Estimating ambiguity preferences and perceptions in multiple prior models: Evidence from the field

Estimating ambiguity preferences and perceptions
in multiple prior models: Evidence from the field
Stephen G. Dimmock1
& Roy Kouwenberg2,3
&
Olivia S. Mitchell4
& Kim Peijnenburg5
Published online: 16 December 2015
# Springer Science+Business Media New York 2015
Abstract We develop a tractable method to estimate multiple prior models of decision-
making under ambiguity. In a representative sample of the U.S. population, we measure
ambiguity attitudes in the gain and loss domains. We find that ambiguity aversion is
common for uncertain events of moderate to high likelihood involving gains, but
ambiguity seeking prevails for low likelihoods and for losses. We show that choices
made under ambiguity in the gain domain are best explained by the α-MaxMin model,
with one parameter measuring ambiguity aversion (ambiguity preferences) and a
second parameter quantifying the perceived degree of ambiguity (perceptions about
ambiguity). The ambiguity aversion parameter α is constant and prior probability sets
are asymmetric for low and high likelihood events. The data reject several other
models, such as MaxMin and MaxMax, as well as symmetric probability intervals.
Ambiguity aversion and the perceived degree of ambiguity are both higher for men and
for the college-educated. Ambiguity aversion (but not perceived ambiguity) is also
positively related to risk aversion. In the loss domain, we find evidence of reflection,
implying that ambiguity aversion for gains tends to reverse into ambiguity seeking for
losses. Our model’s estimates for preferences and perceptions about ambiguity can be
used to analyze the economic and financial implications of such preferences.
J Risk Uncertain (2015) 51:219–244
DOI 10.1007/s11166-015-9227-2
Electronic supplementary material The online version of this article (doi:10.1007/s11166-015-9227-2)
contains supplementary material, which is available to authorized users.
* Roy Kouwenberg
roy.kou@mahidol.ac.th
1
Nanyang Technological University, Nanyang Business School, Singapore, Singapore
2
College of Management, Mahidol University, 69 Vipawadee Rangsit Road,
Bangkok 10140, Thailand
3
Erasmus School of Economics, Erasmus University Rotterdam, Rotterdam, Netherlands
4
The Wharton School, University of Pennsylvania, NBER, Philadelphia, PA, USA
5
Bocconi University, NETSPAR, and IGIER, Milan, Italy

Keywords Ambiguity. Decision-making under uncertainty. Multiple prior models .
Alpha-MaxMin model
JEL Classification D81 . C93 . C23
1 Introduction
Ambiguity about probability distributions for uncertain outcomes plays a central role in
many important decisions. A common way of modeling choice under ambiguity is to
assume that decision-makers consider multiple possible probability distributions, fol-
lowing Gilboa and Schmeidler (1989). These multiple prior models posit that decision-
makers consider the prior distributions yielding the worst expected utility (MaxMin
model), the highest expected utility (MaxMax model), or some weighted average of
both extremes (α-MaxMin). 1
Although multiple prior models are widely used in
theoretical studies of decision-making under ambiguity,2
relatively little is known about
how to calibrate or estimate the parameters of multiple prior models, particularly those
which distinguish between preferences and perceptions. This paper proposes a simple
and tractable method to estimate the α-MaxMin model. We then estimate the model’s
parameters using choices made in a large representative survey of the U.S. population.
Our method allows us to estimate both preferences towards ambiguity and percep-
tions about the level of ambiguity. We model these with two parameters: α indicates a
decision-maker’s dislike of ambiguity, and δ gauges his beliefs about the degree of
ambiguity relative to a reference probability distribution (i.e., the decision-maker’s
confidence in this distribution). Chateauneuf et al. (2007) introduce the specification of
α-MaxMin used in this paper and they show it is equivalent to a Choquet expected
utility model with a neo-additive capacity, providing a solid axiomatic (i.e., behavioral)
foundation.
Our first major contribution is to develop a simple method to estimate these two
parameters, α and δ, without requiring any information about utility or risk aversion.
The procedure is as follows: We measure ambiguity attitudes using matching proba-
bilities, applying the elicitation method of Dimmock et al. (2015b). Our interactive
questions follow the traditional Ellsberg two-urn problem, presenting respondents with
simple binary choices between betting on the color of a ball drawn from an urn with a
known composition of balls, versus an urn with an unknown composition. We also vary
the likelihood of the ambiguous event from low, to medium, to high, resulting in three
separate measurements. In the context of the α-MaxMin model, we then mathemati-
cally derive how the matching probability of an ambiguous event is a function of that
event’s likelihood, as well as the two key model parameters α and δ. We show how the
resulting equation can be easily estimated empirically with an ordinary least squares
regression, providing estimators for α and δ. Further, we derive how alternative
1
See Ghirardato et al. (2004) for a behavioral foundation of the α-MaxMin model. In this paper we use the
specification of the α-MaxMin model by Chateauneuf et al. (2007).
2
For instance, regarding consumption and investment see Dow and Werlang (1992) and Epstein and Wang
(1994). On healthcare investment see Asano and Shibata (2011).
220 J Risk Uncertain (2015) 51:219–244

assumptions about the prior distribution set translate into different, empirically testable,
constraints on the regression coefficients.
Our second major contribution is to estimate these two key parameters for ambiguity
preferences (α) and perceptions (δ) using a large representative survey of the U.S.
population, so they can be used to calibrate the α-MaxMin model in other studies. For
this purpose we fielded a custom-designed module in the American Life Panel (ALP),
implementing the ambiguity questions with real incentives for all participants. On
average respondents are only slightly ambiguity averse, while there is strong hetero-
geneity in ambiguity attitudes in the population: 52% of the respondents are ambiguity-
averse, 38% ambiguity-seeking, and only 10% are ambiguity-neutral. An important
conclusion is that ambiguity aversion is not universal, suggesting that a fruitful avenue
for future research would be to relax this common assumption made in many of the
ambiguity models used in economics and finance.
In addition to the traditional Ellsberg problem with balls of two different colors, our
survey module also included questions asking the subject to choose between betting on
urns containing 10 different ball colors. With these questions, we measure subjects’
attitudes towards ambiguous events of low likelihood (winning if one of 10 colors is
drawn) and high likelihood (winning if any of nine of 10 colors is drawn). For uncertain
events with high likelihood, we find that most respondents make ambiguity-averse
choices (58% of the ALP sample), but for uncertain events with low likelihoods
ambiguity-seeking choices prevail (60%). These results are consistent with prior
studies, mostly conducted in the laboratory with students, summarized in Trautmann
and van de Kuilen (2015).
These observed choices can be explained with the α-MaxMin model with a
constant ambiguity aversion parameter α and asymmetric prior probability sets for
low and high likelihood events. At first glance, the same person making ambiguity-
averse choices for high likelihood events but ambiguity-seeking choices for low
likelihoods seems inconsistent with the α-MaxMin model, in which the decision-
maker’s degree of ambiguity aversion is constant. Nevertheless, we show that the α-
MaxMin model can explain this pattern of choices while keeping the ambiguity
aversion parameter α constant across likelihoods, by modelling beliefs about the level
of ambiguity with the prior distribution set of Chateauneuf et al. (2007). The key aspect
of the model is that the prior probability set is asymmetric for events of low and high
likelihood. For example, when winning if one out of 10 colors is drawn, the set of prior
probabilities in the calibrated model ranges from six to 46%. Independent estimates of
prior distributions in Andersen et al. (2012) confirm such skewness in beliefs for low
likelihood events. Further, our data clearly reject several other variants of the models,
such as MaxMin and MaxMax, as well as alternative specifications of the prior
distribution set, such as symmetric probability intervals. Again, suggesting future
applications of ambiguity models in economics and finance should use the more
general α-MaxMin model instead of the more commonly used MaxMin model.
Specifically, in a model with one representative decision-maker, we estimate an
ambiguity aversion parameter α of 0.56, combined with 60% confidence in the
reference probability (δ=0.40). Taken together, these estimates imply that, on average,
the U.S. population is slightly ambiguity averse, but with ambiguity-seeking choices
prevailing for low likelihoods, and strong ambiguity-averse behavior for high
likelihoods.
J Risk Uncertain (2015) 51:219–244 221

