United States
Environmental Protection
Agency
Contaminant Candidate List 3
Chemicals: Classification of the
PCCL to CCL
Office of Water (4607M) EPA 815-R-08-004 February 2008 - Draft www.epa.gov/safewater
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
Table of Contents
1.0 INTRODUCTION TO THE CONTAMINANT CANDIDATE LIST (CCL)
CLASSIFICATION PROCESS 1
1.1 Principles of Evaluation 1
1.2 Developing the Classification Approach 2
2.0 ATTRIBUTES 4
2.1 Health Effects Attributes 5
2.1.1 Potency 5
2.1.2 Severity 14
2.2 Occurrence Attributes 18
2.2.1 Prevalence and Magnitude Data Elements 19
2.2.2 Prevalence - Calibrating Scales and Scoring 20
2.2.3 Evaluation of the Prevalence Protocol 21
2.2.4 Magnitude - Calibrating Scales and Scoring 22
2.2.5 Persistence-Mobility as a Surrogate Measure for Magnitude 27
2.2.6 Persistence-Mobility Data - Calibrating Scales and Scoring 28
2.2.7 Evaluation of the Magnitude Protocol 29
2.3 Fine Tuning the Protocols 30
3.0 DEFINITIONS AND OVERVIEW OF THE TRAINING DATA SET 31
3.1 Key Considerations 31
3.2 Developing Key Components of the Training Data Set 31
3.2.1 Attribute Scores 31
3.2.2 Making List-Not list Decisions 35
4.0 PROTOTYPE CLASSIFICATION MODELS AND THE CCL PROCESS 36
4.1 Model Training and Development 37
4.2 Model Sensitivity Analyses 39
4.2.1 Training with sub sets of the IDS 39
4.2.2 Training after Selected "Outliers" Are Removed From the IDS 40
4.2.3 Graphical and Statistical Analyses to Identify Significant Differences in
Attribute "Weights" Or Influence on Model Performance 41
4.3 Model Performance Testing 43
4.4 Evaluating Classification Differences 44
4.4.1 Classification Differences Among the Models 45
4.4.2 Logical Evaluation of the Models - Graphical Analysis 47
4.5 Applying Model Results 53
4.5 Applying Model Results 54
4.5.1 Additive Model Results 54
4.5.2 Additive Rank Order Results 54
i of vi
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
5.0 MODEL OUTCOME AND POST MODEL EVALUATION PROCESS 56
5.1 PCCL Characterization and Model Results 56
5.2 Evaluation of the Modeling Output 57
5.2.1 Procedure 57
5.2.2 Evaluation Results 58
5.3 Post-Model Adjustments to Output 60
5.3.1 Using Supplemental Sources to Identify the Data Most Relevant to Drinking
Water 60
5.3.2 Calculation of a Health-Concentration Ratio for Contaminants with Water
Data 61
5.3.3 Grouping Contaminants based on Data Certainty 62
5.3.4 LDso Values with Limited Documentation 64
5.4 Selecting the Draft CCL 3 64
5.5 Summary 64
6.0 REFERENCES 65
7.0 APPENDICES 66
Appendix A. Attribute Scoring Protocols A-l
Appendix B. Example Blinded Information Sheets from the TDS Exercises B-l
Appendix C. Summary of EPA Team TDS Decisions C-l
Appendix D. Software Sources D-l
Appendix E. Solutions E-l
Appendix F. Chemicals Reviewed by the EPA Evaluation Team: Summary of
Results F-l
Appendix G. PCCL Contaminants with Incomplete Data for Scoring or that had
Parent Compounds Scored G-l
ii of vi
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
Table of Exhibits
Exhibit 1. Developing an Approach to Process PCCL Chemicals 3
Exhibit 2. Decile Distribution of RfD Values 7
Exhibit 3A. Logarithmic Distribution of RfD Values 8
Exhibit 3B. Logarithmic Distribution of NOAEL Values 8
Exhibit 3C. Logarithmic Distribution of LOAEL Values 9
Exhibit 3D. Logarithmic Distribution of LDso Values 9
Exhibit 4. Scoring Equations for Potency 11
Exhibit 5. Logarithmic Distribution of Cancer Potency Values 12
Exhibit 6. Potency Scores for Chemicals in the Learning Set 13
Exhibit 7. Potency Scores for Chemicals Not in the Learning Set 13
Exhibit 8. NRC Severity Scoring Proposal 15
Exhibit 9. Final Nine-Point Scoring Protocol for Severity 16
Exhibit 10. Relationship of Data Elements Used to Score Magnitude and Prevalence 20
Exhibit 11. Comparison of Prevalence Scores for Learning Set Contaminants 22
Exhibit 12. Comparison of the NRC Magnitude Score with the Ratio of the Health
Advisory Guideline to the Concentration in Finished Water 23
Exhibit 13: Magnitude Concentrations and Scores Derived from Potency Doses 24
Exhibit 14A. Equal Bins Drinking Water Magnitude Scale 25
Exhibit 14B. Half Log Option A Drinking Water Magnitude Scale 26
Exhibit 14C. Half Log Option B Drinking Water Magnitude Scale 26
Exhibit 15. Magnitude Attribute Scores: Example Contaminants Scored by their Median
of Detections Using the Various Approaches in Exhibit 14 27
Exhibit 16. Mobility and Persistence Data Elements 28
Exhibit 17. Comparison of Scores derived using the Magnitude Protocol 30
Exhibit 18. Combinations of low and high attribute scores1 for the four attributes using
Latin Hypercube Sampling 33
Exhibit 19. Attribute Space for the 101 IDS compared to that for the 202 IDS 34
Exhibit 20a. QUEST Classifications Based on the Full Training Data Set 38
Exhibit 20b. QUEST Classifications Based on 5-Fold Cross-Validation 38
Exhibit 21. Model-estimated versus Team Average Classification for the TDS 41
Exhibit 22. Relative Weights of Attributes at QUEST Nodes 42
Exhibit 23. Summary Statistics from MCMC Sample 43
iii of vi
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
Exhibit 24. Features of the Three Preferred Models Based on TDS Test Results 44
Exhibit 25. Decision Comparison Matrix; Weight of Differences 45
Exhibit 26. Summary of Quaternary Model Decisions 46
Exhibit 27. Results of 202 Model Classifications and Weighted Misclassifications 46
Exhibit 28. Summary of Individual Quaternary Model Classifications 48
Exhibit 29. ANN Model Predictions for the Four Attribute Space 49
Exhibit 30. MARS Model Predictions for the Four Attribute Space 51
Exhibit 31. Univariate CART Model Predictions for the Four Attribute Space 52
Exhibit 32. Linear Model Predictions for the Four Attribute Space 53
Exhibit 33. Summary Comparison of the Sum of the 3 Model Decisions to the
Distribution of the Workgroup Blinded (TDS) Decisions 55
Exhibit 34. Model Results for the PCCL Chemicals 56
Exhibit 35. Results of the Model Output Evaluation (Total = 129 chemicals) 60
iv of vi
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
List of Abbreviations and Acronyms
< Less than
< Less than or equal to
> Greater than
> Greater than or equal to
jig Microgram, one-millionth of a gram
|ig/L Micrograms per liter
ANN Artificial Neural Network
ATSDR Agency for Toxic Substances and Disease Registry
CART Classification and Regression Tree
CASRN Chemical abstract services registry number
CCL Contaminant Candidate List
CCL 1 EPA's first contaminant candidate list
CCL 2 EPA's second Contaminant Candidate List
CCL 3 EPA's third Contaminant Candidate List
CUS/IUR Chemical update system/inventory update rule
DBF Disinfection byproduct
EDWC Estimated Drinking Water Concentration
EEC Estimated Environmental Concentration
EPA United States Environmental Protection Agency
g/day Grams per day
HRL Health Reference Level
IOC Inorganic compound
IRIS Integrated Risk Information System
kg Kilogram
L Liter
LD50 Lethal dose 50; an estimate of a single dose that is expected to cause the death
of 50 percent of the exposed animals; it is derived from experimental data.
Ibs Pounds
LOAEL Lowest observed adverse effect level
MARS Multivariate Adaptive Regression Splines
MCMC Markov Chain Monte Carlo
mg Milligram, one-thousandth of a gram
mg/kg Milligrams per kilogram body weight
v of vi
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
mg/kg/day Milligrams per kilogram body weight per day
mg/L Milligrams per liter
N Number of samples
NAWQA National water quality assessment (USGS program)
NCOD National contaminant occurrence database
ND Not detected (or non-detect)
NOW AC National Drinking Water Advisory Council
MRS National Inorganic and Radionuclide Survey
NOAEL No observed adverse effect level
NRC National Research Council
OW Office of Water
OPP Office of Pesticide Programs
PBPK Physiologically Based Pharmacokinetic
PCCL Preliminary-CCL
PWS Public water system
QUEST Quick, Unbiased, Efficient Statistical Tree
RTECs Registry of Toxic Effects of Chemical Substances
RfD Reference dose
TDS Training data set
TRI Toxics Release Inventory
UCMR Unregulated Contaminant Monitoring Regulations
UCMR 1 First Unregulated Contaminant Monitoring Regulation
UCMR 2 Second Unregulated Contaminant Monitoring Regulation
UL Tolerable upper intake level
US United States of America
USGS United States Geological Survey
vi of vi
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
1.0 INTRODUCTION TO THE CONTAMINANT CANDIDATE LIST
(CCL) CLASSIFICATION PROCESS
The United States Environmental Protection Agency (EPA) developed a multi-step approach to
select contaminants for the third CCL (CCL 3), which includes the following key steps:
(1) The identification of a broad universe of potential drinking water contaminants
(CCL 3 Universe);
(2) A screening process that uses straightforward screening criteria, based on a
contaminant's potential to occur in public water systems and thereby pose a
potential public health concern, to narrow the universe of contaminants to a
Preliminary-CCL (PCCL); and
(3) A structured classification process (e.g., a prototype classification algorithm
model) that objectively compares data and information as a tool and is evaluated
along with expert judgment to develop a CCL from the PCCL.
Steps 1 and 2 in the process are described in other support documents: CCL 3 Chemicals:
Identifying the Universe (EPA, 2008a); and CCL 3 Chemicals: Screening to a PCCL (EPA,
2008b). The purpose of this document is to describe the methodology used to develop the
classification process (Step 3) and the process used to select chemicals for the CCL 3.
The PCCL consisted of 532 chemicals that were screened from the CCL3 Universe. In order to
select contaminants for the CCL 3, EPA used classification models to handle larger, more
complex assortments of data in a consistent and reproducible manner. Learning from EPA's
experience and expertise, the classification models were trained based on past expert decisions.
The algorithms were used to prioritize chemicals which allowed the final expert evaluation and
review to be more objective and efficient.
1.1 Principles of Evaluation
In developing the first CCL (CCL 1), the Agency utilized readily available occurrence and health
effects information coupled with an expert review process. Following the publication of CCL 1,
the Agency sought the advice of the National Research Council (NRC) and National Drinking
Water Advisory Council (NDWAC). The panels provided recommendations to guide EPA in
creating a more comprehensive and transparent evaluation of potential drinking water
contaminants for developing future CCLs. In the light of the NRC and NDWAC
recommendations, EPA has reviewed and evaluated a large number of contaminants and their
data, developed decision making protocols using classification algorithm approaches, and
included expert review in arriving at decisions to list or not list contaminants on CCL 3. These
steps have provided a decision process that is more transparent and reproducible than approaches
used for previous CCLs. The process is driven by the data on individual contaminants and
minimizes the bias that may occur with expert panels due to the participants' individual
Page 1 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
backgrounds and the confounding effects of group dynamics. As experience is gained, the new
classification process is likely to evolve and improve for application to future CCLs.
To guide the development of the classification process, EPA identified several key features that
the approach addresses.
1. Meaningful Basis for Classification. The classification process must reflect the critical
goals of the CCL; that is, it must consider the potential for occurrence in water, the
potential for causing adverse health effects, and it must prioritize contaminants based on
these criteria. The data supporting the list no-list decision must be linked back to these
three tenets.
2. Incorporating Relevant Data. The most relevant data used for the classification process
include health effects data that are appropriate for drinking water exposures, and
occurrence data that indicate the nature and spatial extent of potential occurrence in
drinking water.
3. Transparent Process for Communication. One goal of the classification approach is to
provide a transparent process that can be reviewed by external experts and the public.
The attributes and data characterizing the contaminants should be easy to understand and
the decision-making process to list or not list a particular contaminant must be conveyed
in a straight forward manner.
4. Reproducibility. A key feature of the classification process is that it should be
reproducible. The classification process should always give the same result for the same
set of input information.
1.2 Developing the Classification Approach
Based on this framework, EPA developed an approach for classifying potential drinking water
contaminants. An overarching premise in using classification models to prioritize contaminants
is that different contaminants can be compared on the basis of similar attributes. The approach
ensures that the contaminant attributes reflect the key decision characteristics in deciding
whether or not to list a contaminant on the CCL. The attributes are properties used to categorize
contaminants for their potential to occur in drinking water and for their potential to cause adverse
health effects. For example, occurrence can be characterized by a contaminant's water
concentration data or potential to occur based on its release to the environment. The adverse
health effects of contaminants can be characterized using preliminary toxicological data such as
median lethal dose (LDso) or more developed values such as oral reference doses (RfDs). To
evaluate, categorize, and prioritize the PCCL contaminants as potential CCL contaminants, EPA
integrated various types of data that represent measures of their attributes. This relative
assessment across data measures normalized the available data by developing a set of attribute
scales for the attribute data, and scoring mechanisms for the various types of data available for
potential drinking water contaminants.
Because of this new approach and its new application, EPA developed, tested, and evaluated the
results of several classification algorithms to assess whether they are useful, and which ones
Page 2 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
might provide the best decision support tools. To test and evaluate the process, EPA developed a
data set and used it to "train" the classification algorithms. Once the modeling was completed,
an Evaluation Team evaluated the model output based on the compilation of data for a subset of
the modeled contaminants and assisted in developing a process to utilize the model output to
generate the CCL 3. The following chapters describe the steps EPA used to develop the
components of the classification process, as displayed in Exhibit 1.
Exhibit 1. Developing an Approach to Process PCCL Chemicals
Develop Attribute
Scoring Protocol
Select Training Data Set
Contaminants and Make
Listing Decisions
Score Training Data Set
Contaminants with Final Attribute
Scoring Protocols
Train and Validate Classification
Approaches using Training Data Set
Iterative Process -
The results of training
and validation will
indicate if areas need
further evaluation and
refinement. The iterative
process may or may not
go back to the primary
assumptions.
Post-model
evaluation of PCCL
chemicals
Chapter 2 describes the attributes and scoring protocols. Chapter 3 describes the set of chemicals
used to train the classification models, the training data set. Chapter 4 describes how the models
were calibrated using the attributes and training data set. Chapter 5 describes the evaluation of
the model output and post model processes.
PageS of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
2.0 ATTRIBUTES
Attributes are used to characterize different chemicals on the basis of similar qualities or traits.
These qualities or traits represent the anticipated occurrence or adverse health effects of each
contaminant. Occurrence and health effects are both represented by different types of data. To
evaluate contaminants as potential CCL contaminants, one must be able to establish consistent
relationships among the different types of data that represent measures of the attributes. This
process involves the need to normalize the available data by developing scales and scoring
mechanisms that will accept a variety of input data. The attributes are properties used to
categorize contaminants for their potential to occur in drinking water and for their potential to
cause adverse health effects. For example, occurrence may be characterized by water
concentration data or a contaminant's potential to occur based on its release to the environment.
The adverse health effects of contaminants may be characterized using preliminary toxicological
data such as median lethal dose (LDso) or more developed values such as oral reference doses.
The NRC recommended using the attributes Potency and Severity to describe health effects, and
Prevalence and Magnitude to describe occurrence. When occurrence data are not available, they
also suggested that environmental fate properties (i.e. Persistence and Mobility) could be used as
surrogates to estimate potential for occurrence. The EPA workgroup agreed that the
recommended attributes are appropriate and consistent with data used in the past decision-
making efforts by EPA's Office of Water (OW).
Throughout the process of evaluating the attributes, it was recognized that a wide range of data
elements would have to be used to characterize each attribute. The CCL process involves
classifying relatively new and emerging contaminants and most will not have complete dossiers
of data. If the same data were available for all chemicals their comparison and prioritization
would be relatively straight forward. However, the types of data available for unregulated
chemical contaminants varies. To enable comparisons among chemicals with differing types of
data and information, a scaling system that accepts a variety of input data, yet provides a
consistent comparative framework, is needed. In concert with NRC and NDWAC
recommendations, EPA identified the following principles to guide development of the attribute
scoring process:
• Attribute scores should increase with concern (e.g., a 10 is of greater concern, 1 of lesser
concern);
• There should be sufficient scoring categories to capture the range of data and to
discriminate among the data;
• The number of categories should not be so great that they create a false sense of
precision;
• Attributes can use different numbers of scoring categories if necessary (i.e., Prevalence
could use 1-10, while Severity could use 1-8);
• The possible range of the scores for a given attribute should be the same regardless of the
data elements that are used to assign the score for that attribute;
• The data source and data element used for each attribute should consider more direct
measures of occurrence or health effects before potential measures; peer reviewed data
before unpublished data, and measured data before modeled data.
Page 4 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
• The calibration scale (i.e., the scale relating the range for a data element to the scoring
categories) should be established using a representative "universe" of data for each
attribute to capture the potential range of values that might be encountered;
• The calibration scale must be set and remain constant throughout the operational process;
and
• The scoring approach should be as simple as possible and data should be used with
minimal transformations.
Section 2.1 describes the development of the process used to score the health effects attributes,
and section 2.2, the approach for the occurrence attributes.
2.1 Health Effects Attributes
Potency and Severity are the two attributes used for evaluating health effects. As defined in
detail below, Potency reflects the lowest dose of a chemical that causes an adverse health effect
in a case study report or in a toxicological or epidemiological study. Severity is the adverse
health effect associated with the dose that is used as the measure of Potency, and is calibrated
based on the health-related significance of the adverse effect (e.g., dermatitis versus cancer).
These two attributes are interrelated, in that the Severity is linked to the measure of Potency.
2.1.1 Potency
Potency is a value that indicates the power of a contaminant to cause adverse health effects. In
the case of chemicals, that power is apparent in the dose required to cause the most sensitive
manifestation of an adverse health effect, or to generate a particular excess cancer risk. Potency
for chemicals is reflected in several standard toxicological parameters that are discussed below.
A number of approaches have the potential to be useful in scoring the Potency attribute.
However, regardless of the approach selected, the methods require calibrating the scores to
normalize the scale. To evaluate the data elements and establish consistent scales, an initial
"learning set" of about two hundred chemicals was developed for use in experimentation with
approaches to calibration. The chemicals considered included regulated chemicals and
unregulated chemicals for which EPA has derived Health Advisories (EPA, 2004). These
chemicals are primarily at the high end of the Potency scale. To ensure that the Potency scale
covers the full range of conditions that may be encountered (from high to low Potency) in a
universe of chemicals, a group of chemicals (nutrients/food additives) that are generally
considered as relatively non-toxic and have toxicity values that can be compared to health
advisories were added to the learning set.
PageS of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
The following toxicity parameters were compiled for the learning set chemicals, and their
numeric distribution across the range of values was examined (see the footnotes below for
definitions of the terms).
• Reference Dose (RfD)1 or equivalent
• Cancer potency2 (concentration in water equivalent to a 10"4 cancer risk)
• No Observed Adverse Effect Level (NOAEL)3 and/or Lowest Observed Adverse Effect
Level (LOAEL)4 associated with the RfD
• Rat oral median Lethal Dose
Several approaches to characterize the distribution of values for the different toxicity parameters
were employed in this exercise. The approaches are described in the following section.
The data for the learning set were obtained from the following sources:
• EPA's Integrated Risk Information System (IRIS)
• EPA's Office of Water Health Advisories Documents6
• Registry of Toxic Effects of Chemical Substances (RTECS) (Mostly LDso values)
• Tolerable Upper Intake Levels (ULs) from the Institute of Medicine Dietary Reference
Intakes.
1 A Reference Dose (RfD) is an estimate (with uncertainty spanning perhaps an order of
magnitude) of a daily exposure to the human population (including sensitive subgroups) that is likely to
be without an appreciable risk of deleterious effects during a lifetime. It is expressed in mg/kg/day. The
Agency for Toxic Substances and Disease Registry (ATSDR) lifetime Minimal Risk Levels (MRLs),
World Health Organization (WHO) Tolerable Daily Intakes (TDIs), WHO and Food and Drug
Administration (FDA) Acceptable Daily Intakes (ADIs), and the Institute of Medicine (IOM) nutrient
Tolerable Upper Intake Levels (ULs) are roughly equivalent to the RfD.
2
For this exercise cancer potency was evaluated as the concentration in drinking water
,-4\
equivalent to an excess cancer risk of one case in 10,000 (10" ). This value is given in the Office of Water
(OW) Drinking Water Standards and Health Advisories Tables and also is included in all Integrated Risk
Information System (IRIS) Summary documents. When the 10"4 risk value is not available, it can be
calculated from a cancer slope factor.
3 NOAEL is a No-Observed-Adverse-Effect Level. It is the highest dose in a toxicological study
or a group of studies that has no observed adverse effect.
LOAEL is a Lowest-Observed-Adverse-Effect Level. It is the lowest dose in a toxicological
study or a group of studies that causes an adverse health effect.
5 An oral median Lethal Dose (LD50) is an estimate of the oral dose that will cause the death of
50 percent of the exposed animals. LD50 data are based on acute exposures with limited post-exposure
observations of the animals for cause of mortality, clinical signs, and gross pathology.
6 The 2002 Edition of the Drinking Water Standards and Health Advisories was used for the
RfD and 10"4 risk values.
Page 6 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
2.1.1.1 Potency Data - Calibrating Scales and Scoring
Once the data for the learning set of chemicals was collected, they were arrayed and graphically
displayed to analyze their range and distribution. For the initial evaluation, the range (in
mg/kg/day) was divided into approximately ten equal units (deciles). This distribution was
found to be highly skewed, with a large majority of the values falling in the decile of highest
toxicity (see Exhibit 2 for an example). Two factors influenced this result. The first factor is
that the range of values covered up to twelve orders of magnitude for the parameters evaluated.
The second factor is that the set of contaminants contained both toxic chemicals as well as those
generally regarded as safe (in keeping with the principles) and there are far more toxicological
data available in the literature on chemicals considered to be toxic than for those, like the
nutrients, that are only weakly toxic. This shifts the volume of data toward the chemicals with
higher potencies. Most chemicals that are generally regarded as safe have limited available
toxicological data, as their nutritional and commercial uses do not indicate a potential hazard at
low to moderate intakes.
Exhibit 2. Decile Distribution of RfD Values
>0.1-0.2 >0.2-0.3 >0.3-0.4 >0.4-0.5 >0.5-0.6 >0.6-0.7 >0.7-0.8 >0.8-0.9
The second distribution evaluated was based on logarithms (base 10) of the toxicity parameters
rounded to the nearest integer (see Exhibit 3 A-D as examples).
Page? of66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 3A. Logarithmic Distribution of RfD Values
-2-101
Round(Log10(RfD))
Exhibit 3B. Logarithmic Distribution of NOAEL Values
-1012
Round(Log10(NOAEL|)
PageS of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 3C. Logarithmic Distribution of LOAEL Values
<=-5 -A
-2-1012345 More
Round(Log10(LOAEL))
Exhibit 3D. Logarithmic Distribution of LDso Values
2345
Round(Log10(LD50))
Page 9 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
The decile distribution (Exhibit 2) was found to be undesirable in developing a protocol for
scoring Potency because almost all of the chemicals are clustered at one end of the distribution.
This does not provide a good distribution of scores for discrimination of differences. With the
decile distribution, almost all of the chemicals in the learning set would have a high Potency
score of 10. Very few chemicals would have lower scores. The distribution based on the
rounded Logio of the toxicity parameter provided a distribution that spread the chemical toxicity
parameters across the range and the most frequent Logio value is approximately in the middle of
the range making the curve roughly log-normal Exhibit 3 A-D). It was for this reason that the
Log10 distribution was selected for development of the scoring equation. The distribution of
toxicity values is still somewhat skewed toward higher toxicity scores; however, this is a product
of limited available data for the weakly toxic chemicals.
The log-based distribution was used to establish a scoring equation for Potency for each measure
of toxicity. This was accomplished by assigning the most frequent (modal) value in the
distribution a score of 5 on a 10 point scale and solving an equation for each type of toxicity
parameter that would make that distributional value equal a score of 5. For example, in Exhibit
3 A (RfD), the most frequent value is a rounded logarithm of -2 (0.01). The scoring equation for
the RfD values was developed as follows:
5 = 10- (most frequent rounded log + X)
5 = 10-(-2 + X)
5 = 10 + 2-X
5 = 12-X
5 - 12 = -X
-7 = -X
7 = X
Accordingly the equation for scoring the RfD values is
Score = 10 - (rounded log of RfD + 7)
The scoring equations for the other measures of toxicity were derived from the modal rounded
logarithm values of their distributions in a similar fashion. As displayed in Exhibit 3, the
position of the modal rounded log differed for the different measures of toxicity, which
necessitated differing equations. The resultant equations are summarized in Exhibit 4.
Page 10 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
Exhibit 4. Scoring Equations for Potency
RfD Score = 10 - (Log10 of RfD + 7)
NOAEL Score = 10 - (Log10 of NOAEL + 4)
LOAEL Score = 10 - (Log10 of LOAEL + 4)
LD50 Score = 10 - (Logio of LD50 + 2)
10"4 cancer risk l Score = 10 - (Logic of the 10"4 cancer risk + 6)
1 The concentration in water for 10"4 cancer risk in water was selected as the measure of potency
for carcinogens because this is the value given in the Standards and Drinking Water Health
Advisories Tables prepared by OW and also is provided in IRIS Summaries. Changing the
reference value to the 10"6 risk would merely shift the rounded log value and the constant by two
integers but would not change the score.
Scores were restricted to whole number values with a maximum of 10 and a minimum of 1.
Some distributions for toxicity parameters span a range greater than ten orders of magnitude.
EPA decided that calculated scores less than 1 would be given scores of 1 and calculated scores
greater than 10 would be given scores of 10, which combine the chemicals at the tails of the
distributions. Conversely, for the distributions that covered less than 10 orders of magnitude, no
attempt was made to normalize the scores across a range often because the learning set is limited
and could have been expanded by searching for chemicals that are more toxic than the most toxic
substance in the learning set (dioxin with an RfD of 1 x 10"9 mg/kg/day) and less toxic than the
least toxic chemical in the learning set (phosphorous with an RfD-equivalent of 57 mg/kg/day
derived from the Institute of Medicine (IOM) UL. However an adjustment was made to
accommodate LDso values that are reported as greater than a specific numerical dose. In such a
case, the highest dose used in the study did not cause death in 50 percent of the tested animals,
indicating that the chemical is less toxic than would be indicated by the highest dose tested.
Accordingly, the LDso equation was modified to accommodate this situation and became:
LD50 Score = 10 - (Log10 of >LD50 + 3)
This change to the LD50 equation decreases the Potency score from that derived from the numeric
value of the LD50 by one to accommodate the "greater than" designation. A similar adjustment
was made for situations where the NOAEL in a critical study was the highest dose tested.
The distribution for cancer effects is the most skewed of those examined (see Exhibit 5). There
are a greater number of chemicals that are more potent carcinogens when compared to those in
the modal grouping than there are those that are less potent. This is not unusual because cancer
bioassays are costly and there is an incentive to invest resources in studying chemicals that have
a high likelihood of being potent carcinogens. No attempt was made to further normalize the
cancer scores across a range of 10. For the chemicals in the learning set, the lowest cancer
Potency score is 3.
Page 11 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 5. Logarithmic Distribution of Cancer Potency Values
-2-10 1
Round(Log10(E4))
2.1.1.2 Evaluation of the Potency Scoring Protocol
All of the chemicals in the learning set were scored for each toxicity parameter to examine the
consistency across scores for the non-cancer measures of Potency. Some examples of this
evaluation are provided in Exhibit 6. Since the mechanisms that lead to the development of
cancer involve some biological responses that are unique to tumors, the 10"4 cancer risk values
were not included in this comparison. The scores for individual chemicals were compared across
the toxicity values, and the agreement between scores was evaluated.
Page 12 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 6. Potency Scores for Chemicals in the Learning Set
Chemical
Calcium (Calcium chloride for LD50)
Cyanazine
Dioxin (2,3,7,8-TCDD)
Hexazinone
Iodine (Sodium iodide for LD50)
Methyl ethyl ketone
Methyl parathion
Naphthalene
Phenol
Vitamin D
RfD
1
6
10
4
5
O
7
5
4
6
NOAEL
ND
6
ND
5
8
3
8
4
4
9
LOAEL
4
6
10
4
8
O
7
4
4
9
LD50
5
6
4
5
4
5
7
5
5
ND
ND = No data
In addition, the scoring equations were applied to selected chemicals that were not in the learning
set using data available in the Agency of Toxic Substances and Disease Registry (ATSDR)
Toxicological Profiles. Those results are summarized in Exhibit 7. The scores were evaluated
for consistency across parameters.
