Disparities in ambient nitrogen dioxide pollution in the United States

Significance

Historic systemic racism, segregation, and White-dominated political institutions in the US have created a public health crisis for communities of color, which includes a higher burden of air pollution. Although outdoor air quality hbas improved, addressing environmental injustices remains a priority. Combining high-resolution ground-level NO₂ conentrations, fine-scale sociodemographic information and source-specific emission, we find that racial/ethnic minorities are exposed to higher ambient NO₂ levels in the US, and disparities have persisted and worsened between 2000 and 2016. Although not the only source, traffic-related emissions are a major contributor. The primary contribution of this work is identifying the source-specific contribtions to NO₂ exposure disparities, enabling targeted actions to reduce environmental injustices among racia/ethinic groups

Abstract

Average ambient concentrations of nitrogen dioxide (NO₂), an important air pollutant, have declined in the United States since the enactment of the Clean Air Act. Despite evidence that NO₂ disproportionately affects racial/ethnic minority groups, it remains unclear what drives the exposure disparities and how they have changed over time. Here, we provide evidence by integrating high-resolution (1 km × 1 km) ground-level NO₂ estimates, sociodemographic information, and source-specific emission intensity and location for 217,740 block groups across the contiguous United States from 2000 to 2016. We show that racial/ethnic minorities are disproportionately exposed to higher levels of NO₂ pollution compared with Whites across the United States and within major metropolitan areas. These inequities persisted over time and have worsened in many cases, despite a significant decrease in the national average NO₂ concentration over the 17-y study period. Overall, traffic contributes the largest fraction of NO₂ disparity. Contributions of other emission sources to exposure disparities vary by location. Our analyses offer insights into policies aimed at reducing air pollution exposure disparities among races/ethnicities and locations.

Socially disadvantaged groups, such as minorities and people of low socioeconomic status, have been documented to be disproportionately exposed to several environmental health risks, and to have a higher prevalence of underlying disease (1–4). Exposure differentials in ambient air pollution is a key example in many locations and is a consequence of complex historical, social, cultural, and political factors (5). In the United States, despite years of effective efforts to regulate ambient air pollutant concentrations overall, there are still substantial disparities in exposure (6–11); eliminating these and other forms of environmental injustice is a high priority for the federal government, including the U.S. Environmental Protection Agency (EPA) (12).

Nitrogen dioxide (NO₂), derived from high-temperature combustion processes and subsequent atmospheric transformation is one of the several routinely monitored “criteria” air pollutants that have been regulated under the National Ambient Air Quality Standards of the Clean Air Act. The main source of anthropogenic NO₂ emissions is the burning of fossil fuels, with transport and power plant sources dominating (13). Airport, commercial marine vessels, residential combustion, and other point and nonpoint source emissions also contribute (14–16). Mounting evidence has linked ambient NO₂ exposure to a variety of adverse health end points, including cardiovascular diseases (17, 18), respiratory conditions (19, 20), neurological disorders (21, 22), and cancers (23, 24), and even at concentrations below current national standards (25–27).

Compared to other criteria air pollutants, such as fine particulate matter (PM_2.5), NO₂ has a shorter atmospheric lifetime and therefore has more spatial heterogeneity, which can lead to marked exposure disparities across subpopulations even within a small area. For example, Chamblissa et al. (6) found NO₂ ethnographic exposure disparities in the San Francisco Bay Area using mobile monitoring measurements, and similar national and city-specific findings have been reported in studies based on land-use regression (LUR) modeling (9, 28) or satellite-derived column concentration, both serving as proxies for surface exposure (8, 29). However, few existing studies are based on continuous fields of ground-level NO₂ exposure, and fewer incorporate emission information that can identify the relative contribution of different pollution sources, which is needed to explain the reasons behind the observed exposure differentials and to design effective policies capable of addressing any air pollution inequities.

Here, we leverage a high spatiotemporal resolution dataset of ambient NO₂ concentrations (at daily, 1 × 1 km² resolution) at ground level for 2000 to 2016 derived from an ensemble machine learning model, combined with fine-scale demographic information at the US Census block-group level (600 to 3,000 people per group) from 2010, to understand inequalities in NO₂ exposure by race and population density across the contiguous United States generally, and within major metropolitan statistical areas (MSAs). In addition to applying a unique, state-of-the-art exposure model, we further add to the literature by investigating the contribution of different emission sources and how these differentials have evolved over time. Our assessment is conducted at the census block-group level (as opposed to coarser county or census tract levels in previous studies), which is the smallest area unit possible for this type of analysis.

Results and Discussion

NO₂ Exposure Disparities Are Prevalent across the Contiguous United States.

