Environ. Sci. Technol. 2008, 42, 4818–4824
Nowcasting and Forecasting Concentrations of Biological Contaminants at Beaches: A Feasibility and Case Study W A L T E R E . F R I C K , * ,† Z H O N G F U G E , ‡ A N D RICHARD G. ZEPP† Ecosystems Research Division, U.S. Environmental Protection Agency, Athens, Georgia 30605, and National Research Council, ERD, U.S. Environmental Protection Agency, Athens, Georgia 30605
Received December 19, 2007. Revised manuscript received April 13, 2008. Accepted April 14, 2008.
Public concern over microbial contamination of recreational waters has increased in recent years. A common approach to evaluating beach water quality has been to use the persistence model which assumes that day-old monitoring results provide accurate estimates of current concentrations. This model is frequently incorrect. Recent studies have shown that statistical regression models based on least-squares fitting often are more accurate. To make such models more generally available, the Virtual Beach (VB) tool was developed. VB is publicdomain software that prescribes site-specific predictive models. In this study we used VB as a tool to evaluate statistical modeling for predicting Escherichia coli (E. coli) levels at Huntington Beach, on Lake Erie. The models were based on readily available weather and environmental data, plus U.S. Geological Service onsite data. Although models for Great Lakes beaches have frequently been fitted to multiyear data sets, this work demonstrates that useful statistical models can be based on limited data sets collected over much shorter time periods, leading to dynamic models that are periodically refitted as new data become available. Comparisons of the resulting nowcasts (predictions of current, but yet unknown, bacterial levels) with observations verified the effectiveness of VB and showed that dynamic models are about as accurate as long-term static models. Finally, fitting models to forecasted explanatory variables, bacteria forecasts were found to compare favorably to nowcasts, yielding adjusted coefficients of determination (adjusted R2) of about 0.40.
Introduction Beaches that meet water quality standards are valued for their aesthetics, recreation, and ensuing financial opportunities. Noncompliant beaches pose health threats to bathers and should be posted. When it comes to knowing current beach water quality, the obvious solution is to sample the water. This has been the focus of most assessments. Analytical results are commonly used as the basis for issuing swimming advisories. However, since Escherichia coli and other indicator bacteria samples require about 24 h to analyze, * Corresponding author e-mail:
[email protected]. † Ecosystems Research Division. ‡ National Research Council. 4818
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 42, NO. 13, 2008
the success of this “persistence model” depends on the steadiness of their concentrations. When conditions are highly variable, this model often produces false positive and false negative predictions (1–5). (A false negative outcome results when a concentration below the health criterion is predicted when it is actually exceeded. This outcome can expose the public to dangerous levels of microbial contamination. False positive outcomes are overly protective, erroneously predicting high concentrations). In 2000, Congress passed the Beaches Environmental Assessment and Coastal Health (BEACH) Act to address the increasing problem of beach contamination. The Act affirmed the role of predictive models for notification of potentially unhealthy conditions at beaches. Indeed, statistical regression models have been used increasingly to predict contamination at beaches (1, 4–15). These models use available independent data, called explanatory variables, to predict the yet-unknown bacteria concentrations, called the response variable. Model development uses the least-squares method to fit to data. Such models also have been successfully developed for marine beaches (2, 7, 16–18). These models are based primarily on variables measured at the beach or nearby weather stations. Recent efforts have attempted to incorporate watershed characteristics such as land use and topography (8, 9). Compared to the persistence approach, these efforts have often proved successful. MLR modeling consists of two distinct phases: diagnosis and prognosis. In the diagnostic phase, known response variable data and explanatory variable data are related systematically to produce mathematical models that optimally fit the data. In the prognostic phase, the models so obtained are used to predict bacteria