Lead occurrence in North Carolina well water: importance of sampling representation and collection techniques

Private wells often lack centralized oversight, drinking water quality standards, and consistent testing methodologies. For lead in well water, the lack of standardized data collection methods can impact reported measurements, which can misinform health risks. Here, we conducted a targeted community science testing of 1143 wells across 17 counties in North Carolina (USA) and compared results to state testing data primarily associated with new well construction compiled in the NCWELL database. The goal of our study was to explore the impacts of sampling methodology and household representation on estimated lead exposures and subsequent health risks. At the household scale, we illustrated how sampling and analytical techniques impact lead measurements. The community science testing first draw samples (characterizing drinking water) had a 90th percentile lead value of 12.8 μg l−1 while the NCWELL database flushed samples (characterizing groundwater) had a value below the reporting level of 5 μg l−1. As lead was associated with the corrosion of premise plumbing, flushing prior to collection substantially reduced lead concentrations. At the community scale, we examined how the lack of representation based on household demographics and well construction characteristics impacted the knowledge of lead and blood lead level (BLL) occurrence. When simulating representative demographics of the well populations, we observed that the 90th percentile lead level could differ by up to 6 μg l−1, resulting in communities being above the USEPA action level. This translated to a 1.0–1.3 μg dl−1 difference in predicted geometric mean BLL among infants consuming reconstituted formula. Further, inclusion of less common well construction types also increased lead in water occurrence. Overall, under- and overestimations of lead concentrations associated with differences in sampling techniques and sample representation can misinform conclusions about risks of elevated BLLs associated with drinking water from private wells which may hinder investigations of waterborne lead exposure.


Introduction
The lead burden in drinking water supplied by private wells in the United States (US) is an issue that transcends households, communities, and state-level governmental agencies.Studies report that up to 19% of private wells tested exceed the US Environmental Protection Agency lead action level of 15 µg l −1 , with concentrations as high as 24 740 µg l −1 [1][2][3].Well water continues to be considered a high-risk drinking water source, especially with case studies reporting that lead concentrations and risks of elevated BLLs are higher for consumers connected to a private well compared to those on a municipal supply [4,5].This is particularly concerning for children and infants, as waterborne exposures can cause elevated blood lead levels (BLLs) [6][7][8].Even low levels in water are considered harmful, as chronic exposures as low as 1 µg l −1 can increase a child's BLL by 35% after 150 d [9].
Lead in well water is associated with the corrosion of lead-bearing plumbing [10], and its release is shaped by geologic setting and system construction [11,12].As most well users do not install corrosion control treatment, the dissolution of lead into drinking water is controlled by the chemistry of the untreated groundwater [2,3,12].To identify lead sources, there has been substantial research investigating sampling techniques [13][14][15][16].Researchers have documented that sampling (e.g.flushing prior to collection) and analytical approaches (e.g.acidification protocols) can result in the over-and underestimation of lead in drinking water [2,10].However, most knowledge is generated based on sampling of municipal water, and there has not been the same level of attention given to characterizing lead in well water.As well water testing is not standardized and methods vary considerably, there are concerns that such differences in data collection methods are influencing our knowledge of the occurrence of lead in well water [17,18].
In this study, we focus on North Carolina (NC) where 2.4 million residents (24% of the population) rely on private wells for their drinking water supply [19].Well water testing data collected by the NC Department of Health and Human Services (DHHS) documents that 2.9% of private wells exceed 15 µg l −1 [20,21].However, this is inconsistent with another study in the state documenting higher lead levels in well water (33.5% of wells had >10 µg l −1 ) [5] and rates reported in nearby states with similar geology (12%-19% of wells had >15 µg l −1 ) [2,3].In addition, DHHS well water samples were primarily from predominately White block groups, despite 26% of well-using households estimated to identify as black, indigenous, and people of color (BIPOC) [22].Representation in sampling is important because lowincome and BIPOC households are more likely to have at-risk drinking water systems and resulting well water contamination [4,[23][24][25][26].
To explore these discontinuities, we conducted a targeted community science testing effort and compared results to the DHHS well water testing data [21].At the household scale, we (1) quantify lead in well water associated with corrosion of premise plumbing and (2) evaluate differences in lead concentrations due to sampling and analysis techniques.At the community scale, we examine how (3) data collection methods influence BLL occurrence and (4) the lack of representation based on household, system, and geology characteristics impact our understanding of lead occurrence within the private well community.

