Chapter 13.1

Measures of hardship compared

Of the 20 tract measures of hardship that track the 2020 Democratic vote before any controls, 3 keep at least 50% of that association once race and ethnicity, education, density and state are held fixed, 9 lose more than 50% of it and 8 reverse sign, by a grouping rule set after a first run of the estimates.

Author: Vishal Singh
Affiliation: NYU Stern
Published: 3 October 2026

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

Still being checked: the license terms of the Area Deprivation Index, the definition and vintage of the ZIP code college variable and of the ZIP code controls, and the EJScreen version (2.32 in the atlas build code, 2.3 in the air pollution chapter). The grouping of measures follows a rule that was set after a first run of the estimates.

Of the 20 tract measures of hardship that go with the 2020 Democratic vote before any controls, 3 keep at least 50% of that association once race and ethnicity, education, population density and state are held fixed, 9 lose more than 50% of it and 8 reverse sign. The raw coefficients run from −10.4 for arthritis, a modeled PLACES estimate, to 13.1 for proximity to hazardous waste facilities, an EJScreen indicator, in points of Democratic share of the two-party vote per standard deviation of the measure. The adjusted ones run from −3.2 for ozone, an EJScreen model estimate, to 7.6 for lack of health insurance, a PLACES model estimate, and their median absolute value is 2.2 points against 4.7 for the raw ones. The rule that sorts the measures into these groups was set after a first run of the estimates, and the last section shows that two measures sit close to its line.

Each measure is signed so that a higher value means more hardship, so a positive coefficient means a more Democratic vote where hardship is higher (social capital, taken up below, is the exception). The estimates compare places. The votes are precinct returns that a land-cover model allocated to census geographies [1], so each area's Democratic share is a modeled estimate and not a count, and a pattern across areas can differ from the pattern among the people in them [2] (ecological inference). Chapter 13.6, Air pollution and the vote covers air pollution in detail, and chapters are planned on deprivation and childhood outcomes (Chapter 13.3, Deprivation and mobility (planned)), health (Chapter 13.4, Health outcomes and the vote (planned)) and social ties (Chapter 13.5, Social capital (planned)). This chapter puts the measures on one footing and does not rebuild those results.

The 2020 vote layer covers the contiguous states and the District of Columbia and holds 83,648 census tracts. The tract atlas has a row for 82,765 of them. The 883 without a row are the tracts of Connecticut, which the atlas codes by planning region and the vote layer by county; they carry 1.8 million votes and voted 60% Democratic. Of the atlas tracts, 81,617 have more than 100 estimated votes, and the common sample keeps the 80,873 of those that have every measure and control. They lie in 3,081 counties in 47 states and carry 154.8 million votes. The 744 tracts that drop out at this point voted 74.3% Democratic, against 52.0% in the sample. The race and college controls are missing for 218 of them, 203 in the District of Columbia, and those tracts voted 93% Democratic. The other 526 lack an Area Deprivation Index, Opportunity Atlas or EJScreen value and voted 59% Democratic (Figure 1).

Figure 1

The Social Capital Atlas drops ZIP code areas that voted 40% Democratic against 53% for those kept, and missing controls drop tracts in the District of Columbia

Units dropped at each step of the tract sample and the ZIP code area sample, with the votes they carry and their Democratic share of the two-party vote, 2020

Sample	Step	Units dropped	Units left	Votes dropped (millions)	Democratic share of dropped votes (%)
Tracts	In the 2020 vote layer	n/a	83,648	n/a	n/a
Tracts	No row in the atlas	883	82,765	1.82	60
Tracts	100 or fewer estimated votes	1,148	81,617	0.01	60
Tracts	Missing a race share, the college share or population density	218	81,399	0.36	93
Tracts	Missing the Area Deprivation Index	321	81,078	0.18	61
Tracts	Missing an Opportunity Atlas estimate	200	80,878	0.28	59
Tracts	Missing an EJScreen indicator	5	80,873	0.01	48
ZIP code areas	In the 2020 vote layer	n/a	32,462	n/a	n/a
ZIP code areas	No row in the atlas	34	32,428	0.00	22
ZIP code areas	100 or fewer estimated votes	3,008	29,420	0.14	30
ZIP code areas	Missing a race share, the college share or population density	96	29,324	0.03	55
ZIP code areas	Missing the Area Deprivation Index	48	29,276	0.02	41
ZIP code areas	Bridge sample: Area Deprivation Index and PLACES present	n/a	29,276	n/a	n/a
ZIP code areas	Missing a Social Capital Atlas value	11,138	18,138	7.66	40
ZIP code areas	Missing an IRS value	99	18,039	0.14	54

Show data

Data for The Social Capital Atlas drops ZIP code areas that voted 40% Democratic against 53% for those kept, and missing controls drop tracts in the District of Columbia
dem_dropped	n_dropped	n_left	sample	step	votes_dropped
n/a	n/a	83,648	Tracts	In the 2020 vote layer	n/a
60.20	883	82,765	Tracts	No row in the atlas	1.82
60.03	1,148	81,617	Tracts	100 or fewer estimated votes	0.008
93.41	218	81,399	Tracts	Missing a race share, the college share or population density	0.365
60.51	321	81,078	Tracts	Missing the Area Deprivation Index	0.180
59.06	200	80,878	Tracts	Missing an Opportunity Atlas estimate	0.282
47.84	5	80,873	Tracts	Missing an EJScreen indicator	0.007
n/a	n/a	32,462	ZIP code areas	In the 2020 vote layer	n/a
21.67	34	32,428	ZIP code areas	No row in the atlas	0.001
29.99	3,008	29,420	ZIP code areas	100 or fewer estimated votes	0.145
54.74	96	29,324	ZIP code areas	Missing a race share, the college share or population density	0.027
40.72	48	29,276	ZIP code areas	Missing the Area Deprivation Index	0.017
n/a	n/a	29,276	ZIP code areas	Bridge sample: Area Deprivation Index and PLACES present	n/a
40.10	11,138	18,138	ZIP code areas	Missing a Social Capital Atlas value	7.66
53.81	99	18,039	ZIP code areas	Missing an IRS value	0.138

The tract sample loses 744 tracts with more than 100 votes after the 883 tracts with no atlas row, and the ZIP-only sample loses 11,138 ZIP code areas for a missing Social Capital Atlas value. Each row counts the units removed at that step, in the order shown, with the units that remain, the millions of estimated votes in the removed units and the Democratic share of their two-party vote.Note: Steps that drop no unit are left out: no tract and no ZIP code area lacks a PLACES value.Note: Alaska and Hawaii are not in the vote layer. The tracts with no atlas row are all in Connecticut, which has votes in the vote layer but is coded by planning region in the atlas.Note: The ZIP code atlas has Connecticut ZIP code areas, so the ZIP sample covers Connecticut.Source: Precinct returns allocated to 2020 tracts and ZIP code tabulation areas (Fekrazad 2025) with the project atlases; 2020 presidential vote; counts of units and votes, unweighted sums

One footing for every measure

The comparison fixes the outcome, the weights, the controls and the interval method, and it gives each measure the same sign and scale.

The outcome is the Democratic share of the 2020 two-party presidential vote, in points of share. The Chapter 13.6, Air pollution and the vote chapter reports the margin, the Democratic share minus the Republican share, and a coefficient of the margin is 2 times the coefficient of the share on the same data. Each area counts in proportion to its estimated votes. A measure is cut at its 1st and 99th percentiles across areas, so that a few extreme values do not set the scale, and then standardized with the vote-weighted mean and standard deviation. A coefficient is the difference in Democratic share, in points, between two areas one standard deviation apart on the measure, and with controls the two areas also match on the controls. The vote-weighted standard deviation of the Democratic share across tracts is 20.9 points.

The controls are the shares of residents who are white, Black, Asian and Hispanic, the college share, the log of population density and a fixed effect for each state, which compares areas inside one state. Race and ethnicity, education and density explain 72% of the variation in the vote within states, so an adjusted comparison is among tracts that resemble each other on traits that predict the vote closely. The controls leave out income, which is itself a measure of hardship, and age; the last section adds both as a variation. Intervals are 95% intervals from standard errors clustered by state, with the critical value of a t distribution that has one fewer degree of freedom than there are clusters, since the number of clusters is small [3].

The measures come from six sources in the project atlas, and the atlas build documentation, not the data files, dates them. The Area Deprivation Index ranks neighborhoods from least to most deprived [4]. The atlas holds the 2023 release of the Neighborhood Atlas, with tract values that are unweighted means of block-group ranks and ZIP code values that aggregate them by area, an approximation. The terms of use may restrict redistribution and are unconfirmed, so this chapter publishes coefficients and no area values. The Opportunity Atlas gives model-based estimates of outcomes in adulthood for children who grew up in a tract [5], here children born between 1978 and 1983 whose parents had low incomes, interpolated from 2010 tracts to 2020 tracts. They do not describe the tract's current residents. EJScreen is the Environmental Protection Agency's screening tool, here the 2024 release as kept in a community mirror after the agency withdrew it.The atlas build code and Chapter 13.6, Air pollution and the vote name different versions of the 2024 release, and the two have not been reconciled; the methods note gives both. The PLACES estimates of adult health are crude prevalence from a model, the 2024 release for tracts and the 2025 release for ZIP code areas, taken up below [6]. The tract controls from the American Community Survey cover 2019 to 2023, so the vote of November 2020 is set against measures dated 2019 to 2024.

Two sources exist for ZIP code areas only. The Social Capital Atlas is a 2022 snapshot built from Facebook friendships, with noise added to protect privacy [7]; the 6 columns used here match the 2022 release value for value. The IRS ZIP code data are for tax year 2022 and describe households that file a tax return. The ZIP code sample has its own counts and is described in its section. Social capital keeps its published direction, so a higher value means more connection, and the other measures are signed toward more hardship.

Before controls, the measures disagree in sign

Figure 2 plots the raw and adjusted coefficient of each of the 24 tract measures. Before controls, 12 measures have a coefficient with an interval above zero, 8 have one with an interval below zero and 4 have an interval that includes zero. The positive ones are mainly environmental: 13.1 (95% CI 11.9 to 14.4) for hazardous waste proximity, 11.9 (95% CI 8.1 to 15.6) for traffic proximity, 11.4 (95% CI 10.1 to 12.7) for nitrogen dioxide and 11.6 (95% CI 8.6 to 14.6) for diesel particulate matter. The negative ones are mainly deprivation and illness: −10.4 (95% CI −11.8 to −8.9) for arthritis, −6.3 (95% CI −8.5 to −4.2) for depression and −5.9 (95% CI −8.2 to −3.5) for the Area Deprivation Index, so more deprived tracts voted less Democratic.

