Abstract:
|
Tree-based survival analysis techniques, such as survival tree and survival forest, have been playing an increasingly important role in biomedical research, especially in studies that have a large number of variables and their relations to survival are complicated. However, the critical issue of confounding has not been satisfactorily addressed in these methods. Without appropriate adjustment of the confounding effects, the estimated survivor function could be biased or the important variables identified could be misleading. In this study we propose a confounder adjustment method in tree-based survival analysis. In our method, Kooperberg's hazard regression is employed to estimate the cumulative hazard function, in which only confounders are included as covariates. Martingale residuals are then calculated and used as continuous outcome in the next stage analysis using tree-based method, in which all variables of interest are included. Simulation studies show that our proposed method has good performance in controlling for the confounding effects. The application of the method in the identification of important prognostic factors for a clinical study will be presented.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.