Epidemiological Approaches to Multivariable Models of Health Inequity: A Study of Race, Rurality, and Occupation During the COVID-19 Pandemic

Abstract

Purpose:

Methods for assessing the structural mechanisms of health inequity are not well established. This study applies a phased approach to modeling racial, occupational, and rural disparities on the county level.

Methods:

Rural counties with disparately high rates of COVID-19 incidence or mortality were randomly paired with in-state control counties with the same rural-urban continuum code. Analysis was restricted to the first six months of the pandemic to represent the baseline structural reserves for each county and reduce biases related to the disruption of these reserves over time. Conditional logistic regression was applied in two phases—first, to examine the demographic distribution of disparities and then, to examine the relationships between these disparities and county-level social and structural reserves.

Results:

In over 200 rural county pairs (205 for incidence, 209 for mortality), disparities were associated with structural variables representing economic factors, healthcare infrastructure, and local industry. Modeling results were sensitive to assumptions about the relationships between race and other social and structural variables measured at the county level, particularly in models intended to reflect effect modification or mediation.

Conclusions:

Multivariable modeling of health disparities should reflect the social and structural mechanisms of inequity and anticipate interventions that can advance equity.

INTRODUCTION

The COVID-19 pandemic constituted a widespread shock throughout the United States (U.S.),^1-3 yet its impacts across the population were uneven. Marginalized racial and ethnic groups, rural populations, frontline workers, and communities with limited pre-pandemic socioeconomic reserves had disparately high rates of COVID-19 and its complications^4-14—a finding that reflects long-standing inequities in health and portends potential inequities in the post-acute sequelae of SARS-CoV-2 for years to come.¹⁵ Despite consistent demonstrations of health disparities in the scientific literature, the science of epidemiology and the practice of public health face a persistent challenge to move beyond descriptive methods and engage in mechanistic studies that identify and mitigate the structural bases of health inequity.

Several methodological hurdles have contributed to this persistent challenge: First, demographic variables such as race are often statistically modeled as individual-level phenomena, rather than as complex signals of systems-level resource distributions that lead, in turn, to group-level and individual-level disparities in health outcomes.^16,17 This convention has persisted despite knowledge that race variables do not represent biological mechanisms. Instead, race variables represent structural racism in the form of inequitable resource distribution or access.^17-24 Second, historically, the methods for integrating social and structural variables into analytic contexts have been better defined for long-range social, economic, and policy applications than for pandemic settings, where social and structural resources may be abruptly strained or exhausted.^1-3 Finally, in rural populations, the mechanisms of health disparities (e.g., distance to healthcare) are often conceived as distinct from those influencing racial and occupational disparities—even though more than one-fifth of the U.S. rural population is non-white and although racial and occupational segregation both exist within rural areas.²⁵

To investigate the structural characteristics of COVID-19 disparities during the initial shock of the pandemic in the U.S., we conducted a phased analysis of county-level disparities in COVID-19 incidence and mortality in rural counties in the first six months of the pandemic. We focused on this early period to explicitly examine the role of pre-pandemic, county-level structural reserves in facilitating or mitigating disease burden at a local level. Thus, we used variables that represent community-level access to resources (such as industry concentration and hospital density), as well as variables that are commonly represented as individual-level phenomenon (such as body mass index) to show how these variables may be reconceived as markers of county-level resources. We also specifically assessed the relationship of racial and occupational disparities within this rural context.

MATERIALS AND METHODS

Data Sources

County-level data depicting COVID-19 incidence and mortality, population characteristics, and rural-urban classification for all 50 states were obtained from publicly available sources (Table 1). County-level data for other social and structural variables were obtained as summarized in the Supplementary Information (Table A1). The Institutional Review Board reviewed the study and determined it not to be human subjects research.

Table 1.

