Screening Analysis of Human Pharmaceutical Compounds in U.S.

Environ. Sci. Technol. 2004, 38, 838-849

Screening Analysis of Human Pharmaceutical Compounds in U.S. Surface Waters PAUL D. ANDERSON,† V I N C E N T J . D ’ A C O , * ,‡ PETER SHANAHAN,§ STEVEN C. CHAPRA,| MARY E. BUZBY,⊥ VIRGINIA L. CUNNINGHAM,O BETH M. DUPLESSIE,† EILEEN P. HAYES,∇ FRANK J. MASTROCCO,] NEIL J. PARKE,# JOHN C. RADER,† JOHN H. SAMUELIAN,+ AND BRADLEY W. SCHWAB† AMEC Earth & Environmental, 239 Littleton Road, Suite 1B, Westford, Massachusetts 01886, Quantum Management Group, Incorporated, 1187 Main Avenue, Clifton, New Jersey 07011, HydroAnalysis, Incorporated, 481 Great Road, Suite 3, Acton, Massachusetts 01720, Tufts University, 113 Anderson Hall, Medford, Massachusetts 02155, Merck & Co., Inc., Two Merck Drive, Whitehouse Station, New Jersey 08889, GlaxoSmithKline, 2200 Renaissance Boulevard Suite 105, King of Prussia, Pennsylvania 19406, Bristol-Myers Squibb Company, P.O. Box 191, New Brunswick, New Jersey 08903, Pfizer Inc, 235 East 42nd Street, New York, New York 10017, Eli Lilly and Company, Lilly Corporate Center, Indianapolis, Indiana 46285, and AMEC Earth & Environmental, 15 Franklin Street, Portland, Maine 04101

The PhATE (Pharmaceutical Assessment and Transport Evaluation) model presented in this paper was developed as a tool to estimate concentrations of active pharmaceutical ingredients (APIs) in U.S. surface waters that result from patient use (or consumption) of medicines. PhATE uses a mass balance approach to model predicted environmental concentrations (PECs) in 11 watersheds selected to be representative of most hydrologic regions of the United States. The model divides rivers into discrete segments. It estimates the mass of API that enters a segment from upstream or from publicly owned treatment works (POTW) and is subsequently lost from the segment via in-stream loss mechanisms or flow diversions (i.e., manmade withdrawals). POTW discharge loads are estimated based on the population served, the API use per capita, the potential loss of the compound associated with human use (e.g., metabolism), and the portion of the API mass removed in the POTW. Simulations using three surrogate compounds show that PECs generated by PhATE are generally within an order of magnitude of measured concentrations * Corresponding author phone: (973) 340-9808; fax: (973) 3409818; e-mail: [email protected]. † AMEC Earth & Environmental, Westford, MA. ‡ Quantum Management Group, Incorporated. § HydroAnalysis, Incorporated. | Tufts University. ⊥ Merck & Co., Inc. O GlaxoSmithKline. ∇ Bristol-Myers Squibb Company. ] Pfizer Inc # Eli Lilly and Company. + AMEC Earth & Environmental, Portland, ME. 838

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 38, NO. 3, 2004

and that the cumulative probability distribution of PECs for all watersheds included in PhATE is consistent with the nationwide distribution of measured concentrations of the surrogate compounds. Model simulations for 11 APIs yielded four categories of results. (1) PECs fit measured data for two compounds. (2) PECs are below analytical method detection limits and thus are consistent with measured data for three compounds. (3) PECs are higher than (i.e., not consistent with) measured data for three compounds. However, this may be the consequence of as yet unidentified depletion mechanisms. (4) PECs are several orders of magnitude below some measured data but consistent with most measured data for three compounds. For the fourth category, closer examination of sampling locations suggests that the field-measured concentrations for these compounds do not accurately reflect human use. Overall, these results demonstrate that PhATE may be used to predict screening-level concentrations of APIs and related compounds in the environment as well as to evaluate the suitability of existing fate information for an API.

Introduction The detection of pharmaceutical compounds (referred to in this paper as active pharmaceutical ingredients, or APIs) in the aquatic environment has received widespread attention in the technical literature (1-3) and the popular press (4-6). APIs and their associated metabolites have been detected in surface waters (rivers, lakes, and marine waters), groundwater, drinking water, and sewage treatment plant effluents (1-3). We define an API as a compound responsible for physiological or pharmacological action and used in the diagnosis, cure, mitigation, treatment, or prevention of disease in humans. Pharmaceuticals for human use can enter the environment by excretion following therapeutic use, discharge of treated wastewater from manufacturing facilities, or disposal of unused medicines by the consumer (3). Flushing unused medicines down the toilet appears to be of minor importance, while patient excretion following therapy is widely considered to be the primary pathway to the environment (3). One of the main challenges in the development of pharmaceuticals is to identify molecules that are resistant to metabolic degradation processes and persist to exert the desired effect at the appropriate anatomical site. This persistence can potentially result in a fraction of an API being excreted unchanged, surviving wastewater treatment processes, and entering surface waters. The PhATE (Pharmaceutical Assessment and Transport Evaluation) model, trademark of the Pharmaceutical Research and Manufacturers of America (PhRMA), presented in this paper was developed to estimate predicted environmental concentrations (PECs) of any API that results from human use of medicines, as well as to evaluate the suitability of existing fate information for an API. Another paper (in preparation) includes the use of surface water and drinking water concentrations estimated by PhATE and evaluates whether there is a potential for these concentrations to affect human health. Model Context. Water-quality modeling has evolved a great deal since the first dissolved oxygen model was developed in the early 20th century (7). This historical evolution has been largely dictated by the water-quality problems being addressed during particular periods (8) and has also been highly influenced by the evolution of computing 10.1021/es034430b CCC: $27.50

 2004 American Chemical Society Published on Web 12/18/2003

FIGURE 1. The locations of the 11 watersheds included in the current version of PhATE are shown. and advances in scientific understanding. Over time, waterquality models have increased significantly in complexity and scope. Based in part on these advancements, a common misconception is that more complex models are necessarily superior to simpler frameworks. In fact, the actual level of model complexity should be dictated by (1) the model’s objectives (i.e., the nature of the questions being asked) and (2) the amount of available information (both observational data and quantifiable scientific knowledge). Thus, simpler models are often more useful than complex models when broad questions are being addressed (9-12). Although observations indicate that certain APIs are detectable in natural waters (1-3), the high-resolution data required for advanced, system-specific model applications have yet to be collected. As a consequence, the present model represents a tool to help evaluate the issue on a national and regional scale in the U.S. The model can be used to generate bounding concentrations (i.e., concentrations that are equal to or higher than any that can be reasonably expected) when conservative fate and transport assumptions are employed. The screening evaluation presented in this paper combines PhATE with more realistic assumptions to generate more representative (within an order of magnitude of expected) exposure concentrations.