Lead sampling efforts
In 2018-2019, our team collected drinking water samples for lead analysis from 1,143 private wells in NC (figure 1; section S1) [27].We refer to this dataset as the 'community science testing.'We hosted well water testing clinics in Chatham (237 samples), Iredell (780 samples), New Hanover (25 samples), Northampton (92 samples), and Robeson (9 samples) counties, and allowed residents in neighboring counties to participate (figure S1).We advertised our free water testing events through local media outlets, cooperative extension, non-profits, and local health departments.Interested well users could receive a kit at the local health department, distribution events hosted by the research team, or through USPS mail.Participants were provided a sampling kit that included: (i) sampling instructions; (ii) a household survey; and (iii) sampling bottles.On a predetermined morning, participants collected a 250 ml first draw sample after 6+ hours of stagnation followed by a 250 ml flushed sample after 5 min of flushing at normal flow (∼8.3 l min −1 [2.2 gal min −1 ]) from a drinking water tap (section S2).After, participants were asked to complete a survey about their well system and household demographics (section S3).We determined the geologic setting of the wells by matching mailing address with geology data from the USGS Ground Water Atlas [28].
Participants returned sampling kits to designated locations, and we analyzed samples immediately for pH per method 4500-H + [29].Samples that were shipped did not receive pH analysis (n = 102).In the laboratory, we acidified samples with 2% v/v concentrated nitric acid and digested for a minimum of 16 h before analysis using inductively coupled plasma-mass spectrometry per method 3125 B [29].Our detection limit was 0.1 µg l −1 for lead, cadmium, and tin, 1 µg l −1 for copper and zinc, and 5 µg l −1 for iron.Blanks and spikes of known concentrations were processed every ten samples for QA/QC purposes.Participation was voluntary and all procedures were approved by Virginia Tech Institutional Review Board (#16-918 and #19-149).Participants received confidential water quality results within 4-6 weeks and were invited to attend a community meeting hosted by the research team.
We retrieved geocoded DHHS well water testing data from the NCWELL database [21], which was work that curated testing records.From 1998-2019, the DHHS tested 110 409 private wells for lead (figure S2; section S1).We refer to this dataset as the 'NCWELL database.'Samples were typically collected by local health department officials at the wellhead (flowrate of ∼56.8 l min −1 [15 gal min −1 ]) and after flushing for several minutes to be representative of the groundwater.Residents paid sampling costs that varied from $25-$200 (median of $90) depending on the health department [30].Samples were analyzed for metals, with a minimum reporting level (MRL) of 1 µg l −1 for cadmium, 5 µg l −1 for lead, 50 µg l −1 for copper and zinc, and 100 µg l −1 for iron.Tin was not analyzed.As before, we determined the geologic setting by matching mailing address with geology data [28].
We compared addresses in the community science testing and the NCWELL database and identified 50 private wells that were included in both.Of these, the NCWELL database samples were collected between 1999 to 2019, with 72% (n = 36) collected in 2010 or after (table S1).There were 14 wells that had multiple testing records (2-7 times), so we used the most recent lead result in our paired analysis.We compared lead concentrations above the MRL in our paired analysis, with the community science testing having a MRL of 1 µg l −1 .

Household demographics
In the community science testing, we asked participants to describe the race and ethnicity of persons within their household (section S3).We classified samples as being from a BIPOC household when at least one member identified as BIPOC and a White household when all members identified as White/Caucasian.We used the well-using household BIPOC percentages at the county and state level from Hayes et al (in review) [22] and well location predictions from Murray et al [19].This analysis was not done in New Hanover or Robeson as the sample sizes were too small.