Summary of county-level variables depicting population characteristics and COVID-19 incidence and mortality for non-metropolitan counties in all 50 U.S. states

Variable	Definition(s)	Source
Age	Two definitions: (1) the percent of the total population ≥65 years of age and (2) estimated average age in the county, computed by calculating the proportion in each Census category within intervals of approximately 5 years of age from birth through >85 yearsa	United States Census, 2019 Projections
Sex	Percent of the total population identified by the Census as female	United States Census, 2019 Projections
Race/ethnicity	Percent of the total population identified by the Census for each 2010 Census category for race and ethnicityb	United States Census, 2019 Projections
Rurality	Rural-urban continuum codes (RUCC, including RUCC codes 4-9 for non-metropolitan counties)	US Department of Agriculture Economic Research Service Codes, 2013 Edition
Population size	Estimated total population size, used as a denominator for COVID-19 incidence and mortality	USAFacts Population Data, 2019 Census projections
COVID-19 incidence and mortality	Incidence and mortality counts from 1/22/2020-7/9/2020c	USA Facts Incidence and Mortality Data

^aEstimated average age was calculated using the projected number of people in each age group weighted by age category (an ordinal variable in the dataset).

^bWe used racial categories that were not mutually exclusive to account for multiracial identities within each county (e.g., persons who identified with two racial categories would be represented in the proportion for both categories). Race variables were described using the conventions from the 2010 Census. We included the Census ethnicity categories (Hispanic or non-Hispanic) within the umbrella of racial categories for the purposes of this analysis, given that the category "Hispanic" is also racialized.⁵⁰ There is little consensus on the best practice for many of these terms, such as Hispanic or American Indian; surveys continue to show that individuals identifying with these groups are diverse in their naming preference, but overall country or tribe-specific names are preferred.^51-53

^cSince date-time misclassifications were anticipated due to reporting lags and data entry errors during this early period of the pandemic, we classified and analyzed all data in this range as belonging to a single period of approximately six months.

Study Period

Our objective was to assess the relationship of pre-pandemic, social and structural reserves in non-metropolitan counties with county-level disparities in COVID-19 incidence and mortality during the initial onset of the pandemic. We restricted the analysis to the first wave of the pandemic period (through 7/9/2020), guided by evidence that many baseline reserves were themselves impacted by infrastructure changes during the pandemic (e.g., lockdowns, emergency responses),^1-3 and that the accuracy and interpretation of such data (Table A1) for depicting county-level reserves may not be reliable or predictable across subsequent periods of the pandemic.³

County Pair Definitions

We limited our analysis to non-metropolitan counties (operationalized as RUCC codes 4-9). For each county, COVID-19 incidence and mortality rates were calculated using the cumulative number of reported infections or deaths attributed to COVID-19 during the study period, divided by the total population in the county and expressed as a rate per 1000 persons. To control for state-to-state variation, counties with COVID-19 inequities were defined using averages for COVID-19 incidence and mortality across all non-metropolitan counties in the same state; and these counties were randomly matched (with a 1:1 ratio) with in-state controls with the same rural-urban continuum code to account for differences in population size and proximity to a metropolitan area.²⁶ Counties with COVID-19 inequities were defined as those with documented COVID-19 incidence or mortality rates greater than or equal to one standard deviation above the in-state, non-metropolitan mean. Counties eligible to constitute in-state controls were defined as counties with rates less than the in-state, non-metropolitan mean.

Demographic, Social, and Structural Characteristics

We conducted hierarchical regression modeling²⁷ in two phases. In the first phase, we sought to identify the most prominent demographic correlates of non-metropolitan county disparities (Table 1; age, sex, and race)—that is, who was adversely affected by COVID-19 disparities. In the second phase, we sought to describe the apparent statistical relationships between the variables identified in the first phase and the social and structural characteristics in the affected counties (Table A1)—that is, what structural factors might have contributed to the observed disparities—guided by previous work examining racial inequities in the workforce,²⁸ sociopolitical arrangements in rural areas,^25,29-31 and scientific frameworks for measuring health inequities.^13,17,32 The phased approach was intended to examine multiple possible mechanistic patterns between variables (Figure 1).

Figure 1.

