Likelihood based frequentist inference when data are missing at random

One of the most often quoted results from the original work of Rubin and Little on the classification of missing value processes is the validity of likelihood based inferences under missing at random (MAR) mechanisms. Although the sense in which this result holds was precisely defined by Rubin, and explored by him in later work, it appears to be now used by some authors in a general and rather imprecise way, particularly with respect to the use of frequentist modes of inference. In this paper an exposition is given of likelihood based frequentist inference under an MAR mechanism that shows in particular which aspects of such inference cannot be separated from consideration of the missing value mechanism. The development is illustrated with three simple setups: a bivariate binary outcome, a bivariate Gaussian outcome and a two-stage sequential procedure with Gaussian outcome and with real longitudinal examples, involving both categorical and continuous outcomes. In particular, it is shown that the classical expected information matrix is biased and the use of the observed information matrix is recommended.