Tree-based recursive partitioning methods for subdividing sibpairs into relatively more homogeneous subgroups

We propose a new splitting rule for recursively partitioning sibpair data i nto relatively more homogeneous subgroups. This strategy is designed to ide ntify subgroups of sibpairs such that within-subgroup analyses result in in creased power to detect linkage using Haseman-Elston regression. We assume that the subgroups can be defined by patterns of non-genetic binary covaria tes measured on each sibpair. The data we consider consists of the squared difference of a quantitative trait measurement on each sibpair, estimates o f identity-by-descent (IBD) values at each genetic marker, and binary covar iate data describing characteristics of the sibpair (e.g., race, sex, famil y history of disease). To test the efficacy of this method in linkage analy sis, we performed two simulation experiments. In the first, we simulated a mixture consisting of 66.6% of the sibpairs with no linkage and 33.3% of th e sibpairs with genetic linkage to one marker. The two groups were distingu ished by the value of a single binary covariate. We also simulated one unli nked marker and one random covariate to include as noise in the data. In th e second experiment, we simulated a mixture consisting of 55% of the sibpai rs with no genetic linkage. 22.5% of the sibpairs with genetic linkage to o ne marker, and 22.5% of the sibpairs with linkage to a different marker. Ea ch subgroup was defined by a distinct pattern of two binary covariates. We also simulated one unlinked marker and two random covariates to include as noise in the data. Our simulation studies found that we can significantly i ncrease the overall power to detect linkage by fitting Haseman-Elston regre ssion models to homogeneous subgroups with only a small increase in the fal se-positive rate. Second, the splitting rule can correctly identify importa nt covariates and linked markers. Third, recursive partitioning of sibpair data using this splitting rule can correctly identify sibpair subgroups. Th ese results indicate that partitioning sibpairs into homogeneous subgroups is feasible and significantly increases the power to detect linkage, thus d emonstrating the practical utility and potential this new methodology holds . (C) 2001 Wiley-Liss, Inc.