County‐Level Contributors to Geographic Variation in Medicare Fee‐for‐Service Stroke Hospitalization Rates: A Cross‐Sectional Study

Abstract

Background

Stroke hospitalization represents underlying stroke incidence and hospital utilization. We sought to identify factors associated with county‐level hospitalization rates (HRs) and counties with HRs above or below expectation using publicly available data.

Methods

This cross‐sectional study is based on the analysis of county‐level 3‐year average stroke HRs (principal International Classification of Diseases, Tenth Revision [ICD‐10], codes I60.X–I69.X) among Medicare fee‐for‐service beneficiaries from 2018 to 2020. Linear mixed models were fitted to investigate 6 sets of factors associated with HRs in a serial additive stepwise fashion (ie, demographics, overall population vascular risk factors, risk factor treatment, health delivery and access, environmental features, and socioeconomic status). We reported on marginal R ², the most impactful factors, and characterized proportional difference between crude and predicted HRs.

Results

The cohort of 3198 (98.6%) counties and county‐equivalents had a mean stroke HR of 11.2 per 100 000 (SD=2.6). The mean characteristics of the included counties were as follows: 19.4% age ≥65 years, 73% White race, 7.6% coronary heart disease prevalence, 38% hyperlipidemia prevalence, and 5.7 primary care physicians per 10 000. In the fully adjusted model, between‐county unexplained variation remained moderately high (R ²=0.57). The most impactful factors associated with stroke HRs were prevalence of coronary heart disease, hypertension, smoking, nonadherence to antihypertensive medication, and elevation above sea level. Counties in the northwest United States generally had lower‐than‐expected HRs.

Conclusions

Considerable unexplained county‐level variance in stroke HRs exists after accounting for a wide variety of known and potential predictors. Future work to clarify the mechanism of known predictors and explain variance may inform interventions to improve systems of care.

Nonstandard Abbreviations and Acronyms

AHRF

Administration Area Health Resources Files

ATLAS

Interactive Atlas of Heart Disease and Stroke

BRFFS

Behavioral Risk Factor Surveillance System

CDC

Centers for Disease Control and Prevention

CMS

Centers for Medicare and Medicaid Services

DNV‐GL

Det Norske Veritas—Germanischer Lloyd

FFS

fee‐for‐service

HFAP

Healthcare Facilities Accreditation Program

MA

Medicare Advantage

MEDPAR

Centers for Medicare and Medicaid Services Medicare Provider Analysis and Review

NIMHD

National Institute on Minority Health and Health Disparities

Research Perspective.

What Is New?

• It is hard to explain the regional variation in stroke hospitalization rates in the United States even with a comprehensive set of known and potential predictors.

• We found that most social determinants of health examined were associated with stroke hospitalization rates in the expected direction. However, several unexpected and harder‐to‐explain findings related to the ecological and systems‐level patterns observed for rurality, Medicare Advantage penetration, education, race, region of residence, history of coronary heart disease, total Medicare reimbursements for cardiovascular disease encounters, number of hospitals per population, and average altitude above sea level suggest that there are multiple knowledge gaps.

What Questions Should be Addressed Next?

• This research identified several hard‐to‐explain associations with stroke hospitalization and generated several hypotheses that future studies should focus their attention on to provide better understanding toward developing interventions to improve systems of care.

In the United States, stroke incidence is estimated at 800 000 patients yearly. ¹ This type of global estimation is potentially insensitive to important regional differences in incidence, which may be clinically and societally important. ² The lack of regional granularity limits our capacity to explore ecological stroke risk factors and to use epidemiologic data to tailor stroke programs and policies toward the specific needs in a community. ³ , ⁴ , ⁵ , ⁶ One strategy to assess regional variation in stroke incidence is to use readily available stroke hospitalization rates, which reflect 2 underlying factors: stroke incidence and the propensity to seek acute medical care for stroke symptoms. ⁴ , ⁷

Regional variation in stroke hospitalization rates is known to exist, with evidence suggesting that differences in population characteristics, health risk profile, socioeconomic status, and access to care contribute to this variation. ⁴ , ⁶ , ⁸ For example, a recent report by the Centers for Disease Control and Prevention (CDC) demonstrated that counties in the top quartile of stroke hospitalization rates for Black older adults clustered in the Midwest, Northeast, and South. In contrast, the top quartile for stroke hospitalization rates in White older adults were concentrated in the Stroke Belt, specifically the Mississippi Delta region and into Oklahoma and Texas. ⁴ In addition, compared with regions with low hospitalization rates, regions with high stroke hospitalization rates were associated with lower median household income, high school education, and number of hospitals at the county level. ⁸ Thus far, no study has provided a comprehensive assessment of this variation using an expanded set of factors or outside a binomial stroke hospitalization rate clustered framework.

