A bootstrap method for assessing uncertainty in Kullback-Leibler discrepancy model selection problems

Abstract

The goal of a model selection problem is to determine which model from a candidate collection best approximates the true or generating model, where separation between the true model and an approximating model is measured by a discrepancy function. The contribution of the current paper is to provide a unique approach to assessing the uncertainty inherent to a model selection problem. We define bootstrap weights as a calculation of the probability on each model in the candidate collection being closest to the true model, as measured by the Kullback-Leibler discrepancy. Examples are presented to show that these bootstrap weights provide new and important information to aid model selection practitioners