Adaptive model selection



Xiaotong Shen

The Ohio State University


Model selection becomes increasingly important for analyses involving a large number of candidate models. In the literature, most model selection procedures use a fixed penalty penalizing an increase in the size of a model, including Akaike Information Criterion (AIC), Mallows's $C_p$, Risk Inflation Criterion (RIC) and Bayesian Information Criterion (BIC). These non-adaptive selection procedures perform well only in one type of situations but not across a variety of situations. In this talk, we will present an adaptive model selection procedure, based on a data-driven complexity penalty defined by the concept of generalized degrees of freedom. The proposed procedure approximates the best performance of a class of model selection procedures, across a variety of different situations. This class includes many well-known procedures such as AIC, $C_p$, RIC and BIC. The proposed procedure is applied to variable selection in least squares regression and wavelet thresholding in nonparametric regression. Simulation results and asymptotic analysis support the effectiveness of the proposed procedure.

Based on the joint work with Jimmy Ye.

Back to statistics seminar home page