Tractable limitations of current polygenic scores do not excuse genetically confounded social science

Abstract

Burt's critique of using polygenic scores in social science conflates the “scientific costs” of sociogenomics with “sociopolitical and ethical” concerns. Furthermore, she paradoxically enlists recent advances in controlling for environmental confounding to argue such confounding is scientifically “intractable.” Disinterested social scientists should support ongoing efforts to improve this technology rather than obstructing progress and excusing genetically confounded research.

Burt promises her readers a dispassionate essay challenging the “value of polygenic scores [(PGS)]… for social science.” She states she will do this “not by questioning the ethical or sociopolitical implications of this work…but by scrutinizing the science,” and on this basis will conclude that the “scientific costs outweigh [the] meager benefits.” But the “scientific costs” she enumerates – “obscuring environmental influences, perpetuating a flawed concept of genetic potential…and wasting resources” – are not scientific critiques at all but precisely the “sociopolitical and ethical concerns” she disavows.

Were this a disinterested critique focused on scientific accuracy, Burt would be as concerned about exaggerating the effects of “structural disadvantages and cultural influences” as “obscuring” them. Instead, she admits to “holding sociogenomic methods to higher standards than standard social science methodologies,” excusing this double-standard on the basis of the “social risks” she pledged to leave aside. Similarly, she argues the “scientific costs” of “promoting PGS as ‘genetic potential’…include…promoting the individualization of social problems.” These are ideological objections, not scientific ones.

Burt cautions against “wasting finite resources searching for ‘genes for educational attainment’” – that is, by performing genome-wide association studies (GWASs) that identify genetic variants associated with individual differences in social science outcomes. But a substantial share of the funding for GWASs comes from private and philanthropic sources who disagree with Burt's assessment. As for the remainder, what could be more “sociopolitical” than the question of how taxpayer dollars should be directed by the government and its agencies? Besides, this puts sociogenomics in a Catch-22: Should we fund research to address some of the limitations of PGSs that Burt raises in her essay, or should we give up in despair? Burt counsels despair: “the production of environmentally confounded genetic associations with complex social outcomes is not simply a tractable empirical problem to be addressed with more sophisticated methods. Rather, such confounding is inevitable.”

However, Burt's four substantive criticisms of using PGSs for behavioral outcomes – “relatedness confounding, downward causation, limited coverage of genetic influences, and context-specificity” – are scientifically tractable issues that have substantially been addressed. Within-family studies that use parent or sibling PGS as control variables largely address issues of population stratification and familial confounding, as Burt essentially acknowledges. Furthermore, constructing PGSs from within-family GWASs can remove confounding biases from the PGS. In addition, if the genetic variants associated with behavioral outcomes are principally expressed in the brain rather than in the skin, hair, or musculoskeletal system (e.g., Lee et al., 2018) this constrains the possibility that reported associations are confounded by “downward causation,” that is, by “social selection on attractiveness, height, weight, [or] colorism.” Conducting GWASs in large samples with whole-exome or whole-genome sequencing can increase the fraction of genetic influences covered by PGSs by capturing the effects of rarer variants and shed light on the basic biology underlying behavioral differences (Chen et al., 2022). Finally, extending GWASs to more historically, geographically, and culturally diverse samples will help to quantify the effects of different social contexts on the strength and direction of genetic associations.

The Catch-22 is, however, inescapable: “Even if the problems with environmental confounding could be solved,” Burt insists, “the justification for incorporating PGS into social science is lacking.” This is because, according to Burt, we know the answers to all the important questions already. We don't need PGSs “to demonstrate that supportive, stimulating parenting is associated with child educational attainment” because “we can observe and measure different …background factors and assess how these affect student progressions through educational systems.” But Burt blurs the distinction between the language of association and the language of causation (“affect”). Environmental causation is precisely what genetically controlled designs help establish in observational research. And although we might not need PGSs to recognize “that children who experience childhood disadvantage are not able to fully realize their educational potential,” they can help us more accurately quantify the extent to which various environmental disadvantages account for observed differences in social outcomes and measure how much these effects differ across contexts and conditions. Burt is keen to emphasize the context-specificity of genetic and environmental influences on social outcomes but, as one psychologist forcefully put it, using this as a pretext for “abandoning quantitative estimates is practically and theoretically bankrupt” (Rowe, 1994, p. 24).

The upshot of Burt's critique seems to be that social scientists can safely ignore genetics so long as they include a boilerplate disclaimer that “genetic differences… matter in a complex, context-sensitive way.” But the extent of genetic confounding is not mysterious or unquantifiable. Although Burt is correct that using current PGSs to control for genetic influences is partial at best, a well-established literature going back to the 1970s has used genetically sensitive study designs to investigate social science outcomes. These not only include conventional twin studies (which consistently show genetic differences account for a substantial share of the observed individual differences in social science outcomes, e.g., Frisell, Pawitan, Långström, & Lichtenstein, 2012; Hyytinen, Ilmakunnas, Johansson, & Toivanen, 2019; Silventoinen et al., 2020) but also a panoply of other genetically sensitive designs, such as adoption designs, extended twin designs, sibling difference designs, and more (Baier, Eilertsen, Ystrom, Zambrana, & Lyngstad, 2022; Björklund & Salvanes, 2011; Holmlund, Lindahl, & Plug, 2011; Sariaslan et al., 2021; Wolfram & Morris, 2022). These various designs show substantially attenuated statistical associations between predictor and outcome after controlling for genetic confounds and sometimes remove the original association altogether. Burt insists her article is a broadside against “the scientific value of adding genetics to social science” generally, and not just an argument “about the value of PGS for social science,” yet she neglects to explain why these older, kinship-based designs can be safely ignored.

Burt is correct that social scientists should include appropriate caveats when incorporating PGSs into their work and take efforts to control for environmental confounding. But for the reasons outlined above, they should also support ongoing scientific endeavors to improve this technology. They should not – as Burt does – use tractable limitations of research incorporating PGSs as a pretext to obstruct progress or to excuse genetic confounding in social science research.