Content uploaded by Stein Terje Holden
Author content
All content in this area was uploaded by Stein Terje Holden on Jan 13, 2020
Content may be subject to copyright.
Endowment Effects and Loss Aversion in the Risky
Investment Game
Stein T. Holden and Mesfin Tilahun
School of Economics and Business, Norwegian University of Life Sciences, P. O Box 5003,
1433 ˚
As, Norway
Email: stein.holden@nmbu.no
Abstract
The risky investment game of Gneezy and Potters (1997) has been a popular
tool used to estimate risk tolerance and myopic loss aversion. We have assessed
whether a simple one-shot version of this game that is attractive as a simple
tool to elicit risk tolerance among respondents with limited education, can lead
to biased estimates of risk aversion due to endowment effects. We use a field
experiment with a pool of young business group members with limited education
to test for the potential bias associated with the initial endowment allocation.
We find a highly significant endowment effect which may explain low investment
levels and exaggerated measures of risk aversion where this game has been used
to estimate risk aversion. We develop and test a more balanced version of the
risk investment game and demonstrate that it gives less bias due to endowment
effects than the standard design and a full risk design that creates an endowment
effect in the opposite direction, indicating that loss aversion may not be the
primary cause of the endowment effect.
Keywords: Endowment effect; loss aversion; gender difference; risky
investment game; field experiment; Ethiopia.
JEL Codes: C93; D91.
1. Introduction
Endowment effects have been used to explain the Willingness to Accept
(WTA) – Willingness to Pay (WTP) gap and the exchange asymmetries often
found in settings where there are supposed to be minimal or no transaction costs
(Kahneman et al., 1990; Knetsch et al., 1989; Horowitz and McConnell, 2002).
The term “endowment effect” was first used by Thaler (1980) and he related this
effect to the fact that losses are weighted more heavily than gains and associated
this with prospect theory and loss aversion. Other reasons for the “stickiness” of
endowments include “status quo bias” (Samuelson and Zeckhauser, 1988; Kah-
neman et al., 1991) and “anchoring effects” (Epley and Gilovich, 2001; Simonson
and Drolet, 2004). Plott and Zeiler (2005, 2007) have demonstrated that with
a set of “priming” tools it is possible to eliminate such WTA-WTP gaps and
Preprint submitted to Elsevier January 10, 2020
exchange asymmetries but such priming is an exception in the real world, and
this may imply that there exist some fundamental behavioral characteristics
that contribute to the widespread occurrence of such exchange asymmetries.
A recent review puts emphasis on expectations affecting reference points in
combination with loss aversion as explanations for the persistent phenomenon
(Marzilli Ericson and Fuster, 2014). Marzilli Ericson and Fuster (2011) dis-
entangle endowment effects from ownership and show that expectations affect
reference points and may thereby trigger endowment effects. Our study provides
new insights on the presence of endowment effects associated with monetary en-
dowments and risky prospects rather than commodities when expected returns
should be the same across treatments.
Many of the standard experiments in behavioral economics start by pro-
viding an initial endowment in form of money. It is generally accepted that
such an endowment can have a wealth effect while endowment effects are asso-
ciated with ownership of commodities. Here we investigate whether a monetary
endowment can create endowment effects that are not simply wealth effects.
Experiments that typically provide an initial monetary endowment include the
dictator game, the ultimatum game, the trust game, the public goods game and
the risky investment game of Gneezy and Potters (1997). The monetary nature
of these endowments may be the reason for endowment effects being ignored
by assumption. However, if endowment effects are caused by loss aversion, one
should perhaps not just assume away endowment effects for money.
In the dictator game economists were initially surprised that respondents
did not hold back more money and this may have taken the attention away
from possible endowment effects in the dictator and ultimatum game as well.
Later studies have shown that it makes a difference whether the endowment
is windfall money or earned money where studies suggest that the endowment
effect is stronger for earned money (Cherry et al., 2002). A similar effect has
also been found for the public goods game (Muehlbacher and Kirchler, 2009).
We focus on the risky investment game of Gneezy and Potters (1997). More
specifically, we use a simple one-shot version of the game that was first used by
Gneezy et al.(2009). The respondents are provided an initial endowment X, of
which they are free to invest any amount 0 ≤x≤Xand they have a 50 %
probability of winning 3xand of losing the amount xinvested.
A rational respondent behaving according to expected utility theory should
be risk-neutral and invest the whole endowment in the typical small amount
gambles used in such experiments (Rabin, 2000; Rabin and Thaler, 2001). How-
ever, already Binswanger (1980, 1981) revealed that experimental decisions over
risky prospects are not integrated with the wealth of respondents. Most peo-
ple therefore appear to be risk averse also in small gambles and particularly
so if losses are included among the possible outcomes. Narrow bracketing with
a combination of risk aversion and loss aversion could potentially explain why
respondents exposed to the risky investment game prefer not to invest the full
initial endowment received in this game.
Charness et al. (2013) noted in a review of the risky investment game that
a weakness is that it does not separate risk-neutral and risk-loving respondents
2
as both these types would invest the whole amount. The beauty of the game
is, however, that it is so simple and easy to apply in the field and variants
of it have for that reason become a popular tool that has been used for the
study of a variety of issues such as gender differences in risk preferences (Char-
ness and Gneezy, 2012), myopic loss aversion among students and professional
traders (Gneezy and Potters, 1997; Haigh and List, 2005), and the correlation
between risk-taking, testosterone levels, and facial masculinity (Apicella et al.,
2008). It has also been tested and found useful in a developing country set-
ting where respondents have limited education and numeracy skills and more
complex elicitation methods such as the Holt and Laury (2002) Multiple Price
List approach may be associated with more cognitive problems and inconsistent
responses (Charness and Viceisza, 2016). This is also the type of environment
that our research is focusing on. We hope our study can contribute to the re-
finement of simple tools for risk preference elicitation in a development context
with non-WEIRD samples.
We combine a pilot field experiment and a large sample field experiment
in rural Ethiopia to investigate the existence of endowment effects in the risky
investment game.Two alternative treatments to the basic design were intro-
duced, one providing a full risk initial lottery endowment (Treatment 2), and
one balanced binary treatment (Treatment 3) that should eliminate or reduce
the potential endowment effects associated with the first two treatments. We
find highly significant and substantial endowment effects in our study.
