Commentary on Roseman (2018)

Note: A large part of the paper and my subsequent commentary discusses the interpretation of the arguments that certain individuals in the scientific literature are making and the implications thereof. I am not any of the individuals in question and thus cannot speak directly for them, but only to my interpretation of their work and what I expect they would say. If anyone is still interested in what these people think, I would recommend contacting them.

A recent article by Charles Roseman (2018) aims to come at the race-IQ debate from a metascientific perspective as to what methodologies and techniques produce valid inferences and forms of evidence. Specifically, Roseman, from an anti-hereditarian perspective, charges many anti-hereditarians with (what amounts to me to be) obscurantism and erecting unnecessary burdens of proof for mechanistic explanations.

There are essentially two arguments that Roseman takes issue with. The first is an a priori argument against genetic group differences from the complexity of genetic outcomes. He primarily cites authors in the public sphere, such as Nathaniel Comfort and Gavin Evans, rather than academic criticisms.

Roseman describes Comfort’s criticism of Plomin:

Similarly, in a review of recent book with a hereditarian outlook (Plomin, 2018), Comfort (2018) describes the volume as “yet another expression of the discredited, simplistic idea that genes alone control human nature seems particularly insidious.”

and Evans critique of the revival of scientific hereditarianism:

Intelligence — even the rather specific version measured by IQ — involves a network of potentially thousands of genes, which probably takes at least 100 millennia to evolve appreciably. Given that so many genes, operating in different parts of the brain, contribute in some way to intelligence, it is hardly surprising that there is scant evidence of cognitive advance, at least over the last 100,000 years

At the conceptual level, too warm an embrace of complexity is a denial of the possibility of theory. All understanding relies on abstraction and simplification. If life is so very complex and we ask that every last detail be accounted for before an explanation is accepted, any attempt at a science of human behavior or anything else that includes genes or environment is an exercise in futility

When the stakes are as high as they are in a question like genetic explanations of group differences (Kitcher 1985), some would argue that a much higher burden of proof is necessary to adjudicate the scientific claims. As such, the complexity of the relationship between genes and behaviors – specifically the numerous confounds (Taylor 2007), context-dependency (Moore 2018) and inability to make high-level inferences (MacKenzie 1980) – goes against hereditarianism by undercutting their core claims. The issue is not, per se, that these critics are arguing that the complexity of phenotypic development needs to be completely solved before we can say anything useful about phenotypic development, but more that the complexity of this development precludes us from making the inferences hereditarians need for their scientific claims (Turkheimer 2017).

It may be that environmental effects swamp out genetic effects
or otherwise stymie random genetic drift and natural selection, but there is nothing inherent to complex traits that will slow their evolution over short time spans.

In my simulations, as polygenicity increases (and behavioral traits are the most polygenic traits there are), the variance of the expected mean differences on drift decreases. If correct, this means it is unlikely that drift will be a sufficient explanans for group differences in behavioral traits. Moreover, given the fact that IQ is almost universally posited to be fitness increasing, the increased efficacy of selection in Africa would mean that deleterious variants would be purged more effectively even with negligible fitness differentials (Aris-Brosou 2019). This means that we expect a leftward shift of the expected mean gap on drift towards Africans having a higher genotypic IQ. That is the opposite of the claim posited by hereditarians.

From a more rhetorical standpoint, critiquing and rejecting hereditarian arguments about the relationships among genome, organism, and environment by pointing to a past set of failures behavior genetics and accusing them of oversimplifying a complex situation is a self-defeating strategy

I do not have access to the most recent version of this paper, but I do not understand either the initial claim made here or the evidence marshaled in support of it. The fact that psychiatric genomics has attempted to estimate the influence of genetic effects on psychiatric traits does not obviate the criticisms that critics of hereditarianism have made in the past with respect to the relationship between environments, genes, organisms and development.

If genetic accounts of behavior begin to mimic their morphological and physiological cousins and as more complex studies of genes and behavior are lavishly funded, ground erodes from beneath the palisades of the complexity argument and it is not clear where antihereditarians of the complexity stripe can further retreat.