Overview of the PhATE Model PhATE uses a mass balance approach to model PECs in 11 watersheds selected to be representative of most watersheds and hydrologic regions of the United States (Figure 1, Table 1). The watersheds represent about 19% of the land area of the contiguous 48 states and contain about 44 260 kilometers of stream length divided into 2710 segments (i.e., reaches) (Table 1). Fourteen percent of the U.S. population lives within the selected watersheds, which contain 1112 publicly owned treatment works (POTWs) (Table 1). Notably absent are watersheds containing major metropolitan centers (e.g., New York City, Los Angeles, Miami). Such watersheds were not included because drinking water for these and several other major metropolitan centers is provided by sources not affected by POTW discharges. The current version of PhATE

TABLE 1. Characteristics of Watersheds Included in the PhATE Model

watershed

watershed total area stream population (square lengthd in the km)c watersheda POTWsb (km)

Merrimack River 2 090 300 Miami River, Ohio 1 809 700 White River, Indiana 2 465 600 Mississippi Headwaters 5 291 500 Kansas River 1 333 700 Trinity River 5 104 300 Lower Colorado River 5 861 200 Sacramento River 2 589 100 Columbia River 6 306 400 Atlanta Headwaters 3 894 400 Schuylkill River 1 950 400 totale 38 696 300

41 58 113 363 99 114 24 19 162 76 43 1112

13 030 645 13 900 1 320 31 600 3 155 174 735 11 840 155 660 6 050 46 540 3 270 350 060 3 900 72 260 1 160 570 135 9 450 52 860 2 915 5 000 500 1 485 800 44 260

a Data from U.S. Census. b Data from BASINS/U.S. EPA Clean Water Needs Survey-1996 and BASINS/U.S. EPA Permit Compliance System. c Data from USGS, Hydrologic Unit Boundaries for the United States. d Data from BASINS/U.S. EPA River Reach File version 1.0 (RF1), complemented with USGS Enhanced RF-1 data. Total stream length located downstream of at least one POTW, which is a subset of the total stream length within the watershed. e Total may disagree with sum of watersheds due to rounding.

does not consider veterinary APIs or septic system discharges to groundwater because these releases are through pathways other than a POTW discharge to surface water. PhATE estimates concentrations of APIs in a given reach of river following basic principles of conservation of mass. The model accounts for the mass of API that flows into the reach from upstream, enters the reach via POTW discharges, and is lost from the reach via in-stream loss mechanisms or flow diversions (i.e., man-made withdrawals). POTW discharge loads are estimated based on the population served by the POTW, the API use per capita, the potential loss of the compound associated with human use (e.g., metabolism), and the API mass removed in the POTW. PhATE was developed using Microsoft Visual C++ and uses Microsoft Access databases to store input and output data. VOL. 38, NO. 3, 2004 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

839

TABLE 2. Summary of Key Inputs for PhATE Model compound use and characteristicsa

publicly-owned treatment works (POTWs)b dams and reservoirsc segment detaild

usage per capita, in-stream first-order loss, human loss (e.g., metabolism), POTW removal efficiency for each POTW treatment type name, location, POTW treatment type, population served, flow rate name, volume, surface area, length, depth EPA reach segment number, system segment number, segment sequencing, mean flow, low flow (7-day, 10-year low flow), mean-flow velocity, low-flow velocity, length, depth, width

a Input parameters provided by the user based on manufacturer or literature sources. b BASINS/U.S. EPA Clean Water Needs Survey-1996 and BASINS/U.S. EPA Permit Compliance System. c BASINS/U.S. Army Corps of Engineers National Inventory of Dams. d RF1, complemented with USGS Enhanced River Reach File 2.0 (ERF1-2) data (15).

Model Inputs. Screening for APIs in rivers necessarily entails the incorporation of many data to define the physical and hydrologic characteristics of the receiving water. It also requires specifying the source of the compounds, which in this version of PhATE is assumed to be patients using the compounds at therapeutic doses and the POTWs through which the compounds pass before entering a surface water. A Geographic Information System (GIS) is used to manage watershed and hydrologic data. For hydrological inputs (e.g., information about river segments, mean- and low-flow rates, see Table 2), the model relies upon data provided with the U.S. Environmental Protection Agency (EPA) Better Assessment Science Integrating Point and Nonpoint Sources (BASINS) database (13). Input data were subject to a number of quality control evaluations (summarized in Table S-1, Supporting Information) to ensure consistency with the principal components for spatial data assessment recommended by the National Institute of Standards and Technology (14). Streamflow Corrections. Upon completion of the initial quality checks (described in Table S-1), two additional data verification steps were taken. The first addressed discrepancies between the river flow information from the U.S. EPA BASINS Reach File, version 1 (RF1), and POTW flows; the second addressed loss of in-stream flow (and mass) due to diversions. Both are described in the Supporting Information. Though RF1 does not indicate where diversions of water are located along a river, PhATE assumes that a large (10% or greater) decrease in mean flow that persists over multiple downstream segments represents a diversion. When a diversion is identified, PhATE assumes that both water and mass of the compound are removed from the river, with the amount of mass removed in proportion to the reduction in flow. Mass Loading. The mass loading from a POTW to a receiving water is based upon the average annual per capita human use of the compound, potential loss of the compound associated with human use (e.g., metabolism), the number of people being served by the POTW, and removal within the POTW during treatment. Annual per capita use of a compound is derived by dividing the annual sales of the compound by the total U.S. population; that is, the model assumes uniform use across the U.S. Regional differences may exist for compound use; however, the annual per capita human use data source does not provide regional differentiation. This represents an area for potential future 840

9


refinement. Losses of the compound associated with human use and metabolism (prior to a compound reaching a POTW) and treatment within a POTW are assumed to be a fixed proportion of a compound’s mass. PhATE assumes that the entire mass loading is treated in a POTW and does not explicitly consider bypasses, such as combined sewer overflows. The mass loading from the POTW to a river segment is calculated using the following equation:

MPOTW ) PMh(1 - Lh)(1 - LPOTW)

