The geography of arthritis-attributable pain outcomes: a county-level spatial analysis

Supplemental Digital Content is Available in the Text.

County-level prevalences of arthritis and arthritis-attributable pain outcomes have significant spatial clustering patterns, and factors shaping these patterns are different for different outcomes.

Abstract

Research on the geographic distribution of pain and arthritis outcomes, especially at the county level, is limited. This is a high-priority topic, however, given the heterogeneity of subnational and substate regions and the importance of county-level governments in shaping population health. Our study provides the most fine-grained picture to date of the geography of pain in the United States. Combining 2011 Behavioral Risk Factor Surveillance System data with county-level data from the Census and other sources, we examined arthritis and arthritis-attributable joint pain, severe joint pain, and activity limitations in US counties. We used small area estimation to estimate county-level prevalences and spatial analyses to visualize and model these outcomes. Models considering spatial structures show superiority over nonspatial models. Counties with higher prevalences of arthritis and arthritis-related outcomes are mostly clustered in the Deep South and Appalachia, while severe consequences of arthritis are particularly common in counties in the Southwest, Pacific Northwest, Georgia, Florida, and Maine. Net of arthritis, county-level percentages of racial/ethnic minority groups are negatively associated with joint pain prevalence, but positively associated with severe joint pain prevalence. Severe joint pain is also more common in counties with more female individuals, separated or divorced residents, more high school noncompleters, fewer chiropractors, and higher opioid prescribing rates. Activity limitations are more common in counties with higher percentages of uninsured people. Our findings show that different spatial processes shape the distribution of different arthritis-related pain outcomes, which may inform local policies and programs to reduce the risk of arthritis and its consequences.

1. Introduction

The prevalence of chronic pain has been increasing among US adults in the past 2 decades.⁴¹ By 2019, more than one-fifth of US adults (50.2 million in total) reported having chronic pain.³⁹ Arthritis, a painful health condition that is also increasing in prevalence over time,¹⁶ is a major contributor to chronic pain,^29,41 as well as a leading cause of disability.⁹ In 2013 to 2015, among individuals with arthritis, 26.8% (14.6 million) reported severe joint pain and 43.5% (23.7 million) had activity limitations caused by the condition.^2,3

There are 2 major gaps in our understanding of arthritis-related pain and functional limitations among US adults. First, research on the geographic variability in such outcomes is lacking, especially at the county level. Analyzing US states, 3 studies have documented great spatial variation in arthritis prevalence,⁴ arthritis-attributable joint pain,¹⁸ and general (all-cause) pain.⁴³ To the best of our knowledge, however, there are no comparable county-level analyses. In fact, we are only aware of 2 published county-level maps of pain-related conditions: one showing prevalence of arthritis⁴ and one showing pain experienced “yesterday” in a subset of counties for working-age White US adults.⁷ These maps suggest substantial within-state county-level variability in arthritis and in pain. County-level analysis, used widely in health research,^21,31,38 is critical for pain research as well because US states are not homogenous units (demographically, economically, or otherwise), and the distribution of arthritis-attributable pain and limitations may vary substantially within states. Moreover, counties are important governmental units that shape population health by allocating state funding, making local policies, and providing public services.^21,31

Second, even less research has examined the determinants of the geographic distribution of pain and related outcomes. According to the subnational spatial inequality framework, place-based characteristics, such as demographic composition and economic structure, play an important role in shaping subnational inequalities in health.²¹ However, previous studies on pain geography in the United States have not only focused on the state level but have also been largely descriptive.^4,43 Only 1 study¹⁸ included contextual state-level predictors of arthritis-attributable pain but conducted no smaller-unit analyses.

This study draws on the subnational inequality framework to document spatial variations in arthritis and 3 arthritis-attributable outcomes—joint pain, severe joint pain, and activity limitations—across the US counties and to identify county-level socioenvironmental determinants of these outcomes. Our findings contribute to the emerging scholarship regarding substate heterogeneity in pain-related outcomes and regarding potentially modifiable regional factors that may shape pain risk.