We aggregated the high-resolution NO₂ exposure dataset onto census block groups by calculating the population-weighted mean NO₂ exposure and linking it with census-based demographic information. The 17-y national mean concentration of NO₂ across block groups was 22.05 ± 9.17 ppb (mean ± SD). Fig. 1A shows the spatial distribution of the 17-y mean NO₂ at block group level across the contiguous United States from 2000 to 2016. Hot spots with high NO₂ values appeared in major metropolitan areas, with radial patterns of high NO₂ along the interstate highway network, consistent with several previous studies (30, 31). The proportion of racial/ethnic minority populations (i.e., Blacks/African Americans, Hispanics, Asians, and others) within NO₂ exposure deciles, increased with increasing exposure (Fig. 1B and SI Appendix, Fig. S2). This result was in part explained by the relatively large percentage of minority populations residing in urban areas of major cities. To illustrate the NO₂ spatial distribution within metropolitan areas, we selected eight regionally representative MSAs (32), as shown in Fig. 1 C–J. The urban core of each MSA suffered the highest NO₂ concentrations, and the concentration decreased in the surrounding suburban territories. Within the MSA regions, NO₂ exposure was spatially significantly anticorrelated with the percentage of non-Hispanic Whites (Fig. 1 C–J), with a Pearson’s correlation coefficient between −0.28 and −0.47 (P < 0.01). These results illustrate that minority-dominant block groups suffered a higher NO₂ exposure compared with White-dominant block groups both nationally and within metropolitan areas.

Fig. 1.

The spatial distribution map of the average NO₂ concentrations and the proportion of non-Hispanic White population across the United States (A) and MSAs (C–J) over 17 y. The Pearson’s correlation coefficients between NO₂ levels and the proportion of non-Hispanic Whites: −0.28 (Denver), −0.40 (Seattle), −0.31 (Chicago), −0.40 (Boston), −0.41 (New York), −0.46 (Atlanta), −0.47 (Houston), and −0.42 (Los Angeles). The division of the total population exposure distribution within deciles (B) by race/ethnicity (non-Hispanic Whites, non-Hispanic Asians, non-Hispanic Blacks, Hispanics and other races, including American Indian, Alaska Native, Native Hawaiian, or Other Pacific Islander), with decile concentration boundaries shown on the y axis. The White, Asian, and Black refer to those identifying as non-Hispanic.

We further applied a cluster analysis for 217,740 block groups based on sociodemographic information, which identified five clusters with distinct sociodemographic characteristics (i.e., high-income non-Hispanic Whites, low-income non-Hispanic Whites, Blacks, Hispanics, and Asians; See Materials and Methods and SI Appendix, Table S1). We also stratified these block groups as high population or low population based on population density and their normalized difference vegetation index (NDVI) (SI Appendix, Tables S2 and S3). The boxplot in Fig. 2A shows the statistics of 17-y mean NO₂ exposure by sociodemographic groups across the United States (overall) and in select MSAs. The overall 17-y mean NO₂ concentrations were lower in the non-Hispanic White block groups—both high and low income—compared to block groups with higher proportions of minority populations. Low-income White block groups had lower exposure levels (mean ± SD: 17.11 ± 7.84 ppb) than high-income White block groups (mean ± SD: 21.89 ± 8.17 ppb), reflecting where these two income groups reside. Low-income White people more commonly live in rural areas with low population and traffic densities and high green cover and are less affected by traffic-related pollution compared to high-income Whites (Fig. 2A).

Fig. 2.

Variation in average NO₂ exposure distributions by racial/ethnic groups from 2000 to 2016 across the United States and MSAs (A). Variation in the percentage change in NO₂ levels over 17 y across the United States/MSAs (B). The distribution shown in the plots includes the median (central line), upper and lower quartiles (box boundaries), and upper and lower 1.5 × interquartile range whiskers.

Block groups with a high share of Asian people had the highest mean exposure level, which was 70% higher than that of the low-income White block groups and 40% higher than that of the high-income White block groups. This disparity was similar in both high- and low-population areas. Although the Asian population in general has higher educational and economic status compared to other minorities, these groups include a larger share of first-generation immigrants, which may lead to greater segregation (33), with Asians concentrated mainly in Asian neighborhoods in city centers and other populated areas (SI Appendix, Fig. S3). The corresponding values for Hispanic block groups were 56% (vs. low-income Whites) and 22% (vs. high-income Whites) higher, and that for Black block-groups was 43% and 12% higher. These disparities were wider in areas of high population than in areas of low population.

The distribution of NO₂ concentration varied greatly by location (Fig. 2A). In general, NO₂ concentrations were always higher in areas of high population density than in low-population areas. Denver, for example, which has long been the site of the development and production of a variety of energy sources, including oil, gas, and electricity, and is frequently affected by wildfires, showed the highest levels of NO₂ pollution. New York, Los Angeles, and Chicago, which have been among the top three MSAs in terms of population and traffic density in the United States (34), have also experienced high NO₂ concentrations for decades.

