Expanding the environmental scope: an environment-wide association study for mental well-being

The study investigates how a broad set of neighbourhood-level environmental characteristics relate to individual well-being and whether observed associations may be confounded by genetic predisposition. Motivated by mixed evidence linking urbanization and environmental context to mental health, and enabled by advances in record linkage, the authors link population-based well-being data from the Netherlands Twin Register to geo-referenced environmental data from GECCO. Prior work often examined a limited set of exposures, risking selective reporting. This study adopts a comprehensive, data-driven Environment-Wide Association Study (EnWAS) to systematically test 139 environmental exposures for association with life satisfaction, and evaluates potential gene–environment correlation using polygenic scores. The purpose is to inform prevention, intervention, and policy by identifying modifiable, neighbourhood-level correlates of well-being.

The introduction summarizes evidence that urbanization has both benefits (e.g., infrastructure, healthcare access) and detriments (e.g., higher prevalence of mental disorders potentially due to reduced green space, social stress, and safety concerns). It notes genetic influences on residential choice and environmental exposure, exemplified by higher genetic risk for schizophrenia among individuals living in urbanized areas. Record linkage efforts like UK Biobank have advanced understanding of mental health determinants. A prior EnWAS identified psychosocial correlates of well-being but did not examine physical environmental characteristics. The authors highlight the need for comprehensive, neighbourhood-level analyses given policy relevance and potential genetic confounding.

Design: Environment-Wide Association Study (EnWAS) linking individual well-being data to neighbourhood environmental exposures at the 4-digit postal code (PC-4) level, with genetic analyses to test gene–environment correlation.

Data sources and sample: Adult participants from the Netherlands Twin Register (NTR), waves 6 (2002/2003; N=9,951) and 8 (2009/2010; N=11,975). Well-being measured using the 5-item Satisfaction With Life (SWL) scale (sum 7–35). Environmental exposures from the Geoscience and Health Cohort Consortium (GECCO), a centralized repository of longitudinal geo-data in the Netherlands. Of 1330 potential variables across 34 domains, 168 were preselected based on temporal availability and representativeness; post–quality control 139 variables remained across 22 domains. Variables were assessed either in 2002/2003, 2009/2010, or both (80 only in 2002/2003; 23 only in 2009/2010; 15 in both). Linkage used PC-4 postal codes.

Genetic data: NTR participants genotyped on multiple SNP arrays and imputed using the Genome of the Netherlands reference panel. Genomic principal components (PCs) were computed to capture ancestry and batch effects. Genetic data and well-being scores were available for 7,527 individuals.

Pre-registered analyses: (1) Regression: For each environmental predictor, generalized estimating equations (GEE) regressed SWL on the predictor, adjusting for sex, age, and age-squared. Familial relatedness was modeled via an exchangeable correlation structure with sandwich-corrected SEs. Given negligible within–postal code ICCs (0.02 in 2002/2003; 0.002 in 2009/2010), GEE was preferred over multilevel models. Multiple testing was controlled using Bonferroni correction at 0.05/139 = 3.6×10^-4. (2) Polygenic score (PGS): Well-being spectrum PGSs were constructed using LDpred from Baselmans et al. GWAS summary statistics (recomputed excluding NTR). PGSs were generated with PLINK allelic scoring. GEE tested associations between PGS (independent variable) and environmental correlates significant in EnWAS (dependent variables), adjusting for age, age-squared, sex, and the first 10 genomic PCs. Additionally, participants were stratified into PGS septiles; mean SWL and environmental values were compared across septiles via CI overlap.

Non–pre-registered analyses: (1) Multicollinearity: Correlations among significant environmental predictors were visualized using chord diagrams (thresholds |r|>0.4 and |r|>0.8). Principal component analysis (PCA) on standardized environmental exposures (prcomp in R) extracted uncorrelated PCs explaining at least 90% of variance (43 PCs for 2002/2003 covering 90.5% of 95 variables; 16 PCs for 2009/2010 covering 90.7% of 38 variables). These PCs predicted residualized SWL in an unrelated sample after regressing out age, age^2, and sex; variance explained was assessed. (2) SES correction: GEE EnWAS models were repeated adding covariates for individual educational attainment (EA) and neighbourhood SES (GECCO status score) to evaluate SES confounding.

Outcomes and reporting: Significant associations, effect sizes (β, SE), p-values, and R^2 were summarized (Tables 3–4; Supplementary Tables). Power and additional analyses (e.g., MZ twin intra-pair differences) are discussed in the article and supplements.

