EPA 600-6-87-007a
PB88-127105
INVESTIGATION OF CANCER RISK ASSESSMENT METHODS:
SUMMARY
Clement Associates, Incorporated
Ruston, LA
Sep 87
-------
TECHNICAL REPORT DATA PB88-127105
tricase read Imuiiclium on the ret trie before completing/
I REPORT NO
EPA/600/6-87/007a
TITLE AND SUBTITLE
Investigation of Cancer Risk Assessment Methods;
Summary
3. RECIPIENT'S ACCESSION N
5 REPORT DATS
September 1987
6. PERFORMING ORGANIZATION CODE
7 AOTMOR1S> Bruce C. Allen, Annette M. Shipp, Kenny S.
Crump, Bryan Kilian, Mary Lee Hogg, Joe Tudor,
Barbara Keller
8. PERFORMING ORGANIZATION REPORT NO
9 PERFORMING ORGANIZATION NAME AND ADDRESS
10. PROGRAM ELEMENT NO.
Clement Associates, Inc.
1201 Gaines Street
Ruston, LA 71270
11. CONTRACT/GRANT NO.
68-01-6807
12 SPONSORING AGENCY NAME AND ADDRESS
Office of Health and Environmental Assessment
Carcinogen Assessment Group (RD-639)
U.S. Environmental Protection Agency
Washington, DC 20460
13. TYPE Of REPORT AN.1 PERIOD COVERED
14. SPONSORING AGENCY CODE
EPA/600/?!
IS. SUPPLEMENTARY NOTES
EPA Project Office*: Chao Chen, Carcinogen Assessment Group
Office of Health and Environmental Assessment, Washington, DC (382-5719)
is.ABSTRACT jhe major focus of this study is upon making quantitetive comparisons of
carcinogenic potency in animals and humans for 23 chemicals for which suitable
animal and human data exists. These comparisons are based upon estimates of risk
related doses (RRDs) obtained from both animal and human data. An RRD represents
the average daily dose per body weigh", of a chemical that would result in an extra
cancer risk of 25%. Animal data on these and 21 other chemicals of interest to the
EPA and the 000 are coded into an animal data base that permits evaluation by
computer of many risk assessment approaches.
This report is the result of a two-year study to examine the assumptions,
other than those involving low dose extrapolation, used in Quantitative cancer risk
assessment. The study was funded by the Department of Defer.se [through an inter-.
agency transfer of funds to the Environmental Protection Agency (EPA)J, the EPA,
the Electric Power Research Institute and, in its latter stages, by the Risk Science
Institute.
KEY WORDS ..NO DOCUMENT ANALYSIS
DESCRIPTORS
b.lDENTlFIEHS/OPEF-' ENDED TERMS C. COSAT' Field/Group
REPRODUCED BV
U.S. DEPORTMENT OF COMMERCE
MATIONAL TECHNICAL
0- f-OBMATION SERVICE
Sf HINGFIELD. VA. 22161
13. DISTRIBUTION STA"6H6NT
Distribute to public
19. SECURITY CLASS
-------
DISCLAIMER
This document has been reviewed in accordance with the U.S. Environmental
Protection Agency's peer and administrative review policies and approved for
publication. Mention of trade names or commercial products does not constitute
endorsement or recommendation for use. The information in this document has
been funded by the U.S. Environmental Protection Agency, the Department of
Defense (through Interagency Agreement Number RW97075101), the Electric
Power Research Institute, and the Risk Science Institute.
-------
CONTENTS
Pogs
B-ief Summary of Major Conclusions 1
Introduction 3
Data Bos* 5
Analysis of Epidemiologicol Onto 9
Analysis of Animal Data 13
Investigation of Component -Specific Uncertainty 18
Methods for Comparison of Animal and Human Results 19
Results of Animal and Human Comparisons 22
Correlation Analyses 22
Prediction Analyses 26
Discussion 32
General Considerations 32
Directions for Future Research 33
m
-------
ILLUSTRATIONS
1 B«ftt Estimates and Upper and ! ower Bounds 53
for RRDs From Each Human Study
2 Correlation of Animal, and Human RRDs - Analysis 0 95
3 Correlation of Animal and Huiitan RRDs - Analysis 3b 56
-------
TABLES
Poqe
1 Approaches to RisK Assessment Components 40
2 Summary of Animal Data by Chemical 42
3 Chemicals for Which Minimal Human and Animal Data 44
Exist for Quantifying Carcinogenic Potency
4 Descriptions of Initial Analyses 45
5 Descriptions of Supplemental Analyses 47
6 Ranks Based on Length of Experiment and 48
Number of Treated Animals
7 Component-Specific Uncertainty: Modes and Dispersion 49
Factors for Ratios of RROs, by Supplemental Analysis
8 Correlation Coefficients and Associated p-values, 5C
by Analysis Method
9 Conversion Factors Corresponding to Various Dose Units, 51
by Method of Analysis
10 Comparison of Results for Selected Analyses 52
-------
BRIEF SUMMARY OF MAJOR CONCLUSIONS
The major focus of this study is upon making quantitative comparisons of
carcinogenic potency in animals end humans for 23 chemicals for which
suitable animals and human data exists. These comparisons are based upon
estimates of "RRDs" obtained from both animal and human data. An RRD
represents th-j average daily dose per body weight of a chemical that
would result in an extra cancer risk of 25%. Animal data on these and
21 other chemicals of interest to the EPA and the DCD are coded into an
animal data base that permits evaluation by computer of many risk
assessment approaches.
The major findings of this study are as follows:
1. Animal and human RRDs are strongly correlated. The knowledge that
this correlation exists between animal and human carcinogenicity
data should strengthen the scientific basis for cancer risk assess-
ment and cause increased confidence to be placed in estimates of
human cancer risk made from animal data.
2. In the majority of cases considered, analysis methods for bioassay
data that utilize lower statistical confidence limits as predictors
yield better predictions of human results than do the same methods
using maximum likelihood estimates.
3. Analysis methods for animal data that utilize median lower bound
RRDs determined from the ensemble of data for a chemical generally
yield better predictions of human results then analyses that utilize
minimum RRDs calculated from all the studies available.
-------
4. Use of the "mg intoke/kg body weight/doy" (body weight) method for
animl-to-human extrapolation generally cause* RRDs estimated from
animal and human data to correspond more closely than th* othsr
methods evaluated, including the "mg intoko/m2 surface area/day"
(surface area) method.
5. The risk assessment approach for animal data that was intended to
mimic that used by the EPA underestimates the RRDs (equivalent to
overestimating human risk) obtained from the human data in this
study by about an order of magnitude, on overage. However, it
should be understood that the risk assessment approaches implemented
in this study are computer automated and do not always utilize the
same data or provide the same result as the EPA approach.
6. Reasonable risk analysis methods can be defined for the chemicals in
this study that reduce the residual loss (roughly the average
multiplicative factor by which the RRD predictors obtained from the
animal data are inconsistent with the ranges of human RRDs consis-
tent with the human data) to 1.7. This is not the srme as saying
that the predictors are accurate to within a factor of 1.7, because
the estimated ranges of human RROs that are consistent with the
human data cover an order of magnitude or more for most chemicals.
7. It has been possible to identify a set of analysis methods using the
median lower bound estimates that are most appropriate for extrapo-
lating risk from animals to humans, given the current state of know-
ledge and data analysis. It is possible to use the information and
results presented in this investigation to calculate ranges of risk
estimates that are consistent with the data and also incorporate
many uncertainties associated with the extrapolation procedure.
-------
8. The many components of risk assessment are interrelated and evalua-
tion of risk assessment methods should focus on the complete risk
assessment process rather than on individual components.
9. The data base and methods used in this study can provide a useful
basis for evaluating various risk assessment methods.
This study only compared human and animal results for a relatively high
risk level. It did not examine the uncertainty inherent in the low dose
extrapolation process.
INTRODUCTION
This report is the result of a two year study to examine the assump-
tions, other than those involving low dose extrapolation, used in
quantitative cancer risk assessment. The study was funded by the
Department of Defense [through an interagency transfer of funds to the
Environmental Protection Agency (EPA)], the EPA, the Electric Power
Research Institute and, in its latter stages, by the Risk Science
Institute. The objectives of the study are as follows:
1. To identify and express quantitatively uncertainties that are
involved in the process of risk estimation, excluding the
uncertainties in the low dose extrapolation model;
2. To examine the impact of the different assumptions that ars
made in risk estimation;
3. To compare results calculated from human and animal data,
including the identification of the assumptions that produce
the best correlation of risk estimates between humans and
animals;
*». To develop guidelines for presenting a range of risk estimates
based on different but scientifically acceptable assumptions or
-------
assumption* that have considerable backing in the scientific
community.
These objectives are pursued using empirical methods in which carcino-
genicity data fcr H chemicals are analyzed systematically in a variety
of ways. Particular attention is placed on those 23 chemicals for which
there exist data from both animal and human studies suitable for making
quantitative comparisons.
