Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Ecta200555 2

Download as pdf or txt
Download as pdf or txt
You are on page 1of 24

Econometrica, Vol. 91, No.

4 (July, 2023), 1393–1416

TESTING HURWICZ EXPECTED UTILITY

HAN BLEICHRODT
Department of Economics (FAE), University of Alicante

SIMON GRANT
Research School of Economics, Australian National University

JINGNI YANG
Research School of Economics, Australian National University

Gul and Pesendorfer (2015) propose a promising theory of decision under uncer-
tainty, they dub Hurwicz expected utility (HEU). HEU is a special case of α-maxmin
EU that allows for preferences over sources of uncertainty. It is consistent with most of
the available empirical evidence on decision under risk and uncertainty. We show that
HEU is also tractable and can readily be measured and tested. We do this by deriving
a new two-parameter functional form for the probability weighting function, which fits
our data well and which offers a clean separation between ambiguity perception and
ambiguity aversion. In two experiments, we find support for HEU’s predictions that
ambiguity aversion is constant across sources of uncertainty and that ambiguity aver-
sion and first order risk aversion are positively correlated.
KEYWORDS: Hurwicz expected utility, probability weighting functions, source pref-
erences.

1. INTRODUCTION
THE MODAL CHOICES first proposed by Ellsberg (1961) show a dislike of ambiguity and
an inconsistency with beliefs that can be represented by a (unique) subjective probabil-
ity measure.1 While Ellsberg’s paradox suggests ambiguity aversion, subsequent research
has found that ambiguity attitudes are richer than simply aversion. For gains on unlikely
events and for losses, people are generally ambiguity seeking (Kocher, Lahno, and Traut-
mann (2018)). Moreover, many experiments have shown that they have preferences over
the sources of uncertainty that go beyond a preference for risk (objective probabilistic
information) over uncertainty (vaguely specified probabilities). People prefer, for exam-
ple, to bet on more familiar sources (French and Poterba (1991), Abdellaoui, Baillon,
Placido, and Wakker (2011)) and on sources about which they consider themselves com-
petent (Heath and Tversky (1991)). A rich variety of models has been developed to ac-
count for Ellsberg-type behavior (see Gilboa and Marinacci (2016) for a review). These
models tend to concentrate on ambiguity aversion, and with a few exceptions, do not al-
low for ambiguity seeking, let alone source preference. Moreover, most of these models

Han Bleichrodt: hanbleichrodt@ua.es


Simon Grant: simon.grant@anu.edu.au
Jingni Yang: jingni.yang@anu.edu.au
We would like to thank four anonymous reviewers for their helpful comments on a previous version of
this paper. We also benefit from discussions with Aurélien Baillon and Peter P. Wakker. We would also like
to thank Evan Calford and Emma Schultz for making it possible for us to run the additional sessions during
Covid. Han Bleichrodt acknowledges financial support from the Generalitat Valenciana (Prometeo/2021/073).
Open access publishing facilitated by Australian National University (ANU), as part of the Wiley-Australian
National University (ANU) agreement via the Council of Australian University Librarians.
1
See Trautmann and Van De Kuilen (2015) for a review.
© 2023 The Authors. Econometrica published by John Wiley & Sons Ltd on behalf of The Econometric Soci-
ety. Simon Grant is the corresponding author on this paper. This is an open access article under the terms of
the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in
any medium, provided the original work is properly cited, the use is non-commercial and no modifications or
adaptations are made.
1394 H. BLEICHRODT, S. GRANT, AND J. YANG

are defined in the Anscombe–Aumann framework and cannot account for the observed
violations of expected utility for decision under risk.
The evidence about which of these ambiguity models best describes people’s prefer-
ences is surprisingly scarce. The main reason for this is that most ambiguity models use
concepts that are hard to measure. The studies that compared (a subset of) the ambigu-
ity models concluded that models that allow for both ambiguity seeking and ambiguity
aversion such as α-maxmin expected utility (Hurwicz (1951), Luce and Raiffa (2012),
Ghirardato, Maccheroni, and Marinacci (2004)), Choquet expected utility (Schmeidler
(1989)), and prospect theory (Tversky and Kahneman (1992)), perform better than mod-
els that account only for ambiguity aversion (Baillon and Bleichrodt (2015)). Moreover,
source-dependent models perform better than models that do not account for source
preference (Chew, Miao, and Zhong (2017)).
Gul and Pesendorfer (2015) provide behavioral foundations for a model, that they call
Hurwicz expected utility (HEU), which can account for many of the observed empirical
deviations from subjective expected utility. HEU is a special case of α-maxmin where the
set of priors is generated from extensions of a single prior μ defined on a σ-algebra of
ideal events. HEU can address Ellsberg-style evidence as well as the joint occurrence of
ambiguity aversion and ambiguity seeking. For a particular source of ambiguity, which
(formally) corresponds to a probability measure defined on a specific σ-algebra of events,
HEU admits a rank-dependent utility representation (Quiggin (1982)). Consequently, it
can also accommodate the main deviations from expected utility that have been observed
for decision under risk. Moreover, while in many applications of rank-dependent utility an
ad hoc specification is used for the probability weighting function, Gul and Pesendorfer
prove that the HEU probability weighting function for a given source is a convex com-
bination of a power series and its dual. Because these power series vary across sources,
HEU allows for source preference. In other words, HEU is the first axiomatic theory that
is consistent with most of the available empirical evidence on decision under risk and
uncertainty.
In this paper, we show that HEU can also be made tractable. The main problem in
measuring HEU is that the probability weighting function has an infinite number of pa-
rameters. We restrict the power series, while still satisfying all requirements of HEU, so
that the probability weighting function is characterized by two parameters, one reflect-
ing ambiguity perception, the other ambiguity aversion. HEU can then be estimated and
tested. Hence, an interesting byproduct of our paper is to propose a new functional form
of the weighting function, which is axiomatically derived (to the extent that it follows from
axioms of HEU). We show that this HEU weighting function fits our data at least as well
as widely used alternatives like those of Prelec (1998) and Goldstein and Einhorn (1987)
with the additional advantage that the parameters have a clear interpretation.
We test two predictions of HEU: first, that the parameter reflecting ambiguity aver-
sion is constant across sources, and second, that ambiguity aversion and first-order risk
aversion are positively correlated. First order risk aversion is a central concept in insur-
ance theory and determines the extent to which people are willing to accept actuarially
unfair insurance. It can also be used to distinguish decision theories. To the best of our
knowledge, we are the first to quantify first-order risk aversion.
We found support for HEU’s predictions in our main experiment where we measured
HEU’s parameters for three distinct sources of uncertainty (Canberra temperature, Paris
temperature, and attendance of the football match FC Barcelona-Atletic Bilbao) and
first- order risk aversion. We could not reject the null that the α’s were the same in the
aggregate analysis and for about 60% of the subjects their α’s were close and HEU de-
scribed their preferences well in the individual analysis. The correlation between α and
TESTING HURWICZ EXPECTED UTILITY 1395

our measure of first-order risk aversion was moderate and highly significant. The second
experiment avoided parametric assumptions, but as a result could not fully separate am-
biguity perception and ambiguity aversion. It confirmed that ambiguity attitudes were the
same across sources, but the positive correlation between ambiguity aversion and first-
order risk aversion was insignificant. This was due to the low number of subjects who
were first order risk averse in this study and the confounding impact of ambiguity per-
ception. It illustrates the importance of separating ambiguity perception and ambiguity
aversion, which is what our HEU weighting function achieves.

