Targeted maximum likelihood based causal inference: Part I.

Area of Science:

• Causal inference
• Statistical modeling
• Epidemiology

Background:

• Estimating causal effects from observational data is crucial for understanding interventions.
• G-computation formula defines counterfactual distributions under causal graph assumptions.
• Marginal structural models are used for nonparametric estimation of causal effects.

Purpose of the Study:

• To develop a targeted maximum likelihood estimator (TMLE) for causal effects of multiple time point interventions.
• To enhance the robustness and accuracy of causal effect estimation in complex scenarios.
• To provide a foundation for implementing TMLE in practical causal inference problems.

Main Methods:

• Utilizes G-computation formula derived from causal graph interventions.
• Employs loss-based super-learning for initial estimation of G-computation components.
• Applies iterative targeted maximum likelihood updating with optimal fluctuation functions for robustness.

Main Results:

• The proposed TMLE is double robust: consistent if either initial or fluctuation function estimators are consistent.
• Correct specification of conditional distributions (treatment and censoring mechanisms) is key for the optimal fluctuation function.
• Selection among TMLEs can be guided by loss-based cross-validation methods.

Conclusions:

• The developed TMLE offers a robust and efficient approach for estimating causal effects of complex interventions.
• This methodology provides a strong basis for advanced causal effect estimation, detailed in a companion article.
• The iterative updating strategy enhances statistical properties, making it suitable for time-dependent treatments.