https://orcid.org/0000-0002-9402-8308
Figures
Abstract
One quarter of the population of Pennsylvania relies on private domestic well water: two-fold greater than the US average. Private well owners are responsible for the maintenance and treatment of their water supply. Targeted interventions are needed to support these well owners to ensure they have access to safe drinking water, free of contaminants. To develop appropriate interventions, an understanding of the characteristics and social vulnerability of communities with high well water use is needed. The purpose of this study was to determine the spatial patterning of social vulnerability in Pennsylvania and assess the association between social vulnerability and private domestic wells using profile regression. Census data and water supply information were used to estimate the proportion of the population using domestic wells. Ten area-level measures of social vulnerability at the census-tract level were investigated, using Bayesian profile regression to link clustering of social vulnerability profiles with prevalence of private domestic wells. Profile regression results indicated 15 distinct profiles of social vulnerability that differ significantly according to the area-level prevalence of domestic well use frequency. Out of these, two profiles of census tracts were identified as socially vulnerable and had a high proportion of well-water users, representing approximately 1.1 million Pennsylvanians or a third of all well water users in the State. High area-level social vulnerability profiles coincide with a high frequency of private well-water use in PA. This study presents a data-driven approach to supporting public health programs aimed at reducing exposure and health risks of chemical and infectious agents in household water supplies by targeting vulnerable populations.
Citation: Wamsley M, Coker ES, Wilson RT, Henry K, Murphy HM (2024) Social vulnerability and exposure to private well water. PLOS Water 3(12): e0000303. https://doi.org/10.1371/journal.pwat.0000303
Editor: Francis Dakyaga, SD Dombo University of Business and Integrated Developmental Studies, GHANA
Received: June 28, 2024; Accepted: October 31, 2024; Published: December 30, 2024
Copyright: © 2024 Wamsley et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: Data sources are all publicly available: Social vulnerability data: 10 measures of social vulnerability that may be related to a household’s exposure to a domestic well or to a community’s vulnerability to waterborne disease were taken from the CDC’s SVI measure as measured in 2018 at the census tract level. Data may be accessed here: https://www.atsdr.cdc.gov/placeandhealth/svi/data_documentation_download.html Measures of children under the age of 5, median construction age of the home, and proportion of households who are renters were taken from the ACS (5-year estimates from 2018 at the census tract level, published 2019). Data may be accessed here: https://www.socialexplorer.com/tables/ACS2018_5yr/R12897432 Census tract boundaries Tiger line shapefiles for the year 2010 may be accessed here: https://www2.census.gov/geo/pvs/tiger2010st/42_Pennsylvania/42001/tl_2010_42001_tract00.zip Public water supply boundaries These data are supplied by the Pennsylvania Department of Environmental Protection. Public water supplier’s service area data from the year 2018 were accessed and used via the Pennsylvania Spatial Data Access and may be accessed here: https://newdata-padep-1.opendata.arcgis.com/datasets/PADEP-1::public-water-systems-public-water-supplier-service-areas/about.
Funding: Funding for this research was provided in part by the Pennsylvania Department of Health’s Pennsylvania Commonwealth Universal Research Enhancement Program (PA CURE) Formula Funds 2019 (Grant # 4100083099 to Temple University). The preparation of the manuscript and analysis was also partially funded by Dr. Murphy’s Tier II Canada Research Chair in One Health funded by the Canada Research Chairs Program (Grant # 950-232787). Miriam Wamsley was partially supported through Temple University’s dissertation completion grant award.
Competing interests: The authors have declared that no competing interests exist.
Introduction
In Pennsylvania (PA), an estimated 3 million people, or 27% of the population, lack access to publicly supplied drinking water and rely on private domestic well water for consumption [1]. Water quality throughout the United States (US) is generally thought to be of high quality, as greater than 90 percent of the US population, reportedly, has access to a safely managed drinking-water supply, aligning with goal 6.1.1 of the United Nations’ Sustainable Development Goals (SDGs) [2, 3].
Most of the US population is served by a community water supply which is defined by the United State Environmental Protection Agency (US EPA) as a water source that serves greater than 25 people or 15 connections [4]. However, an estimated 42 million are served by a domestic well water supply, whose quality is not regulated by the federal government. In Pennsylvania, as in most of the US, the quality and any necessary treatment or maintenance of a domestic well is the responsibility of the well owner [1, 5].
Consequently, those served by domestic wells are exposed to potentially unsafe water with unknown quality [2, 6]. Socially vulnerable communities may be less likely to benefit from community drinking water and wastewater services and therefore may rely on a domestic well [7]. Since upkeep of private water supplies falls completely on the well-owner, community-level and individual-level social deprivation/vulnerability is likely to be an important determinant of whether well-owners can keep up with the requirements to ensure safe drinking water. Importantly, the availability of support and programs for domestic well owners vary by State (and even within a State) meaning many of these homeowners are left to independently navigate the management and treatment of their water supply.
Homeowners are not always aware that their water may be contaminated and contaminants can co-exist with other chemicals which may or may not impart a change in color, taste, or odour [8]. While many well water users will not be concerned unless their water looks or smells “dirty” [9], odour-free and taste free water may not be safe [10]. Pathogens and other contaminants can co-occur with turbidity [11], and a change in turbidity can signal a change in the overall water quality, but pathogen-impacted water can also be clear [12]. Climate change is increasingly impacting the quality and quantity of water, particularly in times of natural disasters including during extreme rain or extreme drought [13–16].
Although the use of domestic well water is common across the United States and in PA (13% and 27% of populations respectively) [1, 17], the socioeconomic and demographic characteristics of the population using domestic wells (that are excluded from public water supplies) are not well described. In an assessment of race and access to public water supply in North Carolina (NC), investigators found that for every 10% increase in the African American population, the odds of exclusion from municipal water services in a census block increased by 3.8% [18]. A later assessment of race, community water, and sewer in NC found that lower income black populations, and higher-income non-black populations were the most unserved. They also reported greater disparity for community drinking water than for public wastewater after looking at rates for the whole state [19]. These findings exemplify a complex relationship of inequalities in regulated drinking water access, whereby the profiles of NC communities with limited access vary by race and income.
In Pennsylvania, little research has been done to examine the relationships between domestic well ownership as it relates to social vulnerability. Clough and Bell (2016) assessed whether people that reside closer to the extraction of oil and gas from the Marcellus shale formation are more likely to be in poverty and/or belong to a minority racial/ethnic group. The authors found no evidence of traditionally defined environmental injustice, however, they did find that “…(the population) suffers the hazards associated with proximity to well sites but are not enjoying substantial economic benefits from shale gas development” [20]. Their study, as well as others [21], suggest complex and spatially heterogeneous relationships between conventional measures of social vulnerably and environmental health hazards. Therefore, any type of index such as the Centers for Disease Control’s Social Vulnerability Index (SVI), that is a composite score of multiple population indicators including race and income, may obscure complex and non-linear relationships between social vulnerability and regulated drinking water access. More flexible statistical methods, which harness such complexity, are needed to assess environmental health disparities in the context of social vulnerability.
