Socioeconomic Disparities in Exposures to PFAS and Other Unregulated Industrial Drinking Water Contaminants in US Public Water Systems

Abstract

Background:

Unregulated contaminants in drinking water, such as per- and polyfluoroalkyl substances (PFAS), can contribute to cumulative health risks, particularly in overburdened and less-advantaged communities. To our knowledge, there has been no nationwide assessment of socioeconomic disparities in exposures to unregulated contaminants in drinking water.

Objective:

The goals of this study were to identify determinants of unregulated contaminant detection among US public water systems (PWSs) and evaluate disparities related to race, ethnicity, and socioeconomic status.

Methods:

We gathered data from the US Environmental Protection Agency’s (US EPA’s) Third Unregulated Contaminant Monitoring Rule (2013–2015), PWS characteristics, sociodemographic data, and suspected pollution sources from regulatory databases. We included four target contaminants (or classes) with industrial sources: PFAS, 1,4-dioxane, 1,1-dichloroethane, and chlorodifluoromethane (HCFC-22). Associations were evaluated with pairwise comparison tests and generalized logistic mixed-effects regression models for six dichotomous outcomes: detection of each of four target contaminants, detection of $\ge 1$ target contaminant, and PWS exceedance of $\ge 1$ US EPA health reference level that was in effect in 2017.

Results:

More than 97 million US residents were served by a PWS with detectable levels of 1,4-dioxane (22% of PWSs), HCFC-22 (5.8%), 1,1-dichloroethane (4.7%), and/or PFAS (4.0%). Unregulated contaminant detection was more frequent among large systems, urban systems, and systems using groundwater or a combination of groundwater and surface water. In comparison with PWSs with no detectable levels of these unregulated contaminants, PWSs with detectable levels served counties with higher proportions of Hispanic residents (17% vs. 13%), as did PWSs that exceeded EPA health reference levels in comparison with PWSs with no exceedances (18% vs. 14%). There were persistent positive associations between proportions of Hispanic residents and detections of target contaminants, even after accounting for pollution sources.

Discussion:

Previously, inequities in exposures to drinking water contaminants were underestimated because prior studies have focused on regulated contaminants. PWSs serving counties with more Hispanic residents, non-Hispanic Black residents, and urban households may benefit from additional resources to proactively mitigate unregulated chemical contamination. Future studies should evaluate factors underlying these disparities to promote actions that protect water quality for all residents. https://doi.org/10.1289/EHP14721

Introduction

Despite improvements in drinking water quality over the last 50 y following the passage of the Safe Drinking Water Act (SDWA), access to safe drinking water in the United States remains a challenge for many communities of color and low-income communities. Accounts of drinking water violations in historically marginalized communities (e.g., Flint, Michigan; Newark, New Jersey; the Central Valley in California; the Navajo Nation; and settlements along the US–Mexico border) illustrate that safe drinking water is an unevenly protected human right.¹ In 2021, 6,586 public water systems (PWSs), representing 4.3% of all PWSs and serving 19.6 million US residents, had violations of maximum contaminant levels (MCLs) and other health-based standards.² Although microbial contaminants such as Campylobacter and Legionella cause acute health effects, most regulated chemical contaminants increase chronic disease risks, so the true impact of poor drinking water quality on long-term population health extends far beyond waterborne disease outbreaks.³

Previous studies have identified race, ethnicity, and socioeconomic status (SES) as predictors of water quality and violations of drinking water standards in the US.^4–11 These demographic factors may shape PWS performance because they can be associated with inadequate and uneven enforcement of regulations, closer proximity to pollution sources, and limited community capacity to maintain and upgrade water infrastructure.¹ Source waters closer to anthropogenic activities may be at greater risk for contamination, and discriminatory policies have disproportionately placed industrial and hazardous waste facilities in areas with lower SES and higher percentages of non-Hispanic Black and Hispanic people in the US.¹² Small water systems and systems serving lower SES communities, non-English speaking communities, and rural communities may struggle to obtain resources for necessary infrastructure to protect source water quality and address contamination.^1,13–15

PWS compliance with US testing requirements and health-based standards has been associated with community demographics for a range of regulated contaminants.^4,6–11 Prior studies have evaluated SDWA violations for contaminants such as nitrate, arsenic, lead, other metals, and disinfection byproducts. Small, rural systems have been found to have more frequent noncompliance with standards and higher contaminant levels.^14–17 In nationwide and regional analyses, communities with more Hispanic residents are served by PWSs with higher nitrate and arsenic concentrations.^9,18,19 Some studies further suggest that the association between race/ethnicity and water quality may vary by SES,^7,13 region,⁹ and system size.^6,18,19 However, associations between environmental justice indicators and water quality have not been consistently observed across the country. For example, Statman-Weil et al.²⁰ did not find associations between the percentage of people of color and health-based SDWA violations among community water systems in Pennsylvania. Cory and Rahman²¹ reported that percentages of Hispanic and Black residents and lower SES were not associated with arsenic contamination of drinking water in Arizona. Studies examining inequities in drinking water vary in their choice of geographic scale of analysis, geographic scope, and statistical methods, which can be influential in detecting inequities or a lack thereof.^17,22

Federal standards have been set for around 90 contaminants; however, this list pales in comparison to the much longer list of unregulated contaminants that have been detected in environmental waters²³ and more than 86,000 potential chemical contaminants used in US commerce.²⁴ Unregulated contaminants, or “contaminants of emerging concern,” refer generally to chemical, physical, and biological contaminants that do not have enforceable standards under the SDWA. Sources of unregulated contaminants include industrial facilities, agricultural and livestock production, landfills, wastewater, and disinfection of drinking water. In lieu of federal drinking water standards, the US Environmental Protection Agency (US EPA) has developed non-enforceable health advisory (HA) levels for $\sim 200$ contaminants²⁵ as guidelines for PWSs and state agencies. In addition, some individual states have promulgated health-based guidelines or enforceable standards for contaminants that are not regulated at the federal level. For instance, prior to April 2024, the US EPA had issued a series of non-enforceable HAs for several per- and polyfluoroalkyl substances (PFAS). In the absence of federal standards, 18 US states had established stricter or more extensive standards and guidelines,²⁶ although this state-by-state approach can lead to unevenly distributed exposures and health risks.²⁷

Exploring the relationship between unregulated contaminants and community demographics can inform our understanding of the extent to which demographics influence water quality. Recent studies have documented disparities in PFAS contamination among water systems based on analyses of datasets from individual US states (New Jersey,²⁸ California²⁹) and a compilation across multiple states.³⁰ Liddie et al.³⁰ examined inequities in PFAS detections from 7,873 community water systems in 18 states that conducted PFAS monitoring between 2016 and 2022. They reported associations in PFAS sources in watersheds by race/ethnicity and associations between PFAS detections $\ge 5\text{\hspace{0.17em}ng}/\mathrm{L}$ in drinking water with higher proportions of Hispanic and Black residents and with poverty in rural areas. However, these prior analyses were based on single states or a subset of US states and did not include contaminants beyond PFAS. To our knowledge, no study has investigated socioeconomic disparities in unregulated contaminants in drinking water at the national scale in the United States.

The goal for this study was to extend the framework from our previous environmental justice research⁹ to evaluate sociodemographic disparities in the presence of unregulated industrial contaminants in US public water systems. We hypothesized that positive associations between target contaminant detections and sociodemographic factors would persist even after accounting for the presence of suspected point sources.

