Parametric and nonparametric propensity score weighting analysis with subgroup covariate balance

Area of Science:

• Observational studies
• Causal inference
• Biostatistics

Background:

• Estimating subgroup causal treatment effects is vital for understanding treatment effect heterogeneity.
• Existing propensity score methods are sensitive to model misspecification, leading to biased estimates and imbalance.
• Propensity score model misspecification can compromise the validity of subgroup analyses.

Purpose of the Study:

• To develop a propensity score method that ensures covariate balance within all subgroups.
• To enhance robustness against propensity score model misspecification.
• To improve the accuracy of subgroup causal treatment effect estimation.

Main Methods:

• Proposed G-SBPS (generalized propensity score analysis with controlled subgroup balance) for covariate mean balance across subgroups.
• Developed kernelized G-SBPS (kG-SBPS) incorporating nonparametric kernel regression for improved balance of covariate transformations.
• Evaluated methods through extensive numerical simulations.

Main Results:

• G-SBPS and kG-SBPS significantly improved subgroup covariate balance compared to existing methods.
• Both proposed methods demonstrated superior subgroup treatment effect estimation.
• kG-SBPS showed increased robustness to propensity score model misspecification.

Conclusions:

• G-SBPS and kG-SBPS offer effective solutions for subgroup analysis in observational studies.
• These methods enhance the reliability of causal effect estimates when treatment effects vary across subgroups.
• The proposed approaches improve upon existing propensity score techniques for subgroup inference.