Disparities in Air Pollution Exposure in the United States by Race/Ethnicity and Income, 1990–2010

Background: Few studies have investigated air pollution exposure disparities by race/ethnicity and income across criteria air pollutants, locations, or time. Objective: The objective of this study was to quantify exposure disparities by race/ethnicity and income throughout the contiguous United States for six criteria air pollutants, during the period 1990 to 2010. Methods: We quantified exposure disparities among racial/ethnic groups (non-Hispanic White, non-Hispanic Black, Hispanic (any race), non-Hispanic Asian) and by income for multiple spatial units (contiguous United States, states, urban vs. rural areas) and years (1990, 2000, 2010) for carbon monoxide (CO), nitrogen dioxide (NO2), ozone (O3), particulate matter with aerodynamic diameter ≤2.5μm (PM2.5; excluding year-1990), particulate matter with aerodynamic diameter ≤10μm (PM10), and sulfur dioxide (SO2). We used census data for demographic information and a national empirical model for ambient air pollution levels. Results: For all years and pollutants, the racial/ethnic group with the highest national average exposure was a racial/ethnic minority group. In 2010, the disparity between the racial/ethnic group with the highest vs. lowest national-average exposure was largest for NO2 [54% (4.6 ppb)], smallest for O3 [3.6% (1.6 ppb)], and intermediate for the remaining pollutants (13%–19%). The disparities varied by U.S. state; for example, for PM2.5 in 2010, exposures were at least 5% higher than average in 63% of states for non-Hispanic Black populations; in 33% and 26% of states for Hispanic and for non-Hispanic Asian populations, respectively; and in no states for non-Hispanic White populations. Absolute exposure disparities were larger among racial/ethnic groups than among income categories (range among pollutants: between 1.1 and 21 times larger). Over the period studied, national absolute racial/ethnic exposure disparities declined by between 35% (0.66μg/m3; PM2.5) and 88% (0.35 ppm; CO); relative disparities declined to between 0.99× (PM2.5; i.e., nearly zero change) and 0.71× (CO; i.e., a ∼29% reduction). Discussion: As air pollution concentrations declined during the period 1990 to 2010, absolute (and to a lesser extent, relative) racial/ethnic exposure disparities also declined. However, in 2010, racial/ethnic exposure disparities remained across income levels, in urban and rural areas, and in all states, for multiple pollutants. https://doi.org/10.1289/EHP8584


Disparities in air pollution exposure in the United States by race-ethnicity and income, 1990 -2010 Introduction
Air pollution is associated with ~100,000 annual premature deaths in the United States (US) (Stanaway et al. 2018) and is linked to cardiovascular disease, respiratory disease, cancers, adverse birth outcomes, cognitive decline, and other health impacts (Cohen et  The objective of our research was to comprehensively and consistently investigate disparities in exposure to Environmental Protection Agency (EPA) criteria air pollutants for the two decades following the 1990 Clean Air Act Amendments in the US. Specifically, we investigated the following questions regarding disparities in exposure to six criteria air pollutants: (1) How do exposures vary by race-ethnicity and income? (2) How do racial-ethnic exposure disparities vary by pollutant? (3) How do racial-ethnic exposure disparities vary by location (state, urban vs. rural areas)? (4) How have racial-ethnic exposure disparities changed over time? To address these questions, we combined demographic data from the US Census (Manson et al. 2019) with predictions of outdoor average levels of six criteria air pollutants from a publicly-available national empirical model derived from satellite, measurement and other types of data (Kim et al. 2020) at the spatial scale of census block groups and census tracts. We then analyzed disparities in exposure to six criteria air pollutants (all criteria air pollutants except lead [Pb]; i.e., carbon monoxide [CO], nitrogen dioxide [NO 2 ], ozone [O 3 ], fine and coarse particulate matter [PM 2.5 , PM 10 ], and sulfur dioxide [SO 2 ]) by race-ethnicity (four racial-ethnic groups: white, Black, Hispanic, Asian) and income (16 household income categories) across time-points (decennial census years: 1990, 2000, and 2010) and spatial units (contiguous US, state, urban vs. rural areas).

Demographic and Air Pollution Datasets
We obtained demographic data (i.e., population estimates by race-ethnicity, household income, and household income disaggregated by race-ethnicity) and map boundaries (e.g., states, census tracts, and census block groups) for the contiguous US from the 1990, 2000, and 2010 decennial census from the IPUMS National Historic Geographic Information System (NHGIS) (Manson et al. 2019).
For analyses by income disaggregated by race-ethnicity in 2010, data from the 2010 NHGIS were available at the census tract level. For each census tract, NHGIS reports the number of census householders in each of the 16 census income categories, disaggregated in eight census racial and/or ethnic groups: (i) not Hispanic or Latino, White alone, (ii) Black or African American alone, (iii) American Indian and Alaska Native alone, (iv) Asian alone, (v) Native Hawaiian and Other Pacific Islander alone, (vi) some other race alone, (vii) two or more races, and (viii) Hispanic or Latino. To best match demographic variables used in race-ethnicity analysis at the census block group level, we reported results for four largest race-ethnicity groups (total covered in 2010: 113 million census householders, 98.5% of householders with data on income by race-ethnicity): not Hispanic or Latino, White alone (71% of householders; hereafter, "non-Hispanic White"), Hispanic or Latino (12%; hereafter, "Hispanic"), Black or African American alone (12%; hereafter, "Black"), and Asian alone (3.8%; hereafter, "Asian"). Thus, for the data used for the household income by race-ethnicity analysis (but not for other analyses), Black and Asian categories included both Hispanic and non-Hispanic individuals; for these analyses (but not others), Hispanic Black populations (~0.40% of the population) would be included in results for Hispanic and for Black populations, and Hispanic Asian populations (~0.08%) would be included in results for Hispanic and for Asian populations. Additionally, for the data used for the household income by race-ethnicity analysis (but not for other analyses), the Asian category does not also include Pacific Islander populations.
The US Census Bureau defined census blocks as "urban" or "rural", based on population density and other characteristics (Ratcliffe et al. 2016). We used 2010 census urban/rural block definitions to define a 2010 census block group for all three years (1990, 2000, and 2010) as rural if all blocks inside it were rural, and we defined the remaining block groups as urban.
Average estimates of ambient air pollution levels for US EPA criteria pollutants were obtained from Center for Air, Climate, and Energy Solutions (CACES) empirical models for the contiguous US (www.caces.us/data). These models incorporate satellite-derived estimates of air pollution, satellite-derived land cover data, land use data, EPA monitoring station data, and universal Kriging (Kim et al. 2020); estimated pollution levels were available by census block at block centroids based on 2010 census boundaries for years from 1990 to 2010 for all pollutants except PM 2.5 (for which monitoring data and exposure models were only available starting in 109  110  111  112  113  114  115  116  117  118  119  120  121  122  123  124  125  126  127  128  129  130  131  132  133  134  135  136  137  138  139  140  141  142  143  144  145  146  147  148  149  150  151  152  153  154 1999). Estimated levels of O 3 from the CACES empirical model are 5-month summer averages (specifically, the average during May through September of the daily maximum 8-hour moving average level); for remaining pollutants, estimated levels are annual averages.