Our ALP module also included ambiguity questions involving hypothetical losses to
test whether ambiguity attitudes towards gains and losses are different, that is, to test for
reference dependence.3
We find evidence of reflection: that is, ambiguity aversion for
gains reverses into ambiguity seeking for losses. This resembles the reflection effect for
decisions under risk documented by Tversky and Kahneman (1992): most people are
risk averse for payoffs involving gains, but risk seeking for losses. To capture such
reference dependence, we adjust the α-MaxMin model by introducing separate ambi-
guity aversion parameters for gains and losses. Our results are consistent with
laboratory evidence, such as Baillon and Bleichrodt (2015) who find that ambiguity
aversion for losses and gains differ, and Kothiyal et al. (2014) who provide support for
reflection.
We then test how ambiguity preferences and perceptions relate to individual char-
acteristics, and find that ambiguity aversion is higher among men and college-educated
individuals. Ambiguity aversion is positively related to risk aversion, but the correlation
is low and it does not subsume ambiguity aversion. Furthermore, ambiguity aversion is
lower for older individuals, which may reflect learning from life experiences. The level
of perceived ambiguity is higher among men, whites, and college-educated individuals.
Overall, these results demonstrate that ambiguity aversion is not merely the result of
poor cognitive skills, since college-educated respondents perceive more ambiguity and
are more averse to it. Another relevant finding is that risk aversion is positively
correlated with ambiguity aversion, both for gains and for losses, but risk aversion is
not related to the level of perceived ambiguity.
Our paper is related to prior studies that empirically examine ambiguity models.
Potamites and Zhang (2012) and Ahn et al. (2014) estimate α-MaxMin models, but
they consider only preferences towards ambiguity and not perceptions about levels of
ambiguity. Further, unlike our method, the specifications employed in those papers
required those authors to specify a utility function and to simultaneously estimate risk
aversion. Andersen et al. (2012) parametrically estimate the shape of the prior distri-
bution, but they do not estimate ambiguity aversion. Only a few empirical papers have
considered both preferences and beliefs in models of ambiguity. Andersen et al. (2009)
examine both ambiguity preferences and prior beliefs for the smooth model of
Klibanoff et al. (2005).4
Hey et al. (2010) compare the predictive ability of several
ambiguity models, including multiple prior models, but without providing details about
the parameter estimates of any single model.5, 6
Baillon et al. (2015) estimate α and δ
3
We chose to implement our ambiguity question involving losses without real incentives to avoid house
money effects (Thaler and Johnson 1990). The house money effect refers to empirical evidence that people’s
risk taking can depend on prior gains and losses. In our setting, if we had given the respondent an initial
endowment to implement real losses, the presence of an endowment could influence people’s subsequent
decisions (e.g., more risk taking after a windfall). Exposing respondents to real losses without giving an initial
endowment raises ethical issues.
4
A related paper by Attanasi et al. (2014) tests the behavioral predictions of the smooth model using several
decision tasks with varying levels of ambiguity (perceived ambiguity).
5
The sophisticated approach of Hey et al. (2010) involves joint estimation of risk and ambiguity preferences,
as well as beliefs; implementation used 135 choice problems per subject (students). As our objective is to
measure ambiguity preferences and perceptions in a survey of the general population, we use a relatively
simple elicitation method involving only 17 choices per subject that does not require joint estimation of risk
preferences.
6
Similarly, Conte and Hey (2013) evaluate the predictive ability of several multiple prior models based on
subjects’ choices between two compound lotteries with known probabilities (i.e., two-stage lotteries).
222 J Risk Uncertain (2015) 51:219–244

for the α-MaxMin model specification of Chateauneuf et al. (2007) in a sample of
64 students. Unlike our method, their approach involves joint estimation of a utility
function with a complex non-linear likelihood maximization. Finally, Dimmock et al.
(2015a) analyze the same dataset as we do, but they focus on the relation between
ambiguity aversion and portfolio choice.
This paper is organized as follows: in Section 2 we describe the elicitation method
for ambiguity attitudes based on matching probabilities, fielded in the American Life
Panel. We also summarize the ambiguity attitudes of the U.S. population, using non-
parametric statistics that do not assume any ambiguity model. In Section 3 we model
ambiguity attitudes with the α-MaxMin specification of Chateauneuf et al. (2007). We
then derive an equation for the matching probability as a function of the two key model
parameters α and δ, and the likelihood of the ambiguous event. In Section 4 we first
show how this equation can be estimated with panel regressions, and then we estimate
α and δ for the U.S. population, using our ALP data. Section 5 concludes.
2 Measuring ambiguity attitudes
The survey module we use to measure ambiguity attitudes in the general population
was fielded in the American Life Panel (ALP), a representative panel of U.S. house-
holds that regularly answer Internet surveys.7
The module offered respondents real
rewards based on their choices. Specifically, at the outset of the survey module, all
subjects were told that one of their choices in the ambiguity questions involving gains
would be randomly selected and played for a chance to win $15. In total, real incentives
worth $23,850 were paid to 1590 of the 3258 ALP subjects.8
2.1 The elicitation procedure
The ALP module posed questions similar to those in a standard Ellsberg experiment
with two urns9
asking respondents to choose between an ambiguous Box U (Unknown)
and an unambiguous Box K (Known), following the elicitation method of Dimmock et
al. (2015b). Both boxes contained 100 balls, which could be purple or orange, as shown
in Fig. 1. The respondent was told that one ball would be randomly drawn from the box
he selected and he would win $15 if that ball was purple. For Box K, the respondent
could see its content on the screen: 50 purple balls and 50 orange balls. The content of
Box U is not shown, creating ambiguity. A preference for Box K over Box U indicates
ambiguity aversion.
Apart from choosing Box K and Box U, respondents could also choose
BIndifferent^. A choice of Indifferent implies ambiguity neutrality, with the subject
treating Box U as if the probability of winning were 50% as in Box K. In cases
where the subject did not select Indifferent in the first round, he was presented
7
For more information about how the ALP recruits respondents see the American Life Panel website at
https://mmicdata.rand.org/alp/. One advantage of the ALP is that respondents who lack Internet access are
provided with either a laptop and Internet access, or a so-called WebTV that allows them to use their television
to participate in the panel.
8
Dimmock et al. (2015a) describe the fielding of the ALP module in more detail.
9
The survey module uses Bbox^ instead of Burn,^ as the word Burn^ might be unfamiliar to some.
J Risk Uncertain (2015) 51:219–244 223

additional question rounds similar to the one in Fig. 1. For example, if the respon-
dent selected Box K in Fig. 1, the known probability of winning was then reduced
to 25% in the second round; if he chose Box U, the known probability of winning
was increased to 75%. This process was repeated for up to four rounds, to closely
approximate the respondent’s indifference point. The known probability that makes
the respondent indifferent when comparing Box K and Box U is defined as the
matching probability.10
Let m50
denote the matching probability for Question 1. We can then measure
ambiguity aversion as follows: AA50
=50%−m50
. Positive values of AA50
imply
ambiguity aversion, negative values indicate ambiguity-seeking behavior, and a value
of zero shows ambiguity neutrality. For example, consider a respondent whose
matching probability is 30% (m50
=0.3), implying he is indifferent between Box K
and Box U when Box K offers a known probability of winning of exactly 30%. He only
prefers the ambiguous Box U when Box K offers a probability of winning less than
30%. This behavior is consistent with ambiguity aversion.
Fig. 1 First ambiguity question: winning for one of two ball colors. Notes: This figure shows the first round in
the ambiguity question sequence with two ball colors. The respondent can win a prize of $15 if a purple ball is
drawn from the box of his preference. Box K contains 50 purple and 50 orange balls, offering 50% initial
known probability of winning. Box U also contains purple and orange balls, but with the proportions
unknown. If the respondent selects BBox K^ or BBox U^, a second question round follows, similar to the
one shown above. If the response is BBox K^, in the second round the probability of winning for Box K is
decreased (fewer purple balls). Vice versa, when the respondent selects BBox U^, in the second round Box K
offers a higher probability of winning (more purple balls). Selecting the BIndifferent^ button takes the
respondent to the next ambiguity sequence, shown in Fig. 2
10
This approach is similar to that of Kahn and Sarin (1988), Baillon et al. (2012), Baillon and Bleichrodt
(2015), and Dimmock et al. (2015b).
224 J Risk Uncertain (2015) 51:219–244

We note that, in theory, a preference for Box K over Box U in the first question
round shown in Fig. 1 can be reconciled with subjective expected utility if a subject
assigned a low subjective probability to drawing a purple ball from Box U (e.g.,
perhaps due to distrust of the surveyor). Nevertheless, the Ellsberg (1961) paradox
then arises, because people are also indifferent between betting on drawing an orange
ball from Box U versus betting on drawing a purple ball from Box U, which implies the
subjective probabilities sum to less than 100%. Both lab experiments and field studies
confirm that people are indeed indifferent about the winning color for Box U.11
To keep the questions as simple as possible for a survey of the general population,
we did not include an option to change the winning color. The ALP survey is
administered by the RAND Corporation, which our respondents trust to provide the
payoffs as they have participated in other previous surveys with this organization. To
test the assumption of no color preferences, we fielded an additional survey in which
250 respondents had the option to select their winning color, and a control group of 250
respondents did not. We find no significant differences in ambiguity attitudes between
the two groups.12
The elicitation method presents subjects with a series of binary choices converging
to the matching probability, instead of directly asking for the matching probabilities, as
prior studies have shown that this produces more reliable measures of preferences (see,
e.g., Bostic et al. 1990). Because eliciting ambiguity preferences is known to be
sensitive to measurement error and within-person inconsistencies (see, e.g., Binmore
et al. 2012; Stahl 2014), the module included two consistency check questions. These
check questions are similar to Fig. 1, but with the known probabilities of winning for
Box K changed to m50
+10% and m50
–10%, respectively, where m50
is the subject’s
matching probability. A preference for Box U in the first check question is inconsistent
with earlier choices, as is a preference for Box K in the second check question.
2.2 Eliciting ambiguity attitudes for low and high likelihood events
Although exact probabilities for ambiguous events are unknown, it is often still
possible to assess whether an event is unlikely (e.g., the U.S. inflation rate exceeding
10% next year) or highly likely (e.g., the U.S. having positive GDP growth next year).
Prior studies reveal that most people are ambiguity seeking for low likelihood ambig-
uous events, while ambiguity aversion is common for high likelihood events. Such
preferences can be interpreted as a tendency to regard all ambiguous events as if they
are 50–50%, or showing too little sensitivity to the likelihood of events (Abdellaoui et
al. 2011). This effect is called ambiguity-generated likelihood insensitivity, or a-
insensitivity.
11
See, for example, Abdellaoui et al. (2011) and Dimmock et al. (2015b). The latter study finds that fewer
than two percent of respondents changed the winning color.
12
In August 2013, we fielded an additional survey (N=500) identical to the original except one: it offered
some respondents a choice for the winning ball color. Specifically, a randomly selected half of the sample was
allowed to select the winning color (purple or orange), while the other half could not. Fewer than one percent
of the respondents in the group allowed to change the color did so, and the mean matching probabilities of the
‘color choice’ and ‘no color choice’ subgroups were not significantly different. Results are available upon
request.
J Risk Uncertain (2015) 51:219–244 225

