We quantify the value of data for the prediction policy problem of reducing antibiotic prescribing to curb antibiotic resistance. Using varying combinations of administrative data, we evaluate machine learning predictions for diagnosing bacterial urinary tract infections and the outcomes of prescription rules based on these predictions. Simple patient demographics improve prediction quality substantially but larger reductions in prescribing can be achieved by making use of rich health data. Our results suggest decreasing returns to data for prediction quality and increasing returns for policy outcomes. Hence, data needs for prediction policy problems must be assessed based on the policy objective and not only on prediction quality.