Exhibit 7. Potency Scores for Chemicals Not in the Learning Set
Chemical/
Potency Scores
Acrylonitrile
Ethion
Malathion
Endosulfan
RfD-equivalent
(mg/kg/day)
4
6
5
6
NOAEL
(mg/kg/day)
5
7
6
7
LOAEL
(mg/kg/day)
5
6
5
ND
LD50
(mg/kg)
6
6
5
5
ND = No Data
The agreement of non-cancer scores across the RfD, NOAEL, LOAEL and LD50 inputs was
evaluated. There were 216 chemicals in the learning set; 13.5 percent of those with multiple non-
cancer scores had identical scores across all parameters (see cyanazine in Exhibit 6). For 54.6
percent, the scores deviated by 1 integer (see hexazinone in Exhibit 6); 20.5 percent deviated by
2 integers (see methyl ethyl ketone in Exhibit 6). There was a 3-integer deviation for only 9.7
percent, and the majority of those were inorganic compounds (see iodine [sodium iodide] in
Exhibit 6). Only 1.6 percent deviated by more than 3 integers (see dioxin in Exhibit 6). Scores
deviated by two integers or less for 88.6 percent of the chemicals. The difference between scores
Page 13 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
for a given compound was greatest for the relatively non-toxic chemicals. In almost all cases the
NOAEL and LOAEL scores were higher than the RfD score, effectively negating the concerns
that the inclusion of uncertainty factors in the calculation of the RfD would inflate the Potency
score. For those chemicals with low uncertainty factors the NOAEL or LOAEL scores were
often 3 or more integers higher than the RfD scores (see calcium chloride and vitamin D in
Exhibit 6).
Since most chemicals with RfD values are also likely to have NOAEL, LOAEL, and/or
values, a policy decision was needed with regard to how one should select the parameter used to
score for a non-cancer endpoint. Since there is a general consistency among scores, the EPA
workgroup determined that a hierarchy of RfD> NOAEL> LOAEL> LD50 would be used. In
cases where a NOAEL is higher than the lowest LOAEL, the LOAEL would be used in its place.
This hierarchy gives preference to the Potency value with the richest supporting data set (the
RfD-or equivalent values) and the lowest ranking to the LD50 because it is a measure of acute
rather than chronic toxicity. When comparing cancer and non-cancer scores, it was determined
that the end point (cancer or non-cancer) that provided the highest measure of Potency would be
used to score the candidate.
These evaluations were used to develop the scales and hierarchy of data used in the Potency
Scoring Protocol, which is presented in Appendix A.
2.1.2 Severity
Severity refers to the relative impact of an adverse physiological change caused by a xenobiotic
chemical in humans or animals on the ability of the human or animal to function and survive in
the environment. The sixteenth century physician, Paracelsus, provided the underlying principle
for the toxicological sciences with the axiom "the dose makes the poison." Just as toxicity
increases with dose, so too does the Severity of the observed effect, in most cases. A low dose
effect could be a simple increase in liver weight while the same chemical at a higher dose could
cause cirrhosis of the liver. For that reason, the measure of Severity that will be used for scoring
in the CCL process is the effect or effects seen at the LOAEL. Restricting Severity scores to the
effects occurring at the LOAEL ties them to the data used to derive the Potency score - the type
of data likely to be available for CCL candidates. This approach is consistent with the advice
provided by the NRC and NOW AC (NRC 2001, NOW AC 2004).
The Severity measures that will be used for CCL scoring differ from those used for Potency,
Prevalence, and Magnitude because they are descriptive rather than quantitative. Accordingly,
they are less amenable to automation and often require more scientific judgment in their
application. The sections that follow describe the approach that was used to derive the scoring
protocol for Severity and to evaluate its performance.
2.1.2.1 Severity - Scales and Scoring
In developing the protocol for scoring Severity, the workgroup began with the system used by
the NRC (2001) for their case study on methods for selecting a CCL from a PCCL. The NRC
Severity scoring protocol was based on the anticipated clinical impact of the most sensitive
endpoint in affected individuals. The NRC prototype for scoring Severity is provided in Exhibit
Page 14 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
Exhibit 8. NRC Severity Scoring Proposal
Score Description
0
No effect
1
Changes in organ weights with minimal clinical significance
Biochemical changes with minimal clinical significance
Pathology of minimum clinical importance (e.g., fluorosis, warts, common cold)
4
Cellular changes that could lead to disease; minimum functional change
5
Significant functional changes that are reversible (e.g., diarrhea)
6
Irreversible changes; treatable disease
7
Single organ system pathology and function loss
Multiple organ system pathology and function loss
Disease likely leading to death
10
Death
In trying to apply the NRC Severity prototype using the critical effects from EPA IRIS Health
Risk Assessments, EPA toxicologists encountered difficulty because of the clinical components
of the prototype. It was difficult to determine clinical outcomes such as function loss,
treatability, or potential for mortality from the critical effects identified in IRIS. In addition,
some of the features of a clinical progression could be influenced by the availability and
affordability of treatment. The workgroup decided that it would not be appropriate to use a
scoring scheme that had economic and environmental justice implications.
The critical effect data for PCCL contaminants will, in most cases, be expressed using
terminology very similar to the terminology found in the IRIS database. Accordingly, critical
effects of 100 IRIS chemicals were compiled and grouped into categories by EPA toxicologists.
These categories were, in turn, used to build a scoring scale that applied some of the rationale
reflected in the NRC prototype, but utilized the critical effects information most likely to be
available from databases such as IRIS, which eliminated outcome judgments that would
confound the scoring process. In this exercise, some difficulties were encountered in scoring
Severity, particularly with assigning the middle score categories (3, 4, 5, and 6) and with
classifying different types of cancer. Accordingly, the scoring protocol was modified again to
try to provide better discrimination between the effects associated with the middle scores and
remove the medical treatment considerations. Two new scoring options were developed. One
was a nine-point scheme and the other a five-point scheme.
Testing of the two new scoring schemes was conducted by EPA toxicologists in the Health and
Ecological Criteria Division of the Office of Water. Each toxicologist was presented with all the
critical effects given in IRIS with no knowledge of the chemical or chemicals to which they were
attached and the revised scoring protocols. They were asked to independently score the large
group of critical effect descriptions. The toxicologists met as a group to compare scores and
reach consensus on the score and category that is best suited for each critical effect. The five-
point scale was compared to the nine-point scale. After completion of this exercise, the nine-
point scale displayed in Exhibit 9 was selected based on its ease of use, more transparent
clustering of effects within scoring categories, and consistency across the individual scores
assigned by toxicologists.
Page 15 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 9. Final Nine-Point Scoring Protocol for Severity
Score
1
2
3
4
5
6
1
8
9
Critical Effect
No adverse effect
Cosmetic effects
Reversible effects; differences in
organ weights, body weights or
changes in biochemical parameters
with minimal clinical significance
Cellular/physiological changes that
could lead to disorders (risk factors
or precursor effects)
Significant functional changes that
are reversible or permanent
changes of minimal toxicological
significance.
Significant, irreversible, non-lethal
conditions or disorders
Developmental or reproductive
effects leading to major
dysfunction
Tumors or disorders likely leading
to death
Death
Interpretation
Considers those effects that alter the
appearance of the body without affecting
structure or functions
Transient, adaptive effects
Considers cellular/physiological changes in
the body that are used as indicators of possible
adverse systemic damage
Considers those disorders in which the
removal of chemical exposure will restore
health back to prior condition
Considers those disorders that persist for over
a long period of time but do not lead to death
Considers those chemicals that cause
developmental effects or that impact the
ability of a population to reproduce
Considers chemical exposures that result in a
fatal disorder and all types of tumors
The consensus judgment of the EPA toxicologists was used to construct a compendium of nearly
250 critical effect descriptions grouped by their severity scores (e.g., "Chronic irritation without
histopathology changes" equals a score of 3). The final Severity protocol and compendium of
critical effects are provided in Appendix A.
The ordering of the nine-point scale, which clusters developmental and reproductive effects at a
score of 7, and assigns tumors or disorders likely leading to death a score of 8 became a point of
discussion. Some reviewers of the protocol felt that a separation of developmental and
reproductive effects by the seriousness of the outcome was better than the clustered approach.
This option was discussed during internal review of model outcomes (Chapter 4) by the Agency
workgroup. The Agency reviewers decided that the benefits of the proposed scale outweighed
potential drawbacks. The ability to clearly identify PCCL chemicals with even a slight
developmental reproductive or tumorigenic effect through their Severity score is benefit of the
Exhibit 9 scoring system.
Page 16 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
The scale's "uneven steps" were also noted as a point of concern. A detailed exploration of
alternative options, which included the collapse or reordering of the categories resulted in a
consensus judgment to retain the current scale. The current Severity scale works well in
providing a meaningful categorization of the array of critical effects. Given the range of critical
effects that result from a given exposure, it is not possible to have a consistent difference in the
Severity of the outcome between each step on the scale.
2.1.2.2 Evaluation of the Severity Scoring Protocol
The Severity scoring protocol was evaluated using the group of chemicals that were included in
the training data set discussed in Chapter 3 of this report. Evaluation criteria included:
• Ease of scoring using the protocol and critical effect compendium
• Correlation of the list or not list decisions made by workgroup members using the written
narrative descriptions of the critical effects with those made with the numeric scores.
• Outcomes from the algorithm list/no-list decisions (discussed in Chapter 4) using the
scored data as compared with workgroup's decisions based on the descriptive data.
During the initial evaluation process several issues were identified. The most challenging issue
related to Severity scores derived from LDso Potency data. According to the scoring protocol,
the Severity score for an LD50 Potency value would be based on the outcome of death in the test
population and result in a Severity score of 9. The same score of 9 would be given to a LOAEL
or RfD from a more chronic study where the critical effect was described as decreased survival
or longevity. When the evaluator's decisions based on descriptive information for both the
Potency and Severity were compared to the decisions based on scores, it was apparent that the
evaluators looked at the two effects differently. A decrease in survival from a standard chronic
study was regarded as a more serious concern than death in a LDso study where death is the
targeted outcome. Several options were considered for solving this problem. The simplest
option was to have no Severity score for an LD50 based Potency value. Another option was to
retrieve the study that was the basis of the LD50 value and use the critical effect and dose for
systemic effects observed rather than death. The last option was to look for a Potency value and
critical effect from a toxicity study other than an LDso study.
Experimentation with the three options for Severity based on LD50 values demonstrated that a
combination of the second and third options provided a feasible alternative to scoring Severity on
the basis of death when the Potency value was an LDso. The option of eliminating the Severity
score for an LD50 value was determined to be a poor choice since it fails to make full use of the
available data. It was decided that only when attempts failed to identify an alternate study and/or
pre-mortality effects in the LDso study, that an LDso based score of 9 would be assigned.
A problem was encountered with critical effect information for LOAELs from the RTECS
database. This database summarized all effects without specifying which one was the critical
effect. In cases where the original data source was available in the supplemental data, it was
consulted to identify which effect was critical. When the supplemental data identified aNOAEL
for the critical study it replaced the RTECS LOAEL. If the original source could not be accessed,
an alternative NOAEL or LOAEL and its critical effect(s) were identified from the supplemental
data and replaced the RTECS LOAEL. Two guidelines were applied when choosing the
Page 17 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
replacement option. In most cases a replacement was made only if the new LOAEL was lower
than the RTECS value. However, in some cases the alternate value, although greater than the
RTECS LOAEL was chosen because it was from a study that was higher in quality, more
accessible and more recent than the RTECS citation. .In any case where the RTECS remained
the only source for the data, the score for Severity was based on the most serious of the cluster of
effects presented.
Some problems with scoring were encountered in cases where critical effects were not included
in the critical effect compendium. The compendium of critical effects descriptors was developed
to allow people who were not toxicologists to score chemicals based on Severity. In cases where
the scorers could not determine a Severity score, the data were submitted to EPA toxicologists.
A minimum of three toxicologists scored the critical effect. The consensus score was determined
and the critical effect descriptor and its score were added to the critical effect compendium.
One Severity scoring factor that may have had an effect on the correlation between the
classification algorithm-based list/no-list decisions (See Chapter 4) and the workgroup decisions
for the Training Data Set was the numeric Severity score of 8 for carcinogens. The only critical
effect to score 8 was carcinogenicity. Workgroup members could easily identify carcinogens by
their Severity score and possibly placed more emphasis on this result than the other numeric
scores. The classification algorithm was less able to do so, particularly for carcinogens with low
Potency values. For example, in some cases, the algorithm made a "no-list" decision when the
Severity Score was 8 and the expert evaluators made a "list" decision primarily because of the
Severity score's linkage to cancer. This was particularly true in a couple of cases where all the
other scored values were identical or close to identical but Severity was a 7 compared to an 8
(cancer). The decisions for the algorithm and the workgroup matched more closely when
Severity was a 7 than when it was an 8 with the workgroup more likely to choose a list decision
for the 8 Severity score than the algorithm.
In most cases, the combination of Potency and Severity scores performed well in the workgroup
exercises used in developing the PCCL to CCL process and the algorithm trials that followed
(Chapter 4). Alternative approaches were adopted for dealing with LD50 based Potency values,
and critical effect terms that were not initially in the critical effects compendium were added.
Finding an alternative to an LD50 Severity score of 9 and consulting supplemental sources for
critical effect information increased the effort required to obtain the Severity data, but appeared
to function well. These changes are reflected in the Severity Scoring Protocol and Compendium
of Critical Effects in Appendix A.
2.2 Occurrence Attributes
The attributes selected to define actual or potential occurrence of contaminants in drinking water
are Prevalence and Magnitude. Magnitude is related to the quantity (e.g., concentration) of a
contaminant that may be in the environment. Prevalence provides a measure of how widespread
the occurrence of the contaminant is in the environment. When direct occurrence data are not
available, Persistence and Mobility data are used as surrogate indicators of potential occurrence
of a contaminant. Persistence-Mobility is defined by chemical properties that measure or
estimate environmental fate characteristics of a contaminant and affect their likelihood to occur
in the water environment.
Page 18 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
Similar to the health effects attributes, the occurrence attributes are interrelated. The data
sources and the learning sets used to define and scale Magnitude, Prevalence, and Persistence-
Mobility, as well as more details about the individual attributes are described in the following
sections. Unlike the health effects attributes, the data elements used to characterize occurrence
are not solely based on a disciplined progressive study of the contaminants. The availability of
data from surveys of contaminants in ambient and drinking water, the detection limits of
analytical methods, limitations in reporting requirements, as well as indirect measures of
potential occurrence needed to be considered and evaluated. Data sources that could provide
occurrence data ranged from direct measures of concentrations in water to annual measures of
environmental release or production.
The most relevant data for characterizing demonstrated occurrence are monitoring studies or
surveys designed to assess national occurrence in drinking water. Finished drinking water
occurrence data sources that have been compiled include the Unregulated Contaminant
Monitoring Regulations (UCMR), the National Drinking Water Contaminant Occurrence
Database (NCOD) (Round 1 and Round 2 unregulated contaminant data), and the National
Inorganic and Radionuclide Survey (MRS).
Finished water occurrence data are often not available for many chemicals, therefore other types
of data that provide the measures of potential occurrence in Public water systems (PWSs) need to
be considered. The workgroup identified national monitoring studies of occurrence in ambient
waters, which may be the eventual source waters for drinking water supplies. Two US
Geological Survey (USGS) data sources provide information on source water occurrence for
CCL: the National Water Quality Assessment Program (NAWQA), and studies related to the
National Reconnaissance of Emerging Contaminants. These sources provide direct measures of
occurrence in potential source water and indicate possible occurrence in PWSs.
Many of the chemicals evaluated through the CCL process will not have direct water
measurements (finished or ambient). Other available sources that provide data about the
potential for drinking water occurrence include:
• the EPA Toxics Release Inventory (TRI), that reports annual volumes of chemicals
released from industrial applications and the number of states in which those releases
occur;
• the National Center for Food and Agricultural Policy's National Pesticide Use Database
that provides estimates of the amount of pesticide applied and the number of states in
which it is applied; and
• EPA's Chemical Update System/Inventory Update Rule (CUS/IUR), a source for annual
production volume data under the Toxic Substances Control Act. Note the CUS/IUR
data are categorical (i.e., chemicals are in categories with a range of production values,
such as 500,000 to 1,000,000 pounds).
2.2.1 Prevalence and Magnitude Data Elements
A learning data set of 207 chemicals was compiled and used to develop and calibrate scales for
scoring the Magnitude and Prevalence attributes. Due to the linkage of the data used, the scaling
and scoring evaluations were performed concurrently. The linkage between Magnitude and
Prevalence measures is shown in Exhibit 10. The Magnitude measure indicates the median
Page 19 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
concentration of detections in water or the total pounds of the chemical released into the
environment. The median was selected over mean because it typically is a more stable estimate
of central tendency in environmental occurrence data. Outliers have strong influence on means,
often to the extent that the mean is greater than all but the maximum value (particularly when
only detections are used in the calculation). The median of detections was selected over the
median of all measurements in water because all measurements would include non-detections.
Non-detections either signify that the chemical is not occurring or the analytical method is
unable to measure the chemical below the detection limit. The inclusion of non-detections
reduces the median value and for the majority of environmental chemicals the median would be
a less than value (i.e., < the reporting or a "non-detect" value). This would provide little
information and limited discrimination among the chemicals. Prevalence uses the same data
source as Magnitude. The linked Prevalence measure provides an indicator of how widely the
contaminant may be present; in general Prevalence shows the proportion of monitoring sites or
states with detections or releases.
Exhibit 10. Relationship of Data Elements Used to Score Magnitude and Prevalence.
Magnitude Data
Median concentration of detections from
finished water systems.
Median concentration of detections from
ambient water sites.
Amount of total releases nationally in TRI;
annual, in pounds.
Prevalence Data
Percent of finished water systems nationally
with detections of a contaminant.
Percent of ambient water sites nationally
with detections of a contaminant.
Number of states reporting releases of the
chemical in the Toxics Release Inventory.
Sections 2.2.2 and 2.2.3 discuss the approach used to develop and calibrate the scales for scoring
Prevalence, and Section 2.2.4 through 2.2.7, discuss the approach for Magnitude including the
use of Persistence and Mobility Scores as a surrogate for Magnitude when Production volume is
used for Prevalence.
2.2.2 Prevalence - Calibrating Scales and Scoring
Prevalence is a measure of a contaminant's occurrence across the United States. It uses
measures such as:
• Contaminant detections from Drinking Water Monitoring Programs
• Contaminant detections from Ambient Water Monitoring
• States where pesticides are applied
• States reporting releases of a given chemical to the environment
• Production of commodity chemicals in pounds per year.
These Prevalence measures have finite ranges such as zero to 100 percent of PWSs or 1 to 50
states depending on the reporting requirements of the available data source. Accordingly,
transformations to log-based distributions are not necessary. The scaling analyses for Prevalence
focused on establishing groupings of the chemicals across the scoring scale.
Page 20 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
The analyses began with equal bin distributions. Both 100 percent of sites with detections and
50 states with releases divide equally into ten bins based on deciles. In the case of Prevalence,
the bins provided a fairly good fit to the distribution. However, they still required some
adjustment because the equal bins had a tendency to segregate contaminants by type.
Contaminants with the highest percent detections scoring a 9 or 10 were ubiquitous inorganics of
geologic origin. For example, in the National Inorganic and Radionuclide Survey for ground
water, ions such as sodium, calcium, and iron were all detected in > 90% of the groundwater
systems sampled. Contaminants with the highest releases were mostly the high-use pesticides
applied in nearly all the agricultural states or high-use commodity chemicals with reported
discharges from manufacturing or distribution sites in a large number of states such as the
Benzene, ethyl benzene, toluene, and xylene impurities in petroleum products.
Creating ten equal bins from the number of states with environmental releases resulted in a scale
where a Prevalence score of 10 meant that releases had to be reported from 45 or more States.
The workgroup revised the scale for release data so that if more than half the states (25) reported
releases the chemical would receive a Prevalence score of 10 and indicate that the contaminants
potential for occurrence was relatively high. The percent of detections in finished and ambient
water (i.e. percent of systems/sites) were also adjusted to ensure that the most widely detected
organic chemicals received more representative scores when compared to the naturally occurring
inorganic compounds (lOCs).
Among occurrence data elements, the linkage between the Prevalence measures and Magnitude
measures works well for the water measurements and environmental release measures. It does
not work well in the cases when only annual Production data are available. The Production data
provide a measure of pounds of a chemical product produced annually in the United States but
these data do not provide a linked measure such as the number of states in which it is produced
or used. This production rate represents the commercial importance of the chemical to some
extent. Since high production tonnage suggests wide use of a commodity chemical, the
workgroup decided that production data would be used as a measure for likely Prevalence across
the country. For example, a chemical produced at a billion pounds per year is more likely to be
used and released more widely than a compound produced at only 10,000 pounds per year.
Experimentation to examine the correlation of Prevalence scores based on measures of detections
in water and the number of states receiving environmental releases, based on production,
supported the workgroup hypothesis. Correlations were only fair to good but justified the use of
production data as a measure of Prevalence when other data on the spatial spread of a
contaminant across the United States are not available.
Following appropriate adjustments to insure that there was adequate representation of organic
and inorganic contaminants across the ten point scale and a reasonable distribution of the scores
based on release data, the Prevalence scoring scales were finalized. The Prevalence scoring
protocol is presented in Appendix A.
2.2.3 Evaluation of the Prevalence Protocol
The relationship between production or even environmental release data and the actual
occurrence in drinking water is complex. Exhibit 11 shows the scores for several contaminants
based on the finalized Prevalence scoring scales. As expected, in some cases the agreement of
Page 21 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
scores across these differing data elements was not good. For example, a chemical like
glyphosate scores very high for environmental release (being perhaps the most widely used
herbicide in the country) but its water occurrence scores are very low, because of the chemical
and physical properties that influence its fate and transport in the environment, restrictions on
use locations and drinking water treatment.
Exhibit 11. Comparison of Prevalence Scores for Learning Set Contaminants
Chemical
Calcium
Atrazine
Glyphosate
Metribuzin
Toluene
Tri chloroethy 1 ene
Tetrachloroethane
1,1,2,2
Potable
water
samples
% PWS
detect.
10
9
2
1
9
9
3
Total TRI
Releases
# states
NA
8
ND
4
10
10
6
Pesticide
Applications
# states
NA
10
10
10
NA
NA
NA
Production
Ibs/year
8
7
NA
NA
9
8
7
The contaminants in Exhibit 11 indicate that, when the correlation between possible Prevalence
scores is weak, the major difference (e.g. glyphosate) is between the finished water score and the
production/release scores. This supported the decision to use a hierarchy of data elements for
Prevalence. Where actual water measurements are available, they are the Prevalence measure of
choice because they are the most direct measures of likely occurrence in drinking water.
The hierarchy selected for use in scoring Prevalence is as follows:
• Percent of PWSs with detections (national scale data)
• Percent of ambient water sites or samples with detections (national scale data)
• Number of states reporting application of the contaminant as a pesticide
• Number of states reporting releases (total) of the chemical
• Production volume in pounds per year
2.2.4 Magnitude - Calibrating Scales and Scoring
To scale the Magnitude attribute, an evaluation to identify possible correlations among data
elements was conducted. First, a comprehensive universe of finished water quality data was
compiled, including the national occurrence database of regulated contaminants (compiled for
the 6-Year Regulatory Review), the historic data from various unregulated contaminant
monitoring programs (noted as NCOD Rounds 1 and 2, above), and the data from NIRS. This
provided a comprehensive array of data covering the expected distribution range of Magnitude
for any new contaminant, ranging from high median concentrations for some naturally occurring
inorganic ions or elements to non-detect values for some trace organic chemicals.
Page 22 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
The NRC (2001) had initially recommended that Magnitude be scored based on its relationship
to Potency. In their pilot study they proposed that the magnitude score be the square root of the
median concentration, (based on its position in a decile distribution) times the potency score. A
median concentration that fell within the lowest decile of the distribution would receive a 1 and
that in the highest decile a 10 for the calculation. The workgroup evaluated the NRC approach to
scoring Magnitude and found that it was not feasible for the following reasons:
• The NRC equation cannot be applied when the Magnitude data are based on
environmental release or chemical/physical properties.
• A decile distribution for the median concentration values results in low scores for almost
all organic chemicals because of the high concentration of geochemical inorganic
contaminants present in water (see Exhibit 12)
• Application of the NRC equation did not provide a good measure of relative Magnitude
(See aldrin and sodium in Exhibit 12). A high concentration, low Potency combination
can receive the same score as a low concentration, high Potency combination.
To examine the efficacy of the NRC approach, the workgroup applied it to six of the chemicals
from CCL 1 for which regulatory determinations had been made and, thus, had the necessary
Potency and occurrence data. The results of that evaluation are summarized in Exhibit 12.
Exhibit 12. Comparison of the NRC Magnitude Score with the Ratio of the Health
Advisory Guideline to the Concentration in Finished Water
Contaminant
Aldrin
Hexachloro-
butadiene
Manganese
Metribuzin
Naphthalene
Sodium
Potency Benchmark
mg/L
0.000002
0.0009
0.3
0.07
0.1
120
Score
10
7
4
5
5
1
Median
Concentration
mg/L
0.0006
0.001
0.01
0.001
0.001
16.4
Score
1
1
1
1
1
10
Magnitude
NRC
score
3.2
2.6
2
2.2
2.2
3.2
Potency
Benchmark:
Concentration Ratio
0.003
0.9
30
70
100
7.3
The Potency Benchmark is the Health Advisory guideline (cancer or non-cancer) for a
lifetime exposure for all chemicals except sodium. The guideline for sodium is derived
from the recommended dietary intake for sodium in adults, 2.4 g/day^-2L/day using a
Relative Source Contribution of 10%
The Potency Scores were derived from the RfD-equivalent or 10"6 cancer risk values..
The concentration scores were obtained by using sodium as the upper level for the range
and dividing the range into deciles as recommended by NRC.
As indicated in Exhibit 12, the NRC score does not display a consistent relationship to the ratio
of the potency-based drinking water guideline to the median finished water concentration.
Aldrin, the contaminant from Exhibit 12 that is present in drinking water at the levels of greatest
concern has the same magnitude score as sodium ion that is only weakly toxic and not present at
a concentration of concern for other than those on very low sodium diets. In addition, as
Page 23 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
mentioned above, the decile distribution of concentrations resulted in a score of 1 for any
contaminant present in water at concentrations lower than 1.6 mg/L (one tenth of the sodium
concentration). Given this distribution, only inorganic contaminants are likely to receive
intermediate scores on the concentration scale. Because of the observed limitations in the NRC
proposed approach the workgroup determined that it was not appropriate for scoring Magnitude.
The second approach that was investigated employed the use of the Health Reference Level
(HRL) to establish the scores for Magnitude. For example, the largest dose that received a
Potency score of 10 was converted to an mg/L equivalent using the HRL methodology.
Anything less than that concentration received a 1 on the Magnitude scale. Each log-based
Potency value was paired with a log-based concentration. A Potency score of 10, when paired
with any Magnitude score, would be suggestive of concern because the concentration was greater
than the Potency. However a Potency score of 8 would only give rise to concern if the
Magnitude score was 3 or greater (see Exhibit 13).
Exhibit 13: Magnitude Concentrations and Scores Derived from Potency Doses
Potency
Score
10
9
8
Potency Range
mg/kg/day
Oto3.16x 10"7
3.17xlO'7 to 3. 16 x 10'6
3.17xlO'6 to 3. 16 x 10'5
Concentration equivalent
mg/L
Oto2.2xlO'6
>2.2xlO'6 to 2.2 x 10'5
>2.2xlO'5 to 2.2 x 10'4
Magnitude
Score
1
2
3
This second approach to relating Potency and Magnitude proved to be unwieldy because the two
scales are inversely related. It was also problematic because it could not be used for Potency
values based on NOAELs, LOAELS, and LDSOs, or Magnitudes that were not expressed in
concentrations terms. It also did not take into account the differences in the HRL determination
process for carcinogens versus non-carcinogens.
The workgroup next explored a variety of potential scales that could be applied to the finished
water concentration data without consideration of Potency. The first Exhibits 14A-C illustrate
the comparisons of three of the approaches evaluated for the organic and inorganic contaminants.
Exhibit 15 shows the differentiation in scores across the three experimental approaches.
The first approach was to develop scales that utilized the array of compiled Magnitude data and
10 bins with approximately equal numbers of contaminants in each bin, referred to as the equal
number bins scale in Exhibit 14A. Equal bins did not provide a good dispersion of scores.
Accordingly, various log-scale options were explored. The Magnitude data do not range across
as many orders of magnitude as the Potency RfD data, so various semi-logarithmic scales were
evaluated to better represent the distribution of values across the scale.
In evaluating and developing the calibration scale, the water occurrence data presented a
particular challenge because the lOCs tended to skew the results. Many lOCs result from
various anthropogenic processes but most are of geologic origin as well, and they have relatively
high measures for both Prevalence and Magnitude compared to most organic chemicals. Hence,
for some of the semi-logarithmic Magnitude scales (e.g., Half-Log Option A), the only chemicals
Page 24 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
that could score high (e.g., a 10 or 9) would be lOCs. Such a scale would depress the score for
organic chemicals that are of equally high concern because of their expectedly lower
concentrations. One approach that EPA evaluated was using different scales for lOCs and
organic chemicals; however, having two scales would make the scoring process complex. To
keep the process simple it was decided to use one scale for all water data. Accordingly, the
scores were distributed across the range of values so that organic contaminants could receive
high scores as well as the lOCs. Comparisons and adjustments were made until the current
protocols, using a semi-logarithmic scale (labeled as Half-Log Option B in Exhibit 14C), were
selected.