For example, using cluster analysis, Molitor et al. 2011 estimated the spatial association between exposure to multiple air pollutants (pollutant profiles) and poverty in Los Angeles (LA), California. Their study showed that LA census tracts with higher exposure to an individual air pollutant were associated with higher exposure to a second different air pollutant (e.g., particulate matter and nitrogen oxides). Similarly, LA census tracts with higher exposures for each air pollutant individually were associated with higher poverty levels. However, the relationship between pollutants and poverty was non-linear when considering multiple exposures in combination. Their analysis suggests that identifying vulnerable populations for combinations of environmental and social factors requires more flexible modelling methods, such as cluster analysis or Bayesian multilevel modeling [21]. Rufat et al. (2019) compared various vulnerability assessment methods in relation to Hurricane Sandy’s adverse outcomes in New York City. Social vulnerability profiling (SVP) using cluster analysis emerged as the most predictive method, consistently aligning with expected relationships with adverse hurricane-related outcomes. In contrast, the SVI approach showed more variable and sometimes counterintuitive associations with hurricane-related outcomes. The flexibility of SVP to characterize distinct vulnerability characteristics of communities across different contexts was highlighted, as opposed to relying on an SVI, which occasionally obscured environmental health hazard and vulnerability relationships [22].
The aim of this study was to determine whether populations in Pennsylvania that rely on domestic wells are more socially vulnerable than those with access to public water supply. We used Bayesian regression profiling to identify distinct social vulnerability profiles of populations served by domestic wells. The information generated can be used to inform public health interventions such as where to allocate funding for a community’s access to public water, private well education campaigns, campaigns during emergency response efforts and education efforts targeted at health professionals in these regions.
Methods
Social vulnerability
The US Centers for Disease Control (CDC) developed a nationwide social vulnerability index (SVI) as one tool to help identify “socially vulnerable” populations at heightened risk to hazardous events such as disease outbreaks. This SVI tool is a summary score which maps socioeconomic status, household composition, race/ethnicity status, housing type, and transportation by comparing these variables to ranked variables within each state by census tract [23, 24]. The SVI variables are described in Fig 1.
In this paper, the term vulnerability is used in the context of “social vulnerability” as defined by the US CDC. According to the CDC’s Agency for Toxic Substances and Disease Registry (CDC/ATSDR), “Social vulnerability refers to the potential negative effects on communities caused by external stresses on human health” [25]. We also define “vulnerable populations” as “groups…at a higher risk for poor health as a result of the barriers they experience to social, economic, political and environmental resources, as well as limitations due to illness or disability” [26].
Drinking water source data
Public water system (PWS) boundaries for PA from 2018 were provided as a shapefile [27] by the Pennsylvania Department of the Environment. For each census tract we estimated the proportion of the population using a domestic well using the following assumptions: 1) census tracks without publicly supplied water rely on domestic wells, 2) the proportional area served by a public water supply is the same as the proportion of homes served by a PWS. To verify whether our assumptions were valid, we compared our maps of proportion of homes served by private well water to population density maps of the state of PA (S1 Fig).
There were 3,193 census tracts from 2010 in PA that had a number of houses greater than zero in the American Community Survey (5-year) from 2018. The median population per census tract was 3772 (range 3 to 12,682) [28]. There were 1,683 census tracts with some proportion of households (> 0–100%) using a domestic well (Fig 2). The area of census tracts using private water sources was determined by removing the area served by public water supply from the TIGER/line shapefiles for both counties and Zip code tabulation areas (ZCTAs) [29] in ArcPro 2.8 (ESRI, 2020) with the erase tool [30]. In the census tracts which are, either 0 or 100%, served by PWS, the estimate of the number of households served is expected to be a good estimate.
Social vulnerability data
For the analyses described herein, we selected 10 parameters of social vulnerability likely related to a household’s exposure to a domestic well or to a community’s vulnerability to waterborne diseases. Measures for our social vulnerability indicators included: proportions of a census tract below the federal poverty level, minority race or ethnicity, unemployed, > 25 years old with no high school diploma (> 25 no HS), > 65 years old, < 17 years old; median per capita income (PCI). The social vulnerability data were obtained from the CDC’s Social Vulnerability Index (SVI), derived from the 2018 American Community Survey 5-year estimates at the census tract level, provided by the US Census Bureau [23, 31]. The CDC’s SVI includes 15 parameters of social vulnerability; 7 of the 15 variables were included in the present study, along with three additional parameters: children under the age of 5, median construction age of home, and proportion of households that rent [32]. We chose a priori to only include the most relevant variables from the SVI instead of using the SVI as a single value because some of the measures (like having a vehicle, living in a mobile home, or living in a single parent household) are not logically related to use of domestic wells or to increased risk of enteric illness.
All values were rescaled into county-specific z-scores before being fit in the Bayesian Profile Regression model. This normalization of indicator data was done because each county has its own baseline level of values for the social vulnerability indicators. Additionally, these indicators vary not only between census tracts but within and between counties. If raw, unscaled values had been used, it would obscure actual inequalities between clusters that are occurring between counties. This similar rescaling was used by Coker et al. (2023) in assessing the spatial relationship between environmental exposure profiles and excess mortality risk due to COVID-19 in Lombardy, Italy [33].
Cluster analysis using bayesian profile regression
Community-level measures of social vulnerability and the prevalence of domestic wells are thought to be colinear because many measures of social vulnerability are higher in rural areas, and rural areas will have more domestic wells [34]. This collinearity can vary spatially. Such collinearity and spatial non-stationarity between social vulnerability and domestic well prevalence may create barriers when performing conventional multivariable regression-based approaches [35–37]. Spatial non-stationarity is when an overall or “global” model is inadequate in describing the relationship between a set of covariates and an outcome. To deal with spatial non-stationarity, the model must be allowed to vary across space [38]. Cluster analysis is an alternative statistical method that groups observations based on similarity between observations; with similarity depending on the nature of correlations between variables. Therefore, cluster analysis may be well suited for dealing with problems of collinearity when evaluating social vulnerability with respect to environmental hazards [39].
Bayesian profile regression (a clustering technique) was used to identify social-vulnerability profiles of co-exposure at the census tract-level. These census tracts were z-score-adjusted to account for within-county differences. This form of profile regression is considered a non-parametric (or data driven) clustering algorithm because the analyst does not need to define the number of clusters prior to analysis [36, 40]. The data-driven nature of profile regression derives from the well-established family of infinite mixture modeling known as the Dirichlet process mixture model (DPMM) [41]. Here, the DPMM implementation is set in a flexible Bayesian framework using stick-breaking priors (described in detail elsewhere[21, 42, 43]). Social vulnerability profiles were comprised of the 10 different census-based indicators of social vulnerability as described above.
To address the skewed distributions of the covariates, the 10 measures of social vulnerability, or ‘x-values’, were first reclassified as quartiles, and added to the model as discrete values. While Bayesian profile regression (BPR) is capable of directly estimating the association between exposure profiles with a response variable, we did not include the response model when determining clusters because domestic well prevalence is clustered by county, and the BPR does not allow for such spatial clustering effects modeling. Instead, the clusters of exposure profiles were fit in a second-stage multilevel regression model to estimate the relationship between social vulnerability profiles with domestic well prevalence (e.g., those without access to public water supply).
