Addressing Extreme Propensity Scores via the Overlap Weights

Area of Science:

• Causal Inference
• Statistical Methods
• Epidemiology

Background:

• Inverse probability weighting is a common causal inference method.
• Extreme propensity scores can lead to biased estimates and high variance.
• Trimming is a common but often suboptimal solution, sensitive to cutoff points.

Purpose of the Study:

• To introduce and evaluate the overlap weighting method.
• To compare overlap weighting against inverse probability weighting with trimming.
• To assess bias, variance, and confidence interval coverage under various propensity score distributions.

Main Methods:

• Developed overlap weights to down-weight extreme propensity scores.
• Conducted simulations comparing overlap weights to trimmed inverse probability weighting.
• Evaluated performance across diverse propensity score distributions, including non-overlapping and extreme values.

Main Results:

• Overlap weighting demonstrated reduced bias and variance compared to trimmed methods.
• Overlap weights provided better confidence interval coverage.
• The method is robust across various propensity score distributions.

Conclusions:

• Overlap weighting is a more effective method for causal inference with extreme propensity scores.
• It mitigates issues associated with trimming, offering improved precision and reliability.
• A consistent standard error estimator for overlap weighting is provided for practical use.