Community intervention trial for estimating risk of acute gastrointestinal illness from groundwater-supplied non-disinfected drinking water

More than 20 million people in the United States drink non-disinfected groundwater from public water systems (PWS), and an estimated 56.8 million people drink groundwater from PWS that is not adequately disinfected (i.e., viruses are not reduced 99.99%) (U.S. EPA 2006). Of the approximately 140,000 PWS supplied by groundwater in the US, estimates range from 60,000 to 100,000 systems that do not provide disinfected drinking water (U.S. EPA 2012; Messner et al. 2017).

Drinking non-disinfected groundwater is responsible for a significant fraction of the waterborne disease outbreaks in the US. Of the 801 deficiencies assigned by the US Centers for Disease Control and Prevention to the 780 drinking water outbreaks between 1971 and 2006, 243 deficiencies (30.3%) were related to drinking non-disinfected groundwater and 179 deficiencies (22.3%) involved groundwater with inadequate treatment or treatment that had been interrupted (Craun et al. 2010). Of the 29 drinking water outbreaks in 2013 and 2014 with a known water source, 14 (48%) were due to groundwater (Benedict et al. 2017). In Denmark, Finland, Norway, and Sweden during the years 1998–2012, there were 163 drinking water outbreaks where the water source was known and of these 124 (76%) were linked to groundwater, which is usually not disinfected in Nordic countries (Guzman-Herrador et al. 2015).

The epidemiological data are corroborated by studies that show the infectious etiological agents of acute gastrointestinal illness (AGI) are detected in groundwater. Of 41 studies on pathogen contamination of groundwater in the USA and Canada, 16.1% of 2,210 wells were positive for an enteric pathogen (Hynds et al. 2014). In 12 studies of public-supply groundwater systems in North America, Italy, Japan, and South Korea, 8% (n = 365) and 30% (n = 611) of wells were positive for culturable and PCR-measured human enteric viruses, respectively (Fout et al. 2017). The parasites Giardia and Cryptosporidium, and bacteria Salmonella, Campylobacter, pathogenic E. coli, and Shigella have been detected in contaminated wells and implicated in groundwater-borne outbreaks (Hynds et al. 2014; Wallender et al. 2014; Guzman-Herrador et al. 2015; Murphy et al. 2017; Stokdyk et al. 2019, 2020).

Pathogen contamination is not restricted to wells with construction or maintenance problems or those in vulnerable aquifers. Among 172 untreated groundwater outbreaks in the US between 1971 and 2008 that had information on outbreak causes only 39 (23%) could be attributed to problems in well design or maintenance, and only 45 (26%) happened in hydrogeological settings that would be considered vulnerable (Wallender et al. 2014). Even a newly constructed well meeting all regulatory criteria can become contaminated and result in a disease outbreak (Borchardt et al. 2011). Human enteric pathogens can be detected in a variety of aquifer types (Fout et al. 2017), even in deep wells cased into confined aquifers whose overlying low permeability geologic units were thought to prevent microbial contamination (Borchardt et al. 2007; Bradbury et al. 2013; Smith et al. 2013).

While there is ample evidence of enteric pathogen occurrence in groundwater supplies and of groundwater-borne disease outbreaks, few studies have investigated the role of groundwater in endemic and sporadic infectious diseases. Our objective in the present study was to conduct a community intervention trial to estimate the fraction of AGI attributable to community supplies of non-disinfected groundwater. The intervention was ultraviolet light (UV) disinfection installed on all wells supplying drinking water in the study communities. Without intervention, all residents living in the communities drank non-disinfected groundwater; with intervention, all residents drank groundwater that had been UV disinfected at the wellhead before entering the distribution system. The reduction in AGI incidence in the communities when UV disinfection was in place represents the fraction of AGI attributable to the groundwater source.

This objective was one of several intended for the Wisconsin Water and Health Trial for Enteric Risks (Wisconsin WAHTER Study). Previous publications from the WAHTER Study have reported on: (1) Tap water concentrations of enteric viruses as related to AGI incidence and attributable risk (Borchardt et al. 2012); (2) Operation and maintenance events in drinking water systems associated with virus contamination (Lambertini et al. 2011); and (3) Distribution system contamination by viruses entering water pipes downstream of supply wells and the resulting AGI attributable risk (Lambertini et al. 2012). The UV intervention study design allowed partitioning of the waterborne AGI risk into two components, distribution system and groundwater source. Here, we report the AGI attributable to groundwater source contamination as measured empirically by the reduction in AGI incidence when the groundwater source is UV disinfected. We place our findings within the context of the lessons learned from the WAHTER study and the AGI risks associated with drinking non-disinfected groundwater from community drinking water systems.