2. HURWICZ EXPECTED UTILITY


Let X be a set of consequences and  an (infinite) state space, with generic elements x
and ω, respectively. We refer to subsets of E ⊂  as events. An act f is a function from 
to X.
A prior is a countably additive, complete, and nonatomic probability measure on some
σ-algebra of events. Let  denote the set of all priors, and for a given π ∈ , let επ denote
the σ-algebra on which π is defined.
A prior π  constitutes an extension of another prior π, if επ  ⊃ επ and π  (E) = π(E) for
all E ∈ επ . Let π be the set of priors that are extensions of π. In other words, the set π
consists of those priors that agree with π on its domain.
We consider a decision maker who has a preference relation  over F , the set of acts. In
Hurwicz Expected Utility (HEU), the decision-maker’s preferences can be characterized by
a prior μ ∈ , an ambiguity index α ∈ [0 1], and a (Bernoulli) utility function v : X → R,
such that  admits the representation:
 
W (f ) = α min v ◦ f dπ + (1 − α) max v ◦ f dπ (1)
π∈μ π∈μ

Both μ and α are unique and v is unique up to unit and scale.


Gul and Pesendorfer (2015) refer to the elements of Eμ as ideal events. The decision
maker sees these events as the least uncertain. Both an ideal event and its complement
satisfy Savage (1954) postulate P2 (the sure-thing principle). Equation (1) shows that if an
act f is εμ -measurable, which Gul and Pesendorfer (2015) call an ideal act, then HEU
reduces to subjective expected utility and the ranking between any pair of such acts does
not depend on α. Thus the utility function v embodies the decision-maker’s risk attitudes
with respect to ideal acts.
For any act f that is not ideal, decision makers apply a Hurwicz criterion to the set of
extensions of μ. In other words, they act as if they are completely ignorant about proba-
bilities beyond those implied by μ. The parameter α reflects the decision-maker’s attitude
toward ambiguity with higher values of α indicating more aversion to ambiguity (Gul and
Pesendorfer (2015, Proposition 3)). If α = 1, then the decision maker is always ambiguity
averse. If 0 < α < 1, then the decision maker exhibits both ambiguity seeking and ambi-
guity aversion. If α = 0, then the decision maker is always ambiguity seeking. HEU is a
special case of α-maxmin for the given set of priors μ .
The events in Eμ are not necessarily the ones for which probabilities are known (deci-
sion under risk). In HEU, sources are subjective and the decision maker may not see a
process generating known probabilities as the least uncertain. Moreover, different pro-
cesses generating known probabilities may be seen as different sources. This is consistent
with the evidence in Armantier and Treich (2016), that risk cannot (always) be treated as
a single source of uncertainty.
1396 H. BLEICHRODT, S. GRANT, AND J. YANG

A bet xA y is a binary act yielding outcome x if event A obtains and y otherwise.


Throughout, we assume that writing xA y implies that x  y. Let
μ∗ (B) = sup μ(E) (2)
E∈Eμ E⊂B

be the inner probability of an event B. For a bet xA y, the least favorable extension of μ
assigns μ∗ (A) to A while the most favorable extension assigns 1 − μ∗ (Ac ) to A.
Putting equation (2) into the HEU formula, equation (1) gives
   
W (xA y) = α μ∗ (A)v(x) + 1 − μ∗ (A) v(y)
     
+ (1 − α) 1 − μ∗ Ac v(x) + μ∗ Ac v(y) (3)
As equation (3) shows, the ranking of bets generally depends on the decision-maker’s
ambiguity aversion, reflected by α. For example, in the Ellsberg two-color example, an
ambiguity averse decision maker prefers drawing a red ball from the known urn to draw-
ing a red ball from the unknown urn whereas an ambiguity seeking decision maker has
the reverse preference. Two events come from the same source if such preference rever-
sals cannot occur: the ranking of the bets depends only on the decision-maker’s prior and
not on their ambiguity attitude. Formally, a prior π ∈  (and its associated σ-algebra επ )
is a source if for every (α ν), A B ∈ επ , xA y  xB y if and only if π(A) ≥ π(B).2 So, in
Ellsberg’s 2-color problem drawing a ball from the known urn and drawing a ball from the
unknown urn are different sources, while we expect that decision makers are indifferent
between betting on red and betting on black when a ball is drawn from the unknown urn
regardless of their ambiguity attitude.
The definition of a source π implies that for all A B ∈ επ , xA y ∼ xB y if and only if
π(A) = π(B). If this condition holds, the decision maker considers A and B equally
likely. Gul and Pesendorfer (2015) define sources in terms of bets for convenience, but
the definition can be extended to any simple act. If (E1 En ) is a finite partition of the
state space with all Ei ∈ επ , then two events Ei and Ej are equally likely if exchanging
the outcomes under events Ei and Ej does not change the indifference class. Or, more
formally, if xEi yEj f denotes the change of the act f ∈ F in which the outcome for any
state in the event Ei is replaced by x and the outcome for any state in the event Ej by y,
then Ei and Ej are viewed as equally likely if xEi yEj f ∼ yEi xEj f for all pairs of outcomes x
and y. This corresponds to the definition of exchangeability proposed by Chew and Sagi
(2008) and tested (and supported) by Abdellaoui et al. (2011).
Proposition 2 in Gul and Pesendorfer (2015) shows that every source is characterized
by a power series.∞That is, for each source π there exists a sequence (a1 a2 an )
with an ∈ [0 1], n=1 an = 1, such that for all A ∈ επ ,

  ∞
 n
μ∗ (A) = γ π(A) = an π(A) (4)
n=1

Let γπ be the power series of source π. Substituting equation (4) into equation (3) gives
      
W (xA y) = α γπ (π (A) v(x) + (1 − γπ π (A) v(y)
       
+ (1 − α) (γπ 1 − π (A) v(x) + (1 − γπ 1 − π (A) v(y) (5)

2
In addition, Gul and Pesendorfer (2015) impose that if An ∈ επ and xAn y converges pointwise to xA y then
W (xA y) is the limit of the sequence W (xAn y).
TESTING HURWICZ EXPECTED UTILITY 1397

If we replace in equation (5) event A by an equally likely source event B, that is, π(A) =
π(B), then W (xA y) = W (xB y). The value of the bets depends only on the likelihoods of
the events and not on ambiguity attitudes. Hence, probabilistic sophistication holds within
sources. Because the ai in a source’s power series are nonnegative (and at least one is
strictly positive), equation (5) is strictly increasing in π(A) and all decision makers prefer
to bet on more likely events.
Proposition 4 in Gul and Pesendorfer (2015) shows that the power series characterizes
the ambiguity perception of decision makers. It states that one source π is more uncertain
than another source π 0 , if γπ , the power series associated with π is dominated by γπ 0 , the
power series associated with π 0 . That is, γπ (p) ≤ γπ 0 (p) for all p.
Because decision makers behave probabilistically sophisticated (Machina and Schmei-
dler (1992)) for sources, they use the prior to reduce acts to “objectively” risky lotteries.
So, for a given source π, any act f for which f −1 (x) ∈ επ for all outcomes x, is evaluated as
the lottery P (x) = π(f −1 (x)). Under HEU (Gul and Pesendorfer (2015), Proposition 6),
these lotteries are evaluated by rank-dependent utility:
  
