Recursive partitioning for the identification of disease risk subgroups: a case-control study of subarachnoid hemorrhage.

Recursive partitioning is a nonparametric technique that produces a classification tree in which subjects are assigned to mutually exclusive subsets according to a set of predictor variables. We examined the utility of recursive partitioning as a supplement to logistic regression for the multivariable analysis of data from case-control studies, demonstrating the technique using data from women enrolled in a population-based study of subarachnoid hemorrhage. The classification tree produced by recursive partitioning consisted of three main risk subgroups: (1) elderly women who had long-standing hypertension and who were not smokers, (2) middle-aged women who were cigarette smokers and frequent binge drinkers, and (3) women in whom risk variables indicate relative estrogen deficiency (i.e., postmenopausal status, no recent exposure to hormone replacement therapy, cigarette smoking). As a supplemental method, recursive partitioning not only identifies subgroups with varying risks, but also may uncover interactions between variables that may be overlooked in the traditional application of logistic regression to case-control data.