Propensity score analysis with local balance.

Area of Science:

• Statistics
• Biostatistics
• Epidemiology

Background:

• Propensity score (PS) analysis is crucial for estimating average treatment effect (ATE) in observational studies.
• Traditional PS methods often rely on correctly specified parametric models, risking biased ATE estimation if misspecified.
• Nonparametric models offer flexibility but may not guarantee sufficient covariate balance.

Purpose of the Study:

• To introduce a novel propensity score methodology that ensures both nonparametric flexibility and robust covariate balance.
• To address the limitations of existing methods that optimize global balance but may fail to achieve the balancing property.
• To provide a reliable tool for unbiased ATE estimation in the presence of potential model misspecification.

Main Methods:

• Proposing the propensity score with local balance (PSLB) methodology.
• Incorporating nonparametric propensity score models to enhance flexibility in treatment assignment.
• Optimizing for local balance, defined as covariate balance within PS-stratified sub-populations, which implies global balance.

Main Results:

• The PSLB methodology demonstrated substantial performance improvements over existing global balance optimization methods in extensive numerical studies.
• The proposed method effectively addresses issues arising from misspecified propensity score models.
• The balancing property, crucial for unbiased ATE estimation, is better achieved through local balance optimization.

Conclusions:

• The propensity score with local balance (PSLB) offers a more robust approach to causal inference from observational data compared to methods relying solely on global balance.
• PSLB enhances the reliability of average treatment effect (ATE) estimation, particularly when propensity score models are misspecified.
• The methodology is accessible through the R package PSLB, facilitating its adoption in research.