Can Self-Preferencing by An Online Retailer Be Detected? A Monte Carlo Simulation

Internet search results must be presented in a sequential manner—there is a first result, a second, a third, and so forth. For some platforms, an independent party may pay a fee for a higher ranking in the sequence. And when a platform offers its own products or services, the platform may favor its offerings over other sellers in its search returns. A paper by Farronato, Fradkin, and MacKay (2023) (hereinafter “FFM Study”) recently evaluated whether Amazon favors its own brands in search returns, reporting a mild preference, though middling rank, for Amazon brands. It is an interesting paper, but unfortunately the results of the FFM Study are invalid. In this POLICY BULLETIN, we demonstrate that the use of Ordinary Least Squares regression when looking at the sequence of search returns provides systematically biased coefficients and is prone to find preferences that do not exist and miss preferences that do. Also, the results of hypothesis tests are incorrect. The problem of omitted variables, which are certain to exist in the study, makes matters worse. A suitable alternative for detecting preference in ranked data has proven elusive