The purpose of this report is to document the collection and analysis of water-quality parameters and E. coli concentrations in NERI and to develop a near-real-time surrogate model to estimate E. coli concentrations and provide timely information to National Park staff and visitors about the risks associated with using the New River recreationally. An assessment of the cross-sectional and longitudinal variability of water-quality parameters and E. coli concentrations will help determine how well the near-real-time surrogate model estimates E. coli concentrations throughout the New River Gorge.

The New River originates in North Carolina and flows northward through Virginia and West Virginia to its confluence with the Gauley River where the two rivers form the Kanawha River, which continues northwest to the Ohio River (fig. 1). The New River Gorge in south-central West Virginia was formed by the New River downcutting though the Appalachian Plateau; the gorge extends for about 60 miles between the City of Hinton and Town of Gauley Bridge, West Virginia. Much of the flow of the New River entering the New River Gorge is regulated by Bluestone Dam, a 165-foot-tall dam operated by the U.S. Army Corps of Engineers just upstream from the confluence of the New River with the Greenbrier River (U.S. Army Corps of Engineers, 2024).

The New River Gorge watershed is mostly forested (Tetra Tech, Inc., 2008) but has several developed areas, including the Cities of Beckley (with a population of 17,270), Mount Hope (1,333), and Oak Hill (8,167) and the Town of Fayetteville (2,898) in the western half of the watershed (U.S. Census Bureau, 2025). Coal mining and the transport of coal by railroad were the dominant economic forces in the New River Gorge from the late 1800s to the 1960s. While coal mining has largely ceased in the gorge, abandoned mines and the many small communities that formerly supported the industry attest to the region’s mining past (Mahan and Young, 2018). Other economic activities in the region include timber harvesting, light manufacturing, agriculture, and animal husbandry, with livestock noted as a minor source of fecal-coliform bacteria (Mathes and others, 2007; Tetra Tech, Inc., 2008).

Tributaries flowing into the New River within the New River Gorge include smaller streams. Piney Creek, the largest tributary within the gorge, has a total drainage area of 136 square miles (U.S. Geological Survey, 2025b). Many of these tributaries are listed as “impaired” by the West Virginia Department of Environmental Protection owing to fecal-coliform bacteria, iron, aluminum, and biological community impairment, among other causes (Tetra Tech, Inc., 2008). Fecal bacteria impairments are largely attributed to inadequate domestic wastewater disposal and treatment because many smaller communities in the area lack sufficient sewerage (Paybins and others, 2000; Mathes and others, 2007; Tetra Tech, Inc., 2008). As a result of inadequate wastewater disposal infrastructure, combined sewer outfalls, failing septic systems, and “straight pipes” that discharge wastewater directly into nearby surface waters are substantial sources of fecal-coliform bacteria (Tetra Tech, Inc., 2008). A part of the New River Gorge watershed was designated as a national river to be managed by the National Park Service by the National Parks and Recreation Act of 1978 (16 U.S.C. 460m-15; Good and Stasick, 2008). In 2020, this New River Gorge National River was redesignated as NERI (National Park Service, 2021). The formal NERI boundary encompasses roughly 72,000 acres of the New River Gorge from the City of Hinton in the south to Hawks Nest in the north. NERI generally includes the New River, the New River Gorge, the immediately surrounding plateau, and the lower reaches of the larger tributaries of the New River. Within the designated NERI boundary, about 74 percent of the land is owned by the Federal Government, and the remaining 26 percent is owned by the State of West Virginia and private landowners (West Virginia GIS Technical Center, 2017).

A multifaceted approach was used to discern patterns in water quality and the occurrence and distribution of E. coli in the New River Gorge. An overview of the study design is provided herein to provide additional context for the reader, with additional detail in the “Methods of Data Collection and Analysis” section of the report.

The monitoring location at the New River at Thurmond, West Virginia (USGS 03185400; hereafter, Thurmond; U.S. Geological Survey, 2025c) was fundamental to this study. Centrally located in the New River Gorge, Thurmond is a streamgage paired with a continuous water-quality monitor, which made the location ideal to gather data for an E. coli surrogate model. Thurmond began operation as a streamgage in February 1981. Continuous measurements of water temperature, pH, dissolved oxygen, specific conductance, and turbidity started at this monitoring location in June 2019. Discrete measurement of water-quality parameters and the sampling for and quantification of E. coli concentrations near Thurmond were done to develop a relation between E. coli concentrations and the continuously measured water-quality parameters.

