Primary care quality and cost for privately insured patients in and out of US Health Systems: Evidence from four states

Abstract

Objective

To characterize physician health system membership in four states between 2012 and 2016 and to compare primary care quality and cost between in‐system providers and non‐system providers for the commercially insured population.

Data Sources

Physician membership in health systems was obtained from a unique longitudinal database on health systems and matched at the provider level to 2014 all‐payer claims data from Colorado, Massachusetts, Oregon, and Utah.

Study Design

Using an observational study design, we compared physicians in health systems to non‐system physicians located in the same state and geography on average cost of care (risk‐adjusted using the Johns Hopkins’ Adjusted Clinical Grouper), five HEDIS quality measures, one measure of developmental screening, and two Prevention Quality Indicator Measures.

Data Collection/Extraction Methods

Patients in commercial health plans were attributed to a primary care physician accounting for the plurality of office visits. A cohort for each quality measure was constructed based on appropriate measure specifications.

Principal Findings

The share of physicians in health systems increased steadily from 2012 to 2016 and ranged from 48% in Colorado to 63% in Utah in 2016. Compared to physicians not in a system, system physicians performed similarly on most HEDIS quality metrics compared to non‐system physicians. Patients attributed to in‐system physicians had about 40% higher rates (P < .05) of Ambulatory Care Sensitive Admissions (measured in admissions per 100 000:921.33 in‐system vs 674.61 not‐in‐system for acute composite; 2540.91 in‐system vs 1972.17 for chronic composite In‐system providers were associated with $29 (P < .05) higher average per member per month costs (453.37 vs 432.93). Overall, differences in performance by system membership were relatively small compared to differences across states and geography.

Conclusion

A growing share of physicians is part of a health system from 2012 to 2016. Providers in health systems are not delivering primary care more efficiently than non‐system providers for the commercially insured.

What is Known on This Topic

• The effect of hospital ownership of physician practices on quality and cost is mixed.

• Most previous studies use survey data or Medicare claims data. None has examined this relationship using commercial claims data.

What This Study Adds

• An increasing share of both primary care physicians and specialists belongs to integrated health systems in CO, MA, UT, and OR from 2012 to 2016.

• Even highly integrated health systems that have many of the components necessary to coordinate care in a wide range of settings do not have higher performance on a number of primary care quality measures. They only have slightly higher risk‐adjusted per member per month cost among commercially insured patients.

• Overall, we do not find that providers in health systems are delivering primary care more efficiently.

1. INTRODUCTION

Health systems are playing an increasingly important role in health care delivery in the United States. In the past decade, a large number of mergers and acquisitions have brought hospitals and physician practices together under the same organization. Industry consultants and professional societies often laud the expansion of health systems as a promising tool for improving quality, ¹ , ² but research findings so far suggest that consolidated systems are driving up cost. ³ , ⁴ , ⁵ , ⁶ , ⁷ , ⁸ There is a growing literature that compares cost and quality of care between care provided by integrated providers and those provided by independent providers. Previous studies on this topic have used hospital‐based quality data, ⁹ survey data, ¹⁰ , ¹¹ , ¹² data from selected organizations, ¹³ or claims data from the Medicare population ¹⁴ to investigate the impact of integration on quality of care. This study is one of the first to study the relationship between health system integration and primary care quality and cost of care using commercial claims data. Claims data have the advantage of broad coverage and consistent reporting of information on a large range of services. Compared to Medicare claims data, data from a commercially insured population will enable us to examine quality metrics related to conditions that are less relevant in the over‐65 population (such as pediatric care). It also enables us to capture cost variation that has a price component. In addition, this study utilizes a comprehensive data set on health systems. We study financially integrated health systems that have the capacity to integrate multiple dimensions of care, including primary care and specialty services in both inpatient and outpatient settings.

Using a novel dataset on health care providers and health systems matched to metrics constructed from all‐payer‐claims data (APCD) from four states—Colorado, Massachusetts, Oregon, and Utah—this paper documents the prevalence and growth of health systems and compares measures of cost and quality of care between health system and non‐system providers in a commercially insured non‐elderly population.

2. DATA AND METHODS

A retrospective observational study design is used to compare the performance of health system providers to non‐system providers based on measures of quality and cost in four states and across three types of geography (ie, rural, metropolitan, and large metropolitan areas).

