ArticlePDF Available

More Research Needed: There is a Robust Causal vs. Confounding Problem for Intelligence-associated Polygenic Scores in Context to Admixed American Populations

  • Ulster Institute for Social Research
  • Ulster Institute for Social Research
  • Ulster Institute for Social Research

Abstract and Figures

Amongst admixed American populations, polygenic scores for educational attainment and intelligence (eduPGS), genetic ancestry, and cognitive ability covary. We argue that this plausibly could be due to either confounding or to causally-relevant genetic differences between ancestral groups. It is important to determine which scenario is the case in order to better assess the validity of eduPGS. We investigate the robustness of the confounding vs. causal concern by examining, in detail, the relation between eduPGS, ancestry, and general cognitive ability in East Coast Hispanic and non-Hispanic samples. EduPGS predicted g among Hispanics (B = 0.175, N = 506) and all other groups (European: B = 0.230, N = 4914; European-African: B = 0.215, N = 228; African: B = 0.126, N = 2179) with controls for ancestry. Path analyses revealed that eduPGS, but not skin color, partially statistically explained the association between g and European ancestry among both Hispanics and the combined sample. Also, we were unable to account for eduPGS differences between the main ancestral populations using common tests for ascertainment bias and confounding related to population stratification. Overall, our results suggest that eduPGS derived from European samples can be used to predict g in American populations. However, owing to the uncertain cause of the ancestry-related differences in eduPGS, it is not yet clear how the effect of ancestry should be handled. We argue that causally-informative research is needed to determine the source of the relation between eduPGS, genetic ancestry, and cognitive ability.
Content may be subject to copyright.
MANKIND QUARTERLY 2021 62:1 151-185
More Research Needed: There is a Robust Causal vs.
Confounding Problem for Intelligence-associated
Polygenic Scores in Context to Admixed American
John G.R. Fuerst*
Cleveland State University, USA
Ulster Institute for Social Research, London, UK
Emil O.W. Kirkegaard, Davide Piffer
Ulster Institute for Social Research, London, UK
* Corresponding author: email
Amongst admixed American populations, polygenic scores for
educational attainment and intelligence (eduPGS), genetic ancestry, and
cognitive ability covary. We argue that this plausibly could be due to either
confounding or to causally-relevant genetic differences between ancestral
groups. It is important to determine which scenario is the case in order to
better assess the validity of eduPGS. We investigate the robustness of the
confounding vs. causal concern by examining, in detail, the relation
between eduPGS, ancestry, and general cognitive ability in East Coast
Hispanic and non-Hispanic samples. EduPGS predicted g among
Hispanics (B = 0.175, N = 506) and all other groups (European: B = 0.230,
N = 4914; European-African: B = 0.215, N = 228; African: B = 0.126, N =
2179) with controls for ancestry. Path analyses revealed that eduPGS, but
not skin color, partially statistically explained the association between g
and European ancestry among both Hispanics and the combined sample.
Also, we were unable to account for eduPGS differences between the
main ancestral populations using common tests for ascertainment bias
and confounding related to population stratification. Overall, our results
suggest that eduPGS derived from European samples can be used to
predict g in American populations. However, owing to the uncertain cause
of the ancestry-related differences in eduPGS, it is not yet clear how the
effect of ancestry should be handled. We argue that causally-informative
research is needed to determine the source of the relation between
eduPGS, genetic ancestry, and cognitive ability.
Keywords: Education, Intelligence, Polygenic scores, Hispanics, African-
Americans, Philadelphia
General intelligence is perhaps the most powerful variable in social science,
as it is not only measurable with relative ease but often strongly predicts
numerous academic, economic, occupational, social, and health-related
outcomes (Deary, 2010; Gottfredson, 2003). Owing to its relation to general
human well-being in contemporary society, understanding the causes of
individual differences in general intelligence is of significant social importance.
Moreover, the search for the causes of these differences has led many to
investigate the environmental and genetic determinants of general intelligence
(Plomin et al., 2014). More recently, large-scale genome-wide association studies
(GWAS) have been conducted to identify the genetic variants underlying the
hereditary contribution to individual differences (see, e.g., Lee et al., 2018;
Sniekers et al., 2017). At present, among Europeans, 4-10% of the variance in
cognitive ability can be explained by intelligence and educational polygenic
scores (eduPGS; Plomin & von Stumm, 2018).
GWAS have mostly been conducted on European-origin populations. Among
certain groups (e.g., African Americans), European-based eduPGS have been
found to display attenuated predictive accuracy with respect to cognitive ability
(Guo, Lin & Harris, 2019; Lasker et al., 2019; Rabinowitz et al., 2019). There are
several possible reasons for this attenuation. One explanation appeals to possible
lower within-group heritability in non-European groups (e.g., Rabinowitz et al.,
2019). This is a theoretically plausible account, since predictive accuracy is a
function of heritability (Daetwyler, Villanueva & Woolliams, 2008). However,
because the heritability of IQ is similar across racial and ethnic groups in the US
(for a meta-analytic review, see Pesta et al., 2020), a more likely possibility is
decay of linkage disequilibrium (LD), which results in different correlations
between SNPs across different ancestry groups (Zanetti & Weale, 2018).
The significance of LD decay in attenuating transethnic predictive accuracy
depends on the specific populations in question, as the impact of LD decay will
be modified by ancestry-assortative mating, selection, admixture, genetic drift and
other factors, which vary across populations. Although the predictive accuracy of
European-derived eduPGS was found to be attenuated among African
Americans, this did not seem to be the case in an admixed Brazilian sample
(Horta, Hartwig & Victora, 2018). Though, if predictive validity is lower in
unadmixed Africans than in Europeans, but an admixed population also has
greater variance in polygenic scores than unadmixed Europeans, this greater
variance in the admixed population can hide the lower predictive validity in
Africans. Nor was it the case for Asian and non-Black Hispanic Americans (Guo,
Lin & Harris, 2019). A similar pattern has been discovered for other, medically
related PGS, with the accuracy of PGS being attenuated the most for African
Americans (Duncan et al., 2019, Figure 2). This is likely because African
Americans are primarily a sub-Saharan African group, and because sub-Saharan
Africans are the continental lineage most genetically distant from Europeans (and
other major races; Duncan et al., 2019). Generally, the validity of eduPGS has to
be tested on a population by population basis and cannot be assumed to
Moreover, among admixed populations, evaluating the validity of PGS is
complicated when ancestry, PGS, and traits of interest covary. In these situations,
the covariance could result from a combination of ascertainment bias and
confounding related to population stratification, or could be a result of genetic
differences causally related to the trait (Lawson et al., 2020). Depending on the
scenario, corrections for ancestry may either be accurate, overcorrect, or
undercorrect. Thus, it is important to evaluate the reason for the intercorrelations.
In the case of admixed American populations, previous research has shown
that ancestry, eduPGS, and cognitive ability scores are inter-correlated (Lasker
et al., 2019). This is found when examining self-identified racial/ethnic (SIRE)
groups separately. For example, Kirkegaard et al. (2019), Lasker et al. (2019),
Warne (2020), and Guo, Lin and Harris (2019) found that cognitive ability was
associated with admixture components within ethnic groups (e.g., self-identifying
African Americans & Hispanics). These findings were robust to controls for SIRE,
despite the fact that, as Fang et al. (2019, p. 764) noted, SIRE “acts as a
surrogate to an array of social, cultural, behavioral, and environmental variables”
and so “stratifying on SIRE has the potential benefits of reducing heterogeneity
of these non-genetic variables and decoupling the correlation between genetic
and non-genetic factors.”
There are two obvious scenarios which could explain the covariance
between ancestry, eduPGS, and cognitive ability. These are depicted in Figure 1.
The first, (a), is the confounding scenario. Here, ancestry is associated with
causally irrelevant eduPGS-related loci as a result of ascertainment bias and
confounding related to population stratification (e.g., Kim et al., 2018; Martin et
al., 2017); coincidentally, ancestry is also associated with cognitive ability by way
of the environment. The environmental differences cause socioeconomic ones
which, in turn, cause cognitive ability ones. In this scenario, eduPGS would have
a spurious relation to g between groups, despite a causal one within groups. The
second, (b), is a causal scenario. In this case, ancestry is associated with causally
relevant eduPGS-related loci due to evolutionary history. These genetic
differences cause cognitive ability ones which, in turn, lead to socioeconomic
ones. In this case, eduPGS would be a component with constitutive explanatory
relevance in the relation between ancestry and g.
Figure 1. Theoretical scenarios depicting the relation between ancestry,
eduPGS, cognitive ability, and socioeconomic differences.
Regarding terminology, since the relation between ancestry and eduPGS is better
characterized as constitutive rather than causal, eduPGS could be better characterized
as being a component with constitutive explanatory relevance (Craver, 2007; Ylikoski,
2013), rather than a mediator as defined by Pearl (2014).
Note: In the confounding scenario (a) global or overall ancestry is correlated with
environmental differences (e.g., due to vertical environmentally-transmitted factors or,
perhaps, racial phenotype-based discrimination); these environmental differences cause
cognitive ones by way of socioeconomic ones; at the same time, global ancestry is
correlated with causally-irrelevant cognitive ability differences. In the alternative causal
scenario (b), global ancestry is correlated with causally-relevant PGS-associated loci
owing to evolutionary divergence in the quantitative genetic trait. These genetic
differences cause cognitive ability ones which in turn cause socioeconomic ones, which
lead to differences in external conditions. Note, this is not intended as a directed acylic
Determining which of the two models comes closer to reality is important for
an accurate interpretation of PGS in admixed American populations (Lawson et
al., 2020). We do not attempt to determine which scenario is correct in this paper.
Rather, we investigate if there is a robust confounding vs. causal problem with
respect to cognitive ability that needs to be solved by future research. This would
be the case if the data was found consistent with either a causal or confound
scenario as illustrated above.
To do this, we used the Trajectories of Complex Phenotypes sample, which
is a large population-representative Philidelphian sample. This sample has a
number of advantages over other available ones. First, it is a local sample and so
the issue of ancestry-related geographic confounding (Kirkegaard et al., 2019;
Lawson et al., 2020) is not of significant concern. Second, the cognitive battery is
well designed and allows for a robust examination of psychometric bias between
ethnic groups. Third, the heritability of cognitive ability has been previously
reported for the two largest ethnic groups (African and European Americans;
Mollon et al., 2021); this is important as the predictive validity of PGS depends on
the trait heritability (Pesta et al., 2020).
We focus here on the relation between ancestry, eduPGS, and general
cognitive ability in the Hispanic sample and compare these results with those from
European, biracial European-African, and African American samples. As far as
we are aware, no previous research has investigated this issue using a largely
Caribbean Hispanic sample. For background, Hispanics residing in Philadelphia
are largely a Caribbean-origin, admixed population, with a predominantly Puerto
Rican component (69% Puerto Rican; US Census Bureau, 2016). Hispanics, in
general, have been found to score significantly below European Americans on
measures of general intelligence, with Roth et al. (2017, Table 12) calculating
meta-analytic effects of d = -0.65 to -1.04.
