R1.1. A critique of PGSs not genetics for social science, in general

Several commentators perceived my article to be a critique of sociogenomics as a field or the incorporation of genetics into social science in any form (Burke; Keller; Moreau & Wiebels; Richters; Zietsch et al.). Rather than sociogenomics in general, my target article was “focus[ed] on the utility of PGSs for social science and the key premises underlying their use as measures of ‘genetic propensities’ for behavioral differences,” as the title also announced. To be sure, I should have better worded a few sentences to reflect my specific focus on PGSs for social science; thus, I take responsibility for inadvertently encouraging this interpretation. Even so, my coverage throughout, including my key recommendations, concentrated on PGSs. This is why – to address Morris, Ritchie, & Young's (Morris et al.) critique – other methods of incorporating genetics into social science were not discussed. This is also why – to address Zietsch et al.'s primary critique – my article does not reflect “a fallacious motte-and-bailey argument” (see Shackel, Reference Shackel2005). My focus, which Fletcher aptly described as being “on a subset of ‘genetics’ [for social science] – the use of polygenic scores” was not a stand-in or “motte” for general opposition to genetics in social science.

Some commentaries interpret my article as implying it is acceptable to “deliberately ignor[e] genetic influences” on social phenomenon (Burke, also Zietsch et al.). Although I do not concede that genetics is relevant to the explanation of all social phenomenon (e.g., the association between being American and driving on right and being British and driving on left), my critique of PGSs was not a call for social scientists to “deliberately ignore genetics” but to recognize that however relevant genetics are to our development and social traits, PGSs do not capture “genetic (vs. environmental) influences” on social traits. By analogy, my air quality indicator is unable to accurately differentiate between carbon dioxide and volatile organic compounds (VOCs). I do not recommend you use it to measure VOCs for that reason, but from that it does not follow that I think VOCs are not important to measure, much less that they be deliberately ignored.

Similarly, Keller depicts me as holding a “black-and-white” position that we “should refrain from researching one of the important factors (genetics) influencing trait variation.” He further implies that my arguments rest on the naïve position that models need to be perfect to be useful. Neither are positions I hold or espoused in the target article. The challenges of PGSs for social science are not merely that they are imperfect as all methods are, but rather that PGSs have specific limitations that vitiate their utility for social science research. As Curtis writes, articulately precising my arguments: “PGSs are so poor at capturing the genetic variation which is biologically relevant while at the same time being profoundly influenced by exactly the kind of confounders social scientists do not want contaminating their research such as race, socioeconomic status and parental characteristics.”

R1.2. Not assuming psychological homogeneity

In a response familiar to critics of behavior genetics, Richters alleges that I, likely ignorantly, endorse a model of “psychological homogeneity” (see, Harden [Reference Harden2021] for an analogous “genetic sameness” argument). I do not (e.g., Burt, Reference Burt2020; Simons & Burt, Reference Simons and Burt2011). My scrutinizing methodological tool, PGSs, as a measure of “genetic influences” or as being useful for enhancing understanding is not the same thing as denying genetic differences or assuming a blank slate view of human psychology. The critique that I assume psychological homogeneity is both wrong and irrelevant. Indeed, we agree that individuals differ genetically and psychologically in a manner that shapes development and social outcomes. The key question at issue, which Richters avoids, is whether PGSs have utility for enhancing understanding of these differences.

R1.3. Recommendation: Use PGSs sparingly and cautiously given limitations

A few commentators interpreted my argument as being that PGSs are useless and should never be used in social science. For example, in their otherwise concurring response, Veit & Browning argue that I overstate my conclusion, which they interpret as being that “sociogenomics is methodologically doomed” and that PGSs are useless for all social science purposes no exceptions. This ostensible disagreement is based on misunderstanding. I specifically recommended that PGSs should be used “sparingly and cautiously” for social science rather than “not at all.”

Analogously, although agreeing that my critique is “mostly correct,” Fletcher takes issue with what he perceives to be my argument that PGSs are worthless and aims to carve out a “clear-eyed middle ground.”Footnote ² Concurring with my arguments, Fletcher writes that studies representing PGSs as “genetic propensity” and which are using PGSs as “genetic influences” (vs. environmental ones) are “overstepping” and “a fool's errand.” Given that most sociogenomics studies use PGSs in this manner, it follows that we agree that most studies use PGSs inappropriately. However, and purportedly disagreeing with my position, Fletcher suggests that “PGSs can be wrong but useful” “in a limited and focused role in social science research.” Contra Fletcher, I did not argue otherwise. The “clear-eyed middle ground” Fletcher aims to carve out was that carved out in my article.

R1.4. Downward causation as a confounder of PGSs

A few commentators interpreted my discussion of downward causation as implying that PGSs “only” or “merely” reflect artificial (social) causation (Trejo & Martschenko; Xia & Hill). For example, Xia & Hill write that I describe “the signal captured by a PGS…on social science traits such as education as being ‘artificial.’” This is a misconception with benefits, as it allows them to apply to my arguments another label familiar to critics of behavior genetics: environmental determinism (in this case downward determinism). I am surprised by this interpretation not only because downward causation is but one of several confounders to PGSs that I describe, but also because I explicitly rejected an environmentally determinist approach. In the service of explanation, I employed simplified examples to illustrate the point that because of downward causation “genetic associations for many complex social behaviors are unavoidably environmentally confounded” not determined (emphasis added). When I wrote: “As is well known, a person's social traits emerge from a complex interplay of environmental and genetic influences over their lifetime,” I meant it.