One of the distinct findings with the risk investment game is that it has
been associated with stronger gender differences in risk preferences than some
alternative approaches to eliciting risk preferences (Charness and Gneezy, 2012;
Charness et al., 2013). In our assessment of potential endowment effects in the
game, we assess whether these can be an explanation for these strong gender
differences. We assess this by comparing the gender differences for each of the
three treatments.
The paper is organized as follows. Part 2 of the paper outlines the experi-
mental design. Part 3 describes the sample characteristics. Part 4 presents the
results and part 5 discusses the findings and draws some tentative conclusions
and suggestions for further work.
2. Experimental design
We combined a pilot field experiment with a large sample field experiment.
The respondents are members of youth business groups located in rural areas in
northern Ethiopia. The experiments were implemented for one group at a time
with up to 12 group members who played the games and were simultaneously
interviewed by 12 experimental enumerators. The baseline treatment was based
on the one-shot version of the game first used by Gneezy et al.(2009). Respon-
dents are given an endowment Xfrom which they can invest a share x/X which
is tripled by the researchers (3x/X) and which can be won with a 50 percent
probability, or otherwise lost such that the respondent only retains X−x. The
lucky winners obtain X−x+ 3x=X+ 2x.
3
The second treatment gave the respondents a lottery prospect of 3Xwith a
50 percent chance of winning. The respondents were then offered to sell all or
part of the lottery prospect and would then get a payment of one-third of the
lottery winning value they would sell. If they sell yout of 3X, they will get y/3
as payment. Losers of the game will get y/3 and winners will get 3X−y+y/3.
Against the H0 hypothesis that there is no significant difference in the
amount invested in Treatments 1 and 2, which implies that there are no endow-
ment or anchoring effects linked to the initial endowment, we test the following
hypotheses:
Hypothesis 1a:An endowment effect implies a bias towards the first en-
dowment allocated and Treatment 2 leads to a larger investment in the risky
option than Treatment 1.
Hypothesis 1b:A larger share of the respondents invest the full amount
in Treatment 2 than in Treatment 1.
The third treatment aimed at striking a balance between the first two treat-
ments that both may lead to a bias towards the initial riskless or risky base
options. This treatment was implemented as a set of binary choices to elicit the
optimal portfolio or balance between a safe and a risky option with the same ex-
pected returns as in treatments one and two. The first binary choice is between
getting Xwith certainty and 3Xwith a 50% probability. The preferred choice
in this first binary choice is then offered in the second binary choice where the
alternative choice is X/2 for sure and a 50% chance of winning the tripled sec-
ond half of the full lottery amount (Expected value: 0.5∗3X/2). Further binary
choices are provided till and optimal mix of safe and lottery amounts are identi-
fied. Details of the experimental protocols for the three treatments are provided
in Appendix 2. Treatment 3 allows us to test the following hypotheses:
Hypothesis 2a:Treatment 3 gives an average investment level that is larger
than for Treatment 1 and lower than for Treatment 2.
Hypothesis 2b:The share investing the full amount in Treatment 3 is
larger than for Treatment 1 and smaller than for Treatment 2
Hypothesis 2c:The average investment levels are the same in Treatments
2 and 3 but these are higher than in Treatment 1.
Hypothesis 2d:The share investing the full amount in Treatments 2 and
3 are the same but these are higher than in Treatment 1.
Hypotheses 2c and 2d are based on the theory that the endowment effect
is driven by loss aversion and an initial endowment in form of a risky prospect
that maximizes the potential loss should therefore not induce an endowment
effect.
The standard game has been shown to give significant gender differences with
women investing significantly less than men in most earlier studies (Charness
and Gneezy, 2012). It is still a mystery why this game tends to give stronger
gender differences than other games used to investigate gender differences in
risk preferences (Filippin and Crosetto, 2016). We wonder whether this could
be associated with an endowment effect bias that may be stronger for women.
We therefore want to test the following hypothesis:
Hypothesis 3:The gender difference is stronger in Treatment 1 than in
4
Treatments 2 and 3 as it is driven by the endowment effect that is stronger in
Treatment 1 and stronger for women than men.
If Hypothesis 3 holds, the gender difference in investment level should be
lower in Treatment 3 than in Treatment 1. If the endowment effect (in Treat-
ment 1) is primarily driven by loss aversion, Treatment 2 should not lead to
a higher investment level than Treatment 3. Treatments 2 and 3 can reveal
whether the gender difference is independent of the endowment effect and loss
aversion. If the endowment effect explains why women invest less in Treatment
1, and the endowment effect is a pure anchoring effect, then women should in-
vest more than men in Treatment 2. If the gender difference in Treatment 1 is
caused primarily by a gender difference in loss aversion that materializes into a
stronger gender difference in the endowment effect, the gender difference should
be reduced in Treatments 2 and 3. If the endowment effect is an anchoring effect
that is independent from loss aversion, but still gender-specific and stronger for
women, women should invest more than men in Treatment 2.
The English versions of the research protocols are included in Appendix 2.
These were translated to the local language Tigrinya which was the language
used in the field. The enumerators were trained with both versions and we en-
sured that the translations were accurate and that the enumerators understood
the questions correctly and used the same exact wording in the local language
for all the questions and explanations.
3. Sampling and implementation
The respondents in the experiment were sampled from rural youth business
groups in northern Ethiopia. The group members were resource-poor rural
youth and young adults that due to their poverty had been found eligible to join
youth business groups in their home communities (tabias) based on their land
poverty, residence, and demonstrated interest in developing a rural livelihood
in their home community. The average age was 31 years and with a standard
deviation of 10 years, giving more age variation than the typical student samples
used in laboratory experiments. The mean level of education was five years but
it varied from no education to 12 years of completed education. Still, financial
and business skills are important for them to succeed in their business activities.
Women constituted close to one third of the group members.
Treatment 1 was used in a baseline survey in the study area in 2016 for a
sample of 1138 youth business group members in 119 business groups in five
districts in the Tigray region of Ethiopia.