2. Data and methods

This is a county-level analysis that encompasses 2 stages. First, county-level prevalences of arthritis and 3 arthritis-attributable conditions—joint pain, severe joint pain, and activity limitation—were established through the small area estimation (SAE) approach. This approach transforms individual-level information into reliable estimates at the county level.⁴⁴ Second, we conducted spatial analysis to visualize the distributions of these county-level estimates and model them in multivariate regression models as a function of county-level characteristics. We describe these methods further after presenting our data.

2.1. Data sources

Our analyses used a combination of individual-level data from the Behavioral Risk Factor Surveillance System (BRFSS) and county-level data from other sources.

The BRFSS is a health-focused annual telephone survey of more than 400,000 residents aged 18 years and older in the 50 US states and the District of Columbia.⁸ Given its wide geographic coverage and large sample size, the BRFSS is the most popular and sometimes the only data source for obtaining small area estimates in the United States.⁴⁴ It has been used for county-level analyses of obesity,²⁸ cancer screening,⁵ and arthritis,⁴ but not for arthritis-related pain outcomes. In this study, we used the 2011 BRFSS data³⁰ because these are the most recent publicly available data with both county code and arthritis-related information.

County-level information comes from the following sources: (1) population counts, total and by demographic subgroups, used in the first stage (SAE) are from the US Census 2010⁶; (2) US county boundaries needed for visualization and multivariate spatial regression are from the US Census Bureau's 2010 TIGER/Line files³⁶; (3) county-level demographic and social and economic variables come from the American Community Survey (ACS) 2007 to 2011 5-year estimates³⁵; (4) variables about the availability of healthcare services are from the Area Health Resources Files (AHRF)¹⁷; and (5) county-level opioid prescribing rates come from the US County Opioid Dispensing Rates Map.¹⁰

2.2. Variables

2.2.1. Outcomes

Our outcomes are the county-level percentages of people who have (1) doctor-diagnosed arthritis, (2) arthritis-attributable joint pain, (3) arthritis-attributable severe joint pain, and (4) arthritis-attributable activity limitations (further referred to as arthritis, joint pain, severe joint pain, and activity limitations, respectively).

These county-level outcomes were constructed in the SAE stage using individual-level variables collected in the BRFSS. The individual-level outcome variables were recoded as dichotomous. The Behavioral Risk Factor Surveillance System asked respondents whether they have ever been told by a doctor or other health professional that they “have some form of arthritis, rheumatoid arthritis, gout, lupus, or fibromyalgia.” We defined those who answer “yes” as having doctor-diagnosed arthritis. Next, the BRFSS asked respondents with arthritis to rate their joint pain intensity from 0 to 10 (“during the past 30 days, how bad was your joint pain on average?”) and to indicate whether they are now limited in any way in any of their usual activities because of arthritis or joint symptoms. We define respondents with pain intensity scores above 0 as having “joint pain”; those with scores of 7 or above as having “severe joint pain” (following previous research²); and those who responded affirmatively to the last question as having arthritis-attributable activity limitations. Among people diagnosed with arthritis, 92.8% reported joint pain, 28.1% reported severe joint pain, and 51% reported activity limitations.

2.2.2. Individual-level covariates used in small area estimation

The individual-level covariates we used for the SAE stage were age, sex, and race/ethnicity. These are commonly used for SAE and are highly related to arthritis and arthritis-related pain outcomes.^2,3 Age was categorized into 13 groups: 18 to 24, 25 to 29, 30 to 34, 35 to 39, etc., through 75 to 79, and 80 years or older. Sex comprised male and female categories, and race/ethnicity was categorized into 8 groups: non-Hispanic White, non-Hispanic Black, Hispanic, Asian, Native Hawaiian/Pacific Islander, American Indian/Alaska Native, other single race, and multiracial.

2.2.3. County-level covariates used in multivariable regression modeling

Our county-level regression models included county-level demographic, social and economic, and healthcare-related characteristics that may shape the distribution of arthritis and related pain outcomes across counties. Arthritis prevalence (as calculated as part of this study) was also included as a covariate in models predicting 3 arthritis-attributable outcomes.

