Linking Industrial Point Sources to PFAS Contamination in Wells: Michigan Case Study

In recent years, concern has grown over widespread environmental contamination of per- and polyuoroalkyl substances (PFAS) in the environment and especially in drinking water. These manmade chemicals were rst synthesized in the mid-20 th century and contain thousands of individual species, some of which are still used today. A wide variety of industries use or have historically used PFAS in their products and processes. PFAS chemicals are recalcitrant by design and thus do not break down in the environment. Federal and state agencies have begun testing of drinking water and other media for PFAS in order to understand the extent of contamination and determine any necessary regulatory actions. Testing is expensive, however, and it would be benecial to be able to prioritize wells for testing based on their likely vulnerability. This study presents a methodology for identifying locations of industrial sites likely to use PFAS, and uses Michigan’s robust well testing data alongside industry and environmental factors to identify any spatial relationships that can be used to help prioritize testing of wells.


Introduction
Per-and poly uorinated alkyl substances, or PFAS, are contaminants of emerging concern due to their widespread occurrence and recalcitrance in the environment, their tendency to bioaccumulate, and the wide variety of health effects associated with exposure at low concentrations (Lindstrom et al., 2011b).
PFAS are known to enter the environment from both industrial and non-industrial waste dumping and e uent, airborne deposition via fumes, re ghting foam use, and disposal of PFAS-containing products in land lls where leaching can occur (Liu et al., 2017;Wang et al., 2014). They have been found extensively in both surface and groundwater drinking water sources water at concentrations ranging from non-detectible to hundreds of parts per trillion (ppt) (Boone et al., 2019). Substantial di culty exists in predicting where high levels of contamination will occur and the causal sources. Identifying PFAS contamination in groundwater is complicated by the wide variety PFAS compounds in use, the lack of transparency and the variability in their use across multiple products, manufacturing processes, and industries. Although they are used in a diverse spectrum of consumer products and industrial processes, PFAS compounds are often only used in small amounts. Finally, lacking government regulation of PFAS chemicals results in limited reporting of use of these persistent chemicals and thus also results in di culty funding and targeting monitoring PFAS contamination in the environment (KEMI, 2015).
Studies have shown that exposure to PFAS is associated with a variety of adverse health outcomes in humans, including hypertension, low birth weight, liver problems, thyroid problems, decreased vaccine response in children, and higher incidence of several kinds of cancer (Agency for Toxic Substances and Disease Registry (ATSDR), 2018; Grandjean et al., 2012;Granum et al., 2013;Kim et al., 2018;Nian et al., 2019; X. W. Zeng et al., 2019;. Although concerns about PFAS chemicals have existed for over a decade, they have only recently become subject to regulation and these regulations are limited in the scope of the types of PFAS they address. In 2009, PFOS was listed as a persistent organic pollutant under the Stockholm Convention, and PFOA was added to the list in 2019 (Stockholm Convention, 2019. In 2019, USEPA issued interim recommendations for groundwater contaminated with PFOA and PFOS and is in early stage of risk assessment for both compounds in biosolids (EPA, Dec 2019).
Lack of information on industrial PFAS use results in the inability of groundwater managers and regulators to identify potential contamination sites e ciently or at a landscape scale for testing and monitoring. A few prior studies have looked at contamination sources of PFAS on various scales. Using data from USEPA's Third Unregulated Contaminant Monitoring Rule (UCMR 3), Hu et al (2016) looked for watershed-level correlation between PFAS contamination and the presence of wastewater treatment plants, military re training areas, and uoropolymer facilities that participated in the USEPA's stewardship program to phase out PFAS production (Hu et al., 2016). Associations were found between all types of sites and elevated PFAS concentrations in water. Guelfo et al (2018) created risk ranking for groundwater in the Providence, RI area based on industrial presence and aquifer vulnerability, listing Department of Defense sites, uoropolymer manufacturers, land lls, and airports among the types of industrial sites contributing the most PFAS risk to water (Guelfo et al., 2018).
In this study we examined whether existing data about the location of industries known to use PFAS compounds can be used to predict future sites for monitoring. This differs from previous studies by incorporating all industry types found in the literature to use PFAS rather than just PFAS manufacturers, and incorporating extensive well testing data to spatially compare PFAS detections to possible point sources. Michigan was chosen as a study case as it began extensive statewide testing of drinking water wells for PFAS in 2018, and continued in 2019 with quarterly testing of wells that had tested at concentrations above 10 ppt the previous year (State of Michigan, 2018). Few states have completed such widespread testing of drinking water sources. Nationwide testing under UCMR 3 has higher detection limits and covers around 5000 drinking water sources nationwide as opposed to over 1000 in Michigan from state-funded testing alone (USEPA, 2012). The goal of this study is to identify characteristics of wells, including proximity to industrial sites, type of bedrock, and well depth, that signi cantly impact the likelihood of PFAS contamination in order to help focus testing efforts on wells that share these characteristics in other areas.

