Recursive Partitioning Using Ranks.
Mark Robert Segal · Defense Technical Information Center (DTIC) · 1985
Replacing the conventional splitting rules used in constructing regression trees by rules based on two sample rank statistics affords many advantages and equally poses some problems. Among the former are computational ease, invariance under monotone transformations of the response and worthwhile extension to censored data. The difficulties involve devising good pruning strategies in the absence of within node loss. These are addressed using look-ahead, bottom-up techniques. Some real-world and simulation performances of the methodology are presented. (Author)