ABSTRACT
This study explores the relationship between linguistic variation and individual attitudes toward risk and uncertainty. We propose an innovative marker that classifies languages according to the number of nonindicative moods in the grammatical contexts involving uncertainty. We find that speakers of languages that use these moods more intensively are on average more risk averse. Our marker is then used to instrument risk aversion in the model for financial asset accumulation. In addition, by using the Chen (2013) future time reference linguistic marker as a proxy for the subjective discount rate, we disentangle the effects of risk aversion and time preferences on asset accumulation.
I. Introduction
Consider a situation in which you have to make a decision about something that has an uncertain prospect, for example, related to sports, health, or financial choices. Your personal characteristics and preferences have an impact on how you evaluate the potential outcomes, but would you also argue that the language you speak might influence the way you perceive risk? In this work, we propose an innovative approach to analyze individual attitudes toward uncertainty and risky behavior based on the hypothesis of linguistic relativity. The concept stems from the idea that differences in grammatical structures and the vocabulary may induce speakers of distinct languages to experience the world differently (Hill and Mannheim 1992). Research in support of this hypothesis has mainly focused on conceptual contents of languages linking, for instance, individual perceptions of space, color, and even cross-country differences in gender political quota, as well as female labor force participation to specific linguistic features (see Davies and Corbett 1997; Gay et al. 2018; Majid et al. 2004; Roberson, Davidoff, and Braisby 1999; Santacreu-Vasut, Shoham, and Gay 2013; Winawer et al. 2007).
If speakers of different languages vary in their worldview depending on the language they use, some dimensions of linguistic structures may also shape individual preferences and their economic decision-making. In a recent paper on the effect of language on economic behavior, Chen (2013) tests a linguistic-savings hypothesis: when people are grammatically required to speak in a distinct way about future events, they take fewer future-oriented actions. The author adopts a criterion that separates languages into two broad categories: weak and strong future time reference (FTR henceforth) according to how they induce speakers to mark the timing of events. Some languages require an explicit verb conjugation in order to distinguish between present and future events (strong FTR languages), while others allow their speakers to talk about the future by using the same verb form as for present events (weak FTR languages). The author then examines how these differences correlate with future-oriented behavior and finds that speakers of weak FTR languages save more, accumulate more wealth by retirement, smoke less frequently, and are more physically active. This evidence remains reasonably robust even after controlling for geographic and historical relatedness of languages (Roberts, Winters, and Chen 2015). On the other hand, Galor and Özak (2016) and Galor, Özak, and Sarid (2017) argue that ancestral characteristics from the parental country of origin might have affected the formation of time preferences and triggered the gradual emergence of grammatical forms that fostered the transmission of these traits across generations. Indeed, these authors show that higher pre-industrial crop yield potential experienced by ancestral populations had a positive impact on the descendants’ future orientation. Nevertheless, the language still had an independent effect on time preferences and economic behavior.
Our approach is conceptually in line with Chen (2013) because we rely on a weak version of the linguistic relativity hypothesis. However, it departs from Chen (2013) for two reasons. First, we propose consideration of the linguistic relativity hypothesis based on a different grammatical property and in a different economic context, namely mood and uncertainty. We develop a new linguistic marker based on the number of grammatical contexts concerned with the expression of uncertainty in which specific nonindicative moods (such as subjunctive and conditional) are used. Since indicative moods are generally used to assert that a fact or a situation is true as of the actual world, we conjecture that the perceived degree of uncertainty is larger with a nonindicative mood compared to an indicative one. Therefore, based on a weak version of the linguistic relativity hypothesis, speakers of languages where these specific grammatical forms are used more often should experience the world as being more mutable and uncertain compared to speakers of languages where these forms are less frequent. Our mapping offers a rigorous and, to the best of our knowledge, the first linguistic mapping related to the grammatical treatment of uncertainty. Second, we analyze the correlation between our linguistic marker and individual self-declared risk aversion and use the marker to instrument the individual attitudes toward risk in order to quantify a direct effect of risk aversion on the probability of holding risky financial assets. In addition, we estimate the separate effects of time preferences and risk attitudes by using the FTR parameter as a proxy for intertemporal choice behavior, alongside our instrument for individual risk preferences.
Using data on a subpopulation of second-generation immigrants from the Survey of Health, Ageing and Retirement in Europe (SHARE) and the European Social Survey (ESS), we show that a more intensive use of nonindicative moods in grammatical contexts involving uncertainty strongly correlates with individual risk preferences. The likelihood of individuals being risk-averse increases with the frequency of use of these forms in their respective languages. This evidence is robust to the inclusion of a rich set of controls related to parental linguistic background and parental ancestral characteristics from Galor and Özak (2016) and Becker, Enke, and Falk (2020). Intensive users are on average 15–24 percent more likely to report a high level of risk aversion compared to individuals speaking languages where these forms are used less frequently or where they are not required at all. The effect of parental linguistic background remains persistent with the effect of the maternal native language being more pronounced compared to the paternal one. This evidence holds for second-generation immigrants coming from culturally heterogeneous couples. Also, the inclusion of parental linguistic background reduces the effect of the local language but does not alter its statistical and economic significance.
As for the separate effects of risk and time preferences on the propensity to invest in financial risky assets, we find that risk-averse individuals are, on average, 16 percent less likely to hold stocks or bonds compared to intermediate risk-takers and to individuals with a low level of risk aversion. When compared to individual time preferences, the effect of risk aversion is five times larger than the effect of the individual discount rate.
In the next section, we introduce the issue of linguistic relativity and mood. In Section III, we exploit the relationship between our linguistic marker and individual attitudes toward risk and estimate a direct effect of risk aversion and time preferences on the probability of investing in risky financial assets. Section IV concludes.
