New Frontiers in Causal Inference: Learning From Experiments in a Connected World

Abstract

The experimental and data-driven approach that was once the purview of science has gained momentum with other actors: large technology companies have adopted systematic A/B testing as a philosophy for driving innovation, while governments rely increasingly on quantitative program evaluation to inform policy decisions. These actors operate on systems and societies that are tightly connected, raising new and exciting challenges for the design and analysis of causal experiments. First, an intervention targeted at a specific unit on a network often spills over to neighboring units -- a phenomenon known as interference -- which may bias the analysis. Second, units that behave similarly tend to congregate and form connections with each other -- a phenomenon known as homophily -- which may inflate the variance of the analysis. In the first three chapters, we explore the problem of interference from different angles: we first explore the theoretical limitations of randomization-based inference under interference, before developing unbiased estimators for a variety of effects, and proposing a framework extending traditional Fisher tests to hypotheses that are no longer sharp. The fourth and last chapter proposes a novel approach to design, leveraging a working-model to reduce the variance of a simple estimator of causal effects under homophily.