We requested study participation from 40 community PWS in Wisconsin, USA and enrolled the first 14 that agreed to participate and met our eligibility criteria: population served > 1,000, water supplied from ≤ four groundwater wells, and no drinking water disinfection. The 14 communities had populations between 1,363 and 8,300, were located throughout Wisconsin, and had supply wells drilled into sandstone or sand and gravel aquifers (Supplementary Figure S1). All had municipally owned drinking water utilities that met State and Federal safe drinking water requirements. Every community lacked a centralized water plant; water was pumped from supply wells directly into water mains and distributed to residents without disinfection. All supply wells were classified by State regulators as free of surface water influence and therefore not subject to treatment requirements.

UV reactors were sized to achieve a UV dose of at least 50 mJ/cm². UV reactors were alarmed to indicate a downward trend in dose and before it dipped below 50 mJ/cm², which then required lamp cleaning or a service call. Engineering plans for UV installations were reviewed and approved by the Wisconsin Department of Natural Resources.

The number of households and participants at each stage of recruitment and enrollment is summarized in Supplementary Figure S2. Participants were followed as a prospective cohort. Households for recruitment were identified from water utility billing addresses. Eligible households had to have drinking water provided by the municipal drinking water utility and at least one child 6 months to 12 years of age available for AGI surveillance. Children older than 12 years were not enrolled. One adult per household was also requested to participate. Participants could not have a chronic gastrointestinal illness, and children who attended school or daycare for more than 20 h/week outside of the drinking water utility service area were excluded. Enrollment at study initiation was 621 households, 1,079 children, and 580 adults. AGI symptoms were recorded weekly on a standardized checklist completed by an adult participant for every participant in the household. Symptom checklists were mailed weekly to the study team; checklists received ≥ 21 days from the end of a weekly reporting period were excluded from analysis (2,821 checklists excluded out of 64,265 submitted (4.4%)). An AGI episode was defined as three or more bowel movements of loose watery stools or one episode of vomiting in a 24-h period. Illness episodes were counted as distinct when separated by ≥ 6 symptom-free days. We estimated person-time, the time considered at-risk for AGI from drinking water exposure in a study community, from participants' daily reports on the checklists of the days they slept away from home. The day immediately following three or more consecutive days away-from-home was defined as not-at-risk for AGI.

AGI surveillance was conducted for four 12-week periods: (a) Period 1, April–June 2006, (b) Period 2, September–November 2006, (c) Period 3, March–May 2007, and (d) Period 4, September–November 2007. Between study years periods were matched by season to minimize seasonal differences in AGI rates. Summertime surveillance was avoided because focus groups conducted during study planning indicated dropout would peak then. Winter was chosen for UV installation and crossover because this season, when construction typically slows, offered the most affordable costs.

Participants completed written surveys at the beginning of the first and second years of surveillance for reporting information on health status, foreign travel, primary drinking water source, raw milk consumption, household characteristics, household water treatment, and daycare attendance.

Free and informed consent of the participants or their legal representatives was obtained and the study protocol was approved by Marshfield Clinic Institutional Review Board, Marshfield, WI, USA, Protocol # BOR20303, approved 2 April 2004.

Study wells (n = 34) were sampled monthly (i.e., three sample times per 12-week surveillance period) for human enteric viruses. Wells pumped for at least one-half hour before sampling by glass wool filtration (Lambertini et al. 2008). Samples were collected by trained technicians from a tap in the main pipe exiting the well, and if UV disinfection was installed, a second sample was collected simultaneously from a tap located immediately downstream from the UV reactor before water entered the distribution system. Filtration flow rate was 4 L/min. Mean sample volume was 849 L (n = 392, range 114–2,067 L) for well water and 864 L (n = 191, range 231–1,666 L) for post-UV disinfection samples. Glass wool filters were sealed, placed on ice, and processed within 48 h of sampling.

Sampling equipment negative controls were performed once per surveillance period; all were negative for the six study viruses. Every batch of constructed glass wool filters was tested for virus contamination; all batches were virus negative. Performance of the glass wool was evaluated every surveillance period by spiking 1 × 10⁴ gene copies/L poliovirus Sabin 3 into 10 liters dechlorinated laboratory tap water. Recovery was performed as described in Lambertini et al. (2008) and ranged from 70 to 96%. Using the same poliovirus spiking and recovery procedure, sampling positive controls were performed with two to four replicate 10-liter volumes of well water from each of the 14 study communities and recovery ranged between 23 and 99%.

