Impact of Inconsistent Imputation Models in Mediation Analysis with Clustered Data

The choice of imputation model is often driven by the focus of the post-imputation analyses. However, this can be impractical especially when multiple imputation (MI) inference is used in large-scale surveys or in applications with complex data structures. This work investigates the impact of the imputation model on mediation analysis with clustered data. Specifically, we consider joint and variable-by-variable imputation models prior to a 1-1-1 multilevel mediation analysis. We provide theoretical and analytical assessment of the bias under each imputation method. A comprehensive simulation study is conducted to understand the performance of imputation methods in a repetitive sampling framework. The choice of the selected imputation model does not have an influence on the multilevel mediation analysis.