Exhibit 14A. Equal Bins Drinking Water Magnitude Scale
o
4567
Category and Break Points
0 Organics Count
Inorganics Count
Page 25 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 14B. Half Log Option A Drinking Water Magnitude Scale
Category and Break Points
HOrganics Count
1 Inorganics Count
Exhibit 14C. Half Log Option B Drinking Water Magnitude Scale
Category and Break Points
EiOrganics Count •Inorganics Count
Page 26 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 15. Magnitude Attribute Scores: Example Contaminants Scored by
their Median of Detections Using the Various Approaches in Exhibit 14.
Chemical
Hexachlorobutadiene
1 , 1 ,2,2-Tetrachloroethane
Boron
Sulfate
Antimony
Ethylbenzene
Endothall
Methyl ethyl ketone
"Bins"
Score
2
3
10
10
9
6
10
5
Half-Log
Option A
Score
2
3
6
10
4
4
6
3
Half-Log
Option B
Score
5
6
10
10
7
6
9
6
When developing the calibration scales for the release data, the ranges of data were similarly
arrayed using a scale based on half-log units with a distribution of scores that reflected the
distribution of the data in the learning set.
2.2.5 Persistence-Mobility as a Surrogate Measure for Magnitude
In cases where production data are the only measure of occurrence, scoring for Prevalence and
Magnitude becomes difficult. The NRC discussed Persistence and Mobility as a fifth attribute
and had suggested they could be used to predict possible occurrence if other direct measures
were not available. In its review, NDWAC suggested that Persistence and Mobility could
provide a surrogate measure of Prevalence with production used as a measure of Magnitude. To
examine the NDWAC proposal, the EPA workgroup carried out a series of exercises in which
scores for Magnitude derived from concentrations in drinking water and environmental releases
were examined to see if they correlated with production scores and with Persistence-Mobility
scores calculated using the scoring equation developed by NDWAC. In no case was the
correlation as good as one might desire, but it was apparent that the Persistence-Mobility
approach showed a better correlation with the Magnitude scores, based on the preferred data
elements (concentration/release), than the production information. Accordingly, the workgroup
chose to use Persistence-Mobility as a surrogate measure for Magnitude.
Persistence and Mobility are environmental fate parameters. They are considered in combination
as a measure of potential occurrence because both transport (i.e. Mobility) and fate (i.e.
Persistence) are important when predicting whether a contaminant is likely to be found in water
at a specific location, in situations where there is an environmental source for the contaminant.
The length of time a chemical remains in the environment before it is degraded (Persistence)
affects its importance as a potential drinking water contaminant. Persistence is generally
expressed as rate of degradation or half-life (ti/2) indicating, in this case, the length of time
required for the chemical to degrade to half its original concentration in the medium of interest
Page 27 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
(e.g. water). Similarly, the Mobility of a chemical, or its ability to be transported to and in water,
affects its potential to reach and dissolve in the source waters for a PWS.
There are a number of data elements that measure the fate of a chemical in the environment. The
physical/chemical parameters that are most relevant to the fate in drinking water are summarized
in Exhibit 16. The first 4 measures of mobility represent the equilibrium ratio for the
partitioning of the contaminant from one medium to another: Koc (sediment: water), Kow
(octanol: water), Kd (soil: water) and Henry's Law Coefficient (air: water). Koc, Kow and Kd are
sometimes expressed as logs of the original measurements. The measures of persistence each
reflect the time the chemical will remain unchanged in the environment.
Exhibit 16. Mobility and Persistence Data Elements
MOBILITY
Organic Carbon Partition Coefficient (Koc)
Octanol/Water Partition Coefficient (Kow)
Soil/Water Distribution Coefficient (Kd)
Henry's Law Coefficient (KH)
Solubility
PERSISTENCE
Half-Life
Measured Degradation Rate
Modeled Degradation Rate
The data elements listed in the table above are arranged in hierarchical order, with the most
desirable at the top (i.e., the first data to be used if available).
Organic Carbon Partitioning Coefficient (Koc) is one of the most common indicators of the
mobility of a chemical in water. A high Koc increases the probability that, once a chemical
reaches a receiving water body, it will remain bound to sediments or adjacent soils, and thus,
slowly partition from the sediment to the water column. A high Koc favors the presence of the
contaminant in water for a long time but at low concentrations since the Koc will favor the
sediment over the water. A high solubility favors rapid dissolution in the water body from a
near-by source and potentially high concentrations if the water source is confined and the
environmental release substantial.
2.2.6 Persistence-Mobility Data - Calibrating Scales and Scoring
Many of the measurements of environmental fate properties vary depending on the actual field or
laboratory conditions. Some are reported in standard data sources only as ranges, or categorical
descriptions. Scoring was further complicated by the fact that two separate environmental fate
parameters were used in the scoring of the one attribute. Accordingly, the EPA workgroup
selected the approach proposed by NRC and supported by the NDWAC for using the
Persistence-Mobility information after experimenting with several other approaches.
The Persistence and Mobility data were arrayed, or partitioned into relatively simple low-
medium-high categories as suggested by NRC. Published definitions for the categories were
used, such as the categories for Koc from Fetter, 1994 and the classifications for the octanol water
partition coefficient (Kow) from Lyman, et al, 1990. The categories are given values of 1, 2, or 3
based on the ranking of the measurement from low to high. The persistence value is averaged
Page 28 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
with the mobility value and a multiplier (10/3) is used to translate the score to a 10 point scale
(see the Persistence-Mobility Protocol in Appendix A, for details).
Since the persistence and mobility data are being used as a measure of Magnitude, a low ranking
(1) for a parameter is one that will minimize the concentration in water and a high ranking (3) is
one that will maximize the concentration. For example, a high Koc means that the distribution
between the water column and sediment favors the sediment and is ranked a low, while a lower
Koc means that the ratio of a contaminant in sediment to that in the water allows a larger portion
of the total to be in the water and is ranked as high.
As mentioned above, the workgroup undertook a series of evaluations to compare the
Persistence-Mobility scores for selected contaminants to the Magnitude scores derived from the
preferred data elements (concentrations in water or environmental releases). Often, data were
not available for a half-life or a measured degradation rate for the Persistence value. In these
cases, EPA's PBT Profiler was tested and added to the Persistence protocol to ensure both
Mobility and Persistence data were used to calculate the attribute score.
The PBT Profiler was developed as a screening tool to identify pollution prevention
opportunities for chemicals without experimental data. Among other endpoints, it estimates
environmental Persistence for organic chemicals.7 In addition to estimating a degradation rate,
the PBT Profiler also estimates the percentage of a chemical that partitions to soil, sediment,
water, and air compartments. As a last option, in cases where other chemical property data are
not available, the amount of a chemical that is predicted to partition to the water phase by the
PBT Profiler (the percent in water, a measure of solubility) is used to score Mobility.
The workgroup recognized that the Persistence-Mobility protocol can result in relatively high
scores (7 to 10) in cases where more direct data elements for scoring are not available. However,
given the uncertainty associated with some of the Persistence-Mobility data elements, the
workgroup decided the somewhat conservative scores were acceptable as surrogate measures for
Magnitude, when only these data were available for scoring.
2.2.7 Evaluation of the Magnitude Protocol
The occurrence data clearly vary in how directly they measure demonstrated or potential
occurrence related to drinking water. Exhibit 17 compares the scores for several chemicals using
the different measures of Magnitude. In all cases the finished water Magnitude score is higher
than the score for ambient water. Scores for pesticide application rates are higher than those for
TRI releases. As was the case for Prevalence, the workgroup determined that a hierarchy would
be used in scoring Magnitude. The hierarchy developed uses finished water occurrence data if
available.
7 http://www.pbtprofiler.net/ The PBT program will not accept inorganics as input, and identifies the elements,
which if present, that prevent the profiling of a particular chemical. The only exceptions to this rule are sodium,
potassium, and ammonium salts of organic acids, which can be profiled. Thus, the PBT profiler cannot be used for
inorganics or organometallics. However, as drinking water ions, inorganic contaminants are generally present as
salts and do not degrade, and thus are assigned a score of "3" - high persistence. See the Appendix A for more
complete review.
Page 29 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 17. Comparison of Scores derived using the Magnitude Protocol
Chemical
Calcium
Atrazine
Glyphosate
Metribuzin
Toluene
Tri chl oroethy 1 ene
1,1,2,2
Tetrachl oroethane
CASRN
7440702
1912249
1071836
21087649
108883
79016
79345
Finished
Water
Concentration
Median (jUg/L)
10
6
2
7
6
7
6
Ambient
Water
Concentration
Median (jUg/L)
—
4
—
o
4
4
5
Pesticide
Release
Data
Lbs/year
—
10
10
8
—
—
..
Total
TRI
Lbs/year
—
8
—
2
7
10
4
Persistence/
Mobility
10
8
7
7
5
10
7
The hierarchy suggested for Magnitude draws on the following data sets:
• Median concentration of detections from finished water systems
• Median concentration of detections from ambient water sites or samples
• Amount of pesticide applied
• Amount of total releases
• Persistence-Mobility data
2.3 Fine Tuning the Protocols
As discussed in the previous sections, the workgroup developed and fine-tuned the Attribute
Scoring Protocols through a step-wise process of data selection, data analysis, calibration of
scales, and evaluation of the functionality of the scores in PCCL to CCL decision-making. The
decision-making component of the process examined the ability of the scored attributes to
adequately represent the level of concern about contaminants. The testing also evaluated
whether or not the scores provide a consistent input to the decision making portion of the CCL
listing process that is relatively independent of the type of input data that provides the basis for
the score.
Quality assurance measures utilized comparisons of list/ no-list determinations by workgroup
experts based on descriptive and quantitative measures of health effects and occurrence (raw
data) compared with determinations based on the scored attributes. Differences in decisions
were identified. The workgroup discussed those differences and the rationale they had used to
reach decisions based on the raw data versus the scored data. Minor adjustments were made to
the scoring protocols based on those discussions.
Using a training data set of contaminants (Chapter 3), blinded test-case decisions made with raw
data versus scored results, or decisions based on one data element in a hierarchy versus another,
were compared. The results provided a high level of confidence that the scores, while not
capturing all information experts used in making decisions based on raw data, adequately
captured the critical relationships that informed the EPA workgroup "list" versus "don't list"
determinations.
Page 30 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
3.0 DEFINITIONS AND OVERVIEW OF THE TRAINING DATA SET
This chapter describes the process used to identify a set of chemicals to train (or calibrate) the
classification models discussed in the next chapter. The raw data, attribute scores, and protocols
discussed in chapter 2 were applied to these contaminants and that information is carried forward
in the evaluation of classification models discussed in Chapter 4.
The training data set (TDS) for chemicals is the set of data used to train (or teach) the
classification models to mimic expert list-not list decisions. The TDS used to train the models
for CCL 3 was comprised of 202 discrete sets of attribute scores for contaminants and consensus
list-not list decisions made by a team of EPA subject matter experts.
Classification models are algorithms that use statistical approaches for pattern recognition and
derive mathematical relationships among input variables (measurements or descriptive data) and
output from a TDS. For the CCL, the classification models are used to develop a relationship
between the contaminant attribute scores (input variables) and the classification of these
contaminants into list-not list categories (output). The mathematical relationship between
attribute scores and list-not list decisions is determined based on the classification decisions on
TDS chemicals and their associated data. Once the TDS is used to train the classification model,
the model is then applied to a larger list of contaminants to predict their likely list-not list
classifications.
The process for developing the TDS utilized EPA subject matter experts familiar with the
technical aspects of the attribute data and the selection of drinking water contaminants for listing
and regulation.
3.1 Key Considerations
EPA considered the following key factors in developing the training data set:
• Selection of contaminants representing a range of outcomes and decisions likely to be
encountered in developing a CCL;
• A variety of input data ensuring adequate coverage of attribute scores and
combinations of scores;
• Chemicals that, when present in drinking water, would present a meaningful
opportunity for public health improvement if regulated; and
• Contaminants that would likely be selected for the PCCL.
3.2 Developing Key Components of the Training Data Set
3.2.1 Attribute Scores
Attribute scores are a critical component of the TDS, as mentioned in Chapter 2. The TDS used
for training the classification models consisted of attribute scores for 202 contaminants. A set of
known chemicals was chosen to develop the TDS and supplemented with a range of attribute
scores that represented hypothetical or artificial contaminants. These artificial contaminants
were developed to fill voids in the space of possible attribute scores and improve classification
model results.
Page 31 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
3.2.1.1 Attribute scores for real contaminants
Initially, EPA selected "data rich" contaminants from among regulated contaminants and
previous CCLs because they had a range of readily available occurrence and health effects
information. EPA drinking water subject matter experts and stakeholders (as part of the
NDWAC process) reviewed the initial list of contaminants and identified candidates for the TDS.
Based upon an NRC and NDWAC recommendation, EPA also added chemicals "generally
regarded as safe" by the U.S. Food and Drug Administration to provide adequate coverage of
possible attribute inputs and a range of list-not list decisions. This initial selection process
identified 51 chemical contaminants for the TDS.
Subsequently, EPA chose 50 additional contaminants from the CCL 3 Universe. These 50
contaminants were randomly selected from those with high health effects toxicity levels that had
occurrence data because they represented contaminants likely to make it to the PCCL. The
addition of these 50 contaminants resulted in 101 contaminants with data to score attributes.
To aid in the review and evaluation, data summary sheets were prepared for each contaminant
that included a range of available health effects, occurrence, and environmental fate data. All the
available health effects and occurrence, use, and fate data that could be used to develop the
attribute scores for Potency, Severity, Magnitude and Prevalence were included on the individual
summary sheets. Samples of the data summary sheets are presented in Appendix B.
While contaminant names were included in the initial evaluations, expert reviewers found that
knowledge of the contaminant name introduced bias into the decision-making process.
Subsequently, EPA "blinded" contaminant names or identifiers in contaminant evaluations to
increase objectivity and force decisions to be made solely on the available data and associated
attribute scores. The names of contaminants were revealed after the "blinded" evaluations. The
attribute scores were developed according to the Attribute Scoring Protocols discussed in
Chapter 2 and presented in Appendix A.
3.2.1.2 Attribute scores for hypothetical contaminants
The performance of the classification models using the initial TDS gave an indication of gaps in
the possible attribute space that the set of 101 TDS contaminants did not adequately cover. This
led EPA to add a set of 101 hypothetical contaminants to the TDS. These contaminants had
specific combinations of attribute scores designed to fill gaps in the space defined by all possible
attribute scores and to improve the performance of the models. The majority of these possible
scores were selected using Latin hypercube sampling from the set of all possible attribute score
combinations, as seen in Exhibit 18 (NIST, 2006). Five contaminants were selected at random
from each of the 16 "cubes" represented by the combinations of high (6-10) and low (1-5) scores
for the four attributes. This selection resulted in 80 hypothetical contaminants. Twenty one
additional contaminants were deliberately selected to fill in some obvious voids in the 4-attribute
space, resulting in 101 artificial contaminants.
Page 32 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 18. Combinations of low and high attribute scores1 for the four
attributes using Latin Hypercube Sampling.
Potency
Low
Low
Low
Low
Low
Low
Low
Low
High
High
High
High
High
High
High
High
Severity
Low
Low
Low
Low
High
High
High
High
Low
Low
Low
Low
High
High
High
High
Prevalence
Low
Low
High
High
Low
Low
High
High
Low
Low
High
High
Low
Low
High
High
Magnitude
Low
High
Low
High
Low
High
Low
High
Low
High
Low
High
Low
High
Low
High
1 Low scores are randomly sampled from the range 1-5.
1 High scores are randomly sampled from the range 6-10.
Exhibit 19 displays the attribute space coverage of the 101 contaminants compared to the
attribute space coverage of the TDS of 202 contaminants. The combination of real and artificial
contaminants resulted in 202 scored candidates that became the TDS. The total attribute space
for a model that includes four attributes with scores from 1 to 10 is 10,000 combinations of
possible attribute scores. Each point plotted in Exhibit 19 represents one chemical in the TDS
and one of the 10,000 possible combinations of attribute scores.
Page 33 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 19. Attribute Space for the 101 TDS compared to that for the 202 TDS
LU
O
z
LU
LU
cr
o_
severity -->
POTENCY -—>
'-1 1
O"1', —
i
CO ^
i
r- ^
_E
•
1
o i
?!
LU -
--=
< AE
LLJ IE
ft; IE
Q. 1=
1
•
1
*
•
: ••
•
•
•
•
•
•
B
•
•
'• .
•
..
• •
•
•
-
"
H
•
V
.
.
."
"
™ • •
•
f
.
.
.
•ft
m
m
m
M
m
m
m
m
m
m
m
m m
m
mm
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
1
•
•
"
• • • " • M ••••! 1 • •
severity --> ' ' i ...5 ... 10
POTENCY ^4 5 6 ' 8 9 10
Page 34 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
This graphical analysis shows five elements of the model results, the four attributes evaluated
and the categorical decision (L, L?, NL?, and NL) in a single graph. Note in Exhibit 19 that the
vertical and horizontal axes show two attributes on each axis. The attribute scores for Potency
are the large squares across the horizontal axis. The corresponding score for Severity is a
separate scale within each larger square. That is, each Potency square has a range of Severity
scores. Similarly the Prevalence and Magnitude scores are plotted on the vertical axis,
Prevalence as the large squares along the vertical axis and Magnitude as a separate square within
each larger square. The decision category assigned each potential attribute is color coded (NL
decisions are denoted by dark blue, NL? by lighter blue, L? by peach, and L decisions by red).
3.2.2 Making List-Not list Decisions
List-not list decisions are the second key component of the TDS, as mentioned in Chapter 3. The
EPA subject matter experts made list-not list decisions on an individual basis and as a group,
based on attribute scores and based on data that had not been converted to attribute scores (actual
or raw data). The development of the list-not list decisions was an iterative process that
incorporated revisions to the attribute scoring protocols, and the final list-not list decisions, as
experience was gained by the EPA experts. Differences between the decisions based on the
scored attributes and the raw data were resolved by revising the scoring protocols to improve the
correlation of scores to the raw data.
After evaluating the health effects and occurrence data for each contaminant, each individual
reviewer made decisions about how to classify the contaminant, and then met as a group to
discuss their decisions. Early in the process the reviewers recognized that clear List or No-List
decisions could easily be made for some contaminants, but not for other contaminants. The
chemicals in the later group were placed into categories of List? (L?) or Not list? (NL?), in which
L? signifies that the decision is leaning towards listing but with some uncertainty, and NL?
signifies that the decision is leaning towards not listing, but with some uncertainty. These
additional two categories were incorporated into the evaluation process.
As part of the iterative process, the reviewers discussed their classification results and made
adjustments to the process, accordingly. When adjustments changed attribute scoring protocols,
TDS contaminants were rescored and reevaluated. Individual decisions were made separately
based upon either the raw data or attribute scores. Decisions based upon raw data utilized health
effects and occurrence data elements, as well as supporting information on fate and uses. For
decisions based on attribute scores, only the numeric individual scores were used. The scores
were developed from the raw data using the protocols, for Potency, Severity, Prevalence, and
Magnitude. In both cases, this evaluation was conducted "blinded," meaning contaminant names
were not shown. Appendix C shows an example of summary decisions based upon raw data and
attribute scores. For each contaminant, comparisons were made between the list - not list
decisions based upon raw data and those based on scores. Reviewers discussed the similarities
and differences on an individual contaminant basis, and revised the attribute protocols to reflect
decisions made on the actual data (see Chapter 2).
Once L/NL classification decisions were made based on the attribute scores using the revised
protocols, consensus among the EPA subject matter experts was used as the final decision for
each contaminant. This consensus decision was used to train the models and is further discussed
in Chapter 4. Consensus decisions were made by averaging the numerical decisions of
Page 35 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
individual reviewers (L = 4, L? = 3, NL? = 2, and NL=1) and rounding to the nearest integer.
The rounded averages became the consensus values used to train and evaluate the models
(Chapter 4). Appendix C also provides the consensus decisions for each IDS contaminant.
4.0 PROTOTYPE CLASSIFICATION MODELS AND THE CCL PROCESS
The NRC recommended EPA use prototype classification models for CCL selection, citing the
limitations of expert processes and other rule-based models. NDWAC agreed that EPA should
use a prototype model, also noting that this should improve the reproducibility and transparency
in the process. This kind of approach does not eliminate subjectivity but rather, makes the
judgments more explicit.
Prototype classification models are often described as pattern recognition models. These models
develop statistical relationships (to recognize the patterns) among input variables (attributes,
discussed in Chapter 2) of drinking water contaminants to predict their classification ("List,"
"List?," "Not List?," and "Not List"). The model determines the relationship or rule that links
the input to the output based on the decisions made on the TDS (Chapter 3) and then uses that
relationship to classify PCCL contaminants based on their attribute scores.
In its study, the NRC experimented with a linear discriminant model and with an artificial neural
network (ANN) model to demonstrate the use of classification approaches. EPA, working with
NDWAC, identified the following classes of models for evaluation:
• Artificial Neural Networks,
• Classification Decision Trees (with univariate and multivariate splitting rules)
• Linear Models, and
• Multivariate Adaptive Regression Splines (MARS)
The model evaluation was a two-step process. First was the evaluation and selection of the most
appropriate ("best-fit") model from within each of the model classes. The second step was the
evaluation of the performance of the best models selected from each class. Following these
evaluations, two classes of models were rejected and three were maintained to inform the final
expert review process.
Artificial Neural Networks (ANNs) - ANNs are information processing models conceptually
based on the human nervous system and its learning processes. ANNs apply flexible and often
very complex parameterization. Their value is that they use flexible, non-linear functions that
can capture almost any kind of underlying relationship between input and output data. For
classification purposes, ANNs apply weighting in non-linear functions and do not specify a strict
functional form (such as quadratic or cubic equations) as do many statistical models.
Classification Decision Trees - The decision tree classifies the sample by devising a series of
tests (or rules, from the TDS) that are mutually exclusive in outcome. The graphical tree is
derived with a test at a node in the tree with outcomes from the test branching from each node.
Hence, in moving through the tree a contaminant encounters the test at a node, and is sent down
one branch or another based on how its attribute meets the test criterion, usually a simple
inequality, such as is Magnitude < 3.5 (true or false). Eventually the contaminant reaches a
Page 36 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
terminal node (the last node, that no longer branches) that assigns the classification (e.g.,
category 2 = NL?). Two types of decision tree models were explored, Classification and
Regression Tree (CART) which utilized univariate (one attribute at a time) tests at nodes, and the
Quick, Unbiased, Efficient Statistical Tree (QUEST) model, which utilized multivariate
(weighted sum of all attributes) tests at all nodes of the tree.
Linear Models - General Linear Models - Two types of linear models were tried. A Logistic
regression model was applied to deal with CCL's categorical data. The Logistic model was only
attempted using two categories (List and Not List). EPA found that the binary approach was not
satisfactory, and moved to a four category approach. Recognizing that the ANN models often
employ logistic regression, to avoid duplication, the Logistic model was dropped from the final
evaluations. Consequently, the data were adapted for use with a regular Linear regression
model. This model estimates the workgroup's average classification (on a scale of 1 to 4; 1 =
Not List, 2 = Not List?, etc.) for each contaminant as a linear combination of the contaminant's
four attribute scores.
Multivariate Adaptive Regression Splines (MARS) - MARS is a non-parametric classification
model sometimes referred to as a statistical neural network model. MARS has become widely
used in data mining and exploratory analysis because it doesn't assume or impose any particular
class of relationship (such as linear or logistic) on all the predictor variables and the outcomes. It
can develop different regression relationships for different input variables.
4.1 Model Training and Development
Some software packages are designed to build, fit, and test models internally, while others
require an expert user to develop the model. Generally, models are evaluated based on:
• the number of attributes that the model is able to consider,
• the types of relationships or mathematical functions that the model utilizes, and
• the model's ability to predict classification of the TDS.
For example, training a model can involve estimating the values of rule coefficients (such as Po
and Pi in the simple linear regression model Y = PO + PiX + s), or determining some other aspect
of model structure (such as the number of splits in a regression tree model) to improve how well
the model classifies the existing data. Ideally, this training process minimizes the model's
predictive error, thereby reducing incorrect model predictions.
"Over-fitting" is a concern when selecting a model. Any of the model classes can be made to fit
a particular data set very well by making the model more complex (this usually means estimating
more model parameters). However, the addition of model complexity can come at the cost of a
loss of general applicability; the added complexity may capture the idiosyncrasies of the specific
data set, but may not be representative of the broader processes that generate the data, and hence,
may not perform well when applied to an unknown sample. Several methods were used as
guidance to avoid over-fitting, depending on the specific type of model being tested.
Software designed specifically for CART, ANN, and MARS were used for those methods.
Appendix D lists the specific software sources that were used. These programs provide the user
with a number of options to control the model building process. For example, QUEST software,
Page 37 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
used to produce a classification decision tree model with linear discriminant nodes, allows the
user to specify the following:
• Minimal node size of the tree
• Splitting method (linear or univariate discriminants)
• Splitting criterion (likelihood ration, Pearson chi-square, etc.)
• Pruning method (by coefficient of variation or by test sample)
• Number of fold for cross-validation
After the user selects the control options, the software does its best to fit the training data set. In
general, the user is not able to view precisely how the software does its job, but is shown the
final model, some statistics regarding its performance, and an indication of other alternatives that
were considered. For example, the QUEST software outputs a list of decision trees and their
summary statistics (numbers of nodes, error rates). QUEST also identifies the optimal tree and
provides the tree's decision rule. In addition, QUEST reports the results of cross-validation tests,
in which subsets of the training data are held back. The algorithm produces a rule to best fit the
remaining data and this rule is then applied to the data that were held back. This gives a slightly
greater error rate because (a) fewer data are used to estimate the model parameters and (b) data
used for checking are independent of those used to estimate the parameters. Exhibits 20a and
20b compare QUEST Classifications based on the full training data set (Exhibit 20a) and 5-fold
cross-validation (Exhibit 20b).
Exhibit 20a. QUEST Classifications Based on the Full Training Data Set
(shaded cells are exact match with Expert Decisions)
Consensus
Blinded
Decisions
4(L)
3(L?)
2 (NL?)
1(NL)
Model Decisions
4(L)
42
13
0
0
3(L?)
0
41
8
0
2 (NL?)
0
2
54
2
1(NL)
0
0
O
37
Exhibit 20b. QUEST Classifications Based on 5-Fold Cross-Validation
(shaded cells are exact match with Expert Decisions)
Consensus
Blinded
Decisions
4(L)
3(L?)
2 (NL?)
1(NL)
Model Decisions
4(L)
41
14
0
0
3(L?)
1
37
10
0
2 (NL?)
0
5
50
8
1(NL)
0
0
5
31
Page 38 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
Unlike other models, the simple linear model did not depend on special software. Under this
model, the average classification of the experts for a contaminant was estimated as a linear
combination of attribute scores. Letting Y[i] be the expert's average classification for training
set contaminant i, the model equation is:
Y[i] = b0 + bpot * Pot[i] + bsev * Sev[i] + bPrev * Prev[i] + bMag * Mag[i] + s
An intercept term (bo) and coefficients for the four attributes (bp0t, bsev, bprev, and bMag) were
selected to maximize the likelihood of the TDS average classifications, given normal error
structure (s is an error term that is normally distributed with mean zero). A residuals plot
revealed that unanimous List and unanimous Not List contaminants were often predicted to have
extreme errors, suggesting that perhaps the subject matter experts would have assigned some of
these to more extreme categories, had they been available. Without censoring, the unanimous
Lists were treated as observations of exactly 4.0 and the unanimous Not Lists were treated as
observations of exactly 1.0. Recognizing that these may be censored values, they are treated as >
4.0 and < 1.0, and the likelihood function is adjusted to include these as probability masses
(probability of at least 4.0 and probability of at most 1.0) rather than probability densities
(probability of exactly 4.0 and exactly 1.0). Maximum likelihood parameters appear to fit the
data very well, and predict most TDS average decisions to within 0.25 units.
4.2 Model Sensitivity Analyses
Some analyses that were performed in the development process may be considered sensitivity
analyses. These included the following:
• Training the models on subsets of the TDS. This included the partial TDS (as it was
being developed) and cross-validation exercises, wherein randomly-selected
contaminants were held back from training to provide independent error checks.
• Training after selected "outliers" are removed from the TDS. Those found to have strong
influence on the overall performance were investigated further to see if there were valid
reasons for excluding them from the TDS.
• Graphical and statistical analyses to identify significant differences in attribute "weights"
or influence on model performance. If any attribute had been found to be insignificant, it
could have been ignored, perhaps saving some data development resources. (Though
attributes were found to have different weights, none was found to be insignificant.)
Rather than detail all of the sensitivity analyses conducted for all classes of models, the
remainder of this chapter illustrates the analyses described above using selected applications.
4.2.1 Training with subsets of the TDS
Cross validation for QUEST is described under 4.1, above. Training with early subsets of the
TDS (50 and 102 contaminants) produced mixed results for the five model classes. QUEST and
linear models exhibited no logical inconsistencies, but ANN, MARS, and CART showed some
serious problems. Most dramatic was MARS, which placed contaminants with the very lowest
health effects and occurrence scores in the List category. Clearly, additional training data was
needed to overcome these difficulties. No class of model was eliminated on the basis of these
findings.
Page 39 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
The final TDS (size 202) allowed all of the classes to improve their performance. ANN was
found to have no logical inconsistencies. Although MARS and CART improved significantly,
both had some areas of non-monotonicity. This means that there were some cases an increasing
attribute score could lead to a decreasing classification for a contaminant. (This inconsistency is
discussed and displayed graphically in Section 4.4.2.)