(1)

where P ) population served by the POTW (persons), Mh ) API usage per capita (kg/person-year), Lh ) fixed loss from human use (e.g., metabolism) of the compound of interest (fraction), LPOTW ) fixed loss within the POTW of the compound of interest (fraction), and MPOTW ) mass loading to receiving water (kg/year). The actual loss that may occur within the POTW is specific to both the compound and the treatment type. The model provides for seven possible treatment types, including no discharge, raw discharge, primary treatment, advanced primary treatment, secondary treatment, and advanced treatment I and II as defined by the U.S. EPA (16). Each POTW is assigned one of these treatment types based on information reported in BASINS or from alternate sources (e.g., telephone contact with the POTW). River Reach Concentrations. PhATE includes river reaches that are modeled as plug-flow segments and reservoirs that are modeled as mixed-flow segments. The following general equation is used to estimate the PEC at the end of a given plug-flow river segment:

CPEC )

∑(M e

(M0e-kτR) +

i

-kτi

)

(2)

Q

where CPEC ) predicted environmental concentration (mg/L), M0 ) mass loading from prior river segment (g/day), k ) first-order loss rate constant (day-1), τR ) travel time of river segment (days), Mi ) mass loading from the ith POTW (g/day), τi ) travel time from the ith POTW to the end of the segment (days), and Q ) flow rate (m3/day). The first-order loss rate constant, k, as used in eq 2 is a composite rate that is calculated as the sum of all relevant first-order loss rate constants as might be associated with biodegradation, hydrolysis, photolysis, evaporation, sedimentation, and so forth. Reservoir Concentrations. Of the 2710 segments within PhATE, a total of 220 segments (8.1%) are within reservoirs. Unlike river segments, which can be effectively modeled as plug-flow transport elements, mixing may be significant in reservoirs and requires explicit consideration. Each reservoir in PhATE is represented as a series of wellmixed tanks where advective flow moves mass from one tank to the next. Such “tanks-in-series” representations are commonly used in environmental engineering to model mixed-flow systems such as reservoirs (17). Mass balances can be written for each tank accounting for flow, POTW loading, and a first-order loss rate. These can be solved for the reservoir outflow concentration as follows:

Cout )

(

Q

Q + kV

)

n

Cin +

n

∑ i)1

[(

Q

Q + kV

)

n-1

Wi

Q + kV

]

(3)

where Cin ) upstream surface water concentration (mg/L), Cout ) outflow surface water concentration (mg/L), Q ) flow rate (m3/day), V ) volume of an individual tank (m3), n ) the number of tanks, and Wi ) the POTW mass inflow into the ith tank (g/day).

TABLE 3. Model Input Parameters for Surrogate Compoundsa

API

chemical functionality at pH 7

Kd or Kp (L/kg)

annual use (kg/year)

human loss (%)

1° only (%)

caffeine LAS triclosan

basic acidic neutral

7.271 30008 47513

2 215 5502 314 000 0009 600 00014

0.03 0.0 0.0

404 2710 3215

a

POTW removal, % 1° + 2° (%) 1° + 3° (%) 935 9811 9016

99.66 9811 9516

in-stream decay (1/day) 0.0057 0.7012 0.2317

Footnotes are in the Supporting Information.

The number of tanks that creates numerical mixing equivalent to the actual diffusion/dispersion can be computed as follows:

n ) 1 + INT

(LU 2E )

(4)

where INT ) integer function (rounds a number down to the nearest integer); L ) reservoir length (m); U ) longitudinal, advective velocity (m/s); and E ) longitudinal turbulent diffusion/dispersion coefficient (m2/s). This is calculated using the formula given by Lawrence et al. (18) as E ) (3.2 × 10-4)LC1.1. LC ) reservoir characteristic length (m). This is calculated as the sum of reservoir mean width and reservoir length, divided by 2. Reservoir Storage Effects. In large reservoirs (i.e., greater than 6 month travel time), the storage capacity of the reservoir can exert a significant averaging effect on concentration, making it inappropriate to use the 7-day, 10-year low flow to calculate the outflow concentration from the reservoir. If reservoir time of travel is greater than 6 months, PhATE uses the mean flow to calculate the outflow concentration under both low- and mean-flow conditions. Thus, under low-flow conditions, the outflow mass is computed as the product of the mean-flow concentration and the 7-day, 10-year low flow. This approach could be considered to not conserve mass in the low-flow simulations if those simulations were considered representative of a true steady-state system. In fact, the storage effects of large reservoirs prevent a true steady state from being achieved during low flow.

Model Corroboration Demonstrating the utility of a computer model as a predictive tool is a process that is sometimes called verification or validation. However, Oreskes et al. (20) have shown that verification or validation is impossible for environmental models and that some lesser degree of demonstrating the model’s utility is all that can be achieved. For this paper, the term corroboration is used to discuss evaluation of the model. The term “corroboration” recognizes that a model can never be shown to be absolutely true (verified) or free of flaws (validated) but that by a variety of comparisons of model predictions with available data, confidence in the ability of the model to make useful predictions can be increased. Reckhow and Chapra (21) call this “measured consistency with empirical evidence”. The U.S. EPA (19) defines screening as “studies where limited calibration and validation data are available and the uncertainty associated with the predicted results is comparatively large, somewhere in the nature of an order of magnitude”. Corroboration of PhATE was directed at determining the performance of the model as a screening tool within this general guideline. Use of Surrogate Corroboration Compounds. Measured concentrations from Kolpin et al. (1), which represent the most extensive published data set of APIs in U.S. surface waters, are used for comparison with model PECs in this paper. Since the detection frequencies for the APIs reported by Kolpin et al. (1) were relatively low, that is, generally less

