The Probability Weighting Function

Author(s): Drazen Prelec

Source: Econometrica , May, 1998, Vol. 66, No. 3 (May, 1998), pp. 497-527
Published by: The Econometric Society

Stable URL: https://www.jstor.org/stable/2998573

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

The Econometric Society is collaborating with JSTOR to digitize, preserve and extend access to
Econometrica

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 Econioniet[rica, Vol. 66, No. 3 (May, 1998), 497-527

THE PROBABILITY WEIGHTING FUNCTION

BY DRAZEN PRELEC 1

A probability weighting finction w(p) is a prominent feature of several non-expected

utility theories, including prospect theoiy and rank-dependent models. Empirical esti-
mates indicate that w(p) is regressive (first w(p)>p, then iv(p)<p), s-shaped (first
concave, then convex), and asymmetrical (intersecting the diagonal at about 1/3). The
paper states axioms for several w(p) forms, including the compound invariant, w(p)=
exp{ - { - ln p}1, 0 < a < 1, which is regressive, s-shaped, and with an invariant fixed
point and inflection point at I/e = .37.

KEYWORDS: Expected utility theory, non-expected utility theory, prospect theory, Allais
paradox.

1. INTRODUCTION

THE COMMON-RATIO EFFECT (Allais (1953)) refers to the obselvation that the
more risky of two simple prospects becomes relatively more attractive when the
probability of winning is reduced by equal proportion in both prospects. Thus a
person who prefers a sure gain of $100,000 over a coin toss for $300,000 or
nothing, might also prefer a one-in-a-million lottery ticket for $300,000 over a
two-in-a-million lottery ticket for $100,000. This contradicts expected utility but
not common sense: There is a world of difference between certainty and a 50-50
shot; the difference between one or two chances in a million is negligible. The
example has all the force and simplicity of classical arguments for nonlinear
utility, where the utility interval from, say $10 and $20, is taken as self-evidently
greater than the utility interval from $1,000,010 and $1,000,020. Both arguments
appeal directly to our intuitions about numbers. A major point of difference is
that the money argument involves a constant money intelval and the probability
argument a constant probability ratio. A minor difference is that in the money
domain a constant interval has less impact as the numbers get larger, while in
the probability domain a constant ratio has less impact as the numbers get
smaller.
Putting the two demonstrations side by side like this raises an interesting
question, namely, whether an account of nonlinearity-in-probabilities can be
somehow adapted from classical utility theory (modulo a log transformation,
converting statements about probability ratios into statements about utility
intervals). Indeed, in their seminal paper on prospect theory, Kahneman and
Tversky (1979) explained the common-ratio effect by means of a nonlinear
transformation of probabilities into "decision weights," p -* w(p), with log(w

lI would like to acknowledge the invaluable critical effort of Peter Wakker, as well as the m
suggestions by anonymous referees at Econometrica, R. Duncan Luce, George Wu, and the late
Amos Tversky, to whom this paper is dedicated.

497

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 498 DRAZEN PRELEC

0.9

0.8

0.7

0.6 - X eq 3.5, gains data (T&K, 1992)

w(p) 0.5 eq 3.5, loss data (T&K, 1992)
------- eq 3.6, (T&F

0.3 - ~ oK ----- eq 3.5, (W&G, 1996)

compound invariance, eq. 3.1

0 0.2 0.4 0.6 0.8 1

FIGURE 1. The compound invariant form (solid line) and several empirical probability weighting
functions. Estimates of the one-parameter equation (3.5) are taken from Tversky and Kahneman
(1992) and Wu and Gonzalez (1996a); estimates of the two-parameter equation (3.6) are taken from
Tversky and Fox (1994).

convex in log(p). They called such functions subproportional. However, they also
listed other compelling expected utility violations that required additional,
apparently unrelated restrictions on w(p).
My objective in this paper is to derive the observed properties of probability
weighting from preference axioms, relying on the common-ratio effect as the
basic building block. A fair summary of current knowledge about w(p) is
displayed in Figure 1, where the dotted and dashed lines reproduce several
recent empirical estimates. Unlike utility functions, which are characterized by a
single property concavity, here we encounter functions that are:
regressive intersecting the diagonal from above,
asymmetric with fixed point at about 1/3,
s-shaped concave on an initial interval and convex beyond that,
reflective assigning equal weight to a given loss-probability as to a given
gain-probability.
Each of these four properties has a distinctive impact on risk attitudes, and, to
complicate matters, each one is independent of subproportionality. The first
property, regressiveness, generates the important "four-fold pattern of risk
attitudes, which is risk-seeking for small-probability gains and large-probability
losses, and risk-aversion for small-probability losses and large-probability gains
(Tversky and Kahneman (1992)). This is because the overweighting of small
probabilities, below the fixed point of the regressive function, enhances the
attraction of small-p gains (e.g., lottery tickets) and the aversion to small-p
losses (e.g., rare accidents), while the underweighting of larger probabilities,
above the fixed point, diminishes the attraction of larger-p gains and the

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 PROBABILITY WEIGHTING FUNCTION 499

aversion to larger-p losses. The asymmetrical depression of the fixed point below
p = .5 further reduces the weight of uncertain relative to certain outcomes, and
so tends to increase risk-aversion for gains and (by reflection) risk-seeking for
losses. The s-shape property "fits naturally" with regressiveness, but is logically
independent one could have a function that is regressive but not s-shaped,
and vice versa. An s-shape implies that changes in probability have less impact
as one moves away from the boundary of the probability interval, making, for
instance, an initial sweepstakes ticket seem more valuable than any additional
ones.2
Proposition 1, given in Section 3 of this paper, axiomatizes a subproportional
function, w(p) = exp( - ( - ln p) ), 0 < a < 1, that satisfies all four target prop-
erties, and that has an invariant fixed point and inflection point at p = 1/e = .37.
This form, shown by the solid line in Figure 1, is the weighting function
counterpart to the power utility function, u(x) = x a. I derive also a more general
form, w(p) = exp(-,3(- ln p)a), which is not constrained to the l/e fixed point
value. Such "compound invariant" functions are simple to estimate, and provide
a consistent ordering of individuals across several categories of expected-utility
violations (because a smaller a-parameter value in: exp( -, ( - ln p)a), impli
more subproportional, more regressive, and a more s-shaped function). In
Section 4.1, I show how Pratt-Arrow analysis can be used to generate other
functional families. Section 5 discusses the complementary work by Tversky and
Wakker (1995) and Wu and Gonzalez (1996b) on common-consequence related
properties of the weighting function.

2. THEORETICAL FRAMEWORK

2.1. Cumulated Prospect Theo;y: Notation and Assumptions

The results to follow are developed in context of a sign- and rank-dependent

representation of preferences (Starmer and Sugden (1989), Luce (1991), Luce
and Fishburn (1991), Tversky and Kahneman (1992), Wakker and Tversky
(1993)), defined over lotteries on a real interval. Specifically, I follow the
notation and assumptions of Cumulated Prospect Theory or CPT, specialized for
the domain of risk (Tversky and Kahneman (1991), Wakker and Tversky (1993)).
Outcomes are elements, x, y, z,..., of a real interval, X = [x-, x+ ], containing
positive and negative values, x- < 0 < x+. The preference relation > is defined
over the set, -9, of probability distributions, P, Q,..., on X. + and ?- are
the sets of distributions concentrated on the nonnegative and nonpositive parts
of X. V(P) represents > if: P > Q < V(P) ? V(Q), for any P, Q in 39. Prospects
are distributions with finite support. CPT preferences are represented by a sign-
and rank-dependent functional, V(P), with a value function v(x) for money

2 The four-fold pattern of risk attitudes was described and documented by Tversky and Kah
man (1992). Preference conditions equivalent to concavity/convexity have been axiomatized by
Wakker (1994), Prelec (1995), and Wu and Gonzalez (1996b), and those equivalent to reflection by
Tversky and Wakker (1995).

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 500 DRAZEN PRELEC

outcomes, and two probability weighting functions, w+( p) and w-(p), for gains
and losses, respectively (equation (2.0), Appendix 1). Throughout the paper we
assume the following axioms, taken with minor modifications from Wakker
(1994):

ASSUMPTION AO: > satisfies axioms W1-W6 in Appendix 1, which support a

sign- and rank-dependent representation (equation (2.0)) with a ratio scale v(x),
continuous and strictly increasing, and unique w-( p), w+( p), continuous on (0, 1),
strictly increasing, and satisfying w+(0) = w-(O) = 0, w+(1) = w-(1) = 1.

Only three types of prospects matter in this paper: Certain prospects, (x, 1),
abbreviated as (x); simple prospects with one nonzero outcome (0, 1 -p; x, p),
abbreviated as (x, p); binary prospects with two nonzero outcomes of the same
sign (0, 1 -p - q; x, p; y, q), 0 < x <Ly, abbreviated as (x, p; y, q). Hence the
only parts of the representation that we need to be concerned with are

(2.1)-w
(2.1) V(((p)) x' {w(p)u
p) w, (PU(X),
(x), x < x
0, > 0,

w+ (p + q)v(x) + w+ (q)(v(y) -v(x),

(2.2) x,py,q) 0 <x <y,0

w- (p + I)v(x) + w- (q)(v(y) - VW))
y <x < O.

Equation (2.2) shows the accounting assumptions of sign- and rank-dependence.

The binary prospect (x, p; y, q) may be interpreted as "a 'p + q' chance of
gaining [losing] at least v(x) and a 'q' chance of gaining [losing] an extra
v(y) - v(x)." The extension of this idea to more complex prospects is then
straightforward (equation (2.0) in Appendix 1). A pure rank-dependent model
arises if: w-(p) = 1 - w+(1 -p).
Because the separable equation for simple prospects (2.1) is common to
several models of risky choice (e.g., Kahneman and Tversky (1979), Quiggin
(1982), Yaari (1987), Rubinstein (1988), Luce (1988, 1991)), those results that
involve only simple prospects will be highlighted by the following assumption:

ASSUMPTION Al: The restriction of > to simple prospects has a separable

representation (2.1) with v(x), w-(p), and w+(p) satisfying the conditions in
Assumption AO.

I will omit the superscripts where the distinction between losses and gains is
not at issue.

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 PROBABILITY WEIGHTING FUNCTION 501

2.2. Supplementa;y Axioms

The mandatory axioms in AO will be supplemented at various points with one

or more optional conditions, which can be selectively "turned on" to further
constrain a particular model. The first of these conditions is boundary continu-
ity:

BOUNDARY CONTINUITY: For any y > x > 0 or 0 > x > y there exist p, q E (0, 1)
such that (x) > (y, p) and (x) < (y, q).

This axiom, adapted from Wakker (1994), combines four conditions that
ensure continuity at p = 0 and p = 1 for gains and losses.
The next condition states that common-ratio violations will be observed at all
probabilities.

SUBPROPORTIONALITY: For any 0 < A < 1, p 0 q, (x, p) (y, q) implies:

(y, Aq) > (x, Ap) if y > x > O, and (y, Aq) < (x, Ap) if 0 > x > y.