Michigan well and industry data
Well testing data in Michigan was gathered from the Department of Environment, Great Lakes, and Energy's (EGLE) report on PFAS testing using a Freedom of Information Act (FOIA) request submitted in February of 2020. This data identi ed wells based on their Water Supply Serial Number (WSSN) and did not include locations. This well test data was merged with the spatially explicit well data from Michigan's GIS portal that also identi ed wells using WSSNs and could thus be merged with the well testing data (Michigan Dept. of Environment, Great Lakes, 2019). The PFAS testing data from Michigan included 1126 sites, of which 1087 were included in this analysis after removal of surface water sources.
The majority of wells were sampled between April and December of 2018, with one well sampled in September 2017 and seven wells sampled from January through September of 2019. Well water was tested using EPA's Method 537 Revised 1.1, which tests for 18 PFAS compounds. Concentrations were reported as a sum totals of PFOA and PFOS, and of all PFAS found; this study used the concentrations that included all 18 compounds. Among the well sites included in this study, 107 tested positive for PFAS with a range of total PFAS concentrations from 2 ppt to 1828 ppt and an average concentration of 28.794 ppt, whereas 974 sites did not contain PFAS at detectable levels. Notably, the site with a concentration of 1828 ppt was an outlier, as the next highest concentration was 72 ppt.
Industries that use or are sources of PFAS were identi ed in a literature search. Prior studies on point sources as well as studies analyzing products that contain PFAS led to the inclusion of the industry groups listed in Table 1 with their relevant NAICS codes and studies that indicated their PFAS use. Brie y, PFAS have a wide variety of industrial uses including nonstick coatings, metal nishing, fabric nishing, plastics manufacturing, lubricants, and pesticides (Buck et al., 2011;Glüge et al., 2020;Guelfo et al. 2018;Herzke et al. 2012;Kissa, 2001;Nascimento et al., 2018). Studies have also indicated PFAS use in medical devices, electronics, cleaning products, and photographic lm processes (Clara et al., 2008;Gosetti et al., 2018;Lin et al., 2009;Xie et al., 2013). Studies indicating PFAS contamination at military sites and land lls are also well documented (Guelfo et al., 2018;Hu et al., 2016;Liu et al., 2017;Oliaei et al., 2013). Industrial sites for Michigan were spatially located using the NAICS codes listed in Tables 1 and 2, and the USEPA's Enforcement and Compliance History Online (ECHO) facility search (USEPA, 2018). Sites include both current and historic facilities as of 8 June, 2020. Industrial and well locations were mapped using a geographic information system (GIS), ESRI ArcMap version 10.7 (Redlands, CA, USA).

