Journal of Experimental Psychology: Copyright 1983 by the

Learning, Memory, and Cognition American Psychological Association, Inc.

1983, Vol. 9, No. 4, 713-717

Risk: The Long and the Short

Amos Tversky Maya Bar-Hillel
Stanford University Hebrew University, Jerusalem, Israel

Normative objections to expected utility theory raised by Lopes (1982) are rebutted,
and a "fallacy of large numbers," discussed by Samuelson (1963), is analyzed from
both mathematical and psychological standpoints.

This note was stimulated by a recent article tossed until tails first comes up, say on trial
by Lopes (1982), "Decision Making in the «. The player then wins $2". This lottery, first
Short Run," which challenges the normative introduced by Nicholas Bernoulli and inves-
adequacy of expected utility theory. In the tigated by his younger cousin Daniel, has in-
present note we address some of the issues trigued many students of probability and de-
raised by Lopes and rebut her main arguments. cision making who sought to explain the ap-
We also propose a new normative treatment parent paradox that people are unwilling to
and a psychological analysis of an interesting pay much for the opportunity of playing a
gambling problem introduced by Samuelson lottery whose expected value is infinite.
(1963) in his article, "Risk and Uncertainty: Lopes reversed the classical question and
A Fallacy of Large Numbers." examined the SPL from the point of view of
Lopes argued that in the short run, at least, the house, not the player. Imagine, she ex-
the probability of winning is a reasonable cri- horted us, a fabulously rich man, called
terion of choice between gambles. However, Scrooge, who sells SPLs for $100 apiece.
the probability of winning cannot serve as a Clearly, in the long run Scrooge is bound to
sole criterion because it leads to intransitivity lose all. However, on the basis of a computer
as well as to absurd choices. For example, a simulation, Lopes concluded that in the course
sure gain of $1 would be preferred to an even of one million transactions, Scrooge has a 90%
chance to win $5,000 or nothing. Lopes ac- chance of being in the black, with an expected
knowledged these difficulties but she did not profit of $56 million. The balance sheet, Lopes
specify what additional criteria should be con- argued, more than justifies Scrooge's venture,
sidered (the probability of loss? the potential despite the risk it entails. "Is Scrooge crazy,
gain? the potential loss?) and how they should then," she asked, "to sell a product for infinitely
be combined. Instead, she discussed three ex- less than it is worth?" (Lopes, 1982, p. 378).
amples involving single and multiple lotteries What is in question, however, is Scrooge's
intended to support her conclusion that the honesty, not his sanity. The trouble with
standard conception of rational choice, based Scrooge's venture is that he cannot guarantee
on the maximization of expected utility, "is to fulfill his obligation to his customers. The
simply not sensible." (p. 385). We address these SPL sets no upper bound on the size of the
examples in turn. prize that a player can win, but Scrooge's abil-
ity to pay is clearly bounded. If the maximal
Bernoulli's Gamble prize that Scrooge can pay a given player is
$M, then the payoff must cease to double and
The St. Petersburg Lottery (SPL) is a game remains constant after the fcth toss, where k =
of chance in which a fair coin is repeatedly log2M. Thus, Scrooge is in fact selling an SPL
truncated at k + 1 steps, whose expected value
This work was supported by Office of Naval Research is k + 1 dollars, not infinity. If, for example,
Grant NR 197-058 to Stanford University. We thank David Scrooge's upper limit is a generous billion dol-
Freedman and Lola Lopes for comments based on an earlier lars, he is actually offering a game whose fair
draft. price is less than $31, because 230 exceeds 109.
Requests for reprints should be sent to Amos Tversky,
Department of Psychology, Building 420, Stanford Uni- When Scrooge's asking price for such an SPL
versity, Stanford, California 94305. is $100, he is selling a product not for "infi-
713
714 AMOS TVERSKY AND MAYA BAR-HILLEL