concentrations either unknown at the time of prediction or deliberately excluded from the diagnoses (fits) for research purposes. The adjusted R2 and other statistics can be used to assess and express both the diagnostic goodness of fit and the prognostic predictive capacities of the models, but in this paper all listed adjusted R2 values quantify prognostic results. If real-time explanatory variables and microbial data are available, then empirical multiple linear regression (MLR) models may be fit to provide a basis for issuing health advisories. A notable example of this approach was the use of E. coli levels and explanatory variables measured by the U.S. Geological Survey (USGS) to develop MLR models to predict bacteria concentrations and issue advisories at Huntington Beach, Ohio on the western shore of Lake Erie (19). USGS used 2000-2005 summer data to fit a unique MLR model that then was used to produce the 2006 advisories. In this paper, such a singular model is called “static”. Using a two-step process, a nowcast (a short-term forecast of near real-time conditions) of E. coli concentration was produced that was analyzed further to establish the probability of the estimate being above the standard 235 colonies per 100 mL (col/100 mL) concentration, defining a violation of the health standard. A supplemental frequency analysis of 2000-2005 data established a probability criterion to optimize predictions, while also addressing the imbalance of false positives and negatives (19, 20). These studies clearly provided beach managers with a powerful tool for beach notification. The Huntington Beach, Ohio study also provided critical data for testing the Virtual Beach tool to evaluate various approaches to MLR modeling (21). This beach also was once used by the U.S. Environmental Protection Agency for conducting an epidemiological study of swimming-associated gastrointestinal illness (22, 23). 10.1021/es703185p CCC: $40.75
2008 American Chemical Society
Published on Web 06/03/2008
TABLE 1. List of Nowcast and Forecast Explanatory Variables Considered in This Study variable turbidity water temperature wave height 24-/48-h rainfall solar radiation (8am, 9am) cloud cover dewpoint temperature (8am) wind direction wind speed alongshore wind component cross-shore wind component
nowcast variable unit source NTU USGSa F USGSa category 1-4 USGSa in USGSa watts/m2 OARDCe category 0-5 (clear NWS KCLEd to overcast) F NWS KCLEd deg NWS KCLEd mph NWS KCLEd mph derived from wind direction and speed mph derived from wind direction and speed
air temperature dewpoint temperature cloud cover precipitation potential wave height wind direction wind speed alongshore wind component
F F % % ft deg mph mph
cross-shore wind component
mph
24 h forecast variable KCLEb KCLEb KCLEb KCLEb NOAAc KCLEb KCLEb derived from wind direction and speed derived from wind direction and speed
transformation natural log none square root square root none or square none or square none none none none or asymmetricf none or asymmetricf
none none none or square square root square root none none none or asymmetricf none or asymmetricf
a www.ohionowcast.info. b National Weather Service, Cleveland-Hopkins International Airport (KCLE): www.erh.noaa.gov/ ifps/MapClick.php?FcstType)digital&textField1)41.492627&textField2)-81.827063&site)cle. c www.weather.gov/forecasts/ graphical/sectors/eastgrtlakes.php. d National Weather Service, Cleveland-Hopkins International Airport (KCLE): www. erh.noaa.gov/data/obhistory/KCLE.html. e www.oardc.ohio-state.edu/newweather. f An asymmetrically transformed real number has the same sign but an absolute value square-root of its original absolute value.
The first objective of the present study was to demonstrate the efficacy of the VB MLR model development tool. The second objective was to evaluate and assess the feasibility of dynamic models, those that are refit periodically to new data as they become available. If short-term, dynamic models can provide satisfactory predictions there would be important implications for “quick-start” programs at many locations. A recent study at two marine beaches indicated that shortterm data sets can be used to develop dynamic models that provide satisfactory predictions of enterococci densities (7), but this approach has not been tested thoroughly at freshwater beaches. The USGS static model and relatively short-duration dynamic VB models were compared using the adjusted R2 statistic as a performance measure. The third objective was to evaluate 24 h forecasts of microbial contamination at beaches. The study also tested the usefulness of publicly available data.