Adenovirus, enterovirus, norovirus genogroups I and II (GI and GII), hepatitis A virus, and rotavirus were measured by qPCR. Virus analytical methods were identical to those described in Borchardt et al. (2012). Viruses in glass wool filters were eluted with beef extract and further concentrated by polyethylene glycol flocculation (Lambertini et al. 2008) to 2 mL final concentrated sample volumes (FCSVs). DNA and RNA were extracted from 280 μL FCSV using the QIAamp DNA blood mini kit and buffer AVL (Qiagen, Valencia, CA), eluted to a final volume of 50 μL. Extracts were analyzed immediately.

Reverse transcription (RT) reaction conditions (two-step with random hexamers) are described in the Supplementary Material.

qPCR with hydrolysis probes (TIB Molbiol, Berlin, Germany) was conducted on a LightCycler 480 instrument (Roche Diagnostics, Mannheim, Germany) using the LightCycler 480 Probes Master kit. The 20 μL final reaction volume consisted of 14 μL master mix and 6 μL extracted DNA (adenoviruses) or cDNA from the RT step (other viruses). Primers (Integrated DNA Technology, Coralville, IA) and TaqMan probes (TIB Molbiol, Berlin, Germany) and their concentrations are reported in Supplementary Table S1. Thermocycling began with 95 °C for 10 min followed by 45 cycles of 15 s at 94 °C and 1 min at 60 °C.

qPCR standards were intact virions of each virus group treated with Benzonase^® (Novagen, Madison, WI), except hepatitis A, which was purchased Armored RNA (Asuragen, Austin, TX). Preparation and quantification of standards is described in the Supplementary Material. Standard curve parameters are reported in Supplementary Table S2.

RT-qPCR inhibition was evaluated for every sample by spiking hepatitis G virus Armored RNA (HGV) (Asuragen) into the RT reaction mixture and performing RT-qPCR as described above (see Supplementary Material). Of the 392 well water samples and 191 post-UV disinfection samples, 38 and 18, respectively, were inhibited and required mitigation by dilution with nuclease-free water.

Negative controls were performed for every batch of the following procedural steps: secondary concentration, extraction, RT, and qPCR. Data associated with failed negative controls (i.e., positive result) were omitted, the contamination source identified and corrected, and the analysis batch repeated. Positive controls were performed for extraction and qPCR for each virus tested. Additional information on controls is in the Supplementary Material.

Duplicate qPCR analyses were performed on all samples. The average of positive replicates is reported; when both duplicates were negative, the virus concentration is zero. Virus concentration calculations are described in the Supplementary Material.

Based on the mean sample volume for well water samples, 849 L, the equivalent volume of sample analyzed by qPCR was 14.3 L for adenovirus and 2.5 L for the RNA viruses.

The completed checklist for Environmental Microbiology Minimum Information (EMMI) for qPCR (Borchardt et al. 2021) is in the Supplementary Material.

Samples qPCR-positive for adenovirus and enterovirus were further evaluated for virus infectivity by culture in two cell lines (Graham 293 and A549) and three cell lines (BGM, RD, and CaCo-2), respectively. Cultures were passaged three times. Cytopathic effect (CPE) was never observed. As an alternative, using the same qPCR methods as for the water samples, we measured the quantity of viral gene copies in the cell lysates from the second passage (4 weeks) and third passage (6 weeks). When the quantity in the lysates was >10 times more than the initial inoculum from the FCSV, indicating virus growth, the sample was designated positive for infectious virus by integrated cell culture-qPCR (ICC-qPCR).

Enteroviruses and adenoviruses in qPCR-positive samples were serotyped by nucleotide sequencing. For enteroviruses, amplicon from a separate PCR using primers OL68-1 and EVP4 (Ishiko et al. 2002) targeting a 656 base pair region encoding one-third of the 5′ UTR (untranslated region), the entire VP4 region, and one-third of VP2 was sequenced. For adenoviruses, the 263 base-pair product from the qPCR that targeted the hexon gene was sequenced. Additional method details (culture and serotyping) are described in the Supplementary Material.

The effect of the study intervention, UV disinfection, was quantified by estimating attributable risk and attributable risk percent (AR%) (Craun et al. 2006). The former is denoted by I_C − I_I and the latter by (I_C − I_I)/I_C = 1 − I_I/I_C, where I_C and I_I are the AGI incidence rates in the control and intervention periods, respectively. Incidence is defined as AGI episodes divided by the corresponding number of person-time units of observation. For analytic purposes, we define ‘exposed’ as the absence of the study intervention and ‘unexposed’ as the presence of the intervention.

