Criteria for Methods Comparison

In this study, we detail a framework for automatic learning of rules to locally match haplotypes and we compare it to an HMM-based method designed for the same purpose. That comparison method is inspired from Howie et al. (2009) and fully described in the section “Hidden Markov Model for Local Haplotype Matching.” In order to quantify the ability of each method to accurately achieve this purpose, we partition the full set of 182 haplotypes in reference and target panels. Haplotypes in the target panel are observed only on the LD map whereas those in the reference panel are observed on both LD and HD maps. Any given target haplotype is locally matched to all reference haplotypes on the LD map. Then based on the quality of these local matches, the target haplotype is inferred as a mosaic of the reference haplotypes (which are observed on the HD map).

The first and main criterion to compare methods is, for any target haplotype, the difference between the inferred and the true haplotypes on the HD map, measured by the metric e_A as the proportion of the 328,045 SNPs whose inferred allele is different from the true allele. Such haplotype-based comparison is possible because we consider the phased haplotypes as correct enough to be the true ones. To get rid of the remaining phasing errors in method comparisons, we used a second criterion based on genotypes rather than on haplotypes: imputation reliability (r²), measured, for any SNP specific to the HD map, as the squared correlation between imputed and observed genotypes of all target individuals (see section “Cross-Validation Plan,” for partitioning the population in reference and target). Details are given in the next sections on how imputation is performed within the random forests framework and the HMM. We also observed the number of switches from a reference haplotype to another one. Such an observation does not reflect the ability of the methods to reach their objective but provides information on their properties (how many segments from reference haplotypes does the method use when modeling a target haplotype as a mosaic).