than 20%, surrogate compounds were identified for model corroboration. Such compounds must (1) be measured and frequently detected at surface water sampling locations included in PhATE, (2) be introduced into the environment predominantly through POTWs, and (3) have necessary model input data available. Based on these criteria, three compounds were identified as surrogates for APIs: caffeine, linear alkyl benzene sulfonates (LAS), and triclosan. These compounds are not commonly considered to be APIs using the definition previously provided; however, they do have certain chemical/physical property and molecular structural similarities to APIs and are thus suitable surrogate compounds regarding environmental fate. In addition, LAS has been investigated by others using a watershed model similar to PhATE (22). Relevant model inputs for the surrogate compounds are summarized in Table 3. The field measurements for triclosan and caffeine were made by the U.S. Geological Survey (USGS) as part of a nationwide reconnaissance (1, 23). Triclosan was detected in 57% of samples, while caffeine was detected in 61.9 and 70.6% of samples depending on the analytical method used. The field measurements for LAS were made by the USGS in the Mississippi River (24). LAS was detected in 17% of the samples collected. Methodology for Comparing Model Results with Field Data. The first phase of model corroboration consisted of comparing field data with model PECs for the same locations on a point-by-point basis. In the second phase, the cumulative probability distributions of field-measured concentrations reported by USGS (23) were compared to the distributions of PECs for the 2710 segments included in PhATE. To conduct the point-by-point comparisons, the latitude and longitude (23) or the river mile (24) for each USGS sampling station was used to identify the stations located within segments included in PhATE. A total of 17 and 23 such stations were identified in the USGS (23) and Tabor and Barber (24) studies, respectively. Since the model calculates predicted concentrations at mean and 7-day, 10year low flow, the model prediction at actual flow was obtained by linear interpolation or extrapolation using the streamflows corresponding to the date(s) when the samples were collected. Streamflows corresponding to field data from Kolpin et al. were estimated from the nearest USGS gauging station (25). Streamflows corresponding to the LAS field data were obtained from Tabor and Barber (24). The field-measured concentrations reported by USGS (23) and Tabor and Barber (24) were compared to the actual-flow PECs generated by PhATE. For field measurements where the compound is detected, the model prediction is considered a favorable match of the field-measured concentrations if the PEC is within a factor of 10 of the field measurement. For field measurements where the compound is not detected, the model prediction is considered a favorable match if the PEC either is below the method detection limit (MDL) or within a factor of 10 above the MDL. This degree of agreement is consistent with use of PhATE for screening purposes. VOL. 38, NO. 3, 2004 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

841

Qualitative comparisons were also made between the cumulative probability distributions of the field-measured concentrations and model PECs for triclosan and caffeine enabling full use of the up-to-104 measurements reported by USGS (23) and not just the 17 stations falling within the watersheds currently included in PhATE. Moreover, the comparison of probability distributions is consistent with use of the model as a tool for nationwide screening. In these comparisons, concentrations for a particular compound at all stations reported by USGS (23) were rank-ordered and plotted versus the cumulative probability of occurrence, which was computed for the mth-highest concentration as m/(n + 1) where n is the total number of observations. Though these field data were collected over a range of flow conditions that generally fall within the low- to mean-flow range, they are presented as a single probability distribution. In contrast, model PECs for the 2710 segments included in PhATE are plotted as two distributions, one for mean flow and one for low flow. An implicit assumption in the comparison of the probability distributions is that the sampling stations selected by the USGS and the river reaches included in PhATE are both representative of the nationwide concentration distribution of stream segments affected by human use compounds and are comparable. Sampling sites selected by USGS (23) were biased toward streams susceptible to contamination, and PhATE includes only river reaches downstream of POTWs; both represent locations where detectable concentrations of human use compounds are most likely to occur. For samples where the compound is detected, the model results are considered to be in favorable agreement with the field data when the distribution of field data falls between the two predicted distributions. Results and Discussion of Point-by-Point Comparisons. A point-by-point comparison of field-measured concentrations (23, 24) to model-generated PECs for caffeine, LAS, and triclosan is shown in Figure 2. Two plots are presented for caffeine corresponding to both analytical methods, that is, method 3 (using filtered water) and method 4 (using whole-water), reported by Kolpin et al. (1). It is important to note that the LAS data are from a series of samples collected along a single river, while the data for caffeine and triclosan are from individual sample locations throughout multiple watersheds. The caffeine concentration by method 3 is nondetect at 6 stations. The model PEC is below or within a factor of 10 of the MDL, and thus consistent with the field measurements, at all of these stations. Of the remaining 10 stations with detected concentrations of caffeine, the model PEC at 7 stations meets the criterion that the model predictions be within a factor of 10 of the measured concentrations (Figure 2a). The caffeine concentration by method 4 is nondetect at 6 stations, and the model PEC is below or within a factor of 10 of the MDL, and thus consistent with the field measurements, at all of these stations. Of the remaining 10 stations with detected concentrations of caffeine, the model PEC at 8 stations meets the criterion that the model predictions be within a factor of 10 of the measured concentrations (Figure 2b). The LAS concentration is nondetect for 8 of the 23 stations, and the model PEC is below or within a factor of 10 of the MDL at all of these stations. Of the remaining 15 stations with detected concentrations of LAS, the model PEC at 14 stations meets the criterion that the model predictions be within a factor of 10 of the measured concentrations (Figure 2c). The triclosan concentration is nondetect for 5 stations, and the model PEC is below or within a factor of 10 of the MDL for all these stations. Of the remaining 11 stations with detected concentrations of triclosan, the model PEC at 9 842

9


stations meets the criterion that the model predictions be within a factor of 10 of the measured concentrations (Figure 2d). Statistical measures of model prediction errors are based on those stations at which the compound is detected. The prediction error is the difference between the flow-adjusted model PEC and the detected concentration. Evaluation measures include the minimum, maximum, arithmetic mean (for prediction error), geometric mean (for measured-tomodeled ratio), median, and standard deviation of the prediction error and the ratio of the measured-to-modeled concentration (Table 4). Consistent with the criterion described above, the ratio of field-measured concentration to modeled concentration should lie in the range of 0.1-10. The geometric mean measured-to-modeled ratios for caffeine of 0.39 and 1.8 for methods 3 and 4, respectively, for LAS of 2.8, and for triclosan of 3.6 meet this criterion (Table 4). The finding that PECs for nondetected levels of surrogate compounds are consistently below or within a factor of 10 of the MDL lends confidence to using the model to examine PECs below MDLs. There is uncertainty in the measured concentrations inherent in the analytical methods. This is demonstrated by comparing the analytical results for caffeine using method 3 (using filtered water) and method 4 (using whole-water) for the same samples. The geometric mean of the method 4 to method 3 concentration ratios is 2.0. Thus the magnitude of the differences in caffeine concentrations reported by these two analytical methods (geometric mean ratio of 2.0) is similar to that of the differences between the measured and modeled concentrations for caffeine (geometric mean ratios of 0.39 and 1.8 for methods 3 and 4, respectively). Results and Discussion of Probability Distribution Comparisons. The cumulative probability distribution of measured concentrations reported by USGS (23) across all sampling stations is compared with the cumulative distribution of PECs across all model segments for caffeine (results of both methods 3 and 4 are shown) and triclosan in Figures 3 and 4, respectively. Note that plotting not-detected samples at the MDL in the probability distributions gives the erroneous visual impression that there is a considerable discrepancy between the model predictions and the field data below roughly the 50th percentile (Figures 3 and 4). In fact, the numerical predictions in this range of the distribution are generally below the MDL and thus consistent with the field data plotted at the MDL. For caffeine, measured concentrations for samples in which the compound is detected are within the range of model PECs calculated for the low- and mean-flow conditions but are closer to the low-flow PECs (Figure 3). For triclosan, the highest measured concentrations are similar to the highest model PECs; however, below the 90th percentile, mean-flow PECs are about 50-fold lower than measured concentrations. The difference between low-flow PECs and measured concentrations gradually increases to about 10-fold near the 50th percentile (Figure 4, top). The generally higher measured concentrations as compared to model PECs may in part be explained by USGS (23) sampling sites being more susceptible to contamination than river segments included in PhATE. However, model sensitivity runs suggest that the difference between measured concentrations and PECs for triclosan may also be explained by an overestimate of in-stream decay. When the in-stream first-order loss rate constant (k) for triclosan is decreased from 0.23 to 0.0 per day, the agreement between measured concentrations and PECs improves substantially; measured concentrations fall between low- and mean-flow PECs instead of exceeding them (Figure 4, bottom). The point-by-point comparison for triclosan also improves when a k of 0.0 per day is used instead of 0.23. All

