The Ancient Origins of Cross-Country
Heterogeneity in Risk Preferences∗
Anke Becker
Benjamin Enke
Armin Falk
May 6, 2015
Abstract
Using novel representative preference data on 80,000 individuals from 76
countries, this paper shows that the migratory movements of our early ancestors thousands of years ago have left a footprint in the contemporary crosscountry distribution of risk preferences. Across a wide range of regression
specifications, differences in risk aversion between populations are significantly
increasing in the length of time elapsed since the respective groups shared common ancestors. This result obtains for various proxies for the structure and
timing of historical population breakups, including genetic and linguistic data
or predicted measures of migratory distance. In addition, we provide evidence
that the within-country heterogeneity in risk aversion significantly decreases in
migratory distance from the geographical origin of mankind. Taken together,
our results point to the importance of very long-run events for understanding
the global distribution of one of the key economic traits.
JEL classification: D01, D03
Keywords: Risk preferences, out of Africa, persistence, genetic distance, genetic
diversity
∗
We are grateful to Quamrul Ashraf, Klaus Desmet, Nathan Nunn, Uwe Sunde, Hans-Joachim
Voth, and Romain Wacziarg for very helpful discussions. Valuable comments were also received
from seminar audiences at Bonn, Munich, the NBER Economics of Culture and Institutions Meeting
2014, SITE Stanford 2014, the North-American ESA conference 2014, and the Annual Congress of
the EEA 2013. We thank Quamrul Ashraf, Oded Galor, and Ömer Özak for generous data sharing.
Ammar Mahran provided oustanding research assistance. Armin Falk acknowledges financial support from the European Research Council through ERC # 209214. Becker, Enke, Falk: University
of Bonn, Department of Economics, Adenauerallee 24-42, 53113 Bonn, Germany. [email protected], [email protected], [email protected]
1
Introduction
Preferences over risk have been shown to vary substantially within populations and
to predict a plethora of individual-level economic decisions. However, do risk preferences also vary across countries? And if so, what explains this variation, i.e., what
are the ultimate determinants of heterogeneity in preferences? Using a novel, globally representative data set on risk preferences, this paper seeks to provide an answer
to these questions. Our key contribution is to show that the structure of mankind’s
ancient migration out of Africa and around the globe has had a persistent impact on
the between-country distribution of risk attitudes as of today. In light of the relevance of risk preferences in predicting important economic and social outcomes, these
results contribute to understanding the ultimate sources of cross-country economic
heterogeneity. Moreover, our findings add to the emerging literature on endogenous
preferences (Fehr and Hoff, 2011) in highlighting historical roots as a key driver in
shaping preferences.
According to the widely accepted “Out of Africa hypothesis”, starting around
50,000 years ago, early mankind migrated out of East Africa and continued to explore and populate our planet through a series of successive migratory steps. Each
of these steps consisted of some sub-population breaking apart from the previous
colony and moving on to found new settlements. This pattern implies that some
contemporary population pairs have spent a longer time of human history apart
from each other than others. As a result, the time elapsed since two groups shared
common ancestors differs across today’s population pairs. The key idea underlying
our analysis is that these differential time frames of separation have affected the
cross-country distribution of risk preferences. First, populations that have spent a
long time apart from each other were exposed to differential historical experiences,
which could affect preferences over risk (Callen et al., 2014). Second, due to, e.g.,
random genetic drift, long periods of separation lead to different population-level
genetic endowments, which might in turn shape risk attitudes.1 Both of these channels imply that populations that have been separated for a long time in the course
of human history, should also exhibit different risk attitudes. Thus, we hypothesize
that the (absolute) difference in risk attitudes between countries reflects the length
of separation of the respective populations, as evoked by the successive population
breakups in ancient times.
To investigate this hypothesis, we use a novel dataset on risk preferences across
countries in combination with proxies for long-run human migration. As part of the
Global Preference Survey (GPS), we collected two survey measures of risk preferences
1
Cesarini et al. (2009) use a twin study to provide evidence for a genetic effect on risk preferences.
1
(see Falk et al., 2015a). The sample of 80,000 people from 76 countries is constructed
to provide representative population samples within each country and geographical
representativeness in terms of countries covered. The survey items – one qualitative
and one quantitative lottery-type measure of risk preferences – were selected and
tested through a rigorous ex ante experimental validation procedure involving real
monetary stakes. The elicitation followed a standardized protocol that was implemented through the professional infrastructure of the Gallup World Poll. These data
allow the computation of a nationally representative level of risk aversion, and hence
the derivation of the absolute difference in risk attitudes within a country pair.2
We combine these data with three classes of proxies for the temporal patterns of
ancient population fissions, i.e., proxies for the length of time since two populations
shared common ancestors. (i) First, we employ the FST and Nei genetic distances
between populations, as originally measured by the population geneticists CavalliSforza et al. (1994) and introduced into the economics literature by Spolaore and
Wacziarg (2009). As population geneticists have long noted, whenever two populations split apart from each other in order to found separate settlements, their
genetic distance increases over time due to random genetic drift. Thus, the genetic
distance between two populations is a measure of temporal distance since separation.
(ii) Second, we use measures of predicted migratory distance between contemporary
populations, which were constructed by Ashraf and Galor (2013b) and Özak (2010),
respectively. The predicted migratory distance variable of Ashraf and Galor (2013b)
is based on a procedure which exploits information on the geographic patterns of
early migratory movements and constitutes a proxy for the predicted length of separation of two populations.3 The human-mobility-index measure of Özak (2010), on
the other hand, explicitly computes the walking time between two countries’ capitals, taking into account topographic, climatic, and terrain conditions, as well as
human biological abilities. (iii) Finally, we make use of the observation that linguistic trees closely follow the structure of separation of human populations and employ
a measure of linguistic distance between two populations as explanatory variable.
In sum, we use various independent sources of data which are known to reflect the
length of separation of populations. Using these measures, we test the hypothesis
that the heterogeneity in risk preferences across contemporary populations is driven
by migratory movements in the distant past.
Our empirical analysis of the relationship between risk preferences and ancient
2
Vieider et al. (2012) and Rieger et al. (2014) also collect data on the prevalence of risk attitudes
in multiple countries. Their samples cover a much smaller set of countries and contain university
students, rendering conclusions about population-level differences in risk attitudes difficult.
3
See Ramachandran et al. (2005).
2
migration patterns starts by establishing that the absolute difference in average risk
attitudes between two countries is significantly increasing in the length of time since
today’s populations shared common ancestors, as proxied for by (observed) genetic
distance, predicted migratory distance, and linguistic distance. In a second step,
using observed genetic distance as main measure, we investigate to what extent
our main result is likely to be driven by contemporary environmental conditions, i.e.,
omitted variables. We establish that the relationship between differences in risk preferences and length of separation is robust to an extensive set of covariates, including
controls for differences in the countries’ demographic composition, their geographic
position, prevailing climatic and agricultural conditions, institutions, and economic
development. In all of the corresponding regressions, the point estimate is rather
stable, suggesting that unobserved heterogeneity is unlikely to drive our results, either (Altonji et al., 2005). Our result also holds when we exclude entire continents or
observations with large genetic distances from the analysis. Final robustness checks
reveal that we obtain similar results when we use alternative genetic distance data
or the predicted migratory distance variables in the conditional regressions.
In sum, irrespective of the measure used, our results indicate the existence of a
relationship between cross-country differences in risk attitudes and the pattern of
very distant migratory movements, which is not explained by differences in the contemporary environments people reside in. However, as established by the so-called
serial-founder effect of population genetics, the structure of successive population
breakups also affects within-population diversity. Intuitively, due to the loss of heterogeneity that is implied by small founder populations breaking apart from previous
parental colonies, the diversity of a population decreases along migratory paths. To
investigate whether such a mechanism also plays out for preferences over risk, we
relate the within-country standard deviation in risk preferences to a population’s
(predicted) migratory distance from East Africa. The results show that, for both
predicted migratory distance variables, the degree of the within-country heterogeneity in individual risk attitudes is indeed significantly decreasing in migratory distance from East Africa, and is hence positively correlated with the diversity of the
respective genetic pool. Thus, along two major dimensions (means and variances),
temporally very distant population fissions explain a significant fraction of today’s
cross-country heterogeneity in one of the key economic traits.
A number of recent contributions argue that the (cultural) diversity caused by
long-run migration thousands of years ago can have real aggregate economic effects.
Spolaore and Wacziarg (2009, 2011) find a strong relationship between genetic distance and income differences across countries and argue that lower cultural distance
3
between two countries facilitates the diffusion of knowledge.4 Ashraf and Galor
(2013b) and Ashraf et al. (2014) document an inverted U-form relationship between
per capita income and genetic diversity (migratory distance from East Africa). They
argue that this pattern reflects the trade-off between more innovation and lower trust
and cooperation that is associated with higher cultural diversity.5 Our paper nicely
dovetails with this set of contributions as it shows that the genetic variables which
proxy for migratory flows indeed capture variation in an economically important
trait (either between or within countries), as is implicitly or explicitly assumed by
these papers.
This paper also forms part of an active recent literature on the historical, biological, and cultural origins of beliefs and preferences. For example, Chen (2013) and
Galor and Özak (2014) show that cross-country differences in future-orientation are
affected by a structural feature of languages and historical agricultural productivity.
Tabellini (2008) and an earlier version of Guiso et al. (2009) relate interpersonal trust
to linguistic features and the genetic distance between two populations. Nunn and
Wantchekon (2011), Voigtländer and Voth (2012) and Alesina et al. (2013) establish
the deep roots of trust, beliefs over the appropriate role of women in society, and
anti-semitism.6 Desmet et al. (2011) show that, within Europe, genetic distance correlates with opinions and attitudes as expressed in the World Values Survey. Desmet
et al. (2014) investigate the relationship between cultural traits and ethnicity. However, presumably given the previous lack of data, this paper is the first contribution
to study the origins of cross-country variation in risk preferences.
The remainder of this paper proceeds as follows. In the next section, we develop
our hypotheses on the relationship between the structure of migratory movements
and risk preferences, while Section 3 presents the data. Section 4 discusses our main
result on the connection between differences in average risk attitudes and length of
separation. In Section 5, we examine the within-country heterogeneity in risk preferences. Section 6 concludes the paper by discussing correlations between country-level
risk attitudes and economic, institutional, and health outcomes.
4
Spolaore and Wacziarg (2014) find a negative relationship between genetic distance and the
occurrence of conflict. Gorodnichenko and Roland (2010) argue that genetic distance to the United
States can be used as an instrument for the cultural “individualism” dimension to show a causal
effect on GDP. Proto and Oswald (2014) investigate the correlation between genetic distance to
Denmark and happiness. Giuliano et al. (2014) show that the negative correlation between genetic
distance and bilateral trade vanishes once transportation costs are accounted for.
5
Arbatli et al. (2013) and Ashraf and Galor (2013a) show that genetic diversity is related to
ethnolinguistic diversity and the emergence of civil conflicts.
6
See also Guiso et al. (2006) and Fernández and Fogli (2006, 2009).
4
2
Risk Attitudes and the Great Human Expansion
According to the widely accepted theory of the origins and the dispersal of early
humans, the single cradle of mankind is to be found in East or South Africa and
can be dated back to roughly 100’000 years ago (see, e.g., Henn et al. (2012) for an
overview). Starting from East Africa, a small sample of hunters and gatherers exited
the African continent around 50’000-60’000 years ago and thereby started what is
now also referred to as the “great human expansion”. This expansion continued
throughout Europe, Asia, Oceania, and the Americas, so that mankind eventually
came to settle on all continents. A noteworthy feature of this very long-run process
is that it occurred through a large number of discrete steps, each of which consisted
of a sub-sample of the original population breaking apart and leaving the previous
location to move on and found new settlements elsewhere (so-called serial founder
effect). The main argument of this paper is that the pattern of successive breakups
affected the distribution of risk preferences we observe around the globe today.
First, the series of migratory steps implied a frequent breakup of formerly united
populations. After splitting apart, these sub-populations often settled geographically
distant from each other, i.e., lived in separation. There are two channels through
which the length of separation of two groups might have had an impact on betweengroup differences in risk attitudes:
First, the fact that two populations have spent a long time apart from each other
implies that they were subject to many differential historical experiences. Recent
work by, e.g., Callen et al. (2014) highlights that risk preferences are malleable by
idiosyncratic experiences or, more generally, by the composition of people’s social
environment. Thus, the differential historical experiences which have accumulated
over thousands of years of separation might have given rise to differential attitudes
towards risk as of today.
Second, whenever two populations spend time apart from each other, they develop
different population-level genetic pools due to, e.g., random genetic drift. Given that
risk attitudes are transmitted across generations (Dohmen et al., 2012) and that part
of this transmission is genetic in nature (Cesarini et al., 2009), the different genetic
endowments induced by long periods of separation could also generate differential
risk attitudes.
Note that both of these channels imply a bilateral (rather than a directional)
statement about between-country differences in risk preferences:
5
Main Hypothesis. The absolute difference in (average) risk aversion between two
countries is increasing in the length of separation of the respective populations in the
course of human history.
The serial founder pattern of ancient migration movements could also have had an
effect on the within-population heterogeneity in risk preferences. A well-established
fact in the population genetics literature is that whenever a sub-population split apart
from its parental colony, those humans breaking new ground took with them only
a fraction of the genetic diversity of the previous genetic pool (intuitively because
they were usually small, and hence non-representative, samples). In consequence,
through the sequence of successive fissions, the total diversity of the gene pool significantly decreases along human migratory routes out of East Africa. Note that the
logic behind this serial founder effect need not be restricted to genetic variation, but
could similarly apply to heterogeneity in risk attitudes:
Ancillary Hypothesis. The within-country heterogeneity in risk preferences is
negatively related to a population’s migratory distance from East Africa.