V (P) = w P (x) − w P (x) · v(z) (6)
z∈R x:v(x)≥z v(x)>z

where v is the (Bernoulli) utility index from equation (1) and w is the probability weight-
ing function given by
w(p) = αγ(p) + (1 − α) γ̂(p) (7)
In equation (7), α is the decision maker’s ambiguity aversion parameter from equation
(1), γ is the power series associated with the source π, and γ̂ is the dual of γ: γ̂(p) = 1 −
γ(1 − p). The right-hand side of equation (6) reduces to expected utility when γ(p) = p,
that is, when w(p) is the identity function.
Equations (6) and (7) show that utility and ambiguity aversion are the same for all
sources. However, the probability weighting function is source-dependent, since it de-
pends on the decision-maker’s perception of ambiguity, γ, which can vary across sources.
As the power series γ is convex, its dual γ̂ is concave, and hence, each probability
weighting function is a weighted average of a convex and a concave function. Therefore, it
can allow for the inverse S-shape that is usually observed in empirical studies of probabil-
ity weighting (Gonzalez and Wu (1999), Abdellaoui (2000), Bleichrodt and Pinto (2000),
Fox and Poldrack (2014)).
Segal and Spivak (1990) showed that probability weighting is related to first order risk
aversion, the tendency to buy full insurance at actuarially unfair prices. More convex
probability weighting (which follows from a higher value of α in HEU) leads to a more
pronounced kink toward the origin in decision-makers’ indifference curves and, conse-
quently, to more first order risk aversion. In particular, for any pair of HEU decision
makers with the same ambiguity perception (for a given source) and the same utility
function, the more ambiguity averse one (the one with higher α) will exhibit more first
order risk aversion. In other words, HEU predicts a positive correlation between α and
first order risk aversion, ceteris paribus.

The HEU Weighting Function. We test these two predictions of HEU: that the ambi-
guity aversion parameter α is source-independent and that α is positively correlated with
first order risk aversion. To do so, we need to measure HEU. The crucial issue to do so
is the shape of the power series in equation (4). Since equation (4) can have an infinite
1398 H. BLEICHRODT, S. GRANT, AND J. YANG

number of positive terms, but experimental data are finite, we have to restrict the number
of terms. We used the following parametric function to describe the power series:

(1 − β)p
γβ (p) = (8)
(1 + β)(1 − p) + (1 − β)p

The following lemma show that γβ in equation (8) is indeed a power series.

LEMMA 1: The coefficients of the Taylor series of γβ (x) = (1−β)x


(1+β)(1−x)+(1−β)x
are always pos-
itive and sum to 1.

The proof of Lemma 1 is in the Appendix. Lemma 1 shows that equation (8) is indeed
a suitable power series for HEU. The function γβ is decreasing in β, so larger values
of β imply that a source is perceived as more ambiguous. Equations (7) and (8) jointly
determine the following two-parameter probability weighting function:

α(1 − β)p (1 − α)(1 + β)p


w(p) = + (9)
(1 + β)(1 − p) + (1 − β)p (1 − β)(1 − p) + (1 + β)p

In equation (9), the parameter α reflects ambiguity aversion and the parameter β am-
biguity perception. Hence, this two-parameter weighting function has an intuitive inter-
pretation. This is an advantage over other weighting functions used in the literature. For
example, in Prelec’s (1998) two-parameter weighting function, the interpretation of the
two parameters is less clear and they interact. In what follows, we will refer to equation
(9) as the HEU weighting function.

3. EXPERIMENTS
In two experiments, we tested whether the ambiguity aversion parameter α is source-
independent and whether it is positively correlated with first order risk aversion. In the
main experiment, the parametric experiment, we used a parametric method to estimate
α. In the second experiment, the nonparametric experiment, we tested the robustness of
the findings from the first experiment by measuring α by a nonparametric method, which
however could not fully separate ambiguity aversion and ambiguity perception.
Prior to the experiments, we performed two pilot studies to fine-tune our experimen-
tal designs. First, we measured HEU using the data of Abdellaoui et al. (2011), which
included as sources risk, Ellsberg urns, home and foreign temperature, and the stock ex-
change. This showed that our method to measure the parameters α and β indeed worked
and that the HEU weighting function fitted well. The data in Abdellaoui et al. (2011) did
not allow us to measure first order risk aversion. We, therefore, performed a second pi-
lot study to explore how we could best measure first order risk aversion. Here, we used
the tasks of Loomes and Segal (1994) and derived a measure of first order risk aversion
from these. We describe these pilot studies and their results in the Online Supplementary
Material (Bleichrodt, Grant, and Yang (2023)) in the Appendix.
The experiments combined the insights from these pilot studies. The first experiment
measured α using the HEU weighting function and tested whether it was equal across
the sources and positively correlated with first order risk aversion as measured by the test
that performed best in our pilot study of first order risk aversion. In the second exper-
iment, we did not use the (parametric) HEU weighting function, but measured α using
TESTING HURWICZ EXPECTED UTILITY 1399

a nonparametric method, and again tested whether it was constant across sources and
correlated with first order risk aversion. The instructions of both experiments are in the
Online Appendix.

3.1. The Parametric Experiment


3.1.1. Subjects and Incentives
We ran the experiment between the 16th and the 24th of February 2022. Subjects were
48 students from the subject pool of the Australian National University (ANU) Behav-
ioral Econ Lab coming from different academic backgrounds. Due to Covid restrictions,
we ran the experiment through Zoom with a maximum of 4 subjects per online session.
Subjects were asked to put their camera on so we could monitor them during the experi-
ment. They could invite the interviewer to a break out room or send private messages to
the interviewer if they had questions. Subjects received a showup fee of 20 AUD (approx-
imately 14 USD). In addition, we randomly selected one subject who played out one of
their randomly selected choices for real. Subjects took on average 40 minutes to complete
the experiment.

3.1.2. Procedure
Measuring Beliefs. The experiment consisted of 3 parts to measure HEU for 3 differ-
ent sources of uncertainty and a fourth part in which we measured subjects’ degree of first
order risk aversion. For each part, we started with an explanation and two comprehension
questions. After they had answered these correctly, subjects moved on to a practice ques-
tion. After this practice question, the actual experiment started.
The sources of uncertainty were the maximum temperature in Canberra on the 5th of
May 2022, the maximum temperature in Paris on the 5th of May 2022 and the attendance
of the football match FC Barcelona against Athletic de Bilbao on the 27th of February
2022. We selected this match because Barcelona has the largest stadium in Europe (ca-
pacity 99,354), Europe was in the middle of the omicron crisis and it was unclear how this
might affect stadium capacity, Bilbao is a good team but not top, and in previous seasons
the match was never sold out, and Barcelona was going through a rough spell after Lionel
Messi, one of the best players in history and seven times winner of the golden ball for
best football player in the world (the last time in 2021) had left and their coach, Ronald
Koeman, had just been fired.3
We did not include risk as a source. According to HEU, if a DM perceives a source as
“objectively” risky, with known probabilities, then they should evaluate it by SEU and α is
not identifiable. If the source risk is not perceived as “objectively” risky, then there is no
particular reason to consider it separately. We performed an overall test of whether α is
constant across sources, on both our data and the data of the experiments of Abdellaoui
et al. (2011), which include risk (see the Online Appendix). The results were consistent
with those presented below.
For each source, we elicited events with probabilities 18 , 14 , 12 , 34 , and 78 using the ex-
changeability method of Abdellaoui et al. (2011). We explain the procedure for the atten-
dance of the football match. The procedures for Canberra and Paris temperature were
similar. We started by splitting the state space into two complementary events with prob-
ability 12 . Figure B.1 in the Appendix shows how we asked the questions. Subjects were

