Challenging the utility of polygenic scores for social science: Environmental confounding, downward causation, and unknown biology

Abstract

The sociogenomics revolution is upon us, we are told. Whether revolutionary or not, sociogenomics is poised to flourish given the ease of incorporating polygenic scores (or PGSs) as “genetic propensities” for complex traits into social science research. Pointing to evidence of ubiquitous heritability and the accessibility of genetic data, scholars have argued that social scientists not only have an opportunity but a duty to add PGSs to social science research. Social science research that ignores genetics is, some proponents argue, at best partial and likely scientifically flawed, misleading, and wasteful. Here, I challenge arguments about the value of genetics for social science and with it the claimed necessity of incorporating PGSs into social science models as measures of genetic influences. In so doing, I discuss the impracticability of distinguishing genetic influences from environmental influences because of non-causal gene–environment correlations, especially population stratification, familial confounding, and downward causation. I explain how environmental effects masquerade as genetic influences in PGSs, which undermines their raison d’être as measures of genetic propensity, especially for complex socially contingent behaviors that are the subject of sociogenomics. Additionally, I draw attention to the partial, unknown biology, while highlighting the persistence of an implicit, unavoidable reductionist genes versus environments approach. Leaving sociopolitical and ethical concerns aside, I argue that the potential scientific rewards of adding PGSs to social science are few and greatly overstated and the scientific costs, which include obscuring structural disadvantages and cultural influences, outweigh these meager benefits for most social science applications.

1. Introduction

Extraordinary techno-scientific advances over the past two decades have transformed human genetics. Scientists are now able to measure several million genetic variants across the genome (i.e., genome-wide) relatively cheaply (<$100) and efficiently with automated pipelines. Consequently, millions of individuals have been genotyped, which is the measurement of preselected variants, across the genome. Over the past decade, genome-wide association studies (GWASs), in which a phenotype (trait) is regressed on each of the millions of genetic variants with a few controls, have become the predominant method to statistically estimate genetic associations with genome-wide data and increasingly large datasets. Thousands of GWASs have been performed, identifying hundreds of thousands of significant associations with a multitude of traits and disease states (e.g., Buniello et al., 2019).

These molecular and computational innovations have launched the new science of sociogenomics, characterized by the application of cutting-edge statistical genetic tools and measures to social outcomes. In recent years, social scientists have teamed with biostatisticians and formed large consortia to conduct GWASs on complex social outcomes, such as educational attainment (Lee et al., 2018), same-sex sexual behavior (Ganna et al., 2019), number of children (Barban et al., 2016), and income (Hill et al., 2019), with large (and growing) genetic datasets. In sociogenomics, as elsewhere, GWAS results are commonly used to create genetic summary scores, known as polygenic scores (PGSs), representing the (additive) genetic propensity for some trait or behavior (e.g., years of educational attainment completed). Preconstructed PGSs have been incorporated into widely used social science datasets, such as the Add Health Study and Health and Retirement Study (HRS), to be dropped into models “just like any other variable,” no genetic expertise required (Braudt, 2018). Given the availability and increased acceptance of genetics in social science, sociogenomics is poised to flourish.

This new “golden age” of sociogenomics filled the void left by the recent demise of the candidate gene × environment era, which was, by and large, a spectacular failure because of methodological limitations and an oversimplified biology (see Charney, 2022; Dick et al., 2015). Suggesting the candidate gene-era “should be a cautionary tale,” psychiatric geneticist Matthew Keller asked: “How on Earth could we have spent 20 years and hundreds of millions of dollars studying pure noise?” (quoted in Yong, 2019; cited in Charney, 2022). With adjustments for multiple testing, attention to statistical power and large samples, and emphasis on replication, among other revisions, this nascent sociogenomics approach has addressed several methodological limitations plaguing the candidate gene approach. As a result, sociogenomics findings are touted as methodologically robust. Advocates are especially bullish about the potential of PGSs, which, they argue “just work” (i.e., are statistically significant genetic predictors) and have several potential social science applications that break through the stale, outdated nature versus nurture debate, on the one hand, and the neglect of genetics (or assumption of “genetic sameness”) on the other (e.g., Belsky & Harden, 2019; Conley, 2016; Conley & Fletcher, 2017; Freese, 2018).

Further still, many sociogenomicists encourage other behavioral scientists to incorporate PGSs into their research (e.g., Braudt, 2018; Cesarini & Visscher, 2017; Harden, 2021b; Mills & Tropf, 2020). Pointing to evidence of ubiquitous heritability, the widening availability of genetic data, and the ease of incorporating PGSs into quantitative research, these scholars urge social scientists to incorporate genetics or risk losing out (e.g., Conley, 2016; Mills & Tropf, 2020). Others take an even stronger stance and emphasize not only the potential but also the necessity of incorporating genetics into social science, arguing that social science research that neglects genetics is, at best, partial and potentially flawed and misleading (e.g., Braudt, 2018; Harden, 2021a; Hart, Little, & van Bergen, 2021; Kweon et al., 2020). In her recent book, The Genetic Lottery, Harden (2021a) contends that social science sans genetics wastes time, resources, attention, and effort; supports misguided models of human behavior; and misinforms policies, causing still further damage. This neglect of unmeasured genetic heterogeneity makes social science research vulnerable to sweeping dismissals from other scientists (Freese, 2008) or political extremists (Harden, 2021a).

Yet it remains the case that only a paucity of behavioral science research includes genetics. This “neglect of genetics” is, some proponents have argued, not because of valid scientific reasons but of an ideologically motivated “tacit collusion” to ignore genetic differences between people among social scientists (Freese, 2018; Harden, 2021a; Wright & Cullen, 2012). Harden (2021b) argues that this alleged tacit collusion is not just misguided or morally “wrong in the way that jaywalking is wrong” but, given the scientific warrant to include genetics, it is “wrong in the way that robbing banks is wrong.” Harden avers that “Failing to take genetics seriously is a scientific practice that pervasively undermines our stated goal of understanding society so that we can improve it” (p. 186). On this view, if progressive social scientists really want to ameliorate inequality, they need to get with the science and add genetics to their research.

Here, I scrutinize proponents' arguments about the significant value of PGSs for social science and with it the need to incorporate genetics into social science models. I do so not by questioning the ethical or sociopolitical implications of this work, as is common, but by scrutinizing the science of sociogenomics. Specifically, I focus on the utility of PGSs for social science and the key premises underlying their use as measures of “genetic propensities” for behavioral differences. Drawing on contemporary statistical genetic research, I explain how methodological limitations produce environmentally confounded PGSs. I emphasize that 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 when attempting to map layered and contingent social behaviors, like educational attainment, to a score representing a linear summation of base-pair differences, which themselves represent an entirely different set of layered contingencies. I explain why this inevitable environmental confounding of PGSs for complex social traits undermines their use as “genetic influences on” or “genetic potential for” social traits and achievements – as is common. After outlining the limitations of current sociogenomics methodologies, I consider the practical implications by examining several existing applications of PGSs to social science and their substantive contributions.

My explicit aim is to challenge the claim that genomics has much to offer social science, so much so that social science sans genetics is fatally flawed, scientifically indefensible, and possibly even morally suspect. I argue that, leaving sociopolitical risks aside, the potential scientific rewards are few and greatly overstated, and the potential scientific costs – obscuring environmental influences, perpetuating a flawed concept of genetic potential for social behaviors and achievements, and wasting resources – outweigh these meager benefits for most applications. I am not alone in my concerns, and not all sociogenomics practitioners are sold on the touted benefits of PGSs; however, cautious and skeptical arguments are invariably drowned out by enthusiastic hype and promissory notes. Much of the excitement around sociogenomics comes from the application of these new measures and techniques without clearly acknowledging limitations or accounting for well-known biases. Given this situation, my goal is to draw attention to and explicate the limitations of sociogenomics methods, especially PGSs, that vitiate their utility in the behavioral sciences.

Before moving forward, a few remarks about the larger backdrop are in order. Most historical and current critiques of social science genetics emphasize sociopolitical or ethical considerations rather than scientific concerns. This focus is because of both socio-historical reasons (racist and eugenicist applications and/or interpretations of this work in the past) and the fact that the advanced biology and statistical genetic methods of sociogenomics are well outside the bailiwick of most social scientists (and thus lack of expertise and skills to critically engage with this research). Here, I do not concentrate on sociopolitical or ethical concerns about sociogenomics research, because existing scholarship addresses these issues, acknowledging historical misuses with some atrocious results and highlighting the potential misrepresentation of sociogenomics findings to support genetic determinist and inferiorizing claims (e.g., Bliss, 2018; Duster, 2015; Harden, 2021a; Herd, Mills, & Dowd, 2021; Martschenko, Trejo, & Domingue, 2019). Although I share these concerns, my current focus is scrutinizing sociogenomics with the aim of fostering a dialogue that focuses squarely on the science.

This critical analysis proceeds in several parts. First, I provide a brief overview of the genetic and statistical genetic fundamentals necessary to understand these models and their limitations, recognizing that sometimes social scientists' lack of expertise in genetics and statistical genetics methods is a key barrier to engagement. (Readers wholly unfamiliar with genetic concepts can see the primer in Appendix A, whereas those familiar with sociogenomics concepts and methodologies may opt to jump to sect. 4.) Next, I describe proponents' key arguments for the value of adding genetics to social science. I then discuss and critique the key premises underlying these arguments, with a particular focus on explicating intractable environmental confounding in GWAS associations and PGSs.¹ I then explain how these challenges undermine the utility of PGSs as measures of genetic influences or potential. I conclude by offering several suggestions for the field.