3
3.1
Data
Representative Cross-Country Data on Risk Preferences
Our data on risk preferences around the globe are part of the Global Preference
Survey (GPS), which constitutes a unique dataset on economic preferences from
representative population samples around the globe. In many countries around the
world, the Gallup World Poll regularly surveys representative population samples
about social and economic issues. In 76 countries, we included as part of the regular
2012 questionnaire a set of survey items which were explicitly designed to measure
a respondent’s risk preferences (for details see Falk et al., 2015a).
Four noteworthy features characterize these data. First, the preference measures have been elicited in a comparable way using a standardized protocol across
countries. Second, contrary to small- or medium-scale experimental work, we use
preference measures that have been elicited from representative population samples
in each country. This allows for inference on between-country differences in preferences, in contrast to existing cross-country comparisons of convenience (student)
samples. The median sample size was 1,000 participants per country; in total, we
collected preference measures for more than 80,000 participants worldwide. Respondents were selected through probability sampling and interviewed face-to-face or via
telephone by professional interviewers. Third, the dataset also reflects geographical
6
representativeness. The sample of 76 countries is not restricted to Western industrialized nations, but covers all continents and various development levels. Specifically,
our sample includes 15 countries from the Americas, 24 from Europe, 22 from Asia
and Pacific, as well as 14 nations in Africa, 11 of which are Sub-Saharan. The set
of countries contained in the data covers about 90% of both the world population
and global income. Fourth, the preference measures are based on experimentally
validated survey items for eliciting preferences. In order to ensure behavioral relevance, the underlying survey items were designed, tested, and selected through an
explicit ex-ante experimental validation procedure (Falk et al., 2015b). In this validation step, out of a large set of preference-related survey questions, those items
were selected which jointly perform best in explaining observed behavior in standard
financially incentivized experimental tasks to elicit preference parameters. In order
to make these items cross-culturally applicable, (i) all items were translated back and
forth by professionals, (ii) monetary values used in the survey were adjusted along
the median household income for each country, and (iii) pretests were conducted in
21 countries of various cultural heritage to ensure comparability. See Appendix A
and Falk et al. (2015a) for details.
The set of survey items included two measures of the underlying risk preference – one qualitative subjective self-assessment and one quantitative measure. The
subjective self-assessment directly asks for an individual’s willingness to take risks:
“Generally speaking, are you a person who is willing to take risks, or are you not
willing to do so? Please indicate your answer on a scale from 0 to 10, where a 0
means “not willing to take risks at all” and a 10 means “very willing to take risks”.
You can also use the values in between to indicate where you fall on the scale.”
The quantitative measure is derived from a series of five interdependent hypothetical binary lottery choices, a format commonly referred to as the “staircase procedure”. In each of the five questions, participants had to decide between a 50-50
lottery to win x euros or nothing (which was the same in each question) and varying
safe payments y. The questions were interdependent in the sense that the choice
of a lottery resulted in an increase of the safe amount being offered in the next
question, and conversely. For instance, in Germany, the fixed upside of the lottery
x was 300 euros, and in the first question, the fixed payment was 160 euros. In
case the respondent chose the lottery (the safe payment), the safe payment increased
(decreased) to 240 (80) euros in the second question (see Figure 3 in Appendix A
for an exposition of the entire sequence of survey items). In essence, by adjusting
the fixed payment according to previous choices, the questions “zoom in” around the
respondent’s certainly equivalent and make efficient use of limited and costly survey
7
time. This procedure yields one of 32 ordered outcomes.
The subjective self-assessment and the outcome of the quantitative lottery staircase were aggregated into a single index which describes an individual’s degree of risk
aversion.7 Despite the lack of financial incentives, there are good reasons to expect
that our measures capture respondents’ risk attitudes. Apart from the experimental
validation procedure, the qualitative subjective self-assessment has previously been
shown to be predictive of both experimental behavior and risk-taking in the field
in a representative sample (Dohmen et al., 2011) as well as of incentivized experimental risk-taking across countries in student samples (Vieider et al., forthcoming).8
Additionally, the quantitative lottery staircase measure is akin to standard experimental lottery-choices and lacks any context, so that it is arguably less prone to
culture-dependent interpretations.
Figure 1 depicts the distribution of the average degree of risk aversion around the
world. Darker colors indicate a higher propensity to take risks. The map indicates
a large heterogeneity across countries, with Portugal being the most risk averse and
South Africa the least risk averse country in the sample. Country averages vary
by almost two standard deviations, relative to the total individual-level standard
deviation of one.9
3.2
Proxies for Ancient Migration Patterns
We use various separate but conceptually linked classes of variables to proxy for the
length of time since two populations split apart: (i) Observed genetic distance, (ii)
predicted pairwise migratory distance, and (iii) linguistic distance.
Observed Genetic Distance Between Countries
First, whenever populations break apart, they stop interbreeding, thereby preventing a mixture of the respective genetic pools. However, since every genetic pool
is subject to random drift (“noise”), geographical separation implies that over time
the genetic distance between sub-populations gradually became (on average) larger.
Thus, the genealogical relatedness between two populations reflects the length of time
elapsed since these populations shared common ancestors. In fact, akin to a molec7
Aggregation was done by first computing the z-scores of each survey item at the individual
level and then linearly combining these two z-scores using the weights obtained in the experimental
validation procedure, see Falk et al. (2015b). The weights are given by:
Risk aversion = −(0.4729985 × Staircase outcome + 0.5270015 × Qualitative item)
8
For instance, this item correlates with stock market participation, self-employment, risky sports
practices, and smoking, see Dohmen et al. (2011). In addition, this question was used to establish
the intergenerational transmission of risk attitudes (Dohmen et al., 2012).
9
When we compute all possible t-tests in our sample of 2,850 country pairs, 78% of all country
pairs are statistically significantly different from each other at the 5% level.
8
Figure 1: Cross-country heterogeneity in risk aversion. Darker blue (red) indicates higher (lower)
risk aversion. All values expressed in standard deviations around the world mean individual (white).
ular clock, population geneticists have made use of this observation by constructing
mathematical models to compute the timing of separation between groups. This
makes clear that, at its very core, genetic distance constitutes not only a measure of
genealogical relatedness, but also of temporal distance between two populations.
Technically, genetic distance constitutes an index of expected heterozygosity,
which can be thought of as the probability that two randomly matched individuals will be genetically different from each other in terms of a pre-defined spectrum of
genes. Indices of heterozygosity are derived using data on allelic frequencies, where
an allele is a particular variant taken by a gene.10 Intuitively, the relative frequency
of alleles at a given locus can be compared across populations and the deviation in
frequencies can then be averaged over loci. This is exactly the approach pursued in
the work of the population geneticists Cavalli-Sforza et al. (1994). The main dataset
assembled by these researchers consists of data on 128 different alleles for 42 world
populations. By aggregating differences in these allelic frequencies, the authors compute the FST genetic distance, which provides a comprehensive measure of genetic
relatedness between any pair of 42 world populations. In addition, using the same
dataset, Cavalli-Sforza et al. (1994) compute the so-called Nei distance for all population pairs. While this genetic distance measure has slightly different theoretical
properties than FST , the two measures are highly correlated (ρ = 0.95). Since ge10
Such genetic measures are based on neutral genetic markers only, i.e., on genes which are not
subject to selection pressure. Thus, differences in such genes merely reflect random drift and do
not arise from evolutionary fit.
9
netic distances are available only at the population rather than at the country level,
Spolaore and Wacziarg (2009) matched the 42 populations in Cavalli-Sforza et al.
(1994) to countries using ethnic composition data from Fearon (2003). Thus, the genetic distance measures we use measure the expected genetic distance between two
randomly drawn individuals, one from each country, according to the contemporary
composition of the population.
Predicted Migratory Distance Between Countries
Rather than physically measure the genetic composition of populations to investigate their kinship, one can also derive predicted migration measures (Ashraf and
Galor, 2013b; Özak, 2010). Key idea behind using both of these variables is that
populations that have lived far apart from each other (in terms of migratory, not necessarily geographic, distance), must also have spent a large portion of human history
apart from each other. Notably, these data are independent of those on observed
genetic distance and thus allow for an important out-of-sample robustness check.
First, the derivation of the predicted migratory distance variable of Ashraf and
Galor (2013b) follows the methodology proposed in Ramachandran et al. (2005) by
making use of today’s knowledge of the migration patterns of our ancestors. Specifically, Ashraf and Galor (2013b) obtain an estimate of bilateral migratory distance
by computing the shortest path between two countries’ capitals. Given that until
recently humans are not believed to have crossed large bodies of water, these hypothetical population movements are restricted to landmass as much as possible by
requiring migrations to occur along five obligatory waypoints, one for each continent. By construction, these migratory distance estimates only pertain to the native
populations of a given pair of countries. Thus, to the extent that the contemporary
populations in a country pair differ from the native ones, these distance estimates
need to be adjusted for post-Columbian migration flows. In order to derive values of
predicted migratory distance pertaining to the contemporary populations, we combine the dataset of Ashraf and Galor (2013b) with the “World Migration Matrix” of
Putterman and Weil (2010), which describes the share of the year 2000 population
in every country that has descended from people in different source countries as of
the year 1500. Thus, the contemporary predicted migratory distance between two
countries equals the weighted migratory distance between the contemporary populations.11
Thus, this ancestry-adjusted predicted migratory distance between two countries
11
Formally, suppose there are N countries, each of which has one native population. Let s1,i
be the share of the population in country 1 which is native to country i and denote by di,j the
migratory distance between the native populations of countries i and j. Then, the (weighted)
predicted ancestry-adjusted migratory distance between countries 1 and 2 as of today is given by
10
can be thought of as the expected migratory distance between the ancestors of two
randomly drawn individuals, one from each country. Further note that migratory
distance and observed genetic distance tend to be highly correlated (Ramachandran
et al., 2005). Ashraf and Galor (2013b) exploit this fact by linearly transforming
migratory distance into a measure of predicted FST genetic distance. Thus, our
measure of predicted migratory distance might as well be interpreted as predicted
genetic distance. Indeed, the correlation of our predicted (ancestry-adjusted) migratory distance measure with observed FST is ρ = 0.54.
Second, as an additional independent measure of migratory distance, we use
the so-called “human mobility index”-based migratory distance developed by Özak
(2010). This measure is more sophisticated than the raw migratory distance using
the five intermediate waypoints in that it measures the walking time along the optimal route between any two locations, taking into account the effects of temperature,
relative humidity, and ruggedness, as well as human biological capabilities. Given
that the procedure assumes travel by foot (as is appropriate if interest lies in migratory movements thousands of years ago), the data do not include islands, but
assume that the Old World and the New World are connected through the Bering
Strait, over which humans are believed to have entered the Americas. The original
data contain the travel time between two countries’ capitals, which we again adjust
for post-Columbian migration flows using the ancestry-adjustment methodology outlined above. Thus, our final variable measures the expected travel time between the
ancestors of two randomly drawn individuals, one from each country. As we will see
below, adjusting both migratory distance variables for post-Columbian migration has
a substantial positive effect on their explanatory power for cross-country differences
in risk attitudes.
Linguistic Distance Between Countries
Population geneticists and linguists have long noted the close correspondence
between genetic distance and linguistic “trees”, intuitively because population breakups do not only produce diverging gene pools, but also differential languages. Hence,
we employ the degree to which two countries’ languages differ from each other as an
additional proxy for the timing of separation. The construction of linguistic distances
follows the methodology proposed by Fearon (2003). The Ethnologue project classifies all languages of the world into language families, sub-families, sub-sub-families
etc., which give rise to a language tree. In such a tree, the degree of relatedness
Predicted migratory distance1,2 =
N X
N
X
i=1 j=1
11
(s1,i × s2,j × di,j )
between different languages can be quantified as the number of common nodes two
languages share.12 As in the case of predicted migratory distances, for each country
pair, we calculate the weighted linguistic distance according to the population shares
speaking a particular language in the respective countries today.
Note that it is well-know in the population genetics and linguistics literatures
that genetic distance appears to be a higher-quality measure of separation patterns.
First, while languages generally maintain a certain structure over long periods of
time, in some cases they change or evolve very quickly. Thus, in general, the slowmoving nature of aggregate genetic endowment makes genetic distance a more robust
measure of ancient breakups of populations, also see the discussion in Cavalli-Sforza
(1997). In addition, any quantitative measure of linguistic distance suffers from
the fact that there is no natural metric on languages. While language trees are a
useful tool to circumvent this problem, they remain coarse in nature, potentially
introducing severe measurement error.13 We hence expect that genetic distance will
be a more powerful explanatory variable than linguistic distance. Likewise, genetic
distance appears to be a more appropriate measure of length of separation than
the predicted migratory distance variables, which are constructed based on entirely
theoretical procedures.
4
Risk Preferences and Temporal Distance
4.1
Baseline Results
This section develops our main result on the relationship between differences in
the average degree of risk aversion between countries and the temporal distance
between the respective populations. Since temporal distance is an inherently bilateral
variable, this analysis will necessitate the use of a dyadic regression framework, which
takes each possible pair of countries as unit of observation. Accordingly, we match
each of the 73 countries with every other country into a total of 2,628 country pairs
and relate our proxies for temporal distance to the (absolute) difference in average
12
If two languages belong to different language families, the number of common nodes is 0. In
contrast, if two languages are identical, the number of common nodes is 15. Following Fearon
(2003), who argues that the marginal increase in the degree of linguistic relatedness is decreasing
in the number of common nodes, we transformed these data according to
r
# Common nodes
Linguistic distance = 1 −
15
to produce distance estimates between languages in the interval [0, 1]. We restricted the Ethnologue
data to languages which make up at least 5% of the population in a given country.
