A Reconsideration of Gender Differences in Risk Attitudes
I
Antonio Filippin
a,b
, Paolo Crosetto
c
a
University of Milan, Department of Economics, Via Conservatorio 7, 20122 Milano, Italy
b
Institute for the Study of Labor (IZA), Schaumburg-Lippe-Str. 5-9, 53113 Bonn, Germany
c
INRA, UMR 1215 GAEL, University of Grenoble, 38000 Grenoble, France
Abstract
This paper reconsiders the wide agreement that females are more risk averse than males.
We survey the existing experimental literature, finding that significance and magnitude of
gender differences are task-specific. We gather data from 54 Holt and Laury (2002) repli-
cations, involving more than 7000 subjects. Gender differences appear in less than 10% of
the studies, and are significant but negligible in magnitude once all the data are pooled. We
exclude that this result is driven by noisier HL data. Gender differences appear to correlate
with the presence of a safe option and fixed probabilities in the elicitation method.
JEL Classifications: C81; C91; D81
Keywords: Gender, Risk, Survey
I
We are grateful to the Max Planck Institute of Economics (Jena) for financial and logistic support and to
Janna Heider for excellent research assistance. We would like to thank all the authors that kindly contributed
their data, the members of the ESA mailing list for useful references, Ainhoa Aparicio Fenoll, Tore Ellingsen,
Andrea Ichino for useful suggestions, together with the participants, in 2013, to: IMEBE, Madrid; BEELAB
workshop, Florence; ESA World, Zurich; SIE Conference, Bologna; MPI Autumn Workshop, Jena; in 2014, to:
ASFEE, Besançon; FUR, Rotterdam; EALE, Ljublijana; Workshop on Gender, Alicante; and to seminars in EUI
Florence, Milan, Oxford, Göttingen, Prague and Padua. All remaining errors are ours.
Contact: antonio.filippin@unimi.it (Antonio Filippin), paolo.crosetto@gmail.com (Paolo Crosetto)
January 30, 2015
1. Introduction
Gender differences in risk preferences are often regarded as a stylised fact in the eco-
nomics and psychology literature. Many studies as well as the available meta-analyses find
that women display a more risk averse behaviour than men when confronted with decisions
under risk. In economics, for instance, surveys made by Eckel and Grossman (2008c) and
Croson and Gneezy (2009) find mostly supporting evidence and investigate the robustness
of this result along several dimensions, such as the characteristics of the subject pool, the
strength of incentives, the gain vs. loss domain, the abstract vs. contextual framework. These
surveys, though, are based on a relatively small sample of studies (16 and 10, respectively,
3 of which in common) given the variety of designs covered. As noted by Charness and
Gneezy (2012) and Holt and Laury (2014), the differences in the methods used to measure
the preferences can act as an additional source of heterogeneity. Consequently, Charness and
Gneezy (2012) focus on a single elicitation method, the Investment Game, and find strong
evidence that females are less willing to take risk. In psychology, Byrnes et al. (1999) pro-
vide a meta-analysis including 150 studies, using a broad definition of risk, from smoking to
driving to gambling, and analyzing self- reported, incentivised, as well as observed choices.
The study finds that males take more risks than females in most of the risk categories, even
though the magnitude of the effect is usually small, seldom significant, and some studies
find contrary evidence.
Despite the apparently wide agreement that females are more risk averse than males,
we believe that the evidence supporting this view cannot be considered conclusive for two
reasons. First, there are important branches of the literature still largely unexplored. For
instance, the Holt and Laury (2002) (henceforth, HL) task has never been the subject of a
comprehensive analysis by gender, despite being by far the most popular elicitation method
in economics according to the number of citations. Moreover, in the Bomb Risk Elicitation
Task (Crosetto and Filippin, 2013a) no gender difference was found. Second, no attempt
has been made yet to investigate whether and how the elicitation methods play a role in
shaping the observed results by gender. Risk attitudes are a latent construct that can only be
indirectly and imperfectly measured: their measurement is by construction a combination
of the latent preferences and the measurement error induced by the tool used to elicit them.
Crosetto and Filippin (2013b) analyse to what extent, and in which direction the measured
risk preferences are shaped by the characteristics of the elicitation task adopted. In this
paper we aim to extend this exercise along a gender dimension.
We first provide a thorough survey of the literature, finding mixed results. We then focus
on the unexplored HL task. Unfortunately, only a small fraction of contributions explicitly
report about gender differences, because HL is usually a companion task in unrelated ex-
perimental studies. Only twenty papers, out of the more than five hundred citing Holt and
Laury (2002), provide data on the gender breakdown of risk preferences. Contrary to the
widespread consensus, out of these twenty papers, only three report significant gender dif-
ferences in risk preferences.
