Geographic Variation in Anticoagulant Use and Resident, Nursing Home, and County Characteristics Associated with Treatment Among US Nursing Home Residents

Abstract

Objectives

To quantify geographic variation in anticoagulant use and explore what resident, nursing home, and county characteristics were associated with anticoagulant use in a clinically complex population.

Design

A repeated cross-sectional design was used to estimate current oral anticoagulant use on December 31st, 2014, 2015, and 2016.

Setting and Participants

Secondary data for United States nursing home residents during the period 2014–2016 were drawn from the Minimum Data Set 3.0 and Medicare Parts A and D. Nursing home residents (≥65 years) with a diagnosis of atrial fibrillation and ≥6 months of Medicare fee-for-service enrollment were eligible for inclusion. Residents in a coma or on hospice were excluded.

Methods

Multilevel logistic models evaluated the extent to which variation in anticoagulant use between counties could be explained by resident, nursing home, and county characteristics, and state of residence. Proportional changes in cluster variation (PCV), intraclass correlation coefficients (ICC), and adjusted odds ratios (aOR) were estimated.

Results

Among 86,736 nursing home residents from 11,860 nursing homes and 1,694 counties, 45% used oral anticoagulants. The odds of oral anticoagulant use were 18% higher in 2016 than 2014 (aOR: 1.18; 95% confidence interval: 1.14–1.22). Most states had counties in the highest (51.3–58.9%) and lowest (31.1%-41.4%) deciles of anticoagulant use. Compared with the null model, adjustment for resident characteristics explained one-third of the variation between counties (PCV: 34.8%). The full model explained 65.5% of between-county variation. Within-county correlation was a small proportion (ICC<2.2%) of total variation.

Conclusions and Implications

In this older adult population at high-risk for ischemic stroke, less than half of residents received treatment with anticoagulants. Variation in treatment across counties was partially attributable to the characteristics of residents, nursing homes, and counties. Comparative evidence and refinement of predictive algorithms specific to the nursing home setting may be warranted.

Brief Summary:

Less than half of older nursing home residents with atrial fibrillation receive treatment. Large variability exists in anticoagulant use for individuals with similar characteristics and across geographic areas.

Introduction

Anticoagulation is highly effective for ischemic stroke prevention for individuals with atrial fibrillation.¹ The real-world safety and effectiveness of anticoagulation with warfarin has been demonstrated in older adults.² Despite clinical trial evidence and additional therapeutic options, less than one-half of nursing home residents receive anticoagulants,³ lower than in community-dwelling older adults.⁴ Anticoagulation decisions for older adults with atrial fibrillation residing in nursing homes are informed by the limited life expectancy of residents⁵ and are complicated by the presence of both vascular and bleeding risk factors.³ Shared-decision making, as is recommended by current practice guidelines,⁶ is challenging for residents with cognitive impairment^3,7 and for whom there is a dearth of evidence on the absolute risk of ischemic stroke and bleeding (and the consequences of each) under alternative treatment scenarios.

Recognizing the limited availability of evidence to guide anticoagulation decisions, our objective was to explore what resident, nursing home, and county characteristics were associated with anticoagulant use for this clinically complex population. Fundamental to this objective was the goal of identifying sociodemographic, clinical, and health system factors that may be amenable to clinical and health policy interventions. Considering that absolute differences on clinical risk scores^8–10 predicting ischemic stroke and bleeding risk are small between treated and untreated residents,³ we hypothesized that in addition to those risk factors, other sociodemographic (e.g., age, Medicaid enrollment), clinical (e.g., medication use), and functional characteristics of residents would be associated with anticoagulant use. Furthermore, because of the large role for patient preference and clinician judgement in current clinical practice, we expected concordance in treatment patterns within local areas (i.e., counties) with shared personal values and healthcare providers.

Methods

Data

Medicare beneficiary enrollment and vital status (Master Beneficiary Summary File), hospital and skilled nursing facility (SNF) utilization (Medicare Part A), pharmacy claims (Medicare Part D), and nursing home assessments (Minimum Data Set (MDS) 3.0) were accessed through a data use agreement with the Centers for Medicare and Medicaid Services. The MDS 3.0 is a mandatory assessment performed at regular intervals in Medicare/Medicaid certified nursing homes that produces valid data.¹¹ Nursing home characteristics were obtained from the Nursing Home Compare and the Provider of Services files. County characteristics were linked from the Area Health Resources File. The Institutional Review Board approved this study.

