How the site and research work

1. 1

   Match the system name

   A ZIP code can contain several water systems. Find the utility name on your water bill and use the result for that system. The ZIP search alone cannot identify the supplier for a specific address.

2. 2

   Check the sampling period and results

   Each water-system card shows when EPA monitored that system. For each PFAS, the card reports the highest yearly average among the system’s EPA sampling locations. These are system monitoring results, not measurements from household faucets.

3. 3

   Read the study benchmark

   This label compares a complete yearly average with a fixed, unrounded research cutoff. The cutoff reproduces EPA’s January 2026 technical-assistance classification, which was based on the April 2024 federal levels. It is not a current federal or state standard, and the label is not a compliance finding.

The research paper asks a separate national question. It tests whether PFAS results across community water systems vary with the racial, ethnic, economic, and rural composition of their service areas after accounting for system size, water source, ownership, and region. The analysis is ecological and cross-sectional. It does not assign exposure or health outcomes to individuals.

• PFAS monitoring: EPA UCMR 5 occurrence records received through January 15, 2026 and EPA’s January 2026 Average PFAS Results export.
• System characteristics: EPA SDWIS Federal Reporting Services.
• Service areas: EPA Public Water System Service Areas version 3, including system-sourced and modeled boundaries.
• Demographics: 2020–2024 American Community Survey block-group estimates and Census Planning Database urban/rural fields.
• ZIP lookup only: EPA’s PWS-reported UCMR service-ZIP associations.

EPA described the January 2026 file as containing approximately 95% of expected UCMR 5 results. EPA released the final UCMR 5 dataset in August 2026. The submitted manuscript and this research release still use the January snapshot. The final federal data are available through EPA’s Data Finder, but they are not substituted into this fixed analysis.

Complete monitoring means that, for at least one sampling location, a water system had every required sample needed to calculate a yearly average. It does not mean every location or home was tested, and it does not show that the system was safe or compliant. UCMR 5 sought to monitor all eligible systems serving at least 3,300 people. It also selected 800 smaller systems using a population-weighted design stratified by water source and system size to support national occurrence estimates. Public row-level inclusion probabilities or analysis weights were not identified. For that reason, the primary inferential cohort includes active community water systems serving at least 3,300 people that had complete monitoring, an eligible service-area boundary, and sufficient demographic coverage.

The 632 monitored smaller community systems with eligible demographic records are reported separately in an explicitly unweighted sensitivity analysis. They do not represent all small U.S. systems. Nontransient noncommunity systems are excluded from the demographic models because residential Census characteristics do not describe the populations of schools, workplaces, hospitals, and other nonresidential systems.

1. 1

   Use complete sampling sets

   Annual averages are defined at an EPA sampling location using its complete required monitoring set. The EPA reporting level is the lowest concentration UCMR 5 reports as a number for that compound. EPA assigns results below that level a value of zero for the yearly-average calculation, but they are not measured zero concentrations and do not prove that a compound was absent.

2. 2

   Reproduce EPA classifications

   The primary source is EPA’s January 2026 Average PFAS Results export. A record-level reconstruction is retained as a validation analysis rather than as a substitute for unresolved replacement-record logic.

3. 3

   Classify the combined primary outcome

   A system enters the combined outcome when at least one complete-set location has a yearly average at or above an April 2024 individual-compound or qualifying-mixture technical cutoff.

These cutoffs reconstruct a rounded federal comparison; they are not newly proposed toxicological thresholds. The manuscript and website use “benchmark” or “technical comparison,” never “violation” or “noncompliance.” Compound-specific outcomes are secondary. Sparse or nonconvergent adjusted models are reported descriptively rather than forced into unstable estimates.

EPA’s water-system-to-2020-Census-block-group crosswalk links each system service area with block groups. The primary aggregation uses residential-building-footprint weights; area weighting is a sensitivity. Demographic records are retained when block-group links with valid core measures cover at least 90% of the crosswalk-estimated population.

The joint model includes service-area percentages Hispanic/Latino, non-Hispanic Black, non-Hispanic American Indian/Alaska Native, below the federal poverty level, and rural. System covariates are log₂ population served, source-water category, ownership, and EPA region.

Connecticut’s planning-region transition changed block-group identifiers in current Census products. For Connecticut only, the analysis uses the official 2023 Planning Database to recover geography-aligned 2017–2021 ACS characteristics and 2020 urban/rural counts on the original 2020 identifiers used by EPA. All 56 primary-cohort systems using that fallback are flagged; a latest-vintage-only sensitivity excludes them.

• Adjusted prevalence ratios are estimated with modified Poisson regression.
• EPA-region fixed effects address broad regional differences; leave-one-region-out analyses assess influence.
• Residual spatial autocorrelation is tested using permutation-based Moran’s I at 4, 8, and 12 nearest neighbors.
• Because the prespecified diagnostic remained positive, primary uncertainty uses a Conley-style spatial heteroskedasticity-and-autocorrelation-consistent sandwich estimator with a 250-km Bartlett kernel.
• State-clustered variance and 100- and 500-km spatial cutoffs are uncertainty sensitivities.
• Secondary compound-specific coefficients are subjected to event-count and convergence gates and Benjamini–Hochberg false-discovery-rate control.

The spatial amendment was logged after the residual diagnostic rather than presented as prespecified. At eight neighbors, the primary-model Pearson residual Moran’s I was 0.119 (999-permutation two-sided P=.001).

