We extrapolated the results in Table 2 to larger sample sizes by limiting the simulations to subsets of chromosomes, as in previous work (Vilhjálmsson et al., 2015) (Fig 1 and S5 Table). EUR+LAT+ANC was the best performing method in each of these experiments. We also performed simulations using predictions constructed using all SNPs including the causal SNPs (S1 Fig and S6 Table). In these experiments, EUR+LAT+ANC was once again the best performing method, and EUR performed much better than LAT, consistent with the larger training sample size for EUR and the fact that differential tagging of causal SNPs is of reduced importance when causal SNPs are typed.
