New Developments on Identification and Inference for Regression Discontinuity Designs
The regression discontinuity (RD) design has become one of the most widely used non-experimentalstrategies in empirical work in the social, behavioral and related sciences. It provides a simpleand intuitive approach to estimating the effect of a treatment on an outcome when unobservedconfounders may threaten the validity of the analysis. Despite its popularity among empiricalresearchers, there is only a handful of methodological results available in the RD literature, andmany important methodological, theoretical and practical problems remain unresolved. In this proposal, we seek to develop novel methodological tools for the analysis of RD designs and also plan to apply these results to concrete empirical problems of substantive interest in Economics and other social sciences.
Intellectual Merit. This proposal outlines several research projects specifically tailored to improve empirical work involving RD designs. We seek to develop new methodological and practical tools for estimation, inference and falsification of RD designs, based on novel theoretical developments in econometrics and statistics. In some cases, the latter theoretical results obtained in the context of this grant may also be of independent interest.
This grant outlines four main research projects. The first main project is motivated by the fact that traditional nonparametric local-linear RD estimators are very sensitive to the choice of bandwidth. We propose a novel confidence interval formula that is based on a bias-corrected version of the RD estimators together with a new standard-error formula for a large class of RD estimands.Our innovation is to consider an alternative distributional approximation for the bias-corrected estimator that explicitly accounts for the contribution of the estimated bias to the (finite-sample) variability of the statistic. Our main goal in this project is to extend the scope of our recent work described in Calonico, Cattaneo, and Titiunik (2014c).
The second main research project seeks to develop new methods for falsification and specification testing in RD designs (e.g., see our recent preliminary working paper Calonico, Cattaneo, and Titiunik (2014a)), as well as to develop valid inference procedures when the running variable is discretely valued. Our third main research project is concerned with identification and inference in cases where the RD design involves multiple cutoffs, a situation that is quite common in empirical work (despite the fact that most methodologies are developed for the special case of a single RD cutoff). Finally, our last research project outlines a new theoretical and practical approach to analyzing RD designs based on the idea of nearest-neighbors (NN) but without requiring the number of NN to grow with the sample size when deriving the results. We also plan to undertake several related, spin-off research projects as part of this grant.
Educational and Broader Impacts. We hope our results will provide researchers with new methodological tools to improve the way empirical work is carried out. Since RD designs are widely employed across disciplines (e.g., economics, education, political science, psychology, public health), we expect our results to be used by a large number of scholars. To broaden the impact of our agenda, we plan to distribute computer code for commonly used platforms (Stata, R, Matlab, etc.). This research will also involve our students, giving them an opportunity to learn new methodologies and develop first-hand research experience.