Sendhil Mullainathan, Ziad Obermeyer

Bibliographic Information

NBER Working Paper No. 26168
Issued in August 2019
NBER Program(s):AG, HC, HE, LS, PR

Available Formats


In deciding whether to test for heart attack (acute coronary syndromes), physicians implicitly judge risk. To assess these decisions, we produce explicit risk predictions by applying machine learning to Medicare claims data. Comparing these on a patient-by-patient basis to physician decisions reveals more about low-value care than the usual approach of measuring average testing results. It more precisely quantifies over-use: while the average test is marginally cost-effective, tests at the bottom of the risk distribution are highly cost-ineffective. But it also reveals under- use: many patients at the top of the risk distribution go untested; and they go on to have frequent adverse cardiac events, including death, in the next 30 days. At standard clinical thresholds, these event rates suggest they should have been tested. In aggregate, 42.8% of the potential welfare gains of improving testing would come from addressing under-use. Existing policies though are too blunt: when testing is reduced, for example, both low-value and high-value tests fall. Finally, to understand physician error we build a separate algorithm of the physician and find evidence of bounded rationality as well as biases such as representativeness. We suggest models of physician moral hazard should be expanded to include ‘behavioral hazard’.

National Bureau of Economic Research
1050 Massachusetts Ave.
Cambridge, MA 02138

Twitter RSS

View Full Site: One timeAlways