To measure a-insensitivity the survey module included two additional sets of
questions with boxes containing balls of 10 different colors. For instance, Fig. 2 shows
the second ambiguity question, where the respondent won $15 if a purple ball was
drawn. Here Box K offers a known probability of winning of 10%, while for Box U the
probability is unknown. The third question is similar, but with the outcomes reversed:
now, the subject won $15 if any of the nine colors other than purple was drawn, and
thus for Box K the initial probability of winning is 90%. We define m10
and m90
as the
matching probabilities for these two questions, which are calculated after a maximum
of four rounds of bi-section.
The ambiguity aversion measures for these two questions are defined as: AA10
=
10% –m10
, and AA90
=90% –m90
. Subjects who selected Indifferent in the first round
are ambiguity neutral, with AA10
=0 and AA90
=0. A-insensitivity predicts that subjects
are ambiguity seeking for low likelihood ambiguous events (AA10
<0), and
ambiguity averse for high likelihood events (AA90
>0). Hence, we describe a
respondent as a-insensitive if AA90
–AA10
>0, a-neutral if AA90
–AA10
=0, and a-
oversensitive if AA90
–AA10
<0.
Fig. 2 Second ambiguity question: winning for one of ten ball colors. Notes: This figure shows the first round
in the second ambiguity question sequence, with 10 ball colors. Here the respondent wins if a purple ball, 1 out
of 10 colors, is drawn from the box of his preference. Box K contains 10 balls of each color and offers a 10%
probability of winning. Box U also contains balls with 10 different colors, but with the proportions unknown.
If the respondent selects BBox K^ or BBox U^, a second question round follows, similar to the one shown
above. If the response is BBox K^, in the second round the probability of winning for Box K is decreased
(fewer purple balls). Vice versa, when the respondent selects BBox U^, in the second round Box K offers a
higher probability of winning (more purple balls). Selecting the BIndifferent^ button takes the respondent to
the next ambiguity sequence: wining for nine out of ten ball colors
226 J Risk Uncertain (2015) 51:219–244

2.3 Eliciting ambiguity attitudes for losses
The last module question assesses ambiguity aversion when the outcomes involve losses.
Onceagain,subjects hada choicebetween Box K with50 purple and50 orangeballs,and Box
U with purple and orange balls in unknown proportions. The subject would now, hypothet-
ically, lose $15 if a purple ball were drawn from the chosen box. For losses, we use m−50
to
denote the matching probability, and the ambiguity aversion measure is: AA−50
=m−50
–50%.
Subjects are ambiguity averse in the loss domain if AA−50
>0, implying they are prepared to
accept a relatively high known probability of losses to avoid Box U.
When choices in the loss domain are incentivized, the standard approach is to give all
respondents an initial endowment (e.g., $15) and then subtract subsequent losses. A draw-
back of this approach is that it can introduce a house money effect. Additionally, rational
respondents who evaluate outcomes in terms of terminal wealth will not experience any
losses but only gains and zero outcomes. For these reasons, we do not implement real losses
in our module. Etchart-Vincent and l’Haridon (2011) have extensively tested the effect of real
incentives in controlled experiments for both gains and losses, and they found that real and
hypothetical choices differ significantly only in the gain domain, but not in the loss domain.13
2.4 Summary of ambiguity attitudes in the U.S. population
Table 1 summarizes ambiguity attitudes for our representative sample of the U.S.
population. Panel A reports the proportions of people whose responses indicate ambi-
guity aversion, seeking, or neutrality. For events of moderate and high likelihood
involving gains, most people’s choices were consistent with ambiguity aversion. For
example, for the first ambiguity question with two ball colors over half of the respon-
dents (52%) make ambiguity-averse choices, only 10% were ambiguity neutral, and
38% were ambiguity seeking. Hence, ambiguity attitudes are heterogeneous, and
ambiguity aversion is far from universal. This finding is consistent with other field data
studies,14
and experimental studies summarized in Trautmann and van de Kuilen (2015).
For the high likelihood event (winning if any of nine of the 10 colors is selected), 58% of
our respondents were ambiguity averse and only 30% sought ambiguity. By contrast, for the
low likelihood event (winning if one of the 10 colors is selected), 60% of respondents were
ambiguity seeking and only 19% were ambiguity averse. This provides strong evidence of
a-insensitivity, the tendency to treat all ambiguous events as equally probable. Indeed, Panel
B of Table 1 shows that 81% of the respondents were a-insensitive (AA90
>AA10
).
Panel C provides summary statistics for the matching probabilities and the ambigu-
ity aversion measures. For example, the median respondent was indifferent between a
known probability of 47% winning for Box K versus betting on one of two colors in
Box U, indicating a low degree of ambiguity aversion (median AA50
=0.50–0.47=
0.03). For the high likelihood event, the median attitude indicates relatively strong
13
Similarly, Baillon and Bleichrodt (2015) find no significant differences between ambiguity attitudes in the
loss domain measured with real incentives and with hypothetical losses.
14
Butler et al. (2014) report that 52% of 1686 Italian bank customers are ambiguity averse (without real
incentives), while in Akay et al. (2012) 57% of 92 Ethiopian farmers are ambiguity averse. In a sample of 666
subjects from the Dutch population, Dimmock et al. (2015b) find that 68% were ambiguity averse, 10%
neutral, and 22% seeking, suggesting that ambiguity aversion is more common among the Dutch.
J Risk Uncertain (2015) 51:219–244 227

ambiguity aversion (median AA90
=0.13), while for the low likelihood event the median
implies ambiguity-seeking behavior (median AA10
=−0.075).
Turning to the loss domain, Panel A reveals that ambiguity seeking is the most
common response (40%), followed by aversion (33%), and neutrality (27%). The
Table 1 Descriptive statistics for ambiguity attitudes
Panel A. Ambiguity attitudes (Proportion of respondents for each question)
Ambiguity Question Gains 10% Gains 50% Gains 90% Losses 50%
Ambiguity averse 18.5 52.4 57.7 33.6
Ambiguity neutral 21.5 9.9 12.8 26.9
Ambiguity seeking 60.0 37.7 29.5 39.6
Panel B. A-Insensitivity Percent of respondents
A-Insensitive (AA90
–AA10
>0) 80.5
Neutral (AA90
–AA10
=0) 7.5
A-Oversensitive (AA90
–AA10
<0) 12.0
Panel C: Summary statistics of matching probabilities and ambiguity aversion
Matching probabilities Mean Std. dev. Min Median Max
m10
0.241 0.207 0.015 0.175 0.850
m50
0.482 0.212 0.030 0.470 0.940
m90
0.710 0.258 0.055 0.770 0.990
m−50
0.486 0.198 0.060 0.500 0.970
Ambiguity aversion
AA10
−0.141 0.207 −0.750 −0.075 0.085
AA50
0.018 0.212 −0.440 0.030 0.470
AA90
0.190 0.258 −0.090 0.130 0.845
AA−50
−0.014 0.198 −0.440 0.000 0.470
Panel D: Correlations
Ambiguity aversion AA10
AA50
AA90
AA−50
AA10
1.00 0.44 0.20 0.26
AA50
0.44 1.00 0.32 0.25
AA90
0.20 0.32 1.00 0.19
AA−50
0.26 0.25 0.19 1.00
Notes: This table shows ambiguity attitudes in the U.S. population measured in our ALP module (see text). We
posed three ambiguity gains questions with ambiguity-neutral probabilities of 10%, 50%, and 90% for Box U,
where subjects could win $15, and one 50% question with hypothetical losses of -$15. Panel A displays the
proportion of respondents who are ambiguity averse, ambiguity seeking, or ambiguity neutral, based on their
choice between Box K and Box U in the first round of each ambiguity question. Panel B summarizes a-
insensitivity. Panel C shows summary statistics for the matching probabilities and ambiguity aversion
measures for each question separately. Panel D presents correlations of the four ambiguity aversion measures.
The sample consists of 2991 people who answered all ambiguity questions and spent at least two minutes of
time. All results use ALP survey weights
228 J Risk Uncertain (2015) 51:219–244

average degree of ambiguity aversion for losses is close to zero, due to the strong
heterogeneity in attitudes. To examine reference dependence, we test the null hypoth-
esis that mean AA50
=meanAA−50
. A paired samples t-test strongly rejects the null
hypothesis of ‘no reference dependence’ (p-value<0.01); accordingly, ambiguity
aversion differs for gains and for losses in the general population, consistent with the
experimental results of Cohen et al. (1987), Abdellaoui et al. (2005), and Baillon and
Bleichrodt (2015). We next test for reflection, or mean AA50
=−meanAA−50
, which
implies that ambiguity aversion for gains is reflected into ambiguity-seeking behavior
for losses. In this case the null hypothesis ‘reflection’ cannot be rejected (p-value=
0.411).15
Using experimental data, Vieider et al. (2012) and Kothiyal et al. (2014) also
found reflection at the aggregate and individual level for decisions under ambiguity.
Finally, Panel D of Table 1 reports positive and significant correlations between all
four question-specific measures of ambiguity attitudes. Hence, although ambiguity
aversion does differ across likelihoods and for gains versus losses, the estimates are
all positively related. This is consistent with a common underlying factor driving
ambiguity aversion across domains.
In results not reported in detail here (see Online Appendix A), we also examine
respondent consistency with the two check questions mentioned above. We find that
41% of the respondents answer both check questions correctly, while 17% give both
consistent and indifferent answers. The remaining 42% of our respondents contradict
an earlier choice once (39%) or twice (3%). The consistency rates are low, but similar to
earlier findings in the Netherlands (Dimmock et al. 2015b) and in experimental settings
(e.g., Trautmann and van de Kuilen 2015). As an important robustness check, all
estimates in Section 4 are repeated in the full sample as well as the subset of
respondents who passed both consistency checks (see Section 4.3).
3 Multiple prior models for decision-making under ambiguity
We next specify a tractable version of the α-MaxMin model, which can be estimated
using the matching probabilities observed in the survey data. The model has two
parameters: α, which measures ambiguity aversion; and δ, which measures the degree
of perceived ambiguity relative to a reference distribution in which the decision-maker
has limited confidence.
3.1 The α-MaxMin model
We model uncertainty using a state space S consisting of a finite number of future states
of the world (s1,s2,…,sn). One state s ϵ S is true and the other states are not, but which
state is true is uncertain. Subsets of the state space are called events. For example, event
E={s1,s3,s6} is true if the state is s1, s3, or s6. The set containing all possible events is
denoted by Θ. A probability measure P is a function P:Θ→ℝ that assigns real
numbers between 0 and 1 to events, with the following properties: P(Ø)=0,P(S)=1,
P(E)≤P(F) for all events E,F∈Θ with E⊆F, and P(E∪F)=P(E)+P(F)−P(E∩F). Let Γ
15
A signed rank test for the median gives similar results: p-value<0.01 for no reference dependence (median
AA50
= median AA−50
); p-value=0.13 for reflection (median AA50
=− median AA−50
).
J Risk Uncertain (2015) 51:219–244 229

