- Email: [email protected]
Efficacy of monitoring and empirical predictive modeling at improving public health protection at Chicago beaches
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
Available at www.sciencedirect.com
journal homepage: www.elsevier.com/locate/watres
Efficacy of monitoring and empirical predictive modeling at improving public health protection at Chicago beaches Meredith B. Nevers*, Richard L. Whitman U.S. Geological Survey, Great Lakes Science Center, Lake Michigan Ecological Research Station, 1100 N. Mineral Springs Road, Porter, IN 46304, USA
article info
abstract
Article history:
Efforts to improve public health protection in recreational swimming waters have focused
Received 17 August 2010
on obtaining real-time estimates of water quality. Current monitoring techniques rely on
Received in revised form
the time-intensive culturing of fecal indicator bacteria (FIB) from water samples, but
3 December 2010
rapidly changing FIB concentrations result in management errors that lead to the public
Accepted 6 December 2010
being exposed to high FIB concentrations (type II error) or beaches being closed despite
Available online 13 December 2010
acceptable water quality (type I error). Empirical predictive models may provide a rapid solution, but their effectiveness at improving health protection has not been adequately
Keywords:
assessed. We sought to determine if emerging monitoring approaches could effectively
E. coli
reduce risk of illness exposure by minimizing management errors. We examined four
Fecal indicator bacteria
monitoring approaches (inactive, current protocol, a single predictive model for all bea-
Recreational water quality
ches, and individual models for each beach) with increasing refinement at 14 Chicago
Lake Michigan
beaches using historical monitoring and hydrometeorological data and compared
Swimming
management outcomes using different standards for decision-making. Predictability (R2) of
Risk
FIB concentration improved with model refinement at all beaches but one. Predictive models did not always reduce the number of management errors and therefore the overall illness burden. Use of a Chicago-specific single-sample standarddrather than the default 235 E. coli CFU/100 ml widely useddtogether with predictive modeling resulted in the greatest number of open beach days without any increase in public health risk. These results emphasize that emerging monitoring approaches such as empirical models are not equally applicable at all beaches, and combining monitoring approaches may expand beach access. Published by Elsevier Ltd.
1.
Introduction
In recent years, efforts to improve public health protection in recreational swimming waters have focused on obtaining realtime estimates of water quality. Current monitoring techniques rely on the culturing of fecal indicator bacteria (FIB)d
such as Escherichia coli or enterococcidfrom water samples, a process that requires an incubation time often in excess of the rate of change of bacteria concentrations in the water (Boehm et al., 2002; Whitman et al., 1999). Because of the lapse in results availability, the public are often either unknowingly swimming in contaminated beach water or are prohibited from
Abbreviations: FIB, fecal indicator bacteria; CFU, colony-forming units; MPN, most probable number; IA, inactive monitoring program model; CM, current model; RM, regional predictive model for all study beaches; IM, individual beach predictive model. * Corresponding author. Tel.: þ1 219 926 8336x425; fax: þ1 219 929 5792. E-mail address: [email protected] (M.B. Nevers). 0043-1354/$ e see front matter Published by Elsevier Ltd. doi:10.1016/j.watres.2010.12.010
1660
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
swimming in water that meets the public health criteria. Efforts have been focused on two means of correcting this shortcoming: shorten the analytical time for the current indicator or find an alternate, faster way to assess water quality. To accomplish the latter, empirical predictive models have been attempted with various levels of success and application. Predictive models have been suggested by numerous authors as a potential means for minimizing errors in beach closings (Hou et al., 2006; Kim and Grant, 2004; Nevers and Whitman, 2005). These models range from simple models that associate weather conditions with direct bacteria loadingsesuch as rainfall and associated runoff (McPhail and Stidson, 2009) e to more advanced models that integrate multiple hydrometeorological variables (Kim and Grant, 2004). Model accuracy at predicting FIB concentration depends on beach location, instrument accuracy, wealth of available data, and level of effort, but predictive models can be successfully incorporated into beach management (Nevers and Whitman, 2005). Beaches at which models have been attempted tend to be high profile beaches with heavy visitor use (Boehm, 2007; Hou et al., 2006), directly or strongly impacted by a large point source (He and He, 2008), or having frequent swimming advisories (Olyphant and Whitman, 2004). The accuracy or success of a given modeling approach has typically been assessed by analyzing the amount of variation in the target FIB explained by the model, the error explained by the model, or the specificity (the percent of false negatives, or type II errors) and sensitivity (the percent of false positives, or type I errors) of the model. The first two calculations determine the accuracy of the model at predicting all FIB concentrations. The third, error-based calculation is used due to the use of a binary approach in beach management policies: beaches are either open or closed to swimming, depending on where the FIB concentration falls relative to a designated standard (acceptable health risk); errors occur when the predicted concentration is not equal to actual concentration. Errors result in either inadvertent exposure of the public to high concentrations of FIB (type II error) or exclusion of swimmers from water that meets the exposure standard (type I error). More type II errors result in more swimmers exposed to high concentrations of FIB and therefore a higher public health risk; decreasing the instances of type II errors is necessary to increase public health protection. Current water quality standards for freshwater were developed using epidemiological studies and based on historical acceptable illness rate (Pru¨ss, 1998, US EPA, 1986). Within the monitoring guidance, however, some measure of flexibility was provided for beach managers, including choice of application of two mathematical estimates of illness risk, based on the concentration of indicator bacteria (US EPA, 1986). Generally, beach managers have applied the single-sample maximum for an individual water sample because of its ease of use and interpretation (Nevers and Whitman, 2010, US EPA, 1986), but others use the 5-day geometric mean, both of which should theoretically provide equal levels of health protection. In this paper, we examine four potential monitoring approaches with increasing refinement at 14 Chicago beaches: inactive, current monitoring model, use of one predictive model for all beaches, and use of individual predictive models for each beach. Further, we examine alternate applications of
monitoring standards under these four approaches to assess the health and management outcome possibilities. Using historical monitoring and beach attendance data we compare the accuracy of each model with several calculations and also the relative public health protection provided by each of these models. Specifically, we sought to determine whether predictive modeling at Chicago beaches could be used as a monitoring tool to increase public health protection over traditional monitoring practices.
2.
Materials and methods
2.1.
Study site
Chicago beaches in general are not impacted by a major point source of contamination. Urban sewage is regularly discharged through the Chicago River and a series of man-made or modified canals to the Mississippi River. In events of extreme precipitation, the system override leads to sewage being directed to Lake Michigan (<1 per year); all beaches are then preemptively closed to swimming. Sources of FIB at the Chicago beaches are unknown but likely include beach sand, birds, and algae (Whitman and Nevers, 2003; Whitman et al., 2003). Beaches included in the current study were (from north to south) Loyola, Albion, Hollywood, Foster, Montrose, North Avenue, Oak, 12th Street, 31st Street, 57th Street, 63rd Street, South Shore, Rainbow, and Calumet.
2.2.
Beach monitoring data
E. coli monitoring data, measured as most probable number (MPN)/100 ml of water, were obtained from the Chicago Park District for 2000e2004. Beaches were sampled at least five days a week; replicate samples (up to three) were averaged. E. coli concentrations above or below detection limits were set at detection limits after determining that occurrences were rare (Boehm et al., 2002; Whitman and Nevers, 2008). Missing data points for individual beaches were calculated; values were estimated from the nearest 6 values (average of three previous and three subsequent readings). Beach management is a binary decision: if E. coli concentration <235 MPN/100 ml, the beach is open for swimming; if E. coli >235 a swimming advisory is issued. This model assumes that E. coli concentration today ¼ E. coli concentration yesterday. Inaccurate predictions, therefore, result in a type I or type II error (Table 1). Type I errors occur when the model predicts an E. coli concentration >235 when the actual concentration is <235, resulting in an unnecessary swimming advisory. A type II error occurs when the model predicts E. coli concentration <235 when the actual concentration is >235, resulting in swimmers being exposed to high concentrations of indicator bacteria and associated pathogens. A simple characterization is that type I errors are associated with economic losses because swimmers are denied access and type II errors are associated with greater public health risk, as swimming occurs in the presence of excessive FIB concentrations (Rabinovici et al., 2004).
1661
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
2.3.
Predictive models
Data for developing predictive models were accumulated from several existing hydrometoerological stations, as described in Whitman and Nevers (2008). Predictors included solar insolation and precipitation (24 h); air temperature, barometric pressure, and wave height (mean for 4e10 AM); and day of year. All E. coli results were log-transformed prior to analysis. Predictor parameters were transformed to z-scores prior to regression analyses. Models were developed using multiple linear regression of the available predictors on the independent variable E. coli; Mallows Cp was used to compare resulting models. For individual models, Akaike’s Information Criterion (AIC) was used to select the best fit model for each individual beach. The regional model was developed previously (Whitman and Nevers, 2008) using barometric pressure, wave height, and day of year as the predictors; here, it was applied separately to each study beach. Model comparisons were made using SPSS (Chicago, IL) and SAS (Cary, NC) software: the adjusted coefficient of determination (R2), which describes the amount of variation in E. coli; and percent of type I and type II errors. The four models compared included inactive model (IA): beach is always open to swimming, regardless of microbiological water quality; current model (CM): E. coli concentration for day 1 is used to make a management decision for day 2; regional model (RM): a universal model used for all beaches; and an individual model (IM): a separate predictive model developed for each beach. Comparisons between models for estimating excess illnesses associated with type II errors were only made on days for which there were results for all four models.
2.4.
Calculating illness rates
Overall beach attendance rates were compared for 20 years, and an average per year estimate was calculated (Chicago Park District, unpublished data). With an annual reported range of 14e31 million visitors to all Chicago beaches, 20 million was used as an average. An estimated 91% of all visits, or 18.2 million, were associated with the 14 beaches included in this study. Number of visitors per beach was calculated based on two years for which individual beach data were available (1999 and 2007). It was assumed that 50% of beach visitors had full-body contact with the water as defined by US EPA (Dufour, 1984).
