Physician network connections to specialists and HIV quality of care

Abstract

Objective

To estimate novel measures of generalist physicians’ network connectedness to HIV specialists and their associations with two dimensions of HIV quality of care.

Data Sources

Medicare and Medicaid claims and the American Medical Association Masterfile data on people living with HIV (PLWH) and the physicians providing their HIV care in California between 2007 and 2010.

Study Design

I construct regional patient‐sharing physician networks from the shared treatment of PLWH and calculate (a) measures of network connectedness to all physician types and (b) specialty‐weighted measures to describe connectedness to HIV specialists. Two HIV quality of care outcomes are then evaluated: medication quality (prescribing antiretroviral drugs from at least two drug classes) and monitoring quality (at least two annual HIV virus monitoring scans). Linear probability models estimate the associations between network statistics and the two dimensions of HIV quality of care, and a policy simulation demonstrates the importance of these statistical relationships. These analyses include 16 124 PLWH, 3240 generalists, and 1031 HIV specialists.

Data Collection/Extraction Methods

PLWH are identified from claims for patients with any indication of HIV using an existing algorithm from the literature.

Principal Findings

Generalists’ network connectedness to HIV specialists is positively related with their own HIV medication quality; one additional HIV specialist connection is associated with a 1.46 percentage point (SE 0.42, P < .01) increase in generalist's medication quality. Based on the estimated associations, a simulated policy that increases connectedness between generalists and HIV specialists reduces the annual rate of HIV infections by up to 6%, roughly 290 fewer infections per year. Only network connectedness to all physician types is associated with improved monitoring quality.

Conclusions

Network connectedness to HIV specialists is positively associated with generalists’ HIV medication quality, which suggests that specialists provide clinical support through patient‐sharing for complex treatment protocol.

What is Known on the Topic

• Specialized training and disease‐specific patient volume (experience) are two significant predictors of physicians’ treatment quality in many clinical settings.

• HIV specialists (physicians with Infectious Disease training and/or large annual caseloads of people living with HIV) are known to provide higher quality HIV care.

• Professional networks are key channels for treatment advice and disseminating new medical information among physicians, and greater network connectedness has been shown to influence physicians’ quality of care.

What This Study Adds

• The results show that increased network connectedness to HIV specialists is associated with a higher quality of HIV medication prescribing among generalist physicians.

• Policy simulations demonstrate how increasing the network connectedness of generalists to HIV specialists within a region could significantly increase the number of people living with HIV who receive an appropriate antiretroviral medication regmien.

1. INTRODUCTION

Specialized training and disease‐specific patient volume (hereby called experience) are two of the physician‐level characteristics most strongly associated with the provision of high‐quality medical care. ¹ , ² , ³ The rapid pace of medical innovation is one reason why a specialist's disease‐specific training and frequent conference attendance are associated with higher rates of evidence‐based care. ⁴ Accordingly, patients may be channeled to specialists in pursuit of higher quality of care, which will result in specialists having more patient‐sharing connections within a regional physician network. Similarly, experience is known to improve nonspecialist (or generalist) physicians’ quality of care, ⁵ and greater experience may also increase a generalist's connectedness within their regional patient‐sharing network. This raises the question of whether patient‐sharing connections to specialists are related to generalists’ quality of care independently from a generalist's own experience. Physician networks are a known source of treatment advice and new medical information, ⁶ , ⁷ , ⁸ so greater connectedness to specialists may be an important mechanism for promoting higher quality of care on complex treatment protocols among generalists. Using network analytic methods, I construct novel measures of generalists’ connectedness to specialists based on observations of the patient‐sharing patterns between physicians in the same Hospital Referral Region. These patient‐sharing connections have been shown to identify a large proportion of physicians’ underlying professional relationships, ⁹ , ¹⁰ but the association between patient‐sharing among generalists and specialists and a generalist's quality of care has yet to be investigated.

I examine these relationships in regard to generalists’ human immunodeficiency virus (HIV) quality of care. Physicians who complete a residency program in Infectious Disease or non‐Infectious Disease physicians who have large caseloads of people living with HIV (PLWH) are considered HIV specialists; ¹ , ² all other physicians treating PLWH are identified as generalists. I quantify a generalist's patient‐sharing connectedness to HIV specialists in the same region through three commonly employed measures of network connectedness: degree, betweenness, and eigenvector centrality. While these network centrality measures have previously been estimated across all physicians types and significantly associated with quality of care and patients’ health outcomes, ¹¹ , ¹² , ¹³ , ¹⁴ , ¹⁵ this is the first study to estimate generalists’ connectedness to specialists and associate these specialty‐weighted measures with quality of care. Understanding the relationship between HIV specialist connections and generalists’ HIV quality of care will help to optimally position specialists within regional patient‐sharing networks.