Materials and Methods Site Description. Data obtained by the USGS from Huntington Beach, Ohio, on Lake Erie west of Cleveland, were used in this study. The site is described in several reports by Francy and co-workers (19–21) and also by Haugland et al. (22). Data. An important subobjective of this VB study was to evaluate the usefulness of publicly available data. For this purpose over 30 real-time and forecasted environmental data were obtained (mined) from the NOAA weather station at the Cleveland-Hopkins International airport (KCLE), NOAA Buoy 45005 in Lake Erie, the agricultural station at Ashtabula, Ohio, and other sources to form a large set of candidate explanatory variables. A subset of the more important explanatory variables is given in Table 1. E. coli concentrations and several explanatory variables were downloaded from the USGS Web site (19). In the USGS
study, water samples were obtained from May 30 to Aug 31, 2006 each morning at a depth of about 30 cm in approximately one meter of water. The sampling time was nominally 0900 but ranged from 0752 to 0952. Samples were analyzed for E. coli concentration as described elsewhere (19, 21). The explanatory variables turbidity, wave height, and water temperature were measured simultaneously. Like the VB study, 24 h rainfall was obtained from the NOAA KCLE Web site. (See Table 1). Thus, models prescribed by VB were based on more explanatory variables than were used by Francy et al. (19–21). This expanded set of explanatory variables potentially could yield improved predictive capacity. Among the numerous variables tested, dew point temperature, wind speed and direction, and cloud cover proved especially valuable. While the MLR approach is not dependent on process-based theory, good potential explanatory variables presumably are related to transport and transformation mechanisms. The Virtual Beach Tool. This section presents relevant and necessary components and functionalities of VB. The VB Multiple Linear Regression (MLR) Tool. VB is comprised of several components, of which the MLR tool for developing statistical models is the most advanced. VB helps develop statistical models (eq 1) for predicting beach E. coli concentrations (denoted byEC): p
E[ln(EC)] ) β0 +
∑βx
(1)
i i
i)1
where E[ln(EC)] represents the natural logarithm of the mean of the random variable (EC), β0 and βi the regression coefficients, xi the explanatory variables, and p the number of variables used in the model. MLR analysis is based on the least-squares method to fit models and is subject to several considerations, notably variable interactions, multicollinearVOL. 42, NO. 13, 2008 / ENVIRONMENTAL SCIENCE & TECHNOLOGY
9
4819
ity and model selection (24, 25). Their relevance to beach bacteria modeling is described further by Ge and Frick (26). VB uses a backward elimination process, described in the parsimonious model section below, to help the user select the most promising models from a number of candidate models that rapidly increase with the number of explanatory variables in the analysis. This process offers a way to rank the usefulness of the selected explanatory variables, the best variable being the last recommended for elimination. VB facilitates model development, making it possible to test systematically the idea that dynamic models based on limited data sets could produce useful models. Thus, as part of the diagnostic phase of the study, a specific number of daily observation sets, n, were used to fit the various different models; n varied from 21 to 49 days, in weekly intervals. In the subsequent prognostic phase, the predicted E. coli concentrations for times outside the diagnostic data sets were then compared to the observed concentrations to evaluate model predictive performance. Data Transformations. By definition MLR equations (eq 1) are linear. This property can limit the value of explanatory variables if the response variable is not a linear function of the variable. Variables transformations overcome this limitation. In fact, the response variable itself is routinely naturallog transformed, a transformation that alone can increase the adjusted R2 value of the model substantially. VB offers a number of common transformations, including square root, square, and others that, when selected, automatically serve to transform the values of the explanatory variables. Transformations can greatly increase the predictive value of vector quantities, such as wind and current (27). For example, wind speed and direction are two independent variables that, untransformed, often rank low as MLR explanatory variables (28). The problem is that wind and current are vector quantities and direction is a harmonic function. However, by transforming vector variables into their components, e.g., into alongshore and cross-shore components, their value is often enhanced. VB includes the trigonometric transformations for converting wind (or current) speed and direction into their vector components, including axis rotation. It also offers asymmetric transformations, as the effects of wind speed may be completely different for onshore winds than offshore winds, for example. Thus the offshore component might be square root transformed while the onshore component remains unchanged. Variable transformations used in the present study are also listed in Table 1. The Parsimonious Model. Most importantly, VB helps to identify the best explanatory variables from the suite of potential variables available for fitting. In a typical study dozens of candidate variables might be considered (28, 29). Each variable tends to increase the explained variance, but in successively smaller increments compared the best variables. Fit to noise or chance occurrences present in the data, and risking multicollinearity, the marginal variables can degrade prognostic performance and model maintenance also increases. In practice a smaller set of variables (Table 1) is preferred from which a parsimonious model is finally selected. In this work, the number of variables recommended for the parsimonious model ranged from two to five, but, for uniformity and comparability, four variables were retained. To help the user identify the most significant variables, VB offers an automated model selection facility. VB uses backward elimination with Mallows’ Cp (30) 2 2 Cp ) p + (n - p)(σ2 - σfull ) ⁄ σfull
(2)
as the selection criterion, where n is the number of samples, 2 p is the number of included variables, and σ2 and σfull the residual variances of the reduced and the full models. The parsimonious model is defined as the one with the minimum 4820
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 42, NO. 13, 2008
Cp value among the full (all variables) and all reduced (fewer variables) models. The “automate model selection” command computes the Cp statistic for all possible models and ranks the variables for elimination. If the default parsimonious model recommended by VB is not selected, a process of guided variable elimination can be performed to arrive stepwise at the final model, based on other factors that might influence variable selection. VB offers other features, such as warnings when the data matrix is singular and when residuals demonstrate significant serial correlations. Users can also check for influential cases (those that greatly influence the values of the regression coefficients) and data outliers. Model Evaluation. It is important to differentiate between fits (diagnostics) and forecasts (prognostics). There are two common indicators of model performance, goodness of fit and predictive capacity, the latter being the only measure of predictive performance. But, whether diagnostic or prognostic, the adjusted coefficient of determination (adjusted R2) is an appropriate measure of model performance (21, 27). The root-mean-square-error of prediction (RMSEP) has been advocated for this purpose (7), however R2 and RMSEP are essentially complementary, representing explained and unexplained variance, respectively. Here, the adjusted R2, a slightly more conservative measure of performance, is used.
