We emphasize that IMPUTE v1 was among the fastest methods in this comparison only because it was assigned a much smaller problem: its reference panel contained 120 phased haplotypes, while every other method confronted a panel with 1,956 chromosomes, most of which were unphased. Using the knowledge that IMPUTE v1's computational burden grows quadratically with the number of chromosomes in the reference panel, we can project that it would have required over 1,300 minutes and 69,000 MB of RAM to run the program on a single 7.5 Mb analysis chunk with a reference panel of that size (which would also have needed to be phased ahead of time). This highlights a major advantage of our new modeling strategy: whereas IMPUTE v1 becomes computationally intractable as the reference panel grows, IMPUTE v2 remains competitive (both in computational burden and imputation accuracy) while allowing more flexibility (such as multiple, unphased, and/or incomplete reference panels).