Linking industrial sites to PFAS contamination
Previous studies have shown that a 0.5 km radius buffer around a groundwater source and potential contaminants is a su ciently sized area to study signi cant correlation between land use and contamination. Johnson and Belitz (2009) found that a 500m circle, though a highly simpli ed model, was su cient to show correlation between volatile organic compound (VOC) contamination of groundwater and urban land use (Johnson and Belitz, 2009). Lindstrom et al. (2011a) found that 82% of wells within 0.5 km of elds that received contaminated biosolids tested positive for PFAS (Lindstron et al., 2011a). Finally, impervious surfaces are known to cause rapid runoff that carries heavy loads pollutants into waterways, which can ultimately impact groundwater sources (Frazer, 2005). Use of PFAS at airports and other places where AFFF foams may be used in paved areas could contribute to pollution from such runoff (Awad et al, 2011). One study has linked land ll, military, and PFAS-producing facilities to drinking water contamination on a hydrologic unit basis using UCMR3 testing data, but did not include industries that use PFAS or link speci c sites to contamination (Hu et al. 2016). Based on these studies, we hypothesized that location within the 0.1 km, 0.5 km, and 1 km buffer zones of PFAS-related industrial sites, Coldwater Shale bedrock type, well-drained soils, and impervious land cover would show a signi cant positive correlation with PFAS detection in wells, whereas well depth would show a signi cant negative correlation with PFAS detections. Coldwater Shale bedrock type was chosen because it was the most common bedrock type at contaminated well sites (35%). Based on previous literature and the prevalence of industries within the buffer zones (as indicated in Table 2), we also hypothesized speci cally that the number and percentage of current and historic electroplating facilities, airports, and land lls within the buffer zones would be positively correlated with PFAS detections in wells (Awad et al., 2011;Hepburn et al., 2019;Milley et al., 2018). To test for effects of industry presence on PFAS well contamination in Michigan, we used a maximum likelihood method, Probit analysis and Moran's I as described below.

Probit Models
Using a Probit function, the spatial siting of industries and the bedrock characteristics were used to predict the probability of a positive well test for PFAS. To test for a relationship between industry presence on PFAS detections in wells, buffer zones of 0.1 km, 0.5 km, and 1 km were constructed around each well tested for PFAS. In this study, industry presence within these buffer zones was analyzed as an effect on detection and non-detection in wells while controlling for other factors such as bedrock type, land cover, and well depth. Wells were given a value of 0 if PFAS were not detected (ND), and given a value of 1 if any level of PFAS were detected. A value of 1 was given for wells located within the Coldwater Shale bedrock formation, and a value of 0 given for wells within all other bedrock formation types. Coldwater Shale was selected as it is the most common bedrock formation for wells with detected levels of PFAS. Well depth was based on the numerical value of the well's depth, in feet, and well depth squared is the squared value of the depth of each well in feet squared. A value of 1 was given for wells with a "developed, low intensity" land cover classi cation (NLCD code 22), and a value of 0 given for wells with any other land cover classi cation. A value of 1 was given for wells with "well drained", "excessively drained", or "somewhat excessively drained" soil drainage classi cations. A value of 0 was given for wells with any other soil drainage classi cation. A few speci c industries were chosen as separate variables because they were present at the highest numbers within the well buffers (see Table  2): land ll, electroplating, metal foil manufacture, and airport sites. All models were estimated as either percent, meaning the total percentage of sites within the given radius (0.5 km or 1 km) of a well that belong to the speci c industry, or raw number, which is the number of industry sites classi ed as the speci c industry within the given radius of a well. The percentage of "all others", which includes all the industry sites included in the analysis apart from the speci ed industries just outlined, was also estimated in each model. Lastly, a variable was assigned for the total number of quarters within the past three years that industry sites in a given radius had failed to comply with EPA environmental standards, with a maximum value of 12. A summary of the variables used in the Probit models is shown in Table 1.
Four Probit models were considered to test the effect of land use and type on the probability of detection, such as the number and types of industries within a certain buffer of the well, bedrock, land cover, soil type, well depth, and well depth squared on the probability of PFAS detection in a well using the equation where the dependent variable is assigned values of 0 or 1. The four models separate the amount/percent and type of industries found within each buffer from the data found within other buffers. In this study, the detection of PFAS pollution at well sites is the dependent variable where a value of 1 indicates any detected level of PFAS pollution, otherwise the variable takes a value of 0. The limit of detection was 2 or 4 ppt depending on the analyte (AECOM 2018). Several independent variables were included in the Probit regression model: number of industrial sites (#sites) within each spatial buffer of the wells (0.1 km, 0.5 km, 1 km), well depth in feet (depth), the percentage of sites within 1 km of speci c industry types thought to be more likely to contribute to contamination (airports, land lls, and electroplaters) (percentindustry), and the total number of quarters in which an industrial site had USEPA violations in the past three years [1] (#violations

