In our accuracy benchmarks using the 70 European-ancestry UK Biobank trios, we observed that Eagle2 achieved better accuracy than SHAPEIT2 and Eagle1 for N≤50,000, as expected (Fig. 5b and Supplementary Tables 7 and 8). At N=150,000, Eagle1 achieved a slightly lower switch error rate (0.31%, s.e. 0.02%) than Eagle2 (0.35%, s.e. 0.02%). However, we observed that running Eagle2 with 4x the default number of conditioning haplotypes (i.e., K=40,000) achieved the lowest error rates across all sample sizes tested (0.27%, s.e. 0.02% at N=150,000). Both differences were statistically significant (binomial p=0.0006 or less). Finally, we confirmed that Eagle2 achieved better phasing accuracy than SHAPEIT2 or Eagle1 when used to phase the GERA samples within each GERA sub-cohort (Supplementary Table 9), with switch error rates consistent with our earlier reference-based benchmarks (Figure 3 and Supplementary Table 4). All differences were statistically significant (binomial p=0.002 or less).
