although this study examined a range of alcohol use phenotypes (from a measure of monthly alcohol use to dependence), the cross-sectional design cannot capture potential longitudinal changes in PRS association across the progression of stages of alcohol use. A sixth limitation is that we did not control for any measures of socioeconomic status (SES) in our models, although SES has been shown to be a potential moderator of polygenic risk for alcohol consumption (Clarke et al., 2016; Barr et al., 2018), as the measures available varied widely and inconsistently across our target samples. The imbalance of cases and controls in the UK Biobank target sample is another limitation; this imbalance would be more likely to seriously bias a discovery GWAS, but may still have confounded the regression analyses in our target sample. A final limitation is that the strength of PRS associations varied across p-value thresholds (pT; see Supplementary Tables 1–12); because we chose the best-performing pT for each model, as is typical for the pruning+thresholding (P+T) method of polygenic risk score analyses (Euesden et al., 2014), we run the risk of over-fitting.
