On single versus multiple imputation for a class of stochastic algorithms estimating maximum likelihood (Q1424618)

This paper aims to clarify certain distinctive advantages of the stochastic version of the expecta\-tion-maximization (SEM) algorithm which uses a single imputation in its E-step. Advantages of SEM in non-exponential family applications is highlighted and comparison of its performance with other multiple imputation schemes is carried out. The author proves an ergodicity result showing that the convergence properties of the stochastic algorithm does not dependent upon whether single or multiple imputation is used. The result provides a basis for applying the stochastic EM algorithm to situations where full exploration of the posterior distributions of model parameters is either computationally too expensive or not necessary.