Given these typically large differences, we assess measurement invariance
to ensure that our measure of cognitive ability functions the same for European
and Hispanic Americans. Next, via regression, we examine whether associations
between genetic ancestry and general cognitive ability can be accounted for by
either SIRE, skin color, or parental education. Following this, we evaluate the
predictive validity of eduPGS. We next examine whether the validity of eduPGS
is robust to controls for genetic ancestry and color. Using path modeling, we then
explore the extent to which eduPGS can statistically explain the association
between European genetic ancestry and general cognitive ability.
After, we apply Jensen’s Method of Correlated Vectors (MCV; Jensen, 1998)
to examine the relation between eduPGS and g-loadings. The expectation is that
eduPGS effects will be largest on the most g-loaded subtests since genetic
effects, in contrast to environmental effects, tend to be g-loaded (te Nijenhuis et
al., 2019). This phenomenon, of a positive correlation between the vector of g-
loading and some other vector, is referred to as a “Jensen effect” (Woodley of
Menie, Fernandes & Hopkins, 2015). While EduPGS has been found to be
differentially associated with subtest scores (de la Fuente et al., 2020; Genç et
al., 2021), these samples do not allow for a robust evaluation of whether eduPGS
exhibits a Jensen effect owing to a limited number of subtests or small sample
sizes. Thus, we evaluate the matter here. We further examine the relation
between g-loadings, group differences, and ancestry. Since a simple causal
scenario would predict a positive manifold of Jensen effects for eduPGS,
heritability, ancestry, and group differences, these analyses can show if the data
are consistent with such a scenario.
Finally, we examine the eduPGS scores for common forms of ascertainment
bias and confounding related to population stratification to see if we can easily
account for the effects of population structure. If we can, there may be no
confounding vs. causal scenario in need of resolving. To be clear, though, it is
outside the scope of this paper to investigate all possible forms of confounding.
We therefore restrict consideration to some forms of bias discussed in the
literature. The overall goal is to better understand how eduPGS, ancestry, SIRE,
color, and cognitive ability are associated with one another and to evaluate
whether there is a robust confound vs. causal concern in need of further research.
Materials and Methods
The Trajectories of Complex Phenotypes (TCP) study was conducted by the
Center for Applied Genomics at the Children's Hospital of Philadelphia, and the
Brain Behavior Laboratory at the University of Pennsylvania. Participants were
English-speaking Philadelphians, aged 8-21 years at the time of testing (which
was done primarily between 2010 and 2013). Those with severe cognitive or
medical impairments were excluded from the final sample.
Cognitive ability
Participants were administered the Penn Computerized Neurocognitive
Battery (PCNB; Gur et al., 2010; Moore et al., 2015), a psychometrically robust
cognitive battery that incorporates tasks linked to specific brain systems. This is
a widely used 1-hour neurocognitive battery, which has been previously validated
for this sample (Moore et al., 2015; Satterthwaite et al., 2016), which has an age
range of 8 to 22 years. Measurement invariance for this battery was previously
found to hold between African and European Americans (He & Li, 2021; Lasker
et al., 2019). We created g-scores based on ten subtests for which measurement
invariance (MI) held between our major SIRE groups (European, Hispanic, and
African Americans). Details about the measures, the tests for measurement
invariance, and subtest scores by subgroup can be found in Supplementary File
Parental education
Following Lasker et al. (2019), we computed z-scores individually for paternal
and maternal education and then averaged these. The average score was then
z-scored again. Paternal and maternal education were the only available
measures of socioeconomic status (SES). Parental education is arguably the
most relevant socioeconomic related variable as it has the strongest correlation
with educational-based PGS (Lee et al., 2018; Figure 4) and as it is as highly
correlated with cognitive ability / scholastic achievement as are other commonly
used parental-based indicators, such as occupation and income (Sirin, 2005,
Table 3). That said, parental education is not a complete measure of SES, and
we do not treat it as such.
Self-identified race/ethnicity (SIRE) was based on yes/no questions for which
the participants could select multiple races or ethnicities out of the following set
of choices: Black or African-American; American Indian or Alaskan Native; Asian;
European-American; Hispanic/Latino; Native Hawaiian/Pacific Islander; Other;
and Not available/Pending validation. Those who identified as European only and
did not report being Hispanic were coded as “European American,” and mutatis
mutandis for “African” and “European-African”. Those who identified as Hispanic,
regardless of self-identified race, were coded as such. Among Hispanics, we also
identified individuals by self-identified race: European, African, Other, and a
mixed category of all others (e.g., European-African Hispanic).
Genetic Ancestry Percentages
Different arrays covered different variants, so to obtain overlapping sets of
single nucleotide polymorphisms (SNPs), we imputed variants with the Michigan
Imputation Server ( We used
this server with the Minimac3 imputation algorithm, 1000G Phase 3 v5 as
reference panel, and Eagle v2.3 Phasing. For computational efficiency when
calculating admixture percentages, we filtered the 15.5 M variants available to the
6.5 M with a minor allele frequency (MAF) of at least 0.05. For color scores (based
on a total possible of 35 variants), we did not filter by MAF, so as not to lose
variants. Imputation was done using PLINK v1.90b6.8 (Chang et al., 2015). For
individual ancestry estimates, we used ADMIXTURE version 1.3.0 D.H.
(Alexander, Novembre & Lange, 2009). Prior, we merged the TCP with 1000
Genomes reference samples: European (British, CEU, Finnish, Spanish, and
Tuscan), African (African-American SW, African-Caribbean, Esan, Gambian
Mandinka, Luhya, Mende, and Yoruba), Amerindian (Peruvian), and mixed
American (Puerto Rican, Colombian, and Mexican-American). We did this to gain
a reliable estimate of Amerindian ancestry, since Hispanics in this sample were a
three-way admixed population. We then ran ADMIXTURE with k = 3 genetic
clusters. The results for the 1000 Genomes reference populations appear in
Figure 2, while those for the TCP European, European-African, African, and
Hispanic samples appear in Figure 3. Note that the Hispanics in the sample were
largely an Afro-European descent group, with an ancestry profile matching the
predominantly Puerto Rican origin of the Philadelphia Hispanic population.
Figure 2. Admixture ternary plot for 1000 Genomes reference samples.
Note: Due to clumping, some labels have been removed. Populations are: European
(British, CEU, Finnish, Spanish, Tuscan), African (African-American SW, African-
Caribbean, Esan, Gambian Mandinka, Luhya, Mende, Yoruba), Amerindian (Peruvian),
and mixed American (Puerto Rican, Colombian, and Mexican-American).
Figure 3. Admixture ternary plot for TCP self-identified race-ethnicity (SIRE)
Note: EA = non-Hispanic European American, EA_AA = non-Hispanic European-African
American, AA = non-Hispanic African American, HI_AA= Hispanic African American,
HI_EA= Hispanic European American, HI_OT = Hispanic Other, and Other = any other
individual who also identified as Hispanic.
Skin color
Because the data did not include measures of appearance, we followed
Lasker et al. (2019) and calculated phenotypic scores from genotypic data.
Namely, we imputed phenotype based on genotype using the HIrisPlex-S web
application ( Developed for use by the U.S.
Department of Justice in forensic investigations, and validated on thousands of
people from around the world (Chaitanya et al., 2018), the HIrisPlex-S imputes
skin, hair, and eye color probabilities from 41 SNPs (with overlaps: 6 for eye color,
22 for hair color, and 36 for skin color), with a high degree of accuracy. We focus
on skin color because this trait is given primacy by proponents of race-associated
phenotypic discrimination (“colorism”) models (e.g., Dixon & Telles, 2017; Marira
& Mitra, 2013), and because the other traits have more missing data, owing to
poor tagging of SNPs in some of the microarrays. A detailed discussion of the
color measure and validity are provided in Supplementary File 2.
Cognitive ability-related polygenic scores
Lasker et al. (2019) detailed the rationale for eduPGS selection. Briefly, we
validate four overlapping eduPGS from Lee et al. (2018): The eduPGS with all
variants trained without the 23andMe cohort (with N = 7,762,369 SNPs in the
present dataset), the multi-trait analysis of genome-wide association study
(MTAG) eduPGS 10k SNPs (with N = 8,442 SNPs in the present dataset), the
MTAG eduPGS lead SNPs (with N = 1,558 SNPs in the present dataset), and
finally Lee et al.’s (2018) putatively causal variants (with N = 111 SNPs in the
present dataset).
We used the MTAG 10k eduPGS for further analysis, since it had the highest
validity in our two largest groups (European and African). This was as predicted,
since both the MTAG eduPGS 10k SNPs and MTAGlead SNPs are predicted to
show moderate within-discovery population validity and moderate validity in non-
discovery populations. This is because 'lead' or 'clumped' SNPs with greater
statistical significance are more likely to be causal or very close to a causal
variant, which are also more likely to be transethnically valid (Grinde et al., 2019;
Marigorta & Navarro, 2013; Spencer, Cox & Walters, 2014; Wang & Teo, 2015).
In contrast, the eduPGS score based on all variants (without 23andMe) is
predicted to show high within discovery population validity, but poor validity in
non-discovery populations owing to high LD decay bias as a result of including a
large number of SNPs, irrespective of their significance (Zanetti & Weale, 2018).
And because there are only 111 causal SNPs, the validities are predicted to be
low in both discovery and non-discovery populations (e.g., Lasker et al., 2019).
In a supplementary analysis, we additionally examined the effect of
computing MTAG 10k eduPGS using population-GWAS versus within-family beta
weights for the SNPs. For the population-GWAS weights, we used Lee et al.’s
(2018) published MTAG weights computed based on 1.1 million individuals. For
the within family weights, we contacted Lee et al. (2018), who provided us with
the within family weights from their analyses of 22,000 sibling pairs. For these
analyses, we computed eduPGS using the subset of MTAG SNPs for which there
were both within-family and population-GWAS weights.
Descriptive statistics
Descriptive statistics for all groups appear in Table 1. Self-identifying
European, European-African, African, and Hispanic Americans were ancestrally
98%, 79%, 20%, and 60% European, respectively. Among Hispanics, those who
identified as European were 81% ancestrally European, while those who
identified as African were 29% ancestrally European. This difference in ancestry
is consistent with those reported previously (e.g., Table S5; Bryc et al., 2015).
The Hispanic Other group had the largest percentage of Amerindian ancestry, at
20%, which would be consistent with origins in Central or South America as
opposed to the Caribbean (Table S5; Bryc et al., 2015).
Table 1. Sample characteristics for the participants; mean ± standard deviation,
and sample size in italics.
