Individual Risk Attitudes:
Measurement, Determinants and Behavioral
April 8, 2009
ROA, Maastricht University, and IZA
University of Bonn, IZA, and CEPR
Swarthmore College and IZA
University of St. Gallen, IZA, and CEPR
DIW Berlin, Berlin University of Technology,
Cornell University and IZA Bonn
Gert G. Wagner
DIW Berlin and IZA Bonn
This paper studies risk attitudes using a large representative survey and a complementary
experiment conducted with a representative subject pool in subjects’ homes. Using a question
asking people about their willingness to take risks “in general”, we find that gender, age, height,
and parental background have an economically significant impact on willingness to take risks.
The experiment confirms the behavioral validity of this measure, using paid lottery choices.
Turning to other questions about risk attitudes in specific contexts, we find similar results on the
determinants of risk attitudes, and also shed light on the deeper question of stability of risk
attitudes across contexts. We conduct a horse race of the ability of different measures to explain
risky behaviors such as holdings stocks, occupational choice, and smoking. The question about
risk-taking in general generates the best all-around predictor of risky behavior.
JEL codes: D0, D1 D80, D81, C91, C93, J16, J24, I1
Individual Risk Attitudes:
Measurement, Determinants and Behavioral
April 8, 2009
This paper studies risk attitudes using a large representative survey and a com-
plementary experiment conducted with a representative subject pool in subjects’
homes. Using a question asking people about their willingness to take risks “in
general”, we find that gender, age, height, and parental background have an eco-
nomically significant impact on willingness to take risks. The experiment confirms
the behavioral validity of this measure, using paid lottery choices. Turning to other
questions about risk attitudes in specific contexts, we find similar results on the
determinants of risk attitudes, and also shed light on the deeper question of sta-
bility of risk attitudes across contexts. We conduct a horse race of the ability of
different measures to explain risky behaviors such as holdings stocks, occupational
choice, and smoking. The question about risk-taking in general generates the best
all-around predictor of risky behavior.
(JEL codes: D0, D1 D80, D81, C91, C93, J16, J24, I1)
Risk and uncertainty play a role in almost every important economic decision. As a
consequence, understanding individual attitudes towards risk is intimately linked to the
goal of understanding and predicting economic behavior. A growing literature has made
progress on developing empirical measures of individual risk attitudes, with the aim of
capturing this important component of individual heterogeneity (see, e.g., Bruhin et al.,
2007), but many questions remain unresolved.
One important open question concerns the determinants of individual differences in
risk attitudes. Previous studies have measured risk attitudes using survey questions, and
found mixed evidence on determinants, for example gender.1A second open question,
however, is whether survey questions are really a good method for measuring risk atti-
tudes. Because survey questions are not incentive compatible, economists are skeptical
about whether self-reported personal attitudes and traits are behaviorally meaningful.
Various factors, including self-serving biases, inattention, and strategic motives could
1See, for example, Barsky et al. (1997), Guiso and Paiella (2001) 2005, Donkers et al. (2001), Guiso et
al. (2002), and Diaz-Serrano and O’Neill (2004).
cause respondents to distort their reported risk attitudes (for a discussion, see Camerer
and Hogarth, 1999). Experimental studies, which measure risk-taking behavior with real
money at stake, on the other hand, offer an incentive compatible measure of risk atti-
tudes.2However, a drawback of this technique is that it is costly and difficult to perform
with a large, representative sample, preventing studies at a large scale.3A related, but
largely unexplored issue is how context and question format matter for eliciting risk atti-
tudes. Survey studies have almost always used a question framed in one relatively specific
context: a (hypothetical) decision regarding a financial lottery. It is conceivable, however,
that alternative survey instruments, which are easier to understand than fairly compli-
cated hypothetical lotteries, can deliver similarly reliable information on individual risk
In this paper we address these open questions by investigating responses to a survey
item that asks individuals about a judgment of their own willingness to take risks. This
self-assessment complements traditional lottery-type elicitation of risk attitudes, and we
compare the responses to behavior in paid real-stakes lotteries. The results shed light on
the determinants of risk attitudes in a representative sample, as well as the behavioral
validity of the survey instrument. The paper also compares the responses with multiple
alternative questions about risk attitudes that using different contexts and approaches,
which were asked in the same survey. This allows us to assess the robustness of results
such as a gender difference in risk attitudes to the use of different types of risk measures,
and it is also useful for determining the best way to reliably capture an individual’s risk
taking tendencies in different contexts.
The analysis makes use of different sources of data. The first data source is the
German Socio-Economic Panel (SOEP), which measures the risk attitudes of more than
22,000 individuals. The sample is carefully constructed to be representative of the adult
population living in Germany. The representativeness and statistical power afforded by
2See, e.g., Binswanger (1980), Schubert et al. (1999), Holt and Laury (2002), Barr and Packard (2002),
Eckel et al. (2005), Eckel and Grossman (2007) and Choi et al. (2007).
3A notable exception in the literature is the work by Harrison et al. (2007) who conduct risk experiments
among 253 Danes that were carefully sampled to reflect a representative subject pool of the Danish
the survey allow us to study the determinants of risk attitudes in detail. The survey is
also attractive in that it incorporates a variety of different approaches to measuring risk
attitudes. For example, one question directly asks individual’s to make a global assess-
ment of their willingness to take risks: “How willing are you to take risks, in general?”
Respondents rate their willingness on a scale from 0 to 10. We call this simple, ordinal
measure the “general risk question”. Initially, we use this measure to study heterogeneity
and determinants of risk attitudes in the population. Our first set of results includes
evidence of substantial heterogeneity in risk attitudes, and strong evidence that gender,
age, height, and parental background play an important role in explaining individual
differences in risk attitudes.
A crucial concern is whether survey questions can be meaningfully interpreted in
terms of actual risk-taking behavior. In order to address this question, we use a sec-
ond data source: a field experiment conducted with an additional representative sample
of 450 subjects, drawn from the adult population in Germany using exactly the same
methodology as for the SOEP. Participants in our experiment answer the same general
risk question asked to participants in the SOEP. The respondents also make choices in
a real-stakes lottery experiment. We find that responses to the general risk question are
a reliable predictor of actual risky behavior, even controlling for a large number of ob-
servables. This lends confidence to the behavioral validity of our results for the larger
survey data set, based on the validated survey measure.4More generally, the findings
document that a simple, qualitative survey measure can generate a meaningful measure
of risk attitudes, which maps into actual choices in lotteries with real monetary conse-
quences. This is important because it suggests that surveys can collect information on
risk attitudes using instruments that are easy to use and relatively cheap to administer,
and yet deliver a behaviorally valid measure of risk attitudes. While we are aware that
hypothetical questions are not a perfect substitute for elicitation of risk preferences using
paid lottery experiments, we see considerable value added in documenting the relation-
4In focusing on the consistency of survey responses to a new survey instrument with risk behavior in
lotteries, our investigation about the behavioral validity complements earlier work on hypothetical
bias, see, e.g., Holt and Laury (2002) for a central contribution in the context of risk attitudes.
ship in responses with experimental elicitation instruments, underscoring the usefulness
(and relative advantage) of these measures for elicitation of risk attitudes in large-scale
With the behavioral validity of the general risk question in terms of predicting
behavior in the experiment in mind, we return to an analysis of the larger SOEP data set,
and exploit additional questions about risk attitudes. In particular, the survey includes
five additional questions that use the same scale as the general risk question, but ask
about risk taking in specific contexts: car driving, financial matters, sports and leisure,
health, and career. We find that the results on determinants are robust to using these
alternative risk measures, e.g., females are less willing to take risks in every context. We
also assess the stability of risk attitudes across contexts. In economics it is common to
think of a single trait as governing risk-taking in all contexts, whereas in psychology there
is more controversy on this point.5Our results suggest that risk attitudes are strongly
but not perfectly correlated across contexts. This suggests the presence of a common
underlying risk trait, but also points to some value-added from asking context-specific
Finally, the survey includes self-reported information on several important risky
behaviors: holding stocks, being self-employed, participating in sports, and smoking.
This allows us to compare the relative ability of the different measures of risk attitudes to
explain these risky behaviors. We find that all measures are significantly related to several
behaviors, providing further evidence on their behavioral validity. The best all-around
explanatory variable, however, is the general risk question, which predicts all behaviors.
