Rare Agent Work · Adversarial Audit Edition
Rev 1.1 · Updated September 21, 2026
Free · open access15 minute briefAir-quality researchers, environmental-justice analysts, journalists covering monitoring policy, and agencies weighing network expansion
Air Monitoring Deserts
An adversarial audit of who lives far from America’s PM2.5 monitors
A pre-publication adversarial audit of a national analysis of PM2.5 monitor coverage — which claims survive, which are statistical artifacts, and what a fixed version can honestly say.
What this report gives you
- 01The core finding is real: in rural America, the most socially vulnerable tracts sit 18–22 km farther from the nearest PM2.5 monitor than the least vulnerable — it survived every attack we threw at it
- 02The scariest headline number was an artifact: “217M Americans uncovered” came from a made-up exponential coverage formula; the defensible figure is 79–92M people living >25 km from a monitor (range driven by whether non-regulatory monitors count)
- 03The “smart equity-weighted siting plan” barely differs from simply following population: ≤0.8 percentage points across all sensitivity configurations — and on the rebuilt regulatory-network pipeline the two objectives resolve the >25 km gap identically (88.8M → 55.8M)
- 04Every published confidence interval was 2–8× too narrow — neighboring tracts are spatially autocorrelated (Moran’s I = 0.77), which i.i.d. bootstraps ignore
- 05The claimed “Simpson’s paradox” doesn’t exist in the data: the national correlation is ~zero, not positive — there is confounding by urbanicity, but no sign reversal
- 06Rev 1.1 — fixes applied and pipeline rebuilt: regulatory-only denominator (1,025 FRM/FEM sites), county-clustered CIs, density partial correlations, LOO-validated τ=30 km kernel, unified siting plan. Rural finding strengthens: r=+0.227, clustered CI [+0.204, +0.251], Q4−Q1 penalty 22.1 km; 250 monitors cut the >25 km population 88.8M → 55.8M
The finding that changes your next decision
“The analysis’s most alarming number — 217 million Americans without PM2.5 monitor coverage — was produced by a formula in which no one, anywhere, ever counts as uncovered. The real figure is 79–92 million, and the gap it measures is real but 2–3× smaller than advertised.”
Read the full report — free ↓
✓ All sections open · No sign-up · No paywall
✓Full preview before purchase✓Cited sources✓Updated September 21, 2026✓Human-authored✓No sign-up required
Why this report exists
Before publishing a national analysis of PM2.5 monitoring coverage, we tried to destroy it. The pipeline asked who lives far from America’s ~1,000–1,300 regulatory PM2.5 monitors, whether the gaps track social vulnerability, and where 250 new monitors should go. Our audit re-derived every headline number from the data itself. The central finding survives: rural vulnerable communities genuinely sit 18–22 km farther from a monitor than rural well-off ones — robust to outliers, Alaska, thresholds, fixed effects, and density controls. But four headline claims do not survive: the “217M uncovered” figure is an artifact of an invented exponential coverage kernel (real answer: 79–92M beyond 25 km, depending on monitor definitions); the monitor count silently includes 277 non-regulatory sites; the “equity-weighted siting plan” is statistically indistinguishable from just following population (≤0.8pp difference everywhere); and every confidence interval is 2–8× too narrow because the method ignores spatial autocorrelation. The publishable paper is the rural inequity finding — stated with clustered intervals and honest ranges — not the flashy numbers around it.
🏛️
What's at stake
- →The binding constraint on this kind of study is monitor-denominator definition: whether “supplemental” monitors (IMPROVE, CSN, nephelometers) count changes the headline population figure by ~15% and the rural distance penalty from 18 to 22 km.
- →Population-weighted correlations over census tracts require clustered or spatial inference — i.i.d. bootstraps on spatially autocorrelated data (Moran’s I = 0.77) produce intervals 2–8× too narrow, turning noise into “significance.”
- →Equity weighting in facility location is not free persuasion: when the coverage kernel is smooth and the budget small, weighted and unweighted objectives select nearly the same sites. Claims that an “equity plan” reduces disparity must show the disparity metric move — here it does not (27.4% → 27.5%).
- →Aggregating to higher spatial units strengthens the rural signal (county-level r = +0.33 vs tract-level +0.22), consistent with centroid measurement noise attenuating the tract-level estimate — the true inequity is, if anything, understated.
⚡
Decision sequence
- 01If you run a monitoring-coverage analysis: define the monitor denominator first (regulatory vs all sites), because it moves every downstream number.
- 02Replace i.i.d. bootstraps with county- or state-clustered bootstraps whenever tract-level spatial data is involved; report Moran’s I as a diagnostic.