Results and Discussion Persistence Models. When beach advisories rely on sampling techniques that require about a day to analyze, a closure decision is made factually using a persistence model, namely ^
YJ ) YJ-1
(3)
where Yˆ and Y are the estimated and observed ln(EC), respectively, and J is the Julian day. This model can be very effective in locations where conditions change slowly, i.e., where conditions persist. But Huntington Beach bacteria concentrations were found to be highly uncorrelated with those of the previous day, so the persistence model performs poorly there, a phenomenon investigated by Kim and Grant (31). Using the persistence model on the 2006 data set yielded an adjusted R2 of only 1.2%. There were only seven correct predictions out of a possible 26, 18 false negatives and 18 false positives. At beaches like Huntington Beach, this and other studies (2, 4, 7, 16, 32) reconfirmed that MLR models outperform the persistence model in the accuracy of health advisories. Various Approaches to Developing MLR Models for Nowcasts. MLR models can be fit to data sets of durations ranging from a few days (the actual minimum number is a function of the number of variables included) to a season or more. Some authors have maintained that MLR models should be based on long-term data sets (4, 21), producing models that are referred to as static models in this paper. With a static MLR model the regression coefficients are invariant over the period of its application. The dynamic modeling approach anticipates that models may vary as the regression responds to recent trends in the data. Periods as short as 10 days have been used to fit models at marine beaches (7). Both the model coefficients and best variables can vary over time due to changing environmental and other conditions, such as treatment plant flow rates, land use changes (9), storm events, and even climatic fluctuations, which are not well represented by static models (18). In fact, as the database grows, the best predictive capacity might be obtained using models based on limited subsets of the data record (7). A dynamic MLR model can be defined as p
E[ln(EC), t] ) β0(t, n) +
∑ β (t, n)x i
i)1
i
(4)
FIGURE 1. The variable make up of E. coli nowcasting models: models selected from ten potential explanatory variables. Fits are based on periods of 21, 28, 35, 42, and 49 day durations. Shown below each tier of models are the adjusted R2 values, standard errors of the estimates, correct predictions above the health standard, and numbers of false negatives and false positives. To fit the first 42 and 49 day models and the second 49 day model, requisite season end data were wrapped to the beginning as if they represented data from the previous season (shading). The same procedure was applied to the 6/13 and 6/20 49 day models. The significance of the variables, indicated by the colors, is based on the model selection process using Cp as the selection criterion. where t is time (days) and n is the model fit sample size (perhaps much smaller than the unabridged data set sample size, N) included in the regression. If dynamic models prove useful, beach advisory programs can begin promptly, not waiting for extensive data to accumulate. The Performance of Dynamic Nowcast Models of Variable Duration. To evaluate the efficacy of the VB tool we compared its summary statistics, i.e., the adjusted R2 values, to those of the USGS. For the 2006 season the overall adjusted R2 for the USGS static model was reported to be 42% (21). For comparison, Figure 1 (upper tier) summarizes the predictive capacity of several VB models, variously fit from 21 up to 49 days of data, in terms of this statistic. The adjusted R2 values of predictions are 50.0, 45.7, 61.0, 53.0, and 60.7% respectively. Again, like the USGS static model, these VB models included turbidity in their suites of explanatory variables. Clearly, VB statistics are comparable to the USGS static model statistic. However, the VB nowcast statistics refer to the final 49 days of the 2006 season. Figure 1 categorizes models by their duration of fit to the data and best explanatory variables. The explanatory variables used in each case are listed to the left of the model panels. In each of the two horizontal tiers, different categories of models based on five different fitting durations are represented: dynamic model fits 21, 28, 35, 42, and 49 days in duration. Dates above the upper tier give the start of each model’s implementation. For example, the 21 day model used to nowcast daily bacteria concentrations for the seven days starting 14 July and ending 20 July was fitted to the previous 21 days (23 June to 13 July) of known data. This was followed by the second 21 day model in that category fitted to the 21 days prior to 21 July, and so on. Thus each of seven 21 day models contributed seven predictions for a total of 49 predictions (n ) 49 for each category). For all models, statistics appear below the two tiers: adjusted R2 (%), standard error of the estimates, number of correct predictions above