3
We told subjects the attendance of Barcelona–Bilbao in the previous 10 seasons.
1400 H. BLEICHRODT, S. GRANT, AND J. YANG

presented with two binary acts. In Option A, they win AUD1000 if attendance is at most
x (and nothing otherwise), in Option B they win AUD1000 if attendance is more than x
(and nothing otherwise). We varied x until subjects were indifferent between A and B.
Let a 1 denote the attendance for which this indifference holds and let E 1 = [0 a 1 ] de-
2 2 2
note the event that the attendance is at most a 1 . Assuming that match attendance is a
2
source πf (we check this later), we get by equation (5) that πf (E 1 ) = πf (E c1 ) = 12 .
2 2
We then split the event E 1 into two equally likely events by finding the attendance
2
a 1 such that subjects were indifferent between winning AUD1000 if the attendance was
4
between 0 and a 1 and winning AUD1000 if the attendance was between a 1 and a 1 . It
4 4 2
follows that the events E 1 = [0 a 1 ] and (a 1 a 1 ] have subjective probability 14 .
4 4 4 2
To find the event E 3 with probability 34 , we split the event E c1 into two equally likely
4 2
events by finding the attendance a 3 for which subjects were indifferent between betting
4
on (a 1 a 3 ] and betting on (a 3 99 354]. As both of these events have probability 14 , the
2 4 4
event E 3 = [0 a 3 ] has probability 34 . Finally, to determine the events E 1 and E 7 with
4 4 8 8
probabilities 18 and 78 , we split the events E 1 and E c3 into equally likely events.
4 4
The above procedure implies that if we split the event [a 1 a 3 ] into two equally likely
4 4
events [a 1 a1 ] and (a1 a 3 ], then we should find that a1 = a 1 (except for random er-
4 2 2 4 2 2
ror). This was our test that exchangeability holds and that match attendance is indeed a
source. We used the same test to check whether Canberra temperature and Paris tem-
perature were sources. Subjects first faced the questions for Paris temperature, then for
match attendance, and finally for Canberra temperature. Within the sources, the order of
elicitation was a 1 , a 1 , a 3 , a 1 , a 7 , a1 .
8
2 4 4 8 2
All indifferences were elicited by a bisection process. The possible values for temper-
ature in Paris and Canberra varied between −10 and 40 degrees Celsius. Possible match
attendance varied between 0 and 99 354. The procedures for the starting values and the
step size in the bisection process were the same as those in Abdellaoui et al. (2011). For
any interval (a,b], we always started with a choice in which the interval was split in winning
events (a a/3 + 2b/3] and (2b/3 b] and then one in which it was split in winning events
(a 2a/3 + b/3] and (2a/3 + b/3 b]. After these choices, we used the usual bisection pro-
cedure where in each choice the step size was halved. So, for example, to determine E 1
2
for match attendance, in the first choice subjects chose between betting on [0 67k] and
betting on (67k 100k] and in the second choice between betting on [0 33k] and betting
on (33k 100k]. Depending on their choices, the bisection process then started.4

Measuring Utility and Probability Weighting. Let xp y denote the risky act that gives x
with (known) probability p and y with probability 1 − p. The certainty equivalent of an act
xE y or xp y is the amount c for sure that is equivalent to that act: c ∼ xE y or c ∼ xp y. To
determine the utility function, we measured the certainty equivalents of the five risky acts
500 1 0, 1000 1 500, 500 1 250, 700 1 500, and 1000 1 750. HEU assumes that utility is source-
2 2 2 2 2
independent, and thus, the same for risk and uncertainty. Abdellaoui et al. (2011) found
support for this assumption (see also Abdellaoui, Bleichrodt, l’Haridon, and Van Dolder

4
So, if the subject had chosen to bet twice on Option B (the one with the higher attendance) then in the next
choice they chose between betting on [0 82K] and betting on (82K 100K]. If they now chose Option A, the
next choice was between betting on [0 74K] and betting on (74K 100K]. The process ended when the highest
attendance (temperature) for which A was chosen and the lowest attendance (temperature) for which B was
chosen was less than 2000 (1 degree). We then took the midpoint of these values as the indifference value.
TESTING HURWICZ EXPECTED UTILITY 1401

(2016)). We estimated utility by a power function U (x) = xρ , which is widely used in


empirical research and typically gives a good fit (e.g., Wakker (2008)).
To measure the probability weights, we elicited for each source the certainty equivalents
of five acts 1000E 0 with the events E equal to E 1 , E 1 , E 3 , E 1 , E 7 . Using equation (6) and
2 4 4 8 8
the estimated power utility function, we obtain the probability weights, which we can then
use to estimate α and β in the HEU weighting function equation (9).5
We determined the certainty equivalents by a bisection procedure where the starting
value was the expected value of the binary act. The process ended when the difference be-
tween the lowest amount for which the sure amount was chosen and the highest amount
for which the act was chosen was less than AUD 10. We then took the midpoint of these
values as the indifference value. Subjects first faced the five certainty equivalents ques-
tions for Paris temperature then for match attendance, then for risk, and finally for Can-
berra temperature.

Measuring First Order Risk Aversion. To measure first order risk aversion, we used the
two-color question from Loomes and Segal (1994).6 Unlike Loomes and Segal (1994),
we also needed to quantify first-order risk aversion to compute the correlation between
ambiguity aversion, as measured by α, and first-order risk aversion. Subjects faced an urn
with blue and orange chips in known proportions 0 5 + ε and 0 5 − ε with ε ∈ [0 0 5].
They had to allocate 17.30 AUD between the events “draw a blue chip” and “draw an
orange chip.” The amount 17.30 AUD was the same as in Loomes and Segal (1994) and
served to counteract responding in round numbers.
We elicited the maximal value of ε (εmax ), for which subjects chose an equal division
between blue and orange (8.65 AUD on both). The value of εmax was taken as a measure
of first-order risk aversion. The larger is εmax , the more first-order risk averse subjects
are. Loomes and Segal (1994) show that an expected utility maximizer with differentiable
utility will choose εmax equal to zero, whereas a rank-dependent utility maximizer (as in
HEU) will have εmax > 0.
We elicited εmax through bisection. In the first allocation, ε was set equal to p/2 = 0 25.
Subjects saw a colored bar on their screen of which (0 5 + ε)% was blue and (0 5 − ε)%
was orange. The numbers 0 5 + ε and 0 5 − ε were also displayed to leave no doubt what
the respective probabilities were.
Subjects first saw a screen (see Figure B.2) where the amount they could put on blue7 in-
creased in steps of 2 AUD from 6.65 AUD to 16.65 AUD (and, consequently, the amount
they could put on orange decreased from 10.65 AUD to 0.65 AUD). The allocation with
6.65 AUD on blue tested for comprehension as this allocation was first order stochasti-
cally dominated by the allocation in which 10.65 AUD was put on blue.
The bisection process varied ε until we found the maximum value ε∗ for which subjects
still preferred an equal allocation between blue and orange and the minimum value ε∗
for which they preferred an unequal allocation between blue and orange. The difference
between ε∗ and ε∗ was at most 2%.
Subjects then saw a second screen (see Figure B.3) where the amount they could put
on blue increased in steps of 0.40 AUD from 8.25 AUD (again a first order stochastically
dominated option) to 9.85 AUD. The bisection process again zoomed in at the maximum

5
We estimated α and β by nonlinear least squares.
6
Loomes and Segal (1994) also used two three color questions. We tested these in the pilot, but they led to
less consistent results.
7
In the printing version, dark gray refers to blue chips and light gray refers to orange chips.
1402 H. BLEICHRODT, S. GRANT, AND J. YANG

value ε∗ for which subjects still preferred an equal allocation between blue and orange
and the minimum value ε∗ for which they preferred an unequal allocation between blue
and orange. The difference between these two values was at most 1% and we took their
midpoint as the value of εmax .

3.1.3. Results
Tests of Exchangeability. Figure 1 shows the elicited values of a 1 and a1 . They were
2 2
close for most subjects (the dotted line shows equality), suggesting that exchangeability
held. The exceptions are subjects 12 and 16 (and subjects 5 and 9 for Paris temperature).
Excluding these two subjects did not alter our conclusions. Their estimated α’s varied
widely, violating HEU, and so keeping them in puts the bar somewhat higher for HEU.