nitely less than it is worth," but for consid- routinely purchase lotteries that have a similar
erably more. To make it a fair game Scrooge structure at prices that exceed their expected
should be able to pay $2", which far exceeds value. Ironically, then, the real challenge to
the entire wealth of the world. If one is con- expected utility theory and the risk aversion
cerned with real-life problems, rather than hypothesis is the purchase of lotteries with a
with mathematical puzzles, one cannot ignore long positive tail at unfavorable prices, not the
the inevitable truncation of the game and treat reluctance to purchase such lotteries at fa-
it as if its expected value were infinite. vorable prices.
By failing to admit that he is selling a trun-
cated SPL and to specify his upper limit, Weaver's Objection
Scrooge is being less than honest—much like In his book "Lady Luck," Weaver (1963)
an insurance company that collects premiums remarked that one may not be willing to pay
for potential damages it could not cover in the the expected value in order to play a particular
hope that the worst will not happen. The vi- gamble because "the odds are about 4 to 6
ability of such business practices depends on that he will receive no prize at all, and just
the gullibility of people and the laws of the throw away his investment" (p. 155). Lopes
land; it does not have much bearing on the interpreted this comment as "rejection of the
adequacy of expected utility theory. expected utility principle" (p.;381) and a vi-
Once we eliminate the deceptive aspect of olation of the von Neumann aqd Morgenstern
Scrooge's offer and truncate the infinite tail axioms. Evidently, she inferred from Weaver's
of the game, what remains of Lopes's argu- observation regarding a particular gamble that
ment? Let us assume, for a moment, that he advocates the general principle of maxi-
Scrooge can actually finance a 55-step SPL. mizing the probability of winning, which is
Lopes's simulation shows that Scrooge can ex- inconsistent with expected utility theory. In
pect to make a relatively small amount of fact, however, Weaver did not address the ex-
money (less than one billionth of his total as- pected utility principle in this context and did
sets) by selling one million SPLs at $50 not advocate the rule of always selecting the
apiece—if he is willing to bear the risk of a gamble with the highest probability of winning.
catastrophic loss. Two comments regarding this Hence, his objection to the expected value
observation may be in order. First, the selling model does not constitute an argument against
of SPLs below their expected value is com- expected utility theory.
patible with the expected utility principle and Lopes also raised the standard objection that
a convex utility function for money. Second, arguments based on expectations are appli-
very rich people do not seem eager to finance cable in the long run but not in the short run.
actuarially unfavorable ventures with a po- This objection, however, does not apply to ex-
tential catastrophic loss. Lopes's example, pected utility theory that is derived from ax-
therefore, does not provide either a hypothet- ioms of rational choice that pertain to unique
ical argument against the maximization of ex- choice situations with no reference to repeated
pected utility or a factual argument against plays. Lopes (1982) claimed that the absence
the maximization of expected value. of long run considerations from expected util-
It might be noted that many scholars, from ity theory "is more apparent than real" be-
Poisson to contemporary authors, have argued cause "it is questionable whether probability
that the limit on the house's ability to pay and values every really combine except in the
dissolves much of the paradoxical character long run" (p. 381). But modern utility theory,
of Bernoulli's original problem. As was noted as conceived by von Neumann and Morgen-
by these authors (see, e.g., Shapley, 1977, and stern, does not assume that value and prob-
references therein), a truncated SPL does not ability "really combine." It assumes only that
provide strong evidence against the expected the cash equivalent of a lottery depends on
value criteripn. It does not seem absurd to both the value of the prize and the probability
pay $.31 for a 30-step penny version of the of getting it. Expected utility theory does not
SPL. This game offers a 6.25% chance to come even assume that receiving the lottery or its
up ahead and some chance of winning more cash equivalent have the samejmpact on sub-
than $10 million. Indeed, millions of people sequent choices, as will be illustrated below.
 LONG AND SHORT RISK 715