4.2.2 Training after Selected "Outliers" Are Removed From the TDS
The linear model was most sensitive to selected TDS contaminants. Fortunately, this model
provided a number of tools for identifying outliers. While other models had the objective of
minimizing the count of classification errors (or in the case of QUEST, a weighted sum of
classification errors), the linear model attempted to minimize the deviance between its prediction
and the average classifications for TDS contaminants. When the other models encountered an
outlier, (for example, a contaminant with very high attribute scores, but a classification of NL),
they did not attempt to make the correct classification for the outlier because that would have
meant making other errors for nearby contaminants. Including or not including such an outlier
had no effect on the outcome. The linear model, in essence, attempted to minimize the squared
estimation error, so outliers tended to have some influence on the linear model parameters.
Residuals plots such as Exhibit 21 revealed potentially important outliers for the linear model.
Exhibit 21 shows the model-estimated versus team classification of one important outlier: a
contaminant with scores (4,8,10, and 10) with a team-average classification of 3.17 (L?) and
model-estimated value of 3.88 (L). Another contaminant has as large a residual (model = 1.53
and team = 2.33, both NL?). However, when the model was run first with one and then the other
contaminant removed, only the first outlier was found to have a marked influence on the overall
error rate (number of misclassifications and weighted sum of misclassifications). When EPA's
team was asked about these two contaminants, they agreed that their classification for the first
contaminant was influenced by their belief that it was a ubiquitous inorganic that should
probably not be listed. When asked how the model should treat PCCL contaminants with such
high Severity and occurrence levels, the team agreed that the correct decision would probably be
to List the contaminant, but that the two tens for occurrence suggested that the contaminant was
inorganic biasing them towards the lower decision category. It was decided to drop this
contaminant from training the linear model. Because it had negligible influence on the other
models, it was included for them.
Page 40 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 21. Model-estimated versus Team Average Classification for the TDS
5f"IIIII
8
Q
0
J_
I
I
I
I
1 1.5 2 2.5 3 3.5
Team Mean Decision (doesn't include perfect 1 =NL and 4 = L)
The graphical displays discussed in Section 4.3.2 were used as additional checks for outliers.
The outliers for the linear model were apparent when the training data set was plotted against the
background display. The inorganic contaminant that was eliminated from linear model training
was seen to fall "between" two other contaminants that were both assigned to the List category -
further evidence that its classification of L? may have been inappropriate, at least for the purpose
of training this model.
4.2.3 Graphical and Statistical Analyses to Identify Significant Differences in Attribute
"Weights" Or Influence on Model Performance
Graphical displays of model outputs (Section 4.3.2) revealed that all of the attributes were
important. The ANN graph is the only means of studying the ANN rule, but QUEST and the
linear model provide mathematical expressions that clarify the roles of the four attributes. For
QUEST, each "node" of the tree involves comparing a weighted sum of attribute scores with a
threshold. If the threshold is surpassed, then the "right" path is taken, otherwise, the "left" path
is taken. The QUEST software is capable of using fewer than four attributes, and when trained
with about half of the 202 TDS contaminants, it sometimes used only three of the four. When
the full TDS was used, however, all four attributes were used at each of the final tree's seven
nodes. At each node, the four attributes can be ranked in order of their model coefficient.
Exhibit 22 shows the ranking of attributes for the nodes of the final QUEST tree.
Page 41 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 22. Relative Weights of Attributes at QUEST Nodes
(1 = greatest weight, 4 = least weight)
Node # l
1
2
O
4
5
1
28
Pot
1
1
2
1
1
1
2
Sev
2
2
1
3
2
3
1
Prev
4
4
O
4
4
4
3
Mag
3
3
4
2
3
2
4
N2
202
141
61
52
89
18
23
Numbers as assigned by QUEST.
N = Number of TDS contaminants that are evaluated at the node. All 202 are evaluated at the first node. Of
these, 141 proceed to node 2, while the remaining 61 pass to node 3.
Overall, it appears the Potency carries the most weight, followed by Severity, Prevalence, and
Magnitude.
The linear model assigns a weight to each attribute and the greatest of these is that of Potency,
followed by Severity, Magnitude, and Prevalence. The order of Prevalence and Magnitude is the
reverse of that found for QUEST. The linear model also provides a means of testing the
statistical significance of the intercept and four coefficients. Because the model accounts for
possible censoring, this testing is not as simple as in a least-squares regression. Two methods
were used to approximate the covariance matrix for this model. The first is based on the Fisher
information (J(model parameters 9)), derived using the likelihood function, L(data|9):
J(9) = - E [d2 ln(L(data|9)) / d92 | 9]
The second used a Bayesian posterior sample of parameter values. This sample produced a
covariance matrix that was nearly identical to that derived from the Fisher information,
suggesting that the likelihood and posterior are very nearly multivariate normal. Hypothesis tests
could therefore be conducted using the Markov Chain Monte Carlo (MCMC) sample (10,000
sets of parameter values). Exhibit 23 below shows means, medians, and 95% credible intervals
for the model parameters, bl through b4 are the parameters for the four attributes (Potency,
Severity, Prevalence, and Magnitude, respectively), bO is an intercept term, and Phi is the
precision (inverse of the error variance). The 95% intervals reveal that all of the attribute
parameters are statistically significantly greater than zero.
Page 42 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 23. Summary Statistics from MCMC Sample
Parameter
bO
bl
b2
b3
b4
Phi
Mean
-1.674
0.2410
0.2170
0.1157
0.1699
14.25
2.5%
-1.865
0.3343
0.2002
0.1033
0.1539
11.44
Median
-1.673
0.241
0.2169
0.1157
0.1699
14.22
97.5%
-1.488
0.2591
0.2342
0.1284
0.1858
17.41
Based on the MCMC sample, pair wise comparisons of attribute parameters were all found to be
statistically significant. Separate weights are needed for the two health effects attributes and for
the two occurrence attributes.
4.3 Model Performance Testing
The TDS, Attribute Scoring Protocols, and prototype model test results were linked together in
an iterative process. Testing of the models in the early stages was impacted by changes and
refinements in attribute scales, resulting changes in the scores, and changes in the composition of
the TDS. These changes required iterative reevaluation of the models and resulted in many
improvements that are part of this final analysis. Refinements in scoring are discussed further in
Chapter 2 and development of the TDS in Chapter 3. EPA also evaluated the impact of the
attributes used by the models and the effects of missing data on the performance of the models
during the various stages of development.
During early stages of the model testing, the models were run with various sized TDSs. The
CART and MARS models did not always use all four attributes with some of the smaller TDSs.
However, all models used all four attributes when trained with the final TDS, consisting of 202
contaminants.
Exploratory analysis of the results revealed some additional problems with the CART and
MARS models. When two contaminants have identical attribute scores for all but one attribute,
the contaminant with the higher score for that attribute should logically be classified at least as
high as the contaminant with the lower score. For example, if a contaminant with scores (4, 4, 4,
4) is assigned to the L? Category, then a contaminant with scores (4, 4, 4, 5) should not be
assigned lower, to category NL? or NL. Both CART and MARS rules had this type of
misclassification. Both models did not consistently classify contaminants. Another problem
with the CART and MARS models was their errors across two categories. Both models did not
consistently separate the NL? from the L contaminants or separate the L? from the NL
contaminants. Because of these problems, and because of poor performance with respect to the
training set decisions, these two models were not selected to inform PCCL to CCL EPA
decisions.
Three models, ANN, QUEST and Linear regression consistently demonstrated the best
performance when using the TDS. Exhibit 24 lists the features of these three models.
Page 43 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 24. Features of the Three Preferred Models Based on TDS Test Results
Features
Objective Function
(to be minimized
or maximized)
Prediction
Ranking
Capability
Transparency of
Optimization
Method
Classification Rule
Computation
Speed
Software Cost
Classification Models
Artificial Neural
Network
Minimize count of
training set errors
Rounded average
workgroup
classification
Rank by Probability
(Prof List)
Not transparent
Not clear, but
classifications
available for all
attribute score
combinations.
< 1 Second
Version used is
Freeware.
Classification Tree
with Linear Nodes
(QUEST)
Minimize count of
training set error
loss OR minimize
error loss
Rounded average
workgroup
classification
Rank by
classification and
distance from
discriminant
(requires post-
processing)
Not transparent
Clear. Complex
classification tree
with linear
inequalities for
intermediate nodes
< 1 Second (but
process for deriving
distances for
ranking is not part
of software)
Freeware
Linear Regression
Maximize
likelihood or
minimize error loss
Average workgroup
classification (not
rounded)
Rank by prediction
Simple and
transparent
Clear. Simple linear
function of attribute
scores.
< 1 Second
No special software
4.4 Evaluating Classification Differences
This section describes how the classification models were assessed and compared with respect
to:
• The number of correct and incorrect classifications for the 202 TDS contaminants
• The number of "large" misclassifications (off by more than one category)
• The weighted sum of TDS classification errors
• Ability to identify intermediate classifications
• Consistent behavior (e.g., no decreasing classification as attribute scores increase)
Page 44 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
As described in Section 3.3.1, the approach to classifying the TDS contaminants became a four-
category decision (L, L?, NL?, and NL) to allow the EPA subject experts, experienced in making
L/NL decisions, to identify the decisions that were not strong list or NL decisions. Accordingly,
quantification of model performance as it compared with the decisions of the EPA subject matter
experts had to consider a suite of various misclassification outcomes, (Exhibit 21) such as a
consensus decision that a contaminant should be a L?, but the model classifying it as a L.
However, not all the misclassifications are considered to be equally serious. Of the differences,
the most substantive would be placing a strong "List" contaminant in the "Not List" category.
This might result in missing a key candidate for the CCL. Considering the relative seriousness
of the different kinds of misclassifications, the workgroup represented the classification error
losses in terms of the weights displayed in Exhibit 25. Initially, the table had equal weights for
all misclassifications and these were adjusted until the workgroup was comfortable that they
represented the relative significance of the 12 misclassifications or errors that are possible. The
most serious error (placing a List contaminant in the Not List category) has ten times the weight
(i.e., a 10) of the least substantive difference (placing a contaminant one category too high, such
as placing a List? contaminant in the List category, i.e., a value of 1).
Exhibit 25. Decision Comparison Matrix; Weight of Differences
Model Decisions
Not list
Not list?
List?
List
Subject Matter Expert Decisions
Not list
^
1
2
O
Not list?
2
^
1
2
List?
5
2
^
1
List
10
5
2
^
The Decision Comparison Matrix and the quantitative weighting of differences were used to
compare model results to EPA decisions. This was part of the process to minimize the losses and
cost of the misclassifications.
The models are tools to help classify and prioritize the contaminants for expert review at the end
of the CCL process. After applying the models, EPA plans to scrutinize all of the contaminants
identified as "List," but likely will spend less evaluation time on those placed in the other
categories (particularly the "Not List"). As a result, the EPA workgroup recognized the need to
minimize the likelihood of classifying "List" contaminants as "Not List" or "Not List?" and
applied the Decisions Comparison Matrix as a tool in evaluating model output misclassifications.
4.4.1 Classification Differences Among the Models
Appendix E describes the classification rules or "solutions" that were generated by the different
models. These rules perform differently, when compared with the TDS consensus decisions.
Exhibit 26 summarizes the number of each type of decision by each model compared to the
subject matter expert consensus decisions and Exhibit 27 summarizes the results and the
Page 45 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Weighted Loss Value. The model input (and output) for ANN, CART, MARS, and QUEST
were the integers representing the classes (i.e., 4=L, through 1=NL) while the Linear model
estimated the average classification. When a majority of decision makers favored one
classification for a contaminant, that class was assigned. When the decision makers were evenly
split (for example, if three assigned a contaminant to 1 (NL) and three assigned it to 2 (NL?)), an
agreement was reached to assign the contaminant to the higher of the two categories (2 (NL?) in
the case of even split between 1 and 2). In contrast, the Linear model predicted the average
classification and was trained using average classifications for the TDS. For example, if three
decision makers assigned a contaminant to 1 and three assigned it to 2, the average classification
was 1.5.
Exhibit 26. Summary of Quaternary Model Decisions
Decision
Category
4(L)
3(L?)
2(NL?)
1(NL)
Total
Number of Decisions in Category by Model
Expert
Workgroup
Blinded
Decision
42
56
65
39
202
ANN
42
55
65
40
202
CART
27
68
73
34
202
Linear
27
69
69
37
202
MARS
47
38
81
36
202
QUEST
55
49
58
40
202
Exhibit 27. Results of 202 Model Classifications and
Weighted Misclassifications
ANN
CART
Linear
MARS
QUEST
Number of
Classification
matching
TDS
168
156
160
160
174
Weighted Loss
Value
52
84
72
67
33
While there are important differences, all the models were able to process the TDS and produce
classification rules. All five models produced from 79% to 86% exact matches with the
consensus decisions. Exhibit 28 provides further details on the predicted classifications for each
model. Perhaps most important, no model classed any consensus L(4) or L?(3) decision as a NL
(1). Only CART classed any L(4) candidates (2%) as NL?(2).
The best performance, by these metrics, was that from the QUEST model, while the lowest
performance was by CART. The objective of the QUEST model was to minimize the value loss
of the misclassifications, while the other methods minimized errors with no regard for the
Page 46 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
weights shown in Exhibit 25. As a result, QUEST has the lowest loss, and the highest exact
match rate. Note on Exhibit 28, that QUEST'S misclassifications are all shifted to the "left;"
i.e., QUEST only predicted 2 of consensus L? decisions would be NL?; and 8 of consensus NL?
were predicted to be L?, a more acceptable and conservative difference. ANN attempts to
maximize the likelihood of correct predictions and simply minimize the number of
misclassifications (not their weighted value). Its misclassifications are rather equally distributed
around the exact match categories. The performance of MARS and the Linear model look
similar, but MARS had the highest value of any model for consensus L? decisions that were
predicted as NL? (16).
4.4.2 Logical Evaluation of the Models - Graphical Analysis
As introduced in Section 3.2.1, the testing of the models included evaluation of the total potential
"attribute space." The total "attribute space" for a model that includes four attributes with scores
from 1 to 10, is 10,000 combinations of possible attribute scores. The graphical analysis of
model performance looked at how the models generated decisions on the category to which it
assigned contaminants (L or NL). When applied across the entire attribute space, the
discriminate surfaces that bound the model's decisions on the category to which it assigned any
possible score became apparent. These category boundaries or discriminant surfaces were
reviewed for consistency through the graphical analysis. Five models (ANN, QUEST, MARS,
CART, and Linear Regression) developed with the 202 TDS produced classification rules that
were applied to the 10,000 scores and plotted to evaluate their performance (Exhibits 29 through
32).
Exhibit 29 is another example of the graphic tool introduced in Chapter 3, Exhibit 19, to help
visualize the multi-dimensional space of the CCL classifications. The graphical analysis shows
five elements of the model results, the four attributes evaluated and the categorical decision (L,
L?, NL?, and NL) in a single graph. Note in Exhibit 29 that the vertical and horizontal axes
show two attributes on each axis. The attribute scores for Potency are the large squares across
the horizontal axis. The corresponding score for Severity for each Potency score is a separate
scale within each larger square. That is, each Potency square has a range of Severity scores.
Similarly the Prevalence and Magnitude scores are plotted on the vertical axis with Prevalence
along the primary axis and Magnitude along the axis imbedded in each Prevalence square. The
categorical decision assigned to each potential attribute score combination is color coded. Red
represents a L decision, peach, a L?; light blue represents a NL? and dark blue represents a NL
decision.
Page 47 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 28. Summary of Individual Quaternary Model Classifications
(shaded cells are exact match with Expert Decisions)
Consensus
Blinded
Decisions
4(L)
3(L?)
2 (NL?)
1(NL)
4(L)
3(L?)
2 (NL?)
1(NL)
4(L)
3(L?)
2 (NL?)
1(NL)
4(L)
3(L?)
2 (NL?)
1(NL)
4(L)
3(L?)
2 (NL?)
1(NL)
Model Decisions
ANN
4(L)
37
5
0
0
3(L?)
5
44
6
0
2 (NL?)
0
7
53
5
1(NL)
0
0
6
34
CART
4(L)
26
1
0
0
3(L?)
12
47
9
0
2 (NL?)
4
8
53
8
1(NL)
0
0
O
31
Linear
4(L)
26
1
0
0
3(L?)
16
47
6
0
2 (NL?)
0
8
54
7
1(NL)
0
0
5
32
MARS
4(L)
37
10
0
0
3(L?)
5
30
3
0
2 (NL?)
0
16
59
6
1(NL)
0
0
3
33
QUEST
4(L)
42
13
0
0
3(L?)
0
41
8
0
2 (NL?)
0
2
54
2
1(NL)
0
0
O
37
Page 48 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA 815-R-08-004
February 2008 - DRAFT
Exhibit 29. ANN Model Predictions for the Four Attribute Space
(10,000 possible score combinations)
severity -->
POTENCY
5
8
10
The colors represent the classification decision: List = red; List? = beige; Not List? = light blue, and Not List = dark blue. One TDS
contaminant (Potency = 4, Severity = 8, Prevalence = 5, and Magnitude = 10) is shown in black, though the workgroup's decision for
that contaminant is List (red). This particular contaminant is always shown in contrasting color to help the viewer orient to the details
of the graph and check the scaling and axes.
1 Expressed in RGB format, dark blue is (5 113 176), light blue is (146 197 222), beige is (244 165 130), and red is (202 0 32). These
colors were selected using ColorBrewer, by Cynthia A. Brewer of Perm State University. ColorBrewer can be found online at
www.ColorBrewer.org.
Page 49 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
Exhibit 29 plots the results of the ANN models classifications for the 10,000 combinations of
attribute scores. The patterns clearly show a logical progression from the lower left to upper
right, progressing from Not list predictions (dark blue) for low attribute scores, through NL? and
L?, to List classifications for the highest scores, both within each square and across the entire
matrix. The graphical analysis helped to understand and visualize the logic of the discriminant
approach of models and to visualize the performance with the TDS. The QUEST model
produces a very similar graphic result to the ANN model.
In contrast to Exhibit 29, Exhibit 30 shows the MARS results. The figure shows areas where red
(L) directly touches light blue (NL?) and where dark blue (NL) touches beige (L?). Both are
indications that the model was unable to define the intermediary categories. Another problem
can be seen in the lower right box of the figure, where Potency is 10 and Prevalence is 1. Within
that box, when magnitude is 1 (along the bottom edge of the box), as Severity increases, the
decision can be seen to go directly from NL? to NL (light blue to dark blue). This unacceptable
result also occurs for several other combinations of high Potency and low Prevalence. These
results were not considered logical or acceptable by the EPA workgroup. Exhibit 31 shows that
the univariate CART model exhibited similar problems.
The adapted Linear regression model, shown in Exhibit 32, presents an interesting variant. As
noted, the Linear model predicts average classification of contaminants. In other words in
contrast to ANN or QUEST which predict a classification as an integer of 3 (or L?), the Linear
model predicts the value from the regression model, such as 3.312 (rounded to 3 = L?), so the
colors can be displayed more as a continuous variable. The Linear model again displays a very
logical function across the total attribute space.
As discussed above, the CART and MARS models exhibited inconsistent categorization of
contaminants and poor performance in the decision matrix comparisons, while the other three
models (ANN, Linear, and QUEST) performed very well with respect to TDS error loss, number
of training set errors, and the logic of the classification model. The linear model was generally
able to predict the workgroup average within approximately 0.3 (less than half a category).
Hence, evaluating ways to apply the model results focused on procedures for utilizing the results
from the ANN, Linear, and QUEST models.
Page 50 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 30. MARS Model Predictions for the Four Attribute Space
(10,000 possible score combinations).
IUIII1
Mill!
J.UI11
111
severity -->
POTENCY -—>
See Exhibit 4-10 for the key and text for discussion.
Page 51 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 31. Univariate CART Model Predictions for the Four Attribute Space
(10,000 possible score combinations)
severity -->
POTENCY
See Exhibit 4-7 for the key and text for discussion.
Page 52 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA 815-R-08-004
February 2008 - DRAFT
Exhibit 32. Linear Model Predictions for the Four Attribute Space
(10,000 possible score combinations)
severity -->
POTENCY
- — >
See Exhibit 4-6 for the key and text for discussion.
Page 53 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
4.5 Applying Model Results
From the inception of the development of the CCL classification process, EPA intended to use
classification models as decision support tools. It was envisioned that, after testing and
evaluation, a model(s) might be used to process complex data in a consistent, objective, and
reproducible manner and provide a prioritized listing of candidate contaminants for the last stage
of the CCL process, an expert review and evaluation. This also would help to focus resources
for the review and evaluation of potential contaminants. The use of classification models as a
tool in the CCL process is a new approach, a new application of such tools.
Several factors have been considered in assessing how to utilize the model results. After testing,
EPA determined that three models performed well: the ANN, Linear, and QUEST models.
These are three different classes of models, with three different mathematical approaches, but all
provided similar results and logical determinations. Yet the results of each are unique (e.g.,
Exhibits 29 and 32). Therefore, EPA explored simple ways to combine the results of all three
models, to capture both agreement among models and unique results. Two straight forward
approaches looked most useful and were applied: a simple additive approach, and a collective
rank-order approach.
4.5.1 Additive Model Results
The first step in combining the results of the three models was to simply add the results of their
classifications for each contaminant. A tabulation of all contaminants (in the TDS) was prepared
with their predicted classification from the models. Recall, the model output is as a class
(number), with 4 equaling L through 1 equaling NL. The Linear model output was rounded to its
integer class for this approach). Then the 3 results were simply added. This resulted in 10
"bins" or classes, ranging from 3 (all three models classed the contaminant as a 1) to 12 (all three
models classed the contaminant as a 4). Hence, a contaminant with an additive score of 11, had
two models class it as 4, totaling 8, and one model class it as a 3. A comparison of the sum of
the three models to the TDS workgroup Decisions is shown in Exhibit 33.
Exhibit 33, shows some important features of the additive process. For 142 of the 202
contaminants, the three models were unanimous and in agreement with the TDS. Every
contaminant that the subject matter experts classed as List (by consensus) was predicted as a List
by at least one model. The models do move some NL? into a strong L? positions, but only 2 of
the L? contaminants were placed into the NL? category. The areas where the models differ in
outcome can provide a place to focus some review during the development of future CCLs.
4.5.2 Additive Rank Order Results
To provide a different approach from the 10 additive classes, a simple method to provide a more
continuous rank-order was also developed. The output for each model was used to produce a
rank-ordering for that model; ordering from highest (a L candidate) as number one, to lowest (a
NL) as number 202 for the TDS. Once the ranks for a model were ordered, the contaminants
were simply assigned a number from 1 to 202 (high to low). After this was done for all three
models, the rank numbers were added (resulting in a range from 3 to 606) divided by 3 (just to
stay on the 202 scale), and then reordered by their composite ranks.
Page 54 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit 33. Summary Comparison of the Sum of the 3 Model Decisions to the
Distribution of the Workgroup Blinded (TDS) Decisions
Shaded c
using all
applicatic
Sum 3 Model Results
All 3 = 4 (L)
All 3 = 3 (L?)
All 3 = 2 (NL?)
All 3 = 1 (NL)
Sum
12
11
10
9
8
7
6
5
4
3
Consensus Blinded Decision
4(L)
26
11
5
42
3(L?)
1
4
8
35
1
5
2
56
2 (NL?)
6
2
49
4
2
2
65
1(NL)
2
3
2
32
39
ells are unanimous model decisions that match with the TDS. These analyses were also conducted
models. The analysis reinforced some of the problems discussed for the CART and MARS
ms.
As part of the unique input of the three models, each model produces different output with which
to develop its own prediction and a rank-order. The Linear regression model as applied,
predicted the outcome as a continuous variable from the regression equation (e.g., 3.312), and
these values were simply used to rank-order. ANN produces a probability of a contaminant
being a 4. So, for ANN, the probabilities for each contaminant were used for the rank-ordering.
QUEST does require some processing after the model produces classification predictions to
produce a rank order. For QUEST, the distance from the lower discriminant surface was
computed. The contaminants were then rank-ordered within a classification group (i.e., ranked
within the L? group), then a composite was compiled. QUEST, as a classification decision or
regression tree, produces more ties than the other models, but it still produces enough of a
continuum that it did not present a problem.
The composite provides a nearly continuous rank-ordered list that can further help to prioritize
the analysis for the expert review. Combining the additive results and the rank ordering could
also be useful. Knowing which contaminants get unanimous 4s and Is, or identifying
contaminants that stand out as anomalies in one model was useful in the review of the model
output. Having the rank-ordering within the group that included a L? decision, for example, was
useful for prioritizing additional evaluation.
Page 55 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
5.0 MODEL OUTCOME AND POST MODEL EVALUATION PROCESS
The preceding chapters have described the process that was developed for selecting the CCL
from the PCCL. The companion document, CCL 3 Chemicals: Screening to a PCCL (USEPA,
2008b), describes the approach that was used for screening and selecting the PCCL from the
Universe of chemicals. Once the PCCL screening was executed, the Attribute Scoring Protocols
finalized, and the models trained, all of the PCCL chemicals were scored for their attributes and
run through the models. This chapter describes the results from the modeling and the processes
EPA used in evaluating the model output before selecting the preliminary CCL 3.
The evaluation of model output lead EPA to formulate several post-model refinements that were
added to the CCL selection process, including an approach for considering the certainty reflected
in the differing data elements. The post-model analyses are also described in this Chapter.
5.1 PCCL Characterization and Model Results
The screening process, described in CCL 3 Chemicals: Screening to a PCCL (USEPA, 2008b),
selected the chemicals for the PCCL. The attributes for these chemicals were scored using the
procedures presented in Chapter 2 and evaluated by the three models described in Chapter 4.
Exhibit 34 illustrates the results of the model output for the PCCL contaminants8. The PCCL
consisted of chemicals with variable health effects data, ranging from RfDs to Lethal Dose 50
(LDso), and occurrence data, ranging from measured water concentration data from PWSs to
production volume data.
Exhibit 34. Model Results for the PCCL Chemicals
3-
Models
Decision
L
L-L?
L?
NL7-L?
NL?
NL7-NL
NL
N(all)
%of
PCCL
9%
12%
33%
6%
28%
4%
9%
1 00%
Total
#
PCCL
44
58
163
30
139
20
46
500
Finished
or
Ambient
Water
3
9
26
6
29
7
21
101
Release
24
29
64
11
28
9
7
172
Production
17
20
73
13
82
4
18
227
As described in Chapter 4, three models were used in classifying the PCCL contaminants. The
bolded decision category (i.e. L, L?, NL?, NL) in Exhibit 34 signifies that all of the models were
in 100% agreement with that listing decision. The other categories (e.g., NL7-NL) represent
varied agreement where one or two of the models choose one listing option and one or two
The screening of the CCL 3 Universe, including processing with supplemental data during the nominations
process, resulted in 532 chemical contaminants for the PCCL. These chemicals were scrutinized as part of the
classification and modeling process. Some of the PCCL chemicals had limited data available for scoring and could
not be run through the models process. The 32 contaminants that had limited data remain on the PCCL. They are
identified in Appendix G. Exhibit 34 recaps the model output for the 500 chemicals that were scored and processed.
Page 56 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
models chose a different option. None of the models categorized a contaminant in a category
more than one category higher or lower than the other models. That is, no contaminants were
categorized as an "L" by one model and as an "NL?" by another model, or vice versa. The
models categorized approximately l/2 of the chemicals on the PCCL as L? or above. When
analyzed by data type, the majority of chemicals in the List category had LDso data for health
effects. This was a concern and became an important issue for consideration in the post-model
evaluation process.
5.2 Evaluation of the Modeling Output
As part of the last stage in the CCL classification process, the model output was reviewed by
internal EPA experts. This step involved:
• a more detailed review of the data used,
• a review of supplemental data, and
• deliberations on how the model data should be used to produce a draft proposal for a
CCL.
Specifically, the function of the team was to critically compare the results from the model to the
information in the database dossier for the individual chemicals, and identify any concerns with
the model output. This exercise was conducted for a cross section of the model outcomes and
their associated contaminants.
The Evaluation Team was comprised of the participants (EPA scientists, engineers, and
environmental protection specialists from the OW, Office of Research and Development, Office
of Children's Health, and Office of Pesticide Programs). The Evaluation Team met on a weekly
basis for approximately 8 weeks to discuss the evaluation results.
5.2.1 Procedure
Prior to the initiation of the evaluation effort, all Evaluation Team members received background
descriptions of the CCL process for chemicals (chapters 1-4 of this document), Attribute Scoring
Protocols, and evaluation work sheets. A spread sheet with the attribute scores, the data that
supported the scores, and the model output for each of the chemicals selected for the first review
session was also included in the package. An initiation meeting was held to familiarize the
participants with the contents of their evaluation package and discuss the approach that would be
followed in evaluating the model output for individual contaminants.
Participants on the Evaluation Team received a set of contaminants and their data dossiers for
evaluation. The completed evaluation sheets were submitted so that the results could be
compiled for discussion. The evaluation sheets allowed the participants to:
Page 57 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
• Comment on the model input data for each attribute
• Provide a statement on their level of confidence in the data underlying each attribute
score
• Express agreement or disagreement with the model output
• Indicate their degree of confidence in the model decision
• Provide an explanation for their agreement of lack of agreement with the model decision.