Methods

UCMR 3 Data

We combined data on PFAS and other industrial contaminants from the US EPA’s Unregulated Contaminant Monitoring Rule’s Third Cycle (UCMR 3) with information about potential pollution sources and demographics of the populations served by PWSs. We accessed the complete UCMR 3 dataset in January 2017.³¹ US PWSs are required to periodically monitor for select unregulated contaminants as part of the UCMR program under the SDWA. PWSs are defined as water systems that have $\ge 15$ service connections or serve on average $\ge 25$ people at least 60 d/y.³² Between January 2013 and December 2015, all PWSs serving $>\mathrm{10,000}$ customers (“large systems”) ($n=\mathrm{4,124}$) and a subset of PWSs ($n=799$) serving $\le \mathrm{10,000}$ customers (“small systems”) tested up to 30 unregulated contaminants during UCMR 3. These contaminants included seven hormones, two enteroviruses, and 21 List 1 Assessment Monitoring contaminants: PFAS [perfluorooctane sulfonic acid (PFOS), perfluorooctanoic acid (PFOA), perfluorononanoic acid (PFNA), perfluorohexane sulfonic acid (PFHxS), perfluoroheptanoic acid (PFHpA), perfluorobutane sulfonic acid (PFBS)], volatile organic compounds [1,2,3-trichloropropane, 1,3-butadiene, chloromethane (methyl chloride), 1,1-dichloroethane (1,1-DCA), 1,4-dioxane, bromomethane, chlorodifluoromethane (HCFC-22), bromochloromethane (Halon 1011)], metals (vanadium, molybdenum, cobalt, strontium, chromium, chromium-6), and chlorate.³³ Systems subject to Assessment Monitoring collected samples for these contaminants within a 12-month period in 2013–2015.³³ Monitoring frequencies for List 1 chemicals varied by source water type. Groundwater (GW) sources were required to be sampled twice over a 5- to 7-month period. Surface water (SW) and GW under the direct influence of surface water (GUDI) sources were required to be sampled in each of four consecutive quarters. We categorized each PWS into one of three categories according to source water type: GW systems relied only on GW sources; SW systems relied only on surface water sources; and MIX systems relied on a combination of groundwater and surface water sources and/or relied on at least one GUDI source. There were no missing levels for system size and source water type.

We selected a subset of List 1 contaminants that were a) synthetic compounds with industrial sources not associated with disinfection byproducts and b) detected in $\ge 1\%$ of samples to ensure adequate statistical power. Four contaminants (or contaminant classes) met these criteria: PFAS, 1,4-dioxane, 1,1-DCA, and HCFC-22 (Table 1). A PWS was considered to have a PFAS detection if at least one of six PFAS (PFOS, PFOA, PFNA, PFHxS, PFHpA, or PFBS) was detected because of low detection frequencies of individual PFAS analytes ($<1\%$ of samples) and high reporting limits (10–$90\text{\hspace{0.17em}ng}/\mathrm{L}$).

Table 1.

Reporting limits, detection frequencies, and common sources of UCMR 3 target contaminants among US public water systems (2013–2015).

Contaminant	Reporting limit (ng/L)	Sample detection frequency (%)a	Percentage of systems with $\ge 1$ detectionb	US EPA health-reference level (ng/L)	Percentage of systems with $\ge 1$ HRL exceedancec	Common sources
1,4-dioxane	70	11.6	22.1	350	7.1	Solvent production, consumer products
1,1-dichloroethane	30	2.2	4.7	6,140	0.02	Solvent production
HCFC-22	80	2.3	5.8	—	—	Refrigerant, low-temperature solvent and in fluorocarbon resins
PFAS (any)	—	1.6	4.0	—	—	Firefighting foams, consumer products, fluoropolymer coatings
PFOA	20	1.0	2.4	70d	1.3
PFOS	40	0.8	1.9
PFHpA	10	0.6	1.7	—	—
PFHxS	30	0.5	1.1	—	—
PFNA	20	0.1	0.3	—	—
PFBS	90	0.1	0.1	—	—

Note: —, not applicable; HCFC-22, chlorodifluoromethane; HRL, health reference level; PFAS, per- and polyfluoroalkyl substances; PFBS, perfluorobutane sulfonic acid; PFHpA, perfluoroheptanoic acid; PFHxS, perfluorohexane sulfonic acid; PFNA, perfluorononanoic acid; PFOA, perfluorooctanoic acid; PFOS, perfluorooctane sulfonic acid; PWS, public water system; UCMR 3, Third Unregulated Contaminant Monitoring Rule (2013–2015); US EPA, Environmental Protection Agency.

^a Percentage of samples tested that contained each contaminant at least once.

^b Percentage of PWSs that detected each contaminant at least once.

^c Percentage of PWSs with at least one HRL exceedance.

^d US EPA’s 2016 lifetime health advisory level of $70\text{\hspace{0.17em}ng}/\mathrm{L}$ for PFOA and PFOS combined was used to define an HRL exceedance. Maximum contaminant levels for PFOA, PFOS, and four other PFAS were finalized in 2024.

For our analyses, we also used US EPA health reference levels (HRLs) that were in place when the UCMR 3 dataset was finalized in 2017 (Table 1). In 2013, the US EPA established a drinking water HRL for 1,4-dioxane of $350\text{\hspace{0.17em}ng}/\mathrm{L}$ and 1,1-DCA of $\mathrm{6,140}\text{\hspace{0.17em}ng}/\mathrm{L}$, based on a 1-in-a-million cancer risk level.³⁴ In 2016, the US EPA established a lifetime HA of $70\text{\hspace{0.17em}ng}/\mathrm{L}$ for PFOS and PFOA (individually or combined) based on adverse developmental effects.^35,36 This HA level was used as the HRL in the current study for PFAS because it has been used to interpret UCMR 3 and was in effect in 2017.^37,38 The US EPA established standards in 2024 for PFOA ($4\text{\hspace{0.17em}ng}/\mathrm{L}$), PFOS ($4\text{\hspace{0.17em}ng}/\mathrm{L}$), PFNA ($10\text{\hspace{0.17em}ng}/\mathrm{L}$), and PFHxS ($10\text{\hspace{0.17em}ng}/\mathrm{L}$) that are well below the MRLs in UCMR 3 for these contaminants, so any detection of these four PFAS would exceed the current federal standards.³⁹ In addition, an indeterminate number of PWSs that did not have reportable PFAS levels in UCMR 3 may also have had levels above these new standards. Our analyses were based on six binary outcome measures across all samples analyzed for each PWS: a) detection of $\ge 1$ target contaminant, b) detection of any PFAS, c) detection of 1,4-dioxane, d) detection of 1,1-DCA, e) detection of HCFC-22, and f) exceedance of $\ge 1$ HRL that was in effect in 2017.

Locations of Suspected Contaminant Sources

To account for potential point sources in our regression analyses, we compiled the locations of facilities known to discharge or suspected of discharging each target contaminant. We compiled data on facilities that reported releases of 1,4-dioxane, 1,1-DCA, and HCFC-22 from the US EPA Toxics Release Inventory (TRI) database.⁴⁰ The TRI database provides information on industrial releases of $>600$ chemicals from facilities with $>10$ employees that a) have a qualifying North American Industry Classification System (NAICS) code for a specific compound and b) handle that compound in a quantity above a reporting threshold.⁴⁰ We downloaded TRI basic files for the reporting years 2010–2015 and filtered for facilities that reported any environmental releases (both onsite and offsite) of 1,4-dioxane, 1,1-DCA, and HCFC-22. We also included facilities that reported releases of 1,1,1-trichloroethane and CFC-12 due to environmental transformation to 1,1-DCA⁴¹ and HCFC-22,⁴² respectively. A total of 214 TRI facilities met our inclusion criteria (“relevant TRI facilities”). County Federal Information Processing Standards (FIPS) codes associated with each criteria facility were matched to PWSs by county. We dichotomized TRI facility (present/absent) rather than rely on reported emissions because TRI facilities occurred in $<10\%$ of 1,720 counties served by a UCMR 3 PWS. These dichotomous indicators captured historical emissions and avoided emissions misspecification among relevant TRI facilities that fell below reporting thresholds in the period 2010–2015. For each PWS, a TRI facility was considered present if a relevant facility was a) present in $\ge 1$ county served by that PWS and b) reported any environmental releases of 1,4-dioxane, 1,1-DCA, HCFC-22, or precursors in the period 2010–2015.

TRI data were not available for PFAS in 2010–2015 because PFAS were not added to TRI until 2020, and data were first available in 2021. Instead, we compiled data about all the point sources included in Hu et al.³⁸:16 facilities of companies that participated in the US EPA 2010/2015 PFOA Stewardship Program (“major PFAS industrial facilities”); 521 airports certified in Title 14 Code of Federal Regulations, Part 139, which required operators to use aqueous fire-fighting foams (AFFF) (“AFFF-certified airports”); and 391 military fire-training areas (“MFTAs”).

