On these kinds of imputation datasets, IMPUTE2 shows clear computational advantages over Beagle. Even when using a large, ancestrally diverse reference panel, IMPUTE2 finished in less time on default settings (127 min with khap = 500) than it took Beagle to impute from an ancestrally homogeneous panel with almost five times fewer haplotypes (655 min). IMPUTE2 also required much less RAM: for each reference panel, Beagle needed about 20 times more memory. These results, in combination with our cross-validation results, confirm that IMPUTE2 is both more accurate and more efficient than Beagle in the kinds of imputation datasets that are beginning to drive the field. As with any sophisticated inference method, there are ways to tweak Beagle's settings to achieve better speed, but all of them would reduce imputation accuracy. (Beagle's memory footprint can also be reduced, at the cost of even longer running times.) We explore some of the factors underlying these computational differences, and the implications they hold for future methods development, in the Discussion.