Figure 2

Raw associations with the Democratic vote split in sign across 24 hardship measures, and adjusted ones are smaller and mostly positive

Democratic share of the 2020 two-party presidential vote, points per standard deviation of each measure, tracts, raw and with race and ethnicity, education, density and state held fixed

No controls
Common specification

Show data

Data for Raw associations with the Democratic vote split in sign across 24 hardship measures, and adjusted ones are smaller and mostly positive
adj_hi	adj_lo	adjusted	group	label	n	pattern	raw	raw_hi	raw_lo	view
5.22	2.58	3.90	Deprivation: Area Deprivation Index	Area Deprivation Index rank	80,873	Reverses sign	−5.86	−3.51	−8.22	source
4.34	2.88	3.61	Childhood outcomes: Opportunity Atlas	Low household income rank in adulthood	80,873	Keeps 50 percent or more of the raw coefficient	3.04	4.81	1.27	source
2.70	1.26	1.98	Childhood outcomes: Opportunity Atlas	Low individual income rank in adulthood	80,873	Reverses sign	−2.37	−0.695	−4.04	source
1.28	0.756	1.02	Childhood outcomes: Opportunity Atlas	Incarceration	80,873	Loses more than 50 percent of the raw coefficient	2.46	4.12	0.798	source
2.59	1.13	1.86	Childhood outcomes: Opportunity Atlas	Teen birth among women	80,873	Raw interval includes zero	0.429	2.29	−1.43	source
2.93	−0.774	1.08	Environment: EJScreen	Diesel particulate matter	80,873	Loses more than 50 percent of the raw coefficient	11.63	14.63	8.63	source
3.55	1.17	2.36	Environment: EJScreen	Traffic proximity	80,873	Loses more than 50 percent of the raw coefficient	11.86	15.60	8.12	source
3.14	0.801	1.97	Environment: EJScreen	Nitrogen dioxide	80,873	Loses more than 50 percent of the raw coefficient	11.40	12.69	10.12	source
−0.087	−3.84	−1.96	Environment: EJScreen	Fine particles (PM2.5)	80,873	Reverses sign	4.63	6.19	3.07	source
−2.09	−4.28	−3.18	Environment: EJScreen	Ozone	80,873	Reverses sign	3.83	6.30	1.36	source
0.635	−0.145	0.245	Environment: EJScreen	Toxic releases to air (RSEI)	80,873	Raw interval includes zero	0.532	1.75	−0.691	source
4.33	2.86	3.60	Environment: EJScreen	Housing built before 1960	80,873	Keeps 50 percent or more of the raw coefficient	6.67	7.81	5.54	source
1.13	−0.005	0.564	Environment: EJScreen	Proximity to Superfund sites	80,873	Loses more than 50 percent of the raw coefficient	6.88	8.01	5.75	source
1.31	−0.327	0.492	Environment: EJScreen	Proximity to risk management facilities	80,873	Loses more than 50 percent of the raw coefficient	4.95	6.71	3.20	source
3.86	2.20	3.03	Environment: EJScreen	Proximity to hazardous waste facilities	80,873	Loses more than 50 percent of the raw coefficient	13.14	14.40	11.88	source
5.27	3.71	4.49	Health: CDC PLACES	Fair or poor health	80,873	Raw interval includes zero	0.221	1.60	−1.16	source
3.57	2.59	3.08	Health: CDC PLACES	Poor mental health	80,873	Raw interval includes zero	−0.113	1.29	−1.52	source
4.14	1.98	3.06	Health: CDC PLACES	Depression	80,873	Reverses sign	−6.33	−4.16	−8.49	source
2.68	1.22	1.95	Health: CDC PLACES	Diabetes	80,873	Reverses sign	−1.67	−0.121	−3.22	source
1.58	−1.50	0.036	Health: CDC PLACES	Obesity	80,873	Loses more than 50 percent of the raw coefficient	−4.80	−2.59	−7.02	source
4.12	2.56	3.34	Health: CDC PLACES	Current smoking	80,873	Reverses sign	−4.70	−2.81	−6.58	source
4.13	2.44	3.29	Health: CDC PLACES	No leisure-time physical activity	80,873	Reverses sign	−1.88	−0.282	−3.49	source
8.97	6.21	7.59	Health: CDC PLACES	No health insurance	80,873	Keeps 50 percent or more of the raw coefficient	1.77	3.13	0.411	source
0.385	−0.636	−0.125	Health: CDC PLACES	Arthritis	80,873	Loses more than 50 percent of the raw coefficient	−10.37	−8.95	−11.80	source
4.33	2.86	3.60	Keeps 50 percent or more of the raw coefficient	Housing built before 1960	80,873	Keeps 50 percent or more of the raw coefficient	6.67	7.81	5.54	pattern
4.34	2.88	3.61	Keeps 50 percent or more of the raw coefficient	Low household income rank in adulthood	80,873	Keeps 50 percent or more of the raw coefficient	3.04	4.81	1.27	pattern
8.97	6.21	7.59	Keeps 50 percent or more of the raw coefficient	No health insurance	80,873	Keeps 50 percent or more of the raw coefficient	1.77	3.13	0.411	pattern
3.86	2.20	3.03	Loses more than 50 percent of the raw coefficient	Proximity to hazardous waste facilities	80,873	Loses more than 50 percent of the raw coefficient	13.14	14.40	11.88	pattern
3.55	1.17	2.36	Loses more than 50 percent of the raw coefficient	Traffic proximity	80,873	Loses more than 50 percent of the raw coefficient	11.86	15.60	8.12	pattern
2.93	−0.774	1.08	Loses more than 50 percent of the raw coefficient	Diesel particulate matter	80,873	Loses more than 50 percent of the raw coefficient	11.63	14.63	8.63	pattern
3.14	0.801	1.97	Loses more than 50 percent of the raw coefficient	Nitrogen dioxide	80,873	Loses more than 50 percent of the raw coefficient	11.40	12.69	10.12	pattern
1.13	−0.005	0.564	Loses more than 50 percent of the raw coefficient	Proximity to Superfund sites	80,873	Loses more than 50 percent of the raw coefficient	6.88	8.01	5.75	pattern
1.31	−0.327	0.492	Loses more than 50 percent of the raw coefficient	Proximity to risk management facilities	80,873	Loses more than 50 percent of the raw coefficient	4.95	6.71	3.20	pattern
1.28	0.756	1.02	Loses more than 50 percent of the raw coefficient	Incarceration	80,873	Loses more than 50 percent of the raw coefficient	2.46	4.12	0.798	pattern
1.58	−1.50	0.036	Loses more than 50 percent of the raw coefficient	Obesity	80,873	Loses more than 50 percent of the raw coefficient	−4.80	−2.59	−7.02	pattern
0.385	−0.636	−0.125	Loses more than 50 percent of the raw coefficient	Arthritis	80,873	Loses more than 50 percent of the raw coefficient	−10.37	−8.95	−11.80	pattern
−0.087	−3.84	−1.96	Reverses sign	Fine particles (PM2.5)	80,873	Reverses sign	4.63	6.19	3.07	pattern
−2.09	−4.28	−3.18	Reverses sign	Ozone	80,873	Reverses sign	3.83	6.30	1.36	pattern
2.68	1.22	1.95	Reverses sign	Diabetes	80,873	Reverses sign	−1.67	−0.121	−3.22	pattern
4.13	2.44	3.29	Reverses sign	No leisure-time physical activity	80,873	Reverses sign	−1.88	−0.282	−3.49	pattern
2.70	1.26	1.98	Reverses sign	Low individual income rank in adulthood	80,873	Reverses sign	−2.37	−0.695	−4.04	pattern
4.12	2.56	3.34	Reverses sign	Current smoking	80,873	Reverses sign	−4.70	−2.81	−6.58	pattern
5.22	2.58	3.90	Reverses sign	Area Deprivation Index rank	80,873	Reverses sign	−5.86	−3.51	−8.22	pattern
4.14	1.98	3.06	Reverses sign	Depression	80,873	Reverses sign	−6.33	−4.16	−8.49	pattern
0.635	−0.145	0.245	Raw interval includes zero	Toxic releases to air (RSEI)	80,873	Raw interval includes zero	0.532	1.75	−0.691	pattern
2.59	1.13	1.86	Raw interval includes zero	Teen birth among women	80,873	Raw interval includes zero	0.429	2.29	−1.43	pattern
5.27	3.71	4.49	Raw interval includes zero	Fair or poor health	80,873	Raw interval includes zero	0.221	1.60	−1.16	pattern
3.57	2.59	3.08	Raw interval includes zero	Poor mental health	80,873	Raw interval includes zero	−0.113	1.29	−1.52	pattern

Adjusted coefficients are closer to zero than raw ones for 17 of the 24 measures, and they lie above zero for 16. Each row plots two coefficients of the tract Democratic share of the two-party vote on one measure, signed so that a higher value means more hardship; the open dot has no controls and the filled dot holds race and ethnicity shares, the college share, population density and state fixed. Bars are 95% intervals from state clusters.Note: Low household income rank and low individual income rank in adulthood are the Opportunity Atlas ranks with the sign reversed, so that a higher value means worse outcomes.Note: Each measure is cut at its 1st and 99th percentiles across tracts before it is standardized with the vote-weighted mean and standard deviation.Note: The grouping rule was set after a first run of the estimates, and the methods note gives it.Note: Votes are estimates allocated from precincts to tracts, not reported counts.Source: Precinct returns allocated to 2020 census tracts (Fekrazad 2025) with the ADI, Opportunity Atlas, EJScreen, PLACES and ACS values of the project tract atlas; 2020 presidential vote; weighted by tract votes

Two Opportunity Atlas measures of income in adulthood disagree. Low household income rank has a raw coefficient of 3.0 (95% CI 1.3 to 4.8) and low individual income rank has −2.4 (95% CI −4.0 to −0.7), and the data here cannot say why the signs differ. Before controls, the answer to which kinds of hardship go with the vote depends on the measure: environmental burdens and some childhood outcomes go with a more Democratic vote, and deprivation and illness go with a more Republican one.

With controls, most coefficients are positive and smaller

With race and ethnicity, education, density and state held fixed, 16 measures have an interval above zero, 2 have one below zero and 6 include zero. The two negative ones are ozone, −3.2 (95% CI −4.3 to −2.1), and fine particles, −2.0 (95% CI −3.8 to −0.1). The largest coefficient belongs to lack of health insurance, 7.6 (95% CI 6.2 to 9.0). The next four are fair or poor health, 4.5, the Area Deprivation Index, 3.9, low household income rank, 3.6 and housing built before 1960, 3.6, and their intervals overlap one another, so the data do not rank them. Among alike tracts in one state, the ones with more hardship voted more Democratic on those 16 measures. Adjusted coefficients are closer to zero than raw ones for 17 of the 24 measures.

The measures sort into groups by what the controls do. The group of 3 that keep at least 50% of the raw coefficient is low household income rank, whose adjusted coefficient is 3.6 (95% CI 2.9 to 4.3), housing built before 1960, 3.6 (95% CI 2.9 to 4.3), and lack of health insurance, whose adjusted coefficient is larger than its raw one. The group of 9 that lose more than 50% of it, or whose adjusted interval includes zero, is made up of 6 environmental indicators, incarceration, obesity and arthritis. The group of 8 that reverse sign, with an adjusted interval on the other side of zero, is the Area Deprivation Index, low individual income rank, ozone, fine particles, depression, diabetes, current smoking and no leisure-time physical activity. The remaining 4 have a raw interval that includes zero, and teen birth, fair or poor health and poor mental health are above zero once the controls are in. Both the 50% line and the choice of the 9 PLACES measures, made after a first look at 19 of the 40, followed a first run of the estimates. The last section shows how firm the groups are.

Race and education move the coefficients most

Figure 3 adds the controls one block at a time. Across measures the median absolute change in the coefficient is 3.5 points when race and ethnicity enter, 3.8 when education enters, 1.0 when density enters and 1.6 when state fixed effects enter. The size of a step depends on the order, so the steps do not add up to a share of the change. Leaving out one block from the common specification gives a median change of 3.2 points for race and ethnicity, 3.0 for education and 0.9 for density.