• 21 of 139 neighbourhood-level environmental variables were significantly associated with well-being after Bonferroni correction. Domains included housing stock, income, core neighbourhood characteristics, livability, and SES scores.
• Representative significant associations (GEE β [SE], p, R^2): higher owner-occupied housing % (0.051 [0.01], 1.31×10^-5, R^2=0.003); lower rental (%) and social rental (%) associated with lower SWL; higher income (80–100%) (0.069 [0.01], 4.89×10^-11, R^2=0.005); higher proportions of lower-income categories associated with lower SWL; higher mean house value (0.064 [0.01], 5.21×10^-11, R^2=0.004); higher livability (LBM) scores, population composition, housing, and safety scores positively associated with SWL; higher neighbourhood status score positively associated; rank order (worse) negatively associated; mean house transactions positive (0.051 [0.01], 2.52×10^-6, R^2=0.003).
• Effect sizes of individual environmental predictors were small (R^2 typically 0.2%–0.6%).
• Multicollinearity: Significant variables showed strong inter-correlations (up to |r|≈1.0), clustering into SES/housing and livability/safety groupings across time points.
• PCA: For 2002/2003, 43 PCs explained 90.5% of environmental variance; combined, PCs explained 1.45% of SWL variance (adjusted R^2=0.69%). PC3 (indicative of low-income neighbourhoods) negatively predicted SWL (β=-0.029, SE=0.006, p=2.73×10^-4). For 2009/2010, 16 PCs explained 90.7% of variance; combined, PCs explained 1.11% (adjusted R^2=0.79%). Two PCs significantly predicted SWL: PC1 (high income/livability; β=0.0185, SE=0.005, p=1.0×10^-4) and PC2 (low income/lower livability; β=-0.0240, SE=0.006, p=3.4×10^-5).
• Genetic analyses: The well-being PGS predicted SWL (R^2=0.007; p=5.11×10^-12) but did not predict any environmental correlates (all p>0.05), and no mean differences across PGS septiles were observed for environmental variables, providing no evidence of gene–environment correlation in this context.
• SES adjustment: Adding individual EA had small effects; after additionally adjusting for neighbourhood SES (status score), only neighbourhood safety and the percentage of land devoted to greenhouse horticulture remained significantly associated with SWL, with the latter showing limited geographical variability and warranting cautious interpretation.

By systematically linking individual well-being to a comprehensive set of neighbourhood environmental exposures, the study addresses which modifiable environmental features associate with life satisfaction. The significant correlates largely reflect socioeconomic indicators, with safety emerging as a robust factor. Strong inter-correlations suggest many associations are not independent; adjusting for both individual and neighbourhood SES attenuated most associations, indicating SES likely drives much of the observed pattern. The absence of detectable gene–environment correlation with current PGSs suggests that, at least with present predictive power, genetic liability for well-being neither strongly drives selection into these neighbourhood environments nor explains the associations observed. Environmental effects at the postal-code level are small and context dependent, and some commonly proposed factors (e.g., green space, air pollution) were not significant here, possibly due to geographic scale, objective versus subjective measures, or the specific well-being metric used. The findings underscore the importance of multi-level, data-driven approaches and suggest that improving neighbourhood safety and broader SES conditions may benefit population well-being, while highlighting the need to disentangle causality and refine environmental measurement.

Linking GECCO environmental data with NTR well-being data via an EnWAS identified 21 neighbourhood characteristics associated with life satisfaction, predominantly reflecting SES and livability, with neighbourhood safety standing out after SES adjustment. Effects were small individually and in aggregate, and no evidence was found for gene–environment correlation with current well-being PGSs. The work demonstrates the utility of large-scale record linkage and comprehensive, multi-domain environmental profiling to inform policy. Future research should: (1) examine multiple geospatial scales and incorporate subjective environmental measures; (2) conduct longitudinal analyses to clarify causal directions; (3) develop and validate poly-environmental scores, accounting for complex environmental covariance; (4) leverage genetically informative designs (e.g., MZ twin differences) with larger samples; and (5) explore interactions between genetic liability and environmental contexts.

Key limitations include: use of postal-code (PC-4) level exposures, which may not capture the most relevant spatial scale for all variables; reliance on objective environmental indicators that may differ from subjective perceptions; cross-sectional assessments at two time points limiting causal inference; potential temporal mismatches across environmental datasets; strong multicollinearity among exposures complicating attribution; limited variation for certain exposures (e.g., greenhouse horticulture); relatively low power of current well-being PGSs to detect subtle gene–environment correlations; and limited numbers of complete monozygotic twin pairs for intra-pair difference analyses.