Companion 8.1d

Air pollution and the vote

Tracts with more diesel exposure voted more Democratic in 2020, by 24.8 points per standard deviation of exposure, but only 4.9 points once race, income, education and density are held fixed.

A companion to Chapter 8.1, Five axes of the county vote

Author: Vishal Singh
Affiliation: NYU Stern
Published: 30 September 2026

DraftFigures and numbers may change as the analysis is rebuilt.what is still being checked

Still being checked: the release years of the deprivation and asthma values and of the American Community Survey values in the tract atlas, and which archived copy of EJScreen version 2.3 the atlas was built from. The atlas file records none of them.

In the 2020 presidential election, the tenth of census tracts in 47 states with the most diesel exhaust in their air had a Democratic two-party margin of +49.1 (95% CI 42.4 to 55.8) points. The tenth with the least had a margin of −29.8 (95% CI −32.2 to −27.4) points, a Republican lead. The margin is the Democratic vote minus the Republican vote as a share of the two, so a positive value is a Democratic lead. Diesel exposure is the modeled annual average concentration of diesel particulate matter in a tract's air, from the EPA's EJSCREEN, in micrograms per cubic meter. Across tracts it moves with the EPA's traffic proximity score, vehicles per day on nearby major roads weighted by distance. The two ends of the range also differ in who lives in them, and the gap leaves open how much of it remains among tracts that resemble each other.

The margin rises with each step in exposure

Figure 1 sorts the tracts into ten groups of equal size by diesel exposure and shows the margin in each. The point estimates rise with each step. The deciles below the fifth lean Republican and the deciles above it lean Democratic, and the fifth decile is an even split within its interval, at +1.4 (95% CI −1.8 to 4.7) points. The intervals are widest in the top two deciles, and they allow for tracts in one county resembling each other.

Figure 1

The Democratic margin rises from −29.8 points in the tenth of tracts with the least diesel exposure to +49.1 points in the tenth with the most

Democratic minus Republican share of the two-party 2020 presidential vote, percentage points; vote-weighted mean by decile of tract diesel particulate exposure, each row labeled with the decile and its median exposure in micrograms per cubic meter; 95% intervals clustered by county

Show data

Data for The Democratic margin rises from −29.8 points in the tenth of tracts with the least diesel exposure to +49.1 points in the tenth with the most
conf_high	conf_low	decile	decile_label	diesel_median	margin	n
−27.42	−32.23	1	1 (0.04 µg/m³)	0.042	−29.82	8,114
−26.24	−31.31	2	2 (0.07)	0.068	−28.78	8,113
−19.39	−24.18	3	3 (0.09)	0.092	−21.78	8,113
−7.48	−13.25	4	4 (0.12)	0.118	−10.36	8,114
4.68	−1.83	5	5 (0.14)	0.144	1.42	8,113
12.45	6.15	6	6 (0.17)	0.171	9.30	8,113
21.73	14.99	7	7 (0.20)	0.203	18.36	8,114
30.07	22.62	8	8 (0.24)	0.244	26.34	8,113
40.11	28.24	9	9 (0.31)	0.311	34.18	8,113
55.84	42.39	10	10 (0.48)	0.477	49.12	8,114

The dots climb from a Republican lead in the first row to a Democratic lead in the tenth, and the interval of the fifth row includes an even split. Each dot is the vote-weighted mean Democratic two-party margin of the tracts in one decile of diesel exposure, labeled with the decile and its median exposure in micrograms per cubic meter. The whiskers are 95% intervals that allow tracts in one county to resemble each other. Values in Table 1.

Note: Deciles of diesel exposure hold equal numbers of tracts; ties are broken by census tract code.Note: Tracts with 100 or fewer votes are left out. Intervals treat counties as the units that vary. They do not include error in the allocation of precinct votes to tracts or in the modeled pollution values.Source: Precinct votes allocated to census tracts (Fekrazad 2025), 2020 presidential election; EPA EJScreen version 2.3 (2024 release, 2020 air data); American Community Survey tract data (vintage not recorded in the atlas); vote-weighted, full counts of tracts, no sampling weights

The tenth decile differs from the first in more than the air. The median tract in the tenth decile holds 5,200 people per square kilometer, against 14 in the first, and the white-alone share of residents is 36% against 79%. Diesel exposure moves with density across tracts. Holding these traits fixed means comparing tracts that resemble each other on them.