2.1. Data

The primary outcome data for the study are cost and quality measures for commercially insured individuals in 2014 computed from all‐payer claims databases (APCDs) for Colorado, Massachusetts, Oregon, and Utah. ¹⁵ The authors did not possess the datasets for the states and did not have permission to pool micro‐level data across states. Member organizations of the Network for Regional Healthcare Improvement in each state computed the quality and cost metrics at the state‐geography‐system level and reported the means to the authors; no variance estimates were provided. We use this aggregated data to perform our analysis. For each outcome, 24 means were provided to the authors. This includes six means from each of the four states: one for in‐system providers and one for non‐system providers in each of three geographies: rural, metropolitan, and large metropolitan. Six of the quality metrics are binary, so the information about their underlying variance is fully embedded in the mean. For each of the non‐binary quality and cost metrics, we assume the within‐category variance for each of the 24 means is the same as the across‐category variance for the 24 categories to assess the statistical significance of our comparisons.

Commercial health insurance can be either fully insured or self‐insured. The raw data from these four APCDs included nearly complete data on the fully insured commercial populations in their states, ranging from 80% in Oregon to nearly 100% in Utah and Massachusetts. Claims from self‐insured plans were included in these analyses to the extent that they were represented in the 2014 APCD data. The inclusion of commercially insured individuals from self‐insured plans varied substantially across states, ranging from 25% of members in self‐insured plans in Colorado and 35% in Oregon to 50% in Utah and Massachusetts. To ensure comparability across states in the study population, claims by supplemental plans, limited liability plans, specific service plans (behavior health, vision, and dental only), and student plans were excluded. Also excluded were plans with incomplete data on enrollment and claims. Our final measures of cost and quality also excluded patients who could not be attributed to a primary care physician because they did not have a primary care visit during the year. After exclusions, our final sample for calculating patient spending includes 20%‐44% of the commercially insured population in these states. The sample for quality metrics varies based on cohort definitions. We present state‐level demographic information using data from the census and other state‐level survey data to provide a context for interpreting our results. This information is presented in Table 1.

TABLE 1.

Sample size description

	CO	MA	OR	UT
# of privately insured patients	3 388 000	4 452 000	2 427 000	2 262 000
Share of population on private insurance	58%	62%	52%	66%
Median HH income	60 940	63 151	58 875	63 383
In APCD sample
N	809 296	1 964 259	500 055	796 412
% Large Metro area	57.91%	62.62%	38.89%	35.89%
% Metro area	27.54%	36.93%	35.44	44.18%
% Rural area	14.55%	0.72%	25.67%	19.93%
# Members included in metric calculation
Cost	809 296	1 964 259	500 055	796 412
Adult avoidance of antibiotics	5854	15 647	6891	5734
ADD initiation phase	1224	11 254	616	1270
AMM acute phase	4827	51 906	5994	5735
Adolescent well care	70 128	269 116	49 408	106 960
Chlamydia screening	18 013	91 671	13 316	21 476
Developmental screening	15 593	62 497	12 383	28 130
PQI acute/chronic composite	32 623	105 583	21 434	56 179

Number of privately insured individuals comes from the National Health Expenditure Data. ²³ Data on the share of population on private insurance are from the Kaiser Family Foundation. ²⁴ Data on Median HH income come from the US Census. ²⁵

Data on delivery system organization come from the Health System and Provider Database (HSPD) created by the National Bureau of Economic Research Center of Excellence for research performed under the AHRQ Comparative Health Systems Performance initiative. ¹⁰ Data from more than 20 sources were combined to identify health systems as groups of commonly owned or managed provider organizations and facilities containing at least one general short‐term acute care hospital, ten primary care physicians, and 50 total physicians co‐located within a single Hospital Referral Region (HRR). This definition aims to capture health systems that have the capability to coordinate services across a broad range of services, including both primary care and specialty care in the inpatient and outpatient settings.

In the HSPD, physicians are identified by their National Provider Identification (NPI) number and assigned to practices based on billing patterns observed in claims data and reassignment of Medicare payments observed in the CMS Provider Enrollment and Chain Ownership System (PECOS), Physicians in the HSPD data were assigned to a practice accounting for the majority of their claims, and physician practices are assigned to health systems. A crosswalk between service provider NPIs and system assignment for 2014 was provided to member organizations of Network for Regional Healthcare Improvement for each state. Claims in 2014 APCD data are then matched to this system assignment based on service provider NPI. We use HSPD panel data to characterize the share of physicians in health systems in four states from 2012 to 2016. A detailed explanation of the data and methods used to create the HSPD may be found at http://data.nber.org/hspd_method/.

