® EPA
EPA/635/R-13/128a
www.epa.gov/iris
Evaluation of the Inhalation Carcinogenicity of
Ethylene Oxide
(CASRN 75-21-8)
In Support of Summary Information on the
Integrated Risk Information System (IRIS)
July 2013
NOTICE
This document is a Revised External Peer Review draft. This information is distributed solely for
the purpose of pre-dissemination peer review under applicable information quality guidelines. It has
not been formally disseminated by EPA. It does not represent and should not be construed to
represent any Agency determination or policy. It is being circulated for review of its technical
accuracy and science policy implications.
National Center for Environmental Assessment
Office of Research and Development
U.S. Environmental Protection Agency
Washington, DC
-------
DISCLAIMER
This document is a preliminary draft for review purposes only. This information is
distributed solely for the purpose of pre-dissemination peer review under applicable information
quality guidelines. It has not been formally disseminated by EPA. It does not represent and
should not be construed to represent any Agency determination or policy. Mention of trade
names or commercial products does not constitute endorsement or recommendation for use.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 ii DRAFT—DO NOT CITE OR QUOTE
-------
CONTENTS
LIST OF TABLES v
LIST OF FIGURES vii
LIST OF ABBREVIATIONS viii
AUTHORS, CONTRIBUTORS, AND REVIEWERS ix
1. EXECUTIVE SUMMARY 1-1
2. INTRODUCTION 2-1
3. HAZARD IDENTIFICATION 3-1
3.1. EVIDENCE OF CANCER IN HUMANS 3-1
3.1.1. Conclusions Regarding the Evidence of Cancer in Humans 3-12
3.2. EVIDENCE OF CANCER IN LABORATORY ANIMALS 3-19
3.2.1. Conclusions Regarding the Evidence of Cancer in Laboratory
Animals 3-23
3.3. SUPPORTING EVIDENCE 3-23
3.3.1. Metabolism and Kinetics 3-23
3.3.2. Protein Adducts 3-26
3.3.3. Genotoxicity 3-28
3.4. MODE OF ACTION 3-38
3.4.1. Analysis of the Mode of Action for Ethylene Oxide Carcinogenicity
under EPA's Mode-of-Action Framework 3-42
3.5. HAZARD CHARACTERIZATION 3-44
3.5.1. Characterization of Cancer Hazard 3-44
3.5.2. Susceptible Life stages and Populations 3-49
4. CANCER DOSE-RESPONSE ASSESSMENT FOR INHALATION EXPOSURE 4-1
4.1. INHALATION UNIT RISK ESTIMATES DERIVED FROM HUMAN
DATA 4-1
4.1.1. Risk Estimates for Lymphohematopoietic Cancer 4-4
4.1.2. Risk Estimates for Breast Cancer 4-21
4.1.3. Total Cancer Risk Estimates 4-41
4.1.4. Sources of Uncertainty in the Cancer Risk Estimates 4-43
4.1.5. Summary 4-55
4.2. INHALATION UNIT RISK DERIVED FROM EXPERIMENTAL ANIMAL
DATA 4-56
4.2.1. Overall Approach 4-56
4.2.2. Cross-Species Scaling 4-57
4.2.3. Dose-Response Modeling Methods 4-58
4.2.4. Description of Experimental Animal Studies 4-60
4.2.5. Results of Data Analysis of Experimental Animal Studies 4-62
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 iii DRAFT—DO NOT CITE OR QUOTE
-------
CONTENTS (continued)
4.3. SUMMARY OF INHALATION UNIT RISK ESTIMATES—NOT
ACCOUNTING FOR ASSUMED INCREASED EARLY-LIFE
SUSCEPTIBILITY 4-64
4.4. ADJUSTMENTS FOR POTENTIAL INCREASED EARLY-LIFE
SUSCEPTIBILITY 4-65
4.5. INHALATION UNIT RISK ESTIMATES—CONCLUSIONS 4-70
4.6. COMPARISON WITH OTHER PUBLISHED RISK ESTIMATES 4-75
4.6.1. Unit Risk Estimates Based on Human Studies 4-75
4.6.2. Unit Risk Estimates Based on Laboratory Animal Studies 4-80
4.7. RISK ESTIMATES FOR SOME OCCUPATIONAL EXPOSURE
SCENARIOS 4-80
REFERENCES R-l
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 iv DRAFT—DO NOT CITE OR QUOTE
-------
LIST OF TABLES
1-1. Summary of major findings 1-6
3-1. Epidemiological studies of ethylene oxide and human
cancer—lymphohematopoietic cancer results 3-13
3-2. Summary of epidemiological results on EtO and breast cancer (all sterilizer
workers) 3-17
3-3. Tumor incidence data in National Toxicology Program Study of B6C3Fi mice
(NTP, 1987) 3-21
3-4. Tumor incidence data in Lynch et al. (1984a; 1984b) study of male F344 rats 3-22
3-5. Tumor incidence data in Snellings et al. (1984) and Garman et al. (1985) reports
onF344rats 3-24
3-6. Cytogenetic effects in humans 3-33
4-1. Considerations used in this assessment for selecting epidemiology studies for
quantitative risk estimation 4-3
4-2. Cox regression results for all lymphohematopoietic cancer and lymphoid cancer
mortality in both sexes in the NIOSH cohort, for the models presented by
Steenland et al. (2004) 4-5
4-3. Exposure-response modeling results for all lymphohematopoietic cancer and
lymphoid cancer mortality in both sexes in the NIOSH cohort for models not
presented by Steenland et al. (2004) 4-11
4-4. Models considered for modeling the exposure-response data for lymphoid cancer
mortality in both sexes in the NIOSH cohort for the derivation of unit risk
estimates 4-13
4-5. ECoi, LECoi, and unit risk estimates for lymphoid cancer 4-14
4-6. ECoi, LECoi, and unit risk estimates for all lymphohematopoietic cancer 4-18
4-7. Cox regression results for breast cancer mortality in females in the NIOSH cohort,
for models presented in Steenland et al. (2004) 4-21
4-8. Exposure-response modeling results for breast cancer mortality in females in the
NIO SH cohort for model snot presented by Steenland etal. (2004) 4-25
4-9. ECoi, LECoi, and unit risk estimates for breast cancer mortality in females 4-27
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 v DRAFT—DO NOT CITE OR QUOTE
-------
LIST OF TABLES (continued)
4-10. Cox regression results for breast cancer incidence in females from the NIOSH
cohort, for the models presented by Steenland et al. (2003) 4-29
4-11. Exposure-response modeling results for breast cancer incidence in females from
the NIOSH cohort for models not presented by Steenland et al. (2003) 4-34
4-12. Models considered for modeling the exposure-response data for breast cancer
incidence in females in the subcohort with interviews from the NIOSH incidence
study cohort for the derivation of unit risk estimates 4-36
4-13. ECoi, LECoi, and unit risk estimates for breast cancer incidence in females—
invasive and in situ 4-39
4-14. Calculation of ECoi fortotal cancer risk 4-42
4-15. Calculation of total cancer unit risk estimate 4-42
4-16. Upper-bound unit risks (per ug/m3) obtained by combining tumor sites 4-61
4-17. Unit risk values from multistage Weibull time-to-tumor modeling of mouse tumor
incidence in the NTP (1987) study 4-63
4-18. Summary of unit risk estimates (per ug/m3) in animal bioassays 4-64
4-19. ECoi, LECoi, and unit risk estimates for adult-only exposures 4-67
4-20. Calculation of ECoi for total cancer risk from adult-only exposure 4-67
4-21. Calculation of total cancer unit risk estimate from adult-only exposure 4-68
4-22. Adult-based unit risk estimates for use in ADAF calculations and risk estimate
calculations involving less-than-lifetime exposure scenarios 4-69
4-23. Adult-based extra risk estimates per ppm based on adult-exposure-only ECois 4-74
4-24. Summary of key unit risk estimates from this assessment (see Section 4.7 for risk
estimates based on occupational exposure scenarios) 4-76
4-25. Comparison of unit risk estimates 4-78
4-26. Extra risk estimates for lymphoid cancer in both sexes for various occupational
exposure levels 4-83
4-27. Extra risk estimates for breast cancer incidence in females for various
occupational exposure levels 4-86
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 vi DRAFT—DO NOT CITE OR QUOTE
-------
LIST OF FIGURES
3-1. Metabolism of ethylene oxide 3-25
3-2. Simulated blood AUCs for EtO following a 6-hour exposure to EtO from the rat,
mouse, and human PBPK models of Fennell and Brown (2001); based on data
presented in Fennell and Brown (2001) 3-27
3-3. Display of 203 data sets, including bacteria, fungi, plants, insects, and mammals
(in vitro and in vivo), measuring the full range of genotoxic endpoints 3-29
4-1. RR estimate for lymphoid cancer vs. mean exposure (with 15-year lag, unadjusted
for continuous exposure) 4-8
4-2. RR estimate for all lymphohematopoietic cancer vs. mean exposure (with 15-year
lag, unadjusted for continuous exposure) 4-17
4-3. RR estimate for breast cancer mortality vs. mean exposure (with 20-year lag,
unadjusted for continuous exposure) 4-23
4-4. RR estimate for breast cancer incidence in full cohort vs. mean exposure (with 15-
year lag, unadjusted for continuous exposure) 4-31
4-5. RR estimate for breast cancer incidence in subcohort with interviews vs. mean
exposure (with 15-year lag, unadjusted for continuous exposure) 4-32
4-6. RR estimate for breast cancer incidence in subcohort with interviews vs. mean
exposure (with 15-year lag, unadjusted for continuous exposure); select models
compared to deciles 4-38
4-7. RR estimates for lymphoid cancer from occupational EtO exposures (with 15-year
lag) 4-84
4-8. RR estimates for breast cancer incidence from occupational EtO exposures (with
15-year lag) 4-89
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 vii DRAFT—DO NOT CITE OR QUOTE
-------
LIST OF ABBREVIATIONS
ADAF age-dependent adjustment factor
AIC Akaike information criterion
AIDS acquired immune deficiency syndrome
AML acute myeloid leukemia
AUC areas under the curve
BEIR Committee on the Biological Effects of Ionizing Radiation
CI confidence interval
DSB double-strand breaks
EC effective concentration
EOIC Ethylene Oxide Industry Council
EPA U.S. Environmental Protection Agency
EtO ethylene oxide
FRG Federal Republic of Germany
GST glutathione S-transferase
HAP hazardous air pollutants
N7-HEG N7-(2-hydroxyethyl)guanine
IARC International Agency for Research on Cancer
ICD International Classification of Diseases
IRIS Integrated Risk Information System
LEC lower confidence limit
MLE maximum likelihood estimate
NCEA National Center for Environmental Assessment
NHL non-Hodgkin lymphoma
NIOSH National Institute for Occupational Safety and Health
NTP National Toxicology Program
O6-HEG O6-hydroxyethylguanine
OBS observed number
OR odds ratios
PBPK physiologically based pharmacokinetic
POD point of departure
RR relative rate, i.e., rate ratio
SCE sister chromatid exchanges
SE standard error
SEER Surveillance, Epidemiology, and End Results
SIR standardized incidence ratio
SMR standard mortality ratios
TWA time-weighted average
UCC Union Carbide Corporation
UCL upper confidence limit
WHO World Health Organization
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
viii DRAFT—DO NOT CITE OR QUOTE
-------
AUTHORS, CONTRIBUTORS, AND REVIEWERS
ASSESSMENT AUTHORS AND CONTRIBUTORS
Jennifer Jinot (Chemical Manager)
National Center for Environmental Assessment
U.S. Environmental Protection Agency
Washington, DC
David Bayliss (retired)
National Center for Environmental Assessment
U.S. Environmental Protection Agency
Washington, DC
Henry D. Kahn (retired)
National Center for Environmental Assessment
U.S. Environmental Protection Agency
Washington, DC
Nagu Keshava
National Center for Environmental Assessment
U.S. Environmental Protection Agency
Washington, DC
Robert McGaughy (retired)
National Center for Environmental Assessment
U.S. Environmental Protection Agency
Washington, DC
Ravi Subramaniam
National Center for Environmental Assessment
U.S. Environmental Protection Agency
Washington, DC
Larry Valcovic (retired)
National Center for Environmental Assessment
U.S. Environmental Protection Agency
Washington, DC
Suryanarayana Vulimiri
National Center for Environmental Assessment
U.S. Environmental Protection Agency
Washington, DC
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 ix DRAFT—DO NOT CITE OR QUOTE
-------
REVIEWERS
This document has been provided for review to EPA scientists, interagency reviewers
from other federal agencies and White House offices, and the public, and peer reviewed by
independent scientists external to EPA. A summary and EPA's disposition of the comments
received from the independent external peer reviewers and from the public is included in
Appendix H.
INTERNAL EPA REVIEWERS
Michele Burgess, OSWER
Deborah Burgin, OPEI
Kerry Dearfield, ORD/OSP
Joyce Donahue, OW
Rebecca Dzubow, AO/OCHP
Michael Firestone, AO/OCHP
Linnea Hansen, OPP
Karen Hogan, ORD/NCEA-IO
Aparna Koppikar, ORD/NCEA-W
Deirdre Murphy, OAR/ESD
Steve Nesnow, ORD/NHEERL
Marian Olsen, Region 2
Brenda Perkovich-Foos, AO/OCHP
Julian Preston, ORD/NHEERL
Santhini Ramasamy, OPP/HED, OW
Nancy Rios-Jafolla, Region 3
Tracey Woodruff, AO/NCEE
The authors would like to acknowledge Julian Preston, David Bussard, Cheryl Scott, and
Paul White of EPA for their contributions during the draft development process.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 x DRAFT—DO NOT CITE OR QUOTE
-------
EXTERNAL PEER REVIEWERS
SCIENCE ADVISORY BOARD ETHYLENE OXIDE REVIEW PANEL
CHAIR
Dr. Stephen Roberts
University of Florida
OTHER SAB MEMBERS
Dr. Timothy Buckley
Ohio State University
Dr. Montserrat Fuentes
North Carolina State University
Dr. Dale Hattis
Clark University
Dr. James Kehrer
Washington State University
Dr. Mark Miller
California Environmental Protection Agency
Dr. Maria Morandi
University of Texas—Houston Health Science Center
Dr. Robert Schnatter
Exxon Biomedical Sciences, Inc.
Dr. Anne Sweeney
TAMU System Health Science Center
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 xi DRAFT—DO NOT CITE OR QUOTE
-------
CONSULTANTS SERVING ON THE PANEL
Dr. Steven Alan Belinsky
University of New Mexico
Dr. Norman Drinkwater
University of Wisconsin Medical School
Dr. Steven Heeringa
University of Michigan
Dr. Ulrike Luderer
University of California
Dr. James Swenberg
University of North Carolina
Dr. Vernon Walker
Lovelace Respiratory Research Institute
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 xii DRAFT—DO NOT CITE OR QUOTE
-------
1 1. EXECUTIVE SUMMARY
2 Ethylene oxide (EtO) is a gas at room temperature. It is manufactured from ethylene and
3 used primarily as a chemical intermediate in the manufacture of ethylene glycol. It is also used
4 as a sterilizing agent for medical equipment and as a fumigating agent for spices.
5 The DNA-damaging properties of EtO have been studied since the 1940s. EtO is known
6 to be mutagenic in a large number of living organisms, ranging from bacteriophage to mammals,
7 and it also induces chromosome damage. It is carcinogenic in mice and rats, inducing tumors of
8 the lymphohematopoietic system, brain, lung, connective tissue, uterus, and mammary gland. In
9 humans employed in EtO-manufacturing facilities and in sterilizing facilities, the greatest
10 evidence of a cancer risk from exposure is for cancer of the lymphohematopoietic system.
11 Increases in the risk of lymphohematopoietic cancer have been seen in most (but not all) of the
12 epidemiological studies of EtO-exposed workers, manifested as an increase either in leukemia or
13 in cancer of the lymphoid tissue. Of note, in one large epidemiologic study conducted by the
14 National Institute for Occupational Safety and Health (NIOSH) of sterilizer workers that had a
15 well-defined exposure assessment for individuals, positive exposure-response trends were
16 reported for lymphohematopoietic cancer mortality, primarily in males and in particular for
17 lymphoid cancer (i.e., non-Hodgkin lymphoma, myeloma, and lymphocytic leukemia), and for
18 breast cancer mortality in females (Steenland et al., 2004). The positive exposure-response trend
19 for female breast cancer was confirmed in an incidence study based on the same worker cohort
20 (Steenland et al., 2003). There is supporting evidence for an association between EtO and breast
21 cancer from other studies, but the database is more limited than that for lymphohematopoietic
22 cancers.
23 Although the evidence of carcinogenicity from human studies was deemed short of
24 conclusive on its own, EtO is characterized as "carcinogenic to humans" by the inhalation route
25 of exposure based on the total weight of evidence, in accordance with EPA's 2005 Guidelines for
26 Carcinogen Risk Assessment (U.S. EPA, 2005a). The lines of evidence supporting this
27 characterization include: (1) strong, but less than conclusive on its own, epidemiological
28 evidence of lymphohematopoietic cancers and breast cancer in EtO-exposed workers,
29 (2) extensive evidence of carcinogenicity in laboratory animals, including lymphohematopoietic
30 cancers in rats and mice and mammary carcinomas in mice following inhalation exposure,
31 (3) clear evidence that EtO is genotoxic and sufficient weight of evidence to support a mutagenic
32 mode of action for EtO carcinogenicity, and (4) strong evidence that the key precursor events are
33 anticipated to occur in humans and progress to tumors, including evidence of chromosome
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 1-1 DRAFT—DO NOT CITE OR QUOTE
-------
1 damage in humans exposed to EtO. Overall, there is strong confidence in the hazard
2 characterization of EtO as "carcinogenic to humans."