denotes the set of all possible probability measures. A prospect x is a function assigning
outcomes to events, x:Θ→ℝ, giving monetary amount x(E) if event E occurs. For
example, in our experiment with ambiguous Box U, there are two states, s1=a purple
ball is drawn, and s2=an orange ball is drawn, with respective outcomes x(s1)=$15 and
x(s2)=$0. Finally, we assume that the decision-maker has an increasing utility function
U:ℝ→ℝ over outcomes.
Ambiguity occurs when the decision-maker does not know the exact probabilities of
all events E in Θ. Multiple prior models assume that the decision-maker considers a
convex set C of possible probability measures P ϵ C . In the multiple prior model of
Gilboa and Schmeidler (1989), the decision-maker uses the worst distribution in C to
evaluate prospects, as follows:
MaxMin x
ð Þ : minP∈CEP U x
ð Þ
½ ¼ minP∈C ∫s∈S
U x s
ð Þ
ð ÞdP s
ð Þ ð1Þ
where EP[U(x)] denotes the expected utility of prospect x under probability measure P,
and x is shorthand notation for x(s).
When choosing between prospects x and y, the decision-maker strictly prefers x over
y if and only if minP∈CEP[U(x)]minP∈CEP[U(y)]. Hence, the decision-maker acts as if
he maximizes the minimum expected utility value, giving rise to the name MaxMin.
The MaxMin model implies that the decision-maker is ambiguity averse or pessimistic,
as he evaluates prospects based on the prior probability distribution giving the lowest
expected utility.
Ambiguity-seeking behavior is consistent with Equation (1) if we replace the
minimum over P by the maximum, which is called the MaxMax model. The α-
MaxMin model combines these two extremes, with weight α ϵ [0,1]:
αMaxMin x
ð Þ : α minP∈CEP U x
ð Þ
½ þ 1−α
ð ÞmaxP∈CEP U x
ð Þ
½ ð2Þ
Thus the α-MaxMin model can accommodate both ambiguity-averse and
ambiguity-seeking behavior.16
Furthermore, the set of prior distributions C now reflects
perceived ambiguity, while α is a measure of aversion to this ambiguity. Maximum
ambiguity aversion occurs at the value α=1 (MaxMin), and maximum ambiguity
seeking at α=0 (MaxMax).
3.2 The multiple prior set
A complication in empirically implementing the α-MaxMin model is the specification
and estimation of the set of priors C. In a real-world setting such as investing, it can be
difficult for decision-makers to indicate exactly which prior probability distributions
they consider plausible. Previous literature on ambiguity often adopts a specification
for C termed ε-contamination (e.g., Epstein and Wang 1994). The approach assumes
that the decision-maker has a reference probability distribution, π, which is an assess-
ment of the probability of events based on his subjective beliefs. The decision-maker is
aware, however, that he does not know the true probabilities, so he considers all
16
Ghirardato et al. (2004) and Eichberger et al. (2011) provide a behavioral (i.e., axiomatic) foundation for the
α-MaxMin model.
230 J Risk Uncertain (2015) 51:219–244

distributions in the larger set Cε={Q∈Γ:Q=(1−ε)π+εP, forP∈D}, with ε ϵ [0,1] and
D is a given set of contaminating probability measures. Hence π is contaminated with
distributions from the set D, with the parameter ε determining the perceived level of
ambiguity. In empirical applications, though, researchers still must specify or estimate
the set of contaminating probability distributions D.
In our specification of perceived ambiguity, we adhere to the concept of a reference
distribution π. Following Chateauneuf et al. (2007), we assume the decision-maker has
a degree of confidence (1−δ) in π, with δ ϵ [0,1]. He then considers the set Cδ of
probability measures P assigning at least probability (1-δ)π(E) to all events E:
Cδ ¼ P∈Γ : P E
ð Þ≥ 1−δ
ð Þπ E
ð Þ; for allE∈Θ
f g; withδ∈ 0; 1
½ ð3Þ
At first glance, Cδ appears to provide only lower bounds on probabilities. But we note
that inequality (3) also holds for each event E’s complement, defined as Ec
=SE.17
Thus
P(Ec
)≥(1−δ)π(Ec
). Using P(Ec
)=1−P(E), we get 1−P(E)≥(1−δ)(1−π(E)), which gives
the following upper bound: P(E)≤1−(1−δ)(1−π(E))=(1−δ)π(E)+δ. Accordingly, the
set of priors Cδ imposes the following bounds on the probability measures P∈Cδ:
0 ≤ 1−δ
ð Þπ E
ð Þ ≤ P E
ð Þ ≤ 1−δ
ð Þπ E
ð Þ þ δ ≤ 1; for allE∈Θ ð4Þ
The prior set Cδ allows the probability P(E) to vary in an interval of length δ around
the reference probability π(E).18
We note that the reference probability π(E) does not
necessarily lie in the middle of the interval: this is only case for π E
ð Þ¼0:5 or δ=0. For
π E
ð Þ0:5, the interval will be longer above π(E) than below π(E), and vice versa when
π E
ð Þ0:5. This feature is important for explaining ambiguity-seeking behavior for low
likelihood events but ambiguity aversion elsewhere, as we will explain later.
The prior set Cδ for α-MaxMin, originally introduced by Chateauneuf et al.
(2007), is useful in empirical applications for several reasons. First, the per-
ceived level of ambiguity is captured by δ and the standard subjective expected
utility framework is a special case when δ=0. Second, instead of eliciting the
decision-maker’s entire set of possible priors C (or the set D), we simply
specify one reference distribution π and estimate the confidence level (1-δ).
In applications with Ellsberg urns, the values for π follow naturally from
symmetry conditions. In stock market applications, a good candidate for π
would be the estimated return distribution based on historical data.19
Third,
Chateauneuf et al. (2007, Remark 3.2) show that the α-MaxMin model with
prior set Cδ is equivalent to their Choquet expected utility model with a neo-
additive capacity, and they provide an axiomatic foundation for the latter.20
In this way, the
17
Ec
contains all states s except those contained in event E: E∪ Ec
=S and E ∩ Ec
=∅.
18
Note that P(E)=0 for E∈N, where N denotes a set of all null events E∈N with π(E)=0.
19
Given sufficient experimental time one could elicit a subjective measure π for uncertain events from
revealed preferences at the individual level as in Abdellaoui et al. (2011).
20
The measures of ambiguity aversion and a-insensitivity introduced by Abdellaoui et al. (2011), named Index
b and Index a, also derive from a neo-additive capacity and can therefore be linked to α and δ. That is, for their
a-insensitivity measure: Index a=δ. For their ambiguity aversion measure: Index b=(2α−1)δ. Hence, Index b
is positive if and only if α 0.5 and δ 0.
J Risk Uncertain (2015) 51:219–244 231

preference ordering implied by the model satisfies important properties such as complete-
ness, reflexivity, and transitivity.
3.3 The α-MaxMin model with prior set Cδ applied to our experiment
For our first survey question, Box U contained 100 balls, each having one of two
possible colors, purple or orange, with the proportions unknown. One ball was drawn
randomly from Box U and the respondent won $15 if the ball was purple, and nothing
otherwise. For S={s1,s2}, state s1 denotes a purple ball was drawn, and similarly s2
denotes an orange ball was drawn. The outcomes are x(s1)=$15 and x(s2)=0, evaluated
with utility function U. The utility function can be rescaled such that U(0)=0 and
U(15)0. The α-MaxMin model with prior distribution set Cδ then evaluates the
prospect x as follows:
α minp∈ 1−δ
ð Þπ; 1−δ
ð Þπþδ
½ pU 15
ð Þ þ 1−α
ð Þmaxp∈ 1−δ
ð Þπ; 1−δ
ð Þπþδ
½ pU 15
ð Þ
¼ α 1−δ
ð ÞπU 15
ð Þ þ 1−α
ð Þ 1−δ
ð Þπ þ δ
ð ÞU 15
ð Þ
¼ 1−δ
ð Þπ þ 1−α
ð Þδ
ð ÞU 15
ð Þ
ð5Þ
where π=π(s1) is the reference probability of drawing a purple ball from Box U.
Box K offers a known probability p of winning the $15 prize and it is evaluated with
expected utility (δ=0), giving: E[U(x)]=pU(15). The matching probability m is the
known probability p that makes the respondent indifferent when comparing Box K and
Box U. Respondents are indifferent between the two when ((1−δ)π+(1−α)δ)U(15)=
pU(15), so the matching probability m is:
m ¼ 1−δ
ð Þπ þ 1−α
ð Þδ ð6Þ
We note that the utility function cancels out in the comparison, so we need not
estimate utility (or risk aversion) in order to measure people’s ambiguity attitudes: this
is a major advantage of the elicitation method.
Similar to what is shown in Section 3.1, once we know the matching probability, we
can define an index of ambiguity aversion as: AA=π−m=(α−(1−π))δ. It follows that
in the special case of no perceived ambiguity, δ=0, the matching probability is π and
respondents are ambiguity neutral (AA=0). For respondents who perceive some ambi-
guity about the exact probability of winning for Box U, δ0, the ambiguity attitudes
are:
AA 0 : if α 1−π and δ 0;
AA ¼ 0 : if α ¼ 1−π and δ 0;
AA 0 : if α 1−π and δ 0:
ð7Þ
We take π¼0:5 as the relevant reference probability for drawing a purple ball from
Box U, assuming indifference between having purple or orange as the winning color.
The result above implies decision-makers perceiving some degree of ambiguity about
the number of purple balls in Box U (δ0) will be ambiguity averse in our first
ambiguity question (AA0) if α0:5, ambiguity seeking (AA0) if α 0:5, and
ambiguity neutral (AA=0) if α¼0:5.
232 J Risk Uncertain (2015) 51:219–244

