Statistical Analysis Methods under the Missing Not At Random Assumption

Abstract

This paper gives an overview of statistical methods used under the Missing Not At Random (MNAR) assumption, with a particular focus on those utilized in drug development. As the missing data mechanism is generally not testable from the observed data, even when we adopt the Missing At Random (MAR) assumption in the primary analysis, we cannot rule out the possibility of the true mechanism being MNAR. Furthermore, the effect of model misspecification may be greater for analyses under the MNAR assumption than those under MAR. Thus careful consideration is needed before analysis under MNAR is specified as the primary analysis. In this paper, we review statistical models and parameter estimation methods under the MNAR assumption from the viewpoint of both primary analysis and sensitivity analysis. For the statistical models, the Selection Model (SM), Shared Parameter Model (SPM), and Pattern-Mixture Model (PMM) are introduced; for the parameter estimation, methods based on maximization of the observed likelihood and those based on Multiple Imputation (MI) techniques are reviewed. In addition to the theoretical descriptions, we also introduce selected references to procedures and publicly available macro programs for the statistical software SAS.