Subgroup balancing propensity score.

Area of Science:

• Statistics
• Econometrics
• Causal Inference

Background:

• Estimating subgroup treatment effects using observational data presents challenges.
• Existing propensity score methods primarily focus on overall treatment effects and may fail to achieve covariate balance within subgroups.
• Subgroup analysis exacerbates the bias-variance tradeoff, potentially increasing variance with improved balance.

Purpose of the Study:

• To introduce a novel method, the subgroup balancing propensity score, designed to ensure covariate balance within subgroups while controlling for variance inflation.
• To address the limitations of existing propensity score methods in subgroup analysis.
• To improve the accuracy and reliability of subgroup treatment effect estimation.

Main Methods:

• Proposing the subgroup balancing propensity score (SBPS) method.
• SBPS selects either the overall sample or a subgroup sample for propensity score estimation within each subgroup to optimize covariate balance and minimize variance.
• Developing two SBPS versions (matching and weighting) and a stochastic search algorithm for large numbers of subgroups.

Main Results:

• Simulations demonstrate that the SBPS method significantly improves the performance of propensity score methods for estimating subgroup treatment effects.
• The proposed method effectively balances covariates within subgroups and controls for variance inflation.
• The SBPS method was successfully applied to real-world data (SHIW) to analyze income group-specific effects of debit card ownership on household consumption.

Conclusions:

• The subgroup balancing propensity score method offers a robust approach for estimating subgroup treatment effects from observational data.
• This method enhances causal inference by addressing critical issues of covariate balance and variance in subgroup analyses.
• The SBPS method provides a valuable tool for researchers investigating heterogeneous treatment effects across different subpopulations.