Following submission of the evaluation results for each set of contaminants, the Evaluation
Team discussed the outcome of the evaluation, concentrating first on those contaminants with the
greatest differences among the reviewers. These discussions identified the issues and steps
described in the following sections of this chapter. The Evaluation team reviewed a subset of 129
chemicals from the PCCL. The contaminants were divided into groups as follows:
• Contaminants with finished and/or ambient water data
• Contaminants with release data (pesticide applications and/or TRI), and
• Contaminants with production data.
The team evaluated all contaminants with finished and/or ambient water data and a randomly
chosen subset of the contaminants with release or production data. The identities of the
contaminants were blinded for the review. This was done so that the team would focus their
review on the data for a contaminant and not its name. The identity of all contaminants was
revealed when the team discussed the evaluation results.
5.2.2 Evaluation Results
Discussion of the model results raised issues that are important to the selection process for CCL
3 and subsequent CCLs. The evaluators represented a variety of disciplines and contributed
important perspectives reflecting their field of specialization. Below are some of the important
issues that were raised by evaluators:
• The ratio between the health reference value and the concentrations observed in finished
and/or ambient water is an important relationship that is not entirely captured by the four
attribute scores. When finished and/or ambient water data were available, this ratio was
most often the reason for not agreeing with the model output. For example, the model
may have classified a chemical as an L?, but when the health value and concentration
data were compared, the outcome indicated that occurrence was one or more orders of
magnitude below the health-based benchmark. In this situation, the evaluators usually
disagreed with the models decision.
• Confidence in the data elements used for attribute scoring varied widely among the
PCCL contaminants. Evaluators noted that there was a considerable difference in the
weight-of-evidence for the differing types of data used to score PCCL contaminants.
Although the scores used a hierarchy in selecting the data elements that best represented
health effects and occurrence, the most highly ranked data element was not equivalent for
every chemical. Individual chemicals used different combinations of data. The type of
data elements used to represent the occurrence and health effects became a subject of
Page 58 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
discussion for the Evaluation Team. Some contaminants had recent UCMR monitoring
data combined with an Office of Pesticide Programs (OPP) RfD and others had TRI
release data combined with an LD50. For some chemicals, the best data came from an
LDso combined with the number of pounds produced per year and environmental fate
properties. The evaluators were more comfortable with the model decisions based on
strong supporting data than on those based on weak data sets.
• Reviewers felt it was important that the occurrence and health values represent the same
form of the chemical. This is particularly important for nonmetals where the common
inorganic form of the element is a complex ion (i.e. phosphate) and not the element (i.e.
phosphorous). This is also important for metals where the occurrence data represent ions
in solution that may have been paired with a toxicity value for the free metal.
• Toxicity data from National Cancer Institute/National Toxicity Program bioassays were
incorporated into the Universe for a number of contaminants that were positive for
tumors, and were tested by way of the inhalation route of exposure. Some of these
contaminants were screened to the PCCL on the basis of their qualitative cancer findings.
They were scored for Potency and Severity based on slope factors that had been derived
for the oral route of exposure, but based on the inhalation data without the use of
Physiologically Based Pharmacokinetic (PBPK) modeling. Some of these very volatile
contaminants received L or L? model designations. Reviewers questioned whether
toxicity data from inhalation studies should be used for scoring cancer Potency.
Therefore, only cancer slope factors that were derived using PBPK modeling for cross
route extrapolation were used to score chemicals. Inhalation data were not used for non-
cancer endpoints.
• Due to the risk assessment policy differences between agencies, the hierarchy for scoring
Potency and Severity considered the agency that established the value. However, some
reviewers questioned whether the date of the assessment rather than the Agency
conducting the assessment should be the basis for the hierarchy.
• Prevalence and Magnitude were given the lowest possible scores ("1") when a
contaminant had been monitored but there were no detections. Since the detection level
for a few chemicals was above the health-based value, some reviewers questioned
whether this was appropriate. They suggested that it might be better to use the detection
limit as the basis of the Magnitude attribute score.
• UCMR 1 screening studies monitored a small number of statistically selected sites (300).
There were cases where there were no finished water detections in the screening surveys,
but the same contaminant had been detected in ambient water by USGS. Reviewers
questioned the placement of finished water above ambient water in the hierarchy in these
cases.
• A number of disinfection byproducts (DBFs) had occurrence data based on production or
release, while some had no occurrence data. Production and release data do not
Page 59 of 66
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
adequately represent the potential occurrence of DBFs and byproducts of other treatment
processes in finished water.
• Reviewers were uniform in feeling that contaminants that had a Potency score based on
an LD50 value and a Severity score of 9 (death), should be returned to the Universe
independent of their other attribute scores.
The quantitative results of the model output evaluation are summarized in Exhibit 35
For Exhibit 35, agreement with the model outcome by a majority of the Evaluation Team
constitutes agreement. Appendix F lists the chemicals reviewed by the Team and the percentage
of the team agreeing with the model outcome for the individual chemicals.
Exhibit 35. Results of the Model Output Evaluation (Total = 129 chemicals)
Number of Contaminants
Agreement with model outcome (>50%)
% where an outcome higher than the
model was recommended
% where an outcome lower than the model
was recommended
% high confidence decisions (avg.)
% medium confidence decisions (avg.)
% low confidence decisions (avg.)
Finished/Ambient
Water Grouping
89
96%
2%
2%
36%
49%
15%
Release
Grouping
28
89%
0%
11%
16%
31%
52%
Production
Grouping
12
67%
0%
33%
7%
17%
76%
5.3 Post-Model Adjustments to Output
Based upon issues identified by the Evaluation Team comments, several post-model refinements
were added to the CCL process. The post-model refinements changed the standing of some of
the chemicals as candidates for CCL 3. The post-model adjustments that were incorporated are
discussed in the following sections.
The simplest of the post-model adjustments was the review of the coupling of occurrence data
with toxicological data for the inorganic contaminants. This problem was a result of some data
being reported for the element of interest (and its CAS number) and other data being reported for
one or more ions and/or salts that contained the element.
5.3.1 Using Supplemental Sources to Identify the Data Most Relevant to Drinking Water
One issue identified by the Evaluation Team was that scoring should be based on the data most
relevant to exposure from drinking water. For example, DBFs were included in the Universe and
many were brought forward to the PCCL. The data used to score these contaminants for
occurrence should be based on their occurrence in drinking water at PWSs, not ancillary data that
may be available such as release or production volume. There are DBF data from the
Page 60 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
Information Collection Rule monitoring and supplemental studies identified in the CCL
Nominations process, These data had not originally been included in the data used for scoring
Prevalence and Magnitude. As part of the post-model processing the data were retrieved, scored
and the chemicals were modeled using the supplemental data. For future CCLs some of these
supplemental data sources may be included in the Universe and used in the attribute scoring
rather than as a post- model adjustment.
5.3.2 Calculation of a Health-Concentration Ratio for Contaminants with Water Data
The models classified chemicals using scores for the four attributes. The Evaluation Team
recognized that the relationship between Potency and Magnitude was important when deciding
whether or not to list a chemical, but only when the Magnitude data represented concentration in
ambient or finished water. Accordingly, calculation of the ratio between the health-based value
and the 90th percentile concentration in finished or ambient water was added as a post-model
process. EPA also sought methods that could be used to model concentration data to develop a
similar ratio for contaminants that did not have direct measurements in water sources. The
health/concentration ratio serves as a benchmark that suggests concern for a contaminant when it
is low, and lesser concern when it is high.
5.3.2.1 Developing a Health Reference Level (HRL)
To calculate the health-concentration ratio, the data that provided the Potency score were
converted to the HRL benchmark that the Agency has used for Regulatory Determination. For a
carcinogen, the HRL is the one-in-a-million cancer risk expressed as a drinking water
concentration. For non-carcinogens, the HRL is equivalent to the lifetime health advisory value.
The lifetime health advisory value is obtained by multiplying the RfD times 70 kg, dividing by a
water intake of 2 L/day and multiplying by a 20% relative source contribution (unless there are
data to suggest that the 20% is inappropriate).
Determining the HRL for chemicals where the Potency value was the NOAEL, LOAEL, or
value from an individual study, required application of an uncertainty factor to adjust the toxicity
value to an RfD approximation. In these cases, the uncertainty factor was based on the
difference in the modal values from the log-based data distributions used to develop the Potency
scoring equations (see Chapter 2). The uncertainty factors applied are as follows:
NOAEL -1,000
LOAEL - 3,000
LD50 - 100,000
The NOAEL and LDso uncertainties were derived from the difference in the constant for the non-
cancer Potency scoring equation (Exhibit 4). For a NOAEL, the difference is 3 (7 - 4 = 3) or
1,000 since the Potency equation is log based. The difference for an LDso is 5 (7-2 = 5) or
100,000. The uncertainty factor (3,000) chosen for the LOAEL is a half log greater than that for
the NOAEL, in recognition that the LOAEL is a level that causes effects rather than no effects.
Page 61 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
5.3.2.2 Developing a HRL - Concentration Ratio
The 90th percentile (of detections) water concentration was selected as the point of comparison
for the ratio, rather than the mean or median. The CCL list is designed to identify contaminants
that may benefit from a Health Advisory, even if they do not merit a positive regulatory
determination. The 90th percentile concentration level was used as a conservative benchmark
that may identify a possible need for a health advisory for areas of the country that may have
higher concentrations in drinking water than others.
The ratio of the heath-value to the 90th percentile concentration detected in water (either ambient
or finished) was calculated for all contaminants with water data. If the ratio was 10 or less the
contaminant was selected for consideration for the draft CCL 3. If the ratio was greater than 10,
the contaminant was eliminated from consideration for CCL 3 and remains on the PCCL. For
chemicals that had been monitored but not detected, and for chemicals that were detected in
ambient waters but not finished water, analytical method detection limits were compared to the
HRL to ensure that the detection accounted for the health effects. Consideration was also given
to whether the ambient water data suggested that the UCMR 1 screening might have been too
limited to identify the contaminant in areas where it might pose a problem. For contaminants that
had limited finished water data, but more robust ambient water monitoring data, the ambient
water concentration was used to develop the ratio.
5.3.2.3 Developing a Ratio for Contaminants Without Concentration Data
OW worked with the OPP to obtain supplemental modeled data on the levels of pesticides
projected to be found in surface and ground water. The modeled concentrations of pesticides in
water are included in the OPP registration and re-registration evaluation documentation, but they
are not readily available in a form that could be used for the Universe database. Once this data
gap was identified, OPP shared the data with OW and they were evaluated in the post-model
process.
For pesticides, the modeled data from OPP were compared with the health reference level. As
part of the pesticide registration process, EPA calculates an Estimated Environmental
Concentration (EEC) in water or Estimated Drinking Water Concentration (EDWC) depending
on the year the last assessment was completed. Both the EEC and EDWC are derived from
models that estimate the pesticide's concentration in an index reservoir used for drinking water.
OPP used the PRZM-EXAMS model for surface water. Ground water concentrations are derived
using the SCI-GROW regression model to represent exposures in shallow ground water. Both
the EEC and the EDWC are equivalent. The modeled EEC values allowed EPA to calculate the
HRL/EEC or EDWC ratio for pesticides and/or their degradates. Pesticides with HRL/EEC
ratios of 10 and lower were selected for the draft CCL 3.
5.3.3 Grouping Contaminants based on Data Certainty
Data certainty was not directly factored into the development of the attribute scoring protocols,
but was indirectly factored into the protocols through the use of the hierarchies of the data used
for health effects and occurrence (Chapter 2). In the evaluation of the model output, data
certainty was an important factor for the Evaluation Team. In cases where the model output
listed a chemical with data from high in the hierarchy (e.g. IRIS RfD, UCMR/NAWQA
Page 62 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
concentration), the team typically agreed with the model decision. The Team confidence ranking
for model decisions based on data from high in the hierarchy was generally high while
confidence for data from low in the hierarchy was generally low (see Exhibit 35). Accordingly,
as part of the post-model evaluation process, EPA tried various approaches for addressing the
certainty issue.
Initially, OW attempted to develop numeric certainty scores for each data element, but decided
not to use this approach because the certainty scores could not be calibrated due to the
subjectivity in assigning the numeric values. For example, it would be difficult to justify that a
chemical evaluated by environmental release data should be assigned a certainty score of 6,
while a chemical evaluated by production volume should be assigned a certainty score of 10
versus 9. Therefore, OW decided to place tags on the chemicals that characterize the certainty.
The chemicals were tagged as high, medium and low certainty based on the combinations of data
elements that were used to score the attributes for health effects and occurrence. The certainty
tags are not calibrated measures of certainty. They were developed to express the relative
certainty associated with the data elements that were used to score a chemical's attributes. The
certainty rankings assigned to the combinations of individual attribute data elements are listed
below:
High Certainty:
Finished Water + RfD/ CSF, NOAEL or LOAEL
Ambient Water + RfD/CSF, NOAEL
Medium Certainty:
Ambient Water + LOAEL
Release/Application + RfD, NOAEL, LOAEL
Production + RfD
Low Certainty:
Finished Water, Ambient Water or Release/Application + LD50
Production + NOAEL, LOAEL, LD50
The high certainty bin consisted of chemicals that had been scored based on the most relevant
data for occurrence in water and with the richest database for health effects. Such contaminants
are expected to be good candidates for regulatory determination with minimal research needs.
Examples of chemicals in the high bin include chemicals with reference doses and measured
water concentration data. The medium bin consists of chemicals that need further occurrence
and/or health effects research. These include chemicals that may have well studied health effects
data but may need additional occurrence data (e.g. chemicals with release data but, no measured
water occurrence data). The low certainty bin consists of chemicals that need extensive health
effects and occurrence research that may take longer than the life cycle of a CCL. Examples
include chemicals with LD50 and/or production volume data. The CCL should consist both of
chemicals that provide sufficient data to support regulatory determinations, as well as chemicals
that are of concern and need to be targeted for additional drinking water research. Contaminants
from each bin were scrutinized separately in selecting which ones should be listed on the CCL 3.
Page 63 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
5.3.4 LDso Values with Limited Documentation
Following the advice from the Evaluation Team, Severity scores based on death from
studies were removed from the modeled PCCL results. This decision applies to contaminants
where no critical endpoint other than death was specified in the source of the LD50 data. These
contaminants were removed from consideration for the CCL. None of the chemicals with
derived health attributes had ambient or finished water data.
5.4 Selecting the Draft CCL 3
The chemicals for the preliminary CCL 3 were selected from within the three uncertainly bins,
described in Section 5.3.3, with the emphasis placed on the source of the occurrence data (e.g.
measured water concentrations, release, and production). Four groups of chemicals were placed
on the draft CCL 3 based on their modeled scores, the potency-concentration ratios analysis,
where available, and the estimate of data certainty. They included:
• 36 chemicals in the high certainty bin, which have finished water data and an FIRL/90th
percentile concentration ratio of < 10.
• 24 pesticide chemicals in the medium certainty bin, which have modeled surface and/or
ground water data that yielded a HRL/concentration ratio of < 10
• 27 chemicals in the medium certainty bin, which have release data that gave modeled L
or L? rankings
• 8 chemicals in the low certainty bin that were nominated and reviewed with supplemental
information that was submitted, were selected for the CCL.
No chemicals with only LDso and production data were selected for the CCL. These chemicals
are viewed as candidates for research and consideration for later CCLs.
Subsequent to placement on the preliminary CCL 3, the list was subject to review by a panel of
qualified external experts and stakeholders. Stakeholder input was considered in determining
which chemicals from among a preliminary CCL 3 grouping were retained for the proposed CCL
3. After publication of the CCL 3 Proposal, EPA will seek consultation from the Science
Advisory Board and consider additional stakeholder comments on the Federal Register proposal,
before finalizing CCL 3.
5.5 Summary
The Draft CCL 3 and the process used to select contaminants was developed and tested to meet
the Safe Drinking Water Act requirements and address recommendations and advice from the
NRC (2001) and NOW AC (2004). The Agency has developed a draft CCL 3 that:
• Considers of a broad Universe of contaminants
• Relies on best available science and information to inform the process
• Evaluates the known or potential health effects and occurrence in screening the
Universe to a PCCL
Page 64 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
• Uses a set of contaminant attributes and prototype classification algorithms as decision
support tools in selecting candidates for the CCL from the PCCL
• Provides an opportunity for nominations and expert judgment.
The first application of the CCL 3 process accomplished many of the specific recommendations
from NRC and NDWAC. During the development of CCL 3, the Agency identified areas for
improvement that can be implemented in the selection of CCL 4 and later CCLs.
6.0 REFERENCES
Fetter, C. W. 1994. Applied Hydrogeology, 3rd Edition, Macmillan College Publishing Co. New
York.
Lyman, W. I, Reehl, W. F., and Rosenblatt, D. H. 1990. Handbook of Chemical Property
Estimation Methods, American Chemical Society, Washington, DC.
National Drinking Water Advisory Council (NDWAC). 2004. National Drinking Water
Advisory Council Report on the CCL Classification Process to the U. S. Environmental
Protection Agency, May 19, 2004.
National Research Council (NRC). 2001. Classifying Drinking Water Contaminants for
Regulatory Consideration. National Academy Press, Washington DC.
NIST. 2006. NIST/SEMATECH e-Hcmdbook of Statistical Methods. Available on the internet at:
http://www.itl.nist.gov/div898/handbook/, (used on May 3, 2007).
USEPA. 2004. Office for Water. Drinking Water Standards and Health Advisories, EPA 822-R-04-005
Washington, DC. Winter 2004.
USEPA. 2008a. Contaminant Candidate List 3 Chemicals: Identifying the Universe. EPA 815-
R-08-002. Draft. February, 2008.
USEPA. 2008b. Contaminant Candidate List 3 Chemicals: Screening to a PCCL. EPA 815-R-
08-003. Draft. February, 2008.
Page 65 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
7.0 APPENDICES
Page 66 of 66
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
Appendix A. Attribute Scoring Protocols
This section provides scoring protocols for the health effects attributes of Potency and Severity as well as
the Occurrence attributes, Magnitude and Prevalence.
A.I Potency Scoring Protocol
This section describes the process for assigning a numerical score for the Potency attribute.
Protocol for Potency Scoring
Step One: Open the spreadsheet for Potency and Severity Scoring (a sample of this spreadsheet is shown
in Exhibit A. 1) and is an alternative to using the computer version of the spread sheet.
Step Two: Enter the name of the chemical in the column labeled contaminant.
Step Three: Identify and score highest-ranked non-cancer data element for potency using the following
hierarchy of values:
Reference Dose (RfD) or equivalent > No-Observed-Adverse-Effect Level (NOAEL) that is
lower than the lowest LOAEL > Lowest-Observed-Adverse-Effect Level (LOAEL) > Toxic DoseLo
(TDLO- RTECS) > Lethal dose (LD50)
• Measured > Modeled
For RfDs (or equivalent) only:
EPA RfD > ATSDR Minimal Risk Level (MRL) (Chronic> Intermediate >Acute) >
RAISHE RfD > Cal EPA Public Health Goal (PHG)a > TDIs from WHO/EU/Health
Canada > UL from IOM
• Office of Pesticide Programs (OPP) > IRIS for Pesticides
Step Four: Enter the selected quantitative measure of non-cancer potency into the appropriate column of
the spread sheet. Make sure that the units are in mg/kg/day. (The spreadsheet formula produces a score in
a corresponding column for the data element on the right side of the sheet.)
-4
Step Five: Select a measure for cancer potency if one is available. The preferable measure will be the 10
risk concentration in drinking water in mg/L. If the risk is expressed at levels other than 10~4, convert the
value to the target risk (10~4). If the cancer potency measure is the slope factor, calculate the 10~4 risk
concentration using the following equation:
10"4 Risk concentration = 0.0001 x 35 kg/dav/L
Slope Factor (mg/kg/day)
a The California PHG will have to be converted from mg/L to a dose by multiplying it by the prinking Water Intake
(L) + (the body weight (kg) x Relative Source Contribution)].
A-1
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Step Six: In a case where the entered potency value is a LD50 value that is reported as greater than a
particular dose, or as a NOAEL with no LOAEL, decrease the score calculated using the spreadsheet by
one integer. Situations where there is a NOAEL with no LOAEL can be identified by the lack of a
critical effect, because the NOAEL was the highest dose tested.
Step Seven: Choose the higher of the non-cancer or cancer potency scores as the measure of potency.
Note: if no value for Potency can be found that qualifies for this protocol, please refer the contaminant for
expert judgment. The only endpoints that may be applied to this protocol are those listed explicitly in the
hierarchy of values. Further, the only endpoints considered as equivalent to an RfD are MRLs from
ATSDR, RAISHE RfDs, Cal EPA RfDs, WHO or HC, TDIs, and IOM ULs.
Exhibit A.I. Potency Scoring Table
SCORE
10
9
8
7
6
5
4
3
2
1
RfD
0 - 0.000000316
0.000000317 - 0.00000316
0.00000317 - 0.0000316
0.0000317 - 0.000316
0.000317 - 0.00316
0.00317 - 0.0316
0.0317 - 0.316
0.317 - 3.16
3.17 - 31.6
31.7 - >31.7
LOAEL/NOAEL
0 - 0.000316
0.000317 - 0.00316
0.00317 - 0.0316
0.0317 - 0.316
0.317 - 3.16
3.17 - 31.6
31.7 - 316
317 - 3,160
3,170 - 31,600
31,700 - >31,700
LD50
0 - 0.0316
0.0317 - 0.316
0.317 - 3.16
3.17 - 31.6
31.7 - 316
317 - 3,160
3,170 - 31,600
31,700 - 316,000
317,000 - 3,160,000
3,170,000 - >31, 700,000
Car
0 - 0.00000316
3.17E-06 - 0.0000316
3.17E-05 - 0.000316
0.000317 - 0.00316
0.00317 - 0.0316
0.0317 - 0.316
0.317 - 3.16
3.17 - 31.6
31.7 - 316
317 - >317
A-2
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
A.2 Severity Scoring Protocol
The score for Severity is based upon the critical effect associated with the data element (RfD, LOAEL,
etc.) used to score Potency. Potency must be scored prior to Severity.
Protocol for Severity Scoring
Step One: Identify the critical effect for the contaminant, based on the data used to score the
attribute of potency, and enter it into the severity scoring worksheet (shown in Exhibit
A.2). If the contaminant has more than one critical effect all of the listed effects should
be included. NOTE: If the critical effect is death and the LD50 data element was used to
score potency, go to Step Four. If the effects are for a LOAEL from RTECS go to Step
Five.
Step Two: Locate the critical effect within the Compendium of Critical Effects Table (see Exhibit
A.3) and enter the severity score associated with that critical effect in the severity scoring
worksheet. If a contaminant has more than one critical effect, choose the highest of the
scores.
NOTE: If the critical effect is not listed in the Table, go to Step Three.
Step Three: If the critical effect is not listed in the Table, the scorer should flag that critical effect as
'not listed.' (Health effects experts should be consulted to score these effects.) Once the
effect is scored it should be added to the compendium for future use and consistent
scoring.
Step Four: If a critical effect is not available, or is "death," use one of the following options for
scoring:
1) Search sources identified as supplemental sources for CCL for additional health
effects data that could be used to score potency and severity for the contaminant. If
data are found that provide a data element from the potency protocol other than LD50
to score the contaminant, then that element can be used for scoring. Sources that may
be most helpful in this search include: Hazardous Substances Data Bank (HSDB),
International Program on Chemical Safety (INCHEM), and the National Toxicology
Program (NTP). The element that is found may be used to rescore the contaminant for
potency, and subsequently severity, using the score associated with the critical effect
endpoint.
2) Search for an alternative critical effect associated with the LD50 determination. Locate
the LD50 study and search for information regarding the types of effects occurring
prior to animal death. If a critical effect other than death is given in the study, it may
be used to score the severity of the contaminant. (The potency score is still given by
the value of the LD50.)
3) If no additional information can be found, recommend that the contaminant be
returned to the Universe.
Step Five If the Potency score is a LOAEL from RTECS, the effects listed represent all effects and
A-3
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
not just the critical effect (s). There are three available options for improving the scoring
in this situation.
1.
If the RTECS data source is included in the supplemental data, review the
supplemental information to identify the critical effect. If the supplemental source
includes aNOAEL for the critical effect, replace the LOAEL with the NOAEL and
rescore potency if necessary.
In cases where the data source for the LOAEL is not in the supplemental data search
the supplemental data for an alternative data source. If the data identified provides a
NOAEL or LOAEL that is the same or lower than that in RTECS or is from a study
of higher quality than the RTECS study , use that NOAEL or LOAEL and its critical
effect to score both potency and severity.
If it is not possible to find better information in the supplemental data sources score
the most serious of the effects listed in RTECS.
Exhibit A.2. Severity Scoring Table
Key
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
Study used to score Potency
Critical Effect(s) for Severity
Severity Score
A-4
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit A.3. Compendium of Critical Effects Table (from Health Advisories & IRIS)
For Scoring Severity
Severity
Score
Score Definition
Compendium of Critical Effects
1
NO ADVERSE EFFECT
No observed effect(s).
No observed adverse effect(s).
Absence of effects.
No critical effect(s) identified.
No effect(s) related to treatment.
Absence of biologically significant adverse effect(s).
Absence of gross light microscopic histopathological
change(s).
Excedance of the Taste Threshold
COSMETIC EFFECT
(Interpretation: Consider those effects
that alter the appearance of the body
without affecting structure or
functions)
Dental fluorosis.
Abnormal appearance.
Facial flushing.
Flushing.
Argyria.
Dermal sensitization.
Skin pigmentation.
Hyperpigmentation.
Alopecia.
Keratosis.
REVERSIBLE EFFECTS;
DIFFERENCES IN ORGAN
WEIGHTS OR SIZE, BODY
WEIGHTS OR CHANGES IN
BIOCHEMICAL
PARAMETERS WITH
MINIMAL CLINICAL
SIGNIFICANCE.
(Interpretation: Transient, adaptive
effects)
Growth and Weight Effects
Decreased body weight and or body-weight gain.
Increased absolute organ weights.
Increased liver weight.
Increased kidney weight.
Increased relative organ weight.
Decreased relative organ weight.
Lower ovarian weight.
Decreased maternal weight gain.
Increased absolute and relative (to body and/or brain)
liver weight.
Increased kidney body weight ratio.
Increase in spleen weight.
Increase in thyroid/body weight ratio.
Changes in thymus weight.
Decreased body weight.
Decreased growth.
Gastrointestinal Disturbances
Decreased stool quantity.
Osmotic diarrhea.
Diarrhea.
Nausea.
Vomiting.
GI irritation.
GI disturbances.
A-5
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit A.3. Compendium of Critical Effects Table (from Health Advisories & IRIS)
For Scoring Severity
Severity
Score
Score Definition
Compendium of Critical Effects
3 (cont.)
Irritation/Irritability
Chronic irritation.
Maternal hyperirritability.
Chronic irritation without histopathology changes.
Biochemical Changes
Decreased glucose.
Increased blood sugar.
Increased enzymes.
Increased triglycerides.
Increase serum concentration of compound.
Clinical serum effects.
Alterations in clinical chemistry.
Increased serum alkaline phosphatase.
Significant elevation of serum calcium levels.
Enzyme inhibition, induction, or change in blood
tissue levels
Decreased ESOD activity.
Decrease in erythrocyte superoxide dismutase
(ESOD) concentration.
Minor alteration in clinical chemistry, e.g., decrease
in erythrocyte superoxide dismutase (ESOD).
Hematological effects
Hematological effects.
Abnormal pigments in blood.
Decreased lymphocyte count.
Decreased blood counts.
Methemoglobinemia.
Increased carboxyhemoglobin.
Hemosiderosis.
Anemia.
Normocytic anemia.
Iron deposits and elevated Heinz bodies in liver.
Decreased hemoglobin and possible erythrocyte
destruction.
Decreased RBC, packed cell volume, and
hemoglobin.
Hematologic, hepatic, and renal toxicity as evidenced
by a statistically significant decrease in
hemoglobin, hematocrit, and RBC levels.
RBC and liver effects as evidenced by increase Heinz
bodies in RBC.
Sporadic decrease in hemoglobin and RBC.
Decreased RBC and hematocrit.
A-6
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit A.3. Compendium of Critical Effects Table (from Health Advisories & IRIS)
For Scoring Severity
Severity
Score
Score Definition
Compendium of Critical Effects
3 (cont.)
Cholinesterase Effects
Reversible PChE (plasma) orRBC-ChE inhibition
without cholinergic symptoms or signs
RBC ChE depression without cholinergic symptoms
or sweating.
Plasma Cholinesterase (ChE) inhibition without
cholinergic symptoms or sweating.
Hormone Changes
Decrease inT3, T4.
Dose-related decrease in T4, T3, and increase TSH.
Elevated thyroid stimulating hormone (TSH)
concentration.
ACTH decrease.
Cellular Vacuolization
Mild to moderate vacuolization
Tubular epithelial vacuolization.
Brain cell vacuolization.
Additional Effects
Changes in teeth and supporting structures.
Sensory organ effects.
Centrilobular eosinophilic liver changes.
Possible vascular complication
CELLULAR/PHYSIOLOGICAL
CHANGES THAT COULD
LEAD TO DISORDERS (risk
factors or precursor effects).
(Interpretation: Considers
cellular/physiological changes in the
body that are used as indicators of
disease susceptibility)
Hematological Effects
Jaundice.
Anemia
Hemolytic anemia.
Erythrocyte destruction.
Hemolysis.
Immunological Effects
Decreased delayed hypersensitivity response.