To address this gap, we designed a hypothesis‐generating study using county‐level Medicare fee‐for‐service (FFS) data and publicly available data from the CDC and other sources to characterize regional variation in stroke hospitalization rates, explore predictors of hospitalization rates, and examine the gap between measured and predicted rates. Studying hospitalization rates at the county level rather than less granular levels such as health referral region increases sensitivity to potentially important regional differences in patterns of stroke hospitalization. Given that counties are also sociopolitical entities, they are reasonable targets for identifying and engaging with policymakers who are most relevant to tailoring stroke prevention and treatment programs and policies to community's needs.

METHODS

Data Availability Statement

This study used publicly available county‐level data that can be downloaded from the links provided in Table S1.

Goal and Hypothesis

Our primary hypothesis was that ecological factors (eg, rurality, number of hospitals, air quality, income, education, community levels of hypertension, and diabetes) would explain only a small proportion (<40%) of the heterogeneity of county‐level stroke hospitalization rates. Our goal was to explore a wide variety of potential stroke‐related factors to generate hypotheses on modifiable stroke risk factors and simultaneously explore the influence of ecological and systems factors on hospitalization rates. We sought to examine the potential stroke factors along with known risk factors to try and identify true novel predictors. Studying all known predictors simultaneously enables comparison of the relative magnitude of both known and unknown risk factors.

Study Population and Data

This cross‐sectional study is based on the analysis of county‐level 3‐year average ischemic and hemorrhagic stroke hospitalization rates (principal International Classification of Diseases, Tenth Revision [ICD‐10], codes I60.X–I69.X]) per 100 000 population among Medicare FFS beneficiaries from 2018 to 2020 using publicly available data from the CDC Interactive Atlas of Heart Disease and Stroke (ATLAS). ⁹ The ATLAS is an online data resource that both collates national data on relevant constructs (ie, stroke, demographics, stroke risk factors, social determinants of health, proximity to care) and allows users to easily create and customize local‐level maps. ⁹ The ATLAS combines data from multiple data sources to this end, including: (1) Medicare FFS part A and D administrative files (stroke hospitalizations); (2) the CDC's Behavioral Risk Factor Surveillance System (BRFFS)—a nationally representative telephone survey that collects state data about US residents regarding their health‐related risk behaviors, chronic health conditions, and use of preventive services; (3) Health Resources and Services Administration Area Health Resources Files (AHRF)—a national data source for information on health facilities, health professions, health status, health training programs, geographic measures, economic indicators, and socioeconomic and environmental attributes; (4) US Census Bureau; (5) American Hospital Association; (6) American Medical Association; and (7) stroke certification agencies (ie, Det Norske Veritas—Germanischer Lloyd [DNV‐GL], the Healthcare Facilities Accreditation Program [HFAP], and the Joint Commission). ⁹

Stroke Hospitalization Rates

ATLAS hospitalization rates are age‐standardized (using the 2000 US standard population) and spatially smoothed using a local empirical Bayes algorithm. ¹⁰ Hospitalization rates are calculated using Centers for Medicare and Medicaid Services Medicare Provider Analysis and Review (MEDPAR) file, Part A. ⁹ During the study period, the database provided data on 98.6% (n=3198) of counties and county‐equivalents in the United States. ⁹

County‐Level Stroke Predictors

The county‐level data obtained directly from ATLAS and its data providers included the following 6 sets of predictors that may be associated with stroke hospitalization rates including: (1) demographic composition, (2) stroke risk factors/population health panel, (3) treatment of medical risk factors, (4) health care delivery and access, (5) environmental features, and (6) socioeconomic status. Detailed information on the included predictors is found in Table S1. In addition to the ATLAS‐provided data, we included data from the National Cancer Institute's Surveillance, Epidemiology, and End Results Program (SEER), Centers for Medicare and Medicaid Services (CMS), HDPulse platform—an ecosystem of minority health and health disparities resources from National Institute on Minority Health and Health Disparities (NIMHD), the US Bureau of Economic Analysis, and Area Health Resources Files—Health Resources and Services Administration data warehouse.

This study used publicly available county‐level data and thus did not fall under the guidelines for human subjects research by The Ohio State University's institutional review board. This study followed Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines.

Statistical Analysis

Our statistical analysis aimed to characterize regional variation in county‐level stroke hospitalization rates, identify factors that influence county‐level hospitalization rates, and characterize residual regional variation in county‐level stroke rates after accounting for probable predictive factors. The unit of analysis was at the county level. We sought to examine the potential stroke factors along with known risk factors to try and identify true novel predictors. Studying all known predictors simultaneously enables comparison of the relative magnitude of both known and unknown risk factors.

