Model Uncertainty and Model Averaging in Regression Discontinuity Designs

Abstract Parametric (polynomial) models are popular in research employing regression discontinuity designs and are required when data are discrete. However, researchers often choose a parametric model based on data inspection or pretesting. These approaches lead to standard errors and confidence intervals that are too small because they do not incorporate model uncertainty. I propose using Frequentist model averaging to incorporate model uncertainty into parametric models. My Monte Carlo experiments show that Frequentist model averaging leads to mean square error and coverage probability improvements over pretesting. An application to [Lee, D. S. 2008. “Randomized Experiments From Non-Random Selection in US House Elections.” Journal of Econometrics 142 (2): 675–697.] shows how this approach works in practice, and how conventionally selected models may not be ideal.