Cluster analysis
A negative Local Moran's I Index was performed to identify any spatial correlation between detections of PFAS in wells in relation to each other, separate from industry presence. This analysis indicates spatial outliers, where the data at one location is independent of the data at nearby locations (Anselin, 1995). In this study, spatial outliers indicate that the level of PFAS pollution at one well is independent of the level of PFAS pollution at nearby wells. A positive Local Moran's I Index value indicates spatial clustering, where the data at one location is similar to the data at nearby locations. In this study, spatial clusters indicate that the level of PFAS pollution at one well signi cantly resembles the levels at nearby wells. Results And Discussion

Industrial and well sites
Using the NAICS codes listed in Table 3, 2696 relevant industry sites were identi ed in Michigan. Figure 1 shows the number of industrial sites present in each county. Among the well sites included in this study, 107 tested positive for PFAS with an average concentration of 28.794 ppt, whereas 974 sites did not contain PFAS at detectable levels. No identi ed industrial sites fell within 0.1 km of a well that tested positive for PFAS, and only two industrial sites fell within 0.1 km of any well identi ed in this study. A total of eight industrial sites were located within 0.5 km of wells that tested positive for PFAS, three of which were land lls. Other industries within half a kilometer of PFAS wells were metal coating, metal foil, plastics manufacturing, paper mill, and oil and grease manufacturing facilities. Table 3 summarizes the industry sites found within each radius. Twenty-six industrial sites fall within 1 km of wells that tested positive for PFAS; of these sites, eight were land lls, six were metal foil manufacturing, three were airport sites, two were paper mills, two were plastics manufacturing, one electroplating, one metal coating, one commercial printing, one oil and grease manufacturing, and one miscellaneous textiles facility. The totals for 1 km are inclusive of the sites found within 0.5 km.

Linking industrial sites to PFAS contamination
We hypothesized that the presence of the selected industries would have signi cant positive correlations with detections of PFAS in wells, as would the depth of the well, and that greater signi cance would be attributed to facilities belonging to industries already shown to contribute to PFAS contamination (i.e., airports, land lls, electroplaters) and to major facilities. In addition to likelihood of contamination decreasing with well depth, we also hypothesized that as the depth of the well increased, the likelihood of contamination would further decrease, indicated by well depth squared. Because of the lack of industrial sites within 0.1 km of wells, no Probit model was estimated for that buffer, but the model was estimated using both the number of each industry in the buffers as well as the percentage each industry in the buffers for 0.5 km and 1.0 km. It should be noted that in the case of the 0.5 km buffers, electroplating was replaced with metal foil due to limited values for metal foil in the 0.5 km buffers. Omnibus Tests determined that our four Probit models were all signi cant at greater than 99% con dence. The model results are shown in Tables 4 and 5. 2.2.1 Probit models, 1 km The 1.0 km buffer models both showed that well depth, well depth squared, and well-drained soil were signi cantly likely to increase the probability of PFAS detection at greater than 99% con dence level (Table 4). In both cases, well depth positively affected the likelihood of detection, while well depth squared had a negative effect. This indicates that the likelihood of PFAS detection increases with deeper wells, but that the rate of increase declines as depth increases. The positive effect of greater well depth on the probability of PFAS detection runs counter to our hypothesized relationship. One explanation for this could be that wells in vulnerable areas are dug deeper as a preventative measure, but further study is necessary to con rm this. There was a positive relationship between percentage of electroplaters, along with percentage of all other industry and the likelihood of PFAS detection in the 1.0 km model, which corroborates our hypothesis.
The relatively high number of metal foil sites (compared to the other industries included in this study) within 1 km of positive-detect wells is surprising. Although this industry appeared on a list of PFAS-using industries created by the Minnesota Pollution Control Agency (MPCA), it has not yet become an industry of focus across literature examining PFAS pollution. Additional research on metal foil practices could help con rm if this industry contributes signi cantly to PFAS contamination in wells near facilities.