Figure 3

Race and education move the coefficients most, and density and state fixed effects move them less

Coefficient of the tract Democratic two-party share on each standardized measure, points per standard deviation, as controls are added from left to right

−12.00.0+12.0RepublicanDemocratic

Heatmap with one row for each of the tract measures and five columns for specifications from no controls to the common specification. Cells shift from large positive or negative values in the left columns toward values near zero or modestly positive in the right columns.
Measure	No controls	Race and ethnicity	Plus education	Plus density	Plus state effects
Area Deprivation Index rank	−5.9, interval −8.2 to −3.5, n = 80,873	−7.2, interval −8.9 to −5.6, n = 80,873	−0.7, interval −3.3 to +1.9, n = 80,873	0.0, interval −2.3 to +2.3, n = 80,873	+3.9, interval +2.6 to +5.2, n = 80,873
Low household income rank in adulthood	+3.0, interval +1.3 to +4.8, n = 80,873	−1.2, interval −2.9 to +0.4, n = 80,873	+2.4, interval +0.9 to +4.0, n = 80,873	+2.1, interval +0.6 to +3.5, n = 80,873	+3.6, interval +2.9 to +4.3, n = 80,873
Low individual income rank in adulthood	−2.4, interval −4.0 to −0.7, n = 80,873	−3.7, interval −5.1 to −2.3, n = 80,873	+0.3, interval −0.9 to +1.5, n = 80,873	+0.3, interval −0.9 to +1.5, n = 80,873	+2.0, interval +1.3 to +2.7, n = 80,873
Incarceration	+2.5, interval +0.8 to +4.1, n = 80,873	−0.6, interval −1.4 to +0.2, n = 80,873	+0.4, interval −0.3 to +1.0, n = 80,873	+0.2, interval −0.3 to +0.7, n = 80,873	+1.0, interval +0.8 to +1.3, n = 80,873
Teen birth among women	+0.4, interval −1.4 to +2.3, n = 80,873	−5.4, interval −7.0 to −3.8, n = 80,873	−1.2, interval −2.7 to +0.3, n = 80,873	−0.6, interval −2.0 to +0.7, n = 80,873	+1.9, interval +1.1 to +2.6, n = 80,873
Diesel particulate matter	+11.6, interval +8.6 to +14.6, n = 80,873	+7.8, interval +5.2 to +10.4, n = 80,873	+5.3, interval +2.7 to +7.8, n = 80,873	+2.5, interval +0.2 to +4.8, n = 80,873	+1.1, interval −0.8 to +2.9, n = 80,873
Traffic proximity	+11.9, interval +8.1 to +15.6, n = 80,873	+8.6, interval +5.9 to +11.3, n = 80,873	+6.5, interval +4.1 to +9.0, n = 80,873	+4.7, interval +2.8 to +6.6, n = 80,873	+2.4, interval +1.2 to +3.5, n = 80,873
Nitrogen dioxide	+11.4, interval +10.1 to +12.7, n = 80,873	+7.6, interval +6.4 to +8.9, n = 80,873	+5.3, interval +3.9 to +6.7, n = 80,873	+2.2, interval +0.1 to +4.3, n = 80,873	+2.0, interval +0.8 to +3.1, n = 80,873
Fine particles (PM2.5)	+4.6, interval +3.1 to +6.2, n = 80,873	+0.7, interval −1.1 to +2.5, n = 80,873	+0.4, interval −1.4 to +2.2, n = 80,873	−0.4, interval −2.2 to +1.4, n = 80,873	−2.0, interval −3.8 to −0.1, n = 80,873
Ozone	+3.8, interval +1.4 to +6.3, n = 80,873	+1.4, interval −1.8 to +4.6, n = 80,873	+0.5, interval −2.2 to +3.2, n = 80,873	−0.6, interval −2.7 to +1.5, n = 80,873	−3.2, interval −4.3 to −2.1, n = 80,873
Toxic releases to air (RSEI)	+0.5, interval −0.7 to +1.8, n = 80,873	−0.2, interval −1.2 to +0.8, n = 80,873	−0.2, interval −1.2 to +0.8, n = 80,873	−0.6, interval −1.4 to +0.2, n = 80,873	+0.2, interval −0.1 to +0.6, n = 80,873
Housing built before 1960	+6.7, interval +5.5 to +7.8, n = 80,873	+6.0, interval +5.0 to +7.0, n = 80,873	+6.3, interval +5.2 to +7.4, n = 80,873	+5.2, interval +4.1 to +6.3, n = 80,873	+3.6, interval +2.9 to +4.3, n = 80,873
Proximity to Superfund sites	+6.9, interval +5.7 to +8.0, n = 80,873	+4.5, interval +3.3 to +5.6, n = 80,873	+3.4, interval +2.2 to +4.5, n = 80,873	+2.1, interval +1.2 to +3.1, n = 80,873	+0.6, interval 0.0 to +1.1, n = 80,873
Proximity to risk management facilities	+5.0, interval +3.2 to +6.7, n = 80,873	+2.0, interval +0.6 to +3.4, n = 80,873	+1.8, interval +0.6 to +2.9, n = 80,873	+0.6, interval −0.4 to +1.7, n = 80,873	+0.5, interval −0.3 to +1.3, n = 80,873
Proximity to hazardous waste facilities	+13.1, interval +11.9 to +14.4, n = 80,873	+10.0, interval +8.9 to +11.1, n = 80,873	+7.7, interval +6.4 to +8.9, n = 80,873	+5.8, interval +4.6 to +7.1, n = 80,873	+3.0, interval +2.2 to +3.9, n = 80,873
Fair or poor health	+0.2, interval −1.2 to +1.6, n = 80,873	−7.1, interval −8.7 to −5.6, n = 80,873	+1.7, interval +0.2 to +3.2, n = 80,873	+2.3, interval +1.0 to +3.6, n = 80,873	+4.5, interval +3.7 to +5.3, n = 80,873
Poor mental health	−0.1, interval −1.5 to +1.3, n = 80,873	−4.3, interval −5.7 to −2.9, n = 80,873	+3.1, interval +1.5 to +4.7, n = 80,873	+1.6, interval +0.1 to +3.2, n = 80,873	+3.1, interval +2.6 to +3.6, n = 80,873
Depression	−6.3, interval −8.5 to −4.2, n = 80,873	−1.7, interval −3.8 to +0.4, n = 80,873	+2.2, interval −0.1 to +4.5, n = 80,873	+1.5, interval −0.6 to +3.6, n = 80,873	+3.1, interval +2.0 to +4.1, n = 80,873
Diabetes	−1.7, interval −3.2 to −0.1, n = 80,873	−8.1, interval −9.6 to −6.6, n = 80,873	−3.0, interval −4.2 to −1.8, n = 80,873	−1.1, interval −2.3 to +0.1, n = 80,873	+1.9, interval +1.2 to +2.7, n = 80,873
Obesity	−4.8, interval −7.0 to −2.6, n = 80,873	−9.4, interval −11.4 to −7.4, n = 80,873	−3.3, interval −6.0 to −0.6, n = 80,873	−1.9, interval −4.3 to +0.6, n = 80,873	0.0, interval −1.5 to +1.6, n = 80,873
Current smoking	−4.7, interval −6.6 to −2.8, n = 80,873	−7.4, interval −8.9 to −5.8, n = 80,873	+1.1, interval −0.6 to +2.7, n = 80,873	+1.3, interval 0.0 to +2.7, n = 80,873	+3.3, interval +2.6 to +4.1, n = 80,873
No leisure-time physical activity	−1.9, interval −3.5 to −0.3, n = 80,873	−8.4, interval −9.7 to −7.1, n = 80,873	−1.3, interval −3.2 to +0.6, n = 80,873	−0.8, interval −2.3 to +0.7, n = 80,873	+3.3, interval +2.4 to +4.1, n = 80,873
No health insurance	+1.8, interval +0.4 to +3.1, n = 80,873	−10.2, interval −12.8 to −7.7, n = 80,873	−4.9, interval −7.0 to −2.8, n = 80,873	−4.0, interval −5.5 to −2.4, n = 80,873	+7.6, interval +6.2 to +9.0, n = 80,873
Arthritis	−10.4, interval −11.8 to −8.9, n = 80,873	−7.8, interval −9.6 to −6.1, n = 80,873	−2.7, interval −3.9 to −1.5, n = 80,873	−0.8, interval −2.4 to +0.8, n = 80,873	−0.1, interval −0.6 to +0.4, n = 80,873