the health standard, false negatives, and false positives. The vertically shaded portions (42 and 49 day models) define instances where the data had to be wrapped at the beginning of the season. There is much more information to be gleaned from Figure 1, however, than just the adjusted R2 statistic of the various models. The figure strives to elucidate aspects of dynamic modeling and facilitate comparison of nowcast and forecast results to achieve two other main objectives of this work. Figure 1 shows that turbidity, with few exceptions, was the last variable recommended for elimination. By this criterion, turbidity was the best explanatory variable in 2006. Thus, for the sake of comparison to USGS results, turbidity had to be included in the VB model fit; however, turbidity is not a forecasted variable. To place subsequent forecast results in perspective, nowcast models that do not depend on turbidity also were developed and evaluated (lower tier, Figure 1). Starting with the 21 day model that includes turbidity, turbidity was the strongest explanatory variable for the week of 14 July. In the second and third weeks turbidity was not among the four strongest variables; however, in the fourth week it appears in second place and thereafter returns to being the strongest variable. Considering all model duration categories, it is the best explanatory variable, followed by dew point temperature and cloud cover. In contrast, the wave height variable, important in the USGS study (19, 21), only appears intermittently. The models’ statistics also show an interesting pattern. First, adjusted R2 values of predicted vs observed E. coli concentrations were generally greater for models including turbidity. Second, adjusted R2 values generally increased as the fitting period for the data sets increased from 21 to 35 days, but were more or less constant, even decreasing as the duration of fit increased further. The standard errors followed VOL. 42, NO. 13, 2008 / ENVIRONMENTAL SCIENCE & TECHNOLOGY
9
4821
FIGURE 2. Five-week model results. Models with turbidity included are indicated by a blue line and stars, models with turbidity omitted by red lines and diamonds. Adjusted R2 values are 61 and 40% respectively. a similar, but inverse, relationship. Overall, models of limited (35 to 42 days) duration appeared to perform best. Also, models incorporating turbidity show great stability; i.e., turbidity, dew point, and cloud cover are consistently among the three best variables, often ranked one, two, three, respectively. The remaining achievable variance derives mainly from the 24 h rain variable, with a smattering representation by the remaining six variables, some of which are related. For models not incorporating turbidity (lower tier, Figure 1), this stability weakens as the selected explanatory variables change from week to week. In the absence of turbidity, wave height, rarely among the four best variables in the top tier, emerges as a strong explanatory variable. This suggests that wave height and turbidity are moderately collinear (or redundant) and other variables contribute better to the overall fit when turbidity is available. A variable like Julian day serves better in short-term models, reflecting its value in predicting short-term trends, especially at the beginning of beach season. Finally, the pattern of stability disappears among longer term models, which suggests that dynamic modeling may be valuable even with longer-term data sets. These varied observations suggest that further research on dynamic modeling is worthwhile. A time-series plot of nowcast results for the 35 day dynamic models, using and omitting turbidity, shows generally how the MLR models follow major patterns in the history of E. coli concentrations (Figure 2). The adjusted R2 values are 61% and 40% for models including and omitting turbidity, respectively (Figure 1). These results show that VB performed as intended, albeit on a part of the season that appears to have been more predictable than earlier in the summer. Operational Considerations. The operational modeling requirements implied here may seem daunting. However, the automated functionalities of VB allow its users to perform daily model updates in about one hour. As described, VB facilitates the rapid assessment of promising explanatory variables and development of models based on the use of Mallows’ Cp criterion. To reduce analytical time further, dynamic models can be updated less frequently, perhaps weekly, as done herein. Other interesting findings emerged from testing dynamic modeling. While turbidity was a valuable explanatory variable 4822
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 42, NO. 13, 2008
