A doubly robust estimator for continuous treatments in high dimensions

Area of Science:

• Causal inference
• Observational studies
• Statistical methodology

Background:

• Generalized propensity score (GPS) methods are widely used for estimating causal effects of continuous treatments in observational studies.
• Rich covariate data strengthens the unconfoundedness assumption in GPS methods.
• Correct specification of treatment distributions is critical for valid GPS analysis.

Purpose of the Study:

• To address limitations in existing GPS methods, particularly in high-dimensional settings.
• To introduce a novel approach, GOALDeR (Generalized Outcome-Adaptive LASSO and Doubly Robust Estimate), for robust causal inference.
• To improve the accuracy and statistical efficiency of causal effect estimation.

Main Methods:

• Extended balance-based approaches to high dimensions.
• Developed GOALDeR, integrating a distribution-misspecification-robust balance method, a model-misspecification-robust doubly robust estimator, and variable selection.
• Employed simulation studies and real-data analysis to evaluate GOALDeR's performance.

Main Results:

• GOALDeR produced nearly unbiased estimates when either the GPS or outcome model was correctly specified.
• GOALDeR demonstrated superior precision and accuracy compared to existing methods.
• Real-data analysis found no significant dose-response between epigenetic age acceleration and Alzheimer's disease.

Conclusions:

• GOALDeR is an advanced, doubly robust GPS method for high-dimensional causal inference.
• GOALDeR offers improved accuracy and precision over existing methods.
• The GOALDeR R package is publicly available for use.