PM2.5 data inputs alter identification of disadvantaged communities

Communities of color and lower income are often found to experience disproportionate levels of fine particulate matter (PM2.5) air pollution in the US (Pope and Dockery 2006 J. Air Waste Manage. Assoc. 56 709–42; Brook et al 2010 Circulation 121 2331–78; Tessum et al 2021 Sci. Adv. 7 eabf4491). The federal and several state governments use relatively coarsely resolved (12 km) PM2.5 concentration estimates to identify overburdened communities. Newly available PM2.5 datasets estimate concentrations at increasingly high spatial resolutions (50 m–1 km), with different magnitudes and spatial patterns, potentially affecting assessments of racial, ethnic, and socioeconomic exposure disparities. We show that two recently available high-resolution datasets from the scientific community and the 12 km dataset are consistent for national and regional average, but not intraurban, PM2.5 concentration disparities in 2019. The datasets consistently indicate that regional average PM2.5 concentrations are higher in the least White (by 3%–65%) and most Hispanic census tracts (2%–47%), compared with in the most Non-Hispanic White tracts. However, in nine of the ten most populous cities, the three datasets differ on the order of least-to-most exposed population subgroups. We identified 1029 tracts (representing ∼4.5 million people) as disadvantaged (⩾65th percentile for poverty and ⩾90th percentile PM2.5 as defined by the Climate and Economic Justice Screening Tool) in all three datasets, 335 tracts (∼1.5 million people) as disadvantaged using both high-resolution datasets but not the 12 km dataset, and 695 tracts (∼2.7 million people) as disadvantaged in the 12 km dataset but not the high-resolution datasets. The 12 km dataset does not capture intraurban disparities and may mischaracterize disproportionately exposed neighborhoods. The high-resolution PM2.5 datasets can be further improved by ground-truthing with observations from rapidly expanding ground and mobile monitoring and by integrating across available datasets.


Introduction
Particulate matter under 2.5 µm (PM 2.5 ) is associated with cardiovascular disease, respiratory infections, asthma, and lung cancer [1,2] and leads to 85 000-200 000 premature deaths in the US annually [3,4].A long history of research demonstrates that PM 2.5 disproportionately impacts communities of color and those with lower income and educational attainment levels [3,5,6].While average PM 2.5 concentrations have decreased in the US over recent decades, PM 2.5 disparities persist [6,7].
The Justice40 Initiative, created by the Biden-Harris Administration, aims to address decades of underinvestment in disadvantaged communities overburdened by pollution, of which PM 2.5 is a major component.The US Environmental Protection Agency's (EPA) EJScreen tool and the Council on Environmental Quality's relatively new Climate and Economic Justice Screening Tool (CEJST) are the main tools used by the federal government to map disparities in environmental pollution, including PM 2.5 .The CEJST is now being applied to implement the Justice40 Initiative's goal that at least 40% of the benefits from federal grants, programs, and initiatives flow to disadvantaged communities that bear disproportionate environmental, health, climate, and socioeconomic burdens.As the Justice40 Initiative guides federal investments, the process and datasets used for identifying communities as disadvantaged could have major implications for how governmental programs are developed and implemented, as well as which communities and population subgroups benefit most from these investments.
Currently, both EJScreen and CEJST use a PM 2.5 concentration dataset that fuses concentrations simulated by the Community Multiscale Air Quality Modeling System (CMAQ) with monitor observations to estimate PM 2.5 at 12 km spatial resolution (hereafter 'CMAQf ').This resolution is too coarse to fully resolve disparities in urban areas [8], where more than 80% of the US population lives [9].
