Social Science Lunch Talk

Abstract: We study average treatment effect estimation and experimental design in settings where outcomes are dyadic. For example, an outcome of interest might be whether agent i=1,...,N purchases (or not) product j=1,...,M, with the treatment being access to a coupon. Coupons may be randomly assigned to agents (for use on any product), to products (available for use by all agents), or directly at the agent-i-by-product-j level (i.e., available for use by agent i on product j alone). The dependence structure across dyadic outcomes complicates analysis relative to the familiar cross-sectional case, but also presents new opportunities (e.g., we observe multiple dependent outcomes for each agent). We explore questions of efficiency and research design in this setting; we also discuss possible applications.