Unpicking environmental consequences of genetically proxied mental health exposures in the UK Biobank

BackgroundAssociations between mental health and residential characteristics of a persons local area are well established in epidemiological literature. However, the clustering of common causes by geographical context (hereafter referred to as confounding via geography), may induce bias into associations for these relationships. Such confounding via geography is also likely to influence genetic exposures of interest, such as polygenic indices (PGI) for mental health. In this study, we investigate the extent to which this may have an impact on PGI analyses by considering within-area and contextual effects. MethodsData from UK Biobank was used to examine this (N=209,391 to 293,851). We conducted analyses contrasting single-level linear models with multilevel Mundlak models to estimate single-level, and geographically decomposed within-area and contextual effects between PGI for depression, wellbeing and schizophrenia and two greenspace outcomes (one measuring greenspace via land use and one measuring greenness via satellite spectroscopy). We used UK Census geography membership as our area level specification. Analyses were conducted with PGI derived at different value thresholds to assess effects for more and less strict p-value thresholds. ResultsSingle-level and within-area estimates were more similar for the most strictly filtered PGI exposures, with single-level effects between genetically predicted depression (at p<5e-8) and lower percentage greenspace (-0.004, 95%CI -0.007 to -0.00023), genetically predicted wellbeing (at p<0.1) and higher percentage greenspace 0.007, 95%CI 0.003 to 0.01) and genetically predicted schizophrenia (p<0.05) and lower percentage greenspace (-0.005, 95%CI -0.008 to -0.0006). Similar, but weaker effects were observed with the greenness outcome. At more relaxed p-value thresholds discordance between the within-area estimate and the single-level estimate is larger, with the within-area estimate being consistent with the more strictly selected PGI. Contextual effects also differed and, although have wide confidence intervals, they appear to in-part drive the single-level effects. We see that the areas in which single-level models perform worst are heavily urbanised areas (e.g., central London) and areas in National Parks (e.g. Peak District and North Yorkshire Moors). ConclusionThis study demonstrates that accounting for confounding via geography meaningfully alters associations between mental health PGI and greenspace outcomes. Given these results, we propose that genetic epidemiologists examining relationships in a spatially clustered population should consider routinely adopting a sensitivity analysis accounting for local context, and consider exploring within- and between-area decomposition of effects.