A population-wide gene-environment interaction study on how genes, schools, and residential areas shape achievement

The study asks how individual genetic propensity for educational attainment (EA-PGI) interacts with multiple levels of social context to shape school achievement. The authors note that child development theories (e.g., bioecological model) posit multilevel environments (family, school, neighbourhood, broader society) that may interact with genetics, but prior empirical work has focused narrowly on family-level factors and has struggled to disentangle intercorrelated contexts and gene–environment correlation (rGE). The Scarr-Rowe interaction hypothesis suggests genetic effects on cognition/achievement are suppressed under disadvantage and amplified under advantage, but evidence is inconsistent across countries and methods. Genomic studies often find null GxE with measured family environments. The authors propose a comprehensive multilevel approach that is agnostic to specific measured features: estimate total interactions between EA-PGI and schools, neighbourhoods, districts, and municipalities using multilevel models, while controlling for passive rGE via parental EA-PGI. They hypothesize, per the bioecological model, that interactions may occur at several levels but diminish at more distal levels (municipality), and per Scarr-Rowe that genetic effects may be weaker in less advantaged environments. Norway provides a relatively egalitarian setting with local public school attendance, allowing a rigorous test of whether small average school/area effects conceal larger effects for certain children and which environments work best for whom.

Prior research on GxE in achievement and cognition has yielded mixed results. Twin studies in the U.S. often report stronger genetic influences in higher-SES families, whereas European/Australian studies report null or opposite patterns; a recent cross-national twin analysis found largely invariant heritability across SES, with some context-specific deviations. Genomic analyses using PGIs typically find minimal interactions with measured family environments (e.g., parental education/income, job loss, chaos at home). Some studies suggest interactions beyond the family: higher neighbourhood income, school quality, and teacher quality may strengthen genetic influences on reading/achievement, consistent with Scarr-Rowe in certain contexts. However, most studies examine single environmental measures, risking confounding from unmeasured, correlated contexts. Two prior studies used multilevel models to estimate total PGI-by-school interactions. A key challenge is rGE: parents select schools/neighbourhoods partly based on heritable traits, and children may select environments through performance. Failure to account for rGE can yield spurious GxE. The within-family genetic design (controlling for parental PGI) helps isolate offspring direct genetic effects from passive rGE and other confounds. This study addresses gaps by jointly modeling multiple environmental levels and using within-family EA-PGI.

Design and data: Linked genome-wide genotypes from the Norwegian Mother, Father, and Child Cohort Study (MoBa) to population administrative records for standardized national test scores (maths, reading at grades 5, 8, 9; English at grades 5, 8), and to identifiers for schools and residential areas (neighbourhoods/grunnkretser, districts/delområde, municipalities/kommune). Final analytic sample included 23,471 students from >23,000 genotyped parent–child trios with non-missing achievement, EA-PGI, school/residential identifiers, and parental variables (education, income, EA-PGI). Students attended 2,578 schools and lived in 7,700 neighbourhoods, 1,440 districts, and 408 municipalities (average cluster sizes: ~11 per school, 3 per neighbourhood, 16 per district, 57 per municipality). Norway context involves comprehensive public schooling with minimal private/ selective schooling. Measures: Achievement scores were residualized for sex, cohort (current age), and test age. A core achievement composite (mean across available subjects at each grade) was standardized (mean 0, SD 1). School and residential identifiers were harmonized as feasible to 2018 boundaries to align nearby areas (especially municipalities/districts). School-level sociodemographic covariates (aggregated from complete registers for all parents of students at each school) included: average parental education (years), average parental pre-tax earned income (ranked within cohort, averaged across ages 11–15), Gini coefficients for parental education and income, and proportion of non-Western immigrants/children of non-Western immigrants. Individual-level covariates were parental education and income. Genotyping and PGI: Post-imputation QC retained high-quality SNPs (imputation quality ≥0.8, call rate >98%, MAF >1%, HWE p<0.001, etc.). Individuals of European ancestries were retained based on PCA with 1000 Genomes reference; principal components (5 maternal, 5 paternal) were computed to adjust for ancestry. EA-PGI were constructed for parents and children using PRSice from GWAS summary statistics (Lee et al., 2018; excluding 23andMe and MoBa), using all SNPs (p≤1) with clumping (kb=500, r2=0.25). Mid-parental PGI (average of maternal and paternal) was computed and standardized; child PGI was standardized. Models included parental PGI and 10 ancestry PCs. Statistical modeling: Multilevel mixed-effects models were fitted using lme4 in R. Time (grade) was included as a fixed effect; individual child ID had a random intercept to pool repeated measures across grades. Model sequence (compared via AIC and likelihood tests): (1) base fixed-effects model of achievement on within-family child EA-PGI (controlling for mid-parent PGI), parental education/income, grade, and ancestry PCs; (2a–d) added random intercepts for contexts: school; +neighbourhood; +district; +municipality (cross-classifying schools with nested residential clusters); (3a–d) added random slopes for child EA-PGI at each context with significant intercept variance to test PGI-by-context interactions; (4) added five school sociodemographic covariates (fixed effects) while retaining child EA-PGI random slopes across schools; (5) added interactions between each school covariate and child EA-PGI to test whether measured covariates explain slope variability. Subject-specific analyses re-estimated best-fitting interaction model for maths, reading, and English separately. rGE control and diagnostics: Including parental EA-PGI ensures child EA-PGI effects represent within-family direct genetic effects, reducing passive rGE. Intraclass correlations showed child EA-PGI clustered 2.6% by school and ≤1.2% by residential areas overall, but clustering dropped to 0% for within-family child EA-PGI (conditional on parental PGI), supporting quasi-random sorting of within-family genetic differences across schools.

