While Figure 4 confirms that a large reference panel can improve imputation accuracy at rare and common SNPs alike, it is instructive to examine where the gains at rare SNPs are made. To evaluate this question, we took the results from Figure 4C and 4D and classified the kinds of errors made by each method's best-guess imputations (i.e., at a calling threshold of 1/3). We focused primarily on the ability of each method to detect copies of the minor allele. This is clearly an important quantity, but it is obscured by gross measures of accuracy, which are inherently dominated by homozygote-common genotypes at rare SNPs. We examined two classifications of erroneous minor allele calls: false positives (homozygous common called as heterozygous) and false negatives (heterozygous called as homozygous common). The results are shown in Table 2, where the false positive and false negative rates are expressed as percentages of the total number of homozygous common and heterozygous genotypes, respectively.