- 03Never publish a single τ (coverage decay) point estimate — hold out monitors and fit the decay from held-out prediction error instead of assuming it.
- 04For siting recommendations, verify the candidate set doesn’t systematically exclude the population the equity objective is supposed to serve.
⚠️
Cost of skipping this
- ✕The audit itself is secondary analysis of one pipeline’s processed data; the underlying raw inputs (2024 AQS annual, 2022 SVI, 2023 PLACES, 2020 Census) were taken as delivered.
- ✕Clustered bootstrap intervals are a sensitivity device, not a certified replacement for a spatial statistical model.
- ✕The rural finding’s magnitude depends on the urbanicity threshold (8–22 km across definitions); it is directionally robust but should be published as a range.
- ✕Centroid-to-monitor distance is a proxy for person-level distance; block-level recomputation is the fix, not yet done.
✋
Who this report is NOT for
- ✕Anyone seeking a finished, publication-ready study — this is the audit and remediation plan; the corrected pipeline is the follow-up
- ✕Regulatory-compliance teams needing NAAQS network-design advice — monitor placement for compliance follows 40 CFR 58 rules this analysis does not model
Honest disqualification. If none of the above matches you, this report was written for you.
What's Inside
4 deliverablesshield
Survives / retracts ledger
Every headline claim graded: robust, inflated, or artifact — with the recomputation that decided it
chart
Robustness attack battery
Threshold sweeps, winsorizing, fixed effects, partial correlations, denominator variants — all rerun from raw data
scale
Feasibility verdict
What fixing the coverage gap actually costs: budgets, timelines, and the sensor-network end run
check
Pre-publication checklist
The exact softened wordings and statistical fixes required before this analysis goes public
Full Report
All 4 sections — scroll down to read.
01
What survives the attack
The pipeline’s central claim is that America’s PM2.5 monitoring network leaves socially vulnerable communities behind. We tried to break the rural version of this claim nine ways. Density threshold swept from 250 to 5,000 people per square mile: correlation stays between +0.12 and +0.22, always positive. Drop Alaska and Hawaii: it strengthens to +0.23. Cap extreme distances at the 90th percentile: +0.25. Add state fixed effects: +0.21. Control flexibly for population density: +0.19. The most vulnerable rural quartile sits 18–22 km farther from the nearest monitor than the least vulnerable quartile, and every reasonable specification agrees on direction and rough magnitude. This is the paper’s real finding, and it is solid — if anything understated, because county-level aggregation (which averages out centroid noise) pushes the correlation to +0.33.
02
The four claims that don’t survive
“217 million Americans uncovered.” An artifact. The pipeline modeled coverage as an exponential decay exp(−distance/10 km), under which nobody ever has zero coverage — a person 25 km from a monitor is scored “8% covered.” The defensible statement is that 79–92 million people live more than 25 km from a monitor, with the range driven almost entirely by whether the 277 non-regulatory “supplemental” sites in the dataset count as monitors at all. “The equity-weighted siting plan reduces disparity.” It doesn’t: across all 48 sensitivity configurations, equity-weighted vs population-only siting differ by at most 0.8 percentage points, and the high-vulnerability share of remaining burden is flat at 27.4→27.5%. “Simpson’s paradox masks the inequity nationally.” The national correlation isn’t positive, it’s zero — there is confounding by urbanicity, but no reversal to explain away. And the confidence intervals — all of them — assume tracts are independent draws. They aren’t: Moran’s I on distance is 0.77. Clustered intervals are 2–8× wider; the rural finding survives that correction, the urban “reversed” association does not.
2 more sections in this report
What unlocks with purchase:
- 03
Can the gap actually be closed?
- 04
The verdict, in one table
One-time purchase · Instant access · No subscription
03
Can the gap actually be closed?
Regulatory path: a FRM/FEM monitoring site costs roughly $20k–$75k to establish and $10k–$25k/yr to operate (agency plans and EPA sensor literature), and sites take 1–2 years to site, permit, and produce design-value data. Closing the headline 250-site plan: ~$10–$30M upfront and ~$3–$6M/yr — trivially affordable for federal/state budgets, but gated by agency staffing and NAAQS siting rules, not money. Full closure of the >25 km gap for all ~90M affected people would need thousands of sites — plausible over a decade, not a funding cycle. The realistic near-term fix is hybrid: regulatory monitors where compliance matters, plus dense low-cost sensor networks (PurpleAir-class, ~$100–$2,500/unit, correctable against FRM/FEM collocations) filling rural gaps for exposure estimation now. That hybrid is being validated by EPA and state agencies already; the analysis’s siting output — after fixing its candidate-set bug — is exactly the siting map such a deployment needs.
