Michael E. Chernew, PhD

 Understanding variation in spending across organizations, rather than across geographic areas, is important because care is delivered by organizations and interventions increasingly focus on organizations. Accountable care organizations (ACOs) are particularly important to study given their incentives to reduce spending. Analyzing spending differences across ACOs may help identify cost savings opportunities.

 Cross-sectional analysis of Medicare claims.

 We stratified ACOs into quartiles based on the deviation between each ACO’s risk-adjusted spending and average risk-adjusted fee-for-service spending in the same market (hospital referral region). We compared spending between top- and bottom-quartile ACOs on each of 7 major service categories and 10 clinical condition groups to identify areas of potential savings. We simulated spending reductions if ACOs with high adjusted spending reduced spending to the levels of lower-spending ACOs.

 In 2016, geographically adjusted and risk-adjusted total per-beneficiary spending for the highest-spending quartile of ACOs was 14% higher than for ACOs in the lowest quartile. Variation between high- and low-spending ACOs was greatest, at 27%, in the use of skilled nursing facilities—a service category in which ACOs have reduced spending by the greatest percentage. Inpatient care was the largest driver of absolute dollar differences in spending, however, accounting for 37% of the total spread. If spending in ACOs above median adjusted spending were brought down to the median, savings would be 3% to 4%.

 By extending the variations literature to focus on ACOs, we illustrated that meaningful further savings opportunities exist both within and across markets.

There was a 14% difference in risk-adjusted total spending between accountable care organizations (ACOs) in the highest and lowest quartiles of spending relative to local fee-for-service Medicare spending, suggesting meaningful opportunities for savings even after 4 years of the ACO program.

If spending in ACOs with above-median adjusted spending were brought down to the median, savings would be 3% to 4%.

Spending variation by service category suggests savings opportunities in inpatient, postacute, and home health care. Variation in spending for patients by select condition categories indicates substantial savings opportunities in aggregate for prevalent chronic conditions, but the magnitude of variation in dollars per patient is small, suggesting that interventions must be low-cost on a per-beneficiary basis to achieve net savings.

Variation in healthcare spending and utilization exists at every level of the healthcare system, ranging from states to individual physicians within a given practice.

 Yet most of the research characterizing variation in healthcare spending has been focused at the geographic level, as exemplified by the Dartmouth Atlas. Results of geographic analyses argue convincingly that a great deal of spending variation across areas is not connected to patient morbidity or outcomes.

 However, considerable differences in spending exist even within geographic areas,

 and geographic areas are not a natural unit for actionable interventions to reduce spending variation.

 The National Academy of Medicine’s 2013 report on spending variation emphasized the importance of extending analyses of practice pattern variation to “the unit of decision-making,” such as individual providers or provider systems.

Evidence characterizing organizational variation remains relatively limited. Studies of organizational variation in Medicare spending typically focus on a specific clinical context, like outpatient cardiology or oncology practices, or examinations of specific services, like low-value care.

 Evidence from the commercial sector focuses on variation in price, rather than total spending, and generally looks at variation across hospitals rather than organizations such as accountable care organizations (ACOs) that are responsible for population-level spending.

 As in the geographic variation literature, research examining practice pattern variation in provider organizations reveals large and unexplained differences, but none of that work examines variation in total spending at the population level.

ACOs are provider groups that agree to take accountability for spending and quality of care for patients in the ACO contract. ACOs represent an important unit of observation because they are CMS’ most important alternative payment model and the foundation of many current efforts to reduce wasteful medical spending. Little research exists on spending variation in ACOs: One early study reported considerable variation in ACO spending, even after controlling for regional variation and case mix, but did not delve into sources of variation.

 Breaking down sources of spending variation among ACOs could inform where organizations might intervene to reduce avoidable spending.

We examined the deviation between each 2016 ACO’s risk-adjusted spending and average risk-adjusted spending in its market, then stratified ACOs into quartiles based on these deviations. We defined top-quartile ACOs—those with the largest positive deviations, in dollars, from adjusted local market fee-for-service (FFS) spending&mdash;as high spending (quartile 4 [Q4]). We defined bottom-quartile ACOs&mdash;those with the largest negative deviations&mdash;as low spending (quartile 1 [Q1]). We compared spending between Q1 and Q4 ACOs on each of 7 major service categories and 10 clinical condition groups to identify areas of potential savings. Finally, we simulated spending reductions if ACOs with high adjusted spending (within their market or nationally) reduced spending to the levels of lower-spending ACOs.

We analyzed a 100% sample of FFS Medicare beneficiaries, refined to include only beneficiaries continuously enrolled in parts A and B for the full year (or until death) and continuously enrolled in the prior year. We further limited the sample to beneficiaries potentially eligible for enrollment in an ACO: individuals with at least 1 qualifying ambulatory evaluation and management visit in the year of interest.

We analyzed ACOs participating in the Medicare Shared Savings Program (MSSP), the largest of Medicare’s ACO initiatives. We assigned beneficiaries to ACOs using the provider identifiers (taxpayer identification numbers [TINs] and CMS Certification Numbers [CCNs]) defining each ACO in the ACO SSP Provider-Level Research Identifiable File for that year. We assigned beneficiaries to an ACO if it accounted for a larger share of qualifying office visits with a primary care physician than any other ACO or non-ACO TIN/CCN during the year (see 

]). Patients assigned to non-ACO TINs served as FFS controls. As in the MSSP, we used retrospective assignment, assigning patients to ACOs based on utilization in the year of interest.