• Best-fitting model included random intercepts and random slopes for child EA-PGI across schools, but only random intercepts (no slopes) for residential levels (neighbourhood, district, municipality). Thus, EA-PGI effects vary by school but not by residential area.
• Residential main effects on achievement were small: variance explained ~1% for municipalities, ~1% for neighbourhoods, and <1% for districts; collectively <2%.
• School-specific EA-PGI effects: average slope of EA-PGI on achievement = 0.22 SD per 1 SD increase in EA-PGI, with between-school SD of slopes = 0.034. In the 2.5% of schools with weakest slopes, EA-PGI effects were <0.15 SD; in the 2.5% with strongest slopes, >0.29 SD. Corresponding variance explained by EA-PGI ranged from ~2% (weakest) to ~8% (strongest), a >4-fold difference.
• Negative slope–intercept correlation across schools: genetic effects (EA-PGI slopes) were weaker in higher-performing schools, indicating a compensatory pattern where high-performing schools reduce the impact of low EA-PGI.
• School importance varies by student EA-PGI: schools explained ~4% of achievement variance among students 2 SD below mean EA-PGI versus ~2% among students 2 SD above mean EA-PGI.
• Subject-specific interactions: between-school SD of EA-PGI slopes was 0.035 for maths, 0.027 for reading, and 0.004 for English, indicating stronger interactions for maths and reading than for English.
• Measured school sociodemographics (average parental education, average parental income, Gini of education, Gini of income, proportion non-Western immigrants) neither reduced the between-school variance in EA-PGI slopes nor improved fit when interacted with EA-PGI, implying they do not explain the latent PGI-by-school interaction.
• rGE control: within-family child EA-PGI showed no school-level clustering after adjusting for parental PGI, supporting causal interpretation of school-specific slopes for within-family genetic effects.

Findings demonstrate that in Norway, the effects of students’ genetic propensity for educational attainment and school contexts are interdependent: EA-PGI effects on achievement vary by school, with weaker genetic effects in higher-performing schools. This compensatory pattern suggests that stronger schools raise overall achievement without leaving behind students with lower EA-PGI, and that school differences matter more for students with lower genetic propensity. The interaction was strongest for mathematics (and reading), aligning with U.S. evidence that higher-status schools can buffer students with lower EA-PGI from adverse academic trajectories in math. The results counter the traditional Scarr-Rowe expectation in this context, showing that genetic effects are relatively stronger in less advantaged schools and attenuated in higher-performing ones—potentially reflecting Norway’s equitable, high-quality schooling that enables compensation. Importantly, residential environments contributed little to achievement and did not interact with EA-PGI, consistent with the idea that academic skills are directly targeted within schools rather than shaped by broader residential contexts. The lack of explanatory power of measured school sociodemographics indicates that latent, unmeasured school features (instructional practices, resources, teacher effectiveness, support systems) likely underpin the interactions, justifying deeper investigation. Policy implications include focusing on equalizing support across schools for students with lower EA-PGI and recognizing that average school effect estimates can mask larger, heterogeneous effects for specific student groups.

This population-wide, family-based genomic and multilevel study shows that schools—but not neighbourhoods, districts, or municipalities—moderate the relationship between children’s EA-PGI and achievement. Higher-performing schools compensate for lower EA-PGI, reducing genetic disparities in outcomes, and school effects are more consequential for students with lower EA-PGI. Measured school sociodemographics did not explain the interaction, highlighting latent school factors as key targets for future research. The study advances a social-genetic framework for understanding which environments work for whom and suggests policy should prioritize reducing inter-school differences that disproportionately affect students with lower genetic propensity for educational attainment. Future work should identify specific school characteristics that drive compensation, test within-school moderators (teachers, peers), examine family investments as moderators, extend analyses to other countries and more diverse ancestries, and explore alternative PGIs (e.g., for environmental sensitivity) within multilevel GxE models.

• Ancestry restriction: analyses included only participants of European ancestries, limiting generalizability.
• Cohort participation: MoBa participation is not random; despite near-complete administrative coverage for outcomes and contexts, selection into the cohort could bias results.
• EA-PGI incompleteness: current PGIs capture only a fraction of genetic influences on education; estimates do not reflect the total magnitude of genetic interactions with schools.
• Context-insensitive PGI weights: EA-PGI are derived from GWAS pooled across contexts and may underrepresent context-sensitive heritability relevant to school/area differences.
• Residual gene–environment correlation: while passive rGE is controlled via parental PGI, children's own genetics could still influence school attendance; however, within-family EA-PGI showed no school-level clustering and Norway lacks selective elementary/middle schools, mitigating concern.
• Measurement of school/area covariates: available sociodemographic aggregates are broad and may miss specific instructional or organizational features driving interactions.