Because wastewater treatment plant (WWTP) effluent is a potential source of all four target contaminants, we compiled effluent discharge data for 14,581 WWTPs included in the 2012 US EPA Clean Watersheds Needs Survey.⁴³ We calculated total daily WWTP effluent discharge for each county and normalized by county area (${10}^6{\mathrm{\hspace{0.17em}L}\text{\hspace{0.17em}per\hspace{0.17em}}\mathrm{km}}^2$) to account for potentially more concentrated impacts on water quality in smaller counties. For PWSs serving $>1$ county, we calculated a population-weighted average WWTP flow as $\sum w_i{p}_i/\sum p_i$, where w_i is wastewater flow (${10}^6{\mathrm{\hspace{0.17em}L}\text{\hspace{0.17em}per\hspace{0.17em}}\mathrm{km}}^2$) in county i and p_i is the county total population.

Sociodemographic Data

PWSs were matched to the county or counties they served as described by Schaider et al.⁹ In brief, counties served information from the US EPA’s Safe Drinking Water Information System (SDWIS) were cleaned and matched to sociodemographic information obtained from the US Census Bureau. We used county-level data because SDWIS does not provide information about cities served for 43% of UCMR 3 PWSs. We used county-level racial/ethnicity data from the 5-y 2010–2014 American Community Survey (ACS),⁴⁴ which best matched the period of water sample collection. Ethnicity was defined as the proportion of residents who identified as of Hispanic or Latino origin of any race (percentage Hispanic), and race was defined as the proportion of residents who identified as Black or African American alone and did not report their ethnicity as Hispanic (percentage Black). To account for the potential influence of urban vs. rural development, we also included urbanicity (percentage urban), which was calculated as the proportion of the county area (defined by census tracts or blocks) with $>\mathrm{50,000}$ people, from 2010 estimates.⁴⁵

We selected the multidimensional deprivation index (MDI) provided by the US Census Bureau as the primary SES variable and calculated a 5-y average based on values published for the period 2010–2014.⁴⁶ The MDI is a multivariate measure constructed using the Alkire-Foster method to capture six dimensions of SES (standard of living, education, health, economic security, housing quality, and neighborhood quality).⁴⁷ The MDI is expressed as a percent of people deprived in two or more dimensions (percentage deprived). For PWSs serving multiple counties, we calculated a population-weighted average for each demographic variable as $\sum x_i{p}_i/\sum p_i$, where x_i is percentage Hispanic, percentage Black, percentage urban, or percentage deprived in county i and p_i is the county total population. As a sensitivity analysis on our selection of MDI as our SES variable, we conducted parallel analyses with unidimensional SES measures from 2010 to 2014 ACS surveys: percentage of families and people whose income in the past 12 months fell below the poverty level (percentage poverty), percent of owner-occupied households (percentage homeownership), and the percentage of the civilian population with no health insurance coverage (percentage uninsured).

Altogether, the main analyses included a study sample of 4,815 PWSs (4,036 large and 779 small) in the 50 states and the District of Columbia that were included in UCMR 3 (2013–2015). The study sample comprised an estimated 3% of all US PWSs in 2013 and only included community water systems or nontransient noncommunity water systems (Table S1). Although large water systems in UCMR 3 represented 96% of all large US water systems in 2013, small UCMR 3 systems represented $<1\%$ of all small US water systems and had a higher proportion of noncommunity nontransient systems than large systems in UCMR 3 (Table S1).

Although 105 PWSs serving Tribal areas and US territories tested for our target contaminants during UCMR 3, we were not able to include these systems in our main analysis. Tribal PWSs had missing data for county or county-equivalent served, and most PWSs in US territories were missing MDI and comparable census data. Instead, we conducted a secondary analysis to compare detection frequencies in these areas in comparison with the areas in the main analysis.

Statistical Analyses

Unequal variance $t$-tests ($\mathrm{\alpha }=0.05$) were used to evaluate differences in average sociodemographic levels between counties with and without suspected point sources and between PWSs with and without contaminant detection or HRL exceedance. Correlations between covariates were evaluated with Spearman’s rank correlation. To identify predictors of target contaminants, we ran logistic regression models evaluating detection or exceedance as a function of each demographic characteristic, PWS characteristic, and the potential contaminant sources as single terms separately. Then, we used generalized logistic mixed-effects models to measure associations between target contaminants and demographic characteristics, PWS characteristics, and potential contaminant sources in the same model. Each mixed-effects model included a state intercept as a random effect to account for demographic clustering and clustering that occurs between systems subject to similar state-specific rules. Fully adjusted models included terms for percentage Hispanic, percentage Black, percentage deprived, percentage urban, system size (small/large), source water type (SW/GW/MIX), number of samples collected, WWTP flow (volume per area), and potential point sources (absent/present; varied by outcome). Potential point sources were included in adjusted models to examine whether race, ethnicity, and SES were associated with contaminant detections independent of correlations between these factors and suspected sources. Models of $\ge 1$ target contaminant detection and $\ge 1$ HRL exceedance included a term for the presence of any relevant TRI facility, and models of detection for specific contaminants included a term for compound-specific sources. We also ran a mixed-effects model stratified by system size to account for the large difference in the number of large and small PWSs sampled in UCMR 3 because we hypothesized that associations between contaminant detection and explanatory variables would differ by system size. Odds ratios (ORs) and confidence intervals (CIs) were derived from exponentiated $\mathrm{\beta }$ coefficients. To facilitate interpretation, we used marginal analysis to estimate the effect of a 1-standard deviation increase in each continuous predictor on the probability of detections. Finally, to explore differences in unregulated contaminant detections between PWSs included in the main analysis and PWSs serving Tribal areas and US territories, we compared the number of systems that detected unregulated contaminants and exceeded HRLs with Fisher’s exact tests ($\mathrm{\alpha }=0.05$). Analyses were conducted using R (version 4.2.2; R Development Core Team).

Results

Contaminant Detections

Of the 4,815 PWSs in this study, 1,277 (27%) systems, together serving more than 97 million people, detected $\ge 1$ target contaminant (Table 2). In this analysis, 1,4-dioxane was detected most frequently, found in 22% of PWSs, followed by HCFC-22 (5.8%), 1,1-DCA (4.7%), and PFAS (4.0%) (Table 1). HRL exceedances for PFAS, 1,4-dioxane, or 1,1-DCA occurred in 378 (7.9%) PWSs, which served an estimated total of 32 million people (Table 2). Exceedance of the 1,4-dioxane HRL was the most frequent type of exceedance (7.1% of systems). In addition, 1,1-DCA was detected in 226 PWSs, all of which were large systems, and the HRL was exceeded by just one PWS.

Table 2.

Characteristics of US PWSs that reported results for target contaminants in UCMR 3 ($n=\mathrm{4,815}$) in 2013–2015 and median demographic values of the counties they serve.

	PWS ($n$)	Population served (in millions)	Percentage of PWSs with:		Number of samples ($\text{mean}\pm \text{SD}$)	Median county demographic characteristics (%)
			$\ge 1$ detection (%)	$\ge 1$ HRL exceedance (%)		Hispanic (Q1, Q3)	Non-Hispanic Black (Q1, Q3)	Deprived (Q1, Q3)	Urban (Q1, Q3)
Overall	4,815	243	26.5	7.9	$14.8\pm 21.4$	7.8 (3.6, 18.9)	5.9 (2.2, 13.5)	16.7 (12.3, 22.3)	76.6 (2.5, 95.0)
Detected a target contaminant
No	3,538	146	—	—	$12.5\pm 10.3$	7.5 (3.3, 18.0)	5.3 (1.8, 13.3)	16.7 (12.4, 22.1)	72.1 (0.0, 93.4)
Yes	1,277	97	—	—	$21.3\pm 37.1$	9.2 (4.2, 22.8)	8.0 (3.2, 14.6)	16.7 (11.6, 23.0)	86.9 (61.0, 96.9)
Exceeded a health-reference level
No	4,437	212	—	—	$13.7\pm 14.6$	7.7 (3.5, 18.8)	5.7 (2.1, 13.7)	16.7 (12.4, 22.1)	74.8 (0.3, 94.5)
Yes	378	32	—	—	$27.4\pm 56.4$	11.4 (4.2, 23.9)	8.0 (3.6, 12.1)	17.2 (11.9, 25.4)	89.8 (64.1, 98.7)
System size
Large	4,036	240	29.2	8.9	$16.1\pm 23.1$	8.1 (3.7, 20.0)	6.1 (2.4, 13.7)	16.7 (12.2, 22.1)	80.6 (38.8, 95.8)
Small	779	3	12.7	2.2	$8.1\pm 5.1$	5.2 (2.3, 15.4)	4.0 (1.1, 12.6)	17.4 (12.5, 23.5)	7.5 (0.0, 74.0)
Source water
GW	2,016	56	25.6	7.1	$14.3\pm 26.9$	7.4 (3.2, 18.3)	5.6 (1.9, 13.7)	16.2 (11.4, 22.2)	72.7 (0.0, 91.0)
MIX	837	67	44.2	15.2	$25.7\pm 26.1$	15.0 (6.1, 30.0)	5.7 (2.4, 9.9)	17.1 (12.4, 23.5)	89.5 (64.1, 97.3)
SW	1,962	119	19.9	5.5	$10.6\pm 5.1$	6.5 (3.3, 16.0)	6.2 (2.3, 14.9)	17.2 (12.8, 22.1)	76.4 (1.6, 95.6)

