B0906
Title: Boosting mixtures of distributional regression models
Authors: Tobias Hepp - University of Erlangen-Nuremberg (Germany) [presenting]
Elisabeth Bergherr - Georg-August-Univerität Göttingen (Germany)
Abstract: Mixture regression models provide a useful modeling framework in the context of data with unobserved heterogeneity that aims to identify latent components that differ in terms of their dependence on one or more covariates. In order to estimate the unknown model parameters, an algorithm based on the component-wise gradient boosting methodology is introduced. Compared to alternative strategies such as the expectation-maximization algorithm or mixture density networks, component-wise boosting does not require the dependence structure to be completely specified in advance while still remaining fully interpretable in terms of the base-learners used. A first version of the algorithm is demonstrated on a laboratory dataset for hemoglobin values and performs on par with alternative strategies. In addition, an outlook on variable selection performance is given using simulated data.