13
Also see the corresponding discussion in Mecham et al. (2006).
12
risk aversion between the respective populations.14 Our regression equation is hence
given by:
|riski − riskj | = α + β × temporal distance proxyi,j + i,j
where riski and riskj represent the average risk aversion in countries i and j, respectively, while i,j is a country pair specific disturbance term. Regarding the latter,
notice that our empirical approach implies that each country will appear multiple
times as part of the (in)dependent variable. Thus, to allow for clustering of the error
terms at the country-level, we employ the two-way clustering strategy of Cameron
et al. (2011), i.e., we cluster at the level of the first and of the second country of a
given pair. This procedure allows for arbitrary correlations of the error terms within
a group, i.e., within the group of country pairs which share the same first country or
which share the same second country, respectively, see the discussion in Appendix II
of Spolaore and Wacziarg (2009).
Columns (1) through (4) of Table 1 provide the results of unconditional OLS regressions of absolute differences in risk aversion on our proxies for temporal distance.
Column (1) shows that the FST genetic distance is a strong predictor of differences
in average risk attitudes. In quantitative terms, the standardized beta is large and
indicates that a one standard deviation increase in genetic distance is associated
with an increase of roughly one-third of a standard deviation in differences in risk
attitudes.15
Column (2) of Table 1 introduces the Nei genetic distance. The coefficient on
this distance measure is very similar to the one on FST and highly statistically
significant. We will continue to use this measure to exemplify the robustness of our
results below. In columns (3) and (4), we establish that both predicted migratory
distance variables are also significantly related to differences in risk attitudes, even
though the coarser measure of predicted migratory distance is only weakly significant.
Column (5) relates the absolute difference in average risk aversion to the linguistic
distance between the respective populations. Again, this proxy for ancient migration
patterns exhibits an unconditional relationship with differences in risk preferences.
Note that the genetic variables explain a considerably larger fraction of the variation
in risk preferences than the predicted migration variables or linguistic distance. This
observation is consistent with the view discussed above that genetic distance is a
14
Due to a lack of data on genetic distance and / or genetic diversity, we excluded Bosnia and
Herzegovina, Serbia, and Suriname from the sample.
15
If the standardized beta equals x, then a one standard deviation increase in the independent
variable is associated with an increase of x% of a standard deviation in the dependent variable.
Interestingly, in this baseline regression, both the standardized beta and the explained variance are
close to the respective values in Spolaore and Wacziarg’s (2009) analysis of the relationship between
differences in national income and genetic distance.
13
Table 1: Risk preferences and temporal distance
Dependent variable:
Absolute difference in average risk aversion
(1)
(2)
(3)
(4)
(5)
Fst genetic distance
0.37∗∗∗
(0.09)
0.40∗∗∗
(0.10)
Nei genetic distance
0.71∗
(0.37)
Predicted migratory distance
0.68∗∗∗
(0.23)
HMI migratory distance
(Özak, 2010)
0.20∗∗∗
(0.07)
Linguistic distance
Constant
Observations
R2
Standardized beta (%)
0.21∗∗∗
(0.03)
0.22∗∗∗
(0.03)
0.28∗∗∗
(0.03)
0.23∗∗∗
(0.04)
0.17∗∗
(0.07)
2628
0.108
32.9
2628
0.117
34.2
2628
0.013
11.5
2080
0.062
24.9
2628
0.021
14.5
OLS estimates, twoway-clustered standard errors in parentheses. The standardized betas refer to the temporal distance proxies. ∗ p < 0.10, ∗∗ p < 0.05,
∗∗∗
p < 0.01.
finer measure of temporal distance compared to the relatively noisy constructions of
migratory and linguistic distance.
In addition, note that the more sophisticated HMI migratory index has substantially more explanatory power for the cross-country variation in risk preferences than
the coarser predicted migratory distance measure. In fact, the results using the migratory distance measures already suggest that the correlation between our temporal
distance proxies and differences in risk preferences is not driven by simple geographic
distance: When we use the non-ancestry adjusted migratory distance measures in
the regressions, the explained variance drops substantially. This pattern suggests
that the precise migration patterns of our ancestors need to be taken into account to
understand the cross-country variation in risk aversion, rather than simple shortestdistance calculations between contemporary populations.
Given the superiority of the genetic data in proxying for temporal distance (both
conceptually and empirically), we proceed by employing the FST genetic distance as
main proxy and use the other variables in the robustness checks.
14
4.2
Controlling for Demographics, Income, Institutions, Geography, and Climate
The argument made in this paper is that our main result reflects the impact of ancient migration patterns and the resulting distribution of temporal distances across
populations, rather than contemporary differences in idiosyncratic country characteristics. To address the issue of omitted variable bias, we subject our result to an
extensive and comprehensive set of control variables, which are commonly used in the
comparative development and economic geography literatures. Since our dependent
variable consists of absolute differences, all of our control variables will also be bilateral variables that reflect cross-country differences along a wide range of dimensions.
In essence, in what follows, our augmented regression specification will be
|riski − riskj | = α + β × temporal distance proxyi,j + γ × di,j + i,j
where dij is a vector of bilateral measures between countries i and j (such as their
geodesic distance or the absolute difference in per capita income). Details on the
definitions and sources of all control variables can be found in Appendix G.
To ensure that our coefficient of interest does not spuriously pick up the effect
of demographic differences, column (2) of Table 2 adds to the baseline specification
the absolute differences in average age and the proportion of females. This does not
affect the coefficient on genetic distance.
Data on genetic, migratory and linguistic distances between populations correlate
with measures of the religious composition of the respective populations. Thus,
analogously to our linguistic distance measure, we use a “religion tree” to construct
an index of religious distance between two countries, following the same methodology
as in deriving linguistic distances.16 Column (3) of Table 2 further controls for this
distance measure. This does not affect our results.
A potentially important determinant of how people perceive risks is the composition of their social environment. For instance, a large religious fractionalization
might increase the occurrence of civic conflicts, potentially altering the way people
cope with risky choice situations. Thus, column (4) introduces as additional covariates the absolute differences in religious fractionalization and the fraction of the
16
Specifically, we first compute the number of common nodes two religions share in the “religion
tree” of Fearon (2006) and transformed this series according to
r
# Common nodes
Distance between two religions = 1 −
5
We then computed weighted religious distance estimates between two countries by weighing the
distances between religions with the respective population shares, so that our final distance measure
represents the expected religious distance between two randomly drawn individuals, one from each
country. This measure exhibits a modest correlation with genetic distance (ρ = 0.18).
15
population who are of European descent. While differences in religious fractionalization are indeed correlated with differences in risk preferences, the coefficient on
the fraction of people of European descent is actually negative.17 However, neither
in terms of the size of the point estimate nor in terms of its statistical significance
does the inclusion of these covariates affect our result.
A potential concern with our baseline specification is that it ignores differences in
comparative development and institutions across countries, in particular given that
genetic distance is known to correlate with differences in national income (Spolaore
and Wacziarg, 2009). Column (5) of Table 2 therefore introduces (log) GDP per
capita. If anything, this causes the coefficient on genetic distance to become slightly
larger.18 Column (6) further establishes that differences in the contemporary institutional environment (proxied by a democracy and a property rights index) as well
as common colonial experiences have little, if any, effect on our main result.19
Recall that our “world map” of risk preferences suggests the presence of geographic patterns in the distribution of risk attitudes. However, human migration
patterns (and hence temporal distance proxies) are correlated with geographic and
climatic variables. Thus, to ensure that effects stemming from variations in geography or climate are not attributed to genetic distance, we now condition on an
exhaustive set of corresponding control variables. Column (1) of Table 3 restates our
baseline regression from Table 1, while column (2) repeats the regression including
all demographic, economic, and institutional covariates from Table 2. Column (3)
introduces four distance metrics as additional controls into this regression. Our first
geographical control variable consists of the geodesic distance (measuring the shortest distance between any two points on earth) between the most populated cities of
the countries in a given pair. Relatedly, we introduce a dummy equal to one if two
countries are contiguous. Finally, we also condition on the “distance” between two
countries along the two major geographical axes, i.e., the difference in the distance
to the equator and the longitudinal (east-west) distance. Again, the introduction
of these variables has virtually no effect on the coefficient of genetic distance. This
pattern suggests that the precise migration patterns of our ancestors, rather than
simple shortest-distance calculations between contemporary populations, need to be
taken into account to understand the cross-country variation in risk aversion.
17
This negative coefficient is probably an artifact of this variable’s correlation with genetic distance (ρ = 0.09). In an unconditional regression, the coefficient on the fraction of European descent
is almost zero and far from being significant (p = 0.98).
18
Appendix C.2 shows that including differences in inequality or the average number of years of
education in a country pair yields very similar results.
19
This insight is robust to using alternative measures of institutional quality such as an index
gauging the constraints imposed on the executive (see Appendix C.2).
16
Given that geographic distance as such does not seem to drive our result, we now
control for more specific information about differences in the micro-geographic and
climatic conditions between the countries in a pair. To this end, we make use of a
wealth of information on the agricultural productivity of land, different features of
the terrain, and climatic factors. As column (4) shows, the inclusion of a large set of
geographic and climatic controls has virtually no effect on the genetic distance point
estimate. These results suggest that it is not geographic distance per se which drives
differential risk attitudes but rather the precise pattern of migration steps and the
distribution of temporal distances associated with these successive breakups.20
As discussed in the concluding remarks below and in Falk et al. (2015a), average risk aversion in a country is correlated with the health risks in the respective
environment. Thus, to ensure that the coefficient on genetic distance is not significantly biased upward due to the omission of these variables, column (6) conditions
on differences in life expectancy and the fraction of people who are at risk of contracting malaria. While life expectancy confers a significant relationship with risk
aversion, the point estimate of genetic distance drops by about 20% in size, but
remains statistically significant.
As in all cross-country regressions, a potential concern is that our main variable
of interest might simply pick up regional effects. For example, the largest genetic
distances occur between countries from different continents, in particular between
African and non-African countries. To account for this possibility, we construct an
extensive set of 28 continental dummies each equal to one if the two countries are
from two given continents. For example, we have a dummy equal to one if both
countries are from Sub-Saharan Africa, and another one equal to one if one country
is from Sub-Saharan Africa and the other one from North America. Conditional
on the other covariates, the inclusion of these fixed effects has no further effect on
the magnitude of the coefficient on genetic distance. In Appendix C.2, we present a
further specification in which we insert a set of 73 country dummies each equal to
one if a given country is part of a country pair. Even under this very conservative
specification, which also includes all covariates, the coefficient of genetic distance is
statistically significant and has a standardized beta of 18.2%.
20
Appendix C.2 provides further robustness checks against geographic features and transportation costs.
17
Table 2: Risk preferences, demographics, and economic structure
(1)
0.36∗∗∗
(0.10)
0.35∗∗∗
(0.09)
0.33∗∗∗
(0.09)
0.35∗∗∗
(0.09)
0.34∗∗∗
(0.09)
∆ Average age
0.0023
(0.00)
0.0020
(0.00)
0.0053∗
(0.00)
0.0096∗∗∗
(0.00)
0.011∗∗∗
(0.00)
∆ Proportion female
-0.0055
(0.74)
-0.037
(0.74)
-0.049
(0.72)
-0.11
(0.71)
-0.11
(0.66)
0.086
(0.06)
0.087
(0.06)
0.083
(0.06)
0.083
(0.06)
0.14∗∗
(0.07)
0.14∗∗
(0.07)
0.14∗∗
(0.07)
-0.053∗∗∗
(0.01)
-0.047∗∗∗
(0.01)
-0.049∗∗∗
(0.02)
-0.030∗∗∗
(0.01)
-0.026∗∗∗
(0.01)
Fst genetic distance
0.37∗∗∗
(0.09)
Dependent variable:
Absolute difference in average risk aversion
(2)
(3)
(4)
(5)
(6)
Religious distance
∆ Religious fractionalization
∆ % Of European descent
∆ Log [GDP p/c PPP]
∆ Democracy index
-0.0017
(0.00)
∆ Property rights
-0.00055
(0.00)
1 if common legal origin
-0.012
(0.02)
1 if ever colonial relationship
-0.0064
(0.03)
1 if colonial relationship post 1945
-0.037
(0.04)
1 if common colonizer post 1945
0.015
(0.03)
Constant
Observations
R2
Standardized beta (%)
0.21∗∗∗
(0.03)
0.21
(0.74)
0.19
(0.74)
0.18
(0.72)
0.26
(0.71)
0.27
(0.67)
2628
0.108
32.9
2628
0.110
32.3
2628
0.115
31.0
2628
0.129
29.4
2628
0.143
31.2
2556
0.146
30.6
OLS estimates, twoway-clustered standard errors in parentheses. The standardized betas refer to
genetic distance. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
18
Table 3: Risk preferences, geography, and climate
(1)
0.32
(0.08)
0.31
(0.09)
0.24
(0.09)
0.26∗∗∗
(0.10)
Log [Geodesic distance]
0.11∗∗∗
(0.04)
0.11∗∗∗
(0.04)
0.095∗∗∗
(0.03)
0.091∗∗∗
(0.03)
1 for contiguity
0.034
(0.04)
0.048
(0.04)
0.038
(0.04)
0.027
(0.03)
∆ Distance to equator
-0.0040∗∗∗
(0.00)
-0.0046∗∗∗
(0.00)
-0.0038∗∗∗
(0.00)
-0.0061∗∗∗
(0.00)
∆ Longitude
-0.0021∗∗∗
(0.00)
-0.0020∗∗∗
(0.00)
-0.0018∗∗∗
(0.00)
-0.00094∗
(0.00)
∆ Log [Area]
0.0016
(0.00)
0.0011
(0.00)
0.0043
(0.00)
∆ Land suitability for agriculture
-0.013
(0.05)
-0.0037
(0.05)
-0.013
(0.04)
0.000034
(0.00)
0.000081
(0.00)
0.00031
(0.00)
∆ Terrain roughness
0.059
(0.15)
0.044
(0.15)
0.11
(0.13)
∆ Mean elevation
-0.0086
(0.03)
-0.0067
(0.03)
-0.073∗
(0.04)
-0.099∗∗∗
(0.02)
-0.086∗∗∗
(0.02)
-0.010
(0.03)
0.039
(0.04)
-0.013
(0.04)
-0.048
(0.04)
∆ Ave precipitation
0.00033
(0.00)
0.00043
(0.00)
0.00095∗∗∗
(0.00)
∆ Ave temperature
0.00041
(0.00)
0.00051
(0.00)
0.0033∗∗∗
(0.00)
0.0076∗∗∗
(0.00)
0.00043
(0.00)
-0.033
(0.05)
-0.074
(0.06)
0.37
(0.09)
∗∗∗
0.34
(0.09)
∆ % Arable land
∆ SD Elevation
∆ Log [Neolithic transition]
∗∗∗
∆ Life expectancy
∆ % Population at risk of malaria
∗∗
(6)
∗∗∗
Fst genetic distance
∗∗∗
Dependent variable:
Absolute difference in average risk aversion
(2)
(3)
(4)
(5)
0.21∗∗∗
(0.03)
0.27
(0.67)
-0.50
(0.69)
-0.54
(0.71)
-0.32
(0.69)
-0.34
(0.63)
Demographic and economic controls
No
Yes
Yes
Yes
Yes
Yes
Continent FE
No
No
No
No
No
Yes
2628
0.108
32.9
2556
0.146
30.6
2556
0.206
36.0
2556
0.230
32.6
2556
0.244
24.2
2556
0.329
23.2
Constant
Observations
R2
Standardized beta (%)
OLS estimates, twoway-clustered standard errors in parentheses. See Table 2 for a complete list of the
demographic and economic control variables. The standardized betas refer to genetic distance. ∗ p < 0.10,
∗∗
p < 0.05, ∗∗∗ p < 0.01.
