ABSTRACT
While likelihood, Bayesian, and semi-parametric methods, under the assumption of MAR, have been embraced as primary analyses for incomplete data (Little et al. 2010), all models make assumptions about the so-called predictive distribution, i.e., the distribution governing the missing data, given the observed ones. Such assumptions are by their very nature unverifiable from the data, while they may have an impact on the inferences drawn. It is therefore imperative to explore how sensitive the conclusions drawn are to the unverifiable assumptions. In this chapter, the focus is on likelihood-based methods. In subsequent chapters, sensitivity analysis in the context of semi-parametric methods (Chapter 17), Bayesian approaches (Chapter 18), and multiple imputation (Chapter 19) is discussed. Many sensitivity analyses will take the form of exploring how deviations from MAR towards NMAR change the conclusions.
