Overall, the variance explained by PRS was low, even when the discovery and target phenotypes were identical (e.g. AUDIT scores in ALSPAC). These estimates are nonetheless consistent with other PRS studies (Mies et al., 2018; Savage et al., 2018), and the proportion of explained variance may improve as the discovery GWAS get larger. In agreement with previous studies (Savage et al., 2018), we demonstrated that ascertainment is likely to be a crucial factor in polygenic prediction analyses. For instance, while the AUDIT-P PRS explained a larger amount of variance in dependence in COGA than in ALSPAC, the AUDIT-C PRS was not a significant predictor for any phenotype in COGA over and above the effect of AUDIT-P PRS, suggesting that the effect sizes for AUDIT-C from a volunteer cohort (UKB) may not be optimal for capturing genetic risk to AUD in high-risk families, such as those in COGA. AUDIT-C associations in COGA using other addiction-enriched samples, such as the Million Veteran Program (Kranzler et al., 2019), where the genetic correlation between AUDIT-C and AUD is high, will likely provide insights into