To supplement the continuous and discrete sampling at Thurmond, two additional monitoring locations were established at bridges near the upper and lower ends of the gorge: the New River at Highway 41 at Prince, West Virginia (USGS 03184905; hereafter, Prince) and the New River at Fayette, West Virginia (USGS 03186000; hereafter, Fayette). These locations encompass the part of the New River where most recreational use happens. Water-quality parameters taken from discrete measurements and water samples taken from Prince and Fayette provided a basis for determining how E. coli concentrations varied throughout the gorge and for being able to assess how well a model developed from monitoring data at Thurmond might represent conditions at other locations within the gorge.

To assess the variability of E. coli concentrations across the width of the New River, cross-sectional water samples, consisting of five single vertical water samples across the width of the New River, were collected at Prince, Thurmond, and Fayette. These cross-sectional samples can also be used to understand any bias or discrepancies in composite sampling techniques or when comparing single (non-composited) samples.

The final component of the study included periodic longitudinal transect sampling of the New River Gorge study area, which included sampling below the mouth of large tributaries, to provide further understanding of the variability in water-quality parameters and E. coli concentrations along this reach of the river.

Collectively, these study-design elements supported the development of a near-real-time surrogate model of estimated E. coli concentrations at Thurmond and an assessment of how well the surrogate model represents E. coli conditions throughout the New River Gorge.

During this study, gage height at Thurmond was measured at 15-minute intervals using a submersible pressure transducer. Streamflow at this location was derived from gage-height data based on a stage-discharge rating model using methods described by Sauer and Turnipseed (2010) and Turnipseed and Sauer (2010).

Turbidity, specific conductance, pH, water temperature, and dissolved oxygen (hereafter collectively called water-quality parameters) have been measured at the Thurmond monitoring location beginning in June 2019. During this study, these water-quality parameters were measured every 15 minutes using a YSI EXO3 sonde (Yellow Springs Instruments, Inc.; Yellow Springs, Ohio), and sensors were operated in accordance with guidelines described in Wagner and others (2006). All continuous-monitoring streamflow and water-quality parameters data were reviewed, approved, and stored in the USGS Water Data for the Nation database (U.S. Geological Survey, 2025c).

Field methods used to collect discrete samples measuring water-quality parameters and water samples, which were subsequently analyzed for E. coli, generally followed standard USGS protocols and methods (Wilde, 2002, 2004; Anderson, 2005; Wagner and others, 2006; U.S. Geological Survey, 2006, 2023; Myers and others, 2014). However, some slight modifications to standard practices were made in specific instances to adapt to the challenges of sampling in the New River Gorge and are detailed in the following sections.

The concentration of E. coli in water samples was determined using the Colilert Quanti-Tray/2000 defined-substrate method (IDEXX Laboratories, Inc., Westbrook, Maine) using a most probable number (MPN) estimate, which is a U.S. Environmental Protection Agency (EPA) approved standard method (Lipps and others, 2018). Aliquots of 100 milliliters of each water sample were decanted and transferred to a sterile mixing vessel to be analyzed using the Colilert Quanti-Tray/2000 system. The contents of a Colilert reagent packet were added to the water sample, which was gently agitated until the reagent completely dissolved. The water sample-reagent solution was poured into a Colilert Quanti-Tray/2000 and incubated for 22–24 hours at 35 degrees Celsius prior to analysis.

Determining E. coli by the Colilert method is a two-step process that relies on taxon-specific enzymes to metabolize provided substrates, ortho-Nitrophenyl β-D-galactopyranoside (ONPG) and 4-methylumbelliferyl β-D-glucuronide (MUG), to produce readily observable changes. First, coliform bacteria use β-D-galactosidase enzyme to metabolize ONPG, which results in a yellow color. Escherichia coli use β-glucuronidase enzyme to metabolize MUG, yielding 4-methylumbelliferone, which fluoresces under ultraviolet light (Lipps and others, 2018). Therefore, Colilert Quanti-Tray/2000 wells that appear yellow in visible light are positive for total coliform bacteria, and those that are yellow and fluoresce under ultraviolet light are positive for E. coli (fig. 4).

The number of small and large wells positive for total coliform bacteria and E. coli were compared with the Colilert Quanti-Tray/2000 MPN table to provide the estimated concentration of E. coli in MPN per 100 milliliters of sample water. The specific methods of this microbiological analysis (without dilution) result in a minimum concentration of less than 1 MPN per 100 milliliters (MPN/100 mL) and a maximum concentration of greater than 2,400 MPN/100 mL. For numerical and statistical analysis, any samples with a censored concentration of less than 1 MPN/100 mL were adjusted to 1 MPN/100 mL.

