B1095
Title: Boosting robust distributional regression
Authors: Jan Speller - University of Bonn (Germany)
Christian Staerk - RWTH Aachen University (Germany) [presenting]
Francisco Gude - Complexo Hospitalario Universitario de Santiago de Compostela (Spain)
Andreas Mayr - University of Bonn (Germany)
Abstract: With the increasing complexity and dimensionality of datasets, it is crucial that statistical approaches are robust against the influence of potentially corrupted observations. A flexible distributional regression approach is proposed that is robust towards outliers in the response variable for generalized additive models for location, scale and shape (GAMLSS). A recently proposed robustification of the log-likelihood is incorporated into the framework of gradient boosting, which is based on trimming low log-likelihood values via a log-logistic function to a boundary value. A data-driven quantile-based choice of the robustness constant is considered and its influence is investigated in a simulation study for low- and high-dimensional data situations. The application of the robust distributional regression approach is illustrated in diverse biomedical data examples, including the modelling of thyroid hormone levels, spatial modelling for functional magnetic resonance brain imaging and high-dimensional modelling of gene expression data for cancer cell lines.