New high-resolution (50 m-1 km) datasets leveraging advances in computational algorithms and satellite observations offer full geospatial coverage and increasingly high spatial resolution, enabling assessments of intraurban PM 2.5 disparities.For example, van Donkelaar et al developed 0.01 • ('VD0.01')and 0.1 • ('VD0.1')spatial resolution PM 2.5 concentration estimates using the observational constraint provided by billions of total column aerosol optical depths from a combined eight different satellite retrievals, millions of ground-based observations, and over a trillion data points calculated from a meteorologically-and chemically-driven global chemical transport model [10].Amini et al used a machine learning approach and ∼82 billion data points as predictors across 20 years to estimate concentrations of PM 2.5 components at 50 m resolution in urban areas and 1 km resolution elsewhere ('Amini').These high-resolution PM 2.5 datasets enable assessment of neighborhood-scale pollution disparities, which is beyond the capability and intention of the US ground monitor network.Studies using high-resolution PM 2.5 concentration datasets find that Black, Asian, Hispanic, and lower income populations have disproportionately high PM 2.5 concentrations [6,7,11,12].However, each of these studies relied on a single PM 2.5 concentration dataset, despite that gridded PM 2.5 datasets often differ in concentration magnitude and spatial patterns, due to differences in methods and data inputs [6,[13][14][15].Previous studies comparing PM 2.5 concentration datasets highlight the need for further exploring discrepancies within cities, which is underscored by the increasing spatial resolutions of available datasets [14].
Here we compare census tract-level disparities in 2019 annual average PM 2.5 concentrations estimated using CMAQf, the coarsely resolved dataset used by EJScreen and CEJST, versus the more highly resolved VD0.01 and Amini datasets.Comparing PM 2.5 disparities using different off the shelf datasets is critical since (1) efforts by governments, community groups, and others to reduce PM 2.5 disparities depend on identifying overburdened communities, and (2) each of the currently available high-resolution PM 2.5 datasets that could be used for identifying overburdened communities carry uncertainties, with no one dataset or methodology considered the 'gold standard' or 'truth' for this purpose.Agreement between datasets on which communities are overburdened by PM 2.5 would provide more confidence when identifying disadvantaged communities.

PM 2.5 data sets
We obtain 2019 annual mean PM 2.5 surface concentrations from three datasets using different methods and spatial resolutions (table S1).Other highresolution datasets are available for earlier years, such as InMAP, a widely-used reduced form model intended to understand health and equity implications of changes in emissions rather than total PM 2.5 mass [3,16].Given large interannual variability in annual mean PM 2.5 , we use only datasets available for 2019.
The CMAQf dataset is available at 12 km resolution annually from 2002 to 2019 across the contiguous US (CONUS).A Bayesian space-time statistical downscaler model is used to 'fuse' PM 2.5 observations from the National Air Monitoring Stations/State and Local Air Monitoring Stations (NAMS/SLAMS) with 12 km gridded output from CMAQ.We consider annual averages for 2019 averaged all cells in the 2010 census tract locations.This CMAQf method at the census tract level is somewhat misleading since a substantial number (63%) of tracts are smaller than the CMAQf grid cell area, resulting in situations where several adjacent census tracts could have the same concentrations based on the coarse resolution of CMAQf.We also use version V5.GL.02 of the dataset described by van Donkelaar et al [10], which estimates concentrations from satellite retrievals of aerosol optical depth, chemical transport modeling, and ground-based observations (e.g.AQS and NAPS and IMPROVE over US).The VD datasets are available from 1998 to 2021 with monthly and annual concentrations and uncertainties.This dataset is available at both 0.01 (∼1 km) and 0.1 • , so we refer to it as VD0.01 and VD0.1, respectively.In this work, we focus on the higher resolution VD dataset, VD0.01, which has consistent similar magnitude and spatial variability with the coarser product, VD0.1; we provide some analysis of it in the SI.The third dataset we use is described by Amini et al in review [17].Amini calculates PM 2.5 components (elemental carbon, organic carbon (OC), nitrate, sulfate, and ammonium) using 987 unique monitoring stations and machine learning in CONUS.They choose whichever machine learning algorithm performs best per chemical component.To translate OC to organic aerosol, the actual component in PM 2.5 , which includes other elements bonded to the carbon like oxygen and nitrogen, we use an organic matter to OC ratio of 2 [18].Because Amini does not include sea salt aerosol (SSA) or dust, which can be large contributors to PM 2.5 , we use data from van Donkelaar et al [19] (V4.NA.03) with compositional information over North America to isolate and estimate the contributions and importance of SSA and dust US regions-for sensitivity purposes only.We use the Amini dataset without the additions of SSA and dust since that is what is available to the general public.The Amini dataset is available as annual averages from 2000 to 2019 with 50 m resolution in urban areas, defined as the urban areas in the 2010 census expanded by a 1 km buffer to account for development after 2010, and 1 km in non-urban areas.
