Who carries the burden, city by city

Seven criteria, three terms

Seven criteria are measured for every census tract (the Census Bureau's neighbourhood-sized unit, about 4,000 people), then collapsed into three terms. Four of the seven are socioeconomic: minority share, limited-English households, income, and a fourth rule for one band of tracts that is described below. The other three are ozone, PM_2.5 and a health composite. Grouping matters: eleven health indicators added one by one would outweigh everything else in the score, and it would stop being the same index. Air quality and health each contribute one averaged term, which is exactly how the published method already treats its two pollutants. Only five quantities are actually added: the three socioeconomic scores, the air quality mean and the health mean.

Term	What goes in	Range
Socioeconomic	Minority share, limited-English-speaking households, median household income; three separate 0–100 terms, added	0–300
Air quality	Mean of normalised ozone and PM2.5. Both are required; a tract missing either is not scored	0–100
Health	Mean of eleven normalised CDC PLACES chronic-disease prevalences	0–100

Turning unlike quantities into one number

Percentages, parts per billion, micrograms per cubic metre and disease prevalences cannot be added as they are. Each indicator is min–max normalised to a 0–100 scale across all tracts in the country:

score = 100 × (value − min) ÷ (max − min)

Income is inverted (100 − score), because low income is the disadvantage. The three socioeconomic scores are added; ozone and PM_2.5 are averaged into the air quality term; the eleven health measures are averaged into the health term. The EJ score is the sum of the three terms, running from roughly 50 to 420 in practice.

For scale: the median city or town scores 173, the middle half falls between 156 and 199, and only one municipality in twenty is above 269. The lowest in the country is Ross, California at 48.0; the highest is Weedpatch, California at 421.2.

The threshold, and what "EJ community" means here

A tract is designated when its score reaches 65% of the summed baseline, where each criterion's baseline is its own policy threshold pushed through the same normalisation. So the comparison is not against other places directly. It is against a fixed, externally defined idea of what constitutes burden.

The individual thresholds are the ones the published Massachusetts method uses, generalised where a state-specific rule had to become a national one:

One inherited quirk, kept on purpose

A tract that is 25–40% minority (14,005 of the 82,297 scored, 17% of the country) has its minority score averaged with the normalised score of its county's median household income. This is criterion (d) of the Massachusetts rule: "25–40% minority and the surrounding municipality's median income is at most 150% of the state median". Generalising "municipality" to "county" is the only sensible national reading.

Two things about the published implementation of that criterion are reproduced rather than corrected, because reproducing the published score was the point of this work:

• the income score entering that average is not inverted, unlike the tract's own income criterion, so within that band a richer county raises the blended score rather than lowering it, which is the opposite of the direction the rule's wording implies;
• the 150% baseline is computed in the published code but never added to the threshold, so it moves scores without moving the bar they are measured against.

Both are noted here rather than silently fixed. The effect is confined to that one band of tracts; corrected variants are a question for the next paper, not something to smuggle into a reproduction.

Where every number comes from

Input	Source

Note the vintages: this is a composite of four periods, not a snapshot. Socioeconomic data is the 2018–2022 five-year American Community Survey, a rolling average, which is how the Census publishes estimates for units this small. Ozone is the 2021–2023 design value: the fourth-highest daily maximum eight-hour concentration, averaged over three years, which is the statistic the national standard is written in. It is measured at 1,090 monitors and statistically interpolated between them, with the interpolation checked by holding out each monitor in turn (mean error 2.57 ppb). PM_2.5 is a single model year, 2019, from EPA's EQUATES: a nationwide reanalysis run with the CMAQ chemical transport model on a 12 km grid. Health is the 2023 round of the CDC's national telephone health survey.

The health measures

Eleven tract estimates from CDC PLACES, which does not survey every tract: it fits a statistical model to a national survey and projects it onto small areas, so these are modelled estimates rather than counts. Each is a crude prevalence: the share of adults with the condition, with no adjustment for how old the local population is. The number beside each is the population-weighted national prevalence, the baseline a tract has to exceed to be "sicker than the country".

From tracts to cities

The score is computed on census tracts, which do not nest inside municipal boundaries: a quarter of all tracts straddle more than one place. A city's score here is the population-weighted mean of the overlapping tracts, weighted by the population actually shared between each tract and that city, from a crosswalk built out of all 8.1 million 2020 census blocks.

"City or town" also means different things by state. In New England and the township states, the functioning municipality is a county subdivision; elsewhere it is a place (an incorporated city, town or village, or a census-designated place). This page uses county subdivisions in the twelve town/township states and places everywhere else, because a places-only table would get all of New England wrong, since Chelsea and Brookline are not places. Population checks reproduce the official 2020 counts exactly.

What the national picture shows that Massachusetts did not

Air pollution disparity runs along race; health disparity runs along income. Nationally the excess-burden index for PM_2.5, where positive means the disadvantaged carry more than their share of the excess, is +0.044 against minority share but −0.006 against income. For health burden the pattern inverts completely: +0.140 against income, −0.004 against race.