19
4.3
Selection on (Un-)Observables
In sum, conditioning on a large set of economic, institutional, geographic, climatic,
and demographic variables, the relationship between genetic distance and differences
in risk preferences is highly significant. Furthermore, the corresponding point estimate is rather robust: While the coefficient is 0.26 in the full regression model, it is
0.37 in the unconditional regression. Following the work of Altonji et al. (2005) and
Bellows and Miguel (2009), this observation can also be used to assess the extent to
which unobservable omitted variables likely bias our result. These authors show that
by comparing the coefficients in the unconditional and the full regression model, the
β̂ f ull
yields an estimate of how much larger the bias arising from
ratio β̂ unconditional
−β̂ f ull
unobservable factors would need to be relative to the bias resulting from observables,
in oder to generate the observed coefficient if the true coefficient was actually zero. If
this ratio equals α, then omitted variables can only explain the observed coefficient
if the bias resulting from unobservables equals at least α times the bias resulting
from observed variables. In our case, comparing the coefficients in columns (1) and
(6) of Table 3, the ratio equals 2.4.21 Thus, if the genetic distance coefficient was
to be explained away by selection on unobservables, then this bias would have to be
at least 2.4 times as strong as selection on the very large and comprehensive set of
covariates that we included in our regressions.
4.4
Robustness
While the majority of the genetic distances in our sample are relatively small, we also
observe a number of large genetic distances. To ensure that our results are not driven
by these large distances, we pursue two distinct strategies. First, using the same
specification as in our full regression model in column (6) of Table 3, we exclude the
largest genetic distances step-by-step. Second, we reduce the magnitude of differences
in genetic distances by estimating a log-log relationship between differences in risk
preferences and genetic distance. Columns (1) through (4) of Table 4 present the
results of the regressions utilizing the restricted samples, which demonstrate that
excluding large distances does not affect the statistical significance of our result.
Furthermore, column (13) shows that the spirit of our results is unaffected in a
log-log specification.
Our control variables already accounted for the presence of differences in continent
fixed effects. To provide further evidence that no single region drives our result,
columns (5) through (12) of Table 4 present the results from regressions which exclude
21
The ratio is the same if we consider the common sample from the specifications from
columns (1) and (6) in Table 3.
20
the four major regions of the world one-by-one. In most specifications, the results
remain statistically significant despite the large drop in the number of observations,
and conditional on the full set of covariates.
In Appendix C.1 we provide further robustness checks by employing different
dependent variables than the absolute difference in average risk attitudes. First, we
show that our main result continues to hold when we use the absolute difference in
the 25th, 50th or 75th percentile of risk aversion as dependent variable. Second,
recall that our index of risk aversion consists of a linear combination of two types of
survey items, a qualitative self-assessment and a series of quantitative binary lottery
choices. By employing both items separately, we show that the effect of temporal
distance on risk preferences is not specific to a particular type of survey question.
4.5
Alternative Measures
Up to this point, all of our multivariate regressions were conducted using the raw
FST genetic distance. We now present a set of regressions which utilize additional
information on the measurement precision of the genetic distance data. Since the
data on allele frequencies are collected from different sample sizes across countries,
the precision of the measurement varies across country pairs. Thus, using bootstrap
analysis, Cavalli-Sforza et al. (1994) compute standard errors of the FST genetic distance for each pair. Following the methodology proposed by Spolaore and Wacziarg
(2009), we utilize this information by linearly downweighing each observation by its
standard error (see Appendix D). The results from the corresponding weighted least
squares regression, are presented in columns (1)-(3) of Table 10. The resulting standardized beta coefficients are highly significant and very similar to those from the
unweighted regressions, providing reassuring evidence that our results are not driven
by measurement error.
Section 4.1 showed that the raw relationship between temporal distance and
differences in risk preferences also holds when employing the Nei genetic distance
or migratory distance measures as proxies. Columns (4)-(12) establish that these
raw correlations are robust to the full set of covariates. Notably, the results obtained
with the predicted measures of migratory distance further emphasize that it is indeed
ancient migration patterns rather than genetic distance as such which is at the core
of our analysis.22
22
In contrast, when we employ linguistic distance in the conditional regressions, we obtain small
and insignificant coefficients. This is consistent with the idea that the coarse and noisy construction
of linguistic distances introduces severe measurement error and is hence an inferior proxy for timing
of separation as compared to genetic or predicted migratory distance data.
21
22
Yes
Yes
2556
0.329
23.2
Continent FE
Observations
R2
Standardized beta (%)
(3)
2258
0.297
29.4
Yes
Yes
Yes
-0.22
(0.64)
0.41∗∗∗
(0.11)
(4)
1925
0.209
24.4
Yes
Yes
Yes
-0.37
(0.70)
0.40∗∗∗
(0.11)
(5)
1653
0.171
32.4
No
No
Yes
0.21
(0.66)
0.36∗∗∗
(0.10)
(6)
1653
0.287
22.4
No
Yes
Yes
-0.90
(0.69)
0.25∗∗
(0.10)
(7)
1081
0.110
20.1
No
No
Yes
0.43
(0.76)
0.22∗∗
(0.09)
(8)
1081
0.207
12.2
No
Yes
Yes
0.39
(0.73)
0.13
(0.10)
(9)
1711
0.183
31.5
No
No
Yes
0.067
(0.77)
0.36∗∗∗
(0.09)
(10)
1711
0.285
29.8
No
Yes
Yes
-0.85
(0.76)
0.34∗∗∗
(0.12)
1326
0.031
12.0
No
No
Yes
0.26
(0.49)
0.15
(0.10)
(11)
1326
0.079
20.5
No
Yes
Yes
0.26
(0.59)
0.25∗∗
(0.11)
(12)
Africa & ME
2556
0.235
16.4
Yes
Yes
Yes
-2.91
(1.90)
0.17∗∗
(0.07)
(13)
Log [Abs. difference
in risk attitudes]
OLS estimates, twoway-clustered standard errors in parentheses. EU = Europe, ME = Middle East, SE Asia = South Asia and South-East Asia. The variables in logs are defined
as ln[0.01 + x], where x denotes the original variable. See Tables 2 and 3 for a complete list of the demographic, economic, geographic, climatic, and health control variables. The
standardized betas refer to genetic distance. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
2483
0.329
25.8
Yes
Yes
Geographic, climate and health controls
Yes
-0.28
(0.64)
Yes
-0.34
(0.63)
(2)
0.31∗∗∗
(0.10)
(1)
0.26∗∗∗
(0.10)
Demographic and economic controls
Constant
Log [Fst genetic distance]
Fst genetic distance
Restricted sample: Fst <
Baseline < 0.85 < 0.7
< 0.5
Dependent variable:
Absolute difference in average risk aversion
Exclude continents
Americas
EU & Central Asia SE Asia & Pacific
Table 4: Robustness: Exclude large genetic distances or continents
23
No
Continent FE
2556
0.186
20.5
No
Yes
Yes
-0.31
(0.71)
0.25
(0.09)
∗∗∗
(2)
2556
0.283
26.9
Yes
Yes
Yes
-0.31
(0.66)
0.33
(0.09)
∗∗∗
(3)
2556
0.244
23.2
No
Yes
Yes
-0.26
(0.69)
0.27∗∗∗
(0.09)
(4)
2556
0.327
22.0
Yes
Yes
Yes
-0.31
(0.64)
0.26∗∗∗
(0.09)
(5)
2628
0.059
24.3
No
No
No
0.23∗∗∗
(0.03)
0.34∗∗∗
(0.10)
(6)
2556
0.162
18.5
No
Yes
Yes
-0.21
(0.71)
0.26∗∗∗
(0.09)
(7)
2556
0.256
23.6
Yes
Yes
Yes
-0.26
(0.68)
0.34∗∗∗
(0.10)
(8)
2556
0.236
14.1
No
Yes
Yes
-0.26
(0.68)
0.87∗∗∗
(0.33)
(9)
(10)
2556
0.328
21.3
Yes
Yes
Yes
-0.23
(0.62)
1.31∗∗∗
(0.41)
OLS
2016
0.268
25.0
No
Yes
Yes
-0.031
(0.84)
0.69∗∗∗
(0.20)
(11)
(12)
2016
0.355
24.5
Yes
Yes
Yes
-0.17
(0.77)
0.67∗∗
(0.28)
OLS
OLS estimates, twoway-clustered standard errors in parentheses. The weighted least-squares regressions linearly downweigh observations by the
standard error of the respective genetic distance estimate, as obtained from bootstrap analysis by Cavalli-Sforza et al. (1994), see Appendix D for
details. See Tables 2 and 3 for a complete list of the demographic, economic, geographic, climatic, and health control variables. The standardized
betas refer to the temporal distance proxies. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
2628
0.074
27.3
No
Geographic, climate and health controls
Observations
R2
Standardized beta (%)
No
0.22∗∗∗
(0.03)
0.34
(0.09)
Demographic and economic controls
Constant
HMI migratory distance
(Özak, 2010)
Predicted migratory distance
Nei genetic distance
Fst genetic distance
∗∗∗
(1)
WLS
Dependent variable:
Absolute difference in average risk aversion
OLS
WLS
Table 5: Alternative measures and estimation techniques
5
Preference Heterogeneity and Migratory Distance
from East Africa
The final step of our analysis considers the relationship between the within-population
variability in risk preferences and migratory distance from Africa. To this end, we
estimate the following equation:
σi = α + β × Predicted migratory distance from Ethiopiai + γ × xi + i
where σi denotes the standard deviation in risk attitudes within a country, xi
a vector of covariates, and i a disturbance term. Table 6 provides an overview of
the results of corresponding OLS regressions. In columns (1)-(4), we use predicted
migratory distance from Ethiopia (constructed as described in Section 3.2) as explanatory variable, while columns (5) through (8) employ the migratory distance
measure which is based on the human mobility index.
Column (1) indicates that the average total length of the migration path of a
given population from East Africa exhibits a raw correlation with the standard deviation in risk attitudes. Columns (2) through (3) add several control variables to the
baseline specification, which capture cross-country differences in sociodemographics,
development, geography, and climate. The details of the corresponding regressions,
i.e., the coefficients of the covariates, can be found in Appendix E. The results show
that, if anything, conditioning on this vector of covariates strengthens the correlation
between preference heterogeneity and migratory distance. However, as column (4)
indicates, the inclusion of continental fixed effects leads the coefficient of migratory
distance to drop in size and to become insignificant. By the serial founder effect and
the paths of human migration, continent dummies are correlated with migratory distance from East Africa and take out a substantial fraction of the variation not only
in migratory distance but also in preference heterogeneity. Columns (5) through (8)
repeat the same specifications using the more sophisticated mobility index-adjusted
measure of migratory distance. Consistent with measurement error in the baseline
migratory distance measure biasing the coefficient towards zero, the results are, if
anything, even stronger, and remain weakly significant even under continent fixed
effects. Figure 5 in Appendix E provides a graphical illustration of the corresponding
raw relationship.
Taken together, our second set of results shows that ancient migration patterns
are not only reflected in cross-country differences in average risk aversion, but also
in the dispersion of the preference pool.
