Intuitively, the RPM is a multi-locus ‘measured genotype’ (Boerwinkle et al. 1986) approach, assessing the mean trait values or prevalence information for different multi-locus genotypes. Genotypes (groups) are iteratively merged if a multiple comparisons test indicates that they are not significantly different. The algorithm halts when either all genotypes are merged (indicating no significant association between these genotypes and the phenotype) or when the remaining groups are significantly different from each other. In this case, the resulting model is evaluated for the proportion of trait variance explained by the groups. Significance of the full model (not just the interaction) is evaluated using permutation tests because the goal is to identify contributors to phenotypic variation, not to identify or test interactions. The benefits of focusing on the full model, rather than on specific interactions, have been demonstrated by extensive evaluation in a variety of simulated models (Marchini et al. 2005). We note that if the RPM model explains none of the variance (i.e. all the genotypes are merged into a single group), then the p value is necessarily = 1, as
