Biostatistics Seminar Series - Robert L. Strawderman, ScD

Title: Censoring Unbiased Trees and Forests Abstract: Survival trees use recursive partitioning to separate patients into distinct risk groups when some observations are right-censored. Survival forests average multiple survival trees derived from bootstrap sample, leading to more flexible and accurate prediction models. The most popular algorithms for trees and forests in the case of uncensored outcomes rely heavily on the specification of a loss function (e.g., squared error loss) that governs all aspects of the decision-making process. Existing algorithms for censored outcomes typically bear little resemblance to what is used when censoring is absent. We unify the treatment of these algorithms through the development of a class of censoring unbiased loss functions. We discuss some of the properties of these loss functions and associated practical issues, and extend them for use with competing risks. We further demonstrate how these new algorithms can be implemented using existing software. The performance of the resulting methods is evaluated through simulation studies and illustrated using data from RTOG 9410, a randomized trial of patients with locally advanced inoperable non-small cell lung cancer. This talk covers joint work with Jon Steingrimsson, Youngjoo Cho, Liqun Diao, Chen Hu and Annette Molinaro.