2.2.3.1. Demographic characteristics

We included county-level percentages of female individuals, non-Hispanic Black individuals, non-Hispanic Asians, and Hispanics in our analyses, based on evidence that women report more pain and arthritis than men; that both non-Hispanic Black individuals and Hispanics have higher prevalences of arthritis-attributable severe pain than White individuals^2,3,14,41,42; and that Hispanics and Asians have significantly lower prevalence of arthritis than White individuals.^2,3

2.2.3.2. Social and economic characteristics

Given the documented association between area-level socioeconomic conditions and population health outcomes,^11,32 we also included county-level social and economic factors: marital status (percentage of residents who are separated or divorced), educational attainment (percentage of residents older than 25 years with highest education less than high school), occupation (percentage of residents older than 16 years who work in agriculture, forestry, fishing and hunting, mining, construction, and manufacturing), poverty rate (percentage of residents who are living below the poverty level in the past 12 months), unemployment rate (percentage of residents who are older than 16 years who are in the labor force but unemployed), and lack of insurance (percentage of residents older than 18 years with no insurance coverage).

2.2.3.3. Healthcare-related characteristics

Because lower availability of healthcare providers is associated with worse population health outcomes,^11,23 our models included the number of primary care providers (ie, primary care physicians, both MDs and DOs) per 10,000 in 2011; the number of chiropractors (with national provider identifier) per 10,000 in 2011; and the number of hospitals per 1,000,000 in 2010. Moreover, because opioid prescribing is highly related to the prevalence of painful conditions,¹⁵ we also included the county-level opioid prescribing rate, measured by the total number of retail opioid prescriptions dispensed per 100 residents in 2011.

2.3. Approach

As noted, our analyses comprise 2 stages: small area estimation and spatial analysis, which have been described in more detail further.

2.3.1. Small area estimation

We used the SAE statistical model–based approach to estimate county-level prevalences of arthritis and arthritis-related conditions from individual-level BRFSS data. This method is widely applied to large national survey data to generate small area estimates when sample sizes in some geographic units are too small for reliable estimates.^28,44 There are 2 steps to SAE.

First, we aggregated counties with fewer than 5 individual observations in the BRFSS to larger county-like areas with adjacent counties that had observations. In our data, 895 counties (of the 3108 in the contiguous United States) had fewer than 5 observations, so were combined with adjacent counties. We used the Geographic Aggregation Tool (GAT, R version 1.61),³³ an iterative aggregation tool, to generate the 2213 county-like areas (hereinafter counties) with at least 5 observations each, imposing the restriction that counties must merge within states. Survey responses, population counts, and county-level variables were combined to match the aggregated counties. Supplemental Figure S1 is a map showing original and aggregated counties (available as supplemental digital content at http://links.lww.com/PAIN/B980). Our analytic sample included 425,166 respondents having complete information on age, race/ethnicity, sex, arthritis diagnosis, and Federal Information Processing Standard code for the county of residence.

Second, we used multilevel regression and poststratification (MRP), a powerful statistical method for SAE,^28,44 to estimate stable prevalences for all 4 conditions for the aggregated 2213 counties. Multilevel models that include individual-level variables (age, sex, and race/ethnicity), random effects for county and state, and rescaled survey weights (by state)⁴⁴ were estimated to obtain individual-level outcome probabilities (results available on request). Then, to perform poststratification, county population counts for each demographic group (208 in total: 13 age groups × 2 sex groups × 8 race/ethnicity groups) were multiplied by the corresponding average probabilities obtained from the multilevel regression models for each group. County-level prevalences were calculated as the summed frequencies for all 208 demographic groups divided by the total population counts in each county and shown as percentages.

Following prior work,²⁸ we checked these model-based estimates for internal validity, and the results were favorable at both the state and county levels (See supplemental Table S1, available at http://links.lww.com/PAIN/B980).

2.3.2. Visualization and modeling of county-level outcomes

After constructing county-level prevalences of 4 conditions with the prior SAE stage, we used spatial analysis to visualize and model the county-level outcomes. In both the visualization and modeling, county-level prevalence estimates were standardized using the 2010 age distribution of US adults to allow for more direct comparability across counties.

