The geography of Medicare's hospital value‐based purchasing in relation to market demographics

Abstract Objective To illustrate the association between the sociodemographic characteristics of hospital markets and the geographic patterns of Medicare hospital value‐based purchasing (HVBP) scores. Data Sources and Study Setting This is a secondary analysis of United States hospitals with a HVBP Total Performance Score (TPS) for 2019 in the Centers for Medicare and Medicaid Services (CMS) Hospital Compare database (4/2021 release) and American Community Survey (ACS) data for 2015–2019. Study Design This is a cross‐sectional study using spatial multivariable autoregressive models with HVBP TPS and component domain scores as dependent variables and hospital market demographics as the independent variables. Data Collection/Extraction Methods We calculated hospital market demographics using ZIP code level data from the ACS, weighted the 2019 CMS inpatient Hospital Service Area file. Principal Findings Spatial autoregressive models using eight nearest neighbors with diversity index, race and ethnicity distribution, families in poverty, unemployment, and lack of health insurance among residents ages 19–64 years provided the best model fit. Diversity index had the highest statistically significant contribution to lower TPS (ß = −12.79, p < 0.0001), followed by the percent of the population coded to “non‐Hispanic, some other race” (ß = −2.59, p < 0.0023), and the percent of families in poverty (ß = −0.26, p < 0.0001). Percent of the population was non‐Hispanic American Indian/Alaskan Native (ß = 0.35, p < 0.0001) and percent non‐Hispanic Asian (ß = 0.12, p < 0.02071) were associated with higher TPS. Lower predicted TPS was observed in large urban cities throughout the US as well as in states throughout the Southeastern US. Similar geographic patterns were observed for the predicted Patient Safety, Person and Community Engagement, and Efficiency and Cost Reduction domain scores but are not for predicted Clinical Outcomes scores. Conclusions The lower predicted scores seen in cities and in the Southeastern region potentially reflect an inherent—that is, structural—association between market sociodemographics and HVBP scores.

• The local and regional geographic impact of these policies across the United States has not been explored.

What this study adds
• We created maps showing how the hospital scores were potentially affected by the sociodemographic characteristics of their market areas.
• Hospitals in the Southeastern states and larger cities with racial and ethnic diversity had lower predicted scores, potentially representing an increased risk of lower hospital valuebased purchasing program scores.
• Patient experience, patient safety, and hospitalization related costs contributed to the lower predicted scores. There was less evidence of regional geographic patterns in clinical quality. In 2019, $1.9B was redistributed among almost 2800 hospitals, with 44% of hospitals having a net loss in Medicare income and 5% losing more than $350K. 2,3 Numerous commentators have noted that as the US payer system moves toward pay-for-performance/value-based purchasing, there is a risk of increased penalization of safety-net hospitals. [4][5][6][7][8][9][10][11][12][13] The "consistent, negative, and significant effect" 13 of safety-net status on HVBP scores implies an association with factors that are outside hospitals' ability to modify, such as market characteristics, government payer distribution, and ownership structure. 8 Lower access to highquality health care is a subset of the inequalities faced by communities of concentrated poverty, and by extension, communities where racial or ethnic minority groups are overrepresented. 12,[14][15][16] At the patient and community level, therefore, there is the potential for HVBP incentive payments to reinforce structural racism.
While the association between safety-net hospital status and lower overall performance on HVBP has been explored in a growing body of literature, the extent and patterns of the impact of the racial and socioeconomic make-up of patient populations on HVBP scores across the US have not been explored from a geospatial perspective.
Our objective is to apply spatial data analysis to map the association between hospital markets and HVBP scores across the US, illustrating both local detail and broader regional trends. Our work fits into the public health tradition of disease mapping wherein the local nature of spatial patterning of data is explicitly used to improve estimation. 17 Our use of spatial models, increased specificity in defining hospital market areas, and national scope represent enhancements to what is known about structural racism and HVBP.

