Presentation (CTI)

OBJECTIVES: Estimating the Average Treatment Effect (ATE) using standard stabilized Inverse Probability of Treatment Weighting (sIPTW) is a foundational biostatistical technique, but it is vulnerable to variance inflation and instability when treatment overlap is poor. This occurs because extreme propensity score (PS) estimates create inflated weights, degrading the estimator's reliability.
METHODS: We propose Regularized stabilized IPTW (Reg-sIPTW), a novel method that introduces a data-adaptive shrinkage parameter (λ) via a power transformation of stabilized weights. Reg-sIPTW adaptively manages the bias-variance trade-off by shrinking extreme weights toward while maintaining consistency for the full-population ATE. The optimal λ is selected by minimizing the robust variance estimate subject to a strict covariate balance constraint.
RESULTS: Our simulations and empirical validation using the NHEFS dataset demonstrate that Reg-sIPTW achieves superior mean squared error (MSE) performance and robust covariate balance compared to standard sIPTW, especially under challenging low-overlap conditions where sIPTW tends to fail. Reg-sIPTW offers a robust alternative for estimating the average treatment effect (ATE), effectively avoiding the estimand shift introduced by methods such as trimming or overlap weighting.
CONCLUSIONS: This study introduces Reg-sIPTW, a novel regularized weighting method designed to improve the stability and efficiency of causal effect estimation in observational studies with limited covariate overlap. By integrating a data-adaptive regularization parameter within a constrained optimization framework, Reg-sIPTW effectively balances covariates while minimizing estimator variance, preserving consistency for the ATE. Simulation and empirical results demonstrate its superiority over standard sIPTW in terms of mean squared error, robustness, and effective sample size. While feasibility challenges remain under extreme violations of the positivity assumption, Reg-sIPTW offers a promising and practical advancement for finite-sample causal inference, with future work aimed at enhancing its flexibility and robustness through non-parametric and doubly robust extensions.