Multilevel risk model
In the second stage of analysis, a response model with the outcome variable set as the count of households exposed to domestic well use. The “hard clusters” output from profile regression, or types of profiles [35, 44], were fit as the exposures of interest, as a fixed effect in a multi-level model (described in turn). Due to the high number of census tracts where there were no private wells according to the data (i.e., zeros), the mean did not equal the variance in the distribution of the number of houses with wells per census tract, making the Poisson distribution, an inappropriate fit. Therefore, counts of houses using domestic wells by census tract were then assumed to follow a negative binomial distribution. After determining the types of clusters from the Bayesian profile regression, the count of houses was modelled as follows:
Where i is the unit of analysis level (census tract, n = 3193) and j is the group level (county, n = 67). The clusters were fit as a fixed effect. The log of the number of houses per census tract were included as an offset while the county was added as a random intercept. This was modeled in R using the glmer.nb function from the lmerTest package [45, 46]. The hard clusters were then ordered by their estimated risk ratio from smallest to largest number of homes using domestic wells. The cluster with the lowest risk ratio was the reference group.
Comparison social vulnerability index model
As a comparison analysis, the CDC overall social vulnerability index (SVI) score (values 1 to 100) where i is the unit of analysis level (census tract, n = 3193) and j is the group level (county, n = 67) was used to model the count of houses with wells as follows:
The log of the number of houses per census tract were included as an offset while the county was added as a random intercept. This was modeled in R using the glmmTMB package to compare types of zero inflated and negative binomial models. Many census tracts have zero houses with wells, causing overdispersion issues with other model types [47].
Results
Social vulnerability clusters
The 10 measures of social vulnerability selected in this study exhibited a range of correlation from high to low negative and positive correlations, as shown in the Pearson correlation plot in Fig 3. From the 3,193 census tracts analyzed, the profile regression resulted in 15 distinct clusters of social vulnerability profiles in PA (Fig 5).
Notably, census tracts comprising the clusters are not always spatially contiguous (Fig 4). The more clusters there are, the more heterogeneous a population is. There were 15 clusters of social vulnerability characteristics identified in this analysis across PA. The clusters vary across the state which suggests that while the multiple measures of social vulnerability are correlated, these correlations are spatially variable.
Out of the 15 clusters, 6 were statistically significant clusters (p< 0.05) in that domestic well exposure in these six clusters was significantly different than the reference cluster (cluster 1) (Table 1). These clusters are illustrated in Fig 4. The clusters are ordered for visualization with higher cluster numbers symbolizing higher counts of homes supplied water from domestic wells (i.e., cluster 1 has the lowest median and mean number of houses served water from a domestic well, while cluster 15 has the highest). In Fig 5, the purple cells illustrate the highest proportion or median value of the vulnerability score attribute (by quartile) and yellow cells have the lowest values (by quartile). For example, cluster 15, with the highest median and mean numbers of houses served by domestic wells, has yellow cells for seven covariates, meaning that it had the lowest quartile of measures of social vulnerability for those seven covariates (i.e., this cluster is less socially vulnerable). Table 1 presents the results of the initial Poisson model along with the final model which is an adjusted negative binomial model.
(Purple cells are the most socially vulnerable, while the yellow cells are the least. The higher the cluster number, the more houses supplied by domestic wells).
Social vulnerability clusters as fixed effects with county as random effect
The negative binomial model, with county as a random effect, was the final chosen model. Results for each cluster are listed in Table 1. The final statistical analysis accounts for baseline differences of domestic well prevalence at the county-level. There are 67 counties in Pennsylvania. The county of residence was found to be an important random effect that needed to be considered. Unsurprisingly, the county that had the most “protective” effect, was the urban county of Philadelphia (Supplementary Information, S2 Fig). As previously stated, the cluster numbers are ordered by incidence rate ratio, with cluster number 1 set as the reference group. The range in values of number of homes served by private wells by cluster are shown in Table 1. The random effects due to county are shown in Fig 6. The estimated proportions of populations using domestic wells by census tracts is listed in S1 Table.
According to Table 1, clusters 10, 11, 12, 13, 14, and 15 were significantly different (p < 0.05) from cluster 1, with respect to the number of households using domestic wells. These six clusters had statistically significant higher incidence rate ratios of domestic wells, after adjusting for county (“cluster 10”: 2.70 (1.14, 6.38); “cluster 11”: 2.96 (1.27, 6.90); “cluster 12”: 4.15(1.54, 11.17); “cluster 13”: 3.21(1.40, 7.37); “cluster 14”: 4.12 (1.75, 9.71); “cluster 15”: 5.17(1.65, 16.18)). The social vulnerability profiles of these elevated domestic well clusters are described in turn.
Clusters where private well ownership corresponded with highest social vulnerability
Cluster 12 included 73 census tracts, with a median of 1681 homes using domestic wells per tract. Cluster 12 had higher measures of social vulnerability. This cluster was characterized as being in the fourth quartile for populations under that age of 17 as well as a population over the age of 25 without a high school diploma or equivalent. Cluster 12 was also in the third quartile for both measures of poverty (per capita income and proportion below the federal poverty limit). To summarize, cluster 12 had a young population that was experiencing high levels of poverty and lack of secondary and higher education when compared to other census tracts in the state.
The exposure profile for Cluster 14 also suggested elevated social vulnerability. Cluster 14 included 262 census tracts with a median of 960 households using domestic wells per tract. Cluster 14’s population was younger (being in the third quartile for the proportion of the population being less than 17 and 5 years of age), had a high proportion of population which was unemployed, and high proportion of the population over the age of 25 without a high school diploma or equivalent. This cluster was also in the third quartile for low median per capita income. In other words, cluster 14 contained a large population of children, was experiencing more than the median level of unemployment, lack of education and income when compared to other census tracts in PA.
Clusters where private well ownership did not correspond with social vulnerability
Cluster 10 included 220 census tracts, with a median of 165 homes using domestic wells per tract. Cluster 10 had lower measures of social vulnerability than clusters 12 and 14 except for age of population. When compared to other census tracts in the state, Cluster 10 had a young population with a large proportion of their population under the age of 17.
The exposure profiles for Cluster 11 and 13 also had lower levels of social vulnerability when compared to the other clusters with high exposure to private wells, except for being in the third or fourth quartile for the proportion of their population that is age 65 or over (cluster 11 and cluster 13 respectively). Cluster 11 included 291 census tracts with a median of 321 households using domestic wells per tract. Cluster 13 included 418 census tracts with a median of 924 households using domestic wells per tract.
The exposure profile for Cluster 15 does not suggest elevated social vulnerability. Cluster 15 included 61 census tracts with a median of 0 households using domestic wells per tract. Cluster 15 also had the highest incidence rate ratio of 5.17 (1.66, 16.18). Cluster fifteen was in the first quartile for all measures of social vulnerability with the exceptions of: 1) age over 65, which was in the second quartile, and 2) older median housing age, of which was also in the second quartile.