For each state, providers were assigned to one of three geographies based on the zip code of their practice location, a zip‐code‐to‐county crosswalk, and CMS’s classification of counties used to assess insurer network adequacy. County types were collapsed into three categories: rural (combining micro, rural, and counties with extreme access considerations), metropolitan, and large metropolitan. ¹⁶

2.2. Variables and definitions

The five HEDIS measures of primary care quality selected for this study include Adolescent Well Care, ADHD Medications Follow‐up Care (initiation phase), Chlamydia Screening, Anti‐Depressant Medication Management (initiation phase), and Adult Avoidance of Antibiotics. ¹⁷ One other measure developed by Oregon Health and Science University, Developmental Screening for Children Ages 1‐3, was also included. Higher rates on these measures indicate higher quality. This set of measures was selected with a focus on younger adults and children, for whom the existing literature on spending and quality using Medicare data could not cover. Although we do not have patient‐level health condition and demographic information to finely risk‐adjust these measures, the HEDIS and developmental screening measures are cohort‐based measures, for which adherence was recommended regardless of other patient characteristics as long as the patient meets the cohort definition. Therefore, unadjusted outcome measures present a meaningful representation of how well the provider is doing in meeting the recommended guidelines. These measures were supplemented with two AHRQ Prevention Quality Indicators (PQIs) that measure the rate of Ambulatory Sensitive Care Hospital Admissions (ASCA) for acute and chronic conditions (PQI 91 and 92 respectively). ¹⁸ For these two measures, lower rates are considered higher quality.

All quality measures were computed from the APCD claims data by Diaz‐Perez and colleagues for individuals attributed to a primary care provider accounting for the plurality of evaluation and management (E&M) visits in 2014. ¹⁵ In the case of a tie, the individual was attributed to the provider with the most recent visit. If an individual had no E&M visits in 2014, attribution was based on 2013 claims data. Providers eligible for attribution included Medical Doctors (MD), Doctors of Osteopathy (DO), Physician Assistants (PA), and Nurse Practitioners (NP) with a primary care focus. Provider specialty is obtained by mapping the service provider NPI listed on a claim to the National Plan and Provider Enumeration System (NPPES) database, which lists all active NPIs along with names, addresses, and taxonomy (specialty) codes. The taxonomy codes are then aggregated to the CMS 2‐digit specialty codes. ¹⁹ A provider is considered to have a primary care focus if their practice maps to one of the following Medicare specialty codes: Physician/General Practice (01), Physician/Family Practice (08), Physician/Internal Medicine (11), Physician/Obstetrics & Gynecology (16), Physician/Pediatric Medicine (37), Physician/Geriatric Medicine (38), Nurse Practitioner (50), Physician Assistant (97).

Cost was measured as the average per member per month (PMPM) allowable charges for all inpatient and outpatient care using the Total Cost of Care software by HealthPartners (https://www.healthpartners.com/hp/about/tcoc/toolkit/index.html). Allowable charges include both the amount the plan paid and the amount the member paid through copay, coinsurance and/or deductible. All administrative claims—for inpatient, outpatient, clinic, ancillary, pharmacy, and all other types of services—contributed to the total cost measure. Comparing cost as well as quality gives us a more holistic understanding of how health system providers may differ from non‐system providers. If we do find a difference in quality, a comparison of cost helps us understand whether the group that delivers higher quality does so with or without higher cost. Although our quality comparisons focus on primary care, quality differences in primary care may affect care in all settings. Therefore, we believe that the Total Cost of Care is the appropriate measure rather than cost incurred in the primary care setting alone. To account for differences across states in demographics of the study populations, the cost measure is risk‐adjusted using the Johns Hopkins’ Adjusted Clinical Grouper (ACG). Details on the computation of cost and quality measures can be found in Diaz‐Perez et al and report by Network for Regional Healthcare Improvement. ¹⁵ , ²⁰

2.3. Analytic approach

We first provide summary statistics at the state level on the share of primary care physicians who are part of a health system, care quality and cost. We then compare the quality and cost of care for privately insured patients who are attributed to health system providers and for those who are attributed to non‐system providers. If patient‐level data were available, we would ideally run the patient‐level regression.

