A new article by Polderman et al. in Nature Genetics, nicely summed up by Jeremy:
— Jeremy Freese (@jeremyfreese) May 25, 2015
is a meta-analysis of essentially every twin-based study of heritability of any trait between 1958 and 2012. The top-line coverage, encouraged by the authors’ press release, is:
One of the great tussles of science – whether our health is governed by nature or nurture – has been settled, and it is effectively a draw.
This is based on the fact that, across 17,804 traits in 28 “general trait domains,” the overall mean heritability was 49%. That prompted me to write:
— Andrew Perrin (@AndrewJPerrin) May 25, 2015
Read on for why I think that, what value there is in the below-the-fold part of the article, and why I think this kind of work is in desperate need of an injection of theory.
In the conclusion, the article claims: “Our results provide compelling evidence that all human traits are heritable: not one trait had a weighted heritability estimate of zero.” But there are obvious selection, measurability, and publication biases that go into the “all human traits” gloss. Traits for which heritability is not plausible to begin with; that are difficult to measure with precision; or that are very rare are unlikely to be in the population of traits investigated for heritability.
This is not just a methodological quibble; it’s an important theoretical point. A reasonable, naive reader could easily read “everything about people” in the “all human traits” phrase. But that’s not a defensible reading of the science here.
The problem is analogous to Gabriel Abend’s excellent critique of moral psychology. Abend argues that moral psychology investigates only “thin” moral concepts, and relies on an unspecified and largely implausible aggregation mechanism to scale up from “thin” to “thick” moral concepts. “At least some moral concepts and properties — thick ones — are ontologically dependent on institutional and cultural facts” (147). By analogy here, at least some traits — thick ones — are ontologically dependent on institutional and cultural facts. In other words, some outcomes of substantive interest are not just agglomerations of simpler traits, each of which displays a certain degree of heritability. (I have suggested [here and here] that we think about these as the result of arbitrary-length causal chains consisting of serial interactions between genetic [G] and environmental [E] effects.)
But wait, I hear you cry. Didn’t you say 17,804 traits? That’s an awful lot!
Well yes, it is an awful lot. But the 49-51 nature-nurture tie claim masks enormous heterogeneity among “general trait domains,” and that variation is the most valuable piece of the article IMHO. The trait domain with the highest overall heritability is ophthalmological, coming in at 71% heritability; the three lowest were environment, reproduction and social values, at 29%, 31%, and 31% respectively (you have to go to page 50 of the supplemental tables to get those numbers). Digging further, supplementary table 33 shows that that Environment category contains everything from “Diazepam Effects” to “Acquired Absence of Organs, Not Elsewhere Classified,” to “Maintaining a Job” and “Intimate Relationships.” “Social Values” includes just three areas: “Individual Attitudes of Strangers,” “Societal Attitudes,” and “Religion and Spirituality.” One of the social attitudes sources is this article, which measures conservatism and religious identity among adolescents; another was this one, with 30 attitudinal measures, ranging from doing crossword puzzles to sweets to the death penalty, being assertive, and riding roller coasters. (Attitude factors “Intellectual Pursuits,” “Treatment of Criminals,” and “Sweets and Games” showed no genetic variance at all, by the way [p. 853].) By contrast, the Ophthalmological category contains things like intraocular pressure, corneal hysteresis, and ocular pulse amplitude–I suspect far more straightforward to conceptualize and to measure.
As I said, I do think there’s significant value in this work. But I think it is in the variance, not the mean. What’s interesting here is the degree of difference in heritability among the different traits and domains. While some of this is probably measurement error, I suspect a lot of it has to do with inadequate theorizing. Specifically, I think there are three theoretical tasks that deserve integration here:
- Complex, hierarchically related traits. These are the “thick” traits referred to above: characteristics of people that cannot be decomposed into collections of isolated traits, but rather are the result of configurations where the structure of the configuration is relevant.
- Complex, sequential causes. These are the indeterminate-length GxE chains I mentioned above. To the extent that the effect of one cause is conditional upon the effect(s) of other cause(s), partitioning the variance between genetic and environmental causes is theoretically problematic (not just empirically).
- Theorizing the environment. Environmental contributions to outcomes are treated as residual; with few exceptions, they are not actually measured, which means they’re not actually theorized either. But environmental causes are likely even more complex and hierarchical than are genetic causes, since there are so many intersecting dimensions of the environments individuals experience.