We repeat this analysis for our second ambiguity question where Box U contained
100 balls with 10 different colors, and the prize of $15 was won if one particular color
was drawn. It is easy to show that Equations (5), (6), and (7) still hold. The only
difference is that the reference probability for winning for one out of 10 possible colors
is π=1/10. Equation (7) now predicts ambiguity-averse responses in the first round of
the second question if α0.9 and δ0. Hence, only individuals with very high levels of
ambiguity aversion would still prefer Box K, and ambiguity-seeking choices are
expected to prevail.
The explanation for this effect is that at π=1/10, the interval of prior probabilities
used by the decision-maker is asymmetric around π: [(1-δ)0.1, (1-δ)0.1+δ]. For
example, with a confidence level of 80% (δ=0.2), the interval is [0.08, 0.28]. The
upper bound of the interval is much further above π than the lower bound is below it.
The result is that the ‘max’ in α-MaxMin will influence the result more than the ‘min’,
and the α-MaxMin value of Box U will exceed πU(15). The decision-maker will prefer
Box U and display ambiguity seeking behavior, unless his ambiguity aversion α is
sufficiently high (precisely, only if α0.9).
The intuition for the asymmetric prior around π=1/10 follows from the fact that the
prospect offers positive skewness: on the positive side the number of winning balls in
Box U can range from 10 to 100, while on the negative side it can only range from zero
to 10. Accordingly, the interval of prior probabilities reflects that perceived ambiguity
should be higher on the upside (p0.1) than on the downside (p0.1). Andersen et al.
(2012, Fig. 6) explicitly estimated the distribution of prior probabilities for an ambig-
uous event with π=1/10 in a sample of students and found that it was indeed highly
positively skewed.
Similarly, Equations (5), (6), and (7) also apply to the third ambiguity question,
where the respondent wins if any of the nine of the 10 ball colors is chosen, with
reference probability π=9/10. We can deduce that ambiguity-averse responses will
occur in the third ambiguity question if α0.1 and δ0. The interval of prior proba-
bilities is again asymmetric. For example, with a confidence level of 80% (δ=0.2), the
interval is [0.72, 0.92]. In this case, it is natural for the decision-maker to perceive more
ambiguity on the downside than on the upside, as the number of winning balls in U
could be far less than the reference number of 90 balls. Experimental results in
Andersen et al. (2012, Fig. 6) support this: the estimated prior distribution for an
ambiguous event with π=8/10 is strongly negatively skewed.
The fourth question involved a hypothetical loss of $15 and two colors of balls. Here
the matching probability is mL
=(1-δ)π+αδ, as detailed in Online Appendix B. For
losses, a subject was ambiguity averse if he was willing to accept a relatively high
known probability of loss to avoid Box U, that is mL
π , corresponding to α0.5 (see
Online Appendix B).
3.4 Alternative multiple prior sets for the α-MaxMin model
We now derive the matching probability for two alternative prior sets, which we will
also estimate and evaluate against our proposed model. First, in the special case of zero
confidence (δ=1), the prior interval is [0, 1], and the decision-maker then considers all
possible probabilities. The matching probability in this case is: m=(1-α). We note that
J Risk Uncertain (2015) 51:219–244 233

m is independent of π, implying that the matching probability is the same for all
ambiguity questions.
Second, we introduce an alternative prior probability set Csym, which is symmetric
around the reference probability π:
Csym ¼ p∈ π−
1
2
d; π þ
1
2
d

; with d≥0 and d ≤min 2π; 2 1−π
ð Þ
f g ð8Þ
where p is a prior probability of winning, and d is the length of the prior probability
interval. Using this prior interval, a prospect paying $15 if a purple ball is drawn from
Box U is evaluated as:
αminp∈ π−0:5d;πþ0:5d
½ pU 15
ð Þ þ 1−α
ð Þmaxp∈ π−0:5d;πþ0:5d
½ pU 15
ð Þ
¼ α π−0:5d
ð ÞU 15
ð Þ þ 1−α
ð Þ π þ 0:5d
ð ÞU 15
ð Þ ¼ π þ 0:5−α
ð Þd
ð ÞU 15
ð Þ
ð9Þ
Accordingly, the matching probability is m=π+(0.5–α)d , and AA=π-m=(α-0.5)d.
The decision-maker is ambiguity neutral (AA=0) if d=0 or α=1/2, ambiguity averse
(AA0) if α1/2 and d0, and ambiguity seeking (AA0) if α1/2 and d0. We note
that the ambiguity attitude (AA) is independent of the reference probability, and thus
this model cannot explain a-insensitivity. Further, the model implies that in a regression
of m on π the slope coefficient of π is equal to 1, a condition that we can test
empirically.
4 Estimating the α-MaxMin model
4.1 Econometric model
For our econometric model, we adopt the following notation: for each respondent
(i=1,2,...,I) and for each of the three ambiguity elicitation questions in the gains domain
(k=1, 2, 3), we measure a matching probability mik.21
Further, let πik denote the reference
probability of winning for Box U in question k. We assume that all respondents are
neutral about the winning color, so that subscript i can be dropped and πk has the
following fixed values: π1=0.5, π2=0.1 and π3=0.9. To capture the general tendency of
ambiguity attitudes in the population, in our model we initially assume that there is only
one representative decision-maker, as is common in the literature on estimating prefer-
ence models. Hence, the parameters for ambiguity aversion α and the perceived level of
ambiguity δ are constant, not depending on i. Equation (6) then implies that the matching
probability mik in the α-MaxMin model is equal to: mik=(1–α)δ+(1–δ)πk, for k=1, 2, 3.
Empirically, we estimate this equation as follows:
mik ¼ c þ sπk þ εik; for i ¼ 1; 2…; I and k ¼ 1; 2; 3 ð10Þ
where c and s are coefficients, and εik is an identically and independently distributed
error term.
21
For now, we defer discussion of the ambiguity loss question (k=4) to Section 4.4.
234 J Risk Uncertain (2015) 51:219–244

Model (10) can be estimated with pooled ordinary least squares (OLS), with k
observations for each respondent. The parameters of the α-MaxMin model are identi-
fied as δ=1–s and α=1–c/δ. Standard errors for δ and α are computed with the delta
method. We can now test and compare the fit of different versions of the α-MaxMin
model using the following restrictions:
EU expected utility
ð Þ : δ ¼ 0 and c ¼ 0 ð11aÞ
αMaxMinCsym : δ ¼ 0 and c ≠ 0 ð11bÞ
MaxMin 0; 1
½ : δ ¼ 1 and α ¼ 1 ð11cÞ
MaxMax 0; 1
½ : δ ¼ 1 and α ¼ 0 ð11dÞ
αMaxMin 0; 1
½ : δ ¼ 1 and 0 α 1 ð11eÞ
MaxMinCδ : 0 δ 1 and α ¼ 1 ð11fÞ
MaxMaxCδ : 0 δ 1 and α ¼ 0 ð11gÞ
αMaxMinCδ : 0 δ 1 and 0 α 1 ð11hÞ
Here Csym is a symmetric prior probability interval (of unknown length d), while
[0,1] denotes the prior interval where the decision-maker considers all possible prob-
abilities, and Cδ the asymmetric prior probability set where the decision-maker has
confidence (1–δ) in the reference probability π. Tests of restrictions (11a–h) allow us to
see which specification of α-MaxMin best describes the general pattern of ambiguity
attitudes in the U.S. population.
To loosen the restrictive assumption of one representative decision-maker and to
take into account heterogeneity in ambiguity preferences (see, e.g., Stahl 2014), we
include a random effect in the model: mik=c+sπk+εik+ui, where ui is a random effect
that is independent of the error term (εik) and uncorrelated between individuals, with
E[ui
2
]=σu
2
. The random effect captures unobserved heterogeneity in the ambiguity
J Risk Uncertain (2015) 51:219–244 235