We use in situ 24 h average PM 2.5 observations (408 monitors) from the EPA's air quality system (AQS) monitoring network [20] for comparison against the datasets in 2019.The AQS network contains air pollution data collected by urban and rural monitors across the US maintained by the EPA, state, local, and tribal agencies using both gravimetric and beta-attenuation techniques.We find that tract-averaged concentrations from CMAQf have the highest correlation with AQS observations (R 2 = 0.65) and those from Amini the lowest (0.37).Both CMAQf and VD0.01 have a small normalized mean bias (NMB) against observations on average across the US, but, in some tracts, concentrations differ from the collocated AQS monitor by up to 4.1 µg m −3 (figure S3).To understand the impact of the Amini dataset not including dust and SSA, we add a climatological version of both from the VD datasets and show that this minimizes the NMB from −0.057 to −0.022 while maintaining the spatial correlation (figure S3).For climatological dust and SSA, we use a five-year average from a speciated PM 2.5 dataset [19] of the most recent available years (2013-2017) because 2019 is not yet available.The previous fiveyear average (2008)(2009)(2010)(2011)(2012) shows the same mean and median and virtually the same spatial pattern.

Sociodemographic data
Demographic information is derived from the American Community Survey (ACS) conducted by the US Census Bureau and maintained by the National Historical Geographic Information System.Data are publicly available at www.nhgis.org[9].We extract 2019 information from the 2015 to 2019 average estimates on race, Hispanic or Latino origin (henceforth 'ethnicity'), educational attainment, and median household income for the 72 539 census tracts in the contiguous United States.To minimize the number of different categorical variables presented in this study, we combine racial groups into four mutually exclusive categories: non-Hispanic White, non-Hispanic Black (includes Black and African American), non-Hispanic Asian, and Hispanic.We acknowledge that this grouping omits people who select 'other' , 'two or more races' , or Hispanic-Asians and Hispanic-Blacks in the US Census Bureau.This number is just over 30 million people or roughly 9% of the US population.Since this manuscript uses census data, the racial-ethnic definitions are theirs.In the case of 'Asian' , Census provides the following definition: 'A person having origins in any of the original peoples of the Far East, Southeast Asia, or the Indian subcontinent including, for example, India, China, the Philippine Islands, Japan, Korea, or Vietnam.It includes people who indicate their race as "Asian Indian," "Chinese," "Filipino," "Korean," "Japanese," "Vietnamese," and "Other Asian" or provide other detailed Asian responses such as Pakistani, Cambodian, Hmong, Thai, Bengali, Mien, etc.' Similarly, for educational attainment, we combine graduate education (includes master's degree, professional school degree, and doctorate degree).For the 'most' educated census tracts, we look at tracts with the largest proportion of graduate degrees; these trends are also robust to using college education and not graduate degrees.

Methods
We select 2019 because it is the most recent year available across datasets.We harmonize the PM 2.5 surface concentrations with tract-level ACS demographics by determining the geographic boundaries of each tract and thereafter calculating a simple arithmetic average over all PM 2.5 grid cells within the tract.Following Kerr et al (2021), for tracts that lack a collocated grid cell (∼8% of dataset), due to their small size (or irregular geometry), we calculate the PM 2.5 levels at their centroids using inverse distance weighting interpolation of levels in the eight neighboring grid cells.If the grid cell center point is within the census tract boundary, then it counts as within.It does not need to be fully contained to be counted.We focus on the census tract resolution to be consistent with EJScreen and CEJST.Tract-level disparities have been shown to match well with both national scale and finer scale (e.g., blocks) disparities [21] where particular racialethnic groups are more or less exposed to pollution.