Our sample included 432 ACOs in 2016. (There were 220 in 2013, 333 in 2014, and 387 in 2015, and 145 ACOs were present all 4 years.) For ACOs present in multiple years, the composition of TINs or CCNs included in the ACO may have changed over time, which we addressed in sensitivity analyses.

Our outcomes of interest were total annual per-beneficiary Part A and Part B Medicare spending and annual per-beneficiary Medicare spending by service subcategory.

In all analyses, we controlled for beneficiary age, sex, race/ethnicity, Medicaid status, disability as the reason for enrollment, and end-stage renal disease. We controlled for comorbidities using prospective hierarchical condition category (HCC) diagnoses and risk scores using CMS 2014 version 12 software for all 4 years, using CMS’ 

International Classification of Diseases, Tenth Revision, Clinical Modification

International Classification of Diseases, Ninth Revision, Clinical Modification

. We identified whether beneficiaries had ever had 1 of 27 chronic conditions identified in the Chronic Conditions Warehouse (CCW). We assigned beneficiaries to a hospital referral region (HRR) based on their residential zip code.

Ranking ACOs by Adjusted Spending Deviations From Their Markets

To rank ACOs by total adjusted spending, we estimated a linear mixed-effects regression model predicting total per-beneficiary spending (eAppendix) as a function of the covariates described previously: HRR fixed effects and ACO random effects. Estimated random effects represent an adjusted mean spending deviation from local market FFS spending for each ACO. In contrast to fixed-effects models, the random-effects model accounts for within-ACO variation in spending in the estimation of the between-ACO variance; without this correction (eg, in a fixed-effects model), the ACO-level variance would be overestimated. We defined an ACO’s market as the HRRs that it serves. ACOs can serve multiple HRRs, so an ACO beneficiary’s spending relative to other FFS beneficiaries in its market is implicitly estimated as a weighted average of the beneficiaries in each HRR served by the ACO. We were thus able to measure each ACO’s spending deviation relative to other ACOs and non-ACO FFS beneficiaries in the ACO’s market.

We grouped ACOs into quartiles based on their risk-adjusted spending deviation from average FFS spending in their market. We used ACOs’ total spending quartile rank in all analyses, including condition or service subcategory analyses.

&#8203;&#8203;&#8203;&#8203;&#8203;&#8203;

We ran a linear regression model with Q1 and Q4 quartile dummies and covariates for beneficiary demographic characteristics, case mix (HCC), and market (HRR). We subtracted the Q4 coefficient (spending attributed to high-spending ACOs) from the Q1 coefficient (spending attributed to low-spending ACOs) to calculate the spread in total spending. We focused on this “spending gap” between the Q1 and Q4 means as a useful summary of the range of the spending deviations.

We repeated the same total spending per beneficiary model described previously for each service category, changing only the dependent variable to subcategory spending rather than total spending. Additionally, we calculated the percentage share of the total spending gap between quartiles attributable to each service category.

We estimated deviations for patient populations with specific conditions, using the 27 CCW conditions to create 10 CCW condition groups (eAppendix). We restricted some conditions, such as hip fracture, to diagnoses within the last year to better isolate spending associated with that diagnosis. We included chronic conditions if the patient had ever received a diagnosis. Using linear regression models with quartile dummies, condition dummies, and interactions between quartiles and conditions, we computed spread in total spending and service category spending between Q1 and Q4 ACOs. We determined the prevalence of conditions in Q4 ACOs to estimate potential savings in aggregate associated with each condition; prevalence was approximately evenly distributed across ACO quartiles and other FFS beneficiaries.

Simulation of Potential Spending Reductions

We simulated cumulative spending reductions in 3 scenarios to crudely measure ACO savings potential. We estimated change in total spending if all ACOs with above-median adjusted spending were able to reduce spending to the level of the median (50th percentile) ACO. We repeated this analysis for thresholds at the 25th and 10th percentiles.

Our primary analyses focused on 2016. We repeated the analyses for 2013 to 2015, as well as for ACOs present in all 4 years. We replicated the service and clinical area analyses (both quartile rankings and spending category variation models) for each year, with and without controlling for geography (HRR). We also simulated savings without HRR controls.

Our sample included approximately 26.6 million eligible beneficiaries in 2016, with 8.3 million beneficiaries attributed to ACOs. Mean ACO membership was 19,179. Adjusted mean total per-beneficiary spending in 2016 was $10,212 for low-spending ACOs and $10,850 for high-spending ACOs. Results for 2013 to 2015 are reported in the eAppendix.

In 2016, mean total risk-adjusted spending was 14% higher for ACOs in the highest quartile relative to those in the lowest (

). Although the relative percentage difference between high and low spending quartiles for inpatient care was only 15%, inpatient spending was the largest service category in dollars; consequently, it accounted for 37% of the absolute spread in total spending (ie, in dollars per beneficiary). Outpatient spending in the upper quartile was 22% greater than that in the lowest quartile, and, because of its large spending share, it was the second largest driver of the overall spending gap, accounting for 32% of the spread in total spending.

