Non-Nested Models and the Likelihood Ratio Statistic: A Comparison of Simulation and Bootstrap Based Tests

Abstract

We consider an alternative use of simulation in the context of using the Likelihood-Ratio statistic to test non-nested models. To date simulation has been used to estimate the Kullback-Leibler measure of closeness between two densities, which in turn 'mean adjusts' the Likelihood-Ratio statistic. Given that this adjustment is still based upon asymptotic arguments, an alternative procedure is to utilise bootstrap procedures to construct the empirical density. To our knowledge this study represents the first comparison of the properties of bootstrap and simulation-based tests applied to non-nested tests. More specifically, the design of experiments allows us to comment on the relative performance of these two testing frameworks across models with varying degrees of nonlinearity. In this respect although the primary focus of the paper is upon the relative evaluation of simulation and bootstrap-based nonnested procedures in testing across a class of nonlinear threshold models, the inclusion of a similar analysis of the more standard linear/log-linear models provides a point of comparison.