A Comparison of Node-Splitting Rules in Recursive Partitioning Analysis of Multivariate Quantitative Structure Activity Data

Abstract

Tree-structured methods have been advocated to model nonlinear relationships in quantitative structure-activity relationship (QSAR) studies. One such algorithm, MultiSCAM, was developed to analyze QSAR data with multivariate continuous responses, classifying objects into similar categories by building a tree with recursive partitioning. Hotelling’s T2 test is the method used to determine splits in MultiSCAM. However, this test is not feasible when the dimension of an observation exceeds the number of observations, which often happens when growing trees. This problem is further exacerbated by the fact that missing values are common in QSAR data. We consider two alternatives, the pooled component test (PCT) and a simple ANOVA F test. To compare the three node-splitting tests, we introduce a comprehensive simulation design that could be used by others to evaluate their methods of tree building. A drug discovery dataset is used to illustrate the methods.