It seems here that Roseman is arguing that the complexity argument is only contingently successful on the historical and current failure of behavioral genetics to succeed at what the critics see as its goals, but that this could very well change in the future, destroying the basis of the complexity argument. While this may be true for some complexity theorists, it is definitely untrue for Turkheimer. He has long stated that he predicted that GWAS would identify SNPs associated with behavioral traits (Turkheimer 2000, 2012) and has even argued that they migh recapitulate the twin heritability estimates – the hypothetical success of behavioral genetic studies would be less of a condemnation of his position than a confirmation. As for other critics, it depends exactly upon what sort of “more complex studies of genes and behaviors” that Roseman is referring to – I cannot fathom new studies that would eradicate the issue with the complexity of behavioral development, but that may be the poverty of my imagination.

Do we need to have a molecule by molecule account of how a gene works to be confident that there are genetic effects on phenotypic variation?

Here Roseman conflates two related but importantly distinct questions: whether there are genetic effects on phenotypic variation and whether a particular set of SNPs or genes are causally associated with phenotypes. The former has been answered, at least for the vast majority of critics of hereditarianism (as noted by Roseman, Turkheimer’s weak genetic explanation satisfies the ‘genetic effects on phenotypic variation’ requirement). The question these critics are focused on is rather the latter, as the particular set of SNPs that are associated with a phenotype are often used for social and epistemological purposes (Aaronovith 2018; Piffer 2019), ones in contrast to the purposes of anti-hereditarians. As such, it is prudent that the correct SNPs be identified and the manner in which they become associated with phenotypes be ascertained.

It is indisputable that nutritional deprivation stunts growth in children and adolescents and we knew that well before we knew anything about the molecular machinery of growth plates

The difference between this example and purported results from GWAS is one of causality. One can use experimental, quasi-experimental and propensity-score matching designs to estimate the causal impact of nutritional deprivation of growth in youth, but there is no such analogous methodology for genome-wide association studies. Moreover, the mechanism by which two things are associated is against important for the social and epistemological questions that plague the hereditarianism debate.

Similarly, the links between smoking and lung cancer were well established before the details of the physiological operations of any cancer were understood

Typical to those wishing to defend causal interpretations of certain associations without sufficient evidence, the author here appeals to the history of the relationship between smoking and lung cancer. Here, he slips between the question the critics are asking – causality – and what GWAS has actually achieved: “links”. During the beginning of the 20th century, there was certainly strong evidence that associations between smoking and lung cancer existed. One could not yet conclude causality in any definitive form, but only try to make a best guess as to the true effect size. Luckily, a large body of research has confirmed the association studies from the beginning of the century using model organisms and experiments (Khang 2015; Lucas & Harris 2018). What to infer from current research results depends on your statistical approach to inference (frequentist vs Bayesian) and a number of other interpretative and adjudicative factors. Anti-hereditarians who base their critiques on the necessity of mechanistic explanations

We can tell similar stories for phenomena as disparate as the realization that sewers and clean water supplies were important for health (Chadwick, 1842) and the efficacy of vaccination (Lombard et al., 2007). In all of these cases, the big picture of how causal influences were flowing through the world were clear before the molecule- or microbe-eye view of the mechanisms of causation were worked out.

The discussion of epidemiology is again particularly revealing. The question of causality has long haunted epidemiology journals, who have struggled with the philosophy and statistics of demonstrating causation (Glymour 1998; Mackenzie & Pearl 2018; Parascandola & Weed 2001). The existence of epidemiology’s Hill criterion, and several modifications for different subfields and disciplines have established clear criterion for how to create a case for the causal impact of one variable on another. Yet again, GWAS do not fit these criterion, at least according to the critics.

