A comparison of underlying prediction models of missing data when using EM algorithm in model selection

Abstract

The Akaike information criterion is a frequently used diganostic to select one or a few regression models from a set of candidate models. An additional and often neglected model selection problem is introduced if part of the sample is missing and the expectation-maximization algorithm is used to obtain maximum likelihood estimates, namely, the choice of the assumed underlying data model. In this paper we discuss two assumed underlying data models that can be employed to deal with missing data in the context of model selection. Simulations show the difference in performance of both models in a variety of settings.