Sensitivity Analysis for Selection Models

In the previous chapters, it was indicated on various occasions (see Sections 15.4, 16.1, 16.5, 17.1, and 17.2) that incomplete longitudinal data pose specific challenges related to sensitivity to modeling assumptions. Even when the linear mixed model would beyond any doubt be the choice of preference to describe the measurement process should the data be complete, then the analysis of the actually observed, incomplete version is subject still to further untestable modeling assumptions. The terminology which is useful to this end has been reviewed in Chapter 15.

The methodologically simplest case is discussed in Chapter 16, where it is assumed that the missing data are MCAR. Simple techniques such as a complete case analysis, simple forms of imputation, and so forth may be advised in some cases. However, the MCAR assumption is a strong one and made too often in practice. Thus, simple forms of analysis are certainly too common in applied statistical practice.