In this report, E. coli concentration results are compared with recommended recreational water-quality criteria from the EPA. Due to the sampling frequency of this project, and in consultation with the National Park Service and relevant references (National Park Service, 2019), the EPA E. coli 90th-percentile statistical threshold value (STV) was selected as the relevant comparison value. For an estimated illness rate of 32 per 1,000 primary contact recreators, the STV for E. coli concentration is 320 MPN/100 mL (U.S. Environmental Protection Agency, 2015); this value will be referred to as the recreational-contact standard hereafter. In this report, bacteria are reported by the whole number for values less than 10 and with 2 significant figures for values greater than 10.

All field equipment that contacted the water sample, including the sample bottle, sampler nozzle cap, nozzle, and churn, was sterilized before use. The sample bottle, nozzle cap, and nozzle were sterilized in an autoclave at 121 degrees Celsius at 15 pounds per square inch for 15 minutes. The churn splitter was too large to fit in the available autoclave and was sterilized by chemical means using a sodium hypochlorite solution (Myers and others, 2014). The churn splitter was filled with a 0.005 percent sodium hypochlorite solution, and pH was adjusted to between 6 and 7 for maximum bactericidal potency, for a minimum of 30 minutes, after which the churn was drained. After the sodium hypochlorite soak, the churn splitter was rinsed with a dilute sodium thiosulfate solution to neutralize any remaining chlorine. After 5 minutes of contact with the sodium thiosulfate solution, the churn splitter was rinsed with sterile deionized water, bagged, and stored until use in the field.

Cleaning and handling methods were confirmed with a series of field blanks during the course of the study. Blank water was created by autoclaving a phosphate-buffered saline solution. During the course of the study, nine field blanks were processed and analyzed in the same manner as field samples. All blanks returned an E. coli concentration of less than 1 MPN/100 mL, the minimum E. coli MPN of the concentration determination method used.

Replicate samples were also used to verify field methods and help quantify variability in the microbiological analysis. Eight replicate samples were obtained during the study: five split replicates were obtained from the churn splitter during the composite depth- and width-integrated sampling, and three sequential replicates were obtained with the single vertical samples during the longitudinal transect. All regular and replicate sample pairs had E. coli concentrations that had overlapping 95-percent confidence intervals of the most probable number and were thus considered not significantly different.

As a final quality-control note, the mid-point sample during the January 10, 2023, cross-sectional sampling event at Prince was lost during processing. Escherichia coli concentration results were unavailable for this water sample.

In addition to the MPN, IDEXX provides upper and lower 95-percent confidence limits for E. coli concentrations based upon the number of positive wells in the sampling method (North Carolina Department of Environmental Quality, 2025). Because upper and lower 95-percent confidence limits for the E. coli MPN are provided, this report uses non-overlapping intervals between samples to indicate statistical significance instead of computing separate statistical tests. Comparing the 95-percent confidence intervals to determine statistical significance will result in an alpha value of approximately 0.0006 (Knoll and others, 2011). This value is more conservative than the alpha value of 0.05, which is more commonly used and will therefore create a higher threshold for statistically significant differences (Goldstein and Healy, 1995).

The linear regression model and associated statistical tests and metrics within this report were computed using the statistical programming software R (ver. 4.4.0; R Core Team, 2023) by modifying the suggested code from Helsel and others (2020). R packages, including tidyverse (Wickham, 2023), dataRetrieval (DeCicco and others, 2025), ggplot2 (Wickham and others, 2025), leaps (Lumley, 2024), car (Fox and others, 2024), olsrr (Hebbali, 2024), ppcc (Pohlert, 2020), and asbio (Aho, 2025), supplemented the capabilities of base R. More specifically, the linear regression used for the surrogate model was computed with the “lm()” function, and confidence intervals were computed using the “predict()” function.

Collecting depth- and width-integrated composite water samples and water-quality parameters provides a detailed spatial and temporal record of E. coli concentrations and water quality in the New River Gorge. A total of 34, 36, and 33 composite water samples were collected at Prince, Thurmond, and Fayette, respectively. Water-quality parameters were recorded with each routine sample. Collectively, these composite water samples and water-quality parameter measurements spanned a range of hydrologic and weather conditions over a 40-km reach of the New River. Kruskal-Wallis and pairwise Wilcoxon rank sum tests (with a Bonferroni correction) were used to compare results among monitoring locations and indicated differences among monitoring locations were not significant (p>0.05) for E. coli concentrations, turbidity, pH, or specific conductance (fig. 5). This finding may be the result of a lack of independence owing to the connectivity of sites within the same river. The consistency of E. coli concentrations and water-quality parameters may also indicate the typical chemostatic behavior of larger rivers. Although the contributions of tributaries in the New River Gorge can affect water quality, these tributary streamflows typically comprise a small fraction of total streamflow in the gorge, and their effects are often diluted by the greater volume of streamflow in the New River.

