Precisely determining why these large differences in performance between the methods exist is difficult. We suspect that the reason resides in the fact that within the SHAPEIT2 method the haplotypes of each individual are explicitly modelled as a mosaic of the underlying haplotypes of other individuals [48]. In other words the underlying haplotype sharing between two individuals can be explicitly captured by allowing each individual to ‘copy’ the haplotypes of another individual over a long stretch of sequence. BEAGLE takes a different approach by collapsing the haplotype information of the sample into a compact graph. Each individual's haplotypes are then updated within the method conditional upon this graph. Thus no direct comparison between pairs of individuals is made and thus the information regarding long stretches of shared sequence between individuals is lost. These comments also apply to HAPI-UR which uses a different graph to encode the haplotypes of the samples. Our results are consistent across a range of cohorts with differing levels of relatedness. Most of these cohorts are isolated cohorts but the Split and GPC cohorts contain levels of relatedness that might be expected in a GWAS cohort.