Rural or urban census tract designations come from the classification in the last decadal census in 2010.Urban tracts lie within the boundaries of an incorporated or census-designated place with >2500 residents; rural tracts are outside these boundaries.We classify suburban regions on the periphery of cities with >2500 residents as 'urban' .We assess disparities in the ten most populous metropolitan statistical areas (MSAs) in the US: New York City-Newark-Jersey City, NY-NJ-PA; Los Angeles-Long Beach-Anaheim, CA; Chicago-Naperville-Elgin, IL-IN-WI; Dallas-Fort Worth-Arlington, TX; Houston-The Woodlands-Sugar Land, TX; Washington-Arlington-Alexandria, DC-VA-MD-WV; Miami-Fort Lauderdale-Pompano Beach, FL; Philadelphia-Camden-Wilmington, PA-NJ-DE-MD; Atlanta-Sandy Springs-Alpharetta, GA; and Phoenix-Mesa-Chandler, AZ.For conciseness, we use colloquial names (e.g., New York City, rather than New York City-Newark-Jersey City, NY-NJ-PA) when discussing them.
We characterize PM 2.5 disparities across racial, ethnic, and sociodemographic population subgroups using two complementary methods: the 'demographic percentile approach' using the extremes of demographic variables as represented by the top and bottom deciles of population fraction for each subgroup and population-weighted means for each subgroup.We partition census tracts by extreme values of demographic variables using the first decile (0 to 10th percentile) and tenth decile (90th to 100th percentile) (i.e.'most Hispanic' equates to those tracts where the fraction of the Hispanic population is in the tenth decile).Deciles are calculated separately for all, urban, rural, and MSA tracts to account for differences between these geographies.When applied across all urban and rural areas, this approach gives a broad array of tracts distributed across the country, avoids absolute thresholds for different variables, and gives consistent sample sizes for the upper and lower extrema [22].We test whether the distributions of tract-level PM 2.5 concentrations associated with the demographic variables selected for the upper and lower extrema are drawn from the same distribution using the Kolmogorov-Smirnov (KS) test.If the pvalue corresponding to the KS test statistic is less than α = 0.05, we declare that there are significant differences in the distributions.
We also calculate population-weighted mean concentrations and use this information in two ways.First, we calculate population-weighted concentrations associated with three combinations of racial and socioeconomic groups: low income white (LIW), low income non-white (LIN), and high income white (HIW) [23].The 'low' and 'high' income classifications are based on the lower and upper quartiles of median household income within either all, urban, rural, or MSA designated tracts.For example, across CONUS, the low-income designation is associated with a median household income of $44 000 and the upper-income threshold with $80 000; these numbers vary by region and MSA.We also calculate PM 2.5 disparities as the ratio of population-weighted PM 2.5 for each population subgroup to the populationweighted overall average for different aggregations (i.e., all, urban, rural, or MSAs).A value of 1 indicates that a particular subgroup has the same PM 2.5 level as the overall population-weighted average for a given aggregation or MSA.
We calculate the tract-level fraction of all-cause mortality that is attributable to PM 2.5 ('attributable fraction' or AF) using a log-linear relative risk (RR) of 1.06 per 10 µg m −3 from Turner et al [24]: )) where PM 2.5 is the annual mean PM 2.5 concentration, and β is the slope of the association between PM 2.5 and all-cause mortality.