In all years, variation on a percentage basis between Q1 and Q4 ACOs was greatest in skilled nursing facility (SNF) and home health. However, because SNF and home health spending accounted for a smaller share of spending than inpatient and outpatient spending, they account for a smaller share of the overall gap in spending.

 summarizes spending and prevalence by condition group for high- and low-spending ACOs. Taking patients with chronic kidney disease (CKD) as an example, the mean risk-adjusted difference in spending for Q4 is $1372 higher than spending for Q1 ACOs. Prevalence of CKD in Q4 ACOs was 22%. If Q4 ACOs reduced incremental spending for patients with CKD to that of Q1 ACOs, Q4 ACOs could achieve cost savings of $302 per beneficiary across the entire ACO population (ie, including people with and without CKD).

We observed substantial spending variation across all condition groups. In general, acute or newly diagnosed CCW condition groups tended to have a high dollar difference, driven by higher mean spending, in per-beneficiary spending between Q1 and Q4 ACOs, but low prevalence. Chronic condition groups tended to have smaller per-beneficiary dollar differences in spending for patients with those diagnoses but high prevalence, and thus they were responsible for greater absolute spending differences between high- and low-performing ACOs when considered over the full ACO population.

If all beneficiaries in ACOs with geographically adjusted and risk-adjusted spending higher than the median ACO were brought down to that threshold, savings would be 3.6% (

Without adjusting for HRR, the between-quartile spread in total risk-adjusted spending deviations for the top and bottom quartiles of ACOs nationally was 10 percentage points larger than the models adjusting for HRR (eAppendix). Our primary analysis is thus conservative in that it assumes that ACOs are more readily able to reduce spending relative to their markets than relative to national benchmarks. On a national basis, if beneficiaries in ACOs above median national ACO spending (adjusted for case mix but not geography) were brought down to that threshold, savings would be 9.2%.

Our results were largely insensitive to the year of analysis (eAppendix): The gap between Q1 and Q4 ACOs in 2016 was $1409 (14% of mean per-beneficiary spending). The analogous figures for 2013, 2014, and 2015 were $1328 (14%), $1364 (14%), and $1475 (15%), respectively. The services contributing to this variation were largely unchanged. ACO quartile rankings were largely stable over time; only 6 ACOs had a year in the highest quartile and a year in the lowest.

This analysis suggests meaningful variation in ACO spending within markets and overall. Despite widespread recognition that high spending in the United States, particularly in the commercial sector, reflects high prices, our savings estimates largely measure variation in utilization because Medicare prices are set administratively and vary much less than commercial prices.

 Thus, our analysis can broadly identify areas of savings opportunity related to utilization. Adjusted within-market spread between Q1 and Q4 ACOs was 14% in 2016, suggesting meaningful opportunities for savings even after 4 years of ACO adoption.

Our analysis suggests that key savings opportunities exist in the inpatient, SNF, and home health care settings, consistent with ACO evaluations reporting spending reductions in these areas.

 ACOs may find savings by changing the site of care (eg, shifting from hospital outpatient department to office settings), which is consistent with observed evidence on ACO savings.

Research involving ACO executives suggests that they have adopted strategies including strengthened primary care, greater care management, and reductions in avoidable ED use—all with an emphasis on high-risk patients.

 However, care management strategies do not appear to be how ACOs, on average, have found savings. Prior research indicates that ACOs have not differentially reduced admissions for ambulatory care—sensitive conditions, nor have reductions in hospitalizations and spending been disproportionately concentrated among high-risk patients.

 ACOs increasingly report focusing on standardizing and systematizing clinical processes.

 And, as ACOs have discovered the savings opportunities in postacute care, they have initiated strategies such as preferred SNF networks.

Spending variation in clinical groups suggests considerable aggregate savings potential among patients with high-prevalence chronic conditions, but the magnitude of spending variation per beneficiary in dollars is small, meaning that interventions aiming for net savings would need to be highly efficient on a per-patient basis. Incremental spending differences between Q1 and Q4 ACOs for people with newly diagnosed and acute conditions are sizeable in dollars per beneficiary, but these conditions are low in prevalence, meaning that aggregate savings would be modest in condition-specific efforts to control costs.

Targeted interventions to improve chronic conditions may not provide a return on investment for ACOs in the short term because per-beneficiary cost savings are modest and the costs of many interventions, such as case management, are high. Further, returns to chronic disease management may be realized in future years, and beneficiary attribution to an ACO is not guaranteed from year to year. Many popular clinical interventions may be cost-effective but not cost-saving.

 The most proximate savings may lie in interventions to reduce utilization of low-value services; this requires developing effective interventions, as existing tools to measure low-value care capture only a small fraction of it.

 Because low-value care is not necessarily confined to the high-risk group, interventions almost certainly need to be broad.

Articles

Analysis of spending differences among accountable care organizations (ACOs) may help identify cost savings opportunities. We examined the magnitude and sources of spending variation among ACOs over 4 years.

Spending Variation Among ACOs in the Medicare Shared Savings Program

To compare supply of and access to physicians and hospitals for low- and high-income communities in the United States following the Affordable Care Act.

Cross-sectional study using 2015 data on physicians and hospitals from CMS, SK&A, commercial claims, and the American Hospital Association.