Due to the low rate of missingness in potential predictors (0.3%–4.5%, see Table S2), multiple imputation of 10 iterations was used to generate missing values and generate pooled analytical estimates. Multiple imputation used fully conditional specification methodology and included all of the predictors used in the analysis. We reported on descriptive statistics (ie, proportions or mean and medians, SD) and bivariate analysis (ie, linear mixed‐model estimates and their CIs and P value) before and after multiple imputation of the included predictors. Preimputation (referred to as complete case analysis) descriptive statistics are reported in Table S2. We did not observe any differences in the associations of predictors with the outcome between the imputed and complete case data.

Linear mixed models were fitted to investigate the 6 sets of factors associated with county‐level hospitalization rates in a serial additive stepwise fashion (Table S3) starting with sets of relatively well‐known factors (eg, demographics and stroke risk factors) to lesser‐known factors (eg, health care delivery and access, environmental features, and socioeconomic status). We opted to use this to explore the additive impact of each set of predictors on the predicted hospitalization rates and the unexplained heterogeneity before we reported on individual predictors. We started by fitting an empty model predicting stroke hospitalization. In every serial additive step, we calculated the marginal R ² of the fixed effects and reported the predicted hospitalization rates using descriptive statistics and box plots. Multivariable model estimates of stroke hospitalizations with sequentially added risk factors are reported (Table S4).

To characterize the quantitative positive or negative potential effect of the most impactful factors on the predicted hospitalization rates, we used the 10th and 90th percentile values of the significantly associated predictors to calculate the average marginal effects range (90th percentile×predictor model estimate – 10th percentile×predictor model estimate) in the fully adjusted model. Finally, we characterized county‐level hospitalization rates proportional variation between the predicted and observed values ([observed – predicted]/observed×100) visually using a map for the fully adjusted model. The generated map characterizes the residual variation that is not accounted for by the model in a heat map design, which designates counties to have above or below expected hospitalization rates. We repeated the fully adjusted model analysis using the complete case data set (n=2847) and we did not find any differences in the predicted hospitalization rates compared with the multiple imputation analysis (correlation=0.988, P<0.001).

We generated variance inflation factors for the 61 included variables in the fully adjusted model and found that 16 of the 61 had a high variance inflation factor score (>10). They were mainly concentrated within predictors of age groups, race groups, stroke risk factors, and number of certified hospitals per population (Table S5). However, this pattern points to structural redundancy, which may cause inflated standard errors rather than biased estimates. To test the effect of multicollinearity of the indicated groups, predictors were summarized into composite indexes (ie, age, race, number of certified hospitals) using principal component analysis. We did not construct a composite index for high variance inflation factor stroke risk factors (coronary heart disease, antihypertensive medication use, and smoking) due to their conceptual importance as confounders. We reevaluated the variance inflation factors using the composite indexes and only coronary heart disease (CHD) was >10. We reran the fully adjusted model using the composite indexes and the model estimates magnitudes, directions, and standard errors were similar to the fully adjusted model that did not use the composite indexes (Table S5). All analyses were conducted in SAS software version 9.4 (SAS Institute Inc). Mapping was performed using ArcGIS Pro version 3.5.1 (Environmental Systems Research Institute, Inc). Two‐sided testing at P<0.05 indicated significance.

RESULTS

Characteristics of Counties and County Equivalents

In a cohort of 3198 counties and county equivalents representing 98.6% of the United States in 2018 to 2020, the population had a mean prevalence of 19.4% who were older than 65 years with those 65 to 69 years old constituting 6.3% of the population (SD 1.4%); 73.1% (SD=22.8%) were of White race, 7.6% (SD=1.5%) had CHD, 38.1% (SD=3.3%) had hyperlipidemia, 8.6% (SD=1.6%) had diabetes, 37.7% (SD=5.5%) reported taking antihypertensive medications, 27.9% (SD=6.6%) were obese, and 18.0% (SD=4.1%) were current tobacco smokers (Table 1). Of the included counties, 62.4% were rural, 87.4% included medically underserved areas, and 39.2% did not have a specific primary economic sector (Table 1).

Table 1.

