TREE-STRUCTURED PREDICTION FOR CENSORED SURVIVAL-DATA AND THE COX MODEL

Prediction trees for the analysis of survival data are discussed. It is shown that trees are useful not only in summarizing the prognostic information contained in a set of covariates (prognostic classification), but also in detecting and displaying treatment-covariates interactions (subgroup analysis). The RECPAM approach to tree-growing is outlined; prognostic classification and subgroup analysis are then formulated within the RECPAM framework and on the basis of the Cox proportional hazards models with apriori strata. Two examples of data analysis are presented. The issue of cross-validation is discussed in relation to computationally cheaper model selection criteria.