The initial endowment of 30 ETB used as the safe amount was equivalent to
a daily rural wage rate in agriculture in the study areas in 2016. This amount
was split in two 10 ETB and two 5 ETB notes which allowed investment levels
of 0, 5, 10, 15, 20, 25 and 30 ETB. We wanted for practical reasons to avoid
the splitting into a finer sub-division which would require the use of coins. This
was also the reason for multiplying the invested amount with three rather than
the 2.5 factor used in the initial Gneezy and Potters study and several other
studies.
5
Local schools were used as field labs. One youth group was interviewed at a
time with 12 enumerators doing the experiments and interviews of 12 members
simultaneously. Three classrooms were used, locating an experimental enu-
merator and a group member in each corner of a classroom. This prevented
communication between group members during the games. It also implied that
the enumerators never interviewed or did experiments with more than one group
member per group, thereby ensuring orthogonality between groups and enumer-
ators, to control for and minimize potential enumerator bias in the estimation.
The low share of respondents investing the full amount in the 2016 exper-
iment led the authors to worry that the design could lead to bias and reveal
respondents as less risk tolerant than they really are. With new funding from
a new project, a follow-up survey was planned in 2019. To test the hypothesis
of an endowment effect, treatment 2 was implemented as a pilot test in one of
the districts together with treatment 3 which should strike a balance between
treatment 1 and 2 which both could create a bias towards their respective safe
and risky initial endowments.
A large share of the sample in this pilot study has also participated in the
2016 experiment, thereby combining a within-subject and between-subject de-
sign. Treatments 2 and 3 were randomized at group level for a sample of youth
business groups and group members (N=243 for Treatment 2 and N=304 for
treatment 3) in the pilot district.
Based on the outcome of this pilot study and the comparison with the base-
line treatment, it was decided to scale up the binary treatment to the full sample
(N=2184).
4. Estimation strategy
The sample from 2016 received the baseline treatment (Treatment 1) while
Treatments 2 and 3 were implemented in in one pilot district in 2019, and Treat-
ment 3 was then scaled up to a large sample of youth business group members in
2019. Treatment 2 was a pilot treatment implemented for a random set of groups
and members in one district. As a first robustness check we assess whether the
responses in this district were systematically different from in other districts for
the baseline treatment in 2016. We also run separate regression models for this
pilot district with the 2016 and 2019 samples jointly and we run full sample
regressions with district fixed effects to control for possible location effects that
may be correlated with the treatment effects. Our design confounds year with
the baseline treatments and there is a risk that the youth have changed their
behavior in the baseline treatment over this three year period. We scrutinized
this by including individual characteristics (sex, age, birth rank and education)
and inspect whether the gain in age over the three years could have changed
their responses. Another potential source of bias could be the enumerators used
in the experiments. While they were doing only one interview per group each,
we had a change in enumerators from 2016 to 2019 based on the quality of their
work and availability (selection of the best available ones for the 2019 survey
and dropping some poor performers). The inclusion of enumerator fixed effects
6
controls for such possible selection bias. We had five enumerators that partic-
ipated in both years and as an additional robustness check we run a separate
model for the sample from the enumerators that were involved in both years.
The share invested from the maximum safe amount (X= 30ET B ) is used
as the dependent variable. This implies that r=x
Xand 0 ≤r≤1.
We use Wilcoxon rank sum tests, also called Mann-Whitney tests (Wilcoxon,
1945; Mann and Whitney, 1947) to compare the distributions of this risk-share
(r) variable across treatments. We also assessed the shares of the samples for
each treatment with r= 1. We used Chi-square tests to compare the frequency
of full investments across the treatment samples.
To further test the treatment effects and to control for other variables, we
estimated linear panel data models with variants of the following specification:
rgi =r1+α2F ullr iskg+α3Binaryg+α4dDd+α5eEd+αgs sgi +gg+gi (1)
where subscript grepresents group, subscript irepresents individual, r1repre-
sents the estimated share invested in the baseline treatment, α2captures the
treatment effect for Treatment 2 as the mark-up share invested in the risky lot-
tery, α3represents the treatment effect for Treatment 3 as the mark-up share
invested, Ddrepresents a vector of district dummy variables, Edrepresents a
vector of enumerator dummy variables, sgi represents a set of individual char-
acteristics (sex, age, birth rank, education), ggrepresents group random effects,
and gi represents the error term.
The following alternative specifications were estimated to test the robustness
of the results:
•1) A parsimonious model that only included the treatment dummies and
the district and enumerator fixed effects on the full sample
•2) A full sample model with additional individual controls,
•3) A model for the sample using the same enumerators in 2016 and 2019,
with additional controls
•4) A model for the pilot district combining 2016 and 2019 data with group
random effects
•5) As d) but with group fixed effects
•6) As d) but with individual fixed effects.
A substantial share of the groups and individuals in the pilot district was
the same in the 2016 and 2019 samples. This facilitated the use of group fixed
effects and individual fixed effects as additional controls for unobserved hetero-
geneity. These are exploited in Table 4 as further robustness checks that allow
control for unobservable time-invariant within-group as well as within-subject
characteristics as well.
The linear panel data models yield coefficients that are marginal effects and
are convenient for that reason. Since our dependent variable is a share with
7
values from zero to one, we also estimated fractional probit models that take
this into account. We have not included the results from these models, however,
because they gave marginal effects that were very close to those from the linear
panel data models.
5. Results
Figure 1 shows the full sample investment distribution for all three treat-
ments. The figure illustrates highly significant differences in distributions across
the three treatments. Figure 2 shows the investment distribution for Treatment
1, comparing the pilot district (Degua Tembien) distribution with that of the
full sample. Degua Tembien was the district where the pilot test of Treatments
2 and 3 took place in 2019. It can be seen that the response distribution in
the pilot district is very similar to that in the full sample. Figure 3 shows the
distribution of investments in the pilot district in 2019 for Treatments 2 (Full
Risk) and 3 (Binary) (243 versus 304 respondents). We see that a substantially
larger share invested the full amount in Treatment 2 than in Treatment 3.
Table 1 presents average shares invested out of the maximum safe amount
that can be obtained for the three treatments in the full sample and in the pilot
district. Table 2 assesses the statistical significance of the treatments using
Wilcoxon ranksum/Mann-Whitney tests for the shares invested. Table 2 also
includes tests for the gender differences for the different treatments to assess
whether these gender differences exist in our sample and are sensitive to the
alternative treatments.
