Fall Seminar Series - October 9, 2003
University of Minnesota
School of Statistics
College of Liberal Arts

Model Selection and Model Combining in Regression

Yuhong Yang
Department of Statistics
Iowa State University

Thursday, October 9, 2003
4:00 PM, 115 Ford Hall
Minneapolis, East Bank Campus
Social at 3:30 PM, 300 Ford Hall

Abstract

  There are a number of model selection criteria. In general, they work well under different conditions or from different considerations. For example, AIC is minimax-rate adaptive in terms of estimating the regression function for both parametric and nonparametric settings and BIC is consistent in terms of selecting the true model in a parametric setting. There are several successful efforts in constructing more sophisticated model selection methods to share the strengths of AIC and BIC in certain sense. However, we show that the aforementioned essential properties of AIC and BIC cannot be combined by any model selection procedure.

In recent years, methods have been proposed to combine models. We compare model selection and model combining methods. The results support the view that when model selection instability is low, model selection methods should be preferred and when model selection instability is high, model combining produces much more accurate estimators of the regression function.