On the other hand, although less successful across contexts, the single best risk measure
in any given context is the measure incorporating the corresponding specific context. For
example, the best predictor of smoking is the question about willingness to take risks in
health matters, rather than the general risk question or questions incorporating different
contexts. These findings indicate that asking for a global assessment of willingness to take
risks reflects a useful all-around measure of risk attitudes. Questions focused on specific
5See, e.g., Slovic (1972a).
contexts do less well as all-around predictors, but provide strong measures within their
particular domain of risky behavior.
The organization of the paper is as follows. Section 1 describes the SOEP and
the risk measures that we use. Section 2 investigates heterogeneity and determinants
of risk attitudes using the general risk question. Section 3 presents results from the
complementary field experiment. Section 4 assesses the stability of risk attitudes across
different contexts. Section 5 compares the predictive power of the different risk measures
and Section 6 discusses the implications of our results.
1 Data Description
The SOEP is a representative panel survey of the resident adult population of Germany
(for a detailed description, see Wagner et al., 1993, and Schupp and Wagner, 2002).
The initial wave of the survey was conducted in 1984. The SOEP surveys the head
of each household in the sample, but also gives the full survey to all other household
members over the age of 17. Respondents are asked for a wide range of personal and
household information, and for their attitudes on assorted topics, including political and
social issues. Our analysis uses the 2004 wave, which includes 22,019 individuals in 11,803
Much of our analysis focuses on the general risk question in the SOEP, which directly
asks respondents to give a global assessment of their willingness to take risks. The exact
wording of which (translated from German) is as follows: “How do you see yourself: are
you generally a person who is fully prepared to take risks or do you try to avoid taking
risks? Please tick a box on the scale, where the value 0 means: ‘not at all willing to
take risks’ and the value 10 means: ‘very willing to take risks’.”6Notably, the measure
is qualitative and does not involve an explicit lottery. Rather, it relies on the subject
6In German, the wording of the question is: “Wie sch¨ atzen Sie sich pers¨ onlich ein: Sind Sie im all-
gemeinen ein risikobereiter Mensch oder versuchen Sie, Risiken zu vermeiden? Bitte kreuzen Sie ein
K¨ astchen auf der Skala an, wobei der Wert 0 bedeutet: “gar nicht risikobereit” und der Wert 10: “Sehr
risikobereit”. Mit den Werten dazwischen k¨ onnen Sie Ihre Einsch¨ atzung abstufen.” German versions
of all risk questions are available online, at www.diw.de/gsoep/.
to give an assessment of willingness to take risks in general, across the various types of
lotteries (some of which may be non-financial) that could be faced in decision making.
This approach is potentially attractive, for the purpose of eliciting a reliable all-around
measure of risk attitudes across contexts, something which we evaluate empirically in our
analysis. Because there are no explicit stakes or probabilities in the question, there is the
potential that factors other than risk preference could lead to variation in responses across
individuals. Specifically, subjective beliefs about the riskiness of the decision environment
could affect someone’s stated willingness to take risks. For this reason, it is informative
whether the measure explains risky behavior in our field experiment, where choices involve
paid lotteries with explicit stakes and probabilities, and thus subjective beliefs about risk
are held constant. A positive result from the validation exercise would confirm that the
measure does not just reflect subjective beliefs. Five additional measures use the same
wording as the general risk question, and the same scale, but ask about willingness to
take risks in a specific context: car driving, financial matters, leisure and sports, career,
Not every respondent answered all questions, but non-response rates are fairly low.
The number of non-responses out of 22,019 for each of the six question is as follows: 72
for general, 1,349 for car driving, 262 for financial matters, 379 for sports and leisure,
2,051 for career, and 85 for health.
Our analysis also incorporates a field experiment, in which the subjects are a random
sample of the population, drawn using exactly the same procedure as for the SOEP
(sampling is done using the targeted random walk method; see Thompson, 2006).7The
experiment involved a separate subject pool, rather than involving participants from the
SOEP panel. In the experiment, subjects answered a questionnaire similar to the SOEP,
which included the general risk question. Subjects then took part in a lottery experiment.
We describe the experiment in detail in the validation section below. The questionnaire
and experiment were conducted by experienced interviewers who were recruited out of the
7For each of 179 randomly chosen primary sampling units (voting districts), one trained interviewer
was given a randomly chosen starting address. Starting at that specific local address, the interviewer
contacted every third household and interviewed one adult person aged 16 or older per household.
pool of interviewers who conduct the regular SOEP survey. Interviews took place face-to-
face at the subjects’ homes. Both answers to the questionnaire and the decisions in the
lottery experiment were typed into a computer by a professional interviewer (Computer
Assisted Personal Interview (CAPI)). The study was run between June 9th and July 4th,
2005, and a total of 450 participants took part.
2 Willingness to Take Risks in General: Heterogene-
ity and Determinants
This section presents the distribution of willingness to take risks in the population, as
measured by the general risk question, and then turns to the investigation of possible
determinants of individual differences in risk attitudes.
The Figure 1 shows the distribution of general risk attitudes in our representative
sample. Each bar indicates the fraction of individuals choosing a given number on the
eleven point risk scale. The figure reveals substantial heterogeneity in risk attitudes
across the population: the modal response is 5, but risk attitudes vary widely over the
entire scale, with mass distributed over the entire support. A relatively small fraction
of respondents chooses a value of 10, indicating that they are very willing to take risks,
while a somewhat larger mass, roughly 7 percent of all individuals, choose 0, indicating
that they are not at all willing to take risks.
Next, we investigate whether some of the heterogeneity in risk attitudes is system-
atic, thus leading to differences in economic decisions across different types of individuals.
We focus on the impact of four personal characteristics: gender, age, height, and parental
background. These characteristics are plausibly exogenous with respect to individual risk
attitudes and behavior, and thus allow us to give a causal interpretation to correlations
and regression results.8There are also important implications if these characteristics have
an impact on risk attitudes. For example, a gender difference in risk attitudes could be
8Note, however, the caveat that age could potentially be endogenous, for example if people who are less
willing to take risks live longer.
part of the explanation for gender differences in social behavior and economic outcomes.
A first look at the data shows that the proportion of individuals who are relatively
unwilling to take risks, i.e., choose low values on the scale, is higher for women and
increases strongly with age.9Likewise, family background in terms of parental education
appears to play a role in determining risk attitudes, indicating a positive correlation
between parental education and willingness to take risks. Finally, height appears to be a
relevant determinant of risk attitudes, with taller individuals being willing to take risks
for both genders.
To determine the joint role of these four exogenous characteristics simultaneously,
we estimate regressions where the dependent variable is an individual’s response to the
general risk question. Because the dependent variable is measured in intervals, on an 11-
point scale, throughout the analysis we use interval regression techniques.10All estimation
results report robust standard errors, corrected for possible correlation of the error term
across individuals from the same household. The only sample restriction is the omission
of individuals with missing values for the variables in a given regression.
Table 1 summarizes our initial regressions. The baseline specification, presented in
column (1), uses gender, age and height as explanatory variables. The resulting coefficient
estimates show that the unconditional results remain robust. Women are significantly less
willing to take risks in general. Willingness to take risks also decreases significantly with
age. Unreported regressions including age in splines, with knots at 30 and 60 years, reveal
9Willingness to take risks appears to decrease steadily with age for men, whereas for women willingness
to take risks decreases more rapidly from the late teens to age thirty, and then remains flat, until it
begins to decrease again from the mid-fifties onwards. A graphical illustration of this finding as well
as the other unconditional correlations can be found in the discussion paper version of this paper, see
Dohmen et al. (2005). Interestingly, this pattern in early life for women is not solely driven by the
occurrence of childbirth; the pattern is similar for the sub-sample of women who never have a child.
10Rather than interpreting responses as continuous variable, this approach treats each value of the
dependent variable as a left and right censored observation coming from an interval with known bounds
and accounts for censoring. The interval regression procedure maximizes a likelihood function that is
a natural generalization of a Tobit. Error terms are assumed to be normally distributed. Reported
coefficients reflect marginal effects. We also estimated all regressions using Ordered Probit models, OLS
regressions, or using a binary Probit classification with a measure of risk attitudes as the dependent
variable, which takes a value of 1 for individuals who report a value above 5 on the general risk scale
and 0 otherwise. In all instances we found the same qualitative results and similarly significant and
robust coefficients. Results based on any of these alternative estimation methods are available upon
that the age effect is particularly strong for young and old ages. The inclusion of splines
leaves the estimates of the other coefficients virtually unchanged.11
Taller people are
more willing to take risks. All of these effects are individually and jointly significant at
the 1-percent level.12Column (2) repeats the same estimation adding indicator variables
for having a mother or father who have completed the Abitur as exogenous characteristic
from the individual point of view.13Having a mother or father with completed Abitur
significantly increases willingness to take risks. Again, this effect is individually and
jointly significant at any conventional level.
