gms | German Medical Science

Survival trees have become a popular method to analyze failure time data. Specifically, survival trees are a valuable alternative to (semi-)parametric survival modelling when the number of predictor variables in a data set is large or when complex interactions between predictor variables are present. Classical survival tree methodology is limited by the fact that algorithms for tree construction are designed for continuous outcome variables only. Hence classical methods cannot be applied to failure time data that are measured on a discrete time scale (as is often the case in surveys or clinical trials, where data are collected, e.g., quarterly or yearly). To overcome this problem, we present a method for discrete-time survival tree construction. The proposed technique is based on the fact that the likelihood of a discrete-time survival model is equivalent to the likelihood of a regression model for binary outcome data. Hence we propose to modify tree construction methods for binary outcomes (such as CART and C4.5) such that they result in optimized partitions for estimating discrete-time survival functions.