Individual behavior is to a large extent shaped through interactions with the environment. The social environment in particular – including peer networks, classmates and workplace colleagues – has been widely studied by labor economists to understand and quantify important drivers of productivity (see, e.g., Sacerdote, 2001, Falk and Ichino, 2006, Mas and Moretti, 2009). In healthcare, a growing literature in health economics has sought to understand causes of variations in physician practice styles and their consequences for care quality, healthcare utilization and overall health system efficiency (see, e.g., Grytten and Sørensen, 2003, Epstein and Nicholson, 2009, Currie et al., 2016, Epstein et al., 2016, Van Parys, 2016, Molitor, 2018, Cutler et al., 2019, Currie and MacLeod, 2020, Fadlon and Van Parys, 2020, Weng et al., 2020).1,2 However, despite calls for effective policies that seek to reduce inappropriate variations in healthcare utilization (see, e.g., OECD, 2014, Brownlee et al., 2017), the scientific evidence on the impact of healthcare practitioners’ social environment on clinical practice behavior and patient outcomes has so far been scant.3

This paper seeks to add to the literature on the determinants of provider practice styles by studying how physicians’ treatment choices are influenced by their practice environment and the consequences these choices have for their patients. To this end, we make two major contributions that so far have been largely overlooked in the literature. First, we propose a method to decompose the environmental effect into a physical and a social component, corresponding to a hospital-specific and a peer group-specific component. As we argue further below, this is an important distinction to make since the two components provide very different implications for policy. Second, by relating data on physicians’ treatment choices to optimal management and associated patient outcomes, we are able to gauge and directly measure the impact of environmentally induced variation in physician treatment behavior on changes in the appropriateness, treatment costs, and the quality of care received by patients. This is in contrast to most existing studies on physician practice styles, which often rely solely on quantitative measures to evaluate practice heterogeneity.

To provide an empirical framework for identification and consistent estimation of causal effects, we apply and extend the physician migration approach used by Molitor (2018) in the important context of hospital treatment of patients suffering from coronary heart disease. Specifically, we focus on stent choice in coronary angioplasty — a medical procedure used to widen blocked or narrowed blood vessels in the heart.4 We identify physicians who move (migrate) across hospitals and relate the variation in the rate of use of a specific stent type between the physician’s pre-move (origin) and post-move (destination) hospitals to changes in the physician’s own stent use across time in a difference-in-differences model.

To estimate the model, we use rich administrative data from the Swedish Coronary Angiography and Angioplasty Register (SCAAR) on all percutaneous coronary interventions (PCI) performed in Sweden between 2004 and 2013 and study how interventional cardiologists’ choices between the bare metal stent (BMS) and the drug-eluting stent (DES) are determined by their environment.5 Since the procedure, PCI, is identically performed irrespective of the stent type used, this context provides an essentially ideal setting to study how the practice environment shapes physician preferences for treatment.

While empirical evidence on the extent to which physician practice styles are influenced by their work environment is informative, it does not per se convey much detail on which environmental factors are the drivers of such changes. Yet, such knowledge could be important. For example, physical, or hospital-specific, factors may be less informative about the malleability of physicians’ underlying preferences if the possibility to operate in line with such preferences is restricted by factors beyond the individual physician’s control, such as resource constraints, staff micromanagement, or hospital-specific guidelines. In contrast, social, or peer group-specific, factors are more directly related to the adjustment of physician beliefs for which much of the economic literature on physician practice styles lies at the heart of (see, e.g., Epstein and Nicholson, 2009).

To address this important question, we propose and implement a method to decompose the combined impact of the environment on physician treatment styles into a hospital-specific and a peer group-specific factor by exploiting quasi-random variation on physicians working together on given days. Specifically, given sufficient practice style variation among migrating physicians’ coworkers (peers) within a hospital, the inclusion of hospital fixed effects in our econometric model will effectively purge all time-invariant hospital-specific variation in practice styles across hospitals from the analysis. Any remaining practice variation will consequently be derived from changes in the migrating physicians’ coworker mix. Thus, resulting estimates of the environmental effect with and without hospital fixed effects gauge the relative magnitude of the adjustment in physician practice style arising from hospital- and peer group-specific factors, respectively.