FIGURE 1.—Tests whether exchangeability holds and a 1 and a1 are equal.
2 2
TESTING HURWICZ EXPECTED UTILITY 1403

Statistical tests confirmed that exchangeability held and Paris temperature, Canberra
temperature, and match attendance can be taken as sources. We could not reject that a 1
2
and a1 were equal (ANOVA, p = 0 79).8 Moreover, as shown in Figure 1, the correla-
2
tions between a 1 and a1 were almost perfect: they exceeded 0.82 for all sources.9 This is
2 2
comparable to the correlations in Abdellaoui et al. (2011).
In the measurement of first order risk aversion, three subjects violated first order
stochastic dominance (FOSD), two of which violated FOSD in every choice. Excluding
these three subjects did not affect our results. Their implied value of εmax was 0, that
is, consistent with expected utility. As we will explain below, HEU makes no prediction
about the correlation between α and first-order risk aversion for expected utility maximiz-
ers and so expected utility maximizers had to be excluded from the correlation analysis
anyhow.
The median power coefficient of the utility function was 0.97 showing that aggregate
utility was approximately linear (t-test, p = 0 72). Abdellaoui et al. (2011) also could
not reject the null that utility was linear. Apparently, utility does not contribute much
to explaining risk aversion once source preference, ambiguity aversion, and probability
weighting are taken into account.

The HEU Weighting Function. Figure 2 shows that the HEU function fitted the data
well for each of the three sources. This also held for most of our subjects individually.
These estimations are in the Online Appendix.
Table I compares the fit of the HEU weighting function with that of popular alterna-
tive parametric forms: the weighting functions of Goldstein and Einhorn (1987), Tver-
sky and Kahneman (1992), Prelec (1998), and the neo-additive weighting function of
Chateauneuf, Eichberger, and Grant (2007).10 We used the same symbols for the param-
eters as in the HEU weighting function to highlight similarities even though, apart from
the neo-additive weighting function, the other functional forms do not give the same clean
separation between ambiguity perception and ambiguity aversion as the HEU weighting
function. We included Tversky Kahneman’s weighting function because of its widespread
use, even though we expected it to fit worse as it has only one parameter.
The table shows that HEU, Prelec, and GE fit about equally well as measured by their
sums of squared residuals (SSR). Indeed, a two-way ANOVA with functional form and
source as factors found no significant differences between these three parametric forms
(p = 0 91). Neo-additive fitted somewhat worse, but adding the neo-additive function to
the ANOVA still gave no significant difference (p = 0 44). However, when adding the
Tversky Kahneman weighting function, Tukey’s HSD test indicated that it fitted clearly
worse (p < 0 001).11
We also determined for each subject separately which function fitted best (in terms
of SSR). Here, HEU, Prelec, and neo-additive performed best. The main thing to note
from Table I is that the HEU function fits as well as other widely used forms, which gives
support for our parametric analysis presented next.

8
We checked the robustness of all of our findings using Bayesian statistics. This led to the same conclusions.
9
In describing the strength of correlations, we use the scheme of Landis, Richard, and Koch (1977).
10
We follow the convention in the literature to restrict the parameters of the Prelec and Goldstein Einhorn
(GE) weighting functions to satisfy subproportionality (Kahneman and Tversky (1979)). The analysis with the
parameters unrestricted led to the same conclusions.
11
Using the data of Abdellaoui et al. (2011) led to the same conclusions: the fit of HEU, Prelec, and GE is
similar and better than neo-additive and Tversky Kahneman. See the Online Appendix for details.
1404 H. BLEICHRODT, S. GRANT, AND J. YANG

FIGURE 2.—HEU weighting functions for the three sources of uncertainty by mean data.

First Test: Are the α’s Source-Independent? Our data support the first prediction of
HEU, that the α’s are equal across the three sources of uncertainty. Figure 2 shows that
the overall values of α were close. Figure 3 shows that the distributions of the elicited

TABLE I
THE FIT OF THE HEU WEIGHTING FUNCTION IN COMPARISON WITH WIDELY USED ALTERNATIVES.

Sources
Paris Match Canberra
Sum of No. Sum of No. Sum of No.
Functional Form Expression SSRs Subjects SSRs Subjects SSRs Subjects

HEU Equation (9) 1.55 21 2.06 10 0.95 12


αpβ
Goldstein Einhorn αpβ +(1−p) β
1.52 5 2.01 6 0.95 7

Tversky Kahneman 1 6.67 0 8.41 0 7.52 0
(pβ +(1−p) β ) β

Prelec exp(−α(− ln p)β ) 1.42 14 1.87 20 0.91 21




⎨0 p=0
Neo-additive β(1 − α) + (1 − β)p p ∈ (0 1) 2.06 13 2.43 13 1.35 9

⎩1 p=1
TESTING HURWICZ EXPECTED UTILITY 1405

FIGURE 3.—Cumulative distribution functions of α for the three sources of uncertainty.

α’s were also close and an ANOVA test could not reject the null that they are equal
(p = 0 57). We have the same result using the data of Abdellaoui et al. (2011), where they
have four sources including risk (see the Online Appendix for details). Figure 4 shows
the correlations between the α’s for the three sources. They are moderate to substantial
and vary from 0.56 (Paris and Canberra temperature) to 0.68 (Paris temperature and
match attendance) to 0.76 (Canberra temperature and match attendance). Correlation
tests showed that they are all significantly different from zero (p < 0 01 in all cases). The
figure shows that for most subjects their α’s were indeed approximately equal.
To get a clearer picture of how many subjects behaved in line with HEU and had source-
independent ambiguity attitude, we computed for each subject the maximum difference
between their α’s. Figure 5 shows this distribution. For around 60% of the subjects, the
maximum difference was less than 0.20, which is about half the standard deviation of
the distribution of α’s that we observed both in our experiment and when using the data
of Abdellaoui et al. (2011).12 There are clearly some subjects for whom the α’s differed
substantially, but for a majority HEU’s assumption of source-independent ambiguity at-
titudes describes their preferences well.
Figure 6 shows that this good individual fit did not occur because subjects perceived the
sources as equally ambiguous. The figure shows that there was no relation between the
difference in α and the difference in β. The correlation was 0.02 and was not significantly
different from zero (p = 0 91). This analysis also shows that the parameters α and β in
the HEU weighting function are independent and measure different concepts. We see
this as an important advantage of the HEU weighting function.

Second Test: Are Ambiguity Aversion and First Order Risk Aversion Correlated? The
mean of the index of first order risk aversion was 3.11, which by a t-test was significantly
different from 0 (p < 0 01). Thus, on average subjects did not behave according to ex-
pected utility. Seventeen subjects (35.4%) had their index equal to zero. This was close to
the 37 2% who behaved in line with expected utility in the pilot.13

12
The proportion of subjects who behaved according to HEU was similar in the analysis of the data in
Abdellaoui et al. (2011). See the Online Appendix for details.
13
In Loomes and Segal (1994), 24 3% of the subjects behaved in line with expected utility. The difference
might be explained by the larger number (16) of tests in their study.
1406 H. BLEICHRODT, S. GRANT, AND J. YANG

FIGURE 4.—Correlations between the measurements of the α’s in the first experiment.