Within each of the MSAs, NO₂ exposure was also distributed unevenly among different racial/ethnic groups (Fig. 2A and SI Appendix, Fig. S4). For all eight MSAs, White block groups (low and high income) tended to have the lowest exposures, though differences between White groups and minority groups were minimal in some cities, including Denver, Seattle, and Chicago. In other MSAs, such as Atlanta and Houston, the exposures are lower in low-income Whites, who represent a larger proportion of the population compared with other MSAs and are predominantly located in areas further away from urban cores with low traffic volumes and population densities (Fig. 2A and SI Appendix, Fig. S3). This association is inconsistent with the common assumption that low-income residents are more likely to be exposed to higher concentration of air pollutants, suggesting that the reality in specific city is more complex and depends on the cities’ historical socioeconomic composition and evolution. Similar findings and conclusions have also been presented in several studies in European cities (35–37). Atlanta and Houston have larger Black and Hispanic populations than other examined MSAs. In the two MSAs, the NO₂ exposure disparities were evident as demonstrated by the 50 to 100% higher exposure rates among minority groups compared to the low-income Whites, and 15 to 25% higher compared to high-income Whites.

In contrast, in several MSAs with larger, Whiter, and richer populations, such as New York and Boston, exposure in the high-income White group was lower than that in the low-income White group. Additionally, regardless of income, White groups in both MSAs experienced less exposure to pollutants than minority groups. Taken together, these results demonstrate location-specific complexities that may result from historical policies and practices, including redlining, a discriminatory mortgage appraisal practice conducted by the Home Owners’ Loan Corporation, that led to high levels of racial/ethnic segregation (38, 39), and which may continue to exacerbate inequities in air pollution exposure to some extent today (28).

NO₂ Exposure Disparities Persist and Even Worsen over Time.

Between 2000 and 2016, the national population-weighted average NO₂ levels showed a significant downward trend, with an overall decrease of nearly 40% (SI Appendix, Fig. S5A), corroborating other studies (9). NO₂ reductions were evident in both high- and low-population areas (SI Appendix, Fig. S5 C and D) and in all regions of the country (SI Appendix, Fig. S5A). However, the Pacific Coast and Islands, Mid Atlantic, and the West regions had higher average NO₂ concentrations than those of the other regions. Within MSAs, dramatic absolute concentration reductions occurred, but the largest percentage declines were typically away from urban cores, in the surrounding suburb (SI Appendix, Fig. S6). The former was where there was high population and traffic density (SI Appendix, Figs. S3 and S6 A–I), and where most racial/ethnic minority groups resided (SI Appendix, Fig. S3 and Fig. 1 C–J), resulting in persistent NO2 exposure disparities throughout the study period (SI Appendix, Fig. S5 B–D).

Across the country, all groups experienced at least a 10% decrease in NO₂ concentrations (Fig. 2B). Seattle and Denver showed the smallest decrease compared to the other MSAs, while Los Angeles had the largest drop of more than 40%. Nationally, there was no significant difference in the reduction of NO₂ exposure between the White groups and the minority groups in general (Fig. 2B). In some MSAs (e.g., Atlanta, Houston, and Los Angeles), however, block groups with high proportions of White individuals had greater concentration reductions than block groups with high proportions of minorities, indicating that disparities have become wider in these areas during the study period.

Source Contributions of NO₂ Exposure Disparity.

To quantify the disparities in NO₂ exposure, we ran a statistical model to test the sensitivity of NO₂ exposure to the proportion of each racial/ethnic group across block groups (Materials and Methods). The sensitivity values, represented as ppb increase of NO₂ exposure per 1% increase in racial/ethnic proportion, can be used as a measure of exposure disparities (Fig. 3 A–D). These values were negative for non-Hispanic White population strata in all MSA analyses (Fig. 3A). The sensitivity values were positive for all racial/ethnic minority populations (Fig. 3 B–D). The highest positive sensitivity values were observed in the Asian population, consistent with our analyses of Asian-dominant clusters that suffered the highest NO₂ exposure nationally (Fig. 2A).

Fig. 3.

Source contributions to racial/ethnic disparity in NO₂ exposure across the United States and MSAs (A–H). Sensitivity value of NO₂ exposure to the proportion of each racial/ethnic group across block groups (A–D). Breakdown of contributions from different emission sources to the emission-associated disparities for different racial/ethnic groups (E–H).

Fig. 3 E–H and SI Appendix, Fig. S7 show the breakdown of contributions from different emission sources to the observed emission-associated disparities for different racial/ethnic groups. Ground transportation emissions contributed to 40 to 80% of the emission-associated disparities in most analyses across MSAs and racial/ethnic groups. This is because ground transportation contributes a major proportion (60%) of total NO₂ emissions in the United States (40) and minority communities are more likely to live in neighborhoods close to major roadways (41–43). Nonpoint sources, consisting mainly of residential combustions (e.g., gas ranges, wood stoves, gas/oil-fired furnaces), were the second-largest contributors, accounting for 15 to 20% of disparities among Asians living in high- and low-population areas and all minorities living in high-population areas. Population- and housing-dense cities like New York and Boston, in particular, seem more affected by nonpoint sources, likely due to combustion in buildings (homes and businesses) (44, 45). The contributions of other source sectors varied greatly across MSAs. For instance, substantial contributions of airport emissions were observed in Black communities of Atlanta and Los Angeles, as well a)s in the Asian communities of Chicago. The impacts of electricity-generating unit and oil-/gas-related emissions on residents in Denver were of most concern compared to the impacts on other MSAs, with a fractional contribution to the total disparity of about 10% for minority communities. This may be caused by the high levels of oil/gas activity and burning electricity in Colorado, particularly in the Denver metropolitan area (46, 47). In Chicago, the contribution of other point sources (including those other than electric-generating units, oil and gas stations, and airports) accounted for almost all the exposure disparities in the Black communities, which was probably related in part to the concentration of Black population near the Illinois and Indiana state lines where steel mills and manufacturing plants are located (48). Commercial marine vessel emissions were more likely to contribute to disparities in areas with ports, such as Seattle and Houston. Given that the contribution of different sources varies greatly across regions, a uniform control strategy nationwide is unlikely to be sufficient; local and regional emission characteristics must be identified, and specific regulations are needed to tackle disparities.