Table 3 presents the results from linear panel data models with youth group
random effects and with standard errors corrected for clustering at the youth
group level. Models (1) and (2) are for the full sample, and Model (2) includes
additional individual controls. Models (1) and (2) also include district fixed
effects and enumerator fixed effects. Model (3) includes the sample for which
the same enumerators were used in 2016 and 2019 as an extra robustness check
for potential enumerator selection bias. We see that the treatment effects re-
main significant and robust. As found in other studies, we also find a highly
significant gender effect with men investing about 5 percentage points more of
the endowment than women do on average.
Table 4 presents models for the pilot district, combining the 2016 and 2019
data and imposing alternative controls for unobserved heterogeneity. Model (4)
includes group random effects. Model (5) includes group fixed effects and Model
(6) includes individual fixed effects. All the models included enumerator fixed
effects.
Result 1:Treatment 2 results in significantly higher average investment
level than Treatment 1.
Result 2:Treatment 2 gives a much larger share of respondents that invest
the full amount than Treatment 1.
Results 1 and 2 imply that we cannot reject Hypotheses 1a and 1b.
Elaboration: These results can be seen by visual inspection of the distri-
butions and the confidence intervals for each investment level in Figure 1 and
8
Tables 1-3. Figure 1 shows that when an initial endowment of 30 ETB is pro-
vided (Treatment 1), a much larger share of the respondents invested only 5 or
10 ETB and a much smaller share invested the full amount of 30 ETB. The find-
ing indicates that we cannot reject Hypothesis 1a that there is an endowment
effect causing the respondents to invest less on average in Treatment 1 than in
Treatment 2. Furthermore, we cannot reject Hypothesis 1b that Treatement 1
is associated with a much smaller share investing the full amount (10.1% of the
sample) than Treatment 2 (37.4% of the sample, see Table 2).
Table 1: Mean shares invested out of the maximum safe amount for alternative treatments
and samples
— Full sample — — Pilot district — Same enumerators
Treatment Mean St.Err N Mean St.Err N Mean St.Err N
T1:Safe Base 0.443 0.007 1138 0.425 0.015 249 0.460 0.011 487
T2:Full Risk 0.691 0.021 243 0.691 0.021 243 0.609 0.035 102
T3:Binary 0.565 0.007 2184 0.611 0.019 330 0.560 0.011 898
Total 0.535 0.005 3565 0.578 0.012 822 0.530 0.008 1487
Figure 1: Distribution of investments in Treatments 1, 2 and 3 (full sample)
Result 3:Treatment 3 resulted in significantly higher average investment
level than Treatment 1 and lower average investment level than Treatment 2.
9
Table 2: Treatment effects and gender differences
Full sample Males Females Wilcoxon Chi-sq.
Treatment Variable Mean St.err. N Mean St.err. N Mean St.err. N p-value p-value
T1 Average share invested 0.443 0.007 1138 0.463 0.009 779 0.399 0.012 359 0.0002
Safe Base Share invest Full amount 0.101 0.009 1138 0.121 0.012 779 0.058 0.012 359 0.001
T2 Average share invested 0.691 0.021 243 0.704 0.026 158 0.669 0.035 85 0.31
Full Risk Share invest Full amount 0.374 0.031 243 0.405 0.039 158 0.318 0.051 85 0.179
T3 Average share invested 0.565 0.007 2184 0.576 0.009 1510 0.540 0.013 674 0.014
Binary Share invest Full amount 0.208 0.009 2184 0.225 0.011 1510 0.171 0.015 674 0.004
All Average share invested 0.535 0.005 3565 0.548 0.007 2447 0.504 0.009 1118 0.0002
Share invest Full amount 0.185 0.007 3565 0.204 0.008 2447 0.146 0.011 1118 0.000
Full sample Degua Tembien Same enumerator
Tests for shares invested z-score P-value z-score P-value z-score P-value
T1 vs. T2 Wilcoxon-Mann-Whitney -10.965 0.0000 -9.078 0.0000 -3.993 0.0001
T2 vs. T3 Wilcoxon-Mann-Whitney 5.744 0.0000 2.770 0.0056 1.425 0.1542
T1 vs, T3 Wilcoxon-Mann-Whitney -10.487 0.0000 -6.448 0.0000 -5.321 0.0000
10
Figure 2: Robustness check (Treatment 1): Pilot district vs full sample
Result 4:Treatment 3 gave a share investing the full amount that is sig-
nificantly larger than that for Treatment 1 and significantly smaller than that of
Treatment 2.
Results 3 and 4 imply that we cannot reject Hypotheses 2a and 2b.
Elaboration: We see from Figures 1 and 3 that the share keeping the full
lottery is reduced significantly when the more balanced binary treatment is used
as compared to Treatment 2. This may also be interpreted as evidence of an
endowment or anchoring effect towards what is initially provided. Treatment
2 may, therefore, give estimates of excessive risk tolerance among respondents.
Treatment 3 may strike a balance and be less biased due to the endowment
effects and thus give the basis for less biased measures of risk aversion.
The Wilcoxon–Mann-Whitney ranksum test results for a comparison of the
investment levels across treatments for the full, the pilot district, and the same
enumerators samples are presented at the bottom of Table 2. The differences
between Treatment 1 versus Treatments 2 and 3 were highly significant in all
samples. The differences between Treatments 2 and 3 were highly significant for
the full and the pilot district samples. However, for the sample utilizing only the
same enumerators in 2016 and 2019 the p-value was only 0.15. The insignificant
test result is associated with the small sample (N=102) that received Treatment
2 and that used the same enumerators in 2016 and 2019. The lack of significance
is thus primarily due to weak statistical power (high standard error) as can be
11
Table 3: Full sample and same enumerator models with controls
(1) (2) (3)
VARIABLES Full sample Full sample Same enumerators
Full Risk 0.206*** 0.210*** 0.161***
(0.029) (0.029) (0.039)
Binary 0.096*** 0.101*** 0.111***
(0.018) (0.018) (0.019)
Male dummy 0.046*** 0.054***
(0.012) (0.018)
Age -0.000 -0.002**
(0.001) (0.001)
Birth rank 0.005** 0.005
(0.002) (0.004)
Education (years) 0.006*** 0.002
(0.001) (0.002)
Constant 0.376*** 0.315*** 0.396***
(0.024) (0.033) (0.046)
Observations 3,565 3,565 1,487
Number of youth groups 308 308 305
All models with district FE and enumerator FE
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Table 4: Robustness checks for pilot district (Degua Tembien) sample
(4) (5) (6)
VARIABLES riskshare riskshare riskshare
Panel controls Group RE Group FE Individual FE
Full Risk 0.185*** 0.197*** 0.174**
(0.040) (0.047) (0.067)
Binary 0.105*** 0.102** 0.114*
(0.039) (0.047) (0.067)
Constant 0.429*** 0.430*** 0.410***
(0.045) (0.044) (0.030)
Observations 822 822 822
R-squared 0.141 0.292
Number of groups 53 53
Number of individuals 593
All models with enumerator FE. Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
12
Figure 3: Robustness check Treatment 2 (Full Risk) and Treatment 3 (Binary) in pilot district
seen from Table 1.
