For our benchmark analyses of N≈150,000 UK Biobank samples, we removed 144 trio parents and phased the remaining 152,105 samples. For our benchmarks on N≈50,000 or 15,000 samples, we phased all 72 trio children along with 1/3 or 1/10 of the remaining non-trio parent samples (50,752 or 15,270 samples in total). We evaluated phasing accuracy in trio children by comparing computational phase calls to trio phase calls (ignoring SNPs with Mendel errors); trio phase was available at ≈80% of heterozygous SNPs. For each child, we computed switch error rate by dividing the number of phase mismatches at consecutive trio-phased SNPs by the total number of trio-phased heterozygous SNPs minus 1 (ref.1), i.e., ≈15% of all SNPs (varying slightly among samples). In our results, we report mean switch error rates over the 70 European-ancestry trio children (according to self-reported ethnicity; see above). We applied an analogous procedure for our GERA benchmarks (differing only in that we removed all known relatives of the trio children—as the data contained a few extended pedigrees—leaving 60,929 samples).
