 112q 2 F91 1 q F922 for
1 1 2 q 2 F1 1 qF2. Definition 1 states that with all UPE selections F9l [ D1, F92 [ D2.
those expectations as her reference point, she
should indeed be willing to choose F1 and F2 A major feature of UPE and PPE is the con-
from choice sets D1 and D2. straint that choice must be optimal given expec-
UPE is closely related to the notion of “loss- tations at the time. This means that the decision
aversion equilibrium” that Shalev (2000) defined maker does not internalize the effect of her choice
for multiplayer games as a Nash equilibrium fix- on expectations, so—as we will illustrate—she
ing each player’s reference point, where the ref- often does not maximize ex ante expected utility
erence point is equal to the player’s (implicitly among the choices available to her.
defined) reference-dependent expected utility. To begin our analysis of UPE and PPE
Although Shalev does not himself pursue this behavior, we note that the results in Section II
direction, reformulating his notion of loss-aver- on expectations fixed independently of the per-
sion equilibrium using our utility function and son’s choice set D1 can be thought of as apply-
applying it to individual decision making cor- ing UPE or PPE to situations where that choice
responds to UPE. set is a surprise: if she had been expecting to
In the decision between a fifty-fifty chance face choice set D2 with near certainty and D1
of having to pay $100 out of current wealth w with very small probability, and to choose
and buying insurance for $55, when is choosing F [ D2 (perhaps because D 2 5 {F}, so that she
the lottery a UPE? If the lottery is the reference thought she would have no choice), her refer-
point, the following inequality indicates when it ence point would be approximately F indepen-
is preferred to paying $55: dently of D1 or what she had been expecting to
choose from D1.
To study decisions for choice situations that
1 1 are anticipated, we first establish that a person
(3) c 1 w 2 100 2 1 w d
2 2 has a strong preference to insure modest-scale
risks—she is first-order risk averse—and then
1 1 show that expecting risk at the start decreases
1 c m 1 100 2 1 m 1 2 100 2 d
4 4 her aversion to additional risk.

$ [w 2 55] PROPOSITION 3: Suppose m[ is linear.

For any w [ R and mean-zero lottery F ? 0
1 1 with bounded support, there exist k{ , t{ . 0 such
1 c m 1 45 2 1 m 1 2 55 2 d .
2 2 that for any positive t , t, k , k{ , the unique
PPE with the choice set {w, w 1 t 1F 1 k 2} is
14
to choose w.
Because each of our solution concepts is an example
of personal equilibrium as first defined in Kőszegi (2005),
Theorem 1 of that paper implies that when 11 2 q 2D1 1 Proposition 3 states that a person will choose
qD 2 is convex and compact, UPE, as well as PPE and CPE, a riskless w over a sufficiently small multiple
exist. of a better-than-fair but insufficiently attractive
VOL. 97 NO. 4 KŐSZEGI AND RABIN: REFERENCE-DEPENDENT RISK ATTITUDES 1057

bet.15 Intuitively, w is a UPE because when the sensitivity and create aversion to risk. But even
decision maker expects it, loss aversion leads for these decisions, the extent of risk aversion
her to turn down the gamble. And w is a PPE often seems quantitatively much stronger than
because the gamble exposes the decision maker predicted by prospect theory. A calibrational
to a sense of loss from bad outcomes that out- problem is immediately apparent in intuitive
weighs the sense of gain from good outcomes, terms if we pose the deductible choices of hom-
and hence yields lower expected utility than w. eowners analyzed by Justin Sydnor (2006) as
In predicting attitudes toward insuring mod- gambles relative to the status quo. For the aver-
erate and high-probability losses, status quo age consumer with a $500 deductible, the pre-
prospect theory says both that loss aversion does mium for a $1,000 deductible is about $600, and
not play a role, and that diminishing sensitivity the premium for a $500 deductible is about $100
pushes people toward not insuring. When it comes higher. The probability of making a claim in any
to insuring expected losses, our model reverses given year is 0.05. Hence, the choice between
this prediction: it says that loss aversion plays the the two deductibles is equivalent to the choice
crucial role and pushes people toward insuring. between the gambles (2600, 0.95; 21600, 0.05)
As a key force behind this prediction, our model and (2700, 0.95; 21200, 0.05).17 We conjecture
captures an intuition regarding the difference that when asked in these terms, most individu-
between “costs” and “losses” that has sometimes als would choose the former gamble, and indeed
been articulated (e.g., in Kahneman and Tversky Sydnor (2006) shows that typical parameteriza-
1984 and Nathan Novemsky and Kahneman tions of prospect theory make the same predic-
2005), but has not been formalized. In our model, tion. Yet the majority of consumers in his dataset
beliefs about wealth take into account a planned chose the latter gamble.
premium payment, so that such a payment is not But the difference between anticipated and
evaluated as a loss. By contrast, whether or not surprise situations extends beyond attitudes
a person had expected to buy insurance, a bad toward losses. As indicated in the second panel
realization of a stochastic lottery is evaluated as of Table 1, the certainty equivalent for Gamble I
a loss. Because loss aversion therefore plays a in our example is lower in a PPE situation with
central role in the decision of whether to insure, no background risk than in any of the sur-
first-order risk aversion results. Indeed, some prise situations. Indeed, if a person is deciding
consumers’ purchase of insurance against high- between taking on a risk and fully insuring it,
probability losses—such as extremely expensive under A3r she is at least as risk averse in PPE
automobile service contracts—seems consistent as in a surprise situation, whatever her expecta-
with this prediction and is in direct contrast to tions may have been in the latter case.
status quo prospect theory.16
For low-probability losses—such as those cov- PROPOSITION 4: Suppose m[ is linear and
ered by extended warranties and low deductibles m[ satisfies A39. If w 1 F is a PPE in the
on homeowners’ insurance—the overweighting choice set {w,w 1 F}, then for any lottery H,
of low probabilities in conventional status quo U1w 1 F0H2 . U1w0H2.
prospect theory can counteract diminishing
Because of the many ways it predicts that
behavior in surprise situations differs from that in
15
Proposition 11 in Appendix A identifies a precise con- expected situations, our theory cautions against
dition of the attractiveness of a vanishingly small lottery extrapolating from experimental results too
that determines whether the decision maker accepts the casually. Insofar as most experimental subjects
lottery.
16
An experiment by Antoni Bosch-Domènech and do not have a clear idea about the tasks they are
Joaquim Silvestre (2006) can also be interpreted as provid- going to face, their behavior does not correspond
ing evidence for risk aversion in high-probability losses. to what they would do in a similar but expected
Subjects received money for their performance in the first
session of the experiment, and were asked to come back
weeks later for a second session. Subjects were warned that
17
they could lose money during the second session. In this Strictly speaking, the two situations are not equivalent
session, a majority of students were risk averse for both if some claims are between $500 and $1,000. But if so, a
high- and low-probability losses. high deductible is all the more attractive.
1058 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2007

situation in real life. Specifically, Proposition 4 IV. CPE Risk Attitudes

says that laboratory measurements of small-scale
risk aversion will often underestimate risk aver- Beyond the distinction between surprise and
sion for expected small-scale risks. anticipated exposure to risky decisions, our
We conclude this section with results on underlying assumptions also imply differences
two ways in which expecting risk at the start in attitudes toward anticipated risks as a func-
decreases aversion to additional risk. Although tion of how far in advance decisions are com-
our formal results identify the effect of facing mitted to. We now analyze risk preferences
a lottery G the decision maker cannot avoid, regarding outcomes that are resolved long after
this can be the reduced-form representation of a all decisions are committed to, a situation that
situation where she could, but (because it is too applies to most insurance choices. In this case,
attractive) in equilibrium does not avoid G. First, the expectations relative to which a decision’s
the decision maker is no less likely to accept a outcomes are evaluated are formed after—and
given lottery if she is already facing risk than if therefore incorporate the implications of—the
she is not. decision.