Modeled votes, modeled air, one election

The votes are estimates and not counts. Precinct returns from the 2020 presidential election were allocated to census tracts by a land-cover model [1], so each tract's margin carries the error of that allocation. The air is modeled too. EJSCREEN gives each tract a modeled annual average for 2020 of diesel particulate matter, fine particles and nitrogen dioxide, along with the traffic proximity score.The EPA removed EJSCREEN from its website in February 2025. Public Environmental Data Partners hosts a reconstruction of the version the chapter uses, released in July 2024 with air data modeled for 2020 [2]. The sample is 81,000 tracts with more than 100 estimated votes and every pollution measure, control and land area that the models need, and each tract counts in proportion to its votes. The intervals are clustered by county, so they allow tracts in one county to resemble each other, and they leave out the error in the vote allocation and in the modeled pollution. Every estimate is descriptive. It says how the margin moves with exposure across tracts in one election. The report gives every estimate with its interval and sample size.

Controls shrink every gap, and fine particles end near zero

Figure 2 compares tracts on who lives in them. Each exposure is standardized, so a coefficient is the difference in the margin, in points, between tracts one standard deviation apart in exposure. With no controls the diesel coefficient is 24.8 points. Holding the white, Black and Hispanic shares of residents, the college share and median household income fixed leaves 42% (95% CI 36% to 48%) of that, and adding the log of population density leaves 20% (95% CI 12% to 27%). Density is closely tied to the Democratic vote across American places [3]. Traffic proximity and nitrogen dioxide also shrink with each set of controls and keep intervals above zero. Fine particles are the exception. With density held fixed their coefficient is −0.7 (95% CI −2.8 to 1.4) points, an interval that includes zero.

The four exposures move together, so the first rows of each panel do not say which one matters. Diesel exposure is correlated with traffic proximity and with nitrogen dioxide across tracts. In the joint row, where all four enter one model with the controls and density, only traffic proximity has an interval above zero. The data cannot say which exposure carries the difference that remains. Companion 8.1c, Measures of hardship compared sets this specification beside a common one used for every hardship measure. That specification reports the Democratic share in place of the margin, leaves out income, standardizes with a vote-weighted standard deviation and adds state fixed effects. That chapter traces the difference between the two diesel estimates mainly to the state effects.

Figure 2

Controls shrink the Democratic gap for every pollutant, and with all four exposures in one model traffic proximity keeps 8.7 points while diesel falls to 1.3

Difference in the Democratic two-party margin (percentage points) per standard deviation of tract exposure, by pollutant and set of controls; 81,134 tracts; 95% intervals clustered by county

No controls
Race, income, education
Plus density
All four together
County fixed effects

Diesel particulate matter

Traffic proximity

Nitrogen dioxide

Fine particles (PM2.5)

Margin per SD of exposure (points)

Show data

Data for Controls shrink the Democratic gap for every pollutant, and with all four exposures in one model traffic proximity keeps 8.7 points while diesel falls to 1.3
conf_high	conf_low	estimate	n	pollutant	specification
27.99	21.52	24.75	81,134	Diesel particulate matter	No controls
12.56	8.11	10.34	81,134	Diesel particulate matter	Race, income, education
7.21	2.60	4.90	81,134	Diesel particulate matter	Plus density
4.64	−2.05	1.29	81,134	Diesel particulate matter	All four together
8.89	5.26	7.07	81,134	Diesel particulate matter	County fixed effects
30.22	19.19	24.70	81,134	Traffic proximity	No controls
15.35	9.04	12.20	81,134	Traffic proximity	Race, income, education
11.29	6.08	8.69	81,134	Traffic proximity	Plus density
11.72	5.59	8.65	81,134	Traffic proximity	All four together
8.56	4.64	6.60	81,134	Traffic proximity	County fixed effects
26.28	22.35	24.31	81,134	Nitrogen dioxide	No controls
11.86	8.29	10.07	81,134	Nitrogen dioxide	Race, income, education
5.80	1.46	3.63	81,134	Nitrogen dioxide	Plus density
1.85	−3.40	−0.777	81,134	Nitrogen dioxide	All four together
7.80	5.16	6.48	81,134	Nitrogen dioxide	County fixed effects
12.24	5.26	8.75	81,134	Fine particles (PM2.5)	No controls
3.09	−1.37	0.859	81,134	Fine particles (PM2.5)	Race, income, education
1.36	−2.80	−0.721	81,134	Fine particles (PM2.5)	Plus density
0.309	−3.76	−1.72	81,134	Fine particles (PM2.5)	All four together
2.53	−2.43	0.050	81,134	Fine particles (PM2.5)	County fixed effects

The dots in every panel move toward zero as controls are added, and fine particles end near it, with an interval that includes zero once controls are added. The diesel coefficient with race, income, education and density held fixed is 4.9 (95% CI 2.6 to 7.2) points. In the joint row, with all four exposures in one model, only traffic proximity has an interval above zero, and diesel falls to 1.3 (95% CI −2.1 to 4.6) points. Each dot is the coefficient from a vote-weighted regression of the tract margin on the standardized exposure, with the controls named on the row. The bottom row compares tracts within their own county, a different comparison and not a further control. Values in Table 3.

