Dirichlet process mixture models for regression discontinuity designs.

Area of Science:

• Econometrics
• Biostatistics
• Causal Inference

Background:

• Regression discontinuity design (RDD) is a quasi-experimental method to estimate causal effects.
• Assignment in RDD is based on a threshold of a continuous variable.
• Optimal bandwidth selection is critical but challenging in RDD analysis.

Purpose of the Study:

• To propose a novel methodology for RDD bandwidth selection.
• To improve causal effect estimation by prioritizing unit exchangeability.
• To address limitations of existing bandwidth selection methods.

Main Methods:

• Utilized a Dirichlet process mixture model for sample clustering.
• Employed posterior similarity matrices to identify exchangeable clusters.
• Included only clusters with strong evidence of exchangeability in the RDD analysis.
• Applied the methodology to simulated data and a real-world example (statin effects on cholesterol).

Main Results:

• The proposed methodology effectively identifies exchangeable units for RDD.
• This approach enhances the reliability of causal effect estimates.
• Demonstrated validity through simulation and a practical application.

Conclusions:

• Unit exchangeability is a robust criterion for RDD subject selection.
• The Dirichlet process mixture model offers a flexible framework for identifying exchangeable clusters.
• This method provides a practical and statistically sound alternative for RDD bandwidth selection.