Optimally Balanced Gaussian Process Propensity Scores for Estimating Treatment Effects.

Area of Science:

• Statistics
• Epidemiology
• Machine Learning

Background:

• Propensity scores are crucial for estimating average treatment effects in observational studies.
• Existing methods for propensity score estimation can be limited in flexibility and covariate balance assessment.

Purpose of the Study:

• To introduce a flexible propensity score modeling approach using Gaussian processes (GP).
• To develop a metric for quantifying covariate imbalance and optimize GP hyperparameters for minimal imbalance.
• To evaluate the performance of the GP method against existing approaches.

Main Methods:

• Modeling treatment probability using a Gaussian process framework.
• Developing a covariate imbalance metric based on the discrepancy between treated and control group covariate distributions.
• Optimizing Gaussian process hyperparameters to minimize the covariate imbalance metric.

Main Results:

• The covariate imbalance metric is shown to be a function of Gaussian process hyperparameters.
• Hyperparameter optimization effectively minimizes covariate imbalance.
• The GP method demonstrates competitive or superior performance in simulations and a real-world policy application.

Conclusions:

• Gaussian process modeling offers a flexible and effective approach for propensity score estimation.
• The developed covariate imbalance metric aids in optimizing propensity score models.
• This method enhances the reliability of average treatment effect estimation in observational studies.