Schooling substantially improves intelligence, but neither lessens nor widens the impacts of socioeconomics and genetics

The study addresses how schooling, socioeconomic status (SES), and genetic predispositions (captured via a cognitive polygenic score, cogPGS) uniquely and jointly contribute to individual differences in intelligence in late childhood. Prior research shows intelligence is both highly heritable and environmentally influenced, with schooling known to improve cognitive performance. However, it remains unclear whether schooling’s effects are independent from SES and genetic factors, and whether schooling amplifies existing differences (rich-get-richer) or mitigates them (catch-up). The authors focus on crystallized intelligence (cIQ), fluid intelligence (fIQ), and working memory (WM), key domains for educational outcomes. They specifically test whether the effects of schooling or SES are moderated by cogPGS and whether a three-way interaction among schooling, SES, and cogPGS exists, thereby probing gene-by-environment interplay relevant to educational contexts.

Multiple lines of evidence indicate schooling positively affects intelligence, including longitudinal studies controlling for prior ability, natural experiments around compulsory schooling changes, and regression discontinuity designs leveraging school entry cutoffs. Estimates suggest one additional year of schooling yields ~1–5 IQ points (0.07–0.3 SD). SES, often measured via parental education, household income, and neighborhood quality, is linked to initial differences in intelligence and may widen gaps over development, though SES also correlates with genetic factors passed from parents (gene–environment correlation). Polygenic scores for cognitive performance predict 7–10% of variance and correlate moderately with SES, but their unique contributions and interactions with schooling for specific cognitive domains are less understood. Prior gene-by-environment studies, often focusing on SES and educational achievement/attainment rather than intelligence per se, have yielded mixed interaction findings (positive, negative, null). The literature highlights theoretical models (multiplier, transactional, bioecological) wherein genetic propensities shape environments that further influence cognition, but empirical results vary across contexts and outcomes.

Design and sample: The study used data from the Adolescent Brain Cognitive Development (ABCD) Study, including 6567 children in grades 3–5 (ages 8.92–11.00; mean 9.88). One child per family was retained to avoid clustering by family; children who repeated a grade were excluded. Recruitment occurred year-round, enabling measurement of schooling in months. A quasi-experimental fuzzy regression discontinuity design exploited age-based grade assignment cutoffs to separate schooling from chronological age effects. Mixed-effects models with random intercepts for collection site were used. Measures: Cognitive outcomes were from the NIH Toolbox: crystallized intelligence (cIQ: picture vocabulary, oral reading), fluid intelligence (fIQ: pattern comparison, list-sorting WM, picture sequence memory, flanker, dimensional change card sort), and working memory (WM: list-sorting WM). Scores were standardized (mean 0, SD 1) after outlier handling. SES was the first component from a probabilistic PCA on total household income, highest parental education, and neighborhood quality (Area Deprivation Index), capturing 65% of variance. Subcomponents were also analyzed individually. Genetics: Saliva genotyping used the Smokescreen array. QC included call rate thresholds, Hardy–Weinberg equilibrium, MAF filters, heterozygosity, and missingness checks; imputation used 1000 Genomes Phase 3 with IMPUTE4, pre-phasing via SHAPEIT2. Population structure was modeled using 20 genetic principal components (PCs). A multi-trait cognitive polygenic score (cogPGS) was constructed with PRSice-2 from MTAG summary statistics (educational attainment, cognitive performance, highest math class, self-reported math ability); clumping (250 kb, r2>0.25) and p<5e−5 threshold yielded 5255 SNPs; scores were standardized. Statistical modeling: Primary analyses used linear mixed-effects models (lme4) with outcomes Y ∈ {cIQ, fIQ, WM}. Equation 1 estimated age and schooling effects. Equation 2 added SES, cogPGS, sex, and 20 ancestry PCs to estimate unique effects while controlling for stratification. Equations 3 and 4 tested two-way interactions (schooling×cogPGS, schooling×SES, cogPGS×SES) and a three-way interaction (schooling×cogPGS×SES), including age interaction terms to avoid spurious findings. P-values used Satterthwaite’s method; FDR was applied where appropriate. Post hoc analyses included: (a) SES subcomponents; (b) sibling analysis (392 families, n=792) estimating within-family (βw) and between-family (βb) cogPGS effects; (c) a general factor (g) from confirmatory factor analysis and subsequent modeling; (d) a European ancestry subset analysis (4-means clustering on PCs; n=3751) to examine cogPGS accuracy; and (e) Bayesian hierarchical mixed-effects modeling (brms) with weakly informative priors, 95% HDIs, and ROPE bounds (0.05 and 0.02 SD) for interaction terms to assess practical null effects. Sensitivity checks compared grades and tested schooling coefficient differences via conservative z-tests.

