G-computation and machine learning for estimating the causal effects of binary exposure statuses on binary outcomes.

Area of Science:

• Clinical Epidemiology
• Biostatistics
• Machine Learning in Healthcare

Background:

• Growing interest in propensity score methods for causal effect estimation in clinical research.
• G-computation offers high statistical power, while machine learning provides robustness to model misspecification.

Purpose of the Study:

• To propose and evaluate a novel approach combining machine learning and G-computation for causal inference.
• Focus on binary outcomes and exposure status, specifically addressing challenges in small sample sizes.

Main Methods:

• Simulation study evaluating penalized logistic regressions, neural networks, support vector machines, boosted trees, and super learner.
• Six scenarios with varying sample sizes, covariates, and relationships were simulated.
• Application to estimate the efficacy of barbiturates in intracranial hypertension.

Main Results:

• The super learner approach within G-computation demonstrated superior performance in minimizing bias and variance, especially in small samples.
• Support vector machines also performed well, though with slightly higher mean bias than the super learner.
• G-computation combined with the super learner proved effective for causal inference with limited data.

Conclusions:

• G-computation integrated with the super learner is a powerful method for robust causal inference, even with small sample sizes.
• This combined approach enhances the reliability of causal effect estimation in clinical research settings.