were broadly consistent for Eagle2, SHAPEIT2, and SHAPEIT2 –no–mcmc: all methods required similarly increased computation time and achieved improved accuracy with larger values of K (Fig. 2c,d and Supplementary Tables 2 and 3). In particular, increasing the number of conditioning haplotypes by a factor of 4x required 2–3x more computation time for both Eagle2 and SHAPEIT2 while achieving similar decreases in switch error rates (12–20% and 17–19%, respectively, for Nref = 15,000–100,000; Supplementary Table 2). All improvements were statistically significant (binomial p=10−8 or less).
