On these plots, lines that are below and to the left of other lines are more desirable. One interpretation is that, for a given level of missing data, an imputation method with a line further to the left has lower discordance with the external genotypes. Such plots allow us to evaluate competing methods in a more nuanced way than just looking at best-guess genotypes (which is equivalent to setting a single calling threshold of 1/3). We strongly emphasize, however, that the point of this exercise is not to determine an “optimal” calling threshold and use this to make hard calls of imputed genotypes for downstream analyses. Imputation results inherently contain more uncertainty than experimental genotype calls, and a host of methods have been developed to appropriately take this uncertainty into account when doing things like association testing [6]. Such methods are implemented in our freely available association testing software, SNPTEST.