3 This document describes the derivation of inhalation unit risk estimates for cancer
4 mortality and incidence based on the human data from the large NIOSH study (Steenland et al.,
5 2004; 2003). This study was selected for the derivation of risk estimates because it was the
6 largest of the available studies and it had exposure estimates for the individual workers from a
7 high-quality exposure assessment. Multiple modeling approaches were evaluated for the
8 exposure-response data, including modeling the cancer response as a function of either
9 categorical exposures or continuous individual exposure levels. Preferred approaches were
10 defined for each cancer endpoint in consideration of both the statistical properties and biological
11 reasonableness of the resulting model forms (see Tables 4-4 and 4-12 for a summary of models
12 investigated in this assessment for lymphoid cancer and breast cancer incidence, respectively,
13 and the considerations used in model selection).
14 Under the common assumption that relative risk is independent of age, an LECoi (lower
15 95% confidence limit on the ECoi, the estimated effective concentration associated with 1% extra
16 risk) was calculated using a life-table analysis and linear modeling of the categorical Cox
17 regression analysis results for excess lymphoid cancer mortality (Steenland et al., 2004;
18 additional results for both sexes combined provided by Dr. Steenland in Appendix D) excluding
19 the highest exposure group to mitigate the supralinearity of the exposure-response data. Linear
20 low-dose extrapolation below the range of observations is supported by the conclusion that a
21 mutagenic mode of action is operative in EtO carcinogenicity. Linear low-dose extrapolation
22 from the LECoi for lymphoid cancer mortality yielded a lifetime extra cancer unit risk estimate
23 of 2.2 x 1Q~4 per ug/m3 (4.0 x 10~4 per ppb)1 of continuous EtO exposure. Applying the same
24 linear regression coefficient and life-table analysis to background lymphoid cancer incidence
25 rates and applying linear low-dose extrapolation resulted in a preferred lifetime extra lymphoid
26 cancer unit risk estimate of 4.8 x 10~4 per ug/m3 (8.8 x 10~4 per ppb), as cancer incidence
27 estimates are generally preferred over mortality estimates.
28 Using the same approach, a unit risk estimate of 2.8 x 10 4 per ug/m3 (5.1 x 10 4 per ppb)
29 was derived from the breast cancer mortality results of the same epidemiology study (Steenland
30 et al., 2004). Breast cancer incidence risk estimates, on the other hand, were calculated from the
31 data from a breast cancer incidence study of the same occupational cohort (Steenland et al.,
32 2003), and, for these data, a two-piece linear spline model was used for the exposure-response
33 modeling. Using the same life-table approach and linear low-dose extrapolation, a unit risk
34 estimate of 9.5 x io~4 per ug/m3 (1.7 x io~3 per ppb) was obtained for breast cancer incidence.
Conversion equation: 1 ppm= 1830 ug/m .
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 1-2 DRAFT—DO NOT CITE OR QUOTE
-------
1 Again, the incidence estimate is preferred over the mortality estimate. Combining the incidence
2 risk estimates for the two cancer types resulted in a total cancer unit risk estimate of 1.2 * 1CT3
3 per ug/m3 (2.3 x 1CT3 per ppb).
4 Unit risk estimates were also derived from the three chronic rodent bioassays for EtO
5 reported in the literature. These estimates, ranging from 2.2 x 1CT5 per ug/m3 to 4.6 x 1CT5 per
6 ug/m3, are over an order of magnitude lower than the estimates based on human data. The
7 Agency takes the position that human data, if adequate data are available, provide a more
8 appropriate basis than rodent data for estimating population risks (U.S. EPA, 2005a), primarily
9 because uncertainties in extrapolating quantitative risks from rodents to humans are avoided.
10 Although there is a sizeable difference between the rodent-based and the human-based estimates,
11 the human data are from a large, high-quality study, with EtO exposure estimates for the
12 individual workers and little reported exposure to chemicals other than EtO. Therefore, the
13 estimates based on the human data are the preferred estimates for this assessment.
14 Because the weight of evidence supports a mutagenic mode of action for EtO
15 carcinogenicity, and as there are no chemical-specific data from which to assess early-life
16 susceptibility, increased early-life susceptibility should be assumed, according to EPA's
17 Supplemental Guidance for Assessing Susceptibility from Early-Life Exposure to
18 Carcinogens-hereinafter referred to as "EPA's Supplemental Guidance" (U.S. EPA, 2005b).
19 This mode-of-action-based assumption of increased early-life susceptibility supersedes the
20 assumption of age independence under which the human-data-based estimates presented above
21 were derived. Thus, using the same approach as for the estimates discussed above but initiating
22 exposure in the life-table analysis at age 16 instead of at birth, adult-exposure-only unit risk
23 estimates were calculated from the human data under an alternate assumption that relative risk is
24 independent of age for adults, which represent the life stage for which the data upon which the
25 exposure-response modeling was conducted pertain. These adult-exposure-only unit risk
26 estimates were then rescaled to a 70-year basis for use in the standard ADAF calculations and
27 risk estimate calculations involving less-than-lifetime exposure scenarios. The resulting
28 adult-based unit risk estimates were 4.35 x 10 4 per ug/m3 (7.95 x 10 4 per ppb) for lymphoid
29 cancer incidence, 8.21 x 10~4 per ug/m3 (1.50 x 10~3 per ppb) for breast cancer incidence in
30 females, and 1.08 x 10 3 per ug/m3 (1.98 x 10 3 per ppb) for both cancer types combined. For
31 exposure scenarios involving early-life exposure, the age-dependent adjustment factors (ADAFs)
32 should be applied, in accordance with EPA's Supplemental Guidance (U.S. EPA, 2005b).
33 Applying the ADAFs to obtain a full lifetime total cancer unit risk estimate yields 1.8 x 10 3 per
34 ug/m3 (3.3 x 10~3 per ppb), and the commensurate lifetime chronic (lower-bound) exposure level
35 of EtO corresponding to an increased cancer risk of 10 6 is 0.0006 ug/m3 (0.0003 ppb).
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 1-3 DRAFT—DO NOT CITE OR QUOTE
-------
1 The major sources of uncertainty in the unit risk estimates derived from the human data
2 include the low-dose extrapolation, the retrospective exposure assessment conducted for the
3 epidemiology study, and the exposure-response modeling of the epidemiological data (see
4 Section 4.1.4 for a discussion of these and other sources of uncertainty in the unit risk estimates).
5 Although there are uncertainties in the unit risk estimate, confidence in the estimate is
6 relatively high. First, there is strong confidence in the hazard characterization of EtO as
7 "carcinogenic to humans," which is based on strong epidemiological evidence supplemented by
8 other lines of evidence. Second, the unit risk estimate is based on human data from a large,
9 high-quality epidemiology study with individual worker exposures estimated using a
10 high-quality regression model. Finally, the use of low-exposure linear extrapolation is strongly
11 supported by the conclusion that EtO carcinogenicity has a mutagenic mode of action.
12 Confidence in the unit risk estimate is particularly high for the breast cancer component,
13 the largest contributor to the total cancer unit risk estimate, which is based on over 200 incident
14 cases for which the investigators had information on other potential breast cancer risk factors.
15 The selected model for the breast cancer incidence data was the best-fitting model of the models
16 investigated as well as the model which provided the best representation of the categorical
17 results, particularly in the lower exposure range of greatest relevance for the derivation of a unit
18 risk estimate. Alternate estimates calculated from other reasonable models suggest that a unit
19 risk estimate for breast cancer incidence fourfold lower (corresponding to a total cancer unit risk
20 estimate of twofold lower) is plausible; however, unit risk estimates notably lower than that are
21 considered unlikely from the available data.
22 There is lower confidence in the lymphoid cancer component of the unit risk estimate
23 because it is based on fewer events (40 lymphoid cancer deaths); incidence risk was estimated
24 from mortality data; and the exposure-response relationship is exceedingly supralinear, such that
25 continuous models yield apparently implausibly steep low-exposure slopes. Although these
26 continuous models provided statistically significant slope coefficients, there was low confidence
27 in such steep slopes, which, particularly for the two-piece spline models, are highly dependent on
28 a small number of cases in the low-exposure range. Thus, a linear regression model of the
29 categorical results for the lowest three quartiles was used to derive the unit risk estimate for
30 lymphoid cancer, and there was greater confidence in the more moderate slope resulting from
31 that model, although it was not statistically significant, because it was based on more data and
32 provided a good representation of the categorical results across this larger data range in the
33 lower-exposure region. So, while there is lower confidence in the lymphoid cancer unit risk
34 estimate than in the breast cancer unit risk estimate, the lymphoid cancer estimate is considered a
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 1-4 DRAFT—DO NOT CITE OR QUOTE
-------
1 reasonable estimate from the available data, and overall, there is relatively high confidence in the
2 total cancer unit risk estimate.
3 The unit risk estimate is intended to provide a reasonable upper bound on cancer risk.
4 The estimate was developed for environmental exposure levels (it is considered valid for
5 exposures up to 140 ug/m3 [75 ppb]) and is not applicable to higher-level exposures, such as may
6 occur occupationally, which appear to have a different exposure-response relationship.
7 However, occupational exposure levels of EtO are of concern to EPA when EtO is used as a
8 pesticide (e.g., sterilizing agent or fumigant). Therefore, this document also presents extra risk
9 estimates for the two cancer types for a range of occupational exposure scenarios (see
10 Section 4.7). Maximum likelihood estimates of the extra risk of lymphoid cancer and breast
11 cancer combined for the range of occupational exposure scenarios considered (i.e., 0.1 to 1 ppm
12 8-hour TWA for 35 years) ranged from 0.047 to 0.14. The overall uncertainty associated with
13 the extra risk estimates for occupational exposure scenarios is less than that associated with the
14 unit risk estimates for environmental exposures. The extra risk estimates are derived for
15 occupational exposure scenarios that yield cumulative exposures well within the range of the
16 exposures in the NIOSH study. Moreover, the NIOSH study is a study of sterilizer workers who
17 used EtO for the sterilization of medical supplies or spices (Steenland et al., 1991); thus, the
18 results are directly applicable to workers in these occupations, and these are among the
19 occupations of primary concern for current occupational EtO exposures.
20 Table 1-1 provides a summary of the major findings in this assessment.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 1-5 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
Table 1-1. Summary of major findings
Hazard Conclusions
Hazard Characterization
Mode of Action
The weight of evidence from epidemiological
studies and supporting information is sufficient to
conclude that ethylene oxide is carcinogenic to
humans.
The weight of evidence is sufficient to conclude
that ethylene oxide carcinogenicity has a
mutagenic mode of action.
Unit Risk Estimates (for environmental exposures)3
Basis
Inhalation unit risk estimate"
(per jig/m3)b
Full lifetime unit risk estimate0
Total cancer risk based on human data — lymphoid cancer
incidence and breast cancer incidence in females
1.80 x 10 3
Adult-based unit risk estimates'1
Total cancer risk based on human data — lymphoid cancer
incidence and breast cancer incidence in females
Lymphoid cancer incidence in both sexes based on human data
Breast cancer incidence in females based on human data
Total cancer risk based on human data — lymphoid cancer
incidence and range of female breast cancer incidence
estimates from three alternate models
Total cancer incidence risk estimate from rodent data (female
mouse)
1.08 x 10 3
4.35 x 1Q-4
8.21 x ID'4
5.64 x I(r4-1.08x ID'3
4.6 x 1(T5
Extra risk estimates for occupational exposure scenarios (see Section 4.7)
Maximum likelihood estimates of the extra risk of lymphoid
cancer and breast cancer combined for the range of
occupational exposure scenarios considered (i.e., 0.1 to 1 ppm
8-hr TWA for 3 Syr)
0.047-0.14
3
4
5
6
7
8
9
10
11
12
13
14
15
16
aThese unit risk estimates are not intended for use with continuous lifetime exposure levels above about 140 ug/m .
See Section 4.7 for risk estimates based on occupational exposure scenarios. Preferred estimates are in bold.
bTo convert unit risk estimates to (ppm)"1, multiply the (ug/m3)"1 estimates by 1,830 (ug/m3)/ppm.
'Because the weight of evidence supports a mutagenic mode of action for EtO carcinogenicity, and because of the
lack of chemical-specific data, EPA assumes increased early-life susceptibility and recommends the application of
ADAFs, in accordance with EPA's Supplemental Guidance (U.S. EPA, 2005b), for exposure scenarios that include
early-life exposures. For the full lifetime (upper bound) unit risk estimate presented here, ADAFs have been
applied, as described in Section 4.4.
dThese (upper bound) unit risk estimates are intended for use in AD AF calculations and less-than-lifetime adult
exposure scenarios (U.S. EPA, 2005b). Note that these are not the same as the unit risk estimates derived directly
from the human data in Section 4.1 under the assumption that RRs are independent of age. Under that assumption,
the key unit risk estimates were 4.8 x 10~4 per ug/m3 for lymphoid cancer incidence, 9.5 x 10~4 per ug/m3 for breast
cancer incidence, and 1.2 x 10~3 per ug/m3 for the combined cancer incidence risk from those two cancers. See
Section 4.4 for the derivation of the adult-based unit risk estimates.
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
1 -6 DRAFT—DO NOT CITE OR QUOTE
-------
1 2. INTRODUCTION
2 The purpose of this document is to provide scientific support and rationale for the hazard
3 and dose-response assessment in IRIS pertaining to carcinogenicity from chronic inhalation
4 exposure to ethylene oxide (EtO) (CASRN 75-21-8). It is not intended to be a comprehensive
5 treatise on the chemical or toxicological nature of EtO. In general, this IRIS Carcinogenicity
6 Assessment provides information on the carcinogenic hazard potential of EtO and quantitative
7 estimates of risk from inhalation exposure. The information includes a weight-of-evidence
8 judgment of the likelihood that the agent is a human carcinogen and the conditions under which
9 the carcinogenic effects may be expressed. Quantitative risk estimates for inhalation exposure
10 (inhalation unit risks) are derived. The definition of an inhalation unit risk is a plausible upper
11 bound on the estimate of risk per ug/m3 air breathed.
12 Development of the hazard identification and dose-response assessments for EtO has
13 followed the general guidelines for risk assessment as set forth by the National Research Council
14 (NRC, 1983). United States Environmental Protection Agency (U.S. EPA) Guidelines and Risk
15 Assessment Forum Technical Panel Reports that were used in the development of this
16 assessment include the following: Guidelines for Mutagenicity Risk Assessment (U.S. EPA,
17 1986), Methods for Derivation of Inhalation Reference Concentrations and Application of
18 Inhalation Dosimetry (U.S. EPA, 1994), Benchmark Dose Technical Guidance (U.S. EPA,
19 2012), Science Policy Council Handbook: Risk Characterization^. S. EPA, 2000), Guidelines
20 for Carcinogen Risk Assessment (U.S. EPA, 2005a), Supplemental Guidance for Assessing
21 Susceptibility from Early-Life Exposure to Carcinogens (U.S. EPA, 2005b), and Science Policy
22 Council Handbook: Peer Review (U.S. EPA, 2006b).
23 An earlier external review draft of this carcinogenicity assessment (U.S. EPA, 2006a)
24 was peer reviewed by a panel of EPA's Science Advisory Board (SAB) in 2007 (SAB, 2007).
25 This revised external review draft is being released for public comment and for additional
26 external peer review to receive comments primarily on the expanded exposure-response
27 modeling of certain epidemiologic data done in response to comments from the 2007 SAB
28 review.
29 The literature search strategy first employed for this assessment was based on the
30 Chemical Abstracts Service Registry Number (CASRN) and at least one common name. Any
31 pertinent scientific information submitted by the public to the IRIS Submission Desk was also
32 considered in the development of this document. References have been added after the first
33 external peer review in response to the reviewers' and public comments. References have also
34 been added for completeness. These references have not changed the overall qualitative or
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 2-1 DRAFT—DO NOT CITE OR QUOTE
-------
1 quantitative conclusions. See Appendix I for a list of the references added after the first external
2 peer review. The cutoff date for literature inclusion into this carcinogenicity assessment was
3 30 June 2010. A systematic literature search was conducted for the time frame from January
4 2006 to May 2013 to ensure that no major studies were missed from the time of the first external
5 review draft in 2006 until the cutoff date and to determine if any significant new studies had
6 been published since the cutoff date that might alter the findings of the assessment. This
7 systematic literature search is described in Appendix J. No new studies were identified that
8 would impact the assessment's major conclusions. Nonetheless, two new studies of high
9 pertinence to the assessment have been published since the cutoff date for literature inclusion,
10 and these studies are reviewed briefly in Appendix J for transparency and completeness. The
11 references considered and cited in this document, including abstracts, can be found on the Health and
12 Environmental Research Online (HERO) website.2
13 On 23 December 2011, the Consolidated Appropriations Act, 2012, was signed into law.3
14 The report language included direction to EPA for the IRIS Program related to recommendations
15 provided by the National Research Council (NRC) in their review of EPA's draft IRIS
16 assessment of formaldehyde. The NRC's recommendations, provided in Chapter 7 of their
17 review report, offered suggestions to EPA for improving the development of IRIS assessments.
18 The report language included the following:
19
20
21 The Agency shall incorporate, as appropriate, based on chemical-specific datasets
22 and biological effects, the recommendations of Chapter 7 of the National
23 Research Council's Review of the Environmental Protection Agency's Draft IRIS
24 Assessment of Formaldehyde into the IRIS process ... For draft assessments
25 released in fiscal year 2012, the Agency shall include documentation describing
26 how the Chapter 7 recommendations of the National Academy of Sciences (NAS)
27 have been implemented or addressed, including an explanation for why certain
28 recommendations were not incorporated.
29
30
31 Consistent with the direction provided by Congress, documentation of how the
32 recommendations from Chapter 7 of the NRC report have been implemented in this assessment
HERO is a database of scientific studies and other references used to develop EPA's risk assessments, which are
aimed at understanding the health and environmental effects of pollutants and chemicals. HERO is developed and
managed in EPA's Office of Research and Development (ORD) by the National Center for Environmental
Assessment (NCEA). The database includes more than 750,000 scientific articles from the peer-reviewed literature.