We computed densities of primary care and specialty physicians in low- and high-income zip code tabulation areas (ZCTAs), further stratified by urban, suburban, and rural categories. We also estimated geodetic distances to the nearest general acute care, teaching, children’s, and psychiatric hospitals for each ZCTA.

Urban low-income ZCTAs were 0.5% more likely to have very low primary care physician density (<5 physicians per 10,000 population) relative to high-income ZCTAs. The disparity was great in rural areas (where the analogous figure was 1.7%) and greatest in suburban areas, where low-income ZCTAs were 7.4% more likely to have very low primary care density relative to high-income ZCTAs. The percentage of urban and suburban low-income ZCTAs with very low specialty physician density was 1.3 to 1.4 times larger than that of high-income ZCTAs. Although the average distance to a general acute care hospital was similar for low- and high-income ZCTAs (median difference range, 0.4-2.1 miles), residents of rural low-income areas had to travel substantially farther to teaching, children’s, and psychiatric hospitals (10.2-35.6 miles).

Our findings document geographic variation in provider supply related to income and suggest that improving access to care may require greater attention to local delivery systems.

The American Journal of Accountable Care. 2019;7(4):11-18

One of the chief goals of the Affordable Care Act (ACA) was to expand access to healthcare to all Americans, particularly through insurance coverage.

 However, while health insurance increases access to care, it does not guarantee such access.

 Specifically, the structural aspects of the local delivery system—supply, location, and organization of healthcare resources&mdash;all vary widely across markets, but they have received much less attention than coverage as a driver of disparate access. Although considerable literature relates availability of primary care to beneficial outcomes and lower costs,

 it does not focus on disparities in access related to income. Although many studies have shown that Medicaid expansions have reduced the number of uninsured Americans,

 few studies have looked at primary care supply after many states’ Medicaid expansions took effect in 2014. The effects of improved financial access for disadvantaged populations depend on the nature of their local delivery systems.

Barriers to access posed by lack of healthcare facilities and trained physicians has been well documented.

 Policies to address scarcity of healthcare resources date back to at least the 1946 Hill-Burton Act. Today, the Health Resources and Services Administration identifies Health Professional Shortage Areas (HPSAs) or Medically Underserved Areas/Populations based on shortage designation criteria that it develops in collaboration with state partners.

However, the HPSA definition of a shortage area is based on context and specific applications, not a broad assessment of community needs. While adequacy of the supply of providers for low-income individuals is largely unknown, dependence on local safety net providers is substantial and potentially precarious

: Today, 1 in 13 Americans obtain care from the 10,000 community health center sites funded by the ACA.

Much of the existing literature relating provider supply to outcomes focuses on rural areas. Prior research suggests that rural populations experience greater difficulties in accessing needed care compared with urban populations.

 Rural areas often encounter challenges in attracting and retaining providers because providers feel “overburdened and underpaid,” among other reasons, compared with those in urban areas.

Yet rural is not synonymous with low income, and quantitative analyses of delivery systems in low-income areas are lacking despite the focus on access expansion policies for low-income populations. Existing research on this topic is largely qualitative and in the form of case studies of high-functioning safety net organizations.

 The large literature on the safety net generally takes the provider as the unit of observation and hence cannot speak to differences among communities in access to care associated with the local availability of providers. There has been little research to characterize healthcare delivery systems in low-income communities nationally.

We use a unique relational database to examine potential barriers to access and to quantify variation in the supply of primary care and specialty physicians, as well as general acute care and specialty hospitals, across the United States. In this paper, we present a large-scale, national description of the local healthcare delivery systems found in low- versus high-income communities following the ACA.

We selected zip code tabulation areas (ZCTAs) as our primary unit of observation, excluding 667 ZCTAs with fewer than 10 residents and 517 ZCTAs with missing income data. Our final database included 32,322 ZCTAs with 316,196,416 residents (99.9% of the US population in 2015).

ZCTAs capture variation in income better than larger units such as primary care service areas (PCSAs). There is often substantial income heterogeneity among the constituent zip codes within a single PCSA, for example. Because individuals may seek care outside of their ZCTAs, we assign to each ZTCA measures of provider supply that encompass a wider area (described below).

 and using income data from the American Community Survey (ACS), we classified a ZCTA as low-income if the percentage of individuals with income below 200% of the federal poverty level (FPL) fell in the top decile of the national distribution (57.4%-100% population <200% FPL; distributions were computed using all ZCTAs with poverty data and after the population size restriction for metropolitan areas). Analogously, we classified a ZCTA as high-income if the percentage of individuals with income above 500% of the FPL fell in the top decile of the national distribution (43.0%-100% of population >500% FPL). Our results were robust to other thresholds, such as percentage of individuals with income below 100% or below 300% of the FPL.

In a sensitivity analysis, we classified low- and high-income ZCTAs based on their state income distributions and refer to these as relative low- and high-income ZCTAs. Specifically, we classified ZCTAs in each state as being a low- (or high-) income community if the percentage of individuals in the ZCTA with low (high) income fell in the top decile of the state distribution. Income distributions vary from state to state, and relative low-income status in a state may be as important for disparities as national low-income status. Implications for state- and federal-level policy may also differ.