Lower estimates have been published (Rabinovici et al., 2004), but based on data from several Ontario lakes (Seyfried et al., 1985), estuaries (Lepesteur et al., 2006), and marine waters (Dwight et al., 2007; Given et al., 2006), 50% may be conservative for a large freshwater lake. Illness rates for each beach were calculated based on a 100-day swimming season, which assumes equal distribution of visits across the summer (Rabinovici et al., 2004). Illness rate (Y ) was calculated from Dufour (1984): Y ¼ 11:74 þ 9:397log10 ðECÞ
(1)
where Y ¼ the rate of swimming-associated gastrointestinal illness symptoms per 1000 swimmers and EC ¼ E. coli CFU/ 100 ml water. Monitoring standards based on the these epidemiological studies recommend a geometric mean of 126 E. coli CFU/100 ml for five samples over 30 days (Dufour, 1984): an acceptable illness rate of approximately 0.008%. A singlesample limit of 235 CFU/100 ml is also provided, which is within the confidence limits of the calculated geometric mean. These calculations were developed for beaches influenced by a point source (Dufour, 1984), but the sources affecting Chicago’s beaches have not been confidently identified. Using the derived regression equation (Dufour, 1984), we calculated our acceptable illness rate for the 235 CFU as 0.01054%, following Rabinovici et al. (2004). It should be noted that the standards were established using membrane filtration analysis. Chicago uses a defined substrate technique (Colilert; IDEXX, Westbrook, Maine); and while studies have favorably compared the two outcomes (Buckalew et al., 2006), differences in confidence intervals may influence results outcome (Gronewold et al., 2008). A daily excess illness rate was calculated following Given et al. (2006), that is, the number of beach swimmers expected to exhibit symptoms of illness beyond the acceptable 0.01054%. GI ¼ ðY Y0 Þðv=dÞf
(2)
Where Y0 ¼ acceptable illness rate within the monitoring criteria of 10.54/1000, v ¼ number of visitors in a swimming season, d ¼ number of swimming days in the season, and f ¼ percent of beach visitors estimated to have full-body contact with the beach water. The illness regression equation has a y-intercept <0, so calculated illness rates <0 were set at 0 for calculating the mean. There is conflicting research on the association of illness risk and traditional fecal indicators at beaches without a point
Table 1 e Outcome possibilities using the current beach management model: management decision based on previous day’s E. coli concentration. Management outcome
Error
Correct open Correct closed Incorrect open Incorrect closed
None None type 2 type 1
Health risk Loss of Use
Low High High Low
Low High Low High
Health protection Accurate Accurate Overly liberal Overly conservative
Range of reported outcome Range of mean illness rate frequency for for Chicago beaches Chicago beaches (swimmers/1000) 49e78% 3e14% 12e21% 9e17%
2.7e5.0 13.6e15.3 14.4e16.1 4.2e6.4
1662
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
source (Calderon et al., 1991; Sinigalliano et al., 2010). Questions about the appropriateness of using E. coli as an indicator have also been raised (Harwood et al., 2005; Wong et al., 2009), but a review of epidemiological studies associate elevated E. coli with higher risk of gastrointestinal illness (Marion et al., 2010; Wade et al., 2003). Because alternate standards have not been established, all beaches are managed using the E. coli standard, despite its shortcomings. Chicago does not issue a swim ban until E. coli concentration exceeds 1000 MPN/100 ml. Currently, there are no estimates of the percent of swimmers who enter the water during a swimming advisory (i.e., 235 < E. coli <1000) (Hou et al., 2006). Here, we elect to use >235 to express excess illnesses and define model errors because it is the more conservative calculation. Rate of excess illness would increase by 0.0059% using the 1000 CFU standard.
2.5. Different management applications of monitoring standard In an attempt to improve model performance, the E. coli standard for the binary outcome was targeted. As described previously, monitoring standards recommended a five-day geometric mean, and a general single-sample maximum was provided (US EPA, 1986). The guidance suggested, however, that single-sample maximums be calculated for individual jurisdictions based on local log standard deviation of E. coli concentration (US EPA, 1986). The calculation provided in the ambient criteria (US EPA, 1986) is ss ¼ 10^ log10 GM þ cl log10 sd
14/day. Overall, for the 14 beaches, a mean of 35 illnesses/day could be expected (range: 0e659 66 SD).
3.2.
Amount of E. coli variation explained by models (R2)
Coefficient of determination was lowest for the CM, with a range of 0.049 (South Shore) to 0.135 (63rd Street) for individual beaches and an overall adjusted R2 of 0.141 (Fig. 2a). The higher R2 for 63rd Street results from the persistent high E. coli concentrations at this location with little variation. Generally, more variation in E. coli concentrations could be explained using the RM, resulting in a range of R2 from 0.111 (63rd Street) to 0.287 (North Avenue). Further improvement in R2 was seen with the IM (Fig. 2a). The model for 63rd Street included the lowest amount of variation explained (0.141); the highest adjusted R2 was at North Avenue (0.313). The general pattern was an increasing R2 with increasing model refinement, seen at all beaches except 63rd Street. The change in R2 between the CM and the RM was greatest for Hollywood and South Shore beaches, with notable improvements at Foster, North Avenue and South Shore. Change in R2 between RM and IM was not as great overall; most improvement was at the south side beaches. Precision of the predictions increased with model refinement as seen in a decreasing RMSE.
3.3. Binary water quality standard outcome (number of prediction errors) Using the CM, the beach was correctly left open to swimming 68% of the time (range ¼ 49e78%) (Fig. 3). This increased to 78%
(3)
where ss ¼ the single-sample maximum; GM ¼ 126, geometric mean E. coli concentration for acceptable illness rate of 8 per 1000; cl ¼ 0.675, the 75% calculated one-sided confidence level for a designated heavily used beach area, and sd ¼ is the calculated standard deviation for a given jurisdiction, 0.718 for Chicago beaches (mean log E. coli ¼ 1.776) Chicago’s singlesample maximum would therefore be 385 CFU/100 ml.
3.
Results
Visits to Chicago’s lakefront beaches are often in excess of 27,000,000 annually, with fully half of the visits associated with two beaches: Oak Street and North Avenue (Chicago Park District, unpublished data).
3.1.
Baseline water quality
Chicago beaches exceed the 235 single-sample maximum 14e35% of the time, with an overall rate of 20%. Given the current water quality and visitation of Chicago’s beaches and the hypothetical absence of any monitoring or associated beach closures (IA model), the majority of expected illnesses were associated with the high visitation beaches (Fig. 1). The highest illness rate was 138/day for North Avenue, and the lowest was 7/day for Rainbow. Despite having the highest mean E. coli concentration, 63rd Street illness rate averaged
Fig. 1 e Comparison of mean E. coli concentration at each beach and mean number of swimmers that can be estimated to develop gastrointestinal illness each day. Gradation of circles indicates the expected individual illness rate. Beaches include Loyola (LY), Albion (AL), Hollywood (HW), Foster (FO), Montrose (MO), North Avenue (NA), Oak (OK), 12th Street (12), 31st Street (31), 57th Street (57), 63rd Street (63), South Shore (SS), Rainbow (RB), and Calumet (CA).
1663
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
Fig. 2 e Comparison of (a) coefficient of variations, (b) percent type I errors, and (c) percent type II errors by modeling approach.
(range ¼ 53e85%) with the RM and was 77% (range ¼ 54e84%) with the IM. Rate of correct closures under the CM is 6% (range ¼ 3e14%). That decreased to 0.8% (0.3e9) with the RM, and with the IM, the rate was 1.4% (0.3e11%). Type I errors, in which a swimming advisory is posted although E. coli concentration is lower than the single-sample maximum, were made 11.5% (9e17%) of the time using the CM for monitoring (Fig. 2b). With the application of an RM, that
inactive
rate decreased to 2.2% (0.4e4) of the time, and with the IM application, the rate was 3.2% (1e11%). Type II errors were common under the CM, occurring 14% (12e21%) (Fig. 2c). That number increased with both predictive models, to 19% (16e27) with the RM and 18.5% (14e24) with the IM. Comparison of the percent of type II errors showed highly variable results. The greatest percent change between models was an increase in percent type II errors at 12th Street and
current
outcome correct closed correct open type I error type II error
region
individual
Fig. 3 e Management outcome using the binary approach, i.e., an E. coli concentration >235 results in a beach closure.
1664
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
Loyola between CM and RM. Minimizing type II errors between the RM and IM was limited.
3.4. Association of monitoring and modeling approaches with illness burden A comparison of the four models reveals, most noticeably at the high visitation beaches North Avenue, Montrose and Oak Street, that the leveling off of number of type II errors was mirrored in the rate of excess illnesses prevented with increasing model refinement. There was no significant difference in rate of excess illnesses associated with model approach at any of the individual beaches. In fact, model refinement between RM and IM showed little improvement in the number of excess illnesses prevented. At several beaches, cumulative illnesses were identical for all models except for CM. Eight of the beaches showed a difference in cumulative excess illnesses associated with the four models (Table 2). The CM was associated with lower cumulative excess illnesses, largely because there were fewer instances of type II errors; however, for all beaches but 63rd Street, the CM was associated with higher mean excess illnesses, indicating that with this model, the instances of very high E. coli concentrations tend to be missed. The IA model excluded type I errors because the beach would never be closed to swimming and therefore swimmers would not be prevented from water contact under any water quality conditions. Increasingly refined models were effective at reducing type I errors because they rarely predicted events when water quality exceeded the single-sample standard; most predictions were low E. coli concentrations. This tendency is apparent in the rapid decrease in number of type I errors with increasing model refinement.
3.5.
Chicago-specific monitoring standard
With the more liberal, jurisdiction-specific single-sample maximum (385 CFU/100 ml), the overall percentage of errors for all model types decreased (Fig. 4). Reduction in percent type II errors was more pronounced with the RM and IM (mean decrease by 35e36%) than for the CM (mean decrease by 22%). Dramatic reductions in percent of type II errors (>45%) occurred at some of the beaches with lower mean E. coli concentrations (North Avenue, Oak, Albion) with the RM and IM. The difference in number of illnesses should theoretically
remain unchanged because the illness rate is within the confidence limits of the original derived regression. Because the predictive models tend to predict low overalldthe tendency leading to type II errorsdthe reduction in type I errors was not as remarkable.
4.
