Both Eagle2 and SHAPEIT2 have an important parameter, K, that specifies the number of conditioning haplotypes used to phase each target sample and thus adjusts the speed-accuracy trade-off. We therefore also investigated the effects of varying K. (We note that the default values and precise meaning of this parameter are different for Eagle2 vs. SHAPEIT2; by default, SHAPEIT2 locally selects K=100 best reference haplotypes in each 2Mb window, while Eagle2 selects a fixed set of K=10,000 best reference haplotypes to use for the entire chromosome. This difference may be responsible for the slightly lower rate of improvement in accuracy of Eagle2 relative to that of SHAPEIT2 as Nref increases at fixed K in Fig. 2b.) We considered a range of values of K from 0.5–4 times the default K, similar to previous benchmarks of SHAPEIT212. The effects of varying K 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