aversion parameter αi across individuals i, with αi=1−(c+ui)/δ. Finally, we use
clustered standard errors to correct for heteroskedasticity between individuals, and
correlation of the error terms within.
4.2 Main results
Table 2 displays our estimates of the constant c and the slope coefficient s. Additionally,
the table shows estimates for the α-MaxMin model parameters α, (1–α), and δ, to
facilitate interpretation and testing of hypotheses. Column (1) reports estimates for our
baseline sample consisting of 2991 respondents who answered all ambiguity questions
and who spent at least two minutes answering the ambiguity questions.22
The depen-
dent variables are the matching probabilities, mik for the first three ambiguity questions
(k=1, 2, 3), involving only gains and implemented with real incentives. We discuss the
results for losses in a later section. The estimates shown in Table 2 are for a random
effects model specification; estimates for the model without random effects in (10) are
similar and therefore not reported.23
Our baseline results show that, overall, the U.S. population is ambiguity averse.
Column (1) of Table 2 reports α=0.56, combined with 60% confidence in the reference
probability (δ=0.40).24
The extent of ambiguity aversion is modest, with only a slight
overweighting of the worst outcome (α1/2). Nevertheless, it is statistically significant
as we can reject α≤1/2 (t=8.47 and p-value0.01). Together, these estimates (α=0.56,
δ=0.40) imply ambiguity-seeking behavior for the low likelihood ambiguous event of
winning if one of 10 possible colors is chosen. Here, our model predicts a matching
probability of 0.24. The estimates also imply relatively strong ambiguity-averse behav-
ior for the high likelihood ambiguous event of winning if any of nine of 10 possible
colors is selected; the model predicts a matching probability of 0.71. Hence, the α-
MaxMin model with prior set Cδ describes the typical pattern of ambiguity attitudes in
the population well, including a-insensitivity.
We use the model estimates in Column (1) to test the alternative multiple prior
models in hypotheses (11a-h) in subsequent steps. First, the data do not support
subjective expected utility with reference probability measure π (11a): a joint test
rejects the restrictions c=0 and δ=0. Second, the use of a pessimistic prior probability
set [0,1] for the α-MaxMin model is not supported: hypotheses (11c-e) impose the
restriction δ=1, which is rejected.25
Third, the symmetric prior probability set Csym for
the α-MaxMin model is also inconsistent with the data: the restriction δ=0 in (11b) is
rejected. Among the models based on the prior set Cδ, the pessimistic MaxMin-Cδ
22
Out of the 3258 original respondents, 3 did not answer any questions, 85 did not complete all of our
ambiguity questions, and 179 spent less than two minutes on answering the ambiguity questions. After
excluding these 267 respondents, we have a final sample of 2991 respondents.
23
Pooled OLS estimates are consistent in the presence of random effects, but the standard errors may be
inefficient. As we use clustered (robust) standard errors, the results of pooled OLS are similar to a random
effects model.
24
Our estimate of α is similar to values of α=0.515 reported in Ahn et al. (2014) for a small sample of
students and α=0.556 in Potamites and Zhang (2012) for Chinese investors. Baillon et al. (2015) estimate α=
0.61 and δ=0.51 in a sample of 64 students, with the source of ambiguity being the returns of an unknown
stock.
25
We test the single restriction δ=1, implied by all three models with [0,1] as the prior set. Joint tests of δ=1,
α=1 for MaxMin-[0,1], or δ=1, α=0 for MaxMax-[0,1] give the same result.
236 J Risk Uncertain (2015) 51:219–244

model (11f) and the optimistic MaxMax-Cδ (11g) are also both rejected, as they imply
α=1 and α=0. Only the α-MaxMin model with prior set Cδ (0δ1) and 0α1) is
consistent with the pattern of ambiguity attitudes in the U.S. population.
In sum, we find support for ambiguity preferences where not all the weight is put on
the worst case (α=1) nor on the best case (α=0), and the prior probability set reflecting
ambiguity perceptions is asymmetric for all π≠0.5.
4.3 Robustness checks
As a robustness check, we next exclude respondents who gave incorrect answers to the
two check questions: Column (2) limits the sample to respondents who answered both
check questions correctly. Interestingly, the coefficient estimates change hardly at all,
and our conclusions remain the same. For example, ambiguity aversion and confidence
in the reference probability are both slightly lower among respondents who answered
both check questions correctly (α=0.55, δ=0.45), but the difference is small compared
to the full sample results.
Statistical tests (available on request) confirm that a random effects model fits our
data better than pooled OLS or fixed effects models. The row labeled BError correlation
Table 2 Alpha-MaxMin model estimates
(1) (2)
Constant c 0.177***
0.200***
[42.56] [32.03]
Slope s 0.596***
0.555***
[88.97] [55.46]
Alpha-Min α 0.562***
0.549***
[77.13] [59.53]
Alpha-Max (1-α) 0.438***
0.451***
[60.19] [48.90]
Delta δ 0.404***
0.445***
[60.40] [44.39]
Restriction on checks No 2 correct
I individuals 2991 1232
R2
0.431 0.449
Error st. dev. (σε) 0.187 0.171
Random effect st. dev. (σu) 0.123 0.106
Error correlation (ρ) 0.302 0.280
Notes: The table shows estimation results for the α-MaxMin model, derived from matching probabilities mik
for the three ambiguity gain questions (k=1, 2, 3). We estimate Equation (10) using a random effects model,
which gives estimates of the slope coefficient s and constant c. Estimates of ambiguity aversion parameter α
and perceived ambiguity δ are then derived from c and s, with standard errors based on the delta method.
Column (1): full sample results. The full sample consists of respondents who answered all ambiguity questions
and spent at least two minutes of time. Column (2): same model, but further limited to 1232 respondents who
answered both check questions correctly. Standard errors are shown in brackets (robust, clustered by
individual). Statistical significance: *
p0.1, **
p0.05, ***
p0.01
J Risk Uncertain (2015) 51:219–244 237

(ρ)^ in Table 2 shows the estimated within-individual correlation of the overall error
term (εik+ui, including the random effect ui): ρ=0.30. The significance of ρ indicates
unobserved heterogeneity in ambiguity aversion at the individual level; the variance of
the errors (σε
2
) drops by 30% after accounting for this unobserved heterogeneity.
In additional results (available on request), we estimate a model with a random effect
added to the slope coefficient (s+vi), to capture unobserved heterogeneity in ambiguity
perceptions (δ). The main coefficient estimates are unchanged (α=0.56, δ=0.40), as
random effects only alter the covariance matrix of the errors. But, the variance of the
errors (σε
2
= 0.1642
) drops by another 23% after taking into account unobserved
heterogeneity in perceived ambiguity. Further, posterior estimates of α and δ show a
positive correlation between ambiguity aversion and perceived ambiguity, ranging from
0.2 to 0.4. Hence, those who perceive more ambiguity also tend to be more ambiguity
averse.
4.4 Ambiguity attitudes for losses
Next we include the matching probability for the ambiguity question involving losses;
our aim is to test whether ambiguity aversion differs for gains and losses. Recall that the
matching probability for the ambiguity gains questions is mik=(1−α)δ+(1−δ)πk, for
k=1, 2, 3. By contrast, the matching probability for the loss question is: mik
L
=αL
δ+(1
−δ)πk, for k=4 (see Online Appendix B). We introduce a separate ambiguity aversion
parameter for losses, αL
, distinct from α, the ambiguity aversion parameter for gains.26
The adapted regression model specification below allows us to test whether α=αL
:
mik ¼ c þ dL
Lk þ sπk þ εik þ ui; for i ¼ 1; 2; …; I and k ¼ 1; 2; 3; 4 ð12Þ
where Lk is a dummy variable for the loss question (Lk=0 for k=1, 2, 3, and Lk=1 for
k=4), and dL
is the corresponding regression coefficient. The parameters of the α-
MaxMin model are identified as follows: δ=1–s and α=1–c/δ, and αL
=(c+dL
)/δ.
Table 3 displays estimation results for Equation (12). The row labeled BTest α=αL
^
in Table 3 shows that the data reject the hypothesis that ambiguity aversion for losses
and gains are equal. Instead, we find ambiguity aversion for gains (α=0.56) and
ambiguity seeking for losses (αL
=0.46). 27
We note that the restriction dL
=0
corresponds to the special case of αL
=(1-α), or reflection of ambiguity aversion for
gains (α½) into ambiguity seeking for losses (αL
½). In both columns in Table 3, we
cannot reject dL
=0 at the five percent significance level.
Reflection implies that the ambiguity attitude for losses is the opposite of that toward
gains. Reflection is often found for decisions under risk, and it is part of prospect theory
(Tversky and Kahneman 1992): a common finding is of risk aversion for gains and risk
26
The equations for mik and mik
L
have different constant terms, (1–α)δ and αL
δ, so the model in Equation (11)
is no longer applicable (it would imply the restriction 1–α=αL
). Introducing a dummy variable for the loss
question permits us to separately estimate and identify α and αL
.
27
A drawback of the model in Equation (12) is that the random effect ui has opposite effects on ambiguity
aversion for gains and losses, an assumption inconsistent with the positive correlation between AA50
and
AA−50
. As a result the estimated correlation of the random effect (ρ) is relatively low in Table 3. We have also
estimated a model with two separate random effects for the constant c and loss dummy dL
, but we find no
difference in the main results concerning α, αL
and δ. Results are available on request.
238 J Risk Uncertain (2015) 51:219–244

