In light of the very low switch error rates achieved by both Eagle and SHAPEIT2 in N≈150,000 analyses, we further investigated the nature of the errors made by each method. We observed that many of the switch errors accrued by Eagle were the result of “blips” involving one or occasionally two adjacent SNPs oppositely phased relative to surrounding SNPs (Fig. 1d); such errors can arise from genotyping error or from recent mutation or gene conversion. We therefore computed an alternative metric, “switch error rate without blips,” in which we ignored errors in which 1–2 SNPs were oppositely phased relative to ≥10 consistently phased SNPs on both sides. This assessment showed that 1–2 SNP blips (previously counted as two switches) accounted for the majority (≈60%) of Eagle's switch errors; similarly, such errors accounted for roughly half of SHAPEIT2's switch errors (Table 1). We further considered the metric “discrepancies within a 10Mb segment”13 and observed that both Eagle and SHAPEIT2 achieved perfect phase in the majority of 10Mb segments phased (Table 1 and Supplementary Table 5).