Note: County-level demographic data were obtained from American Community Surveys (ACS 2010–2014),^44,45 and deprivation rates were based on the US Census Bureau’s Multidimensional Deprivation Index (MDI 2010–2014).⁴⁶ Percentage urban was calculated as the proportion of the county area (defined by census tracts or blocks) with $>\mathrm{50,000}$ people and percentage deprived is the percentage of people deprived in two or more dimensions according to the MDI. —, not applicable; GW, groundwater; HCFC-22, chlorodifluoromethane; HRL, health reference level; MIX, groundwater under the influence of surface water or a combination of groundwater and surface water; PFAS, per- and polyfluoroalkyl substances; PWSs, public water systems; Q1, 25th percentile; Q3, 75th percentile; SD, standard deviation; SW, surface water; UCMR 3, Third Unregulated Contaminant Monitoring Rule (2013–2015).

The majority of UCMR 3 PWSs were large systems (84%) and either GW (42%) or SW (41%) systems (Table 2). Twenty-nine percent of large systems detected $\ge 1$ target contaminant in comparison with 13% of small systems (Table 2). From the crude model, large systems had 2.83 (95% CI: 2.28, 3.55) times the odds of detecting unregulated contaminants in comparison with small systems, with magnitudes of the ORs varying by contaminant (Table 3). The odds of PFAS detection were 6-fold higher among large systems in comparison with small systems. In the fully adjusted model, in comparison with small systems, large systems were more likely to detect $\ge 1$ target contaminant [adjusted OR $\left(\text{aOR}\right)=1.75$; (95% CI: 1.37, 2.24)]. Adjusted odds for large systems were higher than small systems for detecting 1,4-dioxane and PFAS but not HCFC-22.

Table 3.

Crude and adjusted ORs (95% CI) from models evaluating associations between UCMR 3 contaminant detection (2013–2015) in the United States with county-level demographics, PWS characteristics, and potential pollution sources. Adjusted ORs (95% CI) from logistic mixed-effects models evaluating associations between UCMR 3 contaminant detection with county-level demographics, PWS characteristics, and potential pollution sources.

	Detected $\ge 1$ target contaminant	PFAS	1,4-dioxane	1,1-dichloroethane	HCFC-22	Exceeded $\ge 1$ HRL
No. of PWSs	4,815	4,815	4,810	4,807	4,811	4,815
Percentage Hispanica
OR (95% CI)	1.01 (1.01, 1.02)	1.01 (1.00, 1.02)	1.01 (1.01, 1.02)	1.02 (1.01, 1.03)	1.02 (1.02, 1.03)	1.01 (1.01, 1.02)
aOR (95% CI)	1.02 (1.01, 1.03)	1.02 (1.00, 1.04)	1.02 (1.01, 1.03)	1.01 (0.99, 1.03)	1.02 (1.01, 1.04)	1.01 (0.99, 1.02)
Percentage non-Hispanic Black
OR (95% CI)	1.01 (1.00, 1.01)	1.01 (0.99, 1.02)	1.01 (1.00, 1.01)	0.99 (0.98, 1.00)	0.99 (0.97, 1.00)	1.00 (0.99, 1.01)
aOR (95% CI)	1.01 (1.00, 1.01)	0.99 (0.97, 1.01)	1.00 (0.99, 1.01)	1.00 (0.97, 1.02)	1.01 (0.99, 1.03)	0.99 (0.98, 1.01)
Percentage deprived
OR (95% CI)	1.00 (0.99, 1.00)	0.99 (0.97, 1.00)	1.00 (0.99, 1.00)	0.99 (0.98, 1.01)	0.99 (0.98, 1.00)	1.00 (0.99, 1.01)
aOR (95% CI)	1.00 (0.98, 1.01)	0.99 (0.96, 1.02)	1.00 (0.98, 1.01)	1.02 (0.99, 1.05)	0.97 (0.94, 1.00)	1.01 (1.00, 1.03)
Percentage urban
OR (95% CI)	1.01 (1.01, 1.01)	1.01 (1.01, 1.02)	1.01 (1.01, 1.01)	1.02 (1.01, 1.02)	1.01 (1.01, 1.02)	1.01 (1.01, 1.01)
aOR (95% CI)	1.01 (1.01, 1.01)	1.01 (1.00, 1.01)	1.01 (1.01, 1.01)	1.01 (1.00, 1.02)	1.01 (1.00, 1.01)	1.01 (1.00, 1.01)
Large system (Ref: Small system)b
OR (95% CI)	2.83 (2.28, 3.55)	6.26 (3.03, 16.0)	2.79 (2.20, 3.57)	—e	2.48 (1.62, 4.01)	4.40 (2.78, 7.49)
aOR (95% CI)	1.75 (1.37, 2.24)	3.77 (1.62, 8.75)	1.69 (1.29, 2.21)	—e	1.43 (0.88, 2.32)	2.81 (1.69, 4.68)
GW system (Ref: SW system)
OR (95% CI)	1.38 (1.19, 1.61)	1.51 (1.04, 2.20)	1.18 (1.01, 1.39)	21.8 (10.5, 55.7)	4.99 (3.44, 7.46)	1.31 (1.01, 1.70)
aOR (95% CI)	1.57 (1.30, 1.89)	1.75 (1.13, 2.70)	1.43 (1.18, 1.75)	30.3 (12.6, 72.6)	4.40 (2.87, 6.76)	1.75 (1.30, 2.37)
MIX system (Ref: SW system)
OR (95% CI)	3.18 (2.67, 3.80)	3.95 (2.73, 5.78)	2.66 (2.22, 3.19)	41.3 (19.6, 106)	7.62 (5.12, 11.7)	3.07 (2.34, 4.03)
aOR (95% CI)	2.49 (2.00, 3.11)	3.24 (2.07, 5.07)	2.12 (1.68, 2.66)	30.4 (12.4, 74.6)	3.87 (2.43, 6.16)	2.49 (1.79, 3.47)
Number of samplesc
OR (95% CI)	1.04 (1.03, 1.04)	1.02 (1.02, 1.03)	1.03 (1.03, 1.04)	1.04 (1.03, 1.04)	1.04 (1.04, 1.05)	1.03 (1.02, 1.03)
aOR (95% CI)	1.03 (1.02, 1.04)	1.01 (1.01, 1.02)	1.03 (1.02, 1.03)	1.02 (1.02, 1.03)	1.03 (1.03, 1.04)	1.02 (1.01, 1.02)
WWTP flow (${\text{million\hspace{0.17em}}\mathrm{L}/\mathrm{km}}^2$)
OR (95% CI)	1.14 (1.02, 1.28)	1.19 (0.97, 1.40)	1.18 (1.06, 1.33)	1.17 (0.95, 1.37)	1.07 (0.85, 1.27)	1.18 (1.00, 1.35)
aOR (95% CI)	0.91 (0.77, 1.06)	1.18 (0.91, 1.54)	0.98 (0.84, 1.14)	1.07 (0.76, 1.52)	1.15 (0.89, 1.49)	0.91 (0.71, 1.17)
Relevant TRI facility present (Ref: absent)d
OR (95% CI)	1.60 (1.37, 1.87)	1.89 (1.37, 2.59)	1.74 (1.48, 2.05)	1.74 (1.28, 2.34)	1.38 (1.03, 1.83)	2.29 (1.82, 2.87)
aOR (95% CI)	1.13 (0.92, 1.38)	—	—	—	—	1.65 (1.24, 2.20)
1,4-dioxane TRI facility present (Ref: absent)
OR (95% CI)	1.80 (1.47, 2.19)	2.18 (1.47, 3.14)	1.97 (1.60, 2.42)	2.19 (1.52, 3.08)	1.78 (1.25, 2.48)	3.01 (2.30, 3.91)
aOR (95% CI)	—	—	1.57 (1.22, 2.03)	—	—	—
Chlorinated solvent TRI facility present (Ref: absent)
OR (95% CI)	1.56 (1.23, 1.96)	1.37 (0.80, 2.19)	1.62 (1.27, 2.06)	2.10 (1.38, 3.09)	1.37 (0.88, 2.05)	2.06 (1.47, 2.82)
aOR (95% CI)	—	—	—	1.68 (1.00, 2.83)	—	—
Chlorofluorocarbon TRI facility present (Ref: absent)
OR (95% CI)	1.53 (1.28, 1.82)	1.45 (0.97, 2.09)	1.71 (1.42, 2.06)	2.08 (1.49, 2.85)	1.22 (0.86, 1.70)	1.90 (1.45, 2.46)
aOR (95% CI)	—	—	—	—	0.84 (0.55, 1.28)	—
Major industrial PFAS facility present (Ref: absent)
OR (95% CI)	3.13 (1.85, 5.31)	10.1 (5.41, 18.0)	2.99 (1.76, 5.06)	0.73 (0.12, 2.38)	1.22 (0.37, 3.02)	3.56 (1.82, 6.47)
aOR (95% CI)	—	5.16 (2.47, 10.7)	—	—	—	—
AFFF-certified airports or military fire-training areas present (Ref: absent)
OR (95% CI)	1.49 (1.31, 1.69)	1.75 (1.30, 2.37)	1.41 (1.23, 1.61)	2.62 (1.95, 3.57)	2.39 (1.84, 3.13)	1.73 (1.39, 2.15)
aOR (95% CI)	—	1.48 (0.98, 2.25)	—	—	—	—