New studies are added continuously to HERO.
3Pub. L. No. 112-74, Consolidated Appropriations Act, 2012.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 2-2 DRAFT—DO NOT CITE OR QUOTE
-------
1 is provided in Appendix K. This documentation also includes an explanation for why certain
2 recommendations were not incorporated.
3 For general information about this assessment or other questions relating to IRIS, the
4 reader is referred to EPA's IRIS Hotline at (202) 566-1676 (phone), (202) 566-1749 (fax), or
5 hotline.iris@epa.gov (email address).
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 2-3 DRAFT—DO NOT CITE OR QUOTE
-------
1 3. HAZARD IDENTIFICATION
2 This chapter presents the evidence considered in the hazard identification of EtO
3 carcinogenicity and the hazard characterization resulting from the weight-of-evidence evaluation.
4 Section 3.1 summarizes the human evidence (a more detailed discussion of the human cancer
5 studies is presented in Appendix A). Section 3.2 describes the evidence from experimental
6 animal studies. Section 3.3 discusses supporting evidence, in particular evidence regarding the
7 genotoxicity of EtO. Section 3.4 provides the mode-of-action analysis for EtO carcinogenicity.
8 To conclude the chapter, Section 3.5 presents the hazard characterization for EtO carcinogenicity
9 and a discussion of life stages and populations with potentially increased susceptibility.
10
11 3.1. EVIDENCE OF CANCER IN HUMANS
12 The literature from 1988 to present contains numerous epidemiological studies of the
13 carcinogenic effects of EtO in occupational cohorts; some of these cohorts were the subject of
14 multiple reports. The conclusions about the human evidence of carcinogenicity in this
15 assessment are based on the following summary of those studies, which are discussed in more
16 detail and critically reviewed in Appendix A. Table A-5 in Appendix A provides a tabular
17 summary of the epidemiological studies, including some study details, results, and limitations.
18 The strengths and weaknesses of these studies were evaluated individually using standard
19 considerations in evaluating epidemiological studies. The major areas of concern are study
20 design, exposure assessment, and data analysis. General features of study design considered
21 include sample size and assessment of the health endpoint. For case-control studies, design
22 considerations include representativeness of cases, selection of controls, participation rates, use
23 of proxy respondents, and interview approach (e.g., blinding). For cohort studies, design
24 considerations include selection of referent population (e.g., internal comparisons are generally
25 preferred to comparisons with an external population), loss to follow-up, and length of
26 follow-up. Exposure assessment issues include specificity of exposure (exposure
27 misclassification), characterization of exposure (e.g., ever exposed or quantitative estimate of
28 exposure level), and potential confounders. Analysis considerations include adjustment for
29 potential confounders or effect modifiers and modeling of exposure-response relationships.
30 Two primary sources of exposures to EtO are production facilities and sterilization
31 operations. There are two types of production facilities (IARC, 1994b):
32
33
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-1 DRAFT—DO NOT CITE OR QUOTE
-------
1 1. Those using the older chlorohydrin process, where ethylene is reacted with hypochlorous
2 acid and then with calcium oxide to make EtO (this method produces unwanted
3 byproducts, the most toxic of which is ethylene dichloride), and
4 2. Those producing EtO via direct oxidation of ethylene in a pressurized vessel, which
5 involves less EtO exposure and eliminates the chemical byproducts of the chlorohydrin
6 process.
7
8
9 Exposure in the sterilization of medical equipment and in the direct oxidation process is
10 predominantly to EtO, whereas exposure in the chlorohydrin process is to EtO mixed with other
11 chemicals.
12 Hogstedt et al. (1986) and Hogstedt (1988) summarized findings of three Swedish
13 occupational cohorts (539 men and 170 women) exposed in a plant where hospital equipment
14 was sterilized, in a chlorohydrin production facility, and in a direct oxidation production facility.
15 The incidence of leukemia was elevated in all cohorts, although the risk was not statistically
16 significant in the cohort from the direct oxidation facility. For the three cohorts combined there
17 were statistically significantly elevated standard mortality ratios (SMRs) for leukemia
18 (SMR = 9.2; 95% confidence interval [CI] = 3.7-19), based on 7 deaths, and for stomach cancer
19 (SMR = 5.5; 95% CI = 2.6-10), based on 10 deaths. Although this study produced high SMRs
20 for leukemia, stomach cancer, and total cancer, there are some limitations, such as multiple
21 exposures to numerous other chemicals, lack of personal exposure information, and lack of
22 latency analysis. No gender differences were separately analyzed. No dose-response
23 calculations were possible. This study provides suggestive evidence of the carcinogenicity of
24 EtO.
25 Coggon et al. (2004) reported the results of a follow-up study of a cohort originally
26 studied by Gardner et al. (1989). The cohort included workers in three EtO production facilities
27 (two using both chlorohydrin and direct oxidation processes and the third using direct oxidation
28 only); in a fourth facility that used EtO in the manufacture of other chemicals; and in eight
29 hospitals that used EtO in sterilizing units. The total cohort comprised 1,864 men and
30 1,012 women. No statistically significant excesses were observed for any cancer site. Slight
31 increases, based on small numbers, were observed for the various lymphohematopoietic cancers:
32 Hodgkin lymphoma (2 vs. 1 expected), non-Hodgkin lymphoma (NHL) (7 vs. 4.8), multiple
33 myeloma (3 vs. 2.5), and leukemia (5 vs. 4.6). The increases were concentrated in the
34 1,471 chemical-manufacturing workers, of whom all but 1 were male. In the
35 chemical-manufacturing workers with "definite" exposure, 4 leukemias were observed
36 (1.7 expected) and 9 lymphohematopoietic cancers were observed (4.9 expected). A slight
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-2 DRAFT—DO NOT CITE OR QUOTE
-------
1 deficit in the risk of breast cancer deaths (11 vs. 13.2) was observed in the cohort. No individual
2 exposure measurements were obtained from cohort members, and no exposure measurements
3 were available before 1977. Multiple exposures to other chemicals, small numbers of deaths,
4 and lack of individual EtO measurements make this study only suggestive of a higher risk of
5 leukemia from exposure to EtO.
6 A series of retrospective mortality studies of about 2,000 male workers who were
7 assigned to operations that used or produced EtO in either of two Union Carbide Corporation
8 (UCC) chemical production facilities in West Virginia (Valdez-Flores et al., 2010; Swaen et al.,
9 2009; Teta et al., 1999; Benson and Teta, 1993; Teta et al., 1993; Greenberg et al., 1990) have
10 been published. EtO was produced at these facilities until 1971, after which it was imported to
11 the facilities. For EtO production, the chlorohydrin process was used from 1925 to 1957, and the
12 direct oxidation process was used from 1937 to 1971 (during overlapping years, both processes
13 were in use). The cohort was observed from 1940 through 1978 in the original study (Greenberg
14 et al., 1990), through 1988 in the Teta et al. (1999); Teta et al. (1993) and Benson and Teta
15 (1993) studies, and through 2003 in the latter two studies. A large-scale industrial hygiene
16 survey and monitoring of EtO concentrations was carried out in 1976, at which time EtO was in
17 use at the facilities but no longer in production.
18 Greenberg et al. (1990) found elevated but not statistically significant risks of pancreatic
19 cancer (SMR = 1.7) and leukemia (SMR = 2.3) (each based on seven cases) in the entire cohort;
20 most of the cases occurred in the chlorohydrin production unit (note that the chlorohydrin
21 production unit produced primarily ethylene chlorohydrin, which is used in chlorohydrin-based
22 EtO production, but this unit is not where chlorohydrin-based EtO production took place).
23 Limitations of this study included multiple exposures to many different chemicals in the facility
24 through the years and lack of EtO exposure measurements prior to 1976. Three categories of
25 exposure were established for analysis—low, intermediate, and high—based on a qualitative
26 characterization of the potential for EtO exposure. The number of workers in each exposure
27 category was not reported. No significant findings of a dose-response relationship were
28 discernible. No quantitative estimates of individual exposure were made in this study, and no
29 latency analysis was conducted (average follow-up was 20 years). Furthermore, EtO is not the
30 only chemical to which the observed excesses in cancer mortality could be attributed.
31 A follow-up study (Teta et al., 1993) that extended the observation of this cohort
32 (excluding the 278 chlorohydrin production unit workers, who reportedly had low EtO
33 exposures) for an additional 10 years to 1988 found no significant risk of total cancer; there was
34 a slight trend in the risk of leukemia with increasing duration of assignment to departments using
35 or processing EtO, but it was not significant (p = 0.28) and was based on only five cases. The
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-3 DRAFT—DO NOT CITE OR QUOTE
-------
1 average follow-up was 27 years, and at least 10 years had elapsed since first exposure for all
2 workers. The same problems of exposure ascertainment exist for this study as for that of
3 Greenberg et al. (1990), and furthermore, the follow-up did not update work histories for the
4 workers after 1978. EtO production at the plants was discontinued before 1978, as noted by Teta
5 et al. (1993); however, according to Greenberg et al. (1990), certain nonproduction areas had
6 "intermediate" potential for EtO exposure, although estimates of exposure levels suggest that the
7 levels would also be lower during the update period [<1 ppm 8-hour time-weighted average
8 (TWA), according to Teta et al. (1993)]. It appears from the Greenberg et al. (1990) publication
9 that the high potential exposure group was reserved for EtO production workers, and according
10 to Teta et al. (1993), there were only 425 EtO production workers in the cohort. Of these, only
11 118 worked in the chlorohydrin-based production process, where exposures were reportedly
12 highest. Essentially, the study did not support the earlier studies of cancer in EtO workers;
13 however, it was limited by low statistical power and a crude exposure assessment and, thus, is
14 not very informative regarding whether exposure to EtO is causally related to cancer.
15 In a parallel follow-up study through 1988 of only the chlorohydrin production
16 employees, Benson and Teta (1993) found that pancreatic cancer and lymphohematopoietic
17 cancer cases continued to accumulate and that the SMRs were statistically significant for
18 pancreatic cancer (SMR = 4.9; Obs = 8,/? < 0.05) and for lymphohematopoietic cancer
19 (SMR = 2.9; Obs = S,p< 0.05). These investigators interpreted these excesses as possibly due
20 to ethylene dichloride, a byproduct in the chlorohydrin process. Again, this small study of only
21 278 workers was limited by the same problems as the Greenberg et al. (1990) study and the Teta
22 et al. (1993) study. No individual estimates of exposure are available and the workers were
23 potentially exposed to many different chemicals (see Table A-5 in Appendix A). Furthermore,
24 the chlorohydrin production unit was reportedly considered a low potential EtO exposure
25 department. Hence this study has little weight in determining the carcinogenicity of EtO.
26 In a later analysis, Teta et al. (1999) fitted Poisson regression dose-response models to
27 the UCC data (followed through 1988 and excluding the chlorohydrin production workers) and
28 to data (followed through 1987) from a study by the National Institute for Occupational Safety
29 and Health (NIOSH) (described below). Because Teta et al. (1999) did not present risk ratios for
30 the cumulative exposure categories used to model the dose-response relationships, the only
31 comparison that can be made between the UCC and NIOSH data is based on the fitted models.
32 These models are almost identical for leukemia, but for the lymphoid category, the
33 risk—according to the fitted model for the UCC data—decreased as a function of exposure,
34 whereas the risk for the modeled NIOSH data increased as a function of exposure. However, the
35 models are based on small numbers of cases (16 [5 UCC, 11 NIOSH] for leukemia; 22 [3 UCC,
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-4 DRAFT—DO NOT CITE OR QUOTE
-------
1 19 NIOSH] for lymphoid cancers), and no statistics are provided to assess model goodness-of-fit
2 or to compare across models. In any event, this analysis is superseded by the more recent
3 analysis by the same authors (Valdez-Flores et al., 2010) of the results of more recent follow-up
4 studies of these cohorts (see below).
5 Swaen et al. (2009) studied the same UCC cohort identified by Teta et al. (1993), i.e.,
6 without the chlorohydrin production workers, but extended the cohort enumeration period from
7 the end of 1978 to the end of 1988, identifying 167 additional workers, and conducted mortality
8 follow-up of the resulting cohort of 2,063 male workers through 2003. Work histories were also
9 extended through 1988 (exposures after 1988 were considered negligible compared to earlier
10 exposure levels). Swaen et al. (2009) used an exposure assessment based on the qualitative
11 categorizations of potential EtO exposure in the different departments developed by Greenberg et
12 al. (1990) and time-period exposure estimates from Teta et al. (1993). This exposure assessment
13 was relatively crude, based on just a small number of department-specific and time-period-
14 specific categories, and with exposure estimates for only a few of the categories derived from
15 actual measurements (see Appendix A.2.20 for details).
16 At the end of the 2003 follow-up, 1,048 of the 2,063 workers had died (Swaen et al.,
17 2009). The all-cause mortality SMR was 0.85 (95% CI = 0.80, 0.90) and the cancer SMR was
18 0.95 (95% CI = 0.84, 1.06). None of the SMRs for specific cancer types showed any statistically
19 significant increases. In analyses stratified by hire date [pre- (inclusive) or post-1956], the SMR
20 for leukemia was elevated but not statistically significant (1.51; 95% CI 0.69, 2.87) in the
21 early-hire group, based on nine deaths. In analyses stratified by duration of employment, no
22 trends were apparent for any of the lymphohematopoietic cancers, although in the 9+ years of
23 employment subgroup, the SMR for NHL was nonsignificantly increased (1.49; 95% CI 0.48,
24 3.48), based on five deaths. In SMR analyses stratified by cumulative exposure, no trends were
25 apparent for any of the lymphohematopoietic cancers and there were no notable elevations for
26 the highest cumulative exposure category. Note that only 27 lymphohematopoietic cancer deaths
27 (including 12 leukemias and 11 NHLs) were observed in the cohort.
28 Swaen et al. (2009) also did internal Cox proportional hazards modeling for some disease
29 categories (all-cause mortality, leukemia mortality, and lymphoid cancer [NHL, lymphocytic
30 leukemia, and myeloma] mortality [17 deaths]), using cumulative exposure as the exposure
31 metric. These analyses showed no evidence of an exposure-response relationship. Alternate
32 Cox proportional hazard analyses and categorical exposure-response analyses of the UCC data
33 conducted by Valdez-Flores et al. (2010) for a larger set of cancer endpoints similarly reported
34 an absence of any exposure-response relationships. Each of these cancer analyses, however,
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-5 DRAFT—DO NOT CITE OR QUOTE
-------
1 relies on small numbers of cases and a crude exposure assessment, where there is a high potential
2 for exposure misclassification.
3 In a study of 2,658 male workers at eight chemical plants where EtO is produced
4 (manufacturing process not stated), Kiesselbach et al. (1990) found slightly increased SMRs for
5 cancers of the stomach, esophagus, and lung. A latency analysis was done only for stomach
6 cancer and total mortality. The investigators considered 71.6% of the cohort to be "weakly"
7 exposed; only 2.6% were "strongly exposed." No data were provided to explain how these
8 exposure categories were derived. The workers were followed for a median 15.5 years. Without
9 additional information on exposure to EtO, this study is of little help at this time in evaluating the
10 carcinogenicity of EtO.
11 NIOSH conducted an industry-wide study of 18,254 workers (45% male and 55%
12 female) in 14 plants where EtO was used (Steenland et al., 2004; Stayner et al., 1993; Steenland
13 et al., 1991). Most of the workers were exposed while sterilizing medical supplies and treating
14 spices and in the manufacture and testing of medical sterilizers. Individual exposure estimates
15 were derived for workers from 13 of the 14 plants. The procedures for selecting the facilities and
16 defining the cohort are described in Steenland et al. (1991), and the exposure model and
17 verification procedures are described in Greife et al. (1988) and Hornung et al. (1994). Briefly, a
18 regression model was developed, allowing the estimation of exposure levels for time periods,
19 facilities, and operations for which industrial hygiene data were unavailable. The data for the
20 model consisted of 2,700 individual time-weighted exposure values for workers' personal
21 breathing zones, acquired from 18 facilities between 1976 and 1985. The data were divided into
22 two sets, one for developing the regression model and the second for testing it. Seven out of
23 23 independent variables tested for inclusion in the regression model were found to be significant
24 predictors of EtO exposure and were included in the final model. This model predicted 85% of
25 the variation in average EtO exposure levels. (See Appendix A, Section A.2.8, for more details
26 on the NIOSH exposure assessment and its evaluation.) Results of the original follow-up study
27 through 1987 are presented in Steenland et al. (1991) and Stayner et al. (1993). The cohort
28 averaged 26.8 years of follow-up in the extended follow-up study through 1998, and 16% of the
29 cohort had died (Steenland et al., 2004).
30 The overall SMR for cancer was 0.98, based on 860 deaths (Steenland et al., 2004). The
31 SMR for (lympho)hematopoietic cancer was 1.00, based on 79 cases. Exposure-response
32 analyses, however, revealed exposure-related increases in hematopoietic cancer mortality risk,
33 although the effect was primarily in males, when analyzed by sex. In categorical life-table
34 analysis, men with >13,500 ppm-days of cumulative exposure had an SMR of 1.46 (Obs = 13).
35 In internal Cox regression analyses (i.e., analyses in which the referent population is within the
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-6 DRAFT—DO NOT CITE OR QUOTE
-------
1 cohort) with exposure as a continuous variable, statistically significant trends in males for all
2 hematopoietic cancer (p = 0.02) and for "lymphoid" cancers (NHL, lymphocytic leukemia, and
3 myeloma;/' = 0.02) were observed using log cumulative exposure (ppm-days) with a 15-year
4 lag.4 In internal categorical analyses, statistically significant odds ratios (ORs) were observed in
5 the highest cumulative exposure quartile (with a 15-year lag) in males for all hematopoietic
6 cancer (OR = 3.42; 95% CI = 1.09-10.73) and "lymphoid" cancer (OR = 3.76; 95%
7 CI = 1.03-13.64). The exposure metrics of duration of exposure, average concentration, and
8 maximum (8-hour TWA) concentration did not predict the hematopoietic cancer results as well
9 as did the cumulative exposure metric.