II. Linguistic Relativity and Economic Behavior
The idea that language categories can influence thought has come to be known as the Sapir–Whorf hypothesis, after Sapir (1921) and Whorf and Carroll (1964), and boasts a long history in the philosophy of language and linguistics, which can be traced back at least to von Humboldt’s (1836) idea of Innere Sprachform. Following Geeraerts and Cuyckens (2010), the hypothesis of linguistic relativity encompasses two basic notions—the first is that languages are relative as they vary in their expression of concepts, and the second is that the semantic expression of concepts influences, at least to some extent, conceptualization at the cognitive level. Therefore, speakers of distinct languages may perceive reality differently.
The linguistic relativity hypothesis has generally been interpreted according to two versions. The strong one, also known as linguistic determinism, states that linguistic categories control general cognitive processes. This version of the hypothesis, however, has generally been refuted (Pinker 1994). The weak version claims that linguistic categories have some effect on cognitive habits, particularly with respect to memory and categorization. The latter version of the Sapir–Whorf hypothesis was taken to be more feasible and has inspired research on topics such as color perception, shape classification, conditional reasoning, and number, space, and time categorization.
If speakers of different languages tend to think and behave differently depending on the language they use, some dimensions of linguistic structures may also shape individual preferences. Chen (2013) represents the first attempt to analyze the impact of language differences on the cognitive domain and consequently on several aspects of individual economic behavior. The empirical analysis in Chen (2013) uses a typological distinction discussed in Dahl (2000) and Thieroff (2000) whereby there are languages that employ specific verb morphology for FTR, whereas other languages do not. By adopting the weak version of the Sapir–Whorf hypothesis, Chen (2013) hypothesized that this typological divide affects how speakers conceive time. Specifically, speakers of languages that separate the future from the present tense (strong FTR languages) are more prone to dissociate the future from the present compared to speakers of languages that do not employ that specific verb morphology when referring to future events (weak FTR or “futureless” languages). As a consequence, this may induce people to perceive the future as being more distant and, hence, to undertake fewer future-oriented actions, such as saving, smoking, using condoms, accumulating wealth before retirement, and taking initiatives to enhance long-run health. The association between weak FTR and future-oriented behavior in Chen (2013) is strong—speakers of weak FTR languages save more, accumulate more wealth by retirement, smoke less frequently, and are more physically active.
We propose a reconsideration of the Chen (2013) idea of linking language features to economic behavior through the linguistic relativity hypothesis. Following a weak version of the hypothesis, we conjecture that individual levels of risk aversion are influenced by differences in the intensity of use of indicative versus nonindicative (irrealis) moods as they assign a different degree of uncertainty to possible situations. In other words, when describing possible or hypothetical situations, the displacement of the actual from the alternative state of facts is perceived as larger when an irrealis mood is used. According to this conjecture, in the two sentences below, for example, the leaving event should be perceived as less uncertain by an English speaker than by an Italian speaker, even though they describe the same possible situation.
The former expresses the leaving situation by resorting to the indicative mood, while the latter has to use a subjunctive (irrealis) mood. In general, by using irrealis more intensively, speakers move from the region of certainty to that of uncertainty; in other words, their latent area of the unknown is greater with respect to their peers who speak a less irrealis-intensive language. As a consequence, they are expected to be more risk averse, as the semantic salience of their region of uncertainty increases.
Within the economics literature, the discussion of the dichotomy between risk and uncertainty has a long-standing history. Knight (1921) defines risk as “measurable uncertainty,” or a perceived likelihood, which can be represented by probabilities. What he calls uncertainty, on the other hand, describes a situation in which information about probabilities is imprecise or even unavailable. To overcome the limitations of the Knightian juxtaposition, Ellsberg (1961) introduces the concept of ambiguity for “unmeasurable uncertainty,” or unknown risks, substituting the Knightian interpretation of the term uncertainty with a smoother concept according to which the degree of uncertainty becomes a function of ambiguity. In light of these considerations, we interpret risk aversion as the preference for certain outcomes over probabilistic outcomes, within the domain of the aforementioned “measurable uncertainties.” An underlying assumption to be made is that decision-makers are able to assign objective or subjective probabilities to uncertain outcomes (Abdellaoui et al. 2011; Epstein and Zhang 2001; Machina and Schmeidler 1992). In that sense, while we assume that the use of the irrealis mood makes situations with variable outcomes generally more notable, we restrict our measurement of risk aversion to unambiguous prospects for which subjective probabilities can be formed.
For this purpose, we develop a specific linguistic marker defined on the number of nonindicative moods used in irrealis contexts, that is, contexts that involve grammatical categories concerned with the expression of uncertainty, and we relate it to the individual’s perception of risk and risky behavior. In what follows we describe the definitions of displacement, modality, and mood in more depth and provide some applied examples and contexts that define our linguistic marker.
A. Displacement and Modality
By displacement, semanticists mean the specific characteristic of human language whereby language expressions not only refer to the here and now, but are able to range over future, past, potential, possible, and even impossible situations (Hockett 1960; Hockett and Altmann 1968). In that sense, futurity is an instance of displacement within the temporal dimension. Another crucial dimension of displacement is modality, the grammatical category that indicates whether a sentence expresses a fact, a command, a condition, an opinion, or a desire. Consider for instance the following sentences:1
By observing Sentences 1, 3, and the embedded clause in Sentence 2, we notice that they do not describe actual facts. The truth or falsity of the expressions cannot be decided simply by considering whether the state of facts described in the sentences is true (or false). Sentence 1 does not assert that it is sunny and that the speaker is having a walk; Sentence 2 does not assert that the meeting is finished. It may be finished, and the speaker probably believes that it has, but one’s belief may turn out to be wrong when actual states of facts are taken into consideration. Sentence 3 does not assert that the speaker is at home having a coffee with the hearer. Sentences 1–3 do not refer to actual facts, differently from sentences like, “It is sunny today,” “I am having a walk,” “The meeting has finished,” and “I’m having a coffee at home with a friend.” They refer to possible situations or “possible worlds” (Carnap 1947), not to real ones. Possible worlds represent alternative states of facts, which cannot be asserted as of the world we actually live in (the “actual world”), and as such they involve the notion of uncertainty.