VOL. 38, NO. 3, 2004 / ENVIRONMENTAL SCIENCE & TECHNOLOGY 9

843

FIGURE 2. Field-measured (23, 24) and PhATE-predicted concentrations of surrogate compounds are shown for each segment that had both a field-measured and PhATE-predicted concentration. Sampling stations (or river mile for LAS) are identified on the x-axis, and concentration (in ng/L) is shown on the y-axis. Model PECs are represented as a square (9). Field-measured concentrations above the MDL are presented as diamonds ()). Estimated concentrations below the MDL are represented by a plus sign (+). If the surrogate compound was not detected, the MDL is represented by an X (×). The distance between the PEC and field-measured concentration represents how closely the measured and predicted values correspond. Smaller distances represent closer correspondence. Results are presented for caffeine ((a) method 3 and (b) method 4), (c) LAS, and (d) triclosan.

TABLE 4. Statistics of Model Prediction Errors caffeine (method 3)

counta minimum maximum meanb median std dev

caffeine (method 4)

LAS

triclosan

prediction error (ng/L)

measured-tomodeled ratio






measured- tomodeled ratio

10 -1046 269 -65 31 360

10 0.0037 9.2 0.39 0.23 3.5

10 -746 180 -75 0.69 253

10 0.33 31 1.8 1.1 9.9

8 -1050 130 -325 -247 369

8 0.76 13 2.8 3.2 3.9

11 -173 167 -37 -44 92

11 0.66 15 3.6 4.4 4.8

a Count only includes samples in which the compound was detected. b The arithmetic mean is calculated for the prediction error, and the geometric mean is calculated for the measured-to-modeled ratio (26).

FIGURE 3. The cumulative probability distribution of all caffeine concentrations in U.S. surface waters reported by the USGS (23) and PECs generated by PhATE for all model segments are shown. Caffeine concentrations measured using method 3 and method 4 are shown as triangles and squares, respectively. Filled symbols (i.e., (2) DM3 [detected, method 3] and (9) DM4) represent detected caffeine concentrations, and open symbols (i.e., (4) NDM3 [not detected, method 3] and (0) NDM4) represent nondetects. 17 sampling locations (instead of 15 out of 17) meet the screening criterion, and the geometric mean measured-tomodeled ratio improves by about 40% (the geometric mean decreases from 3.6 to 2.1). Note that decreasing k has a greater effect on the low-percentile concentrations than on the high percentiles (Figure 4) because the high-percentile concentrations generally represent locations close to a POTW, where travel times are too short to allow significant in-stream decay. In contrast, changing the POTW removal efficiency or changing estimated annual per capita use would affect all concentrations equally. However, measured triclosan concentrations are similar to model PECs at highpercentile concentrations suggesting that the POTW removal efficiency used in the model is appropriate. Note that adjusting the in-stream first-order loss rate to improve the agreement between measured concentrations and PECs is more properly considered model calibration than model corroboration. Systematic bias introduced by fundamental model assumptions, as opposed to user-defined input parameters (e.g., POTW removal efficiency and in-stream decay rate constant), may also explain differences between field-measured and PEC cumulative probability distributions. One possible source of systematic bias that may cause PhATE to underpredict PECs in certain areas is release of compounds in 844

9


untreated wastewater during and following heavy rainfall events, that is, combined sewer overflows. This could be especially important for compounds that have high POTW removal rates because PhATE would assume the entire mass of such compounds is being treated when, in fact, some fraction is being released to the river without treatment. Most assumptions inherent in PhATE (such as assuming a constant POTW removal efficiency, a constant in-stream loss rate for all segments, or a uniform usage per capita) would not be expected to cause a systematic bias in PECs. Rather, one would expect that these assumptions may cause PhATE to overpredict PECs for some segments and to underpredict PECs for other segments. The point-by-point comparison presented above found that such potential deviations do occur but also found that such deviations were not substantial (i.e., the model meets the performance criteria used by this analysis). For caffeine (Figure 4), some of the discrepancy at low percentiles may also be due to variability between the analytical methods (i.e., methods 3 and 4 (1)) at concentrations near and below the MDL. Method 3 uses filtered water samples and has lower reporting limits than method 4, which uses unfiltered samples. Caffeine has a low sorption coefficient and is not expected to sorb significantly to suspended solids in the sample. Thus, the filtering pro-

FIGURE 4. The cumulative probability distribution of all triclosan concentrations in U.S. surface waters reported by the USGS (23) and PECs generated by PhATE for all model segments are shown. Detected triclosan concentrations are shown as filled triangles (2, DM4). Nondetects are shown as open triangles (4, NDM4). Model results are shown for an in-stream k ) 0.23 per day (top figure) and k ) 0.0 (bottom figure). cedure used in method 3 is not expected to remove caffeine from the sample. Method 3 has better accuracy and precision metrics than method 4 (1), suggesting that method 3 may be more reliable than method 4 at lower concentrations (i.e., in the range of 10-100 ng/L). Comparison of field-measured to predicted distributions is consistent with this possibility, since the method 3 concentration distribution below the 50th percentile falls within the range of PECs while the method 4 distribution is above modeled concentrations. Summary of Surrogate Compound Simulations. The point-by-point comparisons have the advantage of comparing PECs to measured concentrations within a segment and examining model prediction error under a common set of inputs and stream conditions. However, because relatively

few segments have field-measured concentrations, the cumulative probability distribution plots provide valuable perspective on the model’s overall predictive ability. The favorable agreement between the model predictions and field measurements at the upper ranges of the distributions shows that PhATE generates PECs for the surrogate compounds that are consistent with concentrations measured in the field. This favorable agreement with field data above the MDL shows that the model can be used with confidence to predict screening-level concentrations of APIs and related compounds in the environment.