Study Design

We used a repeated cross-sectional design whereby we estimated oral anticoagulant use on December 31^st 2014, 2015, and 2016.

Study Population

Nursing home residents (≥65 years of age) with a diagnosis of atrial fibrillation and ≥6 months of Medicare fee-for-service enrollment preceding December 31 of each year were eligible. Residents with an admission, annual, quarterly, or change-in-status MDS assessment that was not during a SNF stay were included. At least one diagnosis of atrial fibrillation or atrial flutter¹² on a Medicare Part A claim and one diagnosis of atrial fibrillation, atrial flutter, or dysrhythmia on an MDS 3.0 assessment were required. For residents eligible in multiple years, a single cross-section was randomly selected. Comatose residents and those without a Part D claim in the 12-month lookback period were excluded. Residents on hospice or in a hospital, SNF, or hospital-based nursing home on December 31 were excluded because medications are not reimbursed by Part D in these settings. Residents in counties with <11 residents were excluded (5,762 residents from 1,046 counties) (Table A1). The final sample included 86,736 residents.

Anticoagulant Use

Current use of an oral anticoagulant (apixaban, dabigatran, edoxaban, rivaroxaban, warfarin) was determined on December 31 of each year using Part D claims after nursing home admission. We used date of the Part D claim and the number of days supply and adjusted for early medication refills, hospitalizations, and SNF stays.

Resident Characteristics

Characteristics of the resident population were derived from the most recent admission, quarterly, annual, or change-in-status MDS 3.0 assessment and from Medicare claims in the 12 months before December 31 in each year. These included sociodemographic characteristics (age, sex, race/ethnicity, marital status, dual Medicare-Medicaid enrollment), number of hospital admissions, recent (≤6 months) inpatient surgery, hospitalizations for certain conditions identified using diagnoses on Part A claims (ischemic stroke,¹³ extracranial bleeding,¹⁴ intracranial hemorrhage,¹⁴ myocardial infarction,¹⁵ venous thromboembolism,¹⁶ or transient ischemic attack) (Table A2), CHA₂DS₂-Vasc ischemic stroke risk score and its components,⁸ ATRIA bleeding risk score and its components,⁹ other conditions associated with risk or perceived risk of bleeding (malnutrition, cancer, coagulopathies, cirrhosis,¹⁰ fall history,^17,18 renal impairment^10,19), total unique medications used in the previous year, specific medication classes associated with stroke and/or bleeding risk (angiotensin converting enzyme inhibitors/angiotensin receptor blockers,²⁰ non-steroidal anti-inflammatory drugs,²¹ antiplatelets,²² selective serotonin reuptake inhibitors,²³ statins²⁴), functional status (activities of daily living score),²⁵ and cognitive impairment (the MDS 3.0 Cognitive Function Scale).²⁶ Medication covariates were operationalized as any use in the prior year because recent medications, as well as current, can affect current anticoagulant use.

Nursing Home Characteristics

Nursing home characteristics were grouped conceptually as either structural, resource, staffing, or quality of care. Structural characteristics included size (number of beds) and specialized services available (specialized rehabilitation, laboratory, hospice). Larger nursing homes and those with rehabilitation services were expected to be associated with a larger volume of residents with atrial fibrillation, potentially developing internal or attracting external expertise.

Characteristics representing resources available to the nursing home included occupancy, for-profit status, and status as an individual or corporate entity. Not-for-profit nursing homes²⁷ and those with greater resources have traditionally achieved better care quality.²⁸ Organizations with greater resources, and those with multiple sites, may be more predisposed to have programs (e.g., quality improvement) and protocols that are associated with guideline adherence. Higher staffing has also been found to be positively associated with care quality.²⁹ Staffing was operationalized as quartiles of the minutes per resident-day of care from nurses and nursing assistants (i.e., all nursing care), registered nurses, prescribers (medical director, physicians, physician’s assistants, nurse practitioners), and pharmacists. Quartiles of the fraction of prescriber minutes per day contributed by physician extenders was included to evaluate if provider type was associated with prescribing. Quality of care was operationalized using the overall 5-star nursing home compare rating, which has been found to be associated with medication safety.³⁰

County Characteristics

Wide regional variation exists within the United States in adherence to guideline recommendations for primary and secondary prevention for atherothrombosis.³¹ To understand the contribution of county-level factors to variation in anticoagulant use, we considered socioeconomic factors, health system supply factors, and the county’s cerebrovascular mortality rate. Counties were grouped using the 2013 Rural/Urban Continuum Codes³². We also categorized as quartiles the proportion of the older adult (≥65 years) population on Medicaid and in deep poverty. Sociodemographic factors included the proportion of the overall population that identified with racial/ethnic minority groups, without a high-school education, and the proportion of single parents.

