Design of Field Experiments under Unknown Interference Structures

Abstract

This paper extends the Rubin Causal Model to a framework that allows for interference of unknown form. In particular, we develop a framework to define estimands in a general way using "reference assignments." We show that a sequence of increasingly restrictive non- interference assumptions yield estimators that converge in expectation to the estimand of interest. This allows the researcher to focus on experimental designs that causally identify quantities of interest in a non-parametric fashion. We apply this approach to two common es- timands, the average direct treatment effect and the average indirect exposure effect, recasting the framework in the language of network analysis. We propose a novel two-stage randomiza- tion algorithm that incorporates the researchers’ beliefs about interference, allowing for more efficient estimation and reducing biases from complex spillovers. The algorithm is evaluated with simulated data.