Concentrations of E. coli did not vary significantly (p>0.05) among Prince, Thurmond, and Fayette (fig. 5). Escherichia coli concentrations ranged from less than 1 MPN/100 mL, the minimum level of quantification, to 770 MPN/100 mL at Prince, 730 MPN/100 mL at Thurmond, and 1,100 MPN/100 mL at Fayette. Exceedances of the recreational-contact standard for E. coli (320 MPN/100 mL) were measured in 11 of the 110 depth- and width-integrated composite samples during the 2-year study period. Of the 11 exceedances of the recreational-contact standard, 5, 3, and 3 exceedances were measured at Fayette, Thurmond, and Prince, respectively. Eight of the 11 exceedances of the recreational-contact standard happened during three sampling efforts associated with storm runoff and elevated streamflow (streamflow greater than 18,000 cubic feet per second [ft³/s] at Thurmond; U.S. Geological Survey, 2025c). The other three recreational-contact standard exceedances were associated with antecedent periods of lower flow in the New River (less than 9,000 ft³/s at Thurmond). These instances suggest that storms can increase tributary contributions of E. coli during periods of low flow in the New River when the river’s capacity to dilute tributary input is diminished.

Cross-sectional point samples were collected during routine sampling events to assess how well the depth- and width-integrated composite samples represented the cross section’s inherent variability. A total of 21 sets of point samples were collected from the 3 primary monitoring stations: 6 sets at Prince, 8 sets at Thurmond, and 7 sets at Fayette. Escherichia coli concentration results from each cross-sectional point sample, with uncertainty represented with 95-percent confidence intervals, were compared with the E. coli concentrations and 95-percent confidence intervals of the composite sample. As previously discussed in the “Methods of Data Collection and Analysis” section, samples were considered not to differ significantly if their 95-percent confidence intervals overlapped. Except for a few nearshore water samples, individual cross-sectional point samples and their corresponding composite sample were not significantly different.

During the 6 cross-sectional sampling events at Prince, a total of 29 point samples were collected (fig. 6). As mentioned previously, the mid-point sample during the January 10, 2023, cross-sectional sampling event was lost during processing, and E. coli analysis results were unavailable. The cross-sectional samples were taken during a range of conditions, with E. coli concentrations in corresponding composite samples ranging from 7 to 290 MPN/100 mL. Of the 29 point samples collected during 6 sampling events, 27 had an overlapping 95-percent confidence interval with the composite sample, indicating no statistically significant differences. The two point samples with no overlap were taken on July 27, 2022: these were the cross-sectional point samples closest to either riverbank and had an E. coli concentration of 140 MPN/100 mL, against a composite sample concentration of 290 MPN/100 mL. This sampling event took place when New River streamflow was at its highest (as indicated by Thurmond) among all cross-sectional sampling events. Although only one sample set indicated differences, these differences may indicate greater cross-sectional variability during high flows.

During the 8 sampling events at Thurmond, a total of 40 cross-sectional point samples were collected (fig. 7). These samples were collected during flows ranging from 3,490 to 22,700 ft³/s. The 95-percent confidence intervals of the cross-sectional point samples overlapped those of the associated composite sample for all but one point sample. The E. coli concentrations of the composite sample and the sample from the 15-percent cross-sectional point on May 24, 2022, were less than 1 MPN/100 mL and 10 MPN/100 mL, respectively.

During the 7 sampling events at New River at Fayette, a total of 35 cross-sectional point samples were collected (fig. 8). In all but one instance, September 29, 2022, the 95-percent confidence intervals of the point samples and the associated composite samples overlapped. The E. coli concentrations of the composite sample and the sample from the 15-percent cross-sectional point were 28 MPN/100 mL and 9 MPN/100 mL, respectively.

Most 95-percent confidence intervals of E. coli concentrations from cross-sectional point samples overlapped with the 95-percent confidence intervals of E. coli concentrations from corresponding composite samples, indicating that the New River had relatively little overall variability in concentrations within its cross sections and that E. coli concentrations at all three monitoring locations are well represented by the depth- and width-integrated composite sampling techniques used in this study. The average concentration of the five cross-sectional point samples was within the 95-percent confidence interval of the composite samples, with the exception of two instances at Prince, which includes the July 27, 2022, sampling event described previously.

