How Ignorable Is Missing At Random ?

Abstract

For over two decades, following the pioneering work of Rubin (1976) and Little (1976), there has been a growing literature on incomplete data, with a lot of emphasis on longitudinal data. Following the original work of Rubin and Little, there has evolved a general view that “likelihood methods” that ignore the missing value mechanism are valid under an MAR process, where likelihood is interpreted in a frequentist sense. The availability of flexible standard software for incomplete data, such as PROC MIXED, and the advantages quoted in Section 17.3 contribute to this point of view. This statement needs careful qualification however. Kenward and Molenberghs (1998) provided an exposition of the precise sense in which frequentist methods of inference are justified under MAR processes.