For health system supply variables, quartiles of the ratio of total physicians to the county population, the fraction of total physicians in primary care, and the ratio of cardiologists and neurologists to the population were considered because provider type has been associated with anticoagulant use.^33,34 Due to collinearity among these physician supply variables (variance inflation factors >4), only the ratio of cardiologists to the population was used in multivariable modeling. Because geriatricians are accustomed to the management of nursing home residents, the presence of ≥1 hospital with a geriatric service was also included (community-based supply was unavailable). The presence of ≥1 hospital with a medical school was included as a potential disseminator of best practices and new technologies, while quartiles of the ratio of hospitals to land area was included as an indicator of access to tertiary care.

Geographic Variation

Variation in prescribing quality for Medicare beneficiaries has been documented across healthcare markets and states,³⁵ and variation in prescribing of opioids has been observed in nursing homes.³⁶ To examine within and between state variation in prescribing of anticoagulants, we grouped residents into counties and states. Because counties do not cross state lines, differences in state policies can be largely excluded as a source of variability in prescribing between counties within states. We also evaluated interstate variation because of the concentration of Medicare Part D plans within states, and the variability in features between plans.³⁷

Statistical Analysis

Descriptive statistics summarized the study population by anticoagulation status for resident, nursing home, and county level characteristics.

We examined county rather than nursing home level variation because 79% of the 11,860 nursing homes with eligible residents contributed ≤10 residents each to the study. We first fit a logistic model with random intercepts of counties adjusted only for calendar year (i.e., the null model) from which we were able to graphically depict the geographic variation. Then, to evaluate the extent to which variation in anticoagulant use between counties could be explained by resident, nursing home, and county characteristics, and state, we fit four additional multilevel logistic models: 1) resident characteristics only, 2) adding nursing home characteristics, and 3) adding county characteristics, and 4) adding random intercepts for states (Methods Appendix). Time (year) was included in all models as a fixed effect. To compare the observed level of anticoagulant use for each county (i.e., the average of predicted probabilities from the null model) with the level of anticoagulant use that would be expected based on the composition of its resident population, we also estimated predicted probabilities from the fixed effects portion of the logistic model adjusted for resident characteristics. We then graphically depicted the difference between the observed and predicted levels of anticoagulant use for each county.

The rescaled proportional change in cluster variation (PCV)³⁸ was estimated across models to characterize between-county variation attributable to the explanatory factors included in each model. Intraclass correlation coefficients (ICCs) quantified the magnitude of correlation between residents within counties (all models) and within states across counties (for the model with random state intercepts).³⁸ To examine associations between specific resident, nursing home, and county characteristics, adjusted odds ratios and 95% confidence intervals were estimated from the full model with random state intercepts. The predictive capacity (c-statistic) of resident characteristics was estimated from a single level logistic model adjusted for time. Analyses were performed using SAS, version 9.4 (SAS Institute, Inc., Cary, NC).

Results

During the 2014, 2015, and 2016 cross-sections, 45% of the 86,736 nursing home residents diagnosed with atrial fibrillation were using oral anticoagulants. Residents were included from 11,860 nursing homes located in 1,694 counties (number of residents in median county: 27, Q1 17, Q3 50). The median age of the population was 85 years and 67.3% were women. The median CHA₂DS₂-Vasc score among treated and untreated residents was 5 (Q1 4, Q3 6), while the proportion with ATRIA scores indicating high bleeding risk were 37.7% and 43.2%, respectively (Table 1). The median nursing home size was 120 beds and the median occupancy was 88.6%. Most counties were metropolitan (77.4%) or urban areas adjacent to metropolitan areas (14.6%).

Table 1.

