Resampling for Order Estimation of Autoregressive Models with Missing Data

Abstract

In this article, we consider the order estimation of autoregressive models with incomplete data using the expectation–maximization (EM) algorithm-based information criteria. The criteria take the form of a penalization of the conditional expectation of the log-likelihood. The evaluation of the penalization term generally involves numerical differentiation and matrix inversion. We introduce a simplification of the penalization term for autoregressive model selection and we propose a penalty factor based on a resampling procedure in the criteria formula. The simulation results show the improvements yielded by the proposed method when compared with the classical information criteria for model selection with incomplete data.

Keywords:

• Autoregressive model
• EM algorithm
• Information criteria
• Missing data
• Resampling

Mathematics Subject Classification:

• 62M10
• 62F07