Abstract:
|
Multiple imputation (MI) for missing data is often used in such a way that the imputation and analysis models are 'uncongenial'. A topical example is that of control or reference based imputation, which is increasingly being used in the analysis of randomized trials as part of a missing not at random (MNAR) sensitivity analysis. In this (and other) case(s), Rubin's variance estimator is biased upwards relative to the repeated sampling variance of the MI point estimator. In this talk I investigate the performance of a recent proposal of von Hippel for bootstrap inference for MI, which is computationally efficient and unbiased under uncongeniality. Moreover, I argue against recent suggestions that use of Rubin's rules variance estimator is preferable to frequentist valid variance estimators, such as the bootstrap, in the setting of MNAR sensitivity analysis of randomized trials.
|