Thematic maps were created to visualize the geographic distribution of the 4 outcomes, color-coded by prevalence quintiles. To examine the spatial dependency of these 4 variables across counties, we calculated Moran's I, which indicates how much the value of an outcome for a focal county is correlated with values in neighboring counties. Higher values of Moran's I indicate stronger spatial clustering patterns. The spatial weight matrix was created based on the first-order queen method, which means that 2 counties will be defined as neighbors if they share the same boundary or vertex.²²

Finally, we estimated multivariable regression models to examine the association of the 4 outcomes with county-level characteristics. We used both nonspatial ordinary least squares (OLS) models and spatial regression models that account for spatial dependency among counties. The latter were important because ignoring spatial dependency can lead to biased coefficients.^24,37 Spatial dependency is conveyed with spatial parameters in such models. Three types of spatial models were estimated. Spatial lag models (SLM) incorporate a spatial lag parameter measuring the degree to which the outcome (eg, prevalence of arthritis) in a focal county is associated with the average values of prevalences in neighboring counties. Spatial error models (SEM) include a spatial error parameter capturing the association between the OLS residual of a county and that of its neighboring counties. Spatial autoregressive combined (SAC) models account for both spatial lag and spatial error parameters. All 3 spatial models were conducted for each outcome, and the estimates from the 3 models were substantively equivalent. In this study, we report only the best-fitting ones (ie, those with the lowest Akaike information criterion [AIC] and significant spatial parameters).

For 218 counties (approximately 10%) that were missing opioid prescribing rates, multiple imputation was used based on a Markov chain Monte Carlo model with 30 imputations. The relative consistency of the imputed values was always above 0.99. The final modeling process was conducted in R with the missing values replaced by the average value across the 30 imputed data sets.

3. Results

3.1. Results from small area estimation

County-level prevalences of arthritis and the 3 arthritis-attributable conditions, constructed through SAE, are summarized in Table 1. Variability for the 4 outcomes is extremely high across counties, with ranges more than 20% for each outcome, and even more than 30% for severe joint pain prevalence. Overall county-level prevalence of doctor-diagnosed arthritis, averaged across the 2213 counties (unweighted), was 26.32%, with individual county prevalences ranging from 14.88% to 43.16%. Arthritis-attributable joint pain prevalence was very similar, averaging 23.60% and ranging from 12.27% to 38.75%, again indicating that most people with arthritis experience joint pain. Arthritis-attributable severe joint pain, by contrast, was less common, with a cross-county average prevalence of 7.89% (range: 2.39%-34.21%). The average prevalence of arthritis-attributable activity limitations was slightly higher than the prevalence of severe pain, at 13.08% (and has a slightly narrower range: 6.61%-27.16%). Table 1 also summarizes descriptive statistics for county-level covariates used in multivariate regression models.

Table 1.

Characteristics of 2213 counties based on multiple resources.

Variable	N of counties	Mean	SD	Min	Max
Prevalence of arthritis and related conditions*
Percent of adults with arthritis	2213	26.32	4.03	14.88	43.16
Percent with arthritis-attributable pain
Joint pain	2213	23.60	3.92	12.27	38.75
Severe joint pain	2213	7.89	2.90	2.39	34.21
Activity limitations	2213	13.08	2.73	6.61	27.16
Demographic characteristics†
Percent female	2213	50.31	1.73	34.39	54.71
Percent non-Hispanic Black	2213	9.75	13.91	0.00	82.78
Percent non-Hispanic Asian	2213	1.32	2.33	0.00	33.29
Percent Hispanic	2213	7.88	11.86	0.04	98.45
Social and economic characteristics†
Percent separated or divorced	2213	13.34	2.30	5.79	22.79
Percent less than high school	2213	16.34	6.93	1.39	53.68
Percent in manual labor occupations	2213	25.39	7.33	4.38	54.00
Percent in poverty	2213	15.77	5.79	3.45	43.18
Percent unemployed	2213	8.29	2.70	1.32	25.25
Percent uninsured	2213	20.90	7.11	3.61	59.11
Healthcare-related characteristics
Primary care providers, per 10,000‡	2213	11.36	6.16	0.00	101.66
Chiropractors, per 10,000‡	2213	2.05	1.39	0.00	14.84
Hospitals, per 1,000,000‡	2213	34.95	28.69	0.00	218.97
Opioid prescribing, per 100§	1995	107.78	73.78	1.10	990.90

^*Prevalence of arthritis and related conditions are age-adjusted, model-based estimates constructed through SAE based on the BRFSS 2011 individual survey data.