2. A primer on genomics

At present preconstructed polygenic scores (PGSs) are available in several accessible social science datasets available to be dropped into models just like any other variable (Braudt, 2018; Mills & Tropf, 2020). Properly interpreting the meaning and challenges of PGSs, however, requires some knowledge of what PGSs capture, what they don't, and what these models assume.

2.1 Basic genetic concepts in sociogenomics

See Table 1.

Table 1. Glossary, acronyms, and definitions for sociogenomics terms

* The term copy number variant used to be applied to all variants that had a variable number of tandem repeats, including short tandem repeats, such as the microsatellite in (D) where there are 12 or 11 copies of the CA dinucleotide. In genome sequencing projects, the term is reserved for large size changes only, such as variable numbers of repeats exceeding 50 nucleotides in the case of the 1000 Genomes Project. (Strachan & Read, 2018)

2.2 Genetic variants, function, and prevalence

Given that sociogenomics focuses on genetic variation among people, understanding the type, prevalence, and distribution of human variation is necessary to understand what is and is not being captured in these studies. Genetic variants can be classified into three types: (1) single-nucleotide variants (SNVs), which are single-base changes (G → A); (2) indels, which are insertions of base pairs or deletions up to 50 bp and often involve tandem repeating units (e.g., GATA repeated 2–8 times); and, (3) structural variants (SVs), which are DNA rearrangements (deletions, duplications, or inversions) ranging from 50 bp to more than a million base pairs (1 Mbp). As discussed below, GWASs and PGSs analyze a subset of “common” single-nucleotide variants, known as single-nucleotide polymorphisms (SNPs), where common usually means present in at least 1% of the population (see Appendix A for more details).

Human genetic variation is extensive – all genetic variants compatible with life are likely represented in some individual living today (McClellan & King, 2010). Comparing the genomes of any two humans around the world, we would typically find between 3 and 4.5 million genetic differences between them or approximately 1 variant every 800 bases.² Most of these genetic variants are SNPs and are non-functional. That is, they have no effects on biological functioning or differences between people. Obviously, only functional variants contribute to differences between people. Although some genetic variation is debilitating, most genetic variation in a given genome is benign, ancient, and common.

In contrast, functional variants are those that either alter gene product (the protein produced) or gene dosage (e.g., the amount of protein produced). As an example of the former, the SLC24A5 gene encodes a protein involved in epidermal melanogenesis and skin pigmentation through its intracellular potassium-dependent exchanger activity (Ginger et al., 2008). Several thousand years ago, a G → A mutation in SLC24A5 occurred among people migrating from Africa to Europe. This variant, which changes the encoded amino acid from alanine to threonine, disrupts melanogenesis and thereby results in lighter skin tone (Lamason et al., 2005). Other variants can affect function not by changing the protein produced but, for example, by affecting the binding sites for various RNAs in a manner that reduces or increases transcription and thereby contributes to trait differences by altering gene dosage (the production of too much or too little of the functional protein).

All three variant types can be functional and contribute to differences between people. Although rare compared to SNVs and indels, evidence suggests that SVs have a disproportionate role in shaping human differences compared to other variants (Chiang et al., 2017; Collins et al., 2020; Takumi & Tamada, 2018). SVs can involve multiple copies of genes or the deletion of a gene and thus influence gene dosage. Sudmant et al. (2015) estimated that SVs were 50 times more likely than SNVs to affect gene expression and three times more likely to be associated with a trait difference than an SNV.

Despite being the extreme minority among the variants we carry, we all have thousands of functional variants in our genomes. A recent deep sequencing study of diverse ancestries identified approximately 11,700 functional variants per individual genome (Taliun et al., 2021). Another study of roughly half a million people in the United Kingdom, Backman et al. (2021) observed an average of ~600 variants, including 50 putative loss-of-function (pLOFs) variants, per gene. Backman et al. (2021) estimated that on average each of us carries 214 pLOF variants as “defective” gene copies. Although this variation is non-trivial, recall that we receive two copies of our genes (excepting the male-specific genes on the Y chromosome). In addition, a host of cellular mechanisms, including those shaping gene expression, compensate for many of these loss-of-function variants and facilitate robustness to functional mutations by, for example, up-regulating transcription (thereby producing more mRNA transcripts) and slowing the rate of mRNA decay (thereby increasing the ability of the cell to generate more polypeptides from the same mRNA transcript) (see Strachan & Read, 2018).

In addition to the several million genetic variants passed down by each of our parents, we inherit roughly 30–80 new mutations that arise during meiosis. The human population explosion over the past several hundred years has produced an abundance of new mutations as rare variants. Rare variants are disproportionately deleterious. Fu et al. (2013) estimated that ~86% of all deleterious SNVs are rare and recent. Many of these variants are found in only a handful of related people and are not represented in population samples. As discussed later, despite their prevalence and disease-relevance, rare variants pose a challenge for GWASs.

2.3 A brief note on ancestry and continental populations

Most sociogenomics studies at least briefly discuss ancestry and issues related thereto. A basic understanding of what this refers to is helpful (for a social science discussion, see Herd et al., 2021). Modern humans are, of course, a single species, which emerged some 550–750 thousand years ago (Fu et al., 2016). Although terminology varies, several population genetic studies classify humans roughly into five continental populations: African (AFR), European (EUR), East Asian (EAS), South Asian (SAS), and American (AMR), differentiated by their continental migration out of Africa within the last 100,000 years (The 1000 Genomes Project Consortium, 2015). Importantly, these populations are abstractions from an underlying continuum of genetic relatedness and should not be thought of as genetically distinct subpopulations (Coop & Przeworski, 2022; Feldman, Lewontin, & King, 2003).

The vast majority of variants in an individual's genome are shared by all continental populations (The 1000 Genomes Project Consortium, 2015). Only a small proportion of the variants in an individual genome are restricted to one continental population, and these tend to be recent mutations that are also rare in the populations in which they are found. However, allele frequencies for common variants do differ across groups because of population patterns of migration and mating, shaped by physical boundaries and sociocultural influences. Furthermore, allele frequencies vary in a more fine-grained manner across subgroups of populations, especially for rare variants (Mathieson & Mcvean, 2012). As discussed later, this variation in mostly random allele frequencies across difference groups poses a major challenge for GWASs by inducing or inflating genetic associations through confounding between genotypes and outcomes (e.g., Berg et al., 2019; Morris, Davies, Hemani, & Smith, 2020a).

3. Statistical genetic methods of sociogenomics

3.1 What genetic differences are measured?

The complexity of GWASs/PGSs and the way that they are discussed can produce confusion over what is measured in these studies. Readers can be excused from thinking that these studies measure genes and/or causal variants that shape differences through some known biological pathway. The abstract of a recent study, for example, referenced “mothers with more education-related genes” (Armstrong-Carter et al., 2020). Genes are not measured in these studies. Rather, these studies measure and analyze a select subset of one form of variation in the genome: Single-nucleotide polymorphisms (SNPs) that have two alleles (e.g., A or C) (see Appendix A for a detailed discussion).³ In this section, I describe with as much simplicity as possible what is measured in GWASs/PGSs and why. Although intricate, understanding what GWASs/PGSs do measure (SNPs) and that they do not measure (genes or causal variants) is necessary to understand the inherent limitations with this approach.

The GWAS methodology is rooted in the blocklike structure of our genome. Although technical detail is out of scope, we inherit whole chromosomes from each parent, but these chromosomes are composed of unique blends of blocks of our parents' maternal and paternal chromosomes created during the process of “crossing over” (or genetic recombination). Each chromosome we inherit is a unique blend of our parents' matching chromosomes, created when segments are exchanged in meiosis (an average of 1.5 blocks of exchange per chromosome). Helpfully, crossing over does not occur randomly across the genome but tends to occur in 1–2 kb regions, known as recombination hotspots, which occur every 50–100 kb across the genome (Myers, Bottolo, Freeman, McVean, & Donnelly, 2005). Consequently, blocks of chromosomal segments are passed down across many generations unbroken by recombination, and, by dint of being passed down unbroken, contain correlated SNPs (i.e., SNPs that are not inherited independently). These chromosomal segments that exist between recombination hotspots are known as haplotype blocks. The association between SNPs on a haplotype is known as linkage disequilibrium (LD) and exists as a matter of degree (as a correlation).

This haplotype structure of our genome means that there is much less variability between genomes than would occur from the random assortment of SNPs. For example, the average haplotype block contains ~50 SNPs, which would, in theory, allow 2⁵⁰ different combinations. Typically, however, most haplotype (>90%) blocks will be characterized by six or fewer combinations of alleles (The International HapMap Consortium, 2005). The combination of alleles on a haplotype block is known as haplotype and represents ancestral segments defined by common, ancient SNPs. Rarer variation exists as heterogeneity around the SNPs that define haplotype blocks (Strachan & Read, 2018).

This haplotype structure of our genomes undergirds the GWAS methodology. Measuring and testing each of our 3 bn base pairs is impracticable. Instead, GWASs analyze a smaller number of SNPs from across the genome to tag regions of common variation (i.e., haplotypes). Contemporary GWASs scan the genome for associations between several millions of these preselected SNPs, known as “tag SNPs” and a trait. Significant SNP associations mark a genomic region (“genomic risk locus” or quantitative trait locus, QTL) in which an unknown causal variant(s) driving the association is presumed to lie. Tag SNPs are thus usually non-functional, common variants used as proxies for some unknown causal variant(s) in proximity (with which they are in LD). Proximity is relative and varying. Genomic risk loci can range in size from several hundred thousands to more than 1 Mbp.

