Unlike the other imputation methods with access to the expanded reference panel, PLINK achieved lower accuracy than IMPUTE v1; in Figure 4A and 4B, PLINK's best-guess discordances were 7.83% and 2.45%. We tried varying PLINK's settings from their defaults, including settings that were much more computationally rigorous, but these additional runs led to negligible improvements. PLINK is faster than the other methods considered here, which are all based on HMMs, but it also uses a simpler population genetics model. The multinomial haplotype frequency model that PLINK uses for imputation has fared poorly in recent comparisons of phasing methods [21]; its role in this analysis was to see if any accuracy is lost by using a simpler method to speed up imputation in a large and complex dataset.