Permutation tests developed by Gail et al. (1996) were conducted to compare control and intervention AGI incidence rates. To paraphrase a description of permutation tests from Murray & Blistein (2003): ‘The analyst constructs the distribution of all possible intervention effects under the null hypothesis of no intervention effect, conditional on the observed data. The observed intervention effect is treated as but one of the possible intervention effects, and the probability of getting a more extreme result [i.e., the p-value for the test] is the proportion of possible intervention effects that are greater in magnitude than the observed effect’. Both unadjusted permutation tests and those adjusted for potential confounding factors were conducted. In addition, paired and unpaired tests were performed. Paired tests entail within-community comparisons of the control and intervention periods (original study design). Unpaired tests compare means for control and intervention communities within a surveillance period (revised design) (Figure 1(b)).

For a paired unadjusted test, attributable risk was estimated as ⁠, where I_iC is the illness incidence in the ith community during the control period, I_iI is the illness incidence in the ith community during the intervention period, and N is the number of communities. The first step in the corresponding calculation of AR% was to compute ⁠, the mean natural log transformed ratio of intervention to control incidence rates. AR% was then estimated as 100*(⁠⁠) where exp() denotes exponentiation.

For an unpaired unadjusted test, attributable risk was estimated by ⁠, where I_iC is the illness incidence in the ith control community in the surveillance period, I_iI is the illness incidence in the ith intervention community in the same surveillance period, and N_C and N_I are the number of control and intervention communities in the surveillance period. The first step in the corresponding calculation of AR% was to compute ⁠, the difference in mean natural log transformed incidence rates for the intervention and control communities. AR% was then estimated as 100*(⁠⁠).

Adjusted permutation tests employed the same framework as unadjusted tests except that rather than raw AGI incidence rates, the community-level data points were derived from Poisson regression models where the dependent variable was the number of AGI episodes, the model offset was the natural log of the corresponding number of person-time units of observation and the independent variables were factors that could potentially confound the association between UV intervention status and AGI incidence. The source datasets for the regression models were at the level of a person such that person-, household-, and community-level confounding factors could be adjusted for. Specifically, models included age, gender, daycare attendance, household size, surveillance period (paired permutation tests only), and virus concentrations in the communities' wells. They did not include a term for UV intervention status. For attributable risk estimation with confounder adjustment, the data point for a community took the form (O−E)/PT where O is the sum of observed AGI episodes, E is the sum of expected AGI episodes based on the regression model predictions, and PT is total person-time (Hayes & Moulton 2017). To provide context, the data points in the unadjusted tests described above took the form O/PT. The corresponding information for AR% estimation with confounder adjustment is ln(O/E) in contrast to ln(O/PT) for unadjusted tests. The community-level data points for paired permutation tests with confounder adjustment were derived from two data points (one for the control period and one for the intervention period) in the formats described above.

Confidence intervals for attributable risk and AR% estimates for a paired test are described by Manly (2007) as follows: ‘The limits of the 95% confidence interval (CI) are the values that, when subtracted from the paired differences, result in the sample sums falling at the lower 2.5% and upper 2.5% points in the distribution obtained by randomization’ (in our case, we evaluated the full permutation distribution rather than employing randomization). A similar framework is used for identification of confidence limits for unpaired tests. All statistical analyses were conducted using SAS (SAS Institute Inc, Cary, NC).

All 34 wells (100%) were virus-positive at least once. Of 392 samples, 139 (35%) were positive for one or more virus types (Table 1). Detection frequency was highest for adenovirus and enterovirus, while concentrations of viruses in well water were highest for norovirus GI. Norovirus GII and rotavirus were never detected. About one-quarter of the adenovirus and enterovirus qPCR-positive samples were positive for culturable virus by ICC-qPCR (Table 1).

The UV intervention was effective in reducing virus detection frequencies and concentrations in the communities' well water before it entered the drinking water distribution systems (Table 1).

The characteristics of the study households and participants are described in Table 2. Of the 621 households, 92% were a single-family home and 88% did not use a tap water filtration device. Of the 1,659 participants, 93% drank municipal drinking water and the race of 93% was white. Child participants were 51% female and adult participants were 82% female. Drop-out rate by the end of surveillance was 29% of households and 29% of participants.

Total AGI incidence rates in the study communities are aggregated by study period, participant age group, and UV treatment status in Supplementary Table S3. One community was excluded from analysis because its groundwater supplies were unexpectedly chlorinated during surveillance Periods 1, 2, and 4, which would have masked the effect of UV disinfection.

Among all study participants in the 13 communities included in this analysis, we observed 1,695 AGI episodes during 372,966 person-days of follow-up over 48 weeks of surveillance.