Decrease in cellular immune response.
Decrease in humoral immune response.
Liver Effects
Fatty cyst - liver and elevated liver enzymes (i.e..
SGPT, LDH).
Liver cell enlargement or alteration.
Liver cell polymorphism.
Proteinuria.
Renal cytomegaly.
Cholinergic Effects
Cholinesterase inhibition with symptoms.
Cholinergic signs or symptoms.
Other Effects
Hypothermia
Mild CNS Effects
A-7
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit A.3. Compendium of Critical Effects Table (from Health Advisories & IRIS)
For Scoring Severity
Severity
Score
Score Definition
Compendium of Critical Effects
SIGNIFICANT FUNCTIONAL
CHANGES THAT ARE
REVERSIBLE OR
PERMANENT CHANGES OF
MINIMAL TOXICOLOGICAL
SIGNIFICNACE.
(Interpretation: Consider those
disorders in -which the removal of
chemical exposure will restore health
back to prior condition)
Increased cholinergic effects
ChE inhibition with sweating, diarrhea, hypotention,
and/or fishy body odor..
RBC and/or plasma acetylcholinesterase (AChE)
inhibition with cholinergic symptoms or sweating.
Brain acetylcholineesterase inhibition with or without
signs or symptoms
Hematological Effects
GI bleeding.
Coagulation defects.
Tendency to hemorrhage.
Structural Effects
Rachitic bone.
Renal Effects
Renal cytomegaly.
Renal effects/toxicity (increased uric acid levels;
increased urinary coproporphyrins).
Inflammatory foci - kidneys.
Hepatic Effects
Liver function tests impaired.
Fatty-cyst in liver hemosiderosis.
Multiple Organ Effects
Effects on the lungs, liver, kidney, thyroid and
thyroid hormones.
Ocular Effects
Corneal damage.
Neurological Effects
Mild neurological signs.
Alteration of classic conditioning.
Brain ChE inhibition.
Myelin degeneration.
CNS depression.
Brain/ other coverings- recordings from specific
areas of CNS.
Tremors.
Dyspnea.
Changes in motor activity.
Hypoactivity.
Ataxia.
A-8
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit A.3. Compendium of Critical Effects Table (from Health Advisories & IRIS)
For Scoring Severity
Severity
Score
Score Definition
Compendium of Critical Effects
5 (cont.)
Other Effects
Chronic pneumonitis.
Clinical selenosis.
Nonneoplastic lesions - splenic capsule.
Intestinal lesions.
Splenomegaly
SIGNIFICANT,
IRREVERSIBLE,
NONLETHAL CONDITIONS
OR DISORDERS.
(Interpretation: Consider those
disorders that persist for over a long
period of time but do not lead to death)
Multiple Organ Effects
Histopathological effects in liver, kidney, and
thyroid.
Minimal to moderate congestion of liver, kidney, and
lungs.
Liver and kidney pathology.
Kidney and spleen pathology.
Hepatic Effects
Hepatic lesions/necrosis.
Hepatocyte degeneration.
Hepatotoxicity.
Liver cell polymorphism.
Liver effects/toxicity.
Liver lesions.
Renal Effects
Atrophy and degeneration of the renal tubules -
nephropathy (unspecified).
Kidney toxicity.
Mineralization of the kidneys.
Renal dysfunction.
Renal effects/toxicity (increased uric acid levels;
increased urinary coproporphyrins).
Functional and histopathological effects in kidney.
Kidney damage (unspecified).
Kidney lesions (unspecified).
Impaired renal clearance/function.
Tubular epithelial vacuolation.
Sensory and Neurological Effects
Significant decrease in brain and brain to body
weight ratio.
Degenerative changes for brain/ other coverings.
Peripheral neuropathy- neuropathy (unspecified).
Neurotoxicity.
Nerve damage (unspecified).
Optic nerve degeneration/ damage.
Sensory neuropathy.
Minimal lens opacity and cataracts.
Nasal olfactory lesions.
A-9
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit A.3. Compendium of Critical Effects Table (from Health Advisories & IRIS)
For Scoring Severity
Severity
Score
Score Definition
Compendium of Critical Effects
6 (cont.)
Hyperplasia
Thyroid hyperplasia.
Urothelial hyperplasia.
Hyperplasia.
Squamous and basal hyperplasia of the
forestomach.
Epithelial hyperplasia - forestomach.
Cardiac Effects
Cardiac toxicity.
Cardiomyopathy, including infarction.
Vascular complications.
Right atrial dilation.
Convulsions.
Mild histological lesions.
Other Effects
Gastrointestinal necrotic changes.
Chronic irritation with histopathology findings.
Forestomach lesions (unspecified).
Organ atrophy.
Thyroid effects (unspecified).
Spleen toxicity (unspecified).
Bladder toxicity (unspecified).
Bone marrow toxicity (unspecified).
DEVELOPMENTAL OR
REPRODUCTIVE EFFECTS
LEADING TO MAJOR
DYSFUNCTION.
(Interpretation: Considers those
chemicals that cause permanent
developmental effects or that impact
the ability of a population to
reproduce)
Reproductive Organ Effects
Testicular atrophy/damage.
Testicular and uterine effects.
Atrophied seminiferous epithelium.
Histopathological changes in testes.
Lesions observed in reproductive organs.
Decreased testes weight and testes to body weight
ratio, atrophied seminiferous epithelium; and
decreased tubular size in testes.
Endometriosis.
Decreased tubular size in testes.
Decreased ovarian weight and function.
Altered cellular foci.
Maternal Toxicity
Maternal toxicity.
Decreased maternal weight gain.
A-10
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit A.3. Compendium of Critical Effects Table (from Health Advisories & IRIS)
For Scoring Severity
Severity
Score
Score Definition
Compendium of Critical Effects
7 (cont.)
Fertility effects
Spermatogenic arrest.
Reduced numbers of corpora allata.
Reduced or deformed sperms.
Adverse reproductive effects.
Reduction in fertility.
Decreased fertility index.
Decrease in size of litter.
Growth inhibition
Reduced offspring weight gain, total litter weight, or
litter size.
Decreased pup weight
Decreased lactation indices.
Increased runt incidence.
Decreased crown-rump length
Decreased offspring viability
Excessive loss of litters
Increase in number of stillbirths.
Maternal and fetal toxicity.
Increased intrauterine death.
Decreased pup survival or viability.
Increased abortion rate.
Increase in number of stillbirths.
Increased dead pups at birth.
Decreased pup viability index.
Parturition mortality.
Fetal resorptions.
Developmental effects
Fetal toxicity /malformations.
Developmental toxicity (skeletal or visceral
abnormalities).
Delayed ossification.
Neurodevelopmental effects.
Brain cell vacuolization in neonates.
Myelin degeneration.
Skeletal or visceral abnormalities (Extra ribs and
other measures of sexual maturation).
Increased retinal folds in weanlings.
Mixed sexual differentiation (i.e., effeminization or
emasculanization).
Imbalance in sex ratio.
TUMORS OR DISORDERS
LIKELY LEADING TO DEATH
(Interpretation: Considers chemical
exposures that result in a fatal disorder
and all types of tumors).
Cancer.
Suspected carcinogenicity (including short latency
periods and rare tumors).
Any type of cancer.
A-11
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit A.3. Compendium of Critical Effects Table (from Health Advisories & IRIS)
For Scoring Severity
Severity
Score
Score Definition
Compendium of Critical Effects
DEATH.
Increased mortality.
Longevity.
Mortality.
Survival.
Decreased survival.
Increased mortality.
Decreased adult survival.
Decreased adult longevity.
High incidence of mortality at early age (i.e., 25% to
50% by mid-life) in chronic studies.
Maternal death during pregnancy.
Reduced longevity.
Death.
A.3 Prevalence Scoring Protocol
This section describes how to assign a numerical score for the attribute Prevalence.
Step One: Identify highest-ranked data value
When more than one data value is available for a particular contaminant candidate, use the hierarchy in
Exhibit A.4. Use the same type of data to score Prevalence as for Magnitude.
Exhibit A.4. Hierarchy of Prevalence Data Elements
Rank
1
2a
2b
3
4
5
Prevalence Data Element
Finished Drinking Water- Percentage of all
Public Water Systems (PWSs) with Detections
(If data from both NCOD Round 1 and Round 2
are available, use the higher of the values.)
Percentage of all Ambient/Raw/Source
Monitoring Samples or Sites with Detections
Percentage of Ambient/Raw/Source Monitoring
Samples or Sites with Detections (Note: use
combined surface / ground water if available
and higher of SW/GW if not)
Pesticide application data, number of states
where pesticide was applied
Environmental release data, number of states
reporting releases
Production volume data
Type of Data
National scale / representative data (data
from UCMR has highest priority, then
NCOD, then NIRS)
National scale / representative data
(NAWQA)
National scale / representative data (NREC
-first use National Reconnaissance data,
then National Aggregate data)
From NCFAP
From TRI
From Chemical Update System/ Inventory
Update Rule (CUS/IUR)
A-12
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Step Two: Use scoring table to find attribute score for value identified in Step One.
For each element there is a corresponding column in the Prevalence Scoring table (see Exhibit A.5),
which contains a range of data values assigned to a numeric prevalence score between 1 and 10. Once a
data value has been found for a particular element, look up the value in Exhibit A.5 to determine the
prevalence score. For CUS/IUR data, use the most recent year reported. For pesticides, if the compound is
a degradate and does not have its own data, use the parent to score.
Exhibit A.5. Prevalence Scoring Scales
Hierarchy
Prevalence
Score
1
2
3
4
5
6
7
8
9
10
1
% Finished
Water PWSs
with detections
of contaminant
All PWSs
<=0.10
0.11-0.16
0.17-0.25
0.26-0.44
0.45-0.61
0.62-1.00
1.01-1.30
1.31-2.50
2.51-10.00
>10.00
2
% Ambient
water sites
with
detections of
contaminant
All
sites/samples
<=0.10
0.11-0.16
0.17-0.25
0.26-0.44
0.45-0.61
0.62-1.00
1.01-1.30
1.31-2.50
2.51-10.00
>10.00
3
# States Reporting
Pesticide in Use
~
—
Default for any
pesticide in non-
environmental use
~
Default for any
pesticide in
environmental use
without data
<6
6-10
11-15
16-25
>25
4
# of States
Reporting TRI
total releases
1
2
3
4
5
6
7-10
11-15
16-25
>25
5
CUS/IUR
(production data)
Number of pounds
(by category)
produced
<500K
—
>500K-1M
~
>1M-10M
>10M-50M
>50M-100M
>100M-500M
>500M-1B
>1B
Note:
Use data in the highest category to score.
For CUS/IUR data, use the most recent year reported. Not Reported means there has been no change
in production volume since the last report.
For pesticides, if the compound is a degradate and does not have its own data, use the parent to
score.
A-13
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
A.4 Magnitude Scoring Protocol
This section describes how to assign a numerical score for the attribute Magnitude.
Step One: Identify the highest-ranked data element
When more than one data element is available for a particular contaminant, use the hierarchy below to
select the preferred element. Exhibit A. 6 presents the hierarchy of data elements to be used in the
Magnitude scoring process. Note that the Magnitude element should be correlated with the value used to
score the attribute Prevalence, except when production data are used for Prevalence and Persistence-
Mobility is used for Magnitude.
Exhibit A.6. Hierarchy of Magnitude Data Elements
Rank
1
2a
2b
3
4
5
Magnitude Data Element
Finished Drinking Water- Median of
detected concentrations from all Public
Water Systems with detections (If data
from both NCOD Round 1 and Round 2
are available, use the higher of the
values.)
Median of detected concentrations from
all ambient / raw source monitoring sites
with detections
Median of detected concentrations from
ambient / raw / source water samples
with detections (Note: use combined
surface / ground water if available and
higher of SW/GW if not)
Pesticide application data
Environmental release data, total
pounds or tons reported as released
(TRI)
Persistence - Mobility (Environmental
Fate Data)
Type of Data
National scale finished drinking water occurrence
data [data from Unregulated Contaminant
Monitoring Rule (UCMR) has highest priority, then
the National Contaminant Occurrence Database
(NCOD), then the National Inorganics
Reconnaissance Survey (NIRS)]
National scale ambient monitoring data (National
Water Quality Assessment Program - NAWQA)
National scale / representative data (National
Reconnaissance of Emerging Contaminants - NREC
- first use National Reconnaissance data, then
National Aggregate data)
From National Center for Food and Agricultural
policy (NCFAP)
From Toxics Release Inventory (TRI)
Physical chemical properties
Step Two: Use scoring table to fend attribute score for value identified in Step One.
For each data element, there is a corresponding column in the Magnitude Scoring table (Exhibit A.7),
which contains a range of data values assigned to a numerical magnitude score. Locate the column in the
table associated with the highest-ranking data element identified in step one. Use the information in the
column to determine the numerical score associated with the data value for the chemical being scored.
The number corresponding to each "Score" is the maximum in that category, e.g. 0.1 ug/L for finished
A-14
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
water scores 4, not 5. In cases where there are no data for Scoring Magnitude in Exhibit A.7 (e.g.
Prevalence is scored using Production Volume data), use the Persistence-Mobility Scoring approach to
develop a Magnitude Score.
Persistence-Mobility Scoring
The approach for scoring persistence and mobility includes assigning two values, one for persistence and
one for mobility, on a numeric scale of 1 through 3, representing low, medium, and high for each property
as it favors the presence of the contaminant in water. Using a hierarchy of physical property data
elements, each contaminant is scored for both persistence and mobility. The average of these two values
is multiplied by 10/3 to obtain the persistence-mobility score. Exhibit A.8 displays the hierarchy of
available properties for each data element representing either persistence or mobility.
Protocol for Persistence-Mobility Scoring
Step One: Identify and score highest-ranked data value for Persistence
When more than one data element value is available for a particular contaminant candidate, use the
hierarchy below to select the preferred element. Exhibit A.6 describes the hierarchy of data elements to be
used in the Persistence scoring process. When several values for a physical property are available, the
highest scoring value should be used, unless that value is not representative of environmental conditions
in drinking water.
Step Two: Identify and score highest-ranked data value for Mobility
The hierarchy of physical properties for scoring mobility is given in Exhibit A.6. Select the highest
priority data element available for scoring. When several values for a particular physical property are
available, the highest scoring value should be used for scoring, unless that value is not representative of
environmental conditions in drinking water.
Step Three: Multiply the average of the persistence and mobility values by 10/3 for the
magnitude score.
Exhibit A.7. Magnitude Scoring Scales
Hierarchy
Magnitude
Scale
Data Used
to Score
Units
Score
1
2
1
Finished Water
Occurrence
Scale
Median of
detections - all
PWSs
ug/L
2
Ambient Water
Occurrence
Scale
Median of
detections - all
sites/samples
ug/L
3
Pesticide Use
Scale
Number of pounds
applied
Ibs
4
TRI Total
Releases Scale
Total number of
pounds released
Ibs
<0.003
0.003-0.01
<0.003
0.003-0.01
<10,000
<300
301-1,000
5
Persistence/
Mobility
Used when
Production
data are used
to score for
prevalence.
A-15
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit A.7. Magnitude Scoring Scales
Hierarchy
Magnitude
Scale
Data Used
to Score
3
4
5
6
7
8
9
10
1
Finished Water
Occurrence
Scale
Median of
detections - all
PWSs
>0.01 -0.03
>0.03-0.1
>0.1 -0.3
>0.3- 1
>1 -3
>3- 10
>10-30
>30
2
Ambient Water
Occurrence
Scale
Median of
detections - all
sites/samples
>0.01 -0.03
>0.03-0.1
>0.1 -0.3
>0.3-1
>1 -3
>3-10
>10-30
>30
3
Pesticide Use
Scale
Number of pounds
applied
10,000-30,000
30,001-100,000
100,001-300,000
300,001-1M
1M-3M
3M-10M
10M-30M
>30M
4
TRI Total
Releases Scale
Total number of
pounds released
1,001-3,000
3,001-10,000
10,001-30,000
30,001-100,000
100,001-300,000
300,001-1M
1M-3M
>3M
5
Persistence/
Mobility
See
Persistence/
Mobility
protocol
(Exhibit A.6)
Notes:
Use data in the highest category to score.
The number corresponding to each "Score" is the maximum in that category, e.g. 0.1 ug/L scores 4, not 5.
For pesticides, use the parent to score if the compound is a degradate and does not have its own data.
Exhibit A.8. Magnitude Scales for Environmental Fate Data
Magnitude Hierarchy 5
Mobility Scale Value
1
2
3
4
5
6
7
Organic Carbon
Partitioning Coefficient
(Koc)
Log Octa no I/Water
Partitioning Coefficient
(log Kow)
Soil/Water Distribution
Coefficient (Kd)
Henry's Law
Coefficient (KH)
Henry's Law
Coefficient (KH)
Solubility
Percent in water (PBT
Profiler)
Units
mL/g
dimensionless
mL/g
atm-m3/mol
dimensionless
mg/L
dimensionless
1 (Low)
>1,000
>4
>10
>10"3
>0.042
<1
<25
2 (Medium)
100-1,000
1-4
1-10
10"7-10"3
0.042-
4.2x10"6
1-1,000
>25-50
3 (High)
<100
<1
<1
<10"7
<4.2x10"6
>1,000
>50
Persistence Scale
Value
Units
1 (Low) | 2 (Medium)
3 (High)
A-16
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit A.8. Magnitude Scales for Environmental Fate Data
Magnitude Hierarchy 5
Mobility Scale
Value
1
2
3
Half Life (t1/2)
Measured Degradation
Rate1
Modeled Degradation
Rate (PBT Profiler)
Units
time
time
time
1 (Low)
days,
days-
weeks
days,
days-
weeks
(BF
BFA)2
days,
days-
weeks
2 (Medium)
weeks,
weeks-
months
weeks,
weeks-
months
(BS, BSA)
weeks,
weeks-
months
3 (High)
months,
recalcitrant
months,
recalcitrant
(BSD
months,
recalcitrant
When two results are found for a measured degradation rate, the data are "averaged" and then a
value determined.
2 BF = Biodegrades Fast, BFA = Biodegrades Fast with Acclimation, BS = Biodegrades Slow, BST =
Biodegrades Sometimes.
A-17
-------
CCL 3 Chemicals:
EPA-OGWDW Classification of the PCCL to CCL
Appendix B. Example Blinded Information Sheets from the TDS Exercises Contaminant 3
EPA815-R-08-004
February 2008 - DRAFT
Contaminant Name:
Background:
HEALTH EFFECTS DATA
Data Element
Reference Dose
Carcinogen classification (EPA)
Slope Factor
Carcinogen Classification (IARC)
Non EPA Derived Dose1
Critical Effect
File/Issue Date
Lowest Oral Chronic LOAEL1
Lowest Oral LD501
Is contaminant on list of carcinogens?
It is a volatile organic chemical. It is used as a wetting and dispersing agent in textile processing, dye-baths, stain and printing compositions; used in cleaning and
detergent preparations, adhesives, cosmetics, deodorants, fumigants, emulsions and polishing compositions. Used in lacquers, paints, varnishes, paint and varnish
removers. Degreasing agent. It is on the TSCA list. The reportable released quantity of this substance under CERCLA is 1 Ib. It is also subject to RCRA waste
management requirements, and is listed as a hazardous air pollutant by EPA. Several states have drinking water guidelines for this chemical (CA, FL, MA, ME, NC).
Its one-day Health Advisory Level (HAL) is 4,000 ug/L, its 10-day HAL is 400 ug/L, and its 10A-4 cancer risk HAL is 300 ug/L. This is an HPV chemical. It is also on
the CCL. (HSDB, 2005; EPAHA, 2004)
Value
N/A
Units
B2 (probable human
rarrinnnpn)
0.011
2B (possible)
0.1
Hepatic
pffprtQ
10/1/2004
N/A
N/A
Y
1/(mg/kg-d)
mg/kg-d
Y/N
Source
IRIS
IRIS
IARC
ATSDR
MR I
Cal EPA Chemicals Known to the State to
Cause Cancer or Reproductive Toxicity
Notes
9/1/1990
9/1/1990
Chronic
nral
UF=100
1/1/1988
B-l
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCLto CCL
Contaminant 3
EPA815-R-08-004
February 2008 - DRAFT
Is the contaminant on a list of
reproductive toxins?
Risk assessment ongoing?
Health Reference Level (HRL)^
Health Reference Level (HRL)^cancer
Health Reference Level (HRL) cancer
Notes
N
Y
700
3.18
300
Y/N
Y/N
ug/L
ug/L
ug/L
Cal EPA Chemicals Known to the State to Cause
Cancer or Reproductive Toxicity
Based on
MRL
10"4 cancer risk Health Advisory (EPAHA,
1987)
1 Non-EPA toxicology data will be sought if no EPA Reference Dose or carcinogen information available; may require multiple entries; chronic
studies will be prioritized over short term studies.
i Health Reference Level calculated by conversion of RfD or other dose to units of ug/L, assuming 2 liters per day of water consumed by a 70 Kg adult, and a default Relative Source Contribution of 20%. For
carcinogens, the concentration at the 1 0"6 cancer risk level will be converted to units of ug/L and will also be listed.
OCCURRENCE DATA
Water Occurrence Data
Finished Water Occurrence - total
Source Water-Total
Production/Release
Production data
Pesticide Application - total
Pesticide Application - total (# States)
Release - total
#
PWSs/Sites
sampled
No Data
#
PWSs/Sites
sampled
No Data
Value
>1M- 10M
N/A
N/A
1,146,641
#with
Detects
No Data
#with
Detects
No Data
Units
Ibs/yr
Ibs/yr
# States
Ibs/yr
%
Detects
No Data
%
Detects
No Data
Source
CUS-IUR
(2002)
TRI
Minimum of
Detects
(ug/L)
No Data
Minimum of
Detects
(ug/L)
No Data
Maximum
of Detects
(ug/L)
No Data
Median of
Detects
(ug/L)
No Data
Median of
Detects
(ug/L)
No Data
Mean of
Detects
(ug/L)
No Data
Notes
99% of
Detects
(ug/L)
No Data
90% of
Detects
(ug/L)
No Data
Source
95% of
Detects
(ug/L)
No Data
Notes
99% of
Detects
(ug/L)
No Data
Maximum
of Detects
(ug/L)
No Data
Source
B-2
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCLto CCL
EPA815-R-08-004
February 2008 - DRAFT
Contaminant 3
Release - total (# States)
Release - to Surface Water
Release - to SW (# States)
Environmental Fate Parameters
T1/2, Half life
Koc, Organic Carbon Partition
Coefficient
KOW, Octanol Water Partition Coefficient
HLC, Henry's Law Constant
Water Solubility
Kd, Distribution Coefficient
22
75,119
9
Value
No Data
1
Log -0.27
0.000196
1 ,000,000
N/A
No Data = No data found for this contaminant; N/A = Not
applicable to contaminant
# States
Ibs/yr
# States
Units
length of
time
L/kg
unitless
unitless
mg/L
source
specific
TRI
TRI
TRI
Source
RAISCF
RAISCF
RAISCF
RAISCF
Notes
B-3
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCLto CCL
Contaminant 4
EPA815-R-08-004
February 2008 - DRAFT
Contaminant Name:
Background:
HEALTH EFFECTS DATA
Data Element
Reference Dose
Carcinogen classification (EPA)
Slope Factor
Carcinogen Classification (IARC)
Non EPA Derived Dose1
Lowest Oral Chronic LOAEL1
Lowest Oral LD501
Is contaminant on list of carcinogens?
This is a volatile organic chemical. It is used as a food additive, organic intermediate, solvent, and in cosmetic formulations. It is also used as a solvent or
solubilizer in the paint and printing ink sector, as components in textile auxiliaries and pesticides, for hormone extraction, and in the surfactant field as foam
boosters or antifrothing agents. Per the FDA, this food additive is permitted for direct addition to food for human consumption as a synthetic flavoring substance
and adjuvant. (HSDB, 2005)
Value
N/A
N/A
N/A
N/A
N/A
N/A
500
N
Units
mg/kg
Y/N
Source
RTECS
Notes
Cal EPA Chemicals Known to the State to Cause
Cancer or Reproductive Toxicity
B-4
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCLto CCL
EPA815-R-08-004
February 2008 - DRAFT
Contaminant 4
Is the contaminant on a list of reproductive
toxins?
Risk assessment ongoing?
Health Reference Level (HRL)"
Health Reference Level (HRL)^ cancer
Notes
N
N
N/A
N/A
Y/N
Y/N
ug/L
ug/L
Cal EPA Chemicals Known to the State to Cause
Cancer or Reproductive Toxicity
1 Non-EPA toxicology data will be sought if no EPA Reference Dose or carcinogen information available; may require multiple entries;
chronic studies will be prioritized over short term studies.
i Health Reference Level calculated by conversion of RfD or other dose to units of ug/L, assuming 2 liters per day of water consumed by a 70 Kg adult, and a default Relative Source Contribution of 20%. For
carcinogens, the concentration at the 10"6 cancer risk level will be converted to units of ug/L and will also be listed.
OCCURRENCE DATA
Water Occurrence Data
Finished Water Occurrence - total
Source Water-Total
Production/Release
Production data
Pesticide Application - total
Pesticide Application - total (# States)
Release - total
#
PWSs/Sites
sampled
No Data
#
PWSs/Sites
sampled
No Data
Value
>500K - 1 M
N/A
N/A
No Data
#with
Detects
No Data
#with
Detects
No Data
Units
Ibs/yr
Ibs/yr
# States
Ibs/yr
%
Detects
No Data
%
Detects
No Data
Source
CUS-IUR
(2002)
Minimum of
Detects
(ug/L)
No Data
Minimum of
Detects
(ug/L)
No Data
Maximum
of Detects
(ug/L)
No Data
Median of
Detects
(ug/L)
No Data
Median of
Detects
(ug/L)
No Data
Mean of
Detects
(ug/L)
No Data
Notes
99% of
Detects
(ug/L)
No Data
90% of
Detects
(ug/L)
No Data
Source
95% of
Detects
(ug/L)
No Data
Notes
99% of
Detects
(ug/L)
No Data
Maximum
of Detects
(ug/L)
No Data
Source
B-5
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCLto CCL
EPA815-R-08-004
February 2008 - DRAFT
Contaminant 4
Release - total (# States)
No Data # States
Release - to Surface Water
Release - to SW (# States)
Environmental Fate Parameters
T1/2, Half life
Koc, Organic Carbon Partition Coefficient
Kow, Octanol Water Partition Coefficient
HLC, Henry's Law Constant
Water Solubility
Kd, Distribution Coefficient
No Data = No data found for this contaminant;
No Data
No Data
Value
No Data
15
Log 2.62
1 .88E-05
1000
N/A
N/A = Not
applicable to contaminant
Ibs/yr
# States
Units
length of
time
L/kg
unitless
atm-cu
m/mol
mg/L
source
specific
Source
HSDB
HSDB
HSDB
HSDB
Notes
B-6
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCLto CCL
EPA815-R-08-004
February 2008 - DRAFT
Contaminant 5
Contaminant Name:
Background:
HEALTH EFFECTS DATA
Data Element
Reference Dose
Critical Effect
File/Issue Date
Reference Dose
Critical Effect
File/Issue Date
Carcinogen classification (EPA)
Slope Factor
Carcinogen Classification (IARC)
Non EPA Derived Dose1
Lowest Oral Chronic LOAEL1
Lowest Oral LD501
This is a volatile organic chemical registered for use in the U.S. Nematicide. Seventh most commonly used pesticide in U.S. agricultural crop production. Used in
organic synthesis and in manufacture of pesticides. Pre-plant soil fumigant. It is listed on FIFRA and TSCA. The reportable release quantity under CERCLA is 100
Ibs. It is subject to RCRA waste management requirements. It is listed as a hazardous air pollutant and as a hazardous substance by the Federal Water Pollution
Control Act and the Clean Water Act. It has a state drinking water standard in CA. It has a state drinking water guideline in several states (FL, MA, ME, MN, Wl). It
has a DWEL of 1 ,000 ug/L, and its one-day and ten-day Health Advisory Levels (HALs) are 30 ug/L. This is an HPV chemical. (HSDB, 2005; EPAHA, 2004)
Value
0.03
Chronic
irritation
5/25/2000
0.025
Units
mg/kg-d
mg/kg-d
Source
IRIS
OPP
decrease in body weight gain and an increase in the incidence of basal cell
hyperplasia of the nonglandular mucosa of the stomach
1998
B2; inadequate in humans,
sufficient in animals
0.1
2B (possible)
N/A
N/A
N/A
1 /(mg/kg-d)
IRIS
IARC
IRIS
Notes
Basis = BMDL(10) 3.4 mg/kg-d Rat,
UF=100, MF=1
Confidence: Study: High; Database:
High; RfD: High
Basis = NOEL 2.5 mg/kg-d Rat, UF=100,
MF=1
5/25/2000
B-7
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCLto CCL
EPA815-R-08-004
February 2008 - DRAFT
Contaminant 5
Is contaminant on list of carcinogens?
Is the contaminant on a list of
reproductive toxins?
Risk assessment ongoing?