The term "subproportionality" abbreviates the more correct "strict subpro-

portionality in probabilities on every nondegenerate interval."3 Weak subpropor-
tionality holds if the preferences in the definitions are weak. The function w(p)
is [weakly] subproportional if p > q, 0 < A < 1, implies: w(p)w(Aq) >
[? ]w(q)w(Ap).
The third and final optional axiom is new to the literature, and has the effect
of rendering a subproportional w(p) diagonally concave, which is to say, concave
where it is overweighting and convex where it is underweighting. Empirically
estimated weighting functions have approximately satisfied this requirement.
The relevant preference axiom involves mixtures of a binary prospect
(x, p; y, q) with the sure-thing middle outcome (x), in cases where this outcome
also happens to be the certainty equivalent for the binary prospect, (x)
(x, p; y, q). Informally, diagonal concavity of the weighting function precludes
preferences where a person might like (i.e., dislike) mixtures of (x, p; y, q) with
(x), but at the same time dislike (i.e., like) local mixtures of (x, p; y, q) with
other binary prospects near (x, p; y, q).
To state the axiom requires some more notation. Let A'[s, r] = {(x, p; y, q)JO
<x<y, s<q, p+q<r} and A-[s,r]={(x,p;y,q)ly<x<0, s<q, p+q<r},
i.e., the sets of binary prospects where the probability of the extreme outcome is
at least s and the probability of the zero outcome at least 1 - r. Preferences are
quasiconvex [resp. quasiconcave] on A[s, r] if P Q implies: Q ? >P + (1 -
1&)Q [resp. Q ? tP + (1 - t)Q] for any three P, Q, and tP + (1 - I&Q in
MA[s,r], and similarly for A-[s, r]. Preferences are certainty-equivalent (or
CE)-quasiconvex [resp. CE-quasiconcave] on A+[s, r] if P Q implies: Q ? >P

3From a formal standpoint, this definition is unecessarily strong. Most results go through
weak subproportionality and at least one strict common ratio violation.

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 502 DRAZEN PRELEC

+ (1 - ,u)Q [resp. Q ? ,uP + (1 - ,)Q] for any certain outcome Q, and for any
P and ,uP + (1 - ,)Q in A[s, r], and similarly for A-[s, r]. (Note that Q 4
A+[s, r], A-[s, r], unless s = 0, r = 1. In that case, however, CE-quasiconvexity is
subsumed by quasiconvexity. Strict means strict preference unless P = Q.
Proposition 6, proved in Appendix 3, states that the following preference
axiom is equivalent to the diagonal concavity of the probability weighting
function:

DIAGONAL CONCAVITY: There is no nondegenerate interval [s, r] such that > is

quasiconvex and strictly CE-quasiconcave on A + [ s, r ] or A - [ s, r ], nor quasicon-
cave and strictly CE-quasiconvex on A + [ s, r ] or - [ s, r ].

What might constitute a violation of the diagonal concavity axiom? Here is an

example. Consider a prospect (lOp, $100K) consisting of 10 lottery tickets
(p-chances) for a $100K sweepstakes, with each ticket tradable for a 10%
chance at a "consolation prize" of $50. Imagine a person who is:
(i) not willing to exchange one ticket for a 10% chance of $50.
(ii) willing to exchange two tickets for a 20% chance of $50.
(iii) not willing to exchange all 10 tickets for $50.
(i) and (ii) indicate quasiconvexity, or probabilistic risk-aversion, as
(.2, $50; 8p, $100K)> (lOp, $100K)> (.1, $50; 9p, $100K). (ii) and (iii) indicate
CE-quasiconcavity, as (iii) implies that CE > $50, and so from (ii) we have:
(.2,CE; 8p,$lOOK)> (.2,50;8p,$lOOK)> (lOp, $IOOOK)> (CE). Although intu-
itions about probabilistic risk may not be robust, nevertheless the preference
pattern (i)-(iii) is faintly implausible. A person who is willing to trade some but
not all of the lottery tickets for a better chance at a "consolation prize,"
should, in some sense, derive a greater marginal benefit from trading the first
rather than the second ticket.

3. COMPOUND INVARIANCE

3.1. Compound Invariance

Expressed in the indifference mode, an example of a common-ratio violation

is a pair of judgments like the following:

(Schema 1) ($10,000 for sure) (1/2 chance of $30,000),

(1/2 chance of $10,000) (1/6 chance of $30,000),

which show that certainty is to a coin toss as a coin toss is to a die roll, or,
w(1)/w(1/2) = w(1/2)/w(1/6). What might this pattern indicate about other
judgments?
In this section I consider the implications of a particularly simple extrapola-
tion of Schema 1. The idea is to require that common ratio violations be
preserved under probability compounding, as might take place in situations

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 PROBABILITY WEIGHTING FUNCTION 503

where ultimate success requires one or more independent successes in a row.

Suppose that, starting with Schema 1, we could find a dollar value x that
generates a second schema:

(Schema 2) ($x for sure) (1/4 chance of $30,000),

(1/4 chance of $x) (1/36 chance of $30,000).

The person in Schema 1 is indifferent between $10,00 and tossing a coin for
$30,000, and is also indifferent between tossing a coin for $10,000 and rolling a
die for $30,000. The assumption states that requiring two (or N) successive wins,
either on coin toss or die roll, will preserve the common ratio violation
structure, after adjusting the dollar outcomes. (This would be true for expected
utility preferences, except that the die in the example would have to have four
sides, rather than six). More generally, we have the following definition.

DEFINITION 1: > exhibits compound invariance if for any outcomes x, y, x', y'
E X, probabilities q, p, r, s E [0, 1], and compounding integer N ? 1:
If (x, p) - (y, q) and (x, r) (y, s), then (x', pN) (y', qN) implies (x', r.N)
(y', sN).

In combination with the assumptions of Section 2, compound invariance yields

a weighting function that meets the requirements stated in the introduction to
the paper.

PROPOSITION 1: (A) Let > be a preference relation on SD satisfying AO,

diagonal concavity, subproportionality and compound invariance. Then the weight-
ing of probabilities in the representation of ; is characterized by a unique value
a, 0 < a < 1, such that for allp, 0 <p < 1:

(3.1) w+ (p) = w-(p) = exp{-(-ln p) } .

(B) If > satisfies AG or Al, subproportionality and compound invariance, then

the weighting functions are characterized by a, 0 < a < 1, and /, ,- > 0:

(3.2) w (p) = exp{-t3? (-lnp)a},

w - (p) = exp{ -,l- (-ln p)a}.

(C) If > satisfies AO or Al, and compound invariance, then the weighting
futnctions are as in (3.2) with a,( + /- > 0. The special case, a = 1, yields the
power functions:

(3.3) w+(p) =p w-(p) =p;3

(D) If > satisfies A0 or A1, and if compound invariance holds for probabilities
in the open interval (0, 1), then the weighting functions for p < 1 are characterized

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 504 DRAZEN PRELEC

by: a,,8+, 3-> 0, and 0 < y+,y-< 1:

w+(p)= y+ exp{-/3+(lnp)a},

w-(p)= y- exp{-/ i(-lnp)a}.

I will refer to (3.1)-(3.4) as the compound invariant or CI-family of functions.

Because the limit of any CI form as p -> 0, is zero, the functions (3.1)-(3.4) are
all continuous on [0,1), and continuously differentiable on (0,1). If compound
invariance holds for all probabilities, as it does on (3.1)-(3.3), the continuity
extends to the entire interval [0,1]. Boundary continuity is not invoked in the
proposition; it is rather a consequence of compound invariance. Note that (3.1)
requires a full sign- and rank-dependent representation (AO); the other forms
only presuppose a separable representation for simple prospects (Al).

3.2. Expected Utility Plus One Parameter

The axioms behind Proposition 1A define a highly constrained theory, adding

one degree of freedom to EU. For any allowed value of the a-parameter,
probability weighting is regressive and s-shaped. The common loss-gain weight-
ing function is intrinsically asymmetric, with a fixed point and inflection point at
p = 1/e = .37. This particular "signature property" of (3.1) is displayed in the
left panel of Figure 2, which traces (3.1) for different a. The derivative of w(p)
at 1/e equals a, hence the a-value is also the numerical lower bound for w'(p).
This also means that we can visually infer a by the slope of the function at the
inflection point. As a -o 1, probability weighting approximates the linear, ex-
pected utility case. As a -o 0, w(p) approximates a step function, flat every-
where except near the endpoints of the probability interval.

1 0.1

0.9 0.09
0.8 0.08

0.7 0.07 io03

0.6 00
w(p) 0.5 0.05 -
0.4 __ _o__1

0.3 0.03 ----------0--i-

0.2 ~~~~~~~~~~~~~~0.02 -- 0
0.1 0.01

0 ~~~~~~~~~~~~0
0 0.2 0.4 0.6 0.8 1 0 -n
p p over different orders of magnitude

FIGURE 2.-The function: w(p) = exp(- (-in p)a), plotted for different a (left panel) and for
small probabilities with a = .65 (right panel).

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 PROBABILITY WEIGHTING FUNCTION 505

The a-parameter remains a comprehensive index of nonlinear weighting in

the more general CI form, equation (3.2). If two functions in (3.2) are parame-
terized by (ai, /3) and if a1 < a2, it then follows that w, is: (a) more subpr
tional, (b) more s-shaped, and (c) more regressive than w2, in the usual sense
that the composition, w1{w- 1}, is itself a subproportional, s-shaped, and regres-
sive function. These comparative statements derive from the fact that
w1{w'1} is also CI with a = a1/a2. The /3-parameter in (3.2) is a "net" index of
convexity in that increasing /3 increases the convexity of the function without
affecting subproportionality. The -y-parameter, which appears only in (3.4), is an
index of the weight of sure outcomes (p = 1) relative to uncertain ones (p < 1),
holding constant subproportionality and convexity on [0,1).
The behavior of a CI function near zero is perhaps its most interesting
feature, both theoretically and for practical applications. The right panel of
Figure 2 traces out the CI form (a = .65, /3 = 1) over four orders of magnitude,
ranging from one chance in a thousand (top segment) to one chance in a million
(bottom segment). The vertical axis in the figure is probability weight; the
horizontal axis is the probability interval for one order of magnitude, [0, 10-'li,
with n = 3, 4, 5, or 6.
In the picture we see how the CI accommodates three intuitions about the
decision impact of small chances. First, the weight of ever smaller chances tends
to zero (continuity). Second, the slope tends to infinity at zero, giving a
qualitative character to the transition from impossibility to possibility. Third, the
graph becomes relatively flatter at smaller probabilities, capturing the feeling
that, e.g., two chances in a million are really no different from one chance in a
million. Formally, one can show that for any p, q, the limit as A-- 0 of
w(Ap)/w(Aq) equals 1, and that probability weights approximate a step function
at zero. At the level of preferences, this means that for very small chances the
money dimension overrides the probability dimension: for any (x, p) and (y, q)
where x <y, there exists a A > 0 such that the prospect with the higher-prize is
preferred, (y, Aq) > (x, Ap).
The picture at the other endpoint is almost, but not quite the same. w(p) is
unboundedly convex at p = 1, by the Pratt-Arrow measure (w"/w'), and the
slope, dw/dp, again tends to infinity. However, probabilities near p = 1 are not
relatively indistinguishable in the same sense as are probabilities near p = 0,
because the limit as A - 0 of: (1 - w(Ap))/(1 - w(Aq)), p 0 q, does not ap-
proach 1.

