We assessed the accuracy of each phasing method using gold standard data from the 70 European-ancestry trios in the UK Biobank data set (all but one of which self-reported British ethnicity; see Online Methods). Specifically, we included all trio children and excluded all trio parents in each phasing run; we then assessed computational phase accuracy in trio children at all trio-phased sites (i.e., SNPs heterozygous in the child and homozygous in at least one parent, comprising ≈80% of heterozygous SNPs per trio child). We observed that when phasing N≈150,000 samples over the same 1% of the genome as above, Eagle and SHAPEIT2 achieved near-identical, very low (≈0.3%) mean switch error rates (Fig. 2c and Supplementary Table 1). The accuracy of Eagle relative to SHAPEIT2 degraded slightly with decreasing sample size (as expected with limited IBD in an outbred population); interestingly, however, Eagle still achieved better accuracy than all methods except SHAPEIT2 at sample sizes of N≈50,000 and N≈15,000, with only a 9% (s.e. 7%) increase in switch error rate relative to SHAPEIT2 at N≈50,000: 0.78% (0.05%) for Eagle vs. 0.71%