API Simulations PhATE was used to generate PECs for 11 of the APIs reported by Kolpin et al. (1). These APIs were selected because they VOL. 38, NO. 3, 2004 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

845

TABLE 5. Model Input Parameters for APIsa,1 API

chemical functionality at pH 7

Kp2 (L/kg)

annual use (kg/year)3

human loss (%)

cimetidine diltiazem enalaprilat ethinyl estradiol (EE2) fluoxetine gemfibrozil mestranol metformin 19-norethisterone13 paroxetine metabolite ranitidine

base base zwitterion neutral base single acid neutral base neutral base zwitterion

500 970 190 5250 1360 0.6 21000 31 1320 870 190

160 000 214 000 1 0907 1448 22 700 289 000 0.3 1 700 000 921 21 40015 285 000

52 96 10 0 90 24 8512 0 0 0 6

a

6 11 2 30 10 0.1 50 0.4 10 10 2

70 70 30 759 8510 4411 98 7 80 89 30

in-stream decay (1/day) 05 0.0326 05 05 05 05 05 05 0.04414 05 05

Footnotes are in the Supporting Information.

are used only in human health drug products, and thus accurate volume data based on human use are available for model input. Several of the other APIs included in the USGS survey are also used in veterinary drug products, are naturally occurring, or have uses in addition to their pharmaceutical use. These latter APIs are not discussed in this paper because they have pathways of entry into the environment not currently evaluated by PhATE. Model Inputs for APIs. Key model input parameters for the selected APIs are presented in Table 5. The annual use for each API (27, 28) reflects U.S. sales of all products containing the API (including combination products) to retail pharmacies, nonfederal hospitals, federal facilities, long-term care facilities, clinics, and HMOs (Table 5). These data represent highly accurate figures for annual production and sales of each API, eliminating a source of potential uncertainty to the model. By using the total mass of API sold, the analysis assumes that all of the drug product sold is consumed by patients and is available for excretion (i.e., after accounting for metabolism) to wastewater. Human metabolism data reported in Table 5 are the result of clinical drug metabolism and pharmacokinetic studies using 14C-radiolabeled API and Clinical Trial Design and Safety Assessment protocols prescribed by the U.S. Food and Drug Administration (FDA). Pharmaceutical manufacturers holding drug registrations with the FDA supplied all of the human metabolism data. The human loss terms in Table 5 assume that conjugated metabolites are deconjugated in the environment and converted to the active parent (29). POTW removal efficiencies in Table 5 are based on either actual POTW data or modeling estimates using WW-TREAT (30) and the Kp values in Table 5. In-stream decay half-lives are available for two APIs. In-stream decay is assumed to be zero for the other APIs for which no in-stream decay data are available. Results of API Simulations. Comparisons of PhATE PECs to field-measured API concentrations (23) fall into four general categories (Figure 5) as discussed below. To demonstrate the model’s utility as a screening tool on a national scale, the results are focused on the cumulative probability distribution comparisons. Model Results Agree with Field Data. Diltiazem is an example of an API for which fate and transport parameters are available from laboratory and field data and for which measured concentrations are within the range of model PECs calculated for the low- and mean-flow conditions (Figure 5a). PhATE model results for cimetidine (Figure 5a) are in good agreement with the reported concentrations at the higher percentile concentrations, that is, >90th percentile; however, the model PECs overpredict the reported concen846

POTW removal, % 1° only (%) 1° + 2° (%)

9


trations at the lower percentiles. Model sensitivity runs performed with various estimates of POTW removal and instream depletion suggest that a higher in-stream decay rate than the presumed zero decay reported in Table 5 may be more appropriate for cimetidine because increasing the instream decay rate constant results in better agreement between PECs and field-measured concentrations at lower percentiles but does not significantly change the model PECs at the high percentiles. Model Results Are Consistent with Field Data Because MDLs Are Above Model PECs. USGS (23) reports few or no detects for enalaprilat, fluoxetine, and paroxetine metabolite (Figure 5b). In each case, the model results are consistent with the measured concentrations in that any detected concentrations are about the same as the highest concentrations predicted by PhATE and PhATE otherwise predicts concentrations below the MDL. These results suggest again, as they did for the surrogate compounds discussed above, that the model can be used to estimate environmental concentrations where analytical methods are not sufficiently sensitive to detect concentrations in the field. Model Results Are Higher Than Field Data. For metformin, gemfibrozil, and ranitidine, PhATE predicts higher concentrations than reported by USGS (23) (Figure 5c). Also for the majority of segments with nondetects, the model predicts concentrations substantially above the MDL. The degree of overprediction (i.e., several orders of magnitude) is not consistent with the surrogate compound results or with the API results described above. In particular, the maximum model PECs (i.e., those at the right-hand extreme of the distribution) for ranitidine and metformin are significantly (as much as 1000-fold) higher than the maximum field-measured concentrations. This difference cannot be explained by an overestimate of annual use. Rather, this suggests that POTW removal mechanisms may exist for these two APIs that are not reflected by the estimated removal rates used in this analysis (Table 5). In the absence of measured POTW removal rates for these two APIs, removal rates are calculated using the WW-TREAT model developed by Cowan et al. (30). A laboratory study is available for metformin that shows only 0.6% mineralization over 28 days using a filtrate of activated sludge and secondary effluent under aerobic conditions (31). However, a considerable portion of the metformin molecule (C4H11N5) is nitrogen and degradation mechanisms may exist that are not evaluated with standard biodegradation test protocols. For gemfibrozil, a range of POTW removal rates have been reported: 16% in trickling filters (32), 46% in activated sludge (32), and 69% in activated sludge with phosphorus removal (33). An average value of 44% removal was used in the model simulations. The model results (Figure 5c) suggest that the

VOL. 38, NO. 3, 2004 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

FIGURE 5. The cumulative probability distribution of all concentrations in U.S. surface waters reported by the USGS (23) and PECs generated by PhATE for all model segments are shown for each of the 11 APIs. Detected concentrations are shown as filled triangles (2), and nondetects are shown as open triangles (4). Results of the 11 APIs are presented in four groups depending upon how the PECs compare to measured concentrations. PECs for diltiazem and cimetidine agree with field data (a). PECs for enalaprilat, fluoxetine, and paroxetine metabolite are consistent with the finding of mostly nondetects (b). PECs for metformin, gemfibrozil, and ranitidine are higher than field data (c). PECs for mestranol, EE2, and 19-norethisterone are lower than field data (d).

