Handling multivariable missing data in causal mediation analysis

Mediation analysis is commonly used in epidemiological research, but guidance is lacking on how multivariable missing data should be dealt with in these analyses. Multiple imputation (MI) is a widely used approach, but questions remain regarding impact of missingness mechanism, how to ensure imputation model compatibility and approaches to variance estimation. To address these gaps, we conducted a simulation study based on the Victorian Adolescent Health Cohort Study. We considered six missingness mechanisms, involving varying assumptions regarding the influence of outcome and/or mediator on missingness in key variables. We compared the performance of complete-case analysis, seven MI approaches, differing in how the imputation model was tailored, and a "substantive model compatible" MI approach. We evaluated both the MI-Boot (MI, then bootstrap) and Boot-MI (bootstrap, then MI) approaches to variance estimation. Results showed that when the mediator and/or outcome influenced their own missingness, there was large bias in effect estimates, while for other mechanisms appropriate MI approaches yielded approximately unbiased estimates. Beyond incorporating all analysis variables in the imputation model, how MI was tailored for compatibility with mediation analysis did not greatly impact point estimation bias. BootMI returned variance estimates with smaller bias than MIBoot, especially in the presence of incompatibility.