County-Level Associations between Drinking Water PFAS Contamination and COVID-19 Mortality in the United States

Abstract

Background:

Epidemiologic and animal studies both support relationships between exposures to per- and polyfluoroalkyl substances (PFAS) and harmful effects on the immune system. Accordingly, PFAS have been identified as potential environmental risk factors for adverse COVID-19 outcomes.

Objective:

Here, we examine associations between PFAS contamination of U.S. community water systems (CWS) and county-level COVID-19 mortality records. Our analyses leverage two datasets: one at the subnational scale (5,371 CWS serving 621 counties) and one at the national scale (4,798 CWS serving 1,677 counties). The subnational monitoring dataset was obtained from statewide drinking monitoring of PFAS (2016–2020) and the national monitoring dataset was obtained from a survey of unregulated contaminants (2013–2015).

Methods:

We conducted parallel analyses using multilevel quasi-Poisson regressions to estimate cumulative incidence ratios for the association between county-level measures of PFAS drinking water contamination and COVID-19 mortality prior to vaccination onset (Jan-Dec 2020). In the primary analyses, these regressions were adjusted for several county-level sociodemographic factors, days since the first reported case in the county, and total hospital beds.

Results:

In the subnational analysis, detection of at least one PFAS over 5 ng/L was associated with 12% higher [95% CI: 4%, 19%] COVID-19 mortality. In the national analysis, detection of at least one PFAS above the reporting limits (20–90 ng/L) was associated with 13% higher [95% CI: 8%, 19%] COVID-19 mortality.

Significance:

Our findings support a positive association between county-level drinking water PFAS contamination and COVID-19 mortality. These results reinforce the importance of systematic federal drinking water monitoring efforts for PFAS and remediation of contaminated areas for public health protection, potentially including infectious disease epidemics.

INTRODUCTION

Per- and polyfluoroalkyl substances (PFAS) are a diverse group of highly fluorinated anthropogenic chemicals that are widely used in various consumer products and by industry (1,2). Exposures to PFAS can occur via several routes and extensive PFAS contamination of drinking water has been documented throughout the United States (U.S.) (3). More than 200 million U.S. residents are estimated to consume drinking water with perfluorooctanoic acid (PFOA) and perfluorooctanesulfonic acid (PFOS) concentrations at or above 1 ng/L, a proposed benchmark dose for immunotoxicity (4,5). Both animal and epidemiologic studies report associations between PFAS exposure and diverse health outcomes, including harmful effects on the immune system (3). However, few studies have investigated whether PFAS exposures were associated with severe COVID-19 outcomes or mortality during the pandemic period prior to availability of vaccines.

Past work has suggested that PFAS exposures can suppress both innate and adaptive immune responses (6,7,3). Lowered antibody production, increased emaciation, and increased mortality were observed for mice exposed to higher levels of PFOS in response to infection with the influenza A virus (8). In adults, elevated PFOA concentrations in a community with extensive drinking water contamination was associated with reduced antibody titer rise following receipt of the influenza vaccination (9). An analysis of eight survey cycles within the National Health and Nutrition Examination Survey data concluded that individual PFAS and mixtures of PFAS measured in serum were associated with higher pathogen burdens for several persistent infections (10). The U.S. Environmental Protection Agency (U.S. EPA) used epidemiological data on the association between PFAS exposures and immunosuppression in children to develop the 2022 interim drinking water health advisories for PFOA and PFOS. These health advisories were superseded by the 2024 PFOS and PFOA toxicity assessments (11).

There is limited evidence linking PFAS exposure and COVID-19 outcomes, both at the individual and aggregate levels (12,13). In a Danish cohort, elevated perfluorobutanoic acid (PFBA) concentrations in plasma were associated with higher risk of severe COVID-19 infection (14). However, this study did not find evidence of associations for other PFAS. Small scale area- and individual-level studies in Sweden, China, and Italy have documented higher COVID-19 infection risks and mortality in relation to higher serum, urinary, or drinking water concentrations of PFAS (15–17). A pilot study in the U.S. found that PFAS mixtures measured in plasma were associated with decreased severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) antibody titers among pregnant individuals who were SARS-CoV-2 immunoglobulin G anti-spike protein positive (18). Another U.S.-based study including frontline workers found that several PFAS were associated with altered antibody responses after infection (19). However, existing studies have generally not found evidence of an association between either individual PFAS or PFAS mixtures measured in serum and the immune response to COVID-19 vaccination (19–21). No prior study has investigated associations between larger geographic patterns of drinking water contamination by PFAS and COVID-19 mortality or at lower levels of drinking water contamination.

