K-Adaptive Partitioning for Survival Data, with an Application to Cancer Staging

Abstract

In medical research, it is often needed to obtain subgroups with heterogeneous survivals, which have been predicted from a prognostic factor. For this purpose, a binary split has often been used once or recursively; however, binary partitioning may not provide an optimal set of well separated subgroups. We propose a multi-way partitioning algorithm, which divides the data into K heterogeneous subgroups based on the information from a prognostic factor. The resulting subgroups show significant differences in survival. Such a multi-way partition is found by maximizing the minimum of the subgroup pairwise test statistics. An optimal number of subgroups is determined by a permutation test. Our developed algorithm is compared with two binary recursive partitioning algorithms. In addition, its usefulness is demonstrated with a real data of colorectal cancer cases from the Surveillance Epidemiology and End Results program. We have implemented our algorithm into an R package maps, which is freely available in the Comprehensive R Archive Network (CRAN).