^†Data for county-level demographic and social and economic characteristics are from the ACS 2007 to 2011 5-year estimates.

^‡Data for the number of healthcare providers are from the AHRF 2010 or 2011.

^§Data for the opioid prescribing rate are from the US County Opioid Dispensing Rates Map 2011. The variable is the number of opioid prescriptions per 100 persons.

ACS, American community survey; AHRF, area health resources files; BRFSS, Behavioral Risk Factor Surveillance System; SAE, small area estimation.

3.2. Results from visualization

Spatial analysis results are presented in Figure 1 (for the visualization) and summarized in Table 2 and Table 3 (for the multivariate regression modeling). Figure 1 shows the spatial distributions of doctor-diagnosed arthritis and the 3 arthritis-attributable pain outcomes. For all 4 outcomes, the value of Moran's I was above 0.5, indicating significant spatial clustering (ie, values in a focal county are positively associated with values in neighboring counties). As shown in Panel A, counties with higher arthritis prevalence are clustered in the Deep South, Appalachia, and Michigan. Given that more than 90% of respondents with arthritis report joint pain, it is unsurprising that Panel B, showing prevalences of joint pain, looks very similar to Panel A. However, Panel C, showing severe joint pain, reveals more geographical concentration (confirmed by a higher Moran's I): severe joint pain was more concentrated in the South (now including the Southeast) and the Southwest and less in the Upper Midwest, including Michigan. Activity limitations, shown in Panel D, also have a relatively high Moran's I and show “hotspots” not only in the Deep South, Appalachia, and Michigan but also in Maine, the Pacific Northwest, and the Southwest (especially Arizona and New Mexico). Different spatial patterns of the 3 arthritis-attributable outcomes may indicate diverse processes shaping these outcomes.

Figure 1.

Spatial distributions of county-level arthritis and arthritis-attributable pain outcomes (model based and age adjusted), BRFSS 2011. Panel (A): arthritis prevalence. Panel (B): joint pain prevalence. Panel (C): severe joint pain prevalence. Panel (D): activity limitation prevalence. BRFSS, Behavioral Risk Factor Surveillance System.

Table 2.

OLS and spatial regressions of model-based doctor-diagnosed arthritis.^†

	OLS		SAC
	Coef.	S.E.	Coef.	S.E.
Percent female	0.338***	0.072	0.198***	0.057
Percent non-Hispanic Black	−0.829***	0.084	−0.396***	0.060
Percent non-Hispanic Asian	−0.412***	0.081	−0.175**	0.058
Percent Hispanic	−1.578***	0.088	−0.568***	0.074
Percent separated or divorced	0.458***	0.087	0.275***	0.067
Percent less than high school	0.726***	0.123	0.221*	0.091
Percent in manual labor occupations	−0.132	0.089	−0.067	0.064
Percent in poverty	0.683***	0.112	0.288**	0.088
Percent unemployed	0.437***	0.087	0.217***	0.062
Percent uninsured	−0.341**	0.110	−0.188*	0.076
No. of primary care providers	−0.050	0.086	−0.182**	0.069
No. of chiropractors	−0.722***	0.085	−0.310***	0.066
No. of hospitals	−0.257***	0.077	−0.056	0.060
Opioid prescribing rate	0.291***	0.079	0.125+	0.064
(Intercept)	26.319***	0.066	8.888***	0.729
Rho‡			0.663***
Lambda§			−0.431***
AIC	11,333		10,929

Level of significance: +P < 0.1, *P < 0.05, **P < 0.01, ***P < 0.001.

^†The outcomes are model based and age adjusted. All covariates here are standardized.

^‡Rho is the spatial lag parameter, indicating how the prevalence in a focal county is associated with the average prevalence in neighboring counties.

^§Lambda is the spatial error parameter, indicating how the residual/error of a focal county is associated with the average residuals/errors in neighboring counties.

AIC, Akaike information criterion; OLS, ordinary least squares; SAC, spatial autoregressive combined.

Table 3.