The main objective of this study was to evaluate potential relationships between U.S. drinking water PFAS contamination and COVID-19 mortality prior to widespread availability of vaccines (January 2020 – December 2020). We based this analysis on two datasets documenting PFAS occurrence in community water systems (CWS = public water systems serving the same population year-round) at the national (Third Unregulated Contaminant Monitoring Rule: UCMR 3) and subnational (4,372 CWS in 18 states) scales. Our analysis provides insights into potential associations between drinking water PFAS exposures and COVID-19 mortality outcomes.

METHODS

COVID-19 Data

We obtained daily reported probable and confirmed COVID-19 deaths and cases from the New York Times repository, which were compiled from state and local governments and health departments (22). U.S. territories (including Puerto Rico, the U.S. Virgin Islands, and Guam) were excluded due to limited available COVID-19 data that could be disaggregated at the county scale. From these data, cumulative COVID-19 mortality (total deaths) occurring between January 20, 2020 (date of the first reported case in the U.S.) and December 11, 2020 (date of the Emergency Use Authorization for the Pfizer-BioNTech vaccine) (23) was calculated for each county.

PFAS Exposure Data

We conducted parallel analyses using two different datasets on PFAS concentrations in U.S. drinking water from CWS. The first dataset, hereon referred to as the statewide sampling dataset, is a synthesis of PFAS concentrations reported by 9,950 CWS in 24 states from samples that were collected between 2016–2023. Data from sampling campaigns were either made publicly available or were provided by state agencies upon request. These data include a majority of smaller CWS serving less than or equal to 10,000 people (24). Additional details regarding data acquisition and processing for this dataset are described in prior work (24,25). We previously established conservative uniform detection limits of 5 ng/L when using these sampling data (24,25). For this study, we only included samples that were collected between January 2017 and December 2020 to focus on the period prior to the availability of the first COVID-19 vaccine in the U.S. (26) The full monitoring dataset used for this analysis included 17,230 samples from 4,372 CWS in 18 states. We then aggregated PFAS drinking water data to the county level after matching each water system with the county (or counties) that it serves using the U.S. EPA’s Safe Drinking Water Information System (27). For analysis of the statewide sampling data, we generated binary measures indicating whether CWS in each county had concentrations of at least one of five PFAS over 5 ng/L (or parts per trillion). The five PFAS included in this analysis were PFOA, PFOS, perfluorohexanesulfonic acid (PFHxS), perfluorobutanesulfonic acid (PFBS), and perfluorononanoic acid (PFNA). These PFAS were selected because they were measured consistently across all CWS. We also generated a binary measure for whether CWS had PFAS concentrations above the lowest available state-level Maximum Contaminant Levels (MCL) as of October 2022 in the U.S. for the same five PFAS (PFOS: 10 ng/L, PFOA: 8 ng/L, PFNA: 6 ng/L, PFHxS: 18 ng/L, and PFBS: 420 ng/L) (28–30).

The second dataset is from the Third Unregulated Contaminant Monitoring Rule (UCMR 3) reporting. UCMR 3 was a national survey between 2013–2015 conducted by the U.S. EPA of 4,920 public water systems for federally unregulated contaminants, including PFAS. UCMR 3 comprised samples collected from all public water systems serving over 10,000 people and a representative sample of systems serving under 10,000 people (31). Importantly, although systems serving over 10,000 people comprise a minority (approx. 8%) of all active systems in the U.S., they serve a majority of the U.S. population that receives public water (approx. 82%) (32). In contrast, the composition of systems included in the statewide sampling dataset more closely matches all total active systems in the U.S. when broken down by system size (24). To maximize comparability between analyses of the statewide sampling dataset and UCMR 3, we excluded non-community water systems, resulting in a sample of 4,798 CWS in 50 states. For the analysis of UCMR 3, we generated a binary measure indicating whether a CWS in each county had concentrations of at least one of the five PFAS over the reporting limits in UCMR 3, which ranged from 20–90 ng/L (31). Across all CWS, the detection frequency of at least one of the five PFAS in the statewide sampling dataset is greater than the frequency of concentrations above the reporting limits in UCMR 3 (approximately 16% versus 4% of systems, respectively) (24,25,31). Counties in the aggregated UCMR 3 dataset had a median of 8 samples (interquartile range: 4, 18) and counties in the aggregated statewide sampling dataset had a median of 9 samples (interquartile range: 3, 22). Our final datasets comprised 621 counties served by CWS in the statewide sampling compilation and 1,677 counties served by CWS in UCMR 3 (Figure S1).