Schematic representing traditional (A) and proposed (B and C) relationships between social and structural exposures, health outcomes, and socially constructed variables such as race. We use “race” in quotes here to signify the heterogenous, composite construct measured by race variables in epidemiological studies. (A) In traditional epidemiological models, “race” is often positioned as either a confounder of the exposure (dotted line plus solid line) or as an exposure itself through which social or structural factors may mediate a relationship between “race” and the inequitable outcome (solid line). In the first of these scenarios, adjustment for “race” as a confounder may control for (and thus exclude) the mechanisms of inequities from resulting analyses. In the second of these scenarios, "race“ is implicitly positioned as a heterogeneously defined origin of the causal pathway, an approach that may lead to recapitulation of disparities with little mechanistic information about how to mitigate these disparities. We found evidence of this in our phased approach, where addition of Phase 2 variables to Phase 1 did not generate evidence of mediation. (B) Socially constructed demographic variables such as “race” may modify or moderate the effect of other social or structural factors on health outcomes, leading to complex patterns of inequity across intersecting categories. This is supported by broader evidence about racial inequities, and we found empirical evidence of this in the interactions between “race” and industry in this study. Evidence for mechanisms may provide actionable information about public health practice and health policy, by identifying contexts in which structural support and resources are most urgently needed. (C) If the inequitable impact of a social or structural factor on a health outcome is largely distributed on the basis of “race” (particularly if distributed onto a single, marginalized group), the “race” variable may be a proxy for this mechanism itself and may appear to statistically mediate the exposure-outcome relationship. We find evidence for this in the relationships between some socioeconomic variables and “race” variables in this study. The pathways in both (B) and (C) acknowledge that the origin of structural mechanisms of health and disease is structural (i.e., “race” signifies a mechanism of structural racism); and both may be overlooked in traditional models like those in (A).

Modeling Procedures

For each phase, multivariable, conditional logistic regression modeling was conducted using SAS 9.4 (Cary, NC). We applied a cutoff of p≤0.20 in an unadjusted model for a variable to be eligible for multivariable modeling and then a cutoff of p≤0.10 to remain in a multivariable model. For Phase 2, we also retained variables if they changed the odds ratio (OR) by 10% or more for any of the Phase 1 variables when added to the Phase 1 model to account for potential confounding.³³ We used a combination of manual and automatic selection procedures, including forward, backward, and stepwise methods, with the goal of identifying models that were most consistent across these methods (and thus, less sensitive to the choice of model-building procedure alone). Where more than one candidate model was identified, we prioritized those with the best fit by the Akaike Information Criterion (AIC) and with the most parsimonious combination of variables. We assessed for interactions among variables in the resulting models. For statistical tests of interaction, we used a p-value cutoff ≤0.05.

Screening of multiple variables for Phase 2 was not intended as independent hypothesis testing for each variable but rather as a method for identifying common patterns across types of variables from multiple datasets and for identifying the best-fit variables that measured similar constructs. Thus, we opted not to apply p-value corrections at this stage of screening. Likewise, since we considered a broad array of variables for potential inclusion in the multivariable models, we expected some redundancy or multicollinearity across the resulting sets of eligible variables. If two variables conceptually approximated the same construct, the variable with the strongest correlation and/or the most directly interpretable definition was retained. We also assessed for multicollinearity across variables that were eligible for multivariable modeling and removed variables that demonstrated strong multicollinear impacts.

The number of matched county pairs was used to guide the maximum number of variables in the multivariable models, expecting a ratio of pairs-to-variables of at least 15:1 (i.e., 200 matched pairs, maximum 13-15 variables).

RESULTS

Descriptive Characteristics

Average rural COVID-19 incidence and mortality rates differed across the 50 U.S. states (Table 2). A total of 217 and 220 counties were identified as having county-level inequities in incidence and mortality, respectively. After matching with in-state controls, a total of 205 and 209 county pairs were available for the incidence and mortality analyses.

Table 2.

Average COVID-19 incidence and mortality rates for non-metropolitan counties, where applicable, in all 50 U.S. states during the first six months of the pandemic’s widespread impact in the U.S.

