Abstract:
|
The fundamental non-verifiable nature of the missing-data mechanism is discussed, in selection, pattern-mixture, and shared-parameter frameworks. It is shown how, in all three frameworks, MAR can be formulated to provide a useful analysis starting point. From there, parametric, semi-parametric, and non-parametric sensitivity analysis routes are discussed. It is indicated how non-parametric bounds can be usefully considered to provide a background for (semi-)parametric sensitivity analyses.
|