[This is a guest post by Keith Chen.]
Mark and Geoffrey were kind enough not only to write thoughtful columns on a recent working paper of mine here and here, but to invite me to write a guest post explaining the work. In the spirit of a non-linguist who’s pleased to be discovering this blog, I wanted to use Mark and Geoffrey’s insightful posts as a springboard to explain my work.
In a nutshell: I find a strong correlation between how a language treats future-time reference (FTR), and the choices that speakers of those languages make when thinking about the future. Specifically, in large data sets that survey families across hundreds of countries, I find a strong and robust negative correlation between the obligatory marking of FTR in the language a family speaks, and a whole host of forward-looking behaviors, like saving, exercising, and refraining from smoking. These correlations hold both across countries and within countries, even when comparing effectively identical families born and living in the same country. While the data I analyze don’t allow me to completely understand what role language plays in these relationships, they suggest that there is something really remarkable to be explained about the interaction of language and economic decision making. These correlations are so strong and survive such an aggressive set of controls, that the chances they arise by random lies somewhere between one in 10,000 and one in 10^32.
Starting with Mark’s post: Mark illustrates beautifully an idea that is really the central concern of all work done in modern econometrics: it can often be difficult to tell the difference between strong correlations produced by causal relationships, and correlations which arise through non-causal factors. Since questions of the connection between language and behavior have historically generated considerable controversy, it seems important to think hard about what exactly these correlations actually suggest. Towards this, I’ll discuss briefly why my analysis suggests that a non-causal story is unlikely, and that a language’s structure is causing its speakers to behave differently.
The basic idea is this: I find a strong negative correlation between a family’s savings rate and whether their language has grammatical obligatory marking of future time (call these strong-FTR languages). Obviously, the savings behavior of families can’t CAUSE their language’s structure: structure precedes behavior by large spans of time. What is possible however, is exactly what Mark’s post illustrates: the possibility of something like co-diffusion (correlated adoption) of both language and non-linguistic values, culture, or institutions. For example, might a value towards education have spread along with certain languages? More generally, might languages have co-diffused with attitudes towards work? With particular institutions or religions?
Most of my paper is spent checking for exactly these types of concerns, and what I find is that the data don’t suggest that they explain the correlations I find. How can we know? Well, if we thought languages co-diffused with proclivities towards education, this would suggest we should compare families with identical levels of education to see if the correlation persists. If we though the co-diffusion was mainly spatial (what Mark’s post is mainly concerned with), we might think to compare families born and living in the same country. Effectively, by trying to control for confounding factors that could be driving a spurious correlation between language and behavior, we can get a sense of what these patterns across families really suggest. This leads to a kind of statistical analysis epidemiologists call conditional-logistic regression. What this does, in effect, is drop families around the world into one of 1.4 billion buckets, where two families fall into the same bucket if and only if they are identical in country of birth and residence, age, sex, income, family structure, number of children, and religion, where the religions of the world are broken up into 74 types. What I then compare, is families who fall into the same bucket, but who report speaking different languages at home.
Now, at this level of detail, even the largest economic data sets only allow me to look at around twenty-five thousand families around the world. These families live in eight countries with enough native linguistic diversity to form sets of nearly identical families who also speak different languages. Those countries are Belgium, Burkina Faso, Estonia, Ethiopia, Malaysia, Nigeria, Singapore, and Switzerland. Surprisingly, what I find is that in every one of those countries, the strong vs. weak FTR language categorization I adopt from Östen Dahl (and the EUROTYP project) seems to have a large and consistent correlation with savings behavior. That is, the direction and the magnitude of these effects are statically identical in every country, despite spanning different regions of the world, and different language families.
These effects can be measured not only in savings behavior, but in many different future-regarding behaviors studied by economists. For example, exercising can be thought of as investing time and effort in exchange for future health. Smoking and overeating are basically the opposite: present pleasure at the expense of future (health) costs. I find exactly the same effect in all three of those behaviors, and only when comparing nearly-identical families. Most surprisingly: language FTR structure isn’t correlated with a self-reporting value of saving, even though both have strong effects on behavior. That is, families that report valuing saving (unsurprisingly) save more, as do families that speak weak-FTR languages. But these effects appear independent of each other: language appears to have the same quantitative effect on families whether or not they report value savings. If language simply co-diffused with a cultural value towards saving, we would not have expected to see this.
In short, the data simply don’t seem to suggest that co-diffusing factors play a significant role in explaining what is an extremely consistent set of correlations between the grammatical marking of future reference in languages and many different future-regarding behavior.