Covariates

Covariates, including potential confounders and predictors, were chosen a priori based on available literature investigating associations between environmental exposures (including PFAS), sociodemographic factors, and COVID-19 mortality (Figure S2) (15,24,33,34). These included the following county-level estimates obtained from the 2014 5-year American Community Survey: the proportion of Hispanic residents, the proportion of non-Hispanic Black residents, the proportion of non-Hispanic White residents, the proportion aged 65 or older, the proportion of residents with less than a high school education, the proportion of homeowning residents, the log of the median home value, the log of median household income, and population density (35). We calculated the total number of hospital beds in each county using the Homeland Infrastructure Foundation-Level Data (36). We also calculated the number of days since the first COVID-19 case reported in each county.

Statistical Analysis

In both sets of analyses, we estimated the associations between county-level measures of PFAS contamination in drinking water and cumulative COVID-19 mortality (total deaths) occurring over the study period using multi-level quasi-Poisson regressions (the Supplemental File includes additional details on the modeling approach). All models included an offset term for the log of the total population in each county. These models were used to estimate cumulative incidence ratios (CIR). The main models using the statewide sampling data included the following as primary exposure measures: 1) a binary measure indicating detection over 5 ng/L for at least one PFAS; and 2) a binary measure indicating detection over the state-level MCL. The main models using the UCMR 3 data included a binary measure indicating concentrations above the reporting limits as the primary exposure measure. In our primary models, we adjusted for the covariates described previously. All continuous covariates were z-transformed (mean-centered and divided by the sample standard deviation). All models included a random intercept for each state to control for time-invariant heterogeneity between states that may have affected COVID-19 deaths and PFAS contamination.

We conducted several analyses to assess the sensitivity of our results to differences in modeling choices, sampling data, exposure and outcome measures, and confounder adjustment (Table S1). Among other analyses described in Table S1, these included adjustment for total MCL violations occurring over 2000 to 2013 (excluding total coliform violations) (27) and adjustment for the county-level proportions of the population aged 15–44 and aged 45–64 (35). For counties included in the statewide sampling dataset, we also substituted the binary measure representing PFAS contamination with the percentage of systems detecting at least one PFAS above 5 ng/L and the population-weighted percentage of systems detecting at least one PFAS above 5 ng/L (37). We conducted these two analyses only for counties included in the statewide sampling dataset due to sparser geographic coverage of systems included in UCMR 3. We also used propensity score matching, which attempts to limit confounding bias and possible model misspecification (38–40). In these analyses, exposed (concentrations > 5 ng/L or over UCMR 3 reporting limits) and unexposed counties were paired using one-to-one nearest-neighbor propensity score matching with exact matching on census division. Propensity scores were estimated using logistic regressions that included each of the potential confounding and predictor variables from the main models. Covariate balance was assessed in the matched and unmatched data using standardized mean differences (SMD) and variance ratios. Following the matching procedure, CIR comparing exposed to unexposed counties were estimated using quasi-Poisson regressions with offset terms for the log of the total population. In an additional set of regression models, we adjusted for covariates with observed residual imbalance after matching (SMD > 0.1).

RESULTS

Descriptive Statistics

The statewide sampling dataset included CWS in counties with a total population of 133.2 million people. The UCMR 3 dataset included counties with 298.5 million people (Table 1). The statewide sampling dataset included more extensive geographic coverage within the 18 states, while UCMR 3 included counties in all 50 states and the District of Columbia (Figures 1 and S3). PFAS detection over 5 ng/L occurred more frequently in the statewide sampling data (38% of total counties) compared to PFAS detection above the reporting limits (20–90 ng/L) in UCMR 3 (6%). In both drinking water monitoring datasets, counties served by CWS with detectable levels of one PFAS also frequently had detectable levels of at least one of the other four PFAS considered (Table S2). For example, 56% of UCMR 3 counties with concentrations above the reporting limits for at least one PFAS detected at least one other PFAS above the reporting limits.