State	Total N Non- Metro Counties	COVID-19 Incidence Mean (Median) ± SD (Diagnoses/1000)	Total N Counties ≥ 1 SD Incidence	COVID-19 Mortality Mean (Median) ± SD (Deaths/1000)	Total N Counties ≥ 1 SD Mortality
Alabama	38	14.36 (12.22) ± 8.68	7	0.403 (0.230) ± 0.418	6
Alaska	25	0.55 (0.33) ± 0.74	4	0.014 (0.000) ± 0.061	1
Arizona	7	21.47 (18.19) ± 17.52	1	0.555 (0.332) ± 0.544	2
Arkansas	55	8.07 (3.40) ± 14.77	4	0.106 (0.056) ± 0.138	9
California	21	1.64 (1.37) ± 1.40	3	0.012 (0.000) ± 0.022	3
Colorado	47	5.42 (3.08) ± 6.18	7	0.130 (0.000) ± 0.275	3
Connecticut	1	8.37	0	0.760	0
Delaware	0	-	-	-	-
Florida	23	11.92 (8.65) ± 8.32	5	0.168 (0.094) ± 0.188	4
Georgia	85	11.76 (9.79) ± 7.27	15	0.488 (0.190) ± 0.753	9
Hawaii	2	0.54 (0.54) ± 0.08	0	0	2
Idaho	32	4.01 (1.75) ± 5.39	5	0.037 (0.000) ± 0.077	4
Illinois	62	2.54 (1.50) ± 3.01	8	0.093 (0.000) ± 0.202	14
Indiana	48	5.68 (4.23) ± 6.30	2	0.231 (0.122) ± 0.288	6
Iowa	78	8.40 (4.64) ± 11.98	6	0.166 (0.053) ± 0.308	8
Kansas	86	4.34 (1.82) ± 9.57	5	0.034 (0.000) ± 0.119	5
Kentucky	85	2.48 (1.82) ± 2.51	9	0.090 (0.000) ± 0.200	7
Louisiana	29	15.27 (11.90) ± 11.91	2	0.480 (0.367) ± 0.438	4
Maine	11	0.72 (0.63) ± 0.45	3	0.048 (0.015) ± 0.104	1
Maryland	5	6.31 (6.23) ± 4.06	1	0.308 (0.108) ± 0.493	1
Massachusetts	3	3.51 (3.17) ± 1.77	1	0.295 (0.088) ± 0.438	1
Michigan	57	2.21 (1.34) ± 2.24	8	0.104 (0.066) ± 0.141	8
Minnesota	60	5.51 (2.59) ± 10.76	4	0.061 (0.000) ± 0.108	9
Mississippi	65	14.20 (12.74) ± 7.36	9	0.632 (0.490) ± 0.530	7
Missouri	81	2.25 (1.26) ± 2.99	7	0.049 (0.000) ± 0.166	4
Montana	51	0.99 (0.54) ± 1.57	6	0.037 (0.000) ± 0.182	2
Nebraska	80	4.43 (1.61) ± 9.61	6	0.039 (0.000) ± 0.088	12
Nevada	13	2.21 (1.49) ± 2.10	2	0.030 (0.000) ± 0.066	1
New	7	1.23 (1.00) ± 0.78	1	0.031 (0.023) ± 0.043	1
Hampshire
New Jersey	0	-	-	-	-
New Mexico	26	4.64 (1.79) ± 9.79	2	0.167 (0.019) ± 0.513	1
New York	24	3.12 (1.89) ± 3.87	2	0.151 (0.080) ± 0.194	3
North Carolina	54	6.83 (6.21) ± 4.92	7	0.161 (0.088) ± 0.199	7
North Dakota	47	1.76 (1.56) ± 1.21	7	0.017 (0.000) ± 0.055	5
Ohio	50	3.33 (1.73) ± 6.03	2	0.159 (0.047) ± 0.240	8
Oklahoma	59	3.77 (2.16) ± 6.84	3	0.093 (0.021) ± 0.198	5
Oregon	23	3.50 (1.66) ± 4.17	4	0.018 (0.000) ± 0.030	3
Pennsylvania	30	1.87 (1.37) ± 1.48	5	0.078 (0.047) ± 0.127	3
Rhode Island	0	-	-	-	-
South Carolina	20	9.63 (9.51) ± 4.20	2	0.325 (0.184) ± 0.374	2
South Dakota	57	4.91 (2.58) ± 7.41	5	0.066 (0.000) ± 0.227	4
Tennessee	53	6.32 (2.93) ± 14.32	2	0.050 (0.020) ± 0.072	7
Texas	172	5.68 (4.10) ± 6.04	18	0.109 (0.000) ± 0.242	16
Utah	19	4.72 (2.23) ± 6.96	3	0.053 (0.000) ± 0.154	1
Vermont	11	1.20 (1.11) ± 0.66	2	0.024 (0.017) ± 0.024	2
Virginia	53	7.33 (3.28) ± 9.89	6	0.301 (0.057) ± 0.618	5
Washington	18	2.23 (1.08) ± 2.85	2	0.028 (0.000) ± 0.046	3
West Virginia	34	1.44 (1.07) ± 1.40	4	0.039 (0.000) ± 0.111	1
Wisconsin	46	1.95 (1.57) ± 1.58	7	0.058 (0.022) ± 0.082	8
Wyoming	21	2.74 (1.96) ± 2.79	3	0.052 (0.000) ± 0.146	2
Total N of Eligible Case Counties			217		220
Total N of Case Counties with Matched Controls			205		209

SD: Standard deviation.