9

847

higher end of the reported range in removal efficiencies may be more appropriate for this API. Model Results Are Lower Than Field Data. The maximum model PECs for ethinyl estradiol (EE2), mestranol, and 19norethisterone are 1-5 orders of magnitude lower than the few detected concentrations reported by USGS (23) (Figure 5d). Most of the reported concentrations for these APIs are nondetect, which is consistent with the model results since the MDL is higher than the model PECs. Simulations not depicted here evaluated the effect of assuming zero removal for all depletion mechanisms (i.e., metabolism, POTW removal, and in-stream depletion). Even when using such conservative assumptions, maximum model PECs for these three APIs remained substantially (more than an order of magnitude) lower than the maximum field-measured concentrations. This indicates that the discrepancy between model PECs and field-measured concentrations does not arise from inappropriately excessive model removal mechanisms. Maximum model PECs for these three APIs are consistent with maximum concentrations in surface water as reported by other investigators. For EE2, the maximum model PEC is 2 ng/L as compared with a maximum reported concentration of 15 ng/L (15 ng/L (34), 10 ng/L (35), 5.1 ng/L (36), 4.3 ng/L (37), 0.52 ng/L (38), 0.07 ng/L (39), and nondetect (40, 41, 42, 43)). For 19-norethisterone, the maximum model PEC is 10 ng/L as compared with a maximum reported concentration of 17 ng/L (17 ng/L (41), 10 ng/L (34), and nondetect (35)). Mestranol has not been detected in surface waters by other investigators (40, 35). This is consistent with the model results, which are below the MDL for this API. Discussion of Synthetic Hormone Model Results. The only APIs for which PhATE underestimated measured concentrations are the synthetic hormones, that is, EE2, mestranol, and 19-norethisterone, which are approved by FDA only for human use in prescription oral contraceptive and hormone replacement therapy products. Synthetic hormones were detected at 12 sampling stations. There are only two sampling stations where all three synthetic hormones were detected, that is, FL02 and MT02. These same two stations also account for the highest concentrations of each of these three APIs (23). Neither MT02 nor FL02 are located downstream of any POTWs or pharmaceutical manufacturing facilities. Since both sampling stations are located in rural areas, it appears unlikely that the reported levels of these three APIs are a result of human use (see mass flux calculations in Table S-6, Supporting Information). Although these locations may be affected by other sources of sanitary wastewater, such as septic tanks or small package treatment systems, such sources are unlikely to contribute the relatively high concentrations reported. Ericson et al. (45) suggest that synthetic hormone concentrations, in particular EE2, reported by Kolpin et al. (1) are substantially higher than anticipated and suggest that this difference may be due to interference by natural organic materials that could not be resolved by the analytical method used. Importance of Appropriate Fate Data. The PhATE model results highlight the importance of having comprehensive fate data in order to estimate concentrations in the environment accurately. In using PhATE as a screening tool, it is often appropriate to assume zero POTW removal or zero in-stream depletion in the absence of fate data. However, as demonstrated by the model results for the APIs, without appropriate fate data, environmental concentrations can be significantly overestimated. Conversely, there are potential mechanisms, such as combined sewer overflows and disposal of unused API directly to wastewater systems by patients, that could cause PhATE to underestimate PECs in certain areas. 848

9


Because PECs can be represented as a cumulative probability distribution, the model provides a more realistic characterization of nationwide environmental concentrations than presenting only average or worst case conditions. This representation is well suited for assessing environmental exposure for the purpose of conducting human health or environmental risk assessments.

Acknowledgments The authors acknowledge Eric Mathis and Doug Tedder, both of AMEC Earth and Environmental, for their important contributions in developing the model code. We also gratefully acknowledge the U.S. Geological Survey and The Proctor & Gamble Company for providing field data and model input parameters used in the model corroboration. The technical assistance of Richard Murray-Smith of AstraZeneca and Corrine L. Kupstas and Daniel E. Sullivan of Pharmacia Corporation in reviewing this manuscript is greatly appreciated as are suggestions made by five anonymous reviewers. The Pharmaceutical Research and Manufacturers of America provided financial support for developing this model and for the preparation of this manuscript. AMEC, Quantum Management Group, and HydroAnalysis have provided additional support for developing the model and preparing this manuscript.

Supporting Information Available Summary of PhATE model input data quality checks (Table S-1), RF1 streamflow corrections, footnotes, and references for surrogate compounds and API input parameters (Tables 3 and 5), synthetic hormone mass flux calculations (Table S-6), and nomenclature used in this paper. This material is available free of charge via the Internet at http://pubs.acs.org.