Under HEU, subjects who are first order risk neutral for a source, view all events in that
source as “ideal” and are expected utility maximizers for acts measurable for that source.
That is, γ(p) = p and it follows from equation (7) that regardless of α, w(p) = p. Hence,
for expected utility maximizers HEU predicts no relation between ambiguity aversion and
first-order risk aversion and we, therefore, removed these subjects from the analysis about
the correlation between ambiguity aversion and first order risk aversion.
Figure 7 shows the correlation between ambiguity aversion (measured by the mean
value of α across the three sources) and our measure of first order risk aversion for the
subjects who were first order risk averse. In line with the predictions of HEU, there is a
moderate positive correlation (ρ = 0 51, correlation test p < 0 01).14

14
In the Online Appendix, we show the correlations between α and first-order risk aversion for each source
separately.
TESTING HURWICZ EXPECTED UTILITY 1407

FIGURE 5.—Cumulative distribution function of the maximum difference between α’s for the three sources
of uncertainty.

3.1.4. Conclusion First Experiment and Prologue to the Second


Our first experiment confirms both predictions of HEU. We found no significant dif-
ferences between the α’s that we elicited for the three different sources and we found a
positive correlation between α and first order risk aversion. We also found that the HEU
probability weighting function fitted at least as well as widely-used alternatives in the lit-
erature (while the parameters have a clear interpretation).
The first experiment made parametric assumptions about probability weighting and
the utility function. We wanted to explore in a second experiment whether we could repli-
cate our results without these assumptions. We designed a new measurement of α that
did not make any parametric assumption and tested again whether the α’s were source-
independent and positively correlated with first order risk aversion.
The matching probability m(E) of an event E is defined as the probability p such that
xp y ∼ xE y. Dimmock, Kouwenberg, and Wakker (2016, Theorem 3.1) showed how to
measure ambiguity attitudes using matching probabilities without the need to know util-
ity. Their result is also valid under HEU. In the second experiment, we measure am-

FIGURE 6.—The relation between the difference in α and the difference in β.


1408 H. BLEICHRODT, S. GRANT, AND J. YANG

FIGURE 7.—Positive correlation between ambiguity aversion (α) and first order risk aversion..

biguity aversion as the difference between 12 and the matching probability of an event
with subjective probability 12 . Our method shows how ambiguity attitudes can be mea-
sured using matching probabilities while controlling for beliefs. Baillon, Huang, Selim,
and Wakker (2018) showed how this can be done under the neo-additive weighting func-
tion. Our method extends this to general weighting functions. It measures the matching
probabilities of the event and its complement for which subjects have belief 12 and by
induction, it can be used for other values of beliefs.
A limitation of this nonparametric test is that we cannot separate the effects of am-
biguity aversion and ambiguity perception as cleanly as we did in the first experiment.
However, the first experiment showed that at the aggregate level the differences in ambi-
guity perception between the sources were modest and we, therefore, used them again in
the second experiment (except for adjustments for dates, see below).15

3.2. The Nonparametric Experiment


3.2.1. Subjects and Incentives
We ran the experiment between the 1st and the 7th of June 2022. Subjects were 34
students drawn from the ANU Behavioral Econ Lab subject pool. We ran the experiment
in person in sessions with at most 4 subjects per session.16 The incentives were the same

15
Moreover, it can be shown that if subjects are expected utility for risk, then matching probabilities for an
event with subjective probability 12 , as we use in our second experiment, maximally reflect changes in α.
16
Recruiting subjects proved difficult, because even though Covid restrictions were gradually lifted at the
time, we performed the experiment, ANU still did almost everything online, and most students avoided cam-
pus.
TESTING HURWICZ EXPECTED UTILITY 1409

as in the first experiment: subjects received a showup fee of 20 AUD (approximately 14


USD) and, in addition, we randomly selected one subject to play out one of their randomly
selected choices for real. Subjects took on average 35 minutes to complete the experiment.

3.2.2. Procedure
Measuring α Nonparametrically. The experiment consisted of two parts. First, we de-
termined for the three sources of uncertainty the events with subjective probability 12 and
their corresponding matching probabilities. Second, we measured subjects’ first-order risk
aversion using the same method as in the first experiment. The sources were the maxi-
mum temperature in Canberra on the 4th of September 2022, the maximum temperature
in Paris on the 4th of September 2022, and the attendance at the first home match of FC
Barcelona in the 2022–2023 season.17 For both parts, we started with an explanation and
two comprehension questions. After they had answered these correctly, subjects moved
on to a practice question. After this practice question, the actual experiment started.
To determine the events with subjective probabilities 12 , we selected an event and
elicited both its matching probability and that of its complement. For example, for match
attendance we started by determining the matching probabilities of the events that at
most 30,000 people would attend the first home match of FC Barcelona in the 2022/2023
season and that at least 30,000 people would attend this match.18 If the matching prob-
abilities differed, then we adjusted the events until the matching probabilities were the
same. The exact elicitation procedure is described in the Online Appendix.
Under HEU, if the elicited matching probabilities for an event E and its complement
E c are the same, they must be equally likely, and thus, P (E) = P (E c ) = 12 . However, by
equation (5) the elicited matching probabilities can be different from 12 due to ambiguity
attitudes. We took the difference between 12 and the matching probability as our measure
of ambiguity aversion. As mentioned above, this difference is also affected by ambiguity
perception and, consequently, is not a perfect measure, but that is the price to pay for
trying to measure α without making parametric assumptions.

3.2.3. Results
First Test: Are the α’s Source-Independent? Figure 8 shows the correlations between
the measurements of α for the three sources. The correlations are lower than in the
first experiment, but we expected this given the confounding impact of ambiguity percep-
tion. They vary from moderate (Paris–Canberra and Paris–Match) to substantial (Match–
Canberra). An ANOVA could not reject the null that the α’s were equal (p = 0 44). The
data from the second experiment confirm the support for HEU’s prediction of source-
independent ambiguity aversion that we found in the first experiment.

Second Test: Are Ambiguity Aversion and First Order Risk Aversion Correlated? Three
subjects violated first-order stochastic dominance. As these responses probably reflect
confusion, we removed them from the remaining analyses. This left 31 subjects in the

17
It was unknown at the time of the experiment when this match would be and who would be the opponent.
We told subjects what the attendance had been during Barcelona’s first home match for the past 10 years and
who the opponent had been. The actual opponent turned out to be Rayo Vallecano.
18
For Paris temperature, we started with the events “the temperature on the 4th of September 2022 is at most
20.5 degrees celsius” and “the maximum temperature on the 4th of September 2022 is at least 20.5 degrees
celsius. For Canberra temperature, we used the events “the temperature on the 4th of September 2022 is at
most 10.5 degrees celsius” and “the temperature on the 4th of September 2022 is at least 10.5 degrees celsius.
1410 H. BLEICHRODT, S. GRANT, AND J. YANG

FIGURE 8.—Correlations between the measurements of the α’s in the second experiment.

analyses. The mean of our measure of first order risk aversion was 2.24 and differed from
0, the prediction of expected utility (t-test, p < 0 001). 38.7% of the subjects had their
index equal to zero, which is about the same as in the first experiment and in the pilot. As
HEU makes no prediction for these subjects about their correlation between ambiguity
aversion and first-order risk aversion, we removed them from the subsequent analysis.
Figure 9 shows the relation between the measures of ambiguity aversion and first order
risk aversion. The overall correlation was slight and by a correlation test insignificant
(p = 0 65). The lower correlation compared with the first experiment is not surprising
given that ambiguity aversion was measured with some imprecision and there were only
19 subjects with an index of first-order risk aversion exceeding 0.19 In the pilot on first-
order risk aversion, we measured ambiguity aversion using an Ellsberg urn task. Here,
ambiguity perception also interfered and we observed no significant positive correlation

19
A Bayesian test indicated that the data were inconclusive.
TESTING HURWICZ EXPECTED UTILITY 1411

FIGURE 9.—Positive correlation between ambiguity aversion (α) and first order risk aversion.

between ambiguity aversion and first-order risk aversion. See the Online Appendix for
details. The data from experiment 2 and the pilot indicate that the separation between
ambiguity aversion and ambiguity perception that we realized in our main experiment is
important to test HEU properly.