To be clear, my claim that PGSs capture artificial genetic associations does not imply that PGSs only capture artificial genetic associations. We agree that an environmentally determinist approach is untenable.

R1.5. No enthusiasm for the outdated nature versus nurture debate

In another unanticipated response, several commentators (Trejo & Martschenko; Alexander et al.; Richters) charge me with “perpetuating the nature versus nurture debate.” Although sympathetic with some of my critiques, Trejo & Martschenko write that I “desire to separate nature versus nurture” and my arguments encourage attempts at such separation. Richters claims that my arguments “renew the charter” of genetic versus environmental separation. In all cases, this critique is asserted but not explained, and as I do not see how this follows from my arguments, I cannot engage directly with their reasoning.

To clarify, my discussion of environmental confounding was not meant to encourage efforts to differentiate genetic versus environmental influences, which we agree is a futile endeavor (see Burt, Reference Burt2015; Burt & Simons, Reference Burt and Simons2014). On the contrary, by illuminating the fallacy in treating PGSs for complex traits as “genetic influences,” I was arguing against the interminable effort to separate nature and nurture in its contemporary form with PGSs as “nature.” When I wrote that studies using PGSs as genetic influences are “fundamentally and necessarily wedded to an overly simplistic and ultimately misleading (environmentally confounded and biologically implausible) reductionist genes-versus-environments approach,” and the problem is not tractable with advanced statistical methods, as Trejo & Martschenko agree, I meant that too. We “can no more unbraid genetics and environments [on complex social traits] than we can unbraid history and culture, or climate and landscape, or language and thought” (Feldman & Riskin, Reference Feldman and Riskin2022).

R1.6. Unknowns and false dilemmas

From my claims that we don't need PGSs to show well-established social patterns (e.g., “to demonstrate that supportive, stimulating parenting is associated with child educational attainment”), Morris et al. craft a straw man, perhaps for rhetorical effect. They misrepresent me as holding “that we know the answers to all the important questions already.” Obviously, we do not.

In a more reasonable objection, Morris et al. write that: “environmental causation is precisely what genetically controlled designs help establish in observational research.” I anticipated this response, and I refer the reader to sections 5 and 6 of the target article where I discuss why demonstrating environmental causation is not a strength of PGSs. Briefly, because, as we all agree, PGSs do not control for “all genetic differences” and are environmentally confounded, I noted:

Disappointingly, Morris et al. did not engage with these specific arguments. Instead, they pose a dilemma: Support PGSs or support genetically confounded social science research. Fortunately, this is a false dilemma.

R1.7. Not defending “standard social science model” or social measurement

Richters objects to my argument because, in his view, I do not “highlight precisely the same deficiencies in the social science model [I] seek to defend…”Footnote ³ Richters' critique is, however, based on a misunderstanding; my target article is not a defense, much less a “vigorous defense,” of social science research. There is no contradiction in addressing the challenges with PGSs for social science and holding that social science research, in general, has many challenges, even deficiencies.

In a similar critique, Morris et al. complain that were I dispassionate and focused on scientific accuracy, I “would be as concerned about exaggerating the effects of ‘structural disadvantages and cultural influences’ as ‘obscuring them’.” I anticipated this tu quoque, and I point the reader to section 8 where I attempted to dispel such unproductive discussions. Manifestly, my target article was not an overview of “problems with social science” but had a very specific focus on challenges with PGSs.

Focusing on measurement, Moreau & Wiebels interpret me as holding that “because they are already well-measured behaviorally, constructs like academic achievement or cognitive aptitudes have little to benefit from the tools of sociogenomics.” This is a two-part claim, and both are misguided. First, I did not argue that constructs like cognitive aptitudes or psychosocial traits are well-measured. Indeed, I share their concern about the measurement of social constructs (see, e.g., Burt, Reference Burt2012, Reference Burt2020) and agree that we “should refrain from thinking that the measurement of constructs in the behavioral sciences is as good as it can be.” Second, my critique of the utility of PGSs in social science is not based on adequacy of social measurement. If anything, my arguments would lend support to the claim that the inadequacy of measurement of social constructs poses a challenge to GWASs and PGSs. Although I agree that PGSs will be more useful for medical phenotypes defined by the “presence or absence of biological features,” pace Moreau & Wiebels, the fact that “behaviorally assessed constructs remain subjective and far from assumption-free” is, in my view, a barrier to genetic analysis not an argument for its utility.

R1.8. Context-dependency: More than an interpretive problem

In the target article, I discussed the context-dependency of PGSs and outlined the implications for complex social traits (see sect. 5.4). In his commentary, Fletcher briefly depicts this significant challenge as being of a narrower problem: That the “interpretation [of PGSs] is context-dependent” (emphasis added). Although Fletcher is correct in that the interpretation of PGSs themselves – as the aggregate scores – is context-dependent, this framing of the challenge as an “interpretive” one minimizes the complications. The issue is not merely that the interpretation of a PGS effect is context-dependent in the same way that the interpretation of the label “sick” varies from “good” among a group of high school skateboarders, to “disgusting” among people discussing a ghastly crime, to actually “ill,” as traditionally defined. The challenge is much more complicated as contexts can shape which and how – that is, the magnitude and even direction – individual genetic variants matter. This context-dependent variation is missed in PGSs, which are weighted aggregates of the average effect of a tag-SNP in a specific context estimated from disproportionately European-genetic ancestry samples that are frequently not representative of the underlying population (e.g., wealthier and more highly educated; Curtis, also Burt & Munafò, Reference Burt and Munafò2021).