10 Although the overall SMR for female breast cancer was 0.99, based on 102 deaths, the
11 NIOSH mortality follow-up study reported a significant excess of breast cancer mortality in the
12 highest cumulative exposure quartile using a 20-year lag period compared to the U.S. population
13 (SMR = 2.07; 95% CI = 1.10-3.54; Obs = 13). Internal exposure-response analyses also noted a
14 significant positive trend for breast cancer mortality using the log of cumulative exposure and a
15 20-year lag time (p = 0.01). In internal categorical analyses, a statistically significant OR for
16 breast cancer mortality was observed in the highest cumulative exposure quartile with a 20-year
17 lag (OR = 3.13; 95% CI = 1.42-6.92).
18 In summary, although the overall external comparisons did not demonstrate increased
19 risks, the NIOSH investigators found significant internal exposure-response relationships
20 between exposure to EtO and cancers of the hematopoietic system, as well as breast cancer
21 mortality. (Internal comparisons are considered superior to external comparisons in occupational
22 epidemiology studies because internal comparisons help control for the healthy worker effect and
23 other factors that might be more comparable within a study's worker population than between
24 the workers and the general population.) Exposures to other chemicals in the workplace were
25 believed to be minimal or nonexistent. This study is the most useful of the epidemiologic studies
26 in terms of carrying out a quantitative dose-response assessment. It possesses more attributes
27 than the others for performing risk analysis (e.g., good-quality estimates of individual exposure,
28 lack of exposure to other chemicals, and a large and diverse cohort of workers).
29 It should be noted that Steenland et al. (2004) used Cox regression models, which are
30 log-linear relative rate models, thus providing some low-dose sublinear curvature for doses
31 expressed in terms of cumulative exposure. However, the best-fitting dose-response model for
32 both male lymphoid cancers and male all hematopoietic cancers was for dose expressed in terms
33 of log cumulative exposure, indicating supralinearity of the low-dose data. Supralinearity of the
4The sex difference is not statistically significant, however, and the trends for both sexes combined are also
statistically significant [p = 0.01 andp = 0.02, respectively; see Tables D-3e and D-4e in Appendix D].
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-7 DRAFT—DO NOT CITE OR QUOTE
-------
1 dose-response data was also indicated by the categorical exposure results. This is in contrast to
2 the reported results of Kirman et al. (2004) based on the Teta et al. (1999) analysis combining
3 the 1993 UCC leukemia data with the 1993 NIOSH leukemia data, which are claimed by the
4 authors to provide empirical evidence supporting a quadratic dose-response relationship. The
5 2004 NIOSH dose-response data for hematopoietic cancers clearly do not provide empirical
6 evidence in support of a quadratic dose-response relationship. On the contrary, the NIOSH data
7 suggest a supralinear dose-response relationship in the observable range.
8 Wong and Trent (1993) investigated the same cohort as Steenland et al. (1991) but added
9 474 new unexplained subjects and increased the follow-up period by 1 year. They incremented
10 the total number of deaths by 176 and added 392.2 more expected deaths. The only positive
11 finding was a statistically significantly increased risk of NHL among men (SMR = 2.5; Obs = 6;
12 p< 0.05). However, there was a deficit risk of NHL among women. For breast cancer, there
13 was no trend of increasing risk by duration of employment or by latency. This study has maj or
14 limitations, not the least of which is a lack of detailed employment histories, making it
15 impossible to quantify individual exposures and develop dose-response relationships.
16 Furthermore, the addition of more than twice as many expected deaths as observed deaths makes
17 the analysis by the authors questionable.
18 Valdez-Flores et al. (2010) conducted alternative Cox proportional hazards modeling and
19 categorical exposure-response analyses using data from the UCC cohort (Swaen et al., 2009), the
20 NIOSH cohort (Steenland et al., 2004) and the two cohorts combined, analyzing the sexes both
21 separately and together. These investigators reported that they found no evidence of
22 exposure-response relationships for cumulative exposure with either the Cox model or
23 categorical analyses for all of the cohort/endpoint data sets examined (endpoints included all
24 lymphohematopoietic cancers, lymphoid cancers, and female breast cancer, the latter in the
25 NIOSH cohort only). Valdez-Flores et al. (2010) did observe statistically significant increases in
26 response rates in the highest exposure quintile relative to the lowest exposure quintile for
27 lymphohematopoietic and lymphoid cancers in males in the NIOSH cohort, consistent with the
28 categorical results of Steenland et al. (2004), as well as a statistically significant increase in the
29 highest exposure quintile for lymphoid cancers in males and females combined in the NIOSH
30 cohort, consistent with the results in Appendix D. Because the exposure assessment conducted
31 for the UCC cohort is much cruder (see above and Appendix A.2.20), especially for the highest
32 exposures, than the NIOSH exposure assessment (which was based on a validated regression
33 model; see Appendix A.2.8), EPA considers the results of exposure-response analyses of the
34 combined cohort data to have greater uncertainty than those from analyses of the NIOSH cohort
35 alone, despite the additional cases contributed by the UCC cohort (e.g., the UCC cohort
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-8 DRAFT—DO NOT CITE OR QUOTE
-------
1 contributes 17 cases of lymphoid cancer to the 53 from the NIOSH cohort). Furthermore,
2 Valdez-Flores et al. (2010) did not use any log cumulative exposure models, and these were the
3 models that were statistically significant in the Steenland et al. (2004) analyses, consistent with
4 the apparent supralinearity of the NIOSH exposure-response data. See Appendix A.2.20 for a
5 more detailed discussion of the Valdez-Flores et al. (2010) analyses and how they compared with
6 the Steenland et al. (2004) analyses.
7 In a mortality study of 1,971 male chemical workers in Italy, 637 of whom were licensed
8 to handle EtO but not other toxic gases, Bisanti et al. (1993) reported statistically significant
9 excesses of hematopoietic cancers (SMR = 7.1, Obs = 5,p < 0.05). The study was limited by the
10 lack of exposure measurements and by the young age of the cohort. Although this study
11 suggests that exposure to EtO leads to a significant excess of hematopoietic cancer, the lack of
12 personal exposure measurements and the fact that members were potentially exposed to other
13 chemicals in the workplace lessen the study's usefulness for establishing the carcinogenicity of
14 EtO.
15 Hagmar et al. (Hagmar et al., 1995; Hagmar et al., 1991) studied cancer incidence in
16 2,170 Swedish workers (861 male and 1,309 female) in two medical sterilizing plants. They
17 determined concentrations in six job categories and estimated cumulative exposures for each
18 worker. They found hematopoietic cancers in 6 individuals versus 3.4 expected (SMR =1.8) and
19 a nonsignificant doubling in the risk when a 10-year latency period was considered. Even
20 though the cohort was young, the follow-up time was short (median 11.8 years), and only a small
21 fraction of the workers was highly exposed, the report is suggestive of an association between
22 EtO exposure and hematopoietic cancers. The risk of breast cancer was less than expected,
23 although with such short follow-up, the total numbers of cases was small (standardized incidence
24 ratio [SIR] = 0.5, Obs = 5). In the latent category of 10 years or more, the risk was even lower
25 (SIR = 0.4, Obs = 2).
26 In a large chemical manufacturing plant in Belgium (number of employees not stated),
27 Swaen et al. (1996) performed a nested case-control study of Hodgkin lymphoma to determine
28 whether a cluster of 10 cases in the active male work force was associated with any particular
29 chemical. They found a significant association for benzene and EtO. This study is limited by
30 the exclusion of inactive workers and the potential confounding effect of other chemicals besides
31 EtO, and it is not useful for quantitative dose-response assessment.
32 Olsen et al. (1997) studied 1,361 male employees working in the ethylene and propylene
33 chlorohydrin production and processing areas located within the EtO and propylene oxide
34 production plants at four Dow Chemical Company sites in the United States. Although these
35 investigators found a nonsignificant positive trend between duration of employment as
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-9 DRAFT—DO NOT CITE OR QUOTE
-------
1 chlorohydrin workers and lymphohematopoietic cancer (Obs = 10), they concluded that there
2 was no appreciable risk in these workers, in contrast to the findings of Benson and Teta (1993).
3 The small cohort size and the lack of data on EtO exposures limit the usefulness of this study in
4 inferring risks due to EtO.
5 Norman et al. (1995) studied 1,132 workers (204 male and 928 female) in a medical
6 sterilizing plant in the United States. In the women, there was a significant excess incidence of
7 breast cancer (SIR = 2.6, Obs = 12, p < 0.05); no other cancer sites were elevated. The risk of
8 breast cancer was not noted to be excessive in the few previous studies where adequate numbers
9 of females were included and analyzed for breast cancer; however, only one of these studies was
10 also an incidence study. The follow-up time was too short to draw meaningful conclusions at
11 this time. This study lacks the power to determine whether risks for cancers other than breast
12 cancer are statistically significantly elevated. It has no information regarding historical exposure
13 and some breast cancer victims had worked for less than 1 month.
14 Tompa et al. (1999) reported a cluster of eight breast cancers and eight other cancers in
15 98 nurses exposed to EtO in a hospital in Hungary; however, the expected number of cases
16 cannot be identified.
17 The NIOSH investigators used the NIOSH cohort to conduct a study of breast cancer
18 incidence and exposure to EtO (Steenland et al., 2003). The researchers identified 7,576 women
19 from the initial cohort who had been employed in the commercial sterilization facilities for at
20 least 1 year (76% of the original cohort). Breast cancer incidence was determined from
21 interviews (questionnaires), death certificates, and cancer registries. Interviews were obtained
22 for 5,139 women (68% of the study cohort). The main reason for nonresponse was inability to
23 locate the study subject (22% of cohort). The average duration of exposure for the cohort was
24 10.7 years. For the full study cohort, 319 incident breast cancer cases were identified, including
25 20 cases of carcinoma in situ. Overall, the SIR was 0.87 (0.94 excluding the in situ cases) using
26 Surveillance, Epidemiology, and End Results (SEER) reference rates for comparison. Results
27 with the full cohort are expected to be underestimated, however, because of case
28 under-ascertainment in the women without interviews. A significant exposure-response trend
29 was observed for SIR across cumulative exposure quintiles, using a 15-year lag time (p = 0.002).
30 In internal Cox regression analyses, with exposure as a continuous variable, a significant trend
31 for breast cancer incidence was obtained for log cumulative exposure with a 15-year lag
32 (p = 0.05), taking age, race, and year of birth into account. Using duration of exposure, lagged
33 15 years, provided a slightly better fit (p = 0.02), while models with cumulative
34 (nontransformed), maximum or average exposure did not fit as well. In the Cox regression
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-10 DRAFT—DO NOT CITE OR QUOTE
-------
1 analysis with categorical exposures and a 15-year lag, the top cumulative exposure quintile had a
2 statistically significant OR for breast cancer incidence of 1.74 (95% CI = 1.16-2.65).
3 In the subcohort with interviews, 233 incident breast cancer cases were identified.
4 Information on various risk factors for breast cancer was also collected in the interviews, but
5 only parity and breast cancer in a first-degree relative turned out to be important predictors of
6 breast cancer incidence. In internal analyses with continuous exposure variables, the model with
7 duration of exposure (lagged 15 years) again provided the best fit (p = 0.006). Both the
8 cumulative exposure and log cumulative exposure models also yielded significant regression
9 coefficients with a 15-year lag (p = 0.02 andp = 0.03, respectively), taking age, race, year of
10 birth, parity, and breast cancer in a first-degree relative into account. In the Cox regression
11 analysis with categorical exposures and a 15-year lag, the top cumulative exposure quintile had a
12 statistically significant OR of 1.87 (95% CI = 1.12-3.10).
13 Steenland et al. (2003) suggest that their findings are not conclusive of a causal
14 association between EtO exposure and breast cancer incidence because of inconsistencies in
15 exposure-response trends, possible biases due to nonresponse, and an incomplete cancer
16 ascertainment. Although that conclusion seems appropriate, those concerns do not appear to be
17 major limitations. As noted by the authors, it is not uncommon for positive exposure-response
18 trends not to be strictly monotonically increasing, conceivably due to random fluctuations or
19 imprecision in exposure estimates. Furthermore, the consistency of results between the full
20 study cohort, which is less subject to nonresponse bias, and the subcohort with interviews, which
21 should have full case ascertainment, alleviates some of the concerns about those potential biases.
22 In a study of 299 female workers employed in a hospital in Hungary where gas sterilizers
23 were used, Kardos et al. (2003) observed 11 cancer deaths, including 3 breast cancer deaths,
24 compared with slightly more than 4 expected total cancer deaths. Site-specific expected deaths
25 are not available in this study, so RR estimates cannot be determined. However, the observation
26 of 3 breast cancer deaths, with at most 4.4 (with Hungarian national rates as the referent) total
27 cancer deaths expected, is indicative of an increased risk of breast cancer5, and this
28 characterization is supported by the reference of Major et al. (2001) to a cluster of breast cancer
29 cases in female nurses at the same hospital.
5Hungarian age-standardized female cancer mortality rates reported by the International Agency for Research on
Cancer (http://eu-cancer.iarc.fr/country-348-hungary.html,en) suggest that the ratio of breast cancer deaths to total
cancer deaths in Hungarian females is about 0.16 (28.0/100,000 breast cancer mortality rate versus
180.0/100,000 total cancer mortality rate). Although a comparison of this general population ratio with the ratio of
0.68 for breast cancer to total cancer mortality in the Kardos et al. (2003) study is necessarily crude because the
general population ratio is not based on the age-standardized rates that would correspond to the age distribution of
the person-time of the women in the study, which are unknown, the large difference between the ratios (0.68 for the
study versus 0.16 for the general population) indicates an increased risk of breast cancer in the study.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-11 DRAFT—DO NOT CITE OR QUOTE
-------
1 3.1.1. Conclusions Regarding the Evidence of Cancer in Humans
2 Most of the human studies suggest a possible increased risk of lymphohematopoietic
3 cancers, but the total weight of the epidemiological evidence does not provide conclusive proof
4 of causality. Of the eight relevant criteria of causality envisioned by Hill (1965), temporality,
5 coherence, biological plausibility, and analogy are clearly satisfied. There is also evidence of
6 consistency in the response, of a dose-response relationship (biological gradient), and of
7 specificity when the loosely defined blood malignancies are combined under the rubric "cancer
8 of the hematopoietic system." On the other hand, most of the relative risk estimates are not large
9 (strong) in magnitude. See Section 3.5.1 for a more detailed discussion of the Hill criteria as
10 applied to the EtO database.
11 The large NIOSH study (Steenland et al., 2004; Stayner et al., 1993; Steenland et al.,
12 1991) of workers at 14 chemical plants around the country provides the strongest evidence of
13 carcinogenicity. A statistically significant positive trend was observed in the risk of
14 lymphohematopoietic neoplasms with increasing (log) cumulative exposure to EtO, although the
15 results for this model were reported only for males (the sex difference is not statistically
16 significant, however, and the trend for both sexes combined is statistically significant; see
17 Appendix D). Despite limitations in the data, most other epidemiologic studies have also found
18 elevated risks of lymphohematopoietic cancer from exposure to EtO (summarized briefly in
19 Section 3.1 and Table 3-1; see also Appendix A for more details, in particular Table A-5 for a
20 summary of study results and limitations). Furthermore, when the exposure is relatively pure,
21 such as in sterilization workers, there is an elevated risk of lymphohematopoietic cancer that
22 cannot be attributed to the presence of confounders such as those that could potentially appear in
23 the chlorohydrin process. Moreover, the studies that do not report a significant
24 lymphohematopoietic cancer effect from exposure to EtO have major limitations, such as small
25 numbers of cases and inadequate exposure information (see Table A-5 in Appendix A).
26 In addition, there is evidence of an increase in the risk of both breast cancer mortality and
27 incidence in women who are exposed to EtO. Studies have reported increases in the risk of
28 breast cancer in women employees of commercial sterilization plants (Steenland et al., 2004;
29 Steenland et al., 2003; Norman et al., 1995) as well as in Hungarian hospital workers exposed to
30 EtO (Kardos et al., 2003). In several other studies where exposure to EtO would be expected to
31 have occurred among female employees, no elevated risks were seen (Coggon et al., 2004;
32 Hagmar et al., 1991) or breast cancer results were not reported (Hogstedt, 1988; Hogstedt et al.,
33 1986). However, these studies had far fewer cases to analyze than the NIOSH studies, and most
34 did not have individual exposure estimates and relied on external comparisons (see Table 3-2 for
35 a brief summary and Table A-5 in Appendix A for more details). The Steenland et al. (2004) and
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-12 DRAFT—DO NOT CITE OR QUOTE
-------
to
OJ
Table 3-1. Epidemiological studies of ethylene oxide and human cancer—lymphohematopoietic cancer results3
Study/Population/
Industry
Hogstedt (1988) and
Hogstedtetal. (1986).
Sterilizers, production
workers, Sweden.
Coggon et al. (2004).
Update of Gardner et al.
(1989).
Sterilizing workers in 8
hospitals and users in 4
companies, Great
Britain.
Kiesselbach et al.
(1990).
Production workers
(methods unspecified)
from 8 chemical plants
in West Germany.
Benson and Teta (1993).
Follow-up of only the
chlorohydrin-exposed
employees from
Greenberg et al. (1990)
cohort.
Production workers at a
chemical plant in West
Virginia.