4. CONCLUSION
Since the Schmeidler (1989) classic paper, the literature has produced a wide vari-
ety of models to account for ambiguity attitudes. Gul and Pesendorfer (2015) HEU is
a promising theoretical model that can accommodate many known behavioral phenom-
ena. It accounts for the joint occurrence of ambiguity aversion and ambiguity seeking and
for source-dependence of ambiguity attitudes. For decision under risk, HEU is equivalent
to rank-dependent utility. Consequently, it is consistent with all the deviations of expected
utility that rank-dependent utility and prospect theory for gains can explain.
In this paper, we show an additional advantage of HEU: it can be measured empirically.
One challenge of ambiguity models is that they are hard to measure and test, because they
involve abstract concepts like sets of priors (Gilboa and Schmeidler (1989), Ghirardato,
Maccheroni, and Marinacci (2004), Maccheroni, Marinacci, and Rustichini (2006)) or
(subjective) second-order distributions (Klibanoff, Marinacci, and Mukerji (2005), Nau
(2006)). This challenge also makes them hard to apply to real-world problems. HEU also
uses an abstract concept (the ideal sets), but we show that this does not inhibit the possi-
bility to measure it. The crucial step in our measurement of HEU was to operationalize
the power series that characterizes each source of uncertainty. We show how this can be
1412 H. BLEICHRODT, S. GRANT, AND J. YANG

done by adding one parameter that measures ambiguity perception to the model. This
leads to a new two-parameter specification of the probability weighting function, in which
the parameters have a clear interpretation: one measures ambiguity aversion, the other
ambiguity perception. We showed that the HEU weighting function fitted our data at
least as well as popular alternatives. Because risk can be seen as a source of uncertainty
on its own, the HEU weighting function can also be applied in decision under risk and is
an attractive alternative for the weighting functions that are commonly used in empirical
studies.
We tested two predictions of HEU: (i) that the ambiguity aversion parameter α is con-
stant across sources of uncertainty, and (ii) that α is positively correlated with first-order
risk aversion. To test the second prediction, we had to quantify first-order risk aversion,
and proposed a measure to do so.
We observed support for both predictions. Both in our main experiment and in a robust-
ness check that used a nonparametric measure of ambiguity aversion, we could not reject
the null that the α’s were the same across sources of uncertainty. Around 60% of our
subjects behaved in line with HEU and had their α’s close. For the remaining subjects,
ambiguity aversion depended on the source. To model their behavior, we might need a
more general α-maxmin model with α source-dependent. This, in turn, would be a special
case of the model in Ghirardato, Maccheroni, and Marinacci (2004) where α can depend
on the act.
We found that α and first order risk aversion were moderately positively correlated. In
the robustness check, the correlation was slight and insignificant, but this was caused by
the low number of first-order risk averse subjects and the fact that we could not fully sep-
arate ambiguity aversion and ambiguity perception in the nonparametric measurement.
Our data suggest that to properly measure HEU, ambiguity aversion and ambiguity per-
ception should be separated. As we showed in our main experiment, the HEU weighting
function can achieve this. Capturing ambiguity perception by a single parameter may be
ambitious, and sometimes a simplification. On the other hand, it offers substantial advan-
tages for applying HEU and we found that the HEU weighting function fitted the data of
most individual subjects well.
Ambiguity is rife and many real-world decisions are made in the face of uncertainty.
In spite of this, ambiguity is often ignored in such decisions. One reason is that ambi-
guity models are hard to apply. The main contribution of our paper is to bridge the gap
between the theoretical and applied literature. We show that a promising model, HEU,
that captures the main findings from the empirical literature, can be measured, and thus
applied. Moreover, we find support for its main predictions. We hope that in so doing our
paper contributes to helping people and governments make better decisions in the face
of ambiguity.

APPENDIX A: PROOFS
PROOF OF LEMMA 1: First, γβ (1) = 1. Notice that

1
γβ (x) =
1+β 1
−1 +1
1−β x
TESTING HURWICZ EXPECTED UTILITY 1413

If we set
1+β 1  
gβ (x) := − 1 + 1 = 1/γβ (x)
1−β x
(n) (n) (n−1)
then we see that gβ (x) = 1+β1−β
(−1) n n!x−(n+1) and gβ (x) = (−1)nx−1 gβ (x) for all
n ≥ 2.
Moreover, since γβ (x)gβ (x) = 1, it follows that
 (n)
γβ (x)gβ (x) =0 for all n ≥ 1 (10)

By the rule of derivatives, (10) is equivalent to: for all n ≥ 1,


n
n i 
n
n!
γ (x)gβ (x) =
n−i
γ i (x)gβn−i (x) = 0 (11)
i β
i=0
i!(n − i)! β i=0

For i ≥ 2, since gβn−i (x) = (−1)(n − i)x−1 gβn−i−1 (x), we have

n! n!
fβi (x)gβn−i (x) = γ i (x)(−1)(n − i)x−1 gβn−i−1 (x) (12)
i!(n − i)! i!(n − i)! β
−n (n − 1)!
= γ i (x)gβn−i−1 (x) (13)
x i!(n − i − 1)! β

Substituting each term when i > 2 in (11) by (13), we have


n
n! (1)
γ i (x)gβn−i (x) = γβn (x)gβ (x) + nfβn−1 (x)gβ (x)
i=0
i!(n − i)! β

−n  (n − 1)!
n−2

+ γ i (x)gβn−i−1 (x)
x i=0 i!(n − i − 1)! β
−n  n−1
(1) 
= γβn (x)gβ (x) + nγβn−1 (x)gβ (x) + −γβ (x)gβ (x)
x
(1) n  n−1 
= γβ (x)gβ (x) + nγβ (x)gβ (x) + γβ (x)gβ (x)
n n−1
x
(1) 1
= γβn (x)gβ (x) + nγβn−1 (x) gβ (x) + gβ (x) = 0
x
(1)
It is straightforward to check that gβ (x) + x1 gβ (x) < 0, and so by induction if γβn−1 (x) > 0,
then γβn (x) > 0, as required. Q.E.D.
1414 H. BLEICHRODT, S. GRANT, AND J. YANG

APPENDIX B: EXPERIMENTAL DISPLAYS

FIGURE B.1.—A screenshot of the experiment for exchangeable event of the match attendance.

FIGURE B.2.—A screenshot of the experiment for first order risk aversion I.
TESTING HURWICZ EXPECTED UTILITY 1415

FIGURE B.3.—A screenshot of the experiment for first order risk aversion II.