The effect of UV disinfection on reducing groundwater-borne AGI was first evaluated by the difference in AGI rates within the same community during the study years with and without UV disinfection installed at the wellheads (Figure 1(a)). This approach is consistent with our study design (i.e., community intervention trial with crossover), in which the crossover accounts for unmeasured factors that could be contributing to intrinsic differences in AGI rates among the communities. From this analysis, AGI attributable risk from groundwater was not distinguishable from zero (Table 3).

Because the quantities and types of viruses in the communities' wells changed from period to period (Figure 2), the potential effect of the UV intervention on reducing virus exposure likely varied by period. Therefore, we revised the analysis to compare AGI rates between control and UV disinfected communities within the same surveillance period (Figure 1(b)).

Using this approach, during Period 1, AGI incidence was higher in control communities than in UV intervention communities for all age groups, the only exception being the unadjusted estimate for children ≤ 12 years old (Table 4). When the analysis was restricted to children less than 5 years old, attributable risk estimates were the highest observed, 0.73 (95% CI, 0.03, 1.43) and 0.65 (95% CI, −0.07, 1.34) episodes/person-year for unadjusted and adjusted analyses, respectively (Table 4). The corresponding estimates for proportion of AGI attributable to pathogen-contaminated groundwater were 21% (95% CI, 1%, 36%) and 19% (95% CI, −4%, 36%) (Table 4). Adults, too, had notable attributable risk estimates, 0.40 (95% CI, −0.22, 1.03) and 0.32 (95% CI, −0.22, 0.85) episodes/person-year, unadjusted and adjusted analyses, respectively, corresponding to AR% of 18% (95% CI, −9%, 38%) and 15% (95% CI, −9%, 33%) (Table 4). Norovirus GI was the predominant pathogen in the groundwater supply wells during surveillance Period 1 (Figure 2(a)).

In Period 2, norovirus GI was still present in well water albeit at lower frequency than in Period 1 (Figure 2(a)). Reduction in AGI by UV disinfection appeared only for children < 5 years old, however the estimates had wide 95% confidence intervals that overlapped with zero (Table 5). Attributable risk was 0.58 episodes/person-year (95% CI, −0.54, 1.65, P-value = 0.284, unadjusted estimate) and AR% was 17% (95% CI, −51%, 51%). The adjusted test yielded a lower attributable risk estimate, 0.20 episodes/person-year (95% CI, −0.65, 1.07, P-value = 0.615). The corresponding AR% was 3% (95% CI, −58%, 37%).

In Period 3, UV disinfection appeared to reduce AGI incidence only in adults, but again the estimates had wide confidence intervals (Table 6). Attributable risk, unadjusted, was 0.20 episodes/person-year (95% CI, −0.42, 0.79, P-value = 0.472) with a corresponding AR% of 6% (95% CI −52%, 39%). Attributable risk, adjusted, was 0.28 episodes/person-year (95% CI, −0.27, 0.78, P-value = 0.270) with a corresponding AR% of 10% (95% CI −35%, 38%). The other age groups had higher AGI risk in the intervention communities, particularly children < 5 years (Table 6).

In Period 4, UV disinfection did not reduce AGI risk in any age group (i.e., all AGI risk estimates were higher in UV disinfected communities (Table 7)).

The WAHTER study was designed to estimate the risk of AGI from non-disinfected groundwater-supplied municipal drinking water and to partition the risk into two components: (1) risk from pathogen contamination of the drinking water distribution system and (2) risk from pathogen contamination of the groundwater supply. By installing UV disinfection at the wellheads, we inactivated enteric pathogens as they were pumped from the wells. Unlike chemical disinfection (e.g., chlorine), UV treatment does not provide residual disinfection to inactivate any pathogens that enter the distribution system downstream from the source water entry point. Direct contamination of distribution systems can happen via maintenance activities (Lambertini et al. 2011), plumbing mistakes, and transient negative pressure events in which pathogen-contaminated sediments outside of the pipes are aspirated inward through leak openings (LeChevallier et al. 2003). Because none of the study communities used chemical residual disinfection, the only points in their drinking water systems for pathogen inactivation were the UV reactors located at the wellheads. Therefore, any reduction in AGI incidence measured during UV intervention represents the fraction of AGI contributed by pathogen-contaminated groundwater.

