ABSTRACT
This chapter describes a range of practical approaches for sensitivity analysis via multiple imputation. Following a general introduction, Section 19.2 briefly reviews the theory underlying the analysis of data when missing values are NMAR, in particular focusing on the contrast between the pattern-mixture and selection model approaches. This has been dealt with in some detail in earlier chapters, especially 4 and 16. Section 19.3 describes a pattern mixture approach, illustrating its application with missing covariate and survival data. Section 19.4 targets specific issues raised by longitudinal data in clinical trials describing the “Δ-method” (Section 19.4.3) and “reference-based imputation” approach (Section 19.4.5). The latter has been recently proposed by Carpenter et al. (2013); see also Mallenckrodt (2013) and O’Kelly and Ratitch (2014, Ch. 7). These approaches are very flexible and prac-
advantage being that they avoid direct estimation of an NMAR model. They are contrasted in Section 19.5 with a selection model formulation. Again, direct estimation of an NMAR model is avoided, here through reweighting of the imputations from the MAR model. We conclude with a brief discussion in Section 19.6.