B. Mood and Irrealis Context of Use
Mood is the grammatical category concerned with the expression of situations involving the “world” parameter. What grammarians call indicative is the mood generally used to assert that a proposition is true as of the actual world.3 To express possible situations, languages can use moods other than the indicative. In Sentence 1, for instance, the verbs are in the so-called Konjunktiv II. The embedded clause in Sentence 2 is in the Subjunctive. Sentence 3 is in the Imperative. In Sentence 2, the English language uses an indicative, while Italian uses a nonindicative mood (subjunctive).
Some languages have a wide range of morphological moods, and some—the most in fact—have a limited number of grammatical categories concerning mood, which are basically the indicative, the imperative, and the subjunctive/conditional. Others do not have any specific morphological markers for mood.4
Most importantly, languages may vary in the contexts of use of different moods. While in all languages the indicative is the mood used to assert a state of fact and imperative the one to command, the other moods (subjunctive, conditional, etc.) have different functions and may be used in contexts that vary from language to language. The contexts where irrealis moods are used more consistently from a cross-linguistic viewpoint include the following:
complements of modal predicates (for instance, to be possible, to be likely, to be necessary, to be probable);
It is probable that these events were coincidences.
complements of volitional predicates (for instance, to want, to wish, to desire);
I wish I had not been late for school.
complements of epistemic (nonfactive) predicates (for instance, to think, to believe, to doubt);
I think we should keep a diverse energy portfolio.
complements of emotive factive predicates (for instance, to regret, to be happy, to be sad);
I regret that this joke has garnered so much attention.
complements of declarative predicates (for instance, to say, to tell, to announce);
I said that one day in my career bad results will come.
the protasis (the if-clause) in conditional sentences;
If he had studied harder, he would have passed the exam.
the apodosis (the main clause) in conditional sentences.
If he had studied harder, he would have passed the exam.
For the purpose of our index, we take the extent of use of different irrealis moods in these syntactic contexts. We assign a value of one to the occurrence of a nonindicative mood in a particular syntactic environment and zero otherwise. By addition we obtain an indicator (IRR) of how frequently irrealis forms are used in a language, so that languages can be ranked according to the intensity of their use of irrealis moods.5 Finally, languages that do not require irrealis moods in any of the context above are called “moodless” languages.
Our linguistic mapping covers 38 languages as listed in Online Appendix Table A1. Data on grammatical moods were mainly collected from Rothstein and Thiero (2010), as it is the most comprehensive typological survey on grammatical moods in the languages of Europe. Since not all the data we needed were included in Rothstein and Thiero (2010), we also collected some additional primary data. For this purpose, we worked out a questionnaire compiled by a number of linguists throughout Europe.6 We contacted linguistic experts, who were asked to provide a translation of various sentences into their native language and to produce, for each sentence, explanations on which mood they were using in their versions (indicative versus other nonindicative moods to be described).
Online Appendix Table A1 shows that six languages are moodless, whereas three languages use irrealis moods in all of the six contexts. The remaining 29 languages range from two to four contexts in which they employ irrealis moods. Thus, significant variation of IRR across languages may represent a good platform for testing the linguistic relativity hypothesis in the context of economic behavior involving risk and uncertainty.
III. Linguistic Variation, Risk Attitudes, and Investment Behavior
Our main hypothesis is based on the assumption that speakers of languages in which nonindicative moods are used more intensively to express potential situations (high-intensity, or H-type) perceive the world as being more mutable and hence more uncertain. As a consequence, they are expected to be more risk averse than those speaking a low-intensity IRR language (L-type) and to engage less in activities with an uncertain outcome. In this section, we empirically validate the following hypotheses:
Hypothesis 1. Direct Effect of Irrealis on Risk Aversion
Speakers of languages characterized by a more pronounced displacement into uncertainty (H-type) are more risk averse with respect to speakers of languages with a weaker displacement into uncertainty (L-type), ceteris paribus.
Hypothesis 2. Indirect Effect of Irrealis on Investment Behavior
Speakers of languages with a more pronounced displacement into uncertainty invest less in risky assets, ceteris paribus.
The intensity of displacement into uncertainty, hence, directly influences the individuals’ attitudes toward risk and indirectly their propensity to invest in risky assets.
The empirical examination of the proposed mechanism proceeds in two steps. We first estimate a direct association between our linguistic marker and the individuals’ attitudes toward risk, Then we specify a two-stage empirical model and use our linguistic marker to instrument risk aversion in the equation for the probability of holding risky assets. In addition, we disentangle the effects of risk and time preferences by using the FTR linguistic marker as a proxy for the individuals’ subjective discount rate.