24
Table 6: Preference dispersion and migratory distance from East Africa
(1)
Migratory distance from Ethiopia
(2)
∗
-0.52
(0.28)
Dependent variable: SD in risk aversion
(3)
(4)
(5)
(6)
(7)
∗∗
-0.83
(0.32)
∗
-0.64
(0.36)
-0.15
(0.47)
HMI migratory distance from Ethiopia
(Özak, 2010)
Constant
Socioeconomic controls
(8)
-0.52∗∗
(0.21)
-0.77∗∗∗
(0.26)
-0.73∗∗
(0.29)
-0.80∗
(0.45)
0.98∗∗∗
(0.02)
0.11
(0.42)
1.16∗∗
(0.52)
0.80
(0.62)
1.00∗∗∗
(0.02)
-0.47
(0.56)
1.36∗∗
(0.62)
1.51∗∗
(0.70)
No
Yes
Yes
Yes
No
Yes
Yes
Yes
Geographic controls
No
No
Yes
Yes
No
No
Yes
Yes
Continent FE
No
No
No
Yes
No
No
No
Yes
73
0.048
-21.8
73
0.380
-34.6
73
0.509
-26.9
73
0.568
-6.4
64
0.074
-27.2
64
0.412
-40.0
64
0.578
-37.9
64
0.595
-41.5
Observations
R2
Standardized beta (%)
OLS estimates, robust standard errors in parentheses. In columns (2) and (6), the additional controls include
average age, proportion of females, log (GDP p/c) and its square, SD in age, a gender heterogeneity index, SD
in (log) household income, (log) population size, linguistic diversity and life expectancy. Columns (3) and (7)
additionally include distance to the equator, temperature, precipitation, suitability for agriculture, the fraction
of the population at risk of contracting malaria, and the log of the timing of the Neolithic revolution, while
columns (4) and (8) add continent fixed effects. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
6
Conclusion
We analyze the distribution of risk preferences of representative population samples
in countries spanning all continents. The main contribution of this paper is to
establish that, along two dimensions, a significant fraction of the substantial betweencountry heterogeneity has its deep historical origins in ancestral migration patterns.
As for average risk attitudes, we show that three classes of proxies for the length of
time since two populations shared common ancestors are predictive of how similar
two countries’ risk preferences are. Regarding within-country heterogeneity, we show
that the migratory distance from East Africa predicts the dispersion of the preference
distribution.
A growing body of empirical work highlights the importance of heterogeneity in
risk preferences for understanding a myriad of economic and health behaviors at
the individual level. However, it remains an open question whether country-level
risk preferences are related to the cross-country variation in economic, institutional,
and health-related outcomes. In our data, average risk aversion is indeed predictive
of variables that one would intuitively expect to be associated with risk attitudes.
Specifically, more risk-tolerant societies tend to live in more risky economic, institutional, and health environments (see Falk et al., 2015a, and Appendix F). Regarding
economic risks, high risk taking is associated with high income inequality (Gini index,
income share of the top 10%) and a high risk of living in poverty, i.e., on less than
25
$2 per day, conditional on per capita income. Similarly, risk averse populations tend
to live in institutional structures which provide strong labor protection against, e.g.,
unemployment. Similarly, as for health risks, a higher risk tolerance is significantly
related to a higher prevalence of HIV, traffic death rates, and lower life expectancy.
The latter relationship is depicted in the left-hand panel of Figure 2. Thus, in sum,
risk averse socities appear to live in environments which are intuitively less risky, providing suggestive evidence that the cross-country heterogeneity in risk preferences
might have important consequences not just for individual decision making.
However, it is conceivable that not just the level of risk aversion in a given country has economic effects, but also the corresponding variability. Indeed, Ashraf and
Galor (2013b) argue for the importance of cultural diversity in explaining comparative development around the globe. Recall that these authors show that per capita
income is hump-shaped in the diversity of the respective genetic pool (which is a
transformation of migratory distance from East Africa). Given our finding that the
within-country dispersion in the pool of risk preferences decreases in migratory distance from East Africa, and hence increases in the diversity of the genetic pool, it is
a natural question to ask whether there is a relationship between the variability in
risk attitudes and economic development. The right-hand panel of Figure 2 shows
that per capita income monotonically decreases in the standard deviation of risk
attitudes, with a raw correlation of ρ = −0.32, p < 0.01. As we show in Appendix F,
this correlation is very robust and holds up when we condition on continent fxied
effects, a large number of geographic and climatic covariates, genetic diversity and
its square, ethnic or linguistic fractionalization or even institutional quality and average years of schooling. In fact, as we show in the Appendix, this correlation is
much stronger when we consider the standard deviation in the quantitative choice
SD in risk taking and per capita income
12
Risk taking and life expectancy
ISR
CHE
ITAFRA
SWEAUS
ESP
CAN
GRCAUT
GBR NLD
DEU
KOR
CRI
CHL
USA
CZE
ARE
MEX
HRV
BIH
POL
ARG
VNM
SAU
CHN
LKA PERVEN
HUN
SRB
GEO
EST
TUR
THA
COL
ROUJOR
NIC
LTU
BRA
IRN
DZA
GTM
EGY
IRQ
SUR
IDN UKR MAR
BGD MDA
PHL
KHM
RUS
KAZ
PAKBOL
IND
FIN
HTI
RWA
50
CMR
GHA
AFG
KEN
UGA
JPN
TZA
ZAF
Log [GDP p/c PPP]
8
10
Life expectancy at birth
60
70
PRT
CHE
USA
SWE
ARE
CAN
DEU FIN GBR
FRAAUT
ESP
ISRGRC
PRT
SAU
CZE
HUN
HRV
EST
POL
LTU MEX
TUR CHL
VENROU
BWA
RUS
ARG
ZAF
BRA
CRI
COL SUR
SRBKAZ
DZA
BIH
PER
IRN
THA
JOR
GTM
IRQ
UKR MAR CHN
EGY GEO
LKA
PHL
NIC
CMR BOL
MDA
NGA
IND
VNM
PAK
KEN GHA
ZWE
HTI KHM
BGD
TZA
UGA
RWA
AFG
MWI
NLD
AUS
ITA
KOR
IDN
6
80
JPN
NGAMWI
BWA
4
40
ZWE
-1
-.5
0
Willingness to take risks
.5
1
.7
.8
.9
1
SD in willingness to take risks
1.1
1.2
Figure 2: Risk attitudes, health and economic outcomes. The left panel depicts the raw correlation
between life expectancy at birth and risk taking, while the right panel depicts the relationship
between (log) per capita income and the standard deviation in risk aversion.
26
procedure only.
In sum, both the first and second moment of the within-country distribution
of risk preferences are associated with heterogeneous economic, institutional, and
health outcomes. Our results highlight that if we aim to understand the ultimate
roots of the underlying preference heteroegeneity, we might have to consider events
very far back in time.
27
References
Alesina, Alberto, Arnaud Devleeschauwer, William Easterly, Sergio
Kurlat, and Romain Wacziarg, “Fractionalization,” Journal of Economic
Growth, 2003, 8 (2), 155–194.
, Paola Giuliano, and Nathan Nunn, “On the Origins of Gender Roles: Women
and the Plough,” Quarterly Journal of Economics, 2013, 128 (2), 469–530.
Altonji, Joseph G., Todd E. Elder, and Christopher R. Taber, “Selection
on Observed and Unobserved Variables: Assessing the Effectiveness of Catholic
Schools,” Journal of Political Economy, 2005, 113 (1), 151–184.
Arbatli, Cemal Eren, Quamrul Ashraf, and Oded Galor, “The Nature of
Civil Conflict,” Working Paper, 2013.
Ashraf, Quamrul and Oded Galor, “Genetic Diversity and the Origins of Cultural
Fragmentation,” American Economic Review: Papers & Proceedings, 2013, 103 (3),
528–33.
and , “The Out of Africa Hypothesis, Human Genetic Diversity, and Comparative Economic Development,” American Economic Review, 2013, 103 (1), 1–46.
,
, and Marc P. Klemp, “The Out of Africa Hypothesis of Comparative
Development Reflected by Nighttime Light Intensity,” Working Paper, 2014.
Barro, Robert J. and Jong-Wha Lee, “A New Data Set of Educational Attainment in the World, 1950–2010,” Journal of Development Economics, 2012.
Bellows, John and Edward Miguel, “War and Local Collective Action in Sierra
Leone,” Journal of Public Economics, 2009, 93 (11), 1144–1157.
Botero, Juan, Simeon Djankov, Rafael LaPorta, Florencio Lopez
de Silanes, and Andrei Shleifer, “The Regulation of Labor,” Quarterly Journal
of Economics, 2004, 119 (4), 1339–1382.
Callen, Michael, Mohammad Isaqzadeh, James D. Long, and Charles
Sprenger, “Violence and Risk Preference: Experimental Evidence from
Afghanistan,” American Economic Review, 2014, 104 (1), 123–148.
Cameron, A. Colin, Jonah B. Gelbach, and Douglas L. Miller, “Robust
Inference with Multiway Clustering,” Journal of Business & Economic Statistics,
2011, 29 (2).
28
Cavalli-Sforza, Luigi L., “Genes, Peoples, and Languages,” Proceedings of the
National Academy of Sciences, 1997, 94 (15), 7719–7724.
Cavalli-Sforza, Luigi, Paolo Menozzi, and Alberto Piazza, The History and
Geography of Human Genes, Princeton university press, 1994.
Cesarini, David, Christopher T. Dawes, Magnus Johannesson, Paul Lichtenstein, and Björn Wallace, “Genetic Variation in Preferences for Giving and
Risk Taking,” Quarterly Journal of Economics, 2009, 124 (2), 809–842.
Chen, M. Keith, “The Effect of Language on Economic Behavior: Evidence from
Savings Rates, Health Behaviors, and Retirement Assets,” American Economic
Review, 2013, 103 (2), 690–731.
Desmet, Klaus, Ignacio Ortuno-Ortin, and Romain Wacziarg, “Culture,
Ethnicity and Diversity,” Working Paper, 2014.
, Michel Le Breton, Ignacio Ortuno-Ortin, and Shlomo Weber, “The
Stability and Breakup of Nations: A Quantitative Analysis,” Journal of Economic
Growth, 2011, 16, 183–213.
Dohmen, Thomas, Armin Falk, David Huffman, and Uwe Sunde, “The
Intergenerational Transmission of Risk and Trust Attitudes,” Review of Economic
Studies, 2012, 79 (2), 645–677.
,
,
,
, Jürgen Schupp, and Gert G. Wagner, “Individual Risk Attitudes: Measurement, Determinants, and Behavioral Consequences,” Journal of
the European Economic Association, 2011, 9 (3), 522–550.
Falk, Armin, Anke Becker, Thomas Dohmen, Benjamin Enke, Uwe Sunde,
and David Huffman, “The Nature of Human Preferences: Global Evidence,”
Working Paper, 2015.
,
,
, David Huffman, and Uwe Sunde, “An Experimentally Validated
Survey Module of Economic Preferences,” Working Paper, 2015.
Fearon, James, “Is Genetic Distance a Plausible Measure of Cultural Distance?,”
unpublished, Stanford University, 2006.
Fearon, James D., “Ethnic and Cultural Diversity by Country,” Journal of Economic Growth, 2003, 8 (2), 195–222.
29
Fehr, Ernst and Karla Hoff, “Introduction: Tastes, Castes and Culture: the
Influence of Society on Preferences,” Economic Journal, 2011, 121 (556), F396–
F412.
Fernández, Raquel and Alessandra Fogli, “Fertility: The Role of Culture and
Family Experience,” Journal of the European Economic Association, 2006, 4 (2-3),
552–561.
and , “Culture: An Empirical Investigation of Beliefs, Work, and Fertility,”
American Economic Journal: Macroeconomics, 2009, 1 (1), 146–177.
Gallup, John L., Jeffrey D. Sachs, and Andrew D. Mellinger, “Geography
and Economic Development,” International Regional Science Review, 1999, 22 (2),
179–232.
,
et al., “The Economic Burden of Malaria.,” American Journal of Tropical
Medicine and Hygiene, 2000, 64 (1-2 Suppl), 85–96.
Galor, Oded and Ömer Özak, “The Agricultural Origins of Time Preference,”
Working Paper, 2014.
Giuliano, Paola, Antonio Spilimbergo, and Giovanni Tonon, “Genetic Distance, Transportation Costs, and Trade,” Journal of Economic Geography, 2014,
14 (1), 179–198.
Gorodnichenko, Yuriy and Gerard Roland, “Culture, Institutions and the
Wealth of Nations,” Working Paper, 2010.
Guiso, Luigi, Paola Sapienza, and Luigi Zingales, “Does Culture Affect Economic Outcomes?,” Journal of Economic Perspectives, 2006, 20 (2), 23–48.
,
, and , “Cultural Biases in Economic Exchange?,” Quarterly Journal of
Economics, 2009, 124 (3), 1095–1131.
Henn, Brenna M., Luigi L. Cavalli-Sforza, and Marcus W. Feldman, “The
Great Human Expansion,” Proceedings of the National Academy of Sciences, 2012,
109 (44), 17758–17764.
Mecham, R. Quinn, James Fearon, and David Laitin, “Religious Classification
and Data on Shares of Major World Religions,” unpublished, Stanford University,
2006, 1 (5), 18.
Michalopoulosa, Stelios, “The Origins of Ethnolinguistic Diversity,” American
Economic Review, 2012, 102 (4), 1508–1539.
30
Nordhaus, William D., “Geography and Macroeconomics: New data and new
Findings,” Proceedings of the National Academy of Sciences of the United States
of America, 2006, 103 (10), 3510–3517.
Nunn, Nathan and Leonard Wantchekon, “The Slave Trade and the Origins of
Mistrust in Africa,” American Economic Review, 2011, 101 (7), 3221–52.
Özak, Ömer, “The Voyage of Homo-Oeconomicus: Some Economic Measures of
Distance,” Working Paper, 2010.
Proto, Eugenio and Andrew J. Oswald, “National Happiness and Genetic Distance: A Cautious Exploration,” Working Paper, 2014.
Putterman, Louis and David N. Weil, “Post-1500 Population Flows and the
Long-Run Determinants of Economic Growth and Inequality,” Quarterly Journal
of Economics, 2010, 125 (4), 1627–1682.