Mean color
13.76 ± 3.64
0.98 ± 0.07
0.01 ± 0.04
0.34 ± 0.96
14.70 ± 3.84
13.15 ± 3.58
0.79 ± 0.28
0.01 ± 0.03
0.16 ± 0.95
19.18 ± 7.61
14.08 ± 3.75
0.20 ± 0.11
0.02 ± 0.02
-0.57 ± 0.77
30.96 ± 5.87
13.59 ± 3.86
0.60 ± 0.29
0.11 ± 0.15
-0.32 ± 1.01
22.19 ± 7.59
13.62 ± 4.09
0.81 ± 0.14
0.11 ± 0.09
-0.17 ± 1.03
19.45 ± 5.60
13.39 ± 3.94
0.60 ± 0.16
0.20 ± 0.18
-0.53 ± 0.98
21.14 ± 6.50
13.88 ± 3.76
0.29 ± 0.20
0.06 ± 0.11
-0.52 ± 0.88
28.45 ± 6.78
Other, not
13.42 ± 3.72
0.77 ± 0.27
0.09 ± 0.16
-0.02 ± 1.06
18.17 ± 6.31
Figure 4. Regression plot of the relation between g and European genetic
ancestry in the combined sample.
Note: EA = non-Hispanic European American; HI_AA = Hispanic African American,
HI_EA = Hispanic European American, HI_OT = Hispanic Other, and Other = any other
individuals who also identified as Hispanic.
The association between cognitive ability and European ancestry for the
combined sample is depicted in Figure 4. A Bayesian Generalized Additive Model
line (blue) is superimposed on a regression line (orange); moreover, self-
identified racial groups are color coded. As can be seen, the association is nearly
linear. Additionally, European ancestry is positively and significantly associated
with g in the admixed SIRE groups (rEuropean-African = .26, rAfrican = .084, rHispanic = .30),
though not for European Americans (rEuropean = .02) among whom there is little
non-European admixture.
Regression Analysis for Ancestry as a Predictor of g
Multiple regression analysis is preferable to bivariate analysis, since there is
a possibility of confounding with social and environmental factors, particularly
ones correlated with SIRE (Fang et al., 2019) and race-associated phenotype
(Conley & Fletcher, 2017), and also of a non-independence of admixture
components. As such, we relegate the bivariate results to Supplementary File 3.
In the initial regression analysis, shown in Table 2, we deal exclusively with
those who identify as Hispanic American. In the first, Model (1a), we include only
African and Amerindian ancestry as independent variables. In Model 1b, we add
self-identified race to see if ancestry retains predictivity and has independent
explanatory power. In Model 2, we approach the issue from the perspective of
“colorism” (that is, discrimination based on race-associated phenotype,
specifically color). As such, in Model 2a, we include only color. We then add
ancestry in 2b to see which variables retain validity. In Model 3 we further add
parental education to the model with color and ancestry. Note, all values except
genetic ancestry are standardized. This allows the B coefficients for ancestry to
be interpreted as a change in a standard deviation of cognitive ability going from
0% to 100% ancestry. Doing so allows results from different samples with
different variances in ancestry to be compared on the same metric.
As can be observed, African ancestry retained validity when either SIRE or
color were added. In contrast, with the inclusion of SIRE, Amerindian ancestry
became a nonsignificant predictor. However, the direction and magnitude of the
effect was as previously reported (Kirkegaard et al., 2019). As seen in Model 2a,
darker color was a significant predictor of lower cognitive ability, consistent with
previously reported findings (e.g., Hu et al., 2019). However, the effect became
nonsignificant and the sign reversed with the inclusion of ancestry. This is
consistent with the view that the association between color and ability is simply
indexing that between ancestry and ability (e.g., Hu et al., 2019; Lasker et al.,
2019). Note, Model 1 and Model 2 above have different numbers, since there
were fewer cases with color data; running the results with the same sample sizes
(listwise deletion) did not lead to a difference of interpretation.
Table 2. Regression analysis for ancestry as a predictor of g among Hispanic
Americans with controls for skin color (Model 2b), and parental education (Model
2c) added; B coefficients with standard errors in parentheses.
Model 1
Model 1b
Model 2a
Model 2b
Model 3
Ancestry: AFR
Ancestry: AMER
SIRE HI_European
SIRE Other w/ HI
Skin Color
Parental Education
Adjusted R2
Note: *p < .05, **p < .01, ***p < .001. Model 2a shows the results with color as an alternative
predictor. AMR = Amerindian ancestry. AFR = African ancestry. Reference SIRE =
Next, we repeat the analysis with the other groups (EA, EA-AA, and AA)
added. The results appear in Table 3, and are descriptively similar to those for
the Hispanic-only sample. As before, Amerindian ancestry loses validity on
inclusion of SIRE. This time, however, the effect size of Amerindian ancestry is
close to zero. Statistically, this happens because among Philadelphian
Europeans, the association between Amerindian ancestry and ability is non-
significantly positive (B = 0.500, S.E. = 0.391, N = 4914, p = .20), while the
association between African ancestry and ability is significantly negative (B = -
0.616, S.E. = 0.266, N = 4914, p < .05). Why this is the case is not clear. In Model
2a, color alone has validity, but it becomes nonsignificant and changes direction
on inclusion of ancestry (2b). This remains the case when parental education is
added in model 3.
Table 3. Regression analysis for ancestry as a predictor of g in the combined
sample with controls for skin color (Model 2) and parental education (Model 3)
added; B coefficient with standard error in parentheses.
Model 1a
Model 1b
Model 2a
Model 2b
Model 3
Ancestry: African
Ancestry: American
SIRE: African
SIRE: European/African
SIRE: HI_European
SIRE: HI_African
SIRE: HI_Other
SIRE: Other, no Hisp.
Skin Color
Parental education
Adjusted R2
Note: *p < .05, **p < .01, ***p < .001. Model 1b shows the results with color as an alternative
predictor. Reference SIRE = European.
Analyses of the Validity of eduPGS
Again, we relegate discussion of the bivariate results for the four educational
PGS to Supplementary File 3. And we used the MTAG 10k eduPGS for further
analysis. While this variable did not have the highest validity among Hispanics or
among European-Africans, it had the highest predictive validity among the two
largest groups, European and African Americans (Supplementary File 3; Table
S3-S6). Additionally, Lee et al. (2018; Table 4c) shows that MTAG eduPGS has
higher predictive validity than GWAS based eduPGS for cognitive ability in two
samples (the National Longitudinal Study of Adolescent to Adult Health and the
Health and Retirement Study). Thus we are justified in focusing on the MTAG-
based eduPGS.
The relation between this eduPGS and cognitive ability for the four major
groups is shown in Figure 5. The slopes (B) and intercepts based on the
regression equation were: Hispanic (B = 0.294, S.E. = 0.043; Intercept = -0.291;
N = 506), European (B = 0.230, S.E. = 0.014; Intercept = -0.005; N = 4914),
European-African (B = 0.284, 0.058; Intercept = 0.020; N = 228), and African
American (B = 0.151, S.E. = 0.029; Intercept = -0.747; N = 2179). The slope for the
Hispanic (t(5416) = 1.415, N.S.) and the European-African sample (t(5,138) =
0.905, N.S.) was not significantly steeper than the European one, though the
power to detect a significant difference was relatively low. For the African
American sample, the slope was significantly flatter (t(7,089) = 2.45, p <
Figure 5. Regression plot for the predictive validity of MTAG 10k eduPGS with
respect to g in the Hispanic (purple), European (green), European-African (blue),
and African American (red) samples.
Because we saw previously that ancestry is a robust predictor of cognitive
ability in the admixed samples, we next ran multivariate regression to test to what
extent the association between eduPGS and ability may be due to confounding
with either global ancestry or skin color. Model 1, Table 4, shows the effect of
eduPGS alone for Hispanics. EduPGS is significantly related to g (B = 0.294, N =
506, p < .001). When adding ancestry covariates in Model 2, the association
attenuates but remains significant (B = 0.175, N = 506, p < .001). Based on the
formula provided by Clogg, Petkova and Haritou (1995) for comparing betas in
nested models, the z-score for the difference was 1.79, which is statistically
significant (p = .037; one-tailed). Additionally, adding color in Model 3 did not
further attenuate this association (B = 0.194, N = 391, p < .001). Note, the
association in Model 2 for Hispanics was not substantially different from that in
the equivalent model for Europeans (Model 2European: B = 0.230, N = 4,914).
Table 4. Regression results for the effect of eduPGS on cognitive ability among
Hispanic Americans, B coefficient with standard error in parentheses.
Model 1
Model 2
Model 3
-0.291 (0.062)
-0.115 (0.084)
-0.030 (0.096)
0.294*** (0.043)
0.175*** (0.051)
0.194*** (0.051)
Ancestry: African
-0.814*** (0.201)
-0.989*** (0.266)
Ancestry: Amerindian
-0.472 (0.341)
-0.492 (0.412)
0.129 (0.076)
Adjusted R2
Note: *p <. 05, **p < .01, ***p < .001
Next we ran the same analysis on the two other heavily admixed groups,
European-African and African Americans. Table 5, Model 1 shows the effect of
eduPGS alone for European-African Americans (B = 0.284, N = 228, p < .001).
Adding ancestry in Model 2 seems to attenuate the association somewhat but the
effect remains significant (B = 0.215, N = 228, p < .01). When color was added in
Model 3, the association also remained significant (B = 0.181, N = 164, p < .05).
Table 5. Regression results for the effect of eduPGS on cognitive ability among
European-African Americans, B coefficient with standard error in parentheses.
Model 1
Model 2
Model 3
0.020 (0.074)
0.026 (0.086)
0.026 (0.124)
0.284*** (0.058)
0.215** (0.072)
0.181* (0.081)
Ancestry: African
-0.521 (0.296)
-0.743 (0.437)
Ancestry: Amerindian
4.247 (2.632)
5.740* (2.765)
0.140 (0.130)
Adjusted R2
Note: *p < .05, **p < .01, ***p < .001
For African Americans, shown in Table 6, the effect of eduPGS on g (B =
0.151, N = 2,179, p < .001) again seems to be attenuated slightly but remains
significant (B = 0.126, N = 2,179, p < .001) with the inclusion of ancestry in Model
2. As seen in Model 3, adding color did not further attenuate this association (B =
0.133, N = 1,526, p < .001).
Table 6. Regression results for the effect of eduPGS on cognitive ability among
African Americans, B coefficient with standard error in parentheses, B coefficient
with standard error in parentheses.
Model 1
Model 2
Model 3
-0.747 (0.055)
-0.423 (0.171)
-0.229 (0.204)
0.151*** (0.029)
0.126*** (0.030)
0.133*** (0.037)
Ancestry: African
-0.478* (0.219)
-0.532 (0.278)
Ancestry: Amerindian
0.398 (1.198)
-0.455 (1.691)
-0.066 (0.043)
Adjusted R2
Note: *p < .05, **p < .01, ***p < .001
Finally, for the combined sample, Model 1, Table 7, shows the effect of
eduPGS alone. EduPGS is significantly related to g (B = 0.384, N = 7920, p < .001).
When adding ancestry covariates in Model 2, the association is attenuated but
remains significant (B = 0.221, N = 7,920, p < .001). Additionally, adding SIRE and
color, in Model 4, did not further substantially attenuate this association (B =
0.218, N = 5,991, p < .001).
Table 7. Regression results for the effect of eduPGS on cognitive ability for the
combined sample.