3.3. Evidence on the Location of the Fixed and Inflection Points

The main theoretical advantage of the one-parameter form (3.1) over the
two-parameter form (3.2) is the stability of the fixed and inflection point across
different levels of nonlinearity. The exact location of this point is perhaps less
important, provided the 1/e value is not too far off the mark. We can make a

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 506 DRAZEN PRELEC

rough assessment of the 1/e prediction using published estimates of functional

forms that do not have an invariant fixed point. Something like this is attempted
in Figure 1, where the solid line traces equation (3.1) (a = .65), against estimates
of

p
(3.5) w(p)= 1/' 8>0,
(p 8 + (I -p) 8 )

(3.6) ap 8+(p) a

The four functions in the figure are remarkably close.4 The fixed points inferred
from Tversky and Kahneman's (1992) estimates of (3.5) are p0 = .34 for gains
and p0 = .38 for losses. The estimate for equation (3.6) is p0 = .30 (Tversky and
Fox (1994)). Wu and Gonzalez (1996a, p. 1686) report p0 =.39 as the pooled
fixed point estimate, with equation (3.5). Camerer and Ho (1994) conducted a
maximum likelihood estimate of (3.5) for nine previously published data sets.
The mean fixed point value using their nine separate estimates is p0 = .35, with
standard error of .04, which is again close to the conjectured value of .37.
These parametric estimates are not conclusive, relying as they do on possibly
incorrect functional forms for w(p) and for money value. Fully nonparametric
evidence on the location of the inflection point (rather than the fixed point) has
been collected in a recent study by Wu and Gonzalez (1996a). They constructed
p-indexed ladders of "common-consequence" choice pairs, such that preference
for the less risky prospect in a pair increases with the local slope of w(p) near
p, and is therefore minimized near the inflection point of w(p). The estimate of
the inflection point is valid independent of assumptions about money value.
Across the five ladders in the study, the point of minimal preference occurred at
p-values of p = .37, p = .40, p = .32, p = .15, and p = .35, leading them to
conclude that the weighting function, "is concave up to p .40 and convex
beyond that probability."

3.4. Application to the Four-fold Pattem of Risk Attitudes

A concise, global characterization of subproportional CI functions is that they

are regressive not only with respect to the diagonal line, but also with respect to
the entire family of increasing power functions. In other words, the subpropor-
tional CI intersects from above every increasing power function, p :, 3 > 0. I
will call such functions strongly regressive.5
This bears directly on risk-attitudes, at least in the elementary setting where
the choice is between a risky simple prospect and a sure thing. Here one

4Equations (3.5) and (3.6) are not subproportional at small probabilities, however.
5Formally, w(p) is strongly regressive if: Vc > 0, 3pC* E (0, 1) such that p E (0, p*j
and p E (pc*, 1) => w(p) <pc.

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 PROBABILITY WEIGHTING FUNCTION 507

typically observes a risk-seeking for small-p gains and larger-p losses, and a
risk-aversion for small-p losses and larger-p gains (i.e., the "four-fold pattern" of
Tversky and Kahneman (1992)). In a sign- and rank-dependent model with a
linear value function, regressive weighting alone would be a sufficient explana-
tion of the pattern. However, if value is not linear, then probability nonlinearity
must somehow dominate money nonlinearity for the pattern to hold. For
example, to explain risk-seeking behavior with respect to longshot lottery tickets,
the overweighting of the small chance of winning would have to dominate the
concavity of the value function, which works against risk-seeking.
Such dominance arises in representations that combine CI probability weight-
ing with power functions for money value:

xf+ for x > O,

v(x) = l - ( -x)'7- for x < O.

The relative certainty equivalent for a simple prospect is then itself a CI

function of probability, as (c) (x,p) implies cor+=x x+ exp{-,3
or6

C
(3.7) - = exp{ (,3+/o+)(- ln p)a}.
x

Risk attitude is established by comparing the certainty equivalent and the

expected value of the prospect, px. Substituting from (3.7), we find that the risky
prospect (x, p) will be preferred to the EV whenever

(3.8) (for gains, x > 0): c >px < exp{-( 83+/f+?)(lnp)a} >,
(for losses, x < 0): c >px < exp{-( t 3-/ff- )(-lnp) } <p.

For gains, the risky prospect will be preferred at probabilities where the CI
expression in (3.8) is overweighting, while for losses it will be preferred where
the CI is underweighting. Because any CI form with a < 1 is regressive,
overweighting small probabilities and underweighting larger ones, this confirms
the four-fold pattern. Notably, it confirms the pattern independent of the
exponent of the value function, i.e., independent of whether the functions are
concave or convex; the value exponents only affect the sizes of the risk-seeking
regions. The nonlinearity of weighting functions overrides the nonlinearity of
the value functions, in this particular case.

6 The combination of equation (3.1) and a power value function can be estimated with OL
because the plot of relatiLe certainty equivalents, c/x, against p is linear in double-log coordinates.
From (3.7), with , = 1, we have

- ln( - ln(c/x)) = In o- + a(-ln( - ln p)).

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 508 DRAZEN PRELEC

3.5. The Compounding Factor and 8-subproportionality

Having considered one axiom, compound invariance, one might ask whether
this is an anomalous special case or whether there are less restrictive axioms,
conveying similar properties on w(p) without forcing a particular parametric
form. A possibility that immediately comes to mind-the class of all subpropor-
tional preferences-is clearly too broad, containing functions that are neither
regressive nor s-shaped.
A more productive generalization takes its cue from a specific numerical
implication of CI. By computation with equation (3.2), one can show that there
exists a compounding factor, p, such that for any probability p,

(x) (y, p) implies (x, p) (y, p P), p = 21/a,

i.e., W(p)2 = W(p P) for p = 21/a. The CI function in Figure 1, with a= .65,
yields a compounding factor value of p = 2.9. In this case, it would take
(approximately) a cubing of probabilities to produce a squaring of probability
weights. Of course p = 2 is the only value consistent with expected utility
preferences. For subproportional preferences, p > 2, but the exact value of p
will depend on the probability in the definitional schema, (x) (y, p) & (x
(y,pP).
The condition I now state requires that there exist a minimal value 8 < 0 such
that at any base probability level, the compounding factor differs from 2 by at
least ?:

DEFINITION 2: We say that > is 8-subproportional if subproportional and if

there exists an 8> 0 such that for all 0 <x <y, and probabilities, 0 <p < 1,
(x) (y,p) implies (x,p) < (y, p2 8), and for all y <x < 0, (x) (y,p) implies
(X, p) > (y p2+ 8)

Compound invariant preferences satisfy the definition for any 2 < 2 + 8 < 21/ a
provided a < 1.

PROPOSITION 2: Let > be a preference relation on _ satisfying AO or Al, and

8-subproportionality. Then the weighting function of probabilities in the represen
tion of > is strongly regressive.

Although Definition 2 makes much weaker demands on the preference

relation than Definition 1, it still implies a weighting function whose global
shape and boundary nonlinearities are qualitatively similar to the CI form. In
particular, the approximation to a step function at zero, shown in Figure 2,
would also be a property of any strongly regressive weighting function.

4. PRATT-ARROW CLASSIFICATION OF SUBPROPORTIONAL FUNCTIONS

In utility theory, the basic functional forms are introduced as hypotheses

about how risk aversion depends on wealth. Thus the exponential and power

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 PROBABILITY WEIGHTING FUNCTION 509

functions for u(x) define the hypotheses that absolute or proportional risk-aver-
sion is invariant with respect to absolute or proportional changes in wealth.
Stating these hypotheses requires, in turn, a precise definition of greater or
lesser risk aversion, one that permits comparison of two utility functions, or of
the same function before and after a given transformation (Pratt (1964)).
The situation with probability weight is similar. A general theory of functional
forms for w(p) should begin with a definition of greater or lesser divergence
from the linear norm. In Section 5, I will discuss proposals by Tversky and
Wakker (1995) and Wu and Gonzalez (1996b) that take the Allais common-con-
sequence effect as the effective measure of nonlinearity. In this section, how-
ever, I look at comparative nonlinearity in the context of common-ratio viola-
tions, i.e., of subproportional weighting functions. For simplicity, I consider only
positive outcome prospects.

4.1. Comparative Subproportionality

Comparing two preference relations, we use the following definition.

DEFINITION 3: >1 is at least as stibproportional as >? on _9 if for any

probabilities p > r ? q > s, and outcomes 0 < x < y, 0 < x' < y', (x, p) 2 (y, q),
(X, r) '2 (y, s), and (x', p) - (y', q) implies (x', r) < 1(Y', s).

Kahneman and Tversky (1979) pointed out that subproportionality implies

convexity of the weighting function in log-log coordinates, which is to say,
convexity of the w-function,7

(4.1) w(ln p) = ln(w(p)).

Indeed, the ordering of preference relations by subproportionality is the same as

the ordering of the w-functions by convexity:

PROPOSITION 3: Assume that >1 and >2 are two preference relations on 9+
satisfying AO or Al and boundaty continuity, with weighting fiunctions, w1(p) and
w2(p), representing ?I and >2 , respectively. Then, conditions A, B, and C are
equivalent:
(A) ?1 is at least as subproportional as >2
(B) w,(w- t{z}) is a weakly subproportional function of z;
(C) w(jy1{z}) is a convexfuinction of z, with wi(ln p) = ln wi(p).

Proposition 3 offers a general recipe for generating subproportional weighting

functions. Taking any increasing, strictly concave "utility function," u(t), t 2 0,
we can reflect u(t) into the negative domain, w(t) = - u( - t), and create a
weighting function, w(p) = exp( w(ln p)), that is subproportional and continuous

7Tversky and Wakker (1995) derive a result similar to 3AB but for subadditivity, rather than
subproportionality.

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 510 DRAZEN PRELEC

on (0, 1).8 As a practical matter, if we were only interested in empirical

estimation we could borrow such forms and compare w(p) = exp( - u( -ln p))
against data. However, in utility theory the forms are justified by intuitions
about how the risk-premium might depend on wealth. In the probability domain,
the formal counterpart to a risk-premium has little significance. To engage our
intuition we have to go directly to the axioms that support particular functional
forms. I illustrate this with three examples.

4.2. Compound Invariance as the "Constant-Relative" Benchmark

The first example, we have already encountered in Section 3. The w-function

for the CI form, equation (3.2), is W(ln p) = -,3(- ln p)a, which has constant-
relative Pratt-Arrow convexity, (- ln p) w" / w' = a. I now restate the compound
invariance axiom as follows. For any preference relation, ?, define the com-
pounded preferences, N, as

(X, p) ;N(y, q) iff (x, pN) ;> (y, qN).

Compound invariance is just the claim that ?N and > are equally subpropor-
tional, in the sense of Definition 3. Actual preferences might systematically
deviate from invariance in one of two directions. If ?N is observed to be more
subproportional than >, i.e., if common-ratio violations are more common after
compounding, then preferences exhibit increasing-relatiue subproportionality (or
IRS). The other possibility would correspond to the decreasing-relative or DRS
class.

4.3. Conditional Invariance

A second Pratt-Arrow benchmark, "constant-absolute" subproportionality,

also flows from a simple property of preference. Let us return to the schema at
the start of Section 3:

(Schema 1) ($10,000 for sure) (1/2 chance of $30,000),

(1/2 chance of $10,000) (1/6 chance of $30,000),

and require now that common ratio violations be preserved under the propor-
tional reduction of probabilities, as might correspond to a scenario where
ultimate success is made conditional on some other independent event. The idea
is that starting with Schema 1 and a conditional probability (e.g., 1/2) we can
find a dollar value x that generates a second schema:

(1/2 chance of $x) (1/4 chance of $30,000),

(1/4 chance of $x) (1/12 chance of $30,000).
More generally, we have the following definition.

