more power. However, if they are different, then the weights estimated from the validating dataset will be biased toward that dataset and different from the independent datasets, resulting in loss of power. As we noted earlier, AA is a very heterogeneous population. The three AA target datasets in this study were recruited under different ascertainment strategies and in different regions, therefore, meta-analyzed posterior effect sizes were used in this study. In addition, since we only focused on AUD-associated variants implicated in both AA and EA, meta-analysis posterior effect sizes should provide more accurate estimates for those variants. We first selected variants that had P-values < 0.05 in both EA-PAU and AA-AUD (i.e., at least showing marginal associations) and had the same directions of effects (referred to as concordant variants). For our gene-based PRS (PRSgene), only concordant variants located within gene boundaries (defined as within the region containing the gene plus 1 kb upstream of the transcription start site and 1 kb downstream of the transcription end site; annotated using ANNOVAR [45] based on NCBI RefSeq GRCh37) were used. To test whether using any concordant variants regardless of location would do as well, we calculated PRS using concordant variants located outside
