Online Program Home
My Program

Abstract Details

Activity Number: 569 - Theory and Practice for Addressing Asymmetric Measures in Statistical Modeling
Type: Topic Contributed
Date/Time: Wednesday, August 1, 2018 : 2:00 PM to 3:50 PM
Sponsor: WNAR
Abstract #330861
Title: Model Selection Criteria Based on Symmetrized Variants of Asymmetric Divergence Measures
Author(s): Joseph Cavanaugh*
Companies: University of Iowa
Keywords: Akaike information criterion; Mallows' conceptual predictive statistic; variable selection

A model selection criterion is often formulated by constructing an estimator of a divergence measure. In the context of model selection, a divergence is a functional that reflects the disparity between two distributions: one corresponding to the fitted candidate model and the other corresponding to the generating model. Many divergence measures are asymmetric, meaning that an alternate measure is obtained by reversing the roles of the two distributions. As a consequence of this asymmetry, the divergence may differentially assess underspecified and overspecified models. Model selection criteria that estimate such measures may be excessively prone to favoring underfitted or overfitted models, in a manner that violates the philosophy of Occam's razor and the spirit of achieving an optimal bias/variance tradeoff. By symmetrizing such a divergence, and formulating model selection criteria based on the resulting measure, we can often obtain criteria that achieve a more desirable balance in terms of their selection patterns. We illustrate this general methodology through the development and investigation of variants of the Akaike information criterion and Mallow's Cp.

Authors who are presenting talks have a * after their name.

Back to the full JSM 2018 program