Consider an analogy. If I create a weighted scale of 100,000 individual characteristics associated with success in football (context) and call it “athletic propensity” (PGS), and then I apply this “athletic propensity” algorithm to different athletic contexts like soccer, tennis, cycling, and rowing, it will surely perform less adequately in predicting success. The lower predictive ability of this “athletic propensity” scale does not indicate that athleticism matters less for soccer or tennis, but rather follows from the fact that these sports (as contexts) differ and with it the nature and salience of various skills and capacities associated with success. Additionally, like an educational-attainment PGS, using an additive, unidimensional scale of “athletic propensity for football” is misleading (see also Richardson; Sarkar). A variety of traits and combinations thereof facilitate success even within the same context, as even the most cursory comparison of characteristics of football players at different positions would suggest. So too for the skills facilitating educational attainment across contexts and even for different subjects like fine art and music studies compared to sociology and psychology or physics and chemistry. For complex social traits, context is intertwined with almost everything at the phenotypic level; these contingencies are exponentially more complicated at the genetic level.

I reiterate this important point because the context-specificity of PGSs continues to be underappreciated and contributes to misuse (see citations in the target article, Curtis; Moore; Sarkar). In particular, existing studies and claims about the potential utility of PGSs are insufficiently attentive to the implications of the context- and condition-dependent nature of PGSs (but see Mostafavi et al., Reference Mostafavi, Harpak, Agarwal, Conley, Pritchard and Przeworski2020). To reiterate, I was not arguing that this context-dependency makes PGSs useless. Rather, I was highlighting how this context-dependency undermines their utility for certain usages – for example, comparing PGSs across contexts to assess variation in “how much genetics matters.”

R1.9. Mine is a scientific not ethical or sociopolitical critique

Controversies about the ethical and sociopolitical implications of including genetics in social science are longstanding. Distinguishing my target article from extant critiques of sociogenomics, I noted that most existing critical engagement focuses on sociopolitical and ethical concerns.Footnote ⁴ These works address questions such as: Is it ethically responsible to study the genetics of social outcomes profoundly shaped by inequality? How should findings from the field of sociogenomics be used? Who stands to benefit? Who will be harmed (or will not benefit)? And do these ethical concerns about this work outweigh the scientific gains?

These are not the questions addressed in my target article. My article focuses on the scientific challenges with PGSs and the implications for social science, as several commentators recognize (e.g., Trejo & Martschenko). Scientific questions I address include, for example: What do PGSs measure? Do PGSs indicate “genetic influences” on complex social traits as they are often used? Given, as I discuss, they do not, what is their scientific utility for enhancing understanding of social behavior?

Nonetheless, some commentators charge me with being motivated by sociopolitical and ethical concerns. Morris et al. allege that I conflate scientific and ethical concerns, pointing as evidence to my conclusion that the “scientific costs outweigh the meager benefits” (emphasis added).Footnote ⁵ My purported non-scientific concerns about PGSs they cite include “obscuring environmental influences,” “perpetuating a flawed concept of genetic potential,” and “wasting resources.”Footnote ⁶ Given the goal of social science of explaining variation in social behavior, inasmuch as PGSs obscure environmental influences and perpetuate a flawed concept of genetic potential (which is my argument), this impedes scientific advancement (i.e., is a scientific cost). Morris et al. disagree, arguing that the limitations I outline are overstated and/or tractable and thus my recommendation to use PGSs sparingly and cautiously in social science is not justified. However, our disagreement is scientific not sociopolitical or ethical.

R2.1. PGSs and evolutionary insights

Two commentaries draw upon evolutionary perspectives to critique or refine my arguments. Focusing on the utility of PGSs, Hong argues that I overlooked their value for “greatly and uniquely” contributing to “the study of genetic evolution in contemporary societies.” In the target article, I necessarily focused on key arguments about the utility of PGSs for social science. In my reading, enhancing understanding of “natural selection in contemporary human populations” is not a common or touted use of PGSs in social science. This is evidenced by the paucity of such studies and the absence of discussion of the utility of PGSs for such purposes in salient overview articles (e.g., Harden & Koellinger, Reference Harden and Koellinger2020; Mills & Tropf, Reference Mills and Tropf2020). Notably, the limitations of PGSs I discuss also impede their utility for the aim of understanding selection in contemporary human populations (e.g., Berg et al., Reference Berg, Harpak, Sinnott-Armstrong, Joergensen, Mostafavi, Field and Coop2019; Sohail et al., Reference Sohail, Maier, Ganna, Bloemendal, Martin, Turchin and Sunyaev2019).