Discussion
With recent findings that the time-intensive current beach monitoring models are not generally predictive of real-time FIB concentration (Boehm, 2007; Whitman et al., 1999), beach managers need a means of determining water quality rapidly and efficiently in order to protect public health while maximizing beach access. A suite of approaches for expanding monitoring activities and improving timeliness of monitoring results have been proposed and considered, but minimal research has explored their capacity to meet this goal. If a new method does not provide a significant improvement over current management outcomes, it is unlikely that managers will invest the necessary effort to alter the monitoring program. The first step for many jurisdictions is to implement a water quality monitoring program using the currently used model (CM). According to our results, this activity alone decreased the number of excess illnesses at all beaches, largely as a result of keeping swimmers out of the water more often, regardless of actual water quality. Beaches were closed when water quality was within acceptable standards 9e17% of the time, a scenario that can incur a social and economic burden (Rabinovici et al., 2004). With this model, accurate closures and health protection are strictly dependent on the endurance of an E. coli contamination event. With the rapidly changing nature of water quality (Boehm, 2007; Boehm et al., 2002; Whitman and Nevers, 2004), there is high potential to make a management error (Table 1). Multiple day contamination events would result in some correct swimming advisories (correct closed), and extended periods of low E. coli concentrations would result in accurate swimming permission (correct open). The vast majority of contamination events, however, last one day or less (Leecaster and Weisberg, 2001). In Chicago, only 6% of contamination events persisted for more than one sampling day, with the exception of 63rd Street: 17%. Monitoring at beaches with persistent contamination will inevitably decrease the number of type II errors and associated illnesses because swimmers are
Table 2 e Comparison of the cumulative (mean) excess illnesses as a result of the choice of model approach: inactive, current, regional, or individual beach model. Results based on days of type II errors, ranging from 38 to 103 days of the 247e292 days (2000e2004 seasons) considered in the analysis. Inactive Montrose North Ave 31st 57th 63rd South Shore Rainbow Calumet
2492.94 4540.82 366.22 912.68 704.59 551.00 240.73 1049.60
(42.25) (105.60) (5.39) (16.90) (6.84) (10.20) (3.70) (20.58)
Current 1695.49 (42.39) 3659.66 (107.64) 250.73 (5.57) 758.64 (17.64) 389.27 (6.38) 456.19 (11.13) 176.09 (3.91) 869.01 (22.87)
Regional 2407.17 4540.82 348.00 912.51 548.65 551.00 240.73 1049.32
(41.50) (105.60) (5.52) (17.55) (6.94) (10.20) (3.70) (20.99)
Individual 2442.56 4270.00 320.70 894.31 490.04 537.40 217.60 1023.04
(42.85) (101.69) (5.34) (17.54) (6.71) (10.14) (3.69) (20.88)
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
1665
Fig. 4 e Impact on beach advisory decisions if jurisdiction-specific water quality standard is used. The two beach examples include 63rd Street Beach, which has a high mean E. coli, lower visitation rate, and frequent swimming advisories, and North Avenue Beach, which is the most highly visited Chicago beach and has low mean E. coli and infrequent swimming advisories. Use of a Chicago-specific water quality standard would result in fewer errors overall using all three monitoring model approaches.
accurately kept from the water for some percent of the high concentration events. Likewise, beaches with infrequent contamination events will have a limited number of type I errors. To overcome these inconsistencies in actual vs. presumed water quality, predictive models have been used at some beaches to provide a real-time estimate of water quality and overcome these inherent errors. The cornerstone of predictive modeling for beach management has been the individual beach model (Nevers and Whitman, 2005), but the cost of developing such models has been one of the disincentives to widespread use; many of the models use on-site instrumentation that requires installation and continuous maintenance (Francy, 2009). For this reason, regional models have been explored, which examine the hydrometeorological factors similarly affecting groups of beaches and describe background fluctuations in FIB concentrations (Nevers and Whitman, 2008). While more crude in estimations, regional models may provide a cost-effective solution for jurisdictions with numerous monitored beaches while providing insights into source behavior. Perhaps surprising, the RM results in this exercise were quite similar to results from IM, indicating a generally predictable fluctuation in FIB concentrations across Chicago. Overall, increasing refinement of monitoring approach with the use of predictive models was associated with improved accuracy of E. coli predictions. Use of the RM increased the amount of variation explained over the current monitoring approach, and the use of beach-specific IM somewhat further improved this result. With beach-specific refinement, predictive models for all of the study beaches had higher R2 and lower RMSE: more variation in individual E. coli concentrations was
described and there was lower error in this estimation. In this study, geographically widespread predictors were used, which may have limited each model’s ability to detect beach-specific variation, although the low R2 for 63rd Street was somewhat expected due to inherent high variation in E. coli and the complex circulation pattern at this enclosed beach (Ge et al., 2010; Whitman and Nevers, 2004). Models developed for this beach have depended on higher frequency and higher intensity local data than were available for this exercise (Boehm et al., 2007; Olyphant and Whitman, 2004). Predictability improved significantly with the use of the RM for many of the beaches, with the biggest improvement between CM and RM at the north-side beaches, but it was the further refinement to IM that resulted in the greatest improvement at southern beaches. This pattern supports the idea that there is more beach-specific variation at these southern beaches and E. coli concentrations perhaps recover to background concentrations more slowly than at the northern beaches (Whitman and Nevers, 2008). Use of individual predictive models at these beaches can take these factors into account, resulting in better predictability. The number of type II errors for each model was highly variable, lacking the pattern of improvement with increasing refinement shown in the R2. All of the models failed to predict the majority of high E. coli concentrations. Because many beaches have infrequent high E. coli concentration events, it is difficult to detect a pattern of associated hydrometeorological conditions; this is a problem for many predictive modeling attempts but is an important characteristic for eliminating type II errors. The CM generally had the lowest number of type II errors, likely simply because this monitoring approach
1666
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
closes the beach more often, an action that also results in significantly more type I errors (Fig. 3). Calculated excess illnesses were far less variable between models, indicating that more advanced predictive models do not necessarily provide improved public health protection. IM had a lower number of excess illnesses at some beaches (Table 2), but the difference was not nearly as great as might be expected. The results imply that IM detect some instances of extreme high concentrations at these beaches but generally miss these events. The majority of beaches showed a pattern of identical number of estimated cumulative illnesses for all model approaches with the exception of the CM: it was associated with the lowest cumulative excess illnesses specifically because this approach limits overall exposure to the beach water. The CM was therefore associated with lowest excess illnesses due to more closed beaches (both type I errors and correct closed).
4.1.
Application of monitoring standards
The number of excess illnesses associated with a given model is affected by the calculation of the illness rate, and although jurisdictions have the option to calculate single-sample maximums specific to their beach waters, most Great Lakes states use the default 235 CFU/100 ml standard developed using data from the original epidemiological studies for all beaches statewide (Nevers and Whitman, 2010). Use of a higher Chicagospecific standard increases swimming access without impacting illness rates by expanding the range of allowable water contact and therefore the number of accurate predictions (correct open). Percent of type II errors was reduced under the CM and drastically reduced using RM or IM. This was particularly noticeable at beaches with lower mean E. coli concentrations: upwards of 40% reduction in percent type II errors for North Avenue, Montrose, and Oak Street with RM and IM. This significant reduction is mirrored in the potential to significantly reduce the number of illnesses at these high visitation beaches; illness rates are greatly elevated during a single high FIB event at a high use beach (Turbow et al., 2003). The use of the Chicagospecific standard did not greatly reduce the number of type II errors for 63rd Street because of the higher overall mean E. coli concentration. Within the framework of the original monitoring standards, leeway is provided for level of beach use, beach-specific variation in bacteria concentrations, and calculation of the overall water quality (US EPA, 1986), and health protection is assumed to be equally provided over a range of calculated concentrations. Considering this broader range of confidence could increase beach use without influencing health outcome under a variety of monitoring approaches, including the predictive models presented here. These results indicate that the use of a higher standard, along with a predictive model could maximize access at many of the Chicago beaches without increasing public health risk. The monitoring standards recommend use of the 5-day geometric mean, but most managers opt for the single-sample maximum, likely due to ease of use. An examination of Chicago beach monitoring data reveals that use of the 5-day geometric mean results in fewer errors than either the 235 single-sample maximum or the 385 Chicago-specific standard presented here. However, the number of days exceeding the
specified limits (i.e., swimming advisories) increases significantly with use of the geometric mean. The 5-day geometric mean was developed based on studies at beaches influenced by point sources (US EPA, 1986), areas more likely to have persistent high E. coli concentration events; river discharge or sewage releases may create periods of sustained high E. coli concentrations, warranting the extended swimming advisory that results from a running geometric mean. Chicago’s beaches, however, are not influenced by a major point source except during rare events of sewage overflows, during which beaches are preemptively closed for several days.
4.2.
Maximizing public health protection
Predictive modeling results indicate that this monitoring approach would not improve health protection at all Chicago beaches. The best approach for monitoring may differ between beaches, even within an individual jurisdiction such as Chicago. The threshold for level of effort associated with increased model refinement would have to be considered for each beach, perhaps while incorporating economic considerations. Hou et al. (2006) determined that different monitoring policies provided the optimal economic and public health outcomes for each of two beaches. The application of different monitoring strategies may include combining approaches or extending to rapid methods, alternate indicators, or diverse management plans. Recent research to characterize the sources, survival, fate, and transport of FIB and the applications of monitoring programs has perhaps complicated the applicability of different monitoring and management strategies by indicating that one monitoring approach does not fit all beach types. Novel management approaches have included predictive models (Frick et al., 2008; Nevers and Whitman, 2005), rapid tests (Bushon et al., 2009; Lavender and Kinzelman, 2009), new indicators, including hostspecific markers (Bacteroides, Methanobrevibacter, virulence markers etc), and gene-based detection techniques for human pathogens (Griffith et al., 2009). Reconsiderations of monitoring standards have also been explored (Kim and Grant, 2004; Nevers and Whitman, 2010). While epidemiological studies have linked illness rates with outcomes from some of these new analyses (Wade et al., 2006), care will have to be taken to consider whether environmental conditions and sources influence the results outcomes. The ideal method for beach management may differ among beaches and jurisdictions, and it may be desirable to have a variety of monitoring approaches available to beach managers that increase accuracy and reliability at given beaches. In deciphering the best management plans for different beaches, efforts should focus on improving public health protection, perhaps considering a wide variety of available monitoring options.