However, of the few instances where cross-sectional samples did not agree with composite samples, all happened at cross-sectional points at either end of the stream width, closest to the riverbank. This might suggest that the highest variability is in nearshore samples, but the data do not indicate a consistent high- or low-bias in nearshore samples.

Four longitudinal transects were done, two each during the 2022 and 2023 peak recreational-contact seasons from May through October. The July 2022, September 2022, and October 2023 transects were done under relatively stable, seasonally typical low-flow conditions. During these three transects, water samples collected along the longitudinal transect of the New River Gorge had little variability in E. coli concentrations as indicated by the broad overlap of the 95-percent confidence intervals among individual water samples (fig. 9). Furthermore, the overlap of the 95-percent confidence intervals indicates E. coli concentrations at any monitoring location did not have a statistically significant difference from any monitoring location immediately up- or downstream.

The longitudinal transect collected on August 25, 2023, offered insight into conditions during storm events. Although it was only a partial transect of the upper New River, water samples were collected after an intense thunderstorm, with 0.56 inches of rainfall recorded in an hour by Thurmond’s precipitation gage (USGS 03185400; U.S. Geological Survey, 2025c). Although the exact timing, magnitude, and spatial distribution of this storm are difficult to ascertain—particularly as they pertain to streamflow response and sampling timing—the precipitation runoff did produce an increase of New River tributary streamflow. For example, streamflow at Piney Creek at Raleigh, West Virginia (USGS 03185000; hereafter, Piney Creek), the only gaged tributary in the New River Gorge, increased 68 percent (from 4.2 to 7.07 ft³/s) by the end of the day’s longitudinal transect sampling.

The results of the August 2023 longitudinal transect show that E. coli concentrations were 82-percent higher at the New River below Piney Creek at McCreery, West Virginia (USGS 03185208; 52 MPN/100 mL) monitoring location than at Prince (28 MPN/100 mL), which is immediately upstream of the confluence of Piney Creek and the New River; however, the 95-percent confidence interval associated with these two samples narrowly overlaps (fig. 9). The largest difference in the E. coli concentrations of the August 2023 longitudinal transect was observed when comparing the New River below Dunloup Creek at Thurmond, West Virginia (USGS 03185434; 19 MPN/100 mL) and the New River below Arbuckle Creek at Thurmond, West Virginia (USGS 03185440; 390 MPN/100 mL) monitoring locations. This statistically significant difference (no overlap of the 95-percent confidence intervals) in E. coli concentrations indicates the potential effect that tributaries have on the New River during storm events, especially during low-flow conditions when the river has reduced capacity to dilute tributary loads. New River streamflow at Thurmond was less than 2,000 ft³/s during this transect.

Despite the small number of longitudinal transects, E. coli concentrations had less variability during stable conditions but more variability during storm events that increased tributary streamflow. The August 2023 storm-event transect demonstrated the significance of tributary inputs on main stem New River E. coli concentrations. The E. coli input from Arbuckle Creek was sufficient to cause a twentyfold increase in E. coli concentration in the New River, resulting in an exceedance of the recreational-contact standard at the monitoring location below Arbuckle Creek. The August 2023 transect also highlights the challenges of monitoring during storm events or other periods of increased tributary flow that can lead to high variability of tributary E. coli inputs.

A log₁₀ transformation of measured E. coli concentration was used to improve the distribution and linearity of the E. coli data (Helsel and others, 2020). Data for surrogate model regression analysis were compiled for samples collected October 2021 through October 2023. Variables included E. coli samples collected at Thurmond (36 composite and 4 longitudinal transects); streamflow, water-quality parameters (specific conductance, turbidity, pH, dissolved oxygen, water temperature), and precipitation measured at Thurmond; and streamflow at Piney Creek. In addition to the instantaneous measured values, the maximum, mean, and log₁₀-transformed values for streamflow and turbidity were calculated for the antecedent 48 hours in 12-hour intervals. Precipitation included the antecedent 12-, 24-, 36-, and 48-hour sum totals.

All potential variables were analyzed against log₁₀ E. coli concentration in a linear correlation matrix using Pearson’s method in the statistical programming software R (R Core Team, 2023) and are shown in table 2. Coefficients closer to zero indicate a weaker linear relation; coefficients closer to 1 or −1 indicate a stronger positive or negative linear relation, respectively. However, combining variables that are independently uncorrelated with the response variable with other explanatory variables can sometimes increase prediction accuracies for response variables.

