Abstract
An important theme in survival analysis is the investigation of prognosis factors that affect survival time. The tree-structured method has been applied to evaluate covariates; however, this method apparently has poor predictive outcomes. This problem may be improved by modeling many trees in a linear combination, as carried out in ensemble learning. The ensemble learning method has been actively studied in machine learning and statistics. Recently, several ensemble methods have been extended to ensure right-censored survival outcomes. Ishwaran et al. (2008) proposed the random survival forest method, which is constructed using a committee of many survival trees based on logrank statistics or Harrell’s C index. Ridgeway (2008) extended the multivariate additive regression trees (MART) method using the framework of the generalized linear model. Since these ensemble methods construct models having a “black box” nature, the models are difficult to intercept. Friedman and Popescu (2008) proposed the rule ensemble method, in which nodes of tree models are used as base learners. In this paper, we propose the newly developed rule ensemble method to analyze survival data, namely the survival rule ensemble method. The usefulness of the survival rule ensemble method is illustrated by a practical example in oncology data. By carrying out small-scale simulations, we found that the survival rule ensemble method showed better predictive performance compared with existing ensemble methods for survival outcomes.