in the 2006 nowcast test, useful models were developed incorporating off-site explanatory variables. For example, models incorporating KCLE dew point temperature and cloud cover variables routinely outperformed those incorporating on-site water temperature and wave height category variables. This is valuable knowledge as these data can be obtained with little effort. At times VB recommended parsimonious (“best”) models of fewer than four variables, but for demonstrative purposes four were always used to develop Figure 1. Dynamic Forecasting Models. Unlike nowcasts, shortterm predictions that use observations of actual conditions to populate the explanatory variables, forecasts are longterm predictions, typically 24-48 h into the future, using primarily KCLE forecasts as explanatory variables. Theoretically, however, forecasting is not different from nowcasting when MLR models are used. The reason is that all explanatory variables are treated like random dimensionless quantities so that observed and forecast explanatory variables are actually indistinguishable. However, bacteria forecasting accuracy is a function of the accuracy of forecasted explanatory variables. The acquisition of weather forecasts for KCLE, 11 km southeast of the test beach site, began July 21, 2006. The variables are listed in Table 1. As it was impossible to fit the forecast models only to forecast variables and begin 24 h bacteria forecasting on 21 July, it was necessary to use cognates. Cognates of variables are alternative variables of similar content but different character. For example, the observed cloud cover (reported in five categories) and the forecasted cloud cover (reported as the percent of sky covered) are cognates. To fit models that depend on sky cover reported as a category or in percentage, a regression relationship between the two methods of reporting was obtained. In this way the data are rendered interoperable with the same units and model fitting can proceed. Even forecasted dew point temperature expressed in the same units differed systematically from subsequently observed values, affecting performance where fit was based on observations; consequently, its regressed values were used. If long data streams are available, it is usually simplest to use forecast variables for both fitting and
forecasting. However, this work showed that hybrid approaches based on cognates are possible. Forecasting Results. The experimental forecast period covered 42 days in the second half of the 2006 beach season. Based on nowcasting results that show five-week models performed best, no attempt was made to fit forecast models with data sets of other durations. For the 42 day forecast period (21 July to 31 August), the adjustedR2 was 42.3%, similar to nowcast results that exclude turbidity. Eight exceedences were correctly predicted above the standard; there were six false negatives and no false positives. For the first time, this study showed that bacteria concentrations could be forecasted with reasonable accuracy, hastening the day when people will be able to better plan their beach holidays. The 24 h forecast performance compared favorably to the corresponding nowcast performance (omitting turbidity) despite the reduction in available explanatory variables. In an operational mode, where the analyst can draw on experience, even better results might be achieved. For example, all forecast results were based on data nominally valid at 9 am CDT. However, an informed analyst might assess how the actual timing of individual forecasted events, like a frontal passage, might affect predicted bacteria concentrations. Whether models are fitted to observations, cognates, or to the forecasted variables themselves, adjusted R2 values for forecasted results were comparable to values for nowcast results (turbidity excluded) in the present study because they were reported more precisely. For example, KCLE cloud cover forecasts reported by the National Weather Service Forecast Office ranged from 0 to 100%, whereas the NOAA NWS weather observations used for nowcasting were limited to only five categories. Thus, greater precision associated with the forecasted variable could partially compensate for its reduced accuracy. This research benefited from a complete data set, which often is not available in practice. In this work it was assumed that the previous day’s response data were always available for fitting. In reality, over weekends analytical bacteria results were delayed several days before they were reported. Knowing the outcome in advance may have influenced analyses in intangible ways. For example, the data used herein were checked for quality, but manually. Unfortunately, under such circumstances it was easy to focus on episodes where discrepancies were large. The results of this study indicate that the Virtual Beach tool can facilitate and optimize the development of statistical models used for nowcasting and forecasting microbial concentrations at recreational water sites. This tool is particularly valuable for optimizing and updating dynamic models based on short-term data sets. Statistical modeling should increase the demand for frequent bacterial measurements, not just because the data are needed to establish the models, but also because success will increase the demand for data at additional sites. The main benefits of modeling are likely to be increased nowcast and forecast accuracy, not economy.