A. Data Description
Our empirical analysis is run on the pool of individuals in the European Social Survey (ESS) and in the Survey on Health, Ageing and Retirement in Europe (SHARE). SHARE is a rich biennial multidisciplinary and cross-national panel database of microdata on socioeconomic status, social and family networks, individual preferences and choices, and health of individuals aged 50 or older.7 It is suitable for the purposes of our analysis because it contains detailed information on the individuals’ origin and their parental background, as well as on their attitudes toward risky behaviors. In a similar fashion, the ESS collects information on individual attitudes, beliefs, and behavioral patterns, as well as on the respondents’ socioeconomic status, origin, and parental background. However, differently from SHARE, ESS is representative of the entire population in terms of the age structure and contains a larger set of countries. We chose to use both data sets in our analysis because they complement each other mainly in two aspects. First, the features of risk preferences elicited in SHARE fit particularly well with the nature of our research question since they rely on financial and not on adventure risk-taking as in the ESS.8 Second, differently from SHARE, the ESS does not contain any information on the respondents’ asset holdings.9 As for SHARE, we consider the data collected in four different rounds, namely Wave 2, Wave 4, Wave 5, and Wave 6–release 7.0.0.10,11 The sample includes individuals for whom we have a complete information on the self-declared level of risk aversion, asset holdings, as well as on demographic, socioeconomic, family, cognitive and health conditions, origin, and parental background. Similarly, in the ESS we consider six consecutive rounds carried out every two years from 2008 (Round 4) to 2018 (Round 9).12
1. Sample selection and language assignment
To analyze the relationship between our linguistic marker and individual attitudes toward risk, we rely on the subpopulation of second-generation immigrants. Compared to the full sample, or even to a subsample of nonnative individuals (first-generation immigrants), this particular category of natives, with one or both parents who were not born in the respondent’s actual country of residence, allows us to rule out not only any kind of bias due to omitted characteristics from the individual’s country of origin, but also a residual potential bias related to parental cultural and/or linguistic background. The pool of second-generation immigrants in SHARE (ESS) contains 5,302 (8,019) individuals for whom we have a complete information on their risk preferences and parental background, as well as on all the other explanatory and control variables.13
Regarding the language variable treatment in SHARE, respondents are not asked to declare the language they normally speak at home. As a consequence, we assume that the language in which the questionnaire is compiled is also the native individuals’ primary language. Those living in countries with two or more official languages were given the choice between one language or the other. The ESS, on the other hand, lists up to two languages individuals usually speak at home, which allows us to restrict our sample to those respondents who speak the language of the country in which they were born as a first language, excluding native individuals who belong to ethnic enclaves.14
As for the language assignment to the individual mother’s and father’s language of origin, we follow Hicks, Santacreu-Vasut, and Shoham (2015) and consider the official language spoken in their country of origin, if available, or the official language spoken by more than 80 percent of the population in these countries (in all those cases where the country of birth has more than one official language). Individuals whose parents originate from linguistically heterogeneous countries, such as Switzerland, Belgium, or Canada, or who were born in countries (federations) that do not exist anymore (such as USSR, Yugoslavia, Czechoslovakia, etc.), are excluded from the analysis since we are not able to track their original language and/or the information on parental ancestral characteristics is not available.15 The pool of second-generation immigrants in SHARE comes from parents with 32 different linguistic backgrounds originating from 69 different countries for which the information on IRR is available. In the ESS, there are 71 parental birth countries and 39 different native languages for the respondents’ mothers and fathers.
2. Data on risk attitudes and irrealis
As for risk attitudes in SHARE, the respondents were asked to answer a simple question on financial risk tolerance:
When people invest their savings, they can choose between assets that give low return with little risk to lose money, for instance a bank account or a safe bond, or assets with a high return but also a higher risk of losing, for instance stocks and shares. Which of the statements on the card comes closest to the amount of financial risk that you are willing to take when you save or make investments?
(1) Take substantial financial risk expecting to earn substantial returns;
(2) Take above average financial risks expecting to earn above average returns;
(3) Take average financial risk expecting to earn average returns;
(4) Not willing to take any financial risk.
Individuals who answered (1) and (2) have a greater tolerance of volatility of return and hence a lower level of risk aversion. Intermediate risk-takers are those who answered (3), while individuals who answered (4) are highly averse to financial risk. In the SHARE sample, 74.27 percent of individuals declare not to be willing to take financial risks, 21.68 percent are ready to take average financial risk, and only 4.04 percent are willing to take above-average or substantial financial risk. We define risk-averse individuals as those who declare not to be willing to take any financial risk.
In order to elicit risk preferences in the ESS, the respondents were asked the following question on adventure risk-taking:
Now I will briefly describe some people. Please listen to each description and tell me how much each person is or is not like you. Use this card for your answer. She/he looks for adventures and likes to take risks. She/he wants to have an exciting life.
(1) Very much like me;
(2) Like me;
(3) Somewhat like me;
(4) A little like me;
(5) Not like me;
(6) Not at all like me.
An individual is defined as risk averse if they answered (4), (5), or (6) on the above Likert scale. In the ESS sample, 42.69 percent of the respondents consider themselves to have a low aversion to risk, whereas 57.31 percent dislike adventure risk-taking.
Figure 1 shows the distribution of individuals by IRR in SHARE (left-hand side of figure) and in the ESS (right-hand side of figure). In SHARE, 14.23 percent of the respondents are moodless speakers, 59.91 percent are intermediate IRR users, while 25.86 percent are intensive and very intensive IRR users.16 Similarly, the majority of second-generation immigrants in the ESS are intermediate IRR users, there are 17.80 percent moodless speakers, and 31.49 percent classify as intensive and very intensive IRR users.
Distribution of IRR across Second-Generation Immigrants
Source: SHARE Wave 2, Wave 4, Wave 5, and Wave 6; ESS Rounds 4–9. Individuals for whom we have complete information on risk aversion and linguistic background (IRR).
Figures 2 and 3 show the distribution of parental IRR (separately for the mother and for the father). The distribution of parental linguistic background is similar between the two data sets. The only exception relates to moodless parents, which are relatively less represented in SHARE than in the ESS.
Distribution of Parental IRR across Second-Generation Immigrants, SHARE
Source: SHARE Wave 2, Wave 4, Wave 5, and Wave 6; ESS Rounds 4–9. Individuals for whom we have complete information on risk aversion and linguistic background (IRR).
Distribution of Parental IRR across Second-Generation Immigrants, ESS
Source: SHARE Wave 2, Wave 4, Wave 5, and Wave 6; ESS Rounds 4–9. Individuals for whom we have complete information on risk aversion and linguistic background (IRR).
Finally, in Tables 1 and 2 we report the distribution of the three levels of risk aversion over the three different categories of IRR: CatIRR0 contains no IRR (0 IRR), CatIRR1 refers to intermediate IRR usage (2 or 3 IRR), and CatIRR2 represents intensive and very intensive IRR usage (4 or 6 IRR).17
Risk Aversion by CatIRR (Percent), SHARE
Risk Aversion by CatIRR (Percent), ESS
Roughly 37 percent of individuals in SHARE with a low level of risk aversion also speak a moodless language (CatIRR0). Half of the respondents with an intermediate risk attitude also classify as intermediate IRR users (CatIRR1). Finally, those who declare themselves to be averse to any financial risk are mostly either intermediate (CatIRR1) or intensive IRR users (CatIRR2). As for the ESS sample, moodless speakers represent the highest proportion in the category of low risk aversion (23.15 percent) relative to the other risk categories. Also, 52.49 percent of respondents with intermediate risk aversion are intermediate irrealis users, while 33.95 percent of risk-averse respondents also speak a language with an intensive irrealis use.