Other routes of enteric pathogen transmission among residents of the study communities were possible (e.g., foodborne, zoonotic, person-to-person). The amount of transmission from these routes was irrelevant to our analysis because our outcome measure was the reduction in AGI when UV disinfection eliminated the groundwater transmission route. However, any differences among the study communities in their AGI incidence from non-waterborne routes could affect the estimate of groundwater transmission. For example, communities could differ in daycare attendance, affecting person-to-person transmission rates. If that were the case, AGI differences between intervention and control communities attributed to water may in fact be due to other routes. To avoid that problem, the UV reactors were switched halfway through AGI surveillance so that disinfecting communities became non-disinfecting and vice versa (crossover). The crossover allowed the statistical analysis to treat each community as its own control, comparing its AGI incidence with and without UV disinfection.

Using this study design, AGI transmission from the communities' groundwater supplies was not detected. The absence of a UV intervention effect was observed for all age groups and regardless of model adjustments.

There are several possible reasons a UV intervention effect was not observed. First, estimates of attributable risk and AR% could be sensitive to differences in person-observation-time among study communities and to participant drop-out. In the primary analyses, the data points in the paired permutation tests (original crossover design) were weighted equally regardless of person-time. We conducted additional permutation tests where the data points were weighted by 1/[(1/PT_iI) + (1/PT_iC)], where PT_iI and PT_iC are total person-time for community i in the intervention and control periods, respectively. We also conducted permutation tests where the data were restricted to participants followed during all four surveillance periods who had ≤5 missing weekly symptom checklists. Neither modification altered the original outcome.

Second, daily movement by study participants to areas outside of the drinking water service area could have obviated the intervention or control conditions. This may have been true for adult participants traveling for work, but children were not eligible to participate if they attended a school or daycare outside of the drinking water service area for more than 20 h/week. Moreover, 11 of 13 study communities were isolated, bordered by rural regions served by private wells.

Third, UV disinfection could have been inadequate to create an intervention effect. However, measurements immediately downstream from the UV reactors indicated substantial reduction in virus concentrations; some adenoviruses remained likely because they are the enteric virus most resistant to UV irradiation (Yates et al. 2006). To ensure adequate disinfection, UV reactors were set to alarm if the dose dropped below 50 mJ/cm²; fouled UV lamp sleeves were cleaned by utility staff, other issues were resolved by the UV manufacturer's local service company.

Fourth, statistical power could be inadequate. Based on sample size calculations for cluster-randomized trials (Hayes & Bennett 1999), our goal was to follow 62 households in each of 14 communities, which would yield 80% power (type 1 error = 0.05) to detect a 20% reduction in AGI incidence. Participant drop-out over the course of the study (29% of households) and exclusion of one community (because it applied chlorination) reduced statistical power.

Finally, another explanation for the absence of a UV intervention effect is that exposure to groundwater-borne viruses changed between control and intervention periods. This is evidenced by variation in virus occurrence by the surveillance period (Figure 2). For example, intervention communities had UV inactivation of norovirus in Periods 1 and 2, but norovirus had nearly disappeared in well water by the time UV disinfection was removed (i.e., crossover into control periods). With much lower norovirus exposure from well water in Periods 3 and 4, AGI was not elevated during the control period of the former intervention communities. Likewise, the composition of enterovirus and adenovirus serotypes and proportion of infectious adenoviruses changed between the intervention and control periods. The WAHTER study was designed in 2005, and survival studies at that time indicated that viruses may persist in groundwater (John & Rose 2005), suggesting constant exposure from well water. Later studies of repeated sampling of the same well over time showed that virus occurrence is intermittent (Bradbury et al. 2013; Stokdyk et al. 2020; Sorensen et al. 2021). The WAHTER study design was optimal for statistical control, but it did not match the biological reality that contamination of groundwater by enteric viruses and the resulting exposures experienced by drinking water consumers can change rapidly.

Given the variation in virus occurrence among study periods our revised analysis compared AGI rates between intervention and control communities within the same surveillance period. With this analytical approach, UV disinfection reduced AGI incidence for all age groups during surveillance Period 1, especially for children less than 5 years old. AGI reduction by UV disinfection was also observed for children < 5 in Period 2 and for adults in Period 3. However, AGI attributable risk estimates were modest or highly variable, the majority had confidence intervals that overlapped zero, and many were negative. Nonetheless, for those age groups and surveillance periods when AGI reduction by UV disinfection was observed, the AGI attributable risk estimates appear credible because they match the findings of Borchardt et al. (2012), who reported strong statistical associations between virus concentrations in tap water and AGI incidence for these same study communities.

For instance, during Period 1 when norovirus GI was prevalent in well water (Figure 2), UV disinfection reduced the virus to undetectable levels (Table 1), and correspondingly, a reduction in AGI in the UV intervention communities was observed. Corroborating this experimental result, Borchardt et al. (2012) showed that three measures of norovirus GI in the study communities' tap water (mean concentration, maximum concentration, occurrence frequency) were associated with AGI incidence in all age groups. Moreover, when their analysis was restricted to Period 1, norovirus GI concentration in tap water was most strongly associated with AGI incidence in children less than 5 years of age, similar to the observation reported here.