Acknowledgments We acknowledge Donna Francy, Robert Darner, Richard Whitman, David Schwab, Marirosa Molina, Fran Rauschenberg, Mike Cyterski, John Johnston, Candida West, and Robert Swank for their assistance and comments, and the defining role of the National Oceanographic and Atmospheric Administration (NOAA) Oceans and Human Health (OHH) initiative in the early stages of VB development. We acknowledge partial support of this study by the U.S. Environmental Protection Agency’s Advanced Monitoring Initiative. Although this work was reviewed by EPA and approved for publication, it may not necessarily reflect official Agency
policy. Mention of trade names or commercial products does not constitute endorsement or recommendation for use.
Literature Cited (1) Bartram. J. Health-Based Monitoring of Recreational Waters: The Feasibility of a New Approach (The ‘Annapolis Protocol’), WHO/SDE/WSH/99.1; World Health Organization, Geneva, 1999. (2) Boehm, A. B. Enterococci concentrations in diverse coastal environments exhibit extreme variability. Environ. Sci. Technol. 2007, 41 (24), 8227–8232. (3) Nevers, M. B.; Whitman, R. L., Protecting Visitor Health in Beach Waters of Lake Michigan: Problems and Opportunities; Aquatic Ecosystem Health and Management Society 2004. (4) Nevers, M. B.; Whitman, R. L. Nowcast modeling of Escherichia coli concentrations at multiple urban beaches of southern Lake Michigan. Water Res. 2005, 39, 5250–5260. (5) Wymer, L. J.; K. P., B.; Martinson, J. W.; Stutts, W. R.; Schaub, S. A.; Dufour, A. P. , The EMPACT Beaches Project. Result from a Study on Microbiological Monitoring in Recreational Waters, EPA 600/R-04/023; U.S. Environmental Protection Agency, 2005. (6) Boehm, A. B.; Keymer, D. P.; Shellenbarger, G. G. An analytical model of enterococci inactivation, grazing, and transport in the surf zone of a marine beach. Water Res. 2005, 39 (15), 3565– 3578. (7) Hou, D.; Rabinovici, S. J. M.; Boehm, A. B. Enterococci predictions from partial least squares regression models in conjunction with a single-sample standard improve the efficacy of beach management advisories. Environ. Sci. Technol. 2006, 40, 1737–1743. (8) Crowther, J.; Wyer, M. D.; Bradford, M.; Kay, D.; Francis, C. A.; Knisel, W. G. Modelling faecal indicator concentrations in large rural catchments using land use and topographic data. J. Appl. Microbiol. 2003, 94 (6), 962–973. (9) Kay, D.; Wyer, M.; Crowther, J.; Stapleton, C.; Bradford, M.; McDonald, A.; Greaves, J.; Francis, C.; Watkins, J. Predicting faecal indicator fluxes using digital land use data in the UK’s sentinel Water Framework Directive catchment: The Ribble study. Water Res. 2005, 39 (16), 3967–3981. (10) Kleinheinz, G. T.; McDermott, C. M.; Chomeau, V. Evaluation of avian waste and bird counts as predictors of Escherichia coli contamination at Door County, WI beaches. J. Great Lakes Res. 2006, 32, 117–123. (11) Ferguson, C.; Husman, A. M. D.; Altavilla, N.; Deere, D.; Ashbolt, N. Fate and transport of surface water pathogens in watersheds. Crit. Rev. Environ. Sci. Technol. 2003, 33 (3), 299–361. (12) Whitman, R. L.; Horvath, T. G.; Goodrich, M. L.; Nevers; Wolcott, M. J.; Haack, S. K. Characterization of E. coli levels at 63rd Street Beach. Report to the City of Chicago, Department of the Environment and the Chicago Park District, Chicago, IL, 2001; 85 pp. (13) Liu, L.; Phanikumar, M.; Molloy, S.; Whitman, R.; Shively, D.; Nevers, D.; Schwab, D.; Rose, J. Modeling the transport and inactivation of E. coli and enterococci in the near-shore region of Lake Michigan. Environ. Sci. Technol. 