Table 1 shows that IMPUTE v1 was the fastest of the methods considered here, followed by BEAGLE (default), MACH, IMPUTE v2 (k = 40), BEAGLE (50 iterations), fastPHASE (K = 20), IMPUTE v2 (k = 80), and fastPHASE (K = 30). Conversely, fastPHASE required the least computer memory, followed by MACH, IMPUTE v2, IMPUTE v1, and BEAGLE. Note that, while IMPUTE v2 with k = 40 took about six times as long as IMPUTE v1, it needed less than 16% of the RAM; this is mainly a consequence of modeling SNPs in U as haploid in version 2, as opposed to diploid in version 1. We also note that both fastPHASE and MACH spent most of their running time fitting their models to the HapMap, and that both methods could probably decrease running times (via more lenient settings) without sacrificing much accuracy.
