The overall goal of this project is to advance knowledge on the estimation of and inference for causal effects. The researchers propose three specific projects. The first consists of three topics in program evaluation. In the literature on estimation of average treatment effects several efficient estimators have been proposed that require choices for smoothing parameters, e.g., the bandwidth in kernel regression. Typically researchers have not given specific data-driven choices, limiting the applicability of these estimators. Here an optimal data-driven choice for the smoothing parameter will be derived. Second, issues related to overlap between covariate distributions for treated and control units will be analyzed. The lack of such overlap leads to substantial sensitivity of most of the estimators used in the literature and has led researchers to use ad hoc methods for imposing sufficient overlap. A principled and optimal criterion for selecting the sample will be developed. Third, bootstrapping methods for matching estimators will be considered. The simple bootstrap will be shown not to be valid for matching estimators, and alternatives that may have better properties will be studied.
In the second project the plan is to study the aggregate implications of peer effects. If characteristics or actions of, for example, classmates in school affect an individual's outcome, there may be both individual and aggregate effects of reassigning individuals to classes. Whereas much of the literature has focused on the identification and estimation of individual level effects, under the current proposal new methods will be developed for estimation and inference for the aggregate effects. Inference for the effect of full positive assortive matching (where the individuals with the highest values of some characteristic are grouped together), as well as tests for the effect of moving towards more positive assortive matching will be studied.
In the third project the researchers plan to investigate the design and analysis of randomized experiments in the presence of peer effects or interactions between the experimental units. If units interact, standard experimental methods do not apply. The researchers propose to study the equivalent of Fisher exact tests for this setting, as well as the equivalent of Neyman's methods for constructing confidence intervals. Preliminary results show that in general these methods are not as similar to each other in the case with interaction as in the case without.
Broader impacts resulting from the proposed activity are the availability of ready-to-use software for the evaluation of causal effects in both academic areas, including economics and other social sciences, and for policy makers. Software will be developed in Matlab and STATA that will be made available to other researchers.