Note: Controls are the white, Black and Hispanic shares of residents, the college share and median household income; density is the log of people per square kilometer. The joint row enters all four exposures in one model with the controls and density.Note: The county fixed effects row is a different comparison and not a further control: it uses the same controls and density and compares tracts only within their own county.Note: Intervals treat counties as the units that vary. They do not include error in the allocation of precinct votes to tracts or in the modeled pollution values.Source: Precinct votes allocated to census tracts (Fekrazad 2025), 2020 presidential election; EPA EJScreen version 2.3 (2024 release, 2020 air data); American Community Survey tract data (vintage not recorded in the atlas); vote-weighted, full counts of tracts, no sampling weights

Among people, the support gap shrinks within party

Tract margins describe places, and they leave open whether people in more exposed places hold different views on pollution rules. The Cooperative Election Study asked respondents in 2020 whether they support strengthening the Environmental Protection Agency's enforcement of the Clean Air Act and the Clean Water Act even if it costs U.S. jobs. The item names two statutes and a cost in jobs, so it measures more than support for the agency, and it does not ask about local pollution or health. Respondents in the fifth whose counties have the most diesel exposure are more supportive, 70% against 51% in the fifth whose counties have the least, and they are also more often Democrats. Within party the rise in support with exposure shrinks, and once the density of the county is added every group's interval includes zero (Figure 3). The survey assigns each respondent the exposure of their county, so it cannot see how much air pollution a respondent breathes.

Figure 3

Support for EPA enforcement rises with county diesel exposure, but the rise shrinks within party and its interval includes zero with county density held fixed

Percentage-point difference in support for strengthening EPA enforcement of the Clean Air Act and Clean Water Act even if it costs U.S. jobs, per standard deviation of county diesel exposure; 57,411 respondents in 2,630 counties, 2020; 95% intervals clustered by county

All respondents
Democrats
Independents
Republicans

No controls

Party ID and demographics

Plus county density

Support per SD of diesel (points)

Show data

Data for Support for EPA enforcement rises with county diesel exposure, but the rise shrinks within party and its interval includes zero with county density held fixed
conf_high	conf_low	estimate	group	n	specification
7.57	5.30	6.43	All respondents	57,411	No controls
1.74	0.861	1.30	All respondents	57,411	Party ID and demographics
0.369	−1.12	−0.377	All respondents	57,411	Plus county density
0.829	−0.773	0.028	Democrats	28,671	No controls
0.955	−0.620	0.168	Democrats	28,671	Party ID and demographics
0.568	−1.77	−0.603	Democrats	28,671	Plus county density
4.81	1.34	3.08	Independents	8,643	No controls
3.51	−0.177	1.66	Independents	8,643	Party ID and demographics
1.18	−4.11	−1.47	Independents	8,643	Plus county density
6.16	2.39	4.27	Republicans	20,097	No controls
4.78	1.18	2.98	Republicans	20,097	Party ID and demographics
3.40	−1.42	0.991	Republicans	20,097	Plus county density

The interval for all respondents lies above zero in the left panel and includes zero in the right panel, where every group's interval includes it. The Republican interval lies above zero in the first two panels and includes zero in the third. Each dot is the difference in support, in points, per standard deviation of county diesel exposure, from a weighted linear probability model estimated on that group alone. The left panel uses no controls, the middle panel holds party ID and demographics fixed and the right panel adds the log density of the county. Values in Table 6.