Table 3 demonstrates that the treatment effects are robust to the inclusion of
additional controls. The individual control variables were also assessed for their
systematic variation across treatments, see Appendix Table A1. As Treatment
1 was implemented in 2016 it is not surprising to find a significant age difference
between Treatment 1 versus Treatments 2 and 3. Age had, however, very limited
effect on the investment levels as can be seen in Table 3. Age is insignificant
in Model (2) and significant at 5 percent level in Model (3) but with a very
low coefficient. Five years higher age is associated with a 1 percentage point
lower investment share. The difference in age cannot therefore explain the large
differences in investment levels between Treatment 1 versus Treatments 2 and 3.
The age effect even points in opposite direction of the change in mean investment
levels in 2016 compared to 2019, when the group members have become three
years older.
To assess whether the gender differences in the game can be associated
with gender-differentiated endowment effects, we assessed the gender differences
within each treatment. We compared the means as well as the likelihood that
the whole amount was invested across the treatments. The gender differences
are presented in Table 2 with Wilcoxon-Mann-Whitney ranksum tests for dif-
ferences in means and with Chi-square tests for the difference in the probability
that the whole amount is invested.
13
Result 5:The gender difference in average investment level remains robust
and goes in the same direction across the three treatments.
Elaboration: Result 5 implies that we can reject Hypothesis 3. The strong
gender difference in average investment levels is not explained by the endow-
ment effect associated with the initial endowment provided in the game. And,
the endowment effect is not driven by loss aversion as Treatment 2, which en-
dows respondents with a risky prospect with high potential loss, also creates
an endowment effect. Most of the gender difference in the responses is due to
other things than a gender-differentiated endowment effect. For Treatment 2
there was no significant gender difference but the reason for this is related to
the weak statistical power due to the smaller sample. Table 3 also demonstrates
a highly significant gender difference in investment level in the game. The effect
is not reduced when the regression is run for the reduced sample (Model (3) in
Table 3) which used the same enumerators in 2016 and 2019. Using t-tests for
the gender difference in mean investment levels across treatments revealed no
significant differences in the gender differences (detailed results available upon
request).
To further inspect the robustness of the results, the pilot district sample is
used with alternative controls, see models (4) - (6) in Table 4. We utilize the
fact that for this district many of the same youth groups and group members
were included in the 2016 as well as 2019 samples. This allows us to impose
stronger controls for unobserved time-invariant heterogeneity through the use
of group fixed effects or individual fixed effects. We see from Table 4 that the
treatment effects were robust to these alternative specifications. Treatments 2
and 3 give significantly larger average investment levels than Treatment 1 in all
model specifications and the investment levels are 17.4-19.7 percentage points
higher for Treatment 2 than Treatment 1 and 10.2-11.4 percentage points higher
for Treatment 3 than for Treatment 1.
6. Discussion
We will now discuss our findings in light of findings in other relevant studies.
While variants of the risky investment game of Gneezy and Potters (1997) have
been a very popular tool both in lab and field experiments, we are not aware
of any earlier studies that have investigated the possible endowment effects
associated with the initial endowment provided in the game. This may be
because the game initially was associated with loss aversion, more specifically
myopic loss aversion and repeated versions of the game that are more likely to
trigger loss aversion. It is less obvious that the one-shot version of the game
invokes loss aversion as it may be perceived to operate in the gains domain only.
However, if the initial endowment leads to an immediate shift of the reference
point, investing part of the endowment can be perceived to be in the loss domain
and thereby invoke loss aversion.
Marzilli Ericson and Fuster (2011) showed that expectations could affect the
size of the endowment effects by manipulating the probability of being able to
keep a commodity that was initially given to respondents. A higher probability
14
of being able to keep the commodity was associated with a stronger endowment
effect. They concluded that endowment effects are real but operate via expec-
tations instead of formal ownership. Unlike their experiment, our alternative
treatments do not change the rational expected returns across our treatments.
And our experiment did not include any commodity, only monetary prospects,
and still we found highly significant ”endowment effects” associated with the
prospect that was first allocated to the respondents. Such an ”endowment ef-
fect” was even realized when there was a high probability of loss associated with
the initial prospect. This may appear puzzling also in the light of the findings
of Marzilli Ericson and Fuster (2011).
What makes it particularly interesting to study endowment effects with the
risk investment game is that endowment effects have been attempted explained
by loss aversion (Thaler 1980; Kahnemann et al. 1991). We may critically ask
why a lottery prospect like Treatment 2 invokes an endowment effect if this
endowment effect is explained by loss aversion? The more loss averse should
then sell themselves out of the risky lottery prospect and not be triggered to
become more likely to keep (more of) the lottery prospect. Our results therefore
appear to contradict basic endowment effect theory. This may indicate that the
endowment effect, rather than being explained by loss aversion, is explained by
other things.
We argued that loss aversion is also relevant for money and could lead to an
endowment effect for money for that reason. Our study is in a rural economy
where cash is scarce and this could potentially enhance the endowment effect for
money and possibly also risk aversion and loss aversion due to narrow bracketing.
However, the ”endowment effect” we find may rather be a form of starting
point bias or attachment to the initial prospect that may have manipulated the
subjective reference points to differ from the rational expected returns than be
driven by the potential associated loss (in Treatment 2).