Characteristics of Residents with Atrial Fibrillation Treated or Not Treated with an Anticoagulant, 2014 to 2016

	Treated (n=38,693)	Untreated (n=48,043)
Year
2015	32.4	33.4
2016	36.1	32.6
Demographics
Age in years, median (Q1, Q3)	84 (77, 89)	86 (79, 91)
Women, %	67.3	67.3
Married, %	21.6	20.0
Medicaid eligible, %	79.3	78.5
Race/ethnicity, %
Non-Hispanic, White	87.9	86.5
Non-Hispanic, Black	8.5	8.8
Hispanic	1.3	1.7
Asian/Pacific Islander	1.1	1.7
Other/Unknown	1.2	1.3
Days since first observed nursing home admission, median (Q1, Q3)	828 (333, 1473)	810 (325, 1457)
Hospital admissions in prior year, %
1	46.8	48.8
2-3	37.2	36.0
4+	13.0	12.1
Ischemic stroke	9.1	6.0
Transient ischemic attack	1.7	1.2
Extracranial bleeding	5.4	8.2
Intracranial hemorrhage	0.5	1.6
Venous thromboembolism	4.0	1.7
Acute myocardial infarction	3.6	4.6
At least 1 inpatient surgical procedure in past 6 months	24.2	24.6
Unique medications, median (Q1, Q3)	18 (13, 23)	16 (12, 22)
Select prescription medications,* %
Nonsteroidal anti-inflammatory drug	16.2	17.5
Antiplatelet	8.5	19.0
Statin	59.7	50.2
Selective serotonin reuptake inhinitor	52.5	49.9
Angiotensin converting enzyme inhibitor or Angiotensin II receptor blocker	51.4	47.0
Select comorbidities, %
Diabetes mellitus	41.3	35.1
Heart failure	48.1	41.4
Hypertension	87.6	85.8
Coronary artery disease	30.8	32.6
Peripheral vascular disease	15.4	13.4
Anemia	33.7	38.6
Fall history
Fall with fracture in 6 months before last admission	1.0	1.1
Fall since admission	18.2	21.5
Hip fracture	2.4	3.0
Stroke	18.5	14.7
Aphasia	4.9	3.9
Hemiplegia	12.6	7.9
Coagulopathies	3.4	3.1
Malnutrition	2.8	4.0
Cancer†	5.9	6.8
Liver cirrhosis†	0.4	0.7
Renal impairment
Chronic renal insufficiency	30.9	30.2
End-stage renal disease	16.1	16.8
Dialysis	3.0	.3
CHA2DS2-Vasc Risk Score, %
2-3	11.8	4.7
4	22.4	25.6
5	28.2	27.6
6+	37.7	32.2
ATRIA Bleeding Risk Score, %
Low (0-3)	55.4	50.1
Intermediate (4)	7.0	6.8
High (5-10)	37.7	43.2
Level of cognitive impairment, %
Mildly impaired	26.8	26.1
Moderately to severely impaired	30.3	40.8
Activities of daily living score (0-16), median (Q1, Q3)	9 (7, 11)	10 (7, 11)
Nursing home characteristics
Number of beds, median (Q1, Q3)	120 (96, 162)	120 (97, 161)
Occupancy (percentage of beds), median (Q1, Q3)	88.7 (80.0, 94.0)	88.5 (80.0, 94.0)
Ownership, %
Government	5.7	5.8
For profit, individual/partner entity	11.8	12.0
For profit corporation	58.9	59.8
Non-profit church or other non-corporation	5.8	5.5
Non-profit corporation	17.8	16.9
Nursing Home Compare Overall Rating, %
1-2	33.7	34.5
3	19.7	19.4
4-5	46.6	46.1
Clinical lab available, %	81.3	81.3
Medical director, %	89.7	90.2
Physician and extender minutes/resident/day, median (Q1, Q3)	2.4 (0.9, 4.7)	2.5 (0.9, 4.8)
Percentage of minutes from physician extenders, median (Q1, Q3)	30.3 (0.0, 60.8)	30.9 (0.0, 60.9)
Nursing minutes/resident/day, median (Q1, Q3)	344.8 (303.3, 391.3)	346.3 (303.5, 393.1)
Percentage of minutes from registered nurses, median (Q1, Q3)	11.6 (7.7, 16.4)	11.3 (7.4, 16.1)
Pharmacist minutes/resident/day, median (Q1, Q3)	0.9 (0.0, 1.6)	0.9 (0.0, 1.6)
Hospice beds, %	0.9	1.0
Special rehabilitation services, %	3.0	3.0
County characteristics
Area sociodemographics
Proportion ≥65 years in deep poverty, median (Q1, Q3)	2.5 (2.1, 3.1)	2.5 (2.1, 3.2)
Proportion of adults ≥65 eligible for Medicaid, median (Q1, Q3)	12.0 (8.8, 16.6)	12.0 (8.8, 16.6)
Proportion of adults ≥25 years of age without high school diploma, median (Q1, Q3)	7.7 (6.2, 10.1)	7.8 (6.2, 10.2)
Single parent households per 10,000 persons, median (Q1, Q3)	405.3 (342.3, 459.4)	407.3 (347.6, 461.9)
Proportion non-white race/ethnicity, median (Q1, Q3)	18.8 (9.2, 31.7)	20.0 (9.9, 33.7)
Population density (persons per square mile), median (Q1, Q3)	381.2 (104.4, 1,429.6)	405.1 (113.0, 1,433.7)
Urban-rural continuum, %
Metro area	76.8	78.0
Urban area metro adjacent	15.0	14.3
Urban not adjacent to metro	7.0	6.7
Rural	1.3	1.0
Area healthcare resources
Hospitals
Hospitals/100 square miles, median (Q1, Q3)	0.6 (0.2, 1.7)	0.6 (0.2, 1.8)
≥1 hospital with geriatric services, %	68.1	69.1
≥1 medical school affiliated hospital, %	61.0	62.0
Providers
Physicians/10,000 persons, median (Q1, Q3)	24.0 (13.9, 37.3)	24.8 (14.1, 37.3)
Cardiologists/10,000 persons, median (Q1, Q3)	0.6 (0.3, 1.0)	0.6 (0.3, 1.0)
Neurologists/10,000 persons, median (Q1, Q3)	0.4 (0.2, 0.7)	0.4 (0.2, 0.7)
Primary care as fraction of all physicians, median (Q1, Q3)	30.1 (23.7, 37.8)	29.7 (23.4, 37.2)
At least 1 hospice provider, %	83.7	85.0
Cerebrovascular Mortality
Cerebrovascular deaths per 10,000 persons, median (Q1, Q3)	4.4 (3.5, 5.4)	4.4 (3.5, 5.4)