5.
Conclusions
Refinement of monitoring models generally increased predictability of E. coli but did not necessarily result in fewer errors or excess illnesses Regional models provided similar levels of accuracy as individual beach models in many locations
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
Use of a location-specific water quality standard, combined with empirical predictive models, provided the greatest beach access without sacrificing public health protection
Acknowledgments We thank Murulee Byappanahalli (USGS) for his careful review. Research was funded in part by the US Ocean Action Plan: USGS Ocean Research Priorities Plan and by the Great Lakes Restoration Initiative. Any use of trade, product, or firm names is for descriptive purposes only and does not imply endorsement by the U.S. Government. This article is Contribution 1627 of the USGS Great Lakes Science Center.
references
Boehm, A.B., 2007. Enterococci concentrations in diverse coastal environments exhibit extreme variability. Environmental Science & Technology 41 (24), 8227e8232. Boehm, A.B., Grant, S.B., Kim, J.H., Mowbray, S.L., McGee, C.D., Clark, C.D., Foley, D.M., Wellman, D.E., 2002. Decadal and shorter period variability of surf zone water quality at Huntington Beach, California. Environmental Science & Technology 36 (18), 3885e3892. Boehm, A.B., Whitman, R.L., Nevers, M.B., Hou, D., Weisberg, S.B., 2007. Statistical Framework for Recreational Water Quality Criteria and Monitoring. In: Wymer, L.J. (Ed.). Wiley, West Sussex, UK, pp. 179e210. Buckalew, D.W., Hartman, L.J., Grimsley, G.A., Martin, A.E., Register, K.M., 2006. A long-term study comparing membrane filtration with Colilert defined substrates in detecting fecal coliforms and Escherichia coli in natural waters. Journal of Environmental Management 80 (3), 191e197. Bushon, R.N., Likirdopulos, C.A., Brady, A.M.G., 2009. Comparison of immunomagnetic separation/adenosine triphosphate rapid method to traditional culture-based method for E. coli and enterococci enumeration in wastewater. Water Research 43 (19), 4940e4946. Calderon, R.L., Mood, E.W., Dufour, A.P., 1991. Health effects of swimmers and nonpoint sources of contaminated water. International Journal of Environmental Health Research 1, 21e31. Dufour, A.P., 1984. Health Effects Criteria for Fresh Recreational Waters EPA Report EPA-600/1-84-004, Cincinnati, Ohio. Dwight, R.H., Brinks, M.V., SharavanaKumar, G., Semenza, J.C., 2007. Beach attendance and bathing rates for Southern California beaches. Ocean & Coastal Management 50 (10), 847e858. Francy, D.S., 2009. Use of predictive models and rapid methods to nowcast bacteria levels at coastal beaches. Aquatic Ecosystem Health and Management 12 (2), 177e182. Frick, W.A., Ge, Z., Zepp, R.G., 2008. Nowcasting and forecasting concentrations of biological contaminants at beaches: a feasibility and case study. Environmental Science & Technology 42, 4218e4824. Ge, Z., Nevers, M.B., Schwab, D.J., Whitman, R.L., 2010. Coastal loading and transport of Escherichia coli at an embayed beach in Lake Michigan. Environmental Science & Technology 44, 6731e6737. Given, S., Pendleton, L.H., Boehm, A.B., 2006. Regional public health cost estimates of contaminated coastal waters: a case
1667
study of gastroenteritis at southern California beaches. Environmental Science & Technology 40 (16), 4851e4858. Griffith, J.F., Cao, Y., McGee, C.D., Weisberg, S.B., 2009. Evaluation of rapid methods and novel indicators for assessing microbiological beach water quality. Water Research 43 (19), 4900e4907. Gronewold, A.D., Borsuk, M.E., Wolpert, R.L., Reckhow, K.H., 2008. An assessment of fecal indicator bacteria-based water quality standards. Environmental Science & Technology 42 (13), 4676e4682. Harwood, V.J., Levine, A.D., Scott, T.M., Chivukula, V., Lukasik, J., Farrah, S.R., Rose, J.B., 2005. Validity of the indicator organism paradigm for pathogen reduction in reclaimed water and public health protection. Applied and Environmental Microbiology 71 (6), 3163e3170. He, L.M., He, Z.L., 2008. Water quality prediction of marine recreational beaches receiving watershed baseflow and stormwater runoff in southern California, USA. Water Research 42 (10e11), 2563e2573. Hou, D., Rabinovici, S.J.M., Boehm, A.B., 2006. Enterococci predictions from partial least squares regression models in conjunction with a single-sample standard improve the efficacy of beach management advisories. Environmental Science & Technology 40 (6), 1737e1743. Kim, J.H., Grant, S.B., 2004. Public mis-notification of coastal water quality: a probabilistic evaluation of posting errors at Huntington Beach, California. Environmental Science & Technology 38 (9), 2497e2504. Lavender, J.S., Kinzelman, J.L., 2009. A cross comparison of QPCR to agar-based or defined substrate test methods for the determination of Escherichia coli and enterococci in municipal water quality monitoring programs. Water Research 43 (19), 4967e4979. Leecaster, M.K., Weisberg, S.B., 2001. Effect of sampling frequency on shoreline microbiology assessments. Marine Pollution Bulletin 42 (11), 1150e1154. Lepesteur, M., McComb, A.J., Moore, S.A., 2006. Do we all face the same risk when bathing in the estuary? Water Research 40 (14), 2787e2795. Marion, J.W., Lee, J., Lemeshow, S., Buckley, T.J., 2010. Association of illness and recreational water exposure during advisory and non-advisory conditions at an inland US beach. Water Research 44 (16), 4796e4804. McPhail, C.D., Stidson, R.T., 2009. Bathing water signage and predictive water quality models in Scotland. Aquatic Ecosystem Health and Management 12 (2), 183e186. Nevers, M.B., Whitman, R.L., 2005. Nowcast modeling of Escherichia coli concentrations at multiple urban beaches of southern Lake Michigan. Water Research 39 (20), 5250e5260. Nevers, M.B., Whitman, R.L., 2008. Coastal strategies to predict Escherichia coli concentrations for beaches along a 35 km stretch of southern Lake Michigan. Environmental Science & Technology 42 (12), 4454e4460. Nevers, M.B., Whitman, R.L., 2010. Policies and practices of beach monitoring in the Great Lakes, USA: a critical review. Journal of Environmental Monitoring 12 (3), 581e590. Olyphant, G.A., Whitman, R.L., 2004. Elements of a predictive model for determining beach closures on a real time basis: the case of 63rd Street Beach Chicago. Environmental Monitoring and Assessment 98 (1e3), 175e190. Pru¨ss, A., 1998. Review of epidemiological studies on health effects from exposure to recreational water. International Journal of Epidemiology 27 (1), 1e9. Rabinovici, S.J.M., Bernknopf, R.L., Wein, A.M., Coursey, D.L., Whitman, R.L., 2004. Economic and health risk trade-offs of swim closures at a Lake Michigan Beach. Environmental Science & Technology 38 (10), 2737e2745.
1668
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
Seyfried, P.L., Tobin, R.S., Brown, N.E., Ness, P.F., 1985. A prospective study of swimming-related illness I. Swimmingassociated health risk. American Journal of Public Health 75 (9), 1068e1070. Sinigalliano, C.D., Fleisher, J.M., Gidley, M.L., Solo-Gabriele, H.M., Shibata, T., Plano, L.R.W., Elmir, S.M., Wanless, D., Bartkowiak, J., Boiteau, R., Withum, K., Abdelzaher, A.M., He, G., Ortega, C., Zhu, X., Wright, M.E., Kish, J., Hollenbeck, J., Scott, T., Backer, L.C., Fleming, L.E., 2010. Traditional and molecular analyses for fecal indicator bacteria in nonpoint source subtropical recreational marine waters. Water Research 44 (10), 3763e3772. Turbow, D.J., Osgood, N.D., Jiang, S.C., 2003. Evaluation of recreational health risk in coastal waters based on enterococcus densities and bathing patterns. Environmental Health Perspectives 111 (4), 598e603. US EPA, 1986. Ambient Water Quality Criteria for Bacteria 1986, p. 18. US EPA, Office of Water Regulations and Standards, Washington DC. Wade, T.J., Calderon, R.L., Sams, E., Beach, M., Brenner, K.P., Williams, A.H., Dufour, A.P., 2006. Rapidly measured indicators of recreational water quality are predictive of swimming associated gastrointestinal illness. Environmental Health Perspectives 114 (1), 24e28. Wade, T.J., Pai, N., Eisenberg, J.N.S., Colford Jr., J.M., 2003. Do U.S. Environmental Protection Agency water quality guidelines for recreational waters prevent gastrointestinal illness? A
systematic review and meta-analysis. Environmental Health Perspectives 111 (8), 1102e1109. Whitman, R.L., Nevers, M.B., 2003. Foreshore sand as a source of Escherichia coli in nearshore water of a Lake Michigan beach. Applied and Environmental Microbiology 69 (9), 5555e5562. Whitman, R.L., Nevers, M.B., 2004. Escherichia coli sampling reliability at a frequently closed Chicago beach: monitoring and management implications. Environmental Science & Technology 38 (16), 4241e4246. Whitman, R.L., Nevers, M.B., 2008. Summer E. coli patterns and responses along 23 Chicago beaches. Environmental Science & Technology 42 (24), 9217e9224. Whitman, R.L., Nevers, M.B., Gerovac, P.J., 1999. Interaction of ambient conditions and fecal coliform bacteria in southern Lake Michigan waters: monitoring program implications. Natural Areas Journal 19, 166e171. Whitman, R.L., Shively, D.A., Pawlik, H., Nevers, M.B., Byappanahalli, M.N., 2003. Occurrence of Escherichia coli and enterococci in Cladophora (Chlorophyta) in nearshore water and beach sand of Lake Michigan. Applied and Environmental Microbiology 69 (8), 4714e4719. Wong, M., Kumar, L., Jenkins, T.M., Xagoraraki, I., Phanikumar, M.S., Rose, J.B., 2009. Evaluation of public health risks at recreational beaches in Lake Michigan via detection of enteric viruses and a human-specific bacteriological marker. Water Research 43, 1137e1149.