Blood lead modeling
We used the Integrated Exposure Uptake Biokinetic Model (IEUBK; win32 Version 1.1) to predict BLLs based on lead in water concentrations in the community science testing first draw samples and the NCWELL database flushed samples.We modeled our analysis on the approach in Triantafyllidou et al [31], where the authors proposed three exposure scenarios for children and infants.The first scenario was exposure for a 1-2-year-old child drinking tap water.We used IEUBK default values for consumption (430 mL d −1 ), geometric standard deviation (GSD; 1.6 µg dl −1 ), and lead exposures other than water.The second scenario was exposure for 0-1year-olds with an average consumption of infant formula.We used 800 ml d −1 for consumption, a GSD of 1.45 µg dl −1 , a dietary intake of 0 µg d −1 , and set all other lead exposures to default values.The third scenario was exposure for 0-1-year-olds with a high consumption of infant formula.We used the same parameters as scenario 2 but changed the consumption rate to 1,200 ml d −1 .We used the model outputs of geometric means and estimated proportions of children with BLLs over 3.5 µg dl −1 to explore lead exposure outcomes.We expanded this analysis to determine how representation in sampling can impact predicted BLLs for 0-1-year-olds with average and high formula consumption.We repeated our IEUBK modeling for scenarios 2 and 3 but used our simulated lead in water concentrations.Specifically, we used the average 90th percentile values calculated during our bootstrapping analysis as our lead in water model input values.

Statistical analysis
We performed analysis in R version 4.0.4 using an alpha of 0.05.Concentrations below the community science testing detection level and NCWELL database MRL were set to half the level for analysis unless otherwise noted.We used Spearman's correlation test to evaluate associations between lead and other water parameters.We used the Wilcoxon test to compare differences in lead and pH based on construction and household characteristics and across datasets.We used the Kolmogorov-Smirnov test to evaluate the similarity of flushed lead distributions in the community science testing and NCWELL database.
We simulated theoretical sampling events with different demographic compositions using a bootstrapping approach.We performed this analysis for the community science testing first draw data only, and all lead concentrations below detection were set to 0 µg l −1 .For each simulation, we used a total sample size of 100 samples, randomly selected (without replacement) lead concentrations based on the target demographic composition, and then calculated the mean lead concentration.We varied the percentage of BIPOC households from 0%-100% by intervals of 10%.We generated 1000 simulations per target demographic composition and calculated the average 90th percentile and 95th confidence interval.

Corrosion of premise plumbing
In the community science testing, first draw lead concentrations ranged from ⩽0.1 to 1,327 µg l −1 , with a 90th percentile of 12.8 µg l −1 and 8% exceeding 15 µg l −1 (figure 2; table S4).Our results suggest lead was associated with brass corrosion as it was strongly correlated (Spearman's correlation, p < 0.05) with copper (ρ = 0.59) and zinc (ρ = 0.66).There was some indication of corrosion of solder (19% of samples contained tin) and galvanized iron (21% of samples contained cadmium).Lead was negatively correlated with pH (ρ = −0.46),further indicating corrosion of the plumbing components.We did not observe a difference in median lead between bedrock (2.0 µg l −1 ) and sand wells (2.7 µg l −1 ) (Wilcoxon Test, p = 0.17).This was likely attributed to the high percentage of shallow (<100 ft) sand wells sampled (28%; table S2), which had more corrosive water (Wilcoxon Test, p < 0.05).Specifically, the median pH of shallow sand wells was 5.6 compared to 7.7 of deep sand wells.

Lead occurrence in NCWELL database sampling
Lead in the NCWELL database ranged from <5 to 8680 µg l −1 , with a 90th percentile <5 µg l −1 and 2.8% of samples exceeding 15 µg l −1 (figure 2; table S4).Only 8% of samples contained measurable lead (>5 µg l −1 ).Measurable lead was not significantly associated with brass corrosion (ρ Cu = 0.29 and ρ Zn = 0.22), and there was little release from galvanized iron (0.3% of samples contained cadmium).Solder corrosion was not evaluated.Among samples with lead >5 µg l −1 , there was no difference in the distribution of lead concentration between the NCWELL database and the community science testing flushed data (Kolmogorov-Smirnov Test, p = 0.44).There was also no difference in median lead (Wilcoxon test, p = 0.37) between the NCWELL database (11 µg l −1 ) and the community science flushed data (10.2 µg l −1 ) among samples with lead concentrations >5 µg l −1 .
In the 50 private wells that were in both the NCWELL database and the community science testing, lead ranged from <5 to 78 µg l −1 in the NCWELL database flushed samples and <1.0-24 µg l −1 in the community science first draw samples (figure S4).Both testing efforts agreed on the absence of lead in 13 (26%) private wells (figure S5; table S5).For most samples (66%), the NCWELL database reported non-measurable lead (<5 µg l −1 ), while the community science testing reported measurable lead (>1 µg l −1 ).Among these wells, we observed that 64% of the community science first draw samples contained lead between 1-5 µg l −1 , and 36% of samples contained >5 µg l −1 .At four private wells (8%), the NCWELL database reported measurable lead, while the community science testing reported non-measurable.

