Title: Variable selection and allocation in joint models via gradient boosting techniques
Authors: Colin Griesbach - Georg-August-University Goettingen (Germany) [presenting]
Andreas Mayr - University of Bonn (Germany)
Elisabeth Bergherr - Georg-August-Univerität Göttingen (Germany)
Abstract: Modelling longitudinal data and risk for events separately, even though the underlying processes are related to each other, leads to loss of information and bias. Hence, the popularity of joint models for longitudinal and time-to-event data has grown rapidly in the last few decades. Gradient boosting is a statistical learning method that has the inherent ability to select variables and estimate them at the same time. We construct a data-driven allocation algorithm for basic joint models by applying gradient boosting. Instead of specifying beforehand which covariate has an influence on which part of the joint model, the algorithm allocates the covariates to the appropriate sub-model. A simulation study shows that this is possible when using a non-cyclic updating scheme for the boosting algorithm. In addition, recent findings of adaptive step lengths and early stopping based on probing are incorporated in order to improve allocation accuracy and reduce the computational effort.