Model 1
Model 2
Model 3
Model 4
Ancestry: African
Ancestry: Amerindian
SIRE: Hispanic AA
Model 1
Model 2
Model 3
Model 4
SIRE: Hispanic EA
SIRE: Hispanic Other
SIRE: Other
Adjusted R2
Note: *p < .05, **p < .01, ***p < .001. SIRE (self-identified race/ethnicity) categories: AA =
African American, EA = European American.
Path analysis
While fitting cross-sectional data to a path model cannot prove the causal
assumptions, doing so can provide estimates of the effect magnitudes on the
assumption that the model is correct (Bollen & Pearl, 2013). As such, we depict
two sets of path model results fit with the lavaan R package (Rosseel, 2012). We
limited the first path analysis to Hispanics to reduce transethnic bias in eduPGS
validity. In the first model, we include European ancestry, color, and eduPGS as
covariates. As color scores were not missing at random, we did not impute data,
but rather handled missing data with listwise deletion. As with previous analyses,
European ancestry was left unstandardized. The path model is shown in Figure
6. The detailed path estimates are shown in Supplementary File 4 (Table S1). In
the model, eduPGS partially explained the association between European
ancestry and cognitive ability. Independent of eduPGS, European ancestry was
also strongly positively associated with cognitive ability. As expected, European
ancestry was strongly negatively associated with darker color. However, darker
skin color had no significant independent effect on cognitive ability. Moreover, the
sign of the beta here was in the “wrong” direction relative to predictions from a
colorism model. That is, darker color was unexpectedly associated with higher
intelligence when controlling for ancestry. The model indicates that eduPGS is a
plausible component; color, in contrast, does not seem to be a plausible mediator.
In Figure 7, we repeat this analysis with the complete sample. Since SIRE
had little consistent effect, independent of ancestry, we do not include SIRE
variables in the path analysis. The results are comparable, except that the beta
for color is β = -.001 instead of β = .117 (both nonsignificant). Detailed results are
provided in Table S2 of Supplementary File 4.
Figure 6. Path diagram for the relation between European ancestry (EUR), skin
color, education polygenic score (eduPGS), and g in the Hispanic American
sample. N = 391. *Statistically significant at p < .001. Dashed-lines designate
Figure 7. Path diagram for the relation between European ancestry (EUR), skin
color, education polygenic score (eduPGS), and g in the complete sample. N =
5,991. * Statistically significant at p < .001. Dashed lines designate covariance.
As an alternative model, we include parental education instead of color. Color
was dropped as it was not a significant predictor of g and because there were
limited cases with color scores. While most parents are likely biological parents,
not all are, and so the adolescent’s ancestry is not exactly equivalent to the
midparent ancestry in this case. As such, we represent the relationship between
the adolescent’s European ancestry and their parent’s education as covariance
(indicated by dashed lines). For Hispanics, the path model is shown in Figure 8,
and the detailed estimates appear in Table S3 of Supplementary File 4. In the
model, eduPGS again partially explained the association between European
ancestry and g. However, parental education was also a significant predictor. The
covariance between parental education and adolescent eduPGS was significant.
With the current data, however, it is not possible to disentangle the causal paths
between European ancestry, parental education, adolescent eduPGS, and
adolescent g. This is because adolescent eduPGS approximates biological mid-
parent eduPGS. And it is expected that, within populations, mid-parent eduPGS
will be causally related to parental educational levels which, in turn, will be
genetically correlated with adolescent g (see, e.g., Trzaskowski et al., 2014).
Nonetheless, this model also indicates that eduPGS plausibly has constitutive
explanatory relevance regarding the relation between European ancestry and g.
Figure 8. Path diagram for the relation between European ancestry (EUR),
education polygenic score (eduPGS), parental education, and g in the Hispanic
American sample. N = 500. *Statistically significant at p < .001. Dashed lines
designate covariance.
Again, we repeat this analysis with the combined sample. The path model is
shown in Figure 9, and the estimates appear in Table S4 of Supplementary File
4. The results are comparable.
Figure 9. Path diagram for the relation between European ancestry (EUR),
education polygenic score (eduPGS), parental education, and g in the combined
sample. N = 7846. * Statistically significant at p < .001. Dashed-lines designate
The Spearman-Jensen hypothesis
If the association between eduPGS and cognitive ability is primarily genetic
in nature, a Jensen effect between eduPGS effects and subtest g-loadings is
expected, because genetic effects tend to primarily act through g (te Nijenhuis et
al., 2019). Moreover, because the weak version of Spearman’s hypothesis fits the
data well, for both the African-European and Hispanic-European differences, it is
expected that the magnitude of the differences on subtests will positively correlate
with the subtest g-loadings (i.e., there will be a Jensen effect on score
differences). Spearman’s hypothesis can also be extended to associations with
genetic ancestry. If the association between ancestry (not just SIRE) and
cognitive ability is primarily due to ancestry-related differences in g, a Jensen
effect with respect to ancestry would also be expected. The subtest correlations
and the subtest group differences used for this analysis are reported in Tables
S5-S6 of Supplementary File 4.
Table 8a,b shows the vector correlations (using unrounded vectors). The
results based on the ten subtests with measurement invariance appear above the
diagonal; while, those for all 15 subtests appear below. As can be seen, all
associations are moderately to strongly positive. The strong association between
g-loadings and eduPGS is consistent with the finding that Lee et al.’s (2018)
eduPGS is associated with genetic g (de la Fuente, 2020). Consistent with other
research, there is a strong Jensen effect on ethnic differences (te Nijenhuis, van
den Hoek & Dragt, 2019) and on ancestry-related differences within ethnic groups
(Hu et al., 2019; Lasker et al., 2019). Generally, the effect of eduPGS, like that of
ancestry and SIRE group differences, is pronounced on the most g-loaded and
more heritable subtests, consistent with the predictions of a causal scenario.
Table 8a. Results from Method of Correlated Vectors analysis. Correlations of
subtest scores with g (g loading), with % European ancestry and with education
polygenic score (PGS); size of standardized subtest gaps between SIRE groups
(AA = African-American, EA = European-American, HI = Hispanic, EAA = EA-AA
mixed), and subtest heritability (h2). N = 10 subtests above diagonal (for the
measurement invariant subtests), and 15 below diagonal (for all subtests).
g load
Anc. r
Anc. r
Anc. r
g loading
Ancestry r (all)
Ancestry r (AA)
Ancestry r (HI)
PGS r (all)
PGS r (EA)
PGS r (AA)
PGS r (HI)
EA/AA gap
EA/HI gap
Table 8b. Results from Method of Correlated Vectors analysis. Correlations of
subtest scores with g (g loading), with % European ancestry and with education
polygenic score (PGS); size of standardized subtest gaps between SIRE groups
(AA = African-American, EA = European-American, HI = Hispanic, EAA = EA-AA
mixed), and subtest heritability (h2). N = 10 subtests above diagonal (for the
measurement invariant subtests), and 15 below diagonal (for all subtests).
g loading
Ancestry r (all)
Ancestry r (AA)
Ancestry r (HI)
PGS r (all)
PGS r (EA)
PGS r (AA)
PGS r (HI)
EA/AA gap
EA/HI gap
Bias in Education-related PGS (eduPGS)
Evaluation of bias in the TCP sample
PGS may be biased due to the source population with which they were
computed (i.e., ascertainment bias). There are a couple of obvious mechanisms
by which this bias can occur. First, European-based eduPGS may be biased
against non-Europeans due to the inclusion of European-specific variants
(Thomson, 2019). These population-specific variants might have very low
frequencies in non-European populations. Second, out-of-African based eduPGS
may be biased against African populations due to an overrepresentation of
derived (due to new mutations) versus ancestral (shared with other primates)
variants in the out-of-Africa populations (Kim et al., 2018; Thomson, 2019). As a
robustness check, we investigate both possibilities.
First, we computed MTAG eduPGS excluding variants with minor allele
frequency (MAF) < .01 (leaving 7,636 overlapping variants) and < .05 (leaving
7,172 overlapping variants) among African lineages, using the 1000 Genomes
reference samples to determine the African MAF. As a result, these eduPGS
exclude variants not also present in African populations. As seen in Table 9, this
exclusion had no substantive effect on the mean eduPGS differences between
Table 9. eduPGS for European, Hispanic, and African American participants,
using either the complete 10k eduPGS, or the score computed after exclusion of
SNPs with minor allele frequencies of < 1% or < 5% in Africans. Standard
deviations appear in parentheses.
SIRE group
MAF >.01
MAF >.05
Hisp. European
Hisp. African
Hisp. Other
Other, not Hispanic
Kim et al. (2018) demonstrated that when allelic risk scores are based on
out-of-Africa populations, African populations show elevated frequencies of
disease-associated loci for ancestral (shared with other primates) alleles, and
reduced frequencies for derived (due to new mutations after the split with
primates) alleles, even when there are no underlying trait differences. They
conclude that “systematic allele frequency differences between populations need
not be due to any underlying difference in risk” (p. 5) and propose corrections for
bias due to ancestral versus derived allele status. To investigate this, we compute
eduPGS by derived and ancestral status. To be clear, we computed one PGS
with only those SNPs where the enhancing allele is derived, and then another
PGS with only those SNPs where the enhancing allele is ancestral. In this case,
risk alleles and trait-enhancing alleles are the same thing; medical versus
cognitive GWAS studies just use different terminology.
The results are shown in Table 10. As seen, contrary to the findings of Kim
et al. (2018), with eduPGS, non-European populations have both lower derived
and ancestral eduPGS scores. Moreover, the differences are largest for the
derived ones. As a result, when Kim et al.’s (2018, p. 12) correction is applied,
the polygenic score gaps change little.
Table 10. eduPGS for European, European-African, African, and Hispanic
American participants, mean ± standard deviation.
SIRE group
0.02 ± 0.99
0.01 ± 1.00
-0.68 ± 1.15
-0.34 ± 1.12
-2.05 ± 0.85
-1.14 ± 0.79
-1.05 ± 1.21
-0.66 ± 1.02
Hisp.: European
-0.57 ± 1.07
-0.45 ± 1.05
Hisp.: Other
-1.01 ± 1.03
-0.72 ± 1.08
Hisp.: African
-1.84 ± 1.06
-1.06 ± 0.82
Other, not Hispanic
-0.60 ± 1.20
-0.31 ± 0.97
* Corrected following Kim et al.’s (2018) procedure, which involves adjusting the ancestral
PGS down by 0.1902% and adjusting the derived PGS up by 0.1082% and then averaging
the two. Kim et al. report that, in general, "ancestral risk [i.e., trait enhancing] alleles are
found at 9.51% higher frequency in Africa, and derived risk [i.e., trait enhancing] alleles
are found at 5.40% lower frequency in Africa" (p. 1). They then applied this general finding,
with percentages doubled to account for diploidy, to specific traits. We do the same.
Additionally, we computed the validities to see if the corrections affected
these. The validities by eduPGS are reported in Table 11. As seen, the validities
for the different eduPGS were approximately the same for a given SIRE group.
