British samples (Online Methods; see URLs). This benchmarking procedure is commonly used to assess the accuracy of phasing and imputation pipelines5,9. We observed that when imputation was performed using the largest reference panel available (the N≈32,000 HRC), Eagle pre-phasing using all N≈150,000 samples improved imputation R2 by increasing amounts for increasingly rare SNPs, with a gain of 0.020 (0.002) in R2 for MAF 0.1–0.2% SNPs (R2=0.594 (0.012) for Eagle 1×150K vs. R2=0.574 (0.012) for SHAPEIT2 10×15K; Supplementary Table 12). (We caution that these results are based on the preliminary release (r1) of the HRC; development of the HRC is still underway, and performance may improve in future releases.) When imputation was performed using only the N≈4,000 UK10K reference panel (see URLs), gains were roughly half as large (Supplementary Table 13). Finally, to verify that similar improvements could be obtained at genome-wide SNPs (vs. the subsets of SNPs we masked), we ran the 1000 Genomes GBR samples through the same pipeline (after pre-phasing them together with the UK Biobank samples) and again observed a modest improvement using UK10K imputation (Supplementary Table 14). (We were unable to perform this experiment using HRC imputation because the HRC contains the 1000 Genomes data.) These
