variants. Indeed, rare variants are more likely to be population specific and are usually poorly imputed using a small imputation reference panel. Thirdly, our study has focused on PGS calculated from GWS SNPs alone while other methods such as LDpred37 or SBayesR38, which utilise information from the entire genome, have been shown to be more accurate within ancestry. Although we acknowledge that characterizing the theoretical RA of PGS based on genome-wide methods like LDpred or SBayesR deserves further investigation, we also emphasise that the gap in prediction accuracy between GWS SNPs and genome-wide methods is destined to shrink as the sizes of GWAS continue to grow. Nevertheless, we re-analysed our simulated data using these two methods (Supplementary Note 8) and found that their predictive performance relative to our clumping strategy is near proportional in EUR and non-EUR ancestries. Therefore, we see a similar RA across all methods (Supplementary Fig. 13) although SBayesR shows the largest RA across scenarios. The latter observation mirrors our simulation results showing a constant RA as sample size increases (Supplementary Fig. 8). Fourthly, our predictions of the contribution of LD to the RA of PGS can in principle be inflated in the presence of epistatic interactions