To further investigate the contributions of NO₂ emission sources to these observed racial/ethnic exposure disparities, we employed a source-specific NO₂ emission inventory from NEI 2016 (~10 km × 10 km). We incorporated a smoothed all-source emission intensity and proximity field into the aforementioned statistical model as a mediator (Materials and Methods). The smoothing is needed to account for atmospheric transport. Although this treatment of transport is highly simplified compared with computationally expensive 3-D chemical transport model simulations, our analyses indicate that proximity to emission sources generally explains 60 to 100% of the observed disparities (i.e., emission-associated disparities) (Fig. 3 A–D). Including other mediators, such as NDVI and population density in the model, explained additional 0-20% of the observed NO₂ exposure disparities in most locations (Fig. 3 A–D).

The positive correlation between minority proportion and NO₂ concentration does not imply causality, i.e., minority groups in urban areas emit more NO₂ than the White-dominant areas in suburbs. For example, the emissions of NO₂ (and its precursor NO) from traffic depend on vehicle type, fuel type, and vehicle age (49), which can influence the vehicle emission factor. We calculated the spatial distribution of vehicle emission factor, defined as the ratio between traffic-related NO₂ emission and traffic density (SI Appendix, Fig. S8). Within each MSA, the emission factor was generally higher in the suburbs than that in the urban center, indicating that vehicles used in suburban regions were not cleaner than those in the urban areas, in part because heavy-duty, aging, and diesel-powered vehicles were widely used for freight and commuting in suburbs (50). These results suggest that the high concentration in urban areas was mainly driven by the traffic density, rather than emission factors. In addition, our analyses indicate that residents in suburbs relied more on cars to commute (SI Appendix, Fig. S9) and may generally travel longer distances compared with residents in urban cores. The potential equity implication of these findings is that residents of suburban areas may carry emissions as they travel to and through urban areas, as has also been demonstrated in a study of poverty and vehicle ownership and exposure disparities in Europe (51).

Our study provides evidence of racial/ethnic NO₂ exposure disparities in the United States by leveraging an unmatched 17-y high-resolution dataset of NO₂ estimates derived from a spatiotemporal ensemble model. The results reveal that racial/ethnic minorities are disproportionately exposed to high levels of NO₂ pollution, and that disparities persist over time. Furthermore, our results suggest that in some locations, these disparities may have worsened despite a significant reduction in the national average NO₂ concentration from 2000 to 2016. This study attributes the observed NO₂ exposure disparities to emission sectors at a national scale. Our primary finding is that ground transportation-related emissions are the major contributor to NO₂ exposure disparities in most regions, though the contribution of different sources to exposure varies across locations.

Strengths and Limitations.

This study has many strengths and highlights. First, this study adds to the literature on inequities in air pollution exposure, particularly in NO₂ exposure. Most of the previous studies on ambient air pollution exposure disparities focus on PM_2.5 (7, 11), although NO₂ exposure inequities have received an increasing interest in recent years. NO_2, as an important air pollutant brought mainly by traffic, industrial emissions, and residential combustions in most cities, has a relatively short atmospheric lifetime compared with PM_2.5, therefore having a higher spatial heterogeneity and potentially higher disparity among different racial/ethnic communities. This study incorporates the strengths of previous studies and added a deeper and broader discussion of long-term exposure trends and source contributions using high-resolution NO₂ exposure data. The findings of the study will help bridge the gap in research on exposure inequities related to traffic-related air pollutants and provide insights for the future development of NO₂ concentration control strategies and targeted policies to address the exposure disparities.

Second, previous studies of NO₂ exposure disparities lacked discussion of the contribution of different emission sources, which can limit the interpretation of the causes of exposure disparities and the identification of policies to reduce them. The extent of exploration of recent environmental justice-related studies is still based on apparent differences in exposure levels. Even though some studies have incorporated emission-related information (8), these studies mostly discussed the contribution of emissions to NO₂ concentrations rather than quantifying source contributions to racial/ethnic exposure disparities. This study identifies the source contributions of racial/ethnic NO₂ exposure disparities, which can lay the scientific basis for effective policymaking and targeted interventions for different emission sources.

