A0336
Title: Boosting distributional copula regression for right-censored bivariate time-to-event data
Authors: Guillermo Briseno Sanchez - TU Dortmund University (Germany) [presenting]
Nadja Klein - Karlsruhe Institute of Technology (Germany)
Andreas Mayr - University of Bonn (Germany)
Andreas Groll - Technical University Dortmund (Germany)
Abstract: A highly flexible distributional copula regression model is proposed for bivariate right-censored time-to-event data. The joint survival function is constructed using parametric copulas, allowing for separate specifications of the dependence structure between the time-to-event variables and their respective marginal survival distributions. The latter can be specified using well-known parametric distributions, such as log-normal, log-logistic, or Weibull distributions. These results were then in parametric (accelerated failure time, AFT) models for the respective univariate responses. By embedding the model into the framework of generalised additive models for location, scale, and shape (GAMLSS), all parameters of the joint distribution can be modeled as a function of covariates. Using additive predictors thereby enables the account of linear, non-linear, or spatial effects in modelling the dependence structure and the respective marginal distributions. Estimation is proposed by means of component-wise gradient-based boosting to allow for data-driven variable selection. The latter not only renders model building feasible and avoids the manual comparison of different model specifications but also allows the tackling of high-dimensional (p >> n) data structures. To the best of knowledge, this is the first implementation of multivariate AFT models via distributional copula regression and automatic variable selection via statistical boosting.
