B1010
Title: Random forests for prediction of treatment effect and treatment group in survival data
Authors: Ricarda Graf - Universitaet Augsburg (Germany) [presenting]
Sarah Friedrich - University of Augsburg (Germany)
Dennis Dobler - TU Dortmund (Germany)
Abstract: Survival data are often encountered in clinical research. Methods for their analysis incorporate censoring times to avoid biased survival estimates. Random survival forests are a technique for the analysis of survival data with independent right-censoring, usually used for risk prediction and especially suitable for high-dimensional data. Risk prediction models help physicians in making personalized decisions. A modification of the random forest algorithm is presented, suitable for making predictions of the relative treatment effect and of the treatment group, respectively, in survival data. One asset of the method is that the splitting criterion and the prediction/classification outcome are well aligned, as they are based on the same statistical object: the Mann-Whitney effect. A bootstrap version of the Mann-Whitney effect based on normalized Kaplan-Meier estimates is used to compute the difference in treatment effects between potential child nodes. The maximum difference serves as the splitting rule. Estimated treatment effects obtained from the modified random forest model and the Cox proportional hazards model are compared through data simulations based on the Athens multicenter AIDS cohort study (AMACS) and examined classification accuracy of the modified RF model. The method's performance in a real-data example is also demonstrated.