Available at www.sciencedirect.com
journal homepage: www.elsevier.com/locate/watres
Efficacy of monitoring and empirical predictive modeling at improving public health protection at Chicago beaches Meredith B. Nevers*, Richard L. Whitman U.S. Geological Survey, Great Lakes Science Center, Lake Michigan Ecological Research Station, 1100 N. Mineral Springs Road, Porter, IN 46304, USA
article info
abstract
Article history:
Efforts to improve public health protection in recreational swimming waters have focused
Received 17 August 2010
on obtaining real-time estimates of water quality. Current monitoring techniques rely on
Received in revised form
the time-intensive culturing of fecal indicator bacteria (FIB) from water samples, but
3 December 2010
rapidly changing FIB concentrations result in management errors that lead to the public
Accepted 6 December 2010
being exposed to high FIB concentrations (type II error) or beaches being closed despite
Available online 13 December 2010
acceptable water quality (type I error). Empirical predictive models may provide a rapid solution, but their effectiveness at improving health protection has not been adequately
Keywords:
assessed. We sought to determine if emerging monitoring approaches could effectively
E. coli
reduce risk of illness exposure by minimizing management errors. We examined four
Fecal indicator bacteria
monitoring approaches (inactive, current protocol, a single predictive model for all bea-
Recreational water quality
ches, and individual models for each beach) with increasing refinement at 14 Chicago
Lake Michigan
beaches using historical monitoring and hydrometeorological data and compared
Swimming
management outcomes using different standards for decision-making. Predictability (R2) of
Risk
FIB concentration improved with model refinement at all beaches but one. Predictive models did not always reduce the number of management errors and therefore the overall illness burden. Use of a Chicago-specific single-sample standarddrather than the default 235 E. coli CFU/100 ml widely useddtogether with predictive modeling resulted in the greatest number of open beach days without any increase in public health risk. These results emphasize that emerging monitoring approaches such as empirical models are not equally applicable at all beaches, and combining monitoring approaches may expand beach access. Published by Elsevier Ltd.
1.
Introduction
In recent years, efforts to improve public health protection in recreational swimming waters have focused on obtaining realtime estimates of water quality. Current monitoring techniques rely on the culturing of fecal indicator bacteria (FIB)d
such as Escherichia coli or enterococcidfrom water samples, a process that requires an incubation time often in excess of the rate of change of bacteria concentrations in the water (Boehm et al., 2002; Whitman et al., 1999). Because of the lapse in results availability, the public are often either unknowingly swimming in contaminated beach water or are prohibited from
Abbreviations: FIB, fecal indicator bacteria; CFU, colony-forming units; MPN, most probable number; IA, inactive monitoring program model; CM, current model; RM, regional predictive model for all study beaches; IM, individual beach predictive model. * Corresponding author. Tel.: þ1 219 926 8336x425; fax: þ1 219 929 5792. E-mail address: [email protected] (M.B. Nevers). 0043-1354/$ e see front matter Published by Elsevier Ltd. doi:10.1016/j.watres.2010.12.010
1660
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
swimming in water that meets the public health criteria. Efforts have been focused on two means of correcting this shortcoming: shorten the analytical time for the current indicator or find an alternate, faster way to assess water quality. To accomplish the latter, empirical predictive models have been attempted with various levels of success and application. Predictive models have been suggested by numerous authors as a potential means for minimizing errors in beach closings (Hou et al., 2006; Kim and Grant, 2004; Nevers and Whitman, 2005). These models range from simple models that associate weather conditions with direct bacteria loadingsesuch as rainfall and associated runoff (McPhail and Stidson, 2009) e to more advanced models that integrate multiple hydrometeorological variables (Kim and Grant, 2004). Model accuracy at predicting FIB concentration depends on beach location, instrument accuracy, wealth of available data, and level of effort, but predictive models can be successfully incorporated into beach management (Nevers and Whitman, 2005). Beaches at which models have been attempted tend to be high profile beaches with heavy visitor use (Boehm, 2007; Hou et al., 2006), directly or strongly impacted by a large point source (He and He, 2008), or having frequent swimming advisories (Olyphant and Whitman, 2004). The accuracy or success of a given modeling approach has typically been assessed by analyzing the amount of variation in the target FIB explained by the model, the error explained by the model, or the specificity (the percent of false negatives, or type II errors) and sensitivity (the percent of false positives, or type I errors) of the model. The first two calculations determine the accuracy of the model at predicting all FIB concentrations. The third, error-based calculation is used due to the use of a binary approach in beach management policies: beaches are either open or closed to swimming, depending on where the FIB concentration falls relative to a designated standard (acceptable health risk); errors occur when the predicted concentration is not equal to actual concentration. Errors result in either inadvertent exposure of the public to high concentrations of FIB (type II error) or exclusion of swimmers from water that meets the exposure standard (type I error). More type II errors result in more swimmers exposed to high concentrations of FIB and therefore a higher public health risk; decreasing the instances of type II errors is necessary to increase public health protection. Current water quality standards for freshwater were developed using epidemiological studies and based on historical acceptable illness rate (Pru¨ss, 1998, US EPA, 1986). Within the monitoring guidance, however, some measure of flexibility was provided for beach managers, including choice of application of two mathematical estimates of illness risk, based on the concentration of indicator bacteria (US EPA, 1986). Generally, beach managers have applied the single-sample maximum for an individual water sample because of its ease of use and interpretation (Nevers and Whitman, 2010, US EPA, 1986), but others use the 5-day geometric mean, both of which should theoretically provide equal levels of health protection. In this paper, we examine four potential monitoring approaches with increasing refinement at 14 Chicago beaches: inactive, current monitoring model, use of one predictive model for all beaches, and use of individual predictive models for each beach. Further, we examine alternate applications of
monitoring standards under these four approaches to assess the health and management outcome possibilities. Using historical monitoring and beach attendance data we compare the accuracy of each model with several calculations and also the relative public health protection provided by each of these models. Specifically, we sought to determine whether predictive modeling at Chicago beaches could be used as a monitoring tool to increase public health protection over traditional monitoring practices.
2.
Materials and methods
2.1.
Study site
Chicago beaches in general are not impacted by a major point source of contamination. Urban sewage is regularly discharged through the Chicago River and a series of man-made or modified canals to the Mississippi River. In events of extreme precipitation, the system override leads to sewage being directed to Lake Michigan (<1 per year); all beaches are then preemptively closed to swimming. Sources of FIB at the Chicago beaches are unknown but likely include beach sand, birds, and algae (Whitman and Nevers, 2003; Whitman et al., 2003). Beaches included in the current study were (from north to south) Loyola, Albion, Hollywood, Foster, Montrose, North Avenue, Oak, 12th Street, 31st Street, 57th Street, 63rd Street, South Shore, Rainbow, and Calumet.
2.2.
Beach monitoring data
E. coli monitoring data, measured as most probable number (MPN)/100 ml of water, were obtained from the Chicago Park District for 2000e2004. Beaches were sampled at least five days a week; replicate samples (up to three) were averaged. E. coli concentrations above or below detection limits were set at detection limits after determining that occurrences were rare (Boehm et al., 2002; Whitman and Nevers, 2008). Missing data points for individual beaches were calculated; values were estimated from the nearest 6 values (average of three previous and three subsequent readings). Beach management is a binary decision: if E. coli concentration <235 MPN/100 ml, the beach is open for swimming; if E. coli >235 a swimming advisory is issued. This model assumes that E. coli concentration today ¼ E. coli concentration yesterday. Inaccurate predictions, therefore, result in a type I or type II error (Table 1). Type I errors occur when the model predicts an E. coli concentration >235 when the actual concentration is <235, resulting in an unnecessary swimming advisory. A type II error occurs when the model predicts E. coli concentration <235 when the actual concentration is >235, resulting in swimmers being exposed to high concentrations of indicator bacteria and associated pathogens. A simple characterization is that type I errors are associated with economic losses because swimmers are denied access and type II errors are associated with greater public health risk, as swimming occurs in the presence of excessive FIB concentrations (Rabinovici et al., 2004).
1661
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
2.3.
Predictive models
Data for developing predictive models were accumulated from several existing hydrometoerological stations, as described in Whitman and Nevers (2008). Predictors included solar insolation and precipitation (24 h); air temperature, barometric pressure, and wave height (mean for 4e10 AM); and day of year. All E. coli results were log-transformed prior to analysis. Predictor parameters were transformed to z-scores prior to regression analyses. Models were developed using multiple linear regression of the available predictors on the independent variable E. coli; Mallows Cp was used to compare resulting models. For individual models, Akaike’s Information Criterion (AIC) was used to select the best fit model for each individual beach. The regional model was developed previously (Whitman and Nevers, 2008) using barometric pressure, wave height, and day of year as the predictors; here, it was applied separately to each study beach. Model comparisons were made using SPSS (Chicago, IL) and SAS (Cary, NC) software: the adjusted coefficient of determination (R2), which describes the amount of variation in E. coli; and percent of type I and type II errors. The four models compared included inactive model (IA): beach is always open to swimming, regardless of microbiological water quality; current model (CM): E. coli concentration for day 1 is used to make a management decision for day 2; regional model (RM): a universal model used for all beaches; and an individual model (IM): a separate predictive model developed for each beach. Comparisons between models for estimating excess illnesses associated with type II errors were only made on days for which there were results for all four models.
2.4.
Calculating illness rates
Overall beach attendance rates were compared for 20 years, and an average per year estimate was calculated (Chicago Park District, unpublished data). With an annual reported range of 14e31 million visitors to all Chicago beaches, 20 million was used as an average. An estimated 91% of all visits, or 18.2 million, were associated with the 14 beaches included in this study. Number of visitors per beach was calculated based on two years for which individual beach data were available (1999 and 2007). It was assumed that 50% of beach visitors had full-body contact with the water as defined by US EPA (Dufour, 1984).
