We investigated how the prediction accuracy of each method varied as a function of P-value thresholds, by varying either the EUR P-value threshold (Fig 2A and S12A Table) or the LAT P-value threshold (Fig 2B and S12B Table) between 10-8 and 1. In both cases, permissive P-value thresholds performed best, reflecting the relatively small sample sizes analyzed. However, the prediction accuracy of EUR+LAT+ANC was relatively stable, with prediction adjusted R2 > 0.037 across all EUR P-value thresholds (Fig 2A) and adjusted R2 > 0.033 across all LAT P-value thresholds (Fig 2B). In Fig 2A, we observe that as the EUR P-value threshold becomes more stringent, the difference in prediction accuracy between EUR+LAT+ANC and EUR+LAT increases, because EUR is less able to capture polygenic ancestry effects (see above).