PROPOSITION 5: Suppose m 1 # 2 is linear and DEFINITION 3: For any choice set D, F [ D

m 1 # 2 satisfies A39. For any lotteries G and F is a choice-acclimating personal equilibrium
and constant w, if choosing G is a UPE with the 1CPE) if U 1FZF) $ U 1F9ZF9) for all Fr [ D.
choice set {G,G 1 F}, then choosing w is a PPE
with the choice set {w, w 1 F}. If the decision maker makes the choice F [ D
today, this will determine her reference point by
And as a corollary to Proposition 2, a person the time the relevant wealth outcome occurs.
is neutral to risks that are small relative to the Thus, when evaluating her resulting expected
risk she is already expecting. utility, both the reference and outcome lotteries
are equal to F. Note that CPE can naturally be
PROPOSITION 6: Suppose m[ is linear. For extended to decisions where some uncertainty
any lottery F with positive expected value: is resolved today. In this case, immediately
resolved uncertainty gets absorbed into the ref-
(i) There exist A, e . 0 such that for any erence point, so that the decision maker maxi-
lottery G for which PrG 1r [ [k 2 A, k 1 A]2 mizes the expectation of U 1 F 0 F 2 taken over
, e for all constants k, the unique UPE with the possible lotteries F that capture solely the
the decision set {G,G 1 F} is to choose G 1 remaining uncertainty. Hence, for instance, if
F, and the unique UPE with the decision set somebody is confronted with a choice between
{G, G 2 F} is to choose G. getting $10,000 for certain a month later and
getting $25,000 a month later if an immediately
(ii) For any continuously distributed lottery observed coin flip comes up heads, we predict
G, there is a t{ . 0 such that for any t [ 10, t{ ], she would evaluate her options solely in terms
the unique UPE with the decision set {G, G 1 of expected consumption utility.
t ? F} is to choose G 1 t ? F, and the unique Our notion of CPE is related to the models
UPE with the decision set {G, G 2 t ? F} is to of “disappointment aversion” of Bell (1985),
choose G. Loomes and Sugden (1986), and Gul (1991),
where outcomes are also evaluated relative to a
The second panel of Table 1 quantifies Proposi- reference lottery that is identical to the chosen
tions 5 and 6 for our parameterized example, lottery. As we have mentioned, however, these
showing that the decision maker approaches risk theories differ from ours in two important ways.
neutrality for Gamble I even for relatively limited They assume that a person evaluates outcomes
amounts of background risk. An unavoidable relative to the reference lottery’s certainty
risk on the order of a mere $100 decreases the equivalent instead of the full distribution. And
premium she is willing to pay for insurance from because they do not distinguish expectational
$48 to $9, and unavoidable risk on the order of environments at all, they do not share many of
$1,000 makes her virtually risk neutral. our model’s predictions in other situations.
VOL. 97 NO. 4 KŐSZEGI AND RABIN: REFERENCE-DEPENDENT RISK ATTITUDES 1059

Returning to our example of choosing are incorporated into the reference point, so that
between a 50 percent chance of losing $100 and it may be unclear whether a person’s commit-
insuring this risk for $55, selecting the lottery ted decision is made sufficiently early that CPE
is a CPE if rather than PPE is appropriate for modeling her
behavior. In many situations, this will not be a
1 1 major problem, as the appropriate concept will
(4) c 1 w 2 100 2 1 w d be intuitively clear. In addition, because the two
2 2
concepts generate distinctive behavior, observed
1 1 choices can also be used to determine which
1 c m 1 100 2 1 m 1 2 100 2 d
4 4 concept applies.19
As the proofs in Appendix C establish, Prop-
$ [w 2 55] 1[0]. ositions 3 through 6 derived above for PPE also
apply to CPE. Because anticipated risk exposes
The difference between UPE and CPE is in the decision maker to sensations of loss, she is
the right-hand sides of inequalities (3) and (4), first-order risk averse. Furthermore, expecting
which capture the decision maker’s expected risk reduces her aversion to additional risk. This
utilities when deviating from the purported latter result follows partly from the force behind
UPE and CPE, respectively. In UPE, the refer- our analogous results above, that, when expect-
ence point does not adjust to the deviation, so ing and facing stochastic losses to start with,
paying $55 is assessed partly as a loss of $55 and taking further risk does not increase exposure to
partly as a gain of $45. In CPE, the reference losses as much. When a person chooses both her
point does adjust to the deviation, so there is no reference and outcome lotteries, there is an addi-
sensation of gain or loss when paying the $55.18 tional, parallel force acting in the same direc-
As with PPE, except in knife-edge cases there tion: when expecting and facing stochastic losses
will be a unique CPE. But unlike in PPE, where to start with, the expectation of further risk does
the decision maker can choose her favorite plan not increase exposure to losses as much.
only from those that she would follow through Despite these similarities, CPE is consistent
on, in CPE she can commit to her overall favor- with risk aversion that is qualitatively differ-
ite lottery. Hence, there cannot be a divergence ent, not only from standard expected-utility-
between behavior and welfare. over-wealth models and prospect theory, but
It bears emphasizing that UPE/PPE and CPE also from the predictions of UPE, PPE, and
are not different theories of what outcomes peo- any model where the reference point is taken
ple prefer. Indeed, in our example the expected as given at the moment of choice. Whereas in
utility from choosing the lottery (the left-hand these models people never choose stochasti-
side of inequality (3) or (4)) is independent of cally dominated options, they might do so in
whether the choice is determined by UPE or CPE. To illustrate this possibility, suppose m
CPE. Rather than reflecting different notions of satisfies A3r, and consider a lottery F that yields
reference-dependent utility, the two concepts are w 1 g . w with probability p $ 0 and w with
motivated by the same theory of preference, as probability 1 2 p. Then U1w 1 FZw 1 F2 5
manifested differently depending on whether the [p 1w 1 g2 1 11 2 p2 w] 1 [p 11 2 p 2 m 1g2 1
person can commit to her choice ahead of time. p 11 2 p 2 m 12 g2] 5 w 1 pg[1 2 11 2 p 2 h 1l 2
Of course, one weakness of our theory is that it 12]. If h 1 l 2 1 2 . 1, which is a calibrationally
does not specify the lag with which new beliefs
19
For instance, suppose F1 is a small binary lottery that
18
CPE implicitly assumes that the decision maker generates barely enough gains for w 1 F1 to be chosen
maximizes the expected future sensations generated once from {w, w 1 F1}, and let F2 be a different small binary
a reference point determined by the decision is formed lottery with the same property. Let F be a mixture of F1
and outcomes are resolved. As we discuss in the conclu- and F2 with equal weights. Since in PPE the reference point
sion, a person may not appreciate how her decision will is fixed at the moment of choice, it is easy to show that in
affect future sensations of gain or loss, and she may also PPE the person would choose w 1 F from {w, w 1 F}. But
be influenced by current anticipatory utility. Similarly to since in CPE a person influences the reference point by her
other models of reference-dependent utility, it seems to be choice, and she dislikes risky reference points, in CPE she
an appropriate first approximation to ignore these issues. would choose w from {w, w 1 F}.
1060 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2007

