The Association of Insurance Plan Characteristics with Physician Patient-Sharing Network Structure

Abstract

Professional and social connections among physicians impact patient outcomes, but little is known about how characteristics of insurance plans are associated with physician patient-sharing network structure. We use information from commercially insured enrollees in the 2011 Massachusetts All Payer Claims Database to construct and examine the structure of the physician patient-sharing network using standard and novel social network measures. Using regression analysis, we examine the association of physician patient-sharing network measures with an indicator of whether a patient is enrolled in a health maintenance organization (HMO) or preferred provider organization (PPO), controlling for patient and insurer characteristics and observed health status. We find patients enrolled in HMOs see physicians who are more central and densely embedded in the patient-sharing network. We find HMO patients see PCPs who refer to specialists who are less globally central, even as these specialists are more locally central. Our analysis shows there are small but significant differences in physician patient-sharing network as experienced by patients with HMO versus PPO insurance. Understanding connections between physicians is essential and, similar to previous findings, our results suggest policy choices in the insurance and delivery system that change physician connectivity may have important implications for healthcare delivery, utilization and costs.

Introduction:

Professional and social connections among physicians impact patient outcomes (An et al. 2018; Brunson and Laubenbacher 2018); these connections reflect formal and informal information sharing (Barnett et al. 2011), shared practice patterns (Landon et al. 2012; Geissler et al. 2018), and coordination of care (Pollack et al. 2013). Characteristics of physician patient-sharing networks – constructed as physicians connected by the patients they have in common – are associated with utilization patterns (Barnett et al. 2012; O’Malley et al. 2020), geographic areas (Landon et al. 2012), costs (Barnett et al. 2012), treatment choices (Pollack et al. 2012; Tushman and Nadler 1978; Ong et al. 2016), and end of life care (Barnett et al. 2012).

Key differences between the two most common commercial insurance types, Health Maintenance Organizations (HMO) and Preferred Provider Organizations (PPO), may directly influence physician patient-sharing patterns and network structure. HMO enrollees (1) are only covered when they obtain care from a limited set of physicians and (2) maintain a primary care physician (PCP) from whom they must receive specialist referrals. PPO enrollees are generally covered when they see any physician in a typically broader physician set with or without a PCP referral. PPO enrollees still have incentives, in the form of substantially lower patient cost-sharing, to use a preferred set of physicians under contract with the insurer. Recent comparisons of physician choice sets in HMO versus PPO plans in federal health insurance marketplaces show 56% of HMOs had extra small (<10% of physicians participate) or small (10–24% of physicians participate) provider choice sets while only 31% of PPOs had these very limited choice sets (Polsky et al. 2016). Physicians often treat patients with many insurance types in both public and private insurance; for example, many physicians who treat Medicare fee-for-service enrollees also treat Medicare Advantage enrollees in HMO plans (Boccuti and Neuman 2017; Jacobson et al. 2017).

Consistent with varying patient and physician incentives, research from the mid-1990s found significant differences in referral patterns between patients with HMO and indemnity insurance (Forrest and Reid 1997). In recent research, HMO enrollees had lower rates of specialist visits than PPO enrollees, and specialist visits for HMO enrollees were more likely to be in the same health system as the PCP (Barnett et al. 2018). Differences in referral patterns and treatment patterns may lead to variation in physician connectedness and patient-sharing patterns. However, limited research has examined changes in more global connectedness patterns among physicians treating patients with difference insurance types (Geissler et al. 2016; Agha et al. 2019), which may arise from differential referral thresholds and physician selection.

Physician patient-sharing network structure has implications for healthcare cost and quality (Agha et al. 2019; An et al. 2018; Geissler et al. 2018; Landon et al. 2012; Pollack et al. 2013; Uddin et al. 2012); insurance policies and incentives influencing this structure are important, particularly with ongoing delivery system changes (Davis 2007; McClellan et al. 2010; Rittenhouse et al. 2009). Previous work found physicians treating a high proportion of Medicaid patients occupy less globally central positions in the physician patient-sharing network as measured by eigenvector centrality (Geissler et al. 2016); however, this study did not separately examine insurance types with different plan incentives.

We use comprehensive commercial claims data to construct a physician patient-sharing network and calculate measures of physician patient-sharing network structure to capture local and global connectedness patterns. We analyze these network measures at the patient level and examine differences for individuals enrolled in HMO versus PPO plans accounting for patient characteristics, including health status.

Methods:

Overview:

We construct the physician patient-sharing network of Massachusetts physicians treating commercially insured non-elderly adults. In our data, essentially all physicians treat both HMO and PPO patients (98% of physicians have at least 10% of their patient panel enrolled in each insurance type). We construct a single physician patient-sharing network rather than separate networks for HMO versus PPO patients as research shows including only a single payer (e.g., Medicare) misses a significant fraction of patient-sharing relationships and impacts the construction of network measures (Trogdon et al. 2019; Agha et al. 2019).

After constructing the physician patient-sharing network, we calculate standard and novel measures of network centrality and density. We then assign physician patient-sharing network measures to each patient based on visit-weighted averages, which allows us to analyze the association of these network measures with insurance type at the patient level controlling for patient characteristics. We thus include all patient sharing patterns in the construction of the physician patient-sharing network – even patients for whom we do not have full information on their health status due to short enrollment periods (Ericson et al. 2019).

Data and Sample:

We use the 2011 Massachusetts All Payer Claims Database (APCD) v1.0 (Center for Health Information and Analysis 2013), which contains health insurance enrollment and claims data for commercially insured enrollees in Massachusetts. In the construction of the physician patient-sharing network, we include individuals aged 21–64 years in 2011 enrolled for any length of time in any commercial insurance type. In the patient-level analysis, we restrict to patients continuously enrolled in an HMO or PPO plan for 2011 with at least one “face-to-face” visit (Landon et al. 2012) with an included physician in the physician patient-sharing network; this continuous enrollment requirement ensures that patients in the analytic sample have comparable measures of observed health status (Ericson et al. 2019). The study was reviewed and approved by the Boston University Institutional Review Board (Charles River Campus).

Physician Patient-Sharing Network Construction:

We construct a physician patient-sharing network based on links between physicians created by patients they have in common for “face-to-face” visits (Barnett et al. 2011; Landon et al. 2012). We include claims from physicians categorized as PCPs, medical specialists, and surgical specialists (Centers for Medicare and Medicaid Services 2014); this limits our sample to physicians who have a direct relationship with patients. We restrict to physicians whose modal service ZIP code is in Massachusetts. Following previous literature in which physicians recognize a relationship with another physician if they had nine or more patients in common (Barnett et al. 2011; de Choudhury et al. 2010; DuGoff et al. 2018), we only include ties of nine or more patients. This restriction drops physicians who share fewer than nine patients with any other physician. We also drop physicians who are not in the largest connected component of the network.

Using the physician patient-sharing network, we calculate local and global measures of network structure for each physician. As the most local network measure, we calculate degree - the number of connections a given physician has to other physicians. For example, if Dr. A shares at least 9 patients with Dr. B and with Dr. C, Dr. A’s degree would be 2. To better understand the local network position of a physician controlling for panel size, which varies substantially by specialty and across physicians, we introduce a novel local measure termed normalized degree. This new measure accounts for panel size and physician specialty in a more nuanced way than prior literature (Landon et al. 2012). For a given physician, the normalized degree represents the number of connections above or below what would be predicted based on the physician’s panel size and type (Technical Appendix). The third local network measure is clustering coefficient, which is the probability two physicians connected to a given physician are also connected to one another, capturing whether physicians are embedded in highly dense portions of the network over which information flows readily (Watts and Strogatz 1998).

We also include eigenvector centrality to capture the global position of the physician in the physician patient-sharing network. High values of eigenvector centrality occur for physicians who are not only highly connected themselves, but who are connected to other physicians who are also highly connected, and so on recursively.

PCPs have a major role in patient treatment, as well as in selecting and coordinating specialist care (Bodenheimer et al. 1999; Barnett et al. 2009; Mehrotra et al. 2011; Agha et al. 2019; Geissler and Zeber 2020; Geissler 2020) – and their patterns may particularly vary based on patient insurance type (Agha et al. 2019; Forrest and Reid 1997; Barnett et al. 2018). Accordingly, we develop measures for directly measuring the extent to which a PCP works with highly connected specialists who may have access to more information flow due to their network position. Specifically, for each PCP, we calculate an additional set of measures using the weighted average of the network measure of each specialist who shares patients with a given PCP, where the weight is the number of patients the specialist and PCP share. We refer to these statistics as capturing Referral Centrality. We focus on versions of referral centrality defined using normalized degree (local measure) and eigenvector centrality (global measure). An example is a PCP who is connected to highly globally central specialists (measured by eigenvector centrality) would be exposed to much more information from the network than a PCP who is connected to specialists with low global centrality.

Patient-Level Assignment of Physician Patient-Sharing Network Characteristics

We use patient-level analysis to examine the association between insurance type and network structure for physicians visited by a patient. To assign physician patient-sharing network characteristics to the patient level, we use calculated physician patient-sharing network measures and assign to patient-level observations using visit-weighted averages. For referral centrality measures, this patient-level summarization process is much less relevant, as these measures are calculated only for PCPs; patients typically see either a single PCP or a small number of PCPs.

Patient-Level Regression Analyses

We calculate descriptive statistics at the patient-level, with statistical comparisons between insurance types using chi-squared and t-tests. We then conduct patient-level regression analysis with the outcomes of visit-weighted physician patient-sharing network measures and the primary independent variable of the patient’s insurance type (i.e., HMO versus PPO). We do this first showing unadjusted results, without any controls for patient characteristics. We then estimate regressions controlling for fixed effects for insurer, and controls for patient age, patient sex, and patient 3-digit ZIP code. We also include non-linear age measures (age-squared) and an interaction of age times sex to better fit the relationship of patient characteristics. We consider these results to be “partially adjusted” as they do not include measures of patient health status. We then estimate our preferred regression models, which include these controls as well as a detailed vector of indicator variables of patient health status. To measure patient health status, we include 162 categorical indicators from the Massachusetts Hierarchical Condition Categories (HCC) algorithm;(Massachusetts Health Connector) these indicators are for conditions such as diabetes with acute complications, depression, and congestive heart failure. We show summary HCC risk scores in the descriptive statistics (higher values represent sicker patients).

We conduct two sensitivity analyses. In the first, we assign patients to their modal PCP and control for whether the patient’s PCP treats primarily HMO patients and an interaction term with patient insurance type. We report the combined marginal effects of patient HMO enrollment for those treated by a majority HMO PCP versus those with a PCP with a majority of other insurance types. In a second sensitivity analysis, we drop restrictions that a physician must be connected to at least one other physician and be in the largest connected component. For physicians with degree zero, the eigenvector centrality statistic is undefined (Newman 2010); thus we do not include this measure in the sensitivity analysis. We also do not include the referral centrality measures as these measures are calculated based on the specialists connected to a patient’s PCP and are thus not available for the majority of excluded physicians who were dropped due to not being connected to any other physicians.

Statistical analysis was conducted using SAS 9.3 (Cary, NC) and Stata-MP 12.1 (College Station, TX). Network analyses were implemented with the iGraph package in R and Python (Csardi and Nepusz 2006). An alpha of 0.05 was considered statistically significant.

Results:

The patient analytic sample contains 984,470 continuously insured patients with at least one face-to-face visit with a physician meeting the inclusion criteria for the physician patient-sharing network. We identify 19,034 unique Massachusetts physicians in direct patient care specialties and with face-to-face visits for commercial insurance enrollees; this is the vast majority (91%) of physicians active in patient care in Massachusetts as identified in an American Association of Medical Colleges analysis (Center for Workforce Studies 2011). After excluding physicians not connected to another physician by at least 9 patients (37.3%) and physicians who are not in the largest connected component (2.5%), the analytic sample of patients are assigned network measures from 11,639 physicians.

In the analytic sample of continuously enrolled patients with at least one visit, we find patients with HMO and PPO insurance have similar characteristics. Patients with HMO and PPO insurance have nearly identical health status as measured by average HCC risk score (Table 1). Patients with HMO insurance are slightly younger and more likely to be female; they have more face-to-face visits and visit more unique physicians. All of these differences are statistically significant. Ninety percent of patients are insured by an insurer that has both an HMO and PPO plan.

Table 1:

Descriptive statistics of continuously insured commercial insurance enrollees (2011)

	All	PPO	HMO	P-value
HCC Risk Score	1.34 (3.07)	1.34 (3.11)	1.34 (3.06)	0.97
Age in years	45.04 (11.53)	45.35 (11.41)	44.93 (11.58)	<0.001
Female	56.27	54.92	56.76	< 0.001
Face-to-face Visits Per Patient	5.43 (6.54)	5.13 (6.62)	5.54 (6.50)	<0.001
Unique Physicians Per Patient	2.82 (2.39)	2.70 (2.35)	2.86 (2.40)	<0.001
Unique PCPs Per Patient	1.07 (0.88)	0.97 (0.85)	1.11 (0.88)	<0.001
Unique Specialists Per Patient	1.40 (1.69)	1.38 (1.67)	1.40 (1.70)	<0.001
Unique Surgeons Per Patient	0.24 (0.57)	0.24 (0.57)	0.24 (0.57)	0.13
Number of patients	984,470	262,214	722,256

Note: P-Value by t-test for continuous variables and chi-squared test for categorical variables. Standard deviation in parentheses. HCC Risk Scores are calculated assuming both insurance types are in the “gold” metal tier.

We observe significant differences in the visit-weighted average network measures between patients with HMO versus PPO insurance (Table 2). Patients with HMO insurance are treated by physicians with higher mean values of normalized degree, eigenvector, and clustering coefficient than the patients with PPO. Patients with HMO insurance are treated by physicians who are more globally central (i.e., have higher values of eigenvector centrality) and in more dense neighborhoods (i.e., have higher values of clustering coefficient) of the physician patient-sharing network. By contrast, patients with HMO insurance are treated by physicians with lower degree than patients with PPO, indicating patients with HMO insurance see physicians with fewer direct connections to other physicians.

Table 2:

Visit-weighted descriptive statistics for patient analytic sample

	All	PPO	HMO	P-Value
Weighted Average
Degree	40.82 (32.77)	41.24 (33.93)	40.67 (32.34)	<0.001
Normalized Degree	20.27 (16.02)	19.87 (16.51)	20.41 (15.83)	<0.001
Eigenvector	0.0171 (0.0598)	0.0173 (0.0602)	0.0170 (0.0597)	0.03
Clustering Coefficient †	0.5169 (0.1692)	0.5087 (0.1739)	0.5199 (0.1674)	<0.001
Weighted Average Referral Centrality
Normalized Degree Referral Centrality‡	32.81 (16.67)	31.86 (16.16)	33.13 (16.82)	<0.001
Eigenvector Referral Centrality ‡	0.0244 (0.0638)	0.0261 (0.0658)	0.0238 (0.0630)	<0.001
Number of Patients	984,470	262,214	722,256

Note: P-Value by t-test for continuous variables and chi-squared test for categorical variables. Standard deviation in parentheses.

^†Over 978,557 patients

^‡Over 759,149 patients. HCC Risk Scores are calculated assuming both insurance types are in the “gold” metal tier.

Patient-level regression results show insurance type is associated with visit-weighted averages of physician patient-sharing network statistics in our preferred fully-adjusted regression estimates presented in Table 3, as follows. Degree has a negative and statistically significant association with patient enrollment in HMO insurance. Patients with HMO plans see physicians who are more locally connected and more densely embedded in the network, as shown by positive coefficients on normalized degree and clustering coefficient. There is a positive but non-significant association between HMO enrollment and eigenvector centrality, a measure of global centrality. The magnitude of these statistically significant associations are small, with a 0.9% decrease in degree relative to the mean, a 1.4% increase in clustering coefficient relative to the mean, and a 0.9% increase in normalized degree relative to the mean associated with patient enrollment in HMO versus PPO. We see that the partially adjusted results, which omit controls for patient health status, are similar to the preferred specifications meaning that the associations we find are not primarily driven by selection based on observed patient health; given the stability of the coefficients to detailed health status controls, it is unlikely that selection based on unobserved patient health status fully explains our associations (Oster 2019).

Table 3:

Regression analysis of association between network statistics and insurance type

					Referral Centrality
	Degree	Normalized Degree	Eigenvector	Clust. Coeff.	Normalized Degree	Eigenvector
Panel A: Unadjusted Results
Patient has HMO	−0.567*** (0.076)	0.543*** (0.037)	−0.0003* (0.00014)	0.011*** (0.000)	1.273*** (0.043)	−0.002*** (0.000)
R-Squared	0.000	0.000	0.000	0.001	0.001	0.000
Panel B: Partially Adjusted Results including demographics, insurer fixed effects, and patient 3-digit ZIP code
Patient has HMO	−0.384*** (0.085)	0.233*** (0.041)	0.0002 (0.00015)	0.007*** (0.000)	0.266*** (0.045)	−0.001*** (0.000)
R-Squared	0.035	0.050	0.106	0.031	0.181	0.190
Panel C: Fully Adjusted Results
Patient has HMO	−0.363 *** (0.084)	0.177*** (0.040)	0.0002 (0.00015)	0.007*** (0.00043)	0.279*** (0.045)	−0.0014*** (0.00018)
R-Squared	0.047	0.074	0.110	0.043	0.183	0.193
Number of Observations	984,470	984,470	984,470	978,557	759,149	759,149

Note: Robust standard errors in parentheses.

^*:p < 0.05,

^**:p < 0.01,

^***:p < 0.001 Fully adjusted regressions shown in Panel C include controls for demographics (i.e., age, age squared, sex, an interaction of sex and age), 162 indicator variables for hierarchical condition categories (HCC), insurer fixed effects, and patient 3-digit ZIP code.

We find patients with HMO plans see PCPs who refer to specialists who are more locally connected than patients with PPO plans, as measured by the normalized degree referral centrality (Table 3). By contrast, patients with HMO plans see PCPs who refer to specialists who are less globally connected than patients with PPO plans, as measured by the eigenvector referral centrality. These effects are again statistically significant. Patient enrollment in an HMO versus PPO is associated with a 0.9% increase in normalized degree referral centrality relative to the mean and a 5.9% decrease in eigenvector referral centrality relative to the mean.

In the sensitivity analysis where we include a term for whether a patient’s modal PCP sees a majority HMO patients and its interaction with patient insurance, we find that the interaction between patient insurance status and whether PCP patient panel is majority HMO is not statistically significant in three of the four measures (Appendix Table A1). This suggests that patients’ insurance status is a relatively stable predictor of network measures, as the association between patients’ insurance status and network measures does not differ substantially by whether their PCP’s panel’s majority insurance type is HMO or not. In the sensitivity analysis in which we include physicians who do not have any shared patients with other physicians and those not in the largest connected component, we find broadly similar results in terms of direction and magnitude (Table 4).

Table 4:

Sensitivity analysis including network statistics of all physicians

	Degree	Normalized Degree	Clust. Coeff.
Patient has HMO	−0.065 (0.083)	0.276*** (0.039)	0.012*** (0.00048)
R-Squared	0.053	0.071	0.055
Number of Observations	1,000,955	1,000,955	1,000,955

Note: Robust standard errors in parentheses.

^*:p < 0.05,

^**:p < 0.01,

^***:p < 0.001 All regressions include controls for age, age squared, sex, an interaction of sex and age, 162 indicator variables for hierarchical condition categories (HCC), insurer fixed effects, and patient 3-digit ZIP code.

Discussion:

We conduct a detailed analysis of the association of how patient-sharing network characteristics vary between the physicians seen by patients with HMO versus PPO insurance. At the patient-level, insurance type is associated with variation in the visit-weighted averages of local and global physician patient-sharing network measures. Patients enrolled in HMOs have visit-weighted average physician patient-sharing network characteristics of higher local centrality and lower global centrality, and similar results of greater magnitude for the specialists connected to the patient’s PCPs. This is consistent with previous findings (Agha et al. 2019; Geissler et al. 2016) related to the characteristics of HMO plans likely to impact physician patient-sharing networks, including smaller physician choice sets in HMO plans and the requirement to receive referrals from a PCP.

We introduce several new physician patient-sharing network characteristics to better capture connectedness patterns among and between PCPs and specialists. We find patient enrollment in HMO insurance is positively associated with the visit-weighted average of physicians’ normalized degree measure, showing physicians treating HMO patients share with a broader set of co-treatment partners, accounting for physician type and patient panel size. Secondly, we created two referral centrality statistics and find patients with HMO insurance see PCPs who work with specialists who are highly connected locally (measured via normalized degree referral centrality), but less connected globally (measured via eigenvector referral centrality).

Our analysis of referral centrality suggests HMO enrollment may weakly impact the structure of the physician patient-sharing network by changing how referrals are directed to specialists. Moreover, differences in referral centrality statistics suggests information accessible to PCPs likewise differs, and that patients enrolled in HMOs see PCPs who obtain larger amounts of information from local connections and less information from remote parts of the network, potentially reducing awareness of alternative or emerging treatment choices (Shi et al. 2019). The increase in local connectivity (normalized degree) shows these physicians work directly with a larger number of other physicians than expected based on type and panel size. This is consistent with the PCP, for example, referring to a set of specialists from an affiliated or approved set, rather than referring to a specific specialist. Previous work found patients of PCPs who are connected to more specialists within a given specialty have higher costs and utilization with no increase in quality (Agha et al. 2019) and that high quality PCPs are likely to more carefully select the specialists with whom they work (Milstein and Gilbertson 2009), suggesting this increase in local referral centrality may be associated with less carefully chosen specialist use. By contrast, HMO patients see specialists who are less well globally connected. This is consistent with fewer referrals to specialists who are remote in the network based on, for example, personal relationships or outsized reputation.

The lower global centrality of PCPs seen by HMO patients is consistent with prior research that physicians with more Medicaid patients – insurance known to have a smaller physician choice set than commercial insurance (Decker 2018; Neprash et al. 2018) – have lower eigenvector centrality (Geissler et al. 2016). However, the associations we find have small magnitudes. This indicates that potentially, as with converging patient perceptions of HMOs and PPOs (Dugan 2015), HMOs and PPOs have become more similar over time with respect to patient-sharing patterns. Almost 90% of patients in our sample are insured by an insurer that has both an HMO and PPO plan; as we control for insurer, we are capturing differences in plans for the same insurer rather than simply differences in physician choice sets between insurers. Differences we see in the patient-level analyses reflect the different combinations of physicians (type and network position) involved in a patient’s treatment rather than a completely different set of physicians as might be expected based on the perception of HMOs as limited to integrated delivery systems such as Kaiser Permanente (Crosson 2009). This treatment of both HMO and other insurance types, including PPO, enrollees by physicians is common in Massachusetts as it is nationally. Many physicians treat both Medicare fee-for-service patients as well as Medicare Advantage HMO patients (Boccuti and Neuman 2017; Jacobson et al. 2017), as well as patients enrolled in different insurance types for both commercial and Medicaid insurance.

HMOs and PPOs differ in their reimbursement schemes, with HMOs being more likely than PPOs to pay PCPs on a capitated basis (Zuvekas and Cohen 2010) (although full capitation is rare in our data). HMOs may therefore change the way PCPs care for their patients – for instance, by investing more in preventative care – which may also affect the observed physician patient-sharing patterns. Similarly, payment incentives may change PCPs’ willingness to provide a referral, or change how a referral is directed (i.e., to a high or low cost specialist, within or outside of health system) (Milstein and Gilbertson 2009; Bodenheimer et al. 1999; Barnett et al. 2018). Finally, the provider choice set differs between HMOs and PPOs, which affects which specialists are used, since patients are substantially less likely to choose a specialist not covered by their plan (Kyanko and Busch 2012; Rosenthal et al. 2009). Thus, we might expect physicians treating more PPO patients would refer to more globally central specialists – whose network position enables them to obtain more information from the network – and this is what we observe.

There are several limitations to our analysis. First, our analysis is based in Massachusetts, which may limit generalizability. However, most network analysis is done in highly local areas (e.g., hospital referral regions) (Landon et al. 2012) or using a single insurer (Trogdon et al. 2019) and so using a state-level dataset is of value to measure both geographically close and distant patient-sharing patterns (Geissler et al. 2018; O’Malley et al. 2020). We include fixed effects for patient 3-digit ZIP code to account for geographic differences in patient care and physician patient-sharing network measures in areas with fewer physicians. The data also includes physicians from many different organizations, although we do not observe organizational affiliations which may impact the physician patient-sharing network structure. Second, we estimate associations between insurance type and physician patient-sharing network measures rather than causal relationships, and are not able to establish the mechanisms by which these associations develop. Our findings are not driven by observed patient health status, which is similar between insurance types. While unobserved health status may also differ by insurance type, it would have to differ much more than observed health status to account for our associations, which seems unlikely. Third, our data does not include Medicare or Medicaid patients, so these patient-sharing links are omitted from our analysis. However, previous research related to physician patient-sharing networks has been limited in its ability to compare within and across insurance types due to data limitations (Trogdon et al. 2019; Barnett 2019).

Our analysis makes an important contribution in showing referrals are to physicians – including overall and specialists only – with differing positions in the physician patient-sharing network in terms of network connections and centrality, which have been shown to be associated with healthcare costs (Landon et al. 2012; Agha et al. 2019; Geissler et al. 2019). Similar to previous findings (Landon et al. 2013; Geissler et al. 2016), our results suggest policy choices in the insurance and delivery system that change physician connectivity may have important implications for healthcare delivery, utilization and costs. As continued modifications to insurance characteristics (Kaiser Family Foundation 2018; KFF 2019; Jacobson et al. 2018), delivery systems (DeCamp and Lehmann 2015), and provider incentives occur (Rathi and McWilliams 2019), research to understand impacts of health system changes on patient-sharing patterns will be important to provide high quality, coordinated care, at lower costs.