Note: The CES is an opt-in panel, so its intervals are model based. County exposure is a county value assigned to each respondent, not an exposure the respondent was measured to have.Note: It is a different CES item from the one on giving the EPA power to regulate carbon dioxide.Source: Cooperative Election Study 2020 wave, item on strengthening EPA enforcement of the Clean Air Act and Clean Water Act, weighted with the CES survey weight; county diesel exposure from EPA EJScreen version 2.3 averaged over tracts by 2020 votes

What remains after controls is a difference between places in one election, and two other comparisons in the chapter qualify it. Across counties, the margin left over after race, income and education varies strongly between states. Once density also enters the model, the sign of its correlation with county diesel exposure depends on whether large counties count more (Table 5). The chapter also compares modeled asthma prevalence across the deciles of exposure, and that comparison cannot register local air. The CDC's PLACES estimates of current asthma among adults [4] are computed from the age, sex, race and education of survey respondents and the poverty rate of the county, applied to each tract's population. One tract's estimate differs from another's through its demographic mix and its county, and not through its air. The pattern describes how the modeled values line up with diesel exposure and does not test whether exposure affects asthma.

What these tract estimates cannot tell you

The estimates describe places in one election. They do not say how the residents of a tract voted, because a pattern across areas can differ from the pattern among the people in them [5], a limit known as ecological inference. The exposures and the vote are both from 2020, so the data cannot say whether pollution changes how people vote or whether people who vote a certain way settle near sources. Nor can they say whether the siting of sources follows political power, a claim the environmental justice literature makes [6]. Holding traits fixed is partial, the four exposures move together, and the vote and the air are both modeled, so the estimates carry errors that the intervals leave out.

Go deeper

The report: every estimate with its interval and sample size.
Tables: every figure of the report as a table.
Neighborhoods: the explorer for this part.

Data sources

Precinct votes on census geographiesPrecinct votes allocated to 2020 census tracts (Fekrazad 2025) with the ACS demographics of the tract atlas (ACS years not recorded), 2020 presidential electionYears: 2016 and 2020 (contiguous United States only) · Weights: votes cast in the tract (pres_tot); full counts of tracts, no sampling weights · License: CC0 (data doi:10.7910/DVN/Z8TSH3), as stated in the project README.
EJScreenEPA EJScreen version 2.3 (July 2024; modeled air data for 2020), tract values in the atlas; EPA removed the tool in February 2025 and the archived copy is not recordedYears: Version as distributed with the tract atlas file. · Weights: none (tract values) · License: US government data. The EPA removed EJScreen from its website in February 2025, so the archived copy is the source to cite.
Area Deprivation IndexArea Deprivation Index national rank, tract values in the atlas; release year not recordedYears: Vintage as distributed with the tract and ZCTA atlas files. · Weights: none (tract values) · License: The Neighborhood Atlas may restrict commercial use and redistribution; read its terms before publishing derived tract values.
CDC PLACESCurrent asthma among adults, model-based tract estimates in the atlas; release year not recordedYears: Release as distributed with the tract and ZCTA atlas files. · Weights: none (tract values) · License: US government data, open access.
Cooperative Election StudyCooperative Election Study 2020 wave; support for strengthening EPA enforcement of the Clean Air Act and Clean Water Act even if it costs U.S. jobsYears: 2006-2024 (every year; self-reported turnout is available in the 10 even years) · Weights: weight (single wave) · License: See the provider's terms of use.

All data sources

References

Fekrazad, Amir (2025). A dataset of US precinct votes allocated to Census geographies with precision. Scientific Data 12: 794. doi:10.1038/s41597-025-05140-3
US Environmental Protection Agency (2024). EJScreen Technical Documentation for Version 2.3 (revised July 31, 2024). Office of Environmental Justice and External Civil Rights, Washington, DC. epa.gov/system/files/documents/2024-07/ejscreen-tech-doc-version-2-3.pdf
Rodden, Jonathan A. (2019). Why Cities Lose: The Deep Roots of the Urban-Rural Political Divide. Basic Books, New York.
Greenlund, Kurt J.; Lu, Hua; Wang, Yan; Matthews, Kevin A.; LeClercq, Jennifer M.; Lee, Benjamin; Carlson, Susan A. (2022). PLACES: Local Data for Better Health. Preventing Chronic Disease 19: 210459. doi:10.5888/pcd19.210459
Robinson, W. S. (1950). Ecological Correlations and the Behavior of Individuals. American Sociological Review 15(3): 351-357.
Mohai, Paul; Pellow, David; Roberts, J. Timmons (2009). Environmental Justice. Annual Review of Environment and Resources 34: 405-430. doi:10.1146/annurev-environ-082508-094348

How to cite

@misc{singh2026airpollutionandthevote,
  author = {Singh, Vishal},
  title = {Air pollution and the vote},
  year = {2026},
  url = {https://vishalsingh.org/politics/air-pollution-and-the-vote},
  note = {In The Public Mind. Published 30 September 2026.}
}