Table 1:

Summary statistics (mean ± SD) of covariates for counties covered by the statewide sampling dataset (SSD) (n=621) and Third Unregulated Contaminant Monitoring Rule (UCMR 3) (n=1,677)

	Statewide sampling dataset (SSD)			Third Unregulated Contaminant Monitoring Rule (UCMR 3)
	All counties	No detections above 5 ng/L	Detection above 5 ng/L	All counties	No detections above reporting limits	Detection above reporting limits
	(n = 621 counties)	(n = 388)	(n = 233)	(n = 1677)	(n = 1573)	(n = 104)
Total population [millions]	133.2	48.1	85.1	298.5	232.2	66.2
COVID-19 death rate [per 10,000]	8.1 ± 5.7	7.7 ± 5.8	8.7 ± 5.5	8.9 ± 6.3	8.8 ± 6.3	9.6 ± 5.9
% non-Hispanic Black residents	8.2 ± 13.2	7.7 ± 14.0	9.3 ± 11.6	9.8 ± 13.7	9.8 ± 13.9	9.6 ± 8.9
% Hispanic/Latino residents	8.2 ± 11.3	6.9 ± 10.7	10.2 ± 12.0	10.1 ± 13.7	9.8 ± 13.6	14.4 ± 13.5
% non-Hispanic White residents	78.9 ± 18.2	81.5 ± 17.9	74.7 ± 17.9	81.9 ± 15.3	75.2 ± 19.1	68.1 ± 17.8
% Residents aged 65+	18.0 ± 4.1	18.4 ± 4.3	17.3 ± 3.6	17.0 ± 4.0	17.0 ± 4.0	16.6 ± 4.1
% Homeowners	71.3 ± 7.9	72.1 ± 7.9	69.8 ± 7.8	69.5 ± 7.9	69.6 ± 7.8	66.8 ± 8.5
Median income [in USD]	54,800 ± 15,600	51,500 ± 13,200	60,400 ± 17,600	54,300 ± 14,700	53,800 ± 14,400	62,600 ± 17,800
Median household value [USD]	179,000 ± 125,200	157,400 ± 113,600	214,900 ± 135,100	166,100 ± 100,600	161,700 ± 96,700	233,100 ± 130,400
% residents with less than a HS education	8.3 ± 3.6	8.4 ± 3.8	8.0 ± 3.1	8.7 ± 3.9	8.7 ± 3.9	8.1 ± 3.7
Pop. density [people / mile2]	390 ± 1,160	286 ± 1,100	571 ± 1,250	340 ± 1,190	308 ± 1,180	831 ± 1,140
Days since first case	263 ± 17.1	259 ± 19.0	268 ± 11.6	263 ± 14.3	263 ± 14.3	272 ± 12.1
Hospital beds [thousands]	0.7 ± 1.9	0.4 ± 1.4	1.1 ± 2.5	0.6 ± 1.6	0.5 ± 1.3	2.0 ± 3.5

Figure 1:

Counties with concentrations above the Third Unregulated Contaminant Monitoring Rule (UCMR 3) reporting limits or detections above 5 ng/L in statewide sampling data

Associations between PFAS Drinking Water Contamination and COVID-19 Mortality

For the statewide sampling data, adjusted regression results indicated a positive association between county-level COVID-19 mortality and PFAS drinking water contamination. Detection of at least one PFAS over 5 ng/L in the county was associated with 12% higher [95% confidence interval (CI): 4%, 19%] county-level COVID-19 mortality (Table 2). Detection of at least one PFAS over the lowest state-level MCL was associated with 15% higher [95% CI: 7%, 23%] county-level COVID-19 mortality.