Crucially, rare and more likely deleterious variants are not well tagged by SNPs, given that SNPs tag haplotypes defined by shared common variants, and most haplotypes will not contain the rare variants (or they wouldn't be rare)⁴ (Backman et al., 2021; McClellan & King, 2010; Tam et al., 2019). Additionally, other variant forms – indels, copy number variants (CNVs), and SVs – are not measured in GWASs, and many are not well-tagged by common SNPs (Backman et al., 2021; Tam et al., 2019).

Additionally, because different ancestral groups can have different allele frequencies, different patterns of LD, and somewhat different haplotypes, tag SNPs often do not work in the same way across populations, even when the causal variant is the same (Martin et al., 2017; Peterson et al., 2019). This ancestral variation in LD and haplotypes is one biological reason why GWAS findings do not “port well” or generalize across ancestral groups (e.g., Mostafavi et al., 2020).

The haplotype structure of our genome also enables GWASs by facilitating imputation. GWASs rely on large samples; however, studies vary in the genotyping platforms they use, which measure somewhat different SNPs, and contain missing data. Knowledge of haplotypes allows the probabilistic imputation of missing or untyped genotypes at adjacent SNPs using more densely genotyped samples or whole-genome reference panels.⁵ Most genotype arrays now measure between 500,000 and 2 million SNPs, and most contemporary GWASs now include ~10 million SNPs, most imputed (Tam et al., 2019).

The original aim of GWASs was to understand the underlying molecular basis of trait variation by tracing causal pathways from genetic variants to outcomes. The idea was that tag SNPs could be used to mark risk loci that could be followed up with fine-mapping and functional annotation to identify causal variants in genes with well-defined functions. Although GWASs have, in some cases, facilitated the identification of causal variants involved in disease pathogenesis, for reasons that are out of scope, biological interpretation is exceedingly difficult, in general, and even more so for complex social traits with increasingly numerous (>1,000) GWAS hits and miniscule effect sizes (see e.g., Backman et al., 2021; Crouch & Bodmer, 2020; Edwards, Beesley, French, & Dunning, 2013). Hence, sociogenomicists primarily use GWAS results for polygenic prediction, explicitly deemphasizing inquiry into causal variant(s) or biological pathways (but not always, see e.g., Ganna et al., 2019). In what follows, I briefly describe the nuts and bolts of GWASs, this is followed by a discussion of PGS generation.

3.2 GWAS methodology

GWASs are a theory-free analytic approach to scan the genome for trait-associated tag SNPs. This involves testing, for each SNP one at a time, whether an allele (SNP variant) is more common in cases versus controls (or for continuous traits, across different levels). GWASs thus test for independence between genotype and outcome for each SNP, with a few controls (not including other SNPs). The 2018 educational attainment GWASs, for example, assessed whether allele frequencies – for roughly 10 million SNPs – differed (across groups stratified by) years of education (Lee et al., 2018). Typically, the type of effect of interest in GWASs is variant substitution effects, which can be understood as the counterfactual change in an individual outcome that would occur from changing the individual's genotype for a particular SNP at conception (holding all else constant) (Freese, 2008; Morris et al., 2020a). This counterfactual model assumes that genetic associations indicate a causal path from an individual's genotype (or allele dosage) to complex social traits, reflecting a variant substitution effect (Lawson et al., 2020).

The basic form of GWASs is straightforward. Here I focus on these basics, including the familiar linear equation form that underlies the model. This model has been elaborated in recent years, but the underlying logic remains the same. Using bi-allelic SNPs and assuming additive SNP effects, genotypes for a particular SNP (e.g., AA, AC, CC) are translated to numeric allele dosage effects by counting the number of minor (or effect) alleles (0, 1, or 2) for each individual. Allele dosages for each SNP are the focal independent variable in each of these millions of regressions (again, one for each SNP examined separately), which take the following general form:

$$\eqalign{Y =& \beta _0 + \beta _1 \times {\rm SNP} + \beta _{2\;} \times {\rm Sex} + \beta _{3\;} \times {\rm Age} + \beta _{4\;}\cr &\times PC1 \ldots \;\beta _{14\;} \times PC10 + e}$$

where Y is a continuous variable (e.g., years of education), and SNP represents the allele dosage measure, controlling for age, sex, and usually 10–20 genetic ancestry principal components (PCs, discussed shortly). The outcome of interest in this model is β ₁ – the effect size for each SNP – which can be interpreted as the marginal effect of having one more minor allele (a unit increase in allele dosage) and its associated p-value. For binary outcomes, this would just approximate the form of a familiar logistic regression model. These results for the millions of separate regressions are automatically compiled into results by modern computational programming software, such as PLINK (Purcell et al., 2007) and METAL (Willer, Li, & Abecasis, 2010). Focal GWAS results, as the SNP effect size estimates and p-values, are known as summary statistics, which provide the input for further analyses. Summary statistics are often considered the “data” in GWASs even as these are more accurately referred to as the results (of the first step of the analysis) (Burt & Munafò, 2021).

Following the estimation of the GWASs from the primary study sample or the “discovery” sample, a number of diagnostic tests (e.g., Manhattan and QQ-plots, which display p-values on a −log₁₀ scale) are performed (see Choi, Mak, & O'Reilly, 2020; Schaid, Chen, & Larson, 2018). Because of LD (non-independence among SNPs sharing a haplotype) and the examination of each SNP separately, there will invariably be multiple (even dozens of) SNPs marking a risk locus. Thus, follow-up analyses (e.g., clumping and thresholding) are conducted to define clusters of SNPs in high LD (often high LD is defined as r ² > 0.1⁶) and to identify a single “lead SNP,” usually the SNP with the lowest p-value, to represent this clump and mark a risk locus. In this way, risk loci (or QTLs) are defined as trait-associated regions marked by approximately independent (“lead”) SNPs.

As noted, risk loci range in size from ~50 kbp to over 1 Mbp (e.g., Lee et al., 2018). Thus, a GWAS that reports 1,237 lead SNPs can thus be understood as identifying 1,237 approximately independent risk loci defined by a lead SNP and in which the causal variant(s) responsible for the association is presumed to lie. Such risk loci, which often stretch across multiple haplotypes, usually contain thousands of SNVs along with SVs and indels, and often multiple genes (hence the difficulty of biological interpretation).

Importantly, lead SNPs for complex social traits are invariably very weakly associated with an outcome, usually accounting for less than 0.01% of the variation. In their educational attainment study, for example, Lee et al. (2018) reported that “the median effect size of the lead SNPs corresponds to 1.7 weeks of schooling per allele.” Similarly, among the five lead SNPs identified in their study of “non-heterosexuality” in the UK Biobank, Ganna et al. (2019) observed “very small effects”; “males with a GT genotype at the rs34730029 locus had 0.4% higher prevalence of same-sex sexual behavior than those with a TT genotype (4.0 vs. 3.6%)” (p. 3). Given the impracticability of biological interpretation and the weak prediction from any single variant or QTL, researchers have shifted to creating genetic summary scores that aggregate SNPs weighted by their effect sizes, discussed next.

3.3 Polygenic score (PGS) construction

Calculating PGSs (also called polygenic risk scores [PRSs] or genetic risk scores [GRSs], usually when referring to adverse biomedical outcomes) is now a common application of GWASs to predict complex traits (or disease risk) from weight and height to depression and educational attainment (Evans, Visscher, & Wray, 2009; Wray, Goddard, & Visscher, 2007). PGSs operate under a massively polygenic, additive model (Boyle, Li, & Pritchard, 2017). Under this model, summing the GWAS-weighted risk (or effect) allele dosages (0, 1, or 2) usually with several sophisticated statistical adjustments can provide an index of a continuous underlying (additive) genetic liability for a trait.⁷ The human equivalent of the “breeding value” is in selective plant and animal breeding in human populations (Meuwissen, Hayes, & Goddard, 2001), and PGSs have been described as “summariz[ing] the cumulative effects of many variants across the genome and aim[ing] to index an individual's genetic liability for a given trait” (Domingue, Trejo, Armstrong-Carter, & Tucker-Drob, 2020, p. 465) or a “single quantitative measure of genetic predisposition” (Mills, Barban, & Tropf, 2018). The educational attainment PGS has been characterized as measuring “an individual's genetic predisposition for completing [more years of] formal schooling” (Bolyard & Savelyev, 2020) and a “DNA-based indicator[] of propensity to succeed in education” (Harden et al., 2020).

The specific details on PGS construction can, and have, filled articles (see Choi et al. [2020] for more details), but the basic process is as follows: Run GWAS in discovery sample → replicate results in an independent sample → adjust for LD using a reference panel → select SNPs → adjust for LD and winner's curse → construct PGS → test PGS prediction in a target sample → assess PGS with incremental R ². It is worth noting that there are several decisions by researchers involved in PGS construction. In the “select SNPs” phase of PGS construction, researchers decide which SNPs to include in the PGS (via p-value thresholds) based on the success of prediction. Specifically, researchers evaluate several PGSs created at a variety of p-value thresholds and select the best PGS predictor (measured by R ²), which is usually the PGSs created from a p < 1 threshold (i.e., no p-value threshold) (e.g., Belsky et al., 2018; Ganna et al., 2019; Lee et al., 2018).⁸

Thus, in what may come as a surprise to some, most PGSs are constructed from all available SNPs regardless of their statistical significance in the GWAS. Available evidence suggests that these “all SNPs” PGSs are more environmentally confounded than those that use (more stringent) p-value thresholds, such that while these may explain more variance, they do so because they capture environmental influences as well as genetic ones (Berg et al., 2019; Mostafavi et al., 2020).