plausible situation, the decision maker prefers p whereas consumption utility increases linearly.
5 0 over a small p . 0.20 Intuitively, raising And as we discuss in the conclusion, people may
expectations of getting g makes an outcome of underappreciate how coming to expect the gain
no gain feel more painful. To avoid such disap- increases future sensations of loss, and hence
pointments, the person would rather give up the may not be averse to choosing the lottery.
fragile hope of making gains. In fact, if gain- The possibility of rejecting a probabilistic
loss utility is sufficiently important, reducing gain under CPE but not under PPE reflects a
exposure to sensations of loss is the decision more general sense in which people are more
maker’s central concern. risk averse under CPE than PPE. Proposition 8
establishes that under A3r, if the decision maker
PROPOSITION 7: Suppose m 1 # 2 is linear and chooses the less risky one of two lotteries in PPE,
the decision maker faces the finite choice set D she does not choose the riskier one in CPE.
containing the deterministic outcome w and no
greater deterministic outcomes. For any given PROPOSITION 8: Suppose m 1 # 2 is linear, A39
m0 1 # 2 satisfying A2, there is an h such that if holds, and for different lotteries F and Fr in the
m 1 # 2 5 hm0 1 # 2 with h . h, the unique CPE decision maker’s choice set, Fr is a mean-pre-
is to choose w. serving spread of F 1 k for some constant k. If
F is a PPE, Fr is not a CPE.
While the tendency to choose a stochastically
dominated lottery may seem counterintuitive, it The intuition derives directly from the deci-
is consistent both with the flavor of some dis- sion maker’s desire to avoid risky expectations.
cussions in the psychology literature on welfare That F is a PPE implies that holding the refer-
in risky situations, and with some experimen- ence point fixed at F, the person prefers F to F9.
tal results on risk taking. Shane Frederick and When her choice affects the reference lottery in
George Loewenstein (1999) discuss, for exam- addition to the outcome lottery, as in CPE, her
ple, how a prisoner may be made worse off by dislike of risky reference points leads her to pre-
a small chance of being released, because that fer F to F9 even more.
makes the outcome of remaining in prison much The most important implication of Proposi-
more difficult to bear. In the domain being exam- tion 8 is that people will be more risk averse
ined in this paper, Uri Gneezy, John A. List, and when decisions are committed to well in
George Wu (2006) find situations where risky advance than when people are uncommitted.
choices are valued less than their worst pos- But the same results also say that the availabil-
sible outcome. We also feel that the preference ity of some risky options—exactly those that the
for a stochastically dominated lottery captures decision maker takes in PPE but would not take
in extreme form the strong risk aversion con- in CPE—decreases welfare in PPE. Intuitively,
sumers display when purchasing insurance for in PPE the decision maker realizes that she will
long-term modest-scale losses, choosing low take a lottery that is attractive fixing expecta-
deductibles on existing insurance, and selecting tions, and because she incorporates the possibil-
expensive fixed-fee contracts for services. ity of good outcomes into her reference point, her
There are ways in which our result must be sense of loss from low outcomes is increased.
qualified, however. The preference for domi- Both the similarities and differences between
nated lotteries clearly arises only when such PPE and CPE are illustrated in Table 1.
lotteries reduce exposure to gain-loss sensa- Comparing the second and third panels for
tions. In addition, diminishing sensitivity can Gamble I, for any given amount of background
substantially reduce a person’s dislike of risk for risk, CPE choices are more risk averse than
modest stakes: as the gain g in the lottery above PPE choices. Nevertheless, CPE behavior also
increases, the sensation of loss from comparing approaches risk neutrality with even moderate
nothing to g increases more and more slowly, amounts of background risk.

20
It is easy to check that h 1l 2 12 . 1 whenever overall sensitivity to losses, 1 1 hl, is at least twice as high
observed loss aversion is at least two-to-one—whenever as overall sensitivity to gains, 1 1 h.
VOL. 97 NO. 4 KŐSZEGI AND RABIN: REFERENCE-DEPENDENT RISK ATTITUDES 1061

V. Immodest Risk model is due to a systematic misspecification

of that model, and a great deal of the variation
We now illustrate some of our model’s impli- reflects reference-dependent utility. In addition
cations for attitudes toward large-scale risk, to the patterns identified in previous sections, our
showing that for such stakes m 1 # 2 can become model predicts a systematic relationship between
largely irrelevant in determining risk prefer- the inferred coefficient of risk aversion and the
ences. Combined with the previous three sec- scale at which it is measured: when measured on
tions, this means that our model can reconcile large-stakes data, single- and double-digit coef-
highly context-dependent behavior in modest- ficients will be found; when measured on mod-
stakes gambles with context-independent “clas- est-stakes data, triple-digit coefficients will be
sical” predictions for larger stakes. found; and when measured on small-stakes data,
Proposition 9 shows that, when diminishing coefficients too embarrassingly large to report
sensitivity is a significant-enough feature of will be found. Indeed, Raj Chetty (2005) shows
gain-loss utility, the expected utility from very that existing evidence on the income elasticity
risky outcomes is little influenced by the refer- of labor supply comfortably bounds the coeffi-
ence point. cient of relative risk aversion from above by two.
Based on hypothetical choices between large
PROPOSITION 9: Suppose m[ has full range gambles on lifetime wealth, Robert B. Barsky
and lim xS` m91x 2 5 lim xS2` m91x 2 5 0. For et al. (1997) measure an average coefficient of
any r, r9 [ R, r . r9 and e . 0, there is a relative risk aversion of around five. Fitting an
d . 0 such that if F is continuously distrib- unemployment model with consumption and
uted with density less than d everywhere, then search-effort choices to data on unemployment
0 # U 1 F 0 rr 2 2 U 1 F 0 r 2 , e. durations, Chetty (2003) finds a coefficient of
relative risk aversion of around seven. Rajnish
If F is a very risky gamble, most of its out- Mehra and Edward C. Prescott (1985) estimate
comes are far from both r and rr. Since sensi- that to explain the historical equity premium,
tivity of gain-loss utility to changes approaches investors must have a coefficient of relative risk
zero for comparisons far apart, for a typical out- aversion well in the double digits. The deduct-
come of F it makes little difference whether it is ible choices of American homeowners analyzed
being compared to r or rr. by Sydnor (2006) and the risk-taking behav-
Beyond this analytical result, our model’s ior of Paraguayan farmers analyzed by Laura
predictions for large-scale risks can be illus- Schechter (2005) imply coefficients of relative
trated by applying the parameterized version we risk aversion in the triple digits. And although
have considered above to larger stakes. Table 1 risk attitudes are often measured in small-stakes
performs this exercise for two gambles. Gamble laboratory experiments, coefficients calculated
II is a fifty-fifty gamble that yields either using the same mathematics as those above are
$990, 000 or $1, 010, 000, and Gamble III is a typically not reported; if they were, they would
fifty-fifty gamble that yields either $500, 000 be extremely high.21
or $1, 500, 000. While the qualitative patterns
of behavior are similar for the $100-stakes VI. Caveats, Discussion, and Conclusion
Gamble I and the $10,000-stakes Gamble II,
the sensitivity of behavior to the environment Several complications limit the degree to which
is significantly smaller in the latter case. And our model moves us to a full understanding of
for the $500,000 gamble, risk attitudes are close
to what they would be without gain-loss utility,
being largely determined by m 1 # 2 indepen- 21
Many papers that report small coefficients of risk aver-
dently of the environment. sion for small stakes do so by dint of using the same termi-
The wide range of r~’s in Table 1 replicates nology to describe mathematically different measures. By
an observation that an increasing number of variously defining wealth as monthly income or potential
income over the course of a one-hour experiment rather
researchers have become aware of: that the wild than lifetime wealth, these papers report figures that are
variation in measured risk aversion inferred orders of magnitude different from measures using wealth
by use of the prevalent reference-independent levels similar to those in the estimates above.
1062 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2007

