B1151
Title: Assessing the significance of effects in boosted distributional regression models
Authors: Tobias Hepp - Friedrich-Alexander-Universitaet Erlangen-Nuernberg (Germany) [presenting]
Matthias Schmid - University of Bonn (Germany)
Andreas Mayr - University of Bonn (Germany)
Abstract: Generalized Additive Models for Location, Scale and Shape (GAMLSS) can be also estimated via a gradient boosting algorithm. It provides advantages like automatic variable selection and feasibility in high-dimensional settings with more predictors than observations. However, the implicit regularization that allows the shrinkage of effect estimates prevents the computation of standard errors. As a result, the construction of confidence intervals or significance tests is problematic. We discuss the performance of two potential solutions. The first option is based on permutation tests, where the variable of interest is replaced by regression residuals in order to remove possible correlations with covariates. Another option is to draw parametric bootstrap samples from the conditional distribution of the constrained model without the variable of interest. Then, the differences in the quality of fit between the full model and the constrained one are attributable only to the randomness of these samples and should be less distinctive than for the original data.
Journal