- Independent main effects: Schooling, SES, and cogPGS each had significant, independent positive effects on cognition (all p<0.001). Estimated standardized schooling coefficients: cIQ β=0.13, fIQ β=0.10, WM β=0.09. One year (10 months) of schooling corresponded to gains of 0.22 SD (cIQ), 0.14 SD (fIQ), and 0.14 SD (WM). The schooling-to-age effect ratio was 1.1 for cIQ, 0.54 for fIQ, and 2.2 for WM. - SES and cogPGS effects: SES showed larger effects than cogPGS overall, with the largest effects for cIQ (SES β≤0.29; cogPGS β≤0.16). For fIQ and WM, cogPGS β≈0.09; SES β≈0.18 (fIQ) and β≈0.22 (WM). Sex effects: females higher on fIQ (β=0.09) and lower on cIQ (β=−0.05) and WM (β=−0.08). - Magnitude comparison: Two years of schooling (grades 3–5) produced larger differences in intelligence than the lifetime SES or cogPGS inequalities. - SES subcomponents: All three components were significant (FDR<0.001). For cIQ, effects were largest for parental education (β=0.26), then family income (β=0.22), then neighborhood quality (β=0.11). WM showed slightly larger effects than fIQ for each component (e.g., parental education WM β=0.19 vs fIQ β=0.17). - Sibling analysis: Within-family cogPGS effects were smaller than between-family for cIQ (βw=0.08, pFDR=0.009; βb=0.15, pFDR=0.009), suggesting passive genotype–environment correlation. For fIQ, βw=0.08 (pFDR=0.013), βb not significant (βb=0.10, pFDR=0.074). For WM neither βw nor βb was significant. - Gene-by-environment interplay: No significant two-way (schooling×cogPGS, schooling×SES, cogPGS×SES) or three-way (schooling×cogPGS×SES) interactions across cIQ, fIQ, or WM after correction. Bayesian ROPE analyses indicated most interaction posteriors fell entirely within a 0.05 SD ROPE, supporting practical null effects at that boundary; overlap with a 0.02 SD ROPE precluded ruling out very small effects. - General factor (g): Schooling effect on g was 0.136 SD (pFDR<0.001); no significant interactions. - European ancestry subset: Results were generally consistent; cogPGS effects slightly increased (Δβ≈0.005–0.020) and SES effects decreased (Δβ≈−0.046 to −0.080), with interaction terms remaining non-significant.

The findings show that schooling exerts a substantial, independent, and domain-general positive effect on children’s intelligence, comparable to or exceeding the influence of chronological age in this period, particularly for working memory. SES and cogPGS independently predict cognitive performance, with SES contributing more strongly than cogPGS, especially to crystallized abilities. Critically, the absence of significant interactions indicates that, within grades 3–5, schooling neither amplifies nor compensates for preexisting SES- or genetics-based differences in intelligence; schooling raises overall levels without changing rank order. Sibling comparisons support the presence of passive gene–environment correlations affecting between-family estimates, highlighting the intertwined nature of SES and genetic influences. Bayesian analyses suggest any moderation by schooling or SES on genetic effects is likely very small (<0.05 SD), though tiny schooling interactions (<0.02 SD) cannot be excluded and could accumulate over years. Overall, the study addresses the central question by demonstrating unique, additive contributions of schooling, SES, and genetics to cognitive performance, and little evidence for gene-by-schooling or SES-by-genetics moderation within this developmental window.

This study demonstrates that two years of schooling (grades 3–5) produce larger gains in intelligence than the lifetime disparities associated with SES or a cognitive polygenic score, while leaving the relative ordering of children largely unchanged. SES and cogPGS each have sizable, independent associations with cognition, but schooling does not significantly moderate their effects. These results underscore schooling’s robust causal impact on intelligence and suggest that, at least in late childhood, school exposure increases cognitive levels without widening or narrowing SES- or genetics-related gaps. Future research should: (1) examine broader schooling spans (earlier and later grades) to test cumulative interaction effects; (2) investigate diverse ancestries with improved polygenic prediction to enhance generalizability; (3) incorporate richer environmental measures (instructional quality, curriculum, classroom factors); (4) utilize longitudinal within-family and quasi-experimental designs with greater power; and (5) expand multi-domain cognitive assessments (e.g., multiple WM tasks) to refine domain-specific inferences.

Key limitations include: (1) polygenic scores were derived largely from European ancestry GWAS, limiting generalizability and predictive accuracy in non-European groups; (2) GWAS and PGS approaches are optimized for additive effects and may miss certain G×E forms; (3) interpretation of multi-trait GWAS-derived PGS can be influenced by auxiliary phenotypes and sample size differences; (4) controlling for genetic PCs necessitated excluding participants lacking DNA, raising average SES and potentially affecting representativeness; (5) the schooling window was limited to grades 3–5, possibly missing cumulative or developmental phase-specific interactions; (6) working memory was assessed with a single task; (7) sibling analyses had limited power (392 families), especially for fIQ and WM; and (8) the sample had slightly lower SES than the U.S. average, so interactions at the lower SES tail cannot be ruled out.