As noted in the target article, sociogenomics research tends to suffer from a deficit of theory, including evolutionary theory. This manifests in the dearth of theoretically driven models and concepts, including phenotype selection (Boardman & Fletcher, Reference Boardman and Fletcher2021; Burt, Reference Burt2022, Reference Burt2023, also Charney). Drawing on evolutionary theory, Ramus concurs with my argument that complex social traits, like educational attainment, are not well-suited for GWASs and PGSs because, in his view, “these complex social outcomes are not phenotypes that are under direct natural selection.” The solution, according to Ramus, is “redirecting geneticists’ attention to stable traits [that] can be defined and can be the target of selection.” Focusing on cognitive outcomes, he suggests that components underlying specific cognitive abilities, such as verbal ability, working memory, or number sense, as well as character traits, like self-control, intrinsic motivation, and grit are more appropriate phenotypes as they are relatively stable and under direct natural selection.

I concur with Ramus that more narrowly defined, stable traits that are the target of direct natural selection are more appropriate traits for genetic analysis than emergent, social achievements like educational attainment. However, I disagree that the cognitive and character traits he identifies meet these criteria – that is, are appropriately viewed as stable traits that are the target of direct natural selection. Scholarship in evolutionary-developmental behavioral science undermines the notion that such cognitive traits are stable or under direct natural selection (as being uniformly fitness promoting). After all, we did not evolve to maximize wealth, educational attainment, happiness, or even health but to survive and reproduce. Moreover, evolutionary-developmental models direct theoretical attention away from the single-“best” traits (e.g., future time orientation, conscientiousness, working memory, task persistence) and toward context- and condition-dependent optimal traits (see, e.g., Belsky, Steinberg, & Draper, Reference Belsky, Steinberg and Draper1991; Chisholm, Reference Chisholm1999; Ellis et al., Reference Ellis, Abrams, Masten, Sternberg, Tottenham and Frankenhuis2022; West-Eberhard, Reference West-Eberhard2003). For the intelligence or educational-attainment traits of interest to Ramus, this implies rather than a one-context fits all model of intelligence, an ecologically contingent notion of “adaptive intelligence” (also “successful intelligence” or “multiple intelligences”) (Gardner, Reference Gardner2017; Sternberg, Reference Sternberg2019).Footnote ⁷

Regarding stability, evolutionary-developmental models also recognize that contexts are constantly changing and the future is uncertain (Boyce & Ellis, Reference Boyce and Ellis2005). Given this reality, humans have evolved neurobiological mechanisms facilitating adaptive phenotype plasticity in response to external environmental factors and relative condition (e.g., relative health, status) (Del Giudice, Ellis, & Shirtcliff, Reference Del Giudice, Ellis and Shirtcliff2011). Consistent with this model, a growing body of research over the past two decades demonstrates that rather than being stable, many cognitive and character traits are malleableFootnote ⁸ in response to environmental insults (social and physical) (Burt, Lei, & Simons, Reference Burt, Lei and Simons2017; Pepper & Nettle, Reference Pepper and Nettle2017; Shonkoff & Phillips, Reference Shonkoff and Phillips2000) and supports, including interventions, training, and even education (Brinch & Galloway, Reference Brinch and Galloway2012; Brody et al., Reference Brody, McBride Murry, McNair, Chen, Gibbons, Gerrard and Ashby Wills2005; Harrison et al., Reference Harrison, Shipstead, Hicks, Hambrick, Redick and Engle2013; Hegelund et al., Reference Hegelund, Grønkjær, Osler, Dammeyer, Flensborg-Madsen and Mortensen2020; Jaeggi, Buschkuehl, Jonides, & Perrig, Reference Jaeggi, Buschkuehl, Jonides and Perrig2008; Kautz, Heckman, Diris, Ter Weel, & Borghans, Reference Kautz, Heckman, Diris, Ter Weel and Borghans2014), as Charney also notes.

In sum, although Ramus and I agree that several “conveniently available” social outcomes are not appropriate phenotypes for genetic analyses, we disagree that the solution is studying specific cognitive abilities assuming these are relatively context-insensitive stable traits under direct natural selection.

R2.2. Downward causation = social causation

A few commentators disagree with my arguments about downward causation creating what I call artificial – as socially produced – genetic associations. Importantly, commentators do not dispute that PGSs unavoidably capture social forces like discrimination. What is debated is whether these forces are properly interpreted as being social (my argument) or as genetic effects (dissenters' argument). Before tackling this debate, I briefly address an extension of my discussion of downward causation by two commentators who agree with my arguments.

In a concurring commentary, Merchant helpfully offers a more formal definition of artificial genetic associations produced by downward causation as “any association between genomic variants and a given outcome that is forged through social practices rather than biochemical pathways.” In a variation of my argument, both Merchant and Charney suggest that downward causation is a form of population stratification (PS, i.e., population substructure produced phenotype stratification) (see sect. 5.1.1). Traditionally defined, PS reflects random allele frequency differences between subgroups that are associated with, but usually irrelevant to, the trait. In contrast, genetic variant – trait associations reflecting downward causation, which I label artificial genetic associations, need not be differentiated by population subgroups and are relevant to the phenotype, because of (at least in part) social not genetic causation. An example is an association between genetic variants associated with height or perceived attractiveness and income or educational attainment, partly rooted in the social tendency to favor more attractive and taller people. Given this, as I note in the target article (and contra Morris et al.) existing statistical methods designed to mitigate population stratification confounding (e.g., within-family methods) are not able to correct for artificial genetic associations reflecting downward causation. To be sure, the concepts I introduced would benefit from deeper consideration, further revisions, even new labels, and I hope these issues are addressed in future work.