Table 11. Predictive validities of MAF 0.01, MAF 0.05, derived, and ancestral
eduPGS among European, European-African, Hispanic, African and Hispanic
Americans (N = sample size).
SIRE group
MTAG 10k
MAF 0.01
MAF 0.05
Finally, it has been argued that the eduPGS differences may be upwardly
biased by using the discovery sample-based direction of effects for SNPs
(Thompson, 2019). To investigate this we computed the betas for the MTAG
SNPs, used as predictors of g, for European (N = 4,939) and African Americans
(N = 2,228) separately. The SIRE specific betas are provided in the
Supplementary Material. We then examined mean differences and validities for
the variants which showed transracially concordant effects across ethnic groups
as compared to those which showed discordant effects. African Americans had
lower eduPGS than EA based on the concordant SNPs (-3.13 and 0.00,
respectively), but slightly higher eduPGS based on the discordant SNPs (0.49
and 0.00, respectively). Moreover, among African Americans, the concordant
SNPs showed higher validity than the discordant SNPs for g. Cross-validation
confirmed that the concordant eduPGS were more predictive than the discordant
eduPGS among African Americans. A possible interpretation is that the
discordant eduPGS, which are less likely to be causal, contain more LD decay
related effects. Regardless, there is no evidence, based on this sample, that the
eduPGS differences are being inflated by the inclusion of SNPs with
transethnically discordant effects as some have argued (Thompson, 2019).
Evaluation of bias in the 1000 Genomes samples
We additionally ran supplementary analyses which leveraged the 1000
Genomes data to explore the effects of score construction on the population
differences. Methods and detailed results are reported in Supplementary File 5.
First, we examined the effect of using population-GWAS vs. within-family PGS.
In this analysis, we found that the African-European differences based on the
within-family PGS were reduced in size but nonetheless large. Next, we examine
the effect of using trans-ethnically concordant betas. The results were similar to
those reported in our sample. Finally, we examined the predictive validity of
different polygenic score constructions across populations. While the expected
associations remained positive, we found that the construction of eduPGS
affected the magnitude of correlations, consistent with reports by others (e.g.,
Berg et al., 2019, Figure 1; Sohail et al., 2019, Figure 4). These results highlight
Duncan et al.’s (2019) caution that different eduPGS can give markedly different
results, rendering interpretation uncertain.
To better understand how eduPGS function in admixed American
populations, and to determine if there is a robust confound vs. causal problem,
we examined the association between intelligence-related polygenic scores,
global ancestry, and general cognitive ability in Hispanic and non-Hispanic
European, European-African, and African American samples. In this sample, we
were able to confirm full factorial invariance for the cognitive tests in the case of
the European-Hispanic and European-African differences. And, previously, full
factorial invariance had been found to hold along the full range of European
ancestry (Lasker et al., 2019). Thus, the associations found reflect those with
latent mental ability.
Among Hispanics and in the combined Hispanic and non-Hispanic American
sample, the association between European/African ancestry and cognitive ability
was robust to controls for SIRE, color, and parental education. For Amerindian
ancestry, in both the Hispanic-only and the combined samples, the association
with cognitive ability became nonsignificant when SIRE was added as a covariate.
In the Hispanic-only sample, the effect remained directionally consistent with
previously reported results (Kirkegaard et al., 2019; Warne, 2020). Statistically,
the insignificance, in this case, was a result of the higher standard errors of
Amerindian ancestry, as compared to African, which was due to the lower
variance in Amerindian ancestry. Similarly, in a small, “non-Hispanic Caucasian”
sample, with little variance in ancestry, European ancestry versus (apparently)
Mexican/Mexican-American ancestry was only weakly and non-significantly
associated with IQ (r = .09, N = 120; Wang et al., 2016). For the combined sample,
however, the effect of Amerindian ancestry turned positive with SIRE controls.
Statistically this was due to Amerindian ancestry being slightly positively
correlated with general intelligence in the non-Hispanic White sample (r = .014; N
= 4914, N.S.) and to the much larger non-Hispanic White than Hispanic sample
size. Generally, the relation between cognitive ability and Amerindian ancestry
needs to be better explored in predominantly European-Amerindian “Mestizo”
samples. The results to date are ambiguous.
Here, explanations for the association between European ancestry and
cognitive ability by confounding with either geography or racial phenotype (see,
e.g., Conley & Fletcher, 2017) are not viable. This was a local population sample
from the Philadelphia area, so geographic confounding is not a substantial
concern. We did find a significant negative association between cognitive ability
and darker skin color. However, rather than the association between ancestry and
cognitive ability being explained by color, the association between color and
cognitive ability was statistically explained by ancestry.
These latter results are key in that there is a large literature on “colorism”
which purports to demonstrate color or pigment-based discrimination by showing
mere correlations between color and social outcomes (e.g., Marira & Mitra, 2013).
Moreover, it has been argued by some that such discrimination might account for
potential associations between ancestry and cognitive ability (e.g., Conley &
Fletcher, 2017). However, our results concur with the competing distributional
model described by Hu et al. (2019, Figure 1), in which the association between
color and cognitive ability is a proxy (versus a cause) of that between ancestry
and cognitive ability. This finding has implications for genetic research (Lawson
et al., 2020), as it suggests that color-based discrimination is not likely an
additional source of confounding.
We found that eduPGS was significantly associated with cognitive ability
within the European, European-African, African, and Hispanic samples. For the
MTAG 10k eduPGS, which we use for further analyses, the associations in the
admixed populations were attenuated with the inclusion of ancestry. Regardless,
the association remained significant and not substantially different from that in the
European sample (BEuropean = .230 vs. BEuropean-African = .215 and BHispanic = .175),
except in the case of African Americans (BAfrican = .126) where, while statistically
significant, the association was less than that for European Americans.
Moreover, path analysis indicated that eduPGS partially statistically
explained the association between European ancestry and cognitive ability,
whereas color did not. Additionally, the explanatory power of eduPGS was not
fully accounted for by parental education, a variable which is both genetically and
environmentally correlated with adolescent intelligence within groups and thus
may be likewise between groups.
The eduPGS differences may be spurious owing to difficult to control
ascertainment bias and confounding related to population stratification. For
example, Berg et al. (2019) found attenuated effects for height PGS when
applying UK versus pan-European-based PGS to Eurasian samples. The reason
seems to be that controls for ancestry components do not always fully capture
population structure effects, however, some confounding effects can be avoided
by computing PGS using a more homogeneous population and then applying
these to the more heterogeneous populations of interest (see: Berg & Coop,
2014). In our case, though, we start with eduPGS based on European origin
samples and then apply these to samples of different continental ancestry. As
such, we already incorporated this component of Berg et al.’s (2019) analysis into
We further investigated whether differences might be due to biasing effects
of discovery population-specific variants, whether the allele associated with
higher values of the trait is derived or ancestral, failure of SNP sign concordance
between populations, and discrepancy between population-GWAS and within-
family coefficients. The forms of ascertainment bias and confounding related to
population stratification addressed by our procedures do not explain the group
differences in eduPGS. However, it is beyond the scope of this paper to explore
all forms of possible confounding.
The regression and path analysis results raise some possible concerns
regarding the use and interpretation of eduPGS in admixed American
populations. Ancestry covaries with trait and eduPGS scores. This could either
be due to trait-relevant genetic differences between the ancestral groups of the
admixed populations or to a mix of both environmental differences and also
ascertainment bias and confounding related to population stratification. If the
former, controlling for ancestry can attenuate the effect of eduPGS; however, if
the latter, leaving ancestry unadjusted can inflate it.
Additionally, MCV indicated that the effect of eduPGS was strongly g-loaded,
as was the effect of ancestry and group differences. This is consistent with all
these effects acting primarily by way of g. These results for eduPGS are not self-
evident. This is because it has been shown that some of the effect of eduPGS is
a shared environmental effect (Domingue & Fletcher, 2019); however, the effect
of adoption, a shared environmental effect, exhibits an anti-Jensen effect (te
Nijenhuis, Jongeneel-Grimen & Armstrong, 2015). Thus, that eduPGS would act
like typical genetic effects, and be most pronounced on g-loaded subtests, is not
obvious. In addition to eduPGS, both ancestry and group differences exhibited
Jensen effects, which is consistent with differences being primarily in g (i.e.,
Spearman’s hypothesis). A practical implication of this is that good measures of
g are needed to capture the full statistical effects of cognitive ability in context to
research on eduPGS and related variables. Moreover, the positive manifold of
Jensen effects is at least consistent with a causal model. To reconcile it with a
confounding model one would need to additionally propose a mechanism by
which environmentally induced phenotypic differences produced Jensen effects
(e.g., Flynn, 2019).
The differences in eduPGS between ancestral groups are large enough that
this is an issue which future research should try to resolve. While the magnitudes
of differences for the true causal SNPs are unknown, magnitudes can be
calculated for presently known education-related SNPs. We can use the fixation
index, a measure of population differentiation, to do this. Supplementary File 6
shows the Fst for the SNPs from Lee et al. (2018) for 1000 Genomes super-
populations. The Fst value between Europeans and Africans, which are the two
main ancestral groups for the admixed populations here, is .1090 As detailed in
Supplementary File 3, for typically reported heritabilities (h2 = .5; Pesta et al.,
2020; Polderman et al., 2015) this magnitude of population differentiation gives
medium to large expected mean differences (i.e., when environments are equal),
in this case, equivalent to d = 0.68. Thus, if transethnically unbiased eduPGS
differences turn out to be commensurate with those based on Lee et al.’s (2018)
eduPGS, medium to large phenotypic differences are expected under conditions
of environmental equality. Controlling for ancestry might then bias eduPGS
Related to this point, a reader suggested that we should run analyses to
detect polygenic selection as done by Berg et al. (2019). However, whether
differences are due to drift or selection is not necessarily relevant to whether there
are genetic differences between populations. Moreover, given the expected
differences discussed above, selection, in the form of stabilizing or convergent
selection between populations is needed to show trait equality, not trait
differences. That is, the evolutionary default or null expectation would be that trait-
causing SNP frequency differences will be commensurate with random SNP
ones, not that selection acted to homogenize differences in this particular trait
(Edelaar & Björklund, 2011; Edge & Rosenberg, 2015; Leinonen et al., 2013;
Rosenberg et al., 2019). As Rosenberg et al. note: “[P]henotypic differences
among populations are predicted under neutrality to be similar in magnitude to
typical genetic differences among populations” (p. 30). We are not aware of
anyone who has established selection acting between populations such as to
make causally-relevant genetic differences (related to education and intelligence)
substantially smaller than expected by drift. Whether there was divergent
selection is undetermined (e.g., Guo et al., 2018; Racimo, Berg & Pickrell, 2018).
Considering these results, it may be that individual ancestry tracks cognitive
ability in admixed populations across the Americas. If so, proper interpretation of
the predictive accuracy of eduPGS in American admixed populations will require
a better understanding of the causal pathways underwriting this association. We
argue that admixture mapping is the appropriate next step (for a rationale, see
Kirkegaard et al., 2019). That said, different sub-ancestral components (e.g.,
North vs. South Amerindian / European, West versus East African) may yield
different associations between global ancestry and cognitive ability, so additional
global ancestry studies are warranted to better understand the pattern of effects.
