Matching on Generalized Propensity Scores with Continuous Exposures.

Area of Science:

• Causal inference
• Epidemiology
• Biostatistics

Background:

• Matching is established for binary treatments but underdeveloped for continuous exposures.
• Continuous exposure analysis requires robust causal inference methods.

Purpose of the Study:

• To propose an innovative matching approach for estimating average causal exposure-response functions with continuous exposures.
• To address limitations of existing methods in handling model misspecification and extreme propensity score values.

Main Methods:

• Utilized the generalized propensity score (GPS) for matching in continuous exposure settings.
• Introduced the 'local weak unconfoundedness' assumption for theoretical guarantees.
• Developed a method with features like design-analysis separation, robustness, and covariate balance assessment.

Main Results:

• The proposed matching estimator demonstrated point-wise consistency and asymptotic normality under stated assumptions.
• Simulations showed superior performance compared to existing methods, especially with model misspecification or extreme GPS values.
• Application to Medicare data revealed a significant harmful effect of long-term PM2.5 exposure on all-cause mortality.

Conclusions:

• The novel GPS-based matching approach effectively estimates causal exposure-response functions for continuous exposures.
• The method offers theoretical guarantees and practical advantages, outperforming existing techniques.
• Confirmed a substantial link between long-term PM2.5 exposure and increased mortality risk in a large population study.