I investigate the role of HIV specialist connections on two different HIV quality of care outcomes. The first outcome identifies whether a generalist prescribes an antiretroviral medication regimen that adheres to a long‐standing HIV clinical care guideline. Success on this medication quality outcome requires a high degree of specialized HIV treatment knowledge, and thus, it is hypothesized to significantly vary with generalists’ connectedness to HIV specialists independently from a generalist's own experience. The second HIV quality of care outcome identifies whether the PLWH attributed to a generalist receive the recommended number of annual HIV monitoring procedures. This second quality outcome is assigned equally to all physicians that treat a given patient in a calendar year, reflects successful care coordination between physicians, and relies less on the application of HIV‐specific medical knowledge. HIV specialist connectedness is hypothesized to have no association with generalists’ performance on this second HIV quality of care outcome. This paper will estimate the associations between connectedness to HIV specialists and these two dimensions of HIV quality of care (medication and monitoring quality) to test the ability and importance of specialist network connections for transferring complex medical knowledge to generalists. To quantify the health and economic benefits from improvements on the medication quality outcome, this paper concludes with a policy simulation exercise that increases the connectedness between generalists and HIV specialists and estimates the associated increase in the number of appropriately prescribed antiretroviral medication regimen by generalists and the resulting reductions in HIV infections and treatment costs.

2. METHODS

2.1. Data and analytic sample

I used 100% Medicare Part B and D and Medicaid Analytic eXtract research identifiable files from the Centers for Medicare and Medicaid Services for all PLWH in California between 2007 and 2010. Prior research with these data identified PLWH from a larger set of beneficiaries with any indications of HIV, ¹⁶ and this population of PLWH was restricted to those with continuous enrollment in either a Medicare or Medicaid fee‐for‐service plan for at least one calendar year in order to observe the necessary procedural, diagnostic, and drug codes for all filled prescriptions.

The combined sample of Medicare, Medicaid, or dually eligible beneficiaries was used to describe the role of generalists’ connectedness to HIV specialists in shaping their quality of HIV care within the population of publicly insured PLWH, who constituted over 50% of PLWH in the United States during the sample period. ¹⁷ However, existing research with these data has documented significant differences across these payers in the demographics of PLWH, physicians’ treatment practices, and HIV quality of care. ¹⁸ , ¹⁹ , ²⁰ , ²¹ Additionally, recent research has demonstrated that estimates of physician networks differ according to the source of insurance claims data. ²² Thus, I separately estimated the model among Medicare and Medicaid claims (see Appendix S1), and given the stability of the estimates across these samples, I present the results for the fully aggregated sample here, which I also used to inform the simulated policy analysis.

The American Medical Association Masterfile provided additional physician characteristics, including physicians’ specialized training and annual ZIP code of primary practice. Primary practice location allowed me to place physicians within one of the 222 Hospital Service Areas in California, as defined by the Dartmouth Health Atlas. ²³ These areas are defined by grouping ZIP codes such that their Medicare residents are most often hospitalized in the same area, which typically results in only one hospital per service area. To capture physician network connections across hospitals, I also assigned physicians to one of the 24 Hospital Referral Regions in California. These broader regions are defined by grouping Hospital Service Areas according to the region where the greatest proportion of major cardiovascular procedures are performed for Medicare residents of each service area. I used patient‐sharing between physicians in the same Hospital Referral Region to estimate physicians’ regional networks. Additional neighborhood characteristics were combined with these data from the American Community Survey and the California Department of Public Health, Office of AIDS. The institutional review board of the University of California, Los Angeles, approved this study protocol.

Based on the HIV medical literature, physicians were identified as HIV specialists if they either completed a residency program in Infectious Disease or if they had considerable experience treating PLWH. Specifically, non‐Infectious Disease physicians with average annual caseloads above the 95th percentile of PLWH (48 patients per year) were identified as HIV specialists, which closely matches several prior studies that used a cutoff of 49 PLWH. ²⁴ I also tested alternative definitions for large annual caseloads, such as the 90th percentile or 35 PLWH, but the results were qualitatively unchanged, which is also consistent with prior studies that found the association with higher quality of HIV care was not significantly altered by changes in the experience definition for an HIV specialist. ¹ , ²⁵

2.2. HIV quality of care outcomes

HIV clinical care guidelines recommend that all antiretroviral medication regimens consist of three active drugs that should be combined from at least two different drug classes. This guideline has been in place since 1996, ²⁶ , ²⁷ yet performance of this guideline has been found to significantly vary across physician types and regions in California. ²⁰ During the study period (2007‐2010), there were 26 antiretroviral medications approved by the US Food and Drug Administration, producing over 3000 possible antiretroviral medication regimens. Each drug is known to have harmful interactions with different comorbid conditions, and/or PLWH’s own genetic mutations and other medications. As PLWH’s health conditions change over time, so to should their antiretroviral regimen, which is why PLWH had an average of 3.2 (SD 1.4) HIV‐specific Evaluation and Monitoring visits per year in this sample. The prescription of an appropriate antiretroviral regimen is thus a complex medical decision requiring a high degree of HIV‐specific treatment knowledge. HIV specialists are more successful at appropriately combining antiretroviral medications ¹⁶ , ²⁰ because they likely know more about the available antiretroviral medication options and the latest information about potentially harmful drug interactions, so patient‐sharing connectedness to HIV specialists may help generalists adapt this medical expertise for their own patients.

