Sequential regression multiple imputation for incomplete multivariate data using Markov Chain Monte Carlo

Abstract

This paper discusses the theoretical background to handling missing data in a multivariate context. Earlier methods for dealing with item non-response are reviewed, followed by an examination of some of the more modern methods and, in particular, multiple imputation. One such technique, known as sequential regression multivariate imputation, which employs a Markov chain Monte Carlo algorithm is described and implemented. It is demonstrated that distributional convergence is rapid and only a few imputations are necessary in order to produce accurate point estimates and preserve multivariate relationships, whilst adequately accounting for the uncertainty introduced by the imputation procedure. It is further shown that lower fractions of missing data and the inclusion of relevant covariates in the imputation model are desirable in terms of bias reduction.