reference-dependent risk preferences. A weak- both these reasons, the appropriate welfare mea-
ness that our theory shares with other models sure is likely to be closer to consumption utility
is that it takes as one of its primitives the set of than we assume in this paper. This could sig-
decisions and risks a person is considering, as nificantly alter some of our welfare conclusions.
distinct from all the decisions and risks she is For example, although we showed above that
facing. In fact, as Barberis, Huang, and Thaler the availability of a lottery can (in PPE) lower
(2006) emphasize in their setting, Propositions 2 welfare defined to include gain-loss disutility,
and 6 can be interpreted as saying that if people the same analysis implies that this would not be
incorporated all the risks they are facing into possible for welfare based solely on consump-
their expectations, reference dependence would tion utility.
essentially not affect risk attitudes—because The underestimation of changes in the refer-
people would essentially be neutral to small ence point also has behavioral implications. If a
risks. Although psychological evidence does person underestimates the effect of changes in
indicate that people often “narrowly bracket”— her expectations on her preferences, she may not
they isolate individual decisions and risks from appreciate fully how she can rid herself of sensa-
relevant other decisions—relatively little is tions of loss by ridding herself of risky expecta-
known about the extent, patterns, and effects of tions. Because our predictions of first-order risk
such bracketing phenomena.22 aversion are driven partly by a person’s desire to
Another major limitation of our model con- avoid risky expectations, this underappreciation
cerns welfare. Although the analysis empha- can reduce risk aversion.
sized behavioral implications of our model, a Our model also glosses over a set of issues
welfare interpretation was implicit throughout related to what outcomes a person pays attention
and explicit at times. Insofar as the hedonic to. In contrast to our formal model, different
effects of choice include both gain-loss sensa- outcomes resulting from the same choice often
tions and consumption utility, our utility func- differ in salience. As noted in Sydnor (2006), for
tion may provide a useful welfare measure. For instance, having to pay for repairing an unin-
two specific reasons, however, we are more hesi- sured house is a very salient loss, but not having
tant about our model’s welfare implications than to pay is unlikely to result in a salient sensation
about its behavioral implications. First, the nar- of gain. A person who focuses mostly on such
row bracketing discussed above may lead people losses presumably has an even stronger taste for
to care too much prospectively about a gain or insurance than our model predicts.
loss whose effects are likely to be eliminated Finally, our model ignores a source of util-
afterward by an offsetting loss or gain. Second, ity—anticipatory emotions—that seems impor-
evidence indicates that people underestimate tant in many risky situations. For instance, an
how quickly the reference point will adjust to a investor’s anxiety about funding her child’s
choice, and hence put too much weight on gain- education is likely to affect both her welfare
loss sensations when making decisions.23 For and many of her financial decisions. Insofar
as anticipatory feelings are about future con-
22 sumption and gain-loss utilities, our qualitative
For papers highlighting the role of bracketing in
this and other domains, see Kahneman and Dan Lovallo results would not be affected by adding them
(1993), Shlomo Benartzi and Thaler (1995), Daniel Read, to the model. Quantitatively, however, anticipa-
Loewenstein, and Rabin (1999), Thaler (2000), and tory emotions can affect the degree of risk aver-
Barberis, Huang, and Thaler (2006). sion; Andrew Caplin and John Leahy (2001), for
23
For interpretation of evidence along these lines, see
Kahneman’s (2003) discussion of the “transition heuristic,”
instance, show that the attempt to avoid anxiety
and Loewenstein, Ted O’Donoghue, and Rabin’s (2003) about uncertain outcomes can increase a per-
formulation of “projection bias.” son’s preference for riskless options.

Appendix A: Further Definitions and Results

In this appendix we present an array of concepts and results that may be of practical use in apply-
ing our model, but that are not key to any of the main points of the paper. The first result identifies a
VOL. 97 NO. 4 KŐSZEGI AND RABIN: REFERENCE-DEPENDENT RISK ATTITUDES 1063

condition such that with the reference point being the status quo, the decision maker rejects all fair
gambles.

PROPOSITION 10: Suppose m 1 · 2 is linear and the reference point is $0. The decision maker rejects
all fair gambles if and only if lim xS` m912x 2 $ m91 102.

Assumption A3 allows the decision maker’s risk lovingness in losses to be much stronger than her
risk aversion in gains, which may even lead her to accept unfair gambles given a reference point of
$0. Proposition 10 says that when m 1 · 2 is linear, a necessary and sufficient additional condition to
rule out such possibilities is that sensitivity to losses is everywhere greater than sensitivity to gains.
When m 1 · 2 is concave, this condition is of course sufficient, but not necessary.
For our results on PPE and CPE behavior, we introduce three definitions characterizing the riski-
ness of lotteries. Our first definition is the conventional one of second-order stochastic dominance,
except that it allows comparisons of lotteries with different means.

DEFINITION 4: A lottery F is less risky than the lottery F9 if F9 is a mean-preserving spread of

F 1 k for some constant k [ R.

Because it turns out to be an especially pertinent measure of riskiness in our model, we also intro-
duce a more specific concept that is (to our knowledge) undefined and unexplored in the literature on
risk preferences.

DEFINITION 5: The average self-distance of a lottery F is

S 1F2 K Zx 2 yZ dF 1x 2 dF 1y 2.
5

The average self-distance of a lottery is the average distance between two independent draws from
the lottery. A lower self-distance is a necessary but not sufficient condition for one lottery to be unam-
biguously less risky than another.

LEMMA 1: If F is less risky than F9, then F has lower average self-distance than F9.

Finally, we introduce a measure for how a lottery’s possible gains compare to its possible losses.

DEFINITION 6: For a lottery F, let F1 5 EF [max{x,0}] and F25 EF[max{2x,0}]. The favorability
of F is defined as F 1 F 2 ; 1 if F1 5 F2 5 0, F 1F2 5 ` if F1 . 0, and F2 5 0 and F 1F2 K F1/F2
otherwise.

The favorability of a lottery is the ratio of the average gain of the lottery (relative to zero) and the
average loss. Using the concepts above, Proposition 11 precisely identifies the extent of the decision
maker’s first-order risk aversion, generalizing Arrow’s theorem to reference-dependent risky choice.
This extends the limit result of Proposition 3 in the text, which said that small bets that are insuf-
ficiently better than fair will be rejected.

PROPOSITION 11: Suppose w [ R, and F is a lottery with bounded support.

(i) If F 1 F 2 , 1 1 1 mr2 1 0 2 2 / 1 1 1 mr1 1 0 2 2 , then there exists a t . 0 such that for any positive
t , t, the unique PPE in the choice set {w, w 1 t ? F} is to choose w.

If F 1 F 2 . 1 1 1 mr2 1 0 2 2 / 1 1 1 mr1 1 0 2 2 , then there exists a t . 0 such that for any positive
t , t, the unique PPE in the choice set {w, w 1 t ? F} is to choose w 1 t ? F.
1064 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2007

(ii) If 2E[F] , 1m92 102 2 m91 102 2S[F], then there exists a t . 0 such that for any positive t , t, the
unique CPE in the choice set {w, w 1 t ? F} is to choose w. If 2E[F] . 1m92 102 2 m91 102S[F], then
there exists a t . 0 such that for any positive t , t, the unique CPE in the choice set {w, w 1 t ? F}
is to choose w 1 t ? F.

Part (i) states that when applying PPE, small bets will be accepted if and only if their favorability
is greater than the “coefficient of loss aversion” associated with u 1 w 0 r 2 —which is the slope ratio at
the kink in u 1 w 0 r 2 at w 5 r. Part (ii) says that when applying CPE, a small bet will be accepted if and
only if twice its expected value is greater than the product of its average self-distance and the differ-
ence in the decision maker’s sensitivity to small losses and small gains.
Proposition 12 shows that when consumption utility is linear and A3r holds, we can characterize a
person’s CPE attitude toward a lottery purely in terms the lottery’s mean and average self-distance.

PROPOSITION 12: Suppose m 1 w 2 5 w and m 1 # 2 meets A39. Then,

(i) For any lottery F,

1
U 1F|F2 5 E[F] 2 h 1l 2 12S[F].
2
(ii) For a finite choice set D, define y 1D) 5 {F [ D|4F9 [ D, either 1a 2 S[F] , S[F9], or 1b 2 S[F]
5 S[F9] and E[F] $ E[F9]}. For a sufficiently high h, y 1 D 2 is the set of CPE.