8 New functional forms may be generated by taking advantage of the fact that if w1(p) and w2(p
are subproportional, then so are the functions w1(p)w2(p), wl(w2(p)), and max{w1(p),w2(p)}.

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 PROBABILITY WEIGHTING FUNCTION 511

DEFINITION 4: > exhibits conditional invariance if for any outcomes x, y, x',

y' EX, probabilities q, p, r, s E [0, 1], and "conditional probability" A, 0 < A < 1:
If (x, p) (y, q) and (x, r) (y, s), then (x', Ap) (y', Aq) implies (x', Ar)
(y', As).

Actual preferences can systematically diverge from the axiom in two direc-
tions. Let ?A denote the "conditional" preferences, (x, p) ?A (y, q) if (x, Ap) ?
(y, Aq). Then > exhibits increasing-absolute subproportionality or IAS if >? iS
as subproportional as >, and decreasing-absolute subproportionality or DAS if >
is as subproportional as A. The functions implied by the invariant, constant-
absolute case are given below:

PROPOSITION 4: (A) Let > be a preference relation on _9 satisfying AO or Al

and conditional invariance. Then the weighting function for p > 0 in the representa-
tion of > is either an exponential-power fttnction,

(4.2) w(p) =exp o (I P) a '0,3>0,

or a power function,

(4.3) W(P) =PI 8,,t > 0.

(B) If ? satisfies AO orA1, and conditional invariance holds for probabilities in
the open interval (0, 1), then the weighting function for 0 < p < 1 is -y w(p), where
0 < y < 1 and w(p) is either (4.2) or (4.3).
(C) If ? satisfies AO or Al, and conditional invariance, then the axioms:
subproportionality and boundary continuity, are inconsistent.

The exponential-power function is subproportional for a > 0. However, a

sufficiently small /3 can render it concave and overweighting on (0, 1), hence
regressiveness and s-shape are not generic properties in the sense that they are
with CI. A more important difference is noted in part C of the proposition.
Equation (4.2) has a positive limit as p -> 0, which creates discontinuous
weighting at zero (the other possibility, (4.3), is not subproportional). Going
beyond this specific example, any w(p) function whose w-function, equation
(4.1), has a finite lower bound as ln p -> -oo will be discontinuous at zero. In
particular, w-functions that exhibit increasing absolute convexity in the direction
of smaller probabilities will be bounded.9 Therefore, of the three possibilities:
decreasing-, increasing-, or constant-absolute, only the decreasing case can
accommodate subproportionality and continuity at p = 0.
There is no clear argument in favor of increasing vs. decreasing relative
subproportionality as a more likely departure from the constant case-at least
nothing like the continuity argument favoring the decreasing absolute case. It is

9This follows from standard utility theory results (Pratt (1964, Theorems 1 and 8)). The CI
function is decreasing-absolute, constant-relative.

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 512 DRAZEN PRELEC

perhaps interesting that among functions that are diagonally concave, continu-
ous on [0, 1], and twice differentiable on (0, 1), there exists a connection between
the location of the fixed point and the direction of relative subproportionality.
Specifically, a diagonally concave IRS function cannot have an interior fixed
point above 1/e, and a diagonally concave DRS function cannot have one below
1/e (Prelec (1995)).

4.4. Projection Invariance

Our final example illustrates a DAS-IRS function. Let us return again to

Schema 1 and observe that (r, s) are related to the first pair (p, q) by the factors
1/2 and 1/3: r = (1/2)p, s = (1/3)q. In other words, the smaller value in the
initial pair q = 1/2, has to be reduced by a greater factor than the larger value,
p = 1. Suppose that starting with Schema 1, we could then generate a new
indifference judgment by multiplying the left and right side chances by 1/2 and
1/3:

($10,000) (1/2, $30,000),

(1/2, $10,000) (1/6, $30,000),

(1/4, $10,000) (1/18, $30,000).

In the general case, we have the following definition.

DEFINITION 5: > exhibits projection invariance if for any outcomes x, y EX,

and probabilities q, p, r, s E [O, 1]: (x, p) (y, q) and (x, 'p) (y, sq) imply
(X, r2p) (y, S2q).

The axiom is the simplest of the three invariances because it involves two
outcomes, x and y, and three preference judgments, instead of the four out-
comes and four preference judgments required to disconfirm Definitions 1 or 4.

PROPOSITION 5: (A) Let > be a preference relation on 3+ satisfyingAO orAl

and projection invariance. Then > is weakly subproportional, and w(p) is either
power function, (3.3), or a hyperbolic logarithm,

(4.4) w( p) = (1 - ae In p)_ a/ "o, > 0.

(B) If > satisfies AO or Al, and projection invariance holds for probabilit
the open interval (0, 1), then the weighting function for p < 1 is -yw(p), where
0 < -y < 1 and w(p) is either (4.4) or (4.3).

Equation (4.4) is regressive, diagonally concave if: (a + ,3) - ,3 ln(a + ,3) = 1,

a + ,3 ? 1. Under this constraint, the fixed point falls within the interval
(1/e , l/e). This is proved in Appendix 2, after the derivation of (4.4).

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 PROBABILITY WEIGHTING FUNCTION 513

As noted in 5A, weak subproportionality is an implication of projection

invariance. This is a bit surprising, because, on the face of it, Definition 5 does
not preclude a supraproportional common-ratio violation, which would be
diagnosed by the pattern p > q and r < s. However, the axiom has no solutions
that might accommodate such a combination of probabilities. Regarding the
other "essential" properties, the hyperbolic-log is typically regressive and s-
shaped, but, just like the exponential power function, equation (4.2), it becomes
concave and overweighting when ,3 is very small. This failure of generic
regressiveness and s-shape again highlights the special qualities of the CI form.

5. COMPOUND INVARIANCE AND THE ALLAIS COMMON-CONSEQUENCE EFFECT

In addition to the common-ratio effect, the 1953 article by Allais contained a

second class of counterexamples of which the following "certainty effect" is
perhaps the most famous:

(11% chance of $1,000,000) < (10% chance of $5,000,000),

(Schema 4) (1000,000 f ) > 10% chance of $5,000,000

($1,000,000 for sure)> 89% chance of $1,000,000o,

The second pair of prospects are derived from the first pair by adding a
"common-consequence" to both sides, in this case a .89 probability of $1M. In a
rank-dependent representation, Schema 4 leads to the inequality: w(1) - w(.99)
> w(.11) - w(.10), showing that, in the domain of gains, the marginal impact of
the worst centile is greater than that of the 11th best centile. Many such
common-consequence examples have been studied. For instance, the following
schema (Prelec (1990)):

(2% chance of $20,000) > (1% chance of $30,000),

(Schema 5) (34% chance of $20,000) < (1% chance of $30,000 )

32% chance of $20,000,)

generates an inequality in the interior of the probability interval: w(.02) - w(.01)

> w(.34) - w(.33). The generic difference between common-consequence and
common-ratio schemas is that the former tell us about the ordering of intervals,
w(p + A) - w(p) versus w(q + A) - w(q), while the latter tell us about the
ordering of ratios, w(zAp)/w(p) versus w(zAq)/w(q).
Is the ordering of intervals, as related by common-consequence violations,
consistent with the ordering of ratios, as revealed by common-ratio violations of
expected utility? Fortunately, the answer to this question appears to be Yes. The
implications of common-consequence violations for w(p) have been explained
in recent papers by Tversky and Wakker (1995) and Wu and Gonzalez (1996a, b).
Tversky and Wakker show that the generalization of the classic Schema 4, and a
"dual" common-consequence effect, is equivalent to subadditivity of the weight-
ing function, which holds if there exist boundary constants 8 and 8' ? 0 such

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 514 DRAZEN PRELEC

that

.u.t.y w(q) 2 w(p + q) -w(p), p + q < 1- < 1,

(Subadditivity) 1- w(1 - q) ? w(p + q) - w(p), p 2 8' ? 0.

Tversky and Wakker observe that "violations of subadditivity are rare." The
results of Wu and Gonzalez (1996b) are in the same spirit. They show that more
general classes of common-consequence schemas are equivalent to w(p) being
concave or convex on an interval. Schema 5, for instance, supports concavity on
the interior interval (.01,.34). Both papers provide comparative propositions
showing that greater subadditivity (Tversky and Wakker (1995)) or concavity/
convexity (Wu and Gonzalez (1996b)) is equivalent to more common-conse-
quence violations "of the right kind."
A subproportional CI function is subadditive, as is any strongly regressive
function.10 The compound invariance assumption is thus compatible with most,
if not all, current evidence on common-consequence violations of expected
utility. Furthermore, within the family of compound invariant preference rela-
tions, the ordering of relations by more common-ratio violations (in the sense of
Definition 3) coincides with their ordering by more common-consequence viola-
tions (in the sense of conditions (6.3) and (6.4) in Tversky and Wakker (1995)).
This is because a CI form with a smaller a is both more subproportional and
more subadditive.
The coherence of the two Allais paradoxes within the CI framework illus-
trates a more general proposition. If common-ratio violations exceed some
8-minimum (as in Definition 2), then the slope of w(p) at zero and one is
unbounded (Proposition 2). These endpoint nonlinearities are independently
confirmed by the certainty effect and by the four-fold pattern of risk attitudes.
This does not mean that one behavioral paradox is primary and the others
derivative. What it suggests, instead, is that the form of the probability weighting
function is overdetermined by the empirical evidence, with more than one road
leading to the same end.

Sloan School of Management, E56-320, Massachusetts Institute of Technology,

Canmbridge, MA, U.S.A.

Manuscript receited Auigust, 1995; final levisionz received Auigust, 1997.

10 To prove that any strongly regressive function, including the CI, is subadditive, let p0 equal the
fixed point of the function, p0 = w(p? ), and pick p + q < p?. Then w(p + q) > p + q, and w(p + q)
= (p + q)C, for some c < 1. It follows from q <p + q and strong regressivity that w(q) > qc = (q/(p
+ q))C(p + q)C = (q/(p + q))cw(p + q), as w(p + q) = (p + q)C. But c < 1 then indicates that
(q/(p + q))c > q/(p + q), or w(q) > (q/(p + q))w(p + q), and, by the same token, that w(p) >
(p/(p + q))w(p + q), proving: w(p) + w(q) > w(p + q), for boundary constant po. A similar argu-
ment shows that the second line in the definition of subadditivity holds as well. This proves the
formal claim. Tversky and Wakker also indicate that the boundary constants 8, 8' ought to be small,
which cannot be derived from strong regressiveness alone. However, the special CI form, equation
(3.1), does imply small 8, 8'.

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 PROBABILITY WEIGHTING FUNCTION 515

APPENDIX 1: REPRESENTATIONAL ASSUMPTIONS

This appendix collects axioms from Wakker (1994), which together with the results in Wakker and
Tversky (1993) are sufficient for the sign- and rank-dependent representation. Let > denote a
preference relation on the set, SD, of probability distributions, P,Q_., on X =[x-, x], with
x - < 0 <x +. Prospects are distributions with finite support. The following five axioms are assumed to
hold on .P without restriction:

WI. WEAK ORDERING: > is conmplete and transitive.

