We used the ancestry-stratified meta-analysis results to construct ancestry-specific polygenic risk scores in Add Health35, an independent target sample of individuals of diverse ancestries from the United States (n = 2,199 AFR, 1,132 AMR, 525 EAS and 6,092 EUR). To evaluate within-ancestry and across-ancestry performance of polygenic scores, we iteratively fit a multiple regression model and evaluated the incremental predictive accuracy of each ancestry-based score, over and above scores already entered into the model (that is, first including the AMR-based score, then adding the AFR-based, EAS-based and EUR-based scores one at a time to evaluate incremental prediction accuracy). EUR-based scores in EUR ancestries outperformed ancestry-matched scores in non-EUR ancestries (Fig. 2a) and showed significantly stronger associations with most phenotypes in EUR ancestries than in non-EUR ancestries (described by decile plots and tested by modelling an interaction between the EUR-based polygenic risk score and the target sample ancestry group), consistent with expectations36 (Fig. 2b,c). For each ancestry and phenotype, the EUR-based score on its own outperformed the ancestry-matched score on its own (Supplementary Table 12). These results highlight the relative utility