Note: Crude ORs were derived from coefficients in logistic regression models. Adjusted ORs were based on coefficients from logistic mixed-effects models adjusted for percentage Hispanic residents, percentage Black residents, percentage deprived residents, percentage urban areas, system size, source water type, number of samples, wastewater flow, a random state intercept, and the presence of suspected point sources. For detected $\ge 1$ target contaminant and exceeded $\ge 1$ HRL, models were adjusted for TRI facility (present or absent) reporting any environmental release of 1,4-dioxane, 1,1-dichloroethane, 1,1,1-trichloroethane, HCFC-22, or CFC-12 in 2010–2015. For 1,4-dioxane, 1,1-dichloroethane, and HCFC-22, models were adjusted for TRI facility (present or absent) emitting the specific contaminant and relevant precursors, and for PFAS, the model was adjusted for the presence of major industrial PFAS facilities and AFFF sources. County-level demographic data were obtained from American Community Surveys (ACS 2010–2014)^44,45 and deprivation rates were based on the US Census Bureau’s Multidimensional Deprivation Index (MDI 2010–2014).⁴⁶ Percentage urban was calculated as the proportion of the county area (defined by census tracts or blocks) with $>\mathrm{50,000}$ people and percent deprived is the percentage of people deprived in two or more dimensions according to the MDI. Tables of crude and adjusted model presented separately, including corresponding Wald $p$-values, are in shown Supplemental Tables S4–S5. —, not applicable; aOR, adjusted odds ratio; AFFF, aqueous film-forming foam; CI, confidence interval; GW, groundwater; HCFC-22, chlorodifluoromethane; HRL, health reference level; MIX, groundwater under the influence of surface water or a combination of groundwater and surface water; OR, odds ratio; PFAS, per- and polyfluoroalkyl substances; PWS, public water system; Ref, reference; SW, surface water; TRI, Toxics Release Inventory; UCMR 3, Third Unregulated Contaminant Monitoring Rule (2013–2015); WWTP, wastewater treatment plant.

^a For demographic variables, ORs are reported for each one percentage-point increase in each variable.

^b For PWS characteristics, ORs represent odds of detection or exceedance relative to referent group.

^c For number of samples and WWTP flow, ORs are reported for each additional sample tested or each ${1\text{\hspace{0.17em}million\hspace{0.17em}}\mathrm{L}/\mathrm{km}}^2$ increase in flow.

^d For potential pollution sources, ORs represent the odds of detection or exceedance among PWSs serving counties with at least one potential pollution source relative to counties with none.

^e Only large PWSs detected 1,1-dichloroethane.

MIX systems were most likely to detect any target contaminant (44%), followed by GW (26%) and SW systems (20%). Both crude and adjusted models indicated that, in comparison with SW systems, the odds of $\ge 1$ target contaminant detection and $\ge 1$ HRL exceedance were higher for GW and MIX systems (Table 3). MIX coefficients were greater in magnitude in comparison with GW or SW systems in most models. The difference in adjusted odds of detection between GW and MIX systems was highest for 1,4-dioxane and PFAS (Table 3).

Suspected Point Sources and Contaminant Detections

The presence of potential pollutant sources was generally associated with target contaminant detections. For instance, in crude regression models, PWSs that served counties with a major PFAS industry were 10.1 (95% CI: 5.4, 18.0) times more likely to detect PFAS than systems that did not, and systems that served counties with an MFTA or AFFF-certified airport had 1.75 (95% CI: 1.30, 2.37) times the odds in comparison with systems that did not (Table 3). The presence of at least one relevant TRI facility was positively associated with unregulated contaminant detection in the crude model ($\text{OR}=1.60$; 95% CI: 1.37, 1.87) but not the adjusted ($\text{aOR}=1.13$; 95% CI: 0.92, 1.38) (Table 3). In total, relevant TRI facilities were present in 131 out of 1,720 (8%) of counties served by a UCMR PWS (Table S2). TRI facilities reporting chlorofluorocarbon emissions (CFC TRI facilities) were the most prevalent types of facilities, present in 5% of counties included in our study and in counties served by 13% of PWSs in our sample. The PWSs that served counties with a 1,4-dioxane TRI facility had 1.97 (95% CI: 1.60, 2.42) times the odds of detecting 1,4-dioxane than systems that did not, and these associations persisted after adjustment (Table 3). HCFC-22 detection, on the other hand, was not associated with CFC TRI facilities in either crude ($\text{OR}=1.22$; 95% CI: 0.86, 1.70) or adjusted ($\text{aOR}=0.84$; 95% CI: 0.55, 1.28) models, but detection was positively associated with 1,4-dioxane TRI facilities ($\text{OR}=1.78$; 95% CI: 1.25, 2.48) and the presence of AFFF-certified airports or MFTAs ($\text{OR}=2.39$; 95% CI: 1.84, 3.13) (Table 3). The presence of 1,4-dioxane TRI facilities, AFFF-certified airports, and MFTAs were all associated with detecting each target contaminant individually (Table 3).

We observed some associations between WWTP flow and target contaminant detections. Crude ORs indicated each ${\text{million\hspace{0.17em}}\mathrm{L}/\mathrm{km}}^2$ increase in WWTP flow was associated with 14% higher odds of detecting a target contaminant and 18% higher odds of exceeding an HRL (Table 3). Crude models indicated WWTP flow was a predictor of 1,4-dioxane detection, but this association was not significant after adjusting for other covariates (Table 3).

Presence of Suspected Potential Sources and Sociodemographics

The presence of suspected point sources was associated with higher proportions of Hispanic and Black residents and higher urbanicity, but generally not with higher proportions of deprived residents. Counties with at least one relevant TRI facility had higher mean percentage Hispanic residents (12%), percentage Black residents (14%), and percentage urban areas (64%) in comparison with counties without (9%, 10%, and 30%, respectively) (Figure 1; Table S2). Counties with a CFC TRI facility had higher mean percentage Hispanic residents, percentage Black residents, and percentage urban areas. Similarly, the presence of PFAS sources was associated with higher percentage Hispanic residents and percentage urban areas, and marginally with higher percentage Black residents. In contrast, percentage deprived was generally not higher in counties with relevant TRI facilities. In fact, counties with PFAS sources had lower mean percentage deprived (19%) in comparison with counties without (21%). In addition, counties with suspected point sources had both lower rates of homeownership (67%) and lower rates of poverty (15%) in comparison with counties without these sources (70% and 17%, respectively), with no difference by percentage uninsured (Table S2). Wastewater flow was highly correlated with urbanicity ($\mathrm{\rho }=0.80$) and moderately correlated with percentage Hispanic ($\mathrm{\rho }=0.34$) and percentage Black ($\mathrm{\rho }=0.34$) residents (Figure S1).

