Learning and the Efficiency of Expert Referrals

Buyers of complex goods and services often rely on the advice of expert intermediaries who are themselves imperfectly informed about the quality of the available options. How well do these intermediaries learn about that quality and act on it? We study this question in the setting of referrals from primary care physicians (PCPs) to specialists, using data on 4.5 million joint replacement surgeries for Medicare beneficiaries. We first document substantial heterogeneity in specialist quality and costs within geographic markets, and we present design-based evidence showing that PCPs adjust their referrals specifically based on the outcomes of their own patients. We then employ a structural learning model of PCP referral choices to quantify the losses from informational frictions and to simulate possible reallocations with improved information. Beyond learning, the model also accounts for limitations on possible reallocations due to habit persistence and capacity constraints. We find that about one-quarter of patients would be referred to a different specialist in the absence of informational frictions, with small but meaningful improvements in patient outcomes.