Table 1 contains a list of components of a quantitative risk assessment
based upon animal data. Each component requires a decision on the part
of the risk assessor for which there is no unique "correct" choice.
Also listed in Table 1 are various possible approaches to each compo-
nent. The choices that a risk assessor makes for these components
effect the resulting estimates of risk. The choices for these compo-
nents therefore are related to the uncertainty in assessment of risk
from animal data.
Objective 2 is pursued by making different risk estimates for the ^4
chemicals in the study by systematically varying the approaches to the
components listed in Table 1. Examination of the distributions of the
changes in tne estimates associated with different approaches to the
various components permits the examination of the impact of the various
approaches (assumptions). These distributions also relate to the
uncertainties in the process of risk estimation, so this work also
applies to Objective 1.
A major part of the study involves making comparisons between risk
estimates derived from animal data and those derived from human data for
those 23 chemicals for which suitable data are found to exist for both
animals and humans. This work addresses the question of whether
correlations exist between animal and human data and therefore is of
-------
fundamental importance to the scientific validity of quantitative risk
assessment. The practice of making quantitative estimates of human risk
from animal data is based upon the hypothesis (heretofore essentially
untested) that such correlations do in fact exist. If quantitative
correlations can be shown to exist, then these correlations can provide
a stronger scientific basis for risk assessment. Further, evaluation of
the correlations and determination of those approaches to the components
listed in Table 1 that produce the best correlations can suggest better
risk assessment methods and assist in evaluating and presenting the
uncertainty in risk estimates derived using those methods, in accordance
with Objectives 3 and 4.
DATA BASE
At the beginning of the project EPA provided a list of <»0 chemicals for
inclusion in the project that are of interest to the agency. This list
was supplemented by adding additional chemicals for which suitable quan-
titative data are available from both animal and human studies and
deleting a few chemicals from the original list for which suitable
animal bioassay data could not be located, which brought the total
number of chemicals studied to 4
-------
official organizations (IARC monographs, EPA heulth assessments and
similar documents, and NCI and NTP technical reports) and a corcino-
genicity data base complied by Gold «t al. (1). The relevant articles
were obtained and summary information extracted from them was coded into
a computerized data base called the Data Matrix.
The Data Matrix includes information on species, sex, route of exposure,
length of exposure, length of observation, whether a positive carcino-
genic response was observed and whether a data set is suitable for
quantitative risk estimation. Data sets on animal studies that satisfy
this latter condition are coded into a more detailed data base colled
the Animal Data Base. A list of the chemicals included in the study,
the number of carcinogenicity data sets summarized in the Data Matrix,
and the number of those in various categories that are coded into the
Animal Data Base is given in Table 2. As can be seen from this table, a
total of 1233 data sets (a data set is generally composed of all the
dose response data from a given sex and species of animals exposed via a
reasonably common protocol in a study) from 736 studies (a study
generally consists of all of the data in a single primary reference) are
summarized in the Data Matrix.
Animol Data Base: All of the bioassays that are considered to be at
least minimally acceptable for quantitative risk estimation are coded
into the computerized Animol Data Base. The criteria that a data set
needs to satisfy for inclusion are as follows:
* the test species is a non-human mammalian species;
• the protocol includes matched controls', preferably vehicle (or
sham inhalation) treated animals;
• dosing is consistent within a dose group, with dosages and dosing
pattern clearly stated;
-------
• a single route of exposure is employed (early in the project it
was decided not to continue to code experiments that exposed the
animals by skin painting or subcutaneous injection; therefore the
Data Base is not complete with respect to these routes of
exposure);
• the test compound is administered alone or in an acceptable
vehicle, without pretreatment or concurrent treatment of any
kind;
• tumor incidence is reported as number of tumor-bearing animals as
opposed to number of tumors.
Table 2 provides a summary of the date included in the Animal Data Base
for each of the
-------
• the combination of all significantly increased malignant tumors;
• all tumors1;
• all maligno.it tumors1;
• the tumor considered to be the response of interest in humans (if
known).
Early in the study individual animal pathologies were coded whenever
possible, which would make possible time-to-tumor analyses. However,
this work was discontinued due to limited resources after such data had
been coded for about about five chemicals.
Selection of Chemicals for Animal-Human Comparisons: For a chemical to
be included in the analyses comparing results in animals and humans,
data had to be available from both human and animal studies that would
support the quantitative comparisons conducted and for which reasonably
strong positive evidence ef carcinogenicity exists in either the animal
or the human data. A list of the chemicals satisfying these require-
ments and which are therefore included in the comparative analyses is
presented in Table 3 Thirteen industrial chemicals are included in
this list, seven drugs, a food contaminant (aflatoxin), a food additive
(saccharin), and tobacco smoke.
It is neither necessary nor sufficient that a chemical be unequivocally
carcinogenic in humans in order to be included. Thus, a chemicnl such
as saccharin, which has been associated with cancer only in laboratory
rodents, is included while bis(chloromethyl) ether is not included, even
though sufficient evidence apparently exists to establish that
bis(chloromethyl) ether is carcinogenic in humans (2). The reasons such
1Interstitial cell tumors of the testes in male F3M» rats, mammary gland
benign tumors in female Sprague-Dawley rats, malignant lymphomas in AKR
and AKR/J mice, and mammary tumors in MTV+ mice are not included in
these groups. These tumors have o very high background rate of occur-
rence in the indicated species, which would tend to obscure dose-
related effects at other sites.
-------
chemicals are not included generally relate to limitations regarding the
data on human exposure*. Of the 23 chemicals or chemical groups that
IARC considered in 1982 to hove "sufficient" evidence of l-uman carcino-
genicity, 11 are included in this study. Twelve other eht>micals are
included; three are considered to provide "limited* evidence, eight to
provide "inadequate* evidence in support of hurran carcinogenic effects,
and cigarette smoke has not been formally evaluated by IARC.
It was considered important that the study not be limited to chemicals
whose carcinogenicity in humans has been firmly established. One of the
ultimate goals of the study is to compare the predictions of carcino-
genic potency of chemicals derived from animal data with the correspond-
ing potency in humans. If such comparisons are restricted to confirmed
human carcinogens, the ability of the animal data to predict human
results might be overestimated. The same would be true if the study is
restricted to confirmed animal carcinogens. Although a similar study by
the National Academy of Sciences was restricted to conf in.ied human
carcinogens, the authors recognized the potential for bias in this
approach (3).
A thorough search was conducted for useful epidemiological data on the
chemicals selected. Individual researchers were queried regarding
unpublished data that would be helpful in our analyses, possible updates
of their work and, particularly, additional information on exposure.
ANALYSIS OF EPIDEMIOLOGICAL DATA
Calculation of Risk Related Doses (RRDs): The epidemiological data on
the 23 chemicals in Table 3 vary greatly in format and quality. Three
distinct types of studies are represented: prospective cohort studies
(including clinical trials), case-control studies, and (in the case of
-------
aflotoxin) a cross-sectional comparison of concur rotas and levels of
exposure in different populations. Even within one of these categories,
the individual studies differ considerably with respect to such factors
as duration of exposure, latency, and methods for reporting results.
Because of the wicie variations in data from the epidemiological studies,
a systematic, standardized method of recording the human data (like that
developed for the bioascay data base) is not considered feasible.
Instead, the epidemiologic data for each chemical is considered as a
whole and risk estimates are developed using general guidelines whose
purpose is to insure that, to the extent possible, the methodology 1)
can be employed with a minimal amount of data, 2) makes best use of the
data, and 3) ensures that risk estimates made from data of differing
types and quality are comparable.
The majority of epidemiological studies considered are prospective
studies. The minimum amount of information required for an analysis of
a prospective study consists of a single group with known cumulative
dose (expressed in ppm-years, for example) and observed and expected
numbers of cancers. Additional information on observed and expected
responses categorized by exposure group is accommodated whenever
available and may provide better estimates of carcinogenic potency.
Using the linear dose response model for relative risk of RR • 1 + /Jd,
where d is cumulative dose, the potency parameter 0 is estimated by
fitting this model to the epidemiologic data by the method of maximum
likelihood. Comparable linear dose response approaches are applied to
case control and cross-sectional epidemiological studies.
The parameter ft is used in conjunction with a life table analysis that
employs U.S. sex- and age-specific mortality rates for the cancer in
question to estimate the "extra risk" of daoth by cancer from a speci-
fied human exposure pattern. Extra risk is defined as (P - PQ)/(I -
where P is the lifetime probability of death from the cancer under
10
-------
consideration in the presence of the postulated exposure and PO io the
background lifetime probability in the absence of exposure. Extra risk
may be interpreted as the probability of death from the cancer under
consideration, given that without the exposure death Mould have been due
to some other cause.
A constant daily exposure for 45 years beginning at age 20 is used as
the reference human exposure pattern for the calculation of human risk.
