In this study, we investigated factors influencing imputation of missing genotypes for admixed populations, specifically African Americans. To achieve this goal, we tested three different programs representing three classes of fundamentally different underlying models for estimating the conditional distribution of haplotype frequencies. We also explored the decision of which populations to include in the imputation reference panel. For African Americans, obvious choices from the HapMap reference panels include the YRI panel (representing the single most closely related founder population with the highest admixture proportion), a mixture of the CEU and YRI panels (representing the presumed two ancestral founder populations), and the ASW panel (representing the single most closely related population). Although none of the programs explicitly account for admixture, the underlying models differ substantially in their ability to capture patterns of haplotype diversity created by admixture. We found that the best imputation results were achieved by running MACH twice, once with the HapMap CEU reference panel and a second time with the YRI reference panel, and then combining the results. In this way, we can capitalize on the higher accuracy
