The Gloomy Prospect Comes for Genomics

Morris reply, part umpteen

I know I am running on about this, but having started the project of replying to Damien Morris, I figure I should finish.

As a reminder, Morris’ broadside is here, here, and here.

Morris cites derisively a paragraph from Turkheimer (2000), which, if I may say so, I am quite proud of:

There is an interesting parallel between the search for individual genes that influence behavior and the failed attempt to specify the nonshared environment in terms of measured environmental variables. In each case, investigators began with statistically reliable but causally vague sources of variance, and set out to discover the actual causal processes that produced them. The gloomy prospect looms larger for the genome project than is generally acknowledged. (Turkheimer, 2000, p. 164)

Remember: this was eight years before Maher (2008) coined the term missing heritability. At the turn of the millennium, at the height of optimism for the completion of the Human Genome Project, how many people were anticipating how difficult it would turn out to be to translate genetic variance into well-specified genetic causes?

Morris tries once again to label me as a libertarian, as though contending that the difficulties of breaking down aggregate estimates of genetic variance into well-specified, replicable genetic causes necessitates a commitment to non-causal metaphysical free will. It does not. The genetic variants we are dealt, just like the environmental events we are dealt, are embedded in cultural, familial and individual contexts that render their causal consequences highly nonlinear and idiosyncratic. As self-guiding agents, we experience whatever tendencies our genes may confer, and integrate them with our cultural, familial and individual experiences into the construction of a life path. Our behavior is caused by processes, including genetic processes, that are difficult-to-impossible to break down into well-specified causal models, and much of what we can manage to understand in individual people doesn’t generalize to others. The subjective experience of that kind of causation is what we call free will.

Next, Morris argues that I have been unduly pessimistic about the progress of modern genomics. This section reads as though it were written five years ago, in the enthusiasm generated by the educational attainment GWAS series, known as EA1-EA4. He cites the 15% of the variance predicted by a polygenic score in EA4, but leaves out the new technology that EA3-4 introduced. When you estimate heritabilities or polygenic scores within families rather than between them (ie, nonshared as opposed to shared genetics), you control for effects arising in cultures or families rather than from individual inherited genomes, and both heritabilities and the validity of polygenic scores for behavior are greatly reduced. On an R² metric, the within family PGS accounted for 3-4% of the variance in EA4, a quarter of the value Morris triumphantly cites. Since EA4, as statistical genetics has gotten more and more sophisticated about removing non-causal confounds from genetic estimates, those validity coefficients have gotten perilously close to zero; properly controlled behavioral heritability estimates are mostly in the range of .1 to .2.

The important insight is that the problematic relationship between aggregated variance components and granular causal events is isomorphic in the within-family genome and the nonshared environment. Plomin’s question was, Why are siblings raised together so different, and his answer was, because their individual experiences make them different. But the specific experiences that make one sibling a choir boy and the other a delinquent remain elusive, refusing to aggregate into powerful causal mechanisms or generalize from one family to another.

You could call that, “The Missing Environment Problem,” by way of being reminded that the same thing has now happened in the genome, just as I predicted. The genomic equivalent of Plomin’s question about sibling difference becomes, “Why are genetically related people so alike?” The answer, at the level of population variances, is: the genes they share make them alike. But which genes, doing what, remains almost entirely unknown, despite perpetual optimistic predictions from the hereditarians. Identical twins are more likely than other siblings to be concordant as choir boys or delinquents, and the reason, if stated sufficiently abstractly, is that they share more genes. But just as there are no experiences that reliably turn delinquents into choir boys, there aren’t any genes that do it either.

I want to be absolutely clear that I don’t see any of this as leading to old-fashioned environmentalism, much less the kind of mystical libertarianism Morris wants to pin on me. I know “genes matter” for human behavioral development, and if that is all the hereditarians want they can have it. What I do think is that while behavioral variation is easy to “explain” in the very weak sense of analysis of variance, it has the peculiar quality that the more specific one gets about its causal basis, the weaker the phenomenon becomes. Once again, this is not something that has plagued genomics in particular; it is a problem across the non-experimental human behavioral sciences. In my Three Laws paper, I said:

The disconnect between the analysis of variance and the analysis of causes, to use Lewontin’s (1974) phrase, is not a proprietary flaw in behavior genetic methodology; in fact, it is the bedrock methodological problem of contemporary social science.

It’s important to separate two issues here. The first is whether hereditarianism is true. Morris believes that the inexorable progress of science will eventually produce genetic– or let’s say scientific– explanations for most human differences. This view is false, and has been proven false over and over again. We are currently living during the disconfirmation of hereditarianism in its common-variant GWAS modality, just as we lived through it in the candidate gene version and the twin study version. GWAS-world hasn’t fully opened its eyes to this failure yet, but eventually they will, just in time to move on to whatever shiny toy technology comes next.

There is another vision for behavioral genetics that rejects or ignores hereditarianism, and just sees genomic information as a new tool that allows us to conduct old fashioned social science a little bit better, with a few more variables in view. This is what I do for a living, when I am not making pronouncements here. Old fashioned social science is a peculiar beast, grounded in rationalism and empiricism on the one hand, but forever unable to reach law-like conclusions in the mode of the physical sciences or even biology on the other. What exactly social science does do, and how well it does it, are difficult questions that I am in the process of learning more about. I am sometimes prone to say, “We are ethically prohibited from conducting experiments on humans, and therefore we can never establish rigorous causal models.” Both ends of that assertion, I have come to realize, are exaggerated. Sometimes we can randomly assign humans, like schoolchildren to a curriculum, and we can use the experiment to establish the curriculum’s causal effect. But we can’t randomly assign kids to alcoholic parents, or sacrifice them and dissect their brains. Social science isn’t impossible; it is limited. This is deep water, and I am close to once again getting in over my philosophical head.

So if you want to include the EA polygenic score in your evaluation of a school curriculum (maybe because you couldn’t randomize) it’s fine with me. All I ask is that you be transparent about how much work the PGS is doing. I have a part-time hobby of downloading papers that use a PGS as a social scientific tool, and rummaging through the supplemental tables (the carpet under which tiny effect sizes are swept nowadays) to find out how much variance it accounts for. It has been years since I have seen one that did better than 5%. Once again, if you can show that it nevertheless improves the numerical or causal specificity of your estimates, go ahead. But that is social science, not a confirmation of the hereditarian hypothesis.