Escherichia coli concentrations were most correlated with antecedent 24-hour mean turbidity at Thurmond and antecedent 12-hour mean streamflow at Piney Creek. Both variables represent changes in streamflow conditions and particulate-matter transport and were positively correlated with E. coli concentrations. These time-averaged explanatory variables likely help to account for hysteresis and (or) lag between turbidity or streamflow, the measured variable, and E. coli concentrations, the estimated characteristic. The lag is likely caused by the many small tributaries in the gorge, which vary in land use, E. coli concentrations, and the travel time of storm-induced streamflow.

Variables shown in table 2 were further examined for suitability in simple linear regression (single explanatory variable) and multiple linear regression models using the leaps package (Lumley, 2024) and statistical tests in the program R, which included adjusted coefficient of determination (R²), residual standard error, predicted residual error sum of squares (PRESS), Cook’s distance, Akaike information criterion, difference in fits, the Shaprio-Wilks test, and outlier tests. Methods and analysis of linear regression model development, refinement, and selection, as described in Helsel and others (2020), were used to evaluate multiple models and statistical metrics. A selection of the linear regression models that explained the most variability in E. coli concentrations is in table 3.

The two best-performing univariate, simple linear regression models for log₁₀ E. coli concentration used the antecedent 24-hour mean turbidity at Thurmond and the antecedent 12-hour mean streamflow at Piney Creek. Combining these two explanatory variables into a multiple linear regression model resulted in a modest improvement in the adjusted R² and other statistical metrics and allowed for an analysis of variance statistic to be calculated with a p-value of 0.02, indicating an improvement over the results of simple linear regression. Two other multiple linear regressions that included 1) the log₁₀ of maximum 24-hour turbidity at Thurmond plus water temperature at Thurmond and 2) the log₁₀ maximum 36-hour streamflow at Thurmond plus the square root of total 48-hour precipitation at Thurmond had summary statistics and residuals slightly better than the simple linear regressions.

Ultimately, however, the statistical improvements of these multiple linear regression models are relatively small when compared with the simple linear regression model. Any additional variables increase the complexity of the E. coli surrogate model, make predictions more computationally demanding, and introduce a greater possibility for real-world variability to affect surrogate model performance. Therefore, for reasons fully explained in the following “Assumptions and Limitations” section, we recommend that the simple linear regression model produced using antecedent mean 24-hour turbidity at Thurmond (fig. 10) be used.

The linear regression model used to predict E. coli concentrations at Thurmond, developed from antecedent mean 24-hour turbidity measured by a streamgage at the location, is expressed asy = b + mx,(1)where

This linear regression model has an adjusted R² value of 0.556 and a p-value of less than 0.001. A plot of the linear regression model is shown in figure 10. Residuals (fig. 11) had a standard error of 0.469 and were skewed toward lower values but still passed the Shapiro-Wilk’s test for normality with a p-value of 0.0998 (normality p-value greater than 0.05 for this test). The observation made on November 16, 2022, was flagged as a statistical outlier; the observed E. coli concentration was 650 MPN/100 mL during relatively low turbidity. However, this sample had a low effect on model parameters and, without a known reason to exclude it, the point was kept in the dataset (Helsel and others, 2020).

Because E. coli concentration values were log₁₀-transformed, the regression equation (eq. 1) will estimate the median rather than the mean response. To correct for the skew of the mean in a log-normal distribution, a bias correction is needed when back-transforming E. coli concentrations from log₁₀ space to real space (Helsel and others, 2020). Duan smearing (Duan, 1983) was used to calculate a bias correction factor of 2.12 for equation 1 and is also provided for the other equations in table 3. The log-transformation bias correction factor was applied and results in a revised equation used to estimate the mean E. coli concentration as expressed asy = 10^{0.44055 + 0.073643x} × 2.12,(2)where

Equation 2 can be used as a surrogate model to estimate the E. coli concentrations at Thurmond from antecedent mean 24-hour turbidity. Conversely, given a known E. coli concentration, the antecedent mean 24-hour turbidity at Thurmond can be estimated using equation 2. For the E. coli recreational-contact standard of 320 MPN/100 mL, the antecedent mean 24-hour turbidity at Thurmond would be estimated by the surrogate model at 23.6 FNU, with a 95-percent confidence interval estimate of 19.4–30.7 FNU.

