Urban–Rural Disparities in Deaths of Despair: A County-Level Analysis 2004–2016 in the U.S.

Abstract

Introduction:

The purpose of this study is to examine nationwide disparities in drug, alcohol, and suicide mortality; evaluate the association between county-level characteristics and these mortality rates; and illustrate spatial patterns of mortality risk to identify areas with elevated risk.

Methods:

The authors applied a Bayesian spatial regression technique to investigate the association between U.S. county-level characteristics and drug, alcohol, and suicide mortality rates for 2004–2016, accounting for spatial correlation that occurs among counties.

Results:

Mortality risks from drug, alcohol, and suicide were positively associated with the degree of rurality, the proportion of vacant housing units, the population with a disability, the unemployed population, the population with low access to grocery stores, and the population with no health insurance. Conversely, risks were negatively associated with Hispanic population, non-Hispanic Black population, and population with a bachelor’s degree or higher.

Conclusions:

Spatial disparities in drug, alcohol, and suicide mortality exist at the county level across the U.S. social determinants of health; educational attainment, degree of rurality, ethnicity, disability, unemployment, and health insurance status are important factors associated with these mortality rates. A comprehensive strategy that includes downstream interventions providing equitable access to healthcare services and upstream efforts in addressing socioeconomic conditions is warranted to effectively reduce these mortality burdens.

INTRODUCTION

Although improvements in socioeconomic conditions and medical and technologic advancements have contributed to reduced mortality rates over the past century, health disparities across various dimensions (e.g., race/ethnicity, gender, geographic location, age) are becoming larger than ever in the U.S.^1–4 Recent studies have indicated increasing geographic inequalities in life expectancy within the U.S.⁵ The regional variation in health outcomes has been explained by differences in community-level factors such as residential location, access to health care, access to healthy food, and proportion of college graduates.^6–10 Health disparities can arise or be exacerbated when there is unequal access to these opportunities or resources.

There has been a dramatic increase in mortality because of drug overdoses, suicide, and alcohol poisoning—referred to as deaths of despair or stress-related conditions (SRCs). From 2000 to 2015, the number of drug-poisoning deaths increased threefold.^11–14 This recent rise in SRC mortality was first reported among the non-Hispanic (NH) middle-aged White population in rural areas across the U.S.^11,12 A recent report showed that midlife death rates also rose significantly between 1999 and 2016 among people of color that was driven primarily by drug, alcohol, and suicide deaths but also because of increases in dozens of organ diseases such as hypertensive heart disease and liver cancer.^12,15

The sources of stressors that have led to increased mortality are not clear, although several explanations have been offered. Previous studies indicate that poor economic conditions, family distress, and chronic pain/illness along with lax regulations on prescription of opioids to physicians and consumers may have led to higher deaths from the 1990s through the mid-2010s.^16,17 To cope with stress, people may adopt unhealthy behaviors, such as the use of alcohol and drugs (e.g., narcotics, noxious substances, and opioids). From a regional economic development perspective, rural areas within the U.S. experienced a regional economic transformation where a large number of factories closed down and outsourced production to China and Mexico to lower labor costs, thus leading to massive job losses.^13,16,18 In general, rural areas are likely to have fewer socioeconomic resources than urban areas because of their physical location and lack of material and human resources,^19,20 and separation from others who may provide social support also could exacerbate mental health problems and thus increase the risk of drug, alcohol, and suicide mortality.²¹

Previous studies have shown that rural areas in the U.S. had higher age-adjusted death rates than metropolitan areas and have indicated that the risk of adverse health outcomes (e.g., depression, drug over-dose, suicide) is higher in rural areas owing to physical and social isolation.^22–28 Additional studies have examined all-cause mortality or drug-related mortality trends incorporating discrete rural–urban classification (i.e., metropolitan or nonmetropolitan on the basis of population size only), but few studies have investigated urban–rural disparities in drug, alcohol, and suicide mortality by incorporating other operational measures (e.g., population size, population density, remoteness, urban area as a proportion of total land area) to assess the specific characteristics of rurality.^29–33

The spatial distribution of cause-specific mortalities varies across geographic levels, and very little is known about the spatial patterns of inequalities in SRC mortality.^11,12,22,23 Currently, urbanization is continuing to grow at a fast pace globally, and there is a growing need to improve the understanding of the association between SRC mortality and urban/rural status. Thus, identifying and understanding the distribution pattern of SRC mortality across the U.S. is crucial for targeting the locations of greatest need.