This pattern is taken as a compromise between the exposure patterns
found in most of the epidemiological studies (which are of occupation-
ally exposed cohorts for the most part), and constant lifetime exposure
beginning early in life that is typical of animal bioassays. The
endpoint estimated is the daily dose rate in mg/kg/day under this
exposure pattern that will produce an extra risk of 0.25. This daily
dose rate is called a "risk related dose* (RRO). Since the extra risk
measured in most of the epidemiological studies is less than 0.25,
estimation of RRDs will generally require extrapolation beyond the dose
ranges of the epidemiological data. On the other hand, an extra risk of
0.25 can generally be measured directly in standard animal bioassays;
consequently, use of 0.25 as a reference risk should make the analyses
of the animal data robust with respect to the dose response model
selected. The choice of a reference risk of 0.25 therefore represents a
compromise designed to minimize the extrapolation required beyond the
dose and response ranges in the animal and human studies.
Exposures in the epidemiologically studied cohorts are frequently the
source of considerable uncertainty in the analyses. For example,
exposuras in occupational cohorts are often measured infrequently and
those measurements that are made are sometimes of uncertain relevance to
exposures of specific workers. It is considered to be important to
quantify this uncertainty, although such quantification is difficult.
The approach adopted is to estimate uncertainty factors that represent
11
-------
our impression of the uncertainty of the dot* estimates for any given
study. These factor* ore applied to estimate upper and lower bounds for
the exposures in the epidemiological studies. To promote uniformity in
determining these factors, fairly specific guidelines for their calcula-
tion* were adopted a priori and followed consistently for each chemical.
A single investigator (B.A.) developed the bounds for each chemical and
for each study. As additional studies were analyzed, the uncertainty
bounds derived earlier were reviewed and occasionally revised. To
minimize the possibility of unintentional bias, all of the analyses of
the epidamiological data wer« performed independently of the analyses of
the animal data.
The upper and lower bounds on exposures in the epidemiological cohorts
are applied, along with statistical confidence limit procedures, to
estimate upper and lower bounds for ft. These bounds or* then translated
into upper and lower bounds for the RRO. The analysis of each epidemio-
logical study therefore produced a best estimate RUD and corresponding
lower and upper bounds, RRD|_ and RRDy, that reflect both the statistical
uncertainty in the obsarved cancer responses in the epidemiological
studies end the uncertainty in the exposure levels.
In many cases, more than one triple (RRDj_, RRD, RRDg) for a chemical ir.
available from the epidemiologic literature, either because of more than
one study or more than one carcinogenic response analyzed. Rather than
combining results for different responses or from different studies, a
single triple is selected to represent the potency of a given chemical.
The triple that in selected is one that corresponds best with the
consensus of opinion about the carcinogenic effect of the chemical
determined froir all the literature reviewed. Howaver. the results from
. a study or particular response in a study are not used if the dose-
response modal provided o poo* fit to the data or if the study is deeded
to be markodly inferior to other studies providing RRD estimates. In
12
-------
the ease of vinyl chloride, for example, a liver cancer response is
chosen since ongiosarcoma of the liver i« considered to be undeniably
linked to vinyl chloride exposure whereas respiratory cancer, another
endpoint analyzed, is net so cltarly linked. Another example is
provided by isoniazid. Overall, the literature on ieoniar'c d's's not
conclusively demonstrate its carcinogenicity i-» humans le* »lr RRDA' ""DAU^ obtained
from the animal data. If the correlation analysis is positive, then it
is reasonable to ask if particular RRO estimates obtained from animal
data are good predictors of the results obtained directly from epidemio-
logical studios. At this stage one can also examine the magnitude of
errors, i.e. the uncertainty that results from the use of any predictor.
Both correlation and prediction analyses require RRDs from animal data
that ar« similar to those obtained from the epidemiological data.
Calculation of RRDs from Animal Data. For each carcinogenic response
coded from a study testing the chemical of interest, a multistage model
is fit to the dose-response data (<»). The model is fit by an updated
13
-------
When averaging is carried out at every level - over sex, study, and
species (Analyses 12 through 2Ud) the averaging serves to define a
unique triple for each chemical For the remaining analyses, the
collection of RRDs must be further condensed to obtain a unique triple
for each chemical.
For analyses in which no averaging is conducted (Analyses 0 - Be and
25), two predictors from the lower bounds on RRDs are selected: one, LM,
by taking the minimum of the lower bound RRDs, and the other, I-2Q, by
taking the second quartile (median) of the lower bound RRDs, first within a
species, and then taking the median of the species-specific medians.
This approach to computing medians is similar to the method of averaging
described above, and is designed to insure that different species
contribute equally to the RRDs. The maximum likelihood RRDs and upper
bound RRDs are similarly combined and consequently two different types
of triples are produced: (L^, MLE^. UM) and (LZQ- MLEgg, UJQ). For
analyses in which only partial averaging is conducted (Analyses 9 -lib),
the approach token can be roughly described as the some as that just
described for the case of no averaging, except applied to those RRDs
remaining after the appropriate averaging process is complete. Thus two
sots of triples from the animal data are produced for all analyses
except those for which averaging is carried out at every level (Analyses
12 - 2
-------
higher quality. To address this problem, a data sieve was designed such
that, when applied, only higher quality data are used in an analysis.
The sieve is composed of two screens that can operate either separately
or in tandem. The first, the significance screen, examines each data
tet for a statistically significant (p < 0.05) increase in responses at
any treatment group over that in the control group by Fisher's exact
test, or for a statistically significant dose-response trend by the
Cochran-Armitage test. If at least one of the data sets for a chemical
eligible 'or an analysis satisfies this condition, all data sets for
that chemical not satisfying the condition are deleted from the
analysis. If no data sets for a chemical satisfy the condition, then
none of the data sets for that chemical are deleted on the basis of the
significance screen.
The second screen, called the quality screen, screens on the basis of
the length of observation and the number of dosed animals. Each data
set is assigned a rank according to the scheme depicted in Table 6. All
data sets assigned a rank that is higher than the lowest rank of any
data set otherwise eligible for an analysis are excluded from the
analysis.
The sieve is applied to the data sets that would otherwise be eligible
for a particular analysis. When both screens ore employed, the signifi-
cance screen is applied first. The sieve is designed to select the best
data sets pertaining to a chemical among those eligible for a particular
analysis, but not to be the basis for the exclusion of any chemical from
an analysis. Note in this regard that there is no way that use of
either screen can cause all of the data for a chemical to be eliminated
from an analysis.
17
-------
INVESTIGATION OF COMPONENT-SPECIFIC UNCERTAINTV
The importance of individual components and choices for those components
(listed in Table 1) to risk assessment are investigated by constructing
histograms of the ratios RRDx/RRDjo of RRDs obtained from animal data
for the various chemicals, where RRDjg represents on RRD obtained from
Analysis 30, and RRDX represents an RRO obtained from an analysis that
differs from 30 with respect to an approach to a single risk assessment
component. Specifically, x is allowed to range over Analyses 31 to 50,
as each of these differ from Analysis 30 only in the approach to a
single component. Since human data are not required for this investi-
gation, data for all 4
-------
Analyse* *5 - 47 eoeh differ from Analysis 30 only in the manner in
which results from different studies are combined, and each is
associated with a relatively small dispersion factor. This indicates
that the manner used to combine data is relatively unimportant; all
approaches considered give roughly comparable results.
The remaining analyses differ from Analysis 30 with respect to compo-
nents that relate to length of study (Analysis 37), length of dosing
(Analysis 38), exposure route (Analyses 37 and 38), tumor type to use
(Analyses
-------
A stotisticol test was conducted for each of the selected methods of
bioassay analysis to determine if the RRDs estimated from animal data
were significantly correlated with those estimated from human data.
Specifically, the test determined whether the intervals defined by the
upper and lower bounds for the human RRDs were significantly correlated
with the corresponding intervals calculated from the animal data. A
generalization of Spearman's rho statistic (5) was used that applies to
intervals rather than individual points. In -:his statistic, the inter-
val for one chemical was considered to ranK higher than that for a
second chemical if both the lower and upper bounds of the first interval
were larger than the respective bouncis for the second interval. The
statistical significance of a particular analysis was evaluated by
randomly reassigning the human intervals to chemicals while keeping the
animal intervals assigned to the correct chemicals (a permutation test).
The p-value of the statistical test represents the probability that,
given the animal and human intervals calculated, a correlation as large
or larger than that observed could have occurred by ~ random assignment
of these intervals to chemicals.
Prediction Analysis. If the correlation analysis just discussed finds a
positive correldtion between the animal and human RRDs, it is reasonable
to determine which particular estimates derived from the animal data
best predict the results obtained directly from the epidemiological
data, and to determine how well these estimates predict the animal
results. The prediction analysis therefore selects a single estimator
from the bioassay results as the estimate of RRD for each chemical.