| Study population and outcome measures
This is a cross-sectional observational study of all non-federal general medical/surgical hospitals in the United States (US) receiving scores for 2019 discharges in the CMS HVBP program, as of the April 1, 2021 release (n = 2673). 3

| Independent variables
We defined flexible hospital markets based on ZIP code level data using CMS's Market Service Area file for hospitals derived from Medicare fee-for-service recipients. 18 Hospital market was estimated from the 2019 CMS inpatient Hospital Service Area File. 18 This file provides a count of services rendered to fee-for-service Medicare beneficiaries by facility and by the beneficiary's mailing ZIP code, excluding hospital/ZIP codes pairs with 10 or fewer discharges. Population demographics for 5-digit Zip Code Tabulation Areas (ZCTA) were obtained from the American Community Survey (ACS) 5-year estimates for 2015-2019. 19 We assumed a one-to-one correlation between ZCTA and ZIP code. ZIP codes on the market file but not the ACS file (9%) were excluded. These most likely represent PO Box only ZIP codes. Hospital market values were estimated by weighting the ZCTA estimates from the ACS by the percent of the total hospital market in each ZIP code/ZCTA as follows:

| Statistical analysis
We applied spatial autoregressive (SAR) models to each dependent variable to estimate the predicted scores based on individual hospital's independent variables and the observed scores of neighboring hospitals. These spatial models account for the spatial dependence between observations (that is, observations in one location tend to be similar to observations in other locations). These models are more conservative and less biased than ordinary least squares (OLSs) regression models, particularly given the spatial dependency (i.e., overlap) in hospital markets.
A formal model selection process was used to determine the appropriate form of the multivariable spatial regression. A nearestneighbor (k) matrix for each hospital was created for each hospital for k = 6-10 neighbors per hospital. Imputation was used to assign missing values for the HVBP domain scores, which were missing for a small number of hospitals, which allowed the same neighbor matrix to be used for all models.
We used a backward elimination process to select the independent variables used in the analysis. We split the data into training We created dot maps of the observed and predicted HVBP scores to allow comparison of regional patterns. Maps of the predicted values from SAR models show the influence of smoothing the dependent variables based on the local (k-neighbor) patterns of the independent variables. 23 In other words, spatial autoregression reduces the effect of factors not in the spatial model.

| Model selection
The selected model included 11 variables (Table 1). In the SAR models, the number of nearest neighbors per hospital that produced the best model fit was 8 although there was little difference in model fit between 6 and 10 neighbors.
The autoregressive regression coefficient, rho, was statistically significant in all models, indicating all the HVBP variables examined displayed spatial dependence. All models also had statistically significant residual variance (measured by sigma-squared), which can be due to outliers in the predicted HVBP scores or to the presence of regional patterns in addition to the modeled local patterns.

| Model results
There is clear spatial patterning to the predicted TPS that is not as evident in the observed scores ( Figure 1). may be conservative relative to non-spatial models, which run the risk of type I error when spatial dependency is present. 25 Of all the sociodemographic variables examined, the US Census As a cross-sectional ecological study, we need to be careful not to over-interpret these findings. Translating these findings into a causal relationship between HVBP incentive payments and poorer quality care received by individuals, particularly in specific demographic groups or regions, is beyond the scope of this study. 58 We chose to highlight the spatial patterning of the association with hospital market to add to the understanding of potential structural racism in CMS's HVBP program. There are multiple other hospital characteristics, such as size, ownership structure, and teaching status, that also influence hospital quality and have spatial dependence.

| LIMITATIONS
Wider understanding of potential structural racism in access to higher quality care is hindered by the lack of nationwide data on

| CONCLUSION
Our findings highlight the potential for structural racism in the HVBP program, with lower predicted scores for hospitals in the Southeastern states and larger cities in the US that have higher levels of racial and ethnic diversity. In particular, we show a consistent finding for a spatial relationship between market sociodemographic composition and the HVBP Person and Community Engagement and Efficiency and Cost Reduction domains, both of which have been shown in previous studies to be associated with hospital patient mix. We also show this spatial patterning for the safety domain, which has not been as well documented. This exploratory study complements existing understanding of place-based structural racism on value-based payment metrics and further highlights the need for re-evaluation of policy regarding adjustment of incentive payments to reduce disparities among hospitals that serve minority populations.

ACKNOWLEDGMENTS
No funding to report.