To address the association between SRC mortality and urbanicity/rurality, the association between age-adjusted SRC mortality rates and a continuous measure of urban–rural status along with other covariates was examined by assessing the age-adjusted SRC mortality rates by quartiles and the respective average values. Bayesian modeling was used to assess county-level predictors of SRC mortality, and counties with elevated SRC mortality risk were identified by examining the posterior mean of the relative risk (RR). Ultimately, this study will provide insights on urban–rural disparities in SRC mortality and be useful for researchers and policy-makers to develop health interventions aimed at reducing SRC mortality.

METHODS

Study Sample

The study sample comprised deidentified mortality data spanning January 1, 2004 to December 31, 2016 obtained from the National Vital Statistics System after receiving access to restricted-use data from the National Center for Health Statistics. Foreign residents (i.e., non-U.S. residents) and residents in noncontiguous states and territories (e.g., Alaska, Hawaii, American Samoa, Guam, Northern Mariana Islands, Puerto Rico, and the U.S. Virgin Islands) were excluded from analyses.

County-level population data (years 2004–2016 aggregated) in 10 age groups (0–4, 5–14, 15–24, 25–34, 35–44, 45–54, 55–64, 65–74, 75–84, ≥85 years) were obtained from the National Center for Health Statistics Bridged-Race Population Estimates.³⁴ The 2000 U.S. standard population weights for the 10 age groups were used to compute age-adjusted mortality rates through the direct method.³⁵ The total number of counties included was 3,101. This secondary data analysis was deemed exempt from IRB review by Virginia Commonwealth University.

Measures

The observed deaths for this study were aggregated number of SRC deaths (drug, alcohol, or suicide death counts at the county level) as defined in previous literature,^36–38 and the observed deaths were modeled using a Poisson distribution with mean (i.e., e_i · θ_i, where e_i is the expected deaths, and θ_i is the RR in county i). Subsequently, the RRs were modeled on the basis of expected deaths that were derived from the 2004–2016 age-specific U.S. SRC mortality rate. Causes of death were coded on the basis of the 10th revision of the ICD-10. The following ICD-10 codes were used in this study: K70, K76.0, x40–45, Y10–15, Y45, Y47, Y49, x60–84, and Y87.0 (Appendix Table 1, available online, provides a complete list).

The independent variable for this study was the index of relative rurality (IRR). The IRR is a composite measure of urban/rural status that incorporates 4 dimensions (i.e., population size, population density, remoteness, and extent of urbanized area) (Appendix Table 2 available online).^39–41

Additional covariates were obtained from 5-year American Community Survey data for 2008–2012⁴² and were matched to the mortality and population data on the basis of state and county Federal Information Processing Standard Publication code. These county-level covariates included the percentage aged ≥25 years with a bachelor’s degree or higher, the percentage Hispanic, the percentage NH-Black, the percentage male, the percentage with a disability (i.e., difficulty with hearing, vision, cognitive, ambulatory, self-care, or independent living), unemployment rate (aged ≥16 years), the percentage aged ≥15 years who were married, the percentage of housing units that were vacant, and the percentage with no health insurance. In addition, the population with low food access (proportion of population in a county living >1 mile from a supermarket/grocery store in urban area or >10 miles from a supermarket/grocery store in rural area) was obtained from the U.S. Department of Agriculture Economic Research Service data.⁴³

Statistical Analysis

Because counties have different age distributions, age-adjusted mortality rates were computed to examine the spatial variation of SRC mortality rates throughout the U.S. The age-adjusted mortality rates were calculated using direct standardization, using the 2000 U.S. standard population in the calculation. As part of an exploratory data analysis, the association between the age-adjusted SRC mortality rates and their corresponding IRR was examined using locally estimated scatterplot smoothing, which is a nonparametric fitted line. In addition, the age-adjusted SRC mortality rates were grouped by quartiles, and the respective mean values of the county-level factors were examined and tested if there were any significant differences among these average values using 1-way ANOVA.