Show data

Data for Race and education move the coefficients most, and density and state fixed effects move them less
family	hi	lo	measure	n	specification	value
Deprivation: Area Deprivation Index	−3.51	−8.22	Area Deprivation Index rank	80,873	No controls	−5.86
Deprivation: Area Deprivation Index	−5.58	−8.87	Area Deprivation Index rank	80,873	Race and ethnicity	−7.22
Deprivation: Area Deprivation Index	1.91	−3.33	Area Deprivation Index rank	80,873	Plus education	−0.712
Deprivation: Area Deprivation Index	2.34	−2.27	Area Deprivation Index rank	80,873	Plus density	0.037
Deprivation: Area Deprivation Index	5.22	2.58	Area Deprivation Index rank	80,873	Plus state effects	3.90
Childhood outcomes: Opportunity Atlas	4.81	1.27	Low household income rank in adulthood	80,873	No controls	3.04
Childhood outcomes: Opportunity Atlas	0.415	−2.91	Low household income rank in adulthood	80,873	Race and ethnicity	−1.25
Childhood outcomes: Opportunity Atlas	3.98	0.906	Low household income rank in adulthood	80,873	Plus education	2.45
Childhood outcomes: Opportunity Atlas	3.52	0.596	Low household income rank in adulthood	80,873	Plus density	2.06
Childhood outcomes: Opportunity Atlas	4.34	2.88	Low household income rank in adulthood	80,873	Plus state effects	3.61
Childhood outcomes: Opportunity Atlas	−0.695	−4.04	Low individual income rank in adulthood	80,873	No controls	−2.37
Childhood outcomes: Opportunity Atlas	−2.29	−5.08	Low individual income rank in adulthood	80,873	Race and ethnicity	−3.68
Childhood outcomes: Opportunity Atlas	1.51	−0.940	Low individual income rank in adulthood	80,873	Plus education	0.286
Childhood outcomes: Opportunity Atlas	1.53	−0.850	Low individual income rank in adulthood	80,873	Plus density	0.340
Childhood outcomes: Opportunity Atlas	2.70	1.26	Low individual income rank in adulthood	80,873	Plus state effects	1.98
Childhood outcomes: Opportunity Atlas	4.12	0.798	Incarceration	80,873	No controls	2.46
Childhood outcomes: Opportunity Atlas	0.212	−1.43	Incarceration	80,873	Race and ethnicity	−0.610
Childhood outcomes: Opportunity Atlas	1.03	−0.316	Incarceration	80,873	Plus education	0.355
Childhood outcomes: Opportunity Atlas	0.744	−0.326	Incarceration	80,873	Plus density	0.209
Childhood outcomes: Opportunity Atlas	1.28	0.756	Incarceration	80,873	Plus state effects	1.02
Childhood outcomes: Opportunity Atlas	2.29	−1.43	Teen birth among women	80,873	No controls	0.429
Childhood outcomes: Opportunity Atlas	−3.78	−6.98	Teen birth among women	80,873	Race and ethnicity	−5.38
Childhood outcomes: Opportunity Atlas	0.326	−2.69	Teen birth among women	80,873	Plus education	−1.18
Childhood outcomes: Opportunity Atlas	0.688	−1.96	Teen birth among women	80,873	Plus density	−0.636
Childhood outcomes: Opportunity Atlas	2.59	1.13	Teen birth among women	80,873	Plus state effects	1.86
Environment: EJScreen	14.63	8.63	Diesel particulate matter	80,873	No controls	11.63
Environment: EJScreen	10.40	5.16	Diesel particulate matter	80,873	Race and ethnicity	7.78
Environment: EJScreen	7.82	2.73	Diesel particulate matter	80,873	Plus education	5.27
Environment: EJScreen	4.79	0.207	Diesel particulate matter	80,873	Plus density	2.50
Environment: EJScreen	2.93	−0.774	Diesel particulate matter	80,873	Plus state effects	1.08
Environment: EJScreen	15.60	8.12	Traffic proximity	80,873	No controls	11.86
Environment: EJScreen	11.35	5.93	Traffic proximity	80,873	Race and ethnicity	8.64
Environment: EJScreen	8.99	4.06	Traffic proximity	80,873	Plus education	6.52
Environment: EJScreen	6.63	2.79	Traffic proximity	80,873	Plus density	4.71
Environment: EJScreen	3.55	1.17	Traffic proximity	80,873	Plus state effects	2.36
Environment: EJScreen	12.69	10.12	Nitrogen dioxide	80,873	No controls	11.40
Environment: EJScreen	8.87	6.43	Nitrogen dioxide	80,873	Race and ethnicity	7.65
Environment: EJScreen	6.67	3.94	Nitrogen dioxide	80,873	Plus education	5.31
Environment: EJScreen	4.34	0.115	Nitrogen dioxide	80,873	Plus density	2.23
Environment: EJScreen	3.14	0.801	Nitrogen dioxide	80,873	Plus state effects	1.97
Environment: EJScreen	6.19	3.07	Fine particles (PM2.5)	80,873	No controls	4.63
Environment: EJScreen	2.47	−1.15	Fine particles (PM2.5)	80,873	Race and ethnicity	0.662
Environment: EJScreen	2.18	−1.45	Fine particles (PM2.5)	80,873	Plus education	0.366
Environment: EJScreen	1.36	−2.20	Fine particles (PM2.5)	80,873	Plus density	−0.419
Environment: EJScreen	−0.087	−3.84	Fine particles (PM2.5)	80,873	Plus state effects	−1.96
Environment: EJScreen	6.30	1.36	Ozone	80,873	No controls	3.83
Environment: EJScreen	4.61	−1.76	Ozone	80,873	Race and ethnicity	1.43
Environment: EJScreen	3.20	−2.19	Ozone	80,873	Plus education	0.502
Environment: EJScreen	1.52	−2.67	Ozone	80,873	Plus density	−0.577
Environment: EJScreen	−2.09	−4.28	Ozone	80,873	Plus state effects	−3.18
Environment: EJScreen	1.75	−0.691	Toxic releases to air (RSEI)	80,873	No controls	0.532
Environment: EJScreen	0.834	−1.21	Toxic releases to air (RSEI)	80,873	Race and ethnicity	−0.188
Environment: EJScreen	0.752	−1.22	Toxic releases to air (RSEI)	80,873	Plus education	−0.235
Environment: EJScreen	0.163	−1.38	Toxic releases to air (RSEI)	80,873	Plus density	−0.610
Environment: EJScreen	0.635	−0.145	Toxic releases to air (RSEI)	80,873	Plus state effects	0.245
Environment: EJScreen	7.81	5.54	Housing built before 1960	80,873	No controls	6.67
Environment: EJScreen	6.98	4.95	Housing built before 1960	80,873	Race and ethnicity	5.96
Environment: EJScreen	7.40	5.21	Housing built before 1960	80,873	Plus education	6.31
Environment: EJScreen	6.32	4.14	Housing built before 1960	80,873	Plus density	5.23
Environment: EJScreen	4.33	2.86	Housing built before 1960	80,873	Plus state effects	3.60
Environment: EJScreen	8.01	5.75	Proximity to Superfund sites	80,873	No controls	6.88
Environment: EJScreen	5.60	3.34	Proximity to Superfund sites	80,873	Race and ethnicity	4.47
Environment: EJScreen	4.51	2.22	Proximity to Superfund sites	80,873	Plus education	3.36
Environment: EJScreen	3.08	1.18	Proximity to Superfund sites	80,873	Plus density	2.13
Environment: EJScreen	1.13	−0.005	Proximity to Superfund sites	80,873	Plus state effects	0.564
Environment: EJScreen	6.71	3.20	Proximity to risk management facilities	80,873	No controls	4.95
Environment: EJScreen	3.38	0.564	Proximity to risk management facilities	80,873	Race and ethnicity	1.97
Environment: EJScreen	2.88	0.640	Proximity to risk management facilities	80,873	Plus education	1.76
Environment: EJScreen	1.66	−0.431	Proximity to risk management facilities	80,873	Plus density	0.617
Environment: EJScreen	1.31	−0.327	Proximity to risk management facilities	80,873	Plus state effects	0.492
Environment: EJScreen	14.40	11.88	Proximity to hazardous waste facilities	80,873	No controls	13.14
Environment: EJScreen	11.06	8.94	Proximity to hazardous waste facilities	80,873	Race and ethnicity	10.00
Environment: EJScreen	8.90	6.43	Proximity to hazardous waste facilities	80,873	Plus education	7.67
Environment: EJScreen	7.10	4.56	Proximity to hazardous waste facilities	80,873	Plus density	5.83
Environment: EJScreen	3.86	2.20	Proximity to hazardous waste facilities	80,873	Plus state effects	3.03
Health: CDC PLACES	1.60	−1.16	Fair or poor health	80,873	No controls	0.221
Health: CDC PLACES	−5.58	−8.68	Fair or poor health	80,873	Race and ethnicity	−7.13
Health: CDC PLACES	3.23	0.169	Fair or poor health	80,873	Plus education	1.70
Health: CDC PLACES	3.63	1.01	Fair or poor health	80,873	Plus density	2.32
Health: CDC PLACES	5.27	3.71	Fair or poor health	80,873	Plus state effects	4.49
Health: CDC PLACES	1.29	−1.52	Poor mental health	80,873	No controls	−0.113
Health: CDC PLACES	−2.86	−5.70	Poor mental health	80,873	Race and ethnicity	−4.28
Health: CDC PLACES	4.70	1.53	Poor mental health	80,873	Plus education	3.12
Health: CDC PLACES	3.15	0.123	Poor mental health	80,873	Plus density	1.64
Health: CDC PLACES	3.57	2.59	Poor mental health	80,873	Plus state effects	3.08
Health: CDC PLACES	−4.16	−8.49	Depression	80,873	No controls	−6.33
Health: CDC PLACES	0.363	−3.76	Depression	80,873	Race and ethnicity	−1.70
Health: CDC PLACES	4.53	−0.085	Depression	80,873	Plus education	2.22
Health: CDC PLACES	3.63	−0.600	Depression	80,873	Plus density	1.51
Health: CDC PLACES	4.14	1.98	Depression	80,873	Plus state effects	3.06
Health: CDC PLACES	−0.121	−3.22	Diabetes	80,873	No controls	−1.67
Health: CDC PLACES	−6.62	−9.63	Diabetes	80,873	Race and ethnicity	−8.13
Health: CDC PLACES	−1.84	−4.23	Diabetes	80,873	Plus education	−3.04
Health: CDC PLACES	0.097	−2.35	Diabetes	80,873	Plus density	−1.13
Health: CDC PLACES	2.68	1.22	Diabetes	80,873	Plus state effects	1.95
Health: CDC PLACES	−2.59	−7.02	Obesity	80,873	No controls	−4.80
Health: CDC PLACES	−7.44	−11.41	Obesity	80,873	Race and ethnicity	−9.43
Health: CDC PLACES	−0.618	−6.02	Obesity	80,873	Plus education	−3.32
Health: CDC PLACES	0.588	−4.34	Obesity	80,873	Plus density	−1.88
Health: CDC PLACES	1.58	−1.50	Obesity	80,873	Plus state effects	0.036
Health: CDC PLACES	−2.81	−6.58	Current smoking	80,873	No controls	−4.70
Health: CDC PLACES	−5.84	−8.87	Current smoking	80,873	Race and ethnicity	−7.35
Health: CDC PLACES	2.74	−0.592	Current smoking	80,873	Plus education	1.07
Health: CDC PLACES	2.74	−0.049	Current smoking	80,873	Plus density	1.35
Health: CDC PLACES	4.12	2.56	Current smoking	80,873	Plus state effects	3.34
Health: CDC PLACES	−0.282	−3.49	No leisure-time physical activity	80,873	No controls	−1.88
Health: CDC PLACES	−7.07	−9.72	No leisure-time physical activity	80,873	Race and ethnicity	−8.40
Health: CDC PLACES	0.640	−3.25	No leisure-time physical activity	80,873	Plus education	−1.30
Health: CDC PLACES	0.744	−2.34	No leisure-time physical activity	80,873	Plus density	−0.799
Health: CDC PLACES	4.13	2.44	No leisure-time physical activity	80,873	Plus state effects	3.29
Health: CDC PLACES	3.13	0.411	No health insurance	80,873	No controls	1.77
Health: CDC PLACES	−7.70	−12.78	No health insurance	80,873	Race and ethnicity	−10.24
Health: CDC PLACES	−2.84	−6.99	No health insurance	80,873	Plus education	−4.91
Health: CDC PLACES	−2.42	−5.51	No health insurance	80,873	Plus density	−3.96
Health: CDC PLACES	8.97	6.21	No health insurance	80,873	Plus state effects	7.59
Health: CDC PLACES	−8.95	−11.80	Arthritis	80,873	No controls	−10.37
Health: CDC PLACES	−6.06	−9.56	Arthritis	80,873	Race and ethnicity	−7.81
Health: CDC PLACES	−1.48	−3.91	Arthritis	80,873	Plus education	−2.69
Health: CDC PLACES	0.759	−2.40	Arthritis	80,873	Plus density	−0.818
Health: CDC PLACES	0.385	−0.636	Arthritis	80,873	Plus state effects	−0.125

The largest changes come when race and ethnicity enter and when education enters: the median absolute change across measures is 3.5 points at the first step and 3.8 at the second, against 1.0 for density and 1.6 for state fixed effects. Each cell is a separate regression that adds the control named at the head of its column to the ones on its left. Blue cells mean a more Democratic vote where the measure is higher and red cells a more Republican vote.Note: The order of the columns is one of several; the size of each step depends on the order, so the steps are not shares of a total.Note: A coefficient of zero is white, and values beyond the ends of the color scale take the color of the end.Source: Precinct returns allocated to 2020 census tracts (Fekrazad 2025) with the ADI, Opportunity Atlas, EJScreen, PLACES and ACS values of the project tract atlas; 2020 presidential vote; weighted by tract votes

The Area Deprivation Index shows how a control can matter by construction. It is built from income, education, employment and housing measures [4], so holding education fixed removes part of what it measures. Its coefficient is −5.9 (95% CI −8.2 to −3.5) before controls and −7.2 (95% CI −8.9 to −5.6) with race and ethnicity held fixed. With race, education and density held fixed but no state effects it is 0.0 (95% CI −2.3 to 2.3), and with state effects it is 3.9 (95% CI 2.6 to 5.2). In the common specification without the college share it is −3.6 (95% CI −5.2 to −2.1).

