Within each of the 10 panels, we masked a set of SNPs in one individual at a time and used IMPUTE2 to infer the hidden genotypes. We repeated this procedure across a range of khap values and with various reference panels, which were creating by cumulatively adding HapMap 3 panels in the order dictated by genome-wide average FST. One way to think of this procedure is to imagine building a composite reference panel for a population of interest: we start with a population-specific reference panel, and we successively add more-diverged panels to see if they will help (or possibly hurt) the imputation accuracy. These composite panels capture aspects of the Huang et al. (2009a) cross-validation strategy for choosing reference sets, and they provide population-label-informed benchmarks against which to compare our label-free way of using reference data.
