B0691
Title: Boosting joint models
Authors: Andreas Mayr - University of Bonn (Germany) [presenting]
Elisabeth Bergherr - Georg-August-Univerität Göttingen (Germany)
Abstract: Joint models are commonly estimated via likelihood based expectation maximization or Bayesian approaches. Drawbacks of these frameworks are that they do not allow for automated variable selection and are not feasible in high-dimensional data situations. We propose a predictor-wise gradient boosting algorithm overcoming these shortcomings being able to simultaneously estimate and select predictors for joint models in potentially high-dimensional data. The new algorithm cycles through the different additive predictors and applies simple base-learners for each candidate variable while including in each iteration only the best-performing base-learner. When the algorithm is stopped before convergence, base-learners that have never been selected are effectively excluded from the final model, leading to data-driven variable selection.
Journal