Number of
subjects
709
(539 men,
170 women)
2,876
(1,864 men,
1,012
women)
2,658 men
278 men
Lymphohematopoietic cancer results
cancer deaths observed expected SMR (95% CI)
leukemia (ICD-8 204-207) 7 0.8 9.2 (3.7, 19)b
lymphohematopoietic 9 2.0 4.6 (2.1, 8.7)b
(ICD-8 200-208)
cancer deaths observed expected SMR (95% CI)
leukemia (ICD-9 204-208) 5 4.6 1.1(0.35,2.5)
leukemia 5 2.6 1.9 (0.62, 4.5)b
(definite or continual exposure)
NHL (ICD-9 200+202) 7 4.8 1.5 (0.58, 3.0)b
lymphohematopoietic 17 12.9 1.3 (0.77, 2.1)b
(ICD-9 200-208)
cancer deaths observed expected SMR (95% CI)
leukemia (ICD-9 204-208) 2 2.35 0.85(0.10,3.1)
lymphohematopoietic 5 5 1.0(0.32,2.3)
(ICD-9 200-208)
cancer deaths observed expected SMR (95% CI)
leukemia and aleukemia 4 1.14 3.5(0.96,8.9)
lymphosarcoma 1 0.50 2.0(0.05,11)
and reticulosarcoma
lymphohematopoietic 8 2.72 2.9(1.3,5.8)
(ICD NS)
Comments
Insufficient follow-up; only 12.0% of
cohort had died (85 deaths).
Exposure to other chemicals.
Short follow-up; only 19.6% of cohort
had died (565 deaths).
Exposure to other chemicals.
Insufficient follow-up; only 10.1% of
cohort had died (268 deaths).
Exposure to other chemicals.
EtO exposures reported to be low in the
chlorohydrin process.
Exposure to other chemicals.
Very small cohort; thus, small numbers
of specific cancers despite long follow-
up (52.9% had died; 147 deaths).
§•
§
I
^
'TS
o
a
I
a,
8"
O 5
1
3
o
H
W
o
c
o
H
W
-------
to
OJ
Table 3-1. Epidemiological studies of ethylene oxide and human cancer—lymphohematopoietic cancer results"
(continued)
§•
§
I
^
'TS
I
o
a
I
a,
8"
Study/Population/
Industry
Number of
subjects
Lymphohematopoietic cancer results
Comments
Swaen et al. (2009).
Update of Tetaetal.
(1993) [Greenberg et al.
(1990) cohort minus all
chlorohydrin-exposed
employees] plus cohort
enumeration extended
an additional 10 years,
adding 167 workers.
Production workers and
users at 2 chemical
plants in West Virginia.
2,063 men
cancer deaths
observed expected SMR (95% CI)
leukemia 11 11.8 0.93(0.47,1.7)
leukemia 9 NR 1.5(0.69,2.9)
(in workers hired before 1956)
NHL 12 11.5 1.05(0.54,1.8)
lymphohematopoietic 27 30.4 0.89 (0.59, 1.3)
(ICD NS)
Internal Cox regression analyses:
No statistically significant trends were observed for lymphoid or leukemia
cancer categories for continuous cumulative exposure.
Small cohort; thus, small numbers of
specific cancers even though long
follow-up time (50.8% had died; 1,048
deaths).
Crude exposure assessment, especially
for the early time periods.
Exposure to other chemicals.
O 5
1
3
o
H
W
Steenland et al. (2004).
Update of Steenland et
al. (1991), Stayneretal.
(1993).
Sterilizers of medical
equipment and spices;
and manufacturers and
testers of medical
sterilization equipment,
in 14 plants in the
United States.
18,254
(45% male,
55% female)
cancer deaths
leukemia (ICD-9 204-208)
NHL (ICD-9 200+202)
lymphohematopoietic
(ICD-9 200-208)
observed expected
29 NR
31 NR
79 NR
SMR (95% CI)
0.99(0.71, 1.36)
1.00 (0.72, 1.35)
1.00 (0.79, 1.24)
Internal Cox regression analyses:
"lymphoid" cancers (ICD-9 200, 202, 203, 204): OR = 3.0 (p = 0.046)
in highest cumulative exposure group, with 15-yr lag; significant
regression coefficient for continuous log cumulative exposure (p = 0.02).
lymphohematopoietic cancer (ICD-9 200-208): OR = 2.96 (p = 0.03)
in highest cumulative exposure group, with 15-yr lag; significant
regression coefficient for continuous log cumulative exposure (p = 0.009).
Large cohort; thus, substantial number of
deaths (2,852) despite short follow-up
(15.6% had died).
High-quality exposure assessment.
No evidence of exposure to other
occupational carcinogens.
No increase in lymphohematopoietic
cancer risk with increase in exposure in
women.
Results from internal Cox regression
analyses for both sexes combined from
Sections D.3 and D.4 of Appendix D.
o
c
o
H
W
-------
to
OJ
Table 3-1. Epidemiological studies of ethylene oxide and human cancer—lymphohematopoietic cancer results"
(continued)
Study/Population/
Industry
Bisantietal. (1993).
Chemical workers
licensed to handle EtO
and other toxic
chemicals, Italy.
Hagmar et al. (1995) and
Hagmaretal. (1991).
Two plants that
produced disposable
medical
equipment, Sweden.
Norman etal. (1995).
Sterilizers of medical
equipment and supplies
that were assembled at
this plant, New York.
Swaen etal. (1996).
Nested case-control
study; cases and controls
from a large chemical
production plant,
Belgium.
Number of
subjects
1,971 men
2,170
(861 men,
1,309
women)
1,132
(204 men,
928 women)
10 cases of
Hodgkin
lymphoma (7
confirmed)
and 200
controls; all
male
Lymphohematopoietic cancer results
cancer deaths observed expected SMR (95% CI)
leukemia (ICD-9 204-208) 2 1.0 1.9(0.23,7.0)
lymphosarcoma and 4 0.6 6.8(1.9,17)
reticulosarcoma (ICD-9 200)
lymphohematopoietic 6 2.4 2.5 (0.91, 5.5)
(ICD-9 200-208)
in group only licensed to handle EtO (n = 637):
leukemia 2 0.3 6.5(0.79,23)
lymphosarcoma and 3 0.2 17(3.5,50)
reticulosarcoma
lymphohematopoietic 5 0.7 7.0 (2.3, 16)
cancer cases observed expected SIR (95% CI)
leukemia (ICD-7 204-205) 2 0.82 2.4(0.30,8.8)
NHL (ICD-7 200+202) 2 1.25 1.6(0.19,5.8)
lymphohematopoietic 6 3.37 1.8 (0.65, 3.9)
(ICD-7 200-209)
leukemia 2 0.28 7.1(0.87,26)
(among subjects with at least 0.14 ppm-years of cumulative exposure
and 10 years latency)
cancer cases observed expected SIR (95% CI)
leukemia (ICD NS) 1 0.54 1.85 (0.05, 10)b
cancer OR (95% CI)
Hodgkin lymphoma (ICD 201) 8.5 (1.4, 40)
Comments
Insufficient follow-up; only 3.9% of
cohort had died (76 deaths).
Exposure to other chemicals.
Short follow-up period (only 40 cancer
cases).
Short follow-up period and small cohort
(only 28 cancer cases).
Hypothesis-generating study to
investigate a cluster of Hodgkin
lymphomas observed at a chemical plant.
Exposure to other chemicals.
§•
§
I
^
'TS
o
a
I
a,
8"
O 5
O I
2 x
31
-------
to
o
Table 3-1. Epidemiological studies of ethylene oxide and human cancer—lymphohematopoietic cancer results"
(continued)
This document is a draft for review purpos
3-16
Study/Population/
Industry
Olsenetal. (1997).
Four EtO production
plants (chlorohydrin
process) in 3 states.
Kardos et al. (2003).
Female workers from
pediatric clinic of
hospital in Eger,
Hungary.
Number of
subjects
1,361 men
299 women
Lymphohematopoietic cancer results
cancer deaths observed expected SMR (95% CI)
leukemia (ICD-8 204-207) 2 3.0 0.67(0.08,2.4)
lymphosarcoma and 1 1.1 0.91(0.02,5.1)
reticulosarcoma (ICD-8 200)
lymphohematopoietic 10 7.7 1.3 (0.62, 2.4)
(ICD-8 200-209)
1 lymphoid leukemia death of 1 1 cancer deaths; expected number not
reported.
Comments
Short follow-up and small cohort; thus,
small numbers of specific cancers
(22.0% had died; 300 deaths).
Exposure to other chemicals.
Short follow-up period and small cohort
(only 11 cancer deaths).
Possible exposure to natural radium,
which permeates the region.
o
a
I
a,
8"
""Extracted from Table A-5 of Appendix A, with addition of some summary results (e.g., SMRs).
bCalculated by EPA assuming Poisson distribution.
ICONS: ICD codes not specified; NR: not reported
O 5
3
o
H
W
1
o
c
o
H
W
-------
Table 3-2. Summary of epidemiological results on EtO and breast cancer (all sterilizer workers)11
OJ
KO
5^
1
s
3
1
?
a
4>
^
i
^
'•Q
LtJ S
£
a
|
O o
^ 1
f §
o I-
f^
OOo
^
H =s
O ^
HH ^
\ 1 O
W ^
O '
Study
Hogstedt(1988);
Hogstedt et al.
(1986)
Swedish
incidence and
mortality study
Coggon et al.
(2004)
Great Britain
mortality study
Steenland et al.
(2004)
U.S. mortality
study
Steenland et al.
(2003)
U.S. breast
cancer incidence
study
Number of
Women
170
1,011 women
hospital
workers
9,908
1,516
employed for
>lyr; 5,139
with interviews
Breast Cancer Results
not reported
exposure category observed expected SMR (95% CI)
continual 5 7.2
intermittent 0 0.7
unknown 6 5.2
ALL 11 13.1 1.04(0.42,1.51)
SMR in highest quartile of cumulative exposure (with 20-yr lag) = 2.07
(p < 0.05).
significant Cox regression coefficient for log cumulative exposure
(20-yr lag) (p = 0.01).
full cohort results:
Cox regression analysis OR = 1.14(95%CI: 1.16, 2.65) for highest
cumulative exposure quintile (15-yr lag).
p = 0.05 for regression coefficient with log cumulative exposure (15-yr
lag).
subcohort results:
Cox regression analysis OR = 1.87 (95% CI: 1.12, 3.10) for highest
cumulative exposure quintile (15-yr lag).
p = 0.02 for regression coefficient with cumulative exposure (15-yr
lag);p = 0.03 with log cumulative exposure (15-yr lag).
Comments
Only 8 deaths (7 from cancer) had
occurred among the women.
37 incident cancer cases (27
expected) in the total cohort
(including 539 men); 7 were
lymphohematopoietic cancers, rest
not reported.
1 1 breast cancer deaths.
only 14% of the cohort of 1,405
(including males) hospital workers
had died.
103 breast cancer deaths.
319 cases in full cohort.
233 cases in subcohort with
interviews.
0
c
0
H
W
-------
to
OJ
Table 3-2. Summary of epidemiological results on EtO and breast cancer (all sterilizer workers)" (continued)
00
O 5
5^
s?'
o-
o
s
1
§
1-t.
>j'
&
^
1
^*
^
^
'TS
1
$
O
a
a
a
a.
§•
TO
a
o_
1
Study
Hagmar et al.
(1995) and
Hagmar et al.
(1991)
Swedish cancer
incidence study
Norman et al.
(1995)
U.S. cancer
incidence study
Kardos et al.
(2003)
Hungarian
mortality study
Number of
Women
1,309
928
299
Breast Cancer Results
5 cases vs. 10.8 expected SIR = 0.46 (95% CI: 0.15, 1.08).
SIRs ranged from 1.72 (95% CI: 0.99, 3.00) to 2.40 (95% CI: 1.32,
4.37) depending on calendar year of follow-up, assumptions about
completeness of follow-up, and reference rates used.
1 1 cancer deaths observed compared with 4.38, 4.03, or 4.28 expected
(p < 0.01), based on comparison populations of Hungary, Heves
County, and city of Eger, respectively; 3 were breast cancer deaths, i.e.,
3 breast cancer deaths vs. ~4.3 total deaths expected. Although the
expected number of breast cancer deaths was not reported, the number
of breast cancer deaths observed for the total deaths expected is
indicative of an increased risk of breast cancer (see footnote 2 in
Section 3.1).
Comments
5 cases.
12 cases.
3 deaths.
""Extracted from Table A-5 of Appendix A
3
o
H
W
1
o
c
o
H
W
-------
1 Steenland et al. (2003) studies, on the other hand, used the largest cohort of women potentially
2 exposed to EtO and clearly show significantly increased risks of breast cancer incidence and
3 mortality based upon internal exposure-response analyses.
4 In summary, the most compelling evidence of a cancer risk from human exposure to EtO
5 is for cancer of the lymphohematopoietic system. Increases in the risk of lymphohematopoietic
6 cancer are present in most of the studies, manifested as an increase in leukemia and/or cancer of
7 the lymphoid tissue. The evidence of lymphohematopoietic cancer is strongest in the one study
8 (the NIOSH study) that appears to possess the fewest limitations. In this large study, a
9 significant dose-response relationship was evident with cumulative exposure to EtO. However,
10 this effect was observed primarily in males and the magnitude of the effect was not large.
11 Similarly, in most of the other studies, the increased risks are not great, and other chemicals in
12 some of the workplaces cannot be ruled out as possible confounders. Thus, the findings of
13 increased risks of lymphohematopoietic cancer in the NIOSH and other studies cannot
14 conclusively be attributed to exposure to EtO. The few studies that fail to demonstrate any
15 increased risks of cancer do not have those strengths of study design that give confidence to the
16 reported lack of an exposure-related effect.
17 There is also evidence of an elevated risk of breast cancer from exposure to EtO in a few
18 studies. The strongest evidence again comes from the NIOSH studies, which found positive
19 exposure-response relationships for both breast cancer incidence and mortality. Hopefully,
20 future studies will shed more light on this more recent finding.
21
22 3.2. EVIDENCE OF CANCER IN LABORATORY ANIMALS
23 The International Agency for Research on Cancer (IARC) monograph (IARC, 1994b) has
24 summarized the rodent studies of carcinogenicity, and Health Canada (2001) has used this
25 information to derive the levels of concern for human exposure. EPA concludes that the IARC
26 summary of the key studies is valid and is not aware of any animal cancer bioassays that have
27 been published since 1994. The Ethylene Oxide Industry Council (EOIC, 2001) also reviewed
28 the same studies and did not cite additional studies. The qualitative results are described here
29 and the incidence data are tabulated in the unit risk derivation section of this document.
30 One study of oral administration in rats has been published; there are no oral studies in
31 mice. Dunkelberg (1982) administered EtO in vegetable oil to groups of 50 female
32 Sprague-Dawley rats by gastric intubation twice weekly for 150 weeks. There were two control
33 groups (untreated and oil gavage) and two treated groups (7.5 and 30 mg/kg-day). A
34 dose-dependent increase in the incidence of malignant tumors in the forestomach was observed
35 in the treated groups (8/50 and 31/50 in the low- and high-dose groups, respectively). Of the
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-19 DRAFT—DO NOT CITE OR QUOTE
-------
1 39 tumors, 37 were squamous cell carcinomas, and metastases to other organs were common in
2 these animals. This study was not evaluated quantitatively because oral risk estimates are
3 beyond the scope of this document.
4 One inhalation assay was reported in mice (NTP, 1987) and two inhalation assays were
5 reported in rats [(Lynch et al., 1984a; Lynch et al., 1984b) in males; (Garman et al., 1986, 1985;
6 Snellings et al., 1984), in both males and females)]. In the National Toxicology Program (NTP)
7 mouse bioassay (NTP, 1987), groups of 50 male and 50 female B6C3Fi mice were exposed to
8 EtO via inhalation at concentrations of 0, 50, and 100 ppm for 6 hours per day, 5 days per week,
9 for 102 weeks. Mean body weights were similar for treated and control animals, and there was
10 no decrease in survival associated with treatment. A concentration-dependent increase in the
11 incidence of tumors at several sites was observed in both sexes. These data are summarized in
12 Table 3-3. Males had carcinomas and adenomas in the lung. Females had carcinomas and
13 adenomas in the lung, malignant lymphomas, adenocarcinomas in the uterus, and
14 adenocarcinomas in the mammary glands. The NTP also reports that both sexes had dose-related
15 increased incidences of cystadenomas of the Harderian glands, but these are benign lesions and
16 are not considered further here.
17 In the Lynch et al. [Lynch et al. (1984a); Lynch et al. (1984b)] bioassay in male Fischer
18 344 (F344) rats, groups of 80 animals were exposed to EtO via inhalation at concentrations of 0,
19 50, and 100 ppm for 7 hours per day, 5 days per week, for 2 years. Mean body weights were
20 statistically significantly decreased in both treated groups compared with controls (p < 0.05).
21 Increased mortality was observed in the treated groups, and the increase was statistically
22 significant in the 100-ppm exposure group (p < 0.01). Lynch et al. (1984a) suggest that survival
23 was affected by a pulmonary infection alone and in combination with EtO exposure.
24 Concentration-dependent increases in the incidence of mononuclear cell leukemia in the spleen,
25 peritoneal mesothelioma in the testes, and glioma in the brain were observed (see Table 3-4).
26 The fact that the increased incidence of mononuclear cell leukemia was statistically significant in
27 the low-exposure group, but not in the high-exposure group, is probably attributable to the
28 increased mortality in the high-exposure group. The increased incidence in just the terminal kill
29 rats in the 100-ppm group was statistically significant compared with controls.
