B0641
Title: A novel gradient boosting framework for generalized additive mixed models
Authors: Lars Knieper - Georg-August-University of Goettingen (Germany) [presenting]
Elisabeth Bergherr - Georg-August-Univerität Göttingen (Germany)
Torsten Hothorn - University of Zurich (Switzerland)
Nadia Mueller-Voggel - University of Erlangen-Nuremberg (Germany)
Colin Griesbach - Georg-August-University Goettingen (Germany)
Abstract: Mixed models are usually fitted based on the penalised likelihood, while model-based boosting offers a fast and intuitive alternative which additionally enables variable selection and stable performance in high dimensional data. For this purpose, the well-known R-package "mboost" was equipped with a random effects base learner in order to estimate generalised additive mixed models within the framework of component-wise gradient boosting. However, this approach tends to produce biased estimates in the presence of cluster-constant covariates and in addition lacks any parameter estimation for the random components. The new proposed "mermboost" algorithm incorporates a correction step and a separation of estimating fixed and random effects. The latter ensures a removal of competition between fixed and random effects. Both adjustments result not only in unbiased estimates for cluster constant fixed effects but also in unbiased random effects with a reasonable estimate of their covariance. While simulated data with a Poisson distributed response give convincing results, for Bernoulli data a large shrinkage is observed, especially in the random structure. This powerful boosting approach "mermboost" is available as an add-on R-package for "mboost" and enables well-performing estimation of flexible mixed models based on gradient boosting.