Columns (3) to (6) check the robustness of our findings by including other control
variables. Two potentially important controls are income and wealth. High income or
wealth levels may increase the willingness to take risks because they cushion the impact of
bad realizations. A potential problem with adding these variables to the regression is that
they may be endogenous, e.g., a greater willingness to take risks could lead to high wealth
levels. Wealth and income are sufficiently important economic variables, however, that it
is arguably important to know how they affect the baseline results when they are included
in the regression.14
Column (6) is the fullest specification, controlling for household
wealth and income simultaneously, and also adding a large number of other personal
and household characteristics. These additional characteristics, which are all potentially
endogenous, include among others: marital status, socialization in East or West Germany,
nationality, employment status (white collar, blue collar, private or public sector, self-
employed, non-participating), education, subjective health status, and religion.15For the
11Results for spline regressions are available upon request.
12Results from OLS regressions deliver almost identical results. The corresponding R-square measure is
0.12. A likelihood-ratio test reveals that adding interaction terms between all independent variables
improves the fit. The coefficients of interest in the unrestricted specification, however, are very similar
to those from the restricted model, both qualitatively and quantitatively. We prefer the model reported
in column (1) of Table 1 for ease of presentation and interpretation.
13There are two types of high school in Germany, vocational and university-track. The Abitur is an
exam that is completed at the end of university-track high schools and qualifies an individual to attend
university. Thus the Abitur is an indicator of relatively high academic achievement, especially for older
cohorts. Hence, we take whether or not a parent passed the Abitur, an exam that comes at the end of
university-track high school in Germany and is a prerequisite for attending university, as a proxy for
highly-educated parents. In our sample, roughly 7 percent of mothers and 13 percent of fathers have
completed the Abitur.
14For details on the construction of income and wealth measures in the SOEP, see the notes for Table 1.
15Again, OLS results are almost identical, with an R-squared of 0.17 for a specification corresponding to
sake of brevity, Table 1 does not report coefficient estimates for all of the additional
controls. The precise specification and all coefficient estimates are shown in Table A.1 in
A comparison of the results in columns (3) to (6) to results in columns (1) and (2)
shows that the coefficient estimates for gender, age, and height are virtually unchanged,
and remain equally statistically significant when we include the additional income and
wealth controls. Women are less willing to take risks than men.16Increasing age leads to
decreasing willingness to take risks, and increasing height leads to a greater willingness
to take risks. Having a mother who completed the Abitur, and somewhat less strongly
having a father who completed the Abitur, increases an individual’s willingness to take
Importantly, the effects of gender, age, height, and parental education on willingness
to take risks are also quantitatively significant. For example, given that one standard
deviation for the general risk question is about 2.4, the gender effect corresponds to a
substantial decrease in willingness to take risks, about one quarter of a standard deviation.
In the last section of the paper, where we relate risk attitudes to different important risky
behaviors, we return to a discussion of the economic significance of the determinants
identified in this section.
Although causal interpretations are inadvisable, it is interesting that the correlation
between wealth or income and risk attitudes goes in the predicted direction, i.e., these
correlations are invariably positive and significant, indicating that wealthier individuals
are more willing to take risks. For example, the unconditional correlation between general
risk attitudes and log household income is 0.20, and the correlation with log household
wealth is 0.06, significant at any conventional level. The positive relationship between
income or wealth and willingness to take risks remains when we control for other observ-
that reported in column (6).
16We also studied the gender difference in risk attitudes using Oaxaca-Blinder decomposition techniques
(see Blinder, 1973, and Oaxaca, 1973), which allow the separation of differences in observable charac-
teristics from differences in regression coefficients. We found that more than 60 percent of the gender
gap is explained by differences in coefficients rather than characteristics, regardless of the specification
or the reference group chosen.
ables (see coefficient estimates in column (1) of Table A.1 in the Appendix). There are
also a number of other noteworthy correlations, reported in Table A.1. For example, being
widowed, having a bad subjective health status, and being out of the labor force are all
significantly negatively correlated with willingness to take risks. Willingness to take risks
decreases with number of children. People with high life satisfaction are more willing to
3 Experimental Validation of Survey Measures
The previous section identified several exogenous factors that determine individual risk
attitudes. Importantly, these conclusions were drawn from a large and representative
survey. The scope of the results is therefore considerably larger than that of economic
experiments, which typically use a relatively small and often selective sample. A serious
concern with the use of hypothetical questions, however, is that responses are not incentive
compatible. As a result it is unclear to what extent the general risk question is a reliable
indicator for actual risk taking behavior.
In light of this discussion, the researcher who is interested in the measurement of risk
attitudes faces a dilemma. Running an incentive compatible experiment with, say, 22,000
subjects is hardly a feasible option, given the substantial associated administrative and
financial costs. Conducting experiments with affordable but relatively small sample sizes,
on the other hand, leaves the researcher with limited statistical power. In this paper our
solution is to run a large survey including risk measures but also a complementary field
experiment that tests the behavioral validity of the survey measures. This procedure offers
the advantages of both statistical power and confidence in the reliability of the survey
questions.17In order to validate our survey risk measure, we ran a lottery experiment
based on a representative sample of adult individuals living in Germany, and had the same
individuals answer a detailed questionnaire. Of course, it would also be possible to validate
17There is a parallel with the literature on contingent valuation, which has used experiments to explore
the validity of survey questions about preferences for environmental amenities or other goods. See,
e.g., Blackburn, et al. (1994), and Champ and Bishop (2001), among others.
the measure in a lab experiment with undergraduates, a relatively easy and potentially
less expensive option. Strictly speaking, however, this would only allow validation of the
survey questions for this special subgroup of the total population, which is why we decided
on our alternative design.18
In our experimental study, subjects first went through a detailed questionnaire, simi-
lar to the standard SOEP questionnaire. As part of the questionnaire we asked the general
risk question analyzed in the previous section. After completion of the questionnaire, par-
ticipants took part in a paid lottery experiment.19In the experiment participants were
shown a table with 20 rows. In each row they had to decide whether they preferred a safe
option or playing a lottery. In the lottery they could win either 300 Euros or 0 Euros with
50 percent probability (1 Euro ∼ $ US 1.2 at the time of the experiment). In each row
the lottery was exactly the same but the safe option increased from row to row. In the
first row the safe option was 0 Euros, in the second it was 10 Euros, and so on up to 190
Euros in row 20. After a participant had made a decision for each row, it was randomly
determined which row became relevant for the participant’s payoff. Depending on the
subject’s choice in that row, the subject’s payoff would either be the safe payment from
that row, or the outcome of the lottery. This procedure guarantees that each decision
was incentive compatible. Once a respondent preferred the safe option to playing the
lottery, the interviewer asked whether the respondent would also prefer the even higher
safe payments to playing the lottery, and all subjects responded in the affirmative. The
switching point is informative about a subject’s risk attitude. Since the expected value
of the lottery is 150 Euros, weakly risk averse subjects should start to prefer the safe
18Several recent studies have pointed out that student samples are not representative for a broad popu-
lation, and have therefore run experiments on representative samples, see, e.g., Bellemare and Kr¨ oger
(2007) in the context of trust, and Harrison et al. (2007) in the context of risk experiments.
19The general risk question was asked as question 19 in the questionnaire, and the risk experiment was
conducted after subjects had answered question 104. Thus, a considerable amount of time (more than
20 minutes) passed between answering the general risk question and making choices in the lottery
experiment, during which participants answered 85 different questions, many with several sub-items.
This time delay, together with the fact that the lottery choices involved considerable financial incentives,
helps eliminate any spurious relationship between the general risk question and lottery choices, arising
due to a psychological desire to make choices that are consistent with previous survey responses. In
fact, see Saris and van Meurs (1990) and Saris (2003) for evidence that a time delay of 20 minutes
between measures suffices to eliminate memory effects.
option over the lottery for save payments that are less than 150 Euros. They should
also prefer larger safe payments to the lottery. Only risk loving subjects should opt for
the lottery when the offered safe option is 160, 170, 180, or 190 Euros. The increments
in the safe payment, and maximum value of 190, were chosen to allow a relatively fine
grid for categorizing different degrees of risk aversion, risk neutrality, or risk lovingness,
while also keeping the length of the choice table manageable. For example, safe payments
higher than 190 would either require lengthening the table, or using a coarser grid. As
very few individuals are typically extremely risk loving, a maximum of 190 was chosen.