Spatial distribution of clusters
In Fig 4, the areas covered by public water supply are visualized with dark grey. Clusters 1–9, which include clusters with statistically-significant low risk of private well water source are visualized in white. The clusters with high risk of private well water combined with the highest measures of social vulnerability are visualized using dark purple (cluster 12) and dark green (cluster 14). Clusters 10, 11, 13, and 15 are visualized using light purple, light green, medium green, and medium purple, respectively. Interestingly, the socially vulnerable clusters 12 and 14 that have a high proportion of private well users are distributed throughout the state. However, in several cases, they appear on the borders of urban centers where public water supply ends. This apparent spatial pattern may hold value in identifying opportunities where vulnerable communities could be prioritized for connection to public water supplies (assuming communities want to be connected and there are resources allocated to support connection costs).
Social vulnerability index comparison model results
When applying a more simplistic modelling approach, the model that fit best, when fitting the overall CDC Index values by census tract as a predictor of number of wells per census tract as the outcome, was a zero-inflated negative binomial model. The exponentiated estimate was 0.995 (raw estimate was -0.0046576) and was significant with a p value <0.001 and estimated 95% confidence interval of 0.994–0.997 (raw values were -0.00635 to -0.00296). Put another way, for every 1-point increase in the vulnerability index, there is a 0.4% decrease in the number of houses with wells in a census tract or for a 25-point increase in the SVI, there are 10% less houses with wells in a census tract. Model results can be found in S2–S4 Tables.
Discussion
We used a Bayesian cluster profiling approach to identify clusters of social vulnerability linked with a lack of access to public water supply. Certain cluster profiles were significantly more likely to lack access to public water supply than others. For example, the cluster profile with the highest domestic well prevalence was cluster 15 ([IRR = 5.17, (95%CI 1.66, 16.18, p < 0.01)]; range of estimated number of households using wells (mean = 274, 3rd quartile = 12, max = 3131l 7 tracts with more than 1,000 households estimated to be using private wells). This cluster was characterized by relatively low social vulnerability. In contrast, however, at least two clusters (12 and 14) with elevated risk for a lack of public water supply were characterized as having elevated social vulnerability. Approximately 1.1 million well users reside in clusters 12 and 14, meaning a third of PA well users would be considered socially vulnerable from our analyses.
Bayesian cluster model versus simplistic SVI model
Using the Bayesian Cluster Profiling approach, our results suggest that social vulnerability characteristics tend to be on the extremes—i.e., with some areas having high domestic well use and low social vulnerability and other areas with high well use having high social vulnerability. In contrast, when we use the simple comparison model, we find a simple inverse relationship between number of houses with wells in a census tract with the social vulnerability index. This simplistic approach shows that the lower the amount of vulnerability (low SVI), the higher the number of wells per tract. This is not useful or helpful information as it fails to identify that certain locations served by private wells are more socially vulnerable than others and have different vulnerability profiles. Similar studies have also noted that the index is not always useful or helpful for evaluating environmental health-related inequities [22, 48]. Well ownership is spatially complex. The distinct vulnerability profiles identified using the Bayesian cluster profiling approach show heterogenous measures of vulnerability among domestic well owners as well as being spatially heterogeneous. Our result highlights the complex relationships that exist between public water supply access and social vulnerability; a relationship that is otherwise obscured when relying on simpler analyses looking at SVI alone.
Race/Ethnicity & poverty
A lack of access to safe drinking water in the United States has been associated with race/ethnicity, as well as poverty [49, 50]. In a study of environmental injustice regarding a population’s nearness to unconventional gas development in the state of Pennsylvania, Clough and Bell in 2016 found that race and ethnicity were not associated with proximity to unconventional gas development, but poverty was [20]. In our work, we did not find that domestic well ownership was related to race or ethnicity. This differs from what was found in Wake County, North Carolina when MacDonald Gibson et al. found a positive relationship between a population’s proportion of black persons and the odds of being excluding from a municipal water service [18]. In PA, it is not surprising that areas served by private wells, which are rural, have a lower proportion of black populations, given that during World War I, the “Great Migration” began with southern black populations settling predominately in urban centers of northern states, which in PA included Philadelphia and Pittsburgh and other surrounding industrial cities such as Chester and Norristown [51]. This pattern of settlement of black population in the urban areas of PA differs from that of the settlement pattern in North Carolina, where black populations historically predominantly lived in the rural areas of the state [52].
We found no evidence that minority race and/or ethnicity was related to exclusion from public water supply in this analysis. It should be noted here that this analysis did not consider water insecurity or consider lack of access to public water supply due to inability to pay, either a water bill or the fee to hook up to a water supply. It also did not consider whether the water was clean or safe (e.g., lead contamination) [53, 54]. In the report, Closing the Water Access Gap in the US: A National Action Plan, Roller and Gasteyer et al. report that, “…race is the strongest predictor of water and sanitation access.” The analysis in Roller and Gasteyer was completed at the census tract level, but for the entire United States. Also, their definition of access was whether or not a household reported having complete plumbing using data from the 2010–2014 American Community survey, which asks if a home has, “hot and cold running water, a bathtub or shower, a sink with a faucet, or a flush toilet” [50]. The census data utilized by Gasteyer et al. did not include a variable for domestic well water and consequently our definition of “access” to a public water supply contrasts their definition. The data used by Gasteyer et al. could include populations served by public water supplies but that are experiencing water shutoffs, or do not have a water heater, adequate in-home plumbing etc. which is different than the question we were seeking to answer. It is also acknowledged that our analysis is not perfect as a portion of the population that lives within a public water supply area may still not have public drinking water supplied to their homes [55, 56].
In Pennsylvania, the use of a private well (or lack of connection to a public water supply) may not be as strongly related to race as in other states. For clusters 10 through 14, we found that those residing in areas with a high density of domestic wells were more likely to own than rent their homes. People of color have experienced racial inequities when it comes to home ownership and therefore this may explain why race was not as strongly associated with private well density [57]. We found that census tracts in Pennsylvania that had the highest quartile of minority race/ethnicity also had the highest quartile of renters. These clusters of census tracts that have the largest proportion of racial and ethnic minority home renters were also least likely to have a domestic well, meaning, they were connected to publicly supplied water. However, it is important to note that having a connection to a public water supply, does not necessarily equate to access to a water supply, as those experiencing poverty can have their water shut-off. If this analysis was repeated with data on access to public water, including shut-off data, the cluster analysis may yield different results, particularly with respect to minority populations who can disproportionally lack access to water and sanitation services, even in high income countries like the US [58].
Housing age, age of household members, education
Census tracts with younger median age of homes had larger amounts of domestic wells. All census tracts that were more likely to have domestic wells had younger homes than the median age (fell within first or second quartile for median housing age) This does not mean that the houses are new, because housing throughout the state of Pennsylvania is some of the oldest in the United States [59]. Philadelphia and Allegheny counties, which are urban with no domestic wells, have houses with a median age of 73 and 64 years, while Chester and Monroe counties are some of the youngest in the state with median home ages of 39 and 36, respectively [59].
Out of the clusters, cluster 12 had the 4th quartile of proportion of their population with children under the age of 17 years old. Cluster 14 was in the third quartile for under 17 years old. Cluster 13 was in the fourth quartile for their population being over the age of 65 years old. One of the main concerns regarding exposure to domestic wells is exposure to pathogens. Gastroenteritis is the most common illness caused by pathogens found in domestic well water. Children under the age of 5 are highly susceptible to gastroenteritis, as well as those over the age of 65 [60, 61].