4. The utility of PGSs for social science: Proponents' arguments

Touted as a powerful new “tool” for social scientists to incorporate genetics into their research, PGSs are said to offer exciting new opportunities for social science research (Braudt, 2018; Freese, 2018; Harden & Koellinger, 2020; Mills & Tropf, 2020). Below I describe proponents' chief arguments about the utility of PGSs for social science, but first a note on PGSs lack of efficacy in individual prediction.

With few exceptions (e.g., Plomin, 2019; Plomin & Von Stumm, 2018), scholars 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 (see, e.g., Harden & Koellinger, 2020; Morris, Davies, & Smith, 2020b). Although not appropriate for predicting individual outcomes, proponents emphasize myriad of benefits to incorporating PGSs to social science.

4.1 “Getting genetics out of the way”

Perhaps the most hyped value of PGSs in social science is to control for genetic heterogeneity in studies of environmental effects. According to Harden (2021a), many sociogenomicists are most excited about the potential of PGSs as a tool “to make genetics recede into the background, to get it out of the way” so that we can more clearly see the effects of environments (see also Conley, 2016). Given ubiquitous heritability, proponents argue that uncontrolled genetic heterogeneity poses a serious threat to inferences about the effects of specific environments, as these ostensibly environmental causes may be biased or spurious (as actually driven by genetic differences) (Harden & Koellinger, 2020; Hart et al., 2021). For example, rather than health or longevity being influenced by higher educational attainment, scholars have suggested, these relationships may be spurious with genetic endowment being the causal force. Similarly, sociogenomicists have asked, whether parental environments, including early childcare, causally influence educational attainment or whether these are spuriously associated because of shared genetic endowments.

Proponents also argue that incorporating PGSs as control variables into social science research can enhance the precision of environmental estimates (Cesarini & Visscher, 2017; Harden, 2021a, 2021b; Kweon et al., 2020). This enhanced precision may increase the power associated with randomized controlled trials, potentially shrinking their cost (Lee et al., 2018; Rietveld et al., 2013). Controlling for genetic heterogeneity with PGSs, proponents argue, may also reveal previously obscured environmental effects. For example, some environmental influences on educational attainment may only be apparent among those at “high genetic risk” (Herd et al., 2021). For these reasons, proponents suggest, PGSs are valuable as a control for differential genetic propensity to illuminate more clearly and precisely the effects of environmental influences (Harden & Koellinger, 2020; Trejo & Domingue, 2019).

4.2 A powerful, flexible analytic tool for causal inference

Proponents also emphasize the value of PGSs as a powerful tool for causal inference (Belsky & Israel, 2014). This strength of PGSs, proponents argue, draws on several unique advantages of genetic data (Conley, 2016; Harden, 2021a). First, evidence (from twin studies of heritability) suggests that genetic differences matter. Second, “the genetic sequence of each person is fixed at conception and does not change throughout one's lifetime” (Kweon et al., 2020), which means that genotype need only be measured once. Further, once measured, PGSs can be calculated for any outcome, which need not be measured in the study, and as PGSs are updated with larger and more diverse samples, these individual scores can be created and updated (Belsky et al., 2018; Harden, 2021a, 2021b).

Proponents emphasize that this fixity of our DNA sequence means that reverse causality from behavior or environmental exposures to the genome can be ruled out. Given this, genetic data can serve as exogenous measures of individual characteristics, which do not change over the life course, “facilitating the tracing of developmental paths” or as a “fixed point from which to observe child development” (Belsky & Israel, 2014; Harden et al., 2020). Scholars have argued that PGSs can be used as a “molecular tracer”: “Just as a radiologist might administer a radioactive tracer to track the flow of blood within the body, researchers can use genetics as a molecular tracer to get a clearer image of how students progress through the twists and turns of the educational system” (Harden et al., 2020).

4.3 Gene–environment interplay

PGSs are also advertised as a more direct and powerful tool to explore how gene–environment interplay influences social outcomes. Broadly, gene–environment interplay with PGSs can be demarcated into three broad types: (1) PGS–environment interactions (e.g., does gender suppress “genetic potential” for educational attainment; Herd et al., 2019), (2) PGS–environment combinatory effects (e.g., how do “nature” and “nurture” combine to shape children's resemblance to their parents in human capital accumulations over time; Harden & Koellinger, 2020), and (3) PGS-through-environment pathways (e.g., through what social–psychological mechanisms does the education PGS increase educational attainment; Bolyard & Savelyev, 2020).

Proponents have argued that PGSs can reinvigorate the study of gene–environment interactions (G × E) with “robust measures of genotype,” in contrast to the limited candidate G × E approach (Harden & Koellinger, 2020; Martschenko et al., 2019). “By applying the prism of GxE models, it is hoped that the white light of average effects will be refracted into a rainbow of genetically mediated responses that are made clear to the scholar interested in describing human behavior” (Conley, 2016, p. 293). In addition, PGSs may also be gainfully employed in the service of understanding heterogeneous responses to social interventions, in the form of a PGS ×  intervention (Harden & Koellinger, 2020).

4.4 Risk stratification and/or early identification

Although most scholars agree that PGS-based personalized programs or policies are not realistic because of poor individual prediction, PGSs are still advertised as having potential use in risk stratification, particularly for those in the upper and lower deciles of PGSs. On this view, PGSs could be used to identify “at-risk” individuals before problems manifest or become severe through the implementation of an early genetic screening system (Martschenko et al., 2019). Such genetic screening is argued to provide an inexpensive way to more expansively identify those at high genetic risk of problems, such as lower educational attainment or physical inactivity, and intervene in advance with, for example, extra support or placement into a different learning environment (Harden & Koellinger, 2020; Martschenko et al., 2019). Similarly, PGSs could be used to identify “high potential” individuals, who could also be targeted with different learning environments.

In addition to risk stratification, proponents argue that enhanced understanding of the distribution of genetic risks could be used to study the effects of social institutions and programs. For example, in educational systems, studying the distribution of genetic risks “across schools could be used to study inequities in the current ways that the educational system under- and overdiagnoses students… thereby identifying differential diagnoses and treatment across groups” using PGSs as “indicators with some degree of objectivity” (Martschenko et al., 2019).

4.5 Changing worldviews and approaches to social inequalities

Finally, some proponents claim that incorporating genetics into social science will change the way that social scientists think about the world. In the words of Harden and Koellinger (2020, p. 567):

Ultimately, the greatest impact from integrating genetics into the social sciences will probably not come from simply applying new tools to old questions, but from changing how people think about the world around them, allowing them to ask new questions and to pursue new answers that would not have been feasible before. For example, the realization that success in life is partly the result of a genetic lottery raises new questions not only about underlying mechanisms, but also about fairness and what a desirable distribution of wealth in a society should look like.

On this view, GWASs and PGSs reveal the hitherto unrecognized fact that “success in life” is partly shaped by our genetic inheritances. In general, these scholars maintain that incorporating genetics into social science will stimulate new ways of thinking about and investigating our differences and inequalities, which may inform social policies to ameliorate inequalities.

4.6 Summary

Proponents tout several benefits from incorporating PGSs into social science to enhance social science research. In the next section, I scrutinize the science of sociogenomics, highlighting limitations, which I argue, undermine the utility of PGSs into social science. Most of these limitations are acknowledged by sociogenomicists; yet the full implications of these challenges are invariably unheeded in practical applications.

5. Limitations of PGSs that undermine their utility for social science

As is well known, a person's social traits emerge from a complex interplay of environmental and genetic influences over their lifetime. As I have discussed, the goal of GWASs is to identify variant substitution effects as causal genetic effects, and the primary raison d'être of PGSs is to index genetic influences on (differences in) phenotypes. Proponents hype the value of PGSs for “unbraiding” and “disentangling” the effects of genetics and environments in shaping individual differences in complex social outcomes. Naturally, this only works if (a) genetic and environmental influences on traits can be differentiated, and, if so, (b) PGSs are relatively accurate and unbiased estimates of genetic influences (Barton, Hermisson, & Nordborg, 2019). Unfortunately, for a variety of biological, statistical, and developmental reasons, GWASs cannot disentangle “genetic” from “environmental” influences, such that PGSs do not index genetic influences on complex traits (Haworth et al., 2019; Morris et al., 2020a). In particular, dynamic population phenomena induce confounding between genotypes and complex social outcomes at multiple levels, inter alia: family, neighborhood, peer group, region, culture, nation, and historical time (Barton et al., 2019; Lawson et al., 2020). I discuss four primary limitations of PGSs that vitiate their utility for social science as measures of “genetic influences on” or “genetic propensities for” complex social traits: relatedness confounding, downward causation, limited coverage of genetic influences, and context-specificity.

