An exploration into the causal relationships between educational attainment, intelligence, and wellbeing: an observational and two-sample Mendelian randomisation study

Education is linked to improved occupational status, income, health and other life outcomes, but whether education causally improves wellbeing remains unclear due to confounding and its high correlation with intelligence. Prior observational research shows mixed and potentially indirect effects of education on wellbeing via income, employment, marriage, and health, and suggests possible sex differences. Intelligence is strongly correlated with education and is often positively associated with wellbeing in correlational work, but associations can become negative after adjusting for correlated factors. This study asks whether educational attainment and intelligence have causal and independent effects on wellbeing, and whether associations are bidirectional. To address causality and disentangle correlated exposures, the study uses two-sample Mendelian randomisation (MR) with univariable and multivariable approaches, complemented by longitudinal observational analyses in the ALSPAC cohort to examine sex differences, non-linearity, and moderation by intelligence in emerging adulthood.

Observational studies report both positive and negative associations between educational attainment and wellbeing, with indirect effects via income appearing consistently positive for both sexes, and employment-related benefits more evident for males. When indirect paths are not accounted for, associations can appear negative, suggesting multiple mediating channels. Evidence from the UK has been limited and previously suggested little effect of education on happiness or of raising the school-leaving age. Intelligence correlates highly with education; observational associations with wellbeing are often positive but can switch to negative after adjusting for socioeconomic correlates, possibly due to higher expectations among high-ability individuals. Previous MR studies indicate education and intelligence exert independent causal effects on health and economic outcomes, but causal links to wellbeing have been underexplored. New multivariate GWAS instruments capturing a wellbeing spectrum (life satisfaction, positive affect, neuroticism, depressive symptoms) increase power for causal inference on wellbeing. These gaps motivate testing independent and bidirectional causal relationships among education, intelligence, and wellbeing.

Design: Two-sample Mendelian randomisation (MR) was used to test causal relationships among educational attainment (years of schooling), intelligence, and wellbeing using univariable MR (total effects) and multivariable MR (independent effects accounting for the other exposure). Bidirectional MR assessed effects of wellbeing on education and intelligence. Observational analyses in the ALSPAC cohort complemented MR to assess associations at age 26 and examine sex differences, non-linearity, and moderation. MR assumptions and sensitivity: MR relies on instruments strongly associated with the exposure, independent of confounders, and affecting outcomes only via the exposure. Sensitivity to pleiotropy was addressed using multiple MR estimators (IVW as primary; MR-Egger, weighted median, weighted mode as sensitivity), Cochran’s Q for heterogeneity, MR-Egger intercept for directional horizontal pleiotropy, and multiplicative random-effects IVW. SIMEX corrections were applied to MR-Egger where regression dilution was low. Steiger filtering removed SNPs explaining more variance in the outcome than exposure. Multivariable MR used Sanderson–Windmeijer partial F statistics for instrument strength. GWAS data sources: Educational attainment (Years of Schooling) from Okbay et al. discovery GWAS (n≈293,723; ISCED-based; initial 74 genome-wide significant SNPs), with replication including UK Biobank (total n≈405,072; 162 loci). Discovery GWAS was primarily used to reduce sample overlap; replication analyses checked consistency. Intelligence from Savage et al. (n=269,867; 14 cohorts; general intelligence factor; 242 lead SNPs). Wellbeing from Baselmans et al. multivariate GWAMA: N-GWAMA instrument for the wellbeing spectrum (life satisfaction, positive affect, neuroticism, depressive symptoms), identifying 231 independent SNPs; trait-specific MA-GWAMA used in follow-ups. Sample overlaps with education (~11%) and intelligence (~8%) were similar to prior MR work. Instrument selection and harmonisation: SNPs were clumped at r²<0.001 within 10,000 kb and required genome-wide significance (p<5×10⁻⁸). Palindromic SNPs aligned using MAF threshold 0.42. Univariable instrument strengths (F>10) indicated minimal weak instrument bias. Example counts and F-statistics: education→wellbeing (54 SNPs; F=38.88), wellbeing→education (147 SNPs; F=40.78), intelligence→wellbeing (126 SNPs; F=43.35), wellbeing→intelligence (128 SNPs; F=40.83). Multivariable MR used 151 SNPs with conditional F-statistics approximately 7.23 (education) and 7.94 (intelligence). Statistical MR analyses: Primary estimator was IVW; MR-Egger, weighted median, and weighted mode provided pleiotropy-robust sensitivity checks. Heterogeneity (Cochran’s Q) and MR-Egger intercepts assessed pleiotropy. Additional multivariable MR examined: (a) intelligence and wellbeing predicting years of schooling; (b) years of schooling and wellbeing predicting intelligence. Multiple testing used Benjamini–Hochberg FDR. Observational analyses (ALSPAC): Participants from the Avon Longitudinal Study of Parents and Children with education at 26, intelligence at 8, and wellbeing at 26 were included (complete-case n=2844; up to n=3788 for education and n=3179 for intelligence analyses). Measures: educational attainment via self-reported university degree (yes/no) at 26; intelligence via WISC-III total score at age 8; wellbeing via Subjective Happiness Scale and Satisfaction with Life Scale at 26 (z-standardised). Analyses: separate linear regressions for (a) education→wellbeing and (b) intelligence→wellbeing; models with sex main effects and sex interactions; tests for non-linearity in intelligence (quadratic, cubic, quartic polynomials; GAM splines via mgcv; model fit by AIC/BIC); moderation between education and intelligence; and adjustment for family income. Multiple testing controlled via FDR (62 tests). Attrition and missingness addressed with inverse probability weighting (IPW) and multiple imputation by chained equations (MICE; m=60), following prior ALSPAC practice.