Four types of estimates are investigated: the minimum and median of the
lower bound estimates (LM and LJQ) and the minimum and median of the
maximum likelihood estimates (MLE^ and
20
-------
or* not distilled to o single point, but rather the
, and the interval [RRDHL, RRDHU] are used to
•or s. In these evaluations, a straight liae with
-> the base ten logarithmic transform of predictor
e human RRDe. Plots of these fits are produced with
lotted vertically and the predictors derived from the
te
-------
RESUL" c ANIMAL AND HUMAN COMPARISONS
CORRELATION ANALYSES
Table 8 contains the correlation coefficients and their associated
p-volue* corresponding to each of the initial 38 methods of analyzing
the bioassay data studied. Figures 2 through 8 contain graphs of
selected analyses. This summary reports only results from analyses that
applied the data sieve described earlier. Use of the sieve gave a
higher correlation in 28 of the 38 analyses and in each of the 10
exceptions the reduction in the correlation was marginal.
The results in Table 8 provide a strong indication of a positive corre-
lation between the animal and human RRD estimates. Thirty-five of the
38 analyses had a p-value less than 0.05, indicating a statistically
significant positive correlation between the animal and human RRDs.
Fifteen of the analyses had a p-value of 0.0001 or smaller, including
the Base Case analysis which attempts to mimic the analysis method used
by the USEPA. Not only are the correlation coefficients statistically
significantly positive, but they are sizable in an absolute snnse as
well. Twenty-six of the analyses yield a correlation coefficient larger
than 0.7.
Given these results, it is highly unlikely that these correlations are
due to chance. It is also highly unlikely that they are due to bias in
the methods employed. Unlike the earlier study by the National Academy
of Sciences (6), this study was not limited to chemical*, that were
unequivocally carcinogenic in both animals and humans; thus this
potential source of bias was avoided. All animal analyses were
conducted using a computer program that avoided chemical-specific
decisions by an investigator that might perhaps unconsciously be biased
towards improving the correlations. Although the analyses of the human
22
-------
data did require judgement* involving individual chemicals, these
judgements were made blind, without knowledge of the outcome of the
animal analyses. Thus, by any reasonable standard, the animal RROs are
substantially correlated with the- human RROs. This correlation is very
important because it destionstrates that it is scientifically feasible to
estimate human risk from animal data.
Discussed below are highlights of the correlation analysis results as
they relate to specific individual or groups of analyses.
Analyses that Average Over -.ex, Study, and Species (Analyses 12-2
-------
both the cam* rout* and response os in humans has a somewhat larger
correlation than Analysis 0. These mixed results suggest that, given
the uncertainties in the present study with respect to the human RRDs,
it does not appear necessary to base a risk assessment on a lesion known
to result in humans from exposure to the chemical in question.
Similarly, it does not appear to be essential to limit animal data to
experiments employing the same route of exposure as humans experience.
Analyses Based on Only Malignant Tumors (Analyses 7, TO. These
analyses provide essentially the same correlations as their counterparts
(Analyses 0 and 12) that use both benign and malignant tumors, despite
the fact that the human results are for malignant tumors exclusively.
This suggests that there is no claarcut choice between use of malignant
tumors only and use of both benign and malignant in risk assessment and
that reasonable risk assessment methods could be based upon either
approach.
Analyses Restricted to Specific Species (Analyses lib, lie. 11d).
Analycis 11b that averages results from mice and rats provides essen-
tially the same correlation as Analysis 11a that averages results from
all species. This may be a reflection that the vast majority of the
data in the Animal Onto Base is from either mouse or rat studies (cf.
Table 2). RRDs from rat studies (Analyses 11c), mouse studies (Analyses
lid), and both mouse and rat studies (11b) give nearly identical
results.
Choice of Dose Units (Analyses fro, frb. »c. frd. 2»o. 2»b. 2»c. 2<»d).
Selection of dose units for assumed animal-human equivalence has very
little effect upon the correlations; this is expected because relatively
few studies in the Animal Data Base include study-specific data on body
weight, food consumption, and other variables that affect calculation of
the dose measure. However, this choice can have a major effect upon the
-------
actual extrapolated human estimate* derived from animal data. This
important i«su« will be explored in connection with the prediction
analyse* in the next section.
Identification of Analyses Yielding Higher Correlations. Analysis 3b
(Figure 3) yields t:ie highest correlation, f • 0.90. Interestingly,
this analysis is tho least restrictive of all, being the only one that
involves instillation, injection, and implantation studies as wall as
the more standard gevage, inhalation, and oral studieu. This analysis
was the o/ily one that included chlorambucil, chromium, and melphalan,
since data from experiments using the standard routes of exposure were
not available for those chemicals. The correlation analysis was
repeated for Analys.it 3b with these three chemicals omitted to determine
if the high correlation is related to the addition of these chemicals to
the analysis. The resulting correlation was 0.88, which is very close
to the original value, t * 0.90, and is still notably better than the
correlation obtained from any other analysis.
Aside from Analysis 3b, no other analysis stands out from the others.
The next highest correlation is 0.81 (Analysis 25) and another 16
analyses yield correlations between 0.76 and 0.81. The higher
correlation obtained from Analysis 3b which employs routes of exposure
not normally used for risk assessment suggests that inclusion of these
routes may allow improved estimates for some human carcinogens that, for
some reason, are not easily shown to be carcinogen : in animals via
routes through which humans are normally exposed. Further investigation
of this issue may be v ,rranted.
25
-------
PREDICTION ANALYSES
In the prediction analyses a single RRO estimated from the animal data
is used to predict the RRDs obtained from the human data. The fidelity
of the prediction is measured by three loss functions: DISTANCE2,
CAUCHY. AND TANH. Thus, whereas the correlation analyses consider only
whether higher ranked animal RRDs are associated with higher ranked
human RRDs, the prediction analyses examines the ability of the cnimal
bioassays to predict human risk. It also includes an examination of the
magnitude of the errors resulting in prediction of human RRDs from
animal RRDs.
As in the correlation analysis, the use of the sieve to screen the data
appears to be appropriate and useful. This is particularly true when
predictors other than the lower bound median, LJQ, ore used. While
application of the sieve increased average loss for -some analysis
methods when Lgo was the predictor used, this con probably be largely
attributed to confounding associated with use of the sieve and to random
factors. It is concluded that definition and application of some data
screening procedure that eliminates from consideration experiments of
lesser quality should accompany assessments of risk that depend on
animal data.
Evaluation of Animal to Human Conversion Methods. Heretofore, animol-
to-human extrapolation has generally been conducted by assuming that
equal doses /ill produce the same lifetime risks in animals and humans
when both animal and human doses ore measured in the same particular
units. The dose units studied in this report (mg/kg body weight/day,
mg/m2 surface area/day, ppm in air or water, and mg/kg body
weight/lifetime) have all been applied in the past. Because of
differences between animals and humans in body weights, life spans,
etc., use of different units will produce different estimates of human
26
-------
risk. There i« limit»d scientific support for use of any particular
dose unitti (7). However, result* Vrom the present study can be used to
empirically evaluate these different conversion approaches. Specifi-
cally the "conversion factor" 10C, where c is the y-intercept from the
best fitting line on the log-log plots of human and animal RRDs, is an
estimate of che amount the RRD* obtained from the animal data would have
to be multiplied by in order to agr«e, on overage, with the RROs
obtained from the human data. A conversion factor larger than 1
indicates that the RRDs obtained from animal data tend to underestimate
those obtained from human data and vice-verso.
Table 9 contains these conversion factors for two loss functions
(CAUCHY AMD TANK) and for three different sets of analyse* chosen such
that the analyses within o set differ only with respect to the dose
units assumed to yield equivalence between animals and hurrans. These
sets are {ti,d}, C\2,2ka,2<*b,2kc.2kc\), and (31,30,32,33.34).
This table indicates that use of the mg/kg/lifetime dose measure leads
to overestimation of the human risk, for all analysis methods
considered, by estimated factors ranging from 10 to 150. Similarly,
use of mg/tt? surface ar«a/doy also leads to overestimation of risk, by
factors ranging from 1.6 to 12. This is significant because this is the
dose measure generally usea by EPA to estimate human risk. Actually,
the extent of overvstimation by EPA may be greater that indicated in
this table (cf. Tab.Ve 10); EPA'a analysis method generally uses
additional conservative assumptions (such as taking the animal data
indicative of the highest risk rather that using medians or averaging
over studies) not applied in the analysis methods listed in Table 9.
(However it should be kept in mind that non«< of the analyi«s methods
studied will faithfully reproduce EPA's risk assessment results.)
Table 9 indicate* that the dose measure mg/kg/oay provides more nearly
unbiased estimates of human risk wHen the most appropriate analysis
27
-------
method at determined in the prediction analysis (i.e. method 30) is
used. Interestingly, this measure also generally provided about the
smallest loss among the five dose measures, although the differences in
loss were small, as expected.