All clusters with high prevalence of domestic wells fell in the second quartile for unemployment, except for cluster 14, which was in the third quartile. Cluster 14 was also in the fourth quartile for those who do not have a high school diploma.
In summary, we identified six different distinct census-tract cluster types that are at increased risk of lacking access to public water, two of which also have high measures of social vulnerability. A census tract in Pennsylvania is most likely to have a large proportion of houses using domestic wells and also be experiencing social vulnerability if they are a population which, when compared to tracts in the rest of the state: own their home, has a large proportion of children under 17, has a median per capita income less than $30k per year, has a younger median age of homes, and has a higher proportion of those aged 25 or older that do not have a high school diploma or equivalent.
Study strengths and limitations
A major strength of this study is the use of publicly available data to understand area-specific context of social deprivation. Area-level deprivation is a well-validated approach that has been consistently associated with major infectious and chronic disease health outcomes [48].
To our knowledge, this is the first study to use Bayesian profile regression to examine private domestic well use related to social vulnerability. This use of a flexible Bayesian clustering profile regression model is an efficient way to assess which social vulnerability attributes are related to private water supply ownership. The method enabled us to characterize areas in the state of Pennsylvania that are most likely to be served by non-PWS. The Bayesian profile regression was an appropriate choice to determine which covariates were clustering together. A weakness of this method is that it is less able to evaluate the independent effects of any single covariate as would be possible in conventional regression analysis [62]. Traditional multiple regression models can struggle with assessing an outcome when the exposures and covariates are related due to problems of collinearity [39]. Bayesian profile regression is appropriate for observing where patterns overlap, or rather, where certain conditions co-occur, while avoiding issues with collinearity [35]. The problems of lack of access to public water, as well as measures of social vulnerability, are interrelated making it difficult to tease out these relationships. The use of the clustering approach reduced the dimensionality of the measures to cope with the issue of collinearity [39]. Moreover, a clustering approach may be preferable compared to a simple index (e.g., SVI) since co-exposure patterns and their interrelationships may be complex and non-linear in a spatial sense [63].
Although the use of an ecologic study design is generally considered to be a weakness, here our research questions are all at the community level, because any potential interventions, would likely be at the community level. Another limitation is that we relied on the public water service boundary information and made assumptions regarding private well usage using these data. Our approach assumes that an area “not in the PWS area” relies on domestic wells, and that the same proportion of area served by domestic wells is the same proportion as that of families in a census tract. Although this assumption is an imperfect proxy measure, there is not a complete registry of domestic wells available in PA. Although, some drilling records for wells are reported to the Pennsylvania Department of Conservation and Natural Resources, which are compiled and known as the PA Groundwater Information System [64], the drilling data are incomplete as they are not required to be reported in the state of PA, and the system has only been collecting these data since the 1970’s. This means that using data only from this system would be an underestimation of the number of families using domestic wells, both in areas with wells completed before the 1970’s as well as wells completed in the 21st century. Another limitation of this work is that of temporal ambiguity. All we can assess is which factors co-occur, we cannot assess if social vulnerability leads to differences in public water infrastructure or if the opposite is true [39].
In the census tracts with only a proportion being served by public water supply, this estimate likely overestimates the number of households served by private wells, because the PWS areas are likely denser in population. This limitation is acknowledged, and we believe our approach is still reasonable for estimating the proportion of households served by private wells by census tract in the state of Pennsylvania in the absence of better-quality data.
Public health implications
Approximately 90% of the US is considered to have access to safely managed drinking water [17, 49], which in the context of the US is deceiving, as this number includes those served by private, unregulated, untested and untreated domestic wells. If we were to say that only those served by public drinking water in US have access to safely managed drinking water, then this number would be drastically decreased. In the case of PA, an estimated 3 million of 12.79 million people are served by a private well, meaning only 76.5% have access to public drinking water [1, 65]. Additionally, we demonstrated that ~1.1 million domestic well users are socially vulnerable.
For the United States to achieve universal access to “safely managed water” to meet the United Nations Sustainable Development Goals, more progress needs to be made in addressing the gaps in access to safe drinking water, including providing support for homeowners served by a domestic well. Targeted interventions are needed to focus on these vulnerable, underserved populations and more research is needed to understand the social vulnerabilities of these populations so appropriate interventions can be developed [58]. These future interventions could include changes to domestic well installation rules, provision of free testing and treatment systems, the design of educational tools and campaigns, or developing a system for prioritizing which communities receive resources for public water infrastructure. The knowledge gained from our analysis in PA informs us that interventions and education campaigns need to be targeted towards homeowners (and some renters), and populations that are more socially vulnerable (have a low level of education, with an annual income less than $30k, are elderly or have children) as well as those that are less socially vulnerable (those with higher levels of education, an annual income over $30k, and are not a sensitive age). Interventions for these different groups would need to be conceived differently. For example, free testing programs could be made available for populations that are more vulnerable than those with more resources to afford water quality testing. Additionally, there are pockets of clusters 12 and 14 that appear on the outskirts of where public water supply ends. In these situations, these vulnerable populations could benefit from being hooked up to the public water supply. Although these populations would benefit from public water supply and wastewater, the reasons why connections are not extended can be very complex. One reason, and a major issue, is cost. Even if the community has the funding to extend water and wastewater services, the individual households which have never paid a monthly bill may be resistant to the change. The well water users may perceive their water as being safer and better than a public waters supply even though they likely do not have any facts by which to base these opinions [66, 67]. Helping well water users understand the actual health risks, and the costs associated with maintaining a well water safety and an onsite wastewater system over the lifetime of a mortgage may help these populations make choices that are in their best interest.
Conclusions
In conclusion, this study demonstrates that identifying high well-use areas alone is insufficient to identify vulnerable populations clearly. Areas with high well use are paradoxically at both low-socially vulnerable and high- socially vulnerable extremes in PA. Therefore, for public health programs to be most effective, it seems reasonable to focus education interventions on all high private well use areas, and to allocate more resources for infrastructure/grant funding to high private well use areas that are also the most socially vulnerable. We identified six distinct population profiles that are more likely to rely on a domestic well in Pennsylvania. Two of these profiles are also more likely to experience social vulnerability when measured at the census tract level. In general, census tracts with higher proportions of homeowners, lower proportions of those without a high school diploma, lower median per capita income, and higher proportions of children under the age of 17 are more likely to rely on an unregulated drinking water source. These profiles need to be considered when designing interventions focused on increasing access to public water supply or when providing resources and education to homeowners in locations where it is not practical or feasible to connect to a public water supply system.
Supporting information
S1 Fig.
Estimated number of homes served by private wells by census tract (top) compared with population density (middle) for the state of Pennsylvania circa 2020 including overlay comparison (bottom) (greyed out sections on the maps are 100% served by public water). Maps generated with https://www2.census.gov/geo/pvs/tiger2010st/42_Pennsylvania/42/tl_2010_42_tract10.zip, https://www2.census.gov/geo/pdfs/maps-data/data/tiger/tgrshp2010/TGRSHP10SF1.pdf, USA Major Cities—(arcgis.com).
https://doi.org/10.1371/journal.pwat.0000303.s001
(DOCX)
S2 Fig. Random effects due to county for negative binomial model.
https://doi.org/10.1371/journal.pwat.0000303.s002
(DOCX)
S1 Table. Table of proportion of area not served by public water supply by census tract.
https://doi.org/10.1371/journal.pwat.0000303.s003
(DOCX)
S2 Table. Conditional model: Zero-inflated negative binomial model using glmmTMB.
