Robust propensity score estimation via loss function calibration.

Area of Science:

• Statistics
• Machine Learning
• Epidemiology
• Causal Inference

Background:

• Propensity score estimation is crucial for causal inference from observational data.
• Propensity score model misspecification can invalidate average treatment effect estimates.
• Machine learning methods for propensity scores may not guarantee covariate balance.

Purpose of the Study:

• To propose a novel calibration-based method for propensity score estimation.
• To enhance covariate balance and mitigate model misspecification impacts.
• To improve the accuracy and robustness of causal effect estimation.

Main Methods:

• A calibration-based method is proposed, integrating covariate balance into propensity score models.
• Loss functions are calibrated by adding a covariate imbalance penalty.
• The method is applied to both parametric (logistic regression) and machine learning (neural networks) models.

Main Results:

• The proposed method demonstrates robustness against propensity score model misspecification.
• Integrating loss function calibration improves covariate balance and reduces root-mean-square error of causal effect estimates.
• Neural network models with calibration yielded the best performance (less bias, smaller variance) under misspecification.

Conclusions:

• The proposed calibration-based method effectively addresses propensity score model misspecification.
• Explicitly incorporating covariate balance into propensity score estimation improves causal inference validity.
• This approach offers a more reliable way to estimate causal effects from observational data.