04
The verdict, in one table
Rural inequity finding: REAL — publish with clustered CIs and an 18–22 km range. “217M uncovered”: RETRACT — replace with 79–92M beyond 25 km, denominator stated. Equity siting advantage: SOFTEN to “weighting barely matters nationally; where the plans diverge is the actual insight.” Urban “closer to monitors” claim: DROP — vanishes under density control. Simpson’s paradox framing: RETRACT — the aggregate is null, not opposite-signed. Statistical precision everywhere: DOWNGRADE — clustered intervals. The pattern generalizes: the expensive quantitative work (spatial joins, optimization, bootstraps) was sound; the failures were all in translation — proxy metrics described as real quantities, an assumed kernel presented as EPA-derived, and inference that ignored the spatial structure the data obviously had.
Evidence & Citations
Every claim in this report traces to a verifiable source.
Last reviewed September 21, 2026
EPA AQS annual summary files (raw inputs)
https://aqs.epa.gov/aqsweb/airdata/download_files.html
Accessed September 21, 2026
CDC/ATSDR Social Vulnerability Index 2022
https://www.atsdr.cdc.gov/placeandhealth/svi/index.html
Accessed September 21, 2026
CDC PLACES local data 2023
https://www.cdc.gov/places/index.html
Accessed September 21, 2026
EPA PM2.5 parameter codes 88101 vs 88502 (AQS reporting memo)
https://www3.epa.gov/ttnairs1/airsaqsORIG/conference_backup.20151017/AQS2006/Session%20Handouts/Update%20for%20PM2.5%20and%20Precursor%20Gas%20Reporting%20Procedures%20-%20AQS%202006.pdf
Accessed September 21, 2026
40 CFR Part 58 Appendix D — network design (monitor siting scales)
https://www.ecfr.gov/current/title-40/chapter-I/subchapter-C/part-58/appendix-Appendix%20D%20to%20Part%2058
Accessed September 21, 2026
Methodology
Who wrote this, what evidence shaped it, and how the recommendations are framed.
- ●Independent recomputation of every headline statistic from the pipeline’s own processed data (83,848 census tracts, ~331M people, 1,302 monitor sites).
- ●Hostile robustness attacks on each claim: density-threshold sweeps (250–5,000 ppl/sq mi), AK/HI exclusion, distance winsorizing (p90–p99), log and rank transforms, state fixed effects, partial correlation on log density, and monitor-denominator variants (FRM/FEM-only, obs%≥75).
- ●Spatial dependence quantified with Moran’s I (8-NN, permutation p=0.002) and tract/county/state clustered bootstraps (400 replicates) against the pipeline’s i.i.d. intervals.
- ●Optimizer forensics: re-ran the greedy siting over the full 84k-tract candidate universe vs the pipeline’s 6,000-tract shortlist, compared equity-weighted vs population-only objectives, and ablated the asthma multiplier.
Author: Rare Agent Work · Written and maintained by the Rare Agent Work research team. Audit recomputations run on public Census, CDC, and EPA AQS data.
Why This Report Earns Attention
Proof 1
Rural vulnerability–distance correlation r=+0.223 robust across thresholds (0.20–0.22 for 250–2,000 ppl/sq mi), AK/HI exclusion (+0.23), winsorizing (+0.24–0.25), state fixed effects (+0.21), and density controls (+0.19).
Proof 2
Monitor denominator swings headline numbers ~15%: 277 of 1,302 “monitors” are non-regulatory supplemental sites (parameter 88502); FRM/FEM-only raises the >25 km population from 79.3M to 88.8M (92.4M at obs%≥75).
Proof 3
Clustered bootstraps widen the national CI from [−0.011, +0.004] to [−0.059, +0.056] (state-level); the rural finding stays significant [+0.18, +0.26].
Proof 4
The optimizer’s candidate shortlist (top 6,000 tracts by population×distance) excludes 11.9M of 16.7M rural high-vulnerability residents from ever hosting a site; the pipeline ships two different siting plans from two different candidate sets.
Ask the Implementation Guide
Powered by Claude — trained on this report's content. Your first question is free.
Ask anything about implementation, setup, or how to apply the concepts in this report. Your first question is free — then we'll ask you to sign in.
Powered by Claude · First question free
When the report isn't enough
Bring a real problem for direct human review.
Architecture review, implementation rescue, and strategy calls for teams with real blockers. Every intake is read by a human before any next step.