Workshop on Structure Adapting Methods - Abstract

Sperlich, Stefan

An adaptive model class: sliding from fixed to random effects models

In some applications, such as small area studies, using a fixed effect for each region leads to models that are flexible but that have poor estimation accuracy; they are over-parametrized. Regarding region as a random effect reduces the number of parameters, and hence the flexibility, but this introduces assumptions that may be violated, such as that of independence between covariates and the random effects. The proposed class of models constitutes a continuum of models, indexed by a slider, that determines the position of the model between these two extremes. So one can choose a model that is close to the parsimonious random effects case, but far enough away from it to filter out unwanted dependencies.