Analysis of Spillover Effects in Randomized Experiments

Abstract

This dissertation studies identification, estimation, inference and experimental design for analyzing causal spillover effects in randomized experiments. Chapter II provides a nonparametric framework based on potential outcomes to define spillover effects in a setting in which units are clustered and their potential outcomes can depend on the treatment assignments of all units within a group. Using this framework, I provide conditions for identification of average direct and spillover effects when the treatment is randomly assigned. I then study identification under three estimation strategies that are commonly employed in empirical work: a regression of an outcome on a treatment indicator, which calculates a difference in means between treated and controls, a regression that controls for the proportion of treated peers, and a regression exploiting variation in treatment probabilities in two-stage designs.

Chapter III analyzes estimation and inference for spillover effects. I start by illustrating the results from Chapter II using two empirical applications. I then study nonparametric estimation and inference for spillover effects in a setting in which both the number of groups and the group size are allowed to grow. This setting allows me to understand the effect of the number of parameters on the asymptotic properties of the proposed nonparametric estimators. Finally, I discuss the implications of these findings for the design of experiments.

Chapter III discusses some key issues related to the empirical implementation of the results from the previous chapters: the inclusion of covariates, identification of spillover effects in experiments with imperfect compliance and optimal design of experiments.