In our accuracy benchmarks, which we computed using gold standard trio phase calls, we observed that Eagle2 achieved 5–16% lower switch error rates2 compared to SHAPEIT2, with larger gains for lower values of Nref in the 15,000–100,000 range (Fig. 2b and Supplementary Table 2). Eagle2 achieved 4–36% lower switch error rates than Eagle1 and 63–81% lower switch error rates than SHAPEIT2 –no–mcmc in the same Nref range. All differences were statistically significant (binomial p=0.003 or less for each comparison of Eagle2 with SHAPEIT2, p=0.02 or less for each comparison with Eagle1, and p=10−21 for each comparison with SHAPEIT2 –no–mcmc).
