In our primary simulations, we discarded the causal SNPs and used only the non-causal SNPs as input to the prediction methods (i.e. we simulated untyped causal SNPs, which we believe to be realistic). As an alternative, we also considered simulations in which we included the causal SNPs as input to the prediction methods (i.e., a scenario in which causal SNPs are typed). We performed simulations using all available European (WTCCC2) and Latino (SIGMA) training data (approximately a 2:1 ratio). We also performed simulations using training data in which Europeans were subsampled to attain a 1:1 ratio, as the relative performance of different methods may depend on relative training sample sizes; we considered different training sample sizes rather than different validation sample sizes, because the validation sample size does not (in expectation) impact the prediction accuracy.