seeking attitudes for losses. To the best of our knowledge, the results in Table 3 are the
first confirmation of reflection as the typical pattern for decision-making under ambi-
guity in the general population. In line with our results, Kothiyal et al. (2014) find
evidence of reflection in an experiment with students.
The results in Table 3 are subject to two caveats. First, as explained above, the
ambiguity question for losses was implemented without real incentives to avoid house
money effects. Second, in Equation (12) we assume that the perceived level of
Table 3 Testing reference dependence: alpha for gains and losses
(1) (2)
Constant c 0.177***
0.200***
[43.51] [32.02]
Slope s 0.596***
0.555***
[88.97] [55.46]
Loss dummy dL
0.010*
−0.008
[1.87] [1.13]
Delta δ 0.404***
0.445***
[60.40] [44.39]
Gains
Alpha-Min α 0.562***
0.549***
[77.13] [59.53]
Alpha-Max (1-α) 0.438***
0.451***
[60.19] [48.89]
Losses
Alpha-Min αL
0.463***
0.433***
[52.73] [39.27]
Alpha-Max (1-αL
) 0.537***
0.567***
[61.21] [51.49]
Restriction on checks No 2 correct
R2
0.377 0.397
Test α=αL
108.2***
91.4***
Error st. dev. (σε) 0.210 0.188
Notes: The table shows estimation results for the α-MaxMin model, derived from matching probabilities mik
for all four ambiguity questions (k=1, 2, 3, 4), including the question involving losses as outcomes. We
estimate Equation (12), which gives estimates of the slope coefficient s, constant c and the coefficient for the
loss question dummy dL
. Using these estimates, we then derive values for δ (perceived ambiguity), α
(ambiguity aversion for gains), and αL
(ambiguity aversion for losses), with standard errors based on the
delta-method. The row BTest α=αL
^ displays a chi-square statistic for the null hypothesis α=αL
: ambiguity
aversion for gains and losses are equal. Column (1): full sample estimates. Column (2): same model, but
limited to 1232 respondents who answered both check questions correctly. Standard errors are shown in
brackets (robust, clustered by individual). Statistical significance: *
p0.1, **
p0.05, ***
p0.01
J Risk Uncertain (2015) 51:219–244 239

ambiguity δ is equal for gains and losses.28
This is equivalent to assuming that a-
insensitivity is equal for gains and losses, which is supported by experimental evidence
in Baillon and Bleichrodt (2015).
4.5 Ambiguity aversion and observed individual characteristics
Little is known about how ambiguity preferences vary with individual characteristics
such as gender, age and income. Even less is known regarding individual characteristics
and perceptions about ambiguity levels: in this section we are the first to investigate this
relation.29
Let xih denote the value of individual control variable h=1, 2,...,H, for person
i=1,2,...,I. We estimate the following model:
mik ¼ c0 þ ∑H
h¼1chxih þ s0 þ ∑H
h¼1shxih

πk þ εik þ ui;
for i ¼ 1; 2; …I and k ¼ 1; 2; 3; 4
ð13Þ
where c0 is a constant and ch is a coefficient for the effect of variable h=1, 2,...,H on the
constant part of the model. The set of coefficients sh allow the model’s slope coefficient
to depend on the individual attributes, while s0 is the constant part of the slope. The
parameters of the α-MaxMin model are identified as: δi=1−(s0+∑h=1
H
shxih),andαi=1
−(c0+∑h=1
H
chxih)/δi.
The individual characteristics available from the ALP include indicators for male;
White; Hispanic; married; education (highest degree: high school or college); employ-
ment; (ln) family income; (ln) number of children; a financial literacy index; an
indicator of trust in others30
; risk aversion; and question order. Online Appendix C
provides variable definitions and descriptive statistics. The risk aversion metric is the
coefficient of relative risk aversion of a power utility function estimated using questions
similar to those in Tanaka et al. (2010).31
We randomized the order of the ambiguity and
risk questions in the ALP survey, with half of the respondents getting the risk questions
first, and the other half the ambiguity questions first. For this reason, we include an
indicator equal to one if the subject answered the risk questions first, and zero
otherwise.
Table 4 reports the effects of the individual characteristics on ambiguity aversion α
and the perceived level of ambiguity δ, with standard errors derived using the delta
28
If we could measure more matching probabilities for ambiguous events involving loss outcomes with other
likelihoods (e.g., similar to the 10 % and 90 % gains questions), we could also estimate δ separately in the loss
domain. We leave to future research additional refinements of ambiguity surveys and tests for reference
dependence.
29
Borghans et al. (2009) find that men are more ambiguity averse than women in a sample of 347 high school
students. In a study of the Dutch population, Dimmock et al. (2015b) estimate the relation between ambiguity
attitudes and control variables; there, however, few effects are statistically significant (sample size: N=666).
Using our ALP Module, Dimmock et al. (2015a) show in a web appendix that the non-parametric ambiguity
aversion measure AA50
is higher for men than for women, and positively related to risk aversion.
30
This is measured on a reversed scale from 0 to 5, with higher values indicating lower trust.
31
As in Tanaka et al. (2010), utility is defined over the payoffs of the gambles (not integrated with total
wealth), and the power coefficient is limited to the range from 0 to 1.5. Risk aversion, defined as ‘1 – power
function coefficient’, varies from −0.5 (risk seeking) to +1 (strongest level of risk aversion), and a value of
zero implies risk neutrality.
240 J Risk Uncertain (2015) 51:219–244

method. Columns (1) and (2) in Table 4 display marginal effects for ambiguity aversion
α.32
Columns (3) and (4) show the effect of the variables on perceived ambiguity δ.
Turning first to ambiguity aversion (α), we find that men are more ambiguity averse
than women, consistent with the experimental results in Borghans et al. (2009).
Ambiguity aversion is positively related to risk aversion, but the correlation is low and
ambiguity aversion is not subsumed by it. Older people tend to be less ambiguity averse,
which may capture the effect of life experiences, or a cohort effect. College-educated
respondents have higher ambiguity aversion than the less educated. This latter finding is
inconsistent with a potential alternative explanation for ambiguity aversion: that it is
driven by low cognitive ability. Rather, the positive relation with college education
suggests that ambiguity aversion measures preferences rather than cognitive errors.
We also find that ambiguity aversion is higher when the risk aversion questions are
presented to respondents prior to the ambiguity questions. The comparative ignorance
hypothesis of Fox and Tversky (1995), which states that ambiguity aversion is mag-
nified by a comparison to less ambiguous events, predicts such an order effect. In our
survey, the comparison is relative to the preceding risk questions involving only known
probabilities and no ambiguity. For this reason we randomized the survey order of the
risk and ambiguity questions.
Turning to perceived ambiguity (δ), we find that males, whites, and people with
more children tend to perceive higher levels of ambiguity. Further, college-educated
respondents perceive more ambiguity than high school educated respondents.
Receiving the risk aversion questions first also is associated with higher perceptions
of ambiguity. Interestingly, age and risk aversion do not influence perceptions about
ambiguity levels, but only ambiguity aversion. Vice versa, having (more) children is
associated with perceiving more ambiguity, but not with higher ambiguity aversion.
We also investigated how individual attributes are related to ambiguity aversion for
losses, and in results not detailed here, we found that the dependent variable is
positively related to risk aversion but not significantly associated with the other
variables. Thus risk aversion is positively related to both ambiguity aversion for gains
and ambiguity aversion for losses, but not related to the perceived level of ambiguity. A
potential explanation is that preferences towards risk and ambiguity are related (al-
though weakly) as both are preferences, while perceptions about the level of ambiguity
are formed independently from risk preferences.
5 Conclusion
This paper develops a method for estimating and testing multiple prior models of
ambiguity. Using a nationally representative sample of almost 3000 U.S. respondents,
we use matching probabilities to estimate a measure of ambiguity preferences, α, as
well as perceptions about the level of ambiguity, δ. Using our simple and tractable
method, we estimate α to be equal to 0.56 in the gain domain, consistent with mild
ambiguity aversion (α½). We estimate δ to be 0.40, meaning that the typical
respondent has a degree of confidence of 60% in the reference probability.
32
The derivative of αi with of respect to xih is:−ch/δi−sh(c0+∑h=1
H
chxih)/δi
2
, which we evaluate at the mean
values of xih and δi.
J Risk Uncertain (2015) 51:219–244 241

In the loss domain, our estimate of the ambiguity aversion parameter, αL
, is equal
to 0.46, implying ambiguity-seeking behavior (αL
1/2). Not only are ambiguity
attitudes for gains and losses significantly different, they are of opposite sign on
average, with ambiguity aversion for gains being reflected into ambiguity seeking
for losses. This implies that the ambiguity models applied in economics need to be
extended beyond the common assumption of universal ambiguity aversion. One
such alternative is the α-MaxMin model with separate ambiguity aversion param-
eters for gains and for losses.
Table 4 Ambiguity aversion and beliefs explained by individual characteristics
(1) (2) (3) (4)
Alpha-Min α Perceived ambiguity δ
Age −0.002***
−0.002***
−0.001 −0.001
Male 0.072***
0.083***
0.065***
0.063***
White −0.057***
−0.028 0.068***
0.060**
Hispanic 0.018 0.053*
−0.018 0.004
Married 0.017 −0.014 −0.004 −0.043*
Num. of kids (ln) −0.015 −0.013 0.031**
0.050***
High school −0.003 0.042 −0.038 0.015
College 0.083*
0.145**
−0.005 0.063
College−High school 0.087***
0.103***
0.033**
0.047**
Employed −0.022 −0.030 0.002 0.009
Income (ln) 0.004 0.016 0.003 0.007
Risk questions first 0.125***
0.072***
0.046***
0.048**
Financial literacy −0.003 0.012 0.016*
0.027*
Risk aversion 0.147***
0.090***
0.009 −0.009
Trust 0.007 −0.003 −0.007 −0.008
Restriction on checks No 2 correct No 2 correct
R2
0.458 0.483
Error st. dev. (σε) 0.185 0.169
Notes: The table shows estimation results for the α-MaxMin model, including a set of individual-level
explanatory variables for both ambiguity preferences and beliefs. The dependent variable is the matching
probability mik for all three ambiguity gains questions (k=1, 2, 3). We estimate Equation (13), including the 14
explanatory variables shown in the table (Age, Male, …, Trust). See Online Appendix C for definitions of the
explanatory variables. Columns (1) and (2) show marginal effects of the explanatory variables on ambiguity
aversion (α), and model statistics such as R2
. Columns (3) and (4) display the effects of the explanatory
variables on perceived ambiguity (δ). Estimates of the constant (c) and slope coefficient (s) are not shown to
save space. The model includes two dummies for the highest education level achieved: completing high
school, or completing college. The base category for education consists of respondents not completing high
school. The row BCollege−High school^ tests for differences in the groups with college education and high
school education. Columns (1) and (3): full sample results. Columns (2) and (4): same model, but limited to
respondents who answered both check questions correctly. Statistical significance: *
p0.1, **
p0.05, ***
p
0.01
242 J Risk Uncertain (2015) 51:219–244