As discussed below, the fact that the table is bounded at 190 is inconsequential for our
In order to ensure incentive compatibility, subjects were informed that after the
experiment a random device would determine whether they would be paid according to
their decision, and that the chance of winning was 1/7 (see Laury, 2006, for evidence that
this delivers very similar results to the alternative procedure of paying each subject with
probability 1). At the end of the experiment subjects learnt the outcome of the chance
move, and in case they won they were paid by check sent to them by mail.20
Ideally, subjects who take part in the experiment should be as similar as possible to
the SOEP respondents, in particular with respect to the exogenous factors that explain
individual risk attitudes. We tried to make the samples as similar as possible by adopting
the exact same sampling methodology, and using the same surveying company as is used
for the SOEP (for detailed documentation see also Schupp and Wagner, 2007). As the
upper panel of Table 2 shows, the two samples are in fact quite similar. The fraction
of females is 52.7 percent in the experiment and 51.9 percent in the SOEP data. Also,
both mean age and median age of the participants are very similar. The same holds for
height. This congruence reflects the representative character of the experimental subject
pool. Table 2 also shows that the mean and median response to the general risk question
is very similar. While the mean (median) value in the experiment is 4.76 (5), it is 4.42 (5)
20Sending checks is a particularly credible procedure in our case because the interviewers came from one
of Germany’s leading and most distinguished institutes in the field of social science survey research.
None of the interviewers reported any credibility problems in their interviewer reports.
for the people who are interviewed in the SOEP. In addition, the answers to the general
risk question are almost identically distributed.21
Closer investigation of the responses reveals that about 78 percent of the participants
are risk averse. They prefer not to play the lottery, which has an expected value of
150 Euros, when offered a safe payment smaller than 150 Euros.About 13 percent
are arguably risk neutral: 9 percent prefer a safe payment of 150 Euros to the lottery,
but play the lottery at smaller alternative options, and 4 percent play the lottery when
offered a safe payment equal to the expected value of the lottery but do not play the
lottery when the safe payment exceeds the expected value of the lottery. About 9 percent
of the subjects exhibit risk loving behavior, preferring the lottery to safe amounts above
150 Euros. These proportions of risk averse, risk neutral, and risk loving individuals
are closely in line with previous findings.22There is some indication for the commonly
observed tendency for subjects to switch at “prominent” numbers. We verify below that
our results are robust to using smoothed measures of risk aversion that lump together
switching rows into broader categories, and to using various estimation techniques that
correct for potential censoring.
Our main interest in this section is whether survey data can predict actual risk taking
behavior in the lottery experiment. In other words, we want to study whether greater
willingness to take risks in the general risk question maps into a greater willingness to
take risks in the lottery experiment. To test the predictive power of the general risk
question, we ran the regressions reported in the lower panel of Table 2. In the first model,
we simply regress the value of the safe option at the switching point on answers given to
21While a t-test rejects the null hypothesis that the mean response is identical in the two samples, we
cannot reject the null hypothesis that the answers to the survey risk questions in the two samples have
the same distribution on basis of a Kolmogorov-Smirnov-test. Furthermore, an interval regression
using the specification in column (1) of Table 1 shows that determinants of responses to the general
risk question are very similar across the two samples. For example, we cannot reject the null that the
coefficients for gender, age, and height observed in the field experiment are the same as those observed
in the SOEP survey sample.
22For example, in a condition with comparable stakes, Holt and Laury (2002) observe 81 percent risk
averse, 13 percent risk neutral, and 6 percent risk loving. They use a similar choice-table procedure to
obtain incentive compatible measures of risk preference, where switching rows categorize individuals
into different intervals of risk aversion or risk-lovingness, although they vary the probabilities instead
of the safe payment. See also Harrison et al., 2007, who use a similar measure and find a similar
distribution of types in a representative sample of Danes.
the general risk question. The coefficient on general risk is positive and significant at any
conventional level, indicating that the answers given in the survey do predict actual risk
taking behavior. To check robustness, we add controls in columns (2) and (3), which are
essentially the same as the controls in Table 1. Specifically, controls in column (2) include
gender, age, and height, and in Column (3) we control for many additional individual
characteristics, as in Table A.1 in the Appendix. The general risk coefficient becomes
somewhat smaller but stays significant at the one percent level. These results are also
robust to alternative estimation approaches that use a smoothed version of the dependent
variable, or adjust for potential censoring of the dependent variable.23As an additional
robustness check, we also estimated the same regressions as in Table 2, but weighting
responses to the general risk question in the field experiment by the respective fractions
of responses to the general risk question in the SOEP sample. This makes the two samples
exactly comparable in terms of the proportion of individuals choosing a given point on the
general risk scale. We find similar results, indicating that the small differences between
the two samples do not affect the generalizability of the validation results.
In summary, the answers to the general risk attitude question predict actual be-
havior in the lottery quite well. The results are robust even controlling for a wide array
of observable characteristics, and using various estimation techniques, confirming the be-
havioral relevance of this survey measure.
4 Determinants and Stability Across Contexts
In this section we turn to the five questions that ask about willingness to take risks in the
specific contexts of car driving, financial matters, sports and leisure, health matters, and
career, respectively. We investigate the determinants of willingness to take risks in each
23We checked robustness to combining switching row information into six broader categories, each con-
taining one of the prominent number spikes. We find that the general risk question is still a highly
significant predictor of willingness to take risks in this case, confirming that prominent number effects
do not explain the results. As an alternative approach we also estimated a Cox proportional hazard
model, which accounts for the fact that individuals cannot switch beyond values of 190. In this case we
find that a higher value for the general risk scale has a highly significant negative impact on the hazard
of switching to the safe option in the choice table (switching later in the choice table indicates greater
willingness to take risks), which again confirms the behavioral validity of the general risk measure.
context, to assess whether our results so far on gender, age, height, and parental education
are robust to using alternative measures of risk attitudes. The analysis also sheds light
on the deeper question of whether there is a stable trait determining risk-taking across
different domains of life. In economics it is common to think of individuals as having
an underlying risk preference that affects willingness to take risks in all contexts. By
contrast, there is considerable controversy on this point in psychology (see Slovic, 1972a
and 1972b; Weber et al., 2002). We contribute to the discussion in several ways. In a
representative sample, we investigate whether the same factors determine risk attitudes
across contexts. We also assess the strength of the correlation of willingness to take risks
across contexts, and perform a principle component analysis to determine whether a single
principle component determines risk taking in all contexts.
4.1 Determinants of Risk Attitudes in Different Contexts
Table 3 explores the determinants of individual risk attitudes in each of the five specific
contexts identified in the SOEP. For ease of comparison, the first column reports results
for the general risk question, shown previously in column (2) of Table 1. Columns (2)
to (6) show that the impact of the exogenous factors is, for the most part, qualitatively
similar across contexts.24Women are significantly less willing to take risks than men in
all domains. This result is robust when adding the full set of controls. Gender differences
are most pronounced for risk attitudes in car driving and financial matters, and least
pronounced in the career domain. Increasing age reduces willingness to take risks in all
five domains, but has a particularly large impact in the domain of sports and leisure, and
a relatively small impact in financial matters. The table also shows that taller persons
are more willing to take risks, in all domains. The height effect is particularly strong for
car driving, sports and leisure, and career. The relationship between parental education
and risk attitudes is less consistent across domains. Father’s education has a positive and
significant impact on risk taking in all contexts, and mother’s education is important for
risk taking in sports and leisure and career. Adding additional controls to the regressions
24The results are robust to estimating Ordered Probits, or OLS regressions.
has no impact on the qualitative results, and most effects remain similarly significant
across domains, with the main exception of parental education, which becomes somewhat
weaker. Regressions with all controls are reported in Table A.1 in the Appendix.25In
summary, these results provide further support for the importance of gender, age, height,
and to a certain extent parental education, for determining willingness to take risks. The
qualitatively similar impact of these factors across contexts is also suggestive of a common
underlying risk preference.
4.2 Stability of Risk Attitudes Across Contexts
Mean responses for each context-specific question and the general risk question, reported
in Table 4, suggest that context matters for self-reported willingness to take risks.26The
ranking of means in willingness to take risks, from greatest to least, is as follows: general,
career, sports and leisure, car driving, health, and financial matters. The ranking is nearly
identical for both men and women.
