Causal Effect Estimation With TMLE: Handling Missing Data and Near Violations of Positivity

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Missing data is a common challenge in epidemiological studies.
• Positivity violations can bias treatment effect estimates.
• Targeted maximum likelihood estimation (TMLE) is a robust statistical method.

Purpose of the Study:

• Evaluate TMLE performance for average treatment effect estimation with missing data.
• Assess robustness against positivity violations and various missing data mechanisms.
• Compare non-multiple imputation (non-MI) and multiple imputation (MI) approaches.

Main Methods:

• Model- and design-based simulations using the WASH Benefits Bangladesh dataset.
• Considered five missingness-directed acyclic graphs, including not-at-random missingness.
• Compared eight missing data methods with TMLE, including parametric and machine learning MI models.

Main Results:

• Complete cases with TMLE and an outcome-missingness model showed the lowest bias.
• This approach demonstrated greater robustness against positivity violations.
• Multiple imputation with classification and regression trees (CART) yielded lower root mean squared error and maintained nominal coverage.

Conclusions:

• Non-MI methods, specifically complete cases with TMLE, are preferable for bias reduction.
• MI CART is recommended when accurate confidence intervals are the priority.
• Findings highlight the trade-offs between bias and coverage in missing data analysis.