Chen-Pin Wang
University od Texas Health Science Center at San Antonio

Goutis-Robert Kullback-Leibler Divergence in Generalized Linear Models and Beyond: Applications to Studies in Type 2 Diabetes

This talk considers the Kullback-Leibler distance (KLD) by Goutis and Robert (1998) originally proposed for comparing nested generalized linear models. We derive the asymptotic properties of this KLD under certain regularity conditions where neither models in comparison is required to be the true model. We also examine the impact of this asymptotic property when the regularity conditions are not completely satisfied. Furthermore, we establish the connection between Goutis and Robert 's KLD and a weighted posterior predictive p-value (WPPP). Finally, we apply both KLD and WPPP to compare models using simulation studies as well as two type 2 diabetes studies, where only part of the regularity conditions were met.