The next section in the table shows that risk attitudes are not perfectly correlated
across contexts, but the pairwise correlations are large, typically in the neighborhood of
0.5, and all are highly significant. This is suggestive of a stable, underlying risk trait,
but there also appears to be some sensitivity of risk attitudes to context. This could
also reflect variation in risk preference, or other factors, such as subjective beliefs.27For
example, most people may view the typical risk in car driving as more dangerous than
the typical risk in sports, and thus to state a relatively lower willingness to take risks in
25The coefficient estimates are virtually identical in specifications using age splines with knots at 30 and
60 years of age, and are available upon request.
26The different number of observations across domains reflects different response rates. These differences
may arise because individuals feel certain questions do not apply to them, e.g., a 90-year-old without
a driver’s license is free to leave blank the question about taking risks while driving a car.
27Measurement error could also contribute to a less than perfect correlation across questions.
28Evidence from psychology suggests that risk perceptions may vary across individuals and contexts.
For example, women may perceive dangerous events (nuclear war, industrial hazards, environmental
degradation, health problems due to alcohol abuse) as more likely to occur, in conditions where objec-
tive probabilities are difficult to determine (Silverman and Kumka, 1987; Stallen and Thomas, 1988;
Flynn et al., 1994; Spigner et al., 1993). Differences in risk perception could indicate that beliefs are
formed based on systematically different information sets. Alternatively, biases due to emotion (e.g.,
Principal components analysis using the general risk question and the five domain-
specific questions tells a similar story. About 60 percent of the variation in individual
risk attitudes is explained by one principal component, consistent with the existence of a
single underlying trait determining willingness to take risks.29All of the questions capture
this underlying trait to some extent.30Nevertheless, each of the other five components
explains at least five percent of the variation. This again suggests that the individual
measures do capture some additional, context-specific content, as well as capturing a
trait that is common across contexts and pure measurement error.
5 Survey Risk Measures and Risky Behaviors
The last part of our analysis returns to the issue of behavioral validity. Previously we
demonstrated the ability of the general risk measure to predict real-stakes lottery choices
in a field experiment. In this section we study a broader range of risky behaviors, en-
compassing a variety of important economic and social contexts, and test the validity
of all six risk measures. In so doing, we hope to answer three questions. First, are the
survey instruments useful for explaining and predicting risky behavior, in terms of both
statistical and economic significance. Secondly, can the individual risk measures explain
risky behaviors in multiple contexts? Third, how do the different risk measures compare
in terms of explanatory and predictive power? In particular, how does the general risk
question as global measure of risk attitudes perform in comparison to context-specific
measures in a particular context? For example, is smoking best predicted by a health
related risk question or is it equally well explained by the general risk? The answer to
these questions is of obvious importance from both a methodological and a practical point
fear or dread), or overconfidence, could play a role in explaining different risk perceptions (Slovic, 1999;
Loewenstein et al., 2001).
29The eigenvalue associated with this component equals 3.61 while the eigenvalues associated with all
other components are smaller than 0.57. When only one component is retained, none of the off-diagonal
elements of the residual correlation matrix exceeds |0.11|.
30The factor loadings for the different risk questions on the principal component are 0.78 (general), 0.76
(car driving), 0.74 (financial matters), 0.80 (sports and leisure), 0.81 (career), and 0.75 (health).
To address our questions we use a collection of behaviors reported by the SOEP
participants. We choose behaviors that span the different contexts identified by the five
domain-specific questions: portfolio choices (financial context), participation in sports
(sports and leisure), self-employment (career), and smoking (health). All of these risky
behaviors are measured as binary variables and are displayed in Table 5, with the exception
of traffic offenses, which are analyzed separately at the end of the section. As a proxy
for portfolio choice we use information about household stock holdings. The variable
“Investment in Stocks,” shown in column (1), is equal to 1 if at least one household
member holds stocks, shares, or stock options and zero otherwise. Since the question
about stock holdings is typically answered by the household head, we use observations
on risk preferences of household heads in column (1) only. In the context of sports, the
variable “Active Sports” takes a value of 1 if an individual actively participates in any
sports (at least once per month). The variable “Self-Employment” is a binary variable
equal to 1 if an individual is self-employed and zero otherwise. To study risk-taking
behavior in the domain of health, we use information about whether the SOEP participant
smokes or not. The corresponding variable is equal to 1 if the respondent smokes.
Each reported coefficient estimate shown in Table 5 is based on a separate regres-
sion of the respective behavior on this particular risk measure and a set of controls.31
For these regressions we use standardized versions of the risk measures, to aid compari-
son of coefficients across regressions. The regressions are Probit models, and coefficients
are marginal effects showing the impact of a one standard deviation change in the cor-
responding measure of willingness to take risks. We report the standard errors of the
coefficients in brackets, and the log likelihoods in parentheses. For example, the three
entries in the upper left corner in column (1) say that the willingness to take risks in
general is significantly correlated with investments in stocks, the marginal effect being
0.029. The standard error of the coefficient is 0.006 and the log likelihood for this regres-
31In every regression the controls include gender, age, height, and parental education, as in column (2)
of Table 1, but also log household wealth, log household debt, and the log of current gross monthly
household income. One exception is the regression for stock holdings, where we also control for the
number of household members older than 18, because the likelihood that at least one person in the
household holds risky assets is increasing with household size.
sion is -3,993. Coefficient estimates for the controls are not reported but are available on
request. Each column in the table also reports the unconditional probability of observing
the corresponding risky behavior.
Several observations can be made from Table 5. First, all measures are significant
explanatory variables for at least some of the behaviors, providing further confirmation
of their behavioral validity. The marginal effects are also sizeable relative to the uncon-
ditional probabilities, showing the economic significance of the risk attitude measures.
Second, the general risk question is significant in all contexts, with relatively large coeffi-
cients and goodness of fit. Third, each context-specific risk question explains behavior in
its respective context, and is typically the strongest risk measure for this context. Again,
this can been seen by comparing the size of the marginal effects and the log likelihoods
of the different regressions.
Investment in stocks, shown in column (1), is positively correlated with several risk
measures, as expected given the relative riskiness of this kind of financial investment.
The best predictor is the context-specific risk measure on “Financial Matters”. The log
likelihood for this regression is larger than for any other regression based on a differ-
ent risk measure, and the marginal effect is also larger. Notably, the marginal effect is
economically significant: a one standard deviation increase in willingness to take risks
in financial matters is associated with a 34 percent increase in the probability of hold-
ing stocks (relative to the unconditional probability). This implies that the exogenous
determinants of risk attitudes in the financial context, shown in Table 3, are also quan-
titatively important, through the channel of changing risk attitudes. For example, the
gender coefficient in column (3) of Table 3 was -0.77. Given that the standard deviation
of willingness to take risks in financial matters is 2.225 (see Table 4), this is a decrease of
about a third of a standard deviation. The results in Table 5 then imply that a quarter
standard deviation decrease in willingness to take risks is associated with a 12 percent
decrease in the probability of holding stocks.32Analogous calculations show that any one
32Thus, gender has a sizeable indirect effect on the probability of holding stocks through the channel of
risk attitudes. This is also sizeable relative to the direct effect of gender on holding stocks (controlling
for risk attitudes in financial matters and other characteristics). The (unreported) estimated marginal
of the following – 20 fewer years of age, 20 additional centimeters of height, or having two
highly-educated parents – would increases the probability of holding stocks by about 8
percent, through the indirect channel of changing willingness to take risks.33
Column (2) of Table 5 shows that being active in sports is strongly correlated with
several risk measures, but the measure of risk taking in “Sports and Leisure” is the most
important in terms of statistical and economic significance. A one standard deviation
increase in willingness to take risks in the context of sports is associated with about a 22
percent increase in the probability of participating in active sports.34
In column (3) we investigate the relationship between risk attitudes and career
choice. Given the low degree of job security and high income variability associated with
self-employment, we would expect that relatively risk-averse people choose not to become
self-employed. In fact, the coefficients on the risk measures are significant and positive,
consistent with those who are more willing to take risks being more likely to choose self-
employment. Exceptions include risk attitudes in sports and leisure, which are unrelated
to being self-employed.35Notably, willingness to take risks in career matters is the overall
best predictor of self-employment, in terms of significance, fit, and size of the marginal
effect. A one standard deviation increase in willingness to take risks in the career context
is associated with a 43 percent increase in the likelihood of being self-employed.36
In column (4) we turn to risky health behavior in the form of smoking. Willingness
to take risks in general has a strong positive impact on the propensity to smoke, but
effect for gender in Table 5, in the regression involving risk attitudes in financial matters, is -0.083.