Dissenting commentaries argue that what I refer to as artificial genetic associations are appropriately viewed as genetic causes that are environmentally mediated (“indirect genetic effects” or “vertical pleiotropy”) (Keller,Footnote ⁹ Xia & Hill, Trejo & Martschenko, Burke). On their account, again using the example of racial discrimination/colorism, skin pigmentation alleles cause skin pigmentation differences, which cause differences in racial discrimination experiences, which cause disparate outcomes. In their view, this makes the experience of racial discrimination and disparate outcomes caused (in part) by racial discrimination (e.g., higher allostatic load, depression, criminal behavior, income, and educational attainment) caused by and thus “indirect genetic effects” of skin pigmentation alleles.

I disagree. In my view, the cause of racial discrimination based (in part) on skin pigmentation is not genetic variants related to skin pigmentation but social forces (racism/colorism) that act “downward” on genetically influenced differences such as skin pigmentation (see, e.g., Burt, Reference Burt2018; Burt, Simons, & Gibbons, Reference Burt, Simons and Gibbons2012; also Merchant). There is no biological pathway upward from skin pigmentation alleles to racial animus, discriminatory treatment, or racial segregation. In my view, this idea that racial discrimination/colorism is caused by skin pigmentation alleles is rooted in a misguided gene-centric worldview where causality is something that only occurs in one direction: Upward from lower-level parts to higher-level entities. In contrast, I adopt an ontologically pluralist view, which recognizes that higher-level emergent phenomena (social structures) can operate causally on lower-level factors (see, e.g., Burt, Reference Burt2023; Dupré, Reference Dupré2012).Footnote ¹⁰

Thus, although commentators and I agree that PGSs unavoidably capture social forces like discrimination, we will have to agree to disagree whether the resulting genetic associations are properly viewed as artificial (social) or genetic influences. Trejo & Martschenko submit that we need new language to describe such relationships. I do not disagree. In the meantime, however, we likely agree that we would benefit from employing existing concepts with greater care and accuracy.

R2.3. (In)tractability of limitations

As noted, commentaries generally concur with the challenges and limitations that I outline for PGSs as well as the need for appropriate interpretation. Disagreement centers on the tractability of the limitations and my recommendation to use PGSs sparingly and cautiously. Several commentators argue that PGS limitations – especially environmental confounding – are adequately mitigated with sophisticated methods and do not undermine the utility of PGSs for social science (e.g., Keller; Morris et al.; Zietsch et al.). As discussed in the target article, although I agree that several of these methods substantially mitigate confounding, I disagree that the environmental confounding of PGSs is an issue that can be overcome with statistical genetic methods.

I shall not repeat my arguments from the target article explaining why methodological limitations and gene–environment interplay in development result in unavoidably environmentally confounded PGSs (see also Archer & Lavie; Charney; Curtis; Moore; Richardson; Trejo & Martschenko). Nor shall I reiterate, in response to Morris et al. who suggest these problems are “scientifically tractable issues that have been substantially addressed,” the limitations, questionable assumptions, and new challenges accompanying these novel methods (see Zaidi & Mathieson, Reference Zaidi and Mathieson2020; also, Boardman & Fletcher, Reference Boardman and Fletcher2021; Charney, Reference Charney2022; Domingue & Fletcher, Reference Domingue and Fletcher2020; Young et al., Reference Young, Nehzati, Benonisdottir, Okbay, Jayashankar, Lee and Kong2022). Instead, I point readers to section 5 of the target article, where I discuss these issues along with section R.2.4.

Conversely, several commentators suggest that the problems with PGSs are worse than I outline, for example, arguing that PGSs have “negative utility” and that social scientists should “steer clear” of them (e.g., Curtis; Richardson; Sarkar). Although my position is admittedly closer to the “useless” than the “very useful” arguments of some commentators (e.g., Alexander et al.), my arguments fall in the middle. I neither suggest that PGSs should never be used or can never be useful in social science. I argued that their utility is narrow. My position was solidified by advocates' meager examples of utility given limitations.

Rather than tractable, my view and that of many other commentators is that the various difficulties plaguing PGSs as reflecting “genetic influences” for complex social traits are insurmountable. Leaving aside the question of whether disentangling socio-environmental influences from authentic genetic signals influencing complex social traits is possible in principle – and I and many commentators think not – we are nowhere near there yet. Some commentators admit as much but suggest “this is no reason for despair” but rather we should continue to try to develop innovative strategies to overcome these limitations (Keller; Morris et al.). I neither counsel despair nor oppose the development of innovative strategies, as implied. In fact, I encouraged the use of more robust, innovative strategies (e.g., sibling difference GWASs) in the target article. What I oppose is the misuse of PGSs, as, for example, representing genetic (vs. environmental) influences.

R2.4. Challenges with novel advanced methods

Although scrutinizing various specific methods designed to mitigate confounding in GWASs and PGSs is impracticable here, I nevertheless wish to briefly address Keller's response to my questioning strong assumptions underlying a popular contemporary approach to mitigate confounding in GWASs (LDSC; Bulik-Sullivan et al., Reference Bulik-Sullivan, Loh, Finucane, Ripke, Yang, Patterson and Neale2015) and PGSs creation (e.g., LDPred; Vilhjálmsson et al., Reference Vilhjálmsson, Yang, Finucane, Gusev, Lindström, Ripke and Zheng2015). I wrote that in all applications of LDPred that I have seen, studies assume “that all SNPs are causal,” which “is curiously not defended anywhere to [my] knowledge,” and “not consistent with available empirical evidence.” Keller implies that my questioning such assumptions reflects scientific naiveté or unreasonableness, given that “models are not meant to mirror reality.”Footnote ¹¹ In his view, at issue is “the degree to which results are biased and whether this bias matters.”

