A0598
Title: Nonnested model selection based on empirical likelihood ratio
Authors: Jiancheng Jiang - UNC Charlotte (United States) [presenting]
Abstract: An empirical likelihood ratio (ELR) test is proposed for nonparametric model selection, where the competing models may be nested, nonnested, overlapping, misspecified, or correctly specified. It compares the prediction performances of models based on the cross-validation and allows for heteroscedasticity of the errors. We develop its asymptotic distributions for comparing any two supervised learning models under a general framework with convex loss functions. However, for general loss functions, the prediction errors from the cross-validation involve repeatedly fitting the models with one observation held out. An easily implemented approximation is then introduced. It is shown that the approximated test shares the same asymptotics as the original one. We apply the proposed tests to compare additive models and varying-coefficient models. Furthermore, a distributed ELR test is proposed to test the importance of a group of variables in possibly misspecified additive models with massive data, and a fast calculation procedure for the test is introduced. Simulations show that the proposed tests work well and have favorable finite sample performance over some existing approaches. The methodology is validated on an empirical application.