30 In the bioassay conducted by Snellings et al. (1984), 120 male and 120 female F344 rats
31 in each sex and dose group were exposed to EtO via inhalation at concentrations of 0 (2 control
32 groups of 120 rats of each sex were used), 10, 33, and 100 ppm for 6 hours per day, 5 days per
33 week, for 2 years, with scheduled kills at 6 (10 rats per group), 12 (10 rats per group), and
34 18 (20 rats per group) months. Significant decreases in mean body weight were observed in the
35 100-ppm exposure group in males and in the 100-ppm and 33-ppm exposure groups in females.
36
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-20 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
Table 3-3. Tumor incidence data in National Toxicology Program Study of
B6C3F! mice (NTP, 1987)a
Gender/tumor type
EtO concentration
(time-weighted average)1"
0 ppm
50 ppm
(16.3 mg/m3)
100 ppm
(32.7 mg/m3)
EC10
(LEC10)C,
(mg/m3)
Unit risk
(0.1/LEdo)
(per mg/m3)
Males
Lung adenomas plus
Carcinomas
11/49
19/49
26/49d
6.94
(4.51)
2.22 x 10'2
Females
Lung adenomas plus
Carcinomas
Malignant
Lymphoma
Uterine
Carcinoma
Mammary
carcinoma11
2/44
9/44
0/44
1/44
5/44
6/44
1/44
8/44f
22/49e
22/49f
5/498
6/49
14.8
(9.12)
21.1
(13.9)
32.8
(23.1)
9.69
(5.35)
1.1 x 10'2
7.18 x ID'3
4.33 x ID'3
1.87 x ID'2
4
5
6
7
8
9
10
11
12
13
14
15
16
Incidence data were adjusted by eliminating the animals that died prior to the occurrence of the first tumor or prior
to 52 weeks, whichever was earlier.
bAdjusted to continuous exposure from experimental exposure conditions of 6 hr/d, 5 d/wk; 1 ppm = 1.83 mg/m3.
Calculated using Tox_Risk program.
dp < 0.01 (pairwise Fisher's exact test).
ep < 0.001 (pairwise Fisher's exact test).
{p < 0.05 (pairwise Fisher's exact test).
sp = 0.058 by pairwise Fisher's exact test compared to concurrent controls; however, uterine carcinomas are rare
tumors in female B6C3F! mice, andp < 0.0001 by pairwise Fisher's exact test compared to the NTP historical
control incidence of 1/1,077 for inhalation (air) female B6C3Fi mice fed the NTH-07 diet.
hHighest dose was deleted in order to fit a model to the dose-response data.
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
3 -21 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
Table 3-4. Tumor incidence data in Lynch et al. (1984a; 1984b) study of
male F344 rats
Tumor type
Splenic
mononuclear
cell leukemia0
Testicular
peritoneal
mesothelioma
Brain mixed-
cell glioma
Concentration (time-weighted average)3
0 ppm
24/77
3/78
0/76
50 ppm
(19.1 mg/m3)
38/79d
9/79
2/77
100 ppm
(38.1 mg/m3)
30/76
21/79e
5/79e
EC10
(LEC10)b,
(mg/m3)
7.11
(3.94)
16.7
(11.8)
65.7
(37.4)
Unit risk
(0.1/LEdo)
(per mg/m3)
2.54 x 10~2
8.5 x 10~3
2.68 x 10~3
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
""Adjusted to continuous exposure from experimental exposure conditions of 7 hr/d, 5 d/wk; 1 ppm = 1.83 mg/m3.
bCalculated using Tox_Risk program.
'Highest dose deleted while fitting the dose-response data.
dp < 0.05 (pairwise Fisher's exact test).
ep < 0.01 (pairwise Fisher's exact test).
During the 15* month of exposure, an outbreak of viral sialodacryoadenitis occurred, resulting in
the deaths of 1-5 animals per group. Snellings et al. (1984) claim that it is unlikely that the viral
outbreak contributed to the EtO-associated tumor findings. After the outbreak, mortality rates
returned to preoutbreak levels and were similar for all groups until the 20* or 21st month, when
cumulative mortality in the 33-ppm and 100-ppm exposure groups of each sex remained above
control values. By the 22nd or 23rd months, mortality was statistically significantly increased in
the 100-ppm exposure groups of both sexes.
In males, concentration-dependent increases in the incidence of mononuclear cell
leukemia in the spleen and peritoneal mesothelioma in the testes were observed, and in females
an increase in mononuclear cell leukemia in the spleen was seen. These data are summarized in
Table 3-5. Note that these investigators observed the same types of tumors (splenic leukemia
and peritoneal mesothelioma) seen by Lynch et al. (1984a); Lynch et al. (1984b). Snellings et al.
(1984) only report incidences (of incidental and nonincidental primary tumors for all exposure
groups) for the 24-month (terminal) kill. However, in their paper they state that significant
findings for the mononuclear cell leukemias were also obtained when all rats were included and
that a mortality-adjusted trend analysis yielded positive findings for the EtO-exposed females
(p < 0.005) and males (p < 0.05). Similarly, Snellings et al. (1984) report that when male rats
with unscheduled deaths were included in the analysis of peritoneal mesotheliomas, it appeared
that EtO exposure was associated with earlier tumor occurrence, and a mortality-adjusted trend
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3 -22 DRAFT—DO NOT CITE OR QUOTE
-------
1 analysis yielded a significant positive trend (p < 0.005). In later publications describing brain
2 tumors (Garman et al., 1986, 1985), both males and females had a concentration-dependent
3 increased incidence of brain tumors (see Table 3-5). Garman et al. (1986, 1985) report
4 incidences including all rats from the 18- and 24-month kills and all rats found dead or killed
5 moribund. The earliest brain tumors were observed in rats killed at 18 months.
6
7 3.2.1. Conclusions Regarding the Evidence of Cancer in Laboratory Animals
8 In conclusion, EtO causes cancer in laboratory animals. After inhalation exposure to
9 EtO, statistically significant increased incidences of cancer have been observed in both rats and
10 mice, in both males and females, and in multiple tissues (lung, mammary gland, uterus,
11 lymphoid cells, brain, tunica vaginalis testis). In addition, one oral study in rats has been
12 conducted, and a significant dose-dependent increase in carcinomas of the forestomach was
13 reported.
14
15 3.3. SUPPORTING EVIDENCE
16 3.3.1. Metabolism and Kinetics
17 Information on the kinetics and metabolism of EtO has been derived primarily from
18 studies conducted with laboratory animals exposed via inhalation, although some limited data
19 from humans have been identified. Details are available in several reviews (Fennell and Brown,
20 2001; Csanady et al., 2000; Brown et al., 1998; Brown et al., 1996).
21 Following inhalation, EtO is absorbed efficiently into the blood and rapidly distributed to
22 all organs and tissues. EtO is metabolized primarily by two pathways (see Figure 3-1):
23 (1) hydrolysis to ethylene glycol (1,2-ethanediol), with subsequent conversion to oxalic acid,
24 formic acid, and carbon dioxide; and (2) glutathione conjugation and the formation of
25 ,S'-(2-hydroxyethyl)cysteine and N-acetylated derivatives (WHO, 2003). From the available data,
26 the route involving conjugation with glutathione appears to predominate in mice; in larger
27 species (including humans), the conversion of EtO is primarily via hydrolysis through ethylene
28 glycol. Because EtO is an epoxide capable of reacting directly with cellular macromolecules,
29 both pathways are considered to be detoxifying.
30 Among rodent species, there are clear quantitative differences in metabolic rates. The
31 rate of clearance of EtO from the blood, brain, muscle, and testes was measured by Brown et al.
32 (1998); Brown et al. (1996). Clearance rates were nearly identical across blood and other tissues.
33 Following a 4-hour inhalation exposure to 100 ppm EtO in mice and rats, the average blood
34 elimination half-lives ranged from 2.4 to 3.2 minutes in mice and 11 to 14 minutes in rats. The
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-23 DRAFT—DO NOT CITE OR QUOTE
-------
to
Table 3-5. Tumor incidence data in Snellings et al. (1984) and Garman et al. (1985) reports on F344 rats"
Gender/tumor type
Concentration (time-weighted average)1"
Oppmc
10 ppm
(3.27 mg/m3)
33 ppm
(10.8 mg/m3)
100 ppm
(32.7 mg/m3)
ECio
(LEC10)d
(mg/m3)
Unit risk (0.1/LEC10)
(per mg/m3)
Males
Splenic mononuclear cell
leukemia
Testicular peritoneal
mesothelioma
Primary brain tumors
13/97
(13%)e
2/97
(2.1%)
1/181
(0.55%)
9/51
(18%)
2/51
(3.9%)
1/92
(1.1%)
12/39f
(32%)
4/39
(10%)
5/85f
(5.9%)
9/30f
(30%)
4/30f
(13%)
7/87g
(8.1%)
12.3
(6.43)
22.3
(11.6)
36.1
(22.3)
1.56 x ID'2
8.66 x ID'3
4.5 x ID'3
Females
Splenic mononuclear cell
leukemia
Primary brain tumors
11/116
(9.5%)
1/188
(0.53%)
ll/54f
(21%)
1/94
(1.1%)
14/48g
(30%)
3/92
(3.3%)
15/26h
(58%)
4/80f
(5%)
4.46
(3.1)
63.8
(32.6)
3.23 x ID'2
3.07 x ID'3
§•
§
I
^
'TS
o
a
I
a,
8"
O 5
31
o ^
H t
W ??•
"Denominators refer to the number of animals for which histopathological diagnosis was performed. For brain tumors Garman et al. (1985) included animals in
the 18-month and the 24-month sacrifice and found dead or euthanized moribund of those alive at the time of the first brain tumor, whereas for the other sites
Snellings et al. (1984) included animals only at the 24-month sacrifice.
bAdjusted to continuous exposure from experimental exposure conditions of 6 hr/d, 5 d/wk; 1 ppm = 1.83 mg/m3.
°Results for both control groups combined.
dUsing Tox_Risk program.
eNumbers in parentheses indicate percentage incidence values.
!p < 0.05 (pairwise Fisher's exact test).
*p < 0.01 (pairwise Fisher's exact test).
hp < 0.001 (pairwise Fisher's exact test).
o
c
o
H
W
-------
Ethylene oxide
Glutathione
transferase.
Epoxide
hydrolase?
GSCH2CH2OH
S-2-(hydroxyethyl-glutathione)
CYS^CH2CH2OH
S-2-(hydroxy ethyl) cysteine
N-acethyl-S-(2 -hydroxy ethyl)
cysteine
HOCH2CH2OH
1,2-ethanediol
i
HOCH2CHO
hydroxyacetaldehyde
i
HOCH2CO2H
Glycolic acid
i
OHCCO2H
Glyoxylic acid
HCO2H
Formic acid
CO2HCO2H
Oxalic acid
1
2
3
4
5
6
7
8
9
10
11
12
13
14
CO,
Figure 3-1. Metabolism of ethylene oxide.
elimination half-life in humans is 42 minutes (Filser et al., 1992), and the half-life in salt water is
4 days (IARC, 1994b).
In a more detailed study in mice, Brown et al. (1998) measured EtO concentrations in
mice after 4-hour inhalation exposures at 0, 50, 100, 200, 300, or 400 ppm. They found that
blood EtO concentration increased linearly with inhaled concentrations of less than 200 ppm, but
above 200 ppm the blood concentration increased more rapidly. In addition, glutathione levels in
liver, lung, kidney, and testes decreased as exposures increased above 200 ppm. The
investigators interpreted this, along with other information, to mean that at low concentrations
the metabolism and disappearance of EtO is primarily a result of glutathione conjugation, but at
higher concentrations, when tissue glutathione begins to be depleted, the elimination occurs via a
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-25 DRAFT—DO NOT CITE OR QUOTE
-------
1 slower nonenzymatic hydrolysis process, leading to a greater-than-linear increase in blood EtO
2 concentration.
3 Fennell and Brown (2001) constructed physiologically based pharmacokinetic (PBPK)
4 models of uptake and metabolism in mice, rats, and humans, based on previous studies. They
5 reported that the models adequately predicted blood and tissue EtO concentrations in rats and
6 mice, with the exception of the testes, and blood EtO concentrations in humans. Modeling
7 6-hour inhalation exposures yielded simulated blood peak concentrations and areas under the
8 curve (AUCs) that are similar for mice, rats, and humans (human levels are within about 15% of
9 rat and mouse levels; see Figure 3-2). In other words, exposure to a given EtO concentration in
10 air results in similar predicted blood EtO AUCs for mice, rats, and humans.
11 These studies show that tissue concentrations in mice, rats, and humans exposed to a
12 particular air concentration of EtO are approximately equal and that they are linearly related to
13 inhalation concentration, at least in the range of exposures used in the rodent cancer bioassays
14 (i.e., 100 ppm and below).
15
16 3.3.2. Protein Adducts
17 EtO forms DNA (see Section 3.3.3.1) and hemoglobin adducts within tissues throughout
18 the body (Walker et al., 1992a; Walker et al., 1992b). Formation of hemoglobin adducts has
19 been used as a measure of exposure to EtO. The main sites of alkylation are cysteine, histidine,
20 and the N-terminal valine; however, for analytical reasons, the N-(2-hydroxyethyl)valine adduct
21 is generally preferred for measurements (Walker et al., 1990). Walker et al. (1992b) reported
22 measurements of this hemoglobin adduct and showed how the concentration of the adducts
23 changes according to the dynamics of red blood cell turnover. Walker et al. (1992b) measured
24 hemoglobin adduct formation in mice and rats exposed to 0, 3, 10, 33, 100, and 300 (rats only)
25 ppm of EtO (6 hours/day, 5 days/week, for 4 weeks). Response was linear in both species up to
26 33 ppm, after which the slope significantly increased. The exposure-related decrease in
27 glutathione concentration in liver, lung, and other tissues observed by Brown et al. (1998) in
28 mice is a plausible explanation for the increasing rate of hemoglobin adduct formation at higher
29 exposures.
30 In humans, hemoglobin adducts can be used as biomarkers of recent exposure to EtO
31 (IARC, 2008a; Boogaard, 2002; IARC, 1994b), and several studies have reported
32 exposure-response relationships between hemoglobin adduct levels and EtO exposure levels
33 (e.g., van Sittert et al., 1993; Schulte et al., 1992). Hemoglobin adducts are good general
34 indicators of exposure because they are stable (DNA adducts, on the other hand, may be repaired
35
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-26 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
20
40 60 80
EtO exposure concentration (ppm)
100
120
Figure 3-2. Simulated blood AUCs for EtO following a 6-hour exposure to
EtO from the rat, mouse, and human PBPK models of Fennell and Brown
(2001); based on data presented in Fennell and Brown (2001). (Ratl and
rat2 results use different values for pulmonary uptake.)
or fixed as mutations and hence are less reliable measures of exposure). However, Post et al.
(1991) noted that human erythrocytes showed marked interindividual differences in the amounts
of EtO bound to hemoglobin, and Yong et al. (2001) reported that levels of
N-(2-hydroxyethyl)valine were approximately twofold greater in persons with a G<5Tr7-null
genotype than in those with positive genotypes. Endogenous ethylene oxide (see
Section 3.3.3.1) also contributes to hemoglobin adduct levels, making it more difficult to detect
the impacts of low levels of exogenous EtO exposure. In addition, Walker et al. (1993) reported
that hemoglobin adducts in mice and rats were lost at a greater rate than would be predicted by
the erythrocyte life span.
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
3-27 DRAFT—DO NOT CITE OR QUOTE
-------
1 3.3.3. Genotoxicity
2 Since the first report of EtO induction of sex-linked recessive lethals in Drosophila
3 (Rapoport, 1948), numerous papers have been published on the positive genotoxic activity in
4 biological systems, spanning the whole range of assay systems, from bacteriophage to higher
5 plants and animals. Figure 3-3 shows the 203 test entries in the EPA Genetic Activity Profile
6 database in 2001. In prokaryotes and lower eukaryotes, EtO induced DNA damage and gene
7 mutations in bacteria, yeast, and fungi and gene conversions in yeast. In mammalian cells (from
8 in vitro and/or in vivo exposures), EtO-induced effects include unscheduled DNA synthesis,
9 gene mutations, sister chromatid exchanges (SCEs), micronuclei, and chromosomal aberrations.
10 Genotoxicity, in particular increased levels of SCEs and chromosomal aberrations, has also been
11 observed in blood cells of workers occupationally exposed to EtO. Several publications contain
12 details of earlier genetic toxicity studies (e.g., IARC, 2008b; Kolman et al., 2002; Thier and Bolt,
13 2000; Natarajan et al., 1995; Preston et al., 1995; IARC, 1994b; Dellarco et al., 1990; Ehrenberg
14 and Hussain, 1981). This review briefly summarizes the evidence of the genotoxic potential of
15 EtO, focusing primarily on recently published studies that provide information on the mode of
16 action of EtO (see Appendix C for more details from some individual studies).
17
18 3.3.3.1. DNAAdducts
19 EtO is a direct-acting Sn2 (substitution-nucleophilic-bimolecular)-type monofunctional
20 alkylating agent that forms adducts with cellular macromolecules such as proteins (e.g.,
21 hemoglobin, see Section 3.3.2) and DNA (Pauwels and Veulemans, 1998). Alkylating agents
22 may produce a variety of different DNA alkylation products (Beranek, 1990) in varying
23 proportions, depending primarily on the electrophilic properties of the agent. Reactivity of an
24 alkylating agent is estimated by its Swain-Scott substrate constant (s-value), which ranges from
25 0 to 1, and EtO has a high s-value of 0.96 (Beranek, 1990; Golberg, 1986; Warwick, 1963).
