Evaluation of propensity score methods for causal inference with high-dimensional covariates

Area of Science:

• Causal inference
• High-dimensional statistics
• Observational studies

Background:

• Estimating causal effects with many covariates in observational studies is challenging.
• Propensity score (PS) methods are crucial for addressing confounding in such settings.
• This study reviews and evaluates PS methods tailored for high-dimensional data.

Purpose of the Study:

• To systematically compare model-based and balance-based propensity score methods in high-dimensional settings.
• To assess the impact of covariate balancing constraints on estimation performance.
• To identify the most effective methods for causal effect estimation with numerous covariates.

Main Methods:

• Reviewed propensity score (PS) methods, categorizing them into model-based and balance-based approaches.
• Conducted systematic simulation experiments to evaluate methods including GLiDeR, hdCBPS, PDS, OAL, DiPS, balanceHD, and RCAL.
• Compared estimation accuracy, precision, and stability across different high-dimensional PS techniques.

Main Results:

• For model-based methods, GLiDeR showed the highest stability, accuracy, and precision, followed by PDS, OAL, and DiPS.
• Among balance-based methods, hdCBPS performed comparably to GLiDeR and outperformed balanceHD and RCAL.
• Covariate balancing constraints did not significantly improve propensity score method performance in high-dimensional scenarios.

Conclusions:

• Recommends GLiDeR and hdCBPS as preferred methods for estimating causal effects in high-dimensional observational studies.
• Suggests that covariate balancing constraints offer limited benefits in high-dimensional settings.
• Highlights the need for further research into constructing valid confidence intervals for these methods.