Gene-environment interaction analysis of school quality and educational inequality

The study examines whether and how the school environment alters genetic (A) and shared environmental (C) contributions to educational performance, thereby affecting educational inequality. Educational performance in primary school strongly shapes later educational trajectories and life outcomes, making sources of inequality important to identify. Family socioeconomic status (SES) and genetic differences are major contributors to performance variation, and theory suggests that their influence might depend on school characteristics. Competing perspectives predict that high-quality schools could either amplify social and genetic advantages (bioecological model; cumulative advantage) or compensate disadvantages (diathesis-stress; compensatory mechanisms), reducing inequality. The authors ask whether higher-quality schools increase or decrease genetic and family background influences on achievement and whether any moderation attributed to school quality is actually due to school SES composition and/or parental SES. Using a twin design, they estimate the extent to which genetic and environmental variances vary across schools of differing quality and SES to assess implications for educational inequality.

Two strands of theory offer opposing expectations. Behavioral genetics models: (a) the bioecological model (including the Scarr–Rowe hypothesis) predicts stronger genetic influence in more advantaged environments due to enhanced proximal processes that allow genetic potential to manifest; (b) diathesis-stress and compensation models predict reduced genetic influence in supportive contexts that buffer genetic risks for low performance. Sociological arguments similarly diverge: high-quality schools may amplify family background effects because high-SES children are better prepared and culturally aligned with academic norms, or they may compensate by providing resources and climates that particularly benefit low-SES students, reducing background-based disparities. Prior twin and PGI-based studies show mixed evidence of school moderation of genetic and shared environmental influences, with some PGI studies suggesting compensation in higher-SES or higher-achieving schools. However, school effects may be confounded with parental SES because high-SES families select into higher-quality, higher-SES schools. The specific school dimensions that matter (resources, climate/culture) remain unclear, and rigorous measures of school quality versus school composition are needed to parse mechanisms.

Design and data: The study uses linked Dutch administrative microdata from Statistics Netherlands (CBS) covering population cohorts, identifying 29,434 twin pairs (18,384 same-sex; 11,050 opposite-sex) born 1994–2007. Educational performance is measured by the national Cito end-of-primary-school standardized test (scale 501–550; mean ~535; SD ~10). Twins are restricted to those attending the same primary school; school identifiers are linked to Inspectorate of Education and Education Executive Agency data to construct school-level measures. Measures: School quality is an overall factor score derived from Inspectorate indicators (1999–2019), aggregated across available years due to irregular measurement, using factor analyses (FIML in Mplus). Exploratory factor analyses yielded nine dimensions (resources and climate-related): range of educational activities; curriculum (implementation); guidance of educational needs; parental involvement; monitoring/evaluating (special needs) students; learning climate; social climate; safety; quality assurance. A higher-order factor (overall school quality) was constructed; measurement error correction was imposed when incorporating the factor in analytical models. Parental SES is a latent factor from parents' education (ISCED 2011) and percentile personal incomes (year before Cito); missingness handled via FIML. School SES is the aggregation (mean) of parental SES among pupils taking the Cito in the twins' test year. All continuous covariates are z-standardized. Analytical approach: Classical twin ACE structural equation models decompose variance in educational performance into additive genetic (A), shared environmental (C), and non-shared environmental (E) components. Because zygosity is not observed, the genetic correlation among same-sex twins (r_SSG) is estimated via Weinberg’s differential rule and set in models to alternative plausible values: 0.70, 0.75, 0.80; opposite-sex twins are set to genetic correlation 0.50. Models control for sex and birth year and use robust standard errors clustered at the school level. Sequence of analyses: (1) Unmoderated ACE models estimate baseline A, C, E variance components across r_SSG specifications; (2) Models add main effects of school quality, school SES, and parental SES on mean performance; (3) ACE-moderation models (Purcell framework) test linear moderation of a, c, e path coefficients by school quality and by school SES separately (unstandardized and standardized variance components), then jointly include both moderators; (4) Finally, parental SES is added as a moderator to test whether school moderation reflects family vs school processes. Robustness checks: (a) Repeat moderation analyses under r_SSG = 0.75 and 0.80; (b) Assess whether SES-related differences in estimated MZ/DZ ratios among same-sex twins could spuriously generate gene-by-SES findings—evidence suggests if anything more DZ in low-SES families, implying the observed gene-by-SES is not driven by this; (c) Non-parametric multigroup moderation by quantiles of school quality and school SES to explore nonlinearity/thresholds; (d) Alternative operationalizations separating school resources and school culture dimensions, and tests on nine specific dimensions; (e) Consider test score censoring and distributional issues; both standardized and raw scores show decreasing variance with higher SES; prior work correcting for censoring found negligible impact.