B. Irrealis and Risk Aversion
1. Empirical strategy
To test the association between IRR and individual attitudes toward risk (Hypothesis 1), we focus on the subpopulation of second-generation immigrants. We reduce any kind of bias related to parental cultural and/or linguistic background and isolate a direct effect of language by controlling for a rich set of ancestral characteristics in the parental country of origin, which may have influenced the formation of risk preferences, independently from the effect of language. To isolate a direct effect of language on individual risk preferences, we follow the Galor et al. (2017) identification strategy and compare second-generation immigrants with the same parental cultural and linguistic backgrounds but who currently speak a different language.
The dependent variable RAi is equal to one for an individual declaring to be averse to risk-taking and zero otherwise. The empirical problem consists in estimating the following logistic model:
where:
1
Our main variable of interest IRRi denotes the number of nonindicative moods in irrealis contexts in the individual i’s language. Xi is the vector of demographic and socioeconomic characteristics, that is, age, gender, marital status (married or in registered partnerships, divorced or separated, widowed), number of children, occupational status (employed, unemployed, retired, homemaker, or disabled), a control for self-employment in the last occupation, educational attainment (low, medium, high), household income, ownership (housing). Zi contains controls for cognitive ability and literacy (reading and writing abilities), a dummy variable for whether the respondent thinks others are generally trustworthy, and health conditions (self-assessed health, ability to perform daily activities, and number of chronic diseases).18 FEi contains country and wave fixed effects, while Pi includes a full set of controls for parental cultural and linguistic backgrounds, that is, geographical and agricultural factors from Galor and Özak (2016), a measure of linguistic and genetic distance between the individual’s country of residence and the parental country of origin from Becker, Enke, and Falk (2020), parental continent of origin, the irrealis marker associated with parental language together with the corresponding linguistic family, and parental educational attainment.
Ancestral characteristics from the parental country of origin might have influenced both the formation of preferences and triggered the gradual emergence of particular grammatical forms that fostered the transmission of these traits across generations. Linguistic structures, in turn, might have reinforced the effect of transmitted preferences on economic choices and behavior. Indeed, Galor and Özak (2016) show that geographical variation in the pre-industrial return to agriculture in the parental country of origin had a persistent effect on time preferences of second-generation immigrants. Specifically, higher historical crop yields potential experienced by ancestral populations had a positive effect on the descendants’ long-term orientation. In line with this evidence, Galor, Özak, and Sarid (2017) argue that these traits are at the root of existing variations in the presence of the future tense across languages. Nevertheless, their empirical findings suggest that linguistic structures still have an independent effect on time preferences and economic choices and behavior.
Initial conditions experienced by ancestral populations might also have influenced the coevolution of other individual-specific traits, such as the perception of risk, and linguistic structures concerned with the expression of uncertainty. In order to control for potentially confounding historical characteristics of ancestral populations, we consider a set of additional controls from Galor and Özak (2016) for the respondent father’s and mother’s country of origin. This set of variables includes: ancestral pre-1500 potential crop yield and growth cycle, the expansion of suitable crops for cultivation in the post-1500 period (“Columbian Exchange,” Putterman and Weil 2010), absolute latitude, mean elevation above sea level, terrain roughness, distance to coast or river, and landlocked. Even though these controls have been originally thought of as direct and exogenous drivers of individual time preferences, some of them may also be relevant in the context of risk and uncertainty, especially the changes in potential crop yield and growth cycle in the post-1500 period. For any given level of agricultural productivity, higher oscillations in crop productivity may have affected the individuals’ level of uncertainty regarding future rewards. Furthermore, we follow Roberts, Winters, and Chen (2015) and correct standard errors for the relatedness between languages by including controls for language family.
As for the genetic and linguistic distances between the individuals’ country of residence and their parental country of origin, Becker, Enke, and Falk (2020) show that they significantly correlate with differences in preferences between populations, such as risk aversion, time preference, altruism, positive reciprocity, negative reciprocity, and trust, with the effects being particularly pronounced for risk aversion and pro-social traits. Genetic distances reflect the length of time elapsed since the early ancestors of the respective populations (groups) broke apart from each other. Linguistic distances, on the other hand, track the structure of separation of human populations over time. To control for this particular aspect of differences between populations, we use the composite measure of ancestral distance. The index is computed as the unweighted average of the standardized values (z-scores) of linguistic and genetic distances. The construction of linguistic distances is based on the methodology proposed by Fearon (2003) that measures the degree to which two countries’ languages differ from each other. Genetic distances, on the other hand, are drawn from Spolaore and Wacziarg (2009, 2018) and quantify the expected genetic distance between two randomly drawn individuals, one from each country, according to the contemporary composition of the population.19
We consider three alternative definitions of second-generation immigrants: (i) native individuals with one or both parents who are born abroad, (ii) native individuals coming from mixed couples with foreign mother and native father, and (iii) native individuals coming from mixed couples with foreign father and native mother. Specifications (ii) and (iii) allow us to analyze the relative strength of the linguistic backgrounds of foreign-born parents with respect to the language of the country of residence.
2. Results
The empirical evidence in Table 3 for the analysis on ESS respondents suggests that native individuals with the same parental cultural and linguistic heritage, but with a different IRR in their local languages, significantly differ in terms of the probability of being highly averse to risk-taking. A more intensive usage of IRR translates into a higher probability of risk aversion, ceteris paribus. However, we also observe a persistence of the parental linguistic background, with the mother’s language exerting a stronger effect. For instance, speaking a highly intensive IRR language increases the probability of being risk averse by 18.5 percent. At the same time, having a mother with a strong IRR linguistic background boosts the probability of risk aversion by 2.6 percent, while the effect of the father’s language is null (Model RA8). As for the subpopulation of second-generation immigrants in SHARE (Table 4), the effect of local IRR is larger in magnitude with respect to the ESS sample, while the effects of parental linguistic backgrounds are not statistically different from zero.