33The impact through the channel of the general risk attitude is smaller, in any given context, than
through the corresponding context-specific attitude. The total impact of a change in general risk
attitudes, however, is substantial.
34Because gender and age affect willingness to take risks in the context of sports (s.d. = 2.613), being
female, or 20 years older, decreases the probability of participating in sports by about 5 or 10 percent,
respectively. Through the channel of changing risk attitudes, 20 additional centimeters of height
increases the probability of participating in sports by about 6 percent, and having two parents with
the Abitur is associated with an increase of about 7 percent.
35For related evidence on risk aversion and entrepreneurship, see Cramer et al. (2002).
36The impact of being female, or 20 years older, that works through the channel of willingness to take
risks in career matters (s.d. = 2.71), is an 9 percent or 18 percent decrease in the likelihood of being
self-employed, respectively. 20 additional centimeters of height, or having two parents with the Abitur,
increases the probability of being self-employed by 12 percent or 10 percent, respectively, through the
channel of greater willingness to take risks in career matters.
willingness to take risks in the domain of health has an even greater impact, as indicated
by the larger marginal effect and the higher log likelihood. The case of smoking is of
particular interest given that smoking has been used as an instrument for risk attitudes,
in cases where direct measures of risk attitude were not available (e.g., Feinberg, 1977). In
light of results in column (5), however, smoking can only be considered a very imperfect
substitute for more direct measures of risk attitudes. While smoking is strongly associated
with the willingness to take risks in the health domain, it is less correlated or not correlated
at all with risk attitudes in other domains such as financial matters or sports and leisure.
The impact of a one standard deviation increase in willingness to take risks in health
matters is a 20 percent increase in the probability of being a smoker.37
In additional (unreported) regressions we also tested the relative explanatory power
of the different risk measures by regressing a given behavioral outcome on all of the
measures simultaneously. The results are very similar to the ones in Table 5, in the sense
that the corresponding domain specific risk question is the best predictor of investment
in stocks, participation in sports, self-employment, and being a smoker.38
This paper contributes evidence on several open questions regarding the measurement and
nature of individual risk attitudes. Using a simple survey measure that asks people to give
a global assessment of their willingness to take risks in general, we find an economically sig-
nificant impact of gender, age, height, and parental background on individual willingness
to take risks. The behavioral validity of the survey measure is verified in complementary,
incentive compatible field experiment conducted using a representative subject pool. The
results on determinants of risk attitudes are robust to using other survey measures that
37Through the channel of willingness to take risks in health matters (s.d. = 2.465), being female, or
being 20 years older, reduces the probability of smoking by 6 percent, respectively. The change in risk
attitudes due to 20 additional centimeters of height, or the impact of two highly educated parents on
risk attitudes increases the probability of smoking by about 3 percent, respectively.
38We also compared the performance of the different risk measures in terms of prediction accuracy when
splitting the sample and using the estimates from one subsample to predict behavior in the other
subsample. Detailed results are available upon request.
ask about risk taking in specific contexts. Risk attitudes are shown to be relatively stable
across different contexts, shedding light on a deeper question about stability of willingness
to take risks as a personal trait. All of the survey measures are shown to explain various
risky behaviors, including holding stocks, smoking, self-employment, and participation in
active sports. The best all-around predictor is the general risk question. On the other
hand, asking about risk attitudes in a more specific context gives a stronger measure for
the corresponding context.
The findings on the determinants of risk attitudes have important implications.
For example, previous experimental research has documented that women are less will-
ing to sort into relatively risky, tournament compensation schemes (e.g., Niederle and
Vesterlund, 2007). A gender difference in willingness to take risks could be part of the ex-
planation for this important difference in behavior. In fact, another study also finds that
women prefer less risky compensation over tournaments (Dohmen and Falk, 2006), but
also asks our general risk question. They find that lower willingness to take risks among
females explains a substantial part of the gender difference in sorting decisions. An age
profile for risk attitudes could also be important, for explaining behavior at the macroe-
conomic level: demographic changes leading to a more elderly population are predicted
to lead to a more conservative pool of investors and voters, which could substantially
influence macroeconomic performance and political outcomes, increase the resistance to
reforms, and delay necessary but potentially risky policy adjustments. A role for parental
education in shaping the risk attitudes of children highlights a potentially important role
of education policy, and has implications for understanding intergenerational correlations
in economic outcomes and social mobility. The impact of height on risk attitudes could
be relevant for explaining another important puzzle, namely the documented relationship
between height and labor market earnings.39
The results of this paper also have implications for measuring risk attitudes in self-
39Persico et al. (2004) find that height in adolescence, more than height in adulthood, has an impact
on wages later in life. The authors’ hypothesis is that the height effect is due to the impact of height
in adolescence on confidence and self-esteem. Our results suggest that another by-product could be
greater willingness to take risks, which is a plausible channel leading to higher future wages.
reported surveys. They indicate that the approach of asking people for a global assessment
of willingness to take risks in fact generates a useful all-around measure. Asking questions
that include more specific contexts produce measures that are even stronger, for that
given context. On the other hand, focusing on a single context provides less predictive
power across contexts. In light of these findings, the usual practice of only eliciting risk
attitudes in the context of hypothetical financial lotteries would be expected to have
benefits for predicting financial decisions, but be a less effective approach for providing a
summary statistic of risk attitudes across other non-financial contexts. The SOEP does
include a hypothetical lottery question, which asks people how much of an endowment
of 100,000 Euros they would invest in an asset that doubles or returns only half of the
investment with equal probability. In unreported results, available upon request, we find
that answers to the hypothetical lottery are in fact strong predictors for decisions in
the financial domain, in the sense that they predict holding stocks. On the other hand,
responses to the lottery do not predict self-employment, or smoking, consistent with the
hypothesis that the more narrow context produces a measure that is less informative for
risk taking in non-financial domains. It is noteworthy that we also find similarly robust
results on the impact of gender, age, and height using the lottery measure, showing that
these findings also prevail in measures that involve financial lotteries and explicit stakes
and probabilities. Responses to the general risk question are highly correlated with choices
in the hypothetical lottery question, providing an additional indication that the general
risk question has explanatory power for choices in financial lotteries.40
Our validation of the general risk question provides a valuable instrument for future
research using survey data, where a simple and low cost measure of risk attitudes is very
often needed. For example, once one is confident that a measure captures risk attitudes
40In fact, if one assumes CRRA utility, responses to the general risk question can be mapped into CRRA
coefficients using a combination of individuals responses to the lottery question and individual wealth
information. Lottery responses and wealth information imply a distribution of CRRA coefficients
mainly between 1 and 10. The mapping to the general risk question indicates that for an individual with
zero wealth, choosing zero on the general risk scale is equivalent to a CRRA coefficient of approximately
2.9. For a given wealth level, one additional point on the general risk scale is associated with a decrease
in the CRRA coefficient of 0.381, while a larger wealth level implies a larger CRRA coefficient, holding
the degree of risk-taking constant.
in a behaviorally relevant way, it is possible to test theoretical predictions regarding the
relationship between risk attitudes and a given behavior, for example whether people who
are risk averse are more or less likely to be geographically mobile. Alternatively, it may
be important to control for risk attitudes in a regression, if risk attitudes determine some
type of selection process that confounds interpretation of a different variable of interest.
We have taken this next step in some of our own research. For example, Bonin et al.
(2007) use the general risk question, and show that individuals who are willing to take
risks sort into occupations with higher cross-sectional earnings risk. Jaeger et al. (2007)
show that willingness to take risks has a significant impact on geographic mobility. Two
other recent papers have also taken advantage of the experimentally-validated general
risk question. Grund and Sliwka (2006) find support for the theoretical prediction that
risk attitudes determine sorting into performance pay jobs, and Caliendo et al. (2009)
show that willingness to take risks increases the probability that an individual becomes
an entrepreneur in the future. We believe that these studies are the tip of the iceberg, in
terms of applications for validated survey measures of risk attitudes.
Barr, A., and T. Packard (2002): “Revealed Preference and Self-Insurance: Can We
Learn from the Self-Employed in Chile?,” World Bank Policy Research Working Paper
Barsky, R. B., T. F. Juster, M. S. Kimball, and M. D. Shapiro (1997): “Pref-
erence Parameters and Individual Heterogeneity: An Experimental Approach in the
Health and Retirement Study,” Quarterly Journal of Economics, 112(2), 537–579.