County‐Level (N=3198) Descriptive Univariate Statistics of the Risk Factors and Their Association With Stroke Hospitalization Rates Using Multiple Imputed Data^*

Variable†	Description	Mean (SD) or no. (%)	Percentile		Estimate	95% CI	P value
			10th	90th
Demographics, %
Male sex		50.0 (2.8)	47.6	52.8	−0.067	−0.100 to −0.034	<0.01
Age 65–69 y		6.3 (1.4)	4.7	8.1	−0.410	−0.473 to −0.347	<0.01
Age 70–74 y		5.1 (1.3)	3.6	6.7	−0.270	−0.339 to −0.201	<0.01
Age 75–79 y		3.5 (1.0)	2.4	4.7	−0.398	−0.490 to −0.307	<0.01
Age 80–84 y		2.3 (0.7)	1.5	3.1	−0.802	−0.933 to −0.672	<0.01
Age 85–89		1.3 (0.5)	0.8	1.9	−1.565	−1.760 to −1.371	<0.01
Age ≥ 89		0.9 (0.4)	0.5	1.31	−2.386	−2.620 to −2.152	<0.01
Race or ethnicity, %
White (non‐Hispanic)		73.1 (22.8)	40.0	93.9	−0.006	−0.010 to −0.002	<0.01
Black (non‐Hispanic)		8.5 (14.1)	0.1	28.7	0.078	0.072 to 0.084	<0.01
Asian (non‐Hispanic)		1.3 (2.7)	0.0	3.0	−0.069	−0.103 to 0.035	<0.01
American Indian or Alaskan Native (non‐Hispanic)		1.5 (7.0)	0.0	2.0	−0.002	−0.016 to 0.011	0.72
Native Hawaiian or Pacific Islander (non‐Hispanic)		0.1 (0.4)	0.0	0.1	−0.402	−0.631 to −0.172	<0.01
More than 1 (non‐Hispanic)		2.9 (1.9)	1.2	4.8	0.096	0.048 to 0.144	<0.01
Other (non‐Hispanic)		0.5 (0.4)	0.2	0.9	−0.228	−0.465 to 0.009	0.06
Hispanic		12.0 (19.4)	1.6	30.0	−0.033	−0.037 to −0.028	<0.01
Total population	No. of inhabitants in 10 000	9.4 (28.8)	0.5	19.7	0.0002	−0.003 to 0.003	0.92
Rurality (nonmetro counties)	Yes	1996 (62.4%)	No	Yes	−0.732	−0.920 to −0.545	<0.01
	No	1202 (37.6%)			Reference
Stroke risk factors/population health panel, %
Coronary heart disease		7.6 (1.5)	5.6	9.5	0.306	0.246 to 0.367	<0.01
High cholesterol		38.1 (3.3)	33.7	42	0.151	0.123 to 0.178	<0.01
Hypertension (antihypertensive medication use)		37.7 (5.5)	31	44.8	0.210	0.195 to 0.225	<0.01
Diabetes		8.6 (1.6)	7	10.7	0.656	0.601 to 0.711	<0.01
Obesity		27.9 (6.6)	18.4	36.4	0.062	0.048 to 0.076	<0.01
Physical inactivity		19.3 (4.0)	14.3	24.8	0.207	0.186 to 0.229	<0.01
Smoking		18.0 (4.1)	13.2	23.2	0.324	0.305 to 0.344	<0.01
Treatment of medical risk factors
Total Medicare reimbursement per enrollee	Mean of reimbursements in thousands among CVD patients	22.3 (4.0)	18.6	26.7	0.129	0.107 to 0.152	<0.01
Self‐reported medication nonadherence to BP medication	Percent of hypertensive population (Medicare part D)	21.7 (4.2)	16.4	27.1	0.207	0.186 to 0.227	<0.01
Self‐reported medication nonadherence to cholesterol‐lowering medication	Percent of hypercholesterolemia population (Medicare part D)	15.3 (3.6)	11.6	19.2	0.150	0.125 to 0.175	<0.01
Health care delivery and access
Hospital	Per 100 000 population	5.5 (9.5)	0.0	13.3	−0.075	−0.085 to −0.066	<0.01
Hospitals with neurological services	Per 100 000 population	0.6 (1.7)	0.0	1.7	−0.039	−0.095 to 0.016	0.17
Hospitals with CT	Per 100 000 population	3.6 (7.7)	0.0	9.4	−0.072	−0.083 to −0.060	<0.01
Hospitals with medical school affiliation	Per 100 000 population	0.5 (2.0)	0.0	1.4	−0.084	−0.129 to −0.039	<0.01
Hospitals with tele‐stroke care	Per 100 000 population	0.9 (3.1)	0.0	2.5	−0.037	−0.066 to −0.007	0.01
Certified stroke centers	Per 100 000 population	0.3 (0.7)	0.0	1.0	0.351	0.228 to 0.473	<0.01
