Events

Prof. David C. Chan
Associate Professor of Health Policy
Stanford University

Physicians often deviate from medical guidelines, but most guidelines focus on a single disease while ignoring complex interactions that render treatment harmful. We study prescribing decisions covered by a prominent guideline for treating atrial fibrillation in patients at higher risk for stroke. Following the guideline’s introduction, physicians rapidly demonstrated awareness of it, yet they often declined to adhere to it, even for patients with the highest stroke risk. Using machine learning on eight randomized trials of anticoagulation, we find significant stroke treatment effect heterogeneity. However, the trials do not (and were not designed to) reveal heterogeneity in the bleeding side effect of anticoagulation. In the observational data, we find that bleed risk is positively correlated with stroke reduction, raising the possibility that the patients who benefit most from treatment may also suffer greater harm from it. Analyzing the performance of known and optimal treatment rules, we show that known guidelines may perform worse than randomly treating patients. Finally, we develop a guideline under an objective that maximizes the worst outcome under scientific ambiguity and show that this guideline may significantly improve welfare relative to status quo treatment decisions. Analyzing physician behavior under this benchmark, we find that ranking patients according to predicted physician treatment under the “wisdom of the crowds'” performs better than strict adherence to known guidelines.