Figure 1.

Mean demographics among US counties linked to a UCMR 3 public water system ($n=\mathrm{1,720}$) where potential sources of unregulated contaminants were present (Yes) and absent (No) in 2013–2015 results. In all graphs, error bars denote one standard error, and $p$-values above bars were calculated from unequal variances $t$-tests assessing differences in averages between subgroups. Note: County-level demographic data were obtained from American Community Surveys (ACS 2010–2014),^44,45 and deprivation rates were based on the US Census Bureau’s Multidimensional Deprivation Index (MDI 2010–2014).⁴⁶ Percentage urban was calculated as the proportion of the county area (defined by census tracts or blocks) with $>\mathrm{50,000}$ people, and percentage deprived is the percentage of people deprived in two or more dimensions according to the MDI. Corresponding numerical results can be found in Table S2. Relevant TRI facility refers to an industry reporting any environmental release of 1,4-dioxane, 1,1-dichloroethane, 1,1,1-trichloroethane, HCFC-22, or CFC-12 to the Toxics Release Inventory in 2010–2015. Note: CFC-12, dichlorodifluoromethane; HCFC-22, chlorodifluoromethane; MFTA, military fire-training area; PFAS, per- and polyfluoroalkyl substances; UCMR 3, Third Unregulated Contaminant Monitoring Rule.

Target Contaminant Detections and Race/Ethnicity

We also observed associations between contaminant detections and percentage Hispanic and percentage Black residents among UCMR 3 systems. PWSs with $\ge 1$ target contaminant detection served counties with higher average percentage Hispanic residents (17% among PWSs with $\ge 1$ detection vs. 13% among PWSs without detection, $p<0.001$) and modestly higher average percentage Black residents (11% vs. 10%, $p=0.04$) (Figure 2; Table S3). PWSs with an HRL exceedance had higher percentage Hispanic residents (18% vs. 14%, $p<0.001$) but not higher percentage Black residents (10% vs. 10%, $p=0.56$) (Figure 2; Table S3).

Figure 2.

Mean county-level demographics among US PWSs with unregulated contaminant detections (Yes; $n=\mathrm{1,277}$) and without (No; $n=\mathrm{3,538}$), and among PWSs with an HRL exceedance (Yes; $n=378$) and without (No; $n=\mathrm{4,437}$) in UCMR 3 (2013–2015) results. Note: County-level demographic data were obtained from American Community Surveys (ACS 2010–2014),^44,45 and deprivation rates were based on the US Census Bureau’s Multidimensional Deprivation Index (MDI 2010–2014).⁴⁶ Percentage urban was calculated as the proportion of the county area (defined by census tracts or blocks) with $>\mathrm{50,000}$ people and percentage deprived is the percentage of people deprived in two or more dimensions according to the MDI. Corresponding numerical results can be found in Table S3. In all graphs, error bars denote one standard error, and $p$-values above bars were from unequal variances $t$-tests assessing differences in averages between subgroups. Note: HRL, health-reference level; PWS, public water system; UCMR 3, Third Unregulated Contaminant Monitoring Rule (2013–2015).

Overall, percentage of Hispanic residents was associated with small increased odds of detecting a target contaminant and exceeding an HRL across different models. In crude models (Table 3, Tables S4–S5), each one percentage point increase in percentage Hispanic residents was associated with 1.01 (95% CI: 1.01, 1.02) times the odds of detecting any target contaminant or exceeding an HRL. Crude ORs were similar across contaminants, ranging from 1.01 (95% CI: 1.0, 1.02) times the odds of detecting PFAS to 1.02 (95% CI: 1.02, 1.03) times the odds of detecting HCFC-22, per 1 percentage point increase in percentage Hispanic residents. From fully adjusted models that account for other county demographics, PWS characteristics, and suspected point sources, a 1 percentage point increase in percentage Hispanic residents was associated with 1.02 (95% CI: 1.01, 1.03) times the odds of detecting a target contaminant. Adjusted odds ratios for percentage Hispanic residents across models of individual target contaminants were similar in magnitude with minor differences in precision (Table 3). Converting these odds to a linear probability scale, we estimated that a 1 standard deviation increase in percentage Hispanic residents (15.5 percentage points) was associated with a 5 percentage point increase in the likelihood of target contaminant detection.

We observed less consistent associations between percentage Black residents and target contaminant detections. A small positive crude OR was observed between detection of a target contaminant and percentage Black residents [$\text{OR}=1.01$ (95% CI: 1.00, 1.01)] (Table 3). Furthermore, crude ORs for percentage Black residents indicated a positive association with 1,4-dioxane detection [$\text{OR}=1.01$ (95% CI: 1.00, 1.01)] but an inverse association with HCFC-22 detection [$\text{OR}=0.99$ (95% CI: 0.97, 1.00)] (Table 3). Adjusted ORs for percentage Black residents and specific target contaminant detections were similar in magnitude, though imprecise, in comparison with crude ORs (Table 3). In a sensitivity analysis to evaluate whether suspected pollution source terms altered associations between race/ethnicity and contaminant detections, we reran our adjusted mixed-effects models excluding suspected pollution source terms and observed minimal differences in coefficients for percentage Hispanic residents and percentage Black residents (Table S6).

Associations with Urbanicity and SES

In comparison with systems without target contaminant detections, systems with detections on average served counties with higher percentage urban areas (71% vs. 56%, $p<0.001$) but were not different in average percentage deprived (19% vs. 19%, $p=0.15$) (Figure 2). Consistent with the bivariate results, crude ORs for percentage urban were positive and similar in magnitude across detection outcomes (Table 3). Adjusted ORs for percentage urban were somewhat smaller than crude and remained above the null for all outcomes except PFAS detection (Table 3). Furthermore, crude and aORs indicated no relationship between higher percentage deprived and the odds of $\ge 1$ target contaminant detection, although adjusted odds indicated an inverse association with detection of HCFC-22 [$\text{aOR}=0.97$ (95% CI: 0.94, 1.00)] (Table 3). We found similar relationships between other SES measures (percentage in poverty, percentage of home ownership, percentage uninsured) and target contaminant detections in sensitivity analyses of the adjusted models (Tables S7–S9). In parallel adjusted models, two indicators of higher SES (lower percentage in poverty and higher percentage of home ownership but not lower percentage uninsured), were associated with higher odds of $\ge 1$ target contaminant detection. Lower percentage poverty was associated with higher odds of detecting 1,4-dioxane, 1,1-DCA, and HCFC-22, but not PFAS (Table S7), and higher percentage home ownership was associated with higher odds of detecting 1,4-dioxane and 1,1-DCA and exceeding an HRL (Table S8). Higher percentage uninsured was not associated with $\ge 1$ target contaminant detection and was inversely associated with detections of 1,1-DCA and HCFC-22 (Table S9).

System Size-Stratified Associations

Because the UCMR 3 oversampled large systems in the United States relative to small systems, we ran an adjusted model stratified by system size. Size-stratified associations for percentage Hispanic residents and percentage Black residents were similar in magnitude between large and small systems and compared with associations from nonstratified models (Tables 3 and Table 4). In addition, the direction and precision of source water type as a predictor differed by system size. In comparison with large SW systems, large GW and large MIX systems had 1.81 (95% CI: 1.48, 2.20) and 2.74 (95% CI: 2.18, 3.46) times the odds, respectively, of detecting $\ge 1$ target contaminant (Table 4). By contrast, small GW systems were less likely to detect $\ge 1$ target contaminant than small SW systems ($\text{aOR}=0.44$; 95% CI: 0.26, 0.73), and there were no differences in the odds of detection between small MIX and small SW systems ($\text{aOR}=1.03$; 95% CI: 0.49, 2.18) (Table 4).

Table 4.

Adjusted ORs (95% CI) from system size-stratified logistic mixed-effects model assessing the odds of detecting $\ge 1$ target contaminant in the United States during UCMR 3 (2013–2015).