Acute stroke–ready hospitals	Per 100 000 population	0.04 (0.37)	0.0	0.0	0.207	−0.042 to 0.455	0.10
Primary stroke centers	Per 100 000 population	0.2 (0.6)	0.0	0.8	0.402	0.260 to 0.545	<0.01
Thrombectomy capable stroke centers	Per 100 000 population	0.001 (0.03)	0.0	0.0	2.096	−1.083 to 5.275	0.20
Comprehensive stroke centers	Per 100 000 population	0.02 (0.13)	0.0	0.0	0.017	−0.668 to 0.703	0.96
Community health centers‡	Per 100 000 population	9.4 (22.2)	0.0	23.9	0.003	−0.001 to 0.007	0.17
Pharmacies	Per 100 000 population	10.6 (11.3)	0.0	24.2	0.041	0.033 to 0.49	<0.01
Medically underserved area (all or part of county)	Yes	2796 (87.4%)	No	Yes	0.635	0.360 to 0.911	<0.01
	No	402 (12.6%)			Reference
Primary care physician per population	Per 10 000 population	5.7 (4.3)	1.4	10.7	−0.139	−0.160 to −0.119	<0.01
Neurologist	Per 10 000 population	0.2 (0.6)	0	0.6	−0.128	−0.290 to 0.034	0.12
Radiologist	Per 10 000 population	0.1 (0.4)	0	0.4	−0.187	−0.436 to 0.062	0.14
Internists	Per 10 000 population	1.5 (2.2)	0	3.8	−0.054	−0.097 to −0.012	0.01
Environmental features
Air quality	Annual PM2.5 level (μg/m3)	7.5 (1.7)	5.0	9.4	0.763	0.717 to 0.810	<0.01
Elevation above sea level	Mean elevation in 100 ft	13.1 (14.9)	1.0	33.9	−0.088	−0.093 to −0.083	<0.01
Access to parks§	Percent of population	32.1 (27.2)	2.1	74.1	−0.020	0.023 to −0.016	<0.01
Mobility	Percent population that moved in the past year	11.3 (3.9)	7.0	16.0	−0.002	−0.021 to −0.025	0.87
Drinking water violations	Yes	1204 (37.7%)	No	Yes	0.027	−0.216 to 0.162	0.78
	No	1994 (62.3%)			Reference
Crowding	Percent households with >1 person per room	2.5 (2.2)	0.7	4.7	−0.009	−0.049 to 0.032	0.684
Food environmental index||	County‐wide index	7.6 (1.3)	6.0	9.0	−0.389	−0.460 to −0.318	<0.01
Socioeconomic status
Median household income	In the year 2019 in 1000 US dollars	62.9 (16.3)	45	82	−0.041	−0.046 to −0.035	<0.01
Unemployment rate	Percent of population	3.7 (1.4)	2.3	5.2	0.356	0.290 to 0.422	<0.01
Medicare Advantage penetration	Percent of >65 y population in 2019	30.7 (16.8)	8	50.13	0.025	0.019 to 0.030	<0.01
Uninsured rate	Percent of <65 y population	11.7 (5.5)	6	19.5	0.022	0.005 to 0.039	0.01
Living in poverty	Percent of population	14.8 (5.8)	8.6	22.4	0.149	0.134 to 0.164	<0.01
On supplemental nutritional assistance	Percent of population	13.4 (7.4)	5.2	23.8	0.121	0.109 to 0.132	<0.01
Broadband internet coverage	Percent of population	82.1 (7.6)	72.6	90.5	−0.060	−0.072 to −0.048	<0.01
Less than high school education	Percent of population	11.9 (5.9)	5.7	19.8	0.121	0.106 to 0.136	<0.01
Less than college education	Percent of population	76.5 (10.0)	63.1	86.5	0.072	0.063 to 0.081	<0.01
Primary economic sector type¶	Nonspecific	1255 (39.2%)	Nonspecific	Recreation	Reference
	Farming	444 (13.9%)			−2.026	−2.398 to −1.753	<0.01
	Mining	224 (7.0%)			−0.645	−1.002 to −0.287	<0.01
	Manufacturing	505 (15.8%)			0.091	−0.169 to 0.351	0.49
	Government	409 (12.8%)			−0.404	−0.684 to −0.123	<0.01
	Recreation	361 (11.3%)			−1.838	−2.132 to −1.543	<0.01
GDP/capita	In thousands of US dollar	61.8 (104.1)	24.7	89.1	−0.004	−0.005 to −0.003	<0.01