$$Y_{\mathit{ij}}={\beta }_1{\mathrm{SystemIndicator}}_j+{\beta }_2{\mathrm{Geography}}_{\mathit{ij}}+{\beta }_3{\mathrm{State}}_{\mathit{ij}}+{ϵ}_{\mathit{ij}}$$ (1)

where $Y_{\mathit{ij}}$ is a quality metric or cost measure for patient $i$ who is attributed to provider $j$. ${\mathrm{SystemIndicator}}_j$ is a 0/1 variable that indicates whether provider $j$ is in a health system. This regression model assumes that quality or cost is a function of whether the provider is part of a health system, the geography of the provider (rural, metropolitan, large metropolitan) and the provider's state. The geography effect captures average differences in patient and provider characteristics between rural and metro areas. The state effects capture differences in factors that are constant within a state but vary across states, such as state regulations and policies and economic conditions. Our cost measures were risk‐adjusted using the Johns Hopkins’ Adjusted Clinical Grouper (ACG).

Because of data access restrictions, we only have aggregate information at the category level for ${\stackrel{‐}{Y}}_{\mathrm{c}}$, ${\mathrm{SystemIndicator}}_{\mathrm{c}}$, and $X_{\mathrm{c}}$, where ${\stackrel{‐}{Y}}_{\mathrm{c}}$ is the mean quality metric for all providers in Category c for each of the 24 categories, based on state, geography, and whether the provider is in a health system. Observations in each category, by construction, share the same value of ${\mathrm{SystemIndicator}}_{\mathrm{c}}$and the two measures for $X_{\mathrm{c}}$ that we have available: geography (rural, metropolitan, large metropolitan) and state. We could run the aggregate level regression:

$${\stackrel{‐}{Y}}_{\mathrm{c}}={\beta }_1{\mathrm{SystemIndicator}}_{\mathrm{c}}+{\beta }_2{\mathrm{Geography}}_{\mathrm{c}}+{\beta }_3{\mathrm{State}}_{\mathrm{c}}+{ϵ}_{\mathrm{c}}$$ (2)

When weighted with the patient count within each category, this aggregate regression produces the same regression coefficient as the Regression (2). However, the standard errors of the two regressions would be different. The standard errors from Regression (2) would be category‐averaged standard errors and take into account correlated errors at the category level. ²¹ The main drawback of Regression (2) relative to Regression (1) is that we could not include risk adjusters at the individual level.

3. RESULTS

3.1. Cross‐state variation in system participation, cost and quality

The final sample included claims for 20%‐44% of each state's commercially insured population (Table 1). The size of the sample cohorts associated with each of the eight quality measures varies by state, with cohort sizes in Massachusetts being an order of magnitude larger than cohorts in other states. The samples from Massachusetts and Colorado are for populations located predominately in large metro areas, whereas samples from Oregon and Utah include a substantial rural population (over 20%).

Figure 1 shows the share of physicians in each state who were part of a health system in each year from 2012 to 2016. The percentages increased in each state over time and among both primary and specialty care physicians. During the entire time period, Utah had the highest rates of physician system membership, and Colorado had the lowest rates, overall, and for both primary and specialty care. In 2014, 63% of primary care physicians were part of a health system in Utah compared to 53% in Oregon, 52% in Massachusetts and 48% in Colorado.

FIGURE 1.

Share of physicians in health systems. Notes: Data from the Health System and Provider Database. Physicians are assigned to states based on their office location. Data on delivery system organization come from the Health System and Provider Database (HSPD) created by the National Bureau of Economic Research Center of Excellence for research performed under the AHRQ Comparative Health Systems Performance initiative. ¹⁰ Data from more than 20 sources were combined to identify health systems as groups of commonly owned or managed provider organizations and facilities containing at least one general short‐term acute care hospital, ten primary care physicians, and 50 total physicians co‐located within a single Hospital Referral Region (HRR). In the HSPD, physicians are identified by their National Provider Identification (NPI) number and assigned to practices based on billing patterns observed in claims data and reassignment of Medicare payments observed in the CMS Provider Enrollment and Chain Ownership System (PECOS). For this study, physicians were assigned to the practice accounting for the majority of their claims [Color figure can be viewed at wileyonlinelibrary.com]

Variation among states in performance on six cohort‐based measures and two ASCA measures was substantial (Figure 2). For example, the HEDIS score on adult avoidance of antibiotics was equal to 60% for providers in Massachusetts compared to 29% for Utah providers. Performance differences for the highest performing state were approximately twice the rate for the lowest‐performing state for most measures. State performance rankings were relatively stable across measures. Massachusetts led in performance on six out of the eight quality measures.