Environmental burdens lose most of their association

Among the 10 environmental indicators, all have a raw coefficient above zero and 9 have an interval above zero. 6 of them keep at most 23% of the raw coefficient after the controls. Diesel particulate matter goes from 11.6 (95% CI 8.6 to 14.6) to 1.1 (95% CI −0.8 to 2.9), traffic proximity from 11.9 (95% CI 8.1 to 15.6) to 2.4 (95% CI 1.2 to 3.5) and nitrogen dioxide from 11.4 (95% CI 10.1 to 12.7) to 2.0 (95% CI 0.8 to 3.1). The proximity indicators for hazardous waste facilities, Superfund sites and risk management facilities go from 13.1 (95% CI 11.9 to 14.4), 6.9 (95% CI 5.7 to 8.0) and 5.0 (95% CI 3.2 to 6.7) to 3.0 (95% CI 2.2 to 3.9), 0.6 (95% CI 0.0 to 1.1) and 0.5 (95% CI −0.3 to 1.3).

A state bootstrap puts the share of the raw coefficient that remains at 9% (95% CI −1% to 28%) for diesel, 20% (95% CI 14% to 28%) for traffic, 17% (95% CI 7% to 27%) for nitrogen dioxide and 23% (95% CI 17% to 28%) for hazardous waste proximity, and the 6 indicators keep between 8% and 23% of their raw coefficients. Race, education and density explain 47% of the within-state variation in diesel exposure, 59% in nitrogen dioxide and 51% in hazardous waste proximity, so much of the within-state difference in these exposures follows demographics. The other environmental indicators behave differently. Housing built before 1960 keeps 54% (95% CI 42% to 71%) of its raw coefficient and only 9% of its variation is explained by the controls. Ozone and fine particles reverse sign, and toxic releases to air have a raw interval that includes zero. Chapter 13.6, Air pollution and the vote takes up four of these exposures in detail.

These results depend on the level of the fixed effects. With county fixed effects in place of state fixed effects, which compares tracts inside one county, diesel is 4.1 (95% CI 2.7 to 5.6), which is 36% of its raw coefficient against 9% with state effects. Obesity goes from 0.0 (95% CI −1.5 to 1.6) to 6.7 (95% CI 5.7 to 7.7), fine particles from −2.0 (95% CI −3.8 to −0.1) to 0.2 (95% CI −1.5 to 1.9) and ozone from −3.2 (95% CI −4.3 to −2.1) to −1.1 (95% CI −2.1 to −0.2). Whether a measure behaves as a demographic proxy depends on how local the comparison is.

The CDC's PLACES estimates are the output of a multilevel model fitted to survey answers, with the respondent's age, sex, race and ethnicity and education, the county poverty rate and state and county effects as inputs. The fitted model is then applied to each area's population by age, race and ethnicity and sex, and because counts by education are not available for small areas, a bootstrap assigns education to each group [6]. A tract's own college share is therefore not an input, and a tract's estimate differs from another's through its age, sex and racial mix and its county. The atlas holds crude prevalence and not age-adjusted prevalence, so a tract with older residents has higher prevalence of an age-related condition: arthritis correlates +0.62 with the share of older residents. Across tracts, race and ethnicity shares, the shares of older residents and of children, and county effects share between 66% and 89% of the variance of the 9 PLACES measures here. County effects alone share between 33% and 78%, and race and age alone between 25% and 63%. Race, education and density with state effects share between 40% and 80% of their within-state variance (Figure 4).

Figure 4

Race, education and density share 40% to 80% of the within-state variance of the PLACES health measures and 2% of the variance of toxic releases

Share of the vote-weighted variance of each measure shared with the controls, in percent: within states for race and ethnicity shares, the college share and density; across all tracts for race, age and county effects in the second view

Deprivation
Childhood outcomes
Environment
Health
Social capital
Tax returns

Show data

Data for Race, education and density share 40% to 80% of the within-state variance of the PLACES health measures and 2% of the variance of toxic releases
family	label	value	view
Deprivation	Area Deprivation Index rank	59.24	common
Childhood outcomes	Low household income rank in adulthood	38.80	common
Childhood outcomes	Low individual income rank in adulthood	30.35	common
Childhood outcomes	Incarceration	13.36	common
Childhood outcomes	Teen birth among women	41.91	common
Environment	Diesel particulate matter	46.88	common
Environment	Traffic proximity	36.42	common
Environment	Nitrogen dioxide	58.75	common
Environment	Fine particles (PM2.5)	12.93	common
Environment	Ozone	9.23	common
Environment	Toxic releases to air (RSEI)	2.09	common
Environment	Housing built before 1960	9.11	common
Environment	Proximity to Superfund sites	10.08	common
Environment	Proximity to risk management facilities	15.12	common
Environment	Proximity to hazardous waste facilities	50.74	common
Health	Fair or poor health	70.55	common
Health	Poor mental health	59.95	common
Health	Depression	40.34	common
Health	Diabetes	56.23	common
Health	Obesity	71.20	common
Health	Current smoking	71.35	common
Health	No leisure-time physical activity	73.11	common
Health	No health insurance	80.22	common
Health	Arthritis	50.24	common
Social capital	Economic connectedness (ZIP code areas)	71.28	common
Social capital	Economic connectedness among group members (ZIP code areas)	68.35	common
Social capital	Friend clustering (ZIP code areas)	50.74	common
Social capital	Friend support ratio (ZIP code areas)	42.22	common
Social capital	Volunteering rate (ZIP code areas)	35.73	common
Social capital	Civic organizations per capita (ZIP code areas)	9.28	common
Tax returns	Earned income credit share of returns (ZIP code areas)	79.93	common
Tax returns	Unemployment compensation share of returns (ZIP code areas)	23.71	common
Tax returns	Low mean adjusted gross income per return (ZIP code areas)	53.73	common
Health	Fair or poor health	69.26	places
Health	Poor mental health	67.18	places
Health	Depression	87.07	places
Health	Diabetes	76.39	places
Health	Obesity	89.39	places
Health	Current smoking	66.46	places
Health	No leisure-time physical activity	72.75	places
Health	No health insurance	84.86	places
Health	Arthritis	85.16	places

Most PLACES measures and the Area Deprivation Index sit toward the top, and toxic releases, older housing and the proximity indicators sit at the bottom. In the first view each dot is the R-squared of a regression of one measure on race and ethnicity shares, the college share and density with state fixed effects, as a share of the variance within states. The second view covers the PLACES measures only: it leaves out education, density and state effects, adds the shares of older residents and of children and county fixed effects, and takes the share of the total variance across tracts.Note: R-squared values describe the full sample and carry no interval.Note: ZIP code rows use the ZIP code sample that has every Social Capital Atlas and IRS value.Note: Each measure is cut at its 1st and 99th percentiles before the variance is taken, as in the coefficient figures.Source: Precinct returns allocated to 2020 census tracts and ZIP code tabulation areas (Fekrazad 2025) with the project atlases; 2020 presidential vote; weighted by votes in each unit

Before controls the PLACES coefficients are near zero or negative with one exception: 0.2 (95% CI −1.2 to 1.6) for fair or poor health, −0.1 (95% CI −1.5 to 1.3) for poor mental health and −6.3 (95% CI −8.5 to −4.2) for depression, against 1.8 (95% CI 0.4 to 3.1) for lack of health insurance. With the controls in, 7 of the 9 are above zero, from 1.9 (95% CI 1.2 to 2.7) for diabetes to 7.6 (95% CI 6.2 to 9.0) for lack of health insurance. What is left after the controls includes the age and sex structure, which the controls omit, and each county's own level, which are the other inputs of the model. The coefficients are not evidence that illness goes with the vote apart from who lives in a tract. Chapter 13.4, Health outcomes and the vote (planned) is planned to cover the full set of PLACES outcomes.

Deprivation reverses sign and low income rank holds

The Area Deprivation Index changes from −5.9 (95% CI −8.2 to −3.5) to 3.9 (95% CI 2.6 to 5.2), and the share of its variation within states that the controls explain is 59%. Among the Opportunity Atlas measures, low household income rank keeps 119% (95% CI 70% to 234%) of its raw coefficient. Incarceration goes from 2.5 (95% CI 0.8 to 4.1) to 1.0 (95% CI 0.8 to 1.3), a share of 41% (95% CI 21% to 98%), and race, education and density explain only 13% of its variation. Low individual income rank goes from −2.4 (95% CI −4.0 to −0.7) to 2.0 (95% CI 1.3 to 2.7), and teen birth from 0.4 (95% CI −1.4 to 2.3) to 1.9 (95% CI 1.1 to 2.6). These estimates describe children born between 1978 and 1983 who grew up in a tract. Chapter 13.3, Deprivation and mobility (planned) is planned to ask how deprivation and mobility relate to the vote with race held constant.

Social capital and tax returns exist for ZIP code areas only. The ZIP-only sample is 18,039 ZIP code tabulation areas carrying 149.4 million of the 157.2 million votes in ZIP code areas with more than 100 votes. The Social Capital Atlas lacks values for 11,138 areas that otherwise qualify, which carry 7.7 million votes and voted 40% Democratic, against 53% in the sample, so the ZIP-only estimates describe the areas where the atlas has values.