Phase 1: Demographic Variables

In unadjusted models, disparities in COVID-19 incidence were associated with age, sex, and race variables, whereas mortality disparities were associated only with race variables (Table A2). In Phase 1 adjusted models for COVID-19 incidence (Table 3), county-level disparities were correlated with the proportion of residents in the county who identified as Hispanic (7% increase per 1% increase in population who identified as Hispanic; 95% confidence interval, CI: 3-10%), Black/African American (5% increase per 1% of population, 95% CI: 3-8%), or Asian (35% per 1% of population, 95% CI: 0-83%) and inversely associated with the proportion of county residents ≥65 years of age. In the Phase 1 adjusted model for COVID-19 mortality (Table 3), county-level disparities were correlated with the proportion of county residents who identified as Black/African American (8% increase per 1% of population, 95% CI: 4-12%) or Native American/Indigenous (3% increase per 1% of population, 95% CI: 0-6%), whereas the odds of county-level mortality disparities were inversely correlated with the proportion of county residents who identified as white (7% decrease per 1% increase of population, 95% CI: 4-9%) (see also Table 3 footnote about multicollinearity between these variables). Adjustment for age did not significantly affect the other ORs shown in the Phase 1 mortality model.

Table 3.

Phase 1 adjusted models of demographic correlates of COVID-19 incidence and mortality

	COVID-19 Incidence: Phase 1 Model Adjusted OR (95% CI) N=205 pairs	COVID-19 Mortality: Phase 1 Modela Adjusted OR (95% CI) N=209 pairs
Age
Average Age
Percent ≥ 65	0.91 (0.85, 0.98)
Sex
Percent Female
Race
Percent Hispanic	1.07 (1.03, 1.10)
Percent Black/African American	1.05 (1.03, 1.08)	1.08 (1.04, 1.12)
Percent American Indian/Alaska Native		1.03 (1.00, 1.06)
Percent Asian	1.35 (1.00, 1.83)
Percent Native Hawaiian/Other Pacific Island Native
Percent White

^aIn the Phase 1 mortality model, the percent white variable was multicollinear with the percent Black/African American and percent American Indian/Alaska Native variables. The percent white variable was associated with a reduction in the odds of county-level disparity (OR 0.93, 95% CI: 0.91-0.96). We used the model shown here for the Phase 2 analyses (Table 4), but we confirmed the results were similar in a model using the percent white variable. OR: Odds ratio. CI: Confidence interval.

Phase 2: Social and Structural Characteristics

Tables A3 and A4 (Supplementary Information) summarize the results of the screening procedures for the other social and structural variables. For multiple Phase 2 variables, associations with county-level disparities in unadjusted models were attenuated or reversed after addition to the Phase 1 model without a corresponding attenuation or reversal of the associations of the Phase 1 variables with county-level disparities. We viewed this as a potentially important result for contextualizing the relationships of the variables examined in the study, pursuant to our study objective, and examined this further (see Structural Subanalysis below).

The Phase 2 models for COVID-19 incidence and mortality disparities are shown in Table 4. For incidence, hospital closures, hospital size, reliance on mining or manufacturing industries, and a baseline metric related to smoking all correlated with county disparities; and disparities persisted in counties with higher proportions of Black/African American or Hispanic residents. Notably, the disparity for counties with higher proportions of Hispanic residents (i.e., at or above the non-metropolitan U.S. county median proportion) was exacerbated in both manufacturing-predominant and mining-predominant counties, whereas mining was otherwise associated with reduced odds in non-metropolitan counties (Figure 1b and Supplementary Information, Table A5). For mortality, the proportion of residents in the county classified as unemployed and disabled, the proportion classified as obese, the proportion reliant on public transportation, and an index of poverty segregation all correlated with overall non-metropolitan county disparities (Table 4), while disparities persisted in counties with higher proportions of Black/African American or Native American/Indigenous residents.