I agree. This is why much of my target article was focused on explaining the degree and, especially, import of these biases for social outcomes. Although we cannot know the precise amount of bias, given the nature of development and methodological limitations, the evidence we do have suggests it is both substantial (e.g., as little as 1/3 of the EA PGS effect is attributable to “direct genetic effects”; Okbay et al., Reference Okbay, Wu, Wang, Jayashankar, Bennett, Nehzati and Gjorgjieva2022), insufficiently corrected (e.g., Young et al., Reference Young, Nehzati, Benonisdottir, Okbay, Jayashankar, Lee and Kong2022; Zaidi & Mathieson, Reference Zaidi and Mathieson2020), and matters for understanding (Berg et al., Reference Berg, Harpak, Sinnott-Armstrong, Joergensen, Mostafavi, Field and Coop2019; Haworth et al., Reference Haworth, Mitchell, Corbin, Wade, Dudding, Budu-Aggrey and Smith2019; Sohail et al., Reference Sohail, Maier, Ganna, Bloemendal, Martin, Turchin and Sunyaev2019). To be clear, my point was not about any one of the specific dubious assumptions involved in these various methods, but more broadly on the reliance on assumptions, sometimes strong, at almost every level of analysis that are questionable, rarely justified, often obscured, and frequently unknown by non-experts who apply the products of these techniques in social science applications and present the resulting outcomes as being “corrected for” environmental confounding and other issues.

R2.5. Wrong but useful for prediction?

Several commentators argued that the environmental confounding of PGSs does not undermine their utility for risk prediction (e.g., Keller). On this view, PGSs are valuable as they offer practically useful incremental prediction that is independent of traditional social measures (Alexander et al.; Moreau & Wiebels). That the prediction accuracy of PGSs is not undermined by environmental confounding is, of course, true, with the usual caveats (e.g., context-dependency). However, that PGSs have actionable utility for predicting individual risks for complex social traits at the current state of the science is widely recognized to be misguided (but see Plomin & Von Stumm, Reference Plomin and Von Stumm2022). As noted in the target article, most scholars, including ardent supporters of PGSs for social science, “agree that PGSs do not predict complex social outcomes with any degree of efficacy or accuracy and, therefore, should not be used for individual prediction” (citations omitted, see also Moore; Turkheimer, Reference Turkheimer2015, Reference Turkheimer2019). Moreover, as noted in the target article, research suggests that the incremental predictive efficacy of social science PGSs independent of available or easily attainable phenotypic measures, such as prior grades or parental educational attainment, is too weak to be of practical utility (e.g., Morris, Davies, & Smith, Reference Morris, Davies and Smith2020).

Furthermore, we still have the “ancestry-specific” “portability problem” (see sects. 2.3 and 3.1). As Curtis elaborated, because PGSs are tailored to specific ancestral populations in certain contexts, different PGSs would have to be created for different ancestries (Martin et al., Reference Martin, Gignoux, Walters, Wojcik, Neale, Gravel and Kenny2017). Further complicating matters, the human population cannot be neatly demarcated into different ancestral groups. “In reality, there is no bright line demarcating comparisons ‘within’ versus ‘between’ ancestries: there is a giant family tree of humanity, and people who share more ancestral paths through it than others, and more similar environments than others” (Coop & Przeworski, Reference Coop and Przeworski2022, p. 850).

R2.6. Incautious usage

Although equivocating somewhat, some commentators seem to suggest that most social science studies use PGSs appropriately (Alexander et al.; Morris et al.). Given their position, you might think that most social science studies eschew depictions of PGSs as “genetic influences” and adopt rigorous methods to mitigate confounding. You would be mistaken. I refer the reader to sections 5 and 6 of the target article where I highlight several misguided applications of PGSs as “genetic propensity,” “genetic influences,” even “genetic endowment of educational attainment.” On my reading, most – but certainly not all – current social science applications use PGSs as “genetic influences.”

Also noted in the target article, only a paucity of sociogenomics studies employ PGSs using the most robust available methods designed to attenuate environmental confounding (including within-family studies highlighted by Morris et al., Keller, Zietsch et al., and Madole & Harden, Reference Madole and Harden2023).Footnote ¹² Some studies, cited in the target article, fail to employ even basic, necessary adjustments in the creation of PGSs (e.g., LD pruning). In still other cases, studies provide insufficient detail on PGSs necessary for an evaluation, for example, saying only that the PGS was “based on the effect sizes from the most recent GWAS of educational attainment.”

In short, incautious application with insufficient description and/or corrections, overinterpretation, and misinterpretation of PGSs as genetic propensity is a significant problem (see also Fletcher). Often problematic use is accompanied by overinterpretation, hype of weak evidence, and promissory notes. This brings us to the commentary by Alexander et al. whose views starkly diverge from mine.