 Our data on practicing physicians and their practice locations were derived from 2015 data from the CMS Provider Enrollment, Chain, and Ownership file; SK&A physician file; and CMS MAX Provider Characteristics (verified in the National Plan and Provider Enumeration System) file. The total number of physicians included in this study (n = 818,951) compared favorably with numbers in other sources. The American Medical Association Masterfile contains 829,962 physicians in 2013 and the 2014 SK&A data file contains 565,000 unique physician National Provider Identifiers. Nurse practitioners and physician assistants also deliver care to patients, in some states do not require physician supervision, and sometimes serve as patients’ primary care clinicians. However, they were excluded from this study because there was no comprehensive data source on these nonphysician clinicians.

Because some patients cross ZCTA lines for care, we assigned to each ZCTA the physician supply of its PCSA. The ZCTA unit of analysis captures low income more accurately than broader geographic areas, whereas PCSA-level physician supply considers border crossing and more accurately represents the service area in which an individual obtains care. Specifically, we computed primary care and specialist physician density (ie, number of physicians per 10,000 population) for each PCSA and assigned these measures to their corresponding ZCTAs. We counted the 236,575 physicians who practiced at more than 1 site by proportions; for example, a physician who practiced at 2 sites was assigned 0.5 effort at each.

We classified physicians as primary care or specialist. Physicians trained in family medicine, pediatrics, and internal medicine (with no subspecialty training) were classified as primary care physicians. We included only non—hospital-based specialists in our tabulations of specialist supply (ie, not emergency medicine, radiology, anesthesiology, or pathology). Of the 818,951 National Provider Identifiers included in this study, 274,449 (33.5%) were classified as primary care physicians and 396,588 (48.4%) as specialists.

 Our hospital data were derived from CMS’ Provider of Services file, the American Hospital Association survey, and the SK&A hospital file from 2015. We classified 6337 short-term acute care hospitals into general acute care (n = 5033), children’s (n = 187), and psychiatric (n = 2402) hospitals. We identified 1302 (22.0%) teaching hospitals based on the hospital’s receipt of graduate medical education payments from CMS in 2015.

 A hospital could be classified as more than 1 type.

We computed 2 measures of the availability of hospital services. First, we computed geodetic distances (shortest curve between 2 points along the surface of the earth) from each ZCTA centroid to the closest general acute care, psychiatric, children’s, and teaching hospital. To reduce the effects of outliers, we winsorized the distances at 1% and 99% for each respective type of hospital and geographic category. Second, analogous to how we assigned PCSA data to ZTCAs for physician supply, we assigned to each ZCTA the number of beds (by hospital type) in the Hospital Service Area (HSA) in which it was located. We converted zip codes to ZCTAs and then assigned each ZCTA to an HSA using a zip code to HSA crosswalk from Dartmouth Atlas. A total of 170 ZCTAs could not be assigned to an HSA.

Building on methods from a previous study and corresponding to geographical distinctions in the Behavioral Risk Factor Surveillance Survey,

 we classified ZCTAs as urban, suburban, or rural (see 

]). The proportion of individuals in urban areas was high across both low- and high-income areas (

). Although it was 10-fold smaller in terms of individuals, rural areas had nearly twice as many ZCTAs as urban areas. The number of high-income individuals in suburban areas was comparable with that in urban areas, but there were twice as many suburban as urban ZCTAs. Using data from the 2010 US Census

 and rolling 5-year ACS estimates for 2011 to 2015,

 we tabulated socioeconomic characteristics for each ZCTA in our sample.

We first compared the distributions of provider supply in low- and high-income urban, suburban, and rural communities. Next, we used weighted linear regression analysis to examine the relationship between the percentage of low-income ZCTA residents and the supply of physicians (in the ZCTA’s PCSA) and hospital services (in the ZCTA’s HSA).

To assess whether the relationship between physician availability and income varied by urban/suburban/rural designation, we included interaction terms between our measure of low-income (ie, percentage of ZCTA residents with incomes <200% FPL) and urban/suburban/rural indicators. Because physicians affiliated with hospitals may have a wider service area than the PCSA, we controlled for the presence of a general acute care hospital in the PCSA. If we drop this control, the disparities that we measure are attenuated because low-income areas with general acute care hospitals have higher physician supply (see eAppendix). Standard errors were clustered at the PCSA level.

In a sensitivity analysis, we estimated regressions with state fixed effects to explore whether income-related differences in provider supply operate on a national and/or state level.

Our final data set included 32,322 ZCTAs with 316,196,416 residents: 27.6 million (8.7%) in low-income ZCTAs and 46.5 million (14.7%) in high-income ZCTAs. ZCTAs identified as low-income for this study represent socioeconomically disadvantaged populations with high concentrations of poverty and racial and ethnic minorities and substantially lower rates of health insurance (

Low-income ZCTAs were concentrated in the southern region of the United States and high-income ZCTAs in the northeast and mid-Atlantic regions (

). Low-income ZCTAs were disproportionately located in rural areas (57% compared with 21% of high-income ZCTAs), whereas the majority of high-income ZCTAs were located in suburban areas (52% compared with 16% of low-income ZCTAs). Seventy-six percent of people from low-income ZCTAs and 44% of people from high-income ZCTAs lived in urban areas (eAppendix Table 1).