Owing to the sample sizes of the subgroups, some of the ancillary analyses,
such as the comparison of eduPGS validities, were underpowered. Replications
should be attempted with larger samples; until then, caution is warranted
regarding interpretation. Importantly, we did not attempt to determine the cause
of the covariance between eduPGS, ancestry, and g. We can only say that the
cause is presently undetermined and that it needs to be resolved for a proper
interpretation of the predictive accuracy of eduPGS in admixed American
populations. While we suggest local admixture mapping as a way to narrow the
uncertainty, further research should also explore other forms of confounding.
All supplemental files are available online at
Alexander, D.H., Novembre, J. & Lange, K. (2009). Fast model-based estimation of
ancestry in unrelated individuals. Genome Research 19: 1655-1664.
Berg, J.J. & Coop, G. (2014). A population genetic signal of polygenic adaptation. PLoS
Genetics 10(8): e1004412.
Berg, J.J., Harpak, A., Sinnott-Armstrong, N., Joergensen, A.M., Mostafavi, H., Field, Y.,
... & Coop, G. (2019). Reduced signal for polygenic adaptation of height in UK Biobank.
Elife 8: e39725.
Bollen, K.A. & Pearl, J. (2013). Eight myths about causality and structural equation
models. In: S.L. Morgan (ed.), Handbook of Causal Analysis for Social Research, pp.
301-328. Dordrecht: Springer.
Bryc, K., Durand, E.Y., Macpherson, J.M., Reich, D. & Mountain, J.L. (2015). The
genetic ancestry of African Americans, Latinos, and European Americans across the
United States. American Journal of Human Genetics 96: 37-53.
Chaitanya, L., Breslin, K., Zuñiga, S., Wirken, L., Pośpiech, E., Kukla-Bartoszek, M., ...
& Kayser, M. (2018). The HIrisPlex-S system for eye, hair and skin colour prediction
from DNA: Introduction and forensic developmental validation. Forensic Science
International: Genetics 35: 123-135.
Chang, C.C., Chow, C.C., Tellier, L.C., Vattikuti, S., Purcell, S.M. & Lee, J.J. (2015).
Second-generation PLINK: Rising to the challenge of larger and richer datasets.
Gigascience 4(1): 7.
Clogg, C.C., Petkova, E. & Haritou, A. (1995). Statistical methods for comparing
regression coefficients between models. American Journal of Sociology 100: 1261-
Conley, D. & Fletcher, J. (2017). The Genome Factor: What the Social Genomics
Revolution Reveals about Ourselves, Our History, and the Future. Princeton, NJ:
Princeton Univ. Press.
Craver, C. (2007). Constitutive explanatory relevance. Journal of Philosophical
Research 32: 3-20.
Daetwyler, H.D., Villanueva, B. & Woolliams, J.A. (2008). Accuracy of predicting the
genetic risk of disease using a genome-wide approach. PLoS One 3(10): e3395.
de la Fuente, J., Davies, G., Grotzinger, A.D., Tucker-Drob, E.M. & Deary, I.J. (2020). A
general dimension of genetic sharing across diverse cognitive traits inferred from
molecular data. Nature Human Behaviour 5: 49-58.
Deary, I.J. (2010). Cognitive epidemiology: Its rise, its current issues, and its challenges.
Personality and Individual Differences 49: 337-343.
Dixon, A.R. & Telles, E.E. (2017). Skin color and colorism: Global research, concepts,
and measurement. Annual Review of Sociology 43: 405-424.
Domingue, B.W. & Fletcher, J. (2019). Separating measured genetic and environmental
effects: Evidence linking parental genotype and adopted child outcomes. Behavior
Genetics 50: 301-309.
Duncan, L., Shen, H., Gelaye, B., Meijsen, J., Ressler, K., Feldman, M., ... & Domingue,
B. (2019). Analysis of polygenic risk score usage and performance in diverse human
populations. Nature Communications 10(1): 3328.
Edelaar, P.I.M. & Björklund, M. (2011). If FST does not measure neutral genetic
differentiation, then comparing it with QST is misleading. Or is it? Molecular Ecology 20:
Edge, M.D. & Rosenberg, N.A. (2015). Implications of the apportionment of human
genetic diversity for the apportionment of human phenotypic diversity. Studies in History
and Philosophy of Science Part C: Studies in History and Philosophy of Biological and
Biomedical Sciences 52: 32-45.
Fang, H., Hui, Q., Lynch, J., Honerlaw, J., Assimes, T.L., Huang, J., ... & DuVall, S.L.
(2019). Harmonizing genetic ancestry and self-identified race/ethnicity in genome-wide
association studies. American Journal of Human Genetics 105: 763-772.
Flynn, J.R. (2019). A response to te Nijenhuis et al. (2019). Journal of Biosocial Science
51: 913-916.
Genç, E., Schlüter, C., Fraenz, C., Arning, L., Metzen, D., Nguyen, H.P., ... &
Ocklenburg, S. (2021). Polygenic scores for cognitive abilities and their association with
different aspects of general intelligence a deep phenotyping approach. Molecular
Gottfredson, L.S. (2003). g, jobs, and life. In: H. Nyborg (ed.), The Scientific Study of
General Intelligence: Tribute to Arthur R. Jensen, pp. 293-342. New York: Pergamon.
Grinde, K.E., Qi, Q., Thornton, T.A., Liu, S., Shadyab, A.H., Chan, K.H.K., ... & Sofer, T.
(2019). Generalizing polygenic risk scores from Europeans to
Hispanics/Latinos. Genetic Epidemiology 43: 50-62.
Guo, G., Lin, M.J. & Harris, K.M. (2019). Socioeconomic and genomic roots of verbal
ability. bioRxiv.
Guo, J., Wu, Y., Zhu, Z., Zheng, Z., Trzaskowski, M., Zeng, J., ... & Yang, J. (2018).
Global genetic differentiation of complex traits shaped by natural selection in humans.
Nature Communications 9: 1865.
Gur, R.C., Richard, J., Hughett, P., Calkins, M.E., Macy, L., Bilker, W.B., ... & Gur, R.E.
(2010). A cognitive neuroscience-based computerized battery for efficient measurement
of individual differences: Standardization and initial construct validation. Journal of
Neuroscience Methods 187: 254-262.
He, Q. & Li, J.J. (2021). Factorial invariance in hierarchical factor models of mental
disorders in African American and European American youths. Journal of Child
Psychology and Psychiatry 62: 289-298.
Horta, B.L., Hartwig, F.P. & Victora, C.G. (2018). Breastfeeding and intelligence in
adulthood: Due to genetic confounding? Lancet Global Health 6(12): e1276-e1277.
Hu, M., Lasker, J., Kirkegaard, E.O. & Fuerst, J.G. (2019). Filling in the gaps: The
association between intelligence and both color and parent-reported ancestry in the
National Longitudinal Survey of Youth 1997. Psych 1(1): 240-261.
Jensen, A.R. (1998). The g Factor: The Science of Mental Ability. Westport, CT:
Kim, M.S., Patel, K.P., Teng, A.K., Berens, A.J. & Lachance, J. (2018). Genetic disease
risks can be misestimated across global populations. Genome Biology 19: 179.
Kirkegaard, E.O., Woodley of Menie, M.A., Williams, R.L., Fuerst, J. & Meisenberg, G.
(2019). Biogeographic ancestry, cognitive ability and socioeconomic outcomes. Psych
1(1): 1-25.
Lasker, J., Pesta, B.J., Fuerst, J.G. & Kirkegaard, E.O. (2019). Global ancestry and
cognitive ability. Psych 1(1): 431-459.
Lawson, D.J., Davies, N.M., Haworth, S., Ashraf, B., Howe, L., Crawford, A., ... &
Timpson, N.J. (2020). Is population structure in the genetic biobank era irrelevant, a
challenge, or an opportunity? Human Genetics 139: 23-41.
Lee, J.J., Wedow, R., Okbay, A., Kong, E., Maghzian, O., Zacher, M., ... & Fontana,
M.A. (2018). Gene discovery and polygenic prediction from a 1.1-million-person GWAS
of educational attainment. Nature Genetics 50: 1112-1121.
Leinonen, T., McCairns, R.S., O'Hara, R.B. & Merilä, J. (2013). QSTFST comparisons:
Evolutionary and ecological insights from genomic heterogeneity. Nature Reviews
Genetics 14: 179-190.
Marigorta, U.M. & Navarro, A. (2013). High trans-ethnic replicability of GWAS results
implies common causal variants. PLoS Genetics 9(6): e1003566.
Marira, T.D. & Mitra, P. (2013). Colorism: Ubiquitous yet understudied. Industrial and
Organizational Psychology 6(1): 103-107.
Martin, A.R., Gignoux, C.R., Walters, R.K., Wojcik, G.L., Neale, B.M., Gravel, S., ... &
Kenny, E.E. (2017). Human demographic history impacts genetic risk prediction across
diverse populations. American Journal of Human Genetics 100: 635-649.
Mollon, J., Knowles, E.E., Mathias, S.R., Gur, R., Peralta, J.M., Weiner, D.J., ... &
Glahn, D.C. (2021). Genetic influence on cognitive development between childhood and
adulthood. Molecular Psychiatry 26: 656-665.
Moore, T.M., Reise, S.P., Gur, R.E., Hakonarson, H. & Gur, R.C. (2015). Psychometric
properties of the Penn Computerized Neurocognitive Battery. Neuropsychology 29: 235-
Pearl, J. (2014). Interpretation and identification of causal mediation. Psychological
Methods 19: 459-481.
Pesta, B.J., Kirkegaard, E.O., te Nijenhuis, J., Lasker, J. & Fuerst, J.G. (2020). Racial
and ethnic group differences in the heritability of intelligence: A systematic review and
meta-analysis. Intelligence 78: 101408.
Plomin, R. & von Stumm, S. (2018). The new genetics of intelligence. Nature Reviews
Genetics 19: 148-159.
Plomin, R., DeFries, J.C., Knopik, V.S. & Neiderhiser, J.M. (2014). Behavioral Genetics,
6th ed. New York: Worth Publishers.
Polderman, T.J., Benyamin, B., De Leeuw, C.A., Sullivan, P.F., van Bochoven, A.,
Visscher, P.M. & Posthuma, D. (2015). Meta-analysis of the heritability of human traits
based on fifty years of twin studies. Nature Genetics 47: 702-709.
Rabinowitz, J.A., Kuo, S.I.C., Felder, W., Musci, R.J., Bettencourt, A., Benke, K., ... &
Kouzis, A. (2019). Associations between an educational attainment polygenic score with
educational attainment in an African American sample. Genes, Brain and Behavior
18(5): e12558.
Racimo, F., Berg, J.J. & Pickrell, J.K. (2018). Detecting polygenic adaptation in
admixture graphs. Genetics 208: 1565-1584.