^*The unadjusted estimates were generated using a generalized linear mixed model.

^† All variable values are derived for the years of 2019 or 2020.

^‡ Community‐based clinics that primarily focus on preventative and primary care services accessible to people of all ages regardless of insurance status.

^§ Percentage of population living within half a mile of a park.

^|| A healthy food environment measure from 0 (worst) to 10 (best) that accounts for both proximity to healthy foods and income.

^¶ Economic sector that employs the highest percentage of county residents.

BP indicates blood pressure; CT, computed tomography; CVD, cardiovascular disease; and GDP, gross domestic product.

Of the 61 included predictors, 51 (83.6%) were found to be significantly associated with stroke hospitalization rates (Table 1). The complete case analysis descriptive statistics were not different from the multiple imputed data (Table S2).

Variation in Hospitalization Rates

The cohort of 3198 counties and county‐equivalents had a mean observed stroke hospitalization rate of 11.2 per 100 000 (SD=2.6; interquartile range, 3.7 [10th percentile=7.7, 90th percentile=14.5]) (Figure 1). The sequentially adjusted models yielded mean predicted hospitalization rates that were similar to the observed rate, with a slightly reduced variance. The between‐county explained variation increased from an R ² of 0.35 in model 1, which adjusted only for demographics, to an R ² of 0.57 in the fully adjusted model (model 6).

Figure 1. Box plot of sequentially adjusted models predicted stroke hospitalization rates per 100 000 population and their corresponding R ² values.

Model 1 = demographics (ie, % sex+ % age groups + % race groups + total population + rurality). Model 2 = model 1 covariates + population health panel/ stroke risk factors (ie, % history of CVD + % history of high cholesterol + % history of high BP + % history of diabetes + % history of obesity + % physical inactivity + % smoking). Model 3 = model 2 covariates + treatment of risk factors (ie, total Medicare reimbursement per enrollee + % nonadherence to BP medication + % nonadherence to cholesterol lowering medication). Model 4 = model 3 covariates + health care delivery and access per population (ie, number of hospitals + number of hospitals with neurological services + number of hospitals with CT + number of hospitals with tele‐stroke + number of certified stroke centers + number of hospitals + number of pharmacies + number of primary care physicians + number of radiologists + number of neurologists + number of internists + county includes medically underserved areas [yes/no]). Model 5 = model 4 covariates + environmental features (ie, air quality + elevation above sea + access to park + mobility + drinking water violations + crowding + food index). Model 6 = model 5 covariates + socioeconomic status (ie, household income + unemployment rate + Medicare Advantage penetration + uninsured rate + % poverty + % internet coverage + % less than high school and college education + primary economic sector + GDP per population). BP indicates blood pressure; CT, computed tomography; CVD, cardiovascular disease; and GDP, gross domestic product.

Predictors of Stroke Hospitalization Rates

Our fully adjusted model had 17 of 61 (27.9%) significantly associated predictors (Table 2 and Table S4). We found that the most impactful factors in predicting higher hospitalization rates in descending order included having a history of taking antihypertensive medication (a county at the 10th percentile of antihypertensive medication prevalence had a predicted stroke hospitalization rate of 5.47 versus 7.90 at the 90th percentile [difference of 2.43 higher predicted hospitalizations per 100 000 population]), smoking (difference of 1.89), nonadherence to antihypertensive medication (difference of 1.11), unemployment rate (difference of 0.65), total Medicare reimbursement per enrollee (difference of 0.25), air quality (difference of 0.41), number of internists per 10 000 population (difference of 0.37), percentage on nutritional supplement assistance (difference of 0.43), and Medicare Advantage (MA) penetration rates (difference of 0.25) (Table 2).

Table 2.

Range of Predicted Hospitalization Rates in the Fully Adjusted Model Using the 10th and 90th Percentile Values of the Included Predictors^*

Category	Factor	Predicted change in expected hospitalization rate with covariate at the 10th percentile of the covariate level†	Predicted change in expected hospitalization rate with covariate at the 90th percentile of the covariate level†	Difference (90th–10th)	P value
Demographics	Population aged 65–69 y, %	−0.63	−1.09	−0.46	0.04
	Rurality	0.00	−0.44	−0.44	<0.01
Stroke risk factors/population health panel	CHD, %	−3.27	−5.54	−2.28	<0.01
	Hypertension, %	5.47	7.90	2.43	<0.01
	Obesity, %	−0.33	−0.66	−0.32	0.04
	Smoking, %	2.50	4.39	1.89	<0.01
Treatment of medical risk factors	Total Medicare reimbursement per enrollee	0.99	1.42	0.43	<0.01
	Medication nonadherence to antihypertensive medication, %	1.71	2.82	1.11	<0.01
Health care delivery and access	No. of hospitals per 100 000 population	0.00	−0.40	−0.40	<0.01
	Primary care physicians per 10,000 population	−0.08	−0.60	−0.53	<0.01
	Internist per 10,000 population	0.00	0.37	0.37	<0.01
Environmental features	Air quality (annual PM2.5 level μg/m3)	0.46	0.87	0.41	<0.01
	Elevation above sea level (in 100 ft)	−0.03	−1.12	−1.09	<0.01
Socioeconomic status	Unemployment rate, %	0.51	1.16	0.65	<0.01
	Medicare Advantage penetration, %	0.05	0.29	0.25	0.04
	On nutritional supplemental assistance, %	0.12	0.55	0.43	0.02
	Less than college education, %	−1.41	−1.93	−0.52	<0.01

^*The most impactful variables were chosen from the pool of the significantly adjusted factors (n=17) in the fully adjusted model.

