Column (in German): ZEW News 9/2009.

Column (in German): DIW Wochenbericht 14/2014.

Working Papers

  • Machine predictions and human decisions with variation in payoffs and skill (with Michael A. Ribers), DIW Discussion Paper Nr. 1911 (download pdf).

Human decision-making differs due to variation in both incentives and available information. This constitutes a substantial challenge for the evaluation of whether and how machine learning predictions can improve decision outcomes. We propose a framework that incorporates machine learning on large-scale data into a choice model featuring heterogeneity in decision maker payoff functions and predictive skill. We apply this framework to the major health policy problem of improving the efficiency in antibiotic prescribing in primary care, one of the leading causes of antibiotic resistance. Our analysis reveals large variation in physicians’ skill to diagnose bacterial infections and in how physicians trade off the externality inherent in antibiotic use against its curative benefit. Counterfactual policy simulations show that the combination of machine learning predictions with physician diagnostic skill results in a 25.4 percent reduction in prescribing and achieves the largest welfare gains compared to alternative policies for both estimated physician as well as conservative social planner preference weights on the antibiotic resistance externality.

  • Physician effects in antibiotic prescribing: evidence from physician exits (with Shan Huang) – paper draft available on request

Human antibiotic consumption is considered the main driver of antibiotic resistance. Reducing human antibiotic consumption without compromising health care quality poses one of the most important global health policy challenges. A crucial condition for effective policies is to identify who drives antibiotic treatment decisions. We investigate to what extent physician practice style, as opposed to patient-specific factors, determines general practice antibiotic intake and health outcomes. Using linked administrative data from Denmark, a low-prescribing country, we first document that prescriptions in general practice drive large variation in antibiotic consumption. To identify the causal effect of physician practice style on variation in antibiotic prescribing, we exploit quasi-experimental variation in patient-physician relations due to physician exits from clinics in general practice. We estimate that physician practice style accounts for 53 to 56 percent of the cross-practice variation in all antibiotic consumption, and for 74 to 81 percent in broad-spectrum antibiotic consumption. We find little evidence that low prescribing styles adversely affect health outcomes measured as preventable hospitalizations due to infections. Our findings suggest that policies to curb antibiotic resistance are most effective when aimed at improving physician decision-making. We document that high prescribing practice styles are positively associated with physician age and negatively with staff size and the availability of diagnostic tools, suggesting that improvements in the quality of diagnostic information could be an important path to improved decisions.

Column (in German): DIW Wochenbericht 13-14/2021.

Large-scale data show promise to provide efficiency gains through individualized risk predictions in many business and policy settings. Yet, assessments of the degree of data-enabled efficiency improvements remain scarce. We quantify the value of the availability of a variety of data combinations for tackling the policy problem of curbing antibiotic resistance, where the reduction of inefficient antibiotic use requires improved diagnostic prediction. Focusing on antibiotic prescribing for suspected urinary tract infections in primary care in Denmark, we link individual-level administrative data with microbiological laboratory test outcomes to train a machine learning algorithm predicting bacterial test results. For various data combinations, we assess out of sample prediction quality and efficiency improvements due to prediction-based prescription policies. The largest gains in prediction quality can be achieved using simple characteristics such as patient age and gender or patients’ health care data. However, additional patient background data lead to further incremental policy improvements even though gains in prediction quality are small. Our findings suggest that evaluating prediction quality against the ground truth only may not be sufficient to quantify the potential for policy improvements.

  • Battling antibiotic resistance: can machine learning improve prescribing? (with Michael A. Ribers), DIW Discussion Paper Nr. 1803 (download pdf).

Columns: DIW Weekly Report 19/2019 (in English), DIW Wochenbericht 19/2019 and Oekonomenstimme (in German).

Antibiotic resistance constitutes a major health threat. Predicting bacterial causes of infections is key to reducing antibiotic misuse, a leading driver of antibiotic resistance. We train a machine learning algorithm on administrative and microbiological laboratory data from Denmark to predict diagnostic test outcomes for urinary tract infections. Based on predictions, we develop policies to improve prescribing in primary care, highlighting the relevance of physician expertise and policy implementation when patient distributions vary over time. The proposed policies delay antibiotic prescriptions for some patients until test results are known and give them instantly to others. We find that machine learning can reduce antibiotic use by 7.42 percent without reducing the number of treated bacterial infections. As Denmark is one of the most conservative countries in terms of antibiotic use, this result is likely to be a lower bound of what can be achieved elsewhere.

We analyze the effects of a hypothetical payment card fee regulation on bank profits, consumer welfare, and merchant welfare. We model consumers’ and merchants’ bank choices for debit card services, cardholders’ demand for card usage (conditional on bank choice), and how banks account for these in setting card fees to their customers. To estimate the model, we use bank-level data and survey data from the Norwegian debit card scheme, BankAxept. We conduct counterfactual exercises to analyze the effects of interchange fee regulations in the debit card scheme.

Former title: Considerations on partially identified regression models, Centre for European Economic Research Discussion Paper 2012-024, BETA (Bureau d’Economie Théorique et Appliquée) Working Paper No. 2012-07, Online Appendix: Download

Work in progress

My SSRN profile.
My Google Scholar profile.
My arXiv profile.