REFERENCES
ABDELLAOUI, MOHAMMED (2000): “Parameter-Free Elicitation of Utility and Probability Weighting Func-
tions,” Management Science, 46 (11), 1497–1512. [1397]
ABDELLAOUI, MOHAMMED, AURÉLIEN BAILLON, LAETITIA PLACIDO, AND PETER P. WAKKER (2011): “The
Rich Domain of Uncertainty: Source Functions and Their Experimental Implementation,” American Eco-
nomic Review, 101 (2), 695–723. [1393,1396,1398-1400,1403,1405]
ABDELLAOUI, MOHAMMED, HAN BLEICHRODT, OLIVIER L’HARIDON, AND DENNIE VAN DOLDER (2016):
“Measuring Loss Aversion Under Ambiguity: A Method to Make Prospect Theory Completely Observable,”
Journal of Risk and Uncertainty, 52 (1), 1–20. [1400,1401]
ARMANTIER, OLIVIER, AND NICOLAS TREICH (2016): “The Rich Domain of Risk,” Management Science, 62
(7), 1954–1969. [1395]
BAILLON, AURÉLIEN, AND HAN BLEICHRODT (2015): “Testing Ambiguity Models Through the Measurement
of Probabilities for Gains and Losses,” American Economic Journal: Microeconomics, 7 (2), 77–100. [1394]
BAILLON, AURÉLIEN, ZHENXING HUANG, ASLI SELIM, AND PETER WAKKER (2018): “Measuring Ambiguity
Attitudes for All (Natural) Events” Econometrica, 86 (5), 1839–1858. [1408]
BLEICHRODT, HAN, AND JOSE LUIS PINTO (2000): “A Parameter-Free Elicitation of the Probability Weighting
Function in Medical Decision Analysis,” Management Science, 46 (11), 1485–1496. [1397]
BLEICHRODT, HAN, SIMON GRANT, AND JINGNI YANG (2023): “Supplement to ‘Testing Hurwicz Expected
Utility’,” Econometrica Supplemental Material, 91, https://doi.org/10.3982/ECTA19221. [1398]
CHATEAUNEUF, ALAIN, JÜRGEN EICHBERGER, AND SIMON GRANT (2007): “Choice Under Uncertainty With
the Best and Worst in Mind: Neo-Additive Capacities,” Journal of Economic Theory, 137 (1), 538–567. [1403]
CHEW, SOO HONG, AND JACOB S. SAGI (2008): “Small Worlds: Modeling Attitudes Toward Sources of Uncer-
tainty,” Journal of Economic Theory, 139 (1), 1–24. [1396]
CHEW, SOO HONG, BIN MIAO, AND SONGFA ZHONG (2017): “Partial Ambiguity,” Econometrica, 85 (4), 1239–
1260. [1394]
DIMMOCK, STEPHEN G., ROY KOUWENBERG, AND PETER P. WAKKER (2016): “Ambiguity Attitudes in a Large
Representative Sample,” Management Science, 62 (5), 1363–1380. [1407]
ELLSBERG, DANIEL (1961): “Risk, Ambiguity, and the Savage Axioms,” The Quarterly Journal of Economics,
75 (4), 643–669. [1393]
FOX, CRAIG R., AND RUSSELL A. POLDRACK (2014): “Prospect Theory and the Brain,” in Handbook of Neu-
roeconomics, ed. by Glimcher Paul and Fehr Ernst, 533–567. [1397]
FRENCH, KENNETH R., AND JAMES M. POTERBA (1991): “Investor Diversification and International Equity
Markets,” Tech. rep., National Bureau of Economic Research. [1393]
GHIRARDATO, PAOLO, FABIO MACCHERONI, MASSIMO MARINACCI (2004): “Differentiating Ambiguity and
Ambiguity Attitude,” Journal of Economic Theory, 118 (2), 133–173. [1394,1411,1412]
1416 H. BLEICHRODT, S. GRANT, AND J. YANG

GILBOA, ITZHAK, AND MASSIMO MARINACCI (2016): “Ambiguity and the Bayesian Paradigm,” in Readings in
Formal Epistemology. Springer, 385–439. [1393]
GILBOA, ITZHAK, AND DAVID SCHMEIDLER (1989): “Maxmin Expected Utility With Non-Unique Prior,” Jour-
nal of Mathematical Economics, 18 (2), 141–153. [1411]
GOLDSTEIN, WILLIAM M., AND HILLEL J. EINHORN (1987): “Expression Theory and the Preference Reversal
Phenomena,” Psychological Review, 94 (2), 236–254. [1394,1403]
GONZALEZ, RICHARD, AND GEORGE WU (1999): “On the Shape of the Probability Weighting Function,”
Cognitive Psychology, 38 (1), 129–166. [1397]
GUL, FARUK, AND WOLFGANG PESENDORFER (2015): “Hurwicz Expected Utility and Subjective Sources,”
Journal of Economic Theory, 159, 465–488. [1393-1397,1411]
HEATH, CHIP, AND AMOS TVERSKY (1991): “Preference and Belief: Ambiguity and Competence in Choice
Under Uncertainty,” Journal of Risk and Uncertainty, 4 (1), 5–28. [1393]
HURWICZ, LEONID (1951): “Some Specification Problems and Applications to Econometric Models,” Econo-
metrica, 19 (3), 343–344. [1394]
KAHNEMAN, DANIEL, AND AMOS TVERSKY (1979): “Prospect Theory: An Analysis of Decision Under Risk,”
Econometrica, 47, 263–291. [1403]
KLIBANOFF, PETER, MASSIMO MARINACCI, AND SUJOY MUKERJI (2005): “A Smooth Model of Decision Mak-
ing Under Ambiguity,” Econometrica, 73 (6), 1849–1892. [1411]
KOCHER, MARTIN G., AMREI MARIE LAHNO, AND STEFAN T. TRAUTMANN (2018): “Ambiguity Aversion Is
Not Universal,” European Economic Review, 101, 268–283. [1393]
LANDIS, J. RICHARD, AND GARY G. KOCH (1977): “The Measurement of Observer Agreement for Categorical
Data,” Biometrics, 33 (1), 159–174. [1403]
LOOMES, GRAHAM, AND UZI SEGAL (1994): “Observing Different Orders of Risk Aversion,” Journal of Risk
and Uncertainty, 9 (3), 239–256. [1398,1401,1405]
LUCE, R. DUNCAN, AND HOWARD RAIFFA (2012): Games and Decisions: Introduction and Critical Survey.
Courier Corporation. [1394]
MACCHERONI, FABIO, MASSIMO MARINACCI, AND ALDO RUSTICHINI (2006): “Ambiguity Aversion, Robust-
ness, and the Variational Representation of Preferences,” Econometrica, 74 (6), 1447–1498. [1411]
MACHINA, MARK J., AND DAVID SCHMEIDLER (1992): “A More Robust Definition of Subjective Probability,”
Econometrica, 60 (4), 745–780. [1397]
NAU, ROBERT F. (2006): “Uncertainty Aversion With Second-Order Utilities and Probabilities,” Management
Science, 52 (1), 136–145. [1411]
PRELEC, DRAZEN (1998): “The Probability Weighting Function,” Econometrica, 66, 497–528. [1394,1403]
QUIGGIN, JOHN (1982): “A Theory of Anticipated Utility,” Journal of Economic Behavior & Organization, 3 (4),
323–343. [1394]
SAVAGE, LEONARD J. (1954): The Foundations of Statistics. John Wiley & Sons. [1395]
SCHMEIDLER, DAVID (1989): “Subjective Probability and Expected Utility Without Additivity,” Econometrica,
57, 571–587. [1394,1411]
SEGAL, UZI, AND AVIA SPIVAK (1990): “First Order versus Second Order Risk Aversion,” Journal of Economic
Theory, 51 (1), 111–125. [1397]
TRAUTMANN, STEFAN T., AND GIJS VAN DE KUILEN (2015): “Ambiguity Attitudes,” in The Wiley Blackwell
Handbook of Judgment and Decision Making, Vol. 1, 89–116. [1393]
TVERSKY, AMOS, AND DANIEL KAHNEMAN (1992): “Advances in Prospect Theory: Cumulative Representation
of Uncertainty,” Journal of Risk and Uncertainty, 5 (4), 297–323. [1394,1403]
WAKKER, PETER P. (2008): “Explaining the Characteristics of the Power (CRRA) Utility Family,” Health Eco-
nomics, 17 (12), 1329–1344. [1401]

Co-editor Alessandro Lizzeri handled this manuscript.


Manuscript received 6 December, 2020; final version accepted 14 November, 2022; available online 20 December,
2022.

You might also like