Third, our high-resolution pollutant concentration estimates allow us to capture exposure levels on a much smaller scale than most previous studies and has a large spatial coverage that provides more information on exposure in rural and suburban areas than analyses based on traditional monitoring site data. Some other studies filled the exposure assessment gap with satellite-derived data, such as data obtained from the TROPOspheric Monitoring Instrument (TROPOMI). However, the NO₂ column observations provided by TROPOMI are susceptible to varied vertical profiles that may disconnect column measurements from ground-level exposures, and thus bring potential measurement error when linked to sociodemographic and health information (8, 29). The coarse spatial resolution (5 km × 3.5 km) also limits its ability to capture heterogeneous NO₂ levels at finer scales, which may underestimate exposure disparities among racial/ethnic communities in high population density areas. In addition, since TROPOMI is the satellite instrument on board the Sentinel-5 Precursor satellite, which was launched in 2017 (52), it is impossible to collect earlier air pollution data and measure the disparities in long-term exposure. Moreover, compared to exposure assessment using LUR, such as the Center for Air, Climate and Energy Solutions (CACES) (9, 28), our complementary machine learning algorithms improved model performance by integrating neural network, random forest, and gradient boosting, with a variety of predictor variables.

In addition, we integrated the information of all the block groups in the country into group population by using k-means clustering rather than manually defining the cutoff point of classification, as done in some previous studies (8, 9, 53, 54). The unsupervised k-means clustering classification helps create a larger geographic unit of neighborhoods with relatively similar characteristics to achieve a good overall representation of the neighborhood-level information, and the design comprehensively considers a variety of demographic variables and controls for the bias caused by a single variable.

Finally, in contrast to previous studies conducted at the census tract level (8, 53, 54), we use block group as the spatial unit, which is more granular on the scale of census data, improving the resolution of spatial analysis and reducing the possibility of underestimation of exposure disparities at the national level (SI Appendix, Fig. S10). A smaller spatial scale allows for more homogeneous socioeconomic units when analyzing aggregated data. However, analyses at block group level still face potential errors and uncertainties, and the accuracy of the exposure disparity estimation may still be suboptimal; more comparison and evaluation of the accuracy of exposure disparity estimation at different spatial resolutions (coarser or finer) is needed in the future study.

We acknowledge that this study also has some limitations. First, although the NO₂ exposure prediction model we used has excellent prediction accuracy crossvalidated against ground observations, potential measurement error is still inevitable, in particular for regions with sparse observations. Since the concentration estimates obtained from the exposure model are based on the spatial resolution of 1 km × 1 km, while the area size of block groups varies, potential exposure assessment errors may occur during aggregation. The accuracy of the exposure model within cities also needs further evaluation, and the limited number of monitoring sites in cities poses a challenge to assess the accuracy of the modeled NO₂ fields at higher resolution. Although race/ethnicity, as the major exposure of this study, was based on US Census counts and is not prone to introducing measurement error, the potential nondifferential outcome misclassification from NO₂ measurement error may underestimate the relationship between race/ethnicity and NO₂ concentrations. In addition, we are unable to capture individual-level exposure to NO₂ pollution from indoor, commuting, and work environments, so the estimates will deviate somewhat from the true exposure. Furthermore, due to the accessibility and completeness of census sociodemographic data, we are unable to explore the exposure disparities based on finer-resolution demographic data. There is a need for sociodemographic information with high resolution and high integrity to help explore more accurate exposure inequalities in the future. Lastly, the source attribution of disparities relies on the proximity to emission sources. There is a need for high-resolution chemical transport modeling to fully address the effect of transport and chemical transformation.

Our findings offer some insight into policies and interventions aimed at reducing NO₂ disparities among racial-ethnic groups. Although the concentration of NO₂ has been regulated since the enactment of the Clean Air Act and has declined significantly, exposure inequities still require special attention and control by legislation. This situation should be taken seriously, particularly in some of the MSAs with the highest levels of racial/ethnic disparities, such as Atlanta, Los Angeles, and Houston, as well as in some areas with a high proportion of Asian residents. Second, ground transportation emissions have been identified as a major source of emissions and the largest contributor to NO₂ disparities. Efforts to control traffic-related pollution can be strengthened by strictly enforcing emission standards, implementing incentives for electric vehicles, reducing highway congestion, increasing vegetation density along roadsides, and planning for buffer zones between roads and residential areas. Since NO₂ is a short-lived pollutant, widening buffer zones outside neighborhoods and improving indoor filtration and ventilation systems may be potentially effective ways for minority communities to mitigate unequal exposure. In addition, based on our observations, the contributions of emission sources to exposure disparities are variable across regions. Therefore, it is essential to develop policies and interventions tailored to local emission characteristics for each region. Lastly, the root cause of racial/ethnic exposure disparities is the historical racism in employment, income, education, and other aspects of the society. Therefore, to better achieve environmental justice and alleviate racial/ethnic discrimination, efforts should be made on multiple fronts simultaneously, including reducing the racial/ethnic wealth gap, dismantling racial/ethnic residential segregation, and addressing enduring issues that result from historical factors.

