Bayesian additive regression trees for regression discontinuity designs

• Grant number: 21/13137-0
• Support Opportunities: Scholarships in Brazil - Doctorate
• Start date: September 01, 2022
• End date: February 29, 2024
• Field of knowledge: Applied Social Sciences - Economics - Quantitative Methods Applied to Economics
• Principal Investigator: Hedibert Freitas Lopes
• Grantee: Rafael Campello de Alcantara
• Host Institution: Instituto de Ensino e Pesquisa (Insper). São Paulo , SP, Brazil

Abstract

Treatment effects can be estimated in Regression Discontinuity (RD) designs, under cer¬tain assumptions, by the size of a discontinuity in the distribution of the response variable generated by the treatment assignment mechanism. The most common approach to the task is to approximate the distribution linearly around the discontinuity. Despite present¬ing desirable theoretical properties in general, this approach introduces the problems of how to select an appropriate window for the linear approximation and the loss of sample points outside this window. Global approximations avoid these issues and could present reasonable alternatives as long as they have a predictive ability similar or better than the local linear approach. This project investigates whether the Bayesian Additive Regression Trees (BART) algorithm (Chipman et al., 2010) can be adapted to incorporate the RD as¬ sumptions to estimate treatment effects well in this setting. Preliminary results indicate that, even with no adaptation, BART is able to compete with - and generally outperform - the local linear regression. (AU)