Where Is Air Quality Improving, and Who Benefits? A Study of PM_2.5 and Ozone Over 15 Years

Abstract

In the United States, concentrations of criteria air pollutants have declined in recent decades. Questions remain regarding whether improvements in air quality are equitably distributed across subpopulations. We assessed spatial variability and temporal trends in concentrations of particulate matter with an aerodynamic diameter ≤2.5 μm (PM_2.5) and ozone (O₃) across North Carolina from 2002–2016, and associations with community characteristics. Estimated daily PM_2.5 and O₃ concentrations at 2010 Census tracts were obtained from the Fused Air Quality Surface Using Downscaling archive and averaged to create tract-level annual PM_2.5 and O₃ estimates. We calculated tract-level measures of: racial isolation of non-Hispanic Black individuals, educational isolation of non–college educated individuals, the neighborhood deprivation index (NDI), and percentage of the population in urban areas. We fitted hierarchical Bayesian space-time models to estimate baseline concentrations of and time trends in PM_2.5 and O₃ for each tract, accounting for spatial between-tract correlation. Concentrations of PM_2.5 and O₃ declined by 6.4 μg/m³ and 13.5 ppb, respectively. Tracts with lower educational isolation and higher urbanicity had higher PM_2.5 and more pronounced declines in PM_2.5. Racial isolation was associated with higher PM_2.5 but not with the rate of decline in PM_2.5. Despite declines in pollutant concentrations, over time, disparities in exposure increased for racially and educationally isolated communities.

Abbreviations

CrI

credible interval

EI

educational isolation

EPA

Environmental Protection Agency

NHB

non-Hispanic Black

NDI

Neighborhood Deprivation Index

O₃

ozone

PM_2.5

particulate matter with an aerodynamic diameter ≤2.5 μm

RI

racial isolation

SES

socioeconomic status

WAIC

Watanabe-Akaike information criterion

An extensive literature has demonstrated that exposure to ambient air pollution is harmful to health (1). Exposures to ozone (O₃) and particulate matter with an aerodynamic diameter ≤2.5 μm (PM_2.5) are linked with adverse health outcomes, including mortality (2–4), cardiovascular (5, 6) and respiratory disease (7, 8), hospital admissions (9, 10), and adverse pregnancy outcomes (11–13), among others. However, exposure to air pollutants and health impacts at a given level of exposure are not necessarily uniform across all communities and subpopulations.

In the United States, racial/ethnic minorities and low–socioeconomic status (SES) communities may be disproportionately exposed to specific air pollutants (14, 15). Studies have demonstrated that ambient PM_2.5 concentrations tend to be higher in communities with higher proportions of non-Hispanic Black (NHB) or Hispanic residents, higher poverty levels, and greater degrees of racial segregation and urbanicity (16–18); these community characteristics are also linked with health outcomes (19–21). A 2019 study showed a decline in PM_2.5 exposure in the United States between 2005 and 2015 but found that pollution inequity, or the difference between the environmental health damage caused by a racial/ethnic group and the damage that group experiences, remained high for NHB and Hispanic persons (22). In addition to differential exposures, some populations may be more susceptible to health effects associated with air pollution. A recent study observed a more pronounced mortality risk associated with PM_2.5 exposure among racial/ethnic minorities and low-income individuals (23).

Overall concentrations of criteria air pollutants in the United States have declined in recent decades. The US Environmental Protection Agency (EPA) estimated that daily 8-hour maximum O₃ concentrations declined by 22% between 1990 and 2017, while 24-hour average PM_2.5 concentrations declined by 40% between 2000 and 2017 (24). Despite these improvements, 120 million Americans resided in areas with PM_2.5 or O₃ concentrations above the level specified by the National Ambient Air Quality Standards (NAAQS) in 2021 (25, 26). Furthermore, while the increased health risks at concentrations in excess of the NAAQS are well-established (25), there is a growing body of evidence that air pollution is detrimental to health even at lower concentrations (23, 27). Associations between PM_2.5 and O₃ exposure and health outcomes, specifically all-cause, cardiovascular-, and respiratory-related hospitalizations (27) and mortality (23), persist at concentrations below the NAAQS.

Despite national/overall improvements in air quality, it is unclear how consistently the air quality is improving across communities/populations, where the greatest improvements are, and thus who benefits. It is also unclear whether the magnitude/pace of the decline is similar across communities, or whether some populations are experiencing a slower decline in pollution and, therefore, an attenuated risk reduction. Disparities in air pollution exposure may in fact widen if air quality is improving more rapidly in areas that already have lower levels. These issues raise questions of environmental (in)justice, which is a complex concept with multiple definitions involving exposures, outcomes, and processes related to the experience of different subpopulations (28).

