Targeted Maximum Likelihood Estimation for Causal Inference in Observational Studies.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Observational data analysis for causal effects is increasingly common in epidemiology.
• Propensity score methods and G-computation are widely used, but Targeted Maximum Likelihood Estimation (TMLE) presents a robust alternative.
• TMLE is a doubly robust, maximum-likelihood method with a bias-variance optimizing "targeting" step.

Purpose of the Study:

• To provide an accessible guide to TMLE for applied epidemiologic researchers.
• To detail step-by-step instructions for using TMLE to estimate average treatment effects in observational studies.
• To compare TMLE with G-computation and inverse probability weighting using simulation.

Main Methods:

• Targeted Maximum Likelihood Estimation (TMLE) for causal effect estimation.
• Comparison with G-computation and inverse probability weighting.
• Simulation study assessing performance under parametric regression misspecification, utilizing super learning (ensemble machine learning).

Main Results:

• TMLE demonstrated its "doubly robust" property in simulations, performing well even with misspecified parametric models.
• Ensemble machine learning algorithms (super learning) integrated with TMLE showed superior performance compared to traditional parametric regression.
• TMLE provides a valuable, statistically sound alternative for causal inference in observational epidemiology.

Conclusions:

• TMLE is a powerful and flexible method for estimating causal effects from observational data.
• The integration of machine learning, particularly super learning, enhances TMLE's performance and robustness.
• This work facilitates wider adoption of TMLE in epidemiologic research.