Recursive estimation procedures for missing-data problems

Abstract

SUMMARY Titterington (1984) proposes recursive methods for dealing with incomplete data. The present paper concentrates on versions of these for multiparameter problems involving missing data. Theorems are outlined from which asymptotic properties of the recursive procedures can be established and versions of the recursions, some of which are particularly easy to use, are written down for problems in which the missing data are missing at random. After illustration with exponential family models, the case of multivariate normal data is considered in detail. Numerical comparisons of the various methods are obtained using bivariate normal data. Application of the methods as imputation procedures is discussed.