Collaborative Research: Econometric Methods for Models with Clustered Data and Covariate-Adaptive Randomization

University Of Chicago, Chicago IL

Investigators

Abstract

This award funds research that will develop new statistical theories and methods for analyzing three different kinds of data that are used to test economic theories. The first project considers how to analyze data when the individual data points can be grouped into a small number of "clusters". Since statistical methods begin with assumptions about randomization, the PI team will develop a theory of randomization tests that work in these settings. The second project considers problems associated with using randomized controlled methods to test the effects of an intervention, policy, or technology. These methods generally involve assigning people to a treatment or control group based on complicated rules that are designed to assure the two groups are comparable. These rules have implications for the best ways to analyze the resulting data, and the PIs plan to develop a new method that takes this into account. Finally, the team will consider how to best develop a method that can combine several separate statistical tests into a single framework. The project advances science by developing new methods that will be used to test theories from economics and other social sciences. The PIs pursue three different projects in econometrics. First, in joint work with Joseph Romano, they will develop a theory of randomization tests under an approximate symmetry assumption. A leading example where approximate symmetry holds is settings in which the data can be grouped into a small number of clusters. The hope is that their new methodology will work well in such settings, for which inference is known to be challenging. Second, in joint work with Federico Bugni, they will study how complicated treatment assignment rules designed to balance baseline covariates across treatment and control groups impacts inference about the average treatment effect. In preliminary work, they have shown that the usual two-sample t-test can be quite conservative in such settings. We plan to develop methods that do not have this feature. Finally, in joint work with Andres Santos, the PIs will develop methods for testing unions of functional moment inequalities. This framework encompasses many problems that have so far been treated separately in the econometrics literature, including (subvector) inference for (conditional) moment inequalities and tests of stochastic dominance.