Sign up for meetings on the sheet below:
If signing up less than two days before the talk, please also email firstname.lastname@example.org
Existing medical guidelines are simple rules based on a small number of variables. With modern medical records, vastly more information is available on which treatment decisions could be based. We present evidence that incorporating more of this information could lead to extremely large health beneﬁts. Our principle sample is 1.3 million patients: the universe of patients admitted to VA hospitals from 2000 to 2015 with measured Hemoglobin levels less than 10 (the range over which we observe transfusions in the data). We ﬁrst show that treatment decisions are invariant to all patient characteristics except hemoglobin level – roughly 30% of patients receive transfusions regardless of age, past inpatient stays, and comorbities. Treatment eﬀects estimated via OLS nonetheless vary substantially with observables. Using instruments based on regression discontinuity and quasi-random assignment of patients to physicians, we assess the degree of selection on unobservables. We ﬁnd that physicians treat patients with unobservably larger beneﬁts, but that most of the measured heterogeneity in “OLS” treatment eﬀects reﬂects heterogeneity in underlying causal eﬀects rather than selection. In a structural model combining random forest machine learning techniques with a model of selection, we ﬁnd that that better targeting the existing number of transfusions would increase the average treatment eﬀect of transfusions on mortality from 0.06 pp to 0.1 pp, saving 500 lives per year in our sample.