For comparison purposes, we also performed simulations using training data in which Europeans were subsampled to attain a 1:1 ratio (row 2 of Table 1); prediction accuracies and optimal weights for the 5 main methods are reported in Table 2B. On average, the LAT method performed 190% better than the EUR method, again demonstrating the advantages of target-matched LD patterns. EUR+LAT attained 24%-260% relative improvements vs. LAT and EUR respectively (and used a larger weight for LAT than for EUR), again highlighting the advantages of incorporating multiple sources of training data.