Urban low-income ZCTAs were 0.5% more likely than high-income ZCTAs to have very low primary care physician density (we define this to be fewer than 5 physicians per 10,000 population) (

). The disparity was great in rural areas (where the analogous figure was 1.7%) and greatest in suburban areas, where low-income ZCTAs were 7.4% more likely to have very low primary care physician density. The percentage of low-income ZCTAs in urban and suburban areas with very low specialty physician density was 1.3 to 1.4 times larger than that for high-income ZCTAs (63.6% vs 50.1% in urban areas; 92.1% vs 64.4% in suburban areas).

There was considerable variation within states in both physician supply and rates of low-income ZCTAs (

); 82% of total variation in low-income percentage was within states (97% for primary care and 93% for specialty supply). Relative income-related disparities in physician supply might be present if providers chose to locate in higher-income areas within states. To investigate within-state income-related disparities, we added state fixed effects to our regression models. The within-state estimates of disparities were smaller but still significant for primary care physicians (median density was 27% greater for high-income ZCTAs compared with 50% greater) and specialists in suburban areas (70% greater compared with 147% greater) (

Regressions suggested a significant linear relationship between the percentage of a ZCTA’s population with income below 200% of the FPL and the supply of physicians in the ZCTA’s PCSA. We plotted the association between ZCTA percent low-income and physician density separately by urban/suburban/rural category (

 for the full set of regression coefficients). In suburban areas, these regression results implied that moving from the 75th percentile of percent low-income (45.8%) to the 25th percentile (23.4%) was associated with an increase in suburban areas of 0.6 primary care and 1.6 specialist physicians per 10,000 PCSA population. The mean percentage of ZCTA residents living below 200% of the FPL was 35.5% (median, 34.2%; SD, 17.1%; range, 0%-100%).

On average, residents in low-income urban areas lived closer to general acute care, teaching, and psychiatric hospitals than residents in high-income urban areas (

). However, differential median distances were large for teaching, children’s, and psychiatric hospitals in suburban and rural areas (ranging from 15.6 to 30.7 miles).

We modeled the relationship between the percent of ZCTA residents who were low-income and the distance between the ZCTA’s centroid and the closest hospital (

). On average, residents of low-income ZCTAs needed to travel 0.5 and 0.2 miles longer to get to the closest general acute care hospital in suburban and rural areas, respectively. Differential distances were similar for teaching, children’s, and psychiatric hospitals.

In this national study, we documented that physician supply is lower in low- compared with high-income communities and that residents in low-income communities need to travel greater distances to specialty but not necessarily general acute care hospitals. Disparities in physician supply for low- versus high-income communities are greater in suburban areas and for specialty physicians. Disparities were lessened but not eliminated when we examined within-state variation in income and provider supply. Our findings likely underestimate the extent of income disparities in access to healthcare because some physicians do not deliver services to uninsured or Medicaid patients.

Prior work has shown geographic variation in the delivery of healthcare and the importance of the relationship between provider availability and rates of utilization of services such as physician visits.

 Provider availability is also related to quality of care.

 However, much of this work is outdated, and more recent debates on disparities have limited evidence-based discussion of delivery system infrastructure. This study quantifies the current local delivery systems on a national scale by describing the supply of primary care and specialty physicians, as well as access to different types of hospitals. We show that disparities exist between low- and high-income communities, particularly in physician supply in suburban areas and in distance to specialty hospitals for residents in rural areas.

Our study focuses on one piece of the pathway from access to outcomes (ie, delivery system). Although there is support for this relationship in the literature,

 our lack of claims data prevents us from making connections to outcomes in this paper. We cannot be certain that areas with fewer providers are underserved or that places with more abundant providers are overserved. In addition, it is possible that some of the uninsured, particularly those in low-income, urban areas, utilize emergency department or safety net clinics as a default primary care access point. We did not have these data. Nevertheless, given concerns about primary care shortages generally, we think it is unlikely that residents in low-income communities with access to fewer primary care physicians are optimally served and believe that policies to address health disparities would benefit from greater attention to the local delivery system. Similarly, given the increasing interest among policy makers in transportation issues for low-income patients, understanding travel distances is important.

There are also measurement issues worth noting. First, defining markets is challenging because patients can cross boundaries. The measurement of disparities in provider supply related to income may be sensitive to how markets are defined and academic medical centers are treated. Large urban medical centers have many academic physicians who see patients on a part-time basis. These centers are mostly located in urban, often low-income, communities and may artificially inflate provider supply estimates. Second, without recent national Medicaid claims data, we could not examine the extent to which being restricted to providers who accept Medicaid contributed to disparities in access. Third, we did not have cost and quality data for comparing within and across urban, suburban, and rural areas. Fourth, because of data limitations, we could not identify all of the sites in which nonphysician providers such as nurse practitioners delivered care, and therefore did not include them in our analyses. Finally, our results are based on averages across all markets, masking much heterogeneity at the local level. Although the heterogeneity within and across local areas may make it difficult to generalize across the United States, this high-level view is important for informing both federal and local policy.

This study complements other related work. Specifically, recent studies of access focus attention on health insurance expansions (ie, patients’ ability to pay for care).

 The 2016 National Healthcare Quality and Disparities report documented income- and race-related disparities in healthcare utilization and quality.