Samuelson's Theorem
400
Samuelson (1963) told of a distinguished
scholar, unskilled in mathematics, to whom 200
he offered an even chance to win $200 or lose
$ 100, depending on the toss of a coin. Samuel-
0
son's colleague, whom we call SC, declined
the single bet but expressed a willingness to
> -200
play it 100 times. This pattern of preferences
Ij
has some intuitive appeal, but Samuelson
found it normatively unacceptable, noting that 3 -400
multiple play compounds risk rather than re-
duces it. Lopes, on the other hand, justified -600
SC's preferences and regarded them as evi-
dence against the normative adequacy of ex- -800
pected utility theory.
According to Lopes (1982), "Samuelson -1000 -
proves the theorem that no one who wants to
maximize expected utility— which is a goal z-600 z-400 z-200 z z+200 z+400 z»600
WEALTH
his colleague claimed— can agree to a sequence
of bets if each of the single bets is unaccept- Figure 1. A utility function for Samuelson's colleague
able" (p. 382). If SC's behavior is justifiable, (e = .93).
Lopes argued, then expected utility theory
must be wrong. As we show next, however,
this is not the case. There are many utility
functions for wealth that reject the single gam- Contrary to claim, then, SC's preferences
ble and accept the multiple gamble. For ex- are consistent with expected utility theory. A
careful reading of Samuelson's article reveals
amples, define
that the proposition he established was that a
(x -- zf,
f (x if* ^ z person should not accept the multiple game
Hz- if the single game is unacceptable at every asset
position throughout the relevant range of out-,
For some 0 < 0 < 1. Figure 1 displays the comes. Although Samuelson's discussion of
proposed function for 6 = .93. this assumption is informal and brief, he spe-
This function is concave everywhere because cifically cautioned, "I should warn against un-
every cord joining two points on the curve lies due extrapolation of my theorem. It does not
below the curve. Hence, it is risk averse, and say one must always refuse a sequence if one
the degree of risk aversion, r = —u"/u', de- refuses a single venture" (1963, p. 5). For the
creases with the distance from z. A person entire sequence to be unacceptable, therefore,
with this utility function will always prefer a it is not sufficient that each gamble in the
sure thing to a gamble with the same expected sequence is unacceptable at one's present
value. Furthermore, at asset position z, this wealth; each gamble must also be unacceptable
person will decline the single game (equal at any possible level of wealth that can be
chances to win $200 or lose $100) but accept achieved by playing the multiple game. The
two (or more) plays of the game. To verify, utility function of Figure 1, for example, sat-
note that isfies the former but not the latter assumption.
>/2M(z + 200) + '/2M(z - 100) = '/2(200-93) The latter assumption is, indeed, the heart
of the matter. If it holds, Samuelson's conclu-
-'/2(100'-075) = -1.61 <0 = w(z), sion follows from transivity and dominance,
but without assuming expected, utility theory, as
'/4H(z + 400) + '/2u(z + 100) + >/4M(z - 200) we show later. In the next section, however,
= >/4(400-93) + >/2(10093) - 1/4(2001-075) we present illustrative data that cast serious
doubt on the descriptive adequacy of the pre-
= 27.56 > 0 = «(z). vious assumption.
716 AMOS TVERSKY AND MAYA BAR-HILLEL

The present discussion is confined to gam- pected utility theory. However, in conjunction
bles or prospects with monetary outcomes, with Assumption 2, it implies that for every
represented by discrete random variables. As 1 <, k <: 100, Sk must be chosen over S/t+i.
in the standard analysis, we assume that an To prove this proposition, note that for any
individual has a transitive preference, denoted outcome s of Sk, w + s > w + s + Xk+i, by
>, between any pair of (final) asset positions. Assumption 2 and the fact that Xi and Xk+i
Thus, gamble X is chosen over Y by a person have the same distribution. Hence, by As-
with wealth w whenever w + X >- w + Y. We sumption 3,
use lowercase letters, w, x, y, z, to denote mon-
etary values (i.e., constants) and uppercase let- w w k+l w k+i,
ters to denote gambles (i.e., random variables). and by transitivity,
In particular, y is said to be an outcome of Y,
if there is a positive probability that the gamble w> w + Si w
Y results in the monetary outcome y. We as- Given Assumptions 2 and 3, therefore, the
sume that the preference order between asset rejection of S{ implies the rejection of SIQO-
positions depends only on their distributions, Our proof is similar, but not identical, to Sam-
hence, one is indifferent between asset posi- uelson's informal argument. Unlike Samuel-
tions that have the same (marginal) distribu- son's proof, however, the present result does
tion. not rely on expected utility theory. It strength-
Consider a set of gambles played by tossing ens Samuelson's case against his colleague, by
n coins and having the player receive $200 for showing that if Assumption 2 holds, then SC's
each heads and pay $100 for each tails. Let choices must violate transitivity or dominance.
Xk be the gamble on the /rth coin, 1 <, k <, n, For Lopes, who defended SC, the situation is
and let Sk = X\ +. .. + Xk denote the multiple more difficult: If she endorses Assumption 2,
gamble that consists of tossing coins 1 through she has to defend the violation of transitivity
k—in particular, Si = Xi and Sk+i = or dominance; if she does not endorse As-
Sk + X/c-u. Note that a multiple gamble is rep- sumption 2, her argument against expected
resented as a sum of random variables. utility theory is invalid.
Because SC rejects a single toss, we assume We do not wish to argue that expected utility
w > w + X, (1) theory is the only adequate normative frame-
work for decision under risk; there are many
Furthermore, the single toss is unacceptable prescriptive aspects of value and belief that
at any level of wealth that can be achieved by are not readily captured in this framework.
100 tosses, hence we stipulate, with Samuelson, We merely argue that Scrooge's enterprise,
+ Xl for Weaver's objection, and SC's preferences do
not cast serious doubt on the normative ad-
-10,000 <;>> ; 20,000. (2) equacy of expected utility theory.
We also assume the following dominance con-
dition: Psychological Analysis
If X and Y are independent and if for every Let us turn now from normative theory to
psychological analysis. To begin with, it is
outcome y o f Y , w + y>w + y + X, hardly surprising that the 100-fold gamble SIQO
then w + Y > w + Y + X. (3) is attractive: It offers a good possibility for a
significant gain while the chances of a loss are
This condition says that if X is unacceptable less than 1%. It is also not surprising that Sioo
at any level of wealth w + y that might result is preferred to the single gamble Si ; the former
from playing the gamble Y, then X is also appears less risky than the latter in the sense
unacceptable at w + Y where one does not that 5*100 has a mean of $5,000 and a standard
know for sure which of these levels of wealth deviation of $1,500, while Si has a mean of
will obtain. $50 and a standard deviation of $150. What
Assumption 3 seems unobjectionable on puzzled Samuelson and others is the rejection
normative grounds, and it is weaker than ex- of Si despite its favorable expectation by peo-
 LONG AND SHORT RISK 717