Rosenberg, N.A., Edge, M.D., Pritchard, J.K. & Feldman, M.W. (2019). Interpreting
polygenic scores, polygenic adaptation, and human phenotypic differences. Evolution,
Medicine, and Public Health 2019: 26-34.
Rosseel, Y. (2012). lavaan: An R package for structural equation modeling. Journal of
Statistical Software 48(2): 1-36.
Roth, P.L., van Iddekinge, C.H., DeOrtentiis, P.S., Hackney, K.J., Zhang, L. & Buster,
M.A. (2017). Hispanic and Asian performance on selection procedures: A narrative and
meta-analytic review of 12 common predictors. Journal of Applied Psychology 102:
Satterthwaite, T.D., Connolly, J.J., Ruparel, K., Calkins, M.E., Jackson, C., Elliott, M.A.,
... & Qiu, H. (2016). The Philadelphia Neurodevelopmental Cohort: A publicly available
resource for the study of normal and abnormal brain development in youth. Neuroimage
124: 1115-1119.
Sirin, S.R. (2005). The relationship between socioeconomic status and school
outcomes: Meta-analytic review of research. Review of Educational Research 75: 417-
Sniekers, S., Stringer, S., Watanabe, K., Jansen, P.R., Coleman, J.R., Krapohl, E., ... &
Amin, N. (2017). Genome-wide association meta-analysis of 78,308 individuals
identifies new loci and genes influencing human intelligence. Nature Genetics 49: 1107-
Sohail, M., Maier, R.M., Ganna, A., Bloemendal, A., Martin, A.R., Turchin, M.C., ... &
Sunyaev, S.R. (2019). Polygenic adaptation on height is overestimated due to
uncorrected stratification in genome-wide association studies. eLife 8: e39702.
Spencer, A.V., Cox, A. & Walters, K. (2014). Comparing the efficacy of SNP filtering
methods for identifying a single causal SNP in a known association region. Annals of
Human Genetics 78: 50-61.
te Nijenhuis, J., Jongeneel-Grimen, B. & Armstrong, E.L. (2015). Are adoption gains on
the g factor? A meta-analysis. Personality and Individual Differences 73: 56-60.
te Nijenhuis, J., van den Hoek, M. & Dragt, J. (2019). A meta-analysis of Spearman’s
Hypothesis tested on Latin-American Hispanics, including a new way to correct for
imperfectly measuring the construct of g. Psych 1(1): 101-122.
te Nijenhuis, J., Choi, Y.Y., van den Hoek, M., Valueva, E. & Lee, K.H. (2019).
Spearman’s hypothesis tested comparing Korean young adults with various other
groups of young adults on the items of the Advanced Progressive Matrices. Journal of
Biosocial Science 51: 875-912.
Thompson, J. (2019, May 2). Piffer kicks against the pricks. Available online:
Trzaskowski, M., Harlaar, N., Arden, R., Krapohl, E., Rimfeld, K., McMillan, A., ... &
Plomin, R. (2014). Genetic influence on family socioeconomic status and children's
intelligence. Intelligence 42: 83-88.
US Census Bureau (2016). American factfinder Results. US Department of
Commerce, Economics and Statistics Administration. Retrieved from:
Wang, F.L., Pandika, D., Chassin, L., Lee, M. & King, K. (2016). Testing the relations
among family disorganization, delay discounting, and adolescent alcohol use: A
genetically informed study. Alcoholism: Clinical and Experimental Research 40: 846-
Wang, X. & Teo, Y.Y. (2015). Trans-ethnic fine-mapping of rare causal variants. In:
E. Zeggini & A. Morris (eds.), Assessing Rare Variation in Complex Traits, pp. 253-261.
New York: Springer.
Warne, R.T. (2020). Continental genetic ancestry source correlates with global cognitive
ability score. Mankind Quarterly 60: 400-422.
Woodley of Menie, M.A.W., Fernandes, H.B. & Hopkins, W.D. (2015). The more g-
loaded, the more heritable, evolvable, and phenotypically variable: Homology with
humans in chimpanzee cognitive abilities. Intelligence 50: 159-163.
Ylikoski, P. (2013). Causal and constitutive explanation compared. Erkenntnis 78(2):
Zanetti, D. & Weale, M.E. (2018). Transethnic differences in GWAS signals: A
simulation study. Annals of Human Genetics 82(5): 280-286.
... These scores were based on cognitive ability (n = 257,841), hardest math class taken (n = 430,445), and mathematical ability (n = 564,698) (Lee et al., 2018). We use these PGS because previous research has shown them to have trans-ethnic predictive validity in European, Hispanic, and African American populations (Fuerst, Kirkegaard & Piffer, 2021;Lasker et al., 2019). Moreover, common forms of bias were found not to account for the ancestry-related eduPGS differences (Fuerst et al., 2021). ...
... We use these PGS because previous research has shown them to have trans-ethnic predictive validity in European, Hispanic, and African American populations (Fuerst, Kirkegaard & Piffer, 2021;Lasker et al., 2019). Moreover, common forms of bias were found not to account for the ancestry-related eduPGS differences (Fuerst et al., 2021). Thus, we can say that this PGS plausibly captures genetic effects between ancestry groups. ...
... The results are summarized in Table 6. As previously found, the eduPGS by g associations are attenuated among African Americans, but not among Hispanic and Other Americans (Fuerst et al., 2021). Nonetheless, eduPGS are significantly associated with g within all SIRE groups. ...
Full-text available
Black and Hispanic children in the United States have lower mean cognitive test scores than White children. The reasons for this are contested. The test score gap may be caused by socio-cultural factors, but the high heritability of g suggests that genetic variance might play a role. Differences between self-identified race or ethnicity (SIRE) groups could be the product of ancestral genetic differences. This genetic hypothesis predicts that genetic ancestry will predict g within these admixed groups. To investigate this hypothesis, we performed admixture-regression analyses with data from the Adolescent Brain Cognitive Development Cohort. Consistent with predictions from the genetic hypothesis, African and Amerindian ancestry were both found to be negatively associated with g. The association was robust to controls for multiple cultural, socioeconomic, and phenotypic factors. In the models with all controls the effects were as follows: (a) Blacks, African ancestry: b =-0.89, N = 1690; (b) Hispanics, African ancestry: b =-0.58, Amerindian ancestry: b =-0.86, N = 2021), and (c) a largely African-European mixed Other group, African ancestry: b =-1.08, N = 748). These coefficients indicate how many standard deviations g is predicted to change when an individual's African or Amerindian ancestry proportion changes from 0% to 100%. Genetic ancestry statistically explained the self-identified race and ethnicity (SIRE) differences found in the full sample. Lastly, within all samples, the relation between genetic ancestry and g was partially accounted for by cognitive ability and educational polygenic scores (eduPGS). These eduPGS were found to be significantly predictive of g within all SIRE groups, even when controlling for ancestry. The results are supportive of the genetic model.
... These eduPGS, denoted MTAG eduPGS, are used because they were previously validated in European, Hispanic, and African-American populations (Lasker et al., 2019) and also assessed for population-structure-related confounding (Fuerst, Kirkegaard, & Piffer, 2021). ...
Full-text available
Intelligence tests are excellent predictors of school and job performance and racial/ethnic differences in mean IQ are common. Based on five lines of evidence, Warne (2021) builds a case for partly genetic causes of differences in general intelligence (g) across American Self/Parental-identified race or ethnicity. Based on a careful reading of Warne (2021) and the authors he cites, we generated 15 predictions flowing from a partial genetic hypothesis. These predictions concern (1) mean differences, (2) measurement invariance, (3) high within-group heritability, (4) admixture regression for g, (5) polygenic scores, (6) brain volume, and (7) Spearman’s hypothesis. We used the Adolescent Brain Cognitive Development Study sample (N = 10,245) to test these hypotheses using classical and state-of-the-art statistical techniques. Decomposition of variance using twins showed that the heritability of intelligence and of brain/intracranial volume estimates were, respectively, moderate and high for the White and the non-White subsamples. Within all SIRE groups, both genetic ancestry and education-related polygenic scores (eduPGS) predicted both brain volume and g. Moreover, brain volume was weakly but significantly related to g (r = .14 to .25). Path and causal mediation analysis showed that total brain volume explained approximately 15% of the association between European ancestry and g and also explained approximately 8% of that between eduPGS and g. Finally, based on the Method of Correlated Vectors (MCV), a positive manifold was found for genetic, brain volume, and ancestry/SIRE-related variables. We conclude that the results support the hypotheses tested and are in line with a partial genetic hypothesis.
... (Because the sum of ancestry proportions must equal one, the negative-African/positive-European linear estimates are cognate findings.) Lasker et al. (2019) use data from the Philadelphia Neurodevelopment Cohort, n = 7,321; Kirkegaard et al. (2019) use data from the Pediatric Imaging, Neurocognition and Genetics Study, n = 1,369; Warne (2019) uses a convenience sample recruited through Qualtrics, n = 193; Fuerst et al. (2021a) use data from the Philadelphia Neurodevelopment Cohort, n = 7,910; Connor at al. (2021), n = 9,972, andFuerst et al. (2021b), n = 10,370 both use data from the Adolescent Brain and Cognitive Development (ABCD) Study. All six studies find essentially the same negative relationship. ...
Full-text available
Admixture regression involves the regression of a medical or behavioral trait on sub-population admixture proportions, race and ethnic group identifiers, and other explanatory variables, to measure the environmental and genetic components to trait variation across self-identified racial and ethnic groups. Several recent admixture regression studies find that proportion of African genetic ancestry has strong explanatory power for cognitive ability test scores. This paper applies stratified sub-sampling to explore the possibility that potential model mis-specification and/or variable mis-measurement underlies this empirical finding. The estimated coefficient on African ancestry is not significantly changed when the sample is stratified by low socioeconomic status/high socioeconomic status, Hispanic/non-Hispanic, Black-only/Black-White biracial, or Black/White racial group self-identification.
... They demonstrated measurement invariance of the cognitive tests, confirming that the tests measure the same abilities in different groups, and they assessed the extent to which their education polygenic score predicts the g factor of general intelligence rather than other abilities or personality traits. They even tried different polygenic scores, based either on population GWAS or within-family GWAS, to account for possible population stratification in the discovery samples (Fuerst et al., 2021a). ...
Full-text available
In the Fall 2021 issue of Mankind Quarterly (62:1), two papers were published about cognitive ability in genetically admixed African American and Hispanic populations in the United States. The main conclusion of the authors is that in the two cohorts they studied, cognitive ability is substantially influenced by genetic admixture proportions, even after taking account of plausible social and environmental factors such as socioeconomic family background and self-reported discrimination. This commentary critically examines the findings and places this research into a broader context. First, the evolutionary theories under investigation in these two studies are stated and critiqued, and prior probabilities-the probability of different working hypotheses to be true before results are obtained-are estimated; second, the authors' methods and conclusions are assessed; third, avenues for future research in the field are sketched; and fourth, the historical and social context is assessed with a view on the ability or inability of modern societies to profit from the results of this research. The conclusion is that the research appears fundamentally sound, its results can guide efforts to reduce socioeconomic inequalities in the world, but that present-day human societies appear to be incapable of doing so.