Table 4.

Phase 2 adjusted models for COVID-19 incidence and mortality

COVID-19 Incidence: Phase 2 Model N=179 pairs		COVID-19 Mortality: Phase 2 Modela N=141 pairs
Variable	Adjusted OR (95% CI)	Variable	Adjusted OR (95% CI)
Percent ≥ 65 years	0.87 (0.80, 0.95)	Percent Black/African American	1.11 (1.05, 1.17)
Percent Hispanic	1.08 (1.04 1.12)	Percent American Indian/Alaska Native	1.02 (0.98, 1.06)
Percent Black/African American	1.07 (1.04, 1.10)
Percent Asian	1.17 (0.88, 1.56)	Proportion unemployed with a disabilityh	0.90 (0.84, 0.96)
Miningb	0.16 (0.04, 0.60)
Manufacturingc	2.65 (1.10, 6.37)	Proportion reported obesei	654.15 (3.55, >999.99)
Hospital closured	2.99 (1.08, 8.31)
Large hospitale	1.30 (0.99, 1.70)	Segregation of povertyj	>999.99 (103.67, >999.99)
Mid-size hospitalf	0.67 (0.50, 0.92)
Proportion reporting regular smokingg	0.02 (<0.001, 1.51)	Proportion public transit commutersk	0.85 (0.75, 0.95)

^aA Phase 2 multivariable model for mortality including the percent white variable instead of the percent Black/African American variable and the percent American Indian/Alaska Native variable demonstrated that the adjusted ORs for the social and structural variables were not significantly different compared to those shown in the table.

^bMining industry accounts for over 23% or 16% of county earnings or employment, respectively.

^cManufacturing industry accounts for over 23% or 16% of county earnings or employment, respectively.

^dAt least one complete hospital closure since 2005.

^eAt least 1000 employees.

^f500-999 employees.

^gProportion of residents who report regular smoking behaviors.

^hUnemployment rate for persons with any disability.

ⁱProportion of residents whose body mass index is 30.0 or higher. The range of values for this variable was broad, contributing to a high OR and wide CI which are expected to overestimate the association.

^jSegregation of residents with lowest 25% of incomes in county. The range of values for this variable was broad, contributing to a high OR and wide CI which are expected to overestimate the association.

^kProportion of residents reliant on public transit. OR: Odds ratio. CI: Confidence interval.

Structural Subanalysis

For several social and structural variables, the associations with county-level disparities dissipated after adjustment for Phase 1 variables, whereas the corresponding disparities associated with the Phase 1 variables remained fixed throughout the modeling procedures. We considered this to be preliminary evidence that the disparate impacts of some social and structural variables may be mediated or modified by their inequitable distributions across one or more demographic variables (Figure 1c). We undertook a brief subanalysis to examine this further. We selected social or structural variables that met all of the following criteria: (1) positive association with county-level disparity, defined as p≤0.10 in Table A3 or A4; (2) ≥10% change in the OR for the variable after adjustment for the Phase 1 variables, with a corresponding increase above the p-value cutoff; and (3) no change ≥10% in the ORs for any Phase 1 variables after adjustment for the social or structural variable. We then evaluated each variable meeting these criteria with individual Phase 1 variables in separate models to determine whether adjustment for one or more Phase 1 variables appeared to specifically attenuate this association.

For both incidence and mortality, the association of persistent poverty (overall and in childhood) was specifically attenuated by the proportion of Black/African American residents. For the incidence model, the association of cattle plants was specifically attenuated by the proportion of Hispanic residents and, marginally, Asian residents. Stratified analyses of these social or structural variables by dichotomized versions of the attenuating race variables (i.e., at or above the non-metropolitan U.S. county median proportion vs below the non-metropolitan U.S. county median proportion) confirmed that the marked majority of the observed disparities was distributed to the racial group (Supplementary Information, Table A6). The proportion of residents ≥65 attenuated associations for low education, persistent poverty in childhood, and cattle plants.