5.1 Relatedness confounding of PGSs

The most widespread and widely recognized form of environmental confounding is because of (genetic) relatedness and passive gene–environment correlations. Basically, people who are more genetically similar (i.e., more closely related, even distantly) also tend to develop in more similar sociocultural, political, and physical environments, which influence most complex social traits. Thus, genotype and environments are correlated for non-causal reasons. Generally, relatedness confounding is demarcated into population genetic structure and familial confounding. Both are known issues in GWASs/PGSs and steps are taken to mitigate this confounding. However, evidence is mounting that these corrections are insufficient, such that inflated or spurious genetic associations persist (e.g., Barton et al., 2019; Berg et al., 2019; Haworth et al., 2019; Morris et al., 2020a; Mostafavi et al., 2020).

5.1.1 Population (sub)structure and phenotype stratification

With respect to confounding by population structure, the key qualitative difference is between controlling the environment experimentally, and not doing so. Once we leave an experimental setting, we are effectively skating on thin ice, and whether the ice will hold depends on how far out we skate. (Barton et al., 2019, p. 3)

Population (genetic) (sub)structure refers to patterns of genetic variation within populations because of non-random mating. Population structure arises because of complex demographic histories (separation, migration, admixture), which result in mostly random allele frequency differences between population subgroups (Cardon & Palmer, 2003; Lawson et al., 2020). When these coarse population genetic subgroups (shaped by geographic region, race/ethnicity, social class, religion) are differentially exposed to trait-associated sociocultural and physical environmental factors – as they often are – alleles associated with subgroup membership are also associated with trait differences, producing spurious or inflated genetic effect size estimates, known as phenotype stratification (Browning & Browning, 2011; Cardon & Palmer, 2003; Morris et al., 2020a).

The classic example used to illustrate phenotype stratification is a genetic association study of chopstick-eating skills (Hamer, 2000; Lander & Schork, 1994). If we were to conduct a GWAS of using chopsticks in a sample of diverse ancestry, we would no doubt find significant associations. Although there may be some genetic variants affecting our ability to handle chopsticks (e.g., finger dexterity), most genetic associations would be because of cultural differences, namely random variants that differed in frequency between East Asia and the rest of the world and had nothing to do with “genetic propensity” for chopstick use skills. In practical applications, phenotype stratification is most plainly manifest with the geographic patterning of PGSs, which reflects sociocultural and physical environmental influences (Abdellaoui, Verweij, & Nivard, 2022; Haworth et al., 2019; Lawson et al., 2020).

The minimal approach to mitigate phenotype stratification is the examination of an ostensibly homogenous ancestral group. However, population substructure exists within these groups, including populations from a single location, such as “white Europeans” within the United Kingdom, Finland, the Netherlands, and Western France (e.g., Bycroft et al., 2019; Byrne, van Rheenen, van den Berg, Veldink, & McLaughlin, 2020; Haworth et al., 2019; Karakachoff et al., 2015; Kerminen et al., 2017; Leslie et al., 2015). Such finer-scale genetic population structure (known as local or regional population structure) is a function of non-random mating shaped by sociopolitical forces, cultural factors, and different physical environments all of which foster assortative mating (Morris et al., 2020a; Richardson & Jones, 2019; Zaidi & Mathieson, 2020). Consequently, pervasive, albeit often subtle, allele frequency differences between subgroups experiencing many different physical and social environments exist and can be picked up by GWASs as genetic causes, even if functionally unrelated to trait variation. For these reasons, in the presence of population structure, GWAS SNP associations may just be proxies for (or inflated by) an environmental variable that has not been properly corrected (Browning & Browning, 2011; Cardon & Palmer, 2003; Novembre & Barton, 2018).

Several sophisticated statistical methods have been introduced to mitigate or adjust for population structure-confounding, including genomic control (Devlin & Roeder, 1999), genetic principal components (PCs) (Price et al., 2006), linear-mixed models (LMM) (Kang et al., 2010), and LD score regression (LDSC) (Bulik-Sullivan et al., 2015). Although these methods appear to reduce population stratification, evidence from a variety of studies using whole-genome sequence data, simulations, and tests of non-genetic traits (like latitude/longitude of birth, birth order) evince that these methods do not adequately correct for population structure, and this is especially true for complex social traits of interest to sociogenomicists (e.g., Berg et al., 2019; Dandine-Roulland et al., 2016; Mostafavi et al., 2020; Sohail et al., 2019; Zaidi & Mathieson, 2020).

For example, in a recent study, Abdellaoui et al. (2022) demonstrate that controlling for geographic region decreases heritability signals for socioeconomic status (SES)-related traits, especially educational attainment and income, as socioeconomic differences between geographic regions induce gene–environment correlations that are picked up in GWASs and inflate PGSs (see also Leslie et al., 2015; Mostafavi et al., 2020; Sohail et al., 2019). In another study using simulations, Zaidi and Mathieson (2020) show that recent (within the past 100 generations or ~2,500 years) genetic structure with sharp effects pose a particular problem for GWASs/PGSs given the tag SNP methodology. As they explain, recent population structure with sharp local effects, as may result from cultural, language, and/or physical boundaries patterning mating, can only be adequately corrected with rare variants, which are not measured in these studies.⁹

In sum, the evidence is clear that phenotype stratification persists despite sophisticated methods to mitigate such confounding – most obviously in the form of geographic patterning of PGSs (Abdellaoui et al., 2022; Byrne et al., 2020; Haworth et al., 2019) – and its effects (inflating PGSs) appear to be particularly acute for complex behavioral traits related to socioeconomic status (Abdellaoui et al., 2022; Lawson et al., 2020). Crucially, these biases are exacerbated under the very modeling conditions most often used for social science outcomes – when multiple studies are meta-analyzed and millions of SNPs are aggregated in PGSs. Even subtle population stratification can cumulatively generate substantial biases when millions of SNPs are aggregated, especially when less stringent p-values are employed (as is typical) (Barton et al., 2019; Berg et al., 2019; Mathieson & Mcvean, 2012). In short, PGSs for complex social traits capture some non-trivial amount of social environmental effects because of uncorrected population substructure (Abdellaoui et al., 2022; Curtis, 2018; Lawson et al., 2020).

5.1.2 Familial confounding¹⁰

Biological parents not only pass on half of their genome to their children but also their environments, including social status, culture, worldviews, values, habits, and the like (Shen & Feldman, 2020). Therefore, the association between parental and offspring genotypes is often confounded by the association of genotypes with rearing environments, effects which may be amplified over generations via social mechanisms (as “dynastic effects”; Brumpton et al., 2020). Such gene–environment correlations inflate estimates of genetic influences, especially for complex social traits where the transmission of social advantages (e.g., status and wealth) and associated familial practices are significant (e.g., Kong et al., 2018; Morris et al., 2020a).

Several innovative sociogenomics studies have illuminated the extent of familial confounding in PGSs. These studies suggest that roughly half of the effect of the education PGS is because of familial confounding. For example, Kong et al. (2018) found that controlling for an education PGS created from parents' non-transmitted alleles (i.e., the other half of alleles not passed down) reduced the variance explained by the offspring education PGSs by roughly half. If child PGS captures causal genetic effects, then controlling for non-transmitted parental alleles would not substantially reduce the effect of the child PGS on their education. In contrast, Kong et al.'s results suggested significant inflation of ostensibly genetic effects by familial confounding. In another study, Cheesman et al. (2020) compared the predictive effects of an education PGS on years of education in adopted and non-adopted youth. They observed that the PGS was twice as predictive of years of education in non-adopted versus adopted individuals (R ² = 0.074 vs. 0.037), as would be expected if the education PGS captures familial effects. Similarly, Belsky et al. (2018) observed that controlling for parental education reduced the effect of the education PGS on years of education by about half, which “suggests environmental confounding of polygenic score associations with educational attainment” (p. E7277).

As with population structure, practitioners are aware of the issues with familial confounding and have employed statistical techniques to attempt to mitigate this confounding (see, e.g., Trejo & Domingue, 2019; Wu et al., 2021; Young et al., 2018). The most rigorous approach to reduce familial and population structure confounding is a within-family or sibling-difference design. These studies examine how differences between siblings in their genotypes (in GWAS or PGS prediction) explain sibling differences in phenotypes, net of their shared rearing environments using family fixed effects (Belsky et al., 2018; Laird & Lange, 2006). For illustration, Lee et al. (2018) used a sibling-difference study to test the robustness of their (conventionally) unrelated sample education GWAS findings using a sample of ~22,000 sibling pairs. Given differences in statistical power, Lee et al. (2018) examined sign concordances of the GWAS coefficients (i.e., whether the effect direction of the risk alleles matched +/+) rather than their significance or effect sizes across the studies at three different p-value thresholds. By chance, of course, we would expect 50% of the signs to match. Their results showed that for the less stringent p-value threshold (p < 5 × 10⁻³), sign concordances between the discovery GWAS and sibling-difference GWAS were only slightly better than chance at ~56.5%, which improved at more stringent p-value thresholds to ~60% at p < 5 × 10⁻⁵ and ~65% at p < 5 × 10⁻⁸.¹¹ [Aside: Although expecting perfect sign concordance is unrealistic, a sign concordance of <57% at a p-value threshold that was more stringent than the one employed to create the widely used education PGS does not, in my view, demonstrate robustness or constitute replicated findings.] Lee et al. (2018) reported that the within-family effect sizes were, on average, 40% smaller than that from the unrelated GWASs. The just-published updated education GWAS did not present a within-family GWAS replication; however, their within-family PGS analyses indicated that only 30.9% of the PGS effect was a “direct effect” (Okbay et al., 2022; see also Morris et al., 2020a).

