Abstract: Antibiotic misuse due to prescribing under diagnostic uncertainty is a leading driver of antibiotic resistance. We investigate the magnitude and mechanisms by which machine learning predictions can enable policies that reduce antibiotic misuse. Building on predictions from administrative data on urinary tract infections in Denmark, we evaluate counterfactual policies that replace or improve human diagnostic expertise and contrast these to policies manipulating payoffs. Estimating a model of physician decision-making, we find substantial heterogeneity in diagnostic skill and preferences. Consequently, policies combining individual diagnostic skill with predictions achieve the largest effects, reducing antibiotic use by up to 17.8 percent and overprescribing by 33.3 percent.
Joint with Michael Ribers (DIW Berlin & University of Copenhagen)