A quick inspection of the diversity of life makes the difficulty with this position clear. A comparison of Ankylosaurus clubs (Arbour and Currie, 2015), Stegosaurus spikes (Cobb, 2009), and the tailless Manx (Todd, 1961) shows that tails can evolve in elaborate ways. The genetic basis of these tail characteristics (outside of the Manx) are as obscure as that of cognitive characteristics as are the vast majority of traits that have ever evolved. Evolution takes place nonetheless, showing that there are mechanisms that generate genetic variation amenable to sustaining evolutionary responses even if our understanding of relationships between genotype and phenotype in any one generation is less than concrete. This is a simple extension of the Darwinian observation that there is no difference in kind between variation within and among species.

Here, I fear Roseman has misunderstood Turkheimer’s distinction between weak and strong genetic explanations. It is not about whether we currently have access to a mechanistic account of the development of the trait and how it becomes associated with an SNP, but whether there exists such an account at all or whether it is coherent to ask for such an account. Turkheimer would say that most physiological traits fall into the category of traits that do in fact have “strong genetic explanations”, but the question is which behavioral traits do and which do not have strong genetic explanations: the real nature-nurture debate (Turkheimer 2016). Moreover, he is confusing several critiques with one another. I doubt Turkheimer has any issue with the evolutionary claims Roseman lists or the ability to identify selection in other species barring a mechanistic explanation. His claims are about the implications of weak vs strong genetic explanations for sociological questions (like approaches to crime).

It is not clear what the demand for strong genetic explanation is supposed to deliver for anti-hereditarians. A mechanistic explanation is no less context dependent than statements about variation at the population level.

If all that exists for behavioral traits ends up being weak genetic explanation (a la Turkheimer 1998, 2016), then it doesn’t seem like hereditarianism is a scientifically decidable hypothesis like hereditarians argue. Moreover, the demand for a mechanism comes first for a demand for the demonstration of causality. Anti-hereditarians are often skeptics of GWAS and argue that the associations identified are confounded by population structure, associative mating, and other forms of gene-environment correlation (Baverstock 2019; Richardson and Jones 2019). A mechanism goes a long way to demonstrating causality, and in some cases actually ends up proving it. The context-dependency of a mechanism only demonstrates the futility of hereditarianism rather than vindicating variational approaches to understanding phenotypic development. Moreover, what anti-hereditarians who view mechanistic explanation as a requisite factor for adjudicating the hereditarian question argue is that certain mechanisms by which SNPs can become associated with phenotypes are not amenable to hereditarianism. For instance, if it ends up being that height is highly heritable (say 80%) and that height causally impacts IQ  (via some socially mediated pathway), then of course IQ will be statistically heritable. But it isn’t “genetic” in the typical colloquial sense that hereditarians need – if simply changing our social norms will decrease the additive genetic variance of the phenotype, that contradicts a great number of claims in the hereditarian literature (e.g. Jensen 1969; Rushton 2012). Indeed, if it ends up being that (in our hypothetical) all of the additive genetic variation for IQ is mechanistically explained in this manner, then I suspect the ‘so-called’ environmentalists would consider themselves vindicated on the matter, as the proximate developmental mechanism is environmental. So understanding the mechanisms helps us understand the inferences that can be made from the association of a SNP to a phenotype – different types of associations have different biological, epidemiological and social implications.

My fear is that choosing to only accept extreme and well-characterized clinical phenotypes or simple Mendelian characteristics with good molecular level mechanistic explanations as genuinely genetic casts genetics in biased ways.

This is a misrepresentation of Turkheimer’s view on how genetic associations should be understood. His point is that a coherent molecular genetics of say, divorce, are not forthcoming. The proper level of analysis for understanding the origin is not “bottom up from the genes”, but at the phenotypic level itself (Turkheimer et. al 2014). These are not any less legitimate genetic associations, but they do not have simple mechanistic explanations. To be clear, Turkheimer doesn’t think there is a fine line between a “true” genetic association (‘true positive’) and a confounded genetic association (‘false positive’) – the question of causality may not end up having a fact of the matter for GWAS. But that does not dismiss the fact that Turkheimer has never claimed that these associations are not “genuinely genetic” or that the heritability is fictitious.