Figure 4C and 4D mirrors Figure 4A and 4B, respectively, but these results are restricted to imputed SNPs with minor allele frequencies (MAFs) less than 5%—Figure 4C is based on 1,113 SNPs and Figure 4D is based on 1,979 SNPs. The same relative patterns remain, although the discordance and missing data percentages are lower because it is easier to guess most of the genotypes correctly and with high confidence at a rare SNP than a common one, simply because most genotypes at a rare SNP will be homozygous for the common allele. Among the most accurate methods, the best-guess discordances based on Affymetrix genotypes (Figure 4C) were 1.01% (IMPUTE v2, k = 40), 0.84% (IMPUTE v2, k = 80), and 0.97% (BEAGLE), as compared to 1.73% for HapMap-based imputation with IMPUTE v1; the discordances based on Illumina genotypes (Figure 4D) were 0.48% (IMPUTE v2, k = 40), 0.38% (IMPUTE v2, k = 80), and 0.46% (BEAGLE), as compared to 0.93% for IMPUTE v1.