In this paper, the medication quality outcome identifies whether a patient receives an appropriate combination of antiretroviral medications conditional on filling a new antiretroviral medication prescription during the calendar year. This (patient × year × physician) observation of medication quality is assigned to the prescribing physician for any new antiretroviral regimen during the year and does not consider prescription refills or prescriptions not attributed to an individual physician. In this way, generalists are only attributed to medication quality outcomes for their own antiretroviral medication prescriptions and not those of an HIV specialist who may have previously treated a shared patient. Additionally, since the medication quality outcome can be satisfied for any patient using one or more of the greater than 3000 available regimens, this outcome can be assessed without knowing the suitability of a particular antiretroviral regimen for a given patient.

The quality of antiretroviral medication prescriptions was measured using the Medicare Part D and Medicaid Analytic eXtract drug files. In these data, each claim is identified as either a new or refill prescription, and the individual drugs that compose an antiretroviral regimen are identified by their National Drug Codes. Additionally, the prescribing physician is identified by their National Provider Identifier, which allowed me to track physicians over time and merge in information from the American Medical Association. Also, I dropped less than 4% of prescriptions that were not assigned to an individual physician but to a facility identifier.

The second HIV quality of care outcome (monitoring quality) describes the frequency and type of HIV virus monitoring provided over a calendar year. During the study period, HIV clinical care guidelines recommended that physicians perform at least two annual Evaluation and Monitoring visits in order to assess HIV viral load through both CD4 T‐cell count and HIV viral load scans. ²⁶ , ²⁷ Conditional on observing two or more HIV‐specific Evaluation and Monitoring visits in a calendar year for a given PLWH, the monitoring quality outcome identifies whether either type of monitoring scan was performed for that patient twice annually. The monitoring quality outcome was assigned to all physicians that performed an HIV‐specific Evaluation and Monitoring visit for the patient at any point in the calendar year. This dimension of HIV care requires a lower degree of HIV‐specific medical knowledge than the medication quality outcome, and instead describes whether a physician is involved in well‐coordinated annual HIV care.

2.3. Physician network analysis

I constructed regional patient‐sharing networks between physicians in the same Hospital Referral Region if they provide at least one HIV‐specific Evaluation and Monitoring visit for the same patient in the same calendar year. These network ties were estimated annually to capture changes in generalists’ connectedness to HIV specialists over the study period. This method of estimating physicians’ professional networks through insurance claims data has been previously validated ¹⁰ and performed within Hospital Referral Regions. ⁸ Three different measures are commonly employed in the medical and social network literature for quantifying network connectedness: degree, betweenness, and eigenvector centrality. ¹³ , ¹⁴ , ¹⁵ , ²⁸ , ²⁹ , ³⁰ , ³¹ I first calculated these measures across all physician types, and then used a weighting procedure to separately calculate these measures with a greater emphasis on connectedness to HIV specialists.

First, degree measures the number of other physicians who share patients with a given physician in their Hospital Referral Region‐specific network. This measure can be calculated as the sum of all other physicians who share patients with a given physician (unweighted) and as the number of HIV specialists who share patients with the given physician (specialty‐weighted). The second measure, betweenness, is proportional to the number of times a physician lies on the shortest path between all possible physician pairs in the Hospital Referral Region network. The specialty‐weighted betweenness measure is calculated by assigning greater weight to the connections along each shortest path that contain an HIV specialist. Specifically, network connections to an HIV specialist were given a weight of one and connections without an HIV specialist were given a weight of 0.2, which is roughly equivalent to the proportion of HIV specialists treating PLWH in this sample. Finally, eigenvector centrality is calculated as the principal eigenvector of the network adjacency matrix. ³² This centrality measure considers both a physician's degree and the degree of the physician's peers, and can also be calculated by applying a weight of 0.2 to network connections without an HIV specialist. For robustness, I also estimated these specialty‐weighted network statistics using weights of 0.3 and 0.1, and calculated the ratio of specialty‐weighted degree to unweighted degree (adjusted degree), and the results are qualitatively unchanged with similar statistical significance (see Appendix S1).

Both the betweenness and eigenvector centrality specialty‐weighted measures capture indirect network connections to HIV specialists that might rely on intermediate generalist connections. In this context, specialty‐weighted betweenness describes how important a given physician is for connecting other physicians to an HIV specialist, while eigenvector centrality indicates how connected a physician is to highly connected HIV specialists. These network statistics could additionally be weighted by the number of shared patients between physician pairs, which is a common method for measuring the strength of physicians’ professional relationships and is an important extension of this approach that warrants future research.