Our estimation results show that cardiologists’ DES use in angioplasty treatments are strongly determined by the practice style of the hospital they currently work in. Migrating cardiologists rapidly adapt to their prevailing practice environment after relocation by changing their DES use with on average half a percentage point for each percentage point difference in DES utilization rates between the origin and destination hospitals. This result is robust to a set of alternative specifications and close to the corresponding estimate found by Molitor (2018). Furthermore, when decomposing the overall effect into a hospital-specific and a peer group-specific effect, we find that each component is responsible for roughly half of the practice style adjustment. To assess the extent of heterogeneity in response across cardiologists, we also provide results from a series of split sample regressions which reveal that our results are mainly derived from younger cardiologists and cardiologists moving to more DES-intensive environments.

In contrast, we find no empirical evidence to support the hypothesis that environmentally induced changes in migrating physicians’ practice styles had important consequences for the care quality received by patients. In addition to analyzing a set of adverse clinical events related to the medical procedure, we employ a machine learning algorithm to classify appropriate stent choices for each case based on out-of-sample predictions from non-migrating cardiologists and a rich set of patient and clinical characteristics. While our analysis does not reveal important systematic impacts on patient health as a result of changes in their physician’s practice environment, we do find that migrating physicians are more likely to inappropriately apply DES after their move. This result suggests that the environmentally induced changes in physicians’ practice behavior are mainly based on marginal “gray-zone” cases who run little risk of serious adverse medical events as a consequence of such choices. Moreover, a back-of-the-envelope calculation of the potential monetary savings from following the most efficient treatment approach suggests that the average migrating cardiologist incurred an additional cost of USD 1,500 per year from inappropriate stent choices, corresponding to roughly one-quarter of the price of a PCI.

One potential concern with our decomposition approach is that migrating physicians are non-randomly matched with peers after they move. This would introduce bias in our estimated parameters if migrants exert some control over whom they are working with and use this control to select coworkers with matching preferences. While this is unlikely to occur in practice, and would lead our estimates to be a lower bound of the true effect if it did, we nevertheless evaluate the robustness of our results to such endogeneity concerns by replacing our measure of practice environment with a synthetic environment. Based on the synthetic control method (see, e.g., Abadie and Gardeazabal, 2003, Abadie et al., 2010, Abadie et al., 2015), we construct an artificial matched comparison group using the sample of non-migrating cardiologists in our data. This method safeguards against estimation bias by comparing practice styles of migrating cardiologists with non-migrating cardiologists who were exposed to similar peer practice environments prior to the relocation. Reassuringly, we find that our estimates are largely robust to the definition of practice environment. In addition, we also estimate placebo models where we replace our main outcome with indicators for patient case-mix to study whether our peer effects are driven by patient, as opposed to stent, selection. The results from this analysis show – in line with the knowledge transfer hypothesis – that the type of patients that cardiologists are treating is unrelated to their peers’ practice styles.