Materials and Methods

Sociodemographic Information.

Our demographic information was derived from the Decennial Census and the American Community Survey (ACS) conducted by the US Census Bureau. We extracted 2010 census data and 5-y estimate ACS data on race, Hispanic or Latino origin (hereafter “ethnicity”), the proportion of female, proportion of people under age 5, proportion of people over age 65, population density, median household income, proportion of people travel by vehicle, and proportion of households burning fuel. To minimize the number of categorical variables, we focused only on four racial/ethnic groups in the study: non-Hispanic White, non-Hispanic Black (includes Black and African American), non-Hispanic Asian, and Hispanic (includes Hispanic and Latino origin). According to the hierarchy of census geographic entities (55), block groups were selected as the spatial unit of this study after comprehensive consideration of data integrity and accessibility, and the 2010 spatial boundary was regarded as the standard. Block groups are the next level below the census tract in the geographic hierarchy, which is the smallest geographic entity that the decennial census uses to tabulate sample data and control clock numbering. It is generally defined to contain between 600 and 3,000 people. In 2010, there were about 217,740 block groups in the contiguous United States, with the average area and population being 35.38 km² (SD 257.73) and 1,396 (SD 790.67) people, respectively (SI Appendix, Tables S1–S3). All demographical information used in the research is publicly available at https://data.census.gov.

Air Pollution Data.

We obtained daily NO₂ concentration estimates between 2000 and 2016 in the United States using a high-resolution spatiotemporal ensemble model, which integrated three separate machine learning algorithms, including neural networks, random forest, and gradient boosting, and multiple predictor variables (56). Briefly, the satellite retrievals, chemical transport model simulations, meteorological variables, land-use variables, and NO₂ monitoring measurements were fed into the ensemble-based models for calibration and generated daily mean ambient NO₂ concentrations (1-h maximum) at 1 km × 1 km spatial resolution across the United States. The ensemble-based machine learning approach yielded a significant crossvalidated R² of 0.79 and an average rms error of 7.15 ppb overall. We calculated annual averages for NO₂ concentrations for each calendar year at 1-km² grid cells. Gridded population density data at 30-s (~1 km) were obtained from NASA Socioeconomic Data and Applications Center for the years 2000, 2005, 2010, 2015, and 2020, and further linearly interpolated for each year from 2000 to 2016. Then, we calculated the population-weighted annual mean NO₂ concentrations for each block group each year and linked them to the sociodemographic information.

Land-Use Information.

The NDVI information during the study period was derived from Moderate Resolution Imaging Spectroradiometer (MODIS) Vegetation Indices Version 6 data created by the NASA (57), which provided global monthly averages at 1 km grid-level spatial resolution. MODIS’ NDVI complements National Oceanic and Atmospheric Administration’s Advanced Very High-Resolution Radiometer NDVI product and had good time series continuity. As with air pollution data aggregation, the NDVI estimates were aggregated from 1 km grid cells to the block group level with grid centroids located within the boundaries of the block groups. We further averaged the estimates for all months of each year and linked them to sociodemographic information and air pollution information.

Emission Inventory.

Sector-specific NO₂ emissions for the United States were derived from the 2016 National Emissions Inventory (NEI 2016) created by the U.S. EPA (48). The emission inventory contains NO₂ emission fluxes from a variety of point and nonpoint source sectors. To simplify the number of emission sectors and facilitate the identification of major emission impacts, we combined multiple similar emission source types to form a total of seven emission sectors: ground transportation (e.g., on-road vehicle, off-road equipment, and railroad), electric-generating units, airport, commercial marine vessel, oil and gas stations, and other point (point source not included in electric-generating units, oil and gas stations, and airport) and nonpoint sources (e.g., residential combustion). The emission from each sector was gridded at 0.1° × 0.1° (~10 km × 10 km) spatial resolution. We further processed the emission field by applying a 2-D Gaussian smoothing with a SD of 0.3° (~30 km) to account for the effect of atmospheric transport, roughly consistent with a transport distance over the atmospheric NO₂ lifetime of ~4 h (58). The smoothed field represents both vicinity and intensity of local NO₂ emission sources, and a high correlation (r = 0.60 to 0.77) was found with the annual mean NO₂ surface concentration derived from the machine learning model. We used this smoothed NO₂ emission field in statistical analyses for the apportionment of exposure disparities to different sources (Statistical Methods). We then assigned these grid-level emission data to all block groups whose centroids fell within the grid boundaries, aligning the spatial resolution with the exposure data and sociodemographic information.

Statistical Methods.