Figure 1 uses an artificial network to visualize the impact of weighting these three network measures by HIV specialist connections. These plots display a small cluster of 29 physicians composed of 25 generalists (circles) and 4 HIV specialists (squares), where lines between two physicians indicate a patient‐sharing relationship. In Panel A, generalists are represented by circles that are sized relative to the indicated unweighted network connectedness measures. In Panel B, generalist circles are sized relative to the specialty‐weighted version of each network measures using a weight of 0.2 for connections without an HIV specialist for the betweenness and eigenvector centrality measures. All network measures were calculated using the igraph package ³³ in R. ³⁴

FIGURE 1.

Each network graph displays the same small cluster of 29 physicians (25 generalists depicted as circles and 4 HIV specialists depicted as squares), where lines indicate a patient‐sharing relationship. In Panel A, the generalist circles are sized relative to the indicated (un‐weighted) network statistic: degree, betweenness, and eigenvector centrality. Panel B shows the impact of weighting the network statistics for HIV specialist connections (using 0.2 to reduce the weight of connection without an HIV specialist), where generalists are sized relative to the specialty‐weighted degree, betweenness, and eigenvector centrality

2.4. Study variables

In addition to these annual measures of generalists’ network connectedness, I calculated generalists’ experience as the number of HIV‐specific Evaluation and Monitoring visits within a calendar year. I also constructed measures of patients’ physical and mental health status. Using the diagnostic codes associated with each HIV‐specific Evaluation and Monitoring visit for a given patient, I calculated the patient's total number of common comorbidities as classified by the Charlson Comorbidity Index. ³⁵ Since diagnostic codes are not reliably entered on every insurance claim, I calculated this measure annually based on all diagnostic codes observed over a calendar year for each patient. Additional patient‐level covariates include identifiers for anxiety disorders, mood disorders, schizophrenia, and other mental health disorders, which were recorded using the Mental Health and Substance Abuse Clinical Classifications Software ³⁶ and were also based on all diagnostic codes in a calendar year for each patient. These were the top three mental health diagnoses among this sample, and because mental health conditions significantly impact PLWH’s engagement in HIV care and adherence to HIV treatment, ²⁰ , ³⁷ , ³⁸ they may also impact physicians’ ability to provide high‐quality HIV care. Finally, I used patients’ residential ZIP code to merge in annual ZIP code tabulation area‐level measures on median household income from the American Community Survey ³⁹ and county‐level statistics on HIV/AIDS prevalence rates from the California Department of Public Health, Office of AIDS. ⁴⁰

2.5. Statistical analyses

Summary statistics and bivariate analyses were used to compare the two HIV quality of care outcomes and physician‐, patient‐, and Hospital Service Area‐level characteristics between generalists and HIV specialists. The associations of each unweighted and specialty‐weighted network connectedness measure with both dimensions of HIV quality of care are estimated separately using linear probability models. To control for the many unobservable physician‐, patient‐, and neighborhood‐level covariates that could confound these relationships, the models estimate physician‐, patient‐, and Hospital Service Area‐year‐level fixed effects. Additionally, the models include patients’ annual physical and mental health status and annual measures of household income and HIV/AIDS prevalence. As done in similar research, ⁴¹ the HIV quality of care outcomes are evaluated in year t while the network connectedness and experience measures are evaluated in year t−1 to help control for reverse causality. In this way, the model is estimated from the within‐physician, within‐patient, and within‐Hospital Service Area‐year variation in the annual network connectedness measures (between 2007 and 2009) and quality of HIV care outcomes (between 2008 and 2010). That is, holding constant the neighborhood, patient, and physician, changes in a generalists’ connectedness and experience in the previous year are related to subsequent changes in the generalist's HIV quality of care. Thus, the model has the following form:

$$Q_{\mathit{ijat}}={\propto }_o+{\beta }_1{\mathrm{Pat}}_{\mathit{it}}+{\beta }_2{\mathrm{Phys}}_{j(t‐1)}+{\beta }_3{\mathrm{Neighborhood}}_{\mathit{at}}+{\delta }_j+{\theta }_i+{\gamma }_{\mathit{at}}+{\epsilon }_{\mathit{ijat}}$$

where $Q_{\mathit{ijat}}$ is an indicator of either medication or monitoring quality for generalist j treating patient i in Hospital Service Area a in year t. The time‐varying patient‐level health characteristics are contained in ${\mathrm{Pat}}_{it,}$ and time‐varying neighborhood measures are contained in ${\mathrm{Neighborhood}}_{\mathit{at}}$. Fixed effects at the physician level, patient level, and Hospital Service Area‐year level are denoted by ${\theta }_j$, ${\delta }_i$, and ${\gamma }_{\mathit{at}}$, respectively. The two main time‐varying physician‐level characteristics, network connectedness and experience, are contained in ${\mathit{Phys}}_{j(t‐1)}$, and they are measured in the year prior (t−1) to all other covariates. The network connectedness and experience measures were standardized to have a mean of zero and standard deviation of one to better compare model estimates for these characteristics. A linear probably model was estimated instead of nonlinear methods, such as logit or probit, because the use of fixed‐effect parameters at these many levels is known to bias parameter estimates. ⁴² Standard errors estimates were double‐clustered at the patient and physician level. ⁴³ Model estimation was performed with the xtreg command in Stata 16. ⁴⁴

