A Bayesian Nonparametric Causal Model for Regression Discontinuity Designs

• Activity Number: 510
• Type: Invited
• Date/Time: Wednesday, August 7, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): George Karabatsos*+ and Stephen G. Walker
• Companies: University of Illinois-Chicago and University of Kent
• Keywords: Bayesian nonparametric regression ; Causal inference ; Sharp and fuzzy regression discontinuity designs

Abstract

The regression discontinuity (RD) non-randomized design can identify and estimate causal effects for a "locally-randomized" subgroup of subjects, under relatively mild conditions. Though, the accurate estimation of causal effects still relies on the predictive accuracy of the statistical model. We propose a flexible Bayesian nonparametric regression model which can provide predictively-accurate estimates of causal effects, either in terms of the mean, variance, distribution function, quantile, probability density, or any other functional of the outcome variable. We illustrate the model through the analysis of two real educational sets.