The age-adjusted mortality rates were visualized with choropleth maps to compare mortality risk among counties (Appendix 4, available online). The Jenks natural breaks classification method (5 levels) was used because it minimizes variation in each class and can guarantee comparability in counties, which conforms to the mapping objective of direct standardized rate maps.⁴⁴ Although a direct standardized rate is relatively easy to interpret and provides a sense of the mortality risk, it does not provide any statistical inferences about the underlying RR of SRC mortality with respect to multiple county-level factors. Moreover, the rates can be misleading and unreliable in areas with small populations. To address this issue, county-level covariates were included in a full hierarchical Bayesian regression, and the RR of SRC mortality was estimated. In addition, the county-level factors were standardized and included in the regression model to facilitate simpler comparison and interpretation. As for handling the potential nonlinear relationship between the explanatory variables and the outcome variable, the model (i.e., spatial generalized linear mixed model) in this study does not assume a linear relationship.

The hierarchical Bayesian models adopt multiple levels specified in a hierarchical order to estimate the posterior distributions of the parameters on the basis of the Bayes method.⁴⁴ The observation data were combined with the multiple sublevel model specifications (previous distributions) and covariates to estimate the posterior distribution through the Bayes theorem. These models can be used for data that are spatially structured with spatial autocorrelation.⁴⁵

The spatial data for health outcomes often indicates strong spatial autocorrelation such that disease cases or risk factors of nearby areal units tend to have similar values.⁴⁵ To deal with positive autocorrelation, spatially structured random effects were included in the model to account for the residual spatial autocorrelation.^45,46 Model fit was assessed with the Deviance Information Criterion (DIC) values. DIC is a Bayesian tool for model fit comparison, where the model with the smaller DIC is preferred. The DIC is computed by adding the posterior mean of the deviance (i.e., measure of goodness of fit) and the effective number of parameters (i.e., measure of model complexity).⁴⁷ The 3 candidate models for this study were: (1) a crude or unadjusted model that included IRR as the only exposure and accounted for the nonspatial variation; (2) an unadjusted model that included IRR as the only exposure and accounted for spatial correlation between neighboring counties and nonspatial variation by including both structured and unstructured random effects, respectively, with no other covariates; and (3) a fully adjusted model that incorporated both structured and unstructured random effects with IRR and all other covariates.

On the basis of the selected final model, the posterior mean estimates for the covariates were used to assess significance. To obtain the posterior distribution for parameters, Markov Chain Monte Carlo algorithm method is often used but can be intensive in terms of computation and time. Owing to its fast computation, efficient method, and relative accuracy, the integrated nested Laplace approximation (INLA) method was implemented with R package, R-INLA, to obtain the posterior distributions of parameters and random effects.^48,49 The mean RR for each predictor variable was computed by exponentiating the posterior mean estimates.

To identify counties with elevated SRC mortality risk with α % statistical certainty or exceedance probability (threshold probability set at 95%), the posterior mean SRC mortality RR (critical threshold set as the following: θ>1.50) adjusted for the explanatory variables was plotted. The exceedance probability was calculated by subtracting the probability (i.e., p [θ≤1.50]) from 1.

The specification of the statistical model and previous distributions for model parameters are presented in Appendix 3 (available online). Data cleaning and preparation were conducted in SAS 9.4, and Bayesian spatial analysis was conducted in RStudio, Version 1.3.1093.

RESULTS

On the basis of the scatterplot of the age-adjusted SRC mortality rates and the IRR, the mortality rates had a slightly increasing trend as the degree of rurality increased. However, both low and high mortality rates were observed among the more rural counties (i.e., IRR>0.4) (Figure 1).

Figure 1.

Association between county-level age-adjusted SRCs mortality rates and the index of relative rurality, 2004–2016.

Note: The Index of relative rurality ranges from 0 (highly urban) to 1 (highly rural).

SRC, stress-related condition.

Table 1 contains the gradient in SRC mortality across quartiles of each of the county-level social determinants of health. The following variables had a positive gradient, showing increased SRC mortality with larger values: male population, currently married population, the proportion of people with disability, vacant housing units (%), the proportion of the population with no health insurance, unemployment rate, and IRR. By contrast, the following variables had a negative gradient: Hispanic population, NH-Black population, and educational attainment (proportion of people with a bachelor’s degree or higher). An ANOVA test indicated that there were significant differences between the age-adjusted SRC mortality rate quartile groups for all the variables except for the Hispanic population and married percentages (Table 1).

Table 1.

