We performed simulations using real genotypes and simulated phenotypes (row 1 of Table 1). We simulated continuous phenotypes under a non-infinitesimal model with 1% of markers chosen to be causal with the same effect size in all samples and SNP-heritability hg2 = 0.5 (see Methods); we report the average adjusted R2 and standard errors over 100 simulations. We used WTCCC2 (Sawcer et al., 2011; Yang et al., 2014) data (15,622 samples after QC; see Methods) as the European training data, and the SIGMA data (SIGMA Type 2 Diabetes Consortium et al., 2014) (8,214 samples) as the Latino training and validation data (with 10-fold cross-validation). We simulated phenotypes using the 232,629 SNPs present in both data sets and built predictions from these SNPs excluding the causal SNPs, modeling the causal SNPs as untyped (see Methods).