^*At least one Part D claim during the year

^†42 and 40 residents with missing MDS 3.0 data on cancer and cirrhosis, respectively, were categorized as not having an active diagnosis for these conditions

The estimates of anticoagulant use adjusted only for time were plotted by county (Figure 1) and convey variation between counties within states, between states, and across regions. There was a nearly twofold difference between the county with the lowest proportion of residents anticoagulated (McLennan, Texas: 31.1%) and the county with the highest proportion treated (Pottawattamie, Iowa: 58.9%). Most states were composed of counties in the highest (51.3% to 58.9%) and lowest (31.1% to 41.4%) deciles of anticoagulant use.

Figure 1. Percentage* of Nursing Home Residents with Atrial Fibrillation Receiving Treatment with Oral Anticoagulants in the United States, by County (n=86,736 residents within 11,860 Nursing Homes within 1,694 counties in 2014–2016).

*Estimated from a two-level logistic model including a random intercept term for county level variation and adjusted for calendar year. Counties with missing values were those with less than 11 residents.

Differences between observed and predicted values for the proportion of treated residents in each county based on resident characteristics (without county intercepts) were plotted (Figure 2). For counties in the top decile, 5.8% to 17.5% more residents were observed to be receiving treatment than predicted based on resident characteristics in those counties. For counties in the bottom decile, 19.2% to 5.8% fewer residents were receiving treatment than predicted based on resident characteristics. These population average estimates suggest that clusters of counties in certain regions (e.g., Northeast) appeared to have higher treated fractions than would be expected based solely on the characteristics of their residents, while for other regions the inverse may apply (e.g., the Pacific Northwest).

Figure 2. Difference between Observed* and Predicted† Percentage of Nursing Home Residents with Atrial Fibrillation Receiving Treatment with Oral Anticoagulants in the United States, by County (n=86,736 residents within 11,860 Nursing Homes within 1,694 counties in 2014-2016).

*Estimated from a two-level logistic model including a random intercept term for county level variation and adjusting for time. Counties with missing values were those with less than 11 residents

†Estimated from the fixed effects portion of a two-level logistic model adjusted for resident characteristics and time. Counties with missing values were those with less than 11 residents.

The crude and adjusted odds ratios for several resident characteristics underscore the importance of clinical factors for prescribing decisions (Table 2, Table A3). Increasing age, prior intracranial or extracranial bleeding, severe cognitive impairment, cirrhosis, and antiplatelet use each exhibited large inverse associations with treatment. The associations of nursing home and county characteristics were generally modest. Non-profit nursing homes, rural counties, and nursing homes with higher Nursing Home Compare ratings were each associated with increased odds of anticoagulant use.

Table 2.