To classify the 217,740 block groups across the United States and identify the population characteristics of the different classes of block groups, we applied k-means clustering, which is a nonmodel-based approach that is not based on underlying statistical models but corresponds to a discrete optimization algorithm. The clustering and identification of population characteristics were mainly based on sociodemographic information, including the proportion of females, the proportion of non-Hispanic Whites, the proportion of Hispanics, the proportion of Asians, the proportion of Blacks, the proportion of people under age 5, the proportion of people over age 65, median household income, the proportion of people traveling by vehicle, and the proportion of households burning fuel. After attempting to cluster all block groups with 3 to 10 clusters, and comprehensively evaluate the total within sum of square and the characteristics of clusters, five was identified as the optimal number of clusters. The first and second cluster block groups both have a higher proportion of non-Hispanic Whites (83% for both groups), but the median household income of the first cluster is significantly higher than that of the second cluster ($73,147.76 vs. $44,392.87), and they are defined as the high-income White group and the low-income White group, respectively. The remaining three clusters correspond to a higher proportion of Hispanics (59%), a higher proportion of Asians (35%), and a higher proportion of Blacks (70%). K-mean clustering was also used for the identification of geographical characteristics of block groups. We clustered all block groups into two categories based on NDVI and population density, the one with higher NDVI and lower population density was defined as low population area, while the other cluster with low NDVI and high population density was defined as high population area. Using all block groups for clustering, rather than taking extreme percentiles or absolute thresholds, allows for a comprehensive consideration of all regions of the country and can effectively reduce the bias caused by extreme values of various variables.

We further stratified the block groups into eight MSA-level subsets: Atlanta-Sandy Springs-Alpharetta, GA; Boston-Cambridge-Newton, MA-NH; Chicago-Naperville-Elgin, IL-IN-WI; Denver-Aurora-Lakewood, CO; Houston-The Woodlands-Sugar Land, TX; Los Angeles-Long Beach-Anaheim, CA; New York-Newark-Jersey City, NY-NJ-PA; and Seattle-Tacoma-Bellevue, WA. All MSAs are selected based on geographic regions of the United States (32), and these representative MSAs are spatially evenly distributed across seven regions: the Pacific Coast and Islands, the West, Midwest, Mid-Atlantic, New England, the South, and Southwest. We refer to these MSAs by their colloquial abbreviations: Atlanta, Boston, Chicago, Denver, Houston, Los Angeles, New York, and Seattle, respectively, when discussing them.

To investigate the disparities in NO₂ exposure among population clusters and to take spatial autocorrelation into account, we applied conditional autoregressive (CAR) model to compare average NO₂ concentrations between racial/ethnic clusters. The CAR model is commonly used in small-areal estimation applications and captures spatial dependence based on a neighborhood adjacency matrix when analyzing data collected over irregular polygons. Model coefficients were used to measure the magnitude of the disparities with P values less than α = 0.05, indicating significant differences in NO₂ exposure levels among these racial/ethnic populations (Eqs. 1.1A and 1.1B).

$$Y_i={\mathrm{\beta }}_0+{\mathrm{\beta }}_1{\mathrm{Racial}/\mathrm{ethnic} \mathrm{cluster}}_i+{\mathrm{\beta }}_2{\mathit{Cov}}_i+{\epsilon }_i,$$ [1.1 A]

$${\epsilon }_i~N(0,{\tau }^2{(D_w-\rho W)}^{-1}),$$ [1.1 B]

where Y denotes average NO₂ exposure level of different block group i, parameter ${\tau }^2$ denotes the conditional variance, and $W$ denotes the symmetric spatial adjacency matrix, which takes value 1 if two block groups share a border and 0 otherwise.

To understand changes in NO₂ exposure levels in different populations, we calculated the difference in NO₂ concentrations between the average of the first 3 y and the average of the last 3 y of the study period for all block groups based on linear fits, and the same CAR model was used to assess disparities in exposure-level variations in diverse populations.

To understand the effect of emissions on NO₂ exposure disparities, we first regressed all emission sectors using Baron and Kenny’s test (59) to assess whether the emission sector under discussion acts as a mediator influencing the relationship between NO₂ exposure and the proportion of different racial/ethnic groups. After identification of mediating effects, we regressed emissions, NDVI, and population density jointly on NO₂ concentration to evaluate the extent to which each factor influenced exposure inequality in the CAR models (Eqs. 1.2A –1.2F).

$$Y_i={\mathrm{\beta }}_0+{\mathrm{\beta }}_{a1}{\mathrm{P}\mathrm{r}\mathrm{o}\mathrm{p}\mathrm{o}\mathrm{r}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n} \mathrm{o}\mathrm{f} \mathrm{r}\mathrm{a}\mathrm{c}\mathrm{i}\mathrm{a}\mathrm{l}/\mathrm{e}\mathrm{t}\mathrm{h}\mathrm{n}\mathrm{i}\mathrm{c} \mathrm{g}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}}_i+{\mathrm{\beta }}_2{\mathit{Cov}}_i+{\epsilon }_i,$$ [1.2 A]