Furthermore, our estimates of the α-MaxMin model confirm that most Americans
are ambiguity averse for uncertain events of moderate to high likelihood, but ambiguity
seeking for unlikely events. For example, when faced with a chance to win if one of 10
colors is selected, 60% of the people are ambiguity seeking. Our α-MaxMin model can
explain these choices because the prior probability set is asymmetric for low and high
likelihood events. For example, when winning for one out of 10 colors, the set of prior
probabilities in the calibrated model ranges from 6 to 46%.
For future work, our evidence of non-universal ambiguity aversion, especially
ambiguity seeking choices for low likelihood events and losses, implies that under-
standing the economic implications of such preferences is important. The model
developed in this paper offers a good starting point, as it is analytically simple, yet it
can describe actual choices under uncertainty observed in the field.
Acknowledgments The survey module fielded by the authors in the RAND American Life Panel (ALP)
was approved by the Institutional Review Board of the University of Pennsylvania. The authors gratefully
acknowledge financial support from Netspar, and grants to the University of Pennsylvania from the National
Institute on Aging (P30 AG-012836-18), and a grant from the National Institutes of Health–National Institute
of Child Health and Development Population Research Infrastructure Program (R24 HD-044964-9). Support
was also provided by the Pension Research Council/Boettner Center and the Wharton Behavioral Labs at the
University of Pennsylvania. We also thank the ALP teams at RAND and the University of Southern
California. We are grateful to Aurelien Baillon, Peter Wakker and participants at FUR 2014 for helpful
comments, and to Tania Gutsche, Arie Kapteyn, Bart Orriens, and Bas Weerman for assistance with the
survey. Yong Yu provided outstanding programming assistance. The content is solely the responsibility of the
authors and does not represent the official views of the National Institute of Aging, the National Institutes of
Health, or any of the other institutions providing funding for this study or with which the authors are affiliated.
References
Abdellaoui, M., Baillon, A., Placido, L., Wakker, P. P. (2011). The rich domain of uncertainty: source
functions and their experimental implementation. American Economic Review, 101, 695–723.
Abdellaoui, M., Vossmann, F., Weber, M. (2005). Choice-based elicitation and decomposition of decision
weights for gains and losses under uncertainty. Management Science, 51, 1384–1399.
Ahn, D., Choi, S., Gale, D., Kariv, S. (2014). Estimating ambiguity aversion in a portfolio choice
experiment. Quantitative Economics, 5, 192–223.
Akay, A., Martinsson, P., Medhin, H., Trautmann, S. T. (2012). Attitudes towards uncertainty among the
poor: an experiment in rural Ethiopia. Theory and Decision, 73, 453–464.
Andersen, S., Fountain, J., Harrison, G. W., Hole, A. R., Rutström, E. E. (2012). Inferring beliefs as
subjectively imprecise probabilities. Theory and Decision, 73, 161–184.
Andersen, S., Fountain, J., Harrison, G. W., Rutström, E. E. (2009). Estimating aversion to uncertainty.
Working Paper, University of Central Florida.
Asano, T., Shibata, A. (2011). Risk and uncertainty in health investment. European Journal of Health
Economics, 12, 79–85.
Attanasi, G., Gollier, C., Montesano, A., Pace, N. (2014). Eliciting ambiguity aversion in unknown and in
compound lotteries: a smooth ambiguity model experimental study. Theory and Decision, 77, 485–530.
Baillon, A., Bleichrodt, H. (2015). Testing ambiguity models through the measurement of probabilities for
gains and losses. American Economic Journal: Microeconomics, 7, 77–100.
Baillon, A., Bleichrodt, H., Keskin, U., L’Haridon, O., Li, C. (2015). The effect of learning on ambiguity
attitudes. Working paper.
Baillon, A., Cabantous, L., Wakker, P. P. (2012). Aggregating imprecise or conflicting beliefs: an
experimental investigation using modern ambiguity theories. Journal of Risk and Uncertainty, 44, 115–
147.
Binmore, K., Stewart, L., Voorhoeve, A. (2012). How much ambiguity aversion? Journal of Risk and
Uncertainty, 45, 215–238.
J Risk Uncertain (2015) 51:219–244 243

Borghans, L., Heckman, J. J., Golsteyn, B. H. H., Meijers, L. (2009). Gender differences in risk aversion
and ambiguity aversion. Journal of the European Economic Association, 7, 649–658.
Bostic, R., Herrnstein, R. J., Luce, R. D. (1990). The effect on the preference-reversal phenomenon of using
choice indifferences. Journal of Economic Behavior and Organization, 13, 193–212.
Butler, J. V., Guiso, L., Jappelli, T. (2014). The role of intuition and reasoning in driving aversion to risk and
ambiguity. Theory and Decision, 77, 455–484.
Chateauneuf, A., Eichberger, J., Grant, S. (2007). Choice under uncertainty with the best and worst in mind:
neo-additive capacities. Journal of Economic Theory, 137, 538–567.
Cohen, M., Jaffray, J.-Y., Said, T. (1987). Experimental comparisons of individual behavior under risk and
under uncertainty for gains and for losses. Organizational Behavior and Human Decision Processes, 39,
1–22.
Conte, A., Hey, J. D. (2013). Assessing multiple prior models of behaviour under ambiguity. Journal of Risk
and Uncertainty, 46, 113–132.
Dimmock, S. G., Kouwenberg, R., Mitchell, O. S., Peijnenburg, K. (2015a). Ambiguity aversion and
household portfolio choice puzzles: empirical evidence. Journal of Financial Economics, forthcoming.
Dimmock, S. G., Kouwenberg, R., Wakker, P. P. (2015b). Ambiguity attitudes in a large representative
sample. Management Science, forthcoming.
Dow, J., Werlang, S. R. C. (1992). Uncertainty aversion, risk aversion and the optimal choice of portfolio.
Econometrica, 60, 197–204.
Eichberger, J., Grant, S., Kelsey, D., Koshevoy, G. A. (2011). The α-MEU model: a comment. Journal of
Economic Theory, 146, 1684–1698.
Ellsberg, D. (1961). Risk, ambiguity, and the Savage axioms. Quarterly Journal of Economics, 75, 643–669.
Epstein, L. G., Wang, T. (1994). Intertemporal asset pricing under Knightian uncertainty. Econometrica, 62,
283–322.
Etchart-Vincent, N., l’Haridon, O. (2011). Monetary incentives in the loss domain and behavior toward risk:
an experimental comparison of three reward schemes including real losses. Journal of Risk and
Uncertainty, 42, 61–83.
Fox, C. R., Tversky, A. (1995). Ambiguity aversion and comparative ignorance. Quarterly Journal of
Economics, 110, 585–603.
Ghirardato, P., Maccheroni, F., Marinacci, M. (2004). Differentiating ambiguity and ambiguity attitude.
Journal of Economic Theory, 118, 133–173.
Gilboa, I., Schmeidler, D. (1989). Maxmin expected utility with non-unique priors. Journal of
Mathematical Economics, 18, 141–153.
Hey, J. D., Lotito, G., Maffioletti, A. (2010). The descriptive and predictive adequacy of theories of decision
making under uncertainty/ambiguity. Journal of Risk and Uncertainty, 41, 81–111.
Kahn, B. E., Sarin, R. K. (1988). Modeling ambiguity in decisions under uncertainty. Journal of Consumer
Research, 15, 265–272.
Klibanoff, P., Marinacci, M., Mukerji, S. (2005). A smooth model of decision making under ambiguity.
Econometrica, 73, 1849–1892.
Kothiyal, A., Spinu, V., Wakker, P. P. (2014). An experimental test of prospect theory for predicting choice
under ambiguity. Journal of Risk and Uncertainty, 48, 1–17.
Potamites, E., Zhang, B. (2012). Measuring ambiguity attitudes: a field experiment among small-scale stock
investors in China. Review of Economic Design, 16, 193–213.
Stahl, D. O. (2014). Heterogeneity of ambiguity preferences. Review of Economics and Statistics, 96, 609–
617.
Tanaka, T., Camerer, C. F., Nguyen, Q. (2010). Risk and time preferences: linking experimental and
household survey data from Vietnam. American Economic Review, 100, 557–571.
Thaler, R. H., Johnson, E. J. (1990). Gambling with the house money and trying to break even: the effects of
prior outcomes on risky choice. Management Science, 36, 643–660.
Trautmann, S. T., van de Kuilen, G. (2015). Ambiguity attitudes. In G. Wu G. Keren (Eds.), The Wiley-
Blackwell handbook of judgment and decision making. Oxford: Blackwell.
Tversky, A., Kahneman, D. (1992). Advances in prospect theory: cumulative representation of uncertainty.
Journal of Risk and Uncertainty, 5, 292–323.
Vieider, F. M., Martinsson, P., Medhin, H. (2012). Stake effects on ambiguity attitudes for gains and losses.
Working paper.
244 J Risk Uncertain (2015) 51:219–244

Estimating ambiguity preferences and perceptions in multiple prior models: Evidence from the field

More Related Content

What's hot (19)

Similar to Estimating ambiguity preferences and perceptions in multiple prior models: Evidence from the field (20)

More from Nicha Tatsaneeyapan (20)

Recently uploaded (20)

Estimating ambiguity preferences and perceptions in multiple prior models: Evidence from the field