	Large water systems ($n=\mathrm{4,036}$)		Small water systems ($n=779$)
	Adjusted OR (95% CI)	$p$-Value	Adjusted OR (95% CI)	$p$-Value
Percentage Hispanica	1.02 (1.01, 1.03)	$<0.001$	1.02 (0.99, 1.04)	0.14
Percentage Black, non-Hispanic	1.01 (1.00, 1.02)	0.27	1.02 (0.99, 1.04)	0.19
Percentage deprived	1.00 (0.98, 1.01)	0.43	0.99 (0.96, 1.02)	0.53
Percentage urban	1.01 (1.00, 1.01)	$<0.001$	1.01 (1.00, 1.02)	0.01
Source water: GW (Ref: SW)b	1.81 (1.48, 2.20)	$<0.001$	0.44 (0.26, 0.73)	0.00
Source water: MIX (Ref: SW)	2.74 (2.18, 3.46)	$<0.001$	1.03 (0.49, 2.18)	0.93
Number of samplesc	1.03 (1.02, 1.04)	$<0.001$	1.04 (1.00, 1.09)	0.06
WWTP flow (${\text{million}\hspace{0.17em}\mathrm{L}\hspace{0.17em}\text{per}\hspace{0.17em}\text{km}}^2$)	0.91 (0.78, 1.08)	0.28	0.93 (0.33, 2.59)	0.89
Relevant TRI facility present (ref: absent)d	1.11 (0.90, 1.38)	0.32	1.24 (0.61, 2.53)	0.55

Note: Adjusted ORs were based on coefficients from logistic mixed-effects model stratified by size and adjusted for percentage Hispanic residents, percentage Black residents, percentage deprived residents, percentage urban areas, source water type, number of samples, wastewater flow, a random state intercept, and the presence (or absence) of TRI facility reporting any environmental release of 1,4-dioxane, 1,1-dichloroethane, 1,1,1-trichloroethane, HCFC-22, or CFC-12 in 2010–2015. County-level demographic data were obtained from American Community Surveys (ACS 2010–2014)^44,45 and deprivation rates were based on the US Census Bureau’s Multidimensional Deprivation Index (MDI 2010–2014).⁴⁶ Percentage urban was calculated as the proportion of the county area (defined by census tracts or blocks) with $>\mathrm{50,000}$ people and percentage deprived is the percentage of people deprived in two or more dimensions according to the MDI. CI, confidence interval; GW, groundwater; MIX, groundwater under the influence of surface water or a combination of groundwater and surface water; OR, odds ratio; Ref, reference; SW, surface water; TRI, Toxics Release Inventory; UCMR 3, Third Unregulated Contaminant Monitoring Rule (2013–2015); WWTP, wastewater treatment plant.

^a For demographic variables, ORs are reported for each one percentage-point increase in each variable.

^b For PWS characteristics, ORs represent odds of detection or exceedance relative to referent group.

^c For number of samples and WWTP flow, ORs are reported for each additional sample tested or each ${1\text{\hspace{0.17em}million\hspace{0.17em}}\mathrm{L}/\mathrm{km}}^2$ increase in flow.

^d For potential pollution sources, odds ratios represent the odds of detection or exceedance among PWSs serving counties with at least one potential pollution source relative to counties with none.

Tribal PWSs and PWSs in US Territories

A total of 105 PWSs located on Tribal land ($n=29$) or in a US territory ($n=76$) were tested for target contaminants during UCMR 3. PWSs on Tribal lands or in US territories constituted $<2\%$ of PWSs in UCMR 3. US territory-serving PWSs were overrepresented among systems that detected PFAS and HCFC-22. Although only 1.5% of all PWSs served US territories, 2.0% of PWSs that detected PFAS and 2.1% of PWSs that detected HCFC-22 served territories (Table S10). Furthermore, the detection frequency for PFAS was slightly higher among PWSs serving territories (5.4%) in comparison with those serving US states (4.0%), although this difference was not significant ($p=0.61$) (Table S11). By contrast, detection frequencies for other target contaminants among PWSs serving Tribal lands or US territories were similar to or lower than detection frequencies for PWSs serving US states and the District of Columbia (Table S11).

Discussion

This study showed that public water systems serving communities with higher proportions of Hispanic and non-Hispanic Black residents were more likely to detect unregulated industrial contaminants, including PFAS and 1,4-dioxane. Our findings add to growing evidence of sociodemographic disparities in exposures to drinking water contaminants and indicate that communities of color may have elevated exposures to mixtures of both regulated and unregulated contaminants. Furthermore, we observed that potential contaminant sources were associated with higher proportions of Hispanic and Black residents, consistent with long-standing evidence that hazardous sites are located in communities with higher proportions of Hispanic and Black people.¹² In adjusted logistic models, target contaminant detections were associated with proportions of Hispanic residents even after adjusting for other demographic variables, PWS characteristics, and suspected pollution sources. Because these associations persisted even while accounting for suspected sources, the relationship between drinking water quality and the proportion of Hispanic residents may be related to social and political factors that determine PWS capacity to address contamination. Identifying factors associated with unregulated contaminants in drinking water can guide management and policy approaches to support communities in protecting source water quality ahead of federal regulations.

In our analyses, percentage Hispanic residents and percentage Black residents were both associated with unregulated contaminant detections in drinking water in bivariate analyses and crude regression models. Prior studies have found higher odds of PFAS detections among PWSs drawing from watersheds near pollution sources^30,38 and associations with PFAS sources and percentage Hispanic residents and percentage Black residents.³⁰ In our adjusted models, we evaluated the associations of community demographics with contaminant detections while accounting for urbanicity, proximity to pollution sources, and PWS characteristics. Although residual confounding from other factors may be present, the consistent small positive associations of both percentage Hispanic residents and percentage Black residents with contaminant detections in our adjusted models suggest that racial and ethnic disparities in target contaminant detections cannot be explained by proximity to pollution sources alone. In addition, our analysis provides evidence of inequities in exposure to hazardous chemicals via drinking water, particularly among communities of color that may have increased vulnerability to health impacts from other environmental exposures and nonenvironmental stressors.⁴⁸ In contrast to prior studies that investigated disparities in exposures to regulated drinking water contaminants, all target contaminants in our study were unregulated at the time of UCMR 3 monitoring, so our observed associations do not reflect disparities in enforcement, but rather disparities in source water quality and factors related to technical, managerial, and financial capacities of PWSs.

Percentage of Hispanic residents was consistently associated with detection of target contaminants (individually and overall) across crude, adjusted, and stratified models and across various metrics of SES. This finding is consistent with previous studies that found positive associations between percentage Hispanic residents and regulated drinking water contaminants, suggesting that communities with higher proportions of Hispanic residents are exposed to higher levels of both regulated and unregulated contaminants. Higher proportions of Hispanic residents have been linked to more frequent detections and higher concentrations of regulated contaminants such as nitrate^9,19 and arsenic^{10,11,18,21,49} and to more frequent SDWA violations,^8,13 especially in the Southwest region. Recent studies have similarly reported similar trends for PFAS in drinking water.^28–30 Liddie et al.³⁰ examined PFAS detections among 7,873 community water systems in 18 US states where PFAS sampling was conducted. They reported a positive association between detections of any of five PFAS (PFOA, PFOS, PFNA, PFHxS, PFBS) with the proportion of Hispanic residents [$\text{percent\hspace{0.17em}change}=1.7\%$ (95% CI: $-0.1\%$, 3.5%) higher odds per one-unit increase in percentage Hispanic residents after accounting for water system characteristics and pollution sources], which was similar in magnitude to estimates in the current study. Liddie et al. had a lower reporting limit for PFAS detections ($5\text{\hspace{0.17em}ng}/\mathrm{L}$), in comparison with much higher reporting limits in UCMR 3 ($10–90\text{\hspace{0.17em}ng}/\mathrm{L}$) used in the present study, indicating that our observed associations for percentage Hispanic residents and unregulated contaminant detections at the national scale are noteworthy because similar associations have been found across a range of contaminants and among multiple studies using different geographic scales.

Percentage of Black residents was associated with target contaminant detections in crude models, but fewer significant findings were observed in adjusted models. One reason for this difference could be the narrower distribution of county-level percentage Black values [interquartile range (IQR): 2.2%–13.5%] in comparison with county-level percentage Hispanic values (IQR: 3.6%–18.9%), which may have limited our ability to detect an association. Furthermore, among UCMR 3 systems, PWSs serving counties in the US South, which had the highest median percentage Black residents (12%) in comparison with other regions in the United States (5% for the Northeast, 4% for the North Central, and 2% for the West), also had the lowest prevalence of sources (7% of PWSs serving a county with at least one source vs. 11% in Northeast, 10% in North Central, and 13% in West) and lowest mean WWTP flows ($0{.08\text{\hspace{0.17em}million\hspace{0.17em}}\mathrm{L}/\mathrm{km}}^2$ vs. 0.28 for Northeast, 0.32 for North Central, and 0.10 for the West). Thus, regional differences in sources may account for some of the associations in the adjusted models, and associations between race and unregulated contaminants may vary by region.