The simulation exercise used model estimates to quantify the health and economic benefits of a hypothetical policy to optimize the regional network position of HIV specialists. For this simulation, the existing patient‐sharing network structure was held constant (ie, no new connections were formed, and no connections were broken between any two network positions), so repositioning means that physicians would trade their existing patient‐sharing relationships. That is, when an HIV specialist is repositioned to an optimal network position that was occupied by a generalist, they would take over that generalist's caseload of shared PLWH. I repositioned HIV specialists to minimize the average Hospital Referral Region network distance (number of patient‐sharing connections) between HIV specialists and generalists, and then calculated generalists’ simulated specialty‐weighted eigenvector centrality. The model estimates for the medication quality outcome are used to quantify the associated increase in the number of appropriately prescribed antiretroviral regimen, which is additionally based on the generalists’ observed caseload of PLWH in 2010. Finally, since appropriate antiretroviral regimen are necessary for long‐term HIV viral suppression and viral suppression is known to reduce the risk of sexual transmission by up to 96%, ⁴⁵ I estimated the simulated reduction in new HIV infections and avoided HIV treatment costs across all 24 Hospital Referral Regions in California.

3. RESULTS

A total of 31 879 PLWH were identified in either the Medicare or Medicaid claims between 2007 and 2010 in California. The analyses used observations on 16 124 PLWH who had at least one full calendar year of Medicare, Medicaid, or dual enrollment. These PLWH were treated by 4272 physicians, of whom roughly 24% (1031) were identified as HIV specialists. The average annual number of PLWH among non‐Infectious Disease physicians had a highly right‐skewed distribution, with a median of 2, mean of 7.3, standard deviation of 32.6, and a maximum of 629 patients (see Appendix S1 for a histogram of this distribution).

Table 1 presents physicians’ average medication and monitoring quality outcomes over the sample period (2007‐2010). The first panel shows that the average quality of HIV care did not significantly differ among generalist physicians divided into tertiles based on their average annual caseload of PLWH, and the second panel shows that the average quality of care also did not significantly differ among HIV specialists divided into generalist (non‐Infectious Disease) physicians with large caseloads of PLWH and Infectious Disease specialists. The third panel of Table 1 compares the average HIV quality of care between generalists and HIV specialists, and confirms that HIV specialists’ medication quality was significantly higher than generalists’ (91.1 vs 86.2, P < .001; Table 1), but there was not a statistical difference between specialists and generalists on the monitoring quality outcome (P = .167).

TABLE 1.

HIV quality of care outcomes by physician experience and specialty

	Medication Quality (p.p.)	P value a	Monitoring quality (p.p.)	P value a
A: Generalists by patient experience
Generalists, 1‐2 patients w/ HIV (1st tertile)	85.8 (19.3)	.213	85.4 (32.4)	.468
Generalists, 3‐5 patients w/ HIV (2nd tertile)	86.2 (22.7)		85.8 (31.6)
Generalists, 6‐48 patients w/ HIV (3rd tertile)	86.5 (21.6)		85.5 (33.5)
B: HIV Specialists by definition
Generalists, >48 patients w/ HIV	91.7 (21.1)	.547	88.7 (28.7)	.648
Infectious Disease specialty	90.9 (17.3)		88.9 (31.1)
C: Aggregated physician types
Generalists	86.2 (22.3)	<.001	85.6 (34.2)	.167
HIV specialists	91.1 (18.6)		88.8 (30.1)

The medication quality outcome identifies whether a physician writes a new antiretroviral medication prescription for a patient living with HIV that is consistent with HIV clinical care guidelines: two of the three active drugs in an antiretroviral medication regimen should come from different drug classes. In this table, the medication quality measure is averaged for each physician over their new antiretroviral prescriptions written over the entire sample period (2007‐2010) and is displayed in percentage points (standard deviations are in parentheses). The monitoring quality outcome identifies whether a patient living with HIV receives at least two HIV viral load assessments (either viral load or CD4 T‐cell count scans) in a calendar year conditional on observing at least two or more HIV‐specific Evaluation and Monitoring visits for that patient in the calendar year. This outcome is assigned equally to all physicians observed performing an HIV‐specific Evaluation and Monitoring visit for the patient in the given calendar year, and this measure is averaged for each physician over the full sample period and displayed in percentage points. Panel A displays the average of these two outcomes for generalist physicians divided into tertiles based on their average annual number of patients living with HIV. Panel B shows the average outcomes for HIV specialist physicians who are identified as specialists either for averaging more than 48 patients living with HIV (the 95th percentile of patient experience) per year or for having Infectious Disease specialized training. Finally, Panel C compares the outcomes for all generalist and all HIV specialist physicians.