Literature Cited (1) Kolpin, D. W.; Furlong, E. T.; Meyer, M. T.; Thurman, E. M.; Zaugg, S. D.; Barber, L. B.; Buxton, H. T. Environ. Sci. Technol. 2002, 36, 1202-1211. (2) Halling-Sorensen, B.; Nielsen, S. N.; Lanzky, P. F.; Ingerslev, F.; Lutzheft, H. C. H.; Jorgensen, S. E. Chemosphere 1998, 36, 357393. (3) Heberer T. Toxicol. Lett. 2002, 131, 5-17. (4) Household chemicals showing up in streams; Caffeine, cleansers and hormones could harm ecosystems. Chicago Sun-Times, January 27, 2002, Vol. 167, Issue 24. (5) Drug Wastes Pollute Waterways; 80% of Streams Checked by USGS Contain Trace Amounts. The Washington Post, March 13, 2002, p A08. (6) What is in the water? Better detection tools reveal possible ecological ‘villains’ - from hormones to fire retardants - in US streams and rivers. Christian Science Monitor, March 20, 2002. (7) Streeter, H. W.; Phelps, E. B. A Study of the Pollution and Natural Purification of the Ohio River. III. Factors concerned in the phenomenon of oxidation and reaeration. Public Health Bulletin No. 146; U.S. Public Health Service: Washington, DC, Feb 1925. (8) Chapra, S. C. Surface Water-Quality Modeling; WCB McGrawHill: Boston, MA, 1997. (9) Vollenweider, R. A. Schweiz. Z. Hydrol. 1975, 37, 53-84. (10) O’Connor, D. J. J. Environ. Eng. 1988, 114, 507-532. (11) O’Connor, D. J. J. Environ. Eng. 1988, 114, 533-551. (12) O’Connor, D. J. J. Environ. Eng. 1988, 114, 552-574. (13) Better Assessment Science Integrating Point and Nonpoint Source, BASINS, Version 3.0; Report Number EPA-823-B-01-001; Office of Water, U.S. EPA: Washington, DC, June 2001. (14) Standard for Spatial Data Transfer Standard (SDTS); Federal Information Processing Standards Publication 173-1; National Institute of Standards and Technology: Gaitherburg, MD, June 1994. (15) ERF1-2-Enhanced River Reach File 2.0; U.S. Geological Survey Open-File Report 02-40; USGS: Reston, VA, 2002. (16) 1996 Clean Water Needs Survey for the United States and U.S. Territories; U.S. EPA: Washington, DC, 1997. (17) Shanahan, P.; Harleman, D. R. F. J. Environ. Eng. 1984, 110, 42-57.

(18) Lawrence, G. A.; Ashley, K. I.; Yonemitsu, N.; Ellis, J. R. Limnol. Oceanogr. 1995, 40, 1526-1532. (19) Selection criteria for mathematical models used in exposure assessments: Surface water models; Report Number EPA/600/ 8-87/042; Office of Health and Environmental Assessment, U.S. Environmental Protection Agency: Washington, DC, July 1987. (20) Oreskes, N.; Shrader-Frechette, K.; Belitz, K. Science 1994, 263, 641-646. (21) Reckhow, K. H.; Chapra, S. C. Ecol. Modell. 1983, 20, 113-133. (22) GREAT-ER, 2002, URL: http://www.great-er.org/pages/home. cfm. (23) Barnes, K. K.; Kolpin, D. W.; Meyer, M. T.; Thurman, E. M.; Furlong, E. T.; Zaugg, S. D.; Barber, L. B. USGS Open-File Report 02-94; USGS: Reston, VA 2002; url: http://toxics.usgs.gov/pubs/ OFR-02-94/index.html. (24) Tabor, C. F.; Barber, L. B. Environ. Sci. Technol. 1996, 30, 161171. (25) Surface-Water Data for the Nation; USGS: Reston, VA, 2002; url: http://waterdata.usgs.gov/nwis/sw. (26) Sanders, D. H.; Murph, A. F.; Eng, R. J. Statistics: A Fresh Approach, 2nd ed.; McGraw-Hill: New York, 1980. (27) Retail and Provider Perspective data for year 2000; IMS Health Inc.: Fairfield, CT, August 2001. (28) Retail and Provider Perspective data for year 2000; IMS Health Inc.: Fairfield, CT, December 2001. (29) Johnson, A. C.; Sumpter, J. P. Environ. Sci. Technol. 2001, 35, 4697-4703. (30) Cowan, C. E.; Larson, R. J.; Feijtel, T. C.; Rapaport, R. A. Water Res. 1993, 27, 561-573. (31) Bristol-Myers Squibb, 2003, private communication. (32) Stumpf, M.; Ternes, T. A.; Wilken, R.-D.; Rodrigues, S. V.; Baumann, W. Sci. Total Environ. 1999, 25, 135-141. (33) Ternes, T. A. Water Res. 1998, 32, 3245-3260. (34) Aherne, G. W.; Briggs, R. J. Pharm. Pharmacol. 1989, 41, 735736. (35) McQuillan, D.; Parker, J.; Chapman, T. H.; Sherrell, K.; Mills, D. Drug residues in ambient water: initial surveillance in New

(36) (37)

(38)

(39) (40) (41) (42)

(43) (44)

(45)

Mexico, USA. NGWA, 2nd International Conference on Pharmaceuticals and Endocrine Disrupting Chemicals in Water, October 9-11, 2001. Kuch, H. M.; Ballschmiter, K. Environ. Sci. Technol. 2001, 35, 3201-3206. Belfroid. A. C.; Van der Horst, A.; Vethaak, A. D.; Scha¨fer, A. J.; Rijs, G. B. J.; Wegener, J.; Cofino, W. P. Sci. Total Environ. 1999, 225, 101-108. Snyder, S. A.; Keith, T. L.; Verbrugge, D. A.; Snyder, E. M.; Gross, T. S.; Kannan, K.; Giesy, J. P. Environ. Sci. Technol. 1999, 33, 2814-2820. Huang, C.-H.; Sedlak, D. L. Environ. Toxicol. Chem. 2001, 20, 133-139. Ternes, T. A.; Stumpf, M.; Mueller, J.; Haberer, K.; Wilken, R.-D.; Servos, M. Sci. Total Environ. 1999, 225, 81-90. Aherne, G. W.; English, J.; Marks, V. Ecotoxicol. Environ. Saf. 1985, 9, 79-83. Brown, L.; du Preez, J. L.; Meintjies, E. The qualitative and quantitative evaluation of estrogen and estrogen-mimicking substances in the South African water environment: a 19961997 perspective. NGWA, 2nd International Conference on Pharmaceuticals and Endocrine Disrupting Chemicals in Water, October 9-11, 2001. Fawell, J. K.; Sheahan, D.; James, H. A.; Hurst, M.; Scott, S. Water Res. 2001, 35, 1240-1244. Response to Environmental Concerns Regarding The Use of Synthetic Estrogens in Human Drugs; Docket 96N-0057; U.S. Food and Drug Administration: Washington, DC, 1997. Ericson, J. F.; Laenge, R.; Sullivan, D. E. Environ. Sci. Technol. 2002, 36, 4005-4006.

Received for review May 5, 2003. Revised manuscript received September 19, 2003. Accepted October 21, 2003. ES034430B

VOL. 38, NO. 3, 2004 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

849

Screening Analysis of Human Pharmaceutical Compounds in U.S.

Recommend Documents