Mean Values of County-Level Factors by Age-Adjusted SRC Mortality Rate Quartile

County-level factors	Counties grouped by SRC mortalitya				p-value
	Q1 (lowest SRC mortality)	Q2	Q3	Q4 (highest SRC mortality)
Demographic
Hispanic (%)	8.97	8.40	7.92	8.04	0.408
Non-Hispanic Black (%)	13.51	11.44	7.26	3.57	<0.001
Male (%)	49.24	49.07	49.35	49.46	<0.001
Married (%)	55.22	54.67	55.25	55.11	0.295
Disability (%)	13.73	14.46	15.48	17.88	<0.001
Socioeconomic
Bachelor’s degree or higher (%)	21.20	20.34	19.15	17.11	<0.001
Unemployment (%)	4.49	5.02	5.26	5.47	<0.001
Housing
Vacancy (%)	15.56	15.84	17.70	20.84	<0.001
Access to care
No health insurance (%)	13.83	14.21	15.03	17.15	<0.001
Geographic setting
Index of relative rurality (IRR)	0.51	0.49	0.50	0.52	<0.001
Low access to grocery stores (%)	26.54	21.33	22.00	23.19	<0.001

Note: Boldface indicates statistical significance (p<0.05).

IRR, Index of relative rurality; Q, quartile; SRC, stress-related condition.

^aThe Q1, Q2, and Q3 cut off values for SRC mortality were 21.42, 27.21, and 34.35, respectively.

The mean age-adjusted SRC mortality rate for 2004–2016 was 29.3 deaths per 100,000 population (minimum=1.8, Quartile 1=21.4, median=27.2, Quartile 3=34.4, maximum=326.8). Spatial variation in rates was identified, including clusters of high mortality rates presented in portions of Appalachia, Oklahoma, Florida, the Southwest, northern parts of California, the Pacific Northwest, and the Rocky Mountain West (Appendix 4, available online).

The results from the DIC values between the crude and adjusted model indicated that the fully adjusted model had a lower DIC value (26,372.47 vs 26,544.91), which translated to a better model fit (Appendix Table 5, available online).

After inspecting for potential outliers, a comparative assessment of testing a model with and without these extreme values indicated that there was no substantial disagreement or differences in the adjusted RRs. Because the potential outliers did not affect the model results, they were kept in the model. The fully adjusted model indicated that among the statistically significant county-level factors, the proportion of Hispanic and NH-Black population and the proportion of people with a bachelor’s degree or higher were associated with lower SRC mortality risk. For example, the risk of SRC mortality in a county decreases by a factor of 0.88 with a 1 SD increase in the proportion of people with a bachelor’s degree or higher in a county while controlling for other variables. Essentially, the same interpretation follows for other variables. By contrast, the proportion of people with a disability, the proportion of male population, the percentage of vacant housing units, the proportion of the population with low access to grocery stores, higher value of IRR (more rural), and the proportion of people with no health insurance were associated with higher SRC mortality (Table 2). In particular, the risk of SRC mortality in a county increases by a factor of 3.09 with a 1 SD increase in the IRR in a county while controlling for other variables.

Table 2.

Association with SRC Mortality Rate Based on Bayesian Spatial Analysis

County-level factorsa	Crude RR (95% CI)b	Adjusted RR (95% CI)c
Demographic
Hispanic (%)	0.59 (0.54, 0.63)	0.83 (0.80, 0.86)
Non-Hispanic Black (%)	0.69 (0.64, 0.74)	0.93 (0.89, 0.97)
Male (%)	1.46 (1.40,1.52)	1.03 (1.01,1.06)
Married (%)	1.37 (1.32,1.43)	0.99 (0.96, 1.02)
Disability (%)	2.16 (2.07, 2.26)	1.18 (1.14,1.22)
Socioeconomic
Bachelor’s degree or higher (%)	0.50 (0.48, 0.51)	0.88 (0.85, 0.91)
Unemployment (%)	0.93 (0.89, 0.98)	1.01 (0.98, 1.03)
Housing and transit to work
Vacancy (%)	1.98 (1.91, 2.06)	1.22 (1.19,1.25)
Access to care
No health insurance (%)	1.28 (1.21,1.35)	1.08 (1.05,1.12)
Geographic setting
IRR	4.30 (4.17, 4.44)	3.09 (2.97, 3.22)
Low access to grocery stores (%)	1.11 (1.06,1.16)	1.06 (1.04,1.09)

Note: Boldface indicates statistical significance.

CI, credible interval; IRR, Index of relative rurality; SRC, stress-related conditions.

^aAll county-level factors were standardized.

^bUnivariate model included each variable plus structured and unstructured random effects.

^cincluded IRR as the primary exposure variable with all other covariates plus structured and unstructured random effects.