Using quantitative microbial risk assessment (QMRA) for children less than five, Borchardt et al. (2012) showed that the median proportion of AGI in Period 1 from norovirus GI contaminated tap water was 44%. The experimental approach of the present study estimated that UV disinfection in Period 1 reduced AGI incidence for this age group by 19%; the difference between estimates may be due, in part, to the difference in AGI contributions from well water versus tap water, as the latter includes pathogen contamination via distribution systems.

Corroborating findings also support the AGI reduction by UV disinfection observed only for adults in Period 3. Borchardt et al. (2012) showed measures of enteroviruses in the study communities' tap water in Periods 3 and 4 were significantly associated with AGI, again, only in adults. Echovirus 11 was predominant in Period 3, the only time it occurred (Figure 2(d)), and this enterovirus serotype has been linked to adult diarrhea (Klein et al. 1960). UV disinfection greatly reduced enterovirus detection frequency and concentration in well water (Table 1), thereby reducing study participants' exposure to waterborne enterovirus. When enterovirus was present in tap water the adult AGI incidence rate ratio (IRR) was elevated at least 30% (Borchardt et al. 2012), suggesting that enteroviruses were responsible for 30% of AGI in adults. This estimate was derived by statistical association and is similar to the AGI reduction measured experimentally in the present study: UV disinfection of well water in Period 3 was estimated to reduce adult AGI by 10%, although the estimate had large variance.

A UV intervention effect was not detected in Period 4 when adenoviruses were predominant (Figure 2(a)). This absence of an effect is also corroborated by Borchardt et al. (2012), who did not find any associations between measures of adenoviruses in tap water and AGI.

Children <5 appeared to have higher AGI rates in intervention than in control communities during Periods 3 and 4 (Tables 6 and 7). This finding is possible if AGI transmission in this age group during these time periods was unrelated to drinking water, rendering UV disinfection irrelevant. By statistical associations, Borchardt et al. (2012) found no evidence for waterborne AGI for this age group during Periods 3 and 4.

In addition to the high variance of the AGI attributable risk estimates, the study findings should be interpreted with several other limitations in mind. While the revised analysis mitigates the problem of temporal variability in exposure to groundwater viruses, unlike the crossover design, it cannot account for intrinsic differences in AGI rates among communities caused by unmeasured factors. Participant blinding was not possible because the water utilities, as public entities, were required to discuss their study participation in open meetings. Study participants were not informed of the timing for control or intervention conditions in their community nor were UV installation publicized, but some study participants may have known, resulting in reporting bias. Lastly, groundwater supplies were evaluated only for viruses, but bacterial and protozoan pathogens could also have been present.

The goal of the WAHTER study can be viewed conceptually as risk_tap = risk_well + risk_{distribution system}. Diverse approaches were necessary to obtain these three risk estimates: (1) Statistical model building of virus occurrence in tap water related to AGI incidence followed by QMRA (risk_tap, Borchardt et al. 2012); (2) Virus measurements in distribution systems downstream of UV disinfection input into QMRA (risk_{distribution system}, Lambertini et al. 2012); and (3) Experimental intervention, comparing AGI incidence in communities with and without UV disinfection (risk_well, present study). Considered in toto, the three approaches give a picture of groundwater-borne disease transmission in the communities during the study period.

Risk_tap and risk_well were estimated over different time periods (the entire study versus by the surveillance period, respectively). Nevertheless, the estimates were similar, each responsible for approximately 6–20% of the AGI in the communities. Contaminated distribution systems were responsible for a smaller fraction of AGI, 0.1–4.9% (Lambertini et al. 2012). Importantly, AGI risk from drinking non-disinfected groundwater was highly variable, approaching zero in Period 4, and elevated in Period 1 when noroviruses contaminated well water.

The statistical associations reported in Borchardt et al. (2012) cannot identify whether viruses in tap water caused AGI or AGI in the communities caused virus contamination of groundwater. Through virus fecal shedding and leaking sanitary sewers, elevated AGI in the communities could lead to elevated viruses in the underlying aquifers and consequently elevated viruses at the tap. That we observed reduced AGI by UV disinfection for some surveillance periods and age groups suggests that the former causal pathway is likely correct: viruses in the communities' tap water resulted in drinking waterborne AGI.

