Abstract
Inverse probability weighting (IPW) estimators are widely used to estimate average causal effects, primarily because they do not require explicit modeling of the outcome. However, by relying solely on the propensity score, they fail to fully utilize information from the outcome, often resulting in low estimation efficiency in practice. To address these limitations, we propose an improved IPW estimator that incorporates flexible outcome working models with zero marginal treatment contrast to enhance statistical efficiency, extending the existing linear regression adjustment framework. We compare adjustment strategies with and without predictive covariates. Under certain conditions, the proposed estimator exhibits double robustness: it yields consistent estimates if either the propensity score model or the constrained outcome regression model is correctly specified. Moreover, when both models are correctly specified, it achieves asymptotic variance no greater than that of the standard augmented IPW estimator. Extensive simulation studies are conducted to evaluate finite-sample performance. We also apply the proposed method to a diabetes dataset and find strong evidence of a causal link between retinopathy and high HbA1c levels. Thus, incorporating such covariates into both the propensity score and outcome models does not increase optimized asymptotic variance with known or correctly estimated propensity scores.
Funding Information
  • National Natural Science Foundation of China (12401378)

This publication has 31 references indexed in Scilit: