10,000-SNP region to keep per-job run times within the 5-day limit (Table 1). We observed that using K=200 conditioning states achieved accuracy similar to Eagle, while using K=400 states achieved the lowest switch error rate of all methods tested (0.243%, s.e. 0.011%; p=0.007 vs. Eagle, one-sided paired t-test) (Table 1). (We note that SHAPEIT2 K=400 analyses required ≈40% more computation time than default SHAPEIT2 K=100 analyses; while the run time scaling of SHAPEIT2 is asymptotically linear in K (ref.12), the quadratic component of the computation, which is independent of K, dominates at very large N and typical K.) We also considered increasing SHAPEIT2's window size parameter from 2Mb to 4Mb, but results of a pilot experiment indicated that doing so substantially decreased accuracy (Supplementary Table 7).