Health Reference Level (HRL)'
Health Reference Level (HRL)^ cancer
Health Reference Level (HRL) cancer
Notes
Y
N
N
210
0.35
40
Y/N
Y/N
Y/N
ug/L
ug/L
ug/L
Cal EPA Chemicals Known to the State
to Cause Cancer or Reproductive
Toxicity
1/1/1989
Cal EPA Chemicals Known to the State to Cause
Cancer or Reproductive Toxicity
Based on
IRIS RfD
Based on IRIS slope
factor
1 0'4 cancer risk Health Advisory (EPAHA,
1988)
1 Non-EPA toxicology data will be sought if no EPA Reference Dose or carcinogen information available; may require multiple entries; chronic
studies will be prioritized over short term studies.
i Health Reference Level calculated by conversion of RfD or other dose to units of ug/L, assuming 2 liters per day of water consumed by a 70 Kg adult, and a default Relative Source Contribution of 20%. For
carcinogens, the concentration at the 10"6 cancer risk level will be converted to units of ug/L and will also be listed.
OCCURRENCE DATA
Water Occurrence Data
Finished Water Occurrence - total
Finished Water Occurrence - SW
Finished Water Occurrence - GW
# PWSs/Sites
sampled
9,164
898
8,303
#with
Detects
15
5
10
% Detects
0.16%
0.56%
0.12%
Minimum
of Detects
(ug/L)
0.5
1
0.5
Maximum
of Detects
(ug/L)
2
2
1.6
Median of
Detects
(ug/L)
1
1.25
0.5
99% of
Detects
(ug/L)
2
2
1.6
Source
NCOD
Round 1
NCOD
Round 1
NCOD
Round 1
Notes
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCLto CCL
EPA815-R-08-004
February 2008 - DRAFT
Contaminant 5
Finished Water Occurrence - total
Finished Water Occurrence - SW
Finished Water Occurrence - GW
Source Water-Total
Production/Release
Production data
Pesticide Application - total
Pesticide Application - total (# States)
Release - total
Release -total (# States)
Release - to Surface Water
Release - to SW (# States)
Environmental Fate Parameters
T1/2, Half life
Koc, Organic Carbon Partition
Coefficient
16,787
1,609
15,178
# PWSs/Sites
sampled
No Data
Value
>1M-10M
34,717,237
20
10,532
8
85
3
Value
No Data
81
58
10
48
#with
Detects
No Data
Units
Ibs/yr
Ibs/yr
# States
Ibs/yr
# States
Ibs/yr
# States
Units
length of
time
L/kg
0.35%
0.62%
0.32%
% Detects
No Data
Source
CUS-IUR
(2002)
NCFAP
NCFAP
TRI
TRI
TRI
TRI
Source
RAISCF
0.2
0.2
0.2
Minimum
of Detects
(ug/L)
No Data
39
1.6
39
Median of
Detects
(ug/L)
No Data
0.5
0.5
0.5
Mean of
Detects
(ug/L)
No Data
Notes
Notes
39
1.6
39
90% of
Detects
(ug/L)
No Data
NCOD
Round 2
NCOD
Round 2
NCOD
Round 2
95% of
Detects
(ug/L)
No Data
99% of
Detects
(ug/L)
No Data
Maximum
of Detects
(ug/L)
No Data
Source
B-9
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCLto CCL
Contaminant 5
EPA815-R-08-004
February 2008 - DRAFT
Kow, Octanol Water Partition
Coefficient
HLC, Henry's Law Constant
Water Solubility
Kd, Distribution Coefficient
Log 2.03
0.145
2,800
N/A
No Data = No data found for this contaminant; N/A = Not
applicable to contaminant
unitless
unitless
mg/L
source
specific
RAISCF
RAISCF
RAISCF
B-10
-------
EPA-OGWDW
CCL3 Chemicals:
Classification of the PCCLto CCL
EPA815-R-08-004
February 2008 - DRAFT
APPENDIX C. Summary of EPA Team IDS Decisions
Chemical
ID
Blinded
Chemical
Algorithm
Number
INPUT ATTRIBUTE SCORES
Potency
Real Chemicals:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
2
5
5
4
6
3
4
5
4
5
6
4
3
6
8
4
6
7
5
4
3
4
5
5
5
3
7
4
4
4
6
4
8
5
5
7
7
7
4
Severity
7
8
7
3
8
3
9
8
5
6
9
3
3
8
7
5
5
3
6
5
6
3
5
7
3
3
6
3
5
3
5
3
8
8
4
3
3
8
3
Prevalence
1
10
1
7
6
10
10
8
10
1
10
10
10
10
9
2
6
9
9
2
9
6
10
1
10
9
4
10
5
9
1
10
1
9
10
1
1
4
10
Magnitude
1
8
1
6
7
10
10
7
10
6
7
9
10
7
8
6
7
8
8
6
10
6
5
1
9
3
5
9
8
3
10
10
6
7
8
1
1
4
10
Team Consensus Blinded
Decisions
List=4
Mean
1.00
3.67
1.17
1.83
3.50
2.17
3.17
3.67
3.17
1.67
3.67
2.50
2.00
3.83
4.00
1.67
2.83
3.33
3.17
1.50
3.17
1.50
2.67
1.17
3.00
1.50
2.83
2.50
2.33
1.50
2.00
2.50
2.83
3.67
2.83
1.17
1.00
3.00
2.50
Integer
Score
1
4
1
2
4
2
3
4
3
2
4
3
2
4
4
2
3
3
3
2
3
2
3
1
3
2
3
3
2
2
2
3
3
4
3
1
1
3
3
L/NL
NL
L
NL
NL?
L
NL?
L?
L
L?
NL?
L
L?
NL?
L
L
NL?
L?
L?
L?
NL?
L?
NL?
L?
NL
L?
NL?
L?
L?
NL?
NL?
NL?
L?
L?
L
L?
NL
NL
L?
L?
C-l
-------
EPA-OGWDW
CCL3 Chemicals:
Classification of the PCCLto CCL
EPA815-R-08-004
February 2008 - DRAFT
Chemical
ID
Blinded
Chemical
Algorithm
Number
INPUT ATTRIBUTE SCORES
Potency
Real Chemicals:
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
6
5
5
4
3
2
5
4
4
4
2
5
5
6
5
6
3
6
5
8
5
5
7
4
4
7
5
7
7
4
5
7
8
6
3
10
4
5
4
7
4
Severity
3
3
3
4
1
3
8
3
3
8
3
8
8
8
5
9
3
6
8
8
3
7
6
3
6
8
6
8
8
6
5
8
8
8
3
6
7
3
3
3
3
Prevalence
9
1
7
10
3
7
1
9
8
9
10
4
1
1
10
3
10
9
4
3
4
9
1
9
10
10
2
1
7
1
1
3
7
1
10
9
7
2
6
10
10
Magnitude
2
6
6
10
1
8
5
7
8
9
10
6
1
1
10
7
10
6
4
6
8
6
1
7
10
8
7
1
4
1
1
6
6
1
8
8
5
3
7
6
8
Team Consensus Blinded
Decisions
List=4
Mean
1.83
1.50
2.00
2.83
1.00
1.33
2.17
2.17
2.33
3.67
1.83
2.83
1.17
1.17
3.50
3.00
2.00
3.33
2.67
3.33
2.00
3.50
1.33
2.33
3.33
4.00
2.17
1.67
3.33
1.00
1.00
3.33
3.83
1.17
1.83
4.00
2.67
1.17
1.83
2.83
2.33
Integer
Score
2
2
2
3
1
1
2
2
2
4
2
3
1
1
4
3
2
3
3
3
2
4
1
2
3
4
2
2
3
1
1
3
4
1
2
4
3
1
2
3
2
L/NL
NL?
NL?
NL?
L?
NL
NL
NL?
NL?
NL?
L
NL?
L?
NL
NL
L
L?
NL?
L?
L?
L?
NL?
L
NL
NL?
L?
L
NL?
NL?
L?
NL
NL
L?
L
NL
NL?
L
L?
NL
NL?
L?
NL?
C-2
-------
EPA-OGWDW
CCL3 Chemicals:
Classification of the PCCLto CCL
EPA815-R-08-004
February 2008 - DRAFT
Chemical
ID
Blinded
Chemical
Algorithm
Number
INPUT ATTRIBUTE SCORES
Potency
Real Chemicals:
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
3
3
4
3
7
4
6
6
4
5
4
3
4
6
4
6
4
7
3
5
4
Synthetic Chemicals:
63
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
4
3
5
6
2
6
8
2
4
7
4
10
1
9
1
2
7
4
Severity
6
8
8
3
6
6
8
6
6
8
6
3
6
3
4
5
8
9
8
6
8
1
2
3
1
8
9
8
2
1
3
8
8
3
1
1
6
8
5
Prevalence
10
4
10
10
5
3
5
6
6
1
1
5
3
6
10
8
5
5
1
3
8
7
2
1
2
1
1
8
10
8
8
8
8
5
1
1
1
4
10
Magnitude
10
7
6
8
7
5
8
8
7
5
7
7
5
7
7
7
10
8
7
7
7
7
2
2
1
1
4
2
9
1
3
1
1
9
7
1
9
7
9
Team Consensus Blinded
Decisions
List=4
Mean
2.83
2.50
3.33
1.80
3.20
1.80
3.60
3.00
2.40
2.20
1.80
1.20
1.40
2.40
2.40
3.20
3.60
3.80
1.80
2.60
3.00
1.50
1.00
1.00
1.00
1.00
2.33
3.50
2.00
1.00
2.00
2.33
3.50
1.33
1.67
1.00
1.83
3.67
3.17
Integer
Score
3
3
3
2
3
2
4
3
2
2
2
1
1
2
2
3
4
4
2
3
3
2
1
1
1
1
2
4
2
1
2
2
4
1
2
1
2
4
3
L/NL
L?
L?
L?
NL?
L?
NL?
L
L?
NL?
NL?
NL?
NL
NL
NL?
NL?
L?
L
L
NL?
L?
L?
NL?
NL
NL
NL
NL
NL?
L
NL?
NL
NL?
NL?
L
NL
NL?
NL
NL?
L
L?
c-:
-------
EPA-OGWDW
CCL3 Chemicals:
Classification of the PCCLto CCL
EPA815-R-08-004
February 2008 - DRAFT
Chemical
ID
Blinded
Chemical
Algorithm
Number
INPUT ATTRIBUTE SCORES
Potency
Real Chemicals:
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
7
2
6
10
5
9
3
8
10
8
1
10
3
10
5
2
8
1
6
1
6
8
1
9
5
10
5
9
6
2
6
3
6
9
3
8
4
8
3
9
10
Severity
3
6
6
9
1
3
6
7
5
8
5
4
8
8
4
2
5
8
8
2
5
8
8
8
4
4
6
8
6
1
2
7
8
8
1
2
9
8
5
8
2
Prevalence
10
7
7
10
5
3
1
1
1
8
10
8
10
9
3
10
3
5
2
9
6
8
7
7
3
5
3
9
1
8
8
6
8
2
1
2
5
1
6
9
1
Magnitude
10
7
5
7
3
3
3
2
1
8
3
5
1
4
8
7
8
10
6
8
6
6
7
10
3
3
5
9
1
3
2
3
1
3
6
7
7
7
8
4
1
Team Consensus Blinded
Decisions
List=4
Mean
3.67
2.17
2.67
4.00
1.00
2.00
1.00
2.33
2.00
4.00
1.33
3.33
1.83
3.83
2.50
1.17
3.33
2.83
3.00
1.33
2.83
3.83
2.17
4.00
1.17
2.67
2.17
4.00
1.17
1.17
1.50
2.17
2.83
3.17
1.17
2.33
3.33
3.50
2.33
3.67
1.50
Integer
Score
4
2
3
4
1
2
1
2
2
4
1
3
2
4
3
1
3
3
3
1
3
4
2
4
1
3
2
4
1
1
2
2
3
3
1
2
3
4
2
4
2
L/NL
L
NL?
L?
L
NL
NL?
NL
NL?
NL?
L
NL
L?
NL?
L
L?
NL
L?
L?
L?
NL
L?
L
NL?
L
NL
L?
NL?
L
NL
NL
NL?
NL?
L?
L?
NL
NL?
L?
L
NL?
L
NL?
C-4
-------
EPA-OGWDW
CCL3 Chemicals:
Classification of the PCCLto CCL
EPA815-R-08-004
February 2008 - DRAFT
Chemical
ID
Blinded
Chemical
Algorithm
Number
INPUT ATTRIBUTE SCORES
Potency
Real Chemicals:
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
8
4
7
4
9
7
4
7
1
6
2
2
8
5
10
5
8
7
2
8
1
10
4
10
6
1
7
2
8
6
2
8
5
8
4
10
9
4
7
3
9
Severity
4
8
6
4
3
7
7
8
4
3
8
3
8
3
3
8
7
7
4
3
8
8
4
8
4
6
6
3
2
9
8
8
4
1
7
8
7
1
2
8
8
Prevalence
6
8
9
1
4
7
3
1
10
10
9
10
6
3
2
2
2
9
6
10
7
6
4
2
5
5
4
10
7
6
7
8
1
1
4
10
4
8
10
10
7
Magnitude
9
7
8
4
3
7
5
4
2
5
3
5
5
10
6
6
7
4
9
6
8
8
2
9
3
1
2
4
2
6
4
4
6
7
7
3
6
9
6
9
8
Team Consensus Blinded
Decisions
List=4
Mean
3.33
3.67
3.83
1.00
2.33
3.50
2.17
2.67
1.33
2.50
2.17
1.17
3.67
2.50
2.83
2.67
3.33
3.67
2.33
3.50
2.33
3.83
1.00
3.67
2.00
1.00
2.00
1.00
1.50
3.67
2.00
3.83
2.33
1.67
2.83
3.83
3.83
2.00
2.67
3.67
4.00
Integer
Score
3
4
4
1
2
4
2
3
1
3
2
1
4
3
3
3
3
4
2
4
2
4
1
4
2
1
2
1
2
4
2
4
2
2
3
4
4
2
3
4
4
L/NL
L?
L
L
NL
NL?
L
NL?
L?
NL
L?
NL?
NL
L
L?
L?
L?
L?
L
NL?
L
NL?
L
NL
L
NL?
NL
NL?
NL
NL?
L
NL?
L
NL?
NL?
L?
L
L
NL?
L?
L
L
C-5
-------
EPA-OGWDW
CCL3 Chemicals:
Classification of the PCCLto CCL
EPA815-R-08-004
February 2008 - DRAFT
Chemical
ID
Blinded
Chemical
Algorithm
Number
INPUT ATTRIBUTE SCORES
Potency
Real Chemicals:
248
1
Severity
1
Prevalence
10
Magnitude
5
Team Consensus Blinded
Decisions
List=4
Mean
1.17
Integer
Score
1
L/NL
NL
C-6
-------
EPA-OGWDW CCL 3 Chemicals: EPA815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
APPENDIX D. SOFTWARE SOURCES
Artificial Neural Networks - ANN methods packaged in R software libraries "MASS"
and "nnet" are available at no charge from the website http ://www. r-project.org, under
the Free Software Foundation's GNU General Public License.
Univariate Decision Tree - CART - methods packaged in the R software library "rpart"
are available at no charge from the website http ://www.r-project.org, under the Free
Software Foundation's GNU General Public License.
Multivariate Decision Tree - QUEST software is available at no charge from the website
http ://www. stat. wise. eduMoh/quest.html
Linear Modeling - Likelihood function was maximized using MathCAD's built-in
Maximize function (www.mathsoft.com).
Multivariate Adaptive Regression Splines - MARS methods packaged in the R software
library "polspline" are available at no charge from the website http://www.r-project.org,
under the Free Software Foundation's GNU General Public License.
D- 1
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
APPENDIX E. SOLUTIONS
Artificial Neural Network - The software used does not reveal its decision rule. Instead,
it provides classifications for contaminants that have been scored for the four attributes.
When given a complete set of all possible combinations of integer attribute scores, the
software provides classifications. Although not expressed mathematically, this complete
description of the decision rule can be seen in Exhibit 4-4.
Example: Contaminant with scores (3, 4, 5, 6). Exhibit 4-4 shows this as a dark blue
point. Not List.
Simple Linear Model - The maximum likelihood linear model is shown below. Y[i] is
the estimated team-average classification and Pot[i], Sev[i], Prev[i], Mag[i] are the
attribute scores for contaminant i. If Y[i] is less than 1.5, then the classification is Not
List. Similarly, if Y[i] is at least 3.5, then the classification is List.
Y[i] = -1.671 + 0.241 * Pot[i] + 0.217 * Sev[i] + 0.116 * Prev[i] + 0.170 * Mag[i]
Example: Contaminant with scores (3, 4, 5, 6).
Y = -1.671 + 0.241 * 3 + 0.217 * 4 + 0.116 * 5 + 0.170 * 6 = 1.520 -» Not List
Multivariate Tree (QUEST) - The solution involves a number of intermediate nodes and
terminal nodes arranged as shown in Exhibit 4.1.1. When a contaminant encounters an
intermediate node, a weighted sum of attribute scores is compared to a threshold value.
The direction the contaminant moves from the node depends on whether the threshold is
exceeded. Below, vector notation is used below to simplify the description. Letting X[i]
be a column vector of attribute scores, (Pot[i], Sev[i], Prev[i], Mag[i]), then BlT*X[i] is
the vector product of Bl (a column vector of weights) and X[i], which, in turn, is
compared with the threshold. When the contaminant encounters a terminal node (Node
6, 10, 11, 16, 17, 29, 30, or 31), a classification is assigned.
Node 1: If Bl*X[i] < 0.3023, then Node 2, otherwise Node 3.
Node 2: If B2*X[i] < 0.3844, then Node 4, otherwise Node 5.
Node 4: If B4*X[i] < 0.6460, then Node 6, otherwise Node 7.
Node 6: Not List
Node 7: If B7*X[i] < 3.336, then Node 10, otherwise Node 11.
Node 10: Not List
Node 11: Not List?
Node 5: If B5*X[i] < 1.213, then Node 16, otherwise Node 17.
Node 16: Not List?
Node 17: List?
Node 3: If B3*X[i] < 1.181, then Node 28, otherwise Node 29
Node 28: If B28*X[i] < 6.460, then Node 30, otherwise Node 31.
Node 30: List?
Node 31: List
Node 29: List
E- 1
-------
EPA-OGWDW
CCL 3 Chemicals:
Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Exhibit A. 1 - Tree Produced by QUEST (heavy arrows show path of contaminant with
attribute scores 3, 4, 5, 6)
Contaminant Entry:
(3, 4, 5, 6)
Terminal
Node Index
Exhibit A.2 - The column vectors of weights:
Bl
0.01631
0.01315
0.007523
0.01034
B2
0.03008
0.02075
0.01214
0.02043
B3
0.05223
0.06855
0.03516
0.01807
B4
0.06890
0.01756
0.01753
0.05501
B5
0.07779
0.06447
0.03300
0.04850
B7
0.3531
0.1136
0.07560
0.2144
B28
0.2966
0.3174
0.1995
0.1952
Example: Contaminant with scores X = (3, 4, 5, 6)
Node 1: Bl T*X = 0.01631*3 + 0.01315*4 + 0.007523*5 + 0.01034*6 = 0.2012
This is less than 0.3023, so go to Node 2.
Node 2: B2T*X = 0.03008*3 + 0.02075*4 + 0.01214*5 + 0.02043*6 = 0.3565
This is less than 0.3844, so go to Node 4.
Node 4: B4T*X = 0.06890*3 + 0.01756*4 + 0.01753*5 + 0.05501*6 = 0.6947
This exceeds 0.6460, so go to Node 7.
E-2
-------
EPA-OGWDW CCL 3 Chemicals: EPA 815-R-08-004
Classification of the PCCL to CCL February 2008 - DRAFT
Node 7: B7T*X = 0.3531*3 + 0.1136*4 + 0.07560*5 + 0.2144*6 = 3.1781
This is less than 3.336, so go to Node 10.
Node 10: Not List
I1f Y[i] is between 1.5 and 2.5, the classification is NL?; and if Y[i] is between 2.5 and 3.5,
the classification is L?
E-3
-------
EPA-OGWDW
CCL 3 Chemicals
Appendix F. Chemicals Reviewed by
: Classification of the PCCL to CCL
the EPA Evaluation Team: Summary of Results
EPA815-R-08-004
February 2008 - DRAFT
Set 1 Summary
CASR
N
51285
60571
62737
63252
67641
67721
72559
74839
74873
74953
74975
75150
75343
75694
75718
Common Name
2,4-Dinitrophenol
Dieldrin
Dichlorvos
Carbaryl
Acetone
Hexachloroethane
p,p'-DDE
Methyl bromide
Bromomethane
Chloromethane (Methyl
chloride)
Dibromomethane
Halon1011
(bromochloromethane)
Carbon disulfide
1,1-Dichloroethane
CFC-11.
Trichlorofluoromethane
CFC-12.
Dichlorofluoromethane
Model
Decisio
n
NL
L?-L
NL-NL?
NL?
L?
NL
NL-NL?
L?
L?
NL?
NL?
NL? - L?
L?
L?-L
NL?
#
Evaluator
s
18
18
16
16
18
17
16
17
16
14
13
15
12
13
13
%
agreement
100
83
88
69
78
100
88
82
81
71
62
53
67
69
77
Direction - disagree
+/-(+
toward
L)
+/-0
+7
-1
+/-0
-2
+1
+1
+3
+/-0
-1
-2
-2
+1
+1
-4
Value
(L=4;
NL=1)
1.00
3.66
1.63
2.00
2.75
1.06
1.61
3.11
2.88
1.92
1.83
2.21
3.00
3.38
1.71
Categor
y
NL
L
NL?
NL?
L?
NL
NL?
L?
L?
NL?
NL?
NL?
L?
L?
NL?
Overall Confidence
H%
65%
41%
33%
33%
40%
44%
40%
47%
50%
36%
40%
29%
18%
42%
50%
M%
35%
47%
53%
47%
60%
56%
53%
47%
50%
57%
60%
64%
64%
50%
42%
L%
0%
12%
13%
20%
0%
0%
7%
7%
0%
7%
0%
7%
18%
8%
8%
Value
H=3; L=1)
2.647
2.294
2.200
2.133
2.400
2.438
2.333
2.400
2.500
2.286
2.400
2.214
2.000
2.333
2.417
POTENCY Data Element
Element
(L4G)
Reference
Dose (RfD)
Lifetime
Cancer Risk
(10A-4)
Lifetime
Cancer Risk
(10A-4)
Reference
Dose (RfD)
Reference
Dose (RfD)
Reference
Dose (RfD)
Lifetime
Cancer Risk
(10A-4)
Reference
Dose (RfD)
Reference
Dose (RfD)
Reference
Dose (RfD)
Reference
Dose (RfD)
Reference
Dose (RfD)
Slope
Factor
(Oral)
Reference
Dose (RfD)
Reference
Dose (RfD)
Source
IRIS
IRIS
IRIS
OPP
IRIS
IRIS
IRIS
IRIS
EPAHA
RAISH
E
EPAHA
IRIS
OEHHA
IRIS
IRIS
Type
(NCAR
/CAR)
NCAR
CAR
CAR
NCAR
NCAR
NCAR
CAR
NCAR
NCAR
NCAR
NCAR
NCAR
CAR
NCAR
NCAR
PREVALENCE Data Element
Element (L4G)
Percentage of PWSs
(Detects), All Water, Finished
Percentage of PWSs
(Detects), All Water, Finished
Percentage of Samples
(Detects), Surface Water,
Ambient
Percentage of PWSs
(Detects), All Water, Finished
Percentage of Sites
(Detects), All Water, Ambient
Percentage of Sites
(Detects), All Water, Ambient
Percentage of PWSs
(Detects), All Water, Finished
Percentage of PWSs
(Detects), All Water, Finished
Percentage of PWSs
(Detects), All Water, Finished
Percentage of PWSs
(Detects), All Water, Finished
Percentage of PWSs
(Detects), All Water, Finished
Percentage of Sites
(Detects), All Water, Ambient
Percentage of PWSs
(Detects), All Water, Finished
Percentage of PWSs
(Detects), All Water, Finished
Percentage of PWSs
(Detects), All Water, Finished
Source
UCMR
NCODR1
2
NREC
NCODR1
2
NAWQA
NAWQA
UCMR
NCODR1
2
NCODR1
2
NCODR1
2
NCODR1
2
NAWQA
NCODR1
2
NCODR1
2
NCODR1
2
F-1
-------
EPA-OGWDW
CCL 3 Chemicals
Appendix F. Chemicals Reviewed by
: Classification of the PCCL to CCL
the EPA Evaluation Team: Summary of Results
EPA815-R-08-004
February 2008 - DRAFT
Set 1 Summary
CASR
N
79345
80626
86500
87616
87683
88062
91203
94746
95498
95636
Common Name
1,1,2,2-
Tetrachloroethane
Methyl methacrylate
Azinphos-methyl
1 ,2,3-Trichlorobenzene
Hexachlorobutadiene
2,4,6-Trichlorophenol
Naphthalene
MCPA
o-Chlorotoluene
1,2,4-
Trimethyl benzene
Model
Decisio
n
NL?
NL
NL?
NL-NL?
L?
NL
NL?
NL? - L?
NL?
NL?
#
Evaluator
s
14
13
12
13
13
13
13
14
13
13
%
agreement
64
100
100
77
77
92
85
71
77
69
Direction -disagree
+/-(+
toward
L)
+1
+/-0
+1
-3
-2
+1
-1
-1
-4
-1
Value
(L=4;
NL=1)
2.04
1.00
2.15
1.47
2.75
1.08
1.93
2.38
1.71
1.92
Categor
y
NL?
NL
NL?
NL
L?
NL
NL?
NL?
NL?
NL?
Overall Confidence
H%
36%
58%
27%
42%
46%
42%
67%
33%
58%
33%
M%
55%
42%
64%
42%
46%
50%
33%
42%
42%
33%
L%
9%
0%
9%
17%
8%
8%
0%
25%
0%
33%
Value
H=3; L=1)
2.273
2.583
2.182
2.250
2.385
2.333
2.667
2.083
2.583
2.000
POTENCY Data Element
Element
(L4G)
Reference
Dose (RfD)
Reference
Dose (RfD)
Reference
Dose (RfD)
Lowest
Observed
Adverse
Effect Level
(LOAEL)
Reference
Dose (RfD)
Reference
Dose (RfD)
Reference
Dose (RfD)
Reference
Dose (RfD)
Reference
Dose (RfD)
Reference
Dose (RfD)
Source
EPAHA
IRIS
OPP
RTECS
EPAHA
EPAHA
IRIS
OPP
IRIS
RAISH
E
Type
(NCAR
/CAR)
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
PREVALENCE Data Element
Element (L4G)
Percentage of PWSs
(Detects), All Water, Finished
Percentage of Sites
(Detects), All Water, Ambient
Percentage of Sites
(Detects), All Water, Ambient
Percentage of PWSs
(Detects), All Water, Finished
Percentage of PWSs
(Detects), All Water, Finished
Percentage of PWSs
(Detects), All Water, Finished
Percentage of PWSs
(Detects), All Water, Finished
Percentage of Sites
(Detects), All Water, Ambient
Percentage of PWSs
(Detects), All Water, Finished
Percentage of PWSs
(Detects), All Water, Finished
Source
NCODR1
2
NAWQA
NAWQA
NCODR1
2
NCODR1
2
UCMR
NCODR1
2
NAWQA
NCODR1
2
NCODR1
2
F-2
-------
EPA-OGWDW
CCL 3 Chemicals: Classification of the PCCL to CCL
Appendix F. Chemicals Reviewed by the EPA Evaluation Team: Summary of Results
EPA815-R-08-004
February 2008 - DRAFT
CASRN
2212671
2312358
5989275
7439987
7440020
7440097
7440213
7440235
7440246
7440428
7440484
7440564
Set 2 Summary
Common Name
Molinate
Propargite
(D)-Limonene
Molybdenum
Nickel
Potassium
Silicon
Sodium
Strontium
Boron
Cobalt
Germanium
Model
Decision
L?
L?
NL?
L?-L
L?
L?
L
L?
L?
L?
NL? - L?
L?
#
Evaluator
s
19
18
17
18
18
18
18
19
19
18
17
18
%
agreement
84
72
82
78
89
44
61
68
74
61
71
61
Direction - disagree
+/-(+
toward
L)
-1
-5
-4
+/-0
-2
-9
-4
-3
+/-0
+3
-1
-2
Value
(L=4;
NL=1)
3
3
2
3
3
2
3
3
3
3
2
3
Categor
y
L?
L?
NL?
L?
L?
NL?
L?
L?
L?
L?
NL?
L?