This striking result, combined with the presence of large amounts of uncharted HL gen-
der data, spurred us to directly contact the authors of the 94 published HL replications. We
collected the data of 54 published studies, corresponding to almost eight thousand subjects,
and reduced them to a common comparable format.
The resulting dataset increases dramatically the information as compared to that avail-
2
able in published results and allows us to provide conclusive evidence about gender differ-
ences in HL. The results consistently show that gender differences are the exception rather
than the rule in HL replications. Men and women display a similar behaviour, and when a
difference can be detected it is usually small.
The large amount of comparable data also allows us to greatly increase the statistical
power of the analysis. Moreover, access to all microdata allows us to exploit the data of sub-
jects making inconsistent choices using a structural model estimated with maximum likeli-
hood. The results on the pooled data show a comeback of significant gender differences, but
the magnitude of the effect turns out to be economically unimportant. Differences amount
to one sixth of a standard deviation, less than a third of the effect found by other elicitation
methods analysed in this paper (e.g., by Charness and Gneezy, 2012; Eckel and Grossman,
2008b).
Our results indicate that the frequency and the importance of gender differences reflect
specific characteristics of the elicitation methods over and above true differences in the un-
derlying (and latent) risk attitudes. Importantly, such a heterogeneity of the gender pattern
is not due to the fact that HL induces more noise than other tasks, something that, if true,
would make it more difficult to detect the same differences in the underlying preferences.
Observing a gender gap not only depends on the task being contextual or not (Eckel and
Grossman, 2008a), on it having to do with risk or with uncertainty (Wieland and Sarin, 2012),
or on the choices being incentivised, self-reported or observed (Byrnes et al., 1999). Even re-
stricting the analysis to the narrow domain of incentivised lottery choice tasks currently
used in experimental economics, gender differences depend on the details of the task. We
single out two characteristics that jointly correlate with the likelihood of observing gender
differences: a) the presence of a safe option within the choice set, and b) the use of lotteries
with 50% − 50% fixed probabilities in tasks that generate the menu of lotteries changing the
amounts at stake.
Published results as well as our dataset do not allow us to further investigate and to
disentangle the effect of each of these two characteristics. Nevertheless, we believe that this
paper provides a leap forward in the understanding of gender differences in risk preferences
from two points of view. First, it makes clear that, instead of being treated as a fact, gen-
der differences should be analysed jointly with the characteristics of the task used to elicit
risk preferences. Second, it greatly restricts the set of possible determinants. In a compan-
ion paper we analyse this issue by means of controlled experiments, but without finding
conclusive evidence (Crosetto and Filippin, 2014). The availability of a riskless alternative
allows to rationalise gender differences in some but not in all the cases, while the 50% − 50%
fixed probabilities do not play a significant role.
The outline of the paper is as follows. Section 2 summarises the state of the art in the
literature about gender differences in risk preferences and presents the survey of the few
HL published results by gender. Section 3 describes the characteristics of the dataset of
HL replications we built and use. Section 4 analyses our dataset, first paper by paper and
then pooling the data, using both descriptive statistics and structural modeling allowing
for errors in the choices. Section 5 discusses which characteristics of the task could trig-
ger the stark difference in behaviour observed, identifying some candidates, and Section 6
concludes.
3
2. Literature Review
There are more risk elicitation methods than can be mentioned here. Our ambition is
not that of providing an exhaustive survey of the results by gender across the different tasks
used to measure risk preferences. In contrast, the goal of this section is to summarise the
state of the art in the risk and gender literature. Consequently, we limit our analysis to three
representative and widely used methods: the Investment Game, introduced by Gneezy and
Potters (1997), an Ordered Lottery Selection task proposed by Eckel and Grossman (2002,
2008b), and the Holt and Laury (2002) task, the most cited and replicated risk elicitation
method.
In the Investment Game (henceforth IG) subjects decide how to allocate a given endow-
ment E between a safe account and a risky lottery that yields with 50% probability 2.5 times
the amount invested, zero otherwise. The task is framed as an investment decision, and
a risk neutral subject should invest all of her endowment, since the marginal return of the
risky option is greater than one.
In the Eckel and Grossman task (henceforth EG) subjects make a single choice, picking
one out of an ordered set of lotteries. This method has been first introduced in the literature
to specifically measure risk preferences by Binswanger (1981). In the EG implementation
subjects are faced with 5 lotteries characterised by a linearly increasing expected value as
well as greater standard deviation (see Table 1). The task is not framed, and a risk neutral
subject should choose lottery 5, since it has the highest expected value.