Results
We first compare the magnitude and spatial patterns of PM 2.5 concentrations between the three datasets (figures S1 and S2) and the agreement between each dataset and an average across the three to ground observations (figure S3).National population-weighted average PM 2.5 concentrations in the Amini dataset, which excludes dust and SSA, are 2.27 µg m −3 lower than CMAQf and 1.51 µg m −3 lower than VD0.01, and CMAQf is 1.26 µg m −3 larger than VD0.01 (figure S3 and S4).We compare the average annual PM 2.5 concentration uncertainty over CONUS available for VD0.01 (∼0.5 µg m −3 ) to the absolute difference between the VD0.01 dataset and each of the other two datasets.Although the national average uncertainty is smaller than the absolute concentration difference between the datasets, the spatial distribution (and therefore the disparities) exhibit greater variability.Indeed at the regional scale, VD0.01 average annual uncertainties are larger and range from 1.92 µg m −3 in the Northwest to 3.30 µg m −3 in the Midwest.Census tract-average PM 2.5 from CMAQf has the highest correlation with observations from the sparse AQS ground monitor network and Amini the lowest (more details in methods).Averaging the three datasets reduces the NMB to −0.0055 from a range across the datasets of −0.057 (Amini) to 0.013 (CMAQf) and the correlation remains high (R 2 = 0.62) (figure S3).
Focusing on the spatial distribution of PM 2.5 concentration percentiles to account for the absolute magnitude differences, all three datasets have broad geographic consistency with higher concentrations in the Eastern US and California compared with the rest of the Western US (figure S1).However, CMAQf uniquely shows high concentration percentiles in the Pacific Northwest and Amini uniquely shows relatively low concentration percentiles throughout California's Central Valley.Exclusion of dust and SSA in the Amini dataset may explain some of these differences in the Southwest where dust can make up more than 40% of PM 2.5 and along the West Coast where SSA contributes more than 50% (figure S5).PM 2.5 is challenging to estimate in the West owing to the complexities of modeling the unique topography and sources, such as mountains, seas, and fires.For example, CMAQ may be biased high during fire events [25], and we see larger values in the Pacific Northwest with CMAQf where fires can be a large source (figures S1 and S2).
To complement the analysis of disparities between the top and bottom deciles of population subgroups, we next calculate a ratio of the population-weighted PM 2.5 experienced by racial-ethnic subgroups relative to the overall population-weighted average (figure 2).Across all datasets, the NH-White subgroup is the least exposed by 5%-10% except in rural areas.The NH-Black subgroup has the highest concentrations nationally and in urban areas in the CMAQf and Amini datasets, but VD0.01 estimates that the NH-Asian subgroup experiences the highest concentrations, though by a small amount.The datasets are consistent that rural disparities are larger compared with urban disparities and in rural areas NH-Asian appears to experience the lowest concentrations.Amini tends to show larger disparities from the NH-White populations to whichever racial-ethnic group is most exposed with, for example, a spread in the Northern Great Plains of more than 25% relative to CMAQf with 18% and VD0.01 with 15%.
To explore the impact of intersectionality on PM 2.5 disparities, we also calculate populationweighted average concentrations of combined race and economic subgroups (figure S10).LIN experience higher mean PM 2.5 concentrations regionally, in urban areas by 2%-6% (VD0.01 = 2%, CMAQf = 5%, and Amini = 6%), and across the US by 3%-7% (VD0.01 = 3%, CMAQf = 6%, Amini = 7%) with all three datasets, except in the Northwest and the Southwest-and California, where only the Amini product shows LIN as most exposed.In contrast to urban areas, rural HIW show higher concentrations relative to rural LIW by 5%-26% (Amini = 5%, VD0.01 = 17%, CMAQf = 26%) depending on dataset, potentially because: (1) HIW may be more likely to live in suburbs that are closer to higher urban pollution even if they do not live in areas characterized as urban; and (2) HIW may have higher exposure to PM 2.5 in smoke from landscape fires in part due to living closer to the wildland-urban interface.