Blood lead predictions
Using lead measurements from the community science testing, we predicted that consuming first draw water would result in a geometric mean BLL for infants and children of 2.6-7.0 µg dl −1 (table 1).For 1-2-year-old children, BLLs were 3.1 and 4.6 µg dl −1 when consuming water with the 50th percentile and 95th percentile lead in water concentrations.Interestingly, there was only a 1.5 µg dl −1 difference in geometric mean BLLs despite water concentrations differing by 19.3 µg l −1 .When exposed to these lead in water concentrations, 40%-72% of exposed children would be expected to have BLLs greater than 3.5 µg dl −1 .For formula-fed infants 0-1-years old, lead in drinking water had a more substantial impact.There was a 3.0 µg dl −1 difference in geometric mean BLL for average formula consumption and 4.3 µg dl −1 difference for high formula consumption between the 50th and 95th percentile lead in water concentration.Average and high consumption of formula made with water containing 21.3 µg l −1 lead (95th percentile) were predicted to result in BLLs >3.5 µg dl −1 for almost all infants (89% and 97%).
In the NCWELL database, the 95th lead in water percentile of 8 µg l −1 resulted in predicted geometric mean BLLs ranging from 3.6-4.2µg dl −1 , which were 1.0-2.8µg dl −1 lower than the BLLs predicted using the community science 95th lead in water percentile.However, 52%-69% of infants and children were still predicted to have a BLL that exceeded the 3.5 µg dl −1 .
Using a state-level percent BIPOC estimate may be misleading, as the communities where we sampled had varying demographics.In Chatham County, 29% of households reliant on private wells are predicted to be BIPOC, 23% in Iredell County, and 61% in Northampton County.Using these more communityrepresentative percent BIPOC estimates, the average 90th percentile lead varied from 13.7 µg l −1 (7.9-20.7 µg l −1 ) to 16.5 µg l −1 (11.8-22.0µg l −1 ; table S6).This was up to a 3.5 µg l −1 difference compared to the 90th percentile reported in the community science testing and up to 2.5 µg l −1 difference from the 90th percentile reported when simulating a 26% BIPOC population.
We evaluated the impact of representation on predicted geometric mean BLLs using the average 90th percentile lead concentration when 0% to 100% of simulated samples were from BIPOC households.All geometric mean BLLs were above the CDC reference value of 3.5 µg dl −1 , with a predicted 68%-85% and 84%-95% of infants having BLLs above 3.5 µg dl −1 with average and high formula  S6).Although 90th percentile water lead level differed by 6.5 µg l −1 , there was only a 1.0 (4.2-5.2 µg dl −1 ) and 1.3 µg dl −1 (5.1-6.4 µg dl −1 ) difference in predicted BLL for average and high formula consumption.

Discussion
The results from the community science testing illustrate that lead in well water is a public health concern in NC.Lead was associated with corrosion of lead-bearing plumbing components, as indicated by the co-occurrence of corrosion metals and reduction associated with flushing.Our findings are consistent rates observed in neighboring states [2,3], but was almost three times higher than the rate observed in the NCWELL database.While discrepancies in reported lead measurement were not surprising due to differences in data collection techniques [2,17], such discrepancies can impact actions and behaviors at both the household and community level.