3191 census tracts, 67 counties.
https://doi.org/10.1371/journal.pwat.0000303.s004
(DOCX)
S3 Table. Zero-inflation model portion of model listed in S2 Table.
https://doi.org/10.1371/journal.pwat.0000303.s005
(DOCX)
S4 Table. Estimated confidence intervals for estimates in S2 Table using wald method.
https://doi.org/10.1371/journal.pwat.0000303.s006
(DOCX)
References
- 1. PADCNR. Pennsylvania Department of Conservation & Natural Resources. [cited 2019 Apr 3]. Groundwater. Available from: https://www.dcnr.pa.gov:443/Conservation/Water/Groundwater/Pages/default.aspx
- 2. Mattos KJ, Mulhern R, Naughton CC, Anthonj C, Brown J, Brocklehurst C, et al. Reaching those left behind: knowledge gaps, challenges, and approaches to achieving SDG 6 in high-income countries. Journal of Water, Sanitation and Hygiene for Development [Internet]. 2021 Jul 28 [cited 2022 Aug 17];11(5):849–58. Available from: https://doi.org/10.2166/washdev.2021.057
- 3. WHO and UNICEF. Progree on Drinking Water (SDG targe 6.1). 2020 [cited 2023 Feb 23]. Joint Monitoring Programme for /water Supply, Sanitation and Hygiene (JMP) Indicator | SDG 6 Data. Available from: https://sdg6data.org/en/indicator/6.1.1
- 4. USEPA. Summary of the Safe Drinking Water Act | Laws & Regulations | US EPA [Internet]. 1974 [cited 2018 Dec 3]. Available from: https://www.epa.gov/laws-regulations/summary-safe-drinking-water-act
- 5. USEPA Office of Water. The History of Drinking Water [Internet]. 2000 Feb [cited 2021 Mar 17]. Report No.: EPA-816-F-00-006. Available from: https://esa21.kennesaw.edu/modules/water/drink-water-trt/drink-water-trt-hist-epa.pdf
- 6. United Nations. Sustainable Development Goals. 2015 [cited 2022 Sep 8]. Goal 6 | Department of Economic and Social Affairs. Available from: https://sdgs.un.org/goals/goal6
- 7. Mueller JT, Gasteyer S. The widespread and unjust drinking water and clean water crisis in the United States. Nat Commun [Internet]. 2021 Jun 22 [cited 2022 Mar 26];12(1):3544. Available from: http://www.nature.com/articles/s41467-021-23898-z pmid:34158491
- 8. US EPA O. Potential Well Water Contaminants and Their Impacts [Internet]. 2015 [cited 2023 Mar 6]. Available from: https://www.epa.gov/privatewells/potential-well-water-contaminants-and-their-impacts
- 9. Munene A, Hall DC. Factors influencing perceptions of private water quality in North America: a systematic review. Syst Rev [Internet]. 2019 Dec [cited 2019 Nov 18];8(1):111. Available from: https://systematicreviewsjournal.biomedcentral.com/articles/10.1186/s13643-019-1013-9 pmid:31077249
- 10. US EPA O. Secondary Drinking Water Standards: Guidance for Nuisance Chemicals [Internet]. 2015 [cited 2023 Mar 6]. Available from: https://www.epa.gov/sdwa/secondary-drinking-water-standards-guidance-nuisance-chemicals
- 11. De Roos AJ, Gurian PL, Robinson LF, Rai A, Zakeri I, Kondo MC. Review of Epidemiological Studies of Drinking-Water Turbidity in Relation to Acute Gastrointestinal Illness. Environ Health Perspect [Internet]. 2017 Aug 17 [cited 2023 Mar 6];125(8):086003. Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5882241/ pmid:28886603
- 12. CDC NCEZID. Avoid Contaminated Water During Travel | Travelers’ Health | [Internet]. 2022 [cited 2023 Mar 6]. Available from: https://wwwnc.cdc.gov/travel/page/water-disinfection
- 13. Eccles KM, Checkley S, Sjogren D, Barkema HW, Bertazzon S. Lessons learned from the 2013 Calgary flood: Assessing risk of drinking water well contamination. Applied Geography [Internet]. 2017 Mar [cited 2022 Sep 19];80:78–85. Available from: https://linkinghub.elsevier.com/retrieve/pii/S0143622817301534
- 14. Gilliland AE, Pieper KJ, Straif-Bourgeois S, Rhoads WJ, Dai D, Edwards M, et al. Evaluation of Preparedness and Recovery Needs of Private Well Users After the Great Louisiana Flood of 2016. J Public Health Manag Pract. 2021 Dec 1;27(6):577–87. pmid:32332488
- 15. Odimayomi TO, Proctor CR, Wang QE, Sabbaghi A, Peterson KS, Yu DJ, et al. Water safety attitudes, risk perception, experiences, and education for households impacted by the 2018 Camp Fire, California. Natural Hazards: Journal of the International Society for the Prevention and Mitigation of Natural Hazards [Internet]. 2021 [cited 2022 Sep 19];108(1):947–75. Available from: http://econpapers.repec.org/article/sprnathaz/v_3a108_3ay_3a2021_3ai_3a1_3ad_3a10.1007_5fs11069-021-04714-9.htm
- 16. Pieper KJ, Jones CN, Rhoads WJ, Rome M, Gholson DM, Katner A, et al. Microbial Contamination of Drinking Water Supplied by Private Wells after Hurricane Harvey. Environ Sci Technol [Internet]. 2021 Jun 15 [cited 2022 Sep 19];55(12):8382–92. Available from: pmid:34032415
- 17.
