We repeated this experiment for each of the 10 HapMap 3 panels in Table 1 under various conditions. To understand the benefits and drawbacks of different reference panel compositions, we sequentially added HapMap 3 panels to the reference set and re-imputed after each addition. Adding the panels in all possible orders is combinatorically daunting and of dubious interpretive value, and we preferred to have an objective ordering, so we added panels in order of increasing pairwise FST to the cross-validation panel. We calculated FST between each pair of panels as the average across all HapMap 3 SNPs on the autosomes. The panel orderings induced by this criterion are shown in File S1. We also considered different orderings – e.g. adding panels as dictated by a greedy algorithm based on cross-validation accuracy – but these did not change our qualitative results (data not shown).
