Nonparametric Selection Method of Survival Predictors with an Application to Breast Cancer

Mutual information between a survival variable and covariables provides a new tool for selecting covariables with a high predictive value whenever there is no reasonable parametric model with respect to the observed phenomena. The information rate carried out by covariables can be tested by means of a decomposition similar to the analysis of variance. Moreover, a method based on information conservation can be used for aggregating survival curves corresponding to different modalities of the same selected predictor which increases the prediction efficiency. These results are applied to survival data from 1304 patients with breast cancer followed over a period of ten years.