To address these questions, we first described levels and temporal trends in PM_2.5 and O₃ concentrations across North Carolina over 15 years (2002–2016). We then evaluated whether baseline levels and temporal trends in PM_2.5 and O₃ concentrations related to community-level characteristics, including measures of racial isolation (RI), educational isolation (EI), neighborhood deprivation, and urbanicity. Measures of RI, EI, neighborhood deprivation, and urbanicity were selected because previous work has demonstrated associations between: 1) urbanicity, SES, and air pollutants, specifically O₃ and PM_2.5 (29, 30); and 2) RI and PM_2.5, even after controlling for urbanicity (16). As a recently developed measure, less is known about relationships between EI, air pollution, and health, but educational attainment is often used as a proxy for SES that is also associated with health outcomes (30–32). Overall—across all metrics—we hypothesized that communities with higher air pollutant concentrations (of either air pollutant) were more likely to have larger declines in pollutant concentrations, because such communities have more “room for improvement.”

METHODS

Data

Air pollution data.

Estimated concentrations of PM_2.5 and O₃ in North Carolina were obtained for 2002–2016 from the publicly available EPA Fused Air Quality Surface Using Downscaling (“downscaler”) data. The downscaler utilizes a hierarchical Bayesian space-time modeling framework that combines gridded output from the Community Multiscale Air Quality model with monitoring data from the National Air Monitoring Station and State and Local Air Monitoring Station networks to produce daily point-level concentration estimates at the 2010 Census-tract centroids across the United States (33–35). Archived daily downscaler surfaces are available from the US EPA. Detailed descriptions of the downscaler modeling technique and performance are provided elsewhere (34). Downscaler output includes estimates of 24-hour average PM_2.5 and 8-hour maximum O₃ concentrations at census-tract centroids for each day in the study period (2002–2016). We averaged daily values to generate an annual average PM_2.5 and 6-month (April–September) average O₃ concentration estimates for each tract. Six-month averages of O₃ were calculated for April–September because this warm season is when O₃ concentrations tend to be highest, and when the most O₃ monitors are typically operational.

Racial isolation of non-Hispanic Black individuals.

Massey and Denton (36) identified 5 dimensions of racial residential segregation, namely, evenness/dissimilarity, exposure/isolation, concentration, centralization, and clustering. Later, it was determined that clustering, centralization, and concentration were also measures of evenness, thus simplifying the conceptual framework to evenness and exposure/isolation (37, 38). Additionally, a review of segregation and health noted that studies had conceptualized segregation in one of 2 ways (39): 1) a formal measure of geographical segregation of racial groups with indices reflecting either exposure/isolation, evenness, concentration, centralization, or clustering (36); and 2) a proxy measure (e.g., Black racial composition, % NHB). Prior assessments of segregation and health have found it conceptually problematic to conflate formal vs. proxy measures. Thus, we chose to employ a formal measure of RI because of the importance of isolation as one of the 5 domains of racial residential segregation long defined in the literature, and the domain that may be most closely related to adverse and disparate outcomes for racially, economically, and educationally minoritized populations (19, 40).

Using 2010 US Census data on the percentage of tract population self-identifying as NHB, we use a local spatial measure of RI to quantify geographic separation of NHB from other racial groups (41). The RI index is calculated based on the racial composition (e.g., % NHB) of tracts neighboring a given index tract i. We defined “neighbor” by adjacency, such that neighbors are tracts sharing a border or vertex with index tract i. The RI index ranges from 0 to 1 and represents a weighted average proportion of NHB in the local environment. For example, individuals in a neighborhood environment (i.e., index tract and neighboring tracts) that is predominantly non-NHB will have an RI value close to 0. In contrast, individuals living in a neighborhood environment that is nearly all NHB will have an RI value that is close to 1. The RI index was computed for all tracts within the study area with non-zero population in the neighborhood environment of the index tract. Based on previous work (16), we hypothesized that higher RI communities may have higher PM_2.5 concentrations but not necessarily higher O₃ concentrations.

Educational isolation of non–college educated individuals.