Part (ii) represents a lexicographic ranking of lotteries by their lowest average self-distance, and
then the highest mean. There will generally be a unique lottery with minimal average self-distance,
in which case that lottery is chosen when gain-loss utility is very important—even if it is stochasti-
cally dominated by other options.
The role of average self-distance in determining a person’s CPE choices when A39 holds is best
seen with a simple but striking observation. In a personal equilibrium of any sort, every possible
sensation of gain—say from comparing an outcome x to a counterfactual y , x —is matched by an
equally likely and equally large loss—from comparing y to x. Because the losses are more heavily
felt, net gain-loss utility will be proportional to the negative of the average of these distances.
Finally, we show two properties of risky choice in our model that it shares with the standard model.
Although expected risk decreases aversion to risk under either of our solution concepts, without
diminishing sensitivity it never eliminates the risk aversion completely.

PROPOSITION 13: Suppose m 1 # 2 is linear, A39 holds, and F second-order stochastically dominates
Fr 2 F. Then both the unique PPE and the unique CPE from the choice set {F, F9} is to choose F.

We also note that if F first-order stochastically dominates F9, then for any given reference lottery
the decision maker prefers F to F9. Hence, choosing F9 cannot be a UPE. While mathematically
trivial, this result is of interest because it contrasts with some of our results for CPE.

PROPOSITION 14: Suppose the decision maker faces the choice set D. If F [ D first-order stochas-
tically dominates F9, then F9 is not a UPE.24

Appendix B: Extracting m[ and m[ from Behavior

In this appendix, we provide an algorithm that identifies a person’s full utility function from her
behavior (as long as the consumption utility m[ inferred in the first step is continuous). Once this

24
In fact, this result relies solely on Assumption A1, guaranteeing that u 1wZr2 is increasing in w, and on no other feature of m.
VOL. 97 NO. 4 KŐSZEGI AND RABIN: REFERENCE-DEPENDENT RISK ATTITUDES 1065

utility function is identified, our theory provides a prediction for behavior in any decision over mon-
etary risk. To determine m 1 # 2 , our algorithm works whenever m 1 # 2 has unbounded support; minor
modifications cover the opposite case as well.
As we note in Section IV, when a person makes decisions over risk with immediate resolution and
delayed consequences, the natural extension of CPE predicts that she maximizes expected consump-
tion utility. This means that we can identify m 1 # 2 up to an affine transformation using standard
revealed-preference techniques. Once m 1 # 2 is identified and a normalization is chosen (e.g., setting
m 11,000,0012 2 m 11,000,0002 5 1), the next two parts of our procedure allow us to infer m 1 a 2 and
m 1 2 a 2 for any given a . 0. Carrying out these steps for all a . 0 identifies the entire function
m1 # 2.
For a given a, we find wealth levels w0,w1,w2 such that m 1 w2 2 2 m 1 w1 2 5 m 1 w1 2 2 m 1 w0 2 5 a.
Then, we find the wealth level wCE such that in CPE the person is indifferent between a deterministic
wCE and a fifty-fifty gamble that pays either w0 or w1. This certainty equivalent will satisfy

1 1 1
(5) m 1 w0 2 1 m 1 w1 2 1 1 m 1 a 2 1 m 12 a 2 2 5 m 1wCE 2 ,
2 2 4

so that we have identified m 1 a 2 1 m 1 2 a 2 . Intuitively, because the ex ante evaluation of risks

involved in CPE pairs any possible loss with a corresponding equally weighted gain, CPE behavior
allows us to infer the difference between the pain from a loss and the pleasure from the same-sized
gain. But for the same reason, CPE behavior cannot identify how large the pain and pleasure are.
To identify how big they are, we can identify the ratio of m 1 a 2 and 0 m 1 2 a 2 0 by eliciting the prob-
ability p . 1/2 such that a surprise chance p of a gain of a compensates the person for a surprise
chance 1 2 p of a loss of a. Specifically, we find the probability p such that when expecting w1, the
person is indifferent between w1 and a gamble that pays w2 with probability p and w0 with probability
1 2 p. To do this, we tell her that she will get w1 with probability 1 2 e and the choice with prob-
ability e, and we let e S 0. Then, p satisfies

(6) pm 1 w2 2 1 1 1 2 p 2 m 1 w0 2 1 pm 1 a 2 1 1 1 2 p 2 m 1 2 a 2 5 m 1 w1 2 ,

and if A2 is satisfied, p . 1/2. Equation (6) can be rearranged to show that

p
m 1 2 a2 5 1m 1a 2 1 m 12 a 2 2 1 a;
2p 2 1
12p
m 1a 2 5 2 1m 1a 2 1 m 1 2 a 2 2 1 a,
2p 2 1

which identifies m 1 2 a 2 and m 1 a 2 based solely on the person’s observed choices.

Appendix C: Proofs

PROOF OF PROPOSITION 1:
Let F1 5 EF[max{x,0}] and F25 EF[max{2x, 0}]. F1 and F2 are respectively the expected gains
and losses of lottery F relative to 0.
Clearly, U 1 w 1 F 0 w 2 $ U 1 w 0 w 2 if and only if 1 1 1 h 2 F 1 $ 1 1 1 hl 2 F 2 . Now U 1 H 1 F 0 G 2 $
U 1 H 0 G 2 is equivalent to

8 [w 1 w9 1 m 1w 1 w9 2 r2] dG 1r2 dH 1w2 dF 1w92 $ 5 [w 1 m 1w 2 r2] dG 1r2 dH 1w2

1066 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2007

or

8 [w9 1 m 1w 1 w9 2 r2 2 m 1w 2 r2] dG 1r2 dH 1w2 dF 1w92 $ 0.

Notice that for any wr $ 0, m 1 w 1 wr 2 r 2 2 m 1 w 2 r 2 $ hwr, and for any wr # 0, m 1 w 1 wr 2 r 2
2 m 1 w 2 r 2 $ hlwr. Hence

8 [w9 1 m 1w 1 w9 2 r2 2 m 1w 2 r2] dG 1r2 dH 1w2 dF 1w92

$ 8 [w9 1 hmax{w9, 0} 1 hlmin{w9, 0}] dG 1r2 dH 1w2 dF 1w92

5 11 1 h 2F12 11 1 hl2 F2 $ 0.

This completes the proof.

PROOF OF PROPOSITION 2:
For each part of the proposition, we prove the first inequality 1U 1 H 1 F 0 G 2 . U 1 H 0 G 2 in part (i),
and U 1 H 1 t # F 0 G 2 . U 1 H 0 G 2 in part (ii)). The other inequality is analogous.
Let F9 be the mean-zero lottery that satisfies F9 1 v 5 F for a constant v . 0.
(i) We prove that for any e1 . 0, there are A, e . 0 such that if PrG 1r [ [k 2 A, k 1 A]2 , e for
all k [ R then for any a [ R,

(7) 5 1m 1a 1 w 2 r2 2 m 1a 2 r2 2 dF91w2 dG 1r2 . 2 e1.

This is sufficient because it implies that U 1 H 1 Fr 0 G 2 2 U 1 H 0 G 2 . 2 e1, and hence

U 1 H 1 F 0 G 2 2 U 1 H 0 G 2 $ U 1 H 1 Fr 0 G 2 2 U 1 H 0 G 2 1 v . v 2 e1 . 0 for e1 sufficiently small.
Since m is differentiable other than at zero and is concave in gains and convex in losses, both
lim xS` m91x 2 and lim xS2` m91x 2 exist. This implies that for any e2 . 0, there is an A such that

h 1b 2 K 2 1m 1b 1 w2 2 m 1b 2 2 dF91w2 . 2 e2

for any 0 b 0 . A. h is also bounded; let its bound be M. Notice that

5 1m 1a 1 w 2 r2 2 m 1a 2 r2 2 dF91w2 dG 1r2 5 2 h 1a 2 r2 dG 1r2.