Not unexpectedly, sibling-difference studies of non-social (more proximally biological) traits, like height and C-reactive protein, report only minor evidence of familial confounding and slightly reduced effect sizes, whereas sib-studies of social outcomes, like educational attainment and smoking behavior, invariably report appreciably smaller effect size estimates, given the significance of sociocultural forces on these traits (Howe et al., 2022; Lee et al., 2018; Mostafavi et al., 2020). Importantly, this confounding is not simply a minor issue affecting the precise effect size but evidence suggests that this confounding substantively alters sociogenomics findings. For example, Howe et al. (2022) demonstrated that strong genetic correlations between education and height, weight, and C-reactive protein from population genetic studies become “negligible” in sibling-difference analyses.

Given the persistence of genetic relatedness confounding in GWASs and PGSs even with sophisticated methodological “corrections,” research employing PGSs as indicators of genetic influence should, at a minimum (a) control for relevant social environments that are associated with genotype, including geographic location (Abdellaoui et al., 2022), or, preferably, (b) use sibling-study adjusted PGSs through a two-stage model to reduce (if not completely eliminate¹²) relatedness confounding. In the two-stage model, SNP p-values are estimated using a large unrelated GWAS, but the effect sizes are adjusted (downward) using the coefficients from a sibling-difference study (Choi et al., 2020; Zaidi & Mathieson, 2020). Unfortunately, neither is common practice. Estimates used to create the education PGS, now widely available for use in social science datasets, were not adjusted based on the sibling study reduced effects sizes or the sign mismatch in the replication mentioned above. Creditably, the authors (Lee et al., 2018) recognized the persistence of confounding, writing:

[o]ur within-family analyses suggest that GWAS estimates may overstate the causal effect sizes: if educational attainment-increasing genotypes are associated with parental educational attainment-increasing genotypes, which are in turn associated with rearing environments that promote educational attainment, then failure to control for rearing environment will bias GWAS estimates…. Without controls for this bias, it is therefore inappropriate to interpret the polygenic score for educational attainment as a measure of genetic endowment (p. 1116, emphasis added).

Despite this clear caution about using PGSs as genetic potential without controls for confounding, subsequent education PGS studies did not heed these cautions and failed to control for rearing environments while examining PGSs as “genetic propensity” (e.g., Harden et al., 2020; Herd et al., 2019; Wedow et al., 2018).

Notably, even PGSs created from within-family GWASs are not immune to environmental confounding for two key reasons. One has to do with the uniqueness of within-family designs. Because of subtle micro-stratification and complex social–psychological dynamics within families, the extent to which the causes of sibling differences for complex social traits are the same as the causes of general population differences is questionable. Research suggests sibling differences may be amplified or distorted as siblings attempt to create their own niches or fill unique roles in their families (e.g., “the smart one,” “the athlete,” “the funny one,” “the troublemaker,” “the pretty one”) (see, e.g., Healey & Ellis, 2007; Sulloway, 2001) in part through “sibling contrast effects” (Carey, 1986). For other traits and behaviors, differences may be minimized as families tend to socialize children in similar ways and siblings imitate one another. These interactional dynamics influence child identities, expectations, motivations, personality, and developmental outcomes and thus undermine the generalizability of sibling-difference studies.¹³

In addition, genetic associations and PGSs from sib-studies are confounded by broader sociocultural influences. This is because the counterfactual model that underlies genetic association studies does not distinguish between authentic (upward) genetic causes (i.e., from genetic differences to trait differences through biological mechanisms) and artificial downward (social) causation. Both are identified as causes in GWAS's counterfactual variant substation effects approach.

5.2 Downward causation and artificial genetic signals

Downward causation – defined as sociocultural forces that sort and select individuals based on genetically influenced traits, such as skin pigmentation and height, into different environments and exposures that influence social outcomes – creates what I call artificial genetic associations, which are environmental influences masquerading as genetic influences in GWASs. Although the fact that sociocultural environments shape and filter genetic influences is understood by most, less well understood is the extent to which the causal effects of social structural and cultural forces acting on genetically influenced differences are identified as genetic influences in GWASs and PGSs.¹⁴

Jencks' (1972) now classic thought experiment on discrimination by hair color can be used to illustrate downward causation creating artificial genetic associations. Jencks asks us to imagine a system where red-haired children are barred from school. In such a system, genetic variants linked to red hair would be identified by GWASs as genetic causes of educational attainment. However, neither an individuals' red hair, nor the genetic variants contributing to red hair, are appropriately conceived as causes of differences in educational attainment in this hypothetical case, in my view and that of others (Kaplan & Turkheimer, 2021), but see Harden (2021a). The “difference that makes a difference” is not red hair but the social-institutional policies excluding people with red hair, which is why a change in the rules would (over time, I presume) make hair color unrelated to educational attainment (and remove any red-hair genetic associations with education). Although explicit discriminatory exclusionary policies like this one are largely a thing of the past in most developed nations, both ongoing discrimination and the legacy of past discrimination (through intergenerational transmissions of wealth, status, social capital, etc.) continue to influence individual development and trait differences. More broadly, our environments and institutions, educational and otherwise, continue to differentially treat individuals based on a variety of genetically influenced individual traits such as height, body weight, personality, attractiveness, and skin tone into different environments and exposures and thus opportunities, achievements, and developmental outcomes (e.g., Monk, Esposito, & Lee, 2021; Simons, Burt, Barr, Lei, & Stewart, 2014).

GWASs and PGSs capture artificial genetic signals, and these artificial effects are likely to be pervasive given the extent to which we respond to phenotypic cues in our interactions with others in a manner that is unavoidably socioculturally mediated. Although casting such socioculturally driven genetic associations as genetic propensity or even “indirect genetic effects” is misguided, even more concerning is the subsequent framing of such correlations as innate individual propensities (individual “genetic fortune” or “misfortune”). Because of downward causation, genetic associations for many complex social behaviors are unavoidably environmentally confounded and are not appropriately conceived as genetic causes of outcomes.

5.3 Limited coverage of genetic variation

To serve as a control for genetic influences, in addition to not being substantially environmentally confounded, PGSs need to capture genetic influences relatively accurately and comprehensively. They do not.

5.3.1 Low resolution

GWASs and PGSs capture genetic variation at low resolution. As noted, SNPs rarely have functional effects and usually tag large regions of common variation, which may contain numerous causal variants including large effect extremely rare variants (McClellan & King, 2010).¹⁵ The causal variant(s) in the tagged region may often be multiple and rare, and such that only a paucity of individuals with the risk allele (tag SNP) will carry the actual causal variant. Thus, tag SNPs – even if they reflect causal genetic influences – are very imprecise proxies for a causal variant that may only exist on that haplotype for a small minority of individuals.¹⁶ The tag SNP methodology, which excludes rarer and likely functional SNVs, indels, and SVs make GWASs possible, but it also makes PGSs incomprehensive measures of genetic risk (Backman et al., 2021).

PGSs also ignore the X chromosome (given that females have two and one is usually inactivated in a cell), and both GWASs and PGSs invariably ignore the Y chromosome. Mitochondrial DNA is also neglected.

5.3.2 Genetic additivity and interactionism

Finally, GWASs and PGSs usually estimate additive genetic influences. However, because of pervasive gene–gene interactions and interactions between non-coding RNA genes and coding genes, focusing on additive effects from tag SNPs is necessarily misleading (as oversimplified) about the true nature of genetic influences (Belsky & Israel, 2014; Zuk, Hechter, Sunyaev, & Lander, 2012). Almost everything that happens even at the cellular level is because of the combined influences of different molecular mechanisms, such as different proteins and functional RNA molecules. Given that, the idea that genotypes can just be summed together to arrive at a measure of genetic liability seems naïve.

To be sure, evidence for a substantial role of interactionism is lacking; however, the current evidence is primarily based on low-resolution tag SNP methodologies. That low-resolution methods have not yet substantiated the importance of gene–gene interactions, does not suggest they are not biologically important.

In sum, for a variety of methodological reasons, PGSs do not control for genetic heterogeneity. The final limitation of PGSs I consider relates to the neglect of developmental interactionism. As I discuss next, the well-known context-specificity of genetic influences (Feldman & Lewontin, 1975) impedes some of the intended uses of PGSs.

5.4 Context and population specificity

That heritability studies are context- and population-specific – a point made clearly and forcefully by Lewontin (1974) nearly 50 years ago – is now widely appreciated after considerable scholarly effort and some costly misrepresentations (Jensen, 1967). However, that GWASs and PGSs are similarly context- and population-specific is not as widely appreciated in theory or practice (but see Kaplan & Turkheimer, 2021). It should be. This is particularly true for non-biological social behaviors and achievements like educational attainment or same-sex sex, which involve somewhat arbitrary institutional structures (e.g., financial resources and opportunities) as well as cultural norms.¹⁷ For reasons expounded upon below, such genetic associations should not be understood as timeless, context-independent genetic influences. That is, even if we could disentangle the influence of genes from environments for these outcomes, these associations reflect developmental gene–environment interactions under current social arrangements in each context, not what could be in different circumstances (historical periods, social position, cultural context, etc.).