$$\begin{array}{c}{Y}_i={\mathrm{\beta }}_0+{\mathrm{\beta }}_{b1}{\mathrm{P}\mathrm{r}\mathrm{o}\mathrm{p}\mathrm{o}\mathrm{r}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n} \mathrm{o}\mathrm{f} \mathrm{r}\mathrm{a}\mathrm{c}\mathrm{i}\mathrm{a}\mathrm{l}/\mathrm{e}\mathrm{t}\mathrm{h}\mathrm{n}\mathrm{i}\mathrm{c} \mathrm{g}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}}_i+{\mathrm{\beta }}_2{\mathit{Cov}}_i\hfill \\ +{\mathrm{\beta }}_3{\mathrm{E}\mathrm{m}\mathrm{i}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{s}}_i+{\epsilon }_i,\hfill \end{array}$$ [1.2 B]

$$\begin{array}{c}{Y}_i={\mathrm{\beta }}_0+{\mathrm{\beta }}_{c1}{\mathrm{P}\mathrm{r}\mathrm{o}\mathrm{p}\mathrm{o}\mathrm{r}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n} \mathrm{o}\mathrm{f} \mathrm{r}\mathrm{a}\mathrm{c}\mathrm{i}\mathrm{a}\mathrm{l}/\mathrm{e}\mathrm{t}\mathrm{h}\mathrm{n}\mathrm{i}\mathrm{c} \mathrm{g}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}}_i+{\mathrm{\beta }}_2{\mathit{Cov}}_i\hfill \\ +{\mathrm{\beta }}_3{\mathrm{E}\mathrm{m}\mathrm{i}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{s}}_i+{\mathrm{\beta }}_4{\mathrm{NDVI}}_i+{\mathrm{\beta }}_5{\mathrm{Population} \mathrm{density}}_i+{\epsilon }_i,\hfill \end{array}$$ [1.2 C]

$${\epsilon }_i^{~}N\left(0,{\tau }^2{\left(D_w-\rho W\right)}^{-1}\right),$$ [1.2 D]

$$\mathrm{\Delta }{\mathit{NO}}_2\mathit{ }\mathrm{attribute} \mathrm{to} \mathrm{emissions}/1\% \mathrm{Population} \mathrm{increase}={\mathrm{\beta }}_{a1}-{\mathrm{\beta }}_{b1},$$ [1.2 E]

$$\begin{array}{c}{\mathrm{\Delta NO}}_2\mathit{ }\mathrm{attribute} \mathrm{to} \mathrm{NDVI} \mathrm{and} \mathrm{population} \mathrm{density}/1\% \mathrm{Population} \hfill \\ \mathrm{increase}={\mathrm{\beta }}_{b1}-{\mathrm{\beta }}_{c1},\hfill \end{array}$$ [1.2 F]

where Y in the three models denotes average NO₂ exposure level of block group i in different racial/ethnic stratifies in each area/MSA, parameter ${\tau }^2$ denotes the conditional variance, and $W$ denotes the symmetric spatial adjacency matrix, which takes value 1 if two block groups share a border and 0 otherwise.

We further assessed the relationship between various emission sectors and the distribution of racial/ethnic populations through regression models, which in turn allowed us to understand the impact of unequal NO₂ exposure on four racial/ethnic populations in different areas of diverse pollution emission sectors (Eqs. 1.3A–1.3C).

$$Y_i={\mathrm{\beta }}_0+{\mathrm{\beta }}_{d1}{\mathrm{P}\mathrm{r}\mathrm{o}\mathrm{p}\mathrm{o}\mathrm{r}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n} \mathrm{o}\mathrm{f} \mathrm{r}\mathrm{a}\mathrm{c}\mathrm{i}\mathrm{a}\mathrm{l}/\mathrm{e}\mathrm{t}\mathrm{h}\mathrm{n}\mathrm{i}\mathrm{c} \mathrm{g}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}}_i+{\mathrm{\beta }}_2{\mathit{Cov}}_i+{\epsilon }_i,$$ [1.3 A]

$${\epsilon }_i^{~}N\left(0,{\tau }^2{\left(D_w-\rho W\right)}^{-1}\right),$$ [1.3 B]

$${\mathrm{NO}}_2\mathit{ }\mathrm{disparity}=\frac{{\mathrm{\beta }}_{d1}}{{\sum }_k{\mathrm{\beta }}_{d1}}\times 100\%,$$ [1.3 C]

where Y in Eqs. 1.3A–1.3C denotes emission levels of block group i in different emission sectors k for four racial/ethnic groups in each area/MSA, parameter ${\tau }^2$ denotes the conditional variance, and $W$ denotes the symmetric spatial adjacency matrix, which takes value 1 if two block groups share a border and 0 otherwise.

All analyses and data visualizations were performed using R (version 4.0.5).

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (CSV)

Data, Materials, and Software Availability

All other study data are included in the article and/or SI Appendix. A dataset including demographics, concentrations, and emission sources by block group has been deposited in Figshare (https://doi.org/10.6084/m9.figshare.22304899.v1) (60). High-resolution nitrogen dioxide estimates in the U.S. are publicly available at NASA Socioeconomic Data and Applications Center (https://beta.sedac.ciesin.columbia.edu/data/set/aqdh-no2-concentrations-contiguous-us-1-km-2000-2016/data-download) (61).