Required sensitivity analyses include area versus building weights, latest-vintage Census data only, all monitored CWSs as an unweighted sample, system-sourced service boundaries only, alternative spatial cutoffs, leave-one-EPA-region-out refits, and size/source strata. A model is not reported when separation or nonconvergence makes the estimate unreliable.

Of 8,069 primary-cohort systems, 964 (11.9%) had at least one yearly average at or above an April 2024 technical cutoff. In the joint adjusted model, each 10-percentage-point higher Hispanic service-area composition corresponded to a prevalence ratio of 1.15 (95% CI, 1.09–1.22). Estimates for non-Hispanic Black composition (PR, 1.02; 95% CI, 0.93–1.11) and non-Hispanic American Indian/Alaska Native composition (PR, 1.18; 95% CI, 0.91–1.53) were null or imprecise.

Poverty composition (PR, 0.74; 95% CI, 0.64–0.87) and rural composition (PR, 0.94; 95% CI, 0.91–0.97) were inversely associated with the combined benchmark outcome. These directions must be reported as prominently as the positive Hispanic-composition association and investigated—not rationalized away.

The Hispanic-composition estimate was similar across the primary geographic, cohort, uncertainty, regional-influence, and most size/source sensitivities. Stability does not convert an ecological cross-sectional association into a causal or individual-level claim.

The ZIP lookup reads directly from the frozen scientific export. It finds water systems connected with the entered ZIP, identifies each system’s monitoring dates from the same frozen EPA occurrence file, and displays the EPA yearly averages. The website does not recalculate the paper’s results, fill in missing values, or create a health-risk score.

The PFAS Results Assistant is optional. The ZIP lookup itself does not use AI. When someone asks a question, the question and a short summary of the displayed water-system results are sent through the site’s Cloudflare worker to Groq, a third-party AI provider. Visitors should not enter names, medical records, or other private information.

The assistant can explain terms, general health research, sampling locations, utility follow-up, and optional filter guidance. It is limited to a compact evidence set based on EPA, ATSDR, and NSF guidance. Deterministic boundary responses prevent it from treating a system result as a safety, compliance, medical, personal-exposure, or exposure-percentage determination. It cannot change a number or comparison on the page. Its other answers may still contain mistakes. If it is unavailable, the ZIP lookup continues to work.

EPA: Reducing PFAS in Your Drinking Water with a Home Filter →

• Ecological design: service-area composition cannot be assigned to individuals and cannot establish individual exposure.
• Cross-sectional timing: associations do not establish causality or historical exposure.
• Frozen interim release: the research uses the January 2026 UCMR 5 file, not the final federal dataset released in August 2026. Use EPA’s Data Finder for current federal records.
• EPA reporting level: This is the lowest concentration UCMR 5 reports as a number for a compound. A result below that level does not establish that the compound was absent or that its concentration was literally zero.
• Service-area uncertainty: some EPA boundaries are modeled, and demographic interpolation does not eliminate within-area heterogeneity.
• Small-system sampling: unweighted monitored systems below 3,300 people do not support inference to all small systems or demographic subgroups.
• Correlated community measures: compositional predictors are interdependent; adjusted coefficients are not isolated causal effects.
• Benchmark interpretation: technical-assistance classifications based on UCMR monitoring are not regulatory compliance determinations.
• Population totals: summed reported system populations can double-count people served through wholesale-retail networks.
• Policy instability: the federal regulatory configuration may change after this release.

Internal validation status: automated tests cover annual-average comparison rules, Hazard Index classification, prevalence-ratio recovery, and false-discovery-rate behavior. The EPA average-results normalization was checked against the January 2026 federal summary counts. Release files carry cryptographic hashes that can reveal an unintended change.

Independent review status: the manuscript and analysis have not completed independent peer review, and this release has not yet been independently reproduced. Planned or requested review is not represented as completed review, institutional endorsement, or authorship. Authorship must reflect substantive contribution, accountability, final approval, and the ability to defend the work; family relationship, title, or affiliation is not evidence of contribution.

The archived data and code provide the versioned materials needed for an outside reviewer to attempt reproduction. Public availability supports review; it is not evidence that independent validation has already occurred.

Joshua Tamayo-Sarver, MD, PhD, provided early feedback that helped shape the project’s clinical framing. This acknowledgment does not represent current clinical validation, supervision, or institutional endorsement.

Future external review may identify scientific or communication changes. Any method change must be logged and rerun through the same pipeline.

1. Zhang J. PFAS UCMR 5 reproducibility release v0.2.0. Zenodo. doi:10.5281/zenodo.21968954.
2. US Environmental Protection Agency. Fifth Unregulated Contaminant Monitoring Rule Data Finder.
3. US Environmental Protection Agency. Consumer Confidence Reports.
4. US Environmental Protection Agency. Selection of Nationally Representative Public Water Systems for UCMR: 2021 Update. EPA 815-B-21-012.
5. US Environmental Protection Agency. PFAS drinking-water rule implementation.
6. US Environmental Protection Agency. Proposed PFAS rescission rule.
7. US Environmental Protection Agency. Public Water System Service Areas.
8. US Environmental Protection Agency. Reducing PFAS in Your Drinking Water with a Home Filter.
9. Agency for Toxic Substances and Disease Registry. How PFAS Impacts Your Health.
10. Agency for Toxic Substances and Disease Registry. PFAS exposure sources and general information.
11. NSF. PFAS in Drinking Water: Filter Certification Guidance.
12. US Census Bureau. American Community Survey 5-year data.
13. Conley TG. GMM estimation with cross sectional dependence. Journal of Econometrics. 1999;92(1):1–45. doi:10.1016/S0304-4076(98)00084-0.