Primary Submission Category: Causal Inference and SUTVA/Consistencies Violations
Scaling-Up Experiments in Centralized Markets
Authors: Wisse Rutgers, Dmitry Arkhangelsky,
Presenting Author: Wisse Rutgers*
This paper focuses on identification of the marginal policy effect (MPE, Wager and Xu 2021) in centralized markets. Hu, Li, and Wager (2021) show that the MPE can be decomposed in a direct and indirect effect. Where the direct effect is simply equal to a traditional average treatment effect, we show how to leverage knowledge about the market structure in centralized markets to identify the usually hard-to-estimate indirect effect. Furthermore, we show that in certain centralized markets where the mechanism contains a random component we can identify a causal effect not only for the MPE, but for a range of different treatment probabilities. We consider centralized markets similar to that of Munro (2024), where products are allocated by a centralized mechanism based on reports submitted by agents. When agents’ preferences in the market are directly observed from their reports the framework can be applied straightforwardly. Furthermore, when preferences are not directly observed but can be derived from observables the framework still applies.