Prediction accuracies (adjusted R2 on the liability scale (Lee et al., 2012), assuming sample prevalence 15%) and optimal weights for the 5 main methods (EUR, SAS, SAS+ANC, SAS+LAT, EUR+SAS+ANC) are reported in Table 4 (other prediction metrics are reported in S14 Table). In each case, the best prediction accuracy was obtained using LD-pruning threshold RLD2=0.8 (results using different LD-pruning thresholds are reported in S15 Table); the value of the optimal P-value threshold PT was equal to 10-3 for EUR and 0.8 for SAS. EUR performed only 14% better than SAS despite the larger training sample size, again reflecting a tradeoff between sample size and target-matched LD patterns. EUR+SAS attained 72%-95% relative improvements vs. EUR and SAS respectively (and used a slightly larger weight for EUR than for SAS). In addition, EUR+SAS attained a 44% relative improvement vs. EUR-SAS-meta (Table 4), again highlighting the advantages of optimizing mixing weights distinct from meta-analysis weights. Adding an ancestry predictor to EUR+SAS produced an insignificant change in accuracy for EUR+ SAS +ANC compared to EUR+SAS; we note a modest correlation between each prediction method