Dieter CA, Maupin MA, Caldwell RR, Harris MA, Ivahnenko TI, Lovelace JK, et al. Estimated use of water in the United States in 2015 [Internet]. Reston, VA: U.S. Geological Survey; 2018 [cited 2019 Mar 23] p. 76. (Circular). Report No.: 1441. Available from: http://pubs.er.usgs.gov/publication/cir1441
- 18. MacDonald Gibson J, DeFelice N, Sebastian D, Leker H. Racial Disparities in Access to Community Water Supply Service in Wake County, North Carolina. Am J Public Health [Internet]. 2014 Dec [cited 2021 Mar 22];104(12):e45–e45. Available from: http://ajph.aphapublications.org/doi/10.2105/AJPH.2014.10412e45.1
- 19. Leker HG, Gibson JM. Relationship between race and community water and sewer service in North Carolina, USA. PLOS ONE [Internet]. 2018 Mar 21 [cited 2021 Mar 22];13(3):e0193225. Available from: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0193225 pmid:29561859
- 20. Clough E, Bell D. Just fracking: a distributive environmental justice analysis of unconventional gas development in Pennsylvania, USA. Environ Res Lett [Internet]. 2016 Feb [cited 2022 Aug 4];11(2):025001. Available from: https://doi.org/10.1088/1748-9326/11/2/025001
- 21. Molitor J, Su JG, Molitor NT, Rubio VG, Richardson S, Hastie D, et al. Identifying vulnerable populations through an examination of the association between multipollutant profiles and poverty. Environ Sci Technol. 2011 Sep 15;45(18):7754–60. pmid:21797252
- 22. Rufat S, Tate E, Emrich CT, Antolini F. How Valid Are Social Vulnerability Models? Annals of the American Association of Geographers [Internet]. 2019 Jul 4 [cited 2024 Mar 10];109(4):1131–53. Available from: https://doi.org/10.1080/24694452.2018.1535887
- 23. CDC/ATSDR GRASP. CDC Social Vulnerability Index. 2018, 2016, 2014. Pennsylvania by Census Tract [Internet]. 2020 [cited 2021 Mar 17]. Available from: https://www.atsdr.cdc.gov/placeandhealth/svi/data_documentation_download.html
- 24. Flanagan BE, Hallisey EJ, Adams E, Lavery A. Measuring Community Vulnerability to Natural and Anthropogenic Hazards: The Centers for Disease Control and Prevention’s Social Vulnerability Index. J Environ Health [Internet]. 2018 Jun [cited 2021 Mar 17];80(10):34–6. Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7179070/ pmid:32327766
- 25. CDC/ATSDR. CDC/ATSDR Social Vulnerability Index (SVI) [Internet]. 2022 [cited 2023 Jan 18]. Available from: https://www.atsdr.cdc.gov/placeandhealth/svi/index.html
- 26.
National Collaborating Centre for Determinants of Health. Glossary of Essential Health Equity Terms. In Nova Scotia, Canada; 2023 [cited 2023 Jan 23]. Available from: https://nccdh.ca/glossary/entry/vulnerable-populations
- 27. PA DEP. Pennsylvania Spatial Data Access | Full Metadata. 2019 [cited 2019 May 6]. Public Water Supplier’s Service Area. Available from: ftp://ftp.pasda.psu.edu/pub/pasda/dep/PublicWaterSupply2019_04.zip
- 28. US Census Bureau. Social Explorer. 2019 [cited 2021 Mar 15]. ACS 2019 (5-year estimates). Available from: https://www.socialexplorer.com/tables/ACS2019_5yr
- 29. US Census Bureau. TIGER/Line Shapefile, 2017, 2010 state, Pennsylvania, 2010 Census Block State-based [Internet]. Unknown; 2010 [cited 2022 Jan 27]. Available from: https://catalog.data.gov/dataset/tiger-line-shapefile-2017-2010-state-pennsylvania-2010-census-block-state-based
- 30.
ESRI. ArcPRO [Internet]. Redlands, CA; 2020 [cited 2022 Jan 24]. Available from: https://pro.arcgis.com/en/pro-app/2.8/get-started/release-notes.htm
- 31. U. S. Census Bureau. ACS 2018 (5-Year Estimates) [Internet]. 2019 Sep [cited 2022 Jan 31]. Available from: https://www.socialexplorer.com/tables/ACS2018_5yr/R12897432
- 32.
CDC/ATSDR. CDC SVI 2018 Data and Documentation [Internet]. CDC; 2020 [cited 2021 Feb 19]. Available from: https://www.atsdr.cdc.gov/placeandhealth/svi/data_documentation_download.html
- 33. Coker ES, Molitor J, Liverani S, Martin J, Maranzano P, Pontarollo N, et al. Bayesian profile regression to study the ecologic associations of correlated environmental exposures with excess mortality risk during the first year of the Covid-19 epidemic in lombardy, Italy. Environ Res [Internet]. 2023 Jan 1 [cited 2023 Feb 23];216(Pt 1):114484. Available from: https://europepmc.org/articles/PMC9547389 pmid:36220446
- 34. Drakes O, Tate E, Rainey J, Brody S. Social vulnerability and short-term disaster assistance in the United States. International Journal of Disaster Risk Reduction [Internet]. 2021 Feb 1 [cited 2023 Apr 12];53:102010. Available from: https://www.sciencedirect.com/science/article/pii/S2212420920315120
- 35. Coker E, Liverani S, Ghosh JK, Jerrett M, Beckerman B, Li A, et al. Multi-pollutant exposure profiles associated with term low birth weight in Los Angeles County. Environment International [Internet]. 2016 May 1 [cited 2021 Mar 17];91:1–13. Available from: https://www.sciencedirect.com/science/article/pii/S0160412016300460 pmid:26891269
- 36. Coker E, Liverani S, Su JG, Molitor J. Multi-pollutant Modeling Through Examination of Susceptible Subpopulations Using Profile Regression. Curr Envir Health Rpt [Internet]. 2018 Mar 1 [cited 2021 Mar 29];5(1):59–69. Available from: pmid:29427169
- 37. Leyk S, Phillips TP, Smith JM, Nuckols JR. Spatio-temporal Trends of Diarrheal Mortality of Children in Association with Hydrographic Regions of Brazil. Cartography and Geographic Information Science. 2011 Apr;38(2):223–32.
- 38. Brunsdon C, Fotheringham AS, Charlton ME. Geographically Weighted Regression: A Method for Exploring Spatial Nonstationarity. Geographical Analysis [Internet]. 1996 [cited 2023 Apr 12];28(4):281–98. Available from: https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1538-4632.1996.tb00936.x
- 39. Liverani S, Lavigne A, Blangiardo M. Modelling collinear and spatially correlated data. Spatial and Spatio-temporal Epidemiology [Internet]. 2016 Aug 1 [cited 2021 Mar 22];18:63–73. Available from: https://www.sciencedirect.com/science/article/pii/S1877584515300411 pmid:27494961
- 40.
Haining RP, Li G. Regression Modelling wih Spatial and Spatial-Temporal Data: A Bayesian Approach. CRC Press; 2020. 527 p.
- 41. Neal RM. Markov Chain Sampling Methods for Dirichlet Process Mixture Models. Journal of Computational and Graphical Statistics [Internet]. 2000 Jun [cited 2023 Mar 6];9(2):249–65. Available from: http://www.tandfonline.com/doi/abs/10.1080/10618600.2000.10474879
- 42. Liverani S, Hastie DI, Azizi L, Papathomas M, Richardson S. PReMiuM: An R Package for Profile Regression Mixture Models Using Dirichlet Processes. J Stat Soft [Internet]. 2015 [cited 2020 Nov 18];64(7). Available from: http://www.jstatsoft.org/v64/i07/
- 43. Molitor J, Papathomas M, Jerrett M, Richardson S. Bayesian profile regression with an application to the National survey of children’s health. Biostatistics. 2010 Jul 1;11(3):484–98. pmid:20350957
- 44.
Gelman A, Hill J. Data analysis using regression and multilevel/hierarchical models. Cambridge; New York: Cambridge University Press; 2007. 625 p. (Analytical methods for social research).