Table 2:

Adjusted cumulative incidence ratios (CIR) and 95% confidence intervals (CIs) for the association between PFAS drinking water contamination and U.S. county-level COVID-19 cumulative mortality (1/20/2020–12/11/2020)

Model	Primary independent variable	Sample size	CIR
			[95% CI]
Statewide sampling dataa	Detection above 5 ng/L	621	1.12 [1.04, 1.19]
	Detection above lowest state-level MCLb	621	1.15 [1.07, 1.23]
UCMR 3c	Detection above UCMR 3 reporting limits	1677	1.13 [1.08, 1.19]

Results are from quasi-Poisson regressions that included an offset for total population and random intercepts for state, county-level sociodemographic factors (% non-Hispanic Black residents, % Hispanic residents, % non-Hispanic White residents, % residents over age 65, % residents with less than a high school education, % homeowners, log of median household income, log of median owner-occupied household value), total hospital beds, the number of days since the first case in the county, and population density.

^aThe statewide sampling dataset comprises sampling campaigns from 18 states with PFAS samples collected between 2017–2020.

^bThis refers to concentrations above the lowest state-level Maximum Contaminant Levels (MCL) as of October 2022.

^cThe UCMR 3 was a nationwide survey of unregulated contaminants (including PFAS) conducted by the US EPA in 2013–2015.

Results obtained from adjusted models using data from UCMR 3 were similar in magnitude to results for counties included in the statewide sampling dataset (Table 2). Detection of at least one PFAS above the reporting limit in UCMR 3 was associated with 13% higher [95% CI: 8%, 19%] county-level COVID-19 mortality.

Our results were robust to various sensitivity analyses, including the use of different exposure and outcome variables, modeling choices, and confounder adjustments (Table S3). For example, similar estimates were obtained after adjustment for total MCL violations from 2000 to 2013 (CIRs: 1.12 [95% CI: 1.04, 1.19] and 1.13 [95% CI: 1.08, 1.19] for the statewide and UCMR 3 datasets, respectively) and after adjustment for additional age categories (CIRs: 1.08 [95% CI: 1.01, 1.15] and 1.13 [95% CI: 1.08, 1.19] for the statewide and UCMR 3 datasets, respectively). For the statewide sampling data, a one-standard deviation higher unweighted proportion of systems detecting at least one PFAS over 5 ng/L was associated with higher COVID-19 mortality in adjusted models (CIR: 1.08 [95% CI: 1.05, 1.11]). The same was observed for a one-standard deviation higher proportion of the population-weighted proportion of systems detecting at least one PFAS over 5 ng/L (CIR: 1.04 [95% CI: 1.02, 1.07]). A one-standard deviation in the unweighted and weighted proportions of systems corresponds to approximately 28% and 33%, respectively. In addition, main estimates from the propensity score matched analysis (CIR: 1.16, 95% CI: 1.02, 1.32) and from the analysis that adjusted for covariates with residual imbalance after matching (CIR: 1.23, 95% CI: 1.08, 1.40) were comparable or slightly larger in magnitude compared to estimates from the main analyses with the statewide sampling data (Table S4). Confidence intervals for these estimates were wider but overlapped with the main analysis of the statewide sampling data. Across the measured covariates, matched pairs of counties were similar, on average, to all counties that were included in the main analysis of the statewide sampling dataset (Table S5).

For counties included in the UCMR 3 analysis, main estimates after propensity score matching were also similar but slightly larger in magnitude (CIR: 1.21, 95% CI: 1.00, 1.48) compared to estimates from the main model (Table S4). Similar estimates (CIR: 1.18, 95% CI: 1.01, 1.39) were observed after adjustment for covariates with residual imbalance after matching (Figure S4–S5). Confidence intervals were similarly wider for these estimates compared to the main analysis of the UCMR 3 data. Importantly, a larger proportion of counties in UCMR 3 were excluded from the matched analyses because relatively few counties had PFAS concentrations exceeding the reporting limits, which may have impacted precision as well as the obtained estimates. Matched pairs in the UCMR 3 data were similar to all UCMR 3 counties for several measured covariates, but differed in terms of population density, total hospital beds, median income, and owner-occupied household value (Table S5).

DISCUSSION

Results of this analysis suggest drinking water PFAS contamination in the US was associated with higher COVID-19 mortality at the county level. Positive associations between drinking water contamination and COVID-19 mortality were observed in the statewide sampling dataset, which included extensive sub-national geographic coverage, as well as in the national UCMR 3 dataset, which included water systems from all 50 states. These findings were robust to a wide range of sensitivity analyses, including propensity score matching.