Using 2010 US Census data on the percentage of tract population aged 25 years or older with a college degree, we use a local, spatial measure of EI to quantify the geographic separation of non–college educated individuals from college educated individuals. This index is calculated in the same way as RI, and results in a weighted average proportion of non–college educated individuals in the local environment. The EI index ranges from 0 to 1: Individuals living in a neighborhood environment that is predominantly of college educated individuals will have an EI value close to 0. In contrast, individuals living in a neighborhood environment that is nearly all non–college educated individuals will have an EI value that is close to 1. EI was computed for all tracts within the study area with non-zero population in the neighborhood environment of the index tract. We hypothesized that higher EI communities may have lower PM_2.5 and O₃ concentrations, in part because EI may be higher in areas with less industry and lower population density (the correlation between EI and urbanicity is −0.46), in addition to census tracts in central parts of urban areas (32). The correlation between EI and RI in the United States is around 0.21 (32), and in North Carolina, it is 0.23.

The construction of the spatial measures of neighborhood-level RI (41) and EI (32) are described in detail elsewhere and summarized in Web Appendixes 1 and 2 (available at https://doi.org/10.1093/aje/kwac059), respectively.

Neighborhood deprivation.

Tract-level data from the 2010 Census were utilized to calculate a previously developed Neighborhood Deprivation Index (NDI) (42). The NDI was calculated using the first factor loadings from a principal components analysis of the following census variables: percentage of households in poverty, percentage of female-headed households with dependents, percentage with annual household income below $30,000, percentage of households on public assistance, percentage of male persons in management/professional occupation, percentage living in crowded housing, percentage unemployed, and percentage without a high-school education. Larger NDI values indicate more severe deprivation. NDI is typically used in urban settings; it is unclear how relevant it is for statewide analyses like those presented here. Nevertheless, because a measure of socioeconomic status is used so commonly in disparities analyses, we included NDI in our analysis.

Urbanicity.

Urbanicity was estimated based on the tract-level percentage of the population residing in urban settings that is available as a continuous variable (range: 0%–100%) in the 2010 Census data. An urban area comprises a densely settled core of census tracts and/or census blocks that meet minimum population density requirements, along with adjacent territory containing nonresidential urban land uses as well as territory with low population density included to link outlying densely settled territory with the densely settled core. Specifically, to qualify as urban, the territory identified must encompass at least 2,500 people, at least 1,500 of whom reside outside institutional group quarters. “Rural” encompasses all population, housing, and territory not included within an urban area (43).

Statistical analysis

To evaluate whether and how air pollution concentrations were changing over the study period, we plotted tract-level annual average PM_2.5 and annual 6-month (April–September) average O₃ concentrations over time and mapped PM_2.5 and O₃ concentrations at the beginning (2002), midpoint (2009), and end (2016) of the study period. We refer to 2002 levels as “baseline” conditions for the purposes of this study.

To estimate overall (i.e., statewide) and tract-specific baseline air pollutant concentrations as well as statewide and tract-specific temporal trends in air pollutant concentrations, we fitted a model that estimates separate but spatially correlated linear time trends for each census tract. This model is a modification of that proposed by Bernardinelli et al. (44) and takes the form:

(1)

where is the annual average PM_2.5 (or O₃) concentration in census tract i in year t; represents the statewide intercept; is a census tract–level random effect that allows for differences in baseline concentrations; represents the statewide temporal trend in air pollution over the study period; is a tract-level random effect that represents the tract-specific differential temporal trend; and . Thus, each census tract i has its own linear time trend, with a spatially varying intercept, , and a spatially varying slope, .

Each set of random effects and are modeled using independent conditional autoregressive prior distributions proposed by Leroux et al. (45), allowing for the possibility of spatial correlation between these spatially-varying parameters. Specifically, one of these random effect parameters is assumed a priori to be normally distributed, conditional on the parameters from all other tracts, and centered at the weighted average of neighboring values such that

where (i.e., a vector of random effects with the i^th entry removed), and is equal to if tracts i and j are neighbors (i.e., share a border or vertex) and 0 otherwise. The parameters are defined in a similar way, using and , and thus not shown. This flexible specification allows for several spatial patterns during modeling. For example, a small random effect variance, , indicates that there is less variability in that set of parameters with many values near zero, and that a single statewide intercept or slope may adequately describe variability in the air pollution concentrations. In the case of a larger value, , the spatial dependence parameter, indicates whether the variability is closer to independence ( near zero) or spatially correlated ( near 1). To ensure that we do not influence these choices before seeing the data, we specify weakly informative prior distributions for the model parameters to allow the data to drive the inference rather than our prior beliefs. Specifically:

To explore whether and how air pollution levels and temporal trends in pollutant concentrations are related to tract-level characteristics, we extended equation 1 to include a vector of variables for urbanicity, RI of NHB, EI of non–college educated individuals, and NDI that are specific to census tract i () such that

(2)