Abbreviation: p.p., percentage points.

^a P‐values are for tests of equality in each outcome between the physician groups listed within Panels A‐C and are calculated using one‐way ANOVA tests.

Between 2008 and 2010, there was a total of 27 203 (patient × year × physician) observations of generalists’ HIV treatment practices and 16 463 (patient × year × physician) observations of HIV specialists’ treatment practices. Patient, Hospital Service Area, and physician characteristics for these observations of generalists’ and HIV specialists’ HIV care are presented in Table 2. The patients treated by generalists were older (P < .001), more likely to be enrolled in Medicare (P < .001), more likely to have zero comorbidities (P < .001), and less likely to have a diagnosis of anxiety (P < .001), mood disorders (P < .001), schizophrenia (P < .001), or other mental health disorders (P < .001; Table 2). Additionally, generalist physicians were more likely to be female (P < .001), younger (P < .001), and have lower network connectedness measures across all the unweighted and specialty‐weighted network statistics (P < .001 for all two‐way comparisons; Table 2).

TABLE 2.

Characteristics of patients, geographical areas, and physicians by physician specialty

	Generalists	HIV Specialists	P value a
Sample size (patient × year obs.), n	27 203	16 463
Number of unique patients	9852	6272
Patient characteristics
Female, n (%)	4570 (16.8)	2798 (17.0)	.0023
Age, mean (SD) y	53.3 (18.1)	50.7 (17.9)	<.0001
Medicare only, n (%)	5522 (20.3)	2387 (14.5)	<.0001
Medicaid only, n (%)	10 936 (40.2)	7441 (45.2)	<.0001
Dually enrolled, n (%)	10 746 (39.5)	6635 (40.3)	<.0001
No Comorbidities, n (%)	16 703 (61.4)	9910 (60.2)	<.0001
1 Comorbidity, n (%)	7099 (26.1)	4083 (24.8)	<.0001
2 Comorbidities, n (%)	2231 (8.2)	1366 (8.3)	<.0001
3 Comorbidities, n (%)	843 (3.1)	559 (3.4)	<.0001
4+ Comorbidities, n (%)	326 (1.2)	544 (3.3)	<.0001
Diagnosed anxiety, n (%)	1713 (6.3)	1071 (6.5)	<.0001
Diagnosed mood disorders, n (%)	4978 (18.3)	3144 (19.1)	<.0001
Diagnosed schizophrenia, n (%)	1578 (5.8)	1021 (6.2)	<.0001
Other mental health diagnoses, n (%)	1124 (4.5)	839 (5.1)	<.0001
Number of unique Hospital Service Areas	171	153
Hospital service area characteristics b
Household income, median (SD) $	61 842 (4812.4)	60 473 (5418.2)	.0146
HIV/AIDS prevalence, % p. (SD)	0.5 (0.2)	0.6 (0.3)	.0867
Number of unique physicians	3240	1031
Static physician characteristics
Female, n (%)	836 (25.8)	170 (16.5)	<.0001
Age, mean (SD) y	58.1 (12.2)	63.2 (8.8)	<.0001
Infectious disease specialty, n (%)	0 (0)	851 (82.5)	<.0001
Time‐varying physician characteristics b
HIV Eval. & Monitor visits, mean (SD)	1.6 (3.5)	14.2 (41.2)	<.0001
Degree, mean (SD)	2.2 (2.7)	2.6 (3.8)	<.0001
Betweenness, mean (SD)	68.6 (1027.6)	122.7 (1451.2)	<.0001
Eigenvector centrality, mean (SD)	0.18 (1.5)	0.69 (3.2)	<.0001
Specialty‐weighted network statistics b
Degree, mean (SD)	0.5 (0.8)	1.1 (1.8)	<.0001
Betweenness, mean (SD)	14.3 (56.2)	50.8 (314.8)	<.0001
Eigenvector centrality, mean (SD)	0.09 (0.8)	0.42 (2.6)	<.0001

These characteristics are presented for the 43 666 patient × year observations of generalists’ and HIV specialists’ HIV treatment practices between 2007 and 2010. Binary patient‐level characteristics are presented as a percentage of patient‐year observations, and binary physician‐level characteristics are presented as a percentage of physicians. The eigenvector centrality measure is multiplied by 100 for ease of presentation, and all network statistics are standardized to have a mean of zero and standard deviation of one in the main regression analyses.

Abbreviations: % p., percentage points; Eval. & Monitor; evaluation and monitoring; n, number; SD, standard deviation; y, year.

^a P‐values are calculated using one‐way ANOVA and Pearson's chi‐squared tests.

^bTime‐varying statistics are reported for 2009.

