Stabilized Inverse Probability Weighting via Isotonic Calibration.

Area of Science:

• Causal Inference
• Statistical Modeling
• Biostatistics

Background:

• Inverse weighting using propensity scores is standard for adjusting confounding bias in causal inference.
• Directly inverting propensity score estimates can cause instability and bias due to large weights, particularly with limited treatment overlap.

Purpose of the Study:

• To propose a post-hoc calibration algorithm for stabilizing inverse propensity weights.
• To improve the performance of doubly robust estimators for average treatment effect estimation.

Main Methods:

• Developed a post-hoc calibration algorithm for inverse propensity weights.
• Employed a variant of isotonic regression with a tailored loss function.
• Utilized user-supplied, cross-fitted propensity score estimates.

Main Results:

• The proposed isotonic calibration algorithm generates well-calibrated and stabilized weights.
• Demonstrated through theoretical analysis and empirical studies that calibration improves estimator performance.
• Showcased enhanced performance of doubly robust estimators for average treatment effect.

Conclusions:

• Isotonic calibration offers a robust method for stabilizing inverse propensity weights.
• The approach effectively addresses issues of instability and variability in causal inference.
• Improves the reliability of average treatment effect estimation in challenging scenarios.