Since PrG[a 2 r [ [2A, A]] , e and h 1 # 2 . 2 e2 outside this range, the integral above is greater
than 2 eM 2 e2. Therefore, we can choose A, e, and e2 such that if G satisfies the conditions of the
proposition, the integral above is greater than 2 e1.

(ii) Whenever G 1 # 2 is a continuous distribution, the function

h[ K 2 u 1?|r2 dG 1r2

is differentiable everywhere. Hence, for any w [ R,

e 3 h 1 w 1 t # wr 2 2 h 1 w 2 4 dFr 1 wr 2
lim 50
tS` t
VOL. 97 NO. 4 KŐSZEGI AND RABIN: REFERENCE-DEPENDENT RISK ATTITUDES 1067

everywhere. Hence

U 1 H 1 t # Fr|G 2 2 U 1 H|G 2 e 3 h 1 w 1 t # wr 2 2 h 1 w 2 4 dFr 1 wr 2

lim 5 3 ctlim d dH 1 w 2 5 0.
tS` t S`
t

This implies that

U1H 1 t # F0G2 2 U1H0G2

lim . 0,
tS` t
completing the proof.

PROOF OF PROPOSITION 3:
(The proof below establishes the result for both PPE and CPE (Definition 3 in Section IV). The
proof uses the concept of average self-distance defined in Appendix A.)
Let G 5 F 1 k. As in the proof of Proposition 1, let G1 5 EG[max{x,0}] and G2 5 EG[max{2 x,
0}]. Let k{ be defined by G 1 /G 2 5 1 1 1 mr2 1 0 2 2 / 1 1 1 mr1 1 0 2 2 . Clearly, k{ . 0. Then for any k{ , k{
we have G 1 /G 2 , 1 1 1 mr2 1 0 2 2 / 1 1 1 mr1 1 0 2 2 . We prove that for any such G, there is a t{ satisfying
the statement of the proposition.

We have
U 1 w 1 tG 0 w 2 2 U 1 w 0 w 2
lim 5
tS` t

e 3 max 5 0, t wr 6 1 m 1 max 5 0, twr 6 2 4 dG 1 wr 2 1 e 3 min 5 0, t wr 6 1 m 1 min 5 0, twr 6 2 4 dG 1 wr 2

t
5 11 1 m91 102G1 2 11 1 m92 102G2, 0.

Hence, there is a t{ such that for t , t, choosing w is a UPE in the choice set {w, w 1 tG}.
Also:
U1w 1 tG 0w 1 tG 2 U1w 0 w 2 4 m 1 t 1 wr 2 r 2 2 dG 1 wr 2 dG 1 r 2
lim 5 lim w9 dG 1w92 1
tS0 t tS0 2 t

m1 0t 1 wr 2 r 2 0 2 1 m1 2 0t 1 wr 2 r 2 0 2
5 w9 dG 1w92 1 lim dG 1w92 dG 1r2
2 tS0 5 2t

1
5 w9 dG 1w92 1 1m91 102 2 m92 102 2 Z w9 2 rZ dG 1w92 dG 1r2
2 52

1
5 w9 dG 1w92 2 1m92 102 2 m91 102 2S 1G2
2 2

# G12 G22 1m92 102 2 m91 102 2 1G11 G22

, 11 1 m91 102 2G1 2 11 1 m92 102 2G2, 0,

1068 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2007

where the second-to-last inequality is true because G 1 $ G 2 . This establishes that w is both a PPE
and CPE.

PROOF OF PROPOSITION 4:
(The proof below establishes the result for both PPE and CPE (Definition 3 in Section IV).)
By the proof of Proposition 8, if w is a UPE, then U 1 w 0 w 2 . U 1 w 1 F 0 w 1 F 2 , so that w is the
PPE. This means that if w 1 F is a PPE, then it must be the case that w is not a UPE. This is equiva-
lent to U 1 w 1 F 0 w 2 . U 1 w 0 w 2 . Then, by Proposition 1, the result is immediate.
To establish the result for CPE, suppose w 1 F is a CPE. By Proposition 8, w is not a PPE. Then,
by the same logic as above, it must be the case that w is not a UPE. From here, the proof is the same
as above.

PROOF OF PROPOSITION 5:
(The proof below establishes the result for both PPE and CPE (Definition 3 in Section IV).)
By a trivial modification of the proof of Proposition 1, U 1 w 1 F 0 w 2 . U 1 w 0 w 2 implies
U 1 G 1 F 0 G 2 . U 1 G 0 G 2 . Hence, U 1 G 1 F 0 G 2 # U 1 G 0 G 2 implies U 1 w 1 F 0 w 2 # U 1 w 0 w 2 .
Therefore, choosing w is a UPE in the choice set {w, w 1 F}. Then, by the proof of Proposition 8,
U 1 w 0 w 2 . U 1 w 1 F 0 w 1 F 2 , so w is the PPE.
We now prove the same statement for CPE. We want to prove that if U 1 G 0 G 2 $ U 1 G 1 F 0 G 1 F 2 ,
then U 1 w 0 w 2 $ U 1 w 1 F 0 w 1 F 2 . Since w just shifts both sides of the latter inequality by a constant,
it is sufficient to prove for w 5 0. We will prove that U 1 F 0 F 2 # U 1 G 1 F 0 G 1 F 2 2 U 1 G 0 G 2 .
Notice that the two sides are equal in consumption utility (which is equal to the expectation of F).
Hence, we prove the inequality for the gain-loss-utility component. We take advantage of a geometric
analogy: for any distribution H, the negative of the gain-loss utility part of U 1 H 0 H 2 is proportional
to the average self-distance of H. Therefore, the statement above is equivalent to the following: when
the distribution F is added to the distribution G, the increase in the average self-distance is lower than
the average self-distance of F. To show this, consider any two realizations a and b of G, and any two
realizations x and y of F. By the triangle inequality,

0 1a 1 x2 2 1b 1 y2 0 # 0a 2 b0 1 0x 2 y0

or

0 1a 1 x2 2 1b 1 y2 0 2 0a 2 b0 # 0x 2 y0,

completing the proof.

PROOF OF PROPOSITION 6:
(The proof below establishes the result for both PPE and CPE (Definition 3 in Section IV).)
As in the proof of Proposition 2, for each part of the proposition we prove the first part of the
statement. Also as in the proof of that proposition, define Fr as the mean-zero lottery such that
Fr 1 v 5 F for a constant v . 0.
(i) As an obvious implication of Proposition 2, there are A, e . 0 such that the unique UPE, and
hence also the PPE, is to choose G 1 F.
The proof for CPE is only slightly more complicated. By the proof of Proposition 2, for any e1 . 0
there are A, e . 0 such that if PrG 1r [ [k 2 A, k 1 A]2 , e for all k [ R, then U 1 G 1 Fr 0 G 2 2 U 1 G 0 G 2
. 2 e1. Applying a similar argument, there are A, e . 0 such that if PrG 1r [ [k 2 A, k 1 A]2 , e for all
k [ R, U 1 G 1 Fr 0 G 1 Fr 2 2 U 1 G 1 Fr 0 G 2 . 2 e1. Hence, there are A,P . 0 such that if PrG 1r [
[2 A, A]2 , e, U 1 G 1 F 0 G 1 F 2 2 U 1 G 0 G 2 5 U 1 G 1 Fr 0 G 1 Fr 2 2 U 1 G 0 G 2 1 v . v 2 2e1 . 0
for a sufficiently small e1.
(ii) By Proposition 2, there is a t . 0 such that for t , t, the unique UPE, and hence also the PPE,
in the choice set {G, G 1 tF} is to choose G 1 tF.
VOL. 97 NO. 4 KŐSZEGI AND RABIN: REFERENCE-DEPENDENT RISK ATTITUDES 1069

We now prove for CPE. By the same argument as in the proof of Proposition 2,

U 1 G 1 t # F 0 G 1 t # Fr 2 2 U 1 G 0 G 2
lim 5 0,
tS0 t
so that

U1G 1 t # F0G 1 t # F2 2 U1G0G2

lim 5 v . 0.
tS0 t

PROOF OF PROPOSITION 7:
For any lottery F,

(8) U 1FZF2 5 E[F] 1 h 5 1m0 1w9 2 r2 dF 1r2 dF 1w92

1
5 E[F] 1 h 1m 1Zw9 2 rZ2 1 m0 12Zw9 2 rZ2 2 dF 1r2 dF 1w92 .
2 5 0

K 2 n(F)

By A2, n 1F2 . 0 for any nondeterministic lottery F. Let x 5 minF[D, F Z w n 1F2 and y 5 maxF[D, F Z wE 3F4 .
If h . 2 1y 2 w2/x, the unique CPE is to choose w.