2006, 40 (16), 5022– 5028. (14) Mancini, J., Numerical estimates of coliform mortality rates under various conditions. J. Water Pollut. Control Fed. 1978. (15) Schnauder, I.; Bockelmann-Evans, B.; Lin, B. Modelling faecal bacteria pathways in receiving waters. Proc. Inst. Civil Eng., Maritime Eng. 2007, 160 (4), 143–153. (16) Boehm, A.; Whitman, R.; Nevers, M.; Hou, D.; Weisberg, S. NowCasting Recreational Water Quality. In Statistical Framework for Water Quality Criteria and Monitoring; Wymer, L. , Ed.; Wiley: New York, 2007. (17) Boehm, A. B. Model of microbial transport and inactivation in the surf zone and application to field measurements of total coliform in Northern Orange County, California. Environ. Sci. Technol. 2003, 37 (24), 5511–5517. (18) Boehm, A. B.; Grant, S. B.; Kim, J. H.; Mowbray, S. L.; McGee, C. D.; Clark, C. D.; Foley, D. M.; Wellman, D. E. Decadal and shorter period variability of surf zone water quality at Huntington Beach, California. Environ. Sci. Technol. 2002, 36, 3885– 3892. (19) Francy, D. S. Ohio Nowcasting Beach Advisories. http:// www.ohionowcast.info/index.asp. (20) Francy, D. S.; Gifford, A. M.; Darner, R. A. Escherichia coli at Ohio Bathing Beaches-Distribution, Sources, Waste-Water Indicators, And Predictive Modeling; U.S. Geological Survey: Columbus, OH, 2003. VOL. 42, NO. 13, 2008 / ENVIRONMENTAL SCIENCE & TECHNOLOGY
9
4823
(21) Francy, D. S. , R.A., Darner. Procedures for Developing Models to Predict Exceedances of Recreational Water-Quality Standards at Coastal Beaches, U.S. Geological Survey: Columbus, OH, 2006. (22) Haugland, R. A.; Siefring, S. C.; Wymer, L. J.; Brenner, K. P.; Dufour, A. P. Comparison of Enterococcus measurements in freshwater at two recreational beaches by quantitative polymerase chain reaction and membrane filter culture analysis. Water Res. 2005, 39 (4), 559–568. (23) Wade, T. J.; Calderon, R. L.; Sams, E.; Beach, M.; Brenner, K. P.; Williams, A. H.; Dufour, A. P. Rapidly measured indicators of recreational water quality are predictive of swimming-associated gastrointestinal illness. Environ. Health Perspect. 2006, 114, 5. (24) Draper, N. R.; Smith, H., Applied Regression Analysis, 2ed.; John Wiley & Sons Inc.: New York, 1980. (25) Ramsey, F. L.; Schafer, D. W., the Statistical Sleuth: A Course in Methods of Data Analysis, 2nd ed.; Duxbury Thomson Learning: London, 2002. (26) Ge, Z.; Frick, W. E. Some statistical issues related to multiple linear regression modeling of beach bacteria concentrations. Environ. Res. 2007, 103, 358–364.
4824
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 42, NO. 13, 2008
(27) Nevers, M. B.; Whitman, R. L.; Frick, W. E.; Ge, Z. Interaction and influence of two creeks on Escherichia coli concentrations of nearby beaches: exploration of predictability and mechanisms. J. Environ. Qual. 2007, 36, 8. (28) Olyphant, G. A.; Whitman, R. L. Elements of a predictive model for determining beach closures on a real time basis: the case of 63rd Street Beach Chicago. Environ. Monit. Assess. 2004, 98, 175–190. (29) Olyphant, G. A. Statistical basis for predicting the need for bacterially induced beach closures: Emergence of a paradigm? Water Res. 2005, 39, 4953–4968. (30) Mallows, C. L. Choosing variables in a linear regression: A graphic aid In Central Regional Meeting of the Institute of Mathematical Statistics, Manhattan, Kansas; Institute of Mathematical Statistics: Manhattan, KS, 1964. (31) Kim, J. H.; Grant, S. B. Public mis-notification of coastal water quality a probabilistic evaluation of posting errors at Huntington Beach, California. Environ. Sci. Technol. 2004, 38, 2497–2504. (32) Pfister, M., Earth 911 and SwimCast for Lake County, IL Beaches., 2007.
ES703185P
