There are two situations where researchers avoid imputation due to increased error in imputation: (1) SNPs with minor allele frequency less than 1% [1], [6], [10], and (2) association studies where cases and controls are genotyped on different platforms. Imputation accuracy, calculated for each SNP as the proportion of genotypes correctly classified, is the gold standard for evaluating the quality of imputation. Unfortunately, it is an inadequate filter in both of these circumstances. For the majority of SNPs, imputation programs such as IMPUTE [5], MACH[10], and BEAGLE[11], have very high imputation accuracy [5], [11], [12], [13]. However, the use of imputation accuracy in low frequency SNPs to evaluate imputation quality can be misleading. When the minor allele frequency of a SNP is less than 5%, a program could randomly assign the two alleles to the sample only using the minor allele frequency and achieve more than 90% accuracy. Although SNPs with low minor allele frequencies (MAF<5%) are referred to as uncommon SNPs, they represent more than 30% of SNPs in the HapMap Phase II CEU population, and this proportion is