Our results show that endowment effects can introduce biases that will cause
upward bias in estimates of risk aversion from the one-shot version of the game.
Already Charness and Gneezy (2004) found that the game was sensitive to
framing effects. Our Treatments 1 and 2 tested for endowment effects by alter-
natively endowing respondents with a safe amount that can be invested or with
a lottery prospect that the respondents may sell themselves out of at an iden-
tical price in the two treatments. Endowing the respondent with this lottery
prospect treatment dramatically increased the share of the respondents that
accepted the full lottery as well as the average share invested in the game.
An implication, however, is that both these treatments result in bias due to
endowment effects but in opposite directions. We tested for this with our binary
balanced treatment and find highly significant indications of such endowment
or anchoring effects in direction of what is initially provided when comparing
our three treatments. These effects are robust to the alternative econometric
specifications and robustness checks.
Given the tractability of the one-shot game for field experiments, it is of
general interest to know whether the one-shot version of the game can be used
to generate unbiased estimates of risk aversion or risk tolerance.
15
While most studies using the game have not used the game to estimate a
parameter for the risk aversion (capturing the curvature of the utility function)
of respondents, this may be tempting and there are a few studies that also do
this (e.g. Crosetto and Filippin 2016; Dasgupta et al. 2019). An implication
of our study may be that those studies that started by first providing a safe
endowment have overestimated the degree of risk aversion in their samples.
It is possible that the endowment framing has caused risk-neutral persons to
behave as if they were risk averse in the game. However, if also the one-shot
game initiates an immediate change in the reference point and thereby loss
aversion the behavior in the game may be due to a combination of a concave
utility function in gains and loss aversion associated with losses. Loss aversion in
combination with a linear utility function should lead to ”bang-bang” solutions.
This is far from what we observe for all three treatments as the share of interior
solutions was close to 0.9 for Treatment 1, about 0.58 for Treatment 2 and
close to 0.7 for Treatment 3. Table 2 demonstrated large differences in the
shares that invested the full amount in the alternative treatments. This implies
that the endowment effect is caused by other things than loss aversion, such as
psychological transaction costs. Alternatively, moderate changes in loss aversion
and near linear utility curvature contribute to explain the substantial variation
we find in the shares investing the full endowment that we observe in Figure 1
and Table 2.
The fact that the game does not distinguish between risk-loving, risk-neutral
and slightly risk averse individuals has been considered as a weakness of the
design (Crosetto and Filippin 2016). This weakness is also shared with the
Binswanger (1980; 1981) and closely related Eckel and Grossman (2002, 2008)
games. These have been popular methods for the estimation of risk preferences
of poor people in developing countries (Binswanger and Sillers, 1983; Wik et al.,
2004; Yesuf and Bluffstone, 2009). Concerns have been raised recently that
these methods have resulted in biased perspectives on the distribution of risk
preferences in the populations studied with these approaches and that they
have underestimated the share of the populations with near risk-neutral and
risk-loving preferences (Vieider, 2018; Vieider et al., 2019).
We implement a very simple calibration exercise to explore the relationship
between the choices made in the game after endowment effects are eliminated,
assuming this is the case in Treatment 3. We use a Constant Partial Relative
Risk Aversion (CP RRA) utility function; U= (1 −r)(−1) Y(1 −r). We assume
no or limited asset integration and no or moderate levels of loss aversion. With
maximization of expected utility and utility being based only on the outcomes
in the game, the C P RRA −rcoefficient must be very high to lead to optimal
investment levels at or below 10 E T B (38% of the sample investing below 33%
of the endowment). Alternatively, invoking loss aversion and a linear utility
function leads to bang-bang solutions and a switch from investing 30 ET B to
0ET B when the loss aversion parameter increases to 2 or higher. However, a
combination of a reasonably sized CP RRA −rand a lower level of loss aversion
leads to intermediate solutions that also easily can fall in the 1 −10 E T B range
as can be seen in Table 5. We suggest that “weak loss aversion” can be the
16
Table 5: Alternative utility functions and optimal investment level
CP R RA −rAsset integration Loss aversion coeff. Optimal investment amount (ETB)
0.1 No 1 29.9
0.2 No 1 27
0.5 No 1 15
0.7 No 1 11
0.95 No 1 8
0.1 No 1.9 4
0.2 No 1.8 4
0.2 No 1.3 17
0.5 No 1.3 8
0.5 No 1.5 5
0.8 No 1.5 3
0.1 30 ETB base 1.9 8
0.2 30 ETB base 1.9 4
0.2 30 ETB base 1.3 30
0.5 30 ETB base 1.3 16
0.5 30 ETB base 1.5 10
0.8 30 ETB base 1.5 6
result of a weak endowment effect that is invoked for the binary choices. This
would be consistent with the distribution we see. However, this is an area for
possible future investigation. Partial asset integration in combination with weak
risk aversion and weak loss aversion may also yield intermediate outcomes in the
full range of outcomes observed in the Treatment 3 and as calibrated in Table
5.
There are some important variations in the probabilities of winning as well as
the multiplier for the winning outcome compared to the initial game by Gneezy
and Potters (1997) who used it to study the existence of myopic loss aversion
among students and traders. These variations imply that caution has to be
applied when making comparisons across studies. The initial game included a
1/3 probability of winning and a 2/3 probability of losing and the loss was equal
to the amount invested and the amount won was 2.5 times the amount invested.
Our treatments had the same probabilities and multiplier as that of Gneezy
et al. (2009), Gong and Yang (2012) and Dasgupta et al. (2019) who used 50−50
chances of winning and losing and a 3xfactor for the amount won. We narrow
in our comparisons to these studies that have used the same probabilities and
mulitpliers as we used in our study.
Gneezy et al. (2009) applied this design of the game in field experiments in
a matrilineal society in India and a patrilineal society in Tanzania. They found
that on average women in the matrilineal society invested 87% and men 61% of
the initial fund and that women in the patrilineal society invested 61% and men
85%. This can be compared to 40% for women and 46% for men in our baseline
17
treatment which indicates much lower average investment levels and a smaller
average gender difference. Our large sample makes this gender difference highly
significant, however.