Probit models, 0.5 km
For both Probit models estimated at the buffer size of 0.5 km, well depth, well depth squared, and welldrained soil showed signi cant effects on likelihood to detect positive PFAS levels (Table 5). Similar the 1.0 km models, well depth and well-drained soil had a positive effect while well depth squared had a negative effect on the probability of detection. It appears based on these analyses that more permeable soil near a well is more important than presence of speci c industry types when predicting PFAS contamination in wells. Despite the higher numerical value of metal foil sites within the 0.5 km buffers compared to other sites, the number and percent metal foil sites did not prove signi cant in this model. This is likely due to the number of metal foil sites still being low, despite being higher than other industry types. Percent and number of land ll sites were also not signi cant at this level, along with bedrock type, land cover, and quarters with violations.

Cluster analysis
A local Moran's I test examining clusters and outliers of high versus low concentrations of PFAS within wells that tested positive showed eight low-high outliers, three high-high clusters, and one high-low outlier all concentrated in southwest Michigan. A map of clusters and outliers can be seen in Figure 2. It is interesting to note that the high-high clusters and both types of outliers are predominantly located in Kalamazoo and Allegan counties, both of which have relatively high numbers of industry sites compared to other counties. About half of the industrial sites in these counties were land lls, which corresponds to the percentage of land lls compared to total industrial sites in the state as a whole. In these two counties, however, only one of the land lls (1.6% of total land lls) was agged as actively operating, whereas more than 4% of the state's total land lls were labeled active. A higher percentage of inactive land lls may indicate older land lls. Because land lls were not required to be lined until the Resource Conservation and Recovery Act of 1976, older land lls are more likely to cause leaching of contaminants into groundwater (Congress, 1976). Thus, a higher proportion of inactive land lls in Kalamazoo and Allegan counties may explain the clustering found in these counties from the Moran's I test. The presence of outliers, or wells with high concentrations having a signi cant number of wells with low concentrations nearby and vice versa, could be explained by land ll leachate permeating wells to different degrees. Further study of wells in this region, the groundwater ow near these wells, and their proximity to land lls could help explain the outliers found in this study. In addition to the clusters and outliers in southwest Michigan, ten low-low clusters were identi ed throughout the state. All other clusters and outliers determined using Local Moran's I have p-values exceeding the 0.05 threshold, and are not statistically signi cant to include in the study.

Conclusions
Overall, this study underscores the di culty of tying industrial presence to PFAS contamination in drinking water wells as a means of targeting groundwater testing at a landscape scale. Spatially, the Moran's I results pointed to land lls as a possible smoking gun for PFAS detection, due to a high portion of clusters coinciding with counties that host a high number of inactive land lls, but the more detail, sitespeci c probit model did not show signi cance between land ll presence and positive PFAS detections in wells. Future research could investigate the relationship between age of land lls and detections of PFAS, as this data set did not include land ll age. Although multiple variables were investigated in our predictive probit models to predict PFAS well contamination, the only factors that proved signi cant in all models were well depth, well depth squared, and soil type. The positive correlation found between well depth and PFAS detection contradicted our hypothesis, and a possible explanation for this is wells being dug deeper in vulnerable areas. The recalcitrance of PFAS prevents them from breaking down, and this allows them to eventually reach greater depths. It was found in one of the 1.0 km models that the higher the percentage of electroplating facilities within 1 km of the well compared to total industries studied, the more likely a well was contaminated with PFAS. Based on these results, it can be stated that wells constructed in well-drained soil and with electroplating facilities within a kilometer are likelier to be contaminated with PFAS. This information can be useful for identifying wells to prioritize for testing, but additional predictive parameters would be helpful for more accurate targeting. Future research in this area should consider additional parameters that may lead to trends in PFAS well contamination, such as slope, elevation, and groundwater ow direction surrounding wells.

Declarations
Ethical Approval and Consent to Participate Not applicable.

Consent for Publication
Not Applicable

Availability of Data and Materials
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Competing Interests
The authors have no relevant nancial or non-nancial interests to disclose.