For our reference-based phasing benchmarks using GERA data, we applied an analogous procedure with the following minor differences. For each of the four sub-cohorts, we created a single simulated reference panel using all samples not in the same extended pedigree as any trio. We phased the European chip panel using Eagle1 and phased the other three panels using SHAPEIT2 (which is computationally tractable and more accurate than Eagle1 for small cohorts13). In each sub-cohort, we then applied each reference-based phasing method to phase all 22 autosomes of the trio parents in that sub-cohort, and we computed mean switch error rates over all trio parents. We chose to benchmark accuracy using trio parents rather than trio children due to the small numbers of trios in the non-European sub-cohorts. (For the European sub-cohort, we also computed benchmarks using trio children for comparison.) For the European sub-cohort, we computed s.e.m. over samples as before; for the other three sub-cohorts, we computed s.e.m. over 25 SNP blocks due to the small numbers of trios. When used to compare methods, these standard errors are conservative
