Exponential survival trees

Abstract

This paper concerns a recursive partitioning algorithm for incomplete survival data. The splitting criterion employed is exponential log-likelihood loss. The method has been integrated into a computer program similar in structure to the commercial package CART. Special features incorporated into the program include a modification to prevent zero hazard estimates during the the cross-validation procedure and a method based on the chi-square distribution for final tree selection. Simulation results demonstrate the program's capability to identify the underlying hazard structure in survival data. An example utilizing data from patients with diffuse large cell lymphoma is presented.