Household-level impacts
Lead in drinking water can be the primary source of lead exposure that results in elevated BLLs among infants and children [6,17,32].While leaded solders (up to 50% lead) were banned in 1986, lead content in brasses (up to 8% by weight) and galvanized iron (up to 2% in surface coating) was only reduced to a weighted average of 0.25% based on wetted surfaces in 2011 [10].Private wells can be higher-risk for lead in water exposure than municipal water, as lead leaching from 'lead-free' plumbing can be up to 17 times higher than expected when exposed to more well water [33].This is further cemented by work in Wake County, NC documenting that well users were 25% more likely to have elevated BLLs compared to neighbors on municipal supply [4].Thus, characterizing and quantifying lead exposure attributed to well water is imperative to develop communication strategies that promote protective water use behaviors.
Discrepancies in lead measurements between the community science testing and NCWELL database were likely attributed to differences in sampling techniques [2,13,17].As lead release was associated with corrosion, flushing prior to collection and collecting samples at the wellhead can underestimate lead in drinking water as contaminated water within the plumbing system is not assessed [17,18].Results from our study underscore these differences-our community science testing reported 8% of samples exceeded 15 µg l −1 based on first draw sampling whereas the NCWELL database reported 3% exceedance based on flushed samples.These findings are in keeping with prior work in Macon County, NC documenting that first draw lead concentrations at the wellhead (30 µg l −1 ) were significantly higher than at the kitchen sink (10.5 µg l −1 ) [17].However, after flushing at the wellhead for one minute, lead reduced to non-measurable levels (<1 µg l −1 ) in 67% of systems.Difference in sampling techniques between community science and the NCWELL database efforts presented were ultimately linked to original goals of the respective sampling efforts.The community science testing aimed to characterize drinking water, as residents were concerned about water exposures after Hurricanes Florence and Michael.Per the 15 A NC Administrative Codes, the aim of the DHHS well sampling is to characterize the groundwater supplying the homes [34].While there will be subtle to no differences for geogenic contaminants (e.g.arsenic, manganese), there will be potentially large differences for premise plumbing contaminants as observed in our study.
Differences in lead measurements were also attributed to analytical techniques between testing efforts, as there was a MRL difference of 4.0 µg l −1 .When comparing flushed samples, the percent of samples with lead concentrations greater than 5 µg l −1 was consistent between testing efforts-92% of NCWELL database samples and 96% of community science testing samples.However, 19% of the community science testing samples contained 1-5 µg l −1 , which would not be captured at the higher MRL.Low levels of lead release are not uncommon as 'lead-free' materials and components used within the last liter of plumbing can contribute up to 3 or 5 µg lead to a 1-L sample per NSF/ANSI standards [33].Characterizing low lead concentrations is of growing importance as such lead in water levels can result in BLLs over 10 µg dl −1 [9,35].
When comparing lead measurements between testing efforts, these data collection differences resulted in largely incongruent messaging about lead in well water.In two thirds of the private wells sampled by both datasets, the NCWELL database flushed samples contained no measurable lead (i.e.no lead in water), while the community science first draw samples contained measurable lead (i.e.lead was present).More importantly, the community science testing samples contained more than 5 µg l −1 in 36% of wells.Such differences in lead concentrations reported may have altered well stewardship behaviors, from no action to avoidance of the drinking water [10,36,37].Providing educational materials is critical, as studies document that educational materials can increase well water stewardship behaviors that reduce exposures to contaminants [38,39].