Figure 5

Friend clustering, volunteering and the unemployment compensation share change sign once the controls enter, while the earned income credit share keeps its raw sign

Democratic share of the 2020 two-party presidential vote, points per standard deviation of each measure, ZIP code areas, raw and with race and ethnicity, education, density and state held fixed

No controls
Common specification

Show data

Data for Friend clustering, volunteering and the unemployment compensation share change sign once the controls enter, while the earned income credit share keeps its raw sign
adj_hi	adj_lo	adjusted	group	label	n	raw	raw_hi	raw_lo
1.26	−3.09	−0.914	Social capital: Social Capital Atlas (higher means more connection)	Economic connectedness	18,039	0.394	2.28	−1.49
−0.018	−3.67	−1.84	Social capital: Social Capital Atlas (higher means more connection)	Economic connectedness among group members	18,039	1.33	3.03	−0.368
2.11	0.322	1.22	Social capital: Social Capital Atlas (higher means more connection)	Friend clustering	18,039	−9.13	−7.06	−11.21
−1.06	−2.97	−2.01	Social capital: Social Capital Atlas (higher means more connection)	Friend support ratio	18,039	−9.74	−7.81	−11.67
4.59	2.86	3.72	Social capital: Social Capital Atlas (higher means more connection)	Volunteering rate	18,039	−3.05	−1.49	−4.62
2.72	1.79	2.26	Social capital: Social Capital Atlas (higher means more connection)	Civic organizations per capita	18,039	−0.433	0.445	−1.31
3.03	0.129	1.58	Tax returns: IRS ZIP code data	Earned income credit share of returns	18,039	2.03	3.83	0.227
−0.187	−2.28	−1.23	Tax returns: IRS ZIP code data	Unemployment compensation share of returns	18,039	4.80	6.37	3.23
−0.350	−2.96	−1.65	Tax returns: IRS ZIP code data	Low mean adjusted gross income per return	18,039	−2.67	−1.10	−4.25

Of the 9 ZIP-only measures, 4 have an adjusted interval above zero and 4 have one below zero. Each row plots two coefficients, the open dot with no controls and the filled dot with the common specification, with 95% intervals from state clusters. Social capital measures keep their published direction, so a higher value means more connection; tax-return measures are signed so that a higher value means more hardship.Note: Low mean adjusted gross income per return is the mean with its sign reversed, so that a higher value means lower income.Note: Each measure is cut at its 1st and 99th percentiles across ZIP code areas before it is standardized.Note: Votes are estimates allocated from precincts to ZIP code tabulation areas, not reported counts.Source: Precinct returns allocated to ZIP code tabulation areas (Fekrazad 2025) with the ADI, PLACES, Social Capital Atlas, IRS and ACS values of the project ZIP code atlas; 2020 presidential vote; weighted by ZIP code area votes

Economic connectedness, which Chetty and coauthors define as the share of friends with high socioeconomic status among people with low status [7], has a raw coefficient of 0.4 (95% CI −1.5 to 2.3) and an adjusted one of −0.9 (95% CI −3.1 to 1.3), both with intervals that include zero. The two cohesiveness measures (Figure 5) are strongly negative before controls, −9.1 (95% CI −11.2 to −7.1) for friend clustering and −9.7 (95% CI −11.7 to −7.8) for the friend support ratio. With controls, clustering is 1.2 (95% CI 0.3 to 2.1) and the support ratio −2.0 (95% CI −3.0 to −1.1). The volunteering rate goes from −3.1 (95% CI −4.6 to −1.5) to 3.7 (95% CI 2.9 to 4.6). Chapter 13.5, Social capital (planned) is planned to take up these measures and the 2016 to 2020 swing.

The earned income credit share of tax returns keeps its sign, 2.0 (95% CI 0.2 to 3.8) before controls and 1.6 (95% CI 0.1 to 3.0) after. Low mean income per return has the other sign, −2.7 (95% CI −4.3 to −1.1) before controls and −1.7 (95% CI −3.0 to −0.4) after, so ZIP code areas with higher mean income and ZIP code areas with more credit claims both voted more Democratic. Chapter 12.4, Wages, capital and the ZIP code vote documents a U-shaped relation between tax-return shares and the 2020 margin. The share of returns with unemployment compensation goes from 4.8 (95% CI 3.2 to 6.4) to −1.2 (95% CI −2.3 to −0.2).

The adjusted coefficient differs by surface, and the college variable is one reason

Two sets of measures exist on both surfaces, the Area Deprivation Index and the PLACES measures. Figure 6 compares their coefficients on tracts and on ZIP code areas. Without controls the two surfaces differ by at most 1.5 points, with a median of 0.6. With the common specification they differ by up to 3.4 points, with a median of 2.0. No leisure-time physical activity has 3.3 (95% CI 2.4 to 4.1) on tracts and −0.2 (95% CI −1.2 to 0.9) on ZIP code areas. The sign of the adjusted coefficient agrees on 7 of the 10 measures.

Figure 6

The same measure gives raw coefficients within 1.5 points on tracts and ZIP code areas but adjusted ones up to 3.4 points apart, and up to 2.8 points apart without the college share

Democratic share of the 2020 two-party presidential vote, points per standard deviation, Area Deprivation Index and nine PLACES measures, tracts and ZIP code areas

Tracts
ZIP code areas

No controls

Common specification

Without the college share

Democratic share, points per standard deviation of the measure

Show data

Data for The same measure gives raw coefficients within 1.5 points on tracts and ZIP code areas but adjusted ones up to 3.4 points apart, and up to 2.8 points apart without the college share
estimate	hi	lo	measure	n	specification	surface
−5.86	−3.51	−8.22	Area Deprivation Index rank	80,873	No controls	Tracts
3.90	5.22	2.58	Area Deprivation Index rank	80,873	Common specification	Tracts
−3.63	−2.07	−5.19	Area Deprivation Index rank	80,873	Without the college share	Tracts
−7.15	−5.16	−9.14	Area Deprivation Index rank	29,276	No controls	ZIP code areas
1.75	4.00	−0.505	Area Deprivation Index rank	29,276	Common specification	ZIP code areas
−3.78	−1.71	−5.85	Area Deprivation Index rank	29,276	Without the college share	ZIP code areas
0.221	1.60	−1.16	Fair or poor health	80,873	No controls	Tracts
4.49	5.27	3.71	Fair or poor health	80,873	Common specification	Tracts
−3.32	−2.34	−4.30	Fair or poor health	80,873	Without the college share	Tracts
−0.317	1.41	−2.05	Fair or poor health	29,276	No controls	ZIP code areas
2.65	3.67	1.63	Fair or poor health	29,276	Common specification	ZIP code areas
−5.22	−4.07	−6.37	Fair or poor health	29,276	Without the college share	ZIP code areas
−0.113	1.29	−1.52	Poor mental health	80,873	No controls	Tracts
3.08	3.57	2.59	Poor mental health	80,873	Common specification	Tracts
−3.13	−2.38	−3.88	Poor mental health	80,873	Without the college share	Tracts
−1.49	0.106	−3.09	Poor mental health	29,276	No controls	ZIP code areas
1.35	2.54	0.159	Poor mental health	29,276	Common specification	ZIP code areas
−3.18	−2.31	−4.06	Poor mental health	29,276	Without the college share	ZIP code areas
−6.33	−4.16	−8.49	Depression	80,873	No controls	Tracts
3.06	4.14	1.98	Depression	80,873	Common specification	Tracts
−2.24	−1.12	−3.37	Depression	80,873	Without the college share	Tracts
−5.98	−3.10	−8.86	Depression	29,276	No controls	ZIP code areas
3.83	5.57	2.09	Depression	29,276	Common specification	ZIP code areas
−0.727	1.07	−2.53	Depression	29,276	Without the college share	ZIP code areas
−1.67	−0.121	−3.22	Diabetes	80,873	No controls	Tracts
1.95	2.68	1.22	Diabetes	80,873	Common specification	Tracts
−3.07	−1.90	−4.23	Diabetes	80,873	Without the college share	Tracts
−2.36	−0.770	−3.95	Diabetes	29,276	No controls	ZIP code areas
−0.206	0.716	−1.13	Diabetes	29,276	Common specification	ZIP code areas
−4.81	−3.45	−6.17	Diabetes	29,276	Without the college share	ZIP code areas
−4.80	−2.59	−7.02	Obesity	80,873	No controls	Tracts
0.036	1.58	−1.50	Obesity	80,873	Common specification	Tracts
−6.59	−5.20	−7.99	Obesity	80,873	Without the college share	Tracts
−6.04	−3.97	−8.12	Obesity	29,276	No controls	ZIP code areas
−2.63	−0.550	−4.72	Obesity	29,276	Common specification	ZIP code areas
−6.77	−5.14	−8.41	Obesity	29,276	Without the college share	ZIP code areas
−4.70	−2.81	−6.58	Current smoking	80,873	No controls	Tracts
3.34	4.12	2.56	Current smoking	80,873	Common specification	Tracts
−4.83	−3.99	−5.66	Current smoking	80,873	Without the college share	Tracts
−6.19	−4.37	−8.00	Current smoking	29,276	No controls	ZIP code areas
0.890	2.38	−0.598	Current smoking	29,276	Common specification	ZIP code areas
−5.87	−5.05	−6.68	Current smoking	29,276	Without the college share	ZIP code areas
−1.88	−0.282	−3.49	No leisure-time physical activity	80,873	No controls	Tracts
3.29	4.13	2.44	No leisure-time physical activity	80,873	Common specification	Tracts
−4.74	−3.78	−5.69	No leisure-time physical activity	80,873	Without the college share	Tracts
−2.14	−0.438	−3.83	No leisure-time physical activity	29,276	No controls	ZIP code areas
−0.158	0.920	−1.24	No leisure-time physical activity	29,276	Common specification	ZIP code areas
−6.35	−5.27	−7.42	No leisure-time physical activity	29,276	Without the college share	ZIP code areas
1.77	3.13	0.411	No health insurance	80,873	No controls	Tracts
7.59	8.97	6.21	No health insurance	80,873	Common specification	Tracts
−2.86	−0.641	−5.08	No health insurance	80,873	Without the college share	Tracts
1.66	2.98	0.333	No health insurance	29,276	No controls	ZIP code areas
6.04	8.54	3.54	No health insurance	29,276	Common specification	ZIP code areas
−5.69	−3.78	−7.59	No health insurance	29,276	Without the college share	ZIP code areas
−10.37	−8.95	−11.80	Arthritis	80,873	No controls	Tracts
−0.125	0.385	−0.636	Arthritis	80,873	Common specification	Tracts
−3.96	−2.86	−5.06	Arthritis	80,873	Without the college share	Tracts
−10.79	−9.16	−12.42	Arthritis	29,276	No controls	ZIP code areas
−1.53	−0.643	−2.42	Arthritis	29,276	Common specification	ZIP code areas
−4.86	−3.48	−6.23	Arthritis	29,276	Without the college share	ZIP code areas

In the left panel the two surfaces nearly coincide, in the middle panel they separate for several health measures, and in the right panel, which leaves the college share out on both surfaces, they come closer again. Each dot is the coefficient of the Democratic share on one measure, with 95% intervals from state clusters; the middle panel holds race and ethnicity, education, density and state fixed and the right panel holds race and ethnicity, density and state fixed. The ZIP code college variable averages 63% of residents against 37% for the tract college share, and its definition is unresolved. The ZIP code estimates use every ZIP code area with more than 100 votes, the Area Deprivation Index and the PLACES values.Note: Tract and ZIP code area samples differ: ZIP code areas are fewer and include more sparsely populated places, so the comparison mixes the unit and the sample.Note: The surfaces also differ in how the Area Deprivation Index is aggregated and in the PLACES release, so the unit is not the only difference.Note: Estimates for the ZIP-only sample that also requires every Social Capital Atlas and IRS value differ from these by less than the surface difference.Source: Precinct returns allocated to 2020 census tracts and ZIP code tabulation areas (Fekrazad 2025) with the ADI and PLACES values of the project atlases; 2020 presidential vote; weighted by votes in each unit

The surfaces differ in more than the unit. The college variable of the ZIP code atlas has a vote-weighted mean of 63% against 37% for the tract college share, and its definition is unresolved: the atlas metadata call it college degree or higher, and Chapter 12.4, Wages, capital and the ZIP code vote finds that its county average is too high for a bachelor's share. Leaving the college share out on both surfaces brings the coefficients together. The sign agrees on 10 of the 10 measures, the median difference is 1.3 points and the largest is 2.8, and every coefficient is negative on both surfaces. The Area Deprivation Index is aggregated differently, an unweighted mean of block-group ranks for tracts and an area aggregation for ZIP code areas, and the PLACES values of ZIP code areas are the 2025 ZIP code estimates for 78% of the areas and a mix with tract aggregates for the rest, against the 2024 tract release.