Choice Probability Outcome
1
A 50% 16 $
B 50% 16 $
2
A 50% 24 $
B 50% 12 $
3
A 50% 32 $
B 50% 8 $
4
A 50% 40 $
B 50% 4 $
5
A 50% 48 $
B 50% 0 $
Table 1: The 5 lotteries of the orginal Eckel and Grossman (2002) paper
The Holt and Laury (2002) (henceforth HL) risk elicitation method constitutes the most
widely known implementation of a multiple price list format applied to risk. The subjects
face a series of choices between pairs of lotteries, with one lottery safer (i.e., with lower vari-
ance) than the other. At the end of the experiment, one row is randomly chosen for payment,
and the chosen lottery is played to determine the payoff. The lottery pairs are ordered by
increasing expected value. The set of possible outcomes is common to every choice, and the
increase in expected value across rows is obtained by increasing the probability of the ’good’
event (see Table 2).
The subjects make a choice for each pair of lotteries, switching at some point from the
safe to the risky option as the probability of the good outcome increases. The switching
4
Option A Option B
1 1/10 2 $ 9/10 1.6 $ sa 1/10 3.85 $ 9/10 0.1 $
2 2/10 2 $ 8/10 1.6 $ 2/10 3.85 $ 8/10 0.1 $
3 3/10 2 $ 7/10 1.6 $ 3/10 3.85 $ 7/10 0.1 $
4 4/10 2 $ 6/10 1.6 $ 4/10 3.85 $ 6/10 0.1 $
5 5/10 2 $ 5/10 1.6 $ 5/10 3.85 $ 5/10 0.1 $
6 6/10 2 $ 4/10 1.6 $ 6/10 3.85 $ 4/10 0.1 $
7 7/10 2 $ 3/10 1.6 $ 7/10 3.85 $ 3/10 0.1 $
8 8/10 2 $ 2/10 1.6 $ 8/10 3.85 $ 2/10 0.1 $
9 9/10 2 $ 1/10 1.6 $ 9/10 3.85 $ 1/10 0.1 $
10 10/10 2 $ 0/10 1.6 $ 10/10 3.85 $ 0/10 0.1 $
Table 2: The 10 lotteries of the original Holt and Laury (2002) paper
point captures their degree of risk aversion. A risk-neutral subject should start with Option
A, and switch to B from the fifth choice on. The higher the number of safe choices, the
stronger the degree of risk aversion. Never choosing the risky option or switching from B
to A are not infrequent and are regarded as inconsistent choices when modeling the choices
without including a stochastic component.
That women are more risk averse than men is often considered a stylised fact in the
economic literature. This finding is confirmed by some surveys (Croson and Gneezy, 2009;
Eckel and Grossman, 2008a).
1
Among the risk elicitation tasks analysed in this paper, this
state of the art is well captured by both the IG and EG tasks. Both tasks have already been
object of a survey from a gender perspective, and females have been shown to consistently
display a significantly more risk averse average behaviour.
Charness and Gneezy (2012) report that in the IG the gender gap is rather systematic
and quite sizable. Males invest more than females in most of the experiments analysed,
and such a difference is usually about 10 − 15% of the initial endowment (Charness and
Genicot, 2009; Charness and Gneezy, 2004, 2010; Dreber and Hoffman, 2007; Dreber et al.,
2010; Ertac and Gurdal, 2012; Fellner and Sutter, 2009; Gong and Yang, 2012; Langer and
Weber, 2004). Significant differences, but lower than 10% in size, appear in Haigh and List
(2005), Bellemare et al. (2005), and Crosetto and Filippin (2013b), while Gneezy et al. (2009)
is the only contribution in which a gender gap does not appear. Such a result is robust to the
context (lab vs. field) in which data have been gathered as well as to other features (amounts
at stake, geographical location, type of subjects).
Similar findings emerge in the EG task, with sizable gender differences appearing both in
the original experiment and in later replications (Arya et al., 2012; Ball et al., 2010; Crosetto
and Filippin, 2013b; Dave et al., 2010; Eckel et al., 2009, 2011; Grossman and Eckel, 2009; Wik
et al., 2004). Cleave et al. (2010) find a gender gap in a wide sample but not in a subsam-
ple that participated to later experiments, but it is, to the best of our knowledge, the only
1
Surveys also stress how some characteristics of the experiments make gender differences more likely to
appear. For instance, they are usually less likely to be found in contextual experiments (Eckel and Grossman,
2008a; Schubert et al., 1999).
5