Community-level impacts
Under-and overestimations of lead levels can misinform conclusions about risks of elevated BLLs associated with drinking water from private wells which may hinder investigations of lead exposure associated with well water [17,40].The difference in reported lead occurrence between the two testing efforts may impact conversations about BLL risks for infants and children.While the geometric mean BLLs predicted for both testing efforts using the 95th percentile lead values were above the 3.5 µg dl −1 reference threshold, there was a 1.0-2.8µg dl −1 difference in predictions between the NCWELL database and the community science testing.This is particularly concerning for infants consuming restituted formula, as water is the primary lead exposure for formula-fed infants [40,41].Even at the 50th percentile community testing lead value, which was below the 15 µg l −1 , 72%-87% of infants consuming reconstituted baby formula would be expected to have a BLL greater than 3.5 µg dl −1 .As water can be a primary contributor to BLLs for infants and children, accurate representation of the well community in sampling efforts is critical, both in terms of household and system characteristics.
Lead concentration in the community science testing varied by well water chemistry (e.g.pH), well construction characteristics, and source water (i.e.groundwater) quality, which has been observed in prior work [2,3,11,12,17].However, there was no difference in lead concentrations between bedrock and sand wells that participated in the community science testing, which contradicted prior work in Virginia associating lead concentrations with geological region [11].This was likely attributed to the higher percentage of shallow sand wells in the community science testing (21% in this study vs. 16% in Virginia).Our shallow sand wells had higher lead concentrations than our deep sand wells (4.9 vs 0.6 ppb) due to their more acidic groundwater supply (median pH of 5.6 vs 7.7).Inclusion of these shallow systems is important as they have unique construction and water quality characteristics [11].However, many well water testing efforts have higher rates of participation among users with deeper, drilled wells [2,3], and some testing efforts do not allow samples from non-permitted wells and/or wells not meeting current construction codes [30].
Statewide testing efforts typically engage well users who identify as White/Caucasian, affluent, and college educated [2,22,[42][43][44], and this lack of representation can result in skewed lead contamination rates.We did not observe a difference in first draw lead between White and BIPOC households, but only 10% of our participants were from BIPOC households compared to 26% in the NC well population.When simulating a representative well population, the average 90th percentile lead value was 1.0 µg l −1 higher than our community science testing's 90th percentile.However, when evaluating communitylevel impacts, we observed that 90th percentile lead level predictions differed by up to a 6 µg l −1 .While drinking water is often considered a secondary source of exposure, these low concentrations can have substantial health impacts on infants.When simulating infants consuming reconstituted formula, we observed that all predicted that geometric mean BLL would be above the CDC reference value and 68%-95% of infants would have BLLs above 3.5 µg dl −1 .There was a 1.0-1.3µg dl −1 difference when considering the demographics of the population, which is a substantial increase among a population already at risk of exposures.With the known health effects at low BLLs and the high absorption rate among infants [45,46], sampling efforts must strive to ensure that sample populations are representative of the target population.

Conclusions
In this study, we explored the impacts of sampling methodology and household representation on estimated lead exposures and subsequent health risks.Our analysis indicates the following: • The occurrence of lead in well water is associated with the corrosion of premise plumbing.Our results are in keeping with rates observed in neighboring states, which document concerns associated with the corrosion of brass and galvanized iron materials when exposed to untreated corrosive groundwater.While private wells do not have lead service lines like municipal systems, 'lead-free' plumbing manufactured before 2014 can release high lead concentrations when exposed to corrosive well water.• Sampling methods impact our understanding of well water quality and resulting public health outcomes.We compared the commonly used flushed sampling with first draw sampling technique and found flushing prior to collection chronically under reports lead contamination.This underreporting was further compounded by using a higher MRL during sample analysis.To assess lead in drinking water, first draw samples need be collected and analyzed with a MRL of no higher than 1 µg l −1 .• Lead in well water is a public health concern in NC.As an example, we highlight the impact of lead contamination on infant consuming restituted formula.Our results suggest that most infants consuming formula reconstituted with well water would be expected to have a BLL greater than the CDC's new blood lead reference value of 3.5 µg dl −1 .However, waterborne exposure can be mitigated using treatment and/or alternative water supply.• Sampling efforts must be representation of the target population both in terms of household demographics and well construction characteristics.We observed that the lack of inclusive of BIPOC well-using households and less common well types impacted rates of lead occurrence observed.Overall, the mischaracterization of lead in water can impact knowledge of lead exposure associated with well water and hinder lead exposure investigations.

Figure 1 .
Figure 1.Summary of datasets used in study analysis.The datasets used in our analysis (gray boxes) include (1) first draw and flushed samples from the community science testing and (2) flushed samples from the NCWELL database.Different combinations of the first draw and flushed data were used for Objectives 1-4 at the household-(light green boxes) and community-level (dark green boxes) analyses.

Figure 2 .
Figure 2. Lead concentrations measured during the community science testing in the first draw and flushed samples (n = 1143) and the NCWELL database in the flushed sample (n = 110 409).Lead concentrations below the minimum reporting limit (MRL) of 0.1 µg l −1 in the community science and 5 µg l −1 in NCWELL database are not shown.

Figure 3 .
Figure 3. (A) Average 90th lead concentrations calculated by % BIPOC households in bootstrapping analysis using the community science first draw samples.Error bars denote the 95th confidence intervals.(B) Predicted geometric mean blood lead levels for 0-1-year olds consumption with average and high consumption of formula using the simulated 90th percentile lead in water levels by % BIPOC households.

Table 1 .
Blood lead levels of children in various scenarios based on measured lead in water concentrations.