The development and application of a New River E. coli surrogate model includes several assumptions and limitations. The use of linear regression assumes normally distributed data, and the recommended surrogate model that used the antecedent mean 24-hour turbidity passed the Shapiro-Wilk normality test but has a skew toward smaller values. Linear regression should only be used for estimates within the range of the sampled conditions (antecedent mean 24-hour turbidity of 0–30 FNU) used to develop the model. Note that with a y-intercept of 0.44055 in the linear regression equation and a bias correction factor of 2.12, the model’s lowest estimated E. coli concentration will be 6 MPN/100 mL. Furthermore, any near-real-time use of this surrogate model would use provisional USGS data before any additional review, approval, and archival.

The performance of the surrogate model becomes more important as E. coli concentrations near or exceed the recreational-contact standard because model uncertainty could lead to false positives or false negatives. Few water samples collected during this study contained such high concentrations. More samples and associated data at higher E. coli concentrations could help improve the surrogate model. However, figures 11 and 12 indicate that the uncertainty in turbidity is similar over the full range of antecedent mean 24-hour turbidity because the logarithmic fit of the regression equation indicates that the errors are similar on a percentage basis instead of on a set value (arithmetic fit). Figure 12 shows the surrogate model has a slight bias in predicting high values, suggesting that the model may be conservative when used for recreational-contact guidance.

With an adjusted R² value of 0.556, the linear regression model explains about half of the overall variability of E. coli concentrations at Thurmond. The surrogate model used to predict E. coli concentrations at Thurmond performs well when plotted against measured E. coli concentrations as a time-series during the study from October 2021 to October 2023 (fig. 13). This time-series plot of the surrogate model also indicates that exceedances of the recreational-contact standard were infrequent, with antecedent mean 24-hour turbidity exceeding the 23.6 FNU threshold about 18 times during this 2-year period. Plotting mean 24-hour turbidity at Thurmond for 5 years (July 2019–June 2024) as a rank-ordered duration curve indicates that the 23.6 FNU threshold is surpassed approximately 7.5 percent of the time, and the 19.4 FNU lower confidence limit is surpassed 10 percent of the time (fig. 14).

Although the longitudinal monitoring from this study suggests the New River has fairly consistent water quality throughout NERI during stable-flow conditions, the influx of water from tributaries during storms produces heterogeneous changes in water quality. Most of the sampled tributaries of the New River are downstream from Thurmond, including Dunloup Creek, Arbuckle Creek, Manns Creek, Keeney Creek, and Wolf Creek. As such, any turbidity or E. coli contributions from these tributaries will not be represented in the continuous water-quality monitor at Thurmond or the E. coli surrogate model. The August 2023 longitudinal transect highlighted the variability of tributary E. coli inputs during storm events and increased tributary flow and challenged the assumption that a single monitoring location could fully characterize water-quality parameters and E. coli concentrations throughout the entirety of the New River Gorge under all conditions.

Another key consideration is to evaluate how well the sampled conditions used to develop the surrogate model represent longer-term conditions in the New River. The streamflow during sampled conditions was plotted against the flow-duration curve for Thurmond from October 1994 to September 2024. Sample values are generally well distributed along the flow-duration curve, from 8.2 percent at 22,700 ft³/s to 96.5 percent at 1,590 ft³/s (fig. 15). Some sample clustering happens at between 50 and 60 percent of the flow-duration curve (roughly 5,000 to 6,000 ft³/s), and small sampling gaps are present at higher flows. However, data gaps at higher streamflows may also coincide with flows where recreational contact is less likely to happen. The commercial rafting industry has a cutoff of 12 feet at the informal Fayette boaters’ gage, which is associated with a streamflow of 32,000 ft³/s at Thurmond (James Scott, commercial guide, oral commun., 2025). Although more advanced recreational boaters target flows below 20,000 ft³/s, most private recreational boating happens below 9,000 ft³/s (James Scott, commercial guide, oral commun., 2025). Thus, recreational flows were well represented by the sampled conditions.

As previously mentioned, the use of antecedent mean 24-hour turbidity at Thurmond to develop the simple linear regression surrogate model instead of using antecedent mean 12-hour streamflow at Piney Creek or other multivariate linear regressions that produce modestly higher model performance warrants further discussion. Conceptually, an ideal surrogate model would estimate the desired variable with highly correlated input(s) that also fully explain the underlying, causal physical processes with a minimum of variables, assumptions, caveats, and exceptions. An E. coli surrogate model developed using streamflow presents challenges because total streamflow in the New River is a flexible combination of dam release flow at Bluestone Dam, tributary base flow, and tributary stormflow. Additionally, streamflow and water-quality relations often show hysteresis, in that the same streamflow level on the rising and falling limbs of a hydrograph peak can have different water quality because the watershed is “flushed” on the rising limb. Although we acknowledge the significance of tributary streamflow on E. coli concentrations in the New River, using streamflow data from a monitoring location on a single tributary only indicates conditions in a small part of the New River watershed, perhaps missing the effect of a localized storm event elsewhere in the New River Gorge.