R2.7. Hype and disparate interpretations

In their commentary, Alexander et al. offer a “defence of the immediate practical utility of PGSs for maximizing trait prediction” and for “advancing etiological understanding” of complex social traits.Footnote ¹³ In my view, their defense epitomizes what I consider to be an incautious hyping of PGSs based on misinterpretation and which my article is intended to counter. However, I am grateful for their response, which encourages direct engagement with opposing claims for the utility of PGSs. Their arguments sound compelling but are, in my view, partial, misguided, or based on questionable assumptions or weak evidence.

To the utility of PGSs for prediction, Alexander et al. assert that: “it is only legitimate to assume that PGSs [for complex social traits] are just about to unfold their full predictive potential.” For all the reasons I have discussed, I disagree.

Alexander et al.'s defense of the utility of PGSs for etiologic understanding crumbles under scrutiny. Space does not permit a critical citation-by-citation analysis of the support they present for their perspective, so one example will have to suffice. They write:

This evidence is presented to contradict my claim that PGSs lack utility for identifying specific biological pathways to social outcomes. However, Alexander et al. mischaracterize these studies. Neither study examines “structural brain changes”; both studies analyze brain measurements at a single time point. For example, Elliott et al. (Reference Elliott, Belsky, Anderson, Corcoran, Ge, Knodt and Ireland2019) test the hypothesis that an educational-attainment PGS influences individual differences in intelligence by “contributing to the development of larger brains,” which could “constitute a biological pathway linking genetic variation to differences in intelligence and educational attainment” (p. 3497). Despite what Alexander et al. imply, the results are not particularly noteworthy. The educational-attainment PGS explained less than half of 1% of the variance in brain size. Not surprisingly, the mediation analyses revealed extremely weak indirect effect sizes (b = 0.01), with significant effects observed in only two of the four samples. Moreover, Elliott et al. did not follow recommended protocols to mitigate biases in PGSs as discussed in the target article.Footnote ¹⁴ Given the paltry effect sizes, one might reasonably expect that these estimates would be naught if such adjustments for confounding had been implemented.

Leaving aside concerns about methodological limitations, these cited studies illustrate my concern that PGSs are being used in a manner that obscures potentially relevant socio-environmental influences (pace Morris et al., for scientific not ethical reasons). Alexander et al., following Elliott et al., assume that the educational-attainment PGS causes structural brain differences, which cause educational-attainment differences. However, this causal ordering cannot be assumed given the significant environmental confounding of PGSs (see also Coop & Przeworski, Reference Coop and Przeworski2022). Our brains are co-constructed from combined genetic and environmental influences in development. Studies of neuroplasticity in human and rodents demonstrate that a variety of socio-environmental forces alter the structure and function of the brain and with it our ability to respond to ongoing challenges and opportunities (e.g., Glasper & Neigh, Reference Glasper and Neigh2019; Kokras et al., Reference Kokras, Sotiropoulos, Besinis, Tzouveka, Almeida, Sousa and Dalla2019; Leuner, Glasper, & Gould, Reference Leuner, Glasper and Gould2010; Sweatt, Reference Sweatt2016). We all know the life experiences of the average person who gets a Ph.D., J.D., or M.D. and the person who does not graduate high school are very different. These different experiences and contexts of development, which include the experience of education itself, are neurobiologically embodied. For these reasons, interpreting differences in brain structure or cognitive test performances as reflecting causal genetic differences based on PGS associations is unjustified.

In sum, where Alexander et al. dispute my arguments and assert that “PGSs hold great potential for both better prediction and understanding of complex traits in social science” (emphasis in original), I find their evidence problematic and uncompelling, and I strongly, albeit respectfully, disagree.

R3.1. Downplaying wider social forces

Expanding on a brief critique (see note 10 of the target article), Merchant argues that so-called “dynastic effects” (or “indirect genetic effects”) are “undertheorized and underexplored” in GWAS/PGS studies and “often assumed to describe the direct genetic effects of the parents’ genotypes on their parenting” (see sect. 5.1.2; also Coop & Przeworski, Reference Coop and Przeworski2022; Young et al., Reference Young, Nehzati, Benonisdottir, Okbay, Jayashankar, Lee and Kong2022). We agree that the scholarly discourse around this form of environmental confounding, often framed as “genetic nurture,” is overly focused on parenting in a manner that obscures the effect of wider socio-cultural forces. A rich body of social science research highlights the manifold ways that social forces beyond parenting – for example, schools, neighborhoods, social networks – influence important life outcomes (see, e.g., Leventhal & Brooks-Gunn, Reference Leventhal and Brooks-Gunn2000; Sampson, Raudenbush, & Earls, Reference Sampson, Raudenbush and Earls1997; Simons & Burt, Reference Simons and Burt2011). Although several recent studies have usefully explicated and empirically demonstrated how PGSs capture these broader social forces (Abdellaoui, Dolan, Verweij, & Nivard, Reference Abdellaoui, Dolan, Verweij and Nivard2022; Young, Benonisdottir, Przeworski, & Kong, Reference Young, Benonisdottir, Przeworski and Kong2019), a general tendency to narrowly focus on parenting remains.

R3.2. Broader untenable assumptions

Several commentators expand my critique to highlight broader questionable assumptions underlying PGS studies, including those that have long been critiqued. Both Richardson and Archer & Lavie criticize the oversimplified additivity assumption and overall “gene-centric approach” underlying much sociogenomics work. They emphasize the dynamic responsivity of cells and organisms to their environments and the role of the genome as a resource facilitating such responsivity. Both underscore the need to replace the “gene-centric approach” with a more “biologically realistic one” (Richardson; also Richardson & Jones, Reference Richardson and Jones2019).

