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 Supplementary Table 4). All differences were statistically significant (binomial p=10−7 or less). We note that every method had a higher switch error rate in the GERA European sub-cohort compared to the UK Biobank, presumably due primarily to a more diverse set of ancestries represented. In general, absolute switch error rates are not directly comparable among data sets due to differences in demography and genotyping properties (e.g., chip density, allele frequency distribution, and genotype error rate).