There is no obvious a priori reason why any particular dose measure is
the "correct1 one to use for animal-to-human conversions. Results from
the present study can be used empirically to determine appropriate
conversion methods. Specifically, multiplication of trie animal RRO by
the conversion factor, 10C, provides an estimate of the human RRD in
which the bias due to systematic differences in animal and human risk
estimates found in this study have been eliminated. With this approach,
the dose units can be selected on the basis of those that, along with
other facets of an analysis, produced the best correlations between
animals and humans (or smallest losses). Application of the correction
factor 10C eliminates the bias associated with any method by correcting
for any overestimation or underestimation produced, on average, by that
method.
Predictors. Of the four types of predictors investigated
MLEffl, MLEjo). the lower bound median is clearly superior to the others.
This is the indicated by all three loss functions used. Consider the
twenty analyses 0-11d (cf. Table 4). With DISTANCE2 loss, LjQ 9av* °
smaller loss than LM in every case (ML EM and MLEjQ are not considered
with this loss function); with TANH loss, LJQ gave a smaller loss than
the other three typos of predictors in 18 analyses; with CAUCHY loss,
L20 9°v* ° smaller loss then the other three types of predictors in 15
cnolyses.
The superiority of LJQ over the predictors based on maximum likelihood
estimates may be related to the fact that stnall changes in the bioassay
data con result in sizable changes in MLE estimates of RRDs. This
28
-------
suggests that the large-sample theoretical properties of MLEs (such as
consistency and asymptotic efficiency) are not operative to any
practical extent in this situation, given the usual sample sizes
encountered in bioassays. The lack of stability of the MLEs is even more
of a problem when extrapolating to low dose or low risk. Regulatory
agencies have in the past relied more heavily on lower bound RRDs than
on maximum likelihood estimates, mainly in the interest of being
protective of human health. This study provides additional support for
that policy since the lower bound median is, in fact, a better predictor
of human risk estimates than are the MlE predictors (in the sense of
providing smaller lost).
Comparison of Analysis Methods. Given that the superiority of LJQ over
the other predictors has been established, it is desirable to identify
which analysis methods based upon this predictor provide the best
estimates. This task is complicated by the fact that three different
loss functions have been defined, and these do not agree completely with
respect to the analysis yielding smallest loss. Moreover, it seems
unlikely that there would exist a single "best" method. Consequently,
we hove identified a small set of analysis methods that perform
relatively well with respect to all three Iocs functions.
Several such analysis methods, along with others that ore of general
interest are listed in Table 10. All of the results in this table are
from applying the LJQ estimator, except in the one case noted on the
table. The "incremental normalized loss" presented in this table is a
summary loss measure synthesized from all three loss functions. For
each loss function separately, it is possible to determine for a
particular analysis the amount of additional loss over the minimum
contributed by that analysis. The sum of these additional losses over
the three loss functions defines the total incremental normalized loss.
The "corversion factors* listed in Table 10 arc the average factors,
29
-------
10C. by which RRDs obtained from the animal data would have to be
multiplied by in order to agree, on average, with the RRDs obtained from
the human data; these factors-were discussed in an earlier section. The
last column in Tcble 10 contains values of the residual error, which
represents the average distance on a log-log plot from the interval
defined by the human RRDs to the line that fits best, given the animal
RRD predictors and the intervals determined by the human RRDs. This
residual error represents roughly the average multiplicative error in
estimating the human RRDs from the animal data that is not explainable
by the uncertainty in the human RRDs (this uncertainty being expressed
by the intervals [RRD|_, RRDy] estimated from the human data). The
residual error is in essence an additional expression of loss.
The Base Analysis (Analysis 0) employing the minimal lower bound
estimator, LM (second row of Table 10) has both the largest normalized
loss and the largest residual error. Moreover, RRDs derived from this
analysis underestimate the human RRDs on average tv a factor of 12. By
all standards, this method is the poorest of those listed. This method
is also perhaps most like that presently employed by EPA. Modification
of this method by using the median lower bound estimator, LJQ. rather
than LM, as represented in the first row of Table 10, provides an
improvement in terms of normalized loss, residual error, and requiring a
smaller conversion factor. These results illustrate further the finding
discussed earlier that analysis methods that use median lower bound RRDs
as estimators provide smaller losses than analysis methods that use
minimum estimates.
Use of malignant tumors only, rat data only, or mice data only
(Analyses 7, 11c, and 11d, resoectively) did not provide clear improve-
ments over estimates that included data on normalignant tumors and data
from different species.
30
-------
Analyses 30, 31, 42. 45. and 47 or* presented as a group of analyses
thnt generally perform well. All of these analyses use the mg/kg/day
method of extrapolating from animals to humans (except 31, which
utilizes the mg/m^/day method), and all include routes of exposure
(instillation, injection, and implantation) not normally used in
quantitative risk assessment. Analyses 30, 45, and 47 differ only in
the way RRDs are combined and give fairly comparable results; Analysis
45 which averages RRDs from different sexes in the same study, migM be
considered to perform the best over-all, as it has both the smallest
normalized loss and residual uncertainty. This analysis also had the
largest correlation (u.91) of those in Table 10. Analysis 43 employs a
different carcinogenic endpoint than the others, namely total tumor-
bearing animals. Although this analysis has a small normalized loss,
its residual uncertainty factor is 401( larger than any from Analyses 30,
45 and 47.
Options for Presenting a Range of Risk Estimates. Guidelines ore
provided for presenting a range or risk estimates for a risk assessment
based on Analyses 30, 31, 43, 45, and 47. Three options are conuidered.
The first entails selecting, a priori, one method from the recomnended
set. The results of that method, including the uncertainty quantified
by the residual uncertainty factor, are taken as the representative
range of risk estimates. The second option uses all the methods. The
range it produces includes any value that could be obtained from any one
or more of the methods, and so can be considered to give the maximum
range consistent with the recommended set. Although the third option
also considers all methods in the recommended set, it summarizes the
results by the smallest range of estimates that is consistent with the
predictions of all the analyses. As wit* the first option, the last two
incorporate the residual uncertainty factors to define the ranges of
estimates.
31
-------
DISCUSSION
GENERAL CONSIDERATIONS
The animal data base and the methods used in this study provide a useful
basis for evaluating quantitative risk assessment. Their use in the
present context has demonstrated ths strong positive correlation between
the animal and human risk estimates and hence relevance of animal
carcinogenicity experiments to human risk estimation. Moreover, it has
been possible to identify methods of analysis of trie biocscay dato.
including the choice of the median lower bound predictor, that
satisfactorily predict risk-related doses in humans. Application of
these methods has led to suggested guidelines concerning the prediction
of human risks and the presentation of ranges of estimates incorporating
the relevant uncertainties.
There ore, however, certain features of this investigation that should be
borne in mind when evaluating the results of this study. These are
summarized below.
• A risk level of 0.25 is used throughout.
• The b&oassay data is rather crude in several respects. We have
already referred to the data deficiencies and their impact on
the ability to perform some analyses.
• The epidemiological data is of variable quality. Some degree of
subjectivity is inherent in the estimates of uncertainty
associated witti the epidemiological RRDs.
• Different forms (complexes) of some chemicals were grouped
together.
• Other approaches to the components could be defined and
investigated.
-------
• The three loss functions employed in the prediction onolysis
lock an underlying statistical development and so have been used
merely to rank the analysis methods.
• Many other analysis methods could be investigated.
DIRECTIONS FOR FUTURE RESEARCH
In the course of the previous discussion, several proposed extensions of
this project have been mentioned. Several fall under the heading of
sensitivity analyses of the results already obtained. These include
investigation of the robustness of the results to reasonable alternative
choices for the epidemiological estimates; examination of other means to
analyze bioassay data, including time-to-tumor analyses; and
investigation of the effect of using lower levels of risk, say 10~e,
which ere of direct regulatory concern. A detailed statistical
development of the loss functions used hare (or a general development
for certain classes of loss functions) might be of general interest.
The data that is available from this project could provide ?n
interesting and pertinent example to which that development could apply.
Also discussed in connection with component-specific uncertainty are
effort* directed at reducing or explaining that uncertainty. The
greatest uncertainties are related to the components specifying how to
handle experiments of different length*; of dosing, routes of exposure,
or test species and specifying the carcinogenic responses to use. Many
aspects of these components and their uncertainties can be addressed in
an investigation of pharmacokinetics. The data base contains detailed
data on th» timing and intensity of exposure for each bioassay, so a
pharmacokinetic study, which requires such information, is entirely
feasible with the currently collocted data. Two specific proposals are
discussed here.
33
-------
Risk estimates incorporating pharmacokinetic data could be used to
determine appropriate surrogate doses. It is sometimes assumed that a
given dose measured as average concentration of t!ie active metabolite at
the target tissue will produce the same risk in animals and humans.
