A0339
Title: A model-based boosting approach to deal with dependent censoring
Authors: Annika Stroemer - University of Bonn (Germany) [presenting]
Nadja Klein - Karlsruhe Institute of Technology (Germany)
Andreas Mayr - University of Bonn (Germany)
Abstract: A popular model to study the effect of covariates on survival times is the semiparametric proportional hazards model. Estimation in this model is well-established for common right-censored data, assuming independence between survival and censoring time given the covariates. This assumption is mainly held when censoring occurs at the end of the study. However, in medical studies, this assumption is questionable. For example, if a patient's health deteriorates and they choose to withdraw from the trial due to poor prognosis, censoring time depends on their health status. This leads to dependent censoring, as patients with poorer health are more likely to be censored earlier. A model-based boosting approach is proposed to deal with dependent censoring using distributional copula regression. This allows modeling the joint distribution of survival and censoring times by linking appropriately marginal distributions through a parametric copula. Rather than assuming known marginals, all distribution parameters are estimated simultaneously as functions of covariates. A key merit of the boosting approach compared to classical estimation frameworks is that estimation is feasible for high-dimensional data. Additionally, the boosting algorithm includes data-driven variable selection. An extensive simulation study is conducted, and its practical application is illustrated with a biomedical example.
