Learning, Rules, and Discretion: Evidence from Anticoagulation for Atrial Fibrillation

Clinical risk scores are an increasingly common tool to guide clinical decisions with data from patient populations. However, scores typically summarize baseline risk absent treatment while clinicians ideally would like to know the effect of treatment across multiple outcomes, such as the benefit of reducing stroke and the risk of increasing hemorrhage by anticoagulation, on specific patient types. We make use of a large database of detailed clinical records from the Veterans Health Administration to study anticoagulation treatment choices and outcomes for patients with atrial fibrillation. We estimate how anticoagulation effects stroke and hemorrhage across patients with different characteristics; we use variation in which patients are as good as randomly assigned to physicians with different propensities to prescribe anticoagulation in order to estimate heterogeneous treatment effects. We also analyze how current practice deviates from the risk-score based treatment. Using detailed records of patient characteristics, which are not always captured in simple risk scores, we will test whether physicians consider other relevant clinical factors to improve their decision-making. Finally, we compare patient outcomes under actual physician decisions with counterfactual outcomes under strict adherence to treating based on standard risk scores and optimal risk scores, allowing physicians to have both observed and unobserved private information.