FIGURE 2.

Cross‐state comparison of cost and quality. Note: All quality measures were computed from the APCD claims data for individuals attributed to a primary care provider accounting for the plurality of evaluation and management (E&M) visits in 2014. For HEDIS metrics, higher rates suggest higher quality. For PQI metrics, lower rates are usually interpreted as better primary care quality [Color figure can be viewed at wileyonlinelibrary.com]

Colorado and Massachusetts spent about the same amount per member per month and cost in these states was 13% above cost in the two lowest‐cost states, Oregon and Utah, which each spent about the same amount.

3.2. Comparison of cost and quality between health system and non‐system providers

Table 2 shows the regression results for Equation (2) using simulated patient‐level data generated from categorical means. Column (9) shows that controlling for state effects and geographic effects, health system providers had $29 higher PMPM cost than providers not in a system. There is no statistically significant difference between in‐system and non‐system primary care providers for most of the cohort‐based quality metrics, except Adult Avoidance of Antibiotics, where health system providers perform better. Even in this case, the estimated differences in performance on quality measures between system and non‐system providers are small relative to the estimated differences among the states and among different geographies. Massachusetts consistently outperformed other states on most of the quality metrics, and providers in large metro areas consistently achieved higher performance than providers in rural areas. On the Ambulatory Care Sensitive Condition measures, health system providers performed worse (had more ACSC admissions) than non‐system providers, and these differences were comparable in magnitude to differences across states and geographies.

TABLE 2.

Regression results from aggregate data (standard errors in parenthesis)

	Cohort‐based Quality Metrics						AMSC admissions		Cost
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
	AAantibio	ADDinit	AMMacut	ADOwell	Chlamy	DEVscrn	PQIacut	PQIchro	PMPM
In‐system	0.078** (0.020)	−0.003 (0.030)	−0.028 (0.028)	−0.003 (0.019)	0.014 (0.018)	0.047 (0.045)	0.002** (0.001)	0.007*** (0.001)	29.436* (11.573)
MA	0.218*** (0.028)	−0.014 (0.043)	0.474*** (0.056)	0.246*** (0.026)	0.239*** (0.027)	−0.268*** (0.056)	−0.002 (0.001)	−0.003 (0.002)	−2.006 (16.013)
OR	0.097** (0.031)	−0.001 (0.071)	0.302*** (0.071)	−0.198*** (0.036)	−0.035 (0.037)	−0.325*** (0.075)	0.002 (0.001)	0.009** (0.003)	−52.303* (22.025)
UT	−0.062 (0.033)	−0.125* (0.058)	0.234** (0.071)	−0.078* (0.030)	−0.106** (0.033)	−0.480*** (0.063)	0.000 (0.001)	−0.015*** (0.002)	−52.439* (19.482)
Metro	0.047 (0.032)	0.039 (0.071)	−0.038 (0.071)	0.012 (0.033)	0.076 (0.036)	0.084 (0.071)	−0.003* (0.001)	−0.003 (0.002)	−62.988* (22.048)
Large metro	0.142*** (0.032)	0.091 (0.070)	−0.041 (0.071)	0.035 (0.033)	0.093* (0.036)	0.130 (0.071)	−0.004** (0.001)	−0.005 (0.002)	−52.890* (21.811)
Constant	0.249*** (0.033)	0.431*** (0.076)	0.445*** (0.079)	0.497*** (0.036)	0.343*** (0.039)	0.521*** (0.078)	0.009*** (0.001)	0.023*** (0.003)	489.248*** (22.926)
N	24	24	24	24	24	24	24	24	24

Regression results from Equation (2) using simulated patient‐level data based one aggregate means at the state*geography*system level. AAantibio stands for Adult Avoidance of Antibiotics, ADDinit stands for ADD Initiation Phase, AMMacut stands for AMM Acute Phase, ADOwell stands for Adolescent Well Care, Chalmy stands for Chlamydia Screening, DEVscrn stands for Developmental Screening, PQIacut and PQIchro stand for PQI Acute Composite and PQI Chronic Composite, respectively. PMPM stands for per member per month cost.

Figure 3 graphs the state and geography adjusted means for health system providers and non‐system providers across all four states, using predicted results from Equation (2) with the state and geography distribution set at the sample average. It visualizes the same results as Table 2. There is a statistically significant difference between health system providers and non‐system providers for only one out of six cohort‐based quality metrics. Health system providers cost slightly more than non‐system providers after adjusting for state and geography. Health system providers have much higher rates of ACSC admissions for both the acute component and the chronic component.