That income null is not what it looks like, and the page would be misleading if it stopped there. Split the underlying correlation by scale and it reverses: income and PM_2.5 correlate +0.216 between counties, because density produces both income and emissions, so richer counties are dirtier. Within them it is −0.135, which is the classic environmental justice gradient. The two cancel to roughly nothing nationally. The racial disparity does not behave this way: +0.242 between counties and +0.226 within, essentially the same at both scales. So the honest statement is not that income fails to predict exposure, but that it predicts exposure locally and is masked nationally by prosperous metros being polluted metros. The racial null on the health side dissolves the same way: four of the eleven measures change sign once you condition on county, and asthma goes from −0.054 nationally to +0.458 within. Any national correlation computed without conditioning on place is measuring which counties are sick, not which neighbourhoods are.

Gary, Indiana is the case for a cumulative index. Among the 923 municipalities with more than 50,000 residents it ranks 21st overall, but only 116th on the socioeconomic term, the block every threshold rule is built from. It is carried there by air quality and health, where it ranks 35th and 2nd: the worst combined air-and-health burden of any large American city. A rule that checked socioeconomic criteria one at a time would not surface it.

The health term also adds far more independent information nationally than it did in Massachusetts: it correlates just +0.248 with the socioeconomic term here, against +0.558 in Massachusetts, so 94% of its variance is not already explained by the socioeconomic criteria.

Does it agree with the official tools?

Two federal screening tools give an external check. Against EJScreen the score tracks the demographic index closely, at a rank correlation of 0.911, and raw PM_2.5 only weakly, at 0.324. That is a fair description of what this score is, and of EJScreen's own EJ indexes: a demographic instrument with a pollution component. Against CEJST, which designates 35.5% of tracts, the 0.65 cut-off is high-recall and low-precision: it catches 84% of what CEJST designates while flagging nearly twice as many tracts, and agreement peaks at a multiplier of 0.70 rather than 0.65. Where the two disagree is itself structured: CEJST is income-gated, so it cannot see tracts that are poor relative to a wealthy state, and it uniquely flags poor, white, clean-air, sick tracts, the second constituency described above. Both tools were removed from their federal websites during 2025; these comparisons use archived copies.

Choices a reader should be able to argue with

Four defensible variants of this score exist, and they designate very different shares of the country. The explorer above shows the first. The fourth drops cancer and depression from the health composite, for the reason given under the limitations.

Limitations you should know about

• This is research, not a regulatory designation. No agency uses this score. For official screening, use your state's own EJ criteria, and note that EPA removed EJScreen, the long-standing federal screening tool, from its website in February 2025. Version 2.3 is still readable in EPA's own January 2025 snapshot and as a public reconstruction.
• PM_2.5 is modelled, not measured. It is a 12 km CMAQ annual mean for 2019: one number for every place inside a 144 km² cell, with a known cool-season high bias in the Midwest. 2019 was also a very low California wildfire year, so the Ohio Valley rather than the Central Valley holds the national maximum; a different year would move this.
• The air quality term cannot tell neighbourhoods apart. 91.9% of its variation is between counties and only 8.1% within one; for ozone alone it is 96.7% and 3.3%. All 32 tracts in Gary sit within half a part per billion of each other on ozone, and Chelsea and Boston are given the same value. As built here it is a regional label, not a neighbourhood measure. The published Greater Boston version did not have this problem, because its air quality came from 59 sensors deployed across the city; nothing at national scale substitutes for that, which is itself the strongest argument in this work for building those networks.
• Ozone is a design value, not an annual mean. The 70 ppb baseline is the 8-hour standard, so it can only be scored against the same metric. An annual-mean field would put the normalised baseline above every observed value and make the criterion unreachable by construction.
• The health data is partly circular. PLACES tract estimates are model outputs whose covariates include age, race and poverty, and race and poverty are already two of the three socioeconomic criteria. The health term should be read as corroborating burden, not as independent evidence of it.
• Health prevalence is crude, never age-adjusted, because PLACES publishes no age-adjusted rate at tract level. An older neighbourhood scores higher partly for being older.
• Two of the eleven health measures point the wrong way. Cancer and depression correlate negatively with the rest of the composite, so in the sickest places they pull against the other nine: the Gary tract walked through above scores 100 on asthma, COPD, diabetes and general health but 15.8 on cancer. Most of that is age, since cancer prevalence tracks the over-65 share at +0.807 and older neighbourhoods have cleaner air, with diagnosis access as a second channel, since PLACES measures diagnosed prevalence. The fourth row of the variant table shows what dropping both does: the designated share moves from 52.8% to 51.9%, so the headline barely changes while individual places move a great deal.
• Two states carry an older health vintage. Pennsylvania and Kentucky had no 2023 survey sample, so their 4,663 tracts carry 2022 estimates from the previous PLACES release. Their excess burden is real, and they exceed the national average on 10 of 11 measures, but the vintage is not identical to everywhere else.
• Alaska, Hawaii and Puerto Rico are not scored. They fall outside the CONUS air quality modelling domain, so the air quality term cannot be computed for them, and this score requires both pollutants.
• Small places are noisy. A town covered by one tract inherits that tract's values wholesale; tracts under 50 people are dropped from the socioeconomic criteria entirely. The population filters above exist for this reason.