Causal Machine Learning Methods and Use of Cross-Fitting in Settings With High-Dimensional Confounding

Area of Science:

• Epidemiology
• Biostatistics

Background:

• Observational studies aim to estimate causal effects but face challenges from high-dimensional confounding.
• Doubly robust methods like AIPW and TMLE offer potential solutions using data-adaptive techniques.

Purpose of the Study:

• To compare the performance of AIPW and TMLE for estimating average causal effects (ACE) in the presence of high-dimensional confounding.
• To evaluate the impact of cross-fitting and Super Learner library size on method performance.

Main Methods:

• Extensive simulation study using an early-life cohort as motivation.
• Comparison of Augmented Inverse Probability Weighting (AIPW) and Targeted Maximum Likelihood Estimation (TMLE).
• Evaluation of data-adaptive approaches, cross-fitting with varying folds, and Super Learner library variations.

Main Results:

• AIPW and TMLE demonstrated similar point estimate performance for ACE.
• TMLE exhibited superior stability compared to AIPW.
• Cross-fitting enhanced variance estimation and coverage, more so than point estimates.
• A full Super Learner library was crucial for reducing bias and variance in complex scenarios.

Conclusions:

• Both AIPW and TMLE are viable doubly robust methods for high-dimensional confounding.
• TMLE's stability and the benefits of cross-fitting and comprehensive Super Learner libraries are key for reliable causal effect estimation in modern epidemiological research.