We first fitted a glm model for each different PS configuration, each time including as covariates sex, age, age2, age3, genotyping batch and European SNP fraction: these alone define a “base” model. We adopted different sets for this fitting step, including: (a) a balanced set of UK EUR plus admixed sets, (b) a balanced set only including admixed sets, (c) UK EUR as is, (d) a balanced set of UK EAS plus admixed sets, (e) a balanced set of all populations, see Supplementary Fig. 7. We then measured R2 on specific sample sets by predicting their phenotypes with the model fitted above and assessed standard deviation through 5000 bootstrap replications (boot package from R Software (https://www.r-project.org)). Added R2 is defined as realized R2 minus base model R2. Comparisons of nested and non-nested models were performed with likelihood ratio test from lmtest and Vuong test from nonnest2 R packages, respectively. When testing in EstBB we used the genotyping platform as covariate in place of genotyping batch. The incorrect local ancestry was defined as in “Population-wide aspPS analyses”.