Probability of Being Risk Averse—Second-Generation Immigrants Who Speak the Local Language at Home (No Language Enclaves), ESS
Probability of Being Risk Averse—Second-Generation Immigrants, SHARE
This evidence is even more pronounced when we consider second-generation immigrants coming from culturally heterogeneous couples. Risk preferences of individuals with one foreign-born parent are influenced to some extent by the parental linguistic background (Table 5). The effect of parental IRR is slightly higher for individuals with foreign-born mothers compared to those with foreign-born fathers. Among the older individuals in SHARE, the effect of the foreign-born parent’s IRR is null, while the local language IRR among respondents with foreign-born fathers remains statistically significant and larger in magnitude with respect to the ESS sample (Table 6).
Probability of Being Risk Averse—Second-Generation Immigrants Who Speak the Language of the Host Country at Home (No Language Enclaves), Mixed Couples, ESS
Probability of Being Risk Averse—Second-Generation Immigrants, Mixed Couples, SHARE
The effects of language do not change significantly even when we control for linguistic and genetic distances between the individuals’ country of residence and their parental country of origin (Table 7). The inclusion of these additional controls reduces to some extent the effect of IRR associated to the individuals’ local languages (Model RA8). The effects of parental linguistic background persist, and the impact of the paternal language increases in magnitude and statistical significance.
Probability of Being Risk Averse—Second-Generation Immigrants Who Speak the Local Language at Home (No Language Enclaves), Additional Controls for Ancestral Distances, ESS
The effect of local language in SHARE, on the other hand, becomes even more pronounced (Table 8). The coefficients suggest that the probability of risk aversion is 24 percent higher for second-generation immigrants who speak a local language that features a more intensive irrealis use. At the same time, the range between the effects of local and parental linguistic background widens with respect to the representative sample. This is not a surprising result given the age of SHARE respondents (50+). In general, the older second-generation immigrants are, the higher the likelihood that they mostly speak the local language at home. Moreover, as shown by Alesina and Fuchs-Schündeln (2007), it takes about 20–40 years for individual preferences to change, which, in our context, may imply that the actual revealed preferences have already “disengaged” to a large extent from the parental cultural and linguistic influences, making the local language predominant.
Probability of Being Risk Averse—Second-Generation Immigrants, Additional Controls for Ancestral Distances, SHARE
It is important to note that, in general, once the parental linguistic background is taken into account, the effect of the local language decreases but remains statistically and economically significant.
In Table 9 we test the effect of IRR on two types of risk-taking behavior, namely smoking and drinking.20 As for a proxy for the attitudes toward smoking, we distinguish between individuals who have ever smoked daily during their life, and those who have never started smoking. The alcohol abuse, on the other hand, flags individuals who drink almost every day or at least five or six days a week. The likelihood of ever smoking in life is negatively related to a stronger use of IRR in the local language. Individuals with a more intensive usage of IRR in their local languages are 11.4 percent less likely to have ever smoked. Similar evidence emerges for frequent drinking, but here we also observe an additional negative effect of the mother’s language, whereas the paternal linguistic background exerts a positive effect. However, net of parental linguistic influences, speaking a language with an intensive use of IRR translates into 14.9 percent lower probability of everyday alcohol consumption.
Risk-Taking Behavior—Second-Generation Immigrants, SHARE
C. Risk Aversion and Asset Accumulation
1. Empirical strategy
According to our hypotheses, linguistic differences directly influence the individual perception of risk and uncertainty (Hypothesis 1) and indirectly their investment decisions (Hypothesis 2). In other words, language (IRR) affects investment decisions through its direct impact on risk aversion. In light of the empirical evidence presented in the previous section, which strongly supports Hypothesis 1, the IRR linguistic marker may represent a suitable instrument for individual risk preferences in the equation for the propensity of holding risky assets.
The empirical problem of linking risk preferences to investment decision making, hence, consists in estimating the following causal relationship:
2
where RAi denotes the individual i’s risk aversion, while Pi, Xi, Zi, and FEi are the vectors of explanatory and control variables as in Section III.B.1. In the first stage, we estimate the effects of IRR and other covariates on individual self-declared risk aversion:
3
where IRRi denotes the number of nonindicative moods in IRR contexts in the individual i’s language. By plugging the first-stage fitted values into the second-stage equation we obtain the reduced form model for asset accumulation:
4
Since more risk-averse individuals are less prone to take risk and cope with outcomes that are potentially uncertain, the empirical validation of Equation 4 should yield a negative coefficient of 
.
In addition to risk aversion, individual time preferences represent another fundamental driver of intertemporal decision-making. In order to disentangle the effects of time and risk preferences on asset accumulation, we extend Equations 3 and 4 and consider the FTR linguistic marker from Chen (2013) as a proxy for the individual subjective discount rate.
Regardless of the nature of investment choices (investment in risky assets versus savings), separating the effects of risk aversion and intertemporal preferences is not an easy task. Epstein and Zin (1989), for instance, develop a theoretical model flexible enough to allow for the separation between the intertemporal preferences and the attitudes toward risk. The authors propose a class of utility functions that allow each dimension to be parameterized separately and show that this utility representation is equivalent to the CRRA utility whenever the agent’s coefficient of risk aversion is inversely related to the time preference parameter. However, Andreoni and Sprenger (2012a,b) agree that these two aspects of individual preferences cannot be considered as perfect substitutes, but also claim that they cannot be completely separated.21
Even though the separability of risk and time preferences remains an open question from a theoretical point of view, an empirical implementation of FTR and IRR, the first as a proxy for the individual discount rate and the second as an instrument for risk aversion, is an attempt to disentangle the effects of these fundamental aspects of individuals preferences on the propensity to invest in risky assets.