However, given the many differences between animals and humans (size,
life span, and metabolic rates, to mention a few), it is not clear
which, if any, surrogate dose is the most appropriate. This issue is
similar to that of choice of the most appropriate surrogate dose measure
for animal to human extrapolation (e.g. mg/kg/day versus mg/m2/day)
considered in this study and can be studied in a similar manner. Risk
estimates using pharmacokinetic data could be used to determine
empirically the most appropriate surrogate dose. Even though the range
of RRDs consistent with the human data generally cover a range of an
order of magnitude or greater, the potential surrogate doses cover an
even wider range. Just as the present study indicates that certain dose
measures appear to predict human results well in conjunction with
appropriate choices for other risk assessment components, a study using
pharmacokinetic data should allow similar conclusions regarding the
surrogate dose. A preliminary investigation indicates that possibly 18
of the 23 chemicals with suitable human data used in this study might
also have data that would support a risk assessment that incorporates
pharmacokinetic data.
A second potentially useful investigation incorporating pharmacokinetic
data involves using the data in the data bcse on different routes of
exposure to study the best means of extrapolating from route to route in
animal studies. Risk assessment methods, including the ones examined in
this study, often assume a given dose rate involves the same risk,
regardless of route. This clearly is a gross oversimplification. The
animal rtata collected for this study contains numerous examples of
carcinogenicity studies on the same rnemical and animal species, but for
-------
which exposure is through different routes. Those studies could be used
to determine how phornacokinetic data could best be applied to perform
route-to-route extrapolation. Since human data would not be essential
in these investigations, our total data base that encompasses 4ft
chemicals could be used.
The question of different chemical classes and the consistency that may
be apparent within any of the classes is deserving of further study. It
would be reasonable to couple this work with pharmacokir.etic methods.
In the present data base, several classes are represented. However, the
number within any particular class is somewhat limited. An expanded
data base may be necessary for a thorough investigation.
In fact, one desirable goal in and of itself, but one that would enhance
the prospects for successful completion of these other proposals, is the
maintenance and updating of the bioassay data base. All aspects of
this, including accumulation of more data sets for the chemicals already
included and addition of more substances, may be necessary. Some
revamping of the data coding format may also make future analyses easier
and more accurate. Especially for pharmacokinetic studies, for
instance, dose patterns could be recorded on a daily rather than weekly
basis.
As a counterpart to the bioassoy data base enhancement, updating and
augmenting the epidemiological data is essential. Since the
epidemiological data (in particular, data on exposure) is the single
most limiting factor preventing use of human data, any hope of
increasing the size of the sample of chemicals useful in estimating
conversion factors and residual uncertainty must be based on an effort
to acquire such data. For those chemicals already analyzed, more
specific exposure data would reduce the uncertainty bounds surrounding
epidemiological RRD estimates and refine our estimates. As is the case
35
-------
with the bioassay data, much of the limitation or uncertainty is solely
a matter of inadequate reporting of data.
It should be noted in passing that the methods and portions of the
computer programs developed and applied in this project may be useful in
other contexts. Of particular interest is a study of other types of
health effects, e.g. reproductive effects. The investigation of these
issues could include determinations of uncertainty as well as
identification of the most appropriate methods. Other projects,
including investigation of other types of extrapolations, e.g. from one
temporal dosing pattern to another or from rats to mice, could also be
facilitated by use of the data base, methods, and programs developed in
the present work.
Finally, one would like to investigate cancer risk assessment methods
appropriate when data available to a particular assessment are limited.
We have mentioned this problem in connection with component-specific
uncertainty (i.e. noting that confounding like that affecting those
uncertainty calculations will often be present in any given risk
analysis setting) and in connection with the set of recommended bioassay
analysis methods. In the latter instance, it was pointed out that each
analysis in the recommended set, save for Analysis 17, is capable of
being applied to any data base but that data limitations due to
incomplete data presentation may entail that Analyses 20 and 43 are not
possible. The remaining analyses (30, 31, 45, and 47) can be performed
no matter what the data set contains, but they may be seriously affected
by the extent and nature of the contents.
Consequently, the following investigation is proposed as a mt>ans of
studying the effects of the limitations on the data for any chemical of
interest and of determining how best to extrapolate risks to humans.
Pick the data in the data base that most nearly matches the data for the
36
-------
chemical in question. The matching may be based on species, routes of
exposure, and quality of the data. Moreover, one may wish to restrict
attention to chemicals that are in the same class of the substance of
interest. Suppose, for example, a volatile organic chemical is under
investigation and that the only data available are from rat inhalation
studies. Then, the proposed procedure would first select rat inhalation
bioassays conducted using appropriate chemicals (i.e., perhaps limited
to volatile organics). The components of risk assessment not fixed by
the selection could be varied and the method that works best with the
selected data would be the basis for extrapolating to humans risks due
to the chemical in question. Since we also have a recommended set
consisting of methods that appear to perform well for the data and
chemicals considered as a whole, the risks estimated on that basis (i.e.
using the recommended set) would be available for comparison. Those
estimates reveal what would happen if other species, other routes, and
other chemicals are included. The relationship between the estimates
obtained by the two approaches would suggest a general type of
uncertainty attributable to use of a limited data base (in this example,
rat inhalation studies). A pilot study could investigate the
feasibility of such a chemical-specific approach to risk assessment.
37
-------
REFERENCES
1. Gold, 1., Sawyer, C., Magow, R., Bookman, 6., de Veciana, M.,
Levinson, R., Hoooper, N., Havender, W., Bernstein, L., Peto, R.,
Pike, M., and Ames, B. (198<*)- A carcinogenic potency database of
the standardized results of animal bioassays. Environmental Health
Perspectives 58:9-319.
2. International Agency for Research on Cancer (1982). Chemicals,
industrial processes and industries associated with cancer in
humans. IARC Monographs on the Carcinogenic Risk of Chemicals to
Humans Supplement 4.
3. National Academy of Sciences Executive Committee (1975).
Contemporary Pest Control Practices and Prospects. Pest Control:
An Assessment of Present ond Altar-native Technologies, Vol. 1.
4. Howe, R. and Crump, K. (1982). GLOBAL 82: A Computer Program to
Extrapolate Quantal Animal Toxicity data to Low Doses. Prepared for
the Office of Carcinogen Standards, OSHA, U. S. Department of Labor,
Contract 41USC252C3.
5. Ng, T. (1985). A Generalized Ranking on Partially Ordered Sets and
It* Applications to Multivoriote Extensions to Some Nonparametric
Tents. Unpublished report.
6. National Academy of Sciences Executive Committee (1975).
Contemporary Pest Control Practices ond Prospects. Pest Control:
An Assessment of Present ond Alternative Technologies, Vol. 1.
38
-------
7. Crump. K., Silvers, A.. Ricci. P.. and Wyzgo. R. (1985).
Interspecies comparison for carcinogenic pctency to humans.
Principles of Health Risk Assessment. Ricsi, P. (ed.). Prentice
Hall. pp. 321-372.
8. U. S. Environmtneal Protection Agency. Addendum to the Hea?.th
Assessment Document for Dichlo^-omethane (Methylene Chloride).
EPA-600/8-82-004FF.
39
-------
Table 1
APPROACHES TO RISK ASSESSMENT COMPONENTS
1 Length of experimen*
o^ U
-------
Table 1 (continued)
APPROACHES TO RISK ASSESSMENT COMPONENTS
9. Combining data from males and female*
o_._ Use data from each sex within a study separately.
b. Average the results of different sexes within a study.
TO. Combining data from different studies
a_._ Consider every study within a species separately.
b. Average the results of different studies within a species.
11. Combining data from different species
a. Average results from all available species.
b. Average results from mice and rats.
c. Use data from a single, presel«cted species.
(T. Ut» all species separately.
NOTE: Underlines indicate approach used in base analysis (Analysis 0).