Mendelian randomisation (univariable): - Education→Wellbeing (total effect): Positive causal effect; per 1 SD increase in years of schooling (~3.6 years), wellbeing increased by β=0.057 (95% CI 0.042, 0.074; IVW). Evidence also for Wellbeing→Education: β=0.206 SD years (95% CI 0.071, 0.341). MR-Egger did not replicate likely due to measurement error; no evidence of directional pleiotropy (intercepts ~0; balanced funnel plots); Steiger filtering supported direction. - Intelligence→Wellbeing (total effect): No causal effect detected (IVW β≈-0.004, 95% CI -0.028, 0.017). Wellbeing→Intelligence: Positive causal effect β=0.199 SD (95% CI 0.014, 0.390). Sensitivity analyses consistent; no evidence of directional pleiotropy. Mendelian randomisation (multivariable): - Independent effects on wellbeing: Education positive; Intelligence negative. - Education (years of schooling)→Wellbeing: β=0.103 (95% CI 0.047, 0.159; IVW), controlling for intelligence. - Intelligence→Wellbeing: β=-0.044 (95% CI -0.079, -0.009; IVW), controlling for education. MR-Egger showed consistent negative effect (β=-0.075, 95% CI -0.124, -0.026). Conditional F statistics indicated relatively weak multivariable instruments (education F≈7.23; intelligence F≈7.94), but results remained after multiple-testing correction. - Additional multivariable MR (bidirectionality): Wellbeing independently predicted more years of schooling controlling for intelligence (β=0.193, 95% CI 0.07, 0.31). Intelligence independently predicted more years of schooling controlling for wellbeing (β=0.44, 95% CI 0.40, 0.48). - Trait-specific follow-ups: Results were similar using life satisfaction and positive affect components; neuroticism and depression showed opposite directions, as expected. Observational (ALSPAC): - Sample: ≈66.7% had a university degree. Degree holders had higher childhood intelligence (mean 112.21 vs 99.07; Welch t(1879)=22.2, p<0.001). Subjective happiness mean 4.89 (1–7); life satisfaction mean 24.25 (5–35). Degree associated with higher life satisfaction (mean 24.78 vs 23.09; t(2591)=6.99, p<0.001), not with happiness. - Linear models: University degree predicted higher life satisfaction (robust across IPW and MI); no association with subjective happiness overall. Significant sex interactions: degree-by-sex predicted both outcomes. Females with a degree showed greater life satisfaction than females without; for happiness, females benefitted from degree whereas males with degrees showed lower happiness. - Intelligence: Unadjusted, higher intelligence predicted lower happiness and higher life satisfaction. With sex interaction, the association with happiness became positive overall but indicated moderation by sex: males with lower intelligence reported higher subjective happiness; life satisfaction association with intelligence remained positive. - Non-linearity and moderation: No evidence for non-linear intelligence–wellbeing relationships (polynomial or GAM). No strong moderation between education and intelligence on wellbeing. Family income did not explain the observed associations. Findings were robust to multiple-testing correction and to attrition adjustments (IPW, MI).

Findings indicate bidirectional causal relationships: wellbeing increases both years of schooling and intelligence, while educational attainment causally improves wellbeing. Crucially, multivariable MR shows that education has a direct positive effect on wellbeing independent of intelligence, and that intelligence has a direct negative effect on wellbeing independent of education. Observational analyses in young adulthood support these patterns with important sex differences: females derive more wellbeing benefits from higher education than males, and males with lower intelligence reported higher subjective happiness. These results reconcile mixed observational literature by distinguishing total versus independent effects of correlated predictors and by differentiating components of wellbeing (life satisfaction vs subjective happiness). Potential mechanisms include greater socialisation and social capital from extended education, sex-specific health behaviours, and higher expectations or susceptibility to stress among more intelligent individuals. Policy implications include prioritizing retention and engagement in education to promote wellbeing, and targeting wellbeing support to highly intelligent students who may face increased academic stress. The reciprocal relationship between wellbeing and educational attainment suggests school-based wellbeing interventions may yield downstream gains in education and cognition, further reinforcing wellbeing.

This study combines genetic and observational evidence to show that staying in education has a unique, causal, and protective effect on wellbeing beyond intelligence, while intelligence exerts a direct negative effect on wellbeing when accounting for education. Benefits of higher education for wellbeing are likely greater for females than males, suggesting tailored wellbeing support may be warranted, particularly for males and for highly intelligent students. Future research should: (1) use repeated measures to chart how causal effects unfold over time; (2) identify mediators (e.g., socialisation, health behaviours) and differentiate determinants of life satisfaction versus subjective happiness; (3) refine measurement of educational attainment versus achievement and assess roles of non-cognitive skills; (4) employ within-family MR to address assortative mating and dynastic effects; (5) replicate across countries, ancestries, and cohorts, accounting for macroeconomic contexts (e.g., graduation during recessions, post-COVID impacts).

MR analysis may be affected by selection and generalisability issues due to UK Biobank participation (more educated than general population) and potential socioeconomic confounding. Assortative mating and dynastic effects could bias MR estimates; within-family designs were not feasible here. Effect sizes for wellbeing are hard to interpret due to composite GWAMA measures. Intelligence GWAS included samples conditioned on socioeconomic status; sensitivity analyses removing these were consistent but warrant caution. Multivariable MR instruments had modest conditional F values, raising potential weak-instrument concerns. Observational analyses used a binary education measure (degree vs non-degree), precluding assessment of non-linear or cumulative years of schooling effects. Wellbeing was measured at age 26, potentially before gaps by education fully emerge; timing and life transitions may influence happiness differently from life satisfaction. Cohort and period effects (e.g., economic conditions at graduation) limit generalisability despite adjustments for attrition and missingness.