Prediction accuracies (adjusted R2) and optimal weights for the 5 main methods (EUR, AFR, AFR+ANC, EUR+AFR, EUR+AFR+ANC) are reported in Table 5. For EUR and AFR, the best prediction accuracy was obtained using RLD2=0.2 and RLD2=0.8 respectively, thus we used these respective values of RLD2 for EUR and AFR in each PRS in all primary analyses (results using different LD thresholds are reported in S18 Table); the value of the optimal P-value threshold PT was equal to 10-3 for EUR and 0.05 for AFR. EUR performed much better than AFR, consistent with the far larger training sample size. Nevertheless, EUR+AFR attained a 30% improvement vs. EUR (using a larger weight for EUR than for AFR). EUR+AFR also attained a small relative improvement (7%) vs. EUR-AFR-meta (Table 5). Adding an ancestry predictor to EUR+AFR produced an insignificant change in accuracy for EUR+AFR+ANC compared to EUR+AFR; we note a modest correlation between each prediction method and the proportion of European-related ancestry (Reich et al., 2009) within African samples (see S19 Table). We repeated our analyses using stratified 10-fold cross-validation to estimate mixing