Ramachandran, Sohini, Omkar Deshpande, Charles C. Roseman, Noah A.
Rosenberg, Marcus W. Feldman, and L. Luca Cavalli-Sforza, “Support
from the Relationship of Genetic and Geographic Distance in Human Populations
for a Serial Founder Effect Originating in Africa,” Proceedings of the National
Academy of Sciences of the United States of America, 2005, 102 (44), 15942–15947.
Rieger, Marc Oliver, Mei Wang, and Thorsten Hens, “Risk Preferences
Around the World,” Management Science, 2014.
Spolaore, Enrico and Romain Wacziarg, “The Diffusion of Development,” Quarterly Journal of Economics, 2009, 124 (2), 469–529.
and , “Long-Term Barriers to the International Diffusion of Innovations,” in
“NBER International Seminar on Macroeconomics 2011” University of Chicago
Press 2011, pp. 11–46.
and
, “War and Relatedness,” Working Paper, 2014.
Tabellini, Guido, “Institutions and Culture,” Journal of the European Economic
Association, 2008, 6 (2-3), 255–294.
Vieider, Ferdinand M., Mathieu Lefebvre, Ranoua Bouchouicha, Thorsten
Chmura, Rustamdjan Hakimov, Michal Krawczyk, and Peter Martinsson, “Common Components of Risk and Uncertainty Attitudes Across Contexts
and Domains: Evidence from 30 Countries,” Journal of the European Economic
Association, forthcoming.
31
, Thorsten Chmura, and Peter Martinsson, “Risk Attitudes, Development,
and Growth: Macroeconomic Evidence from Experiments in 30 Countries,” Discussion Paper, Social Science Research Center Berlin, 2012.
Voigtländer, Nico and Hans-Joachim Voth, “Persecution Perpetuated: The
Medieval Origins of Anti-Semitic Violence in Nazi Germany,” Quarterly Journal
of Economics, 2012, 127 (3), 1339–1392.
32
ONLINE APPENDIX
A
Details on Data Collection of Risk Preferences
The description of the dataset builds on Falk et al. (2015a).
A.1
Overview
The cross-country dataset covering risk aversion, patience, positive and negative
reciprocity, altruism, and trust, was collected through the professional infrastructure
of the Gallup World Poll 2012. The data collection process essentially consisted
of three steps. First, we conducted a laboratory experimental validation procedure
to select the survey items. Second, Gallup implemented a pre-test in a variety of
countries to ensure the implementability of our items in a culturally diverse sample.
Third, the data were collected through the regular professional data collection efforts
in the framework of the World Poll 2012.
A.2
Experimental Validation
To ensure the behavioral relevance of our preference measures, all underlying survey items were selected through an experimental validation procedure. To this end,
a sample of German undergraduates completed standard state-of-the-art financially
incentivized laboratory experiments designed to measure risk aversion, patience, positive and negative reciprocity, altruism, and trust. The same sample of subjects then
completed a large battery of potential survey items. In a final step, for each preference, those survey items were selected which jointly perform best in predicting
financially incentivized behavior. See Falk et al. (2015a) for details. The resulting
set of items was then adjusted in the process of a pre-test.
A.3
Pre-Test
Prior to including the preference module in the Gallup World Poll 2012, it was
tested in the field as part of the World Poll 2012 pre-test, which was conducted
at the end of 2011 in 22 countries. The main goal of the pre-test was to receive
feedback and comments on each item from various cultural backgrounds in order
to assess potential difficulties in understanding and differences in the respondents’
interpretation of items. Based on respondents’ feedback and suggestions, several
items were modified in terms of wording before running the survey as part of the
World Poll 2012.
33
The pre-test was run in 10 countries in central Asia (Russia, Belarus, Kyrgyzstan, Azerbaijan, Georgia, Tajikistan, Uzbekistan, Kazakhstan, Armenia, and Turkmenistan), 2 countries in South-East Asia (Bangladesh and Cambodia), 5 countries
in Southern and Eastern Europe (Croatia, Hungary, Poland, Romania, Turkey), 4
countries in the Middle East and North Africa (Saudi-Arabia, Lebanon, Jordan, Algeria), and 1 country in Eastern Africa (Kenya). In each country, the sample size
was 10 to 15 people. Overall, more than 220 interviews were conducted. In most
countries, the sample was mixed in terms of gender, age, educational background,
and area of residence (urban / rural).
Participants of the pre-test were asked to state any difficulties in understanding
the items and to rephrase the meaning of items in their own words. If encountering
difficulties in understanding or interpretations of items that were not intended, respondents were asked to make suggestions on how to modify the wording of the item
in order to attain the desired meaning.
Overall, the understanding of both the qualitative items and the quantitative
items was good. In particular, no interviewer received any complaints regarding
difficulties in assessing the quantitative questions. When asked for rephrasing the
qualitative item in their own words, most participants seemed to have understood
the item in exactly the way that was intended.
However, when being confronted with hypothetical choices between monetary
amounts today versus larger amounts one year later, some participants, especially
in countries with current or relatively recent phases of volatile and high inflation
rates, stated that their answer would depend on the rate of inflation, or said that
they would always take the immediate payment due to uncertainty with respect
to future inflation. Therefore, we decided to adjust the wording, relative to the
“original” experimentally validated item, by adding the phrase “Please assume there
is no inflation. i.e. future prices are the same as today’s prices” to each question
involving hypothetical choices between immediate and future monetary amounts.
A.4
Selection of Countries
Our goal when selecting countries was to ensure representativeness for the global population. Thus, we chose countries from each continent and each region within continents. In addition, we aimed at maximizing variation with respect to observables,
such as GDP per capita, language, historical and political characteristics, or geographical location and climatic conditions. Accordingly, we favored non-neighboring
and culturally dissimilar countries. This procedure resulted in the following sample
of 76 countries:
34
East Asia and Pacific: Australia, Cambodia, China, Indonesia, Japan, Philippines,
South Korea, Thailand, Vietnam
Europe and Central Asia: Austria, Bosnia and Herzegovina, Croatia, Czech Republic, Estonia, Finland, France, Georgia, Germany, Greece, Hungary, Italy, Kazakhstan, Lithuania, Moldova, Netherlands, Poland, Portugal, Romania, Russia, Serbia, Spain, Sweden, Switzerland, Turkey, Ukraine, United Kingdom
Latin America and Caribbean: Argentina, Bolivia, Brazil, Chile, Colombia, Costa
Rica, Guatemala, Haiti, Mexico, Nicaragua, Peru, Suriname, Venezuela
Middle East and North Africa: Algeria, Egypt, Iran, Iraq, Israel, Jordan, Morocco,
Saudi Arabia, United Arab Emirates
North America: United States, Canada
South Asia: Afghanistan, Bangladesh, India, Pakistan, Sri Lanka
Sub-Saharan Africa: Botswana, Cameroon, Ghana, Kenya, Malawi, Nigeria, Rwanda,
South Africa, Tanzania, Uganda, Zimbabwe
A.5
A.5.1
Sampling and Survey Implementation
Background
Since 2005, the international polling company Gallup conducts an annual World
Poll, in which it surveys representative population samples in almost every country
around the world on, e.g., economic, social, political, and environmental issues. The
collection of our preference data was embedded into the regular World Poll 2012
and hence made use of the pre-existing polling infrastructure of one of the largest
professional polling institutes in the world.23
A.5.2
Survey Mode
Interviews were conducted via telephone and face-to-face. Gallup uses telephone surveys in countries where telephone coverage represents at least 80% of the population
or is the customary survey methodology. In countries where telephone interviewing
is employed, Gallup uses a random-digit-dial method or a nationally representative
list of phone numbers. In countries where face-to-face interviews are conducted,
households are randomly selected in an area-frame-design.
23
Compare
http://www.gallup.com/strategicconsulting/156923/worldwide-research-methodology.
aspx
35
A.5.3
Sample Composition
In most countries, samples are nationally representative of the resident population
aged 15 and older. Gallup’s sampling process is as follows.
Selecting Primary Sampling Units
In countries where face-to-face interviews are conducted, the first stage of sampling
is the identification of primary sampling units (PSUs), consisting of clusters of households. PSUs are stratified by population size and / or geography and clustering is
achieved through one or more stages of sampling. Where population information is
available, sample selection is based on probabilities proportional to population size.
If population information is not available, Gallup uses simple random sampling.
In countries where telephone interviews are conducted, Gallup uses a randomdigit-dialing method or a nationally representative list of phone numbers. In countries where cell phone penetration is high, Gallup uses a dual sampling frame.
Selecting Households and Respondents
Gallup uses random route procedures to select sampled households. Unless an outright refusal to participate occurs, interviewers make up to three attempts to survey
the sampled household. To increase the probability of contact and completion, interviewers make attempts at different times of the day, and when possible, on different
days. If the interviewer cannot obtain an interview at the initially sampled household, he or she uses a simple substitution method.
In face-to-face and telephone methodologies, random respondent selection is achieved by using either the latest birthday or Kish grid method.24 In a few Middle
East and Asian countries, gender-matched interviewing is required, and probability
sampling with quotas is implemented during the final stage of selection. Gallup implements quality control procedures to validate the selection of correct samples and
that the correct person is randomly selected in each household.
Sampling Weights
Ex post, data weighting is used to ensure a nationally representative sample for each
country and is intended to be used for calculations within a country. First, base sampling weights are constructed to account for geographic oversamples, household size,
24
The latest birthday method means that the person living in the household whose birthday
among all persons in the household was the most recent (and who is older than 15) is selected for
interviewing. With the Kish grid method, the interviewer selects the participants within a household
by using a table of random numbers. The interviewer will determine which random number to use
by looking at, e.g., how many households he or she has contacted so far (e.g., household no. 8) and
how many people live in the household (e.g., 3 people, aged 17, 34, and 36). For instance, if the
corresponding number in the table is 7, he or she will interview the person aged 17.
36
and other selection probabilities. Second, post-stratification weights are constructed.
Population statistics are used to weight the data by gender, age, and, where reliable
data are available, education or socioeconomic status.
A.5.4
Translation of Items
The preference module items were translated into the major languages of each target
country. The translation process involved three steps. As a first step, a translator
suggested an English, Spanish or French version of a German item, depending on
the region. A second translator, being proficient in both the target language and
in English, French, or Spanish. then translated the item into the target language.
Finally, a third translator would review the item in the target language and translate
it back into the original language. If semantic differences between the original item
and the back-translated item occurred, the process was adjusted and repeated until
all translators agreed on a final version.
A.5.5
Adjustment of Monetary Amounts in Quantitative Items
All items involving monetary amounts were adjusted to each country in terms of
their real value. In each country, all monetary amounts were calculated to represent
the same share of the country’s median income in local currency as the share of the
amount in Euro of the German median income since the validation study had been
conducted in Germany. Monetary amounts used in the validation study with the
German sample were round numbers in order to facilitate easy calculations and to
allow for easy comparisons. In order to proceed in a similar way in all countries, we
rounded all monetary amounts to the next “round” number. While this necessarily
resulted in some (if very minor) variation in the real stake size between countries, it
minimized cross-country differences in the difficulty of understanding the quantitative
items due to differences in assessing the involved monetary amounts.
A.5.6
Staircase procedure
The sequence of survey questions that form the basis for the quantitative risk aversion measure is given by the “tree” logic depicted in Figure 3. In Germany, each
respondent faced five interdependent choices between taking part in a 50:50 lottery
of receiving 300 euros or nothing and varying safe amounts of money. The values
in the tree denote the safe amounts of money. The rightmost level of the tree (5th
decision) contains 16 distinct monetary amounts, so that responses can be classified
into 32 categories which are ordered in the sense that the (visually) lowest path /
37
endpoint indicates the highest level of risk aversion. As in the experimental validation procedure in Falk et al. (2015b), we assign values 1-32 to these endpoints, with
32 denoting the highest level of risk aversion.
A.6
A.6.1
Computation of Preference Measures
Cleaning and Imputation of Missings
In order to make maximal use of the available information in our data, missing survey
items were imputed based on the following procedure:
If one survey item was missing, then the missing item was predicted using the
responses to the other item. The procedure was as follows:
• Qualitative question missing: We regress all available survey responses to the
qualitative question on responses to the staircase task, and then use these
coefficients to predict the missing qualitative items using the available staircase
items.
• Staircase item missing: The imputation procedure was similar, but made additional use of the informational content of the responses of participants who
started but did not finish the sequence of the five questions. If the respondent did not even start the staircase procedure, then imputation was done by
predicting the staircase measure based on answers to the qualitative survey
measure using the methodology described above. On the other hand, if the
respondent answered at least one of the staircase questions, the final staircase
outcome was based on the predicted path through the staircase procedure. Suppose the respondent answered four items such that his final staircase outcome
would have to be either x or y. We then predict the expected choice between x
and y based on a probit of the “x vs. y” decision on the qualitative item. If the
respondent answered three (or less) questions, the same procedure was applied,
the only difference being that in this case the obtained predicted probabilities
were applied to the expected values of the staircase outcome conditional on
reaching the respective node. Put differently, the procedure outlined above
was applied recursively by working backwards through the “tree” logic of the
staircase procedure.