Our findings contribute to the scant literature on peer effects and social learning in healthcare. Social learning is broadly defined as the process of information transmission between economic agents when they observe and interact with each other within their social networks (see, e.g., Lin et al., 1981). In line with our results, Huesch (2011) finds evidence for intra-group spillovers in DES use, suggesting that physicians are influenced by their peers. Furthermore, Nair et al. (2010) study peer effects in physicians’ prescribing choices and find that such behavior is particularly influenced by research-active peers within physician groups. Heijmans et al. (2017) find similar results studying peer effects in cardiovascular risk management in networks with and without opinion leaders. On the other hand, Yang et al. (2014) document only small peer effects in prescription behavior for new drugs among physicians working in the same hospital at the same time. Silver (2021) studies the effect of working in a high-pressure peer group environment in an emergency department on physicians’ treatment behavior and patient outcomes and finds that physicians alter their treatment styles in response to their peer group, affecting care quality. Moreover, Epstein and Nicholson (2009) find that physicians’ treatment styles are responsive to changes in treatment styles of other physicians in the same hospital region in the context of Cesarean sections, but the effect dampens when accounting for common shocks at the hospital level. This is in line with our finding that both hospitals and peers are influential in altering physicians’ practice styles. Finally, Burke et al. (2003) find that patients are more likely to receive certain procedures if an attending physician is in a group that performs these procedures more frequently; and Yuan et al. (2020) show that shared beliefs are crucial for successful implementation of new health technology within a peer network. Complementing these findings, the results from our split sample analyses show that the effects are driven by younger cardiologists and cardiologists moving to more DES-intensive environments.

We also add contextual depth to the more general economic literature on peer effects.6 Several papers have investigated the influence of peers on academic performance, yielding mixed results. While some authors find significant peer effects (Sacerdote, 2001, Zimmerman, 2003), others find no effects at all (Foster, 2006, Lyle, 2007), or effects only for particular subgroups (Stinebrickner and Stinebrickner, 2006). In contrast, there exists strong evidence for positive social spillovers on task-oriented work behavior and productivity in non-academic settings. Mas and Moretti (2009) study peer effects at the workplace by analyzing the productivity of coworkers within the same team. They find evidence of positive productivity spillovers when working with highly productive peers, especially when they interact more frequently. Moreover, in an experimental setting, Falk and Ichino (2006) study individuals working on separate tasks within the sight of one another, finding that the productivity of workers is influenced by the productivity of their peers. These results motivate our approach to use physicians working on the same days as relevant peers in the analysis. Finally, Bandiera et al. (2010) study whether workers’ behaviors are affected by the presence of peers that they are socially tied to, finding that a worker’s productivity is positively correlated with friends’ abilities.

Furthermore, we contribute to the emerging health economics’ literature on physician practice styles. Previous research on this topic has documented mixed results. Grytten and Sørensen (2003) find that primary care physicians’ practice styles are largely stable, while for specialists Molitor (2018) and Weng et al. (2020) show that practice styles are malleable to their environment. Variation in physicians’ treatment styles may also have a lasting influence on patient care (see, e.g., Currie et al., 2016, Kwok, 2019, Fadlon and Van Parys, 2020). We complement these findings by providing empirical evidence that physicians equally strongly react to their social as well as their physical environments and relating how these altered practice styles affect care quality.

Lastly, our results have broad implications for healthcare system efficiency. The finding that physicians’ decisions are influenced by their social environment suggests a rationale for why practice styles cluster in certain areas. While such clustering may generate positive productivity and learning spillovers as in Chandra and Staiger (2007), it also implies that patients may receive suboptimal care depending on the dominating practice style at the admitting healthcare provider. In particular, in supply-sensitive areas of healthcare, where the frequency of use of an activity is related to its local capacity, and where the choice of healthcare provider is subject to restrictions, such as place of residence, substantial allocation inefficiencies may exist. If quality of care is largely insensitive to such variations, as this paper shows in the context of cardiac catheterizations, a more integrated system where inappropriate practice variation can be mitigated through enhanced care coordination, monitoring, and follow-up based on evidence-based clinical guidelines could be vastly resource-saving (Wennberg, 2010). However, broadly defined uniform guidelines may not be the most efficient way to reduce inappropriate healthcare variations when patient populations are clinically diverse. For example, Chan et al. (2022) show that diagnosing pneumonia varies substantially across physicians with different skill levels where less skilled physicians are more likely to choose lower thresholds to reduce the risk of failing to correctly diagnose a patient.7 Similarly, we find that younger and less experienced migrating physicians are more likely to inappropriately apply DES after their move. These findings suggest that investments in training of less experienced physicians to increase their skills may be a cost-efficient alternative to national guidelines to reduce unwarranted resource use.