Equation 2 allows us to evaluate how baseline levels of air pollution relate to tract-level RI, EI, NDI, and urbanicity, and whether the variability in the intercepts can be explained by these tract-level covariates. The term for interaction between these covariates and time allows us to assess whether and how tract-level characteristics relate to temporal trends in air pollution, and whether the variability in the slopes can be explained by these covariates. To evaluate whether the covariates included in equation 2 have explained any variability, we can compare the and from equations 1 and 2 and more formally compare the model fits using the Watanabe-Akaike information criterion (WAIC). Smaller WAIC values indicate an improved balance of model fit and complexity among a group of competing models fit to the same data set (46). The same prior distributions from the previous model were used in this analysis, with the newly added regression parameters specified as

Models were fitted separately for PM_2.5 and O₃. For each pollutant, we fitted models with random slopes and random intercepts as well as models with random intercepts only. The WAIC was used to select the most appropriate model specification (i.e., with or without random slopes) (46). For each model, 50,000 posterior samples were collected from the joint posterior distribution after a burn-in period of 10,000 iterations using the car.linear function in CARBayesST (47) or the S.CARmultilevel function in CARBayes (48) for models with and without random slopes, respectively. We thinned the collected samples by a factor of 5 in order to reduce posterior autocorrelation, resulting in 10,000 posterior samples for inference. Visual inspection of individual parameter trace plots and calculation of the Geweke diagnostic (49) suggested no obvious signs of nonconvergence of the models. Statistical analyses were performed using R, version 3.5 (50).

RESULTS

Descriptive statistics

There were 2,163 census tracts in the study area. Average concentrations of PM_2.5 and O₃ in 2002, the start of the study period, were 13.5 μg/m³ and 56.6 ppb, respectively. To visualize how air pollution concentrations changed in North Carolina over the 15-year study period, we plotted tract-level annual average PM_2.5 (Figure 1A) and 6-month average O₃ concentrations (Figure 1B). Both PM_2.5 and O₃ levels decreased between 2002 and 2016, although the decline was not monotonic for either pollutant. For both PM_2.5 and O₃, there were no tracts in which concentrations increased between the beginning and end of the study period. Within-tract declines in concentrations, measured as percent change between the beginning and end of the study period, ranged from 20.8% to 45.1% for PM_2.5 and 19.4% to 32.7% for O₃.

Figure 1.

Downscaler-estimated air pollution concentrations, North Carolina, 2002–2016. A) Tract-specific annual average concentrations of particulate matter with an aerodynamic diameter ≤2.5 μm (PM_2.5), μg/m³; B) tract-specific annual average concentrations of ozone (O₃), ppb.

We also mapped tract-level PM_2.5 and O₃ concentrations at the beginning, midpoint, and end of the study period (Web Figures 1–2, respectively), as well as community-level covariates, namely RI, EI, NDI, and urbanicity (Web Figures 3–6, respectively). Descriptive statistics of community-level covariates (e.g., mean, standard deviation, range) are provided in Web Table 1. Correlations between each community-level characteristic ranged from −0.46 (urbanicity and EI) to 0.58 (EI and NDI) (Web Table 2).

Model selection

For both PM_2.5 and O₃, we fitted models (null and with adjustments) with random slopes and random intercepts as well as models with random intercepts only (WAIC values are provided in Web Table 3). Based on the WAIC results, we present adjusted PM_2.5 models with random intercepts and slopes, and adjusted O₃ models with random intercepts only. Variance inflation factors (VIFs) in the adjusted models, which included all 4 community characteristics (RI, EI, NDI, and urbanicity), ranged from a minimum of 1.45 (RI) to a maximum of 2.10 (EI) in the PM_2.5 and O₃ models. VIFs in excess of 5 are generally a cause for concern (51, 52), and some researchers have argued for use of a more conservative threshold of 2.5 (53).

Statistical analysis

Particulate matter with an aerodynamic diameter ≤2.5 μm.

We estimated tract-level baseline PM_2.5 concentrations and temporal trends in PM_2.5 concentrations after adjustment for RI, EI, NDI, and urbanicity and including a term for interaction between these characteristics and year, as described in equation 2. The tract-specific intercepts () from the adjusted model are shown in Figure 2A (these values also represent the fitted concentrations from 2009, the midpoint of the study period). The fitted, statewide average PM_2.5 concentration was 11.2 μg/m³ in 2009, and the statewide slope (α) was −6.37 μg/m³. The highest PM_2.5 concentrations (yellow) were observed in central and south-central North Carolina, particularly in and extending outward from Charlotte (Mecklenburg County), North Carolina. Lower PM_2.5 concentrations (dark blue) were observed in tracts in far western and eastern parts of the state, representing the Appalachian Mountains and Coastal Plain region, respectively.

