A1752
Title: Variable selection in distributional time-to-event models with dependent and administrative censoring
Authors: Annika Stroemer - Department of Medical Biometry and Statistics, University of Marburg (Germany) [presenting]
Nadja Klein - Karlsruhe Institute of Technology (Germany)
Ingrid Van Keilegom - KU Leuven (Belgium)
Andreas Mayr - Marburg University (Germany)
Abstract: A standard assumption in survival analysis is that the censoring time C and the survival time T are conditionally independent given the covariates. This assumption holds for administrative censoring but is often violated in medical studies, where censoring may depend on the event time, for example when patients withdraw from a trial due to deteriorating health, causing sicker patients to be censored earlier. A model-based boosting approach via distributional copula regression is proposed that accounts for dependent censoring and the heterogeneity of the censoring mechanism. Administrative censoring is treated as independent of the event time, while dependent censoring is jointly modeled with the survival time through a parametric copula. All distribution parameters, including the covariate-dependent copula parameter, are estimated via component-wise gradient boosting. Boosting remains feasible even for high-dimensional data with p > n and enables data-driven variable selection across all distributional components. Performance is investigated in a simulation study covering scenarios with varying proportions of administrative versus dependent censoring, different censoring rates, and varying levels of dependence between T and C, with illustration using a biomedical example.