The model-based mapping of the SRC mortality risk indicated that the counties that have 50% higher SRC mortality risk than the U.S. average with 95% statistical certainty (dark red-shaded areas) were concentrated in Appalachia, northern California, Florida, the Southwest, Oklahoma, and the Pacific Northwest (Figure 2). The map indicates apparent disproportionate disparities in SRC mortality outcomes across the U.S.

Figure 2.

Probability of RR>1.50 for stress-related conditions mortality by county, U.S., 2004–2016.

Note: Scale bar on the right displays the probability (%) of the RR exceeding 1.50.

DISCUSSION

This study examined the extent to which county-level social determinants of health explain SRC mortality across the U.S. and determined whether these vary across the urban–rural continuum. The key findings indicated that mortality risks from drugs, alcohol, and suicide were positively associated with the degree of rurality, the proportion of vacant housing units, the population with a disability, the unemployed population, the population with low access to grocery stores, and the population with no health insurance. Conversely, risks were negatively associated with the Hispanic population, NH-Black population, and population with a bachelor’s degree or higher.

In addition to the findings mentioned earlier, this study on SRC mortality adopted the Bayesian method, which can generate less biased results and better inference on the potential risk factors by incorporating a spatial method that accounted for the correlation between spatial units (e.g., spatial autocorrelation). Incorporating county-level characteristics with spatial structure can produce more reliable estimates of health outcome measures, which is useful for disease surveillance. Moreover, this study identified not only the county-level risk factors and spatial patterns of SRC mortality but also the hot spots with elevated risk.

Although few previous studies have reported mixed findings for suicide, drug poisoning, and alcohol-induced deaths compared with this study, the sensitivity analysis results of the association between the explanatory variables and 3 cause-specific mortalities (i.e., alcohol, drug, and suicide deaths analyzed separately) did not have any substantial differences in adjusted RRs.^31,50 The discrepant results may be attributed to the different analytical frameworks, models, measures of rurality (discrete versus continuous), and data periods incorporated. Moreover, these studies did not account for spatial autocorrelation, which is exhibited when certain measures in neighboring units are likely to be similar.

Consistent with recent research on drug-related mortality trends, this study shows that various domains of social determinants of health are associated with geographic variations in SRC mortality rates. Moreover, study results indicated significant disparities in SRC mortality rate risks. For example, the SRC mortality risk was higher among counties characterized by a higher degree of rurality and greater economic distress, including higher rates of disability, low access to grocery stores, no health insurance, and a lower proportion of the population with a college degree.^{30,36,51–55} By contrast, counties with a larger presence of Hispanic and/or NH-Black population and a higher proportion of the population with a bachelor’s degree or higher had a lower risk of SRC mortality. This is consistent with the findings of previous studies presenting associations between county-level characteristics and drug-related mortality rates.^55,56 Generally, unfavorable socioeconomic conditions can be linked to social isolation or family disintegration and ultimately undermine important supports against SRCs.

Future research should examine other potential contextual factors, such as opioid prescription rates, physician prescribing guidelines, drug surveillance data, and illicit drug trafficking offense trends that may contribute to the variation in county-level SRC mortality rates.

Limitations

This study could not account for decedent characteristics because of its ecologic and aggregated nature (i.e., grouped data by counties). Because counties with a small population typically had few deaths per year, aggregating years of data was applied to obtain reliable counts. This prevented examining potential cohort effects across the study period. A more refined ecologic analysis focusing on smaller geographic units (e.g., census tract) would be helpful in elucidating the association between community-level characteristics and drug, alcohol, and suicide mortality. Because there may be variations in resources and training across jurisdictions, death record reporting across counties may vary based on the ability to classify underlying causes of death.^55,57,58 A series of county-level factors was used to represent social determinants and contextual conditions of the counties, but there may be other relevant factors that were not captured. Finally, the results of this study were based on an ecologic study; thus, inferences or conclusions at the individual level should not be made.

CONCLUSIONS

SRC mortality rates from 2004 to 2016 were not randomly distributed across the contiguous counties in the U.S. Examining the spatial variation of drug, alcohol, and suicide mortality while controlling for county-level demographic and socioeconomic factors is crucial to understanding disparities in mortality outcomes and identifying areas with elevated risk. Findings from this study suggest that educational attainment, disability status, access to health care, and contextual conditions can serve as measures to develop more effective prevention and intervention programs to successfully target specific localities and tackle the deaths of despair epidemic.