In non-European ancestries, we found over 100 replicates that our theory based on Eq. (1) mostly predicts a smaller RA than actually observed on average in the three scenarios (Fig. 2). This therefore makes our approach conservative. We observed a slightly larger downward bias in the predicted RA when using our heuristic approach based on candidate causal SNPs. On average, the latter approach showed an absolute underestimation of the RA of ~2.0% in SAS ancestry (i.e. the difference between the observed and the predicted RA is 92%−90% = 2.0%), 6.5% in EAS ancestry and 8.5% in AFR ancestry, which thus provides a lower bound for the RA. Interestingly, we did not find significant differences in observed and predicted accuracies across scenarios (ANOVA test, p-value >0.05). This somewhat surprising observation is suggestive that when heritability is constant and effect sizes of causal variants are perfectly correlated, differences in strengths of selection between ancestries might have a negligible impact on the RA of PGS. As a consequence, we can expect differences in strengths of selection between ancestries to mainly impact the term
