For each reference panel, we masked and imputed every 25th SNP in the GMB validation set, then repeated this analysis in a sliding window so that every genotyped SNP was imputed exactly once. To mimic a GWAS of unrelated individuals, we treated the nonmasked genotypes in the GMB validation set as unphased. We masked and imputed all available Illumina 650Y SNPs on chromosomes 20 and 11 (we added chromosome 11 to raise the counts of low-frequency SNPs, which are underrepresented in this dataset), except for those that were not typed in or had allele conflicts with HapMap 3, yielding a total of 40,300 SNPs for imputation. As in the HapMap 3 cross-validations, we split the chromosomes into non-overlapping 5-Mb regions to speed up the analysis and support IMPUTE2’s computational approximations.