Full-text available
Frost (2021) and Meisenberg (2021) provide thoughtful commentaries on the two papers, Fuerst, Kirkegaard and Piffer (2021a) and Fuerst, Hu and Connor (2021b). This brief reply discusses a few key points.
Full-text available
Intelligence is a highly polygenic trait and genome-wide association studies (GWAS) have identified thousands of DNA variants contributing with small effects. Polygenic scores (PGS) can aggregate those effects for trait prediction in independent samples. As large-scale light-phenotyping GWAS operationalized intelligence as performance in rather superficial tests, the question arises which intelligence facets are actually captured. We used deep-phenotyping to investigate the molecular determinants of individual differences in cognitive ability. We, therefore, studied the association between PGS of intelligence (IQ-PGS), cognitive performance (CP-PGS), and educational attainment (EA-PGS) with a wide range of intelligence facets in a sample of 557 healthy adults. IQ-PGS, CP-PGS, and EA-PGS had the highest incremental R²s for general (2.71%; 4.27%; 2.06%), verbal (3.30%; 4.64%; 1.61%), and numerical intelligence (3.06%; 3.24%; 1.26%) and the weakest for non-verbal intelligence (0.89%; 1.47%; 0.70%) and memory (0.80%; 1.06%; 0.67%). These results indicate that PGS derived from light-phenotyping GWAS do not reflect different facets of intelligence equally well, and thus should not be interpreted as genetic indicators of intelligence per se. The findings refine our understanding of how PGS are related to other traits or life outcomes.
Full-text available
It has been known since 1904 that, in humans, diverse cognitive traits are positively intercorrelated. This forms the basis for the general factor of intelligence (g). Here, we directly test whether there is a partial genetic basis for individual differences in g using data from seven different cognitive tests (n = 11,263–331,679) and genome-wide autosomal single-nucleotide polymorphisms. A genetic g factor accounts for an average of 58.4% (s.e. = 4.8%) of the genetic variance in the cognitive traits considered, with the proportion varying widely across traits (range, 9–95%). We distil genetic loci that are broadly relevant for many cognitive traits (g) from loci associated specifically with individual cognitive traits. These results contribute to elucidating the aetiology of a long-known yet poorly understood phenomenon, revealing a fundamental dimension of genetic sharing across diverse cognitive traits.
Full-text available
There has been widespread adoption of genome wide summary scores (polygenic scores) as tools for studying the importance of genetics and associated life course mechanisms across a range of demographic and socioeconomic outcomes. However, an often unacknowledged issue with these studies is that parental genetics impact both child environments and child genetics, leaving the effects of polygenic scores difficult to interpret. This paper uses multi-generational data containing polygenic scores for parents (n = 7193) and educational outcomes for adopted (n = 855) and biological (n = 20,939) children, many raised in the same families, which allows us to separate the influence of parental polygenic scores on children outcomes between environmental (adopted children) and environmental and genetic (biological children) effects. Our results complement recent work on “genetic nurture” by showing associations of parental polygenic scores with adopted children’s schooling, providing additional evidence that polygenic scores combine genetic and environmental influences and that research designs are needed to separate these estimated impacts.
Full-text available
Background There is converging evidence that mental disorders are more optimally conceptualized in a hierarchical framework (i.e., the Hierarchical Taxonomy of Psychopathology, HiTOP) that transcends the categorical boundaries of the Diagnostic and Statistical Manual of Mental Disorders (DSM). However, the majority of this evidence comes from studies that draw upon predominantly European American or Caucasian populations. Whether a hierarchical conceptualization of mental disorders generalizes across racial‐ethnic groups, including for African American (AA) populations, is unclear. Methods We tested multidimensional and bifactor models of 15 DSM diagnoses and psychiatric traits in two groups, including AA (n = 3,088) and European American (EA; n = 5,147) youths aged 8–21 from the Philadelphia Neurodevelopmental Cohort (PNC). We also conducted multigroup confirmatory factor analyses to test for factorial invariance between the best fitting AA and EA multidimensional and bifactor models. Results In the multidimensional model tests, a three‐factor model, specifying internalizing, externalizing, and thought dimensions, emerged as the best fitting model for AAs and EAs. In the bifactor model tests, a three‐factor model (i.e., internalizing, externalizing, and thought dimensions) that also specified a general factor emerged as the optimal for both AAs and EAs. The general factor accounted for a significant proportion of the covariation between the secondary factors and the individual disorders and traits. Furthermore, both models were factorially invariant, indicating no significant difference in the factor structure of mental disorders between AAs and EAs in PNC. Conclusions Results suggest that the hierarchical factor structure of mental disorders may be racial‐ethnically robust. This finding has implications for etiological and epidemiological studies focused on racial‐ethnic subgroup comparisons, particularly with respect to identifying similarities and differences in prevalence rates or sociodemographic risk factors for mental disorders.
Full-text available
Via meta-analysis, we examined whether the heritability of intelligence varies across racial or ethnic groups. Specifically, we tested a hypothesis predicting an interaction whereby those racial and ethnic groups living in relatively disadvantaged environments display lower heritability and higher environmentality. The reasoning behind this prediction is that people (or groups of people) raised in poor environments may not be able to realize their full genetic potentials. Our sample (k = 16) comprised 84,897 Whites, 37,160 Blacks, and 17,678 Hispanics residing in the United States. We found that White, Black, and Hispanic heritabilities were consistently moderate to high, and that these heritabilities did not differ across groups. At least in the United States, Race/ Ethnicity × Heritability interactions likely do not exist.
Full-text available
Using data from the Philadelphia Neurodevelopmental Cohort, we examined whether European ancestry predicted cognitive ability over and above both parental socioeconomic status (SES) and measures of eye, hair, and skin color. First, using multi-group confirmatory factor analysis, we verified that strict factorial invariance held between self-identified African and European-Americans. The differences between these groups, which were equivalent to 14.72 IQ points, were primarily (75.59%) due to difference in general cognitive ability (g), consistent with Spearman’s hypothesis. We found a relationship between European admixture and g. This relationship existed in samples of (a) self-identified monoracial African-Americans (B = 0.78, n = 2,179), (b) monoracial African and biracial African-European-Americans, with controls added for self-identified biracial status (B = 0.85, n = 2407), and (c) combined European, African-European, and African-American participants, with controls for self-identified race/ethnicity (B = 0.75, N = 7,273). Controlling for parental SES modestly attenuated these relationships whereas controlling for measures of skin, hair, and eye color did not. Next, we validated four sets of polygenic scores for educational attainment (eduPGS). MTAG, the multi-trait analysis of genome-wide association study (GWAS) eduPGS (based on 8442 overlapping variants) predicted g in both the monoracial African-American (r = 0.111, n = 2179, p < 0.001), and the European-American (r = 0.227, n = 4914, p < 0.001) subsamples. We also found large race differences for the means of eduPGS (d = 1.89). Using the ancestry-adjusted association between MTAG eduPGS and g from the monoracial African-American sample as an estimate of the transracially unbiased validity of eduPGS (B = 0.124), the results suggest that as much as 20%–25% of the race difference in g can be natively explained by known cognitive ability-related variants. Moreover, path analysis showed that the eduPGS substantially mediated associations between cognitive ability and European ancestry in the African-American sample. Subtest differences, together with the effects of both ancestry and eduPGS, had near-identity with subtest g-loadings. This finding confirmed a Jensen effect acting on ancestry-related differences. Finally, we confirmed measurement invariance along the full range of European ancestry in the combined sample using local structural equation modeling. Results converge on genetics as a partial explanation for group mean differences in intelligence.
Full-text available
A historical tendency to use European ancestry samples hinders medical genetics research, including the use of polygenic scores, which are individual-level metrics of genetic risk. We analyze the first decade of polygenic scoring studies (2008-2017, inclusive), and find that 67% of studies included exclusively European ancestry participants and another 19% included only East Asian ancestry participants. Only 3.8% of studies were among cohorts of African, Hispanic, or Indigenous peoples. We find that predictive performance of European ancestry-derived polygenic scores is lower in non-European ancestry samples (e.g. African ancestry samples: t = -5.97, df = 24, p = 3.7 × 10-6), and we demonstrate the effects of methodological choices in polygenic score distributions for worldwide populations. These findings highlight the need for improved treatment of linkage disequilibrium and variant frequencies when applying polygenic scoring to cohorts of non-European ancestry, and bolster the rationale for large-scale GWAS in diverse human populations.
Full-text available
Little research has dealt with intragroup ancestry-related differences in intelligence in Black and White Americans. To help fill this gap, we examined the association between intelligence and both color and parent-reported ancestry using the NLSY97. We used a nationally-representative sample, a multidimensional measure of cognitive ability, and a sibling design. We found that African ancestry was negatively correlated with general mental ability scores among Whites (r = −0.038, N = 3603; corrected for attenuation, rc = −0.245). In contrast, the correlation between ability and parent-reported European ancestry was positive among Blacks (r = 0.137, N = 1788; rc = 0.344). Among Blacks, the correlation with darker skin color, an index of African ancestry, was negative (r = −0.112, N = 1455). These results remained with conspicuous controls. Among Blacks, both color and parent-reported European ancestry had independent effects on general cognitive ability (color: β = −0.104; ancestry: β = 0.118; N = 1445). These associations were more pronounced on g-loaded subtests, indicating a Jensen Effect for both color and ancestry (rs = 0.679 to 0.850). When we decomposed the color results for the African ancestry sample between and within families, we found an association between families, between singletons (β = −0.153; N = 814), and between full sibling pairs (β = −0.176; N = 225). However, we found no association between full siblings (β = 0.027; N = 225). Differential regression to the mean results indicated that the factors causing the mean group difference acted across the cognitive spectrum, with high-scoring African Americans no less affected than low-scoring ones. We tested for measurement invariance and found that strict factorial invariance was tenable. We then found that the weak version of Spearman’s hypothesis was tenable while the strong and contra versions were not. The results imply that the observed cognitive differences are primarily due to differences in g and that the Black-White mean difference is attributable to the same factors that cause differences within both groups. Further examination revealed comparable intraclass correlations and absolute differences for Black and White full siblings. This implied that the non-shared environmental variance components were similar in magnitude for both Blacks and Whites.
Large-scale multi-ethnic cohorts offer unprecedented opportunities to elucidate the genetic factors influencing complex traits related to health and disease among minority populations. At the same time, the genetic diversity in these cohorts presents new challenges for analysis and interpretation. We consider the utility of race and/or ethnicity categories in genome-wide association studies (GWASs) of multi-ethnic cohorts. We demonstrate that race/ethnicity information enhances the ability to understand population-specific genetic architecture. To address the practical issue that self-identified racial/ethnic information may be incomplete, we propose a machine learning algorithm that produces a surrogate variable, termed HARE. We use height as a model trait to demonstrate the utility of HARE and ethnicity-specific GWASs.