26 Acting by the S^2 mechanism and having a high substrate constant both favor alkylation at the
27 N7 position of guanine in the DNA (Walker et al., 1990). The predominant DNA adduct formed
28 by EtO and other SN2-type alkylating agents is N7-(2-hydroxyethyl)guanine (N7-HEG). After in
29 vitro treatment of DNA with EtO, Segerback (1990) identified three adducts, N7-HEG,
30 N3-hydroxyethyladenine, and O6- hydroxy ethyl guanine (O6-HEG), in the ratios 200:8.8:1; two
31 other peaks, suspected of representing other adenine adducts, were also observed at levels well
32 below that of N7-HEG
33 Ethylene, an endogenous precursor of EtO, is produced during normal physiological
34 processes. Such processes reportedly include oxidation of methionine and hemoglobin, lipid
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-28 DRAFT—DO NOT CITE OR QUOTE
-------
2
3
4
5
6
7
8
9
10
1 1
12
13
14
15
16
17
18
19
LED
0.0001
0.001
0.01
0.1
1.0
10
100
1000
10000
10
100
1000
10000
100000
1000000
HID
ETHYLENE OXIDE
75-21-8
s
A
0
_ S
A
5
i-
c
n
GENETIC ACTIVITY PROFILE
s
E E
B S
R A
D H
C'Ht
-^M
Tlr
' N
F C
S
C
G
ss
-1- AA
B 79
P
F
PROKARY
"*"
A
N
G
3
?
I
s
G
R
K
C
F
1
1
LOW EUK
p
L F
M
_ c
-
s
D
M
X
—
D_ D
M H
f_I
j
III
PLNT
INS
IB
B
)
N
p
C
i
T
7
S
C
t
S
"?u $
I u
HH G
_L I
MAMM VITRO
-H
c
B
A
r^ cW _
° v a, t
A;
r c
r
HUMAN
F
NCEA-01
2001
S C
L L
H H
D
V M
HU-V
D T\Eu
L
R
M
" XM^. H
ft -jy(7J: °T
— C
-
F==
-
1 |
1
_
S
L
P
.
" C
_l_
- -~
--
==
V A
A V
A
MAMM VIVO
HU
IARC human carcinogen (group 1: human - limited, animal - sufficient)
Figure 3-3. Display of 203 data sets, including bacteria, fungi, plants, insects,
and mammals (in vitro and in vivo), measuring the full range of genotoxic
endpoints. (This is an updated version of the figure in IARC, 1994b)
See Appendix B for list of references.
peroxidation of fatty acids, and metabolism of intestinal bacteria (reviewed in Thier and Bolt,
2000; IARC, 1994a). EtO is then endogenously produced through the cytochrome
P450-mediated conversion of ethylene (Tornqvist, 1996). This endogenous production of EtO
contributes significantly to background levels of DNA adducts, making it difficult to detect the
impacts of low levels of exogenous EtO exposure on DNA adduct levels. For example, in DNA
extracted from the lymphocytes of unexposed individuals, mean background levels of N7-HEG
ranged from 2 to 8.5 pmol/mg DNA (Bolt, 1996). Using sensitive detection techniques and an
approach designed to separately quantify both endogenous N7-HEG adducts and "exogenous"
N7-HEG adducts induced by EtO treatment in rats, Marsden et al. (2009) reported increases in
exogenous adducts in DNA of spleen and liver consistent with a linear dose-response
relationship (p < 0.05), down to the lowest dose administered (0.0001 mg/kg injected i.p. daily
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
3-29 DRAFT—DO NOT CITE OR QUOTE
-------
1 for 3 days). Note that the whole range of doses studied by Marsden et al. (2009) lies well below
2 the dose corresponding to the lowest LOAEL from an EtO cancer bioassay (see Section C.7 of
3 Appendix C). Marsden et al. (2009) also observed increases in endogenous N7-HEG adduct
4 formation at the two highest doses (0.05 and 0.1 mg/kg), suggesting that, in addition to direct
5 adduct formation via alkylation, EtO can induce N7-HEG adduct production indirectly.
6 (Marsden et al., 2009) hypothesized that this indirect adduct formation by EtO results from the
7 induction of ethylene generation under conditions of oxidative stress.
8 In experiments with rats and mice exposed to EtO at concentrations of 0, 3, 10, 33, 100,
9 or 300 (rats only) ppm for 6 hours per day, 5 days per week, for 4 weeks, Walker et al. (1992a)
10 Walker et al. (1992a) measured N7-HEG adducts in the DNA of lung, brain, kidney, spleen,
11 liver, and testes. At 100 ppm, the adduct levels for all tissues except testis were similar (within a
12 factor of 3), despite the fact that not all of these tissues are targets for toxicity. The study's data
13 on the persistence of the DNA adducts indicate that DNA repair rates differ in different tissues.
14 Although Walker et al. (1992a) suggested that N7-HEG adducts are likely to be removed by
15 depurination forming apurinic/apyrimidinic (AP) sites in DNA, a later study from the same
16 group showed that EtO-induced DNA damage is repaired without accumulation of AP sites or
17 involving base excision repair (Rusyn et al., 2005). Rats exposed to high doses of EtO (300
18 ppm) by inhalation showed steady-state levels of O6-HEG adducts that are -250-300 times
19 lower than the N7-HEG levels (Walker et al., 1992a). Even though low levels of O6-HEG
20 adducts were detected, they are more mutagenic in nature and may contribute to the tumors
21 observed in target organs.
22 Two studies provide evidence of N7-HEG DNA adduct formation in human populations
23 occupationally exposed to EtO, one reporting a modest increase in white blood cells (van Delft et
24 al., 1994) and the other a four- to fivefold increase in granulocytes (Yong et al., 2007) compared
25 to unexposed controls. However, these differences were not statistically significant due to high
26 interindividual variation in adduct levels.
27
28 3.3.3.2. Point Mutations
29 EtO has consistently yielded positive results in in vitro mutation assays from
30 bactedophage, bacteria, fungi, yeast, insects, plants, and mammalian cell cultures (including
31 human cells). For example, EtO induces single base pair deletions and base substitutions in the
32 HPRTgene in human diploid fibroblasts (Kolman and Chovanec, 2000; Lambert et al., 1994;
33 Bastlova et al., 1993) in vitro. The results of in vivo studies on the mutagenicity of EtO have
34 also been consistently positive following ingestion, inhalation, or injection (e.g., Tates et al.,
35 1999). Increases in the frequency of gene mutations in T-lymphocytes (Hprt locus) (Walker et
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-30 DRAFT—DO NOT CITE OR QUOTE
-------
1 al., 1997) and in bone marrow and testes (Lad locus) (Recio et al., 2004) have been observed in
2 transgenic mice exposed to EtO via inhalation at concentrations similar to those in
3 carcinogenesis bioassays with this species (NTP, 1987). At somewhat higher concentrations
4 than those used in the carcinogenesis bioassays (200 ppm, but for only 4 weeks), increases in the
5 frequency of gene mutations have also been observed in the lungs of transgenic mice (Lad
6 locus) (Sisk et al., 1997) and in T-lymphocytes of rats (Hprt locus) (van Sittert et al., 2000; Tates
7 et al., 1999). In in vivo studies with male mice, EtO also causes heritable mutations and other
8 effects in germ cells (Generoso et al., 1990; Lewis et al., 1986).
9 In a study of mammary gland carcinomas in EtO-exposed B6C3Fi mice from the 1987
10 NTP bioassay (NTP, 1987) and 19 mammary gland carcinomas from concurrent controls in the
11 1987 NTP EtO bioassay and a 1986 NTP benzene bioassay, Houle et al. (2006) measured
12 mutation frequencies in exons 5-8 of thep53 tumor suppressor gene and in codon 61 of the Hras
13 oncogene. Mutation frequencies in the mammary carcinomas of EtO-exposed mice were only
14 slightly increased over frequencies in spontaneous mammary carcinomas (33% of the
15 carcinomas in the EtO-exposed mice had Hras mutations versus 26% of spontaneous tumors;
16 67% of the carcinomas in the EtO-exposed mice hadp53 mutations versus 58% of spontaneous
17 tumors); however, the EtO-induced tumors exhibited a distinct shift in the mutational spectra of
18 thep53 and Hras genes and more commonly displayed concurrent mutations of the two genes
19 (Houle et al., 2006). Furthermore, Houle et al. (2006) detected about sixfold higher levels of p53
20 protein expression in the mammary carcinomas of EtO-exposed mice than in spontaneous
21 mammary carcinomas, and there was an apparent dose-response relationship between EtO
22 exposure level and both p53 protein expression andp53 gene mutation (three of the seven tumors
23 in the 50-ppm exposure group and all five tumors in the 100-ppm group had increased protein
24 expression; also, threep53 gene mutations were found in the seven tumors in the 50-ppm
25 exposure group and nine were found in the five tumors in the 100-ppm group). Some of the
26 same investigators conducted a similar study ofKras mutations in lung, Harderian gland, and
27 uterine tumors (Hong et al., 2007). Substantial increases were observed in Kras mutation
28 frequencies in the tumors from the EtO-exposed mice. Kras mutations were reported in 100% of
29 the lung tumors from EtO-exposed mice versus 25% of spontaneous lung tumors (108 NTP
30 control animal tumors, including 8 from the EtO bioassay), in 86% of Harderian gland tumors
31 from EtO-exposed mice versus 7% of spontaneous Harderian gland tumors (27 NTP control
32 animal tumors, including 2 from the EtO bioassay), and in 83% of uterine tumors from
33 EtO-exposed mice (there were no uterine tumors in control mice in the 1986 NTP bioassay and
34 none were examined from other control animals). Furthermore, a specific Kras mutation, a
35 G —> T transversion in codon 12, was nearly universal in lung tumors from EtO-exposed mice
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-31 DRAFT—DO NOT CITE OR QUOTE
-------
1 (21/23) but rare in lung tumors from control animals (1/108). Other specific mutations were also
2 predominant in the Harderian gland and uterine tumors, but too few Kras mutations were
3 available in spontaneous Harderian gland tumors, and no spontaneous uterine tumors were
4 examined; thus, meaningful comparisons could not be made for these sites. Overall, these data
5 strongly suggest that EtO-induced mutations in oncogenes and tumor-supressor genes play a role
6 in EtO-induced carcinogenesis in multiple tissues.
7 Only a few studies have investigated gene mutations in people occupationally exposed to
8 EtO. In one study, HPRT mutant frequency in peripheral blood lymphocytes was measured in a
9 group of 9 EtO-exposed hospital workers, a group of 15 EtO-exposed factory workers, and their
10 respective controls (Tates et al., 1991). EtO exposure scenarios suggest higher exposures in the
11 factory workers, and this is supported by the measurement of higher hemoglobin adduct levels in
12 those workers. HPRT mutant frequencies were 55% increased in the hospital workers, but the
13 increase was not statistically significant. In the factory workers, a statistically significant
14 increase of 60% was reported. In a study of workers in an EtO production facility (Tates et al.,
15 1995), HPRT mutations were measured in three exposed groups and one unexposed group (seven
16 workers per group). No significant differences in mutant frequencies were observed between the
17 groups; however, the authors stated that about 50 subjects per group would have been needed to
18 detect a 50% increase.
19 Major et al. (2001) measured //Permutations in female nurses employed in hospitals in
20 Eger and Budapest, Hungary. This study and an earlier study measuring effects on chromosomes
21 (see Table 3-6) were conducted to examine a possible causal relationship between EtO exposure
22 and a cluster of cancers (mostly breast) in nurses exposed to EtO in the Eger hospital. The
23 Budapest hospital was chosen because there was no apparent increase in cancer among nurses
24 exposed to EtO. Controls were female hospital workers in the respective cities, and nurses in
25 Eger with known cancers were excluded. Mean peak levels of EtO were 5 mg/m3 (2.7 ppm) in
26 Budapest and 10 mg/m3 (5.4 ppm) in Eger. HPRT variant frequencies in both controls and
27 EtO-exposed workers in the Eger hospital were higher than either group in the Budapest hospital,
28 but there was no significant increase among the EtO-exposed workers in either hospital when
29 compared with the respective controls. The authors noted that the //PPrvariant frequencies
30 among smoking EtO-exposed nurses in Eger were significantly higher than among smokers in
31 the Eger controls; however, the fact that the HPRT variant frequency was almost three times
32 higher in nonsmokers than in smokers in the Eger hospital control group raises questions about
33 the basis of the claimed EtO effect.
34
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-32 DRAFT—DO NOT CITE OR QUOTE
-------
to
Table 3-6. Cytogenetic effects in humans
Number exposed
(number of controls)
33(0)
Site I: 13
Site II: 22
Site III: 25-26
(171 total)
12(12)
14(14)
Factory I: 18
Factory II: 10
(20 total)
15 smokers (7)
10 nonsmokers (15)
10 (10)
Low dose: 9(48)
High dose: 27(10)
34 (23)
1 1 smokers
14 nonsmokers
(10 total)
75 (22)
56(141)
Exposure time
(years)
Range
1-14
0.5-8
0.5-8
0.5-10
0.5-10
3-14
1-10
Mean
3.2
1.7
5.7
4.5
3
4
15
8e
7
Ethylene oxide level in air (ppm)a
Range
±0.05-8
0.5C
5-10c
5-20c
±36
<0.07^.3C
20-123
20-123
60-69C
2.7-10.9
2.7-82
<0.1-2.4C
0.5-417f
0.5-208f
2-5c
1-40C
Mean (TWA)
±0.01b
<1
<1
2.7
5.5
0.3
Cytogenetic observations
CA
(+)
+
+
+
+
+
+
-
+
+
SCE
+
+
+
-
-
+
+
+
+
+
-
+
MN
+d
+
Reference
(Clare etal., 1985)
(Stolley et al., 1984)
(Galloway etal., 1986)
(Garry etal., 1979)
(Hansenetal., 1984)
(Hogstedtetal., 1983)
(Laurent etal., 1984)
(Lerda and Rizzi, 1992)
(Major et al., 1996)
(Mayer etal., 1991)
(Poppetal., 1994)
(Ribeiro et al., 1994)
(Richmond et al., 1985)
§•
§
I
^
'TS
o
a
I
a,
8"
O S
31
o ^
H t
w K-
o
c
o
H
W
-------
to
OJ
Table 3-6. Cytogenetic effects in humans (continued)
Number exposed
(number of controls)
22 (22)
19(19)
10(10)
9
o
3
(27 total)
5
5
(10 total)
32
11
(8 total)
9 hospital workers (8)
15 factory workers (15)
7
7
7
(7 total)
Low exposure: 9
High exposure: 5
(13 total)
19
17
(35 total)
Exposure time
(years)
Range
0.6-4
1.5-15
0.5-12
0.1-4
4-12
2-6
3-27
Accidental
<5
>15
1-5
6-14
Mean
o
J
6.8
5
2
8.6
5.1
9.5
4
12
Ethylene oxide level in air (ppm)a
Range
0.2-0.5C
3.7-20c
0-9.3C
0.025-0.38C
>0.38g
-------
to
o
Table 3-6. Cytogenetic effects in humans (continued)
al ppm = 1.83 mg ethylene oxide/m3.
bCalculated by linear extrapolation.
^ CTWA (8-hr).
S^ dPositive for erythroblasts and polychromatic erythrocytes (negative for lymphocytes).
§- eMaximum years exposed.
s fPeak concentrations.
1 gExposed acutely from sterilizer leakage in addition to chronic exposure.
^ hNasal mucosa.
a 'Buccal cells.
a- JAverage 6-month cumulative exposure (mg).
^ CA = chromosomal aberrations
^ MN = micronucleus
~. SCE = sister chromatid exchange
^ TWA = time-weighted average
O
a
I
a,
§
o s
31
o ^
H t
w K-
o
c
o
H
W
-------
1 3.3.3.3. Chromosomal Effects
2 As discussed by Preston (1999) in an extensive review of the cytogenetic effects of EtO,
3 a variety of cytogenetic assays can be used to measure induced chromosome damage. However,
4 most of the assays commonly employed measure events that are detectable only in the first (or in
5 some cases the second) metaphase after exposure and require DNA synthesis to convert DNA
6 damage into a chromosomal aberration. In addition, DNA repair is operating in peripheral
7 lymphocytes to repair induced DNA damage. Thus, for acute exposures, the timing of sampling
8 is of great importance. For chronic studies, the endpoints measure only the most recent
9 exposures, and if the time between last exposure and sampling is long, any induced DNA
10 damage not converted to a stable genotoxic alteration is certain to be missed. The events
11 measured include all types of chromosomal aberrations, micronuclei, SCE, and numerical
12 chromosomal changes. Stable chromosomal aberrations include reciprocal translocations,
13 inversions, and some fraction of insertions and deletions as well as some numerical changes.
14 However, until the development of fluorescent in situ hybridization (FISH), chromosome
15 banding techniques were needed to detect these types of aberrations.
16 In in vitro assays, EtO has consistently tested positive in studies for multiple types of
17 chromosomal effects, including DNA strand breaks, SCEs, micronuclei, and chromosomal
18 aberrations (e.g., see Table 11 of IARC, 2008a). Of note, Adam et al. (2005) measured the
19 sensitivity of different human cell types to EtO-induced DNA damage using the comet assay,
20 which measures direct strand breaks and/or DNA damage converted to strand breaks during
21 alkaline treatment. Adam et al. (2005) reported dose-dependent increases in DNA damage in the
22 concentration range 0-100 uM in each of the cell types examined with no notable cytotoxicity.
23 At the lowest concentration reported (20 uM), significant increases in DNA damage were
24 observed in lymphoblasts, lymphocytes, and breast epithelial cells, but not in keratinocytes or
25 cervical epithelial cells, suggesting that breast epithelial cells may have increased sensitivity to
26 EtO-induced genotoxicity compared to other nonlymphohematopoietic cell types. In addition,
27 Godderis et al. (2006) investigated the effects of genetic polymorphisms on DNA damage
28 induced by EtO in peripheral blood lymphocytes of 20 nonsmoking university students. No
29 significant increases in micronuclei were observed following EtO treatment; however,
30 dose-related increases in DNA strand breaks were seen in the comet assay. GST polymorphisms
31 did not have a significant impact on the EtO-induced effects; however, significant increases in
32 DNA strand breaks were associated with low-activity alleles of two DNA repair enzymes
33 compared to wild-type alleles.