W2. STRICT STOCHASTIC DOMINANCE: P > Q if P 7L Q and P stocliasticalljy dominates Q.

W3. CERTAINTY-EQUIVALENT CONDITION: For aIll P there exists aIn x sIuch thlalt (x) P.

W4. CONTINUITY IN PROBABILITIES: If (y, p) > (x), 0 < p < 1, thlen there exist q, r stuch that
q <p <r, (y, q)> (x), and (y, r)> (x). If (y, p) <(x), 0 <p < 1, theni thlere exist q, r such that
q <p <r, (y, q) < (x) and (y, r) <(x).

Now define S(k, n), 0 < k < n, as the set of all k nonpositive and (n - k) nonnegative rank-
ni-tuples from X, S(k, n) = {(xl,..., xl) E X' : xl < ?<x < O <Xk?+ I ' . l}

W5. SIMPLE-CONTINUITY: For any pr-obability vector (p I,-, p,l) the preference relaitioni indtu
each S(k, n) is conltinuous.

The central (and final) axiom is trade-off consistency, restricted to prospects that are sign- and
rank-order compatible. Let (x, pi; x- , p - _) indicate a prospect with outcome x of rank "i" singled
out, and let .5(k, n, p) denote the set of all sign- and tank-or-der conmpatible prospects that have a
p-chance of yielding a negative outcome: (k n, p) = {(x1, p1; ...I xI P) : (x1.x,) E. S(k, n) and:
PI + +Pk P}.

W6. TRADEOFF CONSISTENCY: There do not exist eighit prospects,

(x', pi; a _, p )., (y', pi; b_j, p_), (x', q; c_j, q _j), (y', q; d
(y, qj; d _j, q _j), such that the first four and the seconid fotur belonig
comnpatible set, and that:

(x, pi;a -j, p -;) >(y, pi; a - j, p - j)

(x', pi; a - j,p -j) <(y' ,pi; b -_i,P -ij),

(x', qj; c _j, q _j) >(y', qj; d _j, c _j),

(x, qj; c _j, q _j) < (y, qj; d _j, q _j).

The necessity of this condition is straightforward to check. Because the first four prospects are
sign- and rank-order compatible, the terms involving pi, a - , b - , and p - will cancel out af
applying (2.1), yielding: v(x) - u(y) > v(x') - v(y'). Because the second four prospects are also sign-
and rank-order compatible, the terms involving qj, c _j,d _j, and q -j will also cancel out, yielding
the contradictory ordering, V(x) - V(y) < V(x') - v(y').
In view of Observations 8.1 and 8.4 of Wakker and Tversky (1993), assumptions W1-W6 conform
to the requirements of their Theorem 6.3, which ensures a CPT representation with unique,
nondecreasing weighting functions, satisfying w(0) = 0, w(1) = 1, and a ratio scale value function.
(Theorem 6.3 of Wakker and Tversky (1993) gives a representation for uncertainty; in the presence
of W2 it extends to the case of risk, as shown by Wakker (1990)). A prospect P = (xI, Pi; .. .; x,, P"),

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 516 DRAZEN PRELEC

xl < -- xk < 0 ?Xk+ I < x,, would be evaluated as:

(2.0) V(P)=E H - PiJ w- pi L) (xi)

1=1 \ \~~~=~,I

+ E W+(p j -W+( E pi LZ (xi)

i = k+ I j= 1 j = i+ I

Distributions are handled by the natural continuous generalization of (2.0). Considering only the
restriction of > to nonnegative and nonpositive prospects, Theorem 12 of Wakker (1994) proves
that W1-W6 imply, first, that w+(p) w-(p) are strictly increasing on [0,1] and continuous on (0,1),
and second, that if W4 is augmented by boundary continuity then they are also continuous at p = 0
or p = 1, respectively.

APPENDIX 2: PROOFS OF PROPOSITIONS 1-5

PROOF OF PROPOSITION 1: I prove the four parts of Proposition 1 in reverse order, D, C, B, and
A. Proposition 6, which is invoked in the proof of part A, is proved in Appendix 3.

PROOF OF PROPOSITION ID (Gains): The proof is based on a functional equation in Aczel (1966,
Theorem 2, p. 153), which gives the only solution to

(Al.1) h1-{li{ hU(t) + (1- u)h(ut )} = vh -1{ h(tl) + (1- g)h(t,)},

for u constant, 0 < u < 1, and v., tl, t, positive, as

(A1.2) h(t) =A ln(t) + B,

1(t) =AtC + B,

with A # 0, C # 0.
To prove Proposition 1, I show that compound invariance gives rise to (ALl) with h(t)=
ln(w(exp(-t))). Once (ALI) is established, w(p) is derived by inserting the solutions in (A.1.2),
solving for w, and refining the constants so that the representational assumptions in AO are satisfied.
Specifically, these assumptions would be violated if w(p) is nonincreasing at some p, or if w(p) = 0
for p #k 0, or if w(p) is unbounded on (0,1).

PROOF: I first show that compound invariance holds for any rational exponent, not just integers.
Define T(p, s, q, r-) = lV(p)W(s) - w(q)v(r). Pick any p, q, r, s such that T(p, s, q, r ) = 0, and choose
positive outcomes x, y such that, v(x)/1t(y) = w(q)/w(p) = w(s)/w(r) (such x, y exist because t' is
a continuous ratio scale). From w(p)v(x) = w(q)v'(y) and w()'(x) = w(s)v(y), we have (x, p) -
(y, q) and (x, r) - (y, s). For any integer N, pick outcomes x', y', such that v(x')/v(y') =
w(qN)/W(pN), or (X, pN) - (y,qN). Compound invariance then implies that (x',} N) _ (y,,SN), or
w(r N)Lu(x ) = w(sN)L)(y, ). Cancelling tu(x') and v(y') from W(pN)Zi(XI) = w(qN)tu,(y') and
(r N)z(xI) = w(sV)z (y ), yields w(pA)W(SV) - w(qN)w(r-N) or ( pN SN, qN r-N) = 0. Hen
any probabilities, p, q, r, s, and any integer N, T(p, s, q, r) = 0 implies T(pN, sN, qN, r'
cause T is strictly monotonic in each argument it follows, also, that T(p, s, q, r) > 0 implies
;(pN, sN, qN, r' N)>0 (if not, and (pN, SN qN, r.N) < 0, then by decreasing the first two argument
and/or increasing the second two one could bring about W(p', s', q', r') = 0 while
;( pN' SN, q/N, r.N) < ;(pN sN, qN, 'N) < 0). Likewise ;(p, s, q, r) < 0 implies ;(pN, s N, q , N) <
0. Consequently, T(p, s, q, r) = 0 if and only if ;(p N, sN, qN, r'N) 0. Therefore, T(p, s, q, ) = 0 also
implies ;(pl N, s IN, q/N , ' /N) = 0, as well as ;(pMl/N sM/N qM/N, 'M/N) = 0, for any positive
integers, M, N. This shows that for any positive rational, A = N/M, (x, p) - (y, q), (X, r) (y, s),
and (x', p A) (y', q A) imply (xI, 7A) ( A).

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 PROBABILITY WEIGHTING FUNCTION 517

Now pick any pair of positive reals tl, t,, and any rati
outcomes x, y, x', y', such that

p exp(-tl),
r exp( -t),

q = w- { w(p)w() }, implying w(q) = w(p)w(r ),

11(x) w(q)

I?(y) w(p)

z'(x') w(q A)

V (y') W(pA)

Correct q exists because w is continuous, strictly increasing on (0, 1). Correct x, y, x', y' exist by
continuity of v.(x) and the fact that t'(0) = 0. Jointly w(p)w(r) = w(q)2 and v,(x)/v1(y) = w(q)/w(p)
imply (x, p) - (y, q), and (x, q) (y, r), while tv(x')/L1(y') = w(q A)/w( pA) implies (x', p A) - (ye, q A).
Invoking compound invariance we extend to (x', qA) _ (y', r A), i.e., v(x')/1v(y') = W(r.A)/w(q A)
which combines with L)(x')/v!(y') =w(qA)/w(p A) to show that w(pA)w(A) =w(q A)2. This proves
that w(p)w(r) = w(q)2 implies w(pA)w(rA) = w(q A)2, or, letting h(t) =ln w(exp(-t)), i.e., w(p)=
exp(h(- ln p)), that h(- ln p) + h(- ln r) = 2h(- ln q) implies h(- A ln p) + h(- A ln r) =
2h(-A ln q). Solving for -ln q and -A ln q,

(l.) - In q = h - l { .5h( - In p) + .5h( - In r)},

(A1) -Aln q =h-1{.5h(-Alnp) + .5h(-Alnr)}.

From the first line of (Al.3), we have: - A In q = Ah {.5h(- In p) + .5h(- In r)}, which we may set
equal to the bottom line of (A1.3) to produce

h - l {.5h1(-A ln p) + .5h(-A ln 1)} =Ah-'{.5h(-ln p) + .5h(-ln r)},

or

h'-{.5h(Atl) + .5h(AtO)} = Ah-1{.5h(tl) + .5h(t2)},

after substituting for tl, t,. Both sides of the equality are continuous in A, as h(t) = ln w(exp( -t))
with w(p) continuous, strictly increasing on (0, 1), and A, t1, and t2 strictly positive. Because the
equality holds for every positive rational A, it follows, by taking limits of both sides, that it holds also
for any positive real A. Hence h(t) = ln w(e t) satisfies the functional equation (Al.l). The first
solution h(t) = AtC + B yields w(p) = exp(A(- ln p)C + B), A, C 0 0. Monotonicity implies that
AC < 0. The case A > 0, C < 0 is unbounded. The case A < 0, C > 0 is given in (3.4), with a = C,
3 =-A, y = exp(B). The normalization w(l) = 1, constrains: -y < 1. The second solution, li(p)
A ln t + B, yields w(p) = eB( - ln p)A, which is unbounded for A < 0 and decreasing for A > 0.

PROOF OF PROPOSITION ID (Gains and Losses): To prove that w+(p) and w-(p) have the same
value of the a-parameter, note that compound invariance allows us to reflect a common-ratio
schema from the positive into the negative domain, and vice versa. Therefore, (x) (y, p) and
(x, p) - (y, q) also implies (z) - (w, p) iff (z, p) - (w, q), even if the sign of z&w differs from the
sign of x&y. Because the probabilities in each schema uniquely determine the value of the a
parameter-and these probabilities are the same-the value of a must be the same in both w+(p)
and w-( p). The remaining parameters, /3, -y are not determined by preferences over simple
prospects; hence they may be distinct in the gains and loss part of the representation.

PROOF OF PROPOSITIONS IC, lB, lA: Direct computation with (3.4) shows that compound
invariance with p = 1 is only possible for -y = 1, proving IC (take (x) - (y, q), (x, q) - (y, s),
(x') - (y', q2), x 7 y, and show that (x', q2) _ (X/, S2) fails unless y = 1). Subproportionality implies
that cw(ln p) = ln w(p) is strictly convex (viz. Proposition 3) which can only be satisfied for a < 1,

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 518 DRAZEN PRELEC

proving lB. By Proposition 6, diagonal concavity precludes w(p) strictly overweighting and convex
on any nondegenerate interval. For the twice differentiable equation (3.2), this means that the fixed
and inflection points coincide. Computation with (3.2) shows that this forces = 1 in equation (3.2),
proving IA.