 A more recent study by Chetty and colleagues went a step further in documenting substantial local area variation in income-related disparities in health outcomes that were uncorrelated with health insurance coverage.

This study on local delivery systems broadens the policy discussion to include nonfinancial barriers to accessing care such as the resources and services available in local communities.

The stratification of our analyses by urban, suburban, and rural areas generated findings consistent with and complementary to those of Schnake-Mahl and Sommers.

 Based on data from a nationally representative survey defining access as having health insurance, a usual source of care, unmet need due to cost, and receipt of a routine checkup, they find that suburban low-income residents face higher barriers to accessing care. Combining their findings with ours suggests that suburban areas may be high-potential places for delivery system interventions to close the divide between the low- and high-income residents. Improving access to care for low-income populations may require greater attention to provider availability and access to specialty hospitals in local delivery systems.

Massachusetts Institute of Technology, Sloan School of Management (CAN), Cambridge, MA; Department of Health Care Policy, Harvard Medical School (CAN, MEC, NDB), Boston, MA; University of California, San Francisco, School of Medicine (IO), San Francisco, CA.

 Research reported in this study was supported by The Commonwealth Fund under award number 20160683. The content is solely the responsibility of the authors and does not necessarily represent the official views of The Commonwealth Fund.

 The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

 Concept and design (CAN, MEC, IO, NDB); acquisition of data (CAN, IO, NDB); analysis and interpretation of data (CAN, MEC, IO, NDB); drafting of the manuscript (CAN, NDB); critical revision of the manuscript for important intellectual content (CAN, MEC, NDB); statistical analysis (CAN, IO, NDB); obtaining funding (MEC); administrative, technical, or logistic support (CAN, IO); and supervision (CAN, NDB).

This large-scale, national study shows geographic variation in provider supply and hospital access for low- and high-income communities following the Affordable Care Act.

Comparison of Healthcare Delivery Systems in Low- and High-Income Communities

Many patients with newly diagnosed colorectal, lung, or breast cancer receive multimodality therapy (ie, surgery, chemotherapy, and/or radiation) within 6 months of diagnosis, and few of those patients received the different types of treatments at a single practice.

For bundled or episode-based payments to work in cancer care, they must be able to hold multiple distinct providers accountable for costs and quality.

Payers are increasingly using episode-based payments to reduce costs and improve quality in several areas, including joint replacements, dialysis, and hospitalizations for certain conditions. In oncology, CMS is implementing the Oncology Care Model, an episode-based alternative payment model for patients receiving chemotherapy.

 Other proposals for oncology episode-based payment models focus on patients with newly diagnosed cancer who may require surgery, radiation, and/or chemotherapy.

In models focused on newly diagnosed cancers, episodes are designed based on guideline-recommended treatments for patients with cancers of a certain type and stage. However, the multidisciplinary nature of cancer care creates a challenge in identifying which provider should receive the bundled payment for such episodes. When a patient with cancer receives multimodality care (ie, surgery, radiation, and/or chemotherapy) from different practices, it may not be readily apparent which practice should be held accountable for the costs and quality of care. We examined the prevalence of multimodality care across different practices to better characterize this challenge in colorectal, lung, and breast cancer, 3 of the most frequently diagnosed cancer types.

Using Surveillance, Epidemiology, and End Results Medicare data, we identified all fee-for-service Medicare beneficiaries with newly diagnosed colorectal, lung, or breast cancer between 2012 and 2015.

For each patient, we used Medicare claims from 2012 to 2016 to identify receipt of surgery, radiation, and chemotherapy within 6 months of diagnosis.

We assigned patients receiving each treatment to the practice (based on Taxpayer Identification Number [TIN]) that submitted the claims for that service in the National Claims History Physician/Supplier file (for surgery, radiation, or chemotherapy) or the Durable Medical Equipment file (for Part B—covered oral chemotherapy).

Among patients who received more than 1 treatment modality, we assessed how often the practice billing for one type of treatment (eg, chemotherapy) was the same as the practice billing for another (eg, surgery). We considered care to be delivered by the same practice if any practice providing one type of treatment also provided another type of treatment.

Among patients newly diagnosed with any stage of colorectal, lung, or breast cancer, the proportions who received multimodality therapy within 6 months of diagnosis were 19.6%, 6.4%, and 16.5%, respectively. Few of these patients received different types of treatment at a single practice. Specifically, of all patients receiving multimodality care for colorectal, lung, and breast cancer, only 5.8%, 17.2%, and 10.7%, respectively, received all of their care from the same practice (

Episode-based payments for patients with newly diagnosed cancer offer a promising opportunity to incentivize cost-efficient and coordinated care. Such models are straightforward for cancers requiring only 1 treatment modality (eg, stage 1 colorectal cancer, for which surgery alone is curative); however, episode-based payment models will be more complicated for patients with cancer who require multimodality care, because such care is infrequently provided by physicians billing together under the same TIN. These more complex patients are likely sicker and have higher costs than patients requiring single-modality care and may represent a real opportunity for alternative payment models to drive savings in cancer care. For episode-based payment reform to be successful in oncology, creative approaches are needed that hold multiple distinct providers accountable for costs and quality.