DISCUSSION

This study offers important methodological perspectives for population-level assessments of health inequity. When complex, socially constructed variables such as race are used to describe the distribution of inequity in populations, epidemiological models must determine the modifiable social and structural factors underlying these disparities. In this analysis focused within rural U.S. counties, the county-level “race” variable is modeled as a heterogeneous composite representing resource distribution mechanisms.³⁴ This aligns with guidance asserting that race is not a modifiable or causal variable but rather an assigned variable by which social inequality is distributed.^35,36 Importantly, this approach also views racial health inequities as preventable outcomes determined by the distribution of resources across groups and not by an essential identity of such groups.^37-39

By explicitly positioning race as a non-random variable by which explanatory exposures may be inequitably distributed, multiple possible relationships can be examined between race variables, social and structural exposures, and outcome variables (Figure 1). Effect modification is one such relationship (Figure 1b). In this analysis, we found evidence of interactions between racial disparities and industry reliance (Supplemental Information, Table A5), with a noticeable exacerbation of disparities in manufacturing-predominant counties and a new, relative disparity in mining-predominant counties with higher proportions of Hispanic populations. Such findings indicate that non-white communities may not only face disproportionate employment in hazardous industries⁴⁰, but that hazardous exposures may also be racially stratified within occupations.

Mediation between variables is also possible—but here, assumptions about these relationships must be closely examined. Traditional models may implicitly assume that exposures, such as inequitable occupational experiences or socioeconomic reserves, may mediate the relationship between a race variable and an outcome (Figure 1a). However, the opposite is also plausible (Figure 1c). We found evidence for this here, in our subanalysis of inequities associated with cattle plants and county-level poverty. As these results suggest, when an inequity is borne almost exclusively by one racial group, traditional analyses may overlook this disparity, by positioning race as a risk factor for inequity instead of as the variable through which a specific social or structural disadvantage is distributed onto a group.

Overall, this study has several strengths: The results demonstrate that assessments of inequity using multivariable models are sensitive to definitions of key variables and assumptions about the relationships between these variables. Here, we have proposed one approach to assessing possible mechanistic relationships between social and structural exposures and inequitable health outcomes. This approach positions race variables (and other variables commonly modeled as individual-level phenomena) as windows into the resource distribution mechanisms that lead to health disparities and, thus, that might be modified to mitigate inequities.

Our study also has several limitations: We acknowledge that our study relies on publicly available, county-level data that prohibits a complete contextual assessment of some variables and that constrains a granular understanding of heterogeneous experiences within and across rural counties.⁴¹ For example, we found that public transit was inversely associated with COVID-19 mortality, in contradiction to other studies.^42,43 However, in rural contexts, public transit may signal a stronger overall network of resources, including better access to transportation for health-related needs.⁴⁴ Likewise, we approach the association of mortality with county-level rates of obesity in this study with caution—both because body mass index is not a reliable health indicator^45,46 and because, like race, this variable may be understood at the county level as an indicator of access to health resources.^47,48 These limitations further affirm the importance of improving and expanding data systems and statistical modeling approaches focused on social and structural determinants of health.

CONCLUSIONS

The advancement of health equity requires rigorous examination of the presumed relationships between socially constructed variables like race, which are frequently used to describe disparities in epidemiological analyses, and the social and structural variables signifying the mechanisms by which disparities occur. As we have shown here (and as summarized in Figure 1), conventional approaches to multivariable models that contain socially constructed variables may systematically overlook or underestimate key structural mechanisms by which inequities persist and, thus, by which such inequities can be mitigated. We demonstrate one possible approach for testing these assumptions in multivariable models, using a phased analysis.

Our results also confirm the intersections of racial, rural, and occupational inequities during the initial onset of the COVID-19 pandemic in 2020 and reinforce the importance of epidemiological analyses that deliberately consider these complex intersections in public health practice and epidemic preparedness in rural areas. Future research is needed to expand both the availability and utility of robust social and structural data at the local and regional level,⁴⁹ including in rural areas and across time periods (e.g., as structural reserves fluctuate), and for working directly with local and regional partners to understand the context of the variables modeled and collaboratively develop interventions for advancing health equity.

Abbreviations and Acronyms:

COVID-19

Coronavirus disease due to SARS-CoV-2 infection

U.S.

United States

OR

Odds ratio

CI

Confidence interval