Differences between the concentrations associated with the top and bottom deciles of population demographics (figure 1) and relative disparities (the ratio of population-weighted PM 2.5 for each population subgroup to the overall population-weighted Here, we use the top and bottom deciles to classify which census tracts are included in the 'most' and 'least' categories.Disparities are shown for three geographies across the contiguous US (all, urban, and rural census tracts), and all tracts are further separated into eight major regions in the US following the Fourth National Climate Assessment regional definitions (figure S6).The difference in highest/lowest decile distributions for all demographic variables is statistically significant (p < 0.05) per the Kolmogorov-Smirnov (KS) test.Figure S7 shows the same plot but also including an average of the three datasets.
average for each geographic aggregation) (figure 2) are generally narrower in individual urban areas compared with in regions.The two high-resolution datasets are generally consistent in terms of relative ordering of PM 2.5 concentrations for different population subgroups across MSAs.PM 2.5 disparities using CMAQf concentrations are much narrower than when using the other two datasets for more than half the MSAs, as expected given the more limited spatial resolution of CMAQf.The concentration range associated with the most and least NH-White (and Hispanic) is largest in Los Angeles and Phoenix, two MSAs covering the largest spatial areas.
Phoenix is the only MSA of the ten most populated where the three datasets show consistent ordering of the population subgroups with the highest to lowest concentrations (highest concentrations for NH-Black populations and lowest concentrations for NH-White populations; figure 2).The datasets also show more consistent ordering of concentrations in population subgroups in Atlanta and Dallas compared with the other MSAs.
To better understand drivers of differences in PM 2.5 disparities across MSAs, we explore spatial variations in PM 2.5 and demographic variables for Los Angeles, Chicago, and Phoenix (figure 3; maps Figure Relative disparities in 2019 annual mean PM2.5 concentrations using each PM2.5 dataset and the population-weighting approach regionally and in the most populous MSAs.Relative disparities are calculated as the ratio of population-weighted PM2.5 for each population subgroup to the overall population-weighted average for each geographic aggregation.Differences in relative disparities across racial-ethnic groups were statistically significant for all regions and all MSAs except Dallas (table S2). Figure S9 shows the same plot but also includes an average of the three datasets.
for all ten MSAs shown in figures S11 and S12), which displayed different patterns of relative disparities in figure 2. As mentioned, all three datasets show consistent ordering of PM 2.5 concentrations among population subgroups in Phoenix.VD0.01 concentrations are larger than those from CMAQf across most of Phoenix, and the Amini dataset shows higher concentrations in the city center.This suggests that the similarities in relative disparities across Phoenix may be driven by concentrations in the city center where all three datasets are more similar.In Los Angeles, the two high-resolution datasets show consistent ordering of PM 2.5 concentrations among population subgroups, but the ordering is different when using CMAQf concentrations.In Los Angeles, it is particularly apparent that higher PM 2.5 concentrations are located in areas with lower NH-White percentages consistent with the literature [26], in this case along major roadways through the center of the city.This may explain why all three datasets are consistent in Los Angeles that non-Whites experience more relative exposure than NH-Whites.However, because the spatial PM 2.5 patterns differ across datasets and racial-ethnic groups are concentrated in different parts of the MSA, the PM 2.5 concentrations lead to inconsistencies in which non-White populations are most exposed.Unlike for Phoenix and Los Angeles, Chicago shows very little differences in PM 2.5 concentrations between population subgroups, regardless of the PM 2.5 dataset used; even though, the PM 2.5 estimates show distinct spatial patterns.CMAQf shows a diffuse mass of high concentrations, VD0.01 appears bimodal, and Amini is higher along the lakeshore and lacks the high concentrations away from the city center.When convolved with population demographics, these concentration differences may be obscured.
Since PM 2.5 concentrations and spatial patterns affect health impact assessments, we quantify the fraction of premature all-cause mortality that is attributable to PM 2.5 , the population AF, using each dataset.Nationally, estimated AFs range from 3.4% (Amini) to 4.7% (CMAQf), with VD0.01 in between at 4.2% (table S3).Approximate magnitudes are consistent with the literature [27,28].The racial-ethnic, income, and educational attainment breakdowns closely follow the exposure differences already discussed with NH-Whites experiencing the smallest contribution of PM 2.5 air pollution to premature all-cause mortality across datasets and NH-Black or NH-Asian the largest (table S3), similar to other work [11].Finer resolution datasets have been shown to lead to larger PM 2.5 AFs by 8%-19% [28], but the underlying average magnitude differences (8%-80%) of PM 2.5 concentrations among the datasets overwhelm that.