Lower estimates have been published (Rabinovici et al., 2004), but based on data from several Ontario lakes (Seyfried et al., 1985), estuaries (Lepesteur et al., 2006), and marine waters (Dwight et al., 2007; Given et al., 2006), 50% may be conservative for a large freshwater lake. Illness rates for each beach were calculated based on a 100-day swimming season, which assumes equal distribution of visits across the summer (Rabinovici et al., 2004). Illness rate (Y ) was calculated from Dufour (1984): Y ¼ 11:74 þ 9:397log10 ðECÞ
(1)
where Y ¼ the rate of swimming-associated gastrointestinal illness symptoms per 1000 swimmers and EC ¼ E. coli CFU/ 100 ml water. Monitoring standards based on the these epidemiological studies recommend a geometric mean of 126 E. coli CFU/100 ml for five samples over 30 days (Dufour, 1984): an acceptable illness rate of approximately 0.008%. A singlesample limit of 235 CFU/100 ml is also provided, which is within the confidence limits of the calculated geometric mean. These calculations were developed for beaches influenced by a point source (Dufour, 1984), but the sources affecting Chicago’s beaches have not been confidently identified. Using the derived regression equation (Dufour, 1984), we calculated our acceptable illness rate for the 235 CFU as 0.01054%, following Rabinovici et al. (2004). It should be noted that the standards were established using membrane filtration analysis. Chicago uses a defined substrate technique (Colilert; IDEXX, Westbrook, Maine); and while studies have favorably compared the two outcomes (Buckalew et al., 2006), differences in confidence intervals may influence results outcome (Gronewold et al., 2008). A daily excess illness rate was calculated following Given et al. (2006), that is, the number of beach swimmers expected to exhibit symptoms of illness beyond the acceptable 0.01054%. GI ¼ ðY Y0 Þðv=dÞf
(2)
Where Y0 ¼ acceptable illness rate within the monitoring criteria of 10.54/1000, v ¼ number of visitors in a swimming season, d ¼ number of swimming days in the season, and f ¼ percent of beach visitors estimated to have full-body contact with the beach water. The illness regression equation has a y-intercept <0, so calculated illness rates <0 were set at 0 for calculating the mean. There is conflicting research on the association of illness risk and traditional fecal indicators at beaches without a point
Table 1 e Outcome possibilities using the current beach management model: management decision based on previous day’s E. coli concentration. Management outcome
Error
Correct open Correct closed Incorrect open Incorrect closed
None None type 2 type 1
Health risk Loss of Use
Low High High Low
Low High Low High
Health protection Accurate Accurate Overly liberal Overly conservative
Range of reported outcome Range of mean illness rate frequency for for Chicago beaches Chicago beaches (swimmers/1000) 49e78% 3e14% 12e21% 9e17%
2.7e5.0 13.6e15.3 14.4e16.1 4.2e6.4
1662
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
source (Calderon et al., 1991; Sinigalliano et al., 2010). Questions about the appropriateness of using E. coli as an indicator have also been raised (Harwood et al., 2005; Wong et al., 2009), but a review of epidemiological studies associate elevated E. coli with higher risk of gastrointestinal illness (Marion et al., 2010; Wade et al., 2003). Because alternate standards have not been established, all beaches are managed using the E. coli standard, despite its shortcomings. Chicago does not issue a swim ban until E. coli concentration exceeds 1000 MPN/100 ml. Currently, there are no estimates of the percent of swimmers who enter the water during a swimming advisory (i.e., 235 < E. coli <1000) (Hou et al., 2006). Here, we elect to use >235 to express excess illnesses and define model errors because it is the more conservative calculation. Rate of excess illness would increase by 0.0059% using the 1000 CFU standard.
2.5. Different management applications of monitoring standard In an attempt to improve model performance, the E. coli standard for the binary outcome was targeted. As described previously, monitoring standards recommended a five-day geometric mean, and a general single-sample maximum was provided (US EPA, 1986). The guidance suggested, however, that single-sample maximums be calculated for individual jurisdictions based on local log standard deviation of E. coli concentration (US EPA, 1986). The calculation provided in the ambient criteria (US EPA, 1986) is ss ¼ 10^ log10 GM þ cl log10 sd
14/day. Overall, for the 14 beaches, a mean of 35 illnesses/day could be expected (range: 0e659 66 SD).
3.2.
Amount of E. coli variation explained by models (R2)
Coefficient of determination was lowest for the CM, with a range of 0.049 (South Shore) to 0.135 (63rd Street) for individual beaches and an overall adjusted R2 of 0.141 (Fig. 2a). The higher R2 for 63rd Street results from the persistent high E. coli concentrations at this location with little variation. Generally, more variation in E. coli concentrations could be explained using the RM, resulting in a range of R2 from 0.111 (63rd Street) to 0.287 (North Avenue). Further improvement in R2 was seen with the IM (Fig. 2a). The model for 63rd Street included the lowest amount of variation explained (0.141); the highest adjusted R2 was at North Avenue (0.313). The general pattern was an increasing R2 with increasing model refinement, seen at all beaches except 63rd Street. The change in R2 between the CM and the RM was greatest for Hollywood and South Shore beaches, with notable improvements at Foster, North Avenue and South Shore. Change in R2 between RM and IM was not as great overall; most improvement was at the south side beaches. Precision of the predictions increased with model refinement as seen in a decreasing RMSE.
3.3. Binary water quality standard outcome (number of prediction errors) Using the CM, the beach was correctly left open to swimming 68% of the time (range ¼ 49e78%) (Fig. 3). This increased to 78%
(3)
where ss ¼ the single-sample maximum; GM ¼ 126, geometric mean E. coli concentration for acceptable illness rate of 8 per 1000; cl ¼ 0.675, the 75% calculated one-sided confidence level for a designated heavily used beach area, and sd ¼ is the calculated standard deviation for a given jurisdiction, 0.718 for Chicago beaches (mean log E. coli ¼ 1.776) Chicago’s singlesample maximum would therefore be 385 CFU/100 ml.
3.
Results
Visits to Chicago’s lakefront beaches are often in excess of 27,000,000 annually, with fully half of the visits associated with two beaches: Oak Street and North Avenue (Chicago Park District, unpublished data).
3.1.
Baseline water quality
Chicago beaches exceed the 235 single-sample maximum 14e35% of the time, with an overall rate of 20%. Given the current water quality and visitation of Chicago’s beaches and the hypothetical absence of any monitoring or associated beach closures (IA model), the majority of expected illnesses were associated with the high visitation beaches (Fig. 1). The highest illness rate was 138/day for North Avenue, and the lowest was 7/day for Rainbow. Despite having the highest mean E. coli concentration, 63rd Street illness rate averaged
Fig. 1 e Comparison of mean E. coli concentration at each beach and mean number of swimmers that can be estimated to develop gastrointestinal illness each day. Gradation of circles indicates the expected individual illness rate. Beaches include Loyola (LY), Albion (AL), Hollywood (HW), Foster (FO), Montrose (MO), North Avenue (NA), Oak (OK), 12th Street (12), 31st Street (31), 57th Street (57), 63rd Street (63), South Shore (SS), Rainbow (RB), and Calumet (CA).
1663
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
Fig. 2 e Comparison of (a) coefficient of variations, (b) percent type I errors, and (c) percent type II errors by modeling approach.
(range ¼ 53e85%) with the RM and was 77% (range ¼ 54e84%) with the IM. Rate of correct closures under the CM is 6% (range ¼ 3e14%). That decreased to 0.8% (0.3e9) with the RM, and with the IM, the rate was 1.4% (0.3e11%). Type I errors, in which a swimming advisory is posted although E. coli concentration is lower than the single-sample maximum, were made 11.5% (9e17%) of the time using the CM for monitoring (Fig. 2b). With the application of an RM, that
inactive
rate decreased to 2.2% (0.4e4) of the time, and with the IM application, the rate was 3.2% (1e11%). Type II errors were common under the CM, occurring 14% (12e21%) (Fig. 2c). That number increased with both predictive models, to 19% (16e27) with the RM and 18.5% (14e24) with the IM. Comparison of the percent of type II errors showed highly variable results. The greatest percent change between models was an increase in percent type II errors at 12th Street and
current
outcome correct closed correct open type I error type II error
region
individual
Fig. 3 e Management outcome using the binary approach, i.e., an E. coli concentration >235 results in a beach closure.
1664
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
Loyola between CM and RM. Minimizing type II errors between the RM and IM was limited.
3.4. Association of monitoring and modeling approaches with illness burden A comparison of the four models reveals, most noticeably at the high visitation beaches North Avenue, Montrose and Oak Street, that the leveling off of number of type II errors was mirrored in the rate of excess illnesses prevented with increasing model refinement. There was no significant difference in rate of excess illnesses associated with model approach at any of the individual beaches. In fact, model refinement between RM and IM showed little improvement in the number of excess illnesses prevented. At several beaches, cumulative illnesses were identical for all models except for CM. Eight of the beaches showed a difference in cumulative excess illnesses associated with the four models (Table 2). The CM was associated with lower cumulative excess illnesses, largely because there were fewer instances of type II errors; however, for all beaches but 63rd Street, the CM was associated with higher mean excess illnesses, indicating that with this model, the instances of very high E. coli concentrations tend to be missed. The IA model excluded type I errors because the beach would never be closed to swimming and therefore swimmers would not be prevented from water contact under any water quality conditions. Increasingly refined models were effective at reducing type I errors because they rarely predicted events when water quality exceeded the single-sample standard; most predictions were low E. coli concentrations. This tendency is apparent in the rapid decrease in number of type I errors with increasing model refinement.
3.5.
Chicago-specific monitoring standard
With the more liberal, jurisdiction-specific single-sample maximum (385 CFU/100 ml), the overall percentage of errors for all model types decreased (Fig. 4). Reduction in percent type II errors was more pronounced with the RM and IM (mean decrease by 35e36%) than for the CM (mean decrease by 22%). Dramatic reductions in percent of type II errors (>45%) occurred at some of the beaches with lower mean E. coli concentrations (North Avenue, Oak, Albion) with the RM and IM. The difference in number of illnesses should theoretically
remain unchanged because the illness rate is within the confidence limits of the original derived regression. Because the predictive models tend to predict low overalldthe tendency leading to type II errorsdthe reduction in type I errors was not as remarkable.
4.