This well-known context- and population-specificity exists for two general reasons. The first is biological: Genes always interact with environments across all levels of development in their effects on complex traits. The second is sociocultural: The individual characteristics influencing traits or achievements, and thus the genetic contributors thereto, vary across historical time, society, and even across structural location. For illustration, the genetically influenced individual traits facilitating educational attainment for a woman in Saudi Arabia in 2000 versus a woman in 1870s in United States, in 2010 in India, in 2002 in Nigeria, in 1950 in Thailand, or in 2021 in United States are likely to be distinct in non-trivial ways. Although a woman going to college in the United States in 2020 would be conforming, a woman going to college in 1870s in United States would be statistically deviant. Because educational attainment reflects numerous genetically influenced traits, filtered by context and relative condition, the idea of a context-invariant “genetic propensity to” complex social outcomes like educational attainment, like crime, smoking, or same-sex sex, is misguided (Burt, 2023).

Moreover, the search for a “winning” genetic endowment that can be measured on a unidimensional scale representing propensity for social success is also misguided, in my view (e.g., Belsky et al., 2016). This is because our DNA is part of an interactional developmental system that responds to context- and condition-dependent stimuli (Burt, 2018; Ellis et al., 2012). Genetic differences influencing complex traits, like traits themselves, are not amenable to facile “good” or “bad,” “winning” or “losing” ratings but rather more like “it depends,” on a host of other factors (e.g., other genetic differences, other traits, historical context, social class, etc.). To use an oversimplified example, while being confident, independent, and talkative may enhance educational attainment and occupational success for an upper-middle class white male, those same traits among a minority youth from a disadvantaged background could very well impede educational attainment. Of course, confidence and independence emerge from a host of influences, but the point of this example is to reveal the oversimplified (theoretically and empirically unwarranted) model underlying an additive genetic index representing a context-independent propensity for complex social behaviors like educational attainment.

The problems with a unidimensional genetic propensity for complex biological traits are even more obvious for a phenotype of (having ever had) same-sex sex. As with people who attain higher levels of educational attainment, people who have ever had same-sex sex display remarkable diversity. From “gold star” lesbians and bisexual women to “femme” women who have same-sex sex only to please their male partners, the search for an additive, context-independent underlying continuum of genetic propensity for “having ever had same-sex sex” is empirically and theoretically unwarranted. Not only is there expansive heterogeneity within these groups, but also same-sex sex, like other social behaviors such as doing ballet, trying ecstasy (MDMA), and playing golf, is not simply the outer manifestation of some inner potentiality. Different sociocultural constraints and opportunities shape the behavioral manifestation of various traits and propensities, however genetic, which are then further altered by social responses in developmental feedback loops (including labeling and self-identification). Of course, we can impose a unidimensional propensity measure – a PGS or otherwise – for such heterogeneous and socially contingent behaviors by estimating the probability of the binary measure of having ever done so. But creating such a continuum statistically does not mean such a propensity exists biologically.

Thus, for yet another reason, PGSs cannot be thought of as “genetic potential,” inasmuch as genetic influences are not static charges where PGS effects sizes can be facilely compared across contexts or conditions. Traits that facilitate educational attainment, and any genetic contributions thereto, are dependent on sociocultural influences. For example, if physical education classes were equally emphasized with non-PE courses and graded not by effort but also by achievement, academic attainment may look noticeably different.

This context-specificity has implications for some prominent applications of PGSs. Following prior behavioral genetics work that examined how heritability estimates varied across contexts or conditions, several recent studies have used PGSs to explore how “genetic influences” are moderated by (often “constrained” or “suppressed” in) different contexts or for different social groups (Harden et al., 2020; Trejo et al., 2018; Wedow et al., 2018). For example, Herd et al. (2019) examined whether “the influence of genetics on educational attainment has changed across cohorts” and “whether this influence varies by gender” by comparing the effect sizes of the education PGS on educational attainment across cohorts (defined by historical time) and by sex. Their focal hypothesis was that among older cohorts, social structures of gender suppressed the “genetic potential for educational attainment” among women but not men, manifest in weaker education PGS prediction among women in older cohorts. To be sure, the Herd et al. study was explicitly sensitive to context, recognizing how genetic effects are “filtered, altered, and shaped by broader complex environments” (p. 1071). Even so, this approach remains insufficiently context-situated and oversimplified. This is because the study rests on the idea that PGSs capture a historically invariant genetic potential for educational attainment, such that weaker PGS prediction can be interpreted as lesser genetic influence and thus suppressed potential. However, for reasons mentioned above, as contexts and opportunities change, so too do the characteristics influencing achievements and social behaviors, and thus their genetic influences. A weaker PGS across contexts may just mean different traits matter (and would be expected in this example for statistical reasons given the lower mean and variance of educational attainment in the earlier cohorts compared to the latter ones). For all these reasons, interpreting effect size differences in PGSs as indicating that “genetic influences matter less” for social traits in different contexts or as evidence that “potential is suppressed” is unsound.

Upon deeper reflection, the extent to which research into how contexts suppress or constrain “genetic potential” (via reductions in PGS effect sizes) advances knowledge is unclear. Leaving aside my objection to the notion of a context-independent genetic potential for social traits, in general, and PGSs as an indicator of such potential, in particular, what, specifically, is the value of assessing whether “genetic potential” is suppressed by these social arrangements? Until well into the twentieth century, the potential for educational attainment for women in the United States was, of course, constrained by structures of gender that limited them to family roles in the household. We already know women's potential was suppressed, in these instances. What would it mean to say that potential was suppressed but not genetic potential? Is the null hypothesis that only “non-genetic potential” was suppressed (and what would that even mean)? Phrased alternatively, given that potential emerges from developmental systems shaped by interacting genetic and environmental forces, is there any argument that can be made that discriminatory arrangements or disadvantages constrain achievement but do not affect genetic potential? How would that work?

6. Questioning substantive value added

Even if the problems with environmental confounding could be solved, the justification for incorporating PGSs into social science is lacking. The scientific warrant to include PGSs to reveal well-established social patterns more precisely or rigorously is, in my view, wanting. Given that we have robust evidence that higher education is associated with higher income, fewer children, and better health, what is the value of demonstrating that an education PGS is associated with fewer children born, household wealth, or health? How could it not be? A recent study with an education PGS investigated whether “parental genetics for educational attainment” are associated with better (i.e., warm, stimulating) parenting, thereby partially explaining the association between parents' education PGS and youth educational attainment (Wertz et al., 2019). Armstrong-Carter et al. (2020) highlighted this study as illustrating how “genes can be used as a lens for the study of social processes through which parents influence their children.” Do we need GWASs, PGSs, and studies of “genetic nurture” to demonstrate that supportive, stimulating parenting is associated with child educational attainment and that higher educated – disproportionately well-off – parents are more likely to engage in such parenting? Or that “children who experience childhood disadvantage are not able to fully realize their educational potential” (Ronda et al., 2020). Or that “that genetic endowments linked to educational attainment strongly and robustly predict wealth at retirement” (Barth, Papageorge, & Thom, 2020). I think not.

Harden et al. (2020) touted the potential of PGSs as “molecular tracers” for social achievements, like educational attainment, that can “measure flows of students through the STEM pipeline and assess how these flows differ across schools” analogous to how “a radiologist might administer a radioactive tracer to track the flow of blood within the body.” However, the reason that radiologists use molecular tracers to trace internal functions is because they cannot observe such internal bodily processes. Unlike the radiologist tracking unobservable internal bodily processes like blood flow, we can observe and measure different student aptitudes, skills, and background factors and assess how these affect student progressions through educational systems. Given that opportunities exist for measuring background factors and proximal behaviors and that we already have a glut of assessments (e.g., grades, cognitive testing), the need for and utility of such a tracer – which those scholars admit is not a useful individual predictor – is surely questionable (Morris et al., 2020b).

In addition to meager benefits, such research has several potential costs. The use of PGSs as molecular tracers is rooted in the misguided idea that PGSs reflect individual propensity – that is, that the potential for educational success resides in our genome. Indeed, the authors argue that “[t]his approach offers a way of diagnosing the extent to which students who have high genetic propensities for success in education leak out of the STEM pipeline by failing to advance in their mathematics training” (Harden et al., 2020; emphasis added). Not only are PGSs flawed as measures of “high genetic potential” but the concern with the “high genetic potential” students “leaking out of the STEM pipeline” seems unjustified given paltry PGS individual prediction and the fact that potential for complex social achievements like years of education cannot be reduced to genotype (which the authors acknowledge). The paper evidences a heightened concern over the “high genetic potential” students leaking out over their “lesser potential” (lower PGS) counterparts, but this concern is never explained. Even more concerningly, this focus on the “high genetic propensity” seems to reflect the privileging of the purportedly “genetically gifted” in a manner that will increase rather than decrease inequalities.

To be sure, Harden et al. (2020) highlight the potential of the education PGS as a molecular tracer to inform school performance evaluations with the explicit aim of ameliorating inequality. However, such applications of school-level “genetic potential” performance assessment would, given existing social arrangements and environmental confounding, identify schools with a much higher proportion of lower income students from less-educated families as having lower genetic potential. Using PGSs as potentials, schools with such lower performing students would thus not be identified as “underperforming” because their students just “lost” in the “genetic lottery” (and we cannot expect much from them on this view). Although this is clearly not the intention of the authors, using PGSs as tracers necessarily rests on the idea of PGSs as indicating genetic potential for educational success – and, as noted, the authors use such terminology.¹⁸ Casting PGSs as “potential” risks reifying genetic differences among groups with different social behaviors and attainments shaped by prior and existing unequal arrangements as “genetic potential” and then excusing future patterns as inevitable because of genetic propensities, even for traits that are substantially driven by social inequalities and malleability.