2. Results
The empirical estimation of the causal relationship in Equation 2 is run on the subpopulation of second-generation immigrants. For the two-stage empirical model in Equations 3 and 4 to work, the IRR linguistic marker must satisfy two basic requirements: (i) it must be correlated with the endogenous variable (instrument relevance), and (ii) it must be uncorrelated with the error term (independence). The exclusion restriction requires that it must not have any direct impact on the probability of holding risky assets other than through its first-stage impact on risk aversion.
Table 10 reports the coefficients from the first-stage estimation. According to the Stock–Yogo rule of thumb (Stock and Yogo 2005), the F-statistics in models RA1–RA4 confirm the strength of our instrument.
First-Stage Linear Estimation and Test Statistics—Second-Generation Immigrants, SHARE
Even though the exogeneity of the instrument cannot be tested directly, there is no reason to suspect that there is any reverse effect of the propensity to invest in risky assets on the instrument. Since we control for country fixed effects (which capture institutional and other country-specific heterogeneities), trust, education, income, occupational status, and health conditions, as well as for the full set of parental cultural and linguistic background variables, the exclusion restriction should not be violated. In other words, it seems reasonable to assume that there are no direct effects of linguistic variation on the propensity to invest in risky assets through omitted variables.
Table 11 shows the second-stage estimates from a recursive bivariate probit model. The dependent variable (asset accumulation) equals one whenever individuals hold some money in stocks or shares (listed or unlisted on the stock market), and zero otherwise. Only marginal effects are reported. In all regression models we control for country and wave fixed effects, cognitive abilities, individual health conditions, and for the full set of controls related to parental cultural and linguistic backgrounds. Models AS1–AS3 report the estimated coefficients for a noncategorized version of the instrument, while AS4 and AS5 refer to a dichotomous version of IRR.
IV Risky Assets (Stocks) by Bivariate Probit—Second-Generation Immigrants, SHARE
To obtain a direct effect of individual time preferences on asset accumulation, we run a separate regression using the FTR linguistic marker as a proxy for the individual subjective discount rate. In order to estimate the separate effects of risk aversion and time preferences on the propensity to invest in risky assets, we reestimate a recursive bivariate probit model using the FTR parameter as a proxy for intertemporal choice preferences and the IRR marker as an instrument for risk aversion. Furthermore, for comparison purposes, in Table 12 we report the coefficients from a simple probit estimation of Table 11 where the individuals’ risk preferences are not instrumented.
Probit Model with Noninstrumented Risk Aversion—Second-Generation Immigrants, SHARE
The instrumented risk aversion is highly significant and larger in magnitude than the noninstrumented one. Without controlling for time preferences, for an individual with average characteristics of the population, being risk averse reduces the probability of holding risky assets by approximately 16 percentage points (Model AS1). As for the individual time preferences, the estimated effect of FTR in Model AS2 shows that individuals with a high subjective discount rate are 3 percent less likely to invest in risky assets. The separate effects of risk aversion and time preferences are shown in Model AS3. The effect of the individual discount rate is negative and highly significant, while the coefficient of risk aversion remains unaltered. The effect of risk aversion is five times larger than the effect of the individual discount rate. Moreover, the coefficients suggest that once the individual risk preferences are taken into account, the effect of FTR reduces only by 0.02 points, which suggests that risk attitudes and time preferences cannot be treated as substitutes. This result is in line with empirical evidence found in the experimental research in the field (Andersen et al. 2008; Andreoni and Sprenger 2012b).
IV. Conclusions
We propose an innovative approach to analyze individual attitudes toward uncertainty and asset accumulation based on a weak version of the Sapir–Whorf hypothesis of linguistic relativity. We develop a specific linguistic marker defined on the basis of the number of nonindicative moods used in irrealis contexts, that is, contexts that involve grammatical categories related to the expression of uncertainty. Our empirical exercise consists of testing the hypothesis that speakers of languages in which nonindicative moods are used more frequently perceive the world as being more mutable and uncertain with respect to speakers of languages where these forms are less common or do not exist at all. The association between our linguistic markers and risk aversion is strong and robust to alternative model specifications, as well as to the inclusion of a rich set of additional controls at the individual and country level.
Second-generation immigrants speaking languages where nonindicative moods are used more often to describe uncertain situations have a significantly higher probability of being averse to risk. The possibility to explore the persistence of linguistic traits among second-generation immigrants allows us to overcome potential concerns due to omitted variables related to parental cultural backgrounds and to isolate a direct effect of language on individual risk preferences.
The approach adopted here, to the best of our knowledge, is also the first nonexperimental attempt to measure a direct effect of risk aversion and individual time preferences on the propensity to invest in risky financial assets. Using our linguistic marker as an instrument for the individuals’ self-declared risk aversion, we show that being risk averse reduces the probability of holding risky financial assets by 16 percentage points. In addition to risk preferences, we run separate regressions using the FTR linguistic marker (Chen 2013) as a proxy for the individual subjective discount rate. We find that both measures are relevant determinants in the decision to invest in stocks. In line with our hypotheses, the level of risk aversion and the preference for current consumption have a negative impact on risky asset holdings. Moreover, by exploiting the orthogonality of the FTR marker and the IRR marker, we were able to show that the impact of risk aversion is five times larger than the impact of the individual discount rate. Since linguistic variation is a trait of individual identity, it can be used as a cultural marker, not only at the individual level, but also at the group level. The results obtained therefore also shed light on the importance of noneconomic factors in shaping individual risk and time preferences, and consequently their economic behavior.
Footnotes
The authors Juliana Bernhofer, Francesco Costantini, and Matija Kovacic certify that they have not received any funding or support for this project from third parties or organizations and that they have no current or past affiliations with or involvement in any organization with any financial or nonfinancial interest in the results of this manuscript. The same applies to the authors’ relatives and partners. No other party reviewed the current version of the manuscript prior to its circulation and our research data (ESS and SHARE) is fully anonymized. This paper uses data from the European Social Survey, which are freely available online. Download links for Rounds 4, 5, 6, 7, 8, and 9 (Norwegian Centre for Research [NSD] 2008, 2010, 2012, 2014, 2016, 2018) are accessible in the reference section. This paper also uses data from SHARE Waves 2, 4, 5, and 6 (Börsch-Supan 2022a,b,c,d); see (Börsch-Supan et al. 2013) for methodological details. Additional instructions and code for replication are provided in an Online Appendix of Replication Materials.
