Abstract:
|
In many analyses it seems very realistic that the missing data follow a Missing Not at Random (MNAR) mechanism. In this case, standard implementations of multiple imputation (MI), which assume Missing at Random (MAR), may not be sufficient and may lead to biased estimates. We will present an approach to multiply impute non-ignorable and hierarchical binary missing data in the framework of Fully Conditional Specification (FCS). The suspected MNAR mechanism will be considered and modeled during the imputation process by applying a censored bivariate probit model as imputation model. For allowing the consideration of a present multilevel structure in the data during the imputation process, the model is expanded by a random intercept term in both equations. In order to assess the performance of this imputation technique, different simulation studies were conducted. The method performed well in all considered situations in terms of coverage and bias and clearly outperforms alternative MAR methods in MNAR settings. In addition, the robustness towards the choice of a suitable exclusion criterion - a crucial condition for the performance of the approach - is investigated more in detail.
|