These studies are in no way unique among sociogenomics studies but instead reflect the implicit “because we can” rationale of much sociogenomics research, often evidenced by the wholly uncompelling justification for some studies. Take the GWAS of “having ever had same-sex sex.” Ganna et al. (2019) explain the value of their study as follows: “With respect to genetic influences [on same-sex sex], several questions arise. First, what genes are involved and what biological processes do they affect? … Identification of robustly associated variants could enable exploration of the biological pathways and processes involved in development of same-sex sexual behavior” (p. 1). Leaving aside the implicit assumption of a molecular pathology underlying ‘non-heterosexuality' indicated by “having ever had same-sex sex,” as we have discussed, GWASs are not at all well suited for identifying genes, underlying causal variants, or tracing biological pathways for complex traits. In short, that scholars can conduct a study, does not mean that they should (i.e., that doing so advances science).¹⁹

From a broader perspective, sociogenomics' ambiguous contributions to knowledge are because of a prevailing deficit of theory, especially as relates to causal theories about developmental processes, which permits a rather shallow approach to the meaning of genetics plus social questions. To be sure, that social science genetics has a deficit of theory is not a novel criticism (e.g., Boardman & Fletcher, 2021; Burt, 2022; Panofsky, 2014), but attention to this neglect of theory and the manner in which this neglect hampers knowledge advancement is scarce. In my view, excitement over our ability to conduct analyses with incredibly advanced statistical and genetic tools appears to overshadow limitations and a sober evaluation of limitations. All too often, the contemporary enthusiasm around applying new genomics tools to social science adds a sheen that glosses over the meager practical and scientific contributions of this work, beyond simply showing that PGSs are statistically significant or have some non-trivial R ².²⁰ At this point, no serious scientist can suggest that genetic differences do not influence – in some complex, context-dependent way – developmental differences. Simply demonstrating that yet again with sophisticated, albeit biased, methods does not advance understanding (see also Turkheimer, 2016).

Finally, as noted, scholars point to PGSs as a control to “get genetics out of the way” to reveal aspects of our environment; however, I have yet to see any sociogenomics findings that change our understanding of environmental influences or suggest different policy or programmatic approaches. Given the limitations mentioned above, I am unable to conceive of any research findings at the present state of the science, which would support such changes in theory or practice. That is, even if the inclusion of PGSs markedly altered an environmental estimate, because PGSs are significantly environmentally confounded, we cannot say that controlling for “genetics” is the cause of such changes. What is more, we cannot say that environments matter “net of genetics” because PGSs only capture a fraction of the ostensible heritability of social outcomes. What, then, can or should we do? Below, I outline suggestions for sociogenomics at the current state of the science.

7. Suggestions

An abundance of genetic data is available for incorporation into social science with increasingly advanced computational methods and enhanced rigor in approach, relative to earlier eras. Given the limitations I have discussed along with my arguments about limited contributions, how should PGSs be used in social science, in my view? My answer is quite possibly unsatisfying: Sparingly and cautiously with caveats placed front and center. Enthusiasm about the opportunities genetics offers behavioral science should be tempered with a more realistic appraisal of current challenges and uncertainties. After all, we have been here – with excitement around genetics, limitations in methodology, and substantial unknown biology – before, quite recently, with the candidate gene era of a few years ago (see Charney, 2022).

Scholars should be more skeptical of the value added of PGSs to social science, and I have several suggestions to this end. First, when considering incorporating PGSs, behavioral scientists should first ask whether the outcome is a sufficiently tightly biologically regulated phenotype amenable to molecular genetic analyses. If so, scholars should explicitly specify how incorporating genetics advances science with a sufficiently high bar, one which acknowledges potential risks and benefits and recognizes that it is already well established that our genetic differences do matter in a complex, context-sensitive way (Turkheimer, 2016). Simply “showcasing the power of genetics” by revealing that PGSs are correlated with some outcome does not advance knowledge. Additionally, sociogenomics research should include controls for social variables associated with complex traits. At present, all too often easily measured and relevant social science predictors are not included in research “showcasing the power of genetics.” This is unsatisfactory.

Importantly, sociogenomics scholarship should eschew terminology that implies that genetic differences are driving behavioral differences given pervasive and unavoidable environmental confounding for all social outcomes. Framing PGSs as “genetic influences” should be avoided, and terminology like “association” or “correlation” should be employed instead. Likewise, I urge scholars to avoid “propensity” terminology or treating genetic endowment as a “lottery” in which there are winners and losers for complex social outcomes. Even if we could identify genetic influences on, for example, the type of intelligence that facilitates educational success and wealth, facilely equating genotypes associated with such capacities to “winning” at genetic inheritance or, conversely a lower education PGS as “an unfavorable genetic endowment” (e.g., Bolyard & Savelyev, 2020), is misguided. That is, of course, not to deny that people with greater wealth have better health and easier times dealing with stressors, on average; rather, it is to say that neither higher education nor greater wealth equals winning “the good life,” whatever that is.

In sum, I urge sociogenomics to think about where the science is, not where it might be (avoid hype and promissory notes); to acknowledge what questions we can answer at the current state of knowledge and which ones we cannot; and, finally, to recognize that just because social scientists can incorporate PGSs into our models, does not mean that we should – that is, that doing so advances knowledge.

8. Summary and discussion

Here, I challenged proponents' claims about the scientific warrant to include PGSs in social science. After outlining proponents' arguments about the utility of PGSs for social science, I argued that these ostensible scientific and practical benefits rely on the misguided notion that PGSs represent “genetic influences” on complex social traits. Instead, I explain that PGSs are unavoidably environmentally confounded because of population stratification, familial confounding, and downward (socio-environmental) causation. Although methods exist to mitigate the former, especially within-family studies, artificial genetic association signals created by downward causation cannot be differentiated from authentic genetic signals with the counterfactual models employed. In addition, I explain why PGSs do not, in fact, accurately or comprehensively control for “genetic influences” on traits because of methodological limitations (e.g., the tag SNP methodology) and biological challenges (including the nature of genetic influences). Finally, I discussed the context-specificity of PGSs, which precludes their use as “genetic potential” in general, and comparisons across contexts and conditions as a means of assessing the suppression of “genetic influences,” in particular. I explained that these models remain fundamentally and necessarily wedded to an overly simplistic and ultimately misleading (environmentally confounded and biologically implausible) reductionist genes-versus-environments approach.

In response to this critique, scholars may point to the fact that “PGSs just work.” By that, they presumably mean that PGSs “predict” the outcomes they were created to predict, even differences within families, albeit weakly in a manner that is inappropriate for individual prediction. However, the potential of PGSs is not rooted in their statistical predictive ability, however meager or substantial, but in their capturing genetic (vs. environmental) influences on trait differences. Furthermore, for complex social traits like education, as Morris et al. (2020b) documented in their evaluation of practical utility, an education PGS “provided little information on [youth] future achievement over phenotypic data that is either available or easily obtainable by educators.”

Others may respond by suggesting that I am holding sociogenomics methods to higher standards than standard social science methodologies.²¹ To that charge I cannot plead “not guilty.” Instead, I justify my scrutiny by pointing to the prior missteps in social science genetics, including the recent spectacular failure of the candidate gene era, the incautious hype, and the potential for misuse (see Dick et al., 2015; Yong, 2019). Moreover, proponents and critics alike have recognized that the scientific and social risks for the misinterpretation of PGSs are real and potentially significant, a situation exacerbated by the media tendency to ignore caveats and uncertainties and social scientists' lack of expertise in genetics (Barton et al., 2019; Richardson, 2017). These risks behoove us to approach the incorporation of genetics into social science with special caution and appropriate scientific skepticism.

Whether and to what extent incorporating genetics can benefit social science theory and research in a manner that may have practical implications remains to be seen. In my view, the payoffs for studying genetic influences on non-disease complex social traits and achievements for most applications are minimal. The potential costs of prematurely and misguidedly promoting PGSs as “genetic potential” are significant, and include, in addition to wasting finite resources searching for “genes for educational attainment,” obscuring social–structural and physical environmental influences and promoting the individualization of social problems.

9. Caveats and conclusion

My critique is intended to promote a dialogue between social and behavioral scientists about the scientific value of adding genetics to social science at the current state of knowledge. I hope this discussion eschews hype, straw man arguments, imputing motives, and ad hominem – all of which foster misunderstanding, polarization, even hostility. If we avoid such discussion-impairing tactics, which characterized some prior efforts to discuss genetics in social science, both science and society will be the better for it.

To avoid misunderstanding, I wish to clarify that my stance does not imply that the incorporation of genetics into social science necessarily involves racist motives and/or tacit support for eugenics; it quite clearly does not. Moreover, this critique is not motivated by a desire to censure scholars by imputing (bad) motives or to censor areas of study for ideological reasons or because of sociopolitical concerns. My aim is to draw attention to limitations of incorporating PGSs into social science and misinterpretations with the aim of promoting better science.

In the end, my argument is simply that the claims made by proponents about the benefits of PGSs and their utility as measures of “genetic influences” or “genetic propensity” are overstated and misguided. Because of these limitations, PGSs cannot be employed as measures of “genetic influences” as they are being used with increasingly regularity. GWASs and PGSs may be powerful tools for identifying genetic associations, but they are not the right tools for understanding complex social traits.