The associations between both unweighted and specialty‐weighted network connectedness measures and generalists’ HIV quality of care from the linear probability models are presented in Table 3. First, none of the unweighted network statistics were significantly associated with medication quality, while all three of the specialty‐weighted network statistics were significantly related to higher medication quality in the subsequent year. Roughly one additional patient‐sharing connection to an HIV specialist (a one standard deviation increase in a generalist's specialty‐weighted degree) was associated with a 1.46 percentage point increase in the likelihood that the generalist appropriately prescribed antiretroviral medications in the subsequent year (P < .01; Table 3). Similarly, an increase in generalists’ network connectedness to well‐connected HIV specialists (a one standard deviation increase in specialty‐weighted eigenvector centrality) was associated with a 1.12 percentage point increase in their likelihood of appropriately prescribing antiretroviral medications the following year (P < .01; Table 3). The estimated association of these network measures and medication quality was slightly larger in magnitude than the association between generalists’ annual caseload (experience) and their medication quality, where an increase in annual caseload of approximately 18 additional PLWH was associated with a 1.08 percentage point higher medication quality in the subsequent year (P < .01; Table 3).

TABLE 3.

Association of HIV specialist network connectedness and generalists’ HIV quality of care

	Medication quality Estimate (SE)	P value	Monitoring quality Estimate (SE)	P value
Mean outcome (SD)	86.2 (22.3)		85.6 (34.2)
Unweighted Network Statistics a (separate models)
Degree	0.37 (0.26)	.154	1.52 (0.40)	<.001
Betweenness	0.11 (0.08)	.189	0.10 (0.07)	.286
Eigenvector centrality	0.32 (0.24)	.212	1.50 (0.32)	<.001
Specialty‐weighted Network Statistics a (separate models)
Degree	1.46 (0.42)	.002	0.86 (0.56)	.140
Betweenness	0.24 (0.08)	.006	0.17 (0.14)	.338
Eigenvector centrality	1.12 (0.30)	<.001	0.54 (0.29)	.077
Control variables
Annual patient health status (averaged over all models)	0.12 (0.06)	.054	0.73 (0.26)	.011
Annual standardized caseload (averaged over all models)	1.08 (0.58)	.007	0.67 (0.28)	.002
Observations	27 203		27 203

This table presents coefficient estimates from linear probability models of each HIV quality of care outcome separately regressed on the indicated physician network statistic using ordinary least squares; standard errors (SE) are double‐clustered over multiple observations of the same generalist and the same patient. Each network statistics is standardized to have a mean of zero and a standard deviation of one, where estimates measure percentage point changes in the HIV medication quality and monitoring quality outcomes that are associated with a one standard deviation increase in the indicated network statistic. All models include physician‐, patient‐, and Hospital Service Area‐year‐level fixed effects, annual measures of Hospital Service Area‐median household income and county‐level HIV/AIDS prevalence, and additionally include annual measures of patients’ physical and mental health status and standardized measures of generalists’ annual caseload of patients living with HIV. The average estimate over all models for annual patients’ health status and standardized generalists’ caseload is presented in this table for comparison.

Abbreviations: HSA, hospital service area; SD, standard deviation; SE, standard error.

^aNetwork statistics are lagged by one year relative to the outcome measurement to control for reverse causality.

The statistical significance between the unweighted and specialty‐weighted network connectedness measures flip when examining the monitoring quality outcome. Both the unweighted degree and eigenvector centrality network statistics were significantly associated with increases in monitoring quality in the subsequent year. Connectedness to well‐connected HIV specialists (the specialty‐weighted eigenvector centrality measure) was weakly related to subsequent monitoring quality (P = .08; Table 3), but both of the other two specialty‐weighted network connectedness measures were statistically unrelated to monitoring quality.

The simulated organizational policy that improves generalists’ connectedness to HIV specialists within Hospital Referral Region networks is presented for a single region in California in 2010 in Figure 2. After repositioning the HIV specialists to reduce the average distance between generalists and HIV specialists, the average specialty‐weighted eigenvector centrality was increased by 4.2 (roughly 2.7 standard deviations), which was associated with a 3.0 percentage point increase in generalists’ medication quality. Based on generalists’ annual caseloads of PLWH in 2010, the simulated 3.0 percentage point increase in medication quality translates into an additional 948 PLWH receiving an appropriate antiretroviral regimen. When performed across all 24 Hospital Referral Regions in California, the aggregate impact of this simulated policy is an additional 3,219 PLWH receiving appropriate antiretroviral medications. I then used estimates of the annual number of new HIV infections by Hospital Referral Region in California ⁴⁶ and estimates of antiretroviral medication adherence rates ⁴⁷ to determine that this simulated policy could have reduced annual HIV infections by as much as 6%, roughly 290 fewer infections. In terms of health care cost savings, this represents a present discounted savings of roughly $95 million. ⁴⁸ Even when the costs of increasing an HIV specialist's annual caseload ⁴⁹ and breaking existing patient‐physician relationships ⁵⁰ are taken into consideration, this is still a highly cost‐effective policy for reducing the economic burden of HIV.