Bellemare, C., and S. Kr¨ oger (2007): “On Representative Social Capital,” European
Economic Review, 51(1), 181–202.
Binswanger, H. P. (1980): “Attitudes Towards Risk: Experimental Measurement in
Rural India,” American Journal of Agricultural Economics, 62(3), 395–407.
Blackburn, M., G. W. Harrison, and E. E. Rutstrom (1994): “Statistical Bias
Functions and Informative Hypothetical Surveys,” American Journal of Agricultural
Economics, 76(5), 1084–1088.
Blinder, A. (1973): “Wage Discrimination: Reduced Form and Structural Estimates,”
Journal of Human Resources, 8(3), 436–455.
Bonin, H., T. Dohmen, A. Falk, D. Huffman, and U. Sunde (2007): “Cross-
Sectional Earnings Risk and Occupational Sorting: The Role of Risk Attitudes,” Labour
Economics, 14(6), 926–937.
Bruhin, A., H. Fehr-Duda, and T. Epper (2007): “Risk and Rationality: Uncovering
Heterogeneity in Probability Distortion,” SOI Working Paper No. 0705.
Caliendo, M., F. Fossen, and A. Kritikos (2009): “Risk Attitudes of Nascent En-
trepreneurs: New Evidence from an Experimentally-Validated Survey,” Small Business
Camerer, C. F., and R. M. Hogarth (1999): “The Effects of Financial Incentives in
Experiments: A Review and Capital-Labor-Production Framework,” Journal of Risk
and Uncertainty, 19(1), 7–42.
Champ, P. A., and R. C. Bishop (2001): “Donation Payment Mechanisms and Con-
tingent Valuation: An Empirical Study of Hypothetical Bias,” Environmental and Re-
source Economics, 19, 383–402.
Choi, S., R. Fisman, D. Gale, and S. Kariv (2007): “Consistency and Heterogeneity
of Individual Behavior Under Uncertainty,” American Economic Review, 97(5), 1921–
Cramer, J. S., Joop Hartog and Nicole Jonker and C. Mirjam van Praag
(2002): “Low Risk Aversion Encourages the Choice for Entrepreneurship: an Empirical
Test of a Truism,” Journal of Economic Behavior and Organization, 48(1), 29–36.
Diaz-Serrano, L., and D. O’Neill (2004): “The Relationship Between Unemploy-
ment and Risk-Aversion,” IZA Discussion Paper No. 1214.
Dohmen, T., and A. Falk (2006): “Performance Pay and Multidimensional Sorting:
Productivity, Preferences and Gender,” IZA Discussion Paper No. 2001.
Dohmen, T., A. Falk, D. Huffman, U. Sunde, J. Schupp, and G. G. Wag-
ner (2005): “Individual Risk Attitudes: New Evidence from a Large, Representative,
Experimentally-Validated Survey,” IZA Discussion Paper No. 1730.
Donkers, B., B. Melenberg, and A. V. Soest (2001): “Estimating Risk Attitudes
Using Lotteries: A Large Sample Approach,” Journal of Risk and Uncertainty, 22(2),
Eckel, C., and P. Grossman (2007): “Men, Women and Risk Aversion: Experimental
Evidence,” in Handbook of Experimental Results, ed. by C. Plott, and V. Smith, North-
Holland. Elsevier Science.
Eckel, C., C. Johnson, and C. Montmarquette (2005): “Saving Decisions of the
Working Poor: Short- and Long-Term Horizons,” in Research in Experimental Eco-
nomics Volume 10: Field Experiments in Economics, ed. by J. Carpenter, G. Harrison,
and J. List, pp. 219–260, Oxford. Elsevier Science.
Feinberg, R. M. (1977): “Risk-aversion, Risk and the Duration of Unemployment,”
Review of Economics and Statistics, 59(3), 264–271.
Flynn, J., P. Slovic, and C. Mertz (1994): “Gender, Race, and Perception of
Environmental Health Risks,” Risk Analysis, 14, 1101–1108.
Grund, C., and D. Sliwka (2006): “Performance Pay and Risk Aversion,” IZA Dis-
cussion Paper No. 2012.
Guiso, L., T. Jappelli, and L. Pistaferri (2002): “An Empirical Analysis of Earn-
ings and Employment Risk,” Journal of Business and Economic Statistics, 20(2), 241–
Guiso, L., and M. Paiella (2001): “Risk-Aversion, Wealth, and Background Risk,”
CEPR Discussion Paper No. 2728.
(2005): “The Role of Risk Aversion in Predicting Individual Behavior,” Bank
of Italy Economic Working Paper No. 546.
Harrison, G. W., M. I. Lau, and E. E. Rutstr¨ om (2007): “Estimating Risk Atti-
tudes in Denmark: A Field Experiment,” Scandanavian Journal of Economics, 109(2),
Holt, C. A., and S. K. Laury (2002): “Risk Aversion and Incentive Effects,” American
Economic Review, 92(5), 1644–1655.
Jaeger, D., H. Bonin, T. Dohmen, A. Falk, D. Huffman, and U. Sunde (2009):
“Direct Evidence on Risk Attitudes and Migration,” Review of Economics and Statis-
Laury, S. L. (2006): “Pay One or Pay All,” EXCEN Working Paper No. 2006-24,
Georgia State University.
Loewenstein, G. F., C. K. Hsee, E. U. Weber, and N. Welch (2001): “Risk as
Feelings,” Psychological Bulletin, 127(2), 267–286.
Niederle, M., and L. Vesterlund (2007): “Do Women Shy Away From Competition?
Do Men Compete Too Much?,” Quarterly Journal of Economics, 122(3), 1067–1101.
Oaxaca, R. (1973): “Male-Female Wage Differentials in Urban Labor Markets,” Inter-
national Economic Review, 14(3), 693–709.
Persico, N., A. Postlewaite, and D. Silverman (2004): “The Effect of Adolescence
Experience on Labor Market Outcomes: The Case of Height,” Journal of Political
Economy, 112(5), 1019–1053.
Saris, W. E. (2003): “Multitrait-Multimethod Studies,” in Cross-Cultural Survey Meth-
ods, ed. by J. Harkness, pp. 265–274, New Jersey. Hoboken.
Saris, W. E., and A. van Meurs (1990): “Memory Effects in MTMM Studies,” in
Evaluation of Measurement Instruments by Meta-Analysis of Multitrait Multimethod
Studies, ed. by W. Saris, and A. van Meurs, pp. 134–146, Amsterdam. North-Holland.
Sch¨ afer, A., and J. Schupp (2006): “Zur Erfassung der Verm¨ ogensbest¨ ande im Sozio-
oekonomischen Panel (SOEP) im Jahr 2002,” DIW Data Documentation, 11.
Schubert, R., M. Brown, M. Gysler, and H. Brachinger (1999): “Financial
Decision-Making: Are Women Really More Risk-Averse?,” American Economic Review
Papers and Proceedings, 89(2), 381–385.
Schupp, J., and G. G. Wagner (2002): “Maintenance of and Innovation in Long-Term
Panel Studies The Case of the German Socio-Economic Panel (GSOEP),” Allgemeines
Statistisches Archiv, 86(2), 163–175.
(2007): “New Sources for Theory Based Socio-Economic Behavioral Analysis:
The 2002-2006 Pre-Tests for the German Socio-Economic Panel Study,” DIW Data
Silverman, J. M., and D. S. Kumka (1987): “Gender Differences in Attitudes Towards
Nuclear War and Disarmament,” Sex Roles, 16, 189–203.
Slovic, P. (1972a): “Psychological Study of Human Judgment: Implications for Invest-
ment Decision Making,” Journal of Finance, 27, 777–799.
(1972b): “Information Procession, Situation Specificity, and the Generality of
Risk-Taking Behavior,” Journal of Personality and Social Psychology, 22, 128–134.
(1999): “Trust, emotion, sex, politics, and science: Surveying the risk-assessment
battlefield,” Risk Analysis, 19(4), 689–701.
Spigner, C., W. Hawkins, and W. Lorens (1993): “Gender Differences in Perception
of Risk Associated with Alcohol and Drug Use Among College Students,” Women and
Health, 20, 87–97.
Stallen, P., and A. Thomas (1988): “Public Concern About Industrial Hazards,”
Risk Analysis, 8, 237–245.