^† The 10th or 90th percentile predicted change in expected hospitalization rate = covariate 10th or 90th percentile×model estimate of the covariate (eg, 10th percentile prevalence CHD prevalence = 5.6, fully adjusted model covariate estimate = −0.583, predicted change = 5.6×−0.584 = −3.27).

CHD indicates coronary heart disease.

The most impactful factors in predicting lower stroke hospitalization rates in descending order included CHD (a county at the 10th percentile of CHD prevalence had a predicted stroke hospitalization rate of −3.27 versus −5.54 at the 90th percentile [difference of −2.28 lower predicted hospitalizations per 100 000 population]), the elevation of the county above sea level [difference of −1.09], number of primary care physicians per 10 000 population [difference of −0.53], proportion with less than college education [difference of −0.52], proportion ages 65 to 69 years old [difference of −0.46], rurality [difference of −0.44], number of hospitals per 100,00 population [difference of −0.40], and proportion with obesity [difference of −0.32] (Table 2). Using the fully adjusted model, ≈49.1% and 50.9% of the examined counties had higher‐than‐expected and lower‐than‐expected hospitalization rates, respectively. Counties in the northwest and New England (far northeast) region generally had lower‐than‐expected stroke hospitalization rates (Figure 2).

Figure 2. Proportional differences between crude and predicted hospitalization rates at the county level.

Proportional difference = ((observed hospitalization rate – predicted hospitalization rate)/observed hospitalization rate)×100%. N/A indicates not available.

DISCUSSION

In this hypothesis‐generating national cross‐sectional study of county‐level stroke hospitalization rates among Medicare FFS beneficiaries, we found that county‐level demographics, population stroke risk factor health panel, treatment, health care delivery and access, environment, and socioeconomic characteristics predicted a substantial proportion (R ²=0.57) of variation in county‐level hospitalization rates. Our findings support the consensus that place of residence is strongly associated with stroke hospitalization and identifies some potential stroke risk factors of interest as well as suggests some possible opportunities to improve stroke systems. ⁸ These findings at the county level may be able to inform public health programs and policies, as such policies are commonly designed and implemented at the local level.

We found that most social determinants of health examined were associated with stroke hospitalization rates in the expected direction. However, the patterns observed for rurality, MA penetration, education, race, and region of residence warrants further investigation. First, the inverse association between rurality and hospitalization may reflect reduced access to care, higher prehospital mortality (selective survival bias), or potentially protective environmental exposures such as lower air pollution or active lifestyle. ¹¹ Second, the positive association between MA penetration and hospitalization may reflect underlying selection effects as MA plans disproportionately enroll lower‐income, Black or Hispanic individuals, those with poorer health status, and those with fewer years of education driven by lower or zero premiums and supplemental benefits. ¹² In addition, MA plans tend to code diagnosis more aggressively compared with FFS, reflecting coding intensity rather than true differences in morbidity. ¹³ To our knowledge, no prior study has directly compared stroke hospitalization rates between MA and FFS beneficiaries. We hypothesize that the combined demographic, health, and socioeconomic profile of MA enrollees may contribute to increased hospitalization. Third, although higher educational attainment is generally associated with reduced stroke incidence (stroke hospitalization is representative of incidence), ¹⁴ our analysis found the opposite direction of effect when evaluating stroke hospitalization rates. We hypothesize that this unexpected inverse association may be attributable to lower stroke literacy, leading to delayed or missed diagnoses ¹⁵ and resulting in differential access to care with the possibility of underdiagnosis among individuals with less education or overdiagnosis among those with more education. ¹⁶ Fourth, despite well‐documented geographic and racial disparities in stroke incidence, including the “stroke belt” and higher rates among Black compared with White individuals, ¹⁷ , ¹⁸ we did not find patterns that would be expected in stroke hospitalization rates. None of the county‐level demographic variables were associated with hospitalization after multivariable adjustment, suggesting that to the extent that regional demographics influence hospitalization, they are likely mediated through other factors captured in the model. This is evident in the findings of Polineni et al where increasing levels of structural racism that includes integration and segregation, residential housing, education, employment, and income domains—which our model represents—were associated with increased county‐level incidence of acute ischemic stroke, regardless of race. ¹⁹ Finally, our observation that the Northwest and New England regions exhibit lower‐than‐expected hospitalization rates is consistent with the findings by Schieb et al, who reported persistently lower hospitalization clusters in these areas based on FFS data from 1995 to 2006. ⁸ Historically, these regions have the most favorable socioeconomic and health care profiles, yet no comprehensive analyses have identified the underlying protective factors. ⁸ , ²⁰ , ²¹