Adjusted Odds of Receiving Treatment with an Oral Anticoagulant for Resident, Nursing Home, and County Characteristics Estimated from Multilevel Logistic Models with Random County and State Intercepts

	Adjusted Odds Ratio*
Year (reference: 2014)
2015	1.04 (1.00-1.08)
2016	1.18 (1.14-1.22)
Demographics
Age (1-year increase from mean 83.7 years)	0.97 (0.97-0.97)
Women	0.98 (0.94-1.02)
Married	1.09 (1.05-1.13)
Medicaid eligible	1.00 (0.96-1.04)
Race/ethnicity (reference: Non-Hispanic White)
Non-Hispanic Black/African American	1.00 (0.94-1.06)
Hispanic	0.90 (0.80-1.02)
Asian/Pacific Islander	0.72 (0.63-0.82)
Other/Unknown	0.92 (0.81-1.05)
Days since first observed nursing home admission (Q4 vs. Q1)	1.06 (1.01-1.10)
Hospital admissions in prior year (reference: 0)
1	0.98 (0.90-1.07)
2	0.98 (0.90-1.07)
3	0.97 (0.88-1.06)
4+	0.98 (0.89-1.08)
Ischemic stroke	1.49 (1.40-1.59)
Transient ischemic attack	1.38 (1.22-1.56)
Extracranial bleeding	0.63 (0.59-0.66)
Intracranial hemorrhage	0.26 (0.22-0.31)
Venous thromboembolism	2.88 (2.63-3.16)
Acute myocardial infarction	0.87 (0.80-0.93)
At least 1 inpatient surgical procedure in past 6 months	0.96 (0.93-0.99)
Unique medications (Q4 vs. Q1)	1.38 (1.32-1.45)
Select Medications, †
Nonsteroidal anti-inflammatory drug	0.87 (0.84-0.91)
Antiplatelet	0.32 (0.30-0.33)
Statin	1.41 (1.37-1.45)
Selective serotonin reuptake inhinitor	1.01 (0.98-1.04)
Angiotensin converting enzyme inhibitor/angiotensin receptor blocker	1.10 (1.07-1.14)
Select comorbidities,‡
Diabetes mellitus	1.09 (1.06-1.13)
Heart failure	1.29 (1.25-1.33)
Hypertension	1.09 (1.05-1.14)
Coronary artery disease	0.92 (0.89-0.95)
Peripheral vascular disease	1.23 (1.18-1.29)
Anemia	0.79 (0.76-0.81)
Fall history
Fall with fracture in six months before last admission	0.92 (0.80-1.06)
Fall since admission	0.86 (0.83-0.89)
Hip fracture	0.99 (0.91-1.09)
Stroke	1.17 (1.12-1.23)
Aphasia	1.14 (1.06-1.24)
Hemiplegia	1.48 (1.40-1.56)
Malnutrition	0.80 (0.74-0.87)
Liver cirrhosis	0.51 (0.42-0.63)
Cancer	0.83 (0.78-0.88)
Coagulopathy	1.09 (1.01-1.18)
Renal impairment (reference: none)
Chronic renal insufficiency	0.96 (0.93-1.00)
End-stage renal disease	0.88 (0.85-0.92)
Dialysis	0.75 (0.69-0.82)
CHA2DS2-Vasc Risk Score,‡ (reference: 2–3)
4	1.28 (1.21-1.34)
5	1.51 (1.44-1.59)
6	1.73 (1.63-1.83)
7+	2.08 (1.96-2.21)
ATRIA Bleeding Risk Score,‡ (reference: low 0-3)
Intermediate (4)	0.69 (0.65-0.73)
High (5-10)	0.75 (0.72-0.77)
Level of cognitive impairment (reference: none)
Mildly impaired	0.87 (0.83-0.90)
Moderately impaired	0.69 (0.66-0.71)
Severely impaired	0.49 (0.45-0.52)
Activites of daily living score (reference: 0-4)
5-8	0.98 (0.94-1.03)
9-12	0.96 (0.92-1.01)
13-16	0.82 (0.77-0.88)
Nursing Home Characteristics
Number of beds (Q4 vs. Q1)	0.98 (0.93-1.03)
Occupancy(Q4 vs. Q1)	1.01 (0.96-1.06)
Ownership (reference: government)
For profit, individual/partner entity	1.05 (0.97-1.13)
For profit corporation	1.06 (0.99-1.14)
Non-profit church or other non-corporation	1.12 (1.02-1.22)
Non-profit corporation	1.14 (1.06-1.23)
Nursing home compare overall rating,§ (reference: 1)
2	0.99 (0.94-1.04)
3	1.03 (0.98-1.09)
4	1.05 (0.99-1.10)
5	1.07 (1.02-1.13)
Clinical lab available	1.02 (0.98-1.06)
Medical director	1.01 (0.95-1.06)
Physician and extender minutes/resident/day (Q4 vs. Q1)	0.96 (0.91-1.02)
Proportion of minutes from physician extenders (Q4 vs. Q1)	1.00 (0.95-1.05)
Nursing minutes/resident/day (Q4 vs. Q1)	1.00 (0.95-1.05)
Proportion of minutes from registered nurses (Quartile 4 vs. 1)	1.06 (1.01-1.11)
Pharmacist minutes/resident/day (Q4 vs. Q1)	1.01 (0.97-1.05)
Hospice beds	0.92 (0.78-1.07)
Special rehabilitation services	0.92 (0.84-1.01)
County characteristics
Area sociodemographics
Proportion of adults ≥65 eligible for Medicaid (Q4 vs. Q1)	1.11 (1.01-1.21)
Proportion of adults ≥25 years of age without high school diploma (Q4 vs. Q1)	1.09 (1.00 -1.18)
Non-white race/ethnicity, (Q4 vs. Q1)	0.87 (0.79-0.96)
Adults ≥65 years of age in deep poverty (Q4 vs. Q1)	0.97 (0.91-1.04)
Single parent households per 10,000 persons (Q4 vs. Q1)	0.94 (0.86-1.01)
Urban-rural continuum (reference: metro area)
Urban area metro adjacent	1.02 (0.96-1.09)
Urban not adjacent to metro	1.05 (0.97-1.13)
Rural	1.18 (1.01-1.39)
Area healthcare resources
Hospitals
Hospitals per 100 square miles (Q4 vs. Q1)	1.08 (0.98-1.19)
At least 1 hospital with geriatric services, Hospitals affiliated with a medical school	0.98 (0.93-1.03)
One medical school affiliated hospital in the county	0.97 (0.92-1.03)
≥2 medical school affiliated hospitals in the county	0.98 (0.91-1.05)
Providers
Cardiologists per 10,000 persons, median (Q4 vs. Q1)	0.99 (0.91-1.07)
At least 1 hospice provider	0.93 (0.88-0.98)
Health
Cerebrovascular deaths per 10,000 persons (Q4 vs. Q1)	0.97 (0.90-1.04)

