Machine learning in causal inference for epidemiology.

Area of Science:

• Causal inference
• Epidemiology
• Machine learning applications in statistics

Background:

• Traditional parametric models in causal inference are susceptible to bias due to incorrect model specification, particularly in high-dimensional data.
• Machine learning (ML) methods can reduce misspecification bias by avoiding assumptions on functional forms, but direct integration risks 'plug-in bias'.

Purpose of the Study:

• To provide an overview of state-of-the-art estimators that leverage Machine Learning for robust causal inference.
• To address the challenge of bias in causal effect estimation, especially in high-dimensional settings.

Main Methods:

• Review of advanced causal inference estimators combining Machine Learning predictive power with statistical inference capabilities.
• Focus on Targeted Maximum Likelihood Estimation (TMLE), Augmented Inverse Probability Weighting (AIPW), and Double/Debiased Machine Learning (DML).

Main Results:

• These advanced estimators aim to overcome 'plug-in bias' associated with directly using ML predictions in causal effect formulas.
• They offer improved asymptotic properties for reliable causal effect estimation.

Conclusions:

• TMLE, AIPW, and DML represent the current state-of-the-art for epidemiologists seeking to utilize ML in causal inference.
• These methods mitigate bias from model misspecification and enhance the reliability of causal effect estimates.