Gong and Yang (2012) applied the game in a field experiment in rural China
in a matrilineal and a patrilineal society and to compare risk preferences of men
and women. Even with such a more favorable expected return function, com-
pared to the initial design, the average share of the initial endowment invested
was as low as 4% for the women in the patrilineal group against 37% for men
in the same group and against 33% and 54% for women and men in the matri-
lineal group. Their samples from each group were quite small. The investment
levels they found in the matrilineal society are similar to ours but their gender
difference is larger. Among the men in the matrilineal society they found a sub-
stantial share (about 30%) that invested the full endowment. This compares
to the average of 12% for men and 6% for women in our baseline (Treatment
1) sample. However, it increased to 41% and 32% for our Treatment 2 (Full
Risk), and to 23% and 17% for the Binary treatment. This demonstrates the
endowment effects in Treatments 1 and 2 while Treatment 3 may not suffer from
such bias. However, more research is needed to investigate this further.
The study by Dasgupta et al. (2019) used a large sample of 2000 students
from India and found that only 6% of the males and 1% of the females invested
the full amount. On average males invested 52% of their endowment and females
44%, demonstrating a highly significant gender effect as well as average levels
of investment not very different from our baseline treatment.
There are a few studies that have compared the risk investment game with
other alternative risk preference elicitation methods. Charness and Viceisza
(2016) compare the game with the well-known Holt and Laury (2002) approach
and an un-incentivized survey-based approach (Willingness-to-take-risk) in a
field experiment in Senegal. While they used the 50% and 3xwinning factor they
used a risky seed framing in their experiment rather than money. This framing
may also have had an effect on the responses compared to if they had used
money. They found that the respondents have more cognitive problems with
the Holt and Laury (2002) approach which gave more inconsistent responses.
The risky investment game performed better as it was easier to comprehend
and had better predictive power. They compared their study with a study by
Dave et al. (2010) that also found that simpler games may be preferable for
respondents with limited numeracy skills. We recommend further studies that
compare alternative methods and their predictive power in field settings where
respondents have limited numeracy skills.
Further research is needed on how alternative framing of the this game and
alternative games used to elicit risk preferences affect respondents’ reference
points. The usefulness of the games depend on their predictive power. The
predictive power of a game may depend on the type of real world phenomena
and decisions they are used to predict. Perhaps the risky investment game is
better at predicting whether to invest and how much to invest decisions while
the Holt and Laury game is better for predicting choices among alternative risky
prospects? Perhaps endowment effects are part of the explanation, in addition
18
to limited asset integration, for the puzzling high risk aversion in small gambles
(Rabin 2000).
Just using the investment share as an indication of risk tolerance may be
less problematic as it does not force the measure to be explained by utility
curvature alone. Even it may be acceptable to allow it to include endowment
effects if this endowment effect is associated with loss aversion. Further studies
are needed, however, to test the correlation between this endowment effect and
measures of loss aversion obtained through alternative experimental methods in
within-subject designs.
7. Conclusion
The one-shot version of the Gneezy and Potters(1997) risky investment game
has gained popularity and has been proposed as particularly useful in field
settings with respondents with limited numeracy skills (Charness and Viceizca
2016). We have investigated whether the game can lead to biased estimates of
risk aversion due to inherent endowment effects. We found strong evidence of
such endowment effects associated with the initial endowment allocated in the
game. Our results indicate that the endowment effect is not a result of loss
aversion as respondents endowed with a risky prospect with high probability
of being lost also created an endowment effect. We suggest that the binary
version of the game that we used may give less biased measures of risk aversion
than the game starting by allocating an endowment that respondents are free
to invest some or all of. We recommend more research to further refine the
tool and test its predictive power with alternative incentives and framing and
to further investigate its correlation with separate measures of loss aversion and
risk aversion (utility curvature).
References
Apicella, C. L., Dreber, A., Campbell, B., Gray, P. B., Hoffman, M., and Little,
A. C. (2008). Testosterone and financial risk preferences. Evolution and
Human Behavior, 29(6):384–390.
Binswanger, H. P. (1980). Attitudes toward risk: Experimental measurement
in rural india. American Journal of Agricultural Economics, 62(3):395–407.
Binswanger, H. P. (1981). Attitudes toward risk: Theoretical implications of an
experiment in rural india. Economic Journal, 91(364):867–890.
Binswanger, H. P. and Sillers, D. A. (1983). Risk aversion and credit constraints
in farmers’ decision-making: A reinterpretation. The Journal of Development
Studies, 20(1):5–21.
Charness, G. and Gneezy, U. (2004). Gender, framing, and investment. Tech-
nical report, mimeo.
19
Charness, G. and Gneezy, U. (2012). Strong evidence for gender differences in
risk taking. Journal of Economic Behavior & Organization, 83:50–58.
Charness, G., Gneezy, U., and Imas, A. (2013). Experimental methods: Eliciting
risk preferences. Journal of Economic Behavior & Organization, 87:43–51.
Charness, G. and Viceisza, A. (2016). Three risk-elicitation methods in the
field-evidence from rural senegal. Review of Behavioral Economics, 3(2):145–
171.
Cherry, T. L., Frykblom, P., and Shogren, J. F. (2002). Hardnose the dictator.
American Economic Review, 92(4):1218–1221.
Dasgupta, U., Mani, S., Sharma, S., and Singhal, S. (2019). Can gender differ-
ences in distributional preferences explain gender gaps in competition? Jour-
nal of Economic Psychology, 70:1–11.
Dave, C., Eckel, C. C., Johnson, C. A., and Rojas, C. (2010). Eliciting risk pref-
erences: When is simple better? Journal of Risk and Uncertainty, 41(3):219–
243.
Eckel, C. C. and Grossman, P. J. (2002). Sex differences and statistical stereo-
typing in attitudes toward financial risk. Evolution and human behavior,
23(4):281–295.
Eckel, C. C. and Grossman, P. J. (2008). Forecasting risk attitudes: An experi-
mental study using actual and forecast gamble choices. Journal of Economic
Behavior & Organization, 68(1):1–17.
Epley, N. and Gilovich, T. (2001). Putting adjustment back in the anchor-
ing and adjustment heuristic: Differential processing of self-generated and
experimenter-provided anchors. Psychological Science, 12(5):391–396.