Thompson, S. K. (2006): “Targeted Random Walk Designs,” Survey Methodology,
Wagner, G. G., R. V. Burkhauser, and F. Behringer (1993): “The English Lan-
guage Public Use File of the German Socio-Economic Panel,” The Journal of Human
Resources, 28(2), 429–433.
Weber, E. U., A. R. Blais, and N. E. Betz (2002): “A Domain-Specific Risk-
Attitude Scale: Measuring Risk Perceptions and Risk Behaviors,” Journal of Behavioral
Decision Making, 15, 263–290.
Table 1: Primary Determinants of General Risk Attitudes
Dependent variable: willingness to take risks in general
Age (in years)
Height (in cm)
Log(household wealth in 2002)
Log(household income 2003)
Log(household income 2004)
Interval regression coefficient estimates. The dependent variable is a measure for general risk attitudes
measured on a scale from 0 to 10, where 0 indicates “not at all willing to take risks” and 10 indicates
“very willing to take risks”. The Abitur exam is completed at the end of university-track high-schools in
Germany; passing the exam is a pre-requisite for attending university or a technical college (we include
Fachabitur, the exam for technical college, in the indicator for Abitur.) The other controls in column (9)
includes demographic, professional and other self-reported information. Wealth and income controls arein logs. Information about individual wealth is taken from the 2002 wave of the SOEP, which contains
detailed information on different assets and property values. Household wealth is constructed by summing
the wealth information of all individuals in the household. Logged absolute values of negative wealth are
added as a separate control variable in the respective specifications. For more detailed information on wealthand income variables see notes to Table A.1 in the Appendix. All specifications include a constant. Robust
standard errors in brackets allow for clustering at the household level; ***, **, * indicate significance at 1-,
5-, and 10-percent level, respectively.
Table 2: Validation of Survey Risk Measure in a Field Experiment
Subjects in experiment
Mean std. dev.
Mean std. dev.Median Median
Age (in years)
Height (in cm)
General risk attitude (survey response)4.762.545 4.422.385
Observations 45045045021,875 21,87521,875
(a) Comparison of SOEP and Experimental Sample
Dependent variable: value of safe option at switch point
Willingness to take risk in general0.611***
Controls for gender, age, height
(b) Predicting Lottery Choices with Survey Measure Risk Attitudes
Interval regression coefficient estimates. The dependent variable is the
value of the safe option at the switching point. Other controls include
controls for marital status, number of dependent children under 16, lived
in GDR in 1989, lived Abroad in 1989, location in 1989 missing, nation-
ality, student, educational achievement, dummies for occupational level
within public and private sector (as in Table A.1) in the Appendix),
health status, body weight, net household income, and life satisfaction,
compare also Table A.1. Robust standard errors in brackets; ***, **, *
indicate significance at 1-, 5-, and 10-percent level, respectively.
Table 3: Primary Determinants of Risk Attitudes in Different Domains of Life
Dependent variable: willingness to take risks in:
Age (in years)
Height (in cm)
Interval regression coefficient estimates. The dependent variable is a measure for risk attitudes in
the respective domain measured on a scale from 0 to 10, where 0 indicates “not at all willing to take risks” and 10 indicates “very willing to take risks”. The Abitur exam is completed at the
end of university-track high-schools in Germany; passing the exam is a pre-requisite for attending
university or a technical college (we include Fachabitur, the exam for technical college, in the
indicator for Abitur.) All specifications include a constant. Robust standard errors in brackets
allow for clustering at the household level; ***, **, * indicate significance at 1-, 5-, and 10-percent
Table 4: Correlations Between Risk Attitudes in Different Contexts
Correlations are based on individuals’ risk attitudes in each context, reported on an 11-point scale.Choosing 0 indicates “not at all willing take risks” and choosing 10 indicates “very willing to take
risks”. Correlations based on 19,043 observations with responses to all risk questions.
Table 5: The Relative Ability of Alternative Measures of Risk Attitudes to Explain Risky
Active sportsSelf-employed Smoking
Willingness to take risks in
Willingness to take risks in
Car driving 0.024***
Sports and leisure
(dependent variable =1) 0.3410.662 0.0840.294
Dependent variables in all columns are binary variables. Willingness to take risks is measured on on a scale
from 0 to 10, where 0 indicates “not at all willing to take risks” and 10 indicates “very willing to take risks”.
All risk measures are standardized. Reported coefficients are Probit marginal effects estimates, evaluated
at the means of independent variables. Each coefficient estimate is based on a separate regression of the
respective dependent variable on this particular risk measure and a set of controls, whose coefficient estimates
are not reported. This set includes the same controls for gender, age, height, and parental education as in
Table 1, as well as controls for log household wealth, log household debt, and the log of current gross
monthly household income in every regression. Additional controls:
Sample excludes individuals that are older than 66 years, or retired, or non-participating in the labor market.
All specifications include a constant. Robust standard errors that allow for clustering at the household level
are reported in brackets below the coefficient estimates. Values of the log pseudo-likelihood of the respective
regression model are reported in parentheses; ***, **, * indicate significance at 1-, 5-, and 10-percent level,
†Number of Adults in household.
Figure 1: Willingness to Take Risks in General
Responses to General Risk Question
(0=not at all willing; 10=very willing)
All Respondents − SOEP 2004
Notes: The Figure shows a histogram of responses to the question about willingness
to take risks “in general”, measured on an eleven-point scale.
Table A.1: Determinants of Risk Attitudes
Willingness to take Risks in the Domain of:
Age (in years)
Height (in cm)
1 Child born after 1987
2 Children born after 1987
3 Children born after 1987
> 3 Children born after 1987
Lived in GDR in 1989
Lived abroad in 1989
Residence in 1989 missing
Lives in East Germany in 2004
Subjective Health Status
Weight (in kg)
Enrolled in School
Enrolled in Continuing Professional
Enrolled in College/University
Table A.1: continued: Determinants of Risk Attitudes
Willingness to take Risks in the Domain of:
General Sports &
Unskilled Blue Collar -0.757**
Skilled Blue Collar
Blue Collar Craftsman
Blue Collar Foreman
Blue Collar Master
Unskilled White Collar
Skilled White Collar
White Collar Technician
White Collar Master
Highly-Skilled White Collar
White Collar Management
Civil Servant Intermediate
Civil Servant High
Civil Servant Executive
Unskilled Blue Collar-0.348*
Skilled Blue Collar
Blue Collar Craftsman
Blue Collar Foreman
Blue Collar Master
Unskilled White Collar
White Collar Technician
Highly-Skilled White Collar
White Collar Master
White Collar Management
Table A.1: continued: Determinants of Risk Attitudes Download full-text
Willingness to take Risks in the Domain of:
Helping Family Member -0.018
No / Missing Occupation
Log(Household Wealth in 2002)
Log(Household Debt in 2002)
Log(Household Income 2004)
Month of Interview:
Pseudo Log Likelihood
Interval regression coefficient estimates.
respective domain measured on a scale from 0 to 10, where 0 indicates “not at all willing to take risks”
and 10 indicates “very willing to take risks”. The Abitur exam is completed at the end of university-track
high-schools in Germany; passing the exam is a pre-requisite for attending university or a technical college
(we include Fachabitur, the exam for technical college, in the indicator for Abitur.) Information on religion
is taken from the 2003 wave. The omitted category is Protestant. The category “No religion” includes
those who are not officially affiliated with any type of church. Wealth and income controls are in logs.
Information about individual wealth is taken from the 2002 wave of the SOEP, which contains detailed
information on different assets and property values, see Sch¨ afer and Schupp (2006) for details. Household
wealth is constructed by summing the wealth information of all individuals in the household, and is treated
as missing if the wealth information is missing for at least one member of the household. An exception is
the case where the individual with missing wealth participated in the survey for the first time in 2004, and is
younger than 20 years of age; in this case the missing value for the (teenage) individual’s wealth is assumed
to indicate zero wealth. Results are qualitatively unchanged if we do not make this assumption. Logged
absolute values of negative wealth are added as a separate control variable in the respective specifications.
Wealth and income controls are in logs. Logged absolute values of negative wealth are added as separate
control. The income data for 2004 are based on answers to questions about current gross monthly income
sources at the time of the interview. We also used the net monthly income measure that is available as a
generated variable in the SOEP; the results (not reported here) are essentially the same. Robust standard
errors in brackets allow for clustering at the household level; ***, **, * indicate significance at 1-, 5-, and
10-percent level, respectively.
The dependent variable is a measure for risk attitudes in the