Combining Demographic and Air Pollution Data
We matched the CACES empirical model results and the Census demographic data using the 2010 census spatial boundary definitions (from finest to coarsest spatial resolution: block, block group and tract boundaries) for the three census years (1990,2000,2010). We matched census block-level CACES model predictions for criteria air pollutants (blocks in 2010 in the contiguous US: n = ~7 million; average: ~44 residents per block) to census block group-level demographic data (block groups: n = ~22,000; ~1400 residents per block group) by calculating population-weighted mean air pollution levels for all census block centroids in that census block group using census block population data in year 2010. Similarly, to match census tract-level demographic data (tracts: n = ~74,000; ~4200 residents per tract), we calculated the populationweighted mean air pollution levels for all census block groups located within that tract.

Estimating Exposures to Pollutants
We estimated annual pollutant-specific exposures for 1990, 2000, and 2010 based on populationweighted mean predicted ambient air pollution levels for each demographic group (raceethnicity, income, and income by race-ethnicity; additional groups described in the Supplemental Material [SM]). This approach (average ambient air pollution level at residential census block group or tract) is broadly consistent with many examples in research and practice, including EPA monitors, the National Ambient Air Quality Standards, many influential epidemiological studies, and the published empirical models employed here. This article focuses on pollution level disparities rather than health outcomes. We used the finest publicly available census spatial boundary data to estimate exposures for each analysis (income by race-ethnicity: tracts; all other analyses: block groups) based on availability of census demographic data.
The national annual (for O 3 , 5-month average; for remaining pollutants, annual-average) exposure for demographic group i ( e i ) was calculated for a given pollutant and year as: where c j is the predicted average ambient pollution level for block group or tract j, p ij is the population of demographic group i in block group or tract j, and n is the number of block groups or tracts in the analyzed spatial level (contiguous US, 49 "states" (defined as the District of Columbia plus 48 states), urban vs. rural areas). Our primary exposure disparity metrics are based on absolute and relative differences in mean pollution levels. We selected metrics based on mean pollution levels for consistency with our focus on broad national average patterns in exposure disparities among multiple pollutants. Absolute disparity metrics are often connect to pollutant-specific health impacts (Harper et al. 2013). Relative disparity metrics (e.g., ratios, relative percent differences) are relevant for quantifying disproportionality in exposure burdens, in a way that can be compared or summarized among different pollutants. An important limitation of these metrics (based on differences in mean exposures) is that they do not include information about disparities across the full exposure distributions (Harper et al. 2013). To address this limitation, we conducted supplemental analyses using inequality metrics accounting for full exposure distributions (Gini Coefficient and between-group Atkinson Index), as described in the SM, as well as sensitivity analyses comparing metrics based on other specific points of the exposure distribution (i.e., comparing specific exposure percentiles) as described below.

National Exposure Disparities Analyses
We calculated the absolute and relative exposure disparity metrics using two different approaches nationally: (1) by race-ethnicity group and/or income category (i.e., the unit of analysis is a national subpopulation defined by race-ethnicity and/or income), and (2) by local demographic characteristics (i.e., the unit of analysis is a set of census block groups defined based on proportion of racial-ethnic minority residents).