Hopefully I’ve conveyed clearly that these correlations are surprisingly strong and survive a surprising amount of attempts to eliminate confounding factors. The working paper goes into much more detail and conducts many more tests than I could describe here, but all analyses I conduct are on publicly available data, and I’d love to talk more with readers interested in replicating, extending, or testing these results in ways I haven’t thought of. The paper also tries to explain why we might have thought that weak-FTR languages would lead to more savings, which comes down to a pretty simple intuition. Every theory of discounting studied in the behavioral sciences is strictly convex (that is, the value of a future reward is a strictly convex function of when you receive it). What I show is that given this, if languages with grammatical marking of the timing of events lead to more precise beliefs about timing of future rewards (an effect similar to many that have been demonstrated by psychologists), then the effect I find is what theory would suggest. Even assuming that language drives savings behavior, though, this analysis does not resolve the question of how it does this: especially because as Geoff points out in his extremely thoughtful and insightful post, it’s not entirely clear what the EUROTYP typology is measuring.
The main point I take away from Geoff’s post is that these types of typological classifications can be exceedingly difficult to interpret, and what they are measuring may not be entirely clear, even after quite a bit of thought. Geoff discusses the example of English, which is characterized in Dahl’s EUROTYP typology as strong-FTR (or more accurately, non weak-FTR). For those readers unfamiliar with this typological distinction, English is considered a strong-FTR language because of observation that unlike most Germanic languages, English generally requires speakers to grammatically mark future events, primarily with either the de-andative construction “be going to” or the de-volative construction “will”. It is this generally tendency toward obligatory grammatical marking of future time that characterizes the strong vs. weak FTR distinction.
What Geoff points out is that there are many exceptions to this, and that the exact nature of these constructions is complicated: neither is purely a tense marker. Indeed, there is substantial evidence that neither construction is a tense marker at all, but instead mark different temporal and modal properties which give rise to future reference in certain contexts (“going to” is prospective aspect, while “will” can be a modal auxiliary). Many linguists may wonder: if the precise function of the EUROTYP classification is not entirely clear, even for deeply studied languages like English, what are we to make of aggregate correlations between this classification and behaviors?
I share this concern, and think it’s worth thinking hard about both what it suggests careful work should look like, and how we might interpret results that survive careful analysis. Though I plan to take more steps in the direction of investigating more fine-grained linguistic distinctions in future work, I think it is important to note that despite these issues we can still learn a lot from correlations between even rough typological distinctions and behaviors.
To see this, note that while the obligatory grammatical marking of future time may collapse several important grammatical features, it is reasonably clear that it measures some distinction in how languages treat time, and that this type of distinction is stable enough to have been discussed by several different authors studying TMA systems. For example, the EUROTYP classification of English as a strong-FTR language and German as a weak-FTR language reflects similar distinctions in earlier work such as Comrie (1985), which goes so far as to use English and German as exemplars:
"Within languages that make a basic present-vs-past distinction, it is worth distinguishing two sub-types… which define end points on a continuum. The one would include languages where the present can always be used with future time reference, the only constraint on this use being avoidance of interpretations with present time reference… German and Finnish would fall into this category. At the opposite extreme would be languages where, although the present can be used with future time-reference, there are several constraints on the use of this form, constraints that are not explainable purely as strategies to avoid conversational implicatures. English would be an example of this category, since the present can be used with future time reference only under highly specific circumstances…"
Copley (2009) also notes this distinction in how English and German make reference to future events, noting a “plannability restriction” in English (on the use of present forms to refer to future events) that is not present in German. None of this is to claim that existing typological classifications are measured without considerable noise, or even that we have a good idea of what a complete typology of future-referring strategies would look like. What is important, though, is that the EUROTYP typology captures something about how future events can and cannot be spoken about across languages. That it also seems to capture an important tendency in future-directed behavior, is both surprising and bears explanation.
Note also, that while typological coarseness may complicate the interpretation of the correlations I find, if anything it strengthens the inference that language and economic behavior are interacting in powerful ways. Why? Well, since these typological distinctions are necessarily noisy, if anything, the correlations we see in data should be an UNDERESTIMATE of the correlations we would find if these characteristics of language could be measured more precisely. To see this, imagine that the EUROTYP FTR classification was effectively no better than a coin flip. Then, we would have expected to find no correlation at all between language and behavior out in the world. More generally, if we think that the classification is by its very nature coarse and messy, then we should also think that if anything, language structure and savings behavior are MORE closely tied than their correlation suggests.
In short, I believe the data suggest a strong and robust relationship between linguistic and economic data, a relationship that bears explaining. Where this leaves us is what I think is an exciting place: one where Economists have a lot to learn from Linguists.
(I gratefully acknowledge the helpful comments of Nicole Palffy-Muhoray. All errors are of course my own.)
[Above is a guest post by Keith Chen.]