PROOF OF PROPOSITION 2: I only prove the proposition for the case of gains. Let W? denote an
s-subproportional weighting function and w3 a power function, w= p 8. The proof involves
establishing two claims. The first claim is that w, can cross w, at nmost once; the second claim is that
w, must cross w, at least once.
Consider the functions (0 and i/, defined by

(A2.1) wj (ln p) ln w(p),

Ir(-ln(- ln p)) -ln(- ln w(p)) -ln(- wj(ln p)).

wj is the graph of w(p) in log-coordinates, and i/ the graph in double-log coordinates. Because w(p)
is continuous, strictly increasing on (0,1), it follows that wj(ln p) is continuous, strictly increasing on
the negative reals, and that i/i( - ln( - ln p)) is continuous, strictly increasing on the entire real line.
If p0 is a fixed point of w, then ln p0 is a fixed point of wj, and - ln( - ln p? ) is a fixed point of i/.
Note also that the +/-graph of a power function is a linear function with slope equal to one:

(A2.2) W X(p) = p =>q8(- In(-In p).) I-n + (-lIn(-lInp)).

To prove the first claim, of at most one intersection, I show that subproportional preferences imply
that the slope of 1/1 on any nondegenerate interval is strictly less than one. By Proposition 3,
subproportional preferences indicate that the w-function is convex, i.e., that for any probabilities
q <p,

&j,(lnp) lnp
w8, (In q) ln q

(Recall that subproportionality is defined as "strict on any nondegenerate subinterval," hence wj is

strictly convex on any subinterval.) Because w8,(ln p) -exp( - q/'( - ln( - ln p))), from (A2.1), this
means that

exp(- q/i(-ln(-lnp))) lnp -lnp

exp( - q( -ln( - ln q))) ln q -ln q

or that - Q?( - ln( - ln p)) + i( - ln( - ln q)) > ln( - ln p) - ln( - ln q), i.e., I(-ln( - ln p)) -
I(-ln( - ln q)) < - ln( - ln p) - (-ln( - ln q)), or

-n(-In p)) - I3( -ln(-ln q))

-ln(-ln p) - (-ln(-ln q)) <

for any - ln( - In p)> - ln( - ln q). Therefore, the slope of '/' on any nondegenerate interval is
strictly less than one. Because i/ia has slope exactly equal to one (A2.2), the two graphs can interse
at most once, with 1/1 intersection i/ia from above.
To prove the second claim, of at least one intersection, I define,

(A2.3) qjN = qj(N In (2 + s)),

where the value of 8 is taken from q/'8 and N is any integer, positive or negative. I the
the difference, q/' - 1N has no upper bound as N - c, nor lower bound as N -* + co

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 PROBABILITY WEIGHTING FUNCTION 519

The preference conditions given by &-subproport

cate that V(X) = w(p)v(y) == w(p))(X ) < w(p2+ 8)L(

W (p) ) <' (p 2 ),

- 21n w,(p) > -In wj(p2+ ),

In 2 + In( -In w,,(p)) > In( - In w, (p 2+ ?

In 2 - j- In(- In p)) > ql,(- In(- In p 2+ ?) (from A4.1),

= -i(- In(-Inp)-In(2 + 8)).

In particular, for - ln( - In p) = N ln(2 + 8), this implies

fr8(N ln(2 + 8)) - qi((N- l)ln(2 + 8)) = qrN -/8N - < n2,

for any integer N. By induction, we have,

q.? - ? < N Iln 2, for N > 1,

(A2.4)
ifrj- -frj > N ln2, for N< -l.

(The second line can also be read as qf - qfN < (-N)ln 2, for N < - 1.)
The difference between "8 and Ifr, can now be expressed as the sum of three terms:

ql~ - _q N = ( q N - qjl) + (j - q N) + (0 - 00)

=(0 v - ipo )-N In(2 + 8) + ( q8? - qjB) (foA2).

For positive N, we substitute for q,N_ - qi/ from the top line

11N - 3N < N ln2-N ln(2 + 8) + (q1r- i 0) (N 2

Because 8 > 0 and &?0 - ,3? is a constant, the difference, t&8N -

large N. For negative N, we substitute for q1,N - q0 from th
inequality,

NfrN > Nln2-Nln(2+ 8) + ( _o-o) (N? -

showing that the difference, .N - must be positive for suff
two graphs have at least one intersection. As we have previously
at most once, this proves that there is a unique intersection point, with the 8-subproportional
function intersecting the power function from above.

PROOF OF PROPOSITION 3-(C) => (A): Assume that (A) is false, and that we have probabilities
p > r 2 q > s, and outcomes x <y, x' <y', such that (x, p) 2 (y, q), (x, r) -, (y, s), and (x', p) -I
(y', q), but (x', r) >1 (y', s). Applying the separable formula to each relation and cancelling v(x) and
v(y), we arrive at w9(p)/w2(r ) = w2(q)/w2(s), and w1(p)/wj(r) < wj(q)/w1(s). In terms of cw-func-
tions, if(ln p) = ln w(p), this implies exp(w,(ln p))/exp(w2(2n r)) = exp(wc,(ln q))/exp(wc2(ln s)), and
exp(cw(ln p))/exp(cw1(ln r)) < exp(col(ln q))/exp(w1j(ln s)), or w,(ln p) - w2(ln r) = 2,(ln q) -
w2^(ln s), and wl(ln p) - co(ln r) < w(ln q) - co(ln s). Consequently,

wjl(ln p) - o1(ln r) 1 c2(ln p) - w2(ln r )

wl(ln q) - w(ln s) wt,2(In q) - c2(In s)

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 520 DRAZEN PRELEC

which, in view of In p > In r 2 In

tion of w.
To prove, (A) => (C), pick any 0 < s < r < q < p < 1, such that

(A3.1) wJ2(In p) - w(ln q) = 02(lnr) - W(2n s).

Substituting w(ln p) = In w(p), we can rewrite this as

(A3.2) w2 ( p)/lw(q) = w2( r)/w9(s)

Because each vi(x) is a continuous ratio scale with v(0) = 0, we can also choose x <y, x' <y', such
that

(A3.3) w9(p)/w2(q) =L2(Y)A1(X)

(A3.4) w1(p)/wI(q) v1(y )1v1(x )-

(A3.3) implies w2(p)v2(x) w2(q)v2(y) or (x, p) - (y, q), (A3.4) implies w1(p)v1(x') = w1(q)v1(y')
or (x', p) I (y', q), while (A3.2) and (A3.3) jointly imply w2(r)v2(x) = W2(s)v2(Y) or (y, s) 2 (X, r
Because >1 exhibits more common-ration violations than >2 , it follows that (y', s) ?1 (x', r), i
w1(s)v1(y') ? w1(r)v1(x'), or v1(y')/v1(x') 2 w1(r)/w1(s). Combining this inequality with (A3.4)
yields w1(p)/wI(q) 2 wl(r)/wl(s), or exp wl(ln p)/exp col(ln q) 2 exp 1),(ln r)/exp wl(ln s), or
co(In p) - to(In q) 2 w (ln r) - to(ln s). In view of (A3.1), this shows that for any nonpositive
In p > In q ? In r > In s, the equality wt2(ln p) - W2(ln q) = 2(ln r) - w9(2n s) implies wl(ln p) -
wl(ln q) 2 wl(ln r) - co(ln s), which proves that w1 is a convex transform of co, on t < 0.
The equivalence of (B) and (C) follows from the definitions. Weak subproportionality, w(p)w(Aq)
? w(q)w(Ap), for p > q, 1 > A > 0, may be rendered as w(p)/w(Ap) ? w(q)/w(Ap), i.e., w(ln p) -
(o(ln p + ln A) > (ln q)-J(ln q + ln A), for 0 > ln p > ln q, 0 > ln A. Hence wJ(ln p) is convex iff
w(p) is weakly subproportional.

PROOF OF PROPOSITIONs 4 AND 5: The proofs are based on a second functional equation in Aczel
(1966, Theorem 2, p. 153), which gives the only solutions to

(A4.1) g- 1 {t(g(V + tl) + (1 - I_0g(V + t9)) = V +g-l{tg(tl) + (1 -)g(t))},

for Ar constant, 0 < ,u < 1, and L), tl, t2 positive as

(A4.2) g(t)=At + B,
g(t) =-AeC" + B,

with A 0 0, C 0 0. To prove Propositions 4 and 5, one shows that both preference patterns give rise
to one of these forms, specifically:
(i) g(t) ln(w(exp( - t))), for conditional invariance,
(ii) g(t) = ln(w - 1 {exp( - t)}), for projection invariance.
Once (A4.1) is established, the corresponding w(p) forms are derived by inserting the solutions in
(A4.2), solving for w, and refining the constants to meet whatever other axioms apply in a particular
case.

PROOF OF 4B: As with Proposition 1, I first prove the more general case (4B), where condition
invariance is assumed to hold only for probabilities in the open interval (0, 1). For any positive

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 PROBABILITY WEIGHTING FUNCTION 521

v, tl, t2, define probabilities A, p, q, r and outcomes x, y,

A = exp(-v ),

p = exp(-t1),

r= exp(-t2),

q = w- V{w(p)w(r) }, implying w(q) = w(p)w(r),

v7(x) w(q)

v(y) w(p)'

v(x') 2(Aq)

v(y') w(Ap)

Correct q exists because w is continuous, strictly increasing on (0, 1). Correct x, y, x', y' exist by
continuity of v(x) and the fact that v(O) = 0.
Now, w(p)w(r) = w(q)2 and v(x)/v(y) = w(q)/w(p) jointly imply (x, p) (y, q), and (x, q)
(y, r), while v(x')/v(y') = w(Aq)/w(Ap) implies (x', Ap) (y', Aq). Invoking conditional invariance
(Definition 4) we can extend this to (x', Aq) - (y', Ar), i.e., v(x')/v(y') = w(Ar)/w(Aq), which
combines with v(x')/v(y') = w(Aq)/w(Ap) to show that w(Ap)w(Ar) =w(Aq)2. This proves that
w(p)w(r) = w(q)2 implies w(Ap)w(Ar) = w(Aq)2, or letting h(p) = ln w(p), that h(p) + h(r)
2h(q) implies h(Ap) + h(Ar) = 2h(Aq). Solving for q and Aq,

(A4.3) q = h -'{.5h(p) + .5h(r)},

Aq = h - 1 {.5h(Ap) + .5h(Ar)}.

From the first line of (A4.3), we obtain Aq = Ah1{.5h(p) + .5h(r)}, which we may set equal to the
bottom line of (A4.3) to produce

h {.5h( Ap) + .5h( Ar)} = Ah1 {.5h(p) + .5h(r)},

or,

g- {.5g(v + tl) + .5g(v + t2)} tL +g-1{.5g(tl) + .5g(t2)},

with g(t) = h(exp(-t)), h -1{t} = exp(-g-1{t}). Hence g(t) = ln w(exp(-t)), or g(-ln p) = ln w(p).
The solution, g(- ln p) = ln w(p) = -A ln p + B, yields w(p) = eBpA, while g(-ln p) ln w(p) =
Ae-C In + B =Ap- c + B yields w(p) = exp(Ap-C + B). The first solution is the power function in
(3.2). For the second solution, monotonicity implies AC < 0. This corresponds to yw(p) where w(p)
is given in equation (4.2) in the text, with a =- C, / =-AC, y = exp(A + B).

