rsei-air: EPA RSEI industrial air scores, recomputed for 2023 and 2024

EPA's Risk-Screening Environmental Indicators (RSEI) model takes the air releases every factory reports to the Toxics Release Inventory (TRI) and turns them into a relative score: how much of each chemical reaches the people living downwind, weighted by how toxic it is. EPA released a new RSEI version every year. The last one, version 2.3.12, covers reporting years through 2022 and came out in March 2024. EPA's own "Current Status and Planned Updates of TRI Tools" page still lists the next RSEI release as "TBD", though TRI data for 2023 and 2024 is already public.

This dataset carries the RSEI air pathway forward to TRI 2023 and 2024, using EPA's own V2312 modeled results and the TRI filings for those years. It also includes EPA's facility scores for 2013-2022, so the series runs unbroken.

Map: https://zer0contextlost.github.io/rsei-air/ · Data: https://huggingface.co/datasets/zer0contextlost/rsei-air · Code: https://github.com/zer0contextlost/rsei-air

RSEI scores are a relative screening tool for comparing facilities, chemicals and places. They are not estimates of health risk, and these recomputed years are not EPA products. Not affiliated with EPA.

What's in it

Path	Rows	What
data/facilities/facility_air_2013_2024.parquet	~186k	One row per facility and year. source = epa_v2312 for 2013-2022 (EPA's published scores), recomputed for 2023-2024
data/releases/releases_2023.parquet, releases_2024.parquet	~81k each	One row per facility, chemical and release type (stack or fugitive), with pounds, score, cancer and noncancer scores, method tier and data-quality flags
data/cells/cells_2023.parquet, cells_2024.parquet	~10M each	RSEI 810 m grid cells: score, toxicity-weighted concentration (total, cancer, noncancer), population, number of contributing facilities
data/tracts/tract_<year>.parquet	~84.6k each	2020 census tracts, 2022 (EPA) and 2023-2024 (recomputed)
data/counties/county_<year>.parquet	~3.2k each	Counties, same years
data/unscored_chemicals.csv		TRI chemicals with on-site air releases that have no RSEI V2312 entry (added to TRI after 2022)
validation/		Backtest results (JSON)

Main columns in releases_<year>: trifid, facility, state, lat, lon, parent, sector, tri_chem_id, rsei_chem, chemical, media (stack/fugitive), pounds, score, cancer_score, noncancer_score, tier, donor_year, flags.

Grid cells use EPA's RSEI grid (grid code, x, y cell index). EPA's grid shapefiles and census crosswalks apply unchanged.

National on-site air score (all facilities):

Year	Score	Source
2018	332M	EPA
2019	323M	EPA
2020	299M	EPA
2021	331M	EPA
2022	369M	EPA
2023	445M	recomputed
2024	360M	recomputed

Much of the 2023 rise traces to a few facilities whose reported pounds jumped sharply (see "Data-quality flags").

Method

RSEI models stack and fugitive air releases with AERMOD and first-order chemical decay. Both are linear in the release rate. So, as long as a facility's location, stack parameters and weather station don't change, EPA's modeled score for a facility, chemical and release type is a fixed score per pound. Each 2023/2024 release gets the closest thing EPA already modeled. The tier is recorded in tier:

Tier	What it uses	Share of 2023 score	Share of 2024 score
T1	EPA's own score per pound for the same facility, chemical and release type, from its most recent modeled year	88.9%	80.3%
T2	Same facility and release type, different chemical: EPA's exposure per pound per toxicity weight from the facility's chemical with the closest air decay rate, times this chemical's toxicity weight	5.2%	8.6%
T3	Same facility, release type EPA never modeled there: footprints borrowed from nearby facilities (below)	0.4%	0.7%
T4	Facility EPA never modeled: the average footprint of its 5 nearest modeled facilities with the same release type, shifted onto its own grid cell and applied to the real population around it. Stack releases also match on stack height (RSEI's NAICS-4 median for facilities without NEI stack data)	5.6%	10.4%
T0	Chemical with no inhalation toxicity weight in RSEI: score 0, as in RSEI	0	0
NM	Chemical not in RSEI V2312 at all: not scored

Then:

• Population growth. EPA interpolates and extrapolates grid population by year, so scores drift about 1% a year even at constant pounds. Each facility's score carries forward its own last observed growth rate, capped at ±3% a year.
• Chromium. RSEI only models the hexavalent fraction (facility- or NAICS-specific, 34% default). The same fractions are used here.
• Dioxins. TRI reports these in grams; converted at EPA's 453.6 g/lb.
• Toxicity weights stay at V2312. EPA advises against mixing versions.
• Pounds come from EPA's current public TRI basic data files for 2023 and 2024.

