Calculating the Appropriate Information Matrix for Log-linear Models When Data Are Missing at Random

Abstract

It is commonly assumed that likelihood based inferences are valid when data are missing at random. In his original work on this topic, Rubin defined precisely the extent to which this statement holds. In particular, the observed but not the expected information matrix can be used for frequentist inference. In the rapidly growing literature on this subject, this fact is not always appreciated. An illustration is given, in the setting of the log-linear model for correlated binary data.

G. Molenberghs is Assistant Professor, Biostatistics, Limburgs Universitair Centrum, Belgium. M.G. Kenward is a Reader in the Institute of Mathematics and Statistics, The University of Kent, UK