FIGURE 2.

The original network graph (on the left) displays a single Hospital Referral Region patient‐sharing network in California in 2010. Generalists are depicted by circles that are sized relative to their shortest network distance (number of patient‐sharing connections) to an HIV specialist. The simulated network repositions HIV specialists to minimize the total distance between all generalists and HIV specialists in the Hospital Referral Region. The reduced distance to HIV specialists is demonstrated by the change in size of the generalist circles

4. DISCUSSION

This paper estimates the association between novel measures of generalists’ connectedness to HIV specialists and two dimensions of generalists’ HIV quality of care. The results show that a significant proportion of generalists’ HIV medication quality (a complex treatment protocol that requires HIV‐specific medical knowledge) can be explained by their network connections to HIV specialists, but not their connections to all physician types. This suggests that generalists’ access to HIV specialist consultations and specialists’ HIV‐specific medical knowledge has a beneficial influence on their HIV treatment quality and that the specialty‐weighted network connectedness measures appear to successfully capture how both direct and indirect connections between generalists and specialists can transfer specialized medical knowledge. Additionally, the larger estimated association of specialty‐weighted eigenvector centrality compared to specialty‐weighted betweenness suggests that the well‐connected HIV specialists are best able to disseminate their medical expertise to generalists, but this finding warrants further investigation.

Interpreting these results as indicative of information transfer between generalists and HIV specialists is further supported by the observed associations between generalists’ monitoring quality (which requires less specialized medical knowledge) and their network connectedness, where the specialty‐weighted measures were not significantly related with this quality measure. Instead, only the unweighted network statistics were significantly associated with monitoring quality. This suggests that annual HIV monitoring is more likely to be completed when the appropriate HIV virus monitoring scans are performed by a greater number of physicians regardless of specialization. These model estimates on both medication and monitoring quality were generated within Hospital Referral Region patient‐sharing networks though, so they capture the benefits of connections across facilities in a local health care market. Future work that is able to identify individual facilities and define network connectedness measures at the facility level will help to better understand the mechanisms underlying these observed associations between network connections and HIV quality of care.

The positive association between network connectedness to specialists and generalists’ performance of a complex HIV quality of care outcome has important implications for the optimal organization of physician networks in this health care setting. ⁴ The complicated HIV medication quality outcome describes generalists’ performance of the primary HIV treatment method, prescribing antiretroviral medications, which has significant repercussions for both the health of PLWH and the spread of the HIV virus. Simulated health care policy that repositions HIV specialists to increase connectedness between generalists and HIV specialists was shown to reduce the annual rate of new HIV infections, and thus prevent additional HIV treatment costs. These improved health outcomes were accomplished by spreading HIV specialists more evenly across each regional network, which resulted in generalists on the periphery of the network having improved access to specializing medical knowledge. Such a policy could be facilitated through telemedicine or other communication methods and does not require physically relocating physicians or other large administrative costs, which makes this a highly cost‐effective policy for improving generalists’ HIV quality of care and reducing the incidence of HIV.

As with many similar network analyses, this study is not without limitations. First, specialist connections were inferred from patient‐sharing relationships observed in administrative data. While these patient‐sharing connections have been validated as a measure of physicians underlying professional relationships, ¹⁰ both the unweighted and specialty‐weighted network connectedness measures may be capturing other confounding factors such as physician or patient preferences. Second, these analyses focused on the provision of HIV care and may not be generalizable to other health care settings and treatment protocols that require different degrees of specialized medical knowledge. Third, the outcomes were constructed at an annual frequency so the timing and sequence of physician and patient interactions within the year were not used in the analyses. Fourth, there may be unobservable time‐varying differences between the patients treated by generalists and HIV specialists that may explain some of the estimated differences in HIV quality of care. Finally, this is an observational study design, so the results should not be interpreted as causal.

Overall, this paper shows that generalists’ connectedness to HIV specialists is associated with HIV medication quality, a treatment outcome that requires specialized medical knowledge and is an important determinant of both patients’ health and HIV transmission. This relationship was estimated conditional on generalists’ experience, which suggests that patient‐sharing network connections to HIV specialists provide additional support for generalists’ HIV medication quality, and also contributes to the known relationship between generalists’ experience and HIV treatment quality. ² , ⁵ Additionally, the simulated policy outlines how networks can be cost‐effectively optimized to improve population health outcomes. These specialty‐weighted network statistics can be readily estimated within other health care settings, and future work should investigate the role of specialist connections on other complex treatment protocols to better understand the mechanisms behind these observed relationships.

CONFLICTS OF INTEREST

There are no conflicts of interest to report.

Supporting information

Author Matrix

Appendix S1

Stecher C. Physician network connections to specialists and HIV quality of care. Health Serv Res. 2021;56:908–918. 10.1111/1475-6773.13628

DATA AVAILABILITY STATEMENT

The data contain confidential, personally identifiable information and are not accessible outside of the Data Use Agreement terms established by CMS.