Handling missing data when estimating causal effects with Targeted Maximum Likelihood Estimation

Targeted Maximum Likelihood Estimation (TMLE) is increasingly used for doubly robust causal inference, but how missing data should be handled when using TMLE with data-adaptive approaches is unclear. Based on Victorian Adolescent Health Cohort Study data, we conducted a simulation study to evaluate missing data methods when using TMLE. Methods compared were complete-case analysis, extended TMLE incorporating outcome-missingness model, a "missing covariate missing indicator" method, and five multiple imputation (MI) approaches using parametric or machine-learning imputation models. We considered simple (exposure/outcome generated from main-effects regressions) and complex (models also included interactions) scenarios, alongside two missingness mechanisms (different in whether the outcome influenced missingness in other variables). Parametric MI including interaction terms performed best across all settings in terms of bias and variance reduction. Non-MI methods had the largest variance and, like MI using machine-learning, had larger biases when the outcome influenced missingness in other variables. Parametric MI without interactions - the default in most MI software - had the largest bias in the complex scenario. When choosing a method to handle missing data in the context of TMLE with data-adaptive approaches, researchers must consider the missingness mechanism. Parametric MI that incorporates interactions and non-linearities appears to be the best approach.