We also found, aligned with prior expectations, that conventional risk factors and risk factor treatment including hypertension, smoking, and nonadherence to antihypertensive therapy were independently associated with increased hospitalization rates. Both hypertension and smoking are established modifiable risk factors for stroke. ²² Poor adherence to antihypertensive medication is also linked to a markedly higher short‐ and long‐term stroke incidence, demonstrating a dose–response effect whereby lower adherence confers greater hospitalization risk. ²³ The negative association between obesity and hospitalization was not as clearly aligned with our expectations. While obesity is not considered an independent stroke risk factor, as its effects are largely mediated by metabolic sequelae such as hypertension, diabetes, and hypercholesterolemia, ²⁴ it was not expected that regions with higher obesity rates would have lower stroke hospitalization rates. Our findings are novel and raise potential hypotheses: (1) receiving an obesity diagnosis may be a marker of health care access and (2) adherence to management of obesity‐related metabolic disorders may mitigate stroke‐related hospitalizations.

We also identified several unexpected and harder‐to‐explain findings related to history of CHD, total Medicare reimbursements for cardiovascular disease encounters, number of hospitals per population, and average altitude above sea level. Notably, we observed an apparently contradictory association between history of CHD that was associated with lower hospitalization rates (inverse to our biological expectation) and total Medicare reimbursement for cardiovascular diseases that was associated with higher hospitalization rates. Several explanations are possible. In spite of the fact that CHD confers a 2‐fold higher risk of stroke, ²⁵ because individuals with CHD tend to experience stroke at an older age than those without CHD, ²² , ²⁶ CHD diagnosis and management before stroke may result in greater health care intensity, attenuating the observed hospitalization risk in this group. ²⁷ However, the possibility that more intense health care is associated with lower stroke hospitalization rates is contradicted by the positive association between total Medicare reimbursements and stroke hospitalization. The tension between those opposing explanations suggests a knowledge gap that warrants further study. In terms of the positive association with number of hospitals, limited evidence suggests that an increase in the number of hospitals and providers may increase stroke hospitalizations by improving stroke diagnosis, reducing missed stroke cases, or potentially increasing false‐positive stroke diagnoses. ⁸ , ²⁸ Finally, the protective association with altitude conformed with findings from 2 reviews and meta‐analyses encompassing studies from the United States, Switzerland, Greece, China, and India. ²⁹ , ³⁰ In addition, being born at higher versus lower altitude appears to exert an independent, additive protective effect. ³¹ However, there is no clear understanding of the mechanisms leading to the altitude protective effect.

Limitations

Studying hospitalization rates at the county level may be insensitive to important regional factors that exist at more granular geographic levels. ⁵ This potential limitation may be partially offset by the idea that local policy levers of action (ie, governmental or administrative regions) often exist at the county level. Further, our approach does not account for patients with stroke who are not hospitalized and will underestimate trends in regions with more outpatient stroke diagnoses and overestimate in regions with a higher proportion of patient diagnoses. Our data relied on ecological measures and do not reflect individual‐level effects. The CDC stroke definition included ICD codes I65.X and I66.X (occlusion or stenosis of precerebral and cerebral arteries without causing an infarct), which might bias the geographic variation, especially in counties with high intensity of testing and procedures. Although 75% of all stroke hospitalizations occur among people aged ≥65 years, ³² by relying on Medicare data, our results are insensitive to regional differences in stroke in the young in general, and given the markedly elevated stroke incidence in younger Black people compared with younger White people, may be insensitive to race differences in hospitalization rates that might be detected with a less age‐constrained sample. Further, our hospitalization rates are underestimated because we included only FFS without MA data; this is likely partially offset by including MA penetration rates in the model. Finally, given the large number of examined predictors, we cannot rule out the possibility that some of the association's magnitude may be affected by collinearity.

Considerable but lower‐than‐expected unexplained county‐level variance in stroke hospitalization rates exists after accounting for a wide variety of known and potential predictors. The findings of this study may inform future work to understand regional variation in stroke hospitalization rates and assist in the development of targeted nuanced interventions to regions that have higher‐than‐expected stroke hospitalization rates and identify regions with lower‐than‐expected rates as a source of assessment to look for protective factors.

Sources of Funding

None.

Disclosures

None.

Supporting information

Tables S1–S5

STROBE Checklist