agreement (i.e., concordance) and subtracting out the chance agreement derived from the SNP’s allele frequencies, IQS is less prone to overestimating the imputation accuracy of low MAF SNPs. Our calculation of the IQS metric demonstrated an advantage of using IMPUTE2 to optimize imputation accuracy, a pattern which was not apparent from the comparison of concordance rates. The average r2hat metric further demonstrated that IMPUTE2 outperformed the other imputation programs, regardless of the reference panel used. Interestingly, unlike IMPUTE2 and the MaCH programs, the overall imputation quality of BEAGLE was not negatively affected by the inclusion of more distantly related reference subjects–a pattern that might reflect the differing algorithms that underlie BEAGLE versus the IMPUTE2 and MaCH programs. BEAGLE is based on localized haplotype clusters, whereas extensions of the IMPUTE and MaCH programs are based on population genetic principles for estimating the conditional distribution of haplotypes [15].
