To assess the robustness of these accuracy benchmarks across genetic ancestries, we performed a similar set of benchmarks using the European, African, East Asian, and Latino GERA sub-cohorts (Online Methods). Because the latter three sub-cohorts were relatively small (Supplementary Table 1), we generated a single simulated reference panel for each sub-cohort containing all samples not belonging to trio pedigrees (Nref = 3,817, 5,164, 7,144, and 61,684 for the African, East Asian, Latino, and European sub-cohorts). We phased the three smaller panels using SHAPEIT2 and phased the European panel using Eagle1. We then benchmarked reference-based phasing accuracy by phasing the trio parents within each sub-cohort using the panel generated from that sub-cohort, running each method with default parameter settings. (We phased trio parents rather than trio children for these benchmarks because the three smaller data sets contained only 3–7 independent trios each; Supplementary Table 1.) These benchmarks confirmed our findings from the UK Biobank data: Eagle2 achieved 5–23% lower switch error rates than SHAPEIT2, and we observed the same relative ordering of accuracies as before across all sub-cohorts (Figure 3 and