However, the AGI leading to groundwater contamination pathway must have also operated, at least initially, as there had to be an original virus source. Accordingly, a reasonable conceptual model for enteric virus transmission in these communities consists of virus introduction by an infected visiting or traveling host, transmission in the community (via person-to-person, food, and fomite fecal-oral routes), shedding of the virus by diarrhea and vomiting into the sewers, and virions from leaking sewers reaching the groundwater supply. This, subsequently, leads to another transmission route, contaminated drinking water, which amplifies the number of infected individuals in the community.

Estimates of attributable risk percent in the present study tended to be higher than previous estimates. Table 8 summarizes 12 studies that estimated AGI risk and disease burden from groundwater-supplied drinking water, disinfected and non-disinfected, from PWS in Canada and the USA. Study designs were diverse and included disinfection intervention, statistical modeling, risk assessment, cross-sectional, and expert elicitation. All studies focused on community water systems; four studies also included non-community systems. The study most similar to ours in terms of water systems (small community systems, non-disinfected groundwater) used pathogen measurements from the literature and by risk assessment estimated AR% was 4.5% (Murphy et al. 2016). The study most similar in terms of design (intervention) added UV disinfection and filtration to households in a chlorinating community supplied by groundwater, possibly under the influence of surface water, and estimated AR% was 12%, within range of our estimates (Colford et al. 2009). In terms of similar geographic location (Upper Midwest of the US), Burch et al. (2022) estimated by QMRA that the well water for community water systems in Minnesota had it not been disinfected, would have been responsible for 0.23 infections/person/year. This is equivalent to 0.115 illnesses/person/year (assuming 50% constant morbidity ratio, Haas et al. 2014), which is the same order of magnitude as our attributable risk estimates.

AGI disease burden attributable to groundwater-supplied PWS in the US has been estimated to be on the order of millions of illnesses per year (Table 8). The lowest estimate, 185,000 illnesses/year, was derived by considering only two pathogens in groundwater, rotavirus and enterovirus (U.S. EPA 2006). Our study investigated a subset of groundwater-supplied PWS, non-disinfecting community water systems. The US population served by such systems is 9.3 million (U.S. EPA 2006). Extrapolating the attributable risk for all ages during surveillance Period 1 (0.16 episodes/person-year) the disease burden among this population is estimated to be 1.49 million AGI episodes/year. Alternatively, using the attributable risks estimated for children ages <5 years in Period 1 (0.65 episodes/person-year) and adults in Period 3 (0.28 episodes/person-year), and accounting for the age distribution in the US population (6% children <5 years, 77.7% older than 18 years (U.S. Census Bureau 2021)), the AGI burden would be 2.39 million episodes/year. These disease burden estimates ignore the high variance we encountered in measuring groundwater-related AGI attributable risk. Moreover, they are likely upper bounds as illness is contingent on AGI-causing viruses being present in community supply wells, which as we showed can be highly variable over time.

Analyzing the AGI incidence data following the original study design (i.e., comparing AGI incidence within the same community with and without UV disinfection), groundwater-borne disease transmission was not evident. This result likely stemmed from changes in virus occurrence in the communities' wells between intervention and control periods. Exposure to viruses in well water was not constant. Insofar as exposure was low during control periods (i.e., no UV disinfection), AGI incidence would not likely be different than during intervention periods (with UV disinfection). In contrast, comparing AGI incidence between intervention and control communities during the same 12-week surveillance period minimized changes in exposure to groundwater-borne viruses. With this revised analysis, AGI risk attributable to groundwater was observed in Period 1 for all age groups, in Period 2 for children less than age 5, and in Period 3 for adults. However, nearly all attributable risk estimates had wide confidence intervals. Corroborating evidence from Borchardt et al. (2012), who found strong statistical associations between viruses in tap water and AGI incidence in the same study communities for the same surveillance periods and age groups as reported here, suggests the attributable risk estimates are credible. Enteric virus contamination of drinking water wells is intermittent. During times when specific AGI-etiologic viruses are present in wells, drinking non-disinfected groundwater can be responsible for an observable fraction of endemic and sporadic AGI.

We thank Greg Harrington and Andrew Jacque for the engineering designs for UV disinfection installed on 34 community wells. Installation of the UV disinfection units was adroitly managed by Kevin Fischer and Aaron Staab at Staab Construction, Marshfield, WI. Epidemiologic and laboratory assistance was provided by Carla Rottscheit, Sandy Strey, Debra Kempf, Vicki Allison, Phil Bertz, and Matt Volenec. We are grateful for the support of EPA project officer, Angela Page. This study is part of the Wisconsin Water and Health Trial for Enteric Risks (WAHTER Study), funded by the U.S. EPA STAR Grant R831630.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.