FIGURE 3.

State and geography adjusted comparisons between in‐system and non‐system providers (* indicates statistically significant differences at the 5% level). Note: Bars show the state and geography adjusted means for in‐system and non‐system providers across all four states, using predicted results from regression results from Equation (2) with the state and geography distribution set at the sample average. AAantibio stands for Adult Avoidance of Antibiotics, ADDinit stands for ADD Initiation Phase, AMMacut stands for AMM Acute Phase, ADOwell stands for Adolescent Well Care, Chalmy stands for Chlamydia Screening, DEVscrn stands for Developmental Screening, PQIacut and PQIchro stand for PQI Acute Composite and PQI Chronic Composite, respectively. PMPM stands for per member per month cost [Color figure can be viewed at wileyonlinelibrary.com]

4. DISCUSSION

This paper contributes to our understanding of the role of health systems in health care cost and quality in the commercially insured population. We link a comprehensive data set on health system ownership to quality and per patient cost metrics generated in all‐payer claims data in four states.

A priori, the integration of providers into systems may lead to improvements in primary care quality metrics through a few different channels. ²² First, integrated organizations may have better coordination of care across multiple providers. This mechanism may be important for metrics such as avoidance of antibiotics and medication management. Another is that integrated organizations may have more standardized procedures and may have both the technological and organizational support to make sure routine processes are always performed consistently. This mechanism may be important for metrics such as wellness visits and disease screening. Health system providers may have broader use of integrated health IT system that can provide updates and alerts when a patient is due for a visit.

We found that there is little difference in quality between Health system and non‐system providers for most cohort‐based measures except for Adult Avoidance of Antibiotics, where health system providers performed better. Compared to performance differences between health system and non‐system providers, differences across states and geographies are much greater. Massachusetts led other states in performance on most cohort‐based measures of quality; average risk‐adjusted cost per patient was the highest in Colorado ($ per member per month). Care in rural areas was of lower quality and higher cost. Other factors that vary by state, such as state regulation and incentive programs, may have more of an impact than system membership on quality and per patient cost as measured in this study (eg, MA’s Blue Cross Blue Shield Alternative Quality Contract and the promotion of developmental screening tools by state agencies in Colorado).

We also found that health system providers have higher admissions rates for Ambulatory Care Sensitive Conditions than non‐system providers. This could be that health system providers have sicker patients even conditional on having the same diagnosis. It could also be that health system providers, by having at least one hospital under the same ownership, have greater access to and greater reliance on inpatient care.

4.1. Limitations

The study has several limitations. First, this study uses cross‐sectional data for quality and cost measures that only permit correlation analysis. The results do not imply a causal effect of systems. Second, data quality and availability mean that the quality and cost metrics were generated on less than 50% of the commercially insured population in each state. A few factors contribute to the limited representation. The first limitation is that claims from self‐insured employers are missing from the APCDs in varied degrees. This limitation may become even more salient for more recent analyses using APCD data after the recent Supreme Court ruling in Gobeille v. Liberty Mutual, which allowed self‐insured plans to be exempt from having to report to state APCDs. Another limitation is that relatively low use of services in a non‐elderly population leads to lower attribution rates to primary care physicians. A third limitation is that we only have aggregated data on quality and cost and could not pool microdata across states. This leaves out valuable information on the amount of variation on these metrics within state‐geography‐system cells for continuous metrics. It also limits our ability to adequately adjust for patient case‐mix differences. Our results should not be interpreted as causal. Many factors that vary within geography, states and system membership can contribute to variations in cost and quality, such as patient characteristics, payment system and organizational structure. A fourth limitation of the study is that it focuses on only four states. As more states establish APCDs, future analyses can investigate whether our findings generalize to a broader population.

5. CONCLUSION

Although a growing share of both primary care and specialty physicians is part of a health system from 2012 to 2016, we do not find that providers in health systems are delivering primary care more efficiently. Future studies should explore the mechanisms through which health systems can change care delivery, such as whether integration into health systems increase the standardization of processes and guidelines, change the matching of patients, or accelerate the adoption of technology and best practices.

Supporting information

Author matrix

Appendix S1

Zhou RA, Beaulieu ND, Cutler D. Primary care quality and cost for privately insured patients in and out of US Health Systems: Evidence from four states. Health Serv Res. 2020;55:1198–1106. 10.1111/1475-6773.13590