The authors would like to acknowledge Laurie Meneades for expert programming assistance and Lauren Riedel for research assistance.

 Department of Health Care Policy, Harvard Medical School (SG, MBL, JZ, MEC), Boston, MA; Division of Population Sciences, Dana-Farber Cancer Institute (AAW), Boston, MA; Division of General Internal Medicine, Brigham and Women’s Hospital (NLK), Boston, MA.

 This work was supported by the Laura and John Arnold Foundation. Dr Keating was additionally supported by the National Cancer Institute (K24CA181510).

 Dr Keating is on the evaluation team for the Oncology Care Model, a CMS alternative payment model. The remaining authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

 Concept and design (SG, AAW, MBL, MEC, NLK); acquisition of data (NLK); analysis and interpretation of data (SG, AAW, MBL, JZ, MEC, NLK); drafting of the manuscript (SG, AAW, NLK); critical revision of the manuscript for important intellectual content (SG, AAW, MBL, JZ, MEC, NLK); statistical analysis (MBL, JZ, MEC, NLK); obtaining funding (NLK); and supervision (NLK).

 Nancy L. Keating, MD, MPH, Department of Health Care Policy, Harvard Medical School, 180 Longwood Ave, Boston, MA 02115. Email: keating@hcp.med.harvard.edu.

1. Oncology Care Model. CMS website. innovation.cms.gov/initiatives/oncology-care. Updated August 2, 2019. Accessed September 16, 2019.

2. Hackensack Meridian Health; Cota Inc. Proposal for PTAC: oncology bundled payment program using CNA-guided care. Office of the Assistant Secretary for Planning and Evaluation website. aspe.hhs.gov/system/files/pdf/255906/OncologyBundledPaymentProgramCNACare.pdf. Published March 24, 2017. Accessed September 16, 2019.

3. Large Urology Group Practice Association. LUGPA APM for initial therapy of newly diagnosed patients with organ-confined prostate cancer. Office of the Assistant Secretary for Planning and Evaluation website. aspe.hhs.gov/system/files/pdf/255906/LUGPAAPM.pdf. Published July 5, 2017. Accessed September 16, 2019.

4. Bach PB, Mirkin JN, Luke JJ. Episode-based payment for cancer care: a proposed pilot for Medicare. 

. 2011;30(3):500-509. doi: 10.1377/hlthaff.2010.0752.

Most patients receiving multimodality cancer care receive care from different practices. Therefore, episode-based payments in oncology must hold multiple providers accountable for costs and quality.

Multimodality Cancer Care and Implications for Episode-Based Payments in Cancer

) strives to provide its readers with information that will support the transition to a more value-oriented healthcare system. Doing so involves knowledge from several different types of studies. At the center are studies that provide clinical information on value, such as the paper by Cohen-Mekelburg et al. This paper quantifies the readmission rates (1 in 7) associated with inflammatory bowel disease and suggests that more attention to postdischarge care for these patients may help reduce one source of readmissions. Oftentimes, the clinical evidence is nuanced: Klink et al study treatment for rheumatoid arthritis, noting the distinction in outcomes for different subsets of patients and arguing that managerial structures will be better suited to produce value if they can be more flexible. In this spirit, Sexton Ward et al assess the social value of prophylaxis of febrile neutropenia using a microsimulation model. Although they find that prophylaxis can add considerable value, a large challenge is determining who captures that value. Finally, Charlie Lin et al make a broader point that for some specialties, defining and measuring value is more complex.

A particular interest of the journal is providing evidence regarding managerial tools and methods that may promote value. For example, Liu et al examine ways to improve screening and follow-up for patients with unhealthy alcohol use. Vandergrift and Gray study the relationship between care quality and practice infrastructure. They note the importance of infrastructure in quality and that this effect is strongest for physicians who scored well on their Maintenance of Certification exam. Efforts to promote better infrastructure may be leveraged if we can find ways to support physicians in the lower tail of exam performance. Schwartz et al provide a more direct examination of the way in which infrastructure promotes value, noting that electronic health record systems are not designed to allow for the efficient ordering of high-value clinical services. We thus have a long way to go to adopt information technology systems that actually promote value. Finally, En-Ju Lin et al explore how artificial intelligence tools can help systems deal with the nuances of value by stratifying patients using sophisticated analytics.

A third category of research examines the financial relationships that drive value. Chambers et al report that health plans vary widely in the extent to which they restrict access to orphan drugs. Given the high price of prescription drugs and imperfection in the market, such restrictions are to be expected, but given this variation, more needs to be learned about their clinical and economic impact. 

 has been continually interested in evaluations of broader value-based payment reforms and, in that spirit, Yoon et al evaluate Oregon’s coordinated care organization program, an example of a population-based payment model in Medicaid, finding some positive impact on reducing preventable hospitalizations, consistent with other work.

’s efforts to provide insights useful to the wide range of stakeholders reached by the journal, hopefully supporting their daily efforts to improve healthcare in the United States.

Author | Michael E. Chernew, PhD

Articles

Spending Variation Among ACOs in the Medicare Shared Savings Program

Comparison of Healthcare Delivery Systems in Low- and High-Income Communities

Multimodality Cancer Care and Implications for Episode-Based Payments in Cancer

From the Editorial Board: Michael E. Chernew, PhD