PROOF OF 4A: Direct computation with yw(p), w(p) either (4.2) or (4.3), shows that conditional
invariance with p = 1 is only possible for y = 1 (take (x) (x, q) (y, s), (x') (y',.5q), x # y, and
show that (x',.5q) (x',.5s) fails unless y = 1).

PROOF OF 4C: By 4A, conditional invariance implies w(p) of the form (4.2) or (4.3). Subpropor-
tionality is inconsistent with (4.3). The co-function for (4.2) is w(ln p) = (-'/8a)(-pa), or w(t) =
(-,/8a)(1 - exp(at)), which is convex for a > 0. Hence subproportionality implies a > 0. The limit
as p -O 0 of equation (4.2) is then exp(- //a) > 0, which violates boundary continuity.

PROOF OF 5B: I take the two parts in the order 5B, 5A. First I will slightly rephrase the statement
of the projection invariance axiom. Note that in the context of our separable representation with
continuous, strictly increasing w(p), the axiom (x, p) (y, q) and (x, ip) (y, sq) => (x, r2p)
(y, s2q), also implies (x, p) (y, q) and (x, r2p) (y, S2q) => (x, rp) (y, sq). Letting r' = pr2, q' =

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 522 DRAZEN PRELEC

sq2, this indicates that, (x, p) (y, q) and (x, r') (y, s') (x, (r'/p)-5p) (y, (s'/q).5q), i.e.,

(x, p) (y, q) and (x, r') (y, s') (x, p 5r' 5) (y, q5s5).

Now for any positive v, t1, t2, define probabilities p, q, r, s and outcomes x, y, such that

p = w- 1{wl exp(- t1)},

r'w- {wlexp(-t2)},

(A5.1) q= w 1{w exp( - (u + t1 ))},

s w- 1 {w1 exp(- (v + t2))},

L)(X)
v() = exp(-u),

where w1 = lim{w(p)} as p 1. Correct x, y exist by continuity of u(x) and the fact that u(O) = 0.
Correct p, q, r, s exist because w is a continuous, strictly increasing map of (0, 1) onto (0, w1). To
show that w(p) is indeed continuous at zero, observe that projection invariance implies, in
particular, that (x) (y, q) and (x, q) (y, sq) (x, q2) (y, S2q), and by induction, for any in
N> 1, that (x, qN) (y,sNq). Hence for any N, w(qN)(x)u =w(sNq)u(y), or u(x)/u(y)=
w(sNq)/w(q N), which indicates that the limit as N - of w(qN) is zero (otherwise th
ratio, w(sNq)/w(q N), would equal one).
Combining lines one, three, and five in (A5.1), yields: w(p)u(x) = w(q)v(y), i.e., (x, p) (y, q);
combining lines two, four, and five yields: w(r)u(x) = w(s)u(y), i.e, (x, r) (y, s). By projection
invariance, we may infer that (x, p.5r" 5) (y, q.5s.5), or

u(x)w(p 5r 5) = w(q4sq5).

Substituting for p, q, rA, s, x, y, from (A5.1),

exp(-u)w(w- 1{w exp(t_ 1)} 5w- 1{wl exp(-t2)} 5)

= w(w {4wl exp( -u - t1)} 5w- 1{wl exp(-u -t2

and writing

(A5.2) exp(g(t)) = w 1 {w1 exp(-t)}, i.e., ln w(p) = g- 1 {ln p} + ln w1,

we substitute first for w - 1 {wi exp( - t):

exp(-u)w(exp(.5g(tl))exp(.5g(t2))) = w(exp(.5g(v + t1))exp(.5g(v + t2)))

Then, taking logs, we substitute for ln w(p) from (A5.2), to obtain the basic functional equation
(A4. 1):

u +g-1{.5g(tl) + .5g(t2)} =g-1{.5g(u + tl) + .5g(u + t2)}.

As g(t) = lnw-1{w exp(-t)}, the first solution, g(t) =At +B, yields the power function: w(p)=
w1 exp(B/A)p- 1I/A. The second solution, g(t) =AeC't + B, yields the log-hyperbola: w(p) = w,((ln
- B)/A)+1 /c. The case, B < 0 is inconsistent with the representational assumptions. Hence B
and A < 0. By monotonicity, C < 0, yielding,

w(p) = y(l - ae In p) -"/Y, ae, 8, y > O,

with a = 1/B, = BC, y= w (-A/B)C. The condition w(1) = 1 constrains y < 1.

PROOF OF 5A: From 5B we know that a > 0 in (4.4). The wo-function for (4.4) is o(ln p) =
(- 8/ac)ln(1 - ae ln p), or w(t) = (- 8/ac)ln(1 - aet), which is convex for a > 0. Hence (4.4) is

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 PROBABILITY WEIGHTING FUNCTION 523

subproportional. The other solution, (4.3), is weakly subproportional. Hence preferences must be
weakly subproportional. To show that y = 1, take (x) (y, q), (x, q) (y, sq), x #y, and show by
direct computation that for either (4.4) or (4.3), the implication (x, q2) (y, S2q) fails unless -y = 1.
To assess the implications of diagonal concavity, I first define the inflection point in terms of the
co-function, (ln p) = ln w(p), which we can rewrite as exp(W(t)) = w(exp(t)), for exp(t) =p, t =
In p. Differentiating once yields w'(t)exp(o)(t)) = exp(t)w'(exp(t)), and twice yields W/(t)2exp(w(t))
+ o'(t)exp(ow(t)) = exp(t)w'(exp(t)) + exp(2)w"(exp(t)). Dividing the left and right terms in the
second-order equality by the left and right terms in the first-order equality, gives

=t1+exp(ptt)
oil (t) + l( )wexp(t) ex(t)),

= 1 if t is an inflection point.

The (v-function associated with (4.4) is w(t) - (/3/a )ln(1 + at). If to is a fixed point, then
a to - /, ln(1 - ato ). If to is also an inflection point, it follows from the above expression that
a + /3 = 1 -a to . Eliminating a to yields the constraint ( a + 3 ) - /3 ln( a + ,B ) = 1. Letting a = a +
/3, we may write a = (a ln(a) - (a - 1))/ln(a), and /3 = (a - 1)/ln(a), with a, ,B > 0 implying a > 1.
To compute the lower and upper fixed point bounds, we express to as

3 13 (a - )ln(a) f(a)
= - - n(l-ato) = - _ _ _ _ _ _ _
a a ln(a) - (a - 1) g(a)

Applying l'Hopital's rule twice yields f'(a)/g"(a) = -(1 + a-1), which, in the two limits, at a = 1,
and a -- + cs, equals -2 and - 1, respectively. Therefore, the fixed point probability lies between
e-2(o= .14) and e-1(= .37).

APPENDIX 3: PROOF OF PROPOSITION 6, ON DIAGONAL CONCAVITY

I derive here a general result on diagonally concave representations, which is needed in the proof
of Proposition 1. As in the text, let A+[s, r] = {(x, p;y, q)I0 <x <y, s < q, p + q ? r} and A-[s, r ] =
{(x, p; y, q)l y < x < 0, s < q, p + q < r), i.e., the sets of binary prospects where the probability of the
extreme outcome is at least s and the probability of the zero outcome at least 1 - r. I first state and
prove the proposition for gains, and then briefly discuss the proposition for losses.

PRoPosITIoN 6 (Gains): Let > be a subproportionalpreference relation on 97+ satisfyingAO. Then

for any interval [s, r], 0 < s < r < 1, the following two conditions are equivalenlt.
(i) > is quasiconvex and strictly CE-quasiconcave on A + [ s r [respectively, > is quasiconcave and
strictly CE-quasiconvex on A + [ s, r ]];
(ii) w+( p) is convex and strictly ovetweighting on [s, r] [respectively, conicave and strictly under-
weighting on [s, r]].

The restriction of [s, r] to the interior of the probability interval takes account of the possible
discontinuities of w+(p) at the endpoints. If w+(p) is continuous on [0, 1] then the statement of
Proposition 6 can include s = 0, r = 0.
The proof requires two separate steps. The first step, in Lemma 1, involves translating concavity
or convexity of the weighting function into the language of preferences. Here I apply Wakker's
(1994) full-interval argument to the subinterval [s, r]. The second step involves translating over- and
under-weighting into the language of preferences. This is done in Lemmas 2 and 3. The key point is
the use of CE-quasiconvexity to derive an inequality, (A6.3), that supports conclusions about the sign
of w(p) -p.

LEMMA 1: Under the assumzlptions in Proposition 6, > is quasiconvex [quasicon7cave] on A+[

w+(p) is convex [concave] on [s, r].

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 524 DRAZEN PRELEC

PROOF: If w+(p) is convex on [s,r] then quasiconvexity follows by direct application of the
rank-dependent formula. Wakker (1994, Theorem 24) proves the converse for the special case of full
probability interval, s = 0, r = 1, and the main idea in his proof can be applied to any [s, r]. Suppose
that > is quasiconvex on A+[s, r] but that w+(p) is not convex. In that case, for some
s p <q <r, and 0< , < 1,

(A6.1) /xw+ (p) + (1 - ,u)w+ (q) < w+ ( ,up + (1-)q).

If q < r, then because w+ is strictly increasing there is a p* E (q, r) at which w+ has finite right
derivative. Define prospects P = (x, q* -p; y, p), Q = (x, p* - q; y, q), for q <p* < q* < r, with
v(x) = v? < 1 = v(y) adjusted so that P Q, i.e.,

(A6.2) (1 - v )(w+ (q) -w+ (p)) = v (w+ (q*) -w+ (p* )).

By quasiconvexity, P, Q > ,8P + (1 - ,u)Q so that ,uV(P) + (1 - ,u)V(Q) ? V( ,uP + (1 - ,u)Q), or

, -v0 )w+ (p) + v? w+ (q*)) + (1 - ,u)((1 -u )w+ (q) + v? w+ (p*))

> (1-v )w ( p + (1-,u)q) + v w+ ( q* + (1- ),

which yields

(1-v? )( ,uw+ (p) + (1 - p_)w+ (q) - w+ ( ,up + (1 - u)q))

2 v? (w+ ( ,ukq* + (1-l,u)p* ) - ( /,tw+ (q* ) + (1 -l,u)w+ (p* ))).

Dividing by the left and right sides of (A6.2) produces a sum of two fractions:

(Aw+ (p) + (1 - ,)w+ (q)) - w+ ( ,p + (1 - ,u)q)

w+ (q)-w+ (p)

(,iw+ (p*) + (1 - ,)w+ (q*)) - w+ (,u-p* + (1-)q*)

w+(q*) -w+(p*) j 0,
that must be nonnegative for all s <p < q <p* < q* < r, if indeed preferences are quasiconvex on
A+[s, r]. By (A6.1) the first fraction is strictly negative, which means that the second fraction must
be positive. However, as w+(p*) has finite right derivative, the second fraction can be brought down
to zero by drawing q* down to p*. Consequently, whenever (A6.1) involves a strict subinterval of
[ p,q] we can construct a contradiction of quasiconvexity. But, if (A6.1) holds for p = s, q = r, it will
also hold for some subinterval, in which case we are again able to create a quasiconvexity violation.
Hence w+ is necessarily convex. This proves Lemma 1.