Discussion
With recent findings that the time-intensive current beach monitoring models are not generally predictive of real-time FIB concentration (Boehm, 2007; Whitman et al., 1999), beach managers need a means of determining water quality rapidly and efficiently in order to protect public health while maximizing beach access. A suite of approaches for expanding monitoring activities and improving timeliness of monitoring results have been proposed and considered, but minimal research has explored their capacity to meet this goal. If a new method does not provide a significant improvement over current management outcomes, it is unlikely that managers will invest the necessary effort to alter the monitoring program. The first step for many jurisdictions is to implement a water quality monitoring program using the currently used model (CM). According to our results, this activity alone decreased the number of excess illnesses at all beaches, largely as a result of keeping swimmers out of the water more often, regardless of actual water quality. Beaches were closed when water quality was within acceptable standards 9e17% of the time, a scenario that can incur a social and economic burden (Rabinovici et al., 2004). With this model, accurate closures and health protection are strictly dependent on the endurance of an E. coli contamination event. With the rapidly changing nature of water quality (Boehm, 2007; Boehm et al., 2002; Whitman and Nevers, 2004), there is high potential to make a management error (Table 1). Multiple day contamination events would result in some correct swimming advisories (correct closed), and extended periods of low E. coli concentrations would result in accurate swimming permission (correct open). The vast majority of contamination events, however, last one day or less (Leecaster and Weisberg, 2001). In Chicago, only 6% of contamination events persisted for more than one sampling day, with the exception of 63rd Street: 17%. Monitoring at beaches with persistent contamination will inevitably decrease the number of type II errors and associated illnesses because swimmers are
Table 2 e Comparison of the cumulative (mean) excess illnesses as a result of the choice of model approach: inactive, current, regional, or individual beach model. Results based on days of type II errors, ranging from 38 to 103 days of the 247e292 days (2000e2004 seasons) considered in the analysis. Inactive Montrose North Ave 31st 57th 63rd South Shore Rainbow Calumet
2492.94 4540.82 366.22 912.68 704.59 551.00 240.73 1049.60
(42.25) (105.60) (5.39) (16.90) (6.84) (10.20) (3.70) (20.58)
Current 1695.49 (42.39) 3659.66 (107.64) 250.73 (5.57) 758.64 (17.64) 389.27 (6.38) 456.19 (11.13) 176.09 (3.91) 869.01 (22.87)
Regional 2407.17 4540.82 348.00 912.51 548.65 551.00 240.73 1049.32
(41.50) (105.60) (5.52) (17.55) (6.94) (10.20) (3.70) (20.99)
Individual 2442.56 4270.00 320.70 894.31 490.04 537.40 217.60 1023.04
(42.85) (101.69) (5.34) (17.54) (6.71) (10.14) (3.69) (20.88)
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
1665
Fig. 4 e Impact on beach advisory decisions if jurisdiction-specific water quality standard is used. The two beach examples include 63rd Street Beach, which has a high mean E. coli, lower visitation rate, and frequent swimming advisories, and North Avenue Beach, which is the most highly visited Chicago beach and has low mean E. coli and infrequent swimming advisories. Use of a Chicago-specific water quality standard would result in fewer errors overall using all three monitoring model approaches.
accurately kept from the water for some percent of the high concentration events. Likewise, beaches with infrequent contamination events will have a limited number of type I errors. To overcome these inconsistencies in actual vs. presumed water quality, predictive models have been used at some beaches to provide a real-time estimate of water quality and overcome these inherent errors. The cornerstone of predictive modeling for beach management has been the individual beach model (Nevers and Whitman, 2005), but the cost of developing such models has been one of the disincentives to widespread use; many of the models use on-site instrumentation that requires installation and continuous maintenance (Francy, 2009). For this reason, regional models have been explored, which examine the hydrometeorological factors similarly affecting groups of beaches and describe background fluctuations in FIB concentrations (Nevers and Whitman, 2008). While more crude in estimations, regional models may provide a cost-effective solution for jurisdictions with numerous monitored beaches while providing insights into source behavior. Perhaps surprising, the RM results in this exercise were quite similar to results from IM, indicating a generally predictable fluctuation in FIB concentrations across Chicago. Overall, increasing refinement of monitoring approach with the use of predictive models was associated with improved accuracy of E. coli predictions. Use of the RM increased the amount of variation explained over the current monitoring approach, and the use of beach-specific IM somewhat further improved this result. With beach-specific refinement, predictive models for all of the study beaches had higher R2 and lower RMSE: more variation in individual E. coli concentrations was
described and there was lower error in this estimation. In this study, geographically widespread predictors were used, which may have limited each model’s ability to detect beach-specific variation, although the low R2 for 63rd Street was somewhat expected due to inherent high variation in E. coli and the complex circulation pattern at this enclosed beach (Ge et al., 2010; Whitman and Nevers, 2004). Models developed for this beach have depended on higher frequency and higher intensity local data than were available for this exercise (Boehm et al., 2007; Olyphant and Whitman, 2004). Predictability improved significantly with the use of the RM for many of the beaches, with the biggest improvement between CM and RM at the north-side beaches, but it was the further refinement to IM that resulted in the greatest improvement at southern beaches. This pattern supports the idea that there is more beach-specific variation at these southern beaches and E. coli concentrations perhaps recover to background concentrations more slowly than at the northern beaches (Whitman and Nevers, 2008). Use of individual predictive models at these beaches can take these factors into account, resulting in better predictability. The number of type II errors for each model was highly variable, lacking the pattern of improvement with increasing refinement shown in the R2. All of the models failed to predict the majority of high E. coli concentrations. Because many beaches have infrequent high E. coli concentration events, it is difficult to detect a pattern of associated hydrometeorological conditions; this is a problem for many predictive modeling attempts but is an important characteristic for eliminating type II errors. The CM generally had the lowest number of type II errors, likely simply because this monitoring approach
1666
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
closes the beach more often, an action that also results in significantly more type I errors (Fig. 3). Calculated excess illnesses were far less variable between models, indicating that more advanced predictive models do not necessarily provide improved public health protection. IM had a lower number of excess illnesses at some beaches (Table 2), but the difference was not nearly as great as might be expected. The results imply that IM detect some instances of extreme high concentrations at these beaches but generally miss these events. The majority of beaches showed a pattern of identical number of estimated cumulative illnesses for all model approaches with the exception of the CM: it was associated with the lowest cumulative excess illnesses specifically because this approach limits overall exposure to the beach water. The CM was therefore associated with lowest excess illnesses due to more closed beaches (both type I errors and correct closed).
4.1.
Application of monitoring standards
The number of excess illnesses associated with a given model is affected by the calculation of the illness rate, and although jurisdictions have the option to calculate single-sample maximums specific to their beach waters, most Great Lakes states use the default 235 CFU/100 ml standard developed using data from the original epidemiological studies for all beaches statewide (Nevers and Whitman, 2010). Use of a higher Chicagospecific standard increases swimming access without impacting illness rates by expanding the range of allowable water contact and therefore the number of accurate predictions (correct open). Percent of type II errors was reduced under the CM and drastically reduced using RM or IM. This was particularly noticeable at beaches with lower mean E. coli concentrations: upwards of 40% reduction in percent type II errors for North Avenue, Montrose, and Oak Street with RM and IM. This significant reduction is mirrored in the potential to significantly reduce the number of illnesses at these high visitation beaches; illness rates are greatly elevated during a single high FIB event at a high use beach (Turbow et al., 2003). The use of the Chicagospecific standard did not greatly reduce the number of type II errors for 63rd Street because of the higher overall mean E. coli concentration. Within the framework of the original monitoring standards, leeway is provided for level of beach use, beach-specific variation in bacteria concentrations, and calculation of the overall water quality (US EPA, 1986), and health protection is assumed to be equally provided over a range of calculated concentrations. Considering this broader range of confidence could increase beach use without influencing health outcome under a variety of monitoring approaches, including the predictive models presented here. These results indicate that the use of a higher standard, along with a predictive model could maximize access at many of the Chicago beaches without increasing public health risk. The monitoring standards recommend use of the 5-day geometric mean, but most managers opt for the single-sample maximum, likely due to ease of use. An examination of Chicago beach monitoring data reveals that use of the 5-day geometric mean results in fewer errors than either the 235 single-sample maximum or the 385 Chicago-specific standard presented here. However, the number of days exceeding the
specified limits (i.e., swimming advisories) increases significantly with use of the geometric mean. The 5-day geometric mean was developed based on studies at beaches influenced by point sources (US EPA, 1986), areas more likely to have persistent high E. coli concentration events; river discharge or sewage releases may create periods of sustained high E. coli concentrations, warranting the extended swimming advisory that results from a running geometric mean. Chicago’s beaches, however, are not influenced by a major point source except during rare events of sewage overflows, during which beaches are preemptively closed for several days.
4.2.
Maximizing public health protection
Predictive modeling results indicate that this monitoring approach would not improve health protection at all Chicago beaches. The best approach for monitoring may differ between beaches, even within an individual jurisdiction such as Chicago. The threshold for level of effort associated with increased model refinement would have to be considered for each beach, perhaps while incorporating economic considerations. Hou et al. (2006) determined that different monitoring policies provided the optimal economic and public health outcomes for each of two beaches. The application of different monitoring strategies may include combining approaches or extending to rapid methods, alternate indicators, or diverse management plans. Recent research to characterize the sources, survival, fate, and transport of FIB and the applications of monitoring programs has perhaps complicated the applicability of different monitoring and management strategies by indicating that one monitoring approach does not fit all beach types. Novel management approaches have included predictive models (Frick et al., 2008; Nevers and Whitman, 2005), rapid tests (Bushon et al., 2009; Lavender and Kinzelman, 2009), new indicators, including hostspecific markers (Bacteroides, Methanobrevibacter, virulence markers etc), and gene-based detection techniques for human pathogens (Griffith et al., 2009). Reconsiderations of monitoring standards have also been explored (Kim and Grant, 2004; Nevers and Whitman, 2010). While epidemiological studies have linked illness rates with outcomes from some of these new analyses (Wade et al., 2006), care will have to be taken to consider whether environmental conditions and sources influence the results outcomes. The ideal method for beach management may differ among beaches and jurisdictions, and it may be desirable to have a variety of monitoring approaches available to beach managers that increase accuracy and reliability at given beaches. In deciphering the best management plans for different beaches, efforts should focus on improving public health protection, perhaps considering a wide variety of available monitoring options.
5.