38
A
A
Risk taking=32
B
Risk taking=31
A
Risk taking=30
B
Risk taking=29
A
Risk taking=28
B
Risk taking=27
A
Risk taking=26
B
Risk taking=25
A
Risk taking=24
B
Risk taking=23
A
Risk taking=22
B
Risk taking=21
A
Risk taking=20
B
Risk taking=19
A
Risk taking=18
B
Risk taking=17
A
Risk taking=16
B
Risk taking=15
A
Risk taking=14
B
Risk taking=13
A
Risk taking=12
B
Risk taking=11
A
Risk taking=10
B
Risk taking=9
A
Risk taking=8
B
Risk taking=7
A
Risk taking=6
B
Risk taking=5
A
Risk taking=4
B
Risk taking=3
A
Risk taking=2
B
Risk taking=1
310
300
B
A
290
280
B
A
A
270
260
B
250
240
A
220
B
B
A
A
230
210
200
B
A
190
180
B
170
160
A
150
140
B
A
130
B
120
B
A
A
110
100
B
90
80
A
70
60
B
B
A
50
40
B
A
30
20
B
10
Figure 3: Tree for the staircase time task (numbers = payment in 12 months, A = choice of “100
euros today”, B = choice of “x euros in 12 months”). The staircase procedure worked as follows.
First, in Germany, each respondent was asked whether they would prefer to receive 160 euros
for sure or whether they preferred a 50:50 chance of receiving 300 euros or nothing. In case the
respondent opted for the safe choice (“B”), the safe amount of money being offered in the second
question decreased to 80 euros. If, on the other hand, the respondent opted for the gamble (“A”),
the safe amount was increased to 240 euros. Working further through the tree follows the same
logic.
39
A.6.2
Computation of Preference Indices at Individual Level
We compute an individual-level index of patience by (i) computing the z-scores of
each survey item at the individual level and (ii) weighing these z-scores using the
weights resulting from the experimental validation procedure of Falk et al. (2015a).
Technically, these weights are given by the coefficients of an OLS regression of observed behavior on responses to the respective survey items, such that the coefficients
sum to one. These weights are given by (see above for the precise survey items):
Risk taking = 0.4729985 × Staircase risk + 0.5270015 × Qualitative item
A.6.3
Computation of Country Averages
In order to compute country-level averages, we weigh the individual-level data with
the sampling weights provided by Gallup, see above.
B
Summary Statistics for Main Variables
Table 7: Summary statistics of main unilateral variables
Obs
Mean
SD
Min
Max
73
73
73
0.01
0.94
0.07
0.31
0.09
0.04
-0.79
0.75
0.01
0.97
1.18
0.19
Average risk aversion
SD in risk aversion
Predicted migratory distance from Ethiopia
Table 8: Summary statistics of main bilateral variables
Abs. difference in avg. risk aversion
Fst genetic distance
Nei genetic distance
Predicted migratory distance
HMI migratory distance
Linguistic distance
40
Obs
Mean
2628
2628
2628
2628
2080
2628
0.34
0.34
0.30
0.08
0.17
0.87
SD
Min
Max
0.27
0
0.24
0
0.23
0
0.04
0
0.10 0.01
0.20
0
1.76
1
1
0.23
0.59
1
0
.5
Density
1
1.5
Average risk aversion
-1
-.5
0
Average risk aversion
.5
1
kernel = epanechnikov, bandwidth = 0.0908
SD in risk aversion
Density
0
0
.5
.5
Density
1
1
1.5
1.5
2
Abs. difference in average risk aversion
0
.5
1
1.5
Abs. difference in average risk aversion
2
-1
kernel = epanechnikov, bandwidth = 0.0488
-.5
0
SD in risk aversion
.5
1
kernel = epanechnikov, bandwidth = 0.0908
Predicted migratory distance from Ethiopia
Density
10
0
0
5
.5
Density
1
15
1.5
2
20
Fst genetic distance
0
.2
.4
.6
Fst genetic distance
.8
1
0
kernel = epanechnikov, bandwidth = 0.0454
.05
.1
.15
Predicted migratory distance from Ethiopia
kernel = epanechnikov, bandwidth = 0.0067
Figure 4: Kernel density estimates of main variables
41
.2
Table 9: Raw correlations among bilateral variables
∆ average risk attitudes
Fst genetic distance
Nei genetic distance
Predicted migratory distance
HMI migratory distance
Linguistic distance
C
C.1
∆ Risk
1
0.343
0.358
0.111
0.249
0.147
Fst dist Nei dist
1
0.945
0.517
0.623
0.457
1
0.509
0.634
0.408
Pred migr dist
HMI dist
Ling. dist
1
0.922
0.105
1
0.210
1
Additional Bilateral Regressions
Alternative Dependent Variable Specifications
This Appendix provides additional robustness checks on our results pertaining to
average risk aversion. Column (1) of Table 10 restates our full regression model,
i.e., the results of column (5) in Table 3. Columns (2) through (4) employ the
same regression specification, yet use the absolute difference in the 25th, 50th, and
75th percentile of risk aversion as dependent variable. Interestingly, the results show
that the relationship between genetic distance and risk attitudes is not specific to
the mean of the distribution of risk preferences, but extends to other parts of the
distribution.
Recall that our main measure of a country’s average risk aversion consists of
a weighted average of two survey questions, a hypothetical staircase lottery task,
and a qualitative question. Columns (5) through (10) demonstrate that our result
extends to using each of these survey questions separately. Specifically, columns (5)
and (6) employ the absolute difference in the average qualitative self-assessment
as dependent variable, while columns (7) through (8) use the average outcome of
the five-step lottery choice procedure. Finally, columns (9) and (10) present two
regressions which utilize only the first choice in the sequential lottery procedure, i.e.,
(in Germany) the choice between 160 euros for sure and a 50-50 bet of receiving 300
euros or nothing. While this measure ignores a lot of information that is contained
in subsequent choices, the corresponding results show that even according to this
very rough measure of risk attitudes does genetic distance explain about 10 % of the
variation in preferences.
C.2
Additional Control Variables
Column (1) of Table 11 reproduces our full between-country regression including all
control variables (i.e., column (5) of Table 3). Columns (2)-(4) introduce additional
42
covariates into the analysis, but the overall picture remains virtually unchanged.
Column (5) presents an additional, very conservative, specification which employs
country fixed effects rather than regional fixed effects. Specifically, we construct
73 dummies each equal to one if a given country is part of a country pair and zero
otherwise. While the resulting point estimate is somewhat lower than in the previous
regressions, it is still significant and features a standardized beta of 18.2 %.
43
44
Yes
Yes
2556
0.329
23.2
Geographic, climate and health controls
Continent FE
Observations
R2
Standardized beta (%)
2556
0.229
25.8
Yes
Yes
Yes
-0.64
(0.89)
0.33
(0.12)
∗∗∗
(2)
2556
0.292
22.5
Yes
Yes
Yes
2556
0.373
18.4
Yes
Yes
Yes
0.30
(1.52)
0.74∗∗∗
(0.08)
2628
0.036
18.9
No
No
2556
0.146
19.3
Yes
Yes
Yes
0.53
(0.21)
No
∗∗
(6)
∗∗∗
(5)
0.24
0.52
(0.10) (0.14)
∗∗
(4)
-0.50 -0.26
(0.68) (0.62)
0.28
(0.11)
∗∗
(3)
2628
0.066
25.6
No
No
No
2.69∗∗∗
(0.34)
3.02
(1.03)
∗∗∗
(7)
2556
0.325
15.9
Yes
Yes
Yes
-4.57
(4.65)
1.83
(0.89)
∗∗
(8)
2628
0.098
31.3
No
No
No
0.094∗∗∗
(0.02)
0.15
(0.05)
∗∗∗
(9)
2556
0.376
19.1
Yes
Yes
Yes
-0.30
(0.19)
0.091∗∗
(0.04)
(10)
Dependent variable: Absolute difference in...
Percentiles
Self-assessment
Lottery choices
25th
50th
75th
(average)
(average outcome) (average 1st choice)
OLS estimates, twoway-clustered standard errors in parentheses. The dependent variables consist of different measures of the absolute
difference in risk attitudes: (1) averages, (2) - (4) percentiles, (5) - (6) average (qualitative) self-assessments, (7) - (8) average outcome of
the staircase lottery sequence, (9) - (10) average decision in first lottery choice (safe amount = 160 euros). See Tables 2 and 3 for a complete
list of the demographic, economic, geographic, and climatic control variables. The standardized betas refer to genetic distance. ∗ p < 0.10,
∗∗
p < 0.05, ∗∗∗ p < 0.01.
Yes
-0.34
(0.63)
Constant
Demographic and economic controls
0.26
(0.10)
Fst genetic distance
∗∗∗
(1)
Average
(baseline)
Table 10: Alternative dependent variables
Table 11: Additional control variables
Dependent variable:
Absolute difference in average risk aversion
(1)
(2)
(3)
(4)
(5)
Fst genetic distance
0.26∗∗∗
(0.10)
Freight cost (surface transport)
0.28∗∗∗
(0.10)
0.32∗∗∗
(0.11)
0.00018 0.00018
(0.00)
(0.00)
∆ Landlocked
0.28∗∗
(0.11)
0.20∗∗
(0.10)
-0.000033
(0.00)
0.00011
(0.00)
-0.033
(0.03)
-0.022
(0.03)
-0.024
(0.03)
-0.0050
(0.02)
0.00027
(0.02)
-0.0060
(0.02)
-0.014
(0.02)
0.031
(0.06)
∆ Traffic death rate
0.0015
(0.00)
0.00011
(0.00)
0.0011
(0.00)
∆ Linguistic fractionalization
0.088
(0.05)
0.13∗∗
(0.05)
0.084
(0.06)
0.0063
(0.01)
0.0014
(0.00)
0.0040∗∗∗
(0.00)
-0.00016
(0.00)
-0.0052
(0.00)
-0.0051
(0.01)
∆ Mean distance to nearest waterway
∆ Executive constraints
∆ Gini coefficient
∆ Years of schooling
Constant
-0.34
(0.63)
-0.29
(0.65)
-0.46
(0.68)
-0.30
(0.68)
0.50
(0.31)
Demographic and economic controls
Yes
Yes
Yes
Yes
Yes
Geographic, climate and health controls
Yes
Yes
Yes
Yes
Yes
Continent FE
Yes
Yes
Yes
Yes
No
Country FE
No
No
No
No
Yes
2556
0.329
23.2
2556
0.331
24.6
2346
0.359
27.3
1953
0.415
25.5
1953
0.671
18.2
Observations
R2
Standardized beta (%)
OLS estimates, twoway-clustered standard errors in parentheses. See Tables 2 and 3 for a
complete list of the demographic, economic, geographic, and climatic control variables. The
standardized beta refers to genetic distance. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
D
Details for WLS Regressions
Following Spolaore and Wacziarg (2009), the regression weights in Table 10 were
computed as follows:
Weight =
Maximum standard error + 1 − standard error of observation
Maximum standard error
Thus, the regression weights are between zero and one and linearly downweigh
45
observations with a high standard error for genetic distance.
E
Details for Analysis of Heterogeneity Within Countries (Table 6)
Tables 12 and 13 provide the details for columns (1) through (8) of Table 6.
Table 12: Correlates of standard deviation in risk attitudes
Dependent variable:
SD in risk aversion
(2)
(3)
(1)
∗
Migratory distance from Ethiopia
-0.52
(0.28)
-0.83
(0.32)
-0.52∗∗
(0.21)
HMI migratory distance from Ethiopia (Ozak, 2010)
Will. to take risks
(4)
∗∗
-0.77∗∗∗
(0.26)
-0.032
(0.04)
-0.040
(0.04)
SD in age
0.00084
(0.01)
0.0041
(0.01)
Average age
0.0040
(0.01)
0.0047
(0.01)
Fraction female
-0.44
(0.77)
0.56
(0.89)
Gender heterogeneity index
31.4∗∗∗
(8.38)
1.41
(14.20)
Log [GDP p/c PPP]
0.19∗∗
(0.08)
0.18∗∗
(0.08)
-0.013∗∗
(0.01)
-0.013∗∗
(0.01)
SD in log [HH income]
0.17∗∗
(0.06)
0.12∗
(0.07)
Log [Population]
0.0018
(0.01)
0.0077
(0.01)
Linguistic diversity
0.049
(0.05)
0.053
(0.05)
Life expectancy
0.0022
(0.00)
0.0025
(0.00)
Log [GDP p/c PPP] squared
Constant
0.98∗∗∗
(0.02)
0.11
(0.42)
1.00∗∗∗
(0.02)
-0.47
(0.56)
73
0.048
73
0.380
64
0.074
64
0.412
Observations
R2
OLS estimates, robust standard errors in parentheses.
46
∗
p < 0.10,
∗∗
p < 0.05,
∗∗∗
p < 0.01.
Table 13: Correlates of standard deviation in risk attitudes
Dependent variable:
SD in risk aversion
(2)
(3)
(1)
∗
Migratory distance from Ethiopia
-0.64
(0.36)
(4)
-0.15
(0.47)
HMI migratory distance from Ethiopia (Ozak, 2010)
-0.73∗∗
(0.29)
-0.80∗
(0.45)
-0.00061
(0.00)
0.00048
(0.00)
Distance to equator
0.00089
(0.00)
0.0022
(0.00)
Average precipitation
-0.00040
(0.00)
-0.00035
(0.00)
Average temperature
-0.00080
(0.00)
0.0013
(0.00)
-0.0021
(0.00)
-0.0012
(0.00)
Log [Timing neolithic revolution]
-0.091∗
(0.05)
-0.088∗
(0.05)
-0.17∗∗∗
(0.05)
-0.15∗∗
(0.06)
% at risk of malaria
-0.081
(0.06)
-0.076
(0.06)
-0.094
(0.06)
-0.080
(0.08)
Land suitability for agriculture
0.043
(0.05)
0.017
(0.06)
0.032
(0.06)
0.024
(0.08)
Constant
1.16∗∗
(0.52)
0.80
(0.62)
1.36∗∗
(0.62)
1.51∗∗
(0.70)
Socioeconimic controls
Yes
Yes
Yes
Yes
Continent FE
No
Yes
No
Yes
Observations
R2
73
0.509
73
0.568
64
0.578
64
0.595
OLS estimates, robust standard errors in parentheses.