^*Adjusted odds ratios were estimated from a logistic model with random intercepts for county and state and fixed effects for resident, nursing home, and county characteristics

^†Individual medication estimates derived from a model omitting the number of unique medications

^‡CHA₂DS₂-Vasc Risk Score and ATRIA Bleeding Risk Score adjusted estimates derived from a separate model omitting the variables included in the scores. Adjusted estimates for comorbidities included in the scores were obtained from models omitting the CHA₂DS₂-Vasc Risk Score and ATRIA Bleeding Risk Score.

^§Nursing home quality estimates derived from a separate model omitting professional staffing variables because staffing is incorporated in the calculation of the quality score

Compared with the model with only a fixed effect for time and random intercepts for counties, resident characteristics explained one-third of between county variation (PCV:34.8%) (Table 3). Notably, age (PCV: -17.2%) was the resident characteristic associated with the greatest decrease in explained variation between counties, while cognitive function was the characteristic associated with the greatest increase (PCV: 24.5%) in explained variation (Table A4). The full model with resident, nursing home, and county characteristics and random state intercepts explained 65.5% of the between-county variation (versus the null model).

Table 3.

The Proportional Change in Between-County Variation in Oral Anticoagulant Use Explained by Resident Characteristics, Nursing Home Characteristics, County Characteristics, and State

Characteristics Included in Multilevel Models*
	Null Model	Resident	Resident & Nursing Home	Resident, Nursing Home, & County	Resident, Nursing Home, County, & State†
PCV (%)	Reference	34.8	46.5	55.8	65.5
ICCCounty (%)	1.8	2.2	1.8	1.5	1.2
ICCState (%)	-	-	-	-	0.6

^*All models included a fixed effect for time (calendar year) and random intercepts for counties

^†Includes random intercepts for counties and states

Abbreviations: proportional change in variance (PCV); intraclass correlation coefficient (ICC);

Within county correlation was weak in each of the models with only county random intercepts (ICC_County: 1.5% to 2.2%). In the model adjusted only for time, 1.8% of the individual variation in the propensity for anticoagulant use was due to systematic variation between counties, while the remaining 98.2% was due to systematic differences between residents. In the fully adjusted model with random county and state intercepts, the propensity for oral anticoagulant use was more closely correlated for two residents in the same county (ICC_county: 1.2%) than for two residents in different counties within the same state (ICC_state: 0.6%). The c-statistic from a logistic regression model adjusted for resident characteristics and time was 0.68.

