Are Guidelines Worth Following? Decision-Making Under Incomplete Evidence

Medical guidelines translate clinical evidence into treatment rules, but the evidence may identify treatment benefits more clearly than harms. We study anticoagulation for atrial fibrillation, combining randomized trials with Veterans Health Administration data. Trial evidence reveals substantial heterogeneity in stroke-prevention benefits but does not reliably identify how treatment-induced bleeding harms vary across patients. In the target population, patients with larger predicted stroke benefits also have higher untreated bleeding risk, making this ambiguity consequential. We evaluate existing guidelines and develop maximin and minimax-regret treatment rules robust to alternative harm structures. At a benchmark preventing 50% of preventable strokes, robust rules induce 33-41% fewer bleeds than random treatment and outperform existing guidelines. Finite-sample uncertainty generates noticeable instability in fine patient rankings but little instability in treatment assignments or welfare. Physician performance and the relationship between guideline adherence and skill also depend substantially on assumptions about how treatment-induced bleeding harms vary across patients.