The results for the TSI panel (Figure 1B) are more surprising. As we would expect, accuracy improved with the addition of each panel that contains recent European ancestry (CEU and MEX). The GIH panel also increased the accuracy, which is reasonable given the relatively modest genetic divergence between this panel and TSI. Perhaps unexpectedly, the maximum imputation accuracy in TSI was not achieved until a set of African haplotypes (MKK) was added to the reference panel. This improvement was also observed when adding MKK haplotypes to the reference set for CEU imputation (File S1). Supplementary results from the HapMap 3 paper (The International HapMap Consortium 2010) show that the MKK panel has an admixture component that could reflect an ancient migration from Europe or the Middle East into eastern Africa, which might explain our results, although we observed a similar increase in accuracy after replacing the MKK panel with the remaining African panels (data not shown). Regardless of the underlying explanation, these results highlight the complexity of human demographic history, which is one motivation for frameworks like ours that use inclusive reference panels without population labels.