Overall Confidence
H%
32%
24%
24%
50%
28%
24%
17%
26%
26%
24%
24%
18%
M%
58%
59%
59%
39%
67%
24%
33%
37%
47%
53%
53%
24%
L%
11%
18%
18%
11%
6%
53%
50%
37%
26%
24%
24%
59%
Value
H=3; L=1)
2.211
2.059
2.059
2.389
2.222
1.706
1.667
1.895
2.000
2.000
2.000
1.588
POTENCY Data Element
Element
(L4G)
Reference
Dose (RfD)
Reference
Dose (RfD)
No
Observed
Effect Level
(NOEL)
UL
Reference
Dose (RfD)
Lowest
Observed
Adverse
Effect Level
(LOAEL)
Lethal Dose
50 (LD50)
Lowest
Observed
Adverse
Effect Level
(LOAEL)
Reference
Dose (RfD)
Reference
Dose (RfD)
MRL-Int
Lowest
Observed
Adverse
Effect Level
(LOAEL)
Source
IRIS
OPP
NTP
IOM
IRIS
NAS
RTECS
RTECS
IRIS
IRIS
ATSDR
RTECS
Type
(NCAR
/CAR)
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
PREVALENCE Data Element
Element (L4G)
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of Sites
(Detects), All Water,
Ambient
Percentage of
Samples (Detects),
Surface Water,
Ambient
Percentage of
Samples (Detects), All
Water, Finished
Percentage of
Samples (Detects), All
Water, Finished
Percentage of
Samples (Detects), All
Water, Finished
Percentage of
Samples (Detects), All
Water, Finished
Percentage of
Samples (Detects), All
Water, Finished
Percentage of
Samples (Detects), All
Water, Finished
Percentage of
Samples (Detects), All
Water, Finished
Percentage of
Samples (Detects), All
Water, Finished
Percentage of
Samples (Detects), All
Water, Finished
Source
UCMR
NAWQA
NREC
NIRS
NIRS
NIRS
NIRS
NIRS
NIRS
NIRS
NIRS
NIRS
F-3
-------
EPA-OGWDW
CCL 3 Chemicals: Classification of the PCCL to CCL
Appendix F. Chemicals Reviewed by the EPA Evaluation Team: Summary of Results
EPA815-R-08-004
February 2008 - DRAFT
CASRN
7440622
7664417
7723140
13071799
13194484
13494809
14797730
16655826
16752775
21087649
21725462
25013165
25057890
27314132
Set 2 Summary
Common Name
Vanadium
Ammonia
White Phosphorus
Terbufos
Ethoprop
Tellurium
Perchlorate
3-Hydroxycarbofuran
Methomyl
Metribuzin
Cyanazine
Butylated hydroxyanisole
Bentazon
Norflurazon
Model
Decision
L?-L
NL?
L
NL
NL?
NL? - L?
NL? - L?
L?
NL?
NL-NL?
NL?
NL?
NL?
NL?
#
Evaluator
s
18
17
19
17
16
16
16
18
16
16
17
15
15
14
%
agreement
78
82
100
82
81
56
50
83
56
69
65
73
53
79
Direction - disagree
+/-(+
toward
L)
-4
-4
-1
+3
+2
+2
+6
+2
-1
+/-0
+/-0
+/-0
+1
+2
Value
(L=4;
NL=1)
3
2
4
1
2
2
3
3
2
2
2
2
2
2
Categor
y
L?
NL?
L
NL
NL?
NL?
L?
L?
NL?
NL?
NL?
NL?
NL?
NL?
Overall Confidence
H%
18%
24%
63%
63%
33%
18%
33%
29%
27%
50%
31%
13%
36%
31%
M%
59%
65%
32%
31%
39%
18%
47%
53%
67%
31%
63%
40%
57%
46%
L%
24%
12%
5%
6%
28%
65%
20%
18%
7%
19%
6%
47%
7%
23%
Value
H=3; L=1)
1.941
2.118
2.579
2.563
2.056
1.529
2.133
2.118
2.200
2.313
2.250
1.667
2.286
2.077
POTENCY Data Element
Element
(L4G)
MRL-Int
Reference
Dose (RfD)
Reference
Dose (RfD)
Reference
Dose (RfD)
Reference
Dose (RfD)
NOAEL
Reference
Dose (RfD)
RfD
Reference
Dose (RfD)
Reference
Dose (RfD)
Reference
Dose (RfD)
Lowest
Observed
Adverse
Effect Level
(LOAEL)
Reference
Dose (RfD)
Reference
Dose (RfD)
Source
ATSDR
RAISHE
IRIS
OPP
OPP
Journal
IRIS
OPP
OPP
OPP
EPAHA
RTECS
IRIS
OPP
Type
(NCAR
/CAR)
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
PREVALENCE Data Element
Element (L4G)
Percentage of
Samples (Detects), All
Water, Finished
Percentage of Sites
(Detects), All Water,
Ambient
Percentage of
Samples (Detects), All
Water, Finished
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of Sites
(Detects), All Water,
Ambient
Percentage of
Samples (Detects), All
Water, Finished
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of Sites
(Detects), All Water,
Ambient
Percentage of
Samples (Detects),
Surface Water,
Ambient
Percentage of Sites
(Detects), All Water,
Ambient
Percentage of Sites
(Detects), All Water,
Ambient
Source
NIRS
NAWQA
NIRS
UCMR
NAWQA
NIRS
UCMR
NCODR12
NCODR12
NCODR12
NAWQA
NREC
NAWQA
NAWQA
F-4
-------
EPA-OGWDW
CCL 3 Chemicals: Classification of the PCCL to CCL
Appendix F. Chemicals Reviewed by the EPA Evaluation Team: Summary of Results
EPA815-R-08-004
February 2008 - DRAFT
CASRN
34014181
34256821
51218452
Set 2 Summary
Common Name
Tebuthiuron
Acetochlor
Metolachlor
Model
Decision
NL-NL?
NL
NL?
#
Evaluator
s
15
16
13
%
agreement
73
69
69
Direction - disagree
+/-(+
toward
L)
-4
+4
-3
Value
(L=4;
NL=1)
1
1
2
Categor
y
NL
NL
NL?
Overall Confidence
H%
53%
67%
38%
M%
33%
20%
54%
L%
13%
13%
8%
Value
H=3; L=1)
2.400
2.533
2.308
POTENCY Data Element
Element
(L4G)
Reference
Dose (RfD)
Reference
Dose (RfD)
Reference
Dose (RfD)
Source
OPP
IRIS
OPP
Type
(NCAR
/CAR)
NCAR
NCAR
NCAR
PREVALENCE Data Element
Element (L4G)
Percentage of Sites
(Detects), All Water,
Ambient
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of PWSs
(Detects), All Water,
Finished
Source
NAWQA
UCMR
NCODR12
F-5
-------
EPA-OGWDW
CCL 3 Chemicals: Classification of the PCCL to CCL
Appendix F. Chemicals Reviewed by the EPA Evaluation Team: Summary of Results
EPA815-R-08-004
February 2008 - DRAFT
CASRN
96184
96333
98066
98953
103651
106434
107028
107131
108054
108861
109999
115968
121142
Set 3 Summary
Common Name
1 ,2,3-Trichloropropane
Methyl acrylate
tert-Butyl benzene
Nitrobenzene
n-Propyl benzene
p-Chlorotoluene
Acrolein
Acrylonitrile
Vinyl acetate
Bromobenzene
Tetrahydrofuran
Trichlorethyl phosphate
2,4-Dinitrotoluene
Model
Decision
NL?
NL
NL?
NL7-L?
NL?
NL?
L?-L
NL?-NL
NL
NL?
L?
NL?-L?
L?-L
#
Evaluators
16
15
16
16
16
15
16
15
15
16
16
14
15
%
agreement
75
93
75
44
94
87
69
73
100
69
75
50
60
Direction -disagree
+/-(+
toward L)
+1
+1
-1
+5
+1
-1
+1
+3
+/-0
+3
-1
-3
+1
Value
(L=4;
NL=1)
2.12
1.07
1.97
2.75
2.03
1.94
3.53
1.78
1.00
2.09
2.93
2.39
3.53
Category
NL?
NL
NL?
L?
NL?
NL?
L
NL?
NL
NL?
L?
NL?
L
Overall Confidence
H%
44%
40%
19%
31%
31%
31%
25%
20%
40%
27%
13%
7%
38%
M%
31%
53%
69%
38%
50%
56%
63%
73%
47%
53%
47%
60%
54%
L%
25%
7%
13%
31%
19%
13%
13%
7%
13%
20%
40%
33%
8%
Value
H=3;
L=1)
2.188
2.333
2.063
2.000
2.125
2.188
2.125
2.133
2.267
2.067
1.733
1.733
2.308
POTENCY Data Element
Element
(L4G)
Reference
Dose (RfD)
Reference
Dose (RfD)
Lowest
Observed
Adverse
Effect Level
(LOAEL)
Reference
Dose (RfD)
Lowest
Observed
Adverse
Effect Level
(LOAEL)
Reference
Dose (RfD)
Reference
Dose (RfD)
Lifetime
Cancer Risk
(10A-4)
Reference
Dose (RfD)
Reference
Dose (RfD)
No
Observed
Effect Level
(NOEL)
Reference
Dose (RfD)
Lifetime
Cancer Risk
(10A-4)
Source
IRIS
RAISHE
RTECS
IRIS
RTECS
EPAHA/
IRIS
RAISHE
EPAHA
RAISHE
RAISHE
Journal
RAISHE
EPAHA
Type
(NCAR/
CAR)
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
CAR
NCAR
NCAR
NCAR
NCAR
CAR
PREVALENCE Data Element
Element (L4G)
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of Sites
(Detects), All Water,
Ambient
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of Sites
(Detects), All Water,
Ambient
Percentage of Sites
(Detects), All Water,
Ambient
Percentage of Sites
(Detects), All Water,
Ambient
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of Sites
(Detects), All Water,
Ambient
Percentage of
Samples (Detects),
Surface Water,
Ambient
Percentage of PWSs
(Detects), All Water,
Finished
Source
NCODR12
NAWQA
NCODR12
UCMR
NCODR12
NCODR12
NAWQA
NAWQA
NAWQA
NCODR12
NAWQA
NREC
UCMR
F-6
-------
EPA-OGWDW
CCL 3 Chemicals: Classification of the PCCL to CCL
Appendix F. Chemicals Reviewed by the EPA Evaluation Team: Summary of Results
EPA815-R-08-004
February 2008 - DRAFT
CASRN
121755
122667
126987
135988
298044
309002
314409
50000
50997
75570
78002
78795
78820
101779
Set 3 Summary
Common Name
Malathion
1 ,2-Diphenylhydrazine
Methacrylonitrile
sec-Butyl benzene
Disulfoton
Aldrin
Bromacil
Formaldehyde
D-Glucose
Tetramethylammonium
chloride
Tetraethyl lead
Isoprene
Isobutyronitrile
Benzenamine, 4,4'-
methylenebis-
Model
Decision
NL
NL-NL?
NL
NL?
NL
L?
NL?
L?-L
NL?-NL
L?
L
L?-L
L?-L
L
#
Evaluators
13
12
14
15
14
15
15
15
14
14
15
15
15
15
%
agreement
77
100
93
93
71
73
73
67
64
57
73
47
33
67
Direction - disagree
+/-(+
toward L)
+3
+/-0
+1
+/-0
+3
+4
-4
-3
-3
-3
-2
-7
-7
-5
Value
(L=4;
NL=1)
1.23
1.50
1.11
2.00
1.35
3.27
1.80
3.27
2.14
2.77
3.88
2.94
3.00
3.40
Category
NL
NL?
NL
NL?
NL
L?
NL?
L?
NL?
L?
L
L?
L?
L?
Overall Confidence
H%
31%
64%
54%
29%
38%
33%
36%
13%
8%
14%
7%
7%
7%
13%
M%
54%
27%
31%
64%
38%
47%
43%
47%
8%
7%
43%
21%
0%
47%
L%
15%
9%
15%
7%
23%
20%
21%
40%
85%
79%
50%
71%
93%
40%
Value
H=3;
L=1)
2.154
2.545
2.385
2.214
2.154
2.133
2.143
1.733
1.231
1.357
1.571
1.357
1.133
1.733
POTENCY Data Element
Element
(L4G)
Reference
Dose (RfD)
Lifetime
Cancer Risk
(10M)
Reference
Dose (RfD)
Lowest
Observed
Adverse
Effect Level
(LOAEL)
Reference
Dose (RfD)
Lifetime
Cancer Risk
(10M)
Reference
Dose (RfD)
Reference
Dose (RfD)
Lethal Dose
50 (LD50)
Lethal Dose
50 (LD50)
Reference
Dose (RfD)
Lowest
Observed
Adverse
Effect Level
(LOAEL)
Lethal Dose
50 (LD50)
Slope
Factor
(Oral)
Source
OPP
IRIS
IRIS
RTECS
OPP,
2002
EPAHA
OPP
IRIS
RTECS
RTECS
IRIS
RTECS
HSDB
OEHHA
Type
(NCAR/
CAR)
NCAR
CAR
NCAR
NCAR
NCAR
CAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
CAR
PREVALENCE Data Element
Element (L4G)
Percentage of Sites
(Detects), All Water,
Ambient
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of Sites
(Detects), All Water,
Ambient
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of PWSs
(Detects), All Water,
Finished
Percentage of Sites
(Detects), All Water,
Ambient
Release, Number of
States
Production Volume
Production Volume
Production Volume
Production Volume
Production Volume
Release, Number of
States
Source
NAWQA
UCMR
NAWQA
NCODR12
UCMR
NCODR12
NAWQA
TRI
CUS/IUR
CUS/IUR
CUS/IUR
CUS/IUR
CUS/IUR
TRI
F-7
-------
EPA-OGWDW
CCL 3 Chemicals: Classification of the PCCL to CCL
Appendix F. Chemicals Reviewed by the EPA Evaluation Team: Summary of Results
EPA815-R-08-004
February 2008 - DRAFT
CASRN
108930
302012
625558
1111780
1335326
3268493
4719044
5216251
6610293
13463406
23422539
71751412
91465086
Set 3 Summary
Common Name
Cyclohexanol
Hydrazine
Isopropyl formate
Ammonium carbamate
Lead acetate
Methional
Hexahydro-1 ,3,5-tris(2-
hydroxyethyl)-s-triazine
4-Chlorobenzotrichloride
Methylthiosem icarbazide
Iron pentacarbonyl
Methanimidamide, N,N-
dimethyl-N'-[3-
[[(methylamino)carbonyl]o
xy]phenyl]-,
monohydrochloride
Avermectin B1
Cyclopropanecarboxylic
acid, 3-2-chloro-3,3,3-
trifluoro-1 -propenyl)-2,2-
dimethyl- cyano(3-
phenoxyphenyl)methyl
ester,
1. alpha. (S*),3.alpha.(Z)-
(.+ -.)-
Model
Decision
L?-L
L
L
L?
L
L?
L
L?-L
L?
L?-L
L?-L
L?-L
L?-L
#
Evaluators
14
15
13
13
14
13
13
14
14
13
14
13
14
%
agreement
64
87
54
77
50
69
38
86
71
62
57
69
71
Direction - disagree
+/-(+
toward L)
-6
-1
-5
-2
-6
-2
-6
-2
-2
-6
-3
-1
-4
Value
(L=4;
NL=1)
2.83
3.79
3.46
2.75
3.35
2.86
3.41
3.21
2.75
2.81
3.27
3.39
3.11
Category
L?
L
L?
L?
L?
L?
L?
L?
L?
L?
L?
L?
L?
Overall Confidence
H%
7%
13%
0%
14%
8%
0%
7%
14%
0%
0%
17%
14%
14%
M%
21%
53%
7%
7%
17%
31%
7%
36%
15%
31%
42%
29%
36%
L%
71%
33%
93%
79%
75%
69%
86%
50%
85%
69%
42%
57%
50%
Value
H=3;
L=1)
1.357
1.800
1.071
1.357
1.333
1.308
1.214
1.643
1.154
1.308
1.750
1.571
1.643
POTENCY Data Element
Element
(L4G)
Lethal Dose
50 (LD50)
Lifetime
Cancer Risk
(10A-4)
Lethal Dose
50 (LD50)
Lethal Dose
50 (LD50)
Slope
Factor
(Oral)
Lethal Dose
50 (LD50)
Lethal Dose
50 (LD50)
NOAEL
Lethal Dose
50 (LD50)
Lethal Dose
50 (LD50)
Reference
Dose (RfD)
ADI
Reference
Dose (RfD)
Source
RTECS
IRIS
RTECS
RTECS
OEHHA
RTECS
RTECS
OPPT
RTECS
HSDB
OPP
JMPR
1997
IRIS
Type
(NCAR/
CAR)
NCAR
CAR
NCAR
NCAR
CAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
NCAR
PREVALENCE Data Element
Element (L4G)
Release, Number of
States
Release, Number of
States
Production Volume
Production Volume
Production Volume
Production Volume
Production Volume
Production Volume
Production Volume
Release, Number of
States
Release, Number of
States
Release, Number of
States
Release, Number of
States
Source
TRI
TRI
CUS/IUR
CUS/IUR
CUS/IUR
CUS/IUR
CUS/IUR
CUS/IUR
CUS/IUR
TRI
NCFAP
NCFAP
NCFAP
F-8
-------
EPA-OGWDW
CCL 3 Chemicals: Classification of the PCCL to CCL
Appendix F. Chemicals Reviewed by the EPA Evaluation Team: Summary of Results
EPA815-R-08-004
February 2008 - DRAFT
CASRN
51796
55630
60355
62533
67561
71363
75218
75569
76879
80159
106990
107211
109864
121448
Set 4 Summary
Common Name
Cobalt compounds
Urethane
Nitroglycerin
Acetamide
Aniline
Methanol
1-Butanol
Ethylene oxide
Propylene oxide
Triphenyltin hydroxide
Cumene hydroperoxide
1 ,3-Butadiene
Ethylene glycol
2-Methoxyethanol
Triethylamine
Model
Decision
L
L
L?-L
L
L?-L
L?-L
L?-L
L
L
L
L
L
L
L
L
#
Evaluate
rs
8
8
9
9
10
10
11
9
9
10
10
11
10
9
7
%
agreement
75
63
78
67
70
60
55
78
89
80
60
73
80
78
43
Direction - disagree
+/-(+
toward L)
-2
-2
+/-0
-2
+2
+1
-1
-2
-1
-2
-3
-2
-2
-3
-4
Value
(L=4;
NL=1)
3.81
3.79
3.50
3.56
3.61
3.45
3.33
3.78
3.78
3.90
3.61
3.80
3.70
3.65
3.36
Category
L
L
L
L
L
L?
L?
L
L
L
L
L
L
L
L?
Overall Confidence
H%
22%
22%
9%
20%
20%
17%
17%
36%
36%
22%
18%
27%
27%
30%
0%
M%
22%
0%
18%
20%
30%
50%
33%
18%
18%
44%
9%
9%
36%
10%
29%
L%
56%
78%
73%
60%
50%
33%
50%
45%
45%
33%
73%
64%
36%
60%
71%
Value
H=3;
L=1)
1.667
1.444
1.364
1.600
1.700
1.833
1.667
1.909
1.909
1.889
1.455
1.636
1.909
1.700
1.286
POTENCY Data Element
Element (L4G)
Lowest Observed
Adverse Effect
Level (LOAEL)
No Observed
Effect Level
(NOEL)
Lowest Observed
Adverse Effect
Level (LOAEL)
Slope Factor
(Oral)
Reference Dose
(RfD)
Reference Dose
(RfD)
Reference Dose
(RfD)
Slope Factor
(Oral)
Slope Factor
(Oral)
Slope Factor
(Oral)
Lowest Observed
Adverse Effect
Level (LOAEL)
Slope Factor
(Oral)
Reference Dose
(RfD)
Reference Dose
(RfD)
Lowest Observed
Adverse Effect
Level (LOAEL)
Source
Journal
Journal
RTECS
OEHHA
RAISHE
IRIS
IRIS
OEHHA
OPP
OPP
RTECS
OEHHA
IRIS
RAISHE
RTECS
Type
(NCAR /
CAR)
NCAR
NCAR
NCAR
CAR
NCAR
NCAR
NCAR
CAR
CAR
CAR
NCAR
CAR
NCAR
NCAR
NCAR
PREVALENCE Data
Element
Element
(L4G)
Release,
Number of
States
Release,
Number of
States
Release,
Number of
States
Release,
Number of
States
Release,
Number of
States
Release,
Number of
States
Release,
Number of
States
Release,
Number of
States
Release,
Number of
States
Release,
Number of
States
Release,
Number of
States
Release,
Number of
States
Release,
Number of
States
Release,
Number of
States
Release,
Number of
States
Source
TRI
TRI
TRI
TRI
TRI
TRI
TRI
TRI
TRI
NCFAP
TRI
TRI
TRI
TRI
TRI
F-9
-------
EPA-OGWDW
CCL 3 Chemicals: Classification of the PCCL to CCL
Appendix F. Chemicals Reviewed by the EPA Evaluation Team: Summary of Results
EPA815-R-08-004
February 2008 - DRAFT
CASRN
123911
133062
137304
319846
330541
330552
333415
541731
542756
Set 4 Summary
Common Name
1 ,4-Dioxane
Captan
Ziram
. alpha. -
Hexachlorocyclohexane
Diuron
Linuron
Diazinon
m-Dichlorobenzene
Telone
Model
Decision
L
L
L
L?
NL?
NL
NL
NL?
L?
#
Evaluate
rs
9
10
8
12
13
12
11
13
13
%
agreement
100
70
88
67
77
92
91
77
62
Direction - disagree
+/-(+
toward L)
+/-0
-2
-1
+1
+3
+/-0
+1
+1
+3
Value
(L=4;
NL=1)
4.00
3.72
3.75
3.00
2.19
1.00
1.09
2.00
3.23
Category
L
L
L
L?
NL?
NL
NL
NL?
L?
Overall Confidence
H%
30%
33%
13%
18%
18%
50%
45%
45%
25%
M%
30%
22%
25%
64%
64%
40%
36%
45%
50%
L%
40%
44%
63%
18%
18%
10%
18%
9%
25%
Value
H=3;
L=1)
1.900
1.889
1.500
2.000
2.000
2.400
2.273
2.364
2.000
POTENCY Data Element
Element (L4G)
Lifetime Cancer
Risk(10A-4)
Slope Factor
(Oral)
Slope Factor
(Oral)
Lifetime Cancer
Risk(10A-4)
Reference Dose
(RfD)
Reference Dose
(RfD)
Reference Dose
(RfD)
Reference Dose
(RfD)
Slope Factor
(Oral)
Source
EPAHA
OPP
OPP
IRIS
OPP
OPP
OPP
EPAHA
OPP
Type
(NCAR /
CAR)
CAR
CAR
CAR
CAR
NCAR
NCAR
NCAR
NCAR
CAR
PREVALENCE Data
Element
Element
(L4G)
Release,
Number of
States
Release,
Number of
States
Release,
Number of
States
Percentage
of Sites
(Detects), All
Water,
Ambient
Percentage
of PWSs
(Detects), All
Water,
Finished
Percentage
of PWSs
(Detects), All
Water,
Finished
Percentage
of PWSs
(Detects), All
Water,
Finished
Percentage
of PWSs
(Detects), All
Water,
Finished
Percentage
of PWSs
(Detects), All
Water,
Finished
Source
TRI
NCFAP
NCFAP
NAWQA
UCMR
UCMR
UCMR
NCODR12
NCODR12
F-10
-------
EPA-OGWDW
CCL 3 Chemicals: Classification of the PCCL to CCL
Appendix F. Chemicals Reviewed by the EPA Evaluation Team: Summary of Results
EPA815-R-08-004
February 2008 - DRAFT
CASRN
630206
759944
944229
1313275
1582098
1610180
1634044
1861321
Set 4 Summary
Common Name
1,1,1 ,2-Tetrachloroethane
S-Ethyl
dipropylthiocarbamate
Fonofos
Molybdenum oxide (MoO3)
Trifluralin
Prometon
Methyl tert-butyl ether
Chlorthal-dimethyl
(Dacthal)
Model
Decision
L?
NL
NL
L
NL-NL?
NL
L?
NL?
#
Evaluate
rs
13
12
12
11
11
12
12
12
%
agreement
77
75
83
45
82
100
58
67
Direction - disagree
+/-(+
toward L)
-1
+3
+/-0
-3
+2
+/-0
+5
+4
Value
(L=4;
NL=1)
2.88
1.38
1.00
3.38
1.59
1.00
3.42
2.25
Category
L?
NL
NL
L?
NL?
NL
L?
NL?
Overall Confidence
H%
27%
55%
60%
0%
56%
40%
10%
33%
M%
64%
45%
40%
25%
44%
40%
70%
56%
L%
9%
0%
0%
75%
0%
20%
20%
11%
Value
H=3;
L=1)
2.182
2.545
2.600
1.250
2.556
2.200
1.900
2.222
POTENCY Data Element
Element (L4G)
Lifetime Cancer
Risk(10A-4)
Reference Dose
(RfD)
Reference Dose
(RfD)
RfD (UL)
Reference Dose
(RfD)
Reference Dose
(RfD)
Slope Factor
(Oral)
Reference Dose
(RfD)
Source
EPAHA
IRIS
IRIS
DRI
OPP
IRIS
OEHHA
OPP
Type
(NCAR /
CAR)
CAR
NCAR
NCAR
NCAR
NCAR
NCAR
CAR
NCAR
PREVALENCE Data
Element
Element
(L4G)
Percentage
of PWSs
(Detects), All
Water,
Finished
Percentage
of PWSs
(Detects), All
Water,
Finished
Percentage
of PWSs
(Detects), All
Water,
Finished
Release,
Number of
States
Percentage
of Sites
(Detects), All
Water,
Ambient
Percentage
of PWSs
(Detects), All
Water,
Finished
Percentage
of PWSs
(Detects), All
Water,
Finished
Percentage
of Sites
(Detects), All
Water,
Ambient
Source
NCODR12
UCMR
UCMR
TRI
NAWQA
UCMR
UCMR
NAWQA
F-11
-------
EPA-OGWDW
CCL 3 Chemicals: Classification of the PCCL to CCL
Appendix F. Chemicals Reviewed by the EPA Evaluation Team: Summary of Results
EPA815-R-08-004
February 2008 - DRAFT
CASRN
1897456
2164172
26471625
Set 4 Summary
Common Name
Chlorothalonil
Fluometuron
Toluene diisocyanate
Model
Decision
NL?
NL
L
#
Evaluate
rs
12
11
10
%
agreement
75
91
80
Direction - disagree
+/-(+
toward L)
+3
+/-0
-1
Value
(L=4;
NL=1)
2.17
1.00
3.89
Category
NL?
NL
L
Overall Confidence
H%
20%
20%
25%
M%
60%
70%
25%
L%
20%
10%
50%
Value
H=3;
L=1)
2.000
2.100
1.750
POTENCY Data Element
Element (L4G)
Reference Dose
(RfD)
Reference Dose
(RfD)
Slope Factor
(Oral)
Source
OPP
IRIS
OEHHA
Type
(NCAR /
CAR)
NCAR
NCAR
CAR
PREVALENCE Data
Element
Element
(L4G)
Percentage
of Sites
(Detects), All
Water,
Ambient
Percentage
of Sites
(Detects), All
Water,
Ambient
Release,
Number of
States
Source
NAWQA
NAWQA
TRI
F-12
-------
EPA-OGWDW
CCL 3 Chemicals: Classification of the PCCL to CCL
EPA815-R-08-004
February 2008 - DRAFT
Appendix G. PCCL Contaminants with Incomplete Data for Scoring or that had Parent Compounds Scored
CASRN
930552
10595956
683181
753731
818086
5160021
7447418
7782992
7783064
7783188
12108133
14808607
75003
75025
75887
102716
106876
115117
116143
127060
7440291
10028156
57018527
1007289
1313275
6190654
7681529
79277671
76578126
56070156
Substance Name
Pyrrolidine, 1-nitroso-
Ethanamine, N-methyl-N-nitroso-
Stannane, dibutyldichloro-
Stannane, dichlorodimethyl-
Stannane, dibutyloxo-
Benzenesulfonic acid, 5-chloro-2-[(2-hydroxy-1-naphthalenyl)azo]-4-methyl-,
barium salt (2:1)
Lithium chloride (LiCI)
Sulfurous acid
Hydrogen sulfide (H2S)
Thiosulfuric acid (H2S2O3), diammonium salt
Manganese, tricarbonyl[(1,2,3,4,5-.eta.)-1-methyl-2,4-cyclopentadien-1-yl]-
Quartz (SiO2)
Ethane, chloro-
Ethene, fluoro-
Ethane, 2-chloro-1 ,1,1 -trifluoro-
Ethanol, 2,2',2"-nitrilotris-
7-Oxabicyclo[4.1 .0]heptane, 3-oxiranyl-
1-Propene, 2-methyl-
Ethene, tetrafluoro-
2-Propanone, oxime
Thorium
Ozone
2-Propanol, 1-(1,1-dimethylethoxy)-
1 ,3,5-Triazine-2,4-diamine, 6-chloro-N-ethyl-
Molybdenum oxide (MoO3)
1 ,3,5-Triazine-2,4-diamine, 6-chloro-N-(1-methylethyl)-
Hypochlorous acid, sodium salt
2-Thiophenecarboxylic acid, 3-[[[[(4-methoxy-6-methyl-1 ,3,5-triazin-2-
yl)amino]carbonyl]amino]sulfonyl]-
Quizalofop
Terbufos-O-analogue sulfone
Diazinon oxygen analog
DCPA mono/di-acid degradate
Common Name
N-nitrosopyrrolidine (NPYR)
N-Nitrosomethylethylamine (NMEA)
Dibutyltin dichloride
Dimethyltin dichloride
Dibutyltin oxide
C.I. Pigment Red 53, barium salt (2:1)
Lithium chloride
Sulfurous acid
Hydrogen sulfide
Ammonium thiosulfate
Methylcyclopentadienyl manganese tricarbonyl
Quartz (SiO2)
Chloroethane
Vinyl fluoride
HCFC-133a
Triethanolamine
1,2-Epoxy-4-(epoxyethyl)cyclohexane
Isobutene
Tetrafluoroethene
2-Propanone oxime
Thorium-232
Ozone
Propylene glycol mono-t-butyl ether
Desisopropylatrazine
Molybdenum trioxide
Desethylatrazine
Sodium hypochlorite
Thifensulfuron
Quizalofop
Terbufos-O-analogue sulfone
Diazinon oxygen analog
Dacthal mono/di-acid degradate
G-1
------- |