Grid cells. Each facility's predicted score is spread over its footprint cells. A footprint is EPA's 2022 footprint for that facility and release type, or the borrowed one for T3/T4. Cells get toxconc = score_fm * conc_per_lb / S_fm, and score = toxconc * pop * k_cell. Here k_cell is EPA's own per-cell exposure factor, which carries the cell's age mix. Cell totals sum exactly to the facility totals.

Tracts and counties use EPA's 2020 cell-to-tract crosswalks. EPA's published crosswalks only carry area-based weights; the population-weighted field is zero throughout. So cell scores and population are split by area share, and toxconc_popw is the resident-weighted mean concentration. EPA's 2022 tract and county files here were built the same way from EPA's own grid, so all years are comparable.

Validation

Every check rebuilds a year EPA did publish, using only earlier EPA results, then compares the rebuild to EPA's published numbers. The checks run through the same code that produced 2023 and 2024.

Model alone (EPA's frozen pounds, tiers T1-T4), release level:

Rebuild	National total vs EPA	Median error	Within 1%	Within 10%
2021 → 2022	+0.4%	0.32%	84%	97%
2020 → 2022 (two years ahead)	+1.5%	0.70%		95%
2019 → 2021	+2.8%	0.65%		96%
2018 → 2020	+0.3%	0.68%		96%

T1 on its own: median error 0.30%, 99.5% of releases within 10%.

Full pipeline (current public TRI 2022 file, EPA's 2021 footprints, all tiers), against EPA's 2022:

Level	Units	National total	Correlation	Correlation (log)	Median error	Within 25%	Top-decile agreement
Release	83k	+3.2%	0.999		0.34%
Facility	15.8k	+3.2%	0.981	0.982	0.39%		top 100: 95 shared, top 10: 9
County score	3,168	+3.3%	0.992	0.991	1.8%	90%	96%
Tract score	84.6k	+3.3%	0.925	0.985	2.1%	92%	97%
Grid cell score	10.1M	+3.2%	0.940	0.994	5.0%	87%	98%

Most of the +3.2% comes from the data, not the model. About 1,200 releases appear in today's TRI 2022 file but not in the copy EPA froze in fall 2023 (late and revised filings), and those releases account for 18M of score. The model on the same pounds is within 0.4%.

New-facility footprints (T4), leave-one-out on EPA's 2022 footprints (exposure sum per pound, median absolute error):

Method	Fugitive	Stack
Median of nearest facilities' exposure factors (first draft)	30%	37%
National distance profile × local population	28%	35%
5 nearest footprints shifted onto the site (used)	3.4%	25% (with stack-height matching)

External check: CalEnviroScreen 5.0 (draft; California's ToxReleases indicator, which OEHHA describes as RSEI modeling of 2021-2023 TRI air releases, plus Mexican border sources). Across 9,055 California tracts, the log correlation with this dataset's average of EPA 2021, EPA 2022 and our 2023 is 0.969. Against EPA's 2021-2022 average alone it is 0.985. Adding our 2023 lowers the correlation slightly. OEHHA hasn't published how it got 2023 RSEI values, so whether its 2023 matches ours can't be checked further. This is reported as an external comparison, not a confirmation.

Data-quality flags

Pounds are as reported to TRI; nothing is corrected. flags (semicolon-separated):

• pounds_jump_vs_prior_year: the release is 50× or more the same facility's prior-year pounds for that chemical and release type. Covers 10.5% of the 2023 score and 3.0% of 2024. Some of these look like reporting errors. For example, one Ohio facility went from 2.5 lb of chromium in 2022 to 94,323 lb of fugitive chromium in 2023 and 96,124 lb in 2024, which makes it the #1 facility both years.
• pounds_identical_to_prior_year: exactly the prior year's pounds (100 lb or more), which can mean a copied estimate.
• facility_not_modeled_by_epa, footprint_imputed, neighbour_footprints: tiers T3/T4, estimated footprint.
• no_inhalation_toxicity_weight (T0), chemical_not_in_rsei_v2312 (NM), no_donor_found (8 releases in 2024, unscored).

