We also benchmarked accuracy in all other 1000 Genomes populations containing >1 trio. We phased trio children in 31 Han Chinese (CHS) trios, 30 Peruvian (PEL) trios, 15 Punjabi (PJL) trios, and 19 Yoruba (YRI) trios using either the 1000 Genomes panel or the HRC panel, and we observed that in all cases Eagle2's accuracy was either slightly better or statistically indistiguishable from SHAPEIT2's (Supplementary Table 5). Specifically, the differences between Eagle2 and SHAPEIT2 were not significant for PEL with either reference panel and for YRI with HRC (p=0.05 or larger); all other differences were significant (binomial p=0.006 or less). Interestingly, all methods achieved lower accuracy using the HRC panel versus the 1000 Genomes panel (Supplementary Table 5). Given that the HRC panel contains the 1000 Genomes panel, this observation suggests that the inclusion of ≈30,000 additional predominantly European samples reduced the ability of each method to model the haplotype structure of non-European populations. However, we did not observe this phenomenon when phasing the two non-European UK Biobank trios using increasing numbers of European reference haplotypes (Supplementary Table 6), so this observation may be specific to the current HRC release (r1.1); development of the HRC is ongoing.
