The estimation of multivariate normal density functions using incomplete data

This paper compares the estimative and predictive methods of estimating a multivariate normal density, and shows that in terms of a goodness-of-fit criterion based on an information measure, the predictive method is to be preferred for small sample sizes. We then extend the predictive method to certain nested patterns of missing data, and consider various approximations to the predictive method for more general patterns of missing data.