We note that the Eagle2 algorithm is very different from the long-range phasing algorithm we recently developed for phasing extremely large cohorts13. (We refer to the previous method as Eagle1.) The basic idea of Eagle1 was to harness identity-by-descent among distant relatives—which is pervasive at very large sample sizes but rare among smaller numbers of samples—to rapidly call phase using a fast scoring approach. In contrast, Eagle2 analyzes a full probabilistic model similar to the diploid Li-Stephens model used by previous HMM-based methods. Consequently, whereas Eagle1 suffered decreased accuracy compared to HMM-based methods when used to phase <50,000 samples, Eagle2 achieves improved accuracy over previous methods for both small and large haplotype reference panel sizes, as we demonstrate below. We note that when a reference panel contains fewer than twice as many samples as the target cohort, Eagle2 iteratively augments the reference panel with inferred target haplotypes (Online Methods); under this paradigm, reference-based phasing should always improve accuracy over cohort-based phasing. We also note that the Eagle1 algorithm was originally only implemented for cohort-based phasing; in this work, we have