OLS and spatial regressions of county-level arthritis-attributable joint pain, severe joint pain, and activity limitations.^†

	Joint pain				Severe joint pain				Activity limitations
	OLS		SAC		OLS		SLM		OLS		SAC
	Coef.	S.E.	Coef.	S.E.	Coef.	S.E.	Coef.	S.E.	Coef.	S.E.	Coef.	S.E.
Prevalence of arthritis	3.731***	0.024	3.679***	0.026	1.591***	0.045	1.426***	0.046	2.050***	0.039	1.870***	0.041
Percent female	0.039+	0.020	0.028	0.021	0.156***	0.038	0.106**	0.036	0.044	0.033	0.030	0.031
Percent non-Hispanic Black	−0.096***	0.024	−0.084**	0.028	0.607***	0.045	0.517***	0.044	0.009	0.039	0.058	0.040
Percent non-Hispanic Asian	−0.125***	0.023	−0.114***	0.025	−0.020	0.043	−0.008	0.041	−0.021	0.037	−0.033	0.037
Percent Hispanic	−0.059*	0.027	−0.026	0.031	−0.014	0.050	0.129**	0.048	−0.116**	0.043	0.006	0.046
Percent separated or divorced	−0.013	0.025	−0.011	0.025	0.149**	0.046	0.113*	0.044	0.102**	0.040	0.060	0.039
Percent less than high school	−0.046	0.035	−0.044	0.038	0.512***	0.065	0.324***	0.063	0.025	0.056	−0.016	0.057
Percent in manual labor occupations	0.054*	0.025	0.039	0.027	−0.050	0.047	−0.009	0.045	−0.132**	0.040	−0.087*	0.040
Percent in poverty	0.118***	0.032	0.086**	0.033	0.234***	0.060	0.183**	0.057	0.243***	0.051	0.159**	0.050
Percent unemployed	−0.042+	0.025	−0.040	0.027	−0.087+	0.046	−0.068	0.044	−0.103**	0.040	−0.080*	0.040
Percent uninsured	0.014	0.031	0.010	0.034	0.136*	0.058	0.088	0.056	0.446***	0.050	0.349***	0.051
No. of primary care providers	−0.018	0.024	−0.020	0.024	−0.001	0.045	−0.014	0.043	0.015	0.039	−0.002	0.037
No. of chiropractors	−0.022	0.025	0.001	0.026	−0.190***	0.046	−0.076+	0.044	−0.072+	0.039	−0.008	0.040
No. of hospitals	0.008	0.022	0.014	0.023	−0.027	0.041	−0.003	0.039	−0.032	0.035	−0.028	0.034
Opioid prescribing rate	0.007	0.023	0.003	0.022	0.125**	0.042	0.080*	0.040	0.027	0.036	0.027	0.034
(Intercept)	23.601***	0.019	22.705***	0.231	7.892***	0.035	5.744***	0.170	13.084***	0.030	10.437***	0.321
Rho‡			0.038***				0.273***				0.203***
Lambda§			0.219***				—				0.157***
AIC	5776.4		5697.1		8494.5		8338		7820.5		7611.4

Level of significance: +P < 0.1, *P < 0.05, **P < 0.01, ***P < 0.001.

^†The outcomes are model based and age adjusted. All covariates here are standardized.

^‡Rho is the spatial lag parameter, indicating how the prevalence in a focal county is associated with the average prevalence in neighboring counties.

^§Lambda is the spatial error parameter, indicating how the residual/error of a focal county is associated with the average residuals/errors in neighboring counties.

AIC, Akaike information criterion; OLS, ordinary least squares; SAC, spatial autoregressive combined; SLM, spatial lag models.

3.3. Results from regression modeling

The results of multivariate regression models predicting arthritis are summarized in Table 2. Smaller AIC values in the SAC model indicate improved model fit compared with OLS. The significant spatial parameters (ie, rho and lambda) show that the prevalences of arthritis are positively spatially correlated across counties but the residuals/errors are negatively spatially correlated. The magnitude of coefficients all decreased in the SAC model compared with OLS. The nonsignificant coefficient for the availability of primary care providers in the OLS model became significant in the SAC model, while number of hospitals ceased to be significant. Counties with more female individuals, fewer racial/ethnic minorities, and more people with marriage instability (ie, separated or divorced) or economic distress (ie, less educated, unemployed, or in poverty) had higher arthritis prevalence. A greater percentage uninsured and greater number of primary care doctors and chiropractors all predicted lower arthritis prevalence, while opioid prescribing rates were positively associated with arthritis. The variance inflation factor for each independent variable was below 4, indicating that multicollinearity is not an issue.