Both animal (8,41) and epidemiological studies (42,43) show that several PFAS can harm immune system health. Mechanisms explaining the role of PFAS for COVID-19 severity are not well understood, although several have been proposed. For example, toxicological studies support that PFAS exposures can affect expression of angiotensin-converting enzyme 2 (ACE2), which is a target of the SARS-CoV-2 virus (44–46). ACE2 is a key activator of vasodilation and anti-inflammatory responses and may also regulate lung damage (47). PFAS can also disrupt cytokine profiles, which can impact ACE2 expression and the severity of respiratory illness (45,48). Associations between PFAS exposures and disruptions of cytokine profiles have been observed in epidemiologic studies (49,50). Further, it is unknown, but hypothesized (14), that previously identified associations between PFAS exposures and other health outcomes, including obesity, diabetes, and hypertension (3,51), may act as mediating factors for the severity of COVID-19 outcomes for affected individuals. Additional comorbidities that influence the progression of COVID-19 illness include other cardiovascular conditions, respiratory diseases, and chronic kidney disease (52). If these comorbidities are unrelated to PFAS exposures, adjustment for them would not be expected to reduce bias.

Our results were robust to adjustments for total MCL violations occurring between 2000 and 2013. However, unmeasured confounding due to the influence of co-contaminants may be a concern, particularly for unregulated contaminants. For example, several other PFAS may co-occur with those considered in this study and may be important environmental risk factors for COVID-19 outcomes. PFBA was not included in this study because it was not consistently measured across all CWS. However, exposure to PFBA has been associated with COVID-19 severity in prior work (14). PFBA is also the one of most commonly detected PFAS in ongoing monitoring as part of the Fifth Unregulated Contaminant Monitoring Rule (UCMR 5) (53), which includes many of the systems studied here. Notably, detection of PFOA, PFOS, PFHxS, and PFBS above UCMR 5 reporting limits are moderately but significantly correlated with detection of PFBA in existing UCMR 5 monitoring results (Table S6). Further evaluation of a mixture of co-occurring drinking water contaminants, including regulated and unregulated contaminants, may require more detailed data on concentrations. This study also focused on available monitoring from counties included in the UCMR 3 and statewide PFAS datasets, which limits the generalizability of these results to counties served by large CWS and areas where PFAS monitoring occurred prior to 2021.

Our study used available area-level mortality data, which restricts study conclusions to area-level associations rather than individual-level exposure-outcome relationships (54). Future research could employ study designs that combine area-level exposure measures with individual-level outcome data (55), which would give a more comprehensive understanding of the relationships between drinking water PFAS exposures and COVID-19 outcomes. Although most of the U.S. population receives public water (56), a study using individual-level data may also be able to better reduce exposure misclassification if it can assess exposure related to consumption of private water, bottled water, or a mixture of sources. We also anticipated that relationships between detection above UCMR 3 reporting limits and COVID-19 mortality would be larger in magnitude than those obtained from lower detection limits, but this was generally not the case. However, sparser coverage of CWS included in UCMR 3, coupled with possible exposure-outcome relationships at lower concentrations, may have resulted in greater exposure misclassification error.

Insights from the Fifth Unregulated Contaminant Monitoring Rule (53), continued efforts by state governments, and monitoring as part of the U.S. EPA’s drinking water regulations for PFAS (57) will further improve our understanding of PFAS occurrence in U.S. drinking water systems. Our results suggest that expanded monitoring will be informative for exposure- and epidemiologic-focused research related to infectious disease epidemics, which in turn is essential for protecting public health.

Impact statement:

Our findings provide evidence for an association between area-level drinking water PFAS contamination and higher COVID-19 mortality in the United States. These findings reinforce the importance of ongoing state and federal monitoring efforts supporting the U.S. Environmental Protection Agency’s 2024 drinking water regulations for PFAS. More broadly, this example suggests that drinking water quality could play a role in infectious disease severity. Future research would benefit from study designs that combine area-level exposure measures with individual-level outcome data.

Data availability

All code, replication datasets, and a data dictionary for this study are available via the following site: https://github.com/SunderlandLab/covid-pfas