LAUSR

A large body of evidence demonstrates substantial area-level variations in intensity of care and health care spending in the United States. This variation is evident both across and within regions and exists for all types of care, including oncology care (1) and care identified to be of low value (2). Although the sources of variation have been long debated, recent evidence suggests that physicians’ beliefs, independent of organizational factors, are the key drivers in explaining area-level variations in health care spending (3). Physicians in the United States have substantial autonomy in decisions about care for patients, and thus their decisions are important drivers of health care utilization. Current alternate payment models target provider organizations in the hope that they create opportunities for physicians to influence and improve the value of care delivered by the organization. However, physician behavior is notoriously difficult to change, and relatively little evidence is available about the ability of physicians to directly influence the clinical decisions of their peers. In this issue of the Journal, Pollack and colleagues used data from the Surveillance, Epidemiology, and End Results–Medicare program to assess the role of physician networks in explaining adoption of perioperative advanced imaging tests for women with breast cancer (4). In a creative analysis, they identified the operating surgeons for women diagnosed with nonmetastatic breast cancer during 2004 to 2009. They specified 2004 to 2006 as the baseline period, during which time magnetic resonance imaging (MRI) and positron emission tomography (PET), two tests with unproven benefits in this setting, were just being adopted. They characterized perioperative use of these tests for all surgeons in the baseline period and specifically identified surgeons who had not used one or the other test during that time. They next assigned each surgeon to a “peer group” or network of physicians who share patients and may thus communicate their beliefs and practice styles with each other. The peer groups were identified using methods from network science (5) that examined patient-sharing patterns of surgeons, medical oncologists, radiation oncologists, primary care physicians, and radiologists caring for breast cancer and noncancer patients. The authors hypothesized that among surgeons who did not use MRI or PET in the baseline period, those whose peer groups ordered more MRIs or PET scans during the baseline period would be more likely to adopt MRI or PET during the follow-up period (2007– 2009). As predicted, women treated by surgeons whose peer physicians had the highest rates of MRI in the baseline period were more likely to receive MRI than women whose surgeons’ peers used fewer or no MRIs in the baseline period. Similarly, women in the follow-up period were more likely to undergo PET scans if their surgeons’ peers were in the highest category of baseline PET use, although if surgeons’ peers were in intermediate categories of baseline use, there was no association. The increasing availability of rich encounter data from insurance claims has allowed researchers to identify physicians who are connected to each other by virtue of sharing patients. Physicians who interact in caring for patients are likely to share their experiences, beliefs, and ideas regarding clinical care. Such exchange of information has the potential to influence the practice patterns of their colleagues. Prior work has demonstrated that identifying relationships between physicians based on sharing of patients is a valid method for identifying meaningful relationships, such as advice seeking and giving/receiving referrals, between physicians, including primary care physicians, medical specialists, and surgical specialists (6). An important consideration in interpreting the findings of this work is the extent to which peer effects can be distinguished from other contextual influences on a physician, such as a shared practice setting, shared financial incentives, use of common guidelines, and availability of new technologies. Although in this study data were not available regarding