We also explore which census tracts would be flagged as 'disadvantaged' using the different PM 2.5 datasets, according to the definition used by the Biden Administration's Justice40 Initiative (figure 4).The Justice40 Initiative defines disadvantaged tracts as those that meet at least one associated socioeconomic threshold and any burden threshold, which in the case of air pollution is above the 90th percentile of PM 2.5 concentrations using the CMAQf dataset and a low income socioeconomic cutoff (⩾65th percentile of the percentage of a census tract's population in households where household income is at or below 200% of the Federal poverty level, not including students enrolled in higher education) [29,30].Using this definition, 2388 tracts are flagged as disadvantaged.Of these, 1029 tracts (43%), representing ∼4.5 million people are also identified in both high-resolution PM 2.5 datasets.VD0.01 identifies 61% of tracts flagged by CMAQf, and Amini flags 53% of those that CMAQf does in 2019.Over a quarter (695; 29%) of tracts identified as disadvantaged using CMAQf, representing ∼2.7 million people, are not flagged as disadvantaged using either of the high-resolution datasets; these areas are on average 14% White, 26% Hispanic, 3% Asian, and 54% Black.On average, tracts identified nationally as disadvantaged are 18% White, 55% Hispanic, 5% Asian, and 18% Black.Conversely, we found that 335 tracts (∼1.5 million people) are identified as disadvantaged using both high-resolution datasets but not CMAQf.Tracts flagged as disadvantaged by the finer resolution datasets are generally near large cities; those flagged by the Amini dataset alone are concentrated in the Southeast and those flagged only by the VD0.01 dataset are often in California, around Phoenix, and in the Midwest and Southeast.The areas flagged by both high-resolution datasets are on average 18% White, 55% Hispanic, 5% Asian, and 18% Black.This suggests that the higher resolution datasets are flagging areas that are more Hispanic, Asian, and White than the coarser resolution dataset (consistent with the nationally identified disadvantaged populations) but missing areas with many Black residents.Results are consistent using only PM 2.5 concentrations without the income cutoff (figure S13).Because of the difference in magnitude between datasets, using absolute values (e.g., comparing concentrations to the National Ambient Air Quality Standard or World Health Organization guideline levels) instead of percentiles, the Amini dataset would flag fewer tracts as disadvantaged compared with the other two datasets.

Summary and discussion
We show that one coarse (EPA's CMAQf) and two finer resolution (VD0.01 and Amini) annual average PM 2.5 concentration datasets show consistent racialethnic and sociodemographic disparities at national and regional but not intraurban scales.All three datasets show that NH-Black, Hispanic, and NH-Asian populations experience larger burdens by 15%-25% than NH-Whites across the US and regionally.At the intraurban scale across the ten largest MSAs, the datasets are generally, but not always, consistent that NH-Whites have 1%-3% (CMAQ = 1%, VD0.01 = 2%, and Amini = 3%) lower concentrations than the overall population-weighted average but are inconsistent in the relative exposure of other racial-ethnic subgroups.In most urban areas we explored, the high-resolution PM 2.5 datasets show wider disparities and greater consistency between each other than the coarsely resolved dataset.The two high-resolution datasets are complementary as they use different methods that help bound our understanding of spatial patterns of PM 2.5 .Other fine resolution PM 2.5 concentration datasets are also available, such as InMAP (though not for our year of analysis) [16,31] and could provide additional information about PM 2.5 spatial patterns.
Our results indicate that the coarsely resolved PM 2.5 concentration dataset used by EJScreen, the Justice40 Initiative, and several other EJ mapping tools [32][33][34][35][36][37][38][39][40][41] adequately captures absolute and relative racial-ethnic and sociodemographic disparities nationally and regionally but not at the intraurban scale.Further, these results suggest that the dataset used in EJScreen and the Justice40 Initiative may miss a substantial number of potentially disadvantaged tracts.More broadly, the choice of PM 2.5 dataset influences identification of overburdened communities, potentially leading to different conclusions about which air pollution mitigation strategies would reduce disparities and which population subgroups would benefit.