-------
Table 2
SUMMARY OF ANIMAL DATA BY CHEMICAL
No. Reviewed
Chemical
Acrylonitrile
Af latoxin
Allyl Chloride
4-Aminobipnenyl
Arsenic
Asbestos
Benzene
fanzidine
Benzc[a]pyrene
Cadmium
Carbon Tetrachloride
Chloromt>jcil
Oh lor done
Chloroform
Chromium
Cigarette Smoke
3. 3-Oichlorober.ridens
1 ,2-Dichloroethane
Dichloroirethane
Diethylstilbestrol
Diphenylhydrazina
Epichlorohydrin
Estrogen
Ethylene Dibromide
Ethylene Oxide
Formoldehyde
Studies0
10
62
2
8
16
39
13
8
42
26
8
3
3
12
12
37
6
5
6
61
2
7
24
7
10
9
Data
Sets0
19
86
6
8
33
84
26
10
51
30
21
6
8
31
16
41
8
14
14
81
2
7
34
19
15
15
Number of
Data Sets Coded in Animal
Oral
Rb
9
23
2
0
7
1
0
0
0
3
0
0
2
1
1
0
4
0
2
0
0
1
1
0
0
0
Mb
0
0
2
2
2
0
0
0
0
0
0
0
6
1
0
3
0
0
2
2
0
0
0
0
0
0
0"
0
3
0
0
0
8
0
0
0
0
0
0
0
2
0
0
1
0
0
0
0
0
0
0
0
0
R
4
3
0
0
0
0
6
1
1
1
2
0
0
3
0
0
0
2
0
0
0
0
0
2
1
0
Gavage
M
0
1
0
0
0
0
2
0
2
1
4
0
0
13
0
0
0
2
0
0
0
0
0
2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
Datnbase
Inhalation
R
6
0
0
0
0
11
2
0
0
1
0
0
0
0
0
3
0
2
6
0
0
2
0
4
9
5
M
0
1
0
0
0
0
4
0
0
0
0
0
0
0
0
0
0
2
2
0
0
0
c
2
0
4
0
0
0
0
0
0
9
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
1
R
0
0
0
0
5
17
0
1
3
1
0
0
0
0
6
1
0
0
0
2
1
0
0
0
0
0
Other
M
0
2
0
0
2
0
1
1
2
1
0
1
0
0
1
0
0
0
0
9
0
0
0
0
0
0
0
0
0
0
0
1
18
0
0
6
0
0
0
0
0
0
1
0
0
0
3
0
0
1
0
0
0
Total
19
33
4
2
17
64
15
3
14
8
6
1
8
20
8
9
5
8
12
16
1
3
2
10
10
10
-------
Table 2 (continued)
SUMMARY OF ANIMAL DATA BY CHEMICAL
No. Reviewed
Chemical
Hexoch lorobenzene
Hy*«-nrine
Isoniazid
Lead
Melphalan
Methotrexate
Mustard Gas
2-Naphthylomine
Nickel
Nitrilotriacetic Acid
Phenacetin
Polychlorinated
Biphenyls
Re.-erpine
Saccharin
2.3.7.8-Tetrochloro-
dibenzo-p Di*>xin
Tetrachloroethylene
Toxaphene
Trichlo»~oethylene
2.4, 6-Tr ichlorophenol
Vinyl Chloride
Vinyl idene Chloride
TOTAL
Studies0
4
15
23
22
4
9
2
23
37
7
13
9
2
19
11
5
1
39
1
35
17
736
Data
Sets0
7
31
66
33
7
16
4
37
77
18
21
12
6
27
19
14
4
34
4
65
46
1233
Number of
Data Sets Coded in Animal
Oral
"R6
2
0
4
9
0
0
0
1
0
9
5
3
2
14
3
0
2
0
2
4
2
119
' M°
2
7
17
2
0
2
0
1
0
6
6
.1
2
1
0
0
2
0
2
0
0
73
Ob
2
0
4
2
0
2
0
9
1
0
0
0
0
1
0
0
0
0
0
0
0
35
R
0
4
0
0
0
0
0
1
0
0
1
0
0
0
2
2
0
4
0
4
3
47
Gavage
M
0
8
11
0
0
0
0
4
0
0
0
0
0
0
3
2
0
8
0
0
2
65
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
1
Database
Inhalation
R
0
0
0
0
0
0
0
0
5
0
0
0
0
0
0
2
0
3
0
23
10
94
M
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
2
0
3
0
20
15
55
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
2
0
1
0
14
R
0
0
0
0
0
2
0
0
20
0
0
0
0
0
0
0
0
0
0
4
0
G3
Other
M
0
0
4
1
1
0
4
4
1
0
0
0
0
2
2
0
0
2
0
0
0
41
0
0
0
0
1
0
0
0
0
1
0
0
0
0
0
6
0
0
0
0
0
0
32
Total
6
19
40
15
1
6
4
21
28
15
12
6
4
18
10
8
4
22
4
56
32
631
°A study is generally comprised of all information contained in a single primary reference on a
simple chemical. A data set generally comprises oil of the dose response data fron a given
sex and species to animals via a common protocol in a study.
bR - rat; M • mouse; 0 > other species.
-------
Table 3
CHEMICALS FOR WHICH MINIMAL HUMAN AND ANIMAL
DATA EXIST FOR QUANTIFYING CARCINOGENIC POTENCY
Evidence for Carcinogenicity
(IARC classification scheme)
Chemical
Aflotoxin (AF)
Arsenic (AS)
Asbestos (AB)
Benzene (BN)
Benzidine (82)
Cadmium (CO)
Chlorambucil (CB)
Chromium (CR)
Cigarette smoke (CS)b
Diethylstilbestrol (DS)
Epichlorohydrin (EC)
Estrogens (ES) (conjugated)
Ethylene oxide (EO)
Isoniazid (IS) (isonicotinic
acid hydrozide)
Melpholon (ML)
Methylene chloride (MC)
Nickel (NC)
Phenacetin (PH) (analgesics
containing phenacetin)
Polychlorinoted biphenyls (PC)
Reserpine (RS)
Saccharin (SC)
Trichloroethylene (TC)
Vinyl Chloride IVC)
Use0
F
1C
1C
1C
1C
1C
D
1C
D
1C
D
1C
D
D
1C
1C
0
1C
D
F
1C
1C
In Humans
Limited
Sufficient
Sufficient
Sufficient
Sufficient
Limited
Sufficient
Sufficient
- -
Sufficient
Inadequate
Sufficient
Inadequate
Inadequate
Sufficient
Inadequate
Limited
Sufficient
Inadequate
Inadequate
Inadequate
Inadequate
Sufficient
In Animals
Sufficient
Inadequate
Sufficient
Limited
Sufficient
Sufficient
Sufficient
Sufficient
- -
Sufficient
Sufficient
Inadequate
Limited
Limited
Sufficient
Sufficient0
Sufficient
Limited
Sufficient
Limited
Limited
Limited
Sufficient
°IC • industrial chemical; 0 • drug; F • food additive or contaminant.
"Not considered in IARC monographs, although tobacco smoke is
acknowledged by IARC as a known human carcinogen.
cAlthough classified as "Inadequate" by IARC (2). results of studies
completed since IARC evaluation indiwate that the evidence for the
carcinogenicity of methylene chloride in animals is now "Sufficient*
(8).
-------
Table <»
DESCRIPTIONS OF INITIAL ANALYSES
Anolysis
0
1
2
3o
3b
ka
4b
kc
ka
5
6
7
80
8b
6c
9
10
110
•Mb
11C
lid
12
13
Ik
15
1ti
17
18
19
20
21
22
23
Tetnplote0
Base
Analytic
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
12
12
1*
12
16
12
18
12
20
12
22
Differences^
[described in Table 1]
limited to experiments of long observation
limited to experiments of long dosing
route like human route only
any route
mg/kg/day
ppm diet
ppm air
mg/kg/lifetime
doses averaged over first 80* of experiment
early deaths eliminated
malignant responses only
combination of significant responses only
total tumor-bearing animals only
response that human get only
results averaged over sex within study
results averaged over study within species
results averaged over all species
results averaged over rats and mice only
rot data only
mouse data only
results averaged over sex, study, and species
limited to experiments of long dosing and
observation
malignant responses only
limited to experiments of long dosing and
observation
combination of significant responses only
limited to experiments of long dosing and
observation
combination of malignant significant responses
only
limited to experiments of long dosing and
observation
total tumor-bearing animals only
limited to experiments of long dosing and
observation
total malign— sy-bearing animals only
limited to experiments of long dosing and
observation
-------
Table k (continued)
DESCRIPTIONS OF INITIAL ANALYSES
Analysis Template0 Differences13
2i*o
2<*b
2
-------
Table 5
DESCRIPTIONS OF SUPPLEMENTAL ANALYSES
Analysis Template0 Differencesb
30
31
32
33
34
35
36
37
38
41
42
43
44
45
46
47
48
49
50
0
30
30
30
30
30
30
30
70
30
30
30
30
30
30
30
3J
30
30
mg/ky/day; any exposure route
mg/m^/day
ppm diet
ppm air
mg/kg/lifetime
limited to experiments of long observation
limited to axperiments of long dosing
route like humans only
oral, gavace, inhalation, or route like humans
malignant responses only
combination of significant responses only
total tumor-bearing, animals only
response that humans get only
results averaged over sex within study
results averaged over study within species
results averaged over all species
results ai/erged over rats and mice only
rat data only
mouse data only
°The template is the analysis which o given analysis most closely
resembles.
bThe differences listed are the ways in which the analysis in question
differs from its template.
47
-------
Table 6
RANKS BASED ON LENGTH OF EXPERIMENT
AND NUMBER OF TREATED ANIMALS
Length of
Experiment0
> 75*
50-75*
< 50*
Number
50+
1
3
6
of Dosed Animals
15-49 <
2
4
8
15
5
7
9
°These values are expressed as percentages of the standard experiment
length of the test specias.