The ZIP code areas in this comparison are the 29,276 that have an Area Deprivation Index and PLACES values, a set larger than the ZIP-only sample that includes more sparsely populated places than the tracts do. Estimates on the smaller ZIP-only sample differ from these by at most 0.7 points. Lack of health insurance has the largest adjusted coefficient on both surfaces. The unit, the sample and the aggregation of the measures cannot be separated from the college variable with these data.

Other chapters use a different footing

Figure 7 sets the specifications side by side. Chapter 13.6, Air pollution and the vote reports a margin, controls for median household income, standardizes exposures with an unweighted standard deviation and clusters by county. Its diesel coefficient with race, college share and income held fixed, converted to points of share, is 5.2 (95% CI 4.1 to 6.3) on 81,134 tracts, and 2.5 (95% CI 1.3 to 3.6) with density added. On the footing of this chapter, with state fixed effects and without income, it is 1.1 (95% CI −0.8 to 2.9). Without state effects but with race, education and density it is 2.5 (95% CI 0.2 to 4.8), close to the air chapter's value with density added, so the difference between the two chapters lies mainly in the state effects. The air chapter's within-county estimate, converted to points of share, is 3.5 (95% CI 2.6 to 4.4), and the county-effects estimate on this chapter's footing is 4.1 (95% CI 2.7 to 5.6). Their intervals overlap, and both are larger than the estimate with state effects, so the diesel result depends on the level of the fixed effects.

Figure 7

Two existing chapters use a different outcome, sample, weights or controls, so their coefficients cannot be set beside the ones here

Outcome, unit, weights, controls and interval method of this chapter and of the two chapters that use the same atlas files

Chapter	Outcome	Unit and sample	Weights	Controls	Intervals
This chapter, tracts	Democratic share of the two-party vote, 2020, in points	80,873 tracts, 3,081 counties	Votes in the tract	White, Black, Asian and Hispanic shares, college share, log density, state effects	State clusters, t distribution
This chapter, ZIP code areas	Democratic share of the two-party vote, 2020, in points	18,039 ZIP code areas (Social Capital Atlas and IRS sample)	Votes in the ZIP code area	Same, with the college variable of the ZIP code atlas	State clusters, t distribution
Air pollution and the vote	Democratic two-party margin, 2020, in points (a linear function of the share)	81,134 tracts, 3,096 counties	Votes in the tract	White, Black and Hispanic shares, college share, median household income; density in one row, county effects in another	County clusters, normal; exposures standardized with the unweighted standard deviation
Wages, capital and the ZIP code vote	Swing: 2020 margin minus 2016 margin, in points	27,144 ZIP code areas, 3,088 counties	2020 votes in the ZIP code area	Poverty, college, Black and Hispanic shares; the average tax rate in some rows	County clusters, t distribution; state clusters as a check

Show data

Data for Two existing chapters use a different outcome, sample, weights or controls, so their coefficients cannot be set beside the ones here
chapter	controls	intervals	outcome	unit	weights
This chapter, tracts	White, Black, Asian and Hispanic shares, college share, log density, state effects	State clusters, t distribution	Democratic share of the two-party vote, 2020, in points	80,873 tracts, 3,081 counties	Votes in the tract
This chapter, ZIP code areas	Same, with the college variable of the ZIP code atlas	State clusters, t distribution	Democratic share of the two-party vote, 2020, in points	18,039 ZIP code areas (Social Capital Atlas and IRS sample)	Votes in the ZIP code area
Air pollution and the vote	White, Black and Hispanic shares, college share, median household income; density in one row, county effects in another	County clusters, normal; exposures standardized with the unweighted standard deviation	Democratic two-party margin, 2020, in points (a linear function of the share)	81,134 tracts, 3,096 counties	Votes in the tract
Wages, capital and the ZIP code vote	Poverty, college, Black and Hispanic shares; the average tax rate in some rows	County clusters, t distribution; state clusters as a check	Swing: 2020 margin minus 2016 margin, in points	27,144 ZIP code areas, 3,088 counties	2020 votes in the ZIP code area

The air pollution chapter reports a margin and controls for income; the wages chapter reports a change between 2016 and 2020 and uses other controls. Each row describes the specification behind the coefficients of one chapter.Source: Chapter data of this book: tract and ZIP code atlas files (Fekrazad 2025) with 2016 and 2020 presidential votes; descriptions of specifications, with unit counts recomputed from the files; counts are unweighted

Chapter 12.4, Wages, capital and the ZIP code vote uses the change in the margin between 2016 and 2020 as its outcome, on 27,144 ZIP code areas, with poverty, college, Black and Hispanic shares as controls. A coefficient there is a change in the margin and a coefficient here is a level of the share, so the two do not compare, even for the earned income credit share that appears in both.

The signs hold under the variations tried

Figure 8 repeats the tract analysis under 9 variations: normal critical values, county clusters, the standard deviation without cutting the tails, percentile ranks in place of standard deviations, every tract counting once, the rural-urban commuting code in place of density, county fixed effects in place of state fixed effects, a higher vote floor, and median household income and the share of older residents added to the controls. The sign of the adjusted estimate agrees with the common specification for at least 21 of the 24 measures in each variation. The groups are less firm than the signs: the pattern agrees for at least 16 of the 24, and it changes most when every tract counts once and when ranks replace standard deviations. Two measures sit close to the line between keeping and losing. Housing built before 1960 keeps 54% (95% CI 42% to 71%) of its raw coefficient and is at or above 50% in 75% of the bootstrap replicates, and incarceration keeps 41% (95% CI 21% to 98%) and is at or above it in 26%.

Figure 8

Under each of 9 changes to the specification the adjusted sign agrees with the common specification for at least 21 of the 24 measures

Tract measures, number of measures whose adjusted interval lies above zero, below zero or includes zero, and agreement with the common specification

Variation	Tracts	Above zero	Below zero	Includes zero	Same sign	Same pattern
Common specification	80,873	16	2	6	24	24
Normal instead of t critical value	80,873	17	2	5	24	24
Intervals clustered by county	80,873	19	2	3	24	24
Standard deviation without cutting the tails	80,873	16	2	6	24	24
Vote-weighted percentile rank in place of the standard deviation	80,873	17	1	6	21	19
Every tract counts once	80,873	18	1	5	24	16
Rural-urban commuting area code in place of density	80,873	19	2	3	23	23
County fixed effects in place of state fixed effects	80,873	21	1	2	23	22
Tracts with more than 500 votes	78,837	16	2	6	23	24
Median household income and the share of older residents added	80,653	17	2	5	23	22

Show data

Data for Under each of 9 changes to the specification the adjusted sign agrees with the common specification for at least 21 of the 24 measures
above	below	includes	n	same_pattern	same_sign	variant
16	2	6	80,873	24	24	Common specification
17	2	5	80,873	24	24	Normal instead of t critical value
19	2	3	80,873	24	24	Intervals clustered by county
16	2	6	80,873	24	24	Standard deviation without cutting the tails
17	1	6	80,873	19	21	Vote-weighted percentile rank in place of the standard deviation
18	1	5	80,873	16	24	Every tract counts once
19	2	3	80,873	23	23	Rural-urban commuting area code in place of density
21	1	2	80,873	22	23	County fixed effects in place of state fixed effects
16	2	6	78,837	24	23	Tracts with more than 500 votes
17	2	5	80,653	22	23	Median household income and the share of older residents added

The count of measures with an adjusted interval above zero stays between 16 and 21, and the pattern agrees with the common specification for at least 16 of the 24 measures. Each row repeats the tract analysis with the one change named in the first column.Note: The pattern is the grouping of the first figure; the group labels there state the rule.Note: The income and age row drops the tracts without a median household income.Source: Precinct returns allocated to 2020 census tracts (Fekrazad 2025) with the project tract atlas; 2020 presidential vote; weighted by tract votes except where the row says every tract counts once

What these data cannot show

The estimates describe areas in one election. They do not say how any resident of a tract voted. The vote in each area is an allocation of precinct returns by a land-cover model, and several of the measures are modeled as well, so the estimates carry errors that the intervals leave out. The measures are not independent of the controls. The Area Deprivation Index contains education, the PLACES values are built from age, sex, race and ethnicity, education and county, and the Opportunity Atlas and EJScreen values are model estimates whose inputs the atlas does not list, so some overlap with the controls is possible. An adjusted coefficient therefore compares areas on the part of a measure that the controls do not explain, and that part can be small. The data cannot say what the remainder is.

The specification is one choice among several. The controls leave out income and age, which the last table adds. The order in which controls enter changes the size of each step, and the level of the fixed effects changes some coefficients. The coefficients are per standard deviation after the tails are cut, and other scalings change the size of some of them. The 24 tract measures overlap and are not entered together, so the coefficients cannot say which of two related measures carries a pattern. Each interval is for one estimate, and there is no adjustment for the many estimates in the chapter.

The sources differ in vintage and in what they describe. The measures are dated 2019 to 2024 and the vote is from November 2020, so most of the releases came after the election; the atlas files do not record their dates, and the dates here come from the atlas build documentation. The atlas also holds firearm, suicide and overdose rates, which this chapter leaves out because the file does not document their source. Social capital and tax returns exist only for ZIP code areas, and the Social Capital Atlas describes Facebook users. The comparison across measures is therefore made on two surfaces with two samples, and Figure 6 shows that the adjusted coefficient differs by surface. Alaska and Hawaii are not in the vote layer, Connecticut's tracts do not join the atlas, and the District of Columbia drops out of the tract sample.

The vote is the 2020 presidential two-party share. The data say nothing about turnout, other offices, other elections or the swing between elections, which Chapter 12.4, Wages, capital and the ZIP code vote takes up for ZIP code areas and Chapter 13.5, Social capital (planned) is planned to take up.

Data sources

Precinct votes on census geographiesPrecinct votes allocated to 2020 census tracts and ZIP code tabulation areas (Fekrazad 2025), contiguous states and the District of Columbia, with the American Community Survey controls of the project atlases (2019 to 2023 five-year for tracts; ZIP code vintage not confirmed)Years: 2016 and 2020 (contiguous United States only) · Weights: votes cast in the unit (pres_tot, pres_tot_2020); units with 100 or fewer votes dropped · License: CC0 (data doi:10.7910/DVN/Z8TSH3), as stated in the project README.
Area Deprivation IndexArea Deprivation Index national rank, Neighborhood Atlas version 4.0.1 (2023); tract values are unweighted means of block-group ranks and ZIP code values are area aggregations; license not confirmedYears: Vintage as distributed with the tract and ZCTA atlas files. · Weights: none (area values) · License: The Neighborhood Atlas may restrict commercial use and redistribution; read its terms before publishing derived tract values.
Opportunity AtlasOpportunity Atlas tract estimates for children born 1978 to 1983 with low-income parents, area-interpolated to 2020 tracts (household and individual income rank, incarceration, teen birth)Years: Birth cohorts and outcome years as defined by Chetty and others (2018). · Weights: none (area values) · License: See the provider's terms of use.
EJScreenEJScreen 2024 tract values (version 2.32 in the atlas build code), from a community mirror of the archived toolYears: Version as distributed with the tract atlas file. · Weights: none (area values) · License: US government data. The EPA removed EJScreen from its website in February 2025, so the archived copy is the source to cite.
CDC PLACESPLACES crude prevalence of nine adult health measures, model-based; 2024 release for tracts and 2025 release for ZIP code areasYears: Release as distributed with the tract and ZCTA atlas files. · Weights: none (area values) · License: US government data, open access.
Social Capital AtlasSocial Capital Atlas ZIP code values (2022 release, Facebook users), six measuresYears: Built from Facebook friendship data; see the provider for the measurement window. · Weights: none (area values) · License: See the provider's terms of use.
IRS SOI ZIP code dataIRS Statistics of Income ZIP code data, tax year 2022: shares of returns with the earned income credit and with unemployment compensation, and mean adjusted gross income per returnYears: Tax year as distributed with the ZCTA atlas file. · Weights: none (area values) · License: See the provider's terms of use.

