al., 2015; Weissbrod et al., 2016; Zhou et al., 2013). Third, our LDpred risk prediction method (Vilhjálmsson et al., 2015), which analyzes summary statistics in conjunction with LD information from a reference panel, is more accurate in European populations than the informed LD-pruning + P-value thresholding approach employed here; we did not employ LDpred due to the complexities of admixture-LD in analyses of admixed populations that explicitly model LD (Bulik-Sullivan et al., 2015), but extending LDpred to handle these complexities could further improve accuracy. Fourth, we note that in our application to real phenotypes adding an ancestry predictor produced insignificant changes in prediction accuracy, primarily because ancestry effects are captured by the polygenic risk scores; adding an ancestry predictor only improves prediction when we use a stringent P-value threshold to build the polygenic risk score (Fig 2). Fifth, we have not considered here how to improve prediction accuracy in data sets with related individuals (Tucker et al., 2015). Sixth, we did not incorporate local ancestry, which could potentially improve prediction accuracy in admixed populations (Seldin, Pasaniuc, & Price, 2011). Seventh, we did not incorporate data from the X chromosome, which is likely to harbor additional heritability that could improve prediction