34 In vivo, several inhalation studies in laboratory animals have demonstrated that EtO
35 exposure levels in the range of those used in the rodent bioassays induce SCEs (see Table 11 of
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-36 DRAFT—DO NOT CITE OR QUOTE
-------
1 IARC, 2008a); however, evidence for micronuclei and chromosomal aberrations from these
2 same exposure levels is less consistent. In particular, studies by van Sittert et al. (2000) and
3 Lorenti Garcia et al. (2001) observed increases in micronuclei and chromosomal aberrations in
4 splenic lymphocytes of rats exposed to 50, 100, or 200 ppm EtO for 6 hours/day, 5 days/week,
5 for 4 weeks compared to levels from control rats, but the increases were not statistically
6 significant. IARC (2008a) noted, however, that "strong conclusions cannot be drawn" from
7 these two studies because the cytogenetic analyses "were initiated 5 days after the final day of
8 exposure, a suboptimal time, and the power of the (FISH) studies were limited by analysis of
9 only a single chromosome and the small numbers of rats per group examined," which was 3 per
10 exposure group in both of the studies, although numerous cells/rat were examined. Moreover, a
11 recent study by Donner et al. (2010) showed clear, statistically significant increases in
12 chromosomal aberrations with longer durations of exposure (>12 weeks) to the concentration
13 levels used in the rodent bioassays.
14 In humans, various studies of occupationally exposed workers have reported SCEs and
15 other chromosomal effects associated with EtO exposure, including micronuclei and
16 chromosomal aberrations. The genotoxicity of EtO was demonstrated in humans as early as
17 1979. Table 3-6 summarizes the cytogenetic effects of EtO on human exposures (see also
18 Appendix C for more details on some of the studies).
19 As illustrated in Table 3-6, numerous studies observed increased SCEs in occupationally
20 exposed workers, especially for workers with the highest exposures (e.g., Major et al., 1996;
21 Sarto et al., 1991; Tates et al., 1991; Sarto et al., 1987). Several studies of occupationally
22 exposed workers have also reported increased micronucleus formation in lymphocytes (Ribeiro
23 et al., 1994; Tates et al., 1991), in nasal mucosal cells (Sarto et al., 1990), and in bone marrow
24 cells (Hogstedt et al., 1983), although this endpoint seems to be less sensitive than SCEs. An
25 association between increased micronucleus frequency and cancer risk has been reported in at
26 least one large prospective general population study (Bonassi et al., 2007). In addition,
27 chromosomal aberrations have been reported in multiple studies of workers occupationally
28 exposed to EtO (Ribeiro et al., 1994; Tates et al., 1991; Sarto et al., 1987). Chromosomal
29 aberrations have been linked to an increased risk of cancer in several large prospective general
30 population studies (e.g., Boffetta et al., 2007; Rossner et al., 2005; Hagmar et al., 2004; Liou et
31 al., 1999).
32
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-37 DRAFT—DO NOT CITE OR QUOTE
-------
1 3.3.3.4. Summary
2 The available data from in vitro studies, laboratory animal models, and epidemiological
3 studies establish that EtO is a mutagenic and genotoxic agent that causes a variety of types of
4 genetic damage.
5
6 3.4. MODE OF ACTION
7 EtO is an alkylating agent that has consistently been found to produce numerous
8 genotoxic effects in a variety of biological systems ranging from bacteriophage to occupationally
9 exposed humans. It is carcinogenic in mice and rats, inducing tumors of the
10 lymphohematopoietic system, brain, lung, connective tissues, uterus, and mammary gland. In
11 addition, epidemiological studies have shown an increased risk of various types of human
12 cancers (see Table A-5 in Appendix A), in particular lymphohematopoietic and breast cancers.
13 Target tissues for EtO carcinogenicity in laboratory animals are varied, and the cancers are not
14 clearly attributable to any specific type of genetic alteration. Although the precise mechanisms
15 by which the multisite carcinogenicity in mice, rats, and humans occurs are unknown, EtO is
16 clearly a mutagenic and genotoxic agent, as discussed in Section 3.3.3, and mutagenicity and
17 genotoxicity are well established as playing a key role in carcinogenicity.
18 Exposure of cells to DNA-reactive agents results in the formation of carcinogen-DNA
19 adducts. The formation of DNA adducts results from a sequence of events involving absorption
20 of the agent, distribution to different tissues, and accessibility of the molecular target (Swenberg
21 et al., 1990). Alkylating agents may induce several different DNA alkylation products (Beranek,
22 1990) with varying proportions, depending primarily on the electrophilic properties of the agent.
23 The predominant DNA adduct formed by EtO is N7-HEG, although other adducts, such as
24 N3-hydroxyethyladenine and O6-HEG, have also been observed, in much lesser amounts
25 (Segerback, 1990). In addition to direct DNA adduct formation via alkylation, Marsden et al.
26 (2009) observed an indirect effect of EtO exposure on endogenous N7-HEG adduct formation
27 and hypothesized that EtO could also indirectly cause adduct formation via oxidative stress (see
28 also Section 3.3.3.1 and Appendix C). The various adducts are processed by different repair
29 pathways, and the subsequent genotoxic responses elicited by unrepaired DNA adducts are
30 dependent on a wide range of variables. The specific adduct(s) responsible for EtO-induced
31 genotoxicity and the mechanism(s) by which this adduct(s) induces the genotoxic damage are
32 unknown.
33 It had been postulated that the predominant EtO-DNA adduct, N7-HEG, although
34 unlikely to be directly promutagenic, could be subject to depurination, resulting in an apurinic
35 site which could be vulnerable to miscoding during cell replication (e.g., Walker and Skopek,
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-38 DRAFT—DO NOT CITE OR QUOTE
-------
1 1993). However, in a study designed to test this hypothesis, Rusyn et al. (2005) failed to detect
2 an accumulation of abasic sites in brain, spleen, and liver tissues of rats exposed to EtO. Rusyn
3 et al. (2005) conclude that the accumulation of abasic sites is unlikely to be a primary
4 mechanism for EtO mutagenicity, although they note that it is also possible that their assay was
5 not sufficiently sensitive to detect small increases in abasic sites or that abasic sites are only
6 mutagenic under conditions of rapid cell turnover, when cell replication may occur before repair
7 of the abasic site (the tissues examined in their study were relatively quiescent). Another
8 potential mechanism for EtO-induced mutagenicity is the direct mutagenicity of the
9 promutagenic adducts such as O6-HEG, although these adducts are generally considered to occur
10 at levels too low to explain all of the observed mutagenicity (IARC, 2008a). In an in vitro study,
11 Tompkins et al. (2009) exposed plasmid DNA to a range of EtO concentrations in water and
12 reported that only the N7-HEG adduct was detectable after exposure to EtO concentrations up to
13 2,000 jiM; at higher EtO concentrations (>10 mM), Nl-hydroxyethyladenine and O6-HEG
14 adducts were also quantifiable but at much lower levels than the N7-HEG adducts. Tompkins et
15 al. (2009) then examined the mutagenicity of these adducts in a supF forward mutation assay and
16 reported that the relative mutation frequencies were significantly elevated only for plasmids
17 exposed to these higher EtO concentrations (see Appendix C, Sections C.I.2 and C.2.2, for a
18 more detailed discussion of this study).
19 The events involved in the formation of chromosomal damage by EtO are similarly
20 unknown. N-alklylated bases are removed from DNA by base excision repair pathways. A
21 review by Memisoglu and Samson (2000) notes that the action of DNA glycosylase and apurinic
22 endonuclease creates a DNA single-strand break, which can in turn lead to DNA double-strand
23 breaks (DSBs). DSBs can also be produced by normal cellular functions, such as during V(D)J
24 recombination in the development of lymphoid cells or topoisomerase II-mediated cleavage at
25 defined sites. A review of mechanisms of DSB repair indicates that the molecular mechanisms
26 are not fully understood (Pfeiffer et al., 2000). This review provides a thorough discussion of
27 both sources (endogenous and exogenous) of DSBs and the variety of repair pathways that have
28 evolved to process the breaks. Although homology-directed repair generally restores the original
29 sequence, during nonhomologous end-joining, the ends of the breaks are frequently modified by
30 addition or deletion of nucleotides. The lack of accumulation of abasic sites observed in the
31 Rusyn et al. (2005) study discussed above argues against a mechanism involving abasic sites as
32 hot spots for strand breaks, although it is possible that abasic sites accumulate more readily in
33 replicating lymphocytes, which were not examined in the study of Rusyn et al. (2005). Another
34 postulated mechanism for EtO-induced strand breaks is via the formation of hydroxyethyl
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-39 DRAFT—DO NOT CITE OR QUOTE
-------
1 adducts on the phosphate backbone of the DNA, but this mechanism requires further study
2 (IARC, 2008a).
3 Lymphohematopoietic malignancies, like all other cancers, are considered to be a
4 consequence of an accumulation of genetic and epigenetic changes involving multiple genes and
5 chromosomal alterations. Although it is clear that chromosome translocations are common
6 features of some hematopoietic cancers, there is evidence that mutations inp53 or NRAS are
7 involved in certain types of leukemia (U.S. EPA, 1997). It should also be noted that
8 therapy-related leukemias exhibiting reciprocal translocations are generally only seen in patients
9 who have previously been treated with chemotherapeutic agents that act as topoisomerase II
10 inhibitors (U.S. EPA, 1997). In NHL, the BCL6 gene is frequently activated by translocations
11 (Chaganti et al., 1998) as well as by mutations within the gene coding sequence (Losses and
12 Levy, 2000). Preudhomme et al. (2000) observed point mutations in theAMLl gene in 9 of
13 22 patients with the MO type (minimally differentiated acute myeloblastic leukemia) of acute
14 myeloid leukemia (AML), and Harada et al. (2003) identified AML1 point mutations in cases of
15 radiation-associated and therapy-related myelodysplastic syndrome (MDS)/AML. In both
16 reports, point mutations within the coding sequence were found in patients with normal
17 karyotypes as well as some with translocations or other chromosomal abnormalities.
18 Zharlyganova et al. (2008) identified AML1 mutations in 7 of 18 radiation-exposed MDS/AML
19 patients but in none of 13 unexposed MDS/AML cases. Other point mutations have also been
20 identified in therapy-related MDS/AML patients, includingp53 gene mutations after exposure to
21 alkylating agents (Christiansen et al., 2001) and mutations in RAS and other genes in the receptor
22 tyrosine kinase signal transduction pathway (Christiansen et al., 2005). Several models have
23 been developed to integrate these various types of genetic alterations. One recent model
24 suggests that the pathogenesis of MDS/AML can be subdivided into at least eight genetic
25 pathways that have different etiologies and different biologic characteristics (Pedersen-Bjergaard
26 et al., 2006).
27 A mode-of-action-motivated modeling approach based solely on chromosome
28 translocations has been proposed by Kirman et al. (2004). The authors suggested a nonlinear
29 dose-response relationship for EtO and leukemia, based on a consideration that "chromosomal
30 aberrations are the characteristic initiating events in chemically induced acute leukemia and gene
31 mutations are not characteristic initiating events." They proposed that EtO must be responsible
32 for two nearly simultaneous DNA adducts, yielding a dose-squared (quadratic) relationship
33 between EtO exposure and leukemia risk. However, as discussed above, there is evidence that
34 does not support the assumption that chromosomal aberrations represent the sole initiating event.
35 In fact, these aberrations or translocations could be a downstream event resulting from genomic
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-40 DRAFT—DO NOT CITE OR QUOTE
-------
1 instability. In addition, it is not clear that acute leukemia is the lymphohematopoietic cancer
2 subtype associated with EtO exposure; in the large NIOSH study, increases in
3 lymphohematopoietic cancer risk were driven by increases in lymphoid cancer subtypes.
4 Furthermore, even if two reactions with DNA resulting in chromosomal aberrations or
5 translocations are early-occurring events in some EtO-induced lymphohematopoietic cancers, it
6 is not necessary that both events be associated with EtO exposure (e.g., background error repair
7 rates or exposure to other alkylating agents may be the cause). Moreover, EtO could also
8 produce translocations indirectly by forming DNA or protein adducts that affect the normally
9 occurring recombination activities of lymphocytes or the repair of spontaneous double-strand
10 breaks. Thus, broader mode-of-action considerations were not regarded as supportive of the
11 hypothesis that the exposure-response relationship is purely quadratic.
12 Breast cancer is similarly considered to be a consequence of an accumulation of genetic
13 and epigenetic changes involving multiple genes and chromosomal alterations (Ingvarsson,
14 1999). Again, the precise mechanisms by which EtO induces breast cancer are unknown. As
15 discussed in Section 3.3.3.2, in a study of mammary gland carcinomas in EtO-exposed mice,
16 Houle et al. (2006) noted that the EtO-induced tumors exhibited a distinct shift in the mutational
17 spectra of thep53 and Hras genes and more commonly displayed concurrent mutations of the
18 two genes. The comet assay results of Adam et al. (2005) suggest that human breast epithelial
19 cells may have increased sensitivity to EtO-induced genotoxicity compared to other
20 nonlymphohematopoietic cell types (see Section 3.3.3.3); however, the basis for any increased
21 sensitivity of breast epithelial cells is similarly unknown.
22 In summary, EtO induces a variety of types of genetic damage. It directly interacts with
23 DNA, resulting in DNA adducts, gene mutations, and chromosome damage. Depending on a
24 number of variables, EtO-induced DNA adducts (1) may be repaired, (2) may result in a
25 base-pair mutation during replication, or (3) may be converted to a DSB, which also may be
26 repaired or result in unstable (micronuclei) or stable (translocation) cytogenetic damage. All of
27 the available data are strongly supportive of a mutagenic mode of action involving gene
28 mutations and chromosomal aberrations (translocations, deletions, or inversions) that critically
29 alter the function of oncogenes or tumor suppressor genes. Although it is clear that chromosome
30 translocations are common features of many hematopoietic cancers, there is evidence that
31 mutations inp53, AML1, or Nras are also involved in some leukemias. The current scientific
32 consensus is that there is very good correspondence between ability of an agent to cause
33 mutations, as does EtO, and carcinogenicity. All of the above scientific evidence provides
34 support for a mutagenic mode of action.
35
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-41 DRAFT—DO NOT CITE OR QUOTE
-------
1 3.4.1. Analysis of the Mode of Action for Ethylene Oxide Carcinogenicity under EPA's
2 Mode-of-Action Framework
3 In this section, the mode of action evidence for EtO carcinogenicity is analyzed under the
4 mode of action framework in EPA's 2005 Guidelines for Carcinogen Risk Assessment (U. S.
5 EPA, 2005a, Section 2.4.3).
6 The hypothesis is that EtO carcinogenicity has a mutagenic mode of action. This
7 hypothesized mode of action is presumed to apply to all of the tumor types.
8 The key events in the hypothesized mutagenic mode of action are DNA adduct formation
9 by EtO, which is a direct-acting alkylating agent, and the resulting genetic damage, including the
10 formation of point mutations as well as chromosomal alterations. Mutagenicity is a well
11 established cause of carcinogenicity.
12
13 1. Is the hypothesized mode of action sufficiently supported in the test animals?
14 Numerous studies have demonstrated that EtO forms protein and DNA adducts, in mice
15 and rats (see Sections 3.3.1 and 3.4 and Figure 3-2). For example, Walker et al. (1992b) and
16 Walker et al. (1992a) demonstrated that EtO forms protein adducts with hemoglobin in the blood
17 and DNA adducts with tissues throughout the body, including in the lung, brain, kidney, spleen,
18 liver, and testes.
19 In addition, there is incontrovertible evidence that EtO is mutagenic (see Section 3.3.3).
20 The evidence is strong and consistent; EtO has invariably yielded positive results in in vitro
21 mutation assays from bacteriophage, bacteria, fungi, yeast, insects, plants, and mammalian cell
22 cultures. The results of in vivo studies on the mutagenicity and genotoxicity of EtO have also
23 been consistently positive following ingestion, inhalation, or injection. Increases in the
24 frequency of gene mutations in the lung, in T-lymphocytes, in bone marrow, and in testes have
25 been observed in transgenic mice exposed to EtO via inhalation at concentrations similar to those
26 in the mouse carcinogenesis bioassays. Furthermore, in a study ofp53 (tumor supressor gene)
27 and Hras (oncogene) mutations in mammary gland carcinomas of EtO-exposed and control
28 mice, Houle et al. (2006) noted that the EtO-induced tumors exhibited a distinct shift in the
29 mutational spectra of thep53 and Hras genes and more commonly displayed concurrent
30 mutations of the two genes, and in a similar study ofKras (oncogene) mutations in lung,
31 Harderian gland, and uterine tumors, substantial increases were observed in Kras mutation
32 frequencies in the tumors from the EtO-exposed mice (Hong et al., 2007).
33 Several inhalation studies in laboratory animals have demonstrated that EtO exposure
34 levels in the range of those used in the rodent bioassays induce SCEs. Evidence for micronuclei
35 and chromosomal aberrations from these same exposure levels has been less consistent;
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3 -42 DRAFT—DO NOT CITE OR QUOTE
-------
1 however, IARC (2008a) has noted analytical limitations with some of these analyses (see
2 Section 3.3.3.3). Moreover, a recent study by Donner et al. (2010) showed clear, statistically
3 significant increases in chromosomal aberrations with exposure durations of >12 weeks to the
4 concentration levels used in the rodent bioassays.
5 Ethylene oxide induces a variety of mutagenic and genotoxic effects, including
6 chromosome breaks, micronuclei, SCEs, and gene mutations; however, the more general effect
7 of mutagenicity/genotoxicity is specific and occurs in the absence of cytotoxicity or other overt
8 toxicity. A temporal relationship is also clearly evident, with adducts and mutagenicity
9 observed in subchronic assays.
10 Dose-response relