Abstract
Background
It is unclear whether Medicare data can be used to identify type and degree of collaboration between primary care providers (PCPs) (MDs, nurse practitioners, and physician assistants) in a team care model.
Methods
We surveyed 63 primary care practices in Texas, and linked the survey results to 2015 100% Medicare data. We identified PCP dyads of two providers in Medicare data and compared the results to those from our survey. Sensitivity, specificity, and positive predictive value (PPV) of dyads in Medicare data at different threshold numbers of shared patients were reported. We also identified PCPs who work in the same practice by social network analysis (SNA) of Medicare data and compared the results to the surveys.
Results
With a cutoff of sharing at least 30 patients, the sensitivity of identifying dyads was 27.8%, specificity was 91.7%, and PPV 72.2%. The PPV was higher for MD-NP/PA pairs (84.4%) than for MD-MD pairs (61.5%). At the same cutoff, 90% of PCPs identified in a practice from the survey were also identified by SNA in the corresponding practice. In 5 out of 8 surveyed practices with at least 3 PCPs, about 20% or fewer PCPs identified in the practices by SNA of Medicare data were not identified in the survey.
Conclusion
Medicare data can be used to identify shared care with a low sensitivity and high PPV. Community discovery from Medicare data provided good agreement in identifying members of practices. Adapting network analyses in different contexts needs more validation studies.
Keywords: Medicare, nurse practitioner, physician assistant, social network, primary care
Introduction
Providing primary care to Medicare patients is complex because of the growing population of elders with multi-morbidity, the proliferation of practice guidelines, the rapid development of diagnostic and interventional technology, the increasing complexity of information, and the greater frequency of inter-professional connections. Team-based primary care is a key component of many healthcare delivery reform programs aimed at improving quality of primary care and lowering cost [1]. In these multidisciplinary team-based primary care models, nurse practitioners (NPs) and physician assistants (PAs) have become important members in healthcare delivery to the clinically complex Medicare beneficiaries [2–5]. However, it is unclear what types of MD-NP/PA team care models exist and which types give the best outcomes.
Ways to investigate characteristics and outcomes of team care models include observation of practices, qualitative interviews of practices, and questionnaire-based approaches to collect practice information [6–10]. These approaches are labor intensive and can be cost prohibitive for population-based studies. Another potential approach is to use Medicare claims data, which contains rich information on beneficiaries’ medical utilization and provider information. However, while the identifier for individual provider is available in the data, there are no specific identifiers for individual practices. Lack of an optimal method to capture shared care among providers has been a challenge for using Medicare data to study team-based care. The Tax Identification Number, used to identify the provider to whom payment is made for services, is the identifier available in Medicare. For some medical centers, however, several hundred providers share the same ID, making it impossible to achieve the granularity needed to identify shared care among providers. In general, the Tax ID does not usually identify whether providers with the same ID actually practice together.
One way to address these challenges is to use Social Network Analysis (SNA). SNA applies the mathematical tools of graph theory to allow visualization of a network of actors (e.g., providers) and quantification of the characteristics of the networks [11–12]. Network analyses using Medicare data have increased in number in the past decade. They have been used to identify physician networks within hospital referral regions [13] and groups of physicians who might be suited to become Accountable Care Organizations [14]. They have also been used to show how characteristics of networks and the position of physicians within networks are associated with overall spending and utilization of services [15], and how network measures are associated with adherence to guidelines [16]. However, SNA of Medicare data has not been employed to study primary care delivery, specifically, the involvement of NPs or PAs in team care, nor the degree of shared care with MDs. As the US faces a growing shortage of primary care physicians, using network analyses to understand the role of NPs and PAs in team-based primary care and their impact on processes and outcomes of care is important. This analysis may provide opportunities to identify the primary care team characteristics that best insure the most cost-effective patient care.
To examine the extent to which Medicare data can be used to identify MD-NP/PA team care, we conducted a survey of primary care practices, and linked the survey results to 100% Texas Medicare data. We addressed two study aims: First, we assessed whether Medicare data can be used to identify shared care between two primary care providers (PCPs) in the same practice including MDs, NPs, and PAs. We identified PCP dyads in Medicare data and compared the findings to the results from our survey. Second, we explored whether Medicare data can be used to identify PCPs who work in the same practice by SNA of Medicare data, and compared the results to the information from the surveys.
Methods
Sources of Data
Primary care practice survey
We obtained a list of PCPs, along with their contact information, who are alumni of the family medicine residence, geriatric fellowship, or nurse practitioner programs of an academic medical center in Texas. We made up to 4 attempts to 232 practices with at least one provider from the alumni lists; 169 practices were either not for primary care, were located outside Texas, could not be located, or declined the interview. In total, we completed 63 semi-structured telephone interviews, with 20 solo practices and 43 group practices, from November 2016 to June 2017 (Appendix Table 1). We collected the names of all PCPs (MD, NP, or PA) in each practice. The office manager or other front office staff at each practice, based on their daily experience and recollections, provided information whether PCPs shared patients with other PCPs, and if so, with whom; how sharing works; the degree of sharing (most, about half, or less than a quarter); and whether NPs/PAs had their own patients (Appendix 2). These results were independently reviewed and coded by two authors (Kuo, Goodwin) and the few disagreements resolved by discussion.
Medicare data
We used 2015 100% Texas Medicare claims, which included Medicare beneficiary summary files, Outpatient Standard Analytic Files, and Medicare Carrier files. First, we selected primary care MDs using Centers for Medicare and Medicaid Services (CMS) provider specialty codes including general practitioners (01), family physicians (08), general internists (11), and geriatricians (38). We selected NPs (50) and PAs (97) and later used provider taxonomy codes (Appendix Table 3) to identify those in primary care. Next, we identified primary care visits through Evaluation & Management (E&M) codes for outpatient visits or preventive medicine services (Appendix Table 3). These services could be rendered by offices, hospital outpatient clinics, rural health clinics, and federally qualified health centers [17]. Only PCPs with bills for primary care visits were included. Two PCPs were defined as sharing a patient if they both submitted Medicare claims for primary care visits for that patient.
We linked the PCPs identified in the survey to Medicare data by their National Provider Identifier (NPI). Of the 256 PCPs (178 physicians, 46 NPs, and 32 PAs) in the 63 surveyed practices, 80% billed at least one primary care service to Medicare in 2015. This percentage varied by type of PCP (Table 1).
Table 1.
Number and proportion of providers and dyads in the survey cohort and in the survey-Medicare linked cohort
| Provider/Dyads Type | Survey Cohort N (%) | Survey – Medicare Cohort* N (%) | % in linked cohort** |
|---|---|---|---|
| Providers | 256 | 205 | |
| Solo practitioner - MD | 20 (7.8) | 17 (8.3) | 85% |
| Group practitioner | 236 | 188 | |
| MD | 158 (61.7) | 144 (70.2) | 91.1% |
| NP | 46 (18.0) | 27 (13.2) | 58.7% |
| PA | 32 (12.5) | 17 (8.3) | 53.1% |
| Dyads in group practice | 391 | 252 | |
| MD-MD | 117 (29.9) | 91 (36.1) | 77.8% |
| MD-NP | 161 (41.2) | 86 (34.1) | 53.4% |
| MD-PA | 110 (28.1) | 75 (29.8) | 68.2% |
| NP-NP | 1 (0.3) | ||
| NP-PA | 2 (0.5) |
Only 205 of the 256 providers identified in the survey were identified as providers in the Medicare data.
The percent of providers identified in the survey who also had charge data in Medicare.
Study cohort
To determine whether shared care can be identified in Medicare, the study cohort comprised those 188 PCPs identified from the survey as being in group practices who were also in the Medicare linked data. In a secondary analysis, we included those 17 PCPs from solo practices based on the survey who were also in the Medicare linked data and 8,233 PCPs in the Texas Medicare data who shared patients with one of the above 205 PCPs (188 in group practices and 17 in solo practices).
To explore whether SNA of Medicare data can identify the primary care practices in our survey, the study cohort included 12,004 PCPs who shared at least 9 patients [18] with at least one other PCP in the Texas 2015 Medicare data. We also varied the number of shared patients to at least 20 and at least 30, resulting in 8,350 and 6,511 PCPs, respectively, who met the selection criteria for the study cohort.
Statistical Analysis
Dyad Analyses
For Aim 1, we examined 252 PCP dyads (91 MD-MD, 86 MD-NP, and 75 MD-PA) identified from the survey of the 188 PCPs in the survey-Medicare linked cohort. We treated the survey results about whether two PCPs shared patients as the gold standard for whether a connection existed between two PCPs in the same practice. Following this definition, none of the physicians in a solo practice was in a dyad. We calculated the sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of the determination of existence of a dyad containing two PCPs using Medicare data against the gold standard of the survey findings. We explored how using different threshold numbers of shared patients to define a dyad in the Medicare data affected the performance.
We performed a secondary analysis using Medicare data to identify all the dyads between the PCPs in the survey-Medicare linked data and all other PCPs in the Medicare data. We then repeated the estimation of PPV and NPV for identifying true dyads. We used the sensitivity and specificity estimated from the primary data analysis and set the prevalence of true dyads at 40% for all PCP dyads, 20% for MD-MD dyads, and 60% for MD-NP/PA dyads suggested in the literature [19–20].
Social Network Analysis
For each threshold number of shared patients, we generated a network graph to identify connections between providers. The network included all eligible Medicare providers (12,004, 8,350 and 6,511 for cutoffs of 9, 20, and 30, respectively). We identified communities (our putative practices), tightly interconnected and mutually disjointed subgraphs of the overall graph, using the Walktrap community finding algorithm [21–23]. We chose the Walktrap algorithm because it has good performance on a variety of types of networks [23]. Four steps for the length of random walk was performed because more than half of identified group practices in the survey had 5 or more providers. We also conducted sensitivity analyses with steps 3 and 6, which respectively corresponded to the 50th and 75th percentile of the number of providers in the survey practices. Both the unweighted algorithm and the algorithm weighted by the number of shared patients between two PCPs were used, in each case assigning each PCP to a unique practice. To assess the optimal set of practices within each network graph, modularity was calculated. This score measures the clustering of communities: networks with high modularity have dense connections between PCPs within practices but sparse connections between PCPs in different practices [24]. To demonstrate the agreement between the surveyed practices and practices identified using the community algorithm applied to Medicare data, we reported some external validation measures including recall (the percent of PCPs in each surveyed practice that the algorithm identified as being in the same practice), one minus purity (within SNA defined practices, the percent of PCPs not found in the survey), and F-measure which harmonizes the precision and recall value of each cluster [25]. To visualize the network of PCPs based on the relationships measured using Medicare data and reported in the survey, we presented figures of the network and constituent communities using the igraph package in R.
Results
Dyad Analysis
In the survey, we identified 391 dyads in 63 practices. A dyad was identified when the respondent noted that the two providers tended to share patients. Overall, 252 of the 391 dyads identified in the survey were also found in Medicare data. Table 2 shows the distribution of the number of shared Medicare patients in these 252 dyads. These “true” dyads are those we hoped to rediscover using Medicare data. Among these dyads, the median number of shared Medicare patients was 11 for MD-MD dyads and 10 for MD-NP or MD-PA dyads. Also, 395 pairs of providers were identified in the survey as not sharing patients. Among those 395 “false” dyads by the survey, the median number of shared patients in the Medicare data was three. In secondary data analysis, we also included PCPs in solo practice and PCPs outside the surveyed practices who shared at least one patient in the Medicare data with the PCPs in the survey. The median number of shared patients was one.
Table 2.
Number of shared patients between various types of provider pairs
| Quantile | True dyads from survey |
False dyads* from survey | Dyads# between solo MD in survey and PCPs in Medicare data | Dyads# between PCPs in survey group practice and PCPs in Medicare data | |
|---|---|---|---|---|---|
| MD-MD | MD-NP/PA | ||||
| N | 91 | 161 | 395 | 1009 | 21136 |
| Max | 376 | 283 | 340 | 83 | 464 |
| 95% | 90 | 89 | 39 | 5 | 6 |
| 90% | 79 | 69 | 21 | 2 | 3 |
| Q3 | 46 | 27 | 10 | 1 | 2 |
| Median | 11 | 10 | 3 | 1 | 1 |
| Q1 | 1 | 1 | 0 | 1 | 1 |
| 10% | 0 | 0 | 0 | 1 | 1 |
| 5% | 0 | 0 | 0 | 1 | 1 |
| Min | 0 | 0 | 0 | 1 | 1 |
The providers in these false dyads were from the same group practice.
Dyads were determined by Medicare E&M bills.
Table 3 presents the sensitivity, specificity, PPV, and NPV for identifying dyads in Medicare data, with the survey results as the gold standard using different minimum numbers of shared patients were used to define shared care in Medicare data. The 252 “true” dyad and 325 “false” dyads from the survey data were used in the analyses. When using a cutoff of at least nine patients, the sensitivity was 52%, specificity was 73.2%, and PPV was 60.1%. When the cutoff for identifying shared care increased to 30 shared patients, the sensitivity of identifying dyads dropped to 27.8%, with specificity of 91.7%, and PPV of 72.2%. The PPV was higher for MD-NP/PA (84.4%) pairs than for MD-MD (61.5%) pairs. At all thresholds, detection of MD-MD pairs had a lower PPV, and detection of MD-NP/PA pairs had a lower specificity.
Table 3.
Sensitivity, specificity, PPV, and NPV for dyads within group practices from survey
| Minimum number of shared patients | Overall (N=577 pairs)* |
MD-MD (N= 386 pairs) |
MD-NP/PA (N= 191 pairs) |
|||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Sensitivity | Specificity | PPV | NPV | Sensitivity | Specificity | PPV | NPV | Sensitivity | Specificity | PPV | NPV | |
| 8 | 55.2 | 69.2 | 58.2 | 66.6 | 53.8 | 71.2 | 36.6 | 83.3 | 55.9 | 50.0 | 85.7 | 17.4 |
| 9 | 52.0 | 73.2 | 60.1 | 66.3 | 52.7 | 75.3 | 39.7 | 83.8 | 51.6 | 53.3 | 85.6 | 17.0 |
| 10 | 51.2 | 75.1 | 61.4 | 66.5 | 51.6 | 77.3 | 41.2 | 83.8 | 50.9 | 53.3 | 85.4 | 16.8 |
| 11 | 49.6 | 79.1 | 64.8 | 66.9 | 50.5 | 81.4 | 45.5 | 84.2 | 49.1 | 56.7 | 85.9 | 17.2 |
| 12 | 48.0 | 81.5 | 66.9 | 66.9 | 49.5 | 83.7 | 48.4 | 84.3 | 47.2 | 60.0 | 86.4 | 17.5 |
| 13 | 46.4 | 82.5 | 67.2 | 66.5 | 49.5 | 84.4 | 49.5 | 84.4 | 44.7 | 63.3 | 86.7 | 17.6 |
| 14 | 44.8 | 83.4 | 67.7 | 66.1 | 47.3 | 85.1 | 49.4 | 83.9 | 43.5 | 66.7 | 87.5 | 18.0 |
| 15 | 42.1 | 84.9 | 68.4 | 65.4 | 45.1 | 86.4 | 50.6 | 83.6 | 40.4 | 70.0 | 87.8 | 17.9 |
| 16 | 40.5 | 86.5 | 69.9 | 65.2 | 44.0 | 88.1 | 53.3 | 83.6 | 38.5 | 70.0 | 87.3 | 17.5 |
| 17 | 38.9 | 87.1 | 70.0 | 64.8 | 41.8 | 88.8 | 53.5 | 83.2 | 37.3 | 70.0 | 87.0 | 17.2 |
| 18 | 36.9 | 87.7 | 69.9 | 64.2 | 40.7 | 89.5 | 54.4 | 83.0 | 34.8 | 70.0 | 86.2 | 16.7 |
| 19 | 34.9 | 88.6 | 70.4 | 63.7 | 40.7 | 90.5 | 56.9 | 83.2 | 31.7 | 70.0 | 85.0 | 16.0 |
| 20 | 34.9 | 89.2 | 71.5 | 63.9 | 40.7 | 90.8 | 57.8 | 83.2 | 31.7 | 73.3 | 86.4 | 16.7 |
| 21 | 33.3 | 89.2 | 70.6 | 63.3 | 38.5 | 90.8 | 56.5 | 82.7 | 30.4 | 73.3 | 86.0 | 16.4 |
| 22 | 32.1 | 89.8 | 71.1 | 63.1 | 36.3 | 91.5 | 56.9 | 82.3 | 29.8 | 73.3 | 85.7 | 16.3 |
| 23 | 31.3 | 90.2 | 71.2 | 62.9 | 36.3 | 91.9 | 57.9 | 82.4 | 28.6 | 73.3 | 85.2 | 16.1 |
| 24 | 31.0 | 90.5 | 71.6 | 62.8 | 36.3 | 92.2 | 58.9 | 82.4 | 28.0 | 73.3 | 84.9 | 15.9 |
| 25 | 31.0 | 91.4 | 73.6 | 63.1 | 36.3 | 92.9 | 61.1 | 82.5 | 28.0 | 76.7 | 86.5 | 16.5 |
| 26 | 30.2 | 91.4 | 73.1 | 62.8 | 36.3 | 92.9 | 61.1 | 82.5 | 26.7 | 76.7 | 86.0 | 16.3 |
| 27 | 29.8 | 91.4 | 72.8 | 62.7 | 36.3 | 92.9 | 61.1 | 82.5 | 26.1 | 76.7 | 85.7 | 16.2 |
| 28 | 28.6 | 91.4 | 72.0 | 62.3 | 35.2 | 92.9 | 60.4 | 82.3 | 24.8 | 76.7 | 85.1 | 16.0 |
| 29 | 28.2 | 91.4 | 71.7 | 62.1 | 35.2 | 92.9 | 60.4 | 82.3 | 24.2 | 76.7 | 84.8 | 15.9 |
| 30 | 27.8 | 91.7 | 72.2 | 62.1 | 35.2 | 93.2 | 61.5 | 82.3 | 23.6 | 76.7 | 84.4 | 15.8 |
Include 252 true dyads and 325 false dyads from survey. 70 NP/PA-NP/PA pairs from the 395 “false” dyads were not excluded because such dyads were not found in the surveys.
In the secondary data analysis, included all dyads involving a PCP in the survey-Medicare linked data and any PCP in the Medicare data, the PPV was 90.7% for all PCP dyads, 76.2% for MD-MD dyads, and 96.4% for MD-NP/PA dyads, using the cutoff of at least nine patients. When the cutoff was set at 30 patients, the PPV was ≥92% for all three types of dyads (Appendix 4a–4c).
Social Network Analysis
We applied SNA to 12,004 PCPs identified in Texas Medicare data. We used various cutoffs of the minimum number of patients shared by two PCPs in order to count as a connection. Table 4 summarizes the results. With a cutoff of at least nine patients shared between two PCPs, the unweighted algorithm identified 1,513 practices among 12,004 PCPs. The number of PCPs per practice ranged from 2 to 576, with a median of 3 and an inter-quartile range of 3. With the algorithm weighted by the number of shared patients between two PCPs, we found 1,780 practices with the number of PCPs per practice ranging from 2 to 678 and the same median and inter-quartile range as from the unweighted algorithm. We also assessed the modularity measuring the clustering of communities. The unweighted algorithm had an average modularity of 0.88, compared to 0.91 for the weighted algorithm. When we increased the cutoff to at least 30 patients for 6,511 PCPs, the unweighted and weighted algorithms identified 1,368 and 1,511 practices with a modularity of 0.94 and 0.95, respectively. The number of PCPs per practice in this weighted analyses ranged from 2 to 329 with a median of 3 and an inter-quartile range of 2.
Table 4.
Number of identified communities from social network analyses (SNA) with different minimum number of shared patients between two PCPs. The comparisons of PCPs at the four selected practices between survey and SNA.
| Surveyed Practice ID |
||||
|---|---|---|---|---|
| 6 | 10 | 13 | 42 | |
| True no. of MDs in surveyed practice | 11 | 4 | 4 | 4 |
| True no. of NP/PAs in surveyed practice | 1 | 3 | 0 | 2 |
| At least 9 shared patients in Medicare data, unweighted (Modularity = 0.88; Total number of provider = 12004; Total identified community = 1513 with no singleton) | ||||
| No. of remaining MDs in surveyed practice# | 11 | 4 | 4 | 4 |
| No. of remaining NP/PAs in surveyed practice# | 1 | 3 | 0 | 1 |
| No. PCPs in both surveyed practice and community identified by SNA | 12 | 7 | 4 | 5 |
| Size (total number of PCPs) of the community identified by SNA | 20 | 43 | 125 | 32 |
| At least 9 shared patients in Medicare data, weighted (Modularity = 0.91; Total number of provider = 12004; Total identified community = 1780 with 6 singleton) | ||||
| No. of remaining MDs in surveyed practice# | 11 | 4 | 4 | 4 |
| No. of remaining NP/PAs in surveyed practice# | 1 | 3 | 0 | 1 |
| No. PCPs in both surveyed practice and community identified by SNA | 12 | 6 | 4 | 5 |
| Size (total number of PCPs) of the community identified by SNA | 20 | 43 | 122 | 35 |
| At least 20 shared patients in Medicare data, unweighted (Modularity = 0.92; Total number of provider = 8350; Total identified community = 1595 with no singleton) | ||||
| No. of remaining MDs in surveyed practice# | 11 | 4 | 4 | 4 |
| No. of remaining NP/PAs in surveyed practice# | 1 | 3 | 0 | 1 |
| No. PCPs in both surveyed practice and community identified by SNA | 12 | 6 | 4 | 5 |
| Size (total number of PCPs) of the community identified by SNA | 15 | 30 | 21 | 17 |
| At least 20 shared patients in Medicare data, weighted (Modularity = 0.93; Total number of provider = 8350; Total identified community = 1708 with no singleton) | ||||
| No. of remaining MDs in surveyed practice# | 11 | 4 | 4 | 4 |
| No. of remaining NP/PAs in surveyed practice# | 1 | 3 | 0 | 1 |
| No. PCPs in both surveyed practice and community identified by SNA | 12 | 6 | 4 | 5 |
| Size (total number of PCPs) of the community identified by SNA | 15 | 8 | 21 | 13 |
| At least 30 shared patients in Medicare data, unweighted (Modularity = 0.94; Total number of provider = 6511; Total identified community = 1368 with no singleton) | ||||
| No. of remaining MDs in surveyed practice# | 10 | 3 | 4 | 4 |
| No. of remaining NP/PAs in surveyed practice# | 1 | 3 | 0 | 1 |
| No. PCPs in both surveyed practice and community identified by SNA | 11 | 5 | 4 | 5 |
| Size (total number of PCPs) of the community identified by SNA | 14 | 14 | 21 | 11 |
| At least 30 shared patients in Medicare data, weighted (Modularity = 0.95; Total number of provider = 6511; Total identified community = 1511 with no singleton) | ||||
| No. of remaining MDs in surveyed practice# | 10 | 3 | 4 | 4 |
| No. of remaining NP/PAs in surveyed practice# | 1 | 3 | 0 | 1 |
| No. PCPs in both surveyed practice and community identified by SNA | 11 | 5 | 4 | 5 |
| Size (total number of PCPs) of the community identified by SNA | 14 | 5 | 21 | 6 |
The true number of MD/NP/PAs in the surveyed practice may change depending on the cutoff of the number of shared patients between providers.
Figure 1 shows the practice structures based on the survey data compared to the structure identified from SNA of the Medicare data for the four surveyed practices with at least four PCPs. The relationships among the five PCPs identified from SNA at practice 10 was similar to that reported in the survey. The SNA result from practice 42 shows the relationships among MDs, and one MD and an NP, which were not reported in the survey. In the other two practices, the practice structures were very different between the survey and SNA.
Figure 1.
Community structures based on a survey of primary care practice (left side) and network analyses of Medicare data with at least 30 shared patients between two PCPs (right side) for selected practices. PCP, primary care provider; MD, medical doctor; NP, nurse practitioner; PA, physician assistant.
Table 4 also summarizes the agreement between surveyed practices and practices identified from SNA of Medicare data. Each Medicare practice which had a majority of PCPs in the correspondent survey practice was selected. Using a cutoff of at least nine shared patients, the sizes of those four practices identified in Medicare were much larger than that of the corresponding practices in our surveyed data. For example, the number of PCPs at practice 6 was identified in survey data as 12; however, SNA identified this practice as having these 12 PCPs and an additional 8. This difference was even larger for the other three practices. When using a cutoff of at least 30 shared patients for the SNA of Medicare, fewer PCPs were identified as in the practice who were not also in the corresponding practice from the survey (three PCPs at practice 6, and one PCP at practice 42).
Most PCPs identified by survey as being in a practice were also in the practice as identified by SNA of Medicare data. All 12 PCPs in practice 6 and all 7 PCPs in practice 10 identified by survey were in their correspondent SNA practice, when we used a cutoff of at least nine patients. When the cutoff was 30, these numbers changed to 11 of 12 PCPs in practice 6 and 5 of 7 PCPs in practice 10. The cutoff did not impact practices 13 and 42. Our results from SNA were based on 4-step random walks. When we repeated the weighted algorithm with a cutoff of 30 patients and different numbers of steps for the random walks, the 4 studied practices were identical using a step of 3, and 3 of 4 studied practices were identical using a step of 6.
We also applied the cutoff of 30 to another four survey practices with 3 PCPs. In 6 of these 8 survey practices, the recall measure was greater than 90% indicating 90% of PCPs identified as in a practice from the survey were found in the corresponding practices identified by SNA. In 5 of 8 survey practices, the purity measure was greater than 80%, indicating about 20% or less of PCPs identified in the practice by SNA were not found in the survey practice. Among 6 practices with enough survey information to identify purity, all had F-measure greater than 75% (Appendix Table 5).
Discussion
This study attempted to build on prior research using Medicare data to characterize physician networks [13–15]. Instead of characterizing overall network patterns, we asked whether Medicare data could identify episodes where two primary care providers shared patients, and also whether Medicare data could accurately identify providers in the same primary care practices. Overall, the sensitivity of using Medicare data to identify shared care is low, but the PPV is high. Setting a cutoff for identifying shared care of at least 30 shared patients yielded a PPV of over 96% for identification of “true” dyads, when the prevalence of shared patients was 40% in the community. However, the corresponding sensitivity was only about 28%. We also found that practices identified from network analyses modestly agree with those identified by survey. Most PCPs identified in a practice from the survey were found in the corresponding practice identified by SNA; a small portion of PCPs identified by SNA as in the practice could not be found in the practice from the survey.
Barnett et al. reported that identifying collaborative relationships within Medicare data using a cutoff of nine shared patients could recapture 82% of actual relationships discovered in a survey of 386 office-based physicians [18]. The sensitivity and PPV corresponding to nine shared patients were 10.9% and 76.9%, respectively, differing from our sensitivity of 52% and PPV of 60.1% using the same threshold. In the previous study, the setting was a large academic medical center and several outlying community hospitals. Seventy-two percent of providers were specialists and only physicians were included. We studied 63 primary care practices that included MDs, NPs, and PAs. Many recent studies have employed SNA using various cutoffs to identify collaborative care for different clinical areas [13, 26–29].
Our previous studies on primary care delivery defined shared care between MDs and NPs as patients receiving care from both types of providers [30–31]. In the current study, we show the potential of using Medicare data to study shared care, with a cutoff of 30 shared patients between providers to indicate collaborative care. Researchers studying local care practices and include other type of providers should consider using 30 shared patients as a cutoff. Although the dyads identified from Medicare data are specific with this cutoff, many examples of shared care will be missed due to low sensitivity.
The study has limitations. First, shared care was identified using Medicare data in predominantly older patients. Patterns of shared care might differ in younger patients, or those with commercial insurance. Second, both surveyed practices and Medicare data were restricted to Texas. Further analyses are thus needed to allow generalization to other states and to examine any similarity or difference in team care characteristics between states. Third, we might not have captured all primary care visits provided by NPs or PAs, because Medicare allows physicians to submit E&M charges for a split or shared visit in which both the physician and NP/PA treated a patient. In such situations, the physicians, rather than the NP/PA, would normally submit the charge, because NPs/PAs have a 15% lower reimbursement rate [32]. This situation presumably contributed to the low sensitivity in identifying MD-NP/PA dyads. Fourth, our surveys were conducted in 2016–2017; however, we used 2015 Medicare data. The gap between sources of data may also explain the low sensitivity observed in dyad analyses. Fifth, the information from the primary care practice survey was obtained from the office managers or front desk staffs, from their memory. We didn’t interview providers directly about whether they shared patients with other providers in their practice. Last, we had only four surveyed practices with at least four providers. Therefore, we were unable to thoroughly study the SNA-specific characteristics of these practices. This topic will require surveying practices identified from SNA of Medicare data in further studies. In addition, we do not know the nature of the “shared care” identified either in the survey or in the Medicare data. For example, NPs and Pas can act as a substitute for a physician in providing primary care or they may bring complementary skills to patient care in a true interdisciplinary care model [33].
In conclusion, our results demonstrate that identifying shared care, defined as sharing at least 30 patients between two providers in Medicare data, had high PPV in identifying dyads verified by survey. Further, community discovery using this cutoff provided good agreement in practices. Studies comparing outcomes of care for patients under such care can be performed; however, due to low sensitivity, one needs to carefully select the comparators deemed not to have received shared care in order to determine comparative effectiveness. Our findings also indicate that adapting network analyses in different contexts needs more validation studies. The interactions are different between PCPs and specialists and between physicians and advanced practitioners. Further, such relationships may also differ between diseases and between networks at local communities and large health care systems.
Supplementary Material
Acknowledgments
The authors have no financial, personal or potential conflicts of interest to disclose. This work was supported by grants R01-HS020642 from the Agency for Healthcare Research and Quality, P30-AG024832 and UL1TR001439 from the National Institutes of Health.
References
- 1.Mitchell P, Wynia M, Golden R, McNellis B, Okun S, Webb CE, Rohrbach V, Von Kohorn I. 2012. Core principles & values of effective team-based health care. Discussion Paper, Institute of Medicine, Washington, DC: www.iom.edu/tbc [Google Scholar]
- 2.Schottenfeld L, Petersen D, Peikes D, Ricciardi R, Burak H, McNellis R, Genevro J. Creating Patient-Centered Team-Based Primary Care. AHRQ Pub. No. 16–0002-EF. Rockville, MD: Agency for Healthcare Research and Quality; March 2016 [Google Scholar]
- 3.Wohler DM, Liaw W. Team-Based Primary Care – Opportunities and Challenges. Starfield Summit; Washington DC: April 23–26, 2016. (Access July 24, 2018, https://www.graham-center.org/content/dam/rgc/documents/publications-reports/reports/StarfieldSummit_Report_TeamBasedPrimaryCare.pdf) [Google Scholar]
- 4.Grumbach K Bodenheimer T Can health care teams improve primary care practices? JAMA. 2004. Mar 10:291(10):1246–51. [DOI] [PubMed] [Google Scholar]
- 5.Howard J, Miller WL, Willard-Grace R, Burger ES, Kelleher KJ, Nutting PA, Hahn KA, Crabtree BF. Creating and sustaining care teams in primary care: perspectives from innovative patient-centered medical homes. Qual Manag Health Care. 2018. Jul/Sep; 27(3): 123–129. [DOI] [PubMed] [Google Scholar]
- 6.Morgan S, Pullon S, McKinlay E. Observation of interprofessional collaborative practice in primary care team: An integrative literature review. Int J Nurs Stud. 2015. July; 52(7):127–30. [DOI] [PubMed] [Google Scholar]
- 7.Mundt MP, Gilchrist VJ, Fleming MF, Zakletskaia LI, Tuan WJ, Beasley JW. Effects of primary care team social networks on quality of care and costs for patients with cardiovascular disease. Ann Fam Med. 2015. March;13(2):139–48. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Quinlan E, Robertson S. The communicative power of nurse practitioners in multidisciplinary primary healthcare teams. J Am Assoc Nurse Pract. 2013; 25(2):91–102 [DOI] [PubMed] [Google Scholar]
- 9.Holtrop JS, Ruland S, Diaz S, Morrato EH, Jones E. Using Social Network Analysis to Examine the Effect of Care Management Structure on Chronic Disease Management Communication Within Primary Care. J Gen Intern Med. 2018. May;33(5):612–620 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Leach B, Morgan P, Strand de Oliveira J, Hull S, Ostbye T, Everett C. Primary care multidisciplinary teams in practice: a qualitative study. BMC Fam Pract. 2017. December 29; 18(1):115. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Wasserman S, Faust K. Social Network Analysis: Methods and Applications. 1994. Cambridge: University Press. [Google Scholar]
- 12.Scott J Social Network Analysis: A Handbook (2nd ed). 2000. SAGE Publications. [Google Scholar]
- 13.Landon BE, Keating NL, Barnett ML, Onnela JP, Paul S, O’Malley AJ, Keegan T, Christakis NA. Variation in patient-sharing networks of physicians across the United States. JAMA. 2012; 308(3):265–73. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Landon BE, Onnela JP, Keating NL, Barnett ML, O’Malley AJ, Keegan T, Christakis NA. Med Care. 2013;51(6):715–21. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Landon BE, Keating NL, Onnela JP, Zaslavsky AM, Christakis NA, O’Malley AJ. Patient-Sharing Networks of Physicians and Health Care Utilization and Spending Among Medicare Beneficiaries. JAMA Intern Med. 2018;178(1):66–73. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Moen L, Bynum JP, Austin AM, Skinner JS, Chakraborti G, O’Malley AJ. Assessing Variation in Implantable Cardioverter Defibrillator Therapy Guideline Adherence with Physician and Hospital Patient-sharing Networks. Med Care. 2018; 56(4):350–57. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Health Resources and Services Administration. Primary Care Service Area Version 3.1 Methods. https://datawarehouse.hrsa.gov/DataDownload/PCSA/2010/PCSA_Version3.1_methods_9_17_2013.pdf. Published September 17, 2013 Accessed September 26, 2017.
- 18.Barnett ML, Landon BE, O’Malley AJ, Keating NL, Christakis NA. Mapping physician networks with self-reported and administrative data. Health Serv Res. 2011;46(5):1592–1609 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Donelan K, DesRoches CM, Dittus RS, Buerhaus P. Perspectives of physicians and nurse practitioners on primary care practice. N Engl J Med. 2013; 368(20): 1898–906. [DOI] [PubMed] [Google Scholar]
- 20.Hing E, Kurtzman E, Lau DT, Taplin C, Bindman AB. Characteristics of Primary Care Physicians in Patient-centered Medical Home Practice: United States, 2013. National Health Statistics Reports. 2017. 101:1–8. [PubMed] [Google Scholar]
- 21.Pons P, Latapy M. Computing Communities in Large Networks Using Random Walks In Computer and Information Sciences - ISCIS 2005. ISCIS 2005. Lecture Notes in Computer Science, vol 3733 Springer, Berlin, Heidelberg [Google Scholar]
- 22.Chejara P, Godfrey WW. Comparative Analysis of Community Detection Algorithms. 2017. Conference on Information and Communication Technology (CICT17) https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8340627 [Google Scholar]
- 23.Yang Z, Algesheimer R, Tessone CJ. A Comparative Analysis of Community Detection Algorithms on Artificial Networks. Scientific Repot. 2016. 6:30750 | DOI: 10.1038/srep30750 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Newman MEJ. Modularity and community structure in networks. PNAS. 2006; 103(23):8577–8582. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Zaki MJ, & Meira W Data Mining and Analysis. 2014. Cambridge University Press. [Google Scholar]
- 26.Carson Matthew B., Scholtens Denise M., Frailey Conor N., Gravenor Stephanie J., Kricke Gayle E., Soulakis Nicholas D.. An Outcome-Weighted Network Model for Characterizing Collaboration. PLoS One. 2016; 11(10): e0163861. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Ong Mei-Sing, Olson Karen L., Chadwick Laura, Liu Chunfu, Mandl Kenneth D.. The Impact of Provider Networks on the Co-prescriptions of Interacting Drugs: A Claims-based Analysis. Drug Saf. 2017. March; 40(3): 263–272 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Geva A, Olson KL, Liu C, Mandl KD. Provider Connectedness to Other Providers Reduces Risk of Readmission After Hospitalization for Heart Failure. Med Care Res Rev. 2017. July 1:1077558717718626. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Pollack CE, Wang H, Bekelman JE, Weissman G, Epstein AE,Liao K, Dugoff EH, Armstrong K. Physician social networks and variation in rates of complications after radical prostatectomy. Value Health. 2014; 17(5):611–8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Xue Y, Goodwin JS, Adhikari D, Raji MA, Kuo YF. Trends in Primary Care Provision to Medicare Beneficiaries by Physicians, Nurse Practitioners, or Physician Assistants: 2008–2014. J Prim Care Community Health. 2017;8(4):256–263 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Kuo YF, Adhikari D, Eke CG, Goodwin JS, Raji MA. Processes and Outcomes of Congestive Heart Failure Care by Different Types of Primary Care Models. J Card Fail. 2018;24(1):9–18 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Medicare claims processing manual: Chapter 12 Physicians/Nonphysician practitioners. (Accessed July 22, 2018 at http://www.cms.gov/manuals/downloads/clm104c12.pdf).
- 33.Everett CM, Thorpe CT, Palta M, Carayon P, Gilchrist VJ, Smith MA. Division of primary care services between physicians, physician assistants, and nurse practitioners for older patients with diabetes. Med Care Res Rev. 2013. October, 70(5):531–41. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.