Conclusions
Refinement of monitoring models generally increased predictability of E. coli but did not necessarily result in fewer errors or excess illnesses Regional models provided similar levels of accuracy as individual beach models in many locations
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
Use of a location-specific water quality standard, combined with empirical predictive models, provided the greatest beach access without sacrificing public health protection
Acknowledgments We thank Murulee Byappanahalli (USGS) for his careful review. Research was funded in part by the US Ocean Action Plan: USGS Ocean Research Priorities Plan and by the Great Lakes Restoration Initiative. Any use of trade, product, or firm names is for descriptive purposes only and does not imply endorsement by the U.S. Government. This article is Contribution 1627 of the USGS Great Lakes Science Center.
references
Boehm, A.B., 2007. Enterococci concentrations in diverse coastal environments exhibit extreme variability. Environmental Science & Technology 41 (24), 8227e8232. Boehm, A.B., Grant, S.B., Kim, J.H., Mowbray, S.L., McGee, C.D., Clark, C.D., Foley, D.M., Wellman, D.E., 2002. Decadal and shorter period variability of surf zone water quality at Huntington Beach, California. Environmental Science & Technology 36 (18), 3885e3892. Boehm, A.B., Whitman, R.L., Nevers, M.B., Hou, D., Weisberg, S.B., 2007. Statistical Framework for Recreational Water Quality Criteria and Monitoring. In: Wymer, L.J. (Ed.). Wiley, West Sussex, UK, pp. 179e210. Buckalew, D.W., Hartman, L.J., Grimsley, G.A., Martin, A.E., Register, K.M., 2006. A long-term study comparing membrane filtration with Colilert defined substrates in detecting fecal coliforms and Escherichia coli in natural waters. Journal of Environmental Management 80 (3), 191e197. Bushon, R.N., Likirdopulos, C.A., Brady, A.M.G., 2009. Comparison of immunomagnetic separation/adenosine triphosphate rapid method to traditional culture-based method for E. coli and enterococci enumeration in wastewater. Water Research 43 (19), 4940e4946. Calderon, R.L., Mood, E.W., Dufour, A.P., 1991. Health effects of swimmers and nonpoint sources of contaminated water. International Journal of Environmental Health Research 1, 21e31. Dufour, A.P., 1984. Health Effects Criteria for Fresh Recreational Waters EPA Report EPA-600/1-84-004, Cincinnati, Ohio. Dwight, R.H., Brinks, M.V., SharavanaKumar, G., Semenza, J.C., 2007. Beach attendance and bathing rates for Southern California beaches. Ocean & Coastal Management 50 (10), 847e858. Francy, D.S., 2009. Use of predictive models and rapid methods to nowcast bacteria levels at coastal beaches. Aquatic Ecosystem Health and Management 12 (2), 177e182. Frick, W.A., Ge, Z., Zepp, R.G., 2008. Nowcasting and forecasting concentrations of biological contaminants at beaches: a feasibility and case study. Environmental Science & Technology 42, 4218e4824. Ge, Z., Nevers, M.B., Schwab, D.J., Whitman, R.L., 2010. Coastal loading and transport of Escherichia coli at an embayed beach in Lake Michigan. Environmental Science & Technology 44, 6731e6737. Given, S., Pendleton, L.H., Boehm, A.B., 2006. Regional public health cost estimates of contaminated coastal waters: a case
1667
study of gastroenteritis at southern California beaches. Environmental Science & Technology 40 (16), 4851e4858. Griffith, J.F., Cao, Y., McGee, C.D., Weisberg, S.B., 2009. Evaluation of rapid methods and novel indicators for assessing microbiological beach water quality. Water Research 43 (19), 4900e4907. Gronewold, A.D., Borsuk, M.E., Wolpert, R.L., Reckhow, K.H., 2008. An assessment of fecal indicator bacteria-based water quality standards. Environmental Science & Technology 42 (13), 4676e4682. Harwood, V.J., Levine, A.D., Scott, T.M., Chivukula, V., Lukasik, J., Farrah, S.R., Rose, J.B., 2005. Validity of the indicator organism paradigm for pathogen reduction in reclaimed water and public health protection. Applied and Environmental Microbiology 71 (6), 3163e3170. He, L.M., He, Z.L., 2008. Water quality prediction of marine recreational beaches receiving watershed baseflow and stormwater runoff in southern California, USA. Water Research 42 (10e11), 2563e2573. Hou, D., Rabinovici, S.J.M., Boehm, A.B., 2006. Enterococci predictions from partial least squares regression models in conjunction with a single-sample standard improve the efficacy of beach management advisories. Environmental Science & Technology 40 (6), 1737e1743. Kim, J.H., Grant, S.B., 2004. Public mis-notification of coastal water quality: a probabilistic evaluation of posting errors at Huntington Beach, California. Environmental Science & Technology 38 (9), 2497e2504. Lavender, J.S., Kinzelman, J.L., 2009. A cross comparison of QPCR to agar-based or defined substrate test methods for the determination of Escherichia coli and enterococci in municipal water quality monitoring programs. Water Research 43 (19), 4967e4979. Leecaster, M.K., Weisberg, S.B., 2001. Effect of sampling frequency on shoreline microbiology assessments. Marine Pollution Bulletin 42 (11), 1150e1154. Lepesteur, M., McComb, A.J., Moore, S.A., 2006. Do we all face the same risk when bathing in the estuary? Water Research 40 (14), 2787e2795. Marion, J.W., Lee, J., Lemeshow, S., Buckley, T.J., 2010. Association of illness and recreational water exposure during advisory and non-advisory conditions at an inland US beach. Water Research 44 (16), 4796e4804. McPhail, C.D., Stidson, R.T., 2009. Bathing water signage and predictive water quality models in Scotland. Aquatic Ecosystem Health and Management 12 (2), 183e186. Nevers, M.B., Whitman, R.L., 2005. Nowcast modeling of Escherichia coli concentrations at multiple urban beaches of southern Lake Michigan. Water Research 39 (20), 5250e5260. Nevers, M.B., Whitman, R.L., 2008. Coastal strategies to predict Escherichia coli concentrations for beaches along a 35 km stretch of southern Lake Michigan. Environmental Science & Technology 42 (12), 4454e4460. Nevers, M.B., Whitman, R.L., 2010. Policies and practices of beach monitoring in the Great Lakes, USA: a critical review. Journal of Environmental Monitoring 12 (3), 581e590. Olyphant, G.A., Whitman, R.L., 2004. Elements of a predictive model for determining beach closures on a real time basis: the case of 63rd Street Beach Chicago. Environmental Monitoring and Assessment 98 (1e3), 175e190. Pru¨ss, A., 1998. Review of epidemiological studies on health effects from exposure to recreational water. International Journal of Epidemiology 27 (1), 1e9. Rabinovici, S.J.M., Bernknopf, R.L., Wein, A.M., Coursey, D.L., Whitman, R.L., 2004. Economic and health risk trade-offs of swim closures at a Lake Michigan Beach. Environmental Science & Technology 38 (10), 2737e2745.
1668
w a t e r r e s e a r c h 4 5 ( 2 0 1 1 ) 1 6 5 9 e1 6 6 8
Seyfried, P.L., Tobin, R.S., Brown, N.E., Ness, P.F., 1985. A prospective study of swimming-related illness I. Swimmingassociated health risk. American Journal of Public Health 75 (9), 1068e1070. Sinigalliano, C.D., Fleisher, J.M., Gidley, M.L., Solo-Gabriele, H.M., Shibata, T., Plano, L.R.W., Elmir, S.M., Wanless, D., Bartkowiak, J., Boiteau, R., Withum, K., Abdelzaher, A.M., He, G., Ortega, C., Zhu, X., Wright, M.E., Kish, J., Hollenbeck, J., Scott, T., Backer, L.C., Fleming, L.E., 2010. Traditional and molecular analyses for fecal indicator bacteria in nonpoint source subtropical recreational marine waters. Water Research 44 (10), 3763e3772. Turbow, D.J., Osgood, N.D., Jiang, S.C., 2003. Evaluation of recreational health risk in coastal waters based on enterococcus densities and bathing patterns. Environmental Health Perspectives 111 (4), 598e603. US EPA, 1986. Ambient Water Quality Criteria for Bacteria 1986, p. 18. US EPA, Office of Water Regulations and Standards, Washington DC. Wade, T.J., Calderon, R.L., Sams, E., Beach, M., Brenner, K.P., Williams, A.H., Dufour, A.P., 2006. Rapidly measured indicators of recreational water quality are predictive of swimming associated gastrointestinal illness. Environmental Health Perspectives 114 (1), 24e28. Wade, T.J., Pai, N., Eisenberg, J.N.S., Colford Jr., J.M., 2003. Do U.S. Environmental Protection Agency water quality guidelines for recreational waters prevent gastrointestinal illness? A
systematic review and meta-analysis. Environmental Health Perspectives 111 (8), 1102e1109. Whitman, R.L., Nevers, M.B., 2003. Foreshore sand as a source of Escherichia coli in nearshore water of a Lake Michigan beach. Applied and Environmental Microbiology 69 (9), 5555e5562. Whitman, R.L., Nevers, M.B., 2004. Escherichia coli sampling reliability at a frequently closed Chicago beach: monitoring and management implications. Environmental Science & Technology 38 (16), 4241e4246. Whitman, R.L., Nevers, M.B., 2008. Summer E. coli patterns and responses along 23 Chicago beaches. Environmental Science & Technology 42 (24), 9217e9224. Whitman, R.L., Nevers, M.B., Gerovac, P.J., 1999. Interaction of ambient conditions and fecal coliform bacteria in southern Lake Michigan waters: monitoring program implications. Natural Areas Journal 19, 166e171. Whitman, R.L., Shively, D.A., Pawlik, H., Nevers, M.B., Byappanahalli, M.N., 2003. Occurrence of Escherichia coli and enterococci in Cladophora (Chlorophyta) in nearshore water and beach sand of Lake Michigan. Applied and Environmental Microbiology 69 (8), 4714e4719. Wong, M., Kumar, L., Jenkins, T.M., Xagoraraki, I., Phanikumar, M.S., Rose, J.B., 2009. Evaluation of public health risks at recreational beaches in Lake Michigan via detection of enteric viruses and a human-specific bacteriological marker. Water Research 43, 1137e1149.