Limits

• Air only. On-site stack and fugitive releases, not direct water releases, transfers to sewage plants, or off-site incineration. RSEI's full score includes those pathways, which are small next to air for most facilities. The water pathway needs stream-reach modeling this method can't borrow.
• Chemicals added to TRI after 2022 are not scored (data/unscored_chemicals.csv). That includes PFAS added since then and diisononyl phthalates (128,594 lb to air in 2024).
• Unchanged facility parameters. Facilities that moved, changed stacks or switched weather stations after 2022 keep EPA's 2022 footprint.
• Population for 2023-2024 is carried forward at each facility's recent growth rate, not new census estimates.
• Stack footprints for new facilities carry about 25% typical error at the facility level.
• If EPA releases version 2.3.13, use it. EPA may also revise toxicity weights, which this dataset doesn't do.

Reproducing

Everything runs on public EPA files (about 47 GB, mostly the two 22 GB disaggregated microdata files). Scripts are in scripts/:

1. download.sh, download_xwalk.sh, to_parquet.py, xwalk_pq.py: fetch and convert.
2. footprints.py <year> and cells.py <year>: per-facility footprints and per-cell totals from EPA's microdata.
3. shapes.py <target year> <shape year>: footprint assignment, including the borrowed T3/T4 footprints.
4. run_year.py <year> [base]: facility-level scores.
5. allocate.py, then tracts.py: grid, tract and county layers.
6. backtest.py, backtest_tri.py, kernel_test.py, kernel_test2.py, validate_geo.py: validation.
7. build_release.py: assemble the release tree.

Peak memory is under 4 GB with DuckDB spilling to disk.

Sources

• EPA RSEI version 2.3.12: public release data, disaggregated and aggregated grid-cell microdata, methodology document (March 2024), census crosswalks (August 2024). https://gaftp.epa.gov/rsei/
• EPA TRI basic data files, 2021-2024. https://www.epa.gov/toxics-release-inventory-tri-program/tri-basic-data-files-calendar-years-1987-present
• OEHHA CalEnviroScreen 5.0 draft, ToxReleases indicator (comparison only).
• US Census Bureau 2020 cartographic boundary files (map).

License

Code: MIT. Data: CC BY 4.0. The underlying EPA data is in the US public domain.

Comparing with airtox and with CDC cancer incidence

The map's "Hotspots 2021" view sets this dataset beside airtox, which holds EPA's AirToxScreen cancer risk from all sources. The comparison year is 2021, the last year both cover.

• Industry is a small share of modeled air-toxics cancer risk. Point sources are 3.7% of population-weighted cancer risk in 2021 (median county 0.9%). Most of the rest is background concentrations and formaldehyde that forms in the air (63%). A few counties are the exception: St. John the Baptist Parish, LA is 70% point sources, and several other Mississippi River parishes and Jefferson County, TX are 40-46%.
• The two models agree moderately on industrial hotspots. County rank correlation between RSEI cancer-weighted concentration and AirToxScreen point-source risk is 0.68. Of RSEI's top 10% of counties, 51% are also in AirToxScreen's. 161 counties are in the top 10% on both.
• Where they disagree, the causes are specific. Some sources missed TRI that year. For example, ethylene oxide sterilizers in Doña Ana County, NM and Henderson County, TX show up in AirToxScreen 2021 but only appear in TRI from 2022, when the number of EtO reporters rose from 120 to 147. Other gaps come from metal speciation and toxicity assumptions: cobalt and chromium plants rank far higher in RSEI than in the emissions inventory.

CDC cancer incidence. County incidence (NCI/CDC State Cancer Profiles, age-adjusted 2019-2023) was checked against both maps, adjusting for adult smoking, obesity and rurality, with melanoma as a negative control. Result: no usable association.

• Total modeled risk correlates weakly with all-cancer and lung cancer incidence. It correlates negatively with leukemia, lymphoma and bladder cancer, and most strongly (negatively) with melanoma. That pattern points to demographic and geographic confounding, not air pollution.
• Industrial exposure shows about 0.5-2% higher all-cancer and lung incidence per doubling of exposure, and none for leukemia, lymphoma or bladder. The melanoma control moves by a similar amount.

This is expected. Modeled air-toxics risk is tens per million over a lifetime, against a baseline lifetime cancer risk of roughly 40%. That is far below what county incidence statistics can resolve. A null result here neither confirms nor rules out local health effects. Scripts and results: analysis/.