Several insights emerge from this table. First, IMPUTE v2 was consistently among the best methods for reducing both false negatives and false positives, suggesting that our new approach is generally more accurate than others at imputing rare SNPs. Second, while most methods were much more likely to make false negative calls than false positive calls, IMPUTE v1 was relatively more inclined toward false positives. This outcome is consistent with the use of a smaller reference panel, which will tend to overestimate the population frequencies of rare alleles. At the same time, IMPUTE v1 missed fewer true heterozygote calls than did fastPHASE or PLINK, despite using a much smaller reference panel. This tendency for fastPHASE and PLINK to mistake rare heterozygotes as homozygous for the major allele is the main factor separating these methods from IMPUTE v2 and BEAGLE in Figure 4C and 4D. Conversely, BEAGLE and IMPUTE v2 with k = 40 showed similar false negative rates, but IMPUTE v2 made half as many false positive rare allele calls.