Table 3 summarizes results of models predicting the 3 arthritis-attributable outcomes. For joint pain and activity limitations, SAC models were the best-fitting models; both spatial parameters were positive and statistically significant, indicating significant spatial dependency for these outcomes and indicating that unobserved factors shaping these outcomes are positively spatially correlated. For severe joint pain, however, only the spatial lag parameter (rho) was significant, and SLM had the best fit. Again, the AIC values were smaller in the spatial models, showing superiority over OLS models.

Results in Table 3 summarize that even after controlling for arthritis prevalence (which necessarily positively and strongly predicts the 3 arthritis-attributable outcomes), several additional county-level factors are significantly predictive. The percentage of female individuals was positively associated with severe joint pain, but not the other outcomes. Racial and ethnic composition factors had significant associations with joint pain and severe joint pain but not activity limitations. Joint pain prevalence was negatively associated with the percentages of non-Hispanic Black individuals and non-Hispanic Asians, while 1 SD increase in the percentages of non-Hispanic Black individuals and Hispanics predicted 0.517 and 0.129 increases in severe joint pain prevalence, respectively. Regarding social and economic factors, county-level poverty rate was significantly positively predictive of all 3 outcomes but yielded larger coefficients for severe joint pain and activity limitations (0.183 and 0.159, respectively). Counties with higher percentages of people who are separated or divorced and people who had less than high school education had, on average, a higher prevalence of severe joint pain, while counties with a higher percentage of uninsured people had a higher prevalence of activity limitations. Of interest, the percentages of people in manual labor occupations and people unemployed were both negatively associated with activity limitations. Lastly, healthcare service factors showed significance only for severe joint pain. The number of chiropractors had a marginal negative impact on prevalence of severe joint pain, and opioid prescribing was positively associated with severe joint pain.

4. Discussion

This is the first study that described the distribution of arthritis and arthritis-attributable joint pain, severe joint pain, and activity limitations across US counties and linked the variability to key county-level characteristics. We found extremely high variability across counties in prevalence of arthritis and all 3 arthritis-related outcomes, even after adjusting for age. For example, the highest prevalence of arthritis-related joint pain was more than 3 times the lowest. All outcomes showed significant spatial clustering patterns, but they had different geographic “hotspots” and were predicted by different county-level characteristics. Counties in the Deep South, Appalachia, and Michigan had higher prevalences of most of the outcomes, while severe consequences of arthritis were also particularly common in counties in the South, Southwest, Pacific Northwest, and Maine. Generally, county-level social and/or economic disadvantages were linked to arthritis-attributable pain outcomes while healthcare-related characteristics played a less salient role.

We highlight 3 findings from our visualizations of the 4 arthritis outcomes. First, consistent with previous studies on the geography of all-cause pain⁴³ and of arthritis,⁴ the Deep South and Appalachia were “hotspots” for all 4 outcomes. We also found that Michigan had high prevalences of arthritis, joint pain, and activity limitations. These areas showed a high need for prevention and treatment of arthritis and arthritis-related outcomes. Second, different spatial patterns were observed for arthritis vs severe joint pain and activity limitations. Counties in the Southwest, the Pacific Northwest, Georgia, Florida, and Maine did not have particularly high arthritis prevalence but did have more high-impact sequelae of arthritis (ie, severe joint pain and/or activity limitations). This may indicate worse/inadequate treatment in the earlier stages of arthritis in these areas. Third, areas in Texas, Arizona, and the South Atlantic showed particularly high spatial heterogeneity in these outcomes, indicating that specific counties may need extra attention and that more localized (substate) policies may be needed.