In his commentary, Sarkar writes that arguments that PGSs have sidestepped the invalid assumptions and environmental confounding of prior eras of social science genetics “are not credible”; I concur. Sarkar helpfully points out that social scientists are not alone in their concerns about the use of GWASs and PGSs for complex social outcomes. After noting that 1970s critics of heritability studies “read like a ‘Who's Who’ of theoretical population and quantitative” geneticists, Sarkar notes that in recent years prominent geneticists have criticized sociogenomics. This echoes Bliss's (Reference Bliss2018) observation that although the behavioral scientists she interviewed “spoke highly of social genomics,” she found an “almost polar opposite response from mainstream genome scientists” (p. 157). For example, Bliss wrote that a recent past president of the American Society for Human Genetics expressed concern “about the ways in which social genomic researchers were characterizing social phenomena as medically relevant traits” and remarked that “he hardly believed that any serious scientist would take social genomics seriously” (p. 158). This recognition that contemporary genomics scholars have published critiques of oversimplified assumptions and/or expressed concerns about the application of these genomic tools to study complex social traits (e.g., Coop & Przeworski, Reference Coop and Przeworski2022; Nelson, Pettersson, & Carlborg, Reference Nelson, Pettersson and Carlborg2013; Rosenberg, Edge, Pritchard, & Feldman, Reference Rosenberg, Edge, Pritchard and Feldman2019) is a rebuttal to the tacit or explicit suggestion that critics of PGSs are invariably naïve and/or politically motivated social scientists.

R3.3. Socio-environmental epigenetics

Moore and Gooding & Auger highlight interesting, important research linking social environmental exposures to epigenetic mechanisms regulating gene expression as a crucial aspect of development and a challenge to PGSs. Both commentaries suggest I was remiss to not discuss environmental epigenetics. Despite their centrality to development, plasticity, and individual differences, epigenetic mechanisms, per se, are not a challenge to PGSs, which is why I did not discuss. Specifically, to the extent that environmental influences inducing epigenetic marks are uncorrelated with genotype, they do not confound PGS associations. In the same way that the correlation between parental income and child educational attainment is not undermined by the fact that parents devote money and other monetary resources to children differently, the correlation between a PGS and some outcome is not undermined by epigenetic mechanisms differentially regulating gene expression. Conversely, if environmentally induced epigenetic marks are correlated with PGSs, they reflect one or more of the forms of environmental confounding that I discuss – population stratification, familial confounding, and downward causation – and/or genetic influences.

R3.4. Alternative approaches to overcome limitations

Several commentators concur with my main arguments but suggest shifts in approach or additional analyses to overcome limitations of PGSs. Highlighting the lack of consideration of GWASs in the context of development, Freitag & Kelsey recommend adopting a “developmental dynamic” approach and the inclusion of “wider age populations” “to gain a holistic understanding of the biology underlying developmental outcomes.” I agree that development is insufficiently considered in GWASs and PGSs. However, for all the reasons I outline in my target article, I believe that the limited utility of GWASs and PGSs for complex social traits remain with a developmental perspective. Even so, when considering the application of GWASs and PGSs to disease traits while also recognizing that these studies will continue in various fields regardless of what I say about them, I agree with Freitag & Kelsey's recommendation that adopting a developmental approach would be beneficial.

Addressing the limitations of PGSs for identifying causal variants and biological pathways, Fox suggests that strategies to identify “rare variants of large effect” might be useful. Insight from such approaches, Fox argues, could provide knowledge on biological pathways and detrimental mutations, which could be used “to repair or counteract the deleterious effects of the mutation.” I agree that rare variants approaches, despite their challenges, have utility for biomedical conditions, like congenital deafness and cystic fibrosis, which reflect a biological dysfunction. However, for normally varying social outcomes, like educational attainment, income, smoking behavior, and same-sex sex, I do not share Fox's enthusiasm for rare variant approaches. This is because this approach necessarily rests on the assumption that these social traits reflect a biological deficit produced by a rare variant of large effect. However, as discussed in the target article, the assumption that normal variation in complex social outcomes – for example, “only” graduating high school versus graduating college – reflects a biological deficit is unjustified (see, e.g., Burt, Reference Burt2023).

Concurring with my arguments that PGSs lack utility for providing biological insights, Nephew, Murgatroyd, Polcari, Santos, & Incollingo Rodriguez (Nephew et al.) argue that augmenting PGS studies with “functional (transcriptome, methylome, metabolome) and/or multimodal genetic data,” can enhance understanding of biological pathways linking genetics and environments to complex traits and knowledge of “the genetics of social phenomena.” I agree with Nephew et al. that sociogenomics studies in social science could be enhanced with such functional genetic data and physiological measurements. However, I remain skeptical of the utility of PGSs even with “additional, more functional assessments of genetic context” for the reasons discussed in the target article. Both because PGSs are environmentally confounded and because they aggregate millions of mostly non-causal variants with miniscule effects, PGSs have limited utility for providing mechanistic insights into complex social traits even when combined with functional and multimodal genetic data. Although I appreciate Nephew et al. drawing attention to these data and possibilities, in my view, these neither rescue PGSs from their limitations nor significantly expand their utility for social outcomes at the current state of the science.