LEMMA 2: Under the assumptions of Proposition 6, if w+( p) is concave and underweighting on

then > is CE-quasiconvex on A + [ s, r]. If w + ( p) is convex and overweighting on [s, r], then > is
CE-quasiconcave on A + [ s, r ]. If either concavity/convexity or under/overweighting is strict, then the
relevant preference property is strict.

PROOF: Consider a sure outcome Q = (x), and prospects P = (x, q -p; y, p) and ,uQ + (1 - )P
= (x, pk + (1 - u,)(q -p); y, (1 - ,)p). Preferences will be CE-quasiconvex with respect to these
three prospects if P Q implies Q > 1P + (1 - ,)Q, which is to say, ,uV(P) + (1 - ,)V(Q) ? V( [uP
+ (1 - px)Q). Let v(y) = 1 > v? = v(x) (note that v is a ratio scale). Applying the rank-dependent
formula for binary prospects, equation (2.2), we have:

Q P implies v? = (1 - v? )w+ (p) + v? w+ (q),

P > Q + (1 - ,)P implies (1 - v )w+ (p) + v? w+ (q)

2 (1 -v )w+ (p - I,up) + v? w+ (q + ,u/(1 - q)).

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 PROBABILITY WEIGHTING FUNCTION 525

Cancelling the (1 - v? )/v? ratio from both equations yields

1 - w(q) w+ (q + ,u(l - q)) - w+ (q)

w+ (p) > w+ (p)-w+ (p-,up)

or

w+ (q + ,uc(i - q)) - w- (q)

(A6.3) -w+ (q) p - j(1- q)

(A6.3) 1 cq w+ (p) w+ (p) -w+ (p -PP)
P-p

Lemma 2 now follows immediately. Assume w+(p) is concave and underweighting on [s,r]. By
assumption, P and ,tP+(1-,t)Q are in A+[s,r], hence s?p- Ip?q+it(1-q)<r. Under-
weighting implies w+(p) ?p and 1 - w+(q) 2 1 - q, so the left side in (A6.3) is not less than 1.
Concavity and p < q imply that

W+ (1 + /,u(l - q)) - w+ (q) w+ ( p) - w+ ( p - 1jp)

,u(l - q) P

which proves the right side in (A6.3) is not greater than 1. Therefore, the inequality in (A6.3) holds,
and preferences are CE-quasiconvex. Strict CE-quasiconvexity arises if the left side in (A6.3) is
strictly less than one, and/or if the right side is strictly greater than one, i.e., for strict underweight-
ing and/or strict concavity of w+(p) on [s, r ].

LEMMA 3: Unider the assumptions of Proposition 6, w+(p)>p [<p] implies that > is not
CE-quasiconvex [not CE-quasiconcave] on any A + [ s, r ] with s <p < r .

PROOF: Note first that under the assumptions of Proposition 6, w+(p) is absolutely continuous
because subproportionality implies that the function, w(ln p) =ln w+(p) is convex, and therefore
absolutely continuous. Because w+ is also strictly increasing, w+(p) >p implies that there exists a
p* near p, such that s <p* < r, w+(p* ) > p*, and w+ has a finite, nonzero derivative at p*. Now
letting p = q =p* in (A6.3), CE-quasiconvexity on A+[s, r] implies that for all ,t such that:
s p-,up, q+i(1 -q)<r,wewill have

w+ (p* + ,u(l -p*)) - w+ (p*)

1w+ (p*) p* ,u(l -p*_)
1p* w+ (p*) w+ (p*) - w + (p - ,p*

Differentiability at p* with nonzero derivative implies that the right side approaches one
the left side, however, is a constant strictly less than one, as w + ( p* ) > p*. Therefore the inequality
reverses for sufficiently small ,t, and > is not CE-quasiconvex on any A + [ s, r ] such that s <p < r .
To prove Proposition 6 it is sufficient to show that for any nondegenerate interval [s, r],
quasiconvexity and strict CE-quasiconcavity of preference on A+ [ s, r] is equivalent to the statement
that w+(p) is convex and strictly overweighting on [s, r]. A similar argument would show that
quasiconcavity and strict CE-quasiconvexity of preference on A+[ s, r-] is equivalent to the statement
that w+(p) is concave and strictly underweighting on [s, r ].

PROOF OF PROPOSITION 6(=>): Quasiconvexity and strict CE-quasiconcavity imply that w+(p) is
convex and strictly overweighting on [s, r ]. We take the two implications in turn:
(i) By Lemma 1, quasiconvexity implies that w+(p) is convex.
(ii) It remains to be shown that w+(p) is strictly overweighting on [s, r]. I prove the contraposi-
tive, i.e., if w+(p) is not strictly overweighting, then either quasiconvexity or strict CE-quasiconcav-
ity must fail. On the assumption that w+( p) is not strictly overweighting, either w+( p) = p on all of

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 526 DRAZEN PRELEC

[s, r] or w+(p) <p for some p E (s, r). If w+(p) =p, then it is also concave and underweighting,
which, by Lemma 2, implies CE-quasiconvexity, hence failure of strict CE-quasiconcavity on
,A[s, r]. If, however, w+(p) <p for some p, then by Lemma 3, preferences are not CE-quasicon-
cave on A[s, r]. Therefore, w+(p) must be strictly overweighting on [s, r].

PROOF (=): w+(p) convex and strictly overweighting on [s, r] implies quasiconvexity and strict
CE-quasiconcavity on A[s, r]. We take the two implications in turn:
(i) By Lemma 1, w+(p) convex implies quasiconvexity on A+[s, r].
(ii) By Lemma 2, w+(p) convex and strictly overweighting implies strict CE-quasiconcavity on
A[ s, r].

The loss-version of Proposition 6 has a parallel statement and proof, but with reversal of
quasiconcavity and quasiconvexity, and of CE-quasiconcavity and CE-quasiconvexity.

PROPOSITION 6 (Losses): Let > be a subproportionalpreference relation on 9 satisfyingAO. Then

for any interval [s, r ], 0 < s < r < 1, the following two conditions are equivalent:
(i) > is quasiconcave and strictly CE-quasiconvex on -[ s, r], [respectively, > is quasiconivex and
strictly CE-quasiconcave on A - [ s, r ]];
(ii) w - ( p) is convex and strictly overweighting on [s, r] [respectively, concave and strictly under-
weighting on [s, r ]].

COMMENTS ON THE PROOF: The proof follows the same sequence in Lemmas 1-3, with parallel
prospect-constructions. In particular, Lemma 1 for losses asserts that > is quasiconcave [quasicon-
vex] on A-(s,r) iff w-(p) is convex [concave] on [s,r]; Lemma 2 for losses asserts that w-(p) is
concave and underweighting on [s, r] implies > is CE-quasiconcave on A - [ s, r], and w - ( p) conve
and overweighting on [s, r] implies > is CE-quasiconvex on A-[s, r]; Lemma 3 for losses asserts
that w - ( p) > p ( <p) implies that > is not CE-quasiconcave [not CE-quasiconvex] on any A - [ s, r],
s <p < r. In terms of constructions, the generic positive binary prospect, P+ = (x, q -p; y, p), with
0 < v? v(x) < 1 v(y), evaluating as V(P+) = (1 - v? )w+(p) + v? w+(q), is replaced by a "re-
flected" counterpart, P- = (x, q - p; y, p), with 0 > - v? v(x) > - 1 v(y), evaluating as V(P-)
= - ((1 - v? )w - ( p) + v? w - (q)). In the proofs, statements about probabilistic risk-attitude for
losses lead to opposite inequalities at the level of the weighting function, e.g., as in the central
(A6.3). The flow of the argument is otherwise the same.

REFERENCES

ACZEL, J. (1966): Lectures on Functional Equations and their Applications. New York: Academic
Press.
ALLAIS, M. (1953): "Le Comportement de l'Homme Rationel Devant le Risque, Critique des
Postulates et Axiomes de l'Ecole Americaine," Econornetrica, 21, 503-546.
CAMERER, C. F., AND T.-H. Ho (1994): "Violations of the Betweeness Axiom and Nonlinearity in
Probability," Journal of Risk and Uncertainty, 8, 167-196.
KAHNEMAN, D., AND A. TVERSKY (1979): "Prospect Theory: An Analysis of Decision Under Risk,"
Econometrica, 47, 263-291.
LUCE, R. D. (1988): "Rank-dependent, Subjective Expected-Utility Representation," Journal of Risk
and Uncertainty, 1, 305-332.
(1991): "Rank- and Sign-dependent Linear Utility Models for Binary Gambles," Journal of
Economic Theory, 53, 75-100.
LUCE, R. D., AND P. C. FISHBURN (1991): "Rank- and Sign-Dependent Linear Utility Models for
Finite First-order Gambles," Journal of Risk and Uncertainty, 4, 29-59.
PRATT, J. W. (1964): "Risk Aversion in the Small and in the Large," Economnetrica, 32, 122-136.

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms
 PROBABILITY WEIGHTING FUNCTION 527

PRELEC, D. (1990): "A Pseudo-endowment Effect, and its Implications for some Recent Non-ex-
pected Utility Models," Journal of Risk and Uncertainty, 3, 247-259.
(1995): "The Probability Weighting Function," MIT Working Paper.
QUIGGIN, J. (1982): "A Theory of Anticipated Utility," Jourcnal of Econonmic Behavior and Organiza-
tion7, 3, 324-343.
RUBINSTEIN, A. (1988): "Similarity and Decision-making Under Risk," Journ1al of Economic Theoiy,
46, 145-153.
STARMER, C., AND R. SUGDEN (1989): "Violations of the Independence Axiom in Common Ratio
Problems: An Experimental Test of Some Competing Hypotheses," Annals of Operations Re-
search, 19, 79-101.
TVERSKY, A., AND C. R. Fox (1994): "Weighting Risk and Uncertainty," Psychological Review, 102,
269-283.
TVERSKY, A., AND D. KAHNEMAN (1992): "Cumulative Prospect Theory: An Analysis of Decision
Under Uncertainty," Journal of Risk anid Uncertainty, 5, 297-323.
TVERSKY, A., AND P. WAKKER (1995): "Risk Attitudes and Decision Weights," Econ7ometrica, 63,
1255-1280.
WAKKER, P. P. (1990): "Under Stochastic Dominance Choquet-Expected Utility and Anticipated
Utility are Identical," Theoty and Decision, 29, 119-132.
(1994): "Separating Marginal Utility and Probabilistic Risk Aversion," Theory and Decision,
36, 1-44.
WAKKER, P. P., AND A. TVERSKY (1993): "An Axiomatization of Cumulated Prospect Theory,"
Journal of Risk and Uncertainty, 7, 147-176.
Wu, G., AND R. GONZALEZ (1996a): "Curvature of the Probability Weighting Function," Maniage-
ment Science, 42, 1676-1690.
(1996b): "Common Consequence Conditions in Decision Making under Risk," Working
Paper.
YAARI, M. E. (1987): "The Dual Theory of Choice Under Risk," Econornetrica, 55, 95-115.

This content downloaded from

165.91.13.230 on Tue, 28 May 2024 20:59:53 +00:00
All use subject to https://about.jstor.org/terms