PROOF OF PROPOSITION 8:
We prove that if F is a UPE, then U 1 F 0 F 2 . U 1 Fr 0 Fr 2 , so that Fr is not a CPE. If Fr 5 F 1 k for
some k, then the result is immediate, since in that case we would otherwise have to have k , 0.
That F is a UPE when Fr is available implies that U 1 F 0 F 2 $ U 1 Fr 0 F 2 . Hence, there is a constant
kr $ 0 such that for Fs 5 Fr 1 kr, we have U 1 F 0 F 2 5 U 1 Fs 0 F 2 . We prove that U 1 F 0 F 2 . U 1 Fs 0 Fs 2 ,
which is sufficient because U 1 Fs 0 Fs 2 $ U 1 Fr 0 Fr 2 .
The condition that U 1 Fs 0 F 2 5 U 1 F 0 F 2 can be written as

(9) 2 w dF0 1w2 1 5 m 1w 2 r2 dF 1r2 dF0 1w2 5 2 w dF 1w2 1 5 m 1w 2 r2 dF 1r2 dF 1w2.

Clearly, we must have 1 w dF 1w2 , 1 w dF0 1w2. Otherwise, because m is strictly increas-
ing and concave and Fs is riskier than F, we would have J m 1w 2 r2 dF 1r2 dF0 1w2 ,
J m 1w 2 r2 dF 1r2 dF 1w2 , a contradiction.
We want to prove that

2 w dF0 1w2 1 5 m 1w 2 r2 dF0 1r2 dF0 1w2 , 2 w dF 1w) 1 5 m 1w 2 r2 dF 1r2 dF 1w2.

Given equation (9), this is equivalent to

5 m 1w 2 r2 dF0 1r2 dF0 1w2 , 5 m 1w 2 r2 dF 1r2 dF0 1w2.

Now, since the mean of Fs is greater than that of F, there is a ks . 0 such that Fs 2 ks and F have
the same mean. Notice that

5 m 1w 2 r2 dF0 1r2 dF0 1w2 , 5 m 1w 2 r2 d 1F0 2 k02 1r2 dF0 1w2 ,

1070 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2007

which in turn is less than or equal to

5 m 1w 2 r2 dF 1r2 dF0 1w2

because Fs 2 ks is a mean-preserving spread of F and m is concave. This completes the proof.

Proof of PROPOSITION 9:
Let M 5 2m 121m 1r2 2 m 1r92 2 2. By properties A2 and A3 of m 1 # 2 , m 1x 1 m 1r2 2 m 1r92 2 2 m 1x 2
# M for any x [ R.
Since lim xSq m91x 2 5 lim xS2q m91x 2 5 0, for any e1 . 0, there is an A such that if 0 x 0 . A, then
mr 1 x 2 , e1. Furthermore, since m 1 # 2 has full range, for all e2 . 0 there is a d . 0 such that if the
density of F is less than d everywhere, PrF [m 1w2 [ [m 1r92 2 A, m 1r2 1 A]] , e2. Denote the interval
[m 1r92 2 A, [m 1r2 2 A] by B. Under these conditions,

U 1FZr92 2 U 1FZr2 5 2 3m 1m 1w2 2 m 1r92 2 2 m 1m 1w2 2 m 1r2 2 4 dF 1w2

5 3 3m 1m 1w2 2 m 1r92 2 2 m 1m 1w2 2 m 1r2 2 4 dF 1w2 1

m 1 w 2 [B

3 3m 1m 1w2 2 m 1r92 2 2 m 1m 1w2 2 m 1r2 2 4 dF 1w2

m 1 w 2 oB

# e2 M 1 e1 1m 1r2 2 m 1r92 2 ,

which is less than e for appropriately chosen e1,e2.

PROOF OF PROPOSITION 10:

Obvious.

PROOF OF LEMMA 1:
Since constant shifts in a distribution clearly leave the average self-distance unchanged, we can
assume without loss of generality that F and Fr have the same mean. Then, using that the absolute-
value function is convex,

5 Zx 2 yZ dF 1x 2 dF 1y 2 # 5 Zx 2 yZ dF91x 2 dF 1y 2

5 5 Zx 2 yZ dF 1y 2 dF91x 2 # 5 Zx 2 yZ dF91y 2 dF91x 2.

Proof of PROPOSITION 11:

Obvious from the proof of Proposition 3.

Proof of PROPOSITION 12:

Part (i) follows from equation (8). Part (ii) is trivial from part (i).

PROOF OF PROPOSITION 13:

Notice that under A3r, the utility function U 1?ZF2 5 eu 1?Zr2 dF 1r2 is weakly concave. Hence,
U 1 F 0 F 2 $ U 1 Fr 0 F 2 , so that choosing F is a UPE.
VOL. 97 NO. 4 KŐSZEGI AND RABIN: REFERENCE-DEPENDENT RISK ATTITUDES 1071

We now prove that U 1 F 0 F 2 . U 1 Fr 0 Fr 2 , both completing the proof that F is a PPE and proving
that it is a CPE. Since the expected consumption utilities cancel, this inequality is equivalent to

5 m 1w 2 r2 dF 1w2 dF 1r2 . 5 m 1w 2 r2 dF91w2 dF91r2 ,

or S 1 F 2 , S 1 Fr 2 . To show this, note that by the convexity of the absolute-value function,

5 Zw 2 rZ dF 1w2 dF 1r2 # 5 Zw 2 rZ dF91w2 dF 1r2 5 5 Zw 2 rZ dF 1r2 dF91w2.

Using the same reasoning and that Fr 2 F, the above is strictly less than

5 Zw 2 rZ dF91r2 dF91w2 5 5 Zw 2 rZ dF91w2 dF91r2.

PROOF OF PROPOSITION 14:

For any reference lottery G, the function

U 1?ZG2 52 u 1?Zr2 dG 1r2

is strictly increasing. Hence, for any two distributions F, F9 such that F first-order stochastically
dominates F9, U 1 F 0 G 2 . U 1 Fr 0 G 2 . This implies that F9 cannot be a UPE when F is available.

