ABSTRACT
This chapter covers many of the central concepts and modeling approaches for dealing with nonignorable missingness and dropout. A prevailing theme is the factorization of the full-data model into the observed data distribution and the extrapolation distribution, where the latter characterizes assumptions about the conditional distribution of missing data given observed information (observed responses, missing data indicators, and covariates). The extrapolation factorization (as we refer to it) factors the full-data distribution into its identified and nonidentified components. We argue that models for nonignorable dropout should be parameterized in such a way that one or more parameters for the extrapolation distribution is completely nonidentified by data. These also should have transparent interpretation so as to facilitate sensitivity analysis or incorporation of prior information about the missing data distribution or missing data mechanism.