All data sources

Methods

Sample. Tracts: the 2020 vote layer covers the contiguous states and the District of Columbia (Alaska and Hawaii are not in it) and holds 83,648 tracts. The tract atlas holds 82,765 of them, in 47 states and the District of Columbia; the 883 tracts with no row in the atlas are all in Connecticut, which the atlas codes by planning region and the vote layer by county. 81,617 atlas tracts have more than 100 estimated presidential votes; the common sample keeps the 80,873 that also have every one of the 24 tract measures, the white, Black, Asian and Hispanic shares of residents, the college share and population density, in 3,081 counties and 47 states, carrying 154.8 million votes. The District of Columbia drops out because the atlas has no white or Asian share for its tracts (203 tracts); 15 tracts in other states lack a race share or the college share, and 526 tracts, carrying 0.47 million votes that went 59 percent Democratic, lack an Area Deprivation Index, Opportunity Atlas or EJScreen value. ZIP code areas: the ZIP code atlas holds 32,428 ZIP code tabulation areas, of the 32,462 in the vote layer; 29,420 have more than 100 estimated votes. The bridge sample keeps the 29,276 that also have the Area Deprivation Index, every PLACES measure and the controls, and the ZIP-only sample keeps the 18,039 that also have all six Social Capital Atlas values and the three IRS values, in 2,892 counties and 48 states and the District of Columbia, carrying 149.4 million votes. The atlas codes a missing American Community Survey value as -1 for some ZIP code controls, which are set to missing. The sample table lists the units dropped at each step in the order given. Votes and outcome. Precinct returns allocated to census geographies by Regionalized Land Cover Regression (Fekrazad 2025): the 2020 returns of the MIT Election Data and Science Lab, allocated in proportion to household population, as fractional counts. The vote layer reproduces the national two-party Democratic share of the contiguous states (52.24 percent in 2020, for tracts and for ZIP code areas); the tract atlas file lacks the 883 Connecticut tracts of the layer, which carry 1.8 million votes and voted 60.2 percent Democratic, and gives 52.15 percent. The outcome is 100 times the Democratic votes over the sum of the Democratic and Republican votes (pres_dem_2pty; pres_dem2pty_2020 for ZIP code areas). The two-party margin used by some chapters is twice this share minus 50. Weights: total presidential votes of the unit (pres_tot; pres_tot_2020), in the regression, in every vote-weighted mean and standard deviation and in the R-squared values. Units with 100 or fewer votes are dropped. Measures and signs. Tracts: Area Deprivation Index national rank (adi_national_rank; higher is more deprived); Opportunity Atlas outcomes of children with low-income parents, kfr_pooled_pooled_p25 (household income rank in adulthood) and kir_pooled_pooled_p25 (individual income rank in adulthood), both with the sign reversed so that higher means a lower rank, jail_pooled_pooled_p25 (incarceration) and teenbrth_pooled_female_p25 (teen birth among women); EJScreen tract values in the atlas (ej_diesel_pm_ug_m3, ej_traffic_proximity, ej_no2_ppb, ej_pm25_ug_m3, ej_ozone_ppb, ej_air_toxics_rsei, ej_pct_pre1960_housing, ej_superfund_pctile, ej_rmp_facility_pctile, ej_haz_waste_pctile, the last three being national percentile values); PLACES health_ghlth, health_mhlth, health_depression, health_diabetes, health_obesity, health_csmoking, health_lpa, health_access2 and health_arthritis, chosen to span current illness, risk and access, plus one condition that rises with age, after an exploratory run that looked at 19 of the 40 PLACES measures; the other PLACES measures are left to the health chapter, and seven of them are missing for every tract in ten states. ZIP code areas: the Area Deprivation Index and the same PLACES measures; the Social Capital Atlas columns sc_econ_connectedness, sc_civic_membership_ec, sc_friend_clustering, sc_support_ratio, sc_volunteering_rate and sc_civic_orgs_per_cap, which match the columns ec_zip, ec_grp_mem_zip, clustering_zip, support_ratio_zip, volunteering_rate_zip and civic_organizations_zip of the 2022 release value for value (sc_civic_membership_ec is economic connectedness measured over group and club memberships, according to the atlas build documentation); IRS irs_pct_eitc_returns, irs_pct_unemp_comp_returns and irs_mean_agi_per_return (sign reversed). The Social Capital Atlas keeps its published direction, and every other measure is signed so that a higher value means more hardship. The firearm, suicide and overdose columns of the tract atlas are not used because their source is undocumented. Vintages. The atlas files record none of the following; it comes from the build code and documentation of the atlas in the birth_death project (the tract atlas file here equals that project's analytic atlas on the five columns compared): American Community Survey 2019 to 2023 five-year estimates for the tract controls (the vintage of the ZIP code controls, which come from a file named Zip_All, is not confirmed); Neighborhood Atlas Area Deprivation Index version 4.0.1 (2023), with tract values the unweighted mean of block-group ranks and ZIP code values an area aggregation of them, an approximation; Opportunity Atlas estimates for children born 1978 to 1983, area-interpolated from 2010 to 2020 tracts; EJScreen 2024, version 2.32 in the build code (the air pollution chapter identifies version 2.3 from the values), a community mirror (DOI 10.5281/zenodo.14767363) because EPA removed the tool in 2025; PLACES crude prevalence, the 2024 release for tracts and the 2025 release for ZIP code areas, with native ZIP code area values for 78 percent of the ZIP code areas of the bridge sample and a mix of native and tract-aggregated values for the rest; IRS Statistics of Income tax year 2022; Social Capital Atlas 2022. The measures are dated 2019 to 2024 and the vote is November 2020. Not confirmed: the license terms of the Area Deprivation Index and the definition of per_college_above. Controls. Tracts: pct_white_alone, pct_black_alone, pct_asian_alone and pct_hispanic, pct_college_plus, and the log of total_population over land area in square kilometers. ZIP code areas: per_white, per_black, per_asian, per_hispanic, per_college_above (labelled college degree or higher in the atlas metadata, but the wages chapter finds a county average of 52.7 percent against 23.6 percent for the county bachelor's share, so its definition is unresolved; its vote-weighted mean is 63 percent in the bridge sample against 37 percent for the tract college share) and the log of population over land area in square miles. State fixed effects: the state of the tract, the state of the county code for ZIP code areas. Median household income and age are not controls in the common specification. Estimator. Each measure is cut at its 1st and 99th percentile across the units of its sample (unweighted percentiles) and standardized with the vote-weighted mean and standard deviation, so a coefficient is points of Democratic share per standard deviation. Weighted least squares of the Democratic share on the measure, with the controls added cumulatively (race and ethnicity, then education, then density, then state fixed effects); the fixed effects are removed by weighted within-state demeaning, which gives the same coefficients as state indicators. The common specification is the last column. Intervals: 95%, standard errors clustered by state (CR1, the statsmodels default), with the critical value of the t distribution with the number of states minus one degrees of freedom (2.01 for 47 tract states, 2.01 for 49 ZIP code states; the normal value is 1.96) (Cameron and Miller 2015). The code reproduces b3.stats.wls to six decimals for the coefficients and standard errors of two measures in three specifications, and pyfixest with state fixed effects reproduces the estimate and standard error of 28 coefficients. Pattern rule, set after a first run of the estimates and not before: a measure keeps if its raw interval excludes zero and the adjusted coefficient has the raw sign, an interval that excludes zero and at least half of the raw size; it loses if the adjusted interval includes zero or if it has the raw sign and less than half of the raw size; it reverses if the adjusted interval excludes zero on the other side from the raw coefficient; the rest have a raw interval that includes zero. Retained share: the common-specification coefficient divided by the raw one, with a percentile interval from 2,000 state bootstrap replicates (seed 12345, whole states resampled, a state drawn twice keeps its own fixed effect); the share of replicates at or above 50 percent measures how close a measure sits to the line between keeping and losing. Variance explained: the weighted R-squared of a regression of the standardized measure on the controls with state fixed effects, as a share of the within-state variance; for PLACES measures also, over all tracts, on race and ethnicity shares, the 65-and-older and under-18 shares and county fixed effects, on county fixed effects alone and on race and age alone, as shares of the total variance of the measure cut at its 1st and 99th percentiles. Variations: normal critical values; county clusters; no cutting of the tails; vote-weighted percentile rank in place of the standard deviation; every tract counting once; the USDA rural-urban commuting area code in place of density; county fixed effects in place of state fixed effects (clusters still by state); tracts with more than 500 votes; median household income and the 65-and-older share added (tracts with an income only). Intervals are model based: they do not include error in the vote allocation or in the modeled measures, and they are not adjusted for the number of estimates. Replications: the air pollution chapter's diesel rows, including its within-county row, computed here on its own sample of 81,134 tracts, and the wages chapter's sample of 27,144 ZIP code areas, all match those chapters as published.

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
Robinson, W. S. (1950). Ecological Correlations and the Behavior of Individuals. American Sociological Review 15(3): 351-357. doi:10.2307/2087176
Cameron, A. Colin; Miller, Douglas L. (2015). A Practitioner's Guide to Cluster-Robust Inference. Journal of Human Resources 50(2): 317-372. doi:10.3368/jhr.50.2.317
Kind, Amy J. H.; Buckingham, William R. (2018). Making Neighborhood-Disadvantage Metrics Accessible: The Neighborhood Atlas. New England Journal of Medicine 378: 2456-2458. doi:10.1056/NEJMp1802313
Chetty, Raj; Friedman, John N.; Hendren, Nathaniel; Jones, Maggie R.; Porter, Sonya R. (2018). The Opportunity Atlas: Mapping the Childhood Roots of Social Mobility. NBER Working Paper 25147. doi:10.3386/w25147
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
Chetty, Raj; Jackson, Matthew O.; Kuchler, Theresa; Stroebel, Johannes; and others (2022). Social capital I: measurement and associations with economic mobility. Nature 608: 108-121. doi:10.1038/s41586-022-04996-4

How to cite

@misc{singh2026measuresofhardshipcompared,
  author = {Singh, Vishal},
  title = {Measures of hardship compared},
  year = {2026},
  url = {https://vishalsingh.org/politics/measures-of-hardship-compared},
  note = {In The Public Mind. Published 3 October 2026.}
}

One footing for every measure

Before controls, the measures disagree in sign

With controls, most coefficients are positive and smaller

Race and education move the coefficients most

Environmental burdens lose most of their association

The PLACES health measures share most of their variance with demographics

Deprivation reverses sign and low income rank holds

ZIP code areas add social capital and tax returns

The adjusted coefficient differs by surface, and the college variable is one reason

Other chapters use a different footing

The signs hold under the variations tried

What these data cannot show

Data sources

Methods

References

How to cite