REFERENCES Puzzle.” Quarterly Journal of Economics,

110(1): 73–92.
Barberis, Nicholas, Ming Huang, and Tano Santos. Birnbaum, Michael H., Gregory Coffey, Barbara
2001. “Prospect Theory and Asset Prices.” A. Mellers, and Robin Weiss. 1992. “Utility
Quarterly Journal of Economics, 116(1): 1–53. Measurement: Configural-Weight Theory and
Barberis, Nicholas, Ming Huang, and Richard H. the Judge’s Point of View.” Journal of Experi-
Thaler. 2006. “Individual Preferences, Mone- mental Psychology: Human Perception and
tary Gambles, and Stock Market Participation: Performance, 18(2): 331–46.
A Case of Narrow Framing.” American Eco- Bombardini, Matilde, and Francesco Trebbi.
nomic Review, 96(4): 1069–90. 2006. Risk Aversion and Expected Utility
Barberis, Nicholas, and Wei Xiong. 2006. “What Theory: A Field Experiment with Large and
Drives the Disposition Effect? An Analysis of Small Stakes. Unpublished.
a Long-Standing Preference-Based Explana- Bosch-Domènech, Antoni, and Joaquim Silvestre.
tion.” National Bureau of Economic Research 2006. “Averting Risk in the Face of Large
Working Paper 12397. Losses: Bernoulli vs. Tversky and Kahneman.”
Barsky, Robert B., F. Thomas Juster, Miles S. Universitat Pompeu Fabra Economics Working
Kimball, and Matthew D. Shapiro. 1997. Paper 932.
“Preference Parameters and Behavioral Het- Bowman, David, Deborah Minehart, and Matthew
erogeneity: An Experimental Approach in the Rabin. 1999. “Loss Aversion in a Consumption-
Health and Retirement Study.” Quarterly Jour- Savings Model.” Journal of Economic Behav-
nal of Economics, 112(2): 537–79. ior and Organization, 38(2): 155–78.
Bell, David, E. 1985. Disappointment in Deci- Breiter, Hans C., Itzhak Aharon, Daniel Kahneman,
sion Making under Uncertainty. Operations Anders Dale, and Peter Shizgal. 2001. Func-
Research, 33(1): 1–27. tional Imaging of Neural Responses to Expec-
Benartzi, Shlomo, and Richard H. Thaler. 1995. tancy and Experience of Monetary Gains and
“Myopic Loss Aversion and the Equity Premium Losses. Neuron, 30(2): 619–39.
1072 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2007

Caplin, Andrew, and John Leahy. 2001. “Psy- Kőszegi, Botond, and Matthew Rabin. 2006.
chological Expected Utility Theory and “A Model of Reference-Dependent Prefer-
Anticipatory Feelings.” Quarterly Journal of ences.” Quarterly Journal of Economics, 121(4):
Economics, 116(1): 55–79. 1133–65.
Chetty, Raj. 2003. “Consumption Commitments, Kőszegi, Botond, and Matthew Rabin. 2007.
Unemployment Durations, and Local Risk “Reference-Dependent Consumption Plans.”
Aversion.” National Bureau of Economic Unpublished.
Research Working Paper 10211. Larsen, Jeff T., A. Peter McGraw, Barbara A.
Chetty, Raj. 2005. “Labor Supply and Risk Aver- Mellers, and John T. Cacioppo. 2004. “The
sion: A Calibration Theorem.” Unpublished. Agony of Victory and Thrill of Defeat: Mixed
Frederick, Shane, and George Loewenstein. Emotional Reactions to Disappointing Wins
1999. “Hedonic Adaptation.” In Well-Being: and Relieving Losses.” Psychological Science,
The Foundations of Hedonic Psychology, 15(5): 325–30.
ed. Daniel Kahneman, Edward Diener, and Loewenstein, George, Ted O’Donoghue, and
Norbert Schwartz, 302–29, New York: Russell Matthew Rabin. 2003. “Projection Bias in Pre-
Sage Foundation. dicting Future Utility.” Quarterly Journal of
Genesove, David, and Christopher Mayer. 2001. Economics, 118(4): 1209–48.
“Loss Aversion and Seller Behavior: Evidence Loomes, Graham, and Robert Sugden. 1986.
from the Housing Market.” Quarterly Journal “Disappointment and Dynamic Consistency
of Economics, 116(4): 1233–60. in Choice under Uncertainty.” Review of Eco-
Gneezy, Uri, John A. List, and George Wu. 2006. nomic Studies, 53(2): 271–82.
“The Uncertainty Effect: When a Risky Pros- Medvec, Victoria Husted, Scott F. Madey, and
pect Is Valued Less Than Its Worst Possible Thomas Gilovich. 1995. “When Less Is More:
Outcome.” Quarterly Journal of Economics, Counterfactual Thinking and Satisfaction
121(4): 1283–309. among Olympic Medalists.” Journal of
Gomes, Francisco. 1991. “Portfolio Choice and Personality and Social Psychology, 69(4):
Trading Volume with Loss-Averse Investors.” 603–10.
Journal of Business, 78(2): 675–706. Mehra, Rajnish, and Edward C. Prescott. 1985.
Gul, Faruk. 1991. “A Theory of Disappointment “The Equity Premium: A Puzzle.” Journal of
Aversion.” Econometrica, 59(3): 667–86. Monetary Economics, 15(2): 145–61.
Kahneman, Daniel. 2003. “Maps of Bounded Mellers, Barbara, Alan Schwartz, and Ilana Ritov.
Rationality: Psychology for Behavioral Eco- 1999. “Emotion-Based Choice.” Journal of
nomics.” American Economic Review, 93(5) Experimental Psychology: General, 128(3):
1449–75. 332–45.
Kahneman, Daniel, and Dan Lovallo. 1993. “Timid Novemsky, Nathan, and Daniel Kahneman. 2005.
Choices and Bold Forecasts: A Cognitive Per- “The Boundaries of Loss Aversion.” Journal
spective on Risk Taking.” Management Sci- of Marketing Research, 42(2): 119–28.
ence, 39(1): 17–31. Odean, Terrance. 1998. “Are Investors Reluctant
Kahneman, Daniel, and Amos Tversky. 1979. to Realize Their Losses?” Journal of Finance,
“Prospect Theory: An Analysis of Decision 53(5): 1775–98.
under Risk.” Econometrica, 47(2): 263–91. Post, Thierry, Martijn J. Van den Assem, Guido
Kahneman, Daniel, and Amos Tversky. 1984. Baltussen, and Richard H. Thaler. Forthcom-
“Choices, Values, and Frames.” American Psy- ing. “Deal or No Deal? Decision Making
chologist, 39(4): 341–50. Under Risk in a Large-Payoff Game Show.”
Knetsch, Jack L., and J. A. Sinden. 1984. “Willing- American Economic Review.
ness to Pay and Compensation Demanded: Rabin, Matthew. 2000. “Risk Aversion and
Experimental Evidence of an Unexpected Expected-Utility Theory: A Calibration Theo-
Disparity in Measures of Value.” Quarterly rem.” Econometrica, 68(5): 1281–92.
Journal of Economics, 99(3): 507–21. Read, Daniel, George Loewenstein, and Matthew
Kobberling, Veronika and Peter P. Wakker. Rabin. 1999. “Choice Bracketing.” Journal of
2005. “An Index of Loss Aversion.” Journal of Risk and Uncertainty, 19(1–3): 171–97.
Economic Theory, 122(1): 119–31. Schechter, Laura. 2005. “Risk Aversion and
Kőszegi, Botond. 2005. “Utility from Anticipa- Expected-Utility Theory: A Calibration Exer-
tion and Personal Equilibrium.” Unpublished. cise.” Unpublished.
VOL. 97 NO. 4 KŐSZEGI AND RABIN: REFERENCE-DEPENDENT RISK ATTITUDES 1073

Shalev, Jonathan. 2000. “Loss Aversion Equilib- Thaler, Richard H. 2000. “Mental Accounting
rium.” International Journal of Game Theory, Matters.” In Choices, Values, and Frames, ed.
29(2): 269–87. Daniel Kahneman and Amos Tversky, 241–68.
Stone, Rebecca. 2005. “Loss Aversion and Self- Cambridge; New York and Melbourne: Cam-
Control.” Unpublished. bridge University Press; New York: Russell
Sugden, Robert. 2003. “Reference-Dependent Sage Foundation.
Subjective Expected Utility.” Journal of Eco- Thaler, Richard H. and Eric J. Johnson. 1990.
nomic Theory, 111(2): 172–91. “Gambling with the House Money and Trying
Sydnor, Justin. 2006. “Sweating the Small Stuff: to Break Even: The Effects of Prior Outcomes
The Demand for Low Deductibles in Home- on Risky Choice.” Management Science, 36(6):
owners Insurance.” Unpublished. 643–60.