Filippin, A. and Crosetto, P. (2016). A reconsideration of gender differences in
risk attitudes. Management Science, 62(11):3138–3160.
Gneezy, U., Leonard, K. L., and List, J. A. (2009). Gender differences in compe-
tition: Evidence from a matrilineal and a patriarchal society. Econometrica,
77(5):1637–1664.
Gneezy, U. and Potters, J. (1997). An experiment on risk taking and evaluation
periods. The Quarterly Journal of Economics, 112(2):631–645.
Gong, B. and Yang, C.-L. (2012). Gender differences in risk attitudes: Field
experiments on the matrilineal mosuo and the patriarchal yi. Journal of
Economic Behavior & Organization, 83(1):59–65.
Haigh, M. S. and List, J. A. (2005). Do professional traders exhibit myopic loss
aversion? an experimental analysis. The Journal of Finance, 60(1):523–534.
20
Holt, C. A. and Laury, S. K. (2002). Risk aversion and incentive effects. Amer-
ican Economic Review, 92(5):1644–1655.
Horowitz, J. K. and McConnell, K. E. (2002). A review of wta/wtp studies.
Journal of Environmental Economics and Management, 44(3):426–447.
Kahneman, D., Knetsch, J. L., and Thaler, R. H. (1990). Experimental tests of
the endowment effect and the coase theorem. Journal of Political Economy,
98(6):1325–1348.
Kahneman, D., Knetsch, J. L., and Thaler, R. H. (1991). Anomalies: The
endowment effect, loss aversion, and status quo bias. Journal of Economic
Perspectives, 5(1):193–206.
Knetsch, J. L. et al. (1989). The endowment effect and evidence of nonreversible
indifference curves. American Economic Review, 79(5):1277–1284.
Mann, H. B. and Whitney, D. R. (1947). On a test of whether one of two random
variables is stochastically larger than the other. The Annals of Mathematical
Statistics, pages 50–60.
Marzilli Ericson, K. M. and Fuster, A. (2011). Expectations as endowments:
Evidence on reference-dependent preferences from exchange and valuation
experiments. The Quarterly Journal of Economics, 126(4):1879–1907.
Marzilli Ericson, K. M. and Fuster, A. (2014). The endowment effect. Annual
Review of Economics, 6(1):555–579.
Muehlbacher, S. and Kirchler, E. (2009). Origin of endowments in public good
games: The impact of effort on contributions. Journal of Neuroscience, Psy-
chology, and Economics, 2(1):59.
Plott, C. R. and Zeiler, K. (2005). The willingness to pay-willingness to accept
gap, the” endowment effect,” subject misconceptions, and experimental pro-
cedures for eliciting valuations. American Economic Review, 95(3):530–545.
Plott, C. R. and Zeiler, K. (2007). Exchange asymmetries incorrectly interpreted
as evidence of endowment effect theory and prospect theory? American
Economic Review, 97(4):1449–1466.
Rabin, M. (2000). Risk-aversion for small stakes: A calibration theorem. Econo-
metrica, 68:1281–1292.
Rabin, M. and Thaler, R. H. (2001). Anomalies: risk aversion. Journal of
Economic Perspectives, 15(1):219–232.
Samuelson, W. and Zeckhauser, R. (1988). Status quo bias in decision making.
Journal of Risk and Uncertainty, 1(1):7–59.
21
Simonson, I. and Drolet, A. (2004). Anchoring effects on consumers’ willingness-
to-pay and willingness-to-accept. Journal of Consumer Research, 31(3):681–
690.
Thaler, R. (1980). Toward a positive theory of consumer choice. Journal of
Economic Behavior & Organization, 1(1):39–60.
Vieider, F. M. (2018). Violence and risk preference: experimental evidence from
afghanistan: comment. American Economic Review, 108(8):2366–82.
Vieider, F. M., Martinsson, P., Nam, P. K., and Truong, N. (2019). Risk pref-
erences and development revisited. Theory and Decision, 86(1):1–21.
Wik, M., Aragie Kebede, T., Bergland, O., and Holden, S. T. (2004). On the
measurement of risk aversion from experimental data. Applied Economics,
36(21):2443–2451.
Wilcoxon, F. (1945). Individual comparisons by ranking methods. biom bull 1:
80–83.
Yesuf, M. and Bluffstone, R. A. (2009). Poverty, risk aversion, and path depen-
dence in low-income countries: Experimental evidence from ethiopia. Amer-
ican Journal of Agricultural Economics, 91(4):1022–1037.
8. Appendix
Appendix Table A1. Individual characteristics by treatment: t-tests
T1 T2 T3 Total t-tests t-tests t-tests
Safe Base Full Risk Binary T1 vs T2 T1 vs. T3 T2 vs. T3
Age, years 29.07 32.78 32.24 31.27 -3.710*** -3.170*** 0.540
(9.796) (9.216) (9.507) (9.696) (0.685) (0.351) (0.641)
Birth rank 3.105 3.198 3.37 3.273 -0.093 -0.265*** -0.172
(2.002) (1.877) (2.183) (2.110) (0.140) (0.078) (0.146)
Education, years 5.345 5.078 4.608 4.875 0.267 0.737*** 0.470
(3.978) (3.747) (3.968) (3.970) (0.278) (0.145) (0.267)
Observations 1138 243 2184 3565 1381 3322 2427
22
Appendix Table A2. Robustness checks for pilot district: With individual controls
(1) (2) (3)
VARIABLES riskshare riskshare riskshare
Panel controls Group RE Group FE Individual FE
Full Risk 0.188*** 0.198*** 0.174**
(0.041) (0.047) (0.067)
Binary 0.105*** 0.100** 0.114*
(0.040) (0.048) (0.067)
Male, dummy 0.041* 0.035
(0.025) (0.027)
Age -0.001 -0.000
(0.001) (0.002)
Birth rank 0.009 0.008
(0.006) (0.006)
Education, years 0.006** 0.006*
(0.003) (0.003)
Constant 0.359*** 0.355*** 0.410***
(0.065) (0.068) (0.030)
Observations 822 822 822
R-squared 0.149 0.292
Number of groups 53 53
Number of individuals 593
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
23