Figure 2.

Particulate matter with an aerodynamic diameter ≤2.5 μm (PM_2.5) concentrations, North Carolina, 2002–2016. A) Tract-specific variations in average PM_2.5 concentration; B) tract-specific variations in change in PM_2.5 concentration.

Modeled temporal trends in PM_2.5, specifically the tract-specific slopes () in the adjusted model, are shown in Figure 2B. Tracts with smaller-than-average declines in PM_2.5 concentrations are shown in yellow, while tracts with larger-than-average declines in PM_2.5 are shown in in dark blue. Tracts with larger declines in PM_2.5 tend to be concentrated in central North Carolina, particularly western-central North Carolina. This coincides to some extent, but not completely, with the areas that had the highest baseline levels of PM_2.5.

Associations between the tract-level variables and PM_2.5 concentrations (baseline levels and change over time) are reported in Table 1. A 1-standard-deviation increase in RI (0.17) was associated with a 0.11 μg/m³ (95% credible interval (CrI): 0.06, 0.15) increase in PM_2.5 at baseline. A 1-standard-deviation increase in the percent of population in urban settings (30.1%) was associated with a 0.21 μg/m³ (95% CrI: 0.17, 0.26) increase in PM_2.5 concentration at baseline. In contrast, 1-standard-deviation increase in EI (0.15) was associated with a 0.10 μg/m³ (95% CrI: 0.04, 0.15) decrease in PM_2.5 concentration at baseline. This combination of results is sensible in that RI tends to be highest in urban areas, and EI tends to be higher in rural areas (e.g., Web Figures 3 and 4).

Table 1.

Associations Between Tract-Level Social and Demographic Variables and Baseline Levels and Time Trends in PM_2.5 Concentrations, North Carolina, 2002–2016^a

Variable b	Main Effect		Interaction With Time
	Posterior Median	95% CrI	Posterior Median	95% CrI
RI of non-Hispanic Black individuals	0.11	0.061, 0.15	−0.015	−0.064, 0.034
EI of non–college educated individuals	−0.10c	−0.15, −0.044	−0.067c	−0.13, −0.0085
Percentage of population in urban settings	0.21c	0.17, 0.26	−0.12c	−0.17, −0.074
NDI	−0.011	−0.039, 0.017	−0.0031	−0.029, 0.024

Abbreviations: CrI, credible interval; EI, educational isolation; NDI, Neighborhood Deprivation Index; PM_2.5, particulate matter with an aerodynamic diameter of ≤2.5 μm; RI, racial isolation.

^a Models adjusted for all community-level characteristics (i.e., RI, EI, urbanicity, and NDI); separate models were not fitted for each community-level characteristic. Values are shown to 2 significant digits.

^b Correlations between demographic variables ranged from −0.46 (between EI and urbanicity) to 0.58 (between EI and NDI); ranges of variable values and correlations between all variables are provided in Web Tables 1–2.

^c Coefficients for which the 95% CrI does not include zero.

Terms for interaction with year were statistically significant and negative for both urbanicity and EI, although the significance on EI was marginal. Interaction terms can be interpreted as location-specific slopes. For example, communities (census tracts) with higher urbanicity values have smaller (more negative) slopes than communities with lower urbanicity values. That is, the decline in PM_2.5 over time is steeper in more-urban areas than in less-urban areas. Similarly, communities with higher EI values have more-negative slopes than communities with lower EI values, such that the decline in PM_2.5 over time is steeper in more educationally isolated communities than in less educationally isolated communities.

For example, an “average” tract would have a baseline PM_2.5 concentration of 13.5 μg/m³ and a change in PM_2.5 over the study period of −6.4 μg/m³. A tract that is 2 standard deviations above the average in terms of urbanicity, holding other variables constant, would have a baseline PM_2.5 concentration of 13.9 μg/m³ and a change in PM_2.5 over the study period of −6.7 μg/m³. That is, for a tract with higher urbanicity, the baseline PM_2.5 concentration is higher, but the decline in PM_2.5 is also steeper. In contrast, a tract that is 2 standard deviations above the average in terms of RI, holding other variables constant, would have a baseline PM_2.5 concentration of 13.7 μg/m³ (higher than average) and a change in PM_2.5 over the study period of −6.4 μg/m³ (same as the average).