-------
Table 7
COMPONENT-SPECIFIC UNCERTAINTY: MODES AND DISPERSION
FACTORS FOR RATIOS OF RRDS°, DY SUPPLEMENTAL ANALYSIS"
Analysis
31
32
33
34
35
36
37
38
41
42
43
44
45
46
47
48
49
50
Number of
Chemicals
44
44
44
44
40
34
24
40
39
29
31
37
44
44
44
43
39
36
Mode of
Histoqram
.05
.2
.2
.02
.8
.8
.8
.8
.8
.8
.8
.8
.8
.8
.8
.8
.8
.8
- .1
- .5
- .5
- .05
-1.25
-1.25
-1.25
-1.25
- 1.25
-1.25
-1.25
- 1.25
- 1.25
- 1 25
- 1.25
- 1.25
- 1.25
- 1.25
Dispersion
Factor0
2.3
1.7
1.8
1.3
28.5
86.0
5.3
33.7
290.6
75.6
39.6
54.1
1.2
1.7
2.2
23.2
39.6
335.6
°The ratios are of the chemical-specific RRD estimates from the
indicated analysis to those of Analysis 30 (cf. Table 5).
bThe analyses were performed with the L2Q predictor and using the full
sieve.
cThe dispersion factor is the average fas-tor by which the ratios differ
from tlie mode.
49
-------
Table S
CORRELATION COEFFICIENTS AND ASSOCIATED
p-VALUES, BY ANALYSIS METHOD"3
Analysis
0
1
2
3a
3b
40
4b
4c
4d
5
6
7
8a
8b
Be
9
10
11a
lib
11c
11d
12
13
14
15
16
17
18
19
20
21
22
23
24a
24b
24c
24d
25
Number of
Chemicals
20
18
19
17
23
20
20
20
20
20
6
19
13
17
18
20
20
20
20
19
13
20
18
19
18
13
11
10
9
17
13
15
13
20
20
20
20
16
*
.78
.68
.49
.73
.90
.78
.76
.78
.78
.79
.79
.76
.56
.66
.76
.76
.77
.76
.76
.79
.76
.75
.43
.71
.46
.49
.58
.73
.79
.63
.38
.35
.18
.75
.74
.74
.75
.81
P-
value
.0001
.0015
.0153
.0007
<.0001
.0001
.0001
<.0001
<.0001
<.0001
.0342
.0001
.0214
.0022
.0001
.0003
.0002
<.0001
<.0001
<.0001
.0023
<.0001
.0416
.0005
.0316
.0436
.0301
.0090
.0058
.0043
.1023
.1036
.2821
.0001
.0001
.0001
<.0001
.0002
°A sieve to screen the data has been used.
50
-------
Table 9
CONVERSION FACTORS0 CORRESPONDING TO VARIOUS
DOSE UNITS, BY METHOD OF ANALYSIS"
Units
Analysis Method
mg/m^/day Restricted routes, unaveraged (0)
Restricted routes, averaged (12)
Unrestricted routes, unaveraged (31)
mg/kg/day Restricted routes, unaveraged (4a)
Restricted routes, averaged (24a)
Unrestricted routes, unaveraged (30)
ppm diet Restricted routes, unaveraged (4b)
Restricted routes, averaged (24b)
Unrestricted routes, unaveraged (32)
ppm air Restricted routes, unaveraged (4c)
Restricted routes, averagod (24c)
Unrestricted routes, unaveraged (33)
mg/kg/life Restricted routes, unaveraged (4d)
Restricted routes, averaged (24c!)
Unrestricted routes, unaveraged (34)
1.58 - 2.07
3.47 - 5.61
8.45 - 12.02
0.28 - 0.40
0.43 - 0.61
1.08 - 1.70
0.59 - 1.17
1.77 - 2.95
4.52 - 5.94
0.83 - 1.06
1.82 - 2.96
1.89 - 6.61
10.40 - 16.67
19.63 - 23.12
72.95 - 79.62
°The multiplicative factor by which bioassay-based RRDs overestimate, on
average, RRDs obtained from human data.
bThe range given is that suggested by the CAUCHY and TANH loss func-
tions; all results based upon median lower oound (LJQ) estimator.
51
-------
Table 10
COMPARISON OF RESULTS FOR SELECTED ANALYSES0
Bias-
Correcting Residual
Number of Correlation Total Incremental Conversion Uncertainty
Analysis Chemicals Coefficient Normalized Lossb Factors0 Factor*3
0
0°
7
11c
11d
17
20
30
31
43
45
47
20
20
19
19
13
11
17
23
23
17
23
23
0.78
0.78
0.76
0.77
0.76
0.58
0.67
0.91
0.90
0.74
0.91
0.89
1.15
1.71
1.40
0.62
1.01
0.27
0.62
0.39
0.53
0.28
0.27
0.28
1.6 -
12
1.6 -
0.81 -
3.7 -
2.8 -
0.69 -
1.1 -
8.5 -
0.18 -
1.2 -
1
2.1
12
3.6
1.9
4.3
2.8
0.78
1.7
12
0.29
1.7
1.7
5.3
16.2
5.4
4.5
3.1
4.2
7.1
2.0
2.0
2.8
1.7
1.8
°The results correspond to the member of the pcir (with sieve, without
sieve) that gives best results. For Analyses 11c, 20, and 43 this is
without the sieve; for other analyses this is -.with the sieve. The
median lower bound predictor, l_2Q. is used in all analyses except for
the exception noted.
bThis value is not the same as that in Table 2-8 because the inclusion
of the supplemental analyses reduced the minimum average loss for two
of the three loss functions and increased the maximum loss for all
three of the functions.
cThese values ore the factors, 10C, based on the y-intercepts from the
CAUCHY and TANK loss functions (cf. Tables 2-13 and 2-17) and represent
the average ratio of human RRDs to animal RROs.
dResidual uncertainty is from Table 2-21 or 2-22. It is the factor
computed for all chemicals and represents the average factor by which a
prediction must be multiplied or divided in order to eliminate
uncertainty not due to uncertainty in the human estimates.
"Using minimal lower bound estimator LM-
52
-------
10(11-13) 1'»»r
AHATOXIN
« (14)
-<0« (15)
ARSENIC
respiratory
" }
« O6)J
ASBESTOS
HO (17) ill
— • - 1(17) long
BENZENE
{(18)
"
.(20)
•rm« I
(18) I
(19)01—
(20)1
BENZIDINE
CADMIUM
«1H24)I-
CHIORAMBUCIL
49 (23) Wukrmu (morb«)i
-------
ETHYLENE OXIDE
• I (40) «tl
I • I (41)
Vuttrni*
ISONIAZID
(42) > « I
•n ( (43)«i • i
(44)« I
(43) »-
MELPHALAN
•I O (45) Vuk.mw (mM-tuli^)
MET HYLENE CHLORIDE
(4«) t-
NICKEL
(33) I
f (31 .32)1
\
. /(3 1 ,32) I -
*' ( '.33) ,-
(30)« I » I r*f?tr<
r»r>»l »«lvi» (m*rbt4pir«t«ry (m«rb>4<(v) (43) I
tw,r (41) « h
-3 -2 -\
L««|0IRRD(m«/M/<««>l
Figure 1. Beat estimates and upper and lover bound) for RRDs
from each human studg.
O harks the data selected to represent the chemical when comparisons
with bioassag-based estimates are made.
* Marks poor fit of linear dose-response model to data.
-------
ui
en
Fir'ire -n
Correlation of Anin.i] find H.inan F'tHI'a - Anrlvris 0
OO A
kl
I"
-2
-4
Base <"ase
Log of AnlMl RRO ErtlM
-------
Ul
o
Firure 3
Correlation of Animal and Hunan PPPs - Analysis 3b
Log «>H AniMl RRO E*tiMt»»
-3
~4
-------
I 2
u
-3
Firur? ^
Correlation of Animal an-1 Hunan f?RDs - Analysis 7
ES--,
-SP
,-EP
/
PC
-4 -3 -2-1 0 1
Log of An I Ml RRO E»t !••!••
-------
Firur*> 5
jrrelation of Anin.-il an-l ffi^nan f-irTlG - Analysis lie
Ul
oo
*•
•
Pit Patn • r.l
-2
-3
-------
"nrrelation of Ar.ir.al .-in i Mum-in "-H';; - Anrilvsir, lid
m
-------
o>
o
Figure T
Correlation of Animal and Human RRDc - Analysis 1?
I"
-2
-3
-4
Average Over f'ex, P
and Species
^SP
/
S
h&*
'£.
*4P
1E2,
I
I
NT
sc
AS
^BS,
_l L.
-5
-3
-2-1 0 1
Log of AnlMl RRO E*tliratM
-------
on
Fipure 8
Correlation of Animal and Human 3PDs - Analysis 25
Pane Route and Pesponse as Humans
"1
-2
-3
^_t0s
j • i
-GP
*9- ,
— *
csT
T! VC
sc
PH
^BS
-6 -4 -3
-2-1 0 1
Log of AnlMl RRO E«tlMU*
IS
m.
4 «
------- |