Supplementary materials are available online at: https://jhr.uwpress.org/.
↵1. “KONJ” stands for the German Konjunktiv; “1SG” for First Singular, and “IMP” for Imperative.
2. Swan (2002, p. 242).
↵3. This does not exclude that the indicative may have modal functions, too.
↵4. With the exclusion of nonfinite moods, like the infinitive or the gerund, most Romance languages have four moods according to traditional grammars: the indicative, the subjunctive, the conditional, and the imperative. Most Slavic languages have three moods: the indicative, the conditional, and the imperative. German has three moods, too: the indicative, the “Konjunktiv,” and the imperative. Northern Germanic languages have only two moods: the indicative and the imperative. Subjunctive mood is also mentioned in some traditional grammars, but it has only residual uses and is no longer productive.
↵5. Since there is no qualitative difference between contexts in defining IRR, the index sum was calculated using a uniform weighting function.
↵6. Full version available upon request.
↵7. In general, the respondents in SHARE are aged 50 or older. However, the survey collects information also on the respondents’ spouses, who may be younger than 50 years old.
↵8. See Weber, Blais, and Betz (2002) and Falk et al. (2018) for the definition and comparison of different risk-taking domains.
↵9. The same question related to adventure risk-taking is also available in the World Value Survey (WVS). However, we do not consider this additional source of data in our analysis mainly for two reasons. First, among the seven regular WVS rounds, only the last two (that is, Rounds 6 and 7) contain detailed information on the respondents’ (and their parents’) country of origin. Moreover, the question on adventure risk-taking as a potential proxy for risk preferences (our risk measure in the ESS is based on the same question), is not available in Wave 7, which leaves us with only one round (Wave 6). Second, our linguistic measure refers (almost entirely) to European languages, some of which are scarcely represented (Albanian, Bosnian, Bulgarian, Croatian, Finnish, Italian, Macedonian, Slovak) or even absent (Catalan, Czech, Danish, Hebrew, Latvian, Lithuanian, Norwegian) in Wave 6. This significantly reduces the coverage of our linguistic marker and, in addition, makes its distribution skewed toward one single category (that is, IRR = 4).
↵10. We do not consider Waves 1 and 7 since they do not contain any information on financial risk preferences. Wave 3, on the other hand, is a retrospective survey with a different methodology.
↵11. The SHARE data collection has been funded by the European Commission through FP5 (QLK6-CT-2001-00360), FP6 (SHARE-I3: RII-CT-2006-062193, COMPARE: CIT5-CT-2005-028857, SHARELIFE: CIT4-CT-2006-028812), FP7 (SHARE-PREP: GA N°211909, SHARE-LEAP: GA N°227822, SHARE M4: GA N° 261982, DASISH: GA N°283646) and Horizon 2020 (SHARE-DEV3: GA N°676536, SHARE-COHESION: GA N°870628, SERISS: GA N°654221, SSHOC: GA N°823782), and by DG Employment, Social Affairs and Inclusion. Additional funding from the German Ministry of Education and Research, the Max Planck Society for the Advancement of Science, the U.S. National Institute on Aging (U01_AG09740-13S2, P01_AG005842, P01_AG08291, P30_AG12815, R21_AG025169, Y1-AG-4553-01, IAG_BSR06-11, OGHA_04-064, HHSN271201300071C), and from various national funding sources is gratefully acknowledged (see www.share-project.org).
↵12. Round 1 of the ESS was excluded as only the parental continents of birth are available and not the countries. Also, the information on parental education of Rounds 1–3 does not match the methodology of later rounds; hence, also Rounds 2 and 3 are not part of the analysis.
↵13. See Online Appendix Tables A2 and A3.
↵14. We turn to this point in Section III.A.1 and offer a more detailed description of language assignments.
↵15. In Belgium, there are two official languages, French and Dutch, and neither is spoken by more than 80 percent of the population. In Canada, the English language is the mother tongue for 60 percent of the population, while 22 percent of all Canadians speaks French. In Switzerland, German, French, and Italian are official languages at the national level, with 62.6 percent German, 22.9 percent French, and 8.2 percent Italian speakers.
↵16. Italian and Portuguese are the only two European languages in our sample with the maximum number of IRR (6). Another European language not included in our sample with IRR = 6 is Icelandic.
↵17. Our linguistic mapping currently does not contain any language with IRR = 5 and IRR = 1, even though these categories exist and would be theoretically plausible. Moreover, in our linguistic categorization, the most intensive category, IRR = 6, is underrepresented in both samples, SHARE and ESS. Consequently, if we were to define the category of intensive irrealis users as those who speak languages characterized by IRR = 5 and IRR = 6, we would end up only with one underrepresented category (IRR = 6) indicating the intensive IRR users. Before proceeding to using our cutoff, which includes the upper three irrealis categories (4–6) against the bottom four (0–3) to describe an intensive use of IRR, we followed up with our collaborating linguistic experts, who agree that this is a correct specification to describe intensive irrealis users from a theoretical viewpoint.
↵18. The information on ownership (housing), cognitive ability and literacy, ability to perform daily activities, and number of chronic diseases is available only in the SHARE data. To control for health status in the ESS, we use a self-assessed measure and distinguish between good and very good against fair, poor, and very poor self-declared health status.
↵19. For more details on the definition and construction of these distance measures, see Becker, Enke, and Falk (2020).
↵20. We report only the results based on the SHARE data since in the ESS these outcomes are available in one round only (Round 7), and the sample coverage is not sufficient to produce reliable estimates.
↵21. In a similar manner, Andersen et al. (2008) stress that the assumption of risk neutrality for individuals that are instead risk averse, result in upward-biased discount rate estimates. They also point out that the parameter values for risk and time preferences must come from the same population.
- Received January 2019.
- Accepted March 2021.
References
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.