We hypothesize that a water-quality parameter measured in the New River is the best surrogate to estimate another water-quality variable in the New River. Turbidity and E. coli relations have been used in this way in many other studies, including Lawrence (2012), Brady and others (2009), and Brady and Plona (2012). A surrogate model based on Thurmond integrates streamflow and water-quality inputs from more sources throughout the New River watershed: the New River upstream from the gorge, the Greenbrier River, and some of the larger tributaries within the gorge.

Other variables that showed promise in predicting E. coli concentration as part of a multivariate linear regression included water temperature and precipitation. Water temperature was negatively correlated to E. coli concentration: as temperatures decrease, E. coli concentration increases. This correlation may help explain the outlier data point: we theorize that lower temperatures in the New River prolong the persistence of E. coli, preventing die-off that may happen at higher temperatures and maintaining elevated concentrations for longer (Korajkic and others, 2019), conceptually mirroring field methods that kept water samples on ice to maintain E. coli concentrations if laboratory processing was delayed. When precipitation data were being analyzed, the three highest E. coli concentrations at Thurmond, all above the recreational-contact standard (320 MPN/100 mL), were measured in samples collected when there was at least 0.25 inches of total precipitation at Thurmond in the antecedent 48 hours. However, precipitation at Thurmond does not fully explain conditions throughout most of the watershed, particularly with isolated storm events. Three different storm events that had more than 0.25 inches of total precipitation at Thurmond in the antecedent 48 hours were associated with E. coli concentrations below 10 MPN/100 mL, indicating that E. coli’s response to precipitation is highly variable, likely because of the complexities of the watershed’s response to the timing of storms, their intensity, and their spatial distribution.

Adding more explanatory variables to a linear regression model increases its complexity. Additional variables can add assumptions and limitations, particularly given the variability of tributary and precipitation effects. For example, isolated storms may hit or miss the limited tributary and precipitation monitoring network in the New River Gorge. Missing or incomplete data are also more likely to happen with additional explanatory variables and would prevent the use of the surrogate model in those instances. Calculating and communicating predictions, thresholds, or guidance criteria is more straightforward with a simple linear regression surrogate model.

Considering the modest improvements in the linear regression E. coli surrogate model performance yielded by the additional variables, we ultimately decided to use a more conservative simple linear regression based on turbidity at Thurmond—one that contains the fewest caveats, conditions, and assumptions—to reduce model complexity and the influence of spatial heterogeneity. However, analyzing additional variables helped clarify the processes that may affect E. coli concentrations in the New River Gorge. Further sampling and monitoring, particularly in other locations throughout the New River watershed, and possibly to include additional precipitation gages, may uncover additional insights to help meaningfully improve the E. coli surrogate model.

The Bluestone Dam on the New River attenuates streamflow through the gorge. Slower streamflow through the dam pool allows some upstream sediment and suspended matter—and any E. coli associated with these suspended sediments—to settle out rather than being transported further downstream to the gorge (Paybins and others, 2000). We theorize that the attenuation of the Bluestone Dam may uncouple some water-quality patterns in the New River Gorge, particularly the influence of streamflow on suspended matter and turbidity, and increase the effect that more proximal sources, such as the Greenbrier River and the smaller tributaries discharging to the New River Gorge, have on E. coli concentrations in the New River. As noted in this study and others, Piney Creek and the other New River tributaries in the New River Gorge are sources of E. coli (Wilson and Purvis, 2000, 2003). Additional tributary streamgages or water-quality monitoring could provide data that could further increase understanding of E. coli concentrations, loads, and trends throughout the gorge and improve the amount of variability in E. coli concentrations explained by the New River E. coli surrogate model.

This report demonstrates the utility of the continuous streamflow and water-quality monitor at Thurmond to NERI park staff and visitors. Similar water-quality monitoring efforts and surrogate modeling relations have supported E. coli recreational advisory tools in partnership with the National Park Service in Georgia’s Chattahoochee River National Recreation Area (Aulenbach and McKee, 2020; U.S. Geological Survey, 2020) and Ohio’s Cuyahoga Valley National Park (Brady and Plona, 2015).