How Effective Are Machine Learning and Doubly Robust Estimators in Incorporating High-Dimensional Proxies to Reduce...

Area of Science:

• Epidemiology
• Biostatistics
• Machine Learning in Health Research

Background:

• Residual confounding is a major challenge in high-dimensional observational studies.
• High-dimensional proxy adjustment methods, like hdPS, use proxies for unmeasured confounders.
• Machine learning and doubly robust estimators have been integrated into hdPS extensions, but their comparative performance is unclear.

Purpose of the Study:

• To evaluate the performance of standard methods, super learner (SL), targeted maximum likelihood estimation (TMLE), and double cross-fit TMLE (DC-TMLE) in confounding adjustment.
• To compare these methods under varying exposure and outcome prevalence using different machine learning learner configurations.
• To assess the impact of high-dimensional proxies and learner complexity on bias, coverage, and variability.

Main Methods:

• Plasmode simulations were conducted to assess method performance.
• Evaluated standard methods, SL, TMLE, and DC-TMLE.
• Compared performance across three learner library sizes: 1, 3, and 4 learners (including logistic regression, MARS, LASSO, and XGBoost).

Main Results:

• Methods without proxies showed the highest bias and lowest coverage.
• Standard methods incorporating high-dimensional proxies demonstrated robust performance with low bias and good coverage.
• TMLE and DC-TMLE reduced bias but had worse coverage, especially with larger, complex learner libraries.
• DC-TMLE underperformed in high-dimensional settings with non-Donsker learners, indicating instability.

Conclusions:

• High-dimensional proxies are crucial for effective confounding adjustment in standard methods.
• Tailoring machine learning learner configurations in SL and TMLE is essential for reliable confounding adjustment.
• Careful selection of learners is vital to avoid instability and ensure accurate results in complex observational studies.