Unmoderated ACE variance (Cito test variance V_ED = 95.34): Across assumptions on same-sex genetic relatedness, educational performance shows substantial genetic influence with modest to small shared environmental influence: r_SSG = 0.70: VA = 86.65 (90.9%), VC = 0 (0%), VE = 8.69 (9.1%); r_SSG = 0.75: VA = 69.61 (73.0%), VC = 8.46 (8.9%), VE = 17.28 (18.1%); r_SSG = 0.80: VA = 58.00 (≈60.8%), VC = 14.26 (≈15.0%), VE = 23.08 (≈24.2%). Main associations with mean performance: School quality positively associates with performance (b = 0.61, β = 0.06, p < 0.001), but this attenuates to b = 0.24 (β = 0.02, p < 0.001) when controlling for school SES and parental SES. Parental SES strongly associates with performance (b = 2.89, β = 0.30, p < 0.001). School SES shows a modest positive association (b = 0.92, β = 0.09, p < 0.001) net of parental SES and school quality. Moderation by school quality and school SES (r_SSG = 0.70): • School quality alone: genetic variance decreases with higher school quality (significant negative moderation of the a path; text reports b_SQ < 0, p ≈ 0.009). With r_SSG = 0.70 and SQ-only model, the a-path moderation is negative and significant; C shows no moderation (VC ≈ 0 in this specification). The total variance falls with higher school quality; standardized heritability changes little. • School SES alone: genetic variance decreases significantly with increasing school SES; non-shared environmental variance also declines with school SES; C shows no moderation. Accounting jointly for school quality, school SES, and parental SES: • When school SES is included alongside school quality, the genetic moderation by school quality is substantially reduced and becomes non-significant; model fit improves versus SQ-only. • Adding parental SES as a moderator further improves fit. The genetic moderation by school SES remains statistically significant but is reduced by about 40% in magnitude, indicating that part of the school SES moderation reflects parental SES, while residual moderation likely reflects school-based processes (e.g., composition/peer effects). • A small, statistically significant moderation of shared environmental variance by school SES appears after controlling for parental SES, but given the small absolute C to begin with, the evidence for a meaningful C moderation is weak. • The initially observed decline in E with school SES is explained by parental SES in the final model. Robustness: • Across r_SSG = 0.75 and 0.80, conclusions are similar: no evidence that school quality moderates genetic or shared environmental variance once SES is considered; under these r_SSG values, the SQ-only genetic moderation is not significant to begin with. The negative moderation of A by school SES persists when controlling for parental SES for r_SSG = 0.75, but not for r_SSG = 0.80. • Non-parametric quantile analyses largely mirror linear results: for school quality, genetic variance is elevated primarily in the lowest-quality quintile; for school SES, a clearer monotonic decline in genetic variance is seen across higher SES quantiles (with wide CIs). • Separating school resources and culture yields positive associations with average performance, similar to the overall quality factor. Both dimensions show negative moderation of genetic variance that disappears when controlling for school SES and parental SES. Of nine specific dimensions, none moderates A robustly; a single learning climate effect on C appears after full controls but is small and not corrected for multiple testing. Overall, school quality per se does not moderate genetic or shared environmental contributions once SES is accounted for; school SES is the salient moderator of genetic variance, partially attributable to parental SES.

The findings indicate that higher-quality schools, as measured by comprehensive resource and climate indicators, do not systematically magnify or attenuate genetic or family background influences on educational performance once school SES composition and parental SES are considered. What initially appeared as reduced genetic variance in higher-quality schools is better explained by SES composition—both at the family and school levels. The remaining school-level moderation after controlling for parental SES suggests school-based processes (e.g., peer composition, norms, instructional rigor tied to SES composition) may buffer genetic risks, aligning with diathesis-stress/compensation perspectives: supportive, higher-SES environments compensate for genetically influenced liabilities (e.g., learning or behavioral problems), thereby shrinking genetic variance and total variance. This implies a double disadvantage for low-SES children: fewer compensatory resources at home and greater exposure to lower-SES school contexts where compensation is less likely. The decline in total variance with higher SES could reflect compensation mechanisms or selection into more homogeneous environments (genetic and/or environmental), consistent with the absence of G×E in standardized components. The small effect of school quality on mean performance and lack of robust moderation could be due to limited between-school quality variation in the Netherlands (equitable funding, additional resources for disadvantaged intakes) or measurement constraints capturing school quality at a coarse level; within-school/classroom processes may be more decisive. Results cohere with Dutch twin and PGI studies documenting reduced genetic differences in more advantaged SES contexts. The study underscores the importance of modeling school quality, school SES, and parental SES jointly to avoid misattribution and to clarify whether observed moderation reflects family or school mechanisms.

Using population administrative twin data and comprehensive school indicators, the study shows that school quality does not moderate genetic or shared environmental influences on primary school achievement once school SES and parental SES are accounted for. Instead, a robust gene–environment interplay is observed for SES: genetic variance in educational performance is smaller in higher-SES family and school contexts, consistent with compensation of genetic risks. This implies that policies narrowly targeting school quality improvements may be insufficient to reduce educational inequality; addressing school SES composition (e.g., reducing segregation) and supporting children’s specific needs may be more impactful. Future research should identify mediating mechanisms (cognitive and non-cognitive traits, specific learning difficulties like dyslexia/ADHD), examine interactions between family and school contexts, investigate within-school/classroom quality, and triangulate twin-based and PGI-based designs for stronger causal inference.

Key limitations include: (1) Unknown zygosity for same-sex twins; genetic relatedness is approximated (r_SSG = 0.70/0.75/0.80) which may affect precision, though conclusions are robust across values. (2) Equal environments assumptions and potential small same-sex vs opposite-sex environmental differences could bias ACE components, though sensitivity checks suggest limited impact. (3) School quality indicators are administrative, irregular over time, and aggregated across years, potentially missing dynamic or within-school quality differences (e.g., teacher/classroom effects). (4) Possible ceiling effects in Cito scores; however, both standardized and raw scores show similar SES-related variance patterns and prior corrections have shown minimal influence. (5) Selection into schools and Cito testing may induce compositional differences; the Netherlands’ context of equitable funding may limit generalizability to systems with wider school quality disparities. (6) Associations of parental SES and school SES with performance are not causal in ACE models due to potential genetic confounding and unobserved factors. (7) The shared environmental component is small, limiting power to detect moderation of C.