These findings indicate that there are existing disparities in PM_2.5 exposure with respect to RI, EI, and urbanicity, with more-urban and higher-RI tracts having higher baseline PM_2.5 exposures, and higher-EI tracts having lower baseline PM_2.5 exposures. Time trends in PM_2.5 concentrations differed by urbanicity and EI only. In contrast to more-urban communities, which have high PM_2.5 levels that are improving more markedly over time, high-RI communities have high PM_2.5 levels that do not show such an improvement.

Posterior median estimates of , reported in Web Table 4, declined from the null model (1.93) to the adjusted model (1.58), suggesting that some of the variability in baseline PM_2.5 values is explained by the covariates in the adjusting model. The posterior median estimate of also declined between the null (0.72) and adjusting (0.69) model.

Ozone.

As with PM_2.5, we estimated baseline O₃ concentrations and temporal trends in O₃ concentrations after adjustment for RI, EI, NDI, and urbanicity, and including a term for interaction between these characteristics and year (as described by equation 2). The tract-specific intercepts () from the adjusted model are shown in Figure 3 (these values also represent the fitted concentrations from 2009, the midpoint of the study period). Tract-specific slopes are not presented in Figure 3 because, for O₃, the model selected based on WAIC included only random intercepts; however, a map of tract-specific slopes from the O₃ model with random slopes and intercepts is provided in the Web Figure 7. The statewide average O₃ concentration was 48.1 ppb in 2009, and the statewide slope (α) was −13.5 ppb (over the entire study period). Similar to PM_2.5, the highest O₃ concentrations were observed in south-central North Carolina, in and extending outward from Charlotte (Mecklenburg County), North Carolina. The lowest O₃ concentrations were observed in the coastal plains, particularly the southeastern part of the state.

Figure 3.

Tract-specific variations in average ozone (O₃) concentrations, North Carolina, 2002–2016.

Associations between the tract-level variables and O₃ concentrations (baseline levels and change over time) are reported in Table 2. Urbanicity and NDI were associated with baseline O₃ concentrations. A 1-standard-deviation increase (30.1%) in the percent of population in urban settings was associated with a 0.38 ppb (95% CrI: 0.30, 0.47) increase in O₃ concentration. A 1-standard-deviation increase in NDI (2.1) was associated with a 0.061 ppb (95% CrI: 0.11, 0.017) decrease in O₃ concentration.

Table 2.

Associations Between Tract-Level Social and Demographic Variables and Baseline Levels and Time Trends in Ozone Concentrations, North Carolina, 2002–2016^a

Variable b	Main Effect		Interaction With Time
	Posterior Median	95% CrI	Posterior Median	95% CrI
RI of non-Hispanic Black individuals	−0.065	−0.15, 0.019	−0.63c	−0.78, −0.47
EI of non–college educated individuals	0.023	−0.077, 0.12	0.54c	0.35, 0.73
Percentage of population in urban settings	0.38c	0.30, 0.47	0.084	−0.079, 0.25
NDI	−0.061c	−0.11, −0.017	0.078	−0.0080, 0.17

Abbreviations: CrI, credible interval; EI, educational isolation; NDI, Neighborhood Deprivation Index; RI, racial isolation.

^a Models adjusted for all community-level characteristics (i.e., RI, EI, urbanicity, and NDI); separate models were not fitted for each community-level characteristic. Values are shown to 2 significant digits.

^b Correlations between demographic variables ranged from −0.46 (between EI and urbanicity) to 0.58 (between EI and NDI); ranges of variable values and correlations between all variables are provided in Web Tables 1–2.

^c Coefficients for which the 95% CrI does not include zero.

Additionally, terms for interaction with year were statistically significant for RI and EI but had differing signs, indicating that the change in O₃ concentration over time is modified by the degree of RI and EI. Specifically, communities (census tracts) with higher RI have smaller (more negative) slopes than locations with lower RI. This indicates that the decline in O₃ over time is steeper in more racially isolated communities than in less racially isolated communities. In contrast, communities with higher EI values have larger (less negative) slopes than communities with lower EI values, such that the decline in O₃ over time is shallower in more educationally isolated communities than in less educationally isolated communities.

As an example, for O₃, an “average” tract would have a baseline O₃ concentration of 56.6 ppb and a change in O₃ over the study period of −13.5 ppb. A tract that is 2 standard deviations above the average in terms of urbanicity, holding other variables constant, would have a baseline O₃ concentration of 57.4 ppb but the same change in O₃ over the study period (−13.5 ppb). A tract that is 2 standard deviations above the average in terms of EI, holding other variables constant, would have a baseline O₃ concentration of 56.6 ppb (same as the average) and a change in O₃ over the study period of −12.4 ppb (below-average improvement). These exposure differences and changes in exposure differences over time may seem small, but they apply to a large number of people, and differences in the rates of change over time suggest that there is potential for disparities to worsen, or develop.

