Title: An adaptive lasso Cox frailty model for time-varying covariates based on the full likelihood
Authors: Andreas Groll - Technical University Dortmund (Germany) [presenting]
Maike Hohberg - University of Goettingen (Germany)
Abstract: A method is proposed to regularize Cox frailty models that accommodates time-varying covariates and is based on the full likelihood. A particular advantage of this framework is the explicit modeling of the baseline hazard in a nonlinear way, e.g. via P-splines. Linear covariate effects are penalized using the lasso penalty. The estimation is based on a Newton-Raphson algorithm and makes use of local quadratic approximations of the penalty terms. Additionally, adaptive weights are included to stabilize the estimation. The full likelihood model can easily be extended by a wide class of frailty distributions including random intercepts and random slopes. The method is implemented in R in the function coxlasso and will be compared to other packages for regularized Cox regression in both simulation scenarios and a real data application.