47
∗
p < 0.10,
∗∗
-0.00038 -0.00028
(0.00)
(0.00)
p < 0.05,
∗∗∗
p < 0.01.
1.2
Preference heterogeneity and migratory distance to Ethiopia
GEO
EGY
BOL
PER
.7
.8
SD in risk aversion
.9
1
1.1
CHN
KAZMAR
MWI
ZWE UKR
MEX
IRN
HUN
RUS
CZE
LTU
TZA
NGA
BGD VNM
MDA
GTM
ZAF
ESTGHA
RWA
KEN
CHL
DZA
POL
THA
ARG
JOR TUR
ESP
FINSWE USA
CRI COL
ROU
AFG
CHE
UGA
PAK
BRA VEN
IND
DEU
KHM
SAU
IRQ
BWA AUT
GRC FRA
ISR
CMR
CAN
NIC
HRV
KOR
NLD
ITA
0
.05
.1
.15
.2
Mobility index-based migratory distance from Ethiopia
Sub-Saharan Africa
South Asia
ME & North Africa
South America
EU & Central Asia
North America
.25
East Asia
Figure 5: Migratory distance from South Africa and preference heterogeneity
F
Relationship Between Risk Preferences and Outcomes
F.1
Average Risk Aversion and Riskiness of the Environment
Table 14 presents correlations between the average level of risk aversion in a given
country and the riskiness of the environment, as proxied for by life expectancy at
birth, HIV prevalence, the traffic death rate, a poverty ratio, the Gini index, the
fraction of income that goes to the top 10%, and a labor protection index. Also see
Falk et al. (2015a).
48
49
70.6∗∗∗
(0.96)
76
0.158
34.7∗∗∗
(3.13)
76
0.692
4.32∗∗∗
(0.34)
-0.86∗∗∗
(0.17)
69
0.150
1.62∗
(0.89)
69
0.245
-0.30∗∗∗
(0.10)
15.3∗∗∗
(0.92)
74
0.111
39.5∗∗∗
(3.49)
74
0.403
-2.90∗∗∗
(0.39)
Log [HIV prevalence] Traffic death rate
(3)
(4)
(5)
(6)
1.96∗∗
1.78∗∗
9.37∗∗∗ 7.66∗∗∗
(0.74)
(0.74)
(2.78)
(2.88)
13.3∗∗∗
(2.08)
50
0.077
96.8∗∗∗
(9.96)
50
0.692
-11.1∗∗∗
(1.23)
Poverty ratio
(7)
(8)
14.0∗
8.90∗∗
(7.83)
(4.23)
∗
∗∗
31.0∗∗∗
(0.86)
68
0.129
p < 0.10,
38.6∗∗∗
(8.74)
52
0.093
0.35
(1.14)
p < 0.05,
∗∗∗
45.2∗∗∗
(3.97)
68
0.252
-1.71∗∗∗
(0.44)
0.49∗∗∗
(0.02)
56
0.135
0.37∗∗∗
(0.14)
56
0.148
0.015
(0.02)
Labor protection
(13)
(14)
-0.24∗∗∗ -0.21∗∗
(0.08)
(0.08)
p < 0.01.
Income share of top 10%
(11)
(12)
9.59∗∗
9.00∗∗
(3.94)
(4.20)
Economic risks
Gini index
(9)
(10)
9.45∗∗ 9.58∗∗
(4.46) (4.53)
41.2∗∗∗
(1.23)
52
0.091
Dependent variable:
OLS estimates, robust standard errors in parentheses. Labor protection is an index taken from Botero et al. (2004).
Observations
R2
Constant
Log [GDP p/c PPP]
Risk taking
Life expectancy
(1)
(2)
-12.3∗∗∗ -9.69∗∗∗
(4.01)
(3.34)
Health risks
Table 14: Average risk taking and riskiness of the environment
F.2
Variability in Risk Attitudes and Economic Development
Table 15 investigates the relationship between the variability in risk attitudes and
per capita income. Across the different specifications, the standard deviation in risk
attitudes is significantly related to income, even conditional on institutions and education. In fact, the relationship becomes even stronger when we focus on the standard
deviation of the quantitative (lottery-choice) risk measure only. The corresponding
raw correlation is ρ = −0.57 (p < 0.01) and is illustrated in Figure 6.
CHE
USA
SWE CAN
NLD
GBR
FIN
DEUAUT
FRA
ITA JPN AUS
ESP
GRC ISR
KOR
PRT
ARE
SAU
CZE
HUN
EST
POL
MEX
LTU
CHL
TUR
BWA
ZAF
BRA CRI RUS
ROU ARGVEN
COL
SRB DZA KAZSUR
PER BIH
THAIRN
JOR
GTMIRQ
MAR
CHN
UKR
GEO
EGY
IDN
LKA
PHL
NIC
BOL
CMR
MDA
NGA
IND
PAK VNM
KEN
GHA
KHM
ZWE
HTI
BGD
UGA
AFG
HRV
RWA
MWI
TZA
4
6
Log [GDP p/c PPP]
8
10
12
SD in quantitative risk measure and per capita income
8
10
12
SD in quantitative risk measure
14
16
Figure 6: Variation in risk attitudes (using the quantitative lottery choice measure only) and
development.
50
Table 15: Variability in risk preferences and development
(1)
(6)
-5.27∗∗∗
(1.61)
-5.41∗∗∗
(1.34)
-4.62∗∗∗
(1.20)
-5.00∗∗∗
(0.96)
-2.87∗∗∗
(0.86)
-0.48
(0.56)
0.89∗
(0.47)
0.22
(0.34)
-0.0085
(0.34)
0.087
(0.35)
Distance to equator
0.034
(0.02)
-0.012
(0.02)
-0.021
(0.01)
Log [Timing neolithic revolution]
-0.057
(0.39)
0.36
(0.44)
0.24
(0.27)
Average temperature
0.059∗∗
(0.03)
0.023
(0.03)
0.015
(0.02)
Average precipitation
0.0044
(0.00)
-0.0045
(0.00)
-0.0038
(0.00)
% at risk of malaria
-2.01∗∗∗
(0.54)
-2.23∗∗∗
(0.60)
-1.64∗∗∗
(0.42)
% living in (sub-)tropical zones
-0.78
(0.73)
-0.069
(0.65)
-0.36
(0.46)
Land suitability for agriculture
-0.73
(0.54)
-0.82
(0.57)
-0.12
(0.37)
Predicted genetic diversity
559.2∗∗∗
(162.10)
477.6∗∗∗
(138.91)
Predicted genetic diversity sqr.
-408.5∗∗∗
(118.63)
-346.0∗∗∗
(101.79)
Ethnic fractionalization
-2.36∗∗∗
(0.64)
-0.48
(0.60)
0.68
(0.45)
0.13
(0.48)
SD in risk taking
-5.39∗∗∗
(1.62)
Dependent variable: Log [GDP p/c PPP]
(2)
(3)
(4)
(5)
Average risk taking
Linguistic fractionalization
0.10∗∗
(0.04)
Average years of schooling
0.022∗∗∗
(0.00)
Property rights
13.4∗∗∗
(1.58)
13.3∗∗∗
(1.58)
15.3∗∗∗
(1.22)
11.7∗∗∗
(3.91)
-176.8∗∗∗
(55.32)
-155.1∗∗∗
(48.00)
Continent FE
No
No
Yes
Yes
Yes
Yes
Observations
R2
Adjusted R2
76
0.100
0.087
76
0.108
0.083
76
0.613
0.566
74
0.756
0.693
72
0.813
0.745
67
0.927
0.893
p < 0.10,
∗∗
Constant
OLS estimates, robust standard errors in parentheses.
51
∗
p < 0.05,
∗∗∗
p < 0.01.
G
G.1
Definitions and Data Sources of Main Variables
Explanatory Variables
Fst and Nei genetic distance. Genetic distance between contemporary populations, taken from Spolaore and Wacziarg (2009).
Linguistic distance. Weighted linguistic distance between contemporary populations. Derived from the Ethnologue project data, taking into account all languages
which are spoken by at least 5% of the population in a given country.
Predicted migratory distance. Predicted migratory distance between two countries’ capitals, along a land-restricted way through five intermediate waypoints (one
on each continent). Taken from Ashraf and Galor (2013b).
HMI migratory distance. Walking time between two countries’ capitals in years,
taking into account topographic, climatic, and terrain conditions, as well as human
biological abilities. Data from Özak (2010).
G.2
Covariates and Outcome Variables
Average age, proportion female, standard deviation age, gender heterogeneity index, standard deviation of household income. Computed from
Gallup’s sociodemographic background data. Gender heterogeneity is computed as
(2 × proportion female − 1)2 .
Religious and linguistic fractionalization. Indices due to Alesina et al. (2003)
capturing the probability that two randomly selected individuals from the same
country will be from different religious / linguistic groups.
Percentage of European descent. Constructed from the “World Migration Matrix” of Putterman and Weil (2010).
Contemporary national GDP per capita. Average annual GDP per capita over
the period 2001 – 2010, in 2005US$. Source: World Bank Development Indicators.
Democracy index. Index that quanties the extent of institutionalized democracy,
as reported in the Polity IV dataset. Average from 2001 to 2010.
52
Property rights. This factor scores the degree to which a country’s laws protect
private property rights and the degree to which its government enforces those laws.
It also accounts for the possibility that private property will be expropriated. In
addition, it analyzes the independence of the judiciary, the existence of corruption
within the judiciary, and the ability of individuals and businesses to enforce contracts.
Average 2001-2010, taken from the Quality of Government dataset, http://www.
qogdata.pol.gu.se/codebook/codebook_basic_30aug13.pdf.
Colonial relationship dummies. Taken from the CEPII Geodist database at
http://www.cepii.fr/CEPII/en/bdd_modele/presentation.asp?id=6.
Geodesic distance, contiguity, longitude, latitude, area, 1 if landlocked
Taken from CEPII GeoDist database. The longitudinal distance between two countries is computed as
Longitudinal distance = min{|longitudei −longitudej |, 360−|longitudei |−|longitudej |}
Suitability for agriculture. Index of the suitability of land for agriculture based
on ecological indicators of climate suitability for cultivation, such as growing degree
days and the ratio of actual to potential evapotranspiration, as well as ecological
indicators of soil suitability for cultivation, such as soil carbon density and soil pH,
taken from Michalopoulosa (2012).
Percentage of arable land. Fraction of land within a country which is arable,
taken from the World Bank Development Indicators.
Terrain roughness. Degree of terrain roughness of a country, taken from Ashraf
and Galor (2013b). Data originally based on geospatial undulation data reported by
the G-ECON project (Nordhaus, 2006).
Mean and standard deviation of elevation. Mean elevation in km above sea,
taken from Ashraf and Galor (2013b). Data originally based on geospatial elevation
data reported by the G-ECON project (Nordhaus, 2006).
Timing of the neolithic transition. The number of thousand years elapsed,
until the year 2000, since the majority of the population residing within a country’s
modern nationalw borders began practicing sedentary agriculture as the primary
mode of subsistence. The measure is weighted within each country, where the weight
represents the fraction of the year 2000 population (of the country for which the
53
measure is being computed) that can trace its ancestral origins to the given country
in the year 1500. Measure taken from Ashraf and Galor (2013b).
Precipitation. Average monthly precipitation of a country in mm per month,
1961-1990, taken from Ashraf and Galor (2013b). Data originally based on geospatial
average monthly precipitation data for this period reported by the G-ECON project
(Nordhaus, 2006).
Temperature. Average monthly temperature of a country in degree Celsius, 19611990, taken from Ashraf and Galor (2013b). Data originally based on geospatial
average monthly temperature data for this period reported by the G-ECON project
(Nordhaus, 2006).
Life expectancy. Life expectancy at birth, average 2001–2010. Taken from the
World Bank Indicators.
Percentage at risk of contracting malaria. The percentage of population in
regions of high malaria risk (as of 1994), multiplied by the proportion of national cases
involving the fatal species of the malaria pathogen, P. falciparum. This variable was
originally constructed by Gallup et al. (2000) and is part of Columbia University’s
Earth Institute data set on malaria.
Freight cost of surface transport Cost of 1,000 kg of unspecified freight transported over sea or land with no special handling, taken from Spolaore and Wacziarg
(2009). Data originally retrieved from http://www4.importexportwizard.com/.
Distance to nearest waterway. The distance, in thousands of km, from a GIS
grid cell to the nearest ice-free coastline or sea-navigable river, averaged across the
grid cells of a country. Source: Ashraf and Galor (2013b), originally constructed by
Gallup et al. (1999).
Executive constraints. The 1960-2000 mean of an index (1-7) that quantifies the
extent of institutionalized conatrsints on the executive, as reported in the Polity IV
data set. Data taken from Ashraf and Galor (2013b).
Average years of schooling. The mean over the 2000-2010 time period, of the
5-yearly figure, reported by Barro and Lee (2012), on average years of schooling
amongst the population aged 25 and over.
54
Population. Average 2001-2010, taken from the World Bank Development Indicators.
Linguistic diversity. Computed akin to linguistic and religious distance indices
by matching a country with itself.
HIV prevalence. Primary data taken from Unicef. Missings replaced by data
from most recent available period of the World Bank Development Indicators.
Traffic death rate. Traffic deaths per 1,000 population, estimate of the World
Health Organization.
Poverty headcount ratio. Fraction of the population living on less than $2 per
day, average 2001-2010. Taken from World Bank Development Indicators.
Gini index. Average 2001-2010. Taken from World Bank Development Indicators.
Income share of top 10%. Average 2001-2010. Taken from World Bank Development Indicators.
Labor protection index. Index capturing the rigidity of employment laws by
Botero et al. (2004). Includes data on employment, collective relations, and social
security laws and measures legal worker protection.
55