These findings suggest that there are existing disparities in O₃ exposure with respect to urbanicity and NDI, with more-urban and lower-NDI tracts having higher baseline O₃ exposures. In contrast, O₃ concentrations at baseline did not differ by RI or EI, but trends over time did. Specifically, declines in O₃ concentrations over time were steeper in high-RI tracts compared with low-RI tracts, and shallower in high-EI tracts compared with low-EI tracts. Should these trends of differential rates of decline in O₃ concentration by community-level EI and RI continue (or accelerate), disparities in O₃ exposure by EI and RI may emerge.

Posterior median estimates of declined from the null model (5.99) to the adjusted model (5.37) (Web Table 4), indicating that some variability in baseline O₃ concentrations was explained by the covariates.

DISCUSSION

Although average levels of PM_2.5 and O₃ have declined over the past 2 decades in the United States (24), it is unclear where improvements in air quality are concentrated and what populations are benefiting most from improvements. Moreover, an overall improvement in air quality could, in fact, mask widening disparities based on geographic, social, or demographic factors. Here, we describe baseline levels and temporal trends in PM_2.5 and O₃ concentrations in the state of North Carolina during 2002–2016. We evaluated whether baseline concentrations and temporal trends related to tract-level community characteristics, including measures of racial and educational isolation, deprivation, and urbanicity.

Areas with higher baseline PM_2.5 concentrations have more “room for improvement” compared with areas that had lower baseline PM_2.5 concentrations, making it intuitive for higher-exposure areas to have more marked declines (improvement) in PM_2.5 compared with lower-exposure areas. In fact, we observed this with urban tracts, which had both higher baseline PM_2.5 values and more pronounced declines in PM_2.5 over time. That is, over the study period, disparities in PM_2.5 exposure were reduced in more-urban compared with less-urban areas. In contrast, a different relationship was observed for RI. More racially isolated tracts also had higher than average baseline PM_2.5 concentrations, but unlike urban tracts, high-RI tracts did not exhibit more marked declines in PM_2.5 over time. This finding suggests a disparity in who benefits from declines in PM_2.5 concentrations and associated improvements in air quality. In fact, it suggests that areas with higher RI may actually experience an increase in PM_2.5 exposure disparity over time, despite an overall decline in PM_2.5 levels. Some tracts, such as more educationally isolated tracts, had lower baseline PM_2.5 values, but also more pronounced declines in PM_2.5 over the study period, compared with less educationally isolated tracts. Thus, areas with low EI may also experience an increase in PM_2.5 exposure disparity over time (compared with areas with high EI).

Different patterns were observed for O₃ compared with PM_2.5. We did not observe that “high exposure” areas had more pronounced declines (improvement) in O₃ concentrations compared with “low exposure” areas. Specifically, more-urban and less-deprived areas had higher baseline O₃ levels but declines in O₃ were similar across urbanicity and deprivation values. RI and EI were not associated with baseline O₃ concentrations but did exhibit interactions with time. Specifically, high-RI tracts exhibited steeper declines in O₃ concentrations over the study period, while high-EI tracts had shallower declines in O₃ levels. That is, there was more improvement in O₃ concentrations in high-RI tracts compared with low-RI tracts over the study period, and less improvement in O₃ concentrations in high-EI tracts compared with low-EI tracts. This suggests that, over time, disparities in O₃ exposure may be widening for high-EI vs. low-EI areas, despite an overall time trend of improving O₃ concentrations.

We observed differences in patterning of PM_2.5 and O₃ exposure according to sociodemographic characteristics, which is interesting but not necessarily unexpected. Previous work has also observed differences in PM_2.5 and O₃ exposure patterns by racial isolation and urbanicity (16) but did not consider how exposures are changing over time. The differences in spatial and/or sociodemographic patterning of these pollutants may relate to the sources of these pollutants and/or their behavior and formation in the atmosphere. PM_2.5 is a primary pollutant that is often emitted as a by-product of combustion (e.g., in automobile engines, industrial processes), while O₃, a secondary pollutant, is formed in light-catalyzed reactions between precursor pollutants, namely nitrogen oxides and volatile organic compounds (54). Ozone concentrations are typically higher where there is a mixture of precursor pollutants in the atmosphere, which may have anthropogenic or biogenic sources (e.g., vegetation such as trees), some of which may be more common in suburban or rural locations as compared with urban areas (55). Moreover, there are documented patterns of siting hazardous waste sites, polluting industrial facilities, and other undesirable activities or contaminated land-use types disproportionately in communities with lower SES and a higher proportion of racial/ethnic minorities (56). Thus, type and density of emission sources of PM_2.5 and O₃ precursors, and resulting ambient concentrations, may differ depending on community-level characteristics such as RI, EI, NDI, and urbanicity, among others.

