Bayesian model averaging for nonparametric discontinuity design.

Area of Science:

• Statistics
• Econometrics
• Causal Inference

Background:

• Quasi-experimental designs like regression discontinuity and interrupted time series enable causal inference when randomized controlled trials are infeasible.
• These methods often rely on strong assumptions and can suffer from overconfidence and model misspecification.

Purpose of the Study:

• To introduce a new framework, Bayesian nonparametric discontinuity design (BNDD), for robust causal inference in quasi-experimental settings.
• To address limitations of existing discontinuity-based designs, specifically overconfidence and model misspecification.

Main Methods:

• Bayesian model averaging combined with Gaussian process regression forms the core of the BNDD framework.
• The approach utilizes appropriate Gaussian process covariance functions to detect discontinuities of any order and in spectral features.

Main Results:

• Simulations demonstrate the utility and robustness of the BNDD framework.
• The framework was successfully applied to diverse real-world scenarios, including political science, public health, and psychology.

Conclusions:

• BNDD offers a flexible and powerful approach to causal inference, improving upon traditional discontinuity-based designs.
• The framework provides a principled way to handle complex data structures and detect nuanced effects in observational studies.