National Exposure Disparity Metrics Based on Racial-Ethnic Group and/or Income Category
Our primary absolute disparity metric for quantifying national racial-ethnic exposure disparities is the pollutant-specific absolute difference in population-weighted average pollution level, as calculated using Equation (1) with block group level data, between the racial-ethnic group with the highest national mean exposure ("most-exposed group") and the racial-ethnic group with the lowest national mean exposure ("least-exposed group") among the four racialethnic groups (non-Hispanic White, non-Hispanic Black, non-Hispanic Asian, Hispanic); here, the unit of analysis is a racial-ethnic group. Our relative exposure disparity metric is the pollutant-specific exposure ratio, calculated as the ratio in population-weighted average pollution level between the most-and least-exposed racial-ethnic group. Both the absolute and relative exposure disparity metrics are constructed based on differences between most and least exposed racial-ethnic groups, to provide a measure of overall racial-ethnic disparities that avoids preselecting two specific groups for comparison and accounts for exposure disparities across multiple groups, in a consistent way for each pollutant (accounting for potential differences in the most-and least-exposed racial-ethnic groups by pollutant). We also report averages in relative disparities across pollutants as a representation of overall average inequalities in exposure to multiple pollutants; not as a representation of inequalities in health risks, which are pollutant-specific and depend on absolute levels of pollution exposure.
To quantify national income-based exposure disparities we calculated the pollutantspecific absolute difference in population-weighted average pollution level, using Equation (1) with block group level data, between the lowest (<$10,000) and the highest (>$200,000) household income categories (of the 16 census categories). Additionally, as a relative disparity metric, we calculated the relative percent difference in mean exposures between the lowest and highest income categories. As a supplementary analysis, we calculated similar absolute and relative exposure disparity metrics between the income category containing the 25 th percentile ($20,000-25,000) and the 75 th percentile ($75,000-100,000) of the income distribution. To quantify national exposure disparities by race-ethnicity and income, we first calculated the absolute difference in population-weighted average pollution level between the most-and least-exposed racial-ethnic group (among the four racial-ethnic groups, as above) within each of the 16 census income categories, and then averaged that income category-specific racial-ethnic exposure disparity across all 16 income categories, for each pollutant. In the analyses for both race-ethnicity and income, we used census data for householders to calculate exposures for the four racial-ethnic groups using Equation (1) with tract level data. Reflecting publicly available census data for racial-ethnic groups by income category, for this section only, the Black and Asian groups include Hispanic and non-Hispanic individuals, and the Asian group does not include Pacific Islander individuals. As a relative disparity metric, we divided the absolute exposure disparity metric by the national mean pollution level, for each of the pollutants.

National Exposure Disparity Metrics Based on Local Demographic Characteristics (i.e., Block Group Bins by Proportion of Racial-Ethnic Minority Residents)
We also investigated exposure disparities based on racial-ethnic minority resident percentages; here, the unit of analysis is bin of census block groups. Each block group bin was defined as single percentile (i.e., 1%) of all block groups stratified by the proportion of racialethnic minority residents. There were approximately 215,000 block groups in 2010, so each block group bin contained approximately 2,150 block groups. To investigate racial-ethnic disparities among block group bins, we rank ordered all census block group bins based on percent of a racial-ethnic minority residents (i.e., people self-reporting any race-ethnicity other than non-Hispanic White alone). For example, the first block group bin was the first percentile, and consisted of all block groups with between 0% and 0.67% racial-ethnic minority residents; the second block group bin was the second percentile, consisting of all block groups with 0.67% -0.97% racial-ethnic minority residents; the third block group bin consisted of all block groups with 0.97% -1.2% racial-ethnic minority residents, and so on through all 100 block group bins. The last block group bin consisted of all block groups with 99% -100% racial-ethnic minority residents. The annual exposure for demographic group i for the p th percentile census block group bin ( e ip ) was calculated for a given pollutant and year as: where c j is the average pollution level for block group j, p ij is the population of demographic group i in block group j, and n p is the number of block groups in the p th percentile block group bin. The absolute disparity is calculated as the exposure difference between block groups with the highest-versus lowest-deciles of proportion racial-ethnic minority residents, and, similarly, the relative disparity is calculated as the exposure ratio between block groups with the highestversus lowest-deciles of proportion racial-ethnic minority residents.

National Analysis of High-End Exposure Disparities in 2010
To quantify racial-ethnic disparities at the highest exposure levels, we analyzed the racial-ethnic composition of census block groups above the 90 th percentiles of the average pollution level among all census block groups. This was done seperately for each pollutant. First, for each of the 243  244  245  246  247  248  249  250  251  252  253  254  255  256  257  258  259  260  261  262  263  264  265  266  267  268  269  270  271  272   273   274  275  276  277  278  279  280  281  282  283  284 four largest racial-ethnic groups, we estimated the proportion of that group's national population that lived in a high exposure block group; here, our unit of analysis is a racial-ethnic group. This calculation reflects the proportion of a racial-ethnic group's total US population that lived in heavily polluted (above the 90 th percentile) block groups. We performed this calculation for each pollutant and each racial-ethnic group, using Equation (3).The second analysis, which was the converse of the first, investigated the racial-ethnic composition of block groups above the 90 th percentile for average pollution level. Here, our unit of analysis is all block groups above the 90 th percentile. This calculation reflects the demographics of only people that lived in heavily polluted block groups. We completed this calculation for each pollutant and each racial-ethnic group using Equation (4).
Here, a i is the percent of racial-ethnic group i living in a block group with concentration above the 90 th percentile for that pollutant, p ij is the population of group i in census block group j, p tota l nationali is the total population for group i in the United States, p tota l block group is the total population of census block groups above the 90 th percentile in the United States for that pollutant, b i is (when considering only the people counted towards P tota l block group ) the percent of people who are in racial-ethnic group i, and n is the number of census block groups above the 90 th percentile.
To explore multi-pollutant aspects of high-end exposure, we investigated the proportion of the total US population living in census block groups that were above the 90th percentile for exposure to 0, 1, 2, 3, and 4+ pollutants, for each racial-ethnic group in 2010, using Equation (3).

Sensitivity Analysis on Robustness of National Exposure Disparity Estimates
We conducted three sensitivity tests to investigate the robustness of conclusions based on estimated exposure disparities. First, as a sensitivity test for conclusions based on comparisons of mean values for exposures between groups, we calculated disparities using different metrics of the exposure distribution (10 th , 25 th , 50 th , 75 th , 90 th percentiles).
The remaining two sensitivity tests investigated whether conclusions here are robust to uncertainty in exposure model predictions. Specifically, in the second sensitivity test, we repeated the analysis of national mean exposures by racial-ethnic group, but for only the population living in a census block group with an EPA monitor in 2010. In this sensitivity test, we used the monitor observations directly as the exposure level, rather than modeling exposures. We then calculated Spearman rank-order correlation of relative disparities by pollutant (between the most-and least-exposed group) between base case and sensitivity test.
In the third sensitivity test, we compared the magnitude of uncertainties in the estimated racial-ethnic exposure disparities with the magnitude of the estimated racial-ethnic exposure disparities. To assess the potential impact of model error on racial-ethnic disparities, we first calculated population-weighted mean error (ME) for each racial-ethnic group, k, using Equation (5):   285  286  287  288  289  290  291  292  293  294   295   296   297  298  299  300  301  302  303  304  305  306  307  308  309  310  311  312  313  314  315  316  317  318  319  320  321  322  323 324 [5] where c im is the modeled average pollution level for block group i, c io is the measured average pollution level across all reporting EPA monitors within census block group i, p ik is the population of demographic group k in block group i, and n is the total number of block groups with EPA monitors. The ME of disparity between two racial-ethnic groups i and j induced by the model was calculated as the difference between populated-weighted ME for the two racial-ethnic groups i and j. Calculated uncertainties are based on comparison with EPA measured pollution level in 2010. We then calculated the ratio between the uncertainties in estimated racial-ethnic exposure disparities (calculated as the difference in population-weighted mean error between the most-and least-exposed racial-ethnic groups) and the estimated racial-ethnic disparities between the most-and least-exposed racial-ethnic groups.

Counterfactual Sensitivity Analysis of Exposure Disparities by Race-ethnicity and Income
To explore interactions between race-ethnicity and income in exposures and absolute exposure disparities, we performed two counterfactual sensitivity analyses for each pollutant in 2010. First, we calculated exposures and exposure disparities by race-ethnicity after controlling for income (i.e., a counterfactual in which each racial-ethnic group has the same income distribution as the national income distribution). To do this, we start with exposure disaggregated by race-ethnicity and income, but then apply the national income distribution (rather than the group's true income distribution) to calculate the (counterfactual) average exposure for each racial-ethnic group (i.e., we held the income distribution the same for each racial-ethnic group). We then calculated the (counterfactual) racial-ethnic exposure disparity between the most-and least-exposed racial-ethnic groups.
Second, we conducted the converse analysis: we calculated exposures and exposure disparities by income after controlling for race-ethnicity (i.e., a counterfactual in which each income category has the same racial-ethnic distribution as the national racial-ethnic distribution). To do this, we apply the national racial-ethnic distribution (rather than that income category's true racial-ethnic distribution) to calculate the (counterfactual) average exposure for each income category (i.e., we held the racial-ethnic distribution the same for each income category). We then calculated the (counterfactual) income-based disparity between the lowest-and highest-income categories.

Counterfactual Analysis of Migration
We investigated whether changes in racial-ethnic exposure disparities over time were mainly attributable to changes in air pollution levels ("air pollution") or changes in where people lived (abbreviated as "migration", but also including immigration and other shifts in demographic patterns) as a sensitivity analysis. To do so, we employed two counterfactual scenarios (Clark et al. 2017) during two decades (1990 to 2000; 2000 to 2010). For each scenario and year, we calculated exposures for the four largest racial-ethnic groups for the contiguous US population using Equation (1) based on census block group data. We then calculated the absolute racialethnic exposure disparity between the most-and least-exposed racial-ethnic groups (referred to in this section as "disparity") for all pollutants with available data (i.e., all except PM 2.5 in 1990).
To analyze 1990 to 2000, we calculated the change in disparity attributable to air pollution changed from 1990 to 2000 levels, with demographics remained constant at 1990 values (counterfactual scenario A), and used 1990 air pollution levels with demographic data changed 325   326  327  328  329  330  331  332  333  334  335  336  337  338  339  340  341  342  343  344  345  346  347  348  349  350  351  352  353  354  355  356  357  358  359  360  361  362  363  364  365  366  367  368  369 from 1990 to 2000 values (counterfactual scenario B). To estimate the separate contribution of changes in air pollution during 1990 to 2000, we divided the disparity-changes from counterfactual scenario A by the "true" calculated disparity-change between 1990 and 2000 (i.e., using 1990 air pollution levels with 1990 demographic data, and using 2000 air pollution levels with 2000 demographic data). Similarly, to estimate the separate contribution of migration during 1990 to 2000, we divided the disparity-changes from counterfactual scenario B by the "true" calculated disparity change between 1990 and 2000. Lastly, we used an analogous approach to analyze the next decade: 2000 to 2010.

Exposure Disparities Comparison Metrics for States
We investigated patterns in absolute exposure disparities among the 48 states of the contiguous US plus the District of Columbia (DC) (hereafter, "states" refers to 48 states and DC, a total of 49 geographic units in state-level related calculations) using two metrics for racial-ethnic exposure disparity. First, we used Equation (6) to calculate a state-specific population-weighted disparity for each of the four racial-ethnic groups and the state average relative to the pollutant's national mean. Second, for each state, we used Equation (7) to calculate a normalized population-weighted disparity between two groups: all racial-ethnic minority groups combined, and the non-Hispanic White group. This metric has the advantage of consistently comparing, for each state, exposures between racial-ethnic minority populations and the majority racial-ethnic group population (non-Hispanic White, 64% of the population). Lastly, we averaged both metrics across the six pollutants. For each state and pollutant, d 1 r is the normalized population-weighted disparity for racialethnic group r, e r is the average exposure for racial-ethnic group r in the state, e state is the average exposure for all people in the state, e national is the average exposure for all people in the contiguous United States, e minority is the average exposure for all racial-ethnic minority populations in the state, and e NH −¿ is the average exposure for the non-Hispanic White population in the state.

By Race-Ethnicity
To investigate national disparities in exposure to criteria air pollution by race-ethnicity, we first compared national population-weighted mean exposures by US census self-reported race-ethnicity in 2010, the most recent decennial census year with available data. We first present results for differences among subpopulations (unit of analysis: racial-ethnic group), then we present differences among locations, depending on the proportion of each racial-ethnic group residents in that location (unit of analysis: census block groups binned by proportion of racialethnic minority residents).
Sensitivity test on robustness of conclusions based on mean values showed that, for all pollutants, the rank-order (i.e., most-to least-exposed racial-ethnic group, among the four racialethnic groups) was consistent throughout the exposure distributions (Fig. 1). The remaining two sensitivity tests investigated whether conclusions here are robust to uncertainty in exposure model predictions. Results reveal that the conclusions are robust to exposure model uncertainty.
Results for analyzing only the population living in a census block group with an EPA monitor in 2010 were essentially the same as results using exposure model predictions: the non-Hispanic White group was the least-exposed group on average for most pollutants (CO, NO 2 , PM 2.5 , PM 10 , and O 3 ), and the relative disparities by pollutant (between the most-and least-exposed group on average) were highly-correlated (Spearman rank-order correlation between base case and sensitivity test: 0.89). The ratio between the uncertainties in estimated racial-ethnic exposure disparities and the estimated racial-ethnic disparities between the most-and least-exposed racialethnic groups were small: on average across the six pollutants, 0.0073 (if using absolute values of the ratio, 0.083). The largest absolute ratio was -0.17 [O 3 ] That result indicated that the uncertainty in the exposure model predictions was always small compared to the predicted racial-ethnic exposure disparities.
We also performed an analysis to determine whether average air pollution levels varied based on the racial-ethnic composition of a given census block group. For CO, NO 2 , PM 2.5 , and PM 10 , average pollution levels were higher in census block groups with higher proportions of racial-ethnic minority residents (Fig. 2). For O 3 , estimated average levels were approximately equal across census block group bins, regardless of census block group racial-ethnic characteristics (Fig. 2). For SO 2 , estimated average levels were generally higher in census block group bins with the highest and lowest proportions of racial-ethnic minority residents (i.e., higher in more racially segregated census block groups) (Fig. 2). This approach also reveals that the disparities were much larger for NO 2 than for other pollutants. The disparity in average air pollution levels between block groups with the highest-versus lowest-deciles of proportion racial-ethnic minority residents (block groups with >88% vs. <4% racial-ethnic minority residents) was larger for NO 2 (absolute disparity: 9.4 ppb, relative disparity [ratio]: 3.1) than for other pollutants (relative disparity [ratio] range: 0.8 -1.4, median: 1.1) Lastly, we investigated racial-ethnic disparities in exposure to the highest air pollution levels. First, for each racial-ethnic group we calculated the proportion of people nationally who lived in a block group with air pollution levels above the 90 th percentile for each pollutant. Averaged across all pollutants, the proportion of people nationally who lived in those highestexposure block groups was: 9.6% for the overall population, 17% for the Hispanic population, 15% for the non-Hispanic Asian population, 12% for the non-Hispanic Black population, and 7.2% for the non-Hispanic White population. Next, we calculated the racial-ethnic composition   413  414  415  416  417  418  419  420  421  422  423  424  425  426  427  428  429  430  431  432  433  434  435  436  437  438  439  440  441  442  443  444  445  446  447  448  449  450  451  452  453  454  455  456  457  458 of the block groups with air pollution levels above the 90 th percentile for each pollutant. Racialethnic minority populations were more likely than non-Hispanic White populations to live in a census block group with air pollution levels above the 90 th percentile for all pollutants (range: 1.1× to 4.1×, median: 2.1×) except SO 2 (0.88×). Racial-ethnic minority populations were also disproportionately likely to live in a census block group having multiple pollutants with levels above the 90 th percentile. For example, the proportion of population living in a census block group with levels above the 90 th percentile for four or more criteria pollutants was 5.2% for the Hispanic population (3.6× the national population average proportion), 2.2% for the non-Hispanic Asian population (1.5× the average), 1.9% for the non-Hispanic Black population (1.3× the average), and 0.36% for the non-Hispanic White population (0.25× the average) (for comparison: 1.4% for the overall US population).

By Income
To investigate national exposure disparities by income, we first compared national mean exposures to criteria air pollution by census income category in 2010. For all pollutants except O 3 , national mean exposures were higher for lower-income than for higher-income households (Fig. 3). Comparing national mean exposures for lowest-income (<$10,000; 7.2% of the households with income data) versus for highest-income (>$200,000; 4.2%) households, exposures for lowest income households were 16% (relative to national mean exposure) higher for SO 2 , 6.6% higher for PM 2.5 , and 5.2% higher for PM 10 . For NO 2 , CO, and O 3 , exposures for lowest-and highest-income households were similar (~2%). Based on differences in average exposures between the approximate 25 th and 75 th percentiles for income ($20,000-25,000 [midpoint: $22,500], $75,000-100,000 [midpoint: $87,500]), a $10,000 increase in income was associated with an average reduction in concentration (expressed as a percent of the national mean concentration) of 0.90% for SO 2 , 0.41% for PM 2.5 , 0.36% for NO 2 , and 0.22% for PM 10 and CO, and an increase of 0.16% for O 3 .

By Both Race-Ethnicity and Income
In this section, we present exposure disparities accounting for both race-ethnicity and income together for census householders (hereafter, "households"). For all six pollutants, the absolute exposure disparity between the most-and least-exposed racial-ethnic groups was larger (on average, ~6× larger; 1.1× for SO 2 , 21× for NO 2 , 1.4×-6.8× for the remaining pollutants) than the absolute exposure disparity between the lowest-and highest-income categories in 2010 (relative disparity: on average, ~ 1.2× larger). For all income levels and pollutants, the mostexposed racial-ethnic group was a racial-ethnic minority group (Fig. 3). For five of the six pollutants (not SO 2 ; Fig. 3), average exposures were higher on average for Black households at the approximate 75 th percentile for income (income category midpoint: $87,500) than for non-Hispanic White households at the approximate 25 th percentile for income (midpoint: $22,500). Racial-ethnic exposure disparities tended to be comparatively smaller at higher incomes than at lower incomes (except for O 3 ), but the size of that effect was modest. For example, the absolute exposure disparity between the most-and least-exposed racial-ethnic groups (Fig. 3) was, on average, 9.5% lower for households at the approximate 75 th percentile than at the approximate 25 th percentile of income.
Income distributions varied by racial-ethnic group. For example, non-Hispanic White households represented 61% of the lowest income category (<$10,000) and 85% of the highest income category (>$200,000), versus 23% and 3.5%, respectively, for non-Hispanic Black   459  460  461  462  463  464  465  466  467  468  469  470  471  472  473  474  475  476  477  478  479  480  481  482  483  484  485  486  487  488  489  490  491  492  493  494  495  496  497  498  499  500  501  502  503  504 households, 13% and 4.3% for Hispanic households, and 3.5% and 6.9% for non-Hispanic Asian households. To quantify racial-ethnic exposure disparities after accounting for racial-ethnic income distribution variation, we calculated the absolute exposure disparity between the mostand least-exposed racial-ethnic groups within each income category in 2010 and then averaged across all 16 income categories. The resulting national absolute exposure disparity between most-and least-exposed racial-ethnic groups averaged across income categories and normalized to national mean exposure (i.e., expressed as a percent of the national mean concentration) was 58% for NO 2 , 4.5% for O 3 , 12% to 17% for the remaining pollutants. Conversely, to quantify income exposure disparities after accounting for race-ethnicity, we calculated the absolute income disparity within each racial-ethnic group and averaged across the four racial-ethnic groups. The resulting national absolute exposure disparity between lowest and highest income categories normalized to national mean exposure was 15% for SO 2 , -2.9% for O 3 , and 2.7% to 6.3% for the remaining pollutants.
We also conducted two counterfactual analyses to explore the interactions between raceethnicity and income in explaining national exposure disparities. Both counterfactuals slightly shifted the calculated exposure disparities, but support the conclusion that racial-ethnic exposure disparities were distinct from, and larger than, exposure disparities by income. The first counterfactual (holding income constant for racial-ethnic groups) shifted the absolute disparity between most-and least-exposed racial-ethnic group normalized to pollutant's national mean exposure by 0.03 percentage units on average (average of absolute values: 1.0 percentage units; range among pollutants: -1.9 to +2.4 percentage units). The second counterfactual (holding raceethnicity constant for income categories) shifted the absolute disparity between highest and lowest income categories normalized to pollutant's national mean exposure by -2.2 percentage units (average of absolute values: 2.2 percentage units; range among pollutants: -7.1 to -0.03 percentage units). In conclusion, exposure disparities by race-ethnicity did not substantially shift after accounting for income differences.

By State
We explored how exposures varied by state, pollutant, and racial-ethnic group in 2010 (Fig. 4). The analysis separately considers the District of Columbia (DC) plus the 48 states of the contiguous US (hereafter, "states" refers to 48 states and DC, a total of 49 geographic units in state-level related calculations). There are 294 pollutant-state combinations (6 pollutants × 49 units and 1176 pollutant-state-groups (294 pollutant-states × 4 racial-ethnic groups). For this section, we define ±5% (all percentages used in this section were expressed as a percent of the national mean exposure in 2010) as "similar to", and therefore report examples where exposures differ from the average by >5% (or, in a sensitivity test, >20%). For example, ">5% lower-thanaverage" means the exposure is lower-than-average by an amount greater than 5%.
Overall, several spatial patterns emerge across states. First, racial-ethnic exposure disparities were ubiquitous among US states. In all 48 states and DC, one or more racial-ethnic groups experienced exposures >5% of the area average exposure in 2010. Second, racial-ethnic minority populations within states were much more likely to have been more-exposed versus less-exposed than average; in contrast, non-Hispanic White populations within states experienced exposures >5% above average in no states. Third, having exposures >5% lowerthan-average within a state was much more likely to happen for non- Hispanic White populations   505  506  507  508  509  510  511  512  513  514  515  516  517  518  519  520  521  522  523  524  525  526  527  528  529  530  531  532  533  534  535  536  537  538  539  540  541  542  543  544  545  546  547  548  549  550 than for racial-ethnic minority populations (Fig. 4, right column). Fourth, racial-ethnic exposure disparities were, on average, largest for NO 2 and smallest for O 3 among states.
Those findings reflect underlying trends across states, pollutants, and racial-ethnic groups. For example, for the non-Hispanic White group, 87% of the 294 pollutant-states had exposures that were similar (5%) to the average, 13% had exposures >5% less than average, and none were >5% greater than average. In contrast, for exposures for the three racial-ethnic minority groups, 42% (of 882 pollutant-state-groups) were >5% greater than average, 55% were ±5% of the average, and only 4% were >5% less than average. Thus, within individual states, the non-Hispanic White group was exposed to pollution levels that were similar to or cleaner than average, whereas the three racial-ethnic minority groups were more likely to be exposed to dirtier rather than cleaner pollution levels. For example, averaged across pollutants, the proportion of the states for which exposures were >5% greater than average is 73% for non-Hispanic Black populations, 57% for Hispanic populations, 35% for non-Hispanic Asian populations, and zero for non-Hispanic White populations.
The three racial-ethnic minority groups were disproportionately likely to be the mostexposed group, and disproportionately unlikely to be the least-exposed group of the four racialethnic groups across states. For example, the most-exposed group (for all cases, not just cases >5% greater than average) was the non-Hispanic Black group for 45% of the 294 pollutant-areas, the Hispanic group for 29%, the non-Hispanic Asian group for 18%, and non-Hispanic White group for 7.5%. In contrast, the least-exposed group was rarely a racial-ethnic minority group (~8% of all 294 pollutant-states for the non-Hispanic Black and for Hispanic group, 15% for the non-Hispanic Asian group) and was usually (70% of 294 pollutant-states) the non-Hispanic White group.
As a sensitivity test, we changed the analysis threshold to exposures >20% greater than average (rather than 5%). Here, we again found that the air pollution disproportionately impacted racial-ethnic minority groups. For example, exposure disparities >20% of national mean exposure for one or more pollutant-groups occurred for 67% of states (Fig. 4, left four columns for six pollutants), further emphasizing that disparities were widespread across states in 2010. Fig. 4 reveals differences among states. For example, the four most populous states (California, Florida, New York, Texas), all have large, racially/ethnically diverse urban areas. However, average disparities between racial-ethnic minority populations and non-Hispanic White populations (Fig. 4 bottom right) were notably larger (on average, 6× larger) for California and New York than for Florida and Texas. Some small, relatively rural states also had substantial exposure disparities. Examples include NO 2 in Nebraska (19%) and PM 2.5 in Nebraska (8.1%).

By Urbanicity
We investigated racial-ethnic and income-based exposure disparities in 2010 separately for block groups that were defined as urban (89% of the population) versus rural (11% of the population).
The most-and least-exposed of the four racial-ethnic groups differed between urban and rural areas for SO 2 and for O 3 . For SO 2 , the most-exposed racial-ethnic group was the non-Hispanic Black group in urban areas and the non-Hispanic White group in rural areas. For O 3 , the most-exposed racial-ethnic group was the non-Hispanic Asian group in urban areas and non-Hispanic White group in rural areas. For the remaining four pollutants, the most-exposed group was a racial-ethnic minority group in both urban and rural areas.
Exposure disparities by income category were also larger in urban than in rural areas. Absolute exposure disparities between lowest and highest income category were 1.1× [PM 2.5 ] to 25× [O 3 ] (median: 3.5×) greater (for relative disparity [ratio], range: 0.98× to 1.1×, median: 1.0×) in urban than in rural areas. Of the 12 pollutant-urbanicity categories (6 pollutants × 2 urbanicities), exposures were higher for the lowest-income category than for the highest-income category in all cases except for O 3 in urban areas and for NO 2 in rural areas.

Changes in National Exposures and Exposure Disparities from 1990 -2010
Criteria air pollution levels have declined in the US in the decades following the 1990 Clean Air Act amendments (US EPA 2020). To investigate if these reductions have led to reductions in racial-ethnic exposure disparities, we compared average exposures by racial-ethnic group from 1990 to 2010, for five of the pollutants. Exposure model results for PM 2.5 were only available from 2000 to 2010, so those results are presented separately.
National mean pollution levels of all six pollutants fell over the study period. For example, from 1990 to 2010, the national mean exposures decreased for all five pollutants by an average of 40% relative to national mean exposures in 1990 (range: -6% [O 3 ] to -71% [SO 2 ]; -34% to -55% for remaining three pollutants). PM 2.5 exposures decreased 29% from 2000 to 2010.
The average exposure disparities also declined from 1990 -2010. The amount of change depends in part on whether one considers absolute or relative disparities. In terms of absolute disparities, the disparities between the most-and least-exposed racial-ethnic groups decreased on average by 69% relative to absolute disparity in 1990 across the five pollutants. The largest change was an 88% decrease for CO disparities (0.40 ppm in 1990, 0.044 ppm in 2010, a 0.35 ppm [i.e., 88%] change) and the smallest change was a 54% decrease for NO 2  Shifting the comparison to the proportion of each racial-ethnic group in a location (unit of analysis: census block group) yields similar conclusions for changes in disparities during 1990 to 2010. For example, absolute air pollution level disparities between census block group bins with the highest versus lowest deciles of proportions of racial-ethnic minority residents (90 th -100 th versus 1 st -10 th percentiles in Fig. 2) decreased for four pollutants, on average by 60% relative to absolute disparity between the highest versus lowest deciles bins in 1990 (CO, NO 2 , PM 10 , SO 2 ). For O 3 , disparities decreased slightly, from near-zero in 1990 to 0.46 ppb (which is 1.0% of the national mean exposure) in 2010. PM 2.5 disparities decreased 58% from 2000 to 2010 .   596  597  598  599  600  601  602  603  604  605  606  607  608  609  610  611  612  613  614  615  616  617  618  619  620  621  622  623  624  625  626  627  628  629  630  631  632  633  634  635  636  637  638 639 640 In addition to national changes, we investigated changes in absolute racial-ethnic exposure disparities from 1990 to 2010 by state and by urban versus rural areas. Most states (>75%) experienced a reduction in racial-ethnic exposure disparities for pollutants except for PM 10 (and, except for PM 2.5 during 2000-2010). Urban areas experienced larger reductions in racial-ethnic exposure disparities than did rural areas for NO 2 and PM 10 (13 larger reductions in urban areas, for both pollutants), CO (2.4), and SO 2 (1.2). Conversely, PM 2.5 (during 2000-2010) and O 3 (during 1990-2010) had larger reductions in absolute racial-ethnic disparities for rural than for urban (2.4 and 3.4 larger in rural areas, respectively).
Finally, we investigated whether the changes in absolute racial-ethnic exposure disparities from 1990 to 2010 were more attributable to changes in air pollution levels or to changes in demographic patterns (migration, immigration, and other factors). Based on a counterfactual analysis, reductions in racial-ethnic exposure disparities between the most-and least-exposed racial-ethnic groups were mainly attributable to changes in air pollution levels rather than to changes in demographic patterns. On average across all pollutants, 87% of the reduction in the absolute racial-ethnic disparity metric was attributable to changes in air pollution levels from 1990 to 2000 (excluding PM 2.5 based on lack of available data), and 97% from 2000 to 2010 (Table S37).

Discussion
Our research provides the first national investigation of air pollution exposure disparities by income and race-ethnicity for all criteria pollutants (except lead). Our results reveal trends by pollutant and across time and space.
In 2010, on average nationally, racial-ethnic minority populations were exposed to higher average levels of transportation-related air pollution (CO, NO 2 ) and particulate matter (PM 2.5 , PM 10 ) than non-Hispanic White populations. This finding, which holds even after accounting for uncertainties in the predictions from exposure models, is consistent with prior national studies of NO 2 , PM 2.5 , and PM 10  Our findings on "which group was most-exposed?" (on average, nationally) varied by pollutant, but in all six cases the most exposed group was a racial-ethnic minority group. That result is consistent with prior national studies, which have reported, for example, highest average NO 2 exposures for Hispanic Black and non-Hispanic Asian populations (Clark et al. 2017), and highest average proximities to industrial PM 2.5 emissions (Mikati et al. 2018) and highest average exposures to industrial air toxins (Ard 2015) for non-Hispanic Black populations.
We found that racial-ethnic minority populations were more than two times as likely than non-Hispanic white populations to live in a census block group with highest air pollution levels (above 90 th percentile) on average. Those results are consistent with existing literature on disproportionate environmental risks for racial-ethnic minority populations (Collins 2016) and on groups or locations with higher risks for one environmental factor having higher risks for other factors too (Morello-Frosch and Lopez 2006; Su et al. 2012).
The findings from this study can be used to compare relative exposure disparities for different criteria air pollutants in a consistent way, providing additional context for previous studies of single pollutants. We found that in 2010, relative racial-ethnic exposure disparities (i.e., ratios of average exposures between the most-and least-exposed groups) were largest for NO 2 and smallest for O 3 . Relative income-based exposure disparities (i.e., ratios of average exposures between the lowest and highest income groups), although smaller than racial-ethnic exposure disparities for each pollutant, were largest for SO 2 and smallest (and similar) for NO 2 , CO, and O 3. (These results provide information on the rank-order of relative disparities in air pollution levels by pollutant; information on the rank-order of relative disparities in associated health impacts by pollutant would require further analysis, as discussed next).
Exposure disparities often connect with health disparities. Based on the magnitude of exposure disparities (e.g., 2010 national average PM 2.5 exposures for non-Hispanic Black people were 1.0 g m -3 higher-than-average), the resulting health disparities may be substantial (thousands of additional premature mortalities per year). Future research could usefully extend our exposure disparity results to provide rigorous, comprehensive investigation of the associated health impacts.
State-level results may be especially useful given the important role that states play in air pollution and environmental policy making (Abel et al. 2015). Exposures >5% greater than the national mean exposure within states were common for racial-ethnic minority populations, but not for non-Hispanic white populations. Exposure disparities varied substantially among states, even among states with similar characteristics (e.g., urbanicity, population, region). Our results emphasize differences among states in the level and makeup of exposure disparities, yet also demonstrate that exposure disparities were ubiquitous, including both large and small states, and states in all regions of the US, in 2010.
Our analyses by urbanicity were in part motivated by, and reflect, urban-rural differences in demographics and air pollution levels (Clark et al. 2017;Mikati et al. 2018;Rosofsky et al. 2018). Racial-ethnic disparities were larger for urban block groups for all pollutants except PM 2.5 . Of the six pollutants, the largest ratio between urban and rural racial-ethnic absolute disparity (5.5× larger) was for NO 2 . The NO 2 results are consistent with prior research (Clark et al. 2017). Over our study period, reductions in absolute racial-ethnic exposure disparity for PM 2.5 and O 3 were larger for rural than for urban areas, likely reflecting the fact that O 3 and most of PM 2.5 are secondary pollutants versus the other pollutants being mainly primary pollutants. Controlling for urbanicity, exposures were mostly higher for the lowest income category than the highest. Absolute income-based exposure disparities were also 7.5 times larger on average in urban than in rural areas.
The results by state and by urbanicity reflect that exposure disparities differ by spatial units (e.g., urban/rural, and by state); future research could explore these aspects further, for example, through a spatial decomposition of national exposure disparities.
Regulations such as the 1990 Clean Air Act Amendments have achieved substantial reductions in the concentrations of many pollutants. Our analysis reveals that, as a co-benefit, falling pollution levels have reduced absolute exposure disparities among racial-ethnic groups .   687  688  689  690  691  692  693  694  695  696  697  698  699  700  701  702  703  704  705  706  707  708  709  710  711  712  713  714  715  716  717  718  719  720  721  722  723  724  725  726  727  728  729  730  731  732 These findings are consistent with previous national research for NO Colmer et al. 2020). We found that a larger share of the racial-ethnic exposure disparity reduction was attributable to air pollution level reduction rather than changes in demographic and residential patterns.
Our study described patterns in exposure disparities but did not investigate aspects such as underlying causes or ethical or legal aspects. Systemic racism and racial segregation are two major causes discussed in multiple previous studies ( Our study has several limitations. The finest spatial scale of publicly-available Census demographic data for race-ethnicity and income is at Census block group level; we were unable to assess disparities at finer spatial scales than what the Census provides. Our disparity estimates do not account for (1) daily mobility for work, shopping, recreation, and other activities, (2) direct indoor exposure to indoor sources such as cigarette smoke, cooking emissions, or incense, (3) indoor-outdoor relationships in pollution levels, such as particle losses during airflow in ducts or ozone losses to indoor surfaces, or (4) occupational exposures. Our exposure disparity estimates were limited by uncertainties in the CACES exposure model predictions and in Census demographic data. Our uncertainty analysis (but not our main analysis) was limited to US EPA monitoring locations; we were not able to test potential exposure errors at locations without monitors on the national scale.