Title: Robust joint modeling of mean and dispersion for GLMs
Authors: Pieter Segaert - KU Leuven (Belgium)
Stefan Van Aelst - University of Leuven (Belgium)
Tim Verdonck - UAntwerp, KU Leuven (Belgium) [presenting]
Abstract: Generalized linear models form a unified way of modeling the mean response under different distributions belonging to the exponential family. Because of their flexibility, they have become a powerful tool in statistics. Real data often show a larger or smaller variability than expected from the model and the dispersion may even change for different observations in the data. It is crucial to properly account for this dispersion. A typical problem in analysing real data is the possible presence of outliers in the data. As classical methods try to fit an optimal model for all observations they are highly susceptible to these atypical observations. Therefore we propose a robust procedure for jointly modeling the mean and dispersion under the GLM framework. Our robust double exponential estimator models both mean and dispersion behaviour based on a possibly different set of predictors. The good performance of our methodology is illustrated in a simulation study and on real data.