Discussion

In this large national study of US nursing home residents with atrial fibrillation, several clinical risk factors for stroke and bleeding were strongly associated with anticoagulant treatment. This study builds on earlier work demonstrating an increase in anticoagulant use among US nursing home residents shortly after the direct-acting oral anticoagulants became available, and confirms that this increase cannot be explained by changes in observed resident characteristics. Residents in different counties were more similar in their use of anticoagulants after accounting for resident factors. Differences in the propensity for anticoagulant use between counties were further reduced after accounting for nursing homes and county factors. However, adjusting for age substantially increased between-county variance. This suggests residents of similar ages were treated differently across counties, and this variability was partially attributable to the characteristics of the nursing homes and counties in which they resided.

The clinical guideline for atrial fibrillation management recommends anticoagulation for patients with a CHA₂DS₂-Vasc risk score of ≥2 for men or ≥3 for women on the basis of Level A evidence⁶ (the recommendation was to treat scores of ≥2 for men and women during our study period).³⁹ Despite a clear recommendation for the use of anticoagulants, many individuals do not receive treatment, even among less medically complex community-dwelling populations.^4,40 The large variability we observed in use of anticoagulants for individuals and geographically suggests that the wealth of clinical information available to nursing home providers may not be used in a systematic manner. While variation introduced through a resident-centered shared decision-making process is appropriate,⁶ predictive information on the probability and functional consequences of outcomes under alternative treatment scenarios would be valuable, but is presently unavailable, to inform such a process.

Residents with cognitive impairment were substantially less likely to receive preventative treatment, consistent with prior nursing home literature.^41,42 Evidence in community-dwelling populations on the association of cognitive impairment with lower anticoagulant use is mixed.^43–47 While the benefit of stroke prevention diminishes with declining function because of a lower ceiling on potential recovery, atrial fibrillation is associated with cognitive decline, and anticoagulation may slow cognitive decline among patients with atrial fibrillation.⁴⁸

Several sociodemographic factors were associated with anticoagulant use. Residents who were currently married, a form a social support associated with more aggressive end of life care,⁴⁹ had slightly higher odds of anticoagulant use. Variation in anticoagulant use was observed between individuals of different races/ethnicities and between counties with varying racial/ethnic composition. Counties with the highest proportion of non-white residents had 13% lower odds of anticoagulant use, consistent with earlier findings of better processes and outcomes in nursing homes with higher proportions of white residents⁵⁰ and better quality of care in nursing homes located in neighborhoods with fewer minority residents.²⁸

Overall nursing home quality was modestly associated with higher odds of anticoagulant use. Prescribing quality has been found to be correlated with overall nursing home quality ratings.³⁰ With the exception of a modest positive association with registered nurse staffing, nursing home staffing was not associated with anticoagulant prescribing. Contextual variables describing the supply of specialist providers were evaluated because of earlier findings of variability in prescribing by provider type.^33,34 Although no strong associations were found, further investigation is needed to understand residents’ access to specialists.

The present study has limitations. Medication use was operationalized using information from Medicare Part D claims. However, in the nursing home, medication administration is overseen by medical personnel mitigating concerns regarding nonadherence. Over-the-counter medications including aspirin are not recorded in Part D. The results should not be generalized to sparsely populated rural areas because counties with fewer than 11 eligible residents during 2014–2016, largely concentrated in the West and Mountain West regions, were excluded. Finally, although we studied a large set of important factors, it is likely that other unavailable individual (e.g., number and proximity of children) and provider/nursing home factors (e.g., pharmacy prescribing policies) influence medication use.

Conclusions and Implications

Anticoagulant use between counties among nursing home residents with atrial fibrillation ranged from a minimum of 31% to a maximum of 59%. In more than a quarter of counties, observed rates of treatment were more than 5% above or below what was predicted based on resident characteristics. Comparative evidence and refinement of predictive algorithms specific to the nursing home setting may be warranted to inform residents, family, and providers making difficult decisions regarding the use of anticoagulants for atrial fibrillation in this vulnerable and clinically complex population.