Leila Agha (Dartmouth College)

Expert decisions often deviate from evidence-based guidelines. If experts are unaware of guidelines, dissemination may improve outcomes. If experts are aware of guidelines but continue to deviate, promoting stricter adherence has ambiguous effects on outcomes, depending on whether experts have information not in guidelines. We study guidelines for anticoagulant use to prevent strokes among atrial fibrillation patients. By text-mining physician notes for guideline mentions, we identify when physicians start using a guideline. After mentioning the guideline, physicians' prescription decisions become more guideline concordant, but adherence remains far from perfect. To evaluate whether non-adherence reflects physicians' superior information, we combine observational data on treatment choices with machine learning estimates of heterogeneous treatment effects from eight randomized trials. When physicians depart from guidelines, they are not responding to measurable treatment effect heterogeneity. Promoting stricter adherence to guidelines could prevent 33% more strokes, producing much larger gains than broader guideline awareness.