Introduction

Although the assumption of MAR is often reasonable in clinical trials, the possibility of MNAR data is impossible to rule out (Verbeke and Molenberghs, 2000). Therefore, analyses valid under MNAR are needed. Analyses in the MNAR framework try in some manner to model or otherwise take into account the missingness process. However, moving beyond MAR to MNAR poses fundamental problems.

In MAR it is assumed that the statistical behavior of the unobserved data is the same as if it had been observed, such that the unobserved data can be predicted from the observed data. As was noted in Section 8.6, the fundamental difficulty with any MNAR method is that the characteristics and statistical behavior of the missing data are unknown.

The inescapable fact is that moving beyond MAR to MNAR can only be done by making assumptions. Conclusions from MNAR analyses are therefore conditional on the appropriateness of the assumed model. While dependence on assumptions is not unique to MNAR analyses, a unique feature with MNAR analyses is that (some of) the assumptions are not testable (Molenberghs, Kenward, and Lesaffre, 1997) because the data about which the assumptions are made are missing (Laird, 1994). Importantly, the consequences of model misspecification are more severe with MNAR methods than with other (e.g.,MAR)methods (Little, 1995; Laird, 1994; Rubin, 1994; Draper, 1995, Kenward, 1998).