In contrast to the associations we found with ethnicity and race, we generally did not find associations between deprivation (an indicator of lower SES) and target contaminant detections. Our analyses primarily relied on MDI, an index that has been used in other environmental health disparity studies^50–53 and as an official measure of poverty in several countries⁵⁴ but has been rarely used in studies of drinking water quality. Overall, percentage deprived was not positively associated with detections of PFAS, 1,4-dioxane, and 1,1-DCA, and many of the model coefficients for these contaminants trended in a negative direction. In prior studies, associations between SES and drinking water contaminants have been mixed.^9,20,30 For example, associations between SES and nitrate levels varied in direction and magnitude according to SES indicators (e.g., poverty or homeownership) and spatial scale (e.g., county or city).⁹ Other studies using composite SES variables concluded that lower-SES communities, in particular systems serving lower SES rural communities, were more likely to incur SDWA violations.^7,13 Similar to communities of color, lower-SES communities may experience higher exposures to drinking water contaminants related to closer proximity to potential sources and limited technical, managerial, and financial capacities. The lack of relationship between lower SES and detection of unregulated industrial contaminants in our study may reflect associations between industrial activity and higher income levels; these associations should be explored in future studies.

In our analysis, we generally found that large PWSs and PWSs with GW or MIX source water types had higher odds of target contaminant detections. Our findings are consistent with prior studies on 1,4-dioxane⁵⁵ and PFAS^37,38 that found that large systems and groundwater-based systems were more likely to detect 1,4-dioxane and PFAS. We also found that large systems were more likely to exceed an HRL compared with small systems. Higher levels of industrial contaminants in groundwater compared with surface water likely reflect source, transport, and fate characteristics of target contaminants.^37,55–57 For instance, groundwater contamination may reflect land releases of target contaminants from disposal or spills from manufacturing facilities, application of certain products (e.g., AFFF) on soil surfaces, or transformation of precursor compounds (e.g., transformation of 1,1,1-trichloroethane to 1,1-DCA) already in GW. Contaminants discharged into SWs may undergo a higher degree of dilution, and sources that discharge into SWs (e.g., WWTPs) may have lower target contaminant concentrations than direct discharges into groundwater.⁵⁵ System size and source water type were interrelated predictors of unregulated contaminant detections. After stratifying by system size, we observed that small GW systems were less likely to detect any target contaminant than small SW systems, whereas large GW systems were more likely to detect any target contaminant than large SW systems. These differences may be attributable to the distribution of these systems among urban and rural areas and the proximity to pollution sources. Our observation that MIX systems had higher target contaminant detection frequencies may be explained by greater urbanicity as well as higher mean number of samples tested (Table 2). Related to demographics, the median percentage of Hispanic residents for MIX systems was more than twice that of GW or SW systems (Table 2), highlighting the need to account for PWS characteristics when evaluating associations between demographic variables and water quality.

PWSs in US territories and Tribal areas may face unique challenges in providing safe drinking water due to differences in regulatory status, political disenfranchisement, higher rates of poverty, and geographic isolation.^14,58,59 We found that the detection frequency for PFAS was slightly higher among PWSs serving territories and that PWSs serving US territories were slightly overrepresented among systems that detected PFAS and HCFC-22. Notably, the highest concentrations of PFOS ($\mathrm{7,000}\text{\hspace{0.17em}ng}/\mathrm{L}$), PFHxS ($\mathrm{1,600}\text{\hspace{0.17em}ng}/\mathrm{L}$), and PFHpA ($410\text{\hspace{0.17em}ng}/\mathrm{L}$) were found in a PWS serving the island territory of Saipan. Overall, our ability to evaluate disparities in exposures to unregulated contaminants among residents in territories and Tribal areas was limited because of the small sample size and lack of comparable demographic information. A smaller proportion of Tribal water systems was tested during UCMR 3 in comparison with PWSs overall because they are more likely to be small systems.⁶⁰ Future monitoring for unregulated contaminants should ensure equitable representation of Tribal areas and territories, and studies on environmental justice and water quality should include these vulnerable communities to the extent possible.

Two major strengths of this study are inclusion of multiple unregulated contaminants beyond PFAS and the characterization of associations between contaminants, sources, and demographics on a national scale, including territories and Tribal areas. We synthesized comprehensive datasets on potential sources of target contaminants to distinguish the role of potential pollution sources from other explanatory factors. By analyzing unregulated contaminants prior to the adoption of any federal drinking water standards, we were able to examine causes of disparities that were independent of inequities in SDWA implementation or enforcement. This focus can help identify and assess built factors that contribute to differences in community exposure to hazardous chemicals, and, in particular, those factors that contribute to differences in source water quality. Our study is timely because, in 2023, the US EPA initiated UCMR 5, which includes 29 PFAS with reporting limits mostly $\le 5\text{\hspace{0.17em}ng}/\mathrm{L}$,⁶¹ so our methodology can be applied to evaluating whether disparities exist in exposures to a broader set of PFAS in drinking water.

Our study has some important limitations. First, we were limited in our ability to characterize the demographics of customers served by each PWS because we had to rely on county-level demographic information on race, ethnicity, SES, and urbanicity, which may not reflect the demographics of specific areas served by each PWS. PWSs do not regularly collect demographic data on the customers they serve, which has been a barrier to assessing demographic disparities in drinking water quality.¹⁷ Thus, our analyses may be missing underlying relationships because of imprecise estimates of the demographics of populations actually served by PWSs. Second, the PWSs included in UCMR 3 were not representative of all PWSs in the United States because monitoring was required by all large PWSs but only a subset of small PWSs (Table S1). UCMR 3 overrepresented large systems by design to characterize water quality among PWSs that serve the majority of public water supply customers in the United States. However, this sample design may limit our ability to generalize our results to smaller PWSs, which constitute the largest proportion of the PWSs in the United States.⁶² In addition, UCMR 3 testing and the SDWA more generally do not include private wells, which serve an estimated 42 million people, approximately 13% of the US population,⁶³ even though these systems may be more susceptible to impacts of contamination. Private well users test less frequently and may have fewer resources to respond to contamination once found. Third, suspected pollution sources were assigned to counties based on emissions reporting in the period 2010–2015, which may miss TRI facilities that did not report during that specific time frame and may miss other facilities that discharged target contaminants (or their precursors) but are not required to report through TRI. In addition, we could not account for contaminant transport into adjacent counties that were not listed in TRI. Further, we were not able to account for the location of pollution sources relative to drinking water sources and localized fate and transport processes. These factors led to possible misclassification of relevant pollution sources within each county and may have biased our estimates toward the null. A final limitation was that the PFAS testing during UCMR 3 underestimated the extent of drinking water contamination because of relatively high reporting limits, leading to low detection frequencies, which limited our statistical power and may have biased our estimates toward the null. In addition, our comparisons of PFAS levels to HRLs was limited because the reporting limits in UCMR 3 were much higher than the 2024 US EPA MCLs. Overall, this study underscores the need to enhance drinking water infrastructure and protect source water quality from contamination to reduce possible downstream inequities of drinking water contaminants.

Conclusions

Our analysis found more frequent detections of unregulated industrial contaminants in public water systems serving higher proportions of Hispanic and non-Hispanic Black residents. These associations persisted even after taking other factors into account (e.g., SES, proximity to pollution sources). Although some of the associations between race, ethnicity, and unregulated contaminants can be explained by disproportionate siting of industrial facilities, our analysis suggests that other factors contribute to disparities in exposures to unregulated drinking water contaminants. Our analysis indicates that drinking water disparities can be overlooked and underestimated when research narrowly focuses on regulated chemicals in drinking water. Elucidating other vulnerability factors and mechanisms by which demographics influence drinking water can promote actions to implement the SDWA more effectively and equitably.

Conclusions and opinions are those of the individual authors and do not necessarily reflect the policies or views of EHP Publishing or the National Institute of Environmental Health Sciences.