The different spatial patterns of the 3 arthritis-attributable outcomes indicate that these outcomes may be shaped by different spatial processes. This is confirmed by results of our spatial models. Net of arthritis prevalence, counties with higher percentages of racial/ethnic minority groups have a lower prevalence of joint pain but a higher prevalence of severe joint pain. This accords with previous findings that racial/ethnic minority groups have lower risk of arthritis and chronic pain compared with White individuals, but this advantage disappears or even reverses when focusing on more severe pain outcomes.^14,40,41 This may reflect minority groups' documented difficulties in getting appropriate and timely treatment for arthritis and pain.²⁵ In addition, areas with a high concentration of racial/ethnic minority groups may be characterized by segregation and/or other aspects of neighborhood disadvantage, which also reduce healthcare quality.^12,20 To improve racial/ethnic health equity, more attention must be paid not only to arthritis diagnosis but also to the severe pain that can result from it and to the local-level factors that shape the risk of such pain.

We also found that counties with higher percentages of socially disadvantaged groups had higher prevalences of more severe outcomes from arthritis. Specifically, counties with more separated or divorced residents or with less educated residents had higher prevalence of severe joint pain, suggesting that marriage instability and lower education are linked to higher pain intensity. Counties characterized by low levels of education usually also have more low-income residents, low-skilled jobs, and unstable labor markets,^1,34 which could contribute to higher prevalence of severe joint pain. The percentage of people without insurance was negatively associated with arthritis prevalence but positively associated with activity limitations. This may be because uninsured people were less likely to see a doctor and hence to receive a diagnosis of arthritis, while for those who do receive a diagnosis, lack of insurance may limit access to care and lead to undertreatment and hence functional limitations. Somewhat unexpectedly, our results also showed that the unemployment rate and the percentage of people in manual labor occupations were both negatively correlated with prevalence of activity limitations. Perhaps people with activity limitations are out of the labor force or cannot take jobs involving manual labor. In addition, people who are unemployed may be less likely to report activity limitations because their usual activities may not be as affected by their arthritis. More research is needed to disentangle the causal associations between activity limitations and work-related factors.

Lastly, a greater availability of primary care doctors and chiropractors is associated with lower arthritis prevalence, and chiropractors are also associated with a lower prevalence of severe joint pain. This echoes previous studies showing the particular benefits of chiropractors for patients with long-lasting and severe pain.²⁶ Net of arthritis prevalence, however, most healthcare resources are not predictive of arthritis-attributable outcomes. In supplementary analyses (available on request), we included measures of the availability of other healthcare providers related to pain management, such as physical therapists, but did not find significant impacts of these variables on the arthritis-attributable outcomes either. This suggests that healthcare services seem to help relatively little in avoiding consequences of arthritis once people are diagnosed with arthritis. In addition, opioid prescribing is positively linked with severe joint pain. This may indicate that people experiencing severe pain obtain more opioids or that higher opioid use leads to more severe pain, as some research suggests.^13,19,27

This study is subject to several limitations. First, the data are from 2011. Spatial patterns may have changed since then; research using more recent data is needed. In addition, our study uses cross-sectional data, which cannot reflect the dynamics of how arthritis-attributable pain outcomes develop across time and space, and hinders causal interpretations of associations between covariates and pain outcomes. Ideally, future research will be conducted using spatiotemporal data sets with data from multiple years. Another limitation is that our pain outcomes are all arthritis related. Our findings thus cannot be generalized to other types of pain, which could be further investigated with different data. In addition, because we used aggregate-level analysis, the study is subject to ecological fallacy and the modifiable areal unit problem. Further research is thus needed to examine whether the results would hold at other levels of analysis.

5. Summary and conclusion

This study provides the most fine-grained geographic estimates of pain outcomes in the United States to date and is among the first to conduct a spatial analysis of arthritis-attributable pain outcomes and to identify their county-level predictors. We find significant spatial clustering patterns across counties in prevalences of all 4 tested examined conditions (arthritis, arthritis-attributable joint pain, severe joint pain, and activity limitations). Our results identify counties at higher risk of increased prevalence of arthritis, chronic pain, and activity limitations, calling for tailored policies at the local government level. Specifically, attention should be directed to counties with a concentration of disadvantaged populations, including racial and ethnic minorities and individuals facing family and economic distress. Future research should further explore how different factors may shape the risk of different arthritis-related outcomes. These efforts are crucial for improving geographic health equity and mitigating the severe consequences of arthritis.

Conflict of interest statement

The authors have no conflict of interest to declare.

Appendix A. Supplemental digital content

Supplemental digital content associated with this article can be found online at http://links.lww.com/PAIN/B980.