pie, like SC, who can surely afford a loss of Si at their current asset position, but most
$100. Evidently, the anticipated psychological subjects accepted it when combined with a
impact of the loss offsets the impact of the gain of $600 or with a loss of $200.
equiprobable larger gain. These data indicate that many individuals,
Following prospect theory (Kahneman & who surely vary in asset positions, reject the
Tversky, 1979) and the analysis of framing gamble in A but accept it in B and C. These
(Tversky & Kahneman, 1981), we propose that observations violate Assumption 2 and thereby
this phenomenon of loss aversion is induced undermine the application of Samuelson's
by framing the choice so that the status quo theorem to the problem under study. It appears
serves as a reference point and the outcomes that the common rejection of the gamble in
are evaluated as a gain of 200 and a loss of A and its almost unanimous acceptance in C
100, not as asset positions of w + 200 and represent a stable pattern of choice that is un-
w - 100. Indeed, if the reference point is affected by small changes (e.g., $200) in one's
changed by framing the outcomes in terms of total wealth. In contrast, the addition of $200
asset positions (as recommended by decision to all outcomes of the offered prospects had
analysts) or by adding a (positive or negative) a marked impact on preferences, as demon-
constant to all outcomes, the extreme aversion strated above. These observations reject a util-
to risk is reduced or eliminated. To illustrate ity function that is based on final asset position
this effect and to demonstrate its relevance to (of the type presented in Figure 1) in favor of
our problem, we presented 230 Stanford un- a theory in which the carriers of value are
dergraduates with a brief questionnaire that gains and losses denned relative to some ref-
included the following problems. erence point. Because the value function tends
A. Suppose you are offered the following option. Would to be concave for gains and convex for losses,
you accept the gamble? the shift of reference point can produce sys-
• 50% chance to win $200 and 50% chance to lose tematic reversals of preferences (Kahneman
$100. & Tversky, 1979; Tversky & Kahneman,
B. Suppose you are offered the following choice.
Which option do you prefer?
1981). In particular, one is less likely to accept
• a sure gain of $600. the single gamble when it is framed as an even
• 50% chance to win $800 and 50% chance to win chance to win 200 or lose 100 than when it
$500. is framed as a choice between w and an even
C. Suppose you are forced to choose between the chance at w + 200 or w - 100. Thus, the
following options. Which do you prefer?
• a sure loss of $200. common aversion to fairly small favorable bets
• 50% chance to lose $300 and 50% chance to lose may be, in part at least, a framing effect.
nothing.
Note that problems B and C are obtained References
from problem A by adding $600 to all out- Kahneman, D., & Tversky, A. Prospect theory: An analysis
comes or subtracting $200 from all outcomes, of decision under risk. Econometrica, 1979, 47, 263-
respectively. Because a gain of $600 or a loss 291.
Lopes, L. Decision making in the short run. Journal of
of $200 might occur in the course of playing Experimental Psychology: Human Learning and Mem-
the multiple gamble, Assumption 2 says that ory, 1982, 7, 377-385.
a person who rejects the gamble in problem Samuelson, P. A. Risk and uncertainty: A fallacy of large
A must also reject the gamble in problems B numbers. Scientia, 1963, 98, 108-113.
Shapley, L. S. The St. Petersburg Paradox: A con game?
and C. Prospect theory, on the other hand, Journal of Economic Theory, 1977, 14, 439-442.
implies that the tendency to select the gamble Tversky, A., & Kahneman, D. The framing of decisions
will be weak in A, intermediate in B, and and the psychology of choice. Science, 1981, 21J, 453-
strong in C. The results support this prediction: 458.
-70% of the respondents rejected the gamble Weaver, W. Lady luck. New York: Anchor Books, 1963.
in A and among these subjects, 60% accepted
the gamble in B, and 95% accepted the gamble Received September 20, 1982
in C. Thus, the majority of subjects rejected Revision received March 23, 1983 •

View publication stats