Our results suggest that the model used by PLINK (which also underpins other imputation methods [7]) may be a liability in a dataset in which a large proportion of genotypes, including those in the reference panel, are unphased. However, we also note that PLINK's imputation functionality is still in beta testing. A recent study of Type 1 Diabetes [14] used a similar method to impute genotypes in a Scenario B dataset. Like PLINK, this method defines a multimarker tag for each SNP to be imputed, although in this case there is no phasing model since the tagging is based on correlations between unphased genotypes. It is not clear how this method would have fared in our comparison, but its similarities with PLINK imply that future studies might be better off using more sophisticated imputation methods.
