Practical Considerations for Gaussian Process Modeling for Causal Inference in Quasi-Experimental Studies With Panel...

Area of Science:

• Statistics
• Econometrics
• Epidemiology

Background:

• Quasi-experimental studies with spatio-temporal panel data often face challenges due to unmeasured confounding.
• Traditional methods may struggle to capture complex spatial and temporal dependencies inherent in such data.

Purpose of the Study:

• To introduce a practical and interpretable framework for applying Gaussian processes (GPs) to causal inference in spatio-temporal panel data.
• To demonstrate how GPs can generalize existing methods like synthetic control and vertical regression.

Main Methods:

• Utilized Gaussian processes (GPs), a nonparametric modeling approach, employing carefully chosen covariance kernels to handle complex spatio-temporal dependencies.
• Represented the GP posterior mean as a weighted average of observed outcomes, with weights indicating spatial and temporal similarity.

Main Results:

• Gaussian processes (GPs) effectively estimate counterfactual outcomes and quantify treatment effects in spatio-temporal settings.
• The study explored the impact of different kernel choices on estimation performance and interpretability, providing guidance for practical application.
• Demonstrated the utility of GP models through simulations and an application to Hurricane Katrina mortality data.

Conclusions:

• Gaussian processes (GPs) present a promising and interpretable tool for mitigating unmeasured spatio-temporal confounding in quasi-experimental research.
• The framework enhances causal inference by offering flexibility and transparency in modeling complex data structures.
• Publicly available code and materials aim to promote reproducibility and wider adoption of this methodology.