Limitations of our study include that the Amini dataset does not include all PM 2.5 components and that the underlying datasets are at different resolutions.The Amini dataset may also be biased low because it may not take into account the contribution of water uptake-component measurements of specific ions tend to be reported as truly dry while PM 2.5 measurements are made at a standard 35% humidity that contains a certain amount of water taken up by the hydrophilic aerosol components.While important to provide our results at the census tract scale to identify disadvantaged communities, our regridding centroid approach may potentially bias results, especially for the coarser resolution data sets.While it is beyond the scope of this study, additional work to elucidate which pollution sources need improvement within each dataset would be helpful to informing policy and management decisions.
Our work shows that studies focused on PM 2.5 disparities that rely on only one dataset may be limited in the strength of the conclusions that can be drawn, since some of the datasets may not fully identify disproportionately high concentrations for some geographies and population subgroups.Our assessment of the fraction of premature deaths attributable to PM 2.5 also revealed differences across datasets considered herein.While differences in the AF appeared small (<2%) among these datasets, when applied to the at-risk US population, these differences could lead to stark differences in estimated health impacts from PM 2.5 .We conclude that efforts to identify disadvantaged communities for additional resources or government intervention should carefully evaluate not just which pollutants should be included, but which datasets are most appropriate for the application.As the currently available higher spatial resolution datasets differ in PM 2.5 concentrations and their spatial distributions at these scales, additional research should seek to improve these datasets and use them concurrently with other information sources.Future work should also investigate the pollution burden and socioeconomic characteristics of areas where the PM 2.5 concentrations estimated by CMAQf are inconsistent with other leading, high-resolution datasets.Ground truthing remains a challenge due to the sparseness of the ground monitor network.Satellites, low-cost sensors, mobile monitoring, and machine learning all present opportunities for leveraging observations that have more complete spatial coverage albeit with their own uncertainties.Integrating across multiple quantitative datasets, as in the average across the datasets we explored here, and qualitative information and citizen science approaches can provide a deeper understanding of how pollution affects the lived experience in localized areas.

Figure 1 .
Figure 1.Disparities in 2019 annual mean PM2.5 concentrations using each PM2.5 dataset and the demographic percentile approach, for (a) most/least non-Hispanic (NH) White regionally, (b) most/least Hispanic regionally, (c) most/least NH-White in the ten largest MSAs, and (d) most/least Hispanic in the ten largest MSAs.Here, we use the top and bottom deciles to classify which census tracts are included in the 'most' and 'least' categories.Disparities are shown for three geographies across the contiguous US (all, urban, and rural census tracts), and all tracts are further separated into eight major regions in the US following the Fourth National Climate Assessment regional definitions (figureS6).The difference in highest/lowest decile distributions for all demographic variables is statistically significant (p < 0.05) per the Kolmogorov-Smirnov (KS) test.FigureS7shows the same plot but also including an average of the three datasets.

Figure 3 .
Figure 3. Maps of census tract-level 2019 annual mean PM2.5 concentrations and percent of the population that is NH-White in Los Angeles, Chicago, and Phoenix.PM2.5 color bars are discretized into deciles, and we provide the PM2.5 concentrations corresponding to each decile in figure S11.Thick black lines indicate major roadways.

Figure 4 .
Figure 4. Locations of census tracts that would be identified as 'disadvantaged' in one or more dataset.Insets of four cities (Los Angeles, Houston, Atlanta, and Chicago) are also plotted with major roadways in light grey and do not obscure any points.Following the Justice40 Initiative's definition for PM2.5, tracts are identified as 'disadvantaged' if they are >90th percentile for PM2.5 concentrations and ⩾65th percentile of the percentage of a census tract's population in households where household income is at or below 200% of the Federal poverty level, not including students enrolled in higher education.See figure S13 for all dataset permutations.