This study has several limitations. The models used here assume a linear trend in air pollution concentrations over time. This is an oversimplification of the temporal trend in pollutant concentrations, but there is a clear decline in concentration of both pollutants between the beginning and end of the study period (e.g., Figure 1). Air pollutant concentrations were obtained from the Fused Air Quality Surface Using Downscaling (“downscaler”) data archive and represent predictions from a statistical model as opposed to observed (monitored) concentrations. These predictions enabled us to conduct this study because: 1) air pollution concentrations were available across the study area, creating a continuous spatial surface for fitting spatial models; and 2) estimates of ambient air pollution concentrations were available across census tracts with differing levels of the sociodemographic covariates of interest. This second point is particularly salient since air pollutant monitors are more often located in urban areas (57). Some evidence also suggests that racial/ethnic minorities may reside closer to pollution sources and farther from monitoring locations (58). Thus, there may be greater uncertainty in downscaler-derived ambient concentration estimates for specific types of communities (e.g., less urban). In forthcoming research, we examine whether uncertainty characterizing the downscaler-estimated concentrations differed by community characteristics such as SES, urbanicity, and RI, among others. Briefly, we observe some disparities in uncertainty of downscaler-derived PM_2.5 and O₃ that relate to community characteristics such as SES and RI.

We selected to adjust for RI, EI, NDI, and urbanicity because these characteristics may be relevant to health outcomes, and there is evidence that some of them are correlated with pollution levels as well. RI and urbanicity are associated with air pollution levels, specifically PM_2.5 (16), as well as health outcomes (59). EI is a newly developed measure of isolation that may be associated with health and developmental outcomes (32). Associations between NDI and health have been observed, even after controlling for air pollution exposure (60). However, there are multiple approaches to assessing neighborhood conditions, including segregation and SES, and no one measure is perfect. Here, we chose to calculate EI comparing college educated with non–college educated groups because there are large and growing gaps in health outcomes between college and non–college educated individuals, and although there is considerable local variability, the majority of the US population has at least a high-school degree (61, 62). We chose to calculate RI with respect to NHB because of the existing literature that examines segregation of Black persons, elevated environmental exposures, and health outcomes (16, 63); a long, particular history of social, political, and economic forces conspiring to specifically segregate NHB from the majority White population (64); and, proportionally, NHB represent the largest racial/ethnic minority group in North Carolina (65). Potentially important avenues for future research include examining relationships between air pollution and: 1) EI calculated for those with vs. without a high-school degree (instead of college degree); 2) RI calculated for other racial/ethnic groups, such as Hispanic persons; and 3) alternative metrics or additional metrics of segregation, racial composition, and socioeconomic status.

Despite limitations, this study has important strengths. It is a statewide analysis of air pollution trends over a period of 15 years. The simulated air pollution data used to evaluate trends in PM_2.5 and O₃ concentrations over time provide excellent temporal and spatial coverage and have been used in previous studies of air pollution exposure and health (12, 66). This work provides information on, geographically, where PM_2.5 and O₃ levels are highest and lowest, and where PM_2.5 and O₃ levels are declining more vs. less rapidly. Moreover, this research identifies the sociodemographic populations in North Carolina with higher vs. lower ambient air pollution exposures, and identifies populations for which ambient air pollution exposures are declining more vs. less rapidly. North Carolina is home to Warren County, considered by many to be the birthplace of the US environmental justice movement in the 1980s (67, 68). It is located in the American South, which is unique among regions in the United States in that there are high-RI communities in both urban and rural areas; in much of the rest of the United States, NHB individuals reside predominantly in urban and suburban communities (32). Compared with the United States overall, North Carolina has a higher than average percentage of the population identifying as NHB (22.2% vs 13.4%) and also has a higher poverty rate (13.6% vs. 11.4%), although both of these variables exhibit considerable heterogeneity across the state (65). However, a key contribution of this work is the approach and framework it provides for examining whether disparities are widening or narrowing over time, and which types of communities have higher air pollution levels and/or more rapidly declining air pollution levels. Importantly, we show that despite overall PM_2.5 declines across North Carolina during the study period, areas with high EI and urbanicity exhibited greater declines, while areas with high RI showed lesser declines. These findings suggest that communities are not all benefiting equally from air quality improvements, and that disparities in exposure may be widening for specific populations.