- 45. Kuznetsova A, Brockhoff PB, Christensen RHB. lmerTest Package: Tests in Linear Mixed Effects Models. J Stat Soft [Internet]. 2017 [cited 2022 May 11];82(13). Available from: http://www.jstatsoft.org/v82/i13/
- 46. Kuznetsova A, Brockhoff PB, Christensen RHB, Jensen SP. lmerTest: Tests in Linear Mixed Effects Models [Internet]. 2020 [cited 2022 May 11]. Available from: https://CRAN.R-project.org/package=lmerTest
- 47. Brooks M E, Kristensen K, Benthem K J, van, Magnusson A, Berg C W, Nielsen A, et al. glmmTMB Balances Speed and Flexibility Among Packages for Zero-inflated Generalized Linear Mixed Modeling. The R Journal [Internet]. 2017 [cited 2023 Jan 31];9(2):378. Available from: https://journal.r-project.org/archive/2017/RJ-2017-066/index.html
- 48. Rollings KA, Noppert GA, Griggs JJ, Melendez RA, Clarke PJ. Comparison of two area-level socioeconomic deprivation indices: Implications for public health research, practice, and policy. PLOS ONE [Internet]. 2023 Oct 5 [cited 2024 Mar 10];18(10):e0292281. Available from: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0292281 pmid:37797080
- 49.
Lee D, Gibson JM, Brown J, Habtewold J, Murphy HM. Burden of Disease from Contaminated Drinking Water in Countries with High Access to Safely Managed Water: A Systematic Review [Internet]. medRxiv; 2022 [cited 2022 Aug 22]. p. 2022.03.03.22271862. Available from: https://www.medrxiv.org/content/10.1101/2022.03.03.22271862v1
- 50.
Roller Z, Gasteyer S. Dig-Deep_Closing-the-Water-Access-Gap-in-the-United-States_DIGITAL_compressed.pdf [Internet]. Washington DC: Dig Deep and US Water Alliance; 2019 [cited 2022 Aug 17]. Available from: https://static1.squarespace.com/static/5e80f1a64ed7dc3408525fb9/t/6092ddcc499e1b6a6a07ba3a/1620237782228/Dig-Deep_Closing-the-Water-Access-Gap-in-the-United-States_DIGITAL_compressed.pdf
- 51. Tolnay SE. The African American “Great Migration” and Beyond. Annual Review of Sociology [Internet]. 2003 [cited 2023 Apr 11];29:209–32. Available from: https://www.jstor.org/stable/30036966
- 52.
Larkins J. Summary of The Negro Population of North Carolina: Social and Economic [Internet]. Raleigh, NC: State Board of Charities and Public Welfare; 1944 [cited 2023 Apr 11]. Available from: https://docsouth.unc.edu/nc/larkins/summary.html
- 53. Meshel T. Environmental Justice in the United States: The Human Right to Water. Wash J Envtl L & Pol’y [Internet]. 2018 [cited 2023 Apr 11];8(2):264–97. Available from: https://heinonline.org/HOL/P?h=hein.journals/wshjoop8&i=276
- 54. Schaider LA, Swetschinski L, Campbell C, Rudel RA. Environmental justice and drinking water quality: are there socioeconomic disparities in nitrate levels in U.S. drinking water? Environmental Health [Internet]. 2019 Jan 17 [cited 2021 Mar 23];18(1):3. Available from: pmid:30651108
- 55. Gaber N, Silva A, Lewis-Patrick M, Kutil E, Taylor D, Bouier R. Water insecurity and psychosocial distress: case study of the Detroit water shutoffs. Journal of Public Health [Internet]. 2021 Dec 1 [cited 2024 Mar 10];43(4):839–45. Available from: pmid:32930795
- 56. Meehan K, Jepson W, Harris LM, Wutich A, Beresford M, Fencl A, et al. Exposing the myths of household water insecurity in the global north: A critical review. WIREs Water [Internet]. 2020 [cited 2024 Mar 10];7(6):e1486. Available from: https://onlinelibrary.wiley.com/doi/abs/10.1002/wat2.1486
- 57. JCHS. Joint Center for Housing Studies of Harvard University. 2020 [cited 2022 Aug 22]. The Racial Disparity in Homeownership Rates Persists. Available from: https://www.jchs.harvard.edu/son-2020-homeownership-gap
- 58. Brown J, Acey CS, Anthonj C, Barrington DJ, Beal CD, Capone D, et al. The effects of racism, social exclusion, and discrimination on achieving universal safe water and sanitation in high-income countries. The Lancet Global Health [Internet]. 2023 Apr 1 [cited 2023 Apr 11];11(4):e606–14. Available from: https://www.thelancet.com/journals/langlo/article/PIIS2214-109X(23)00006-2/fulltext pmid:36925180
- 59. Cusick D. House Method: Studies and Information. 2022 [cited 2022 Aug 22]. The Median Age of Homes in the United States by Build Year. Available from: https://housemethod.com/home-warranty/median-home-age-us/
- 60. Dumic I, Nordin T, Jecmenica M, Stojkovic Lalosevic M, Milosavljevic T, Milovanovic T. Gastrointestinal Tract Disorders in Older Age. Can J Gastroenterol Hepatol [Internet]. 2019 Jan 17 [cited 2022 Sep 9];2019:6757524. Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6354172/ pmid:30792972
- 61. Gorelick MH, McLELLAN SL, Wagner D, Klein J. Water use and acute diarrhoeal illness in children in a United States metropolitan area. Epidemiology & Infection [Internet]. 2011 Feb [cited 2022 Sep 9];139(2):295–301. Available from: http://www.cambridge.org/core/journals/epidemiology-and-infection/article/water-use-and-acute-diarrhoeal-illness-in-children-in-a-united-states-metropolitan-area/830A492C11471CE7ADC5C04044C60EA6
- 62. Coker E, Katamba A, Kizito S, Eskenazi B, Davis JL. Household air pollution profiles associated with persistent childhood cough in urban Uganda. Environment International. 2020;136. pmid:32044526
- 63. Roback P, Legler J. Chapter 8 Introduction to Multilevel Models | Beyond Multiple Linear Regression [Internet]. 2021 [cited 2022 May 11]. Available from: https://bookdown.org/roback/bookdown-BeyondMLR/ch-multilevelintro.html
- 64. PA DCNR. Pennsylvania Department of Conservation & Natural Resources. [cited 2019 May 6]. PA Groundwater Information System. Available from: https://www.dcnr.pa.gov:443/Conservation/Water/Groundwater/PAGroundwaterInformationSystem/Pages/default.aspx
- 65. PA State Data Center. Estimates [Internet]. [cited 2020 Apr 16]. Available from: https://pasdc.hbg.psu.edu/Data/Estimates
- 66. Jones AQ, Dewey CE, Doré K, Majowicz SE, McEwen SA, Waltner-Toews D, et al. Public perception of drinking water from private water supplies: focus group analyses. BMC Public Health [Internet]. 2005 Dec [cited 2023 Apr 11];5(1):1–12. Available from: https://bmcpublichealth.biomedcentral.com/articles/10.1186/1471-2458-5-129 pmid:16336678
- 67. Munene A, Hall DC. Factors influencing perceptions of private water quality in North America: a systematic review. Syst Rev [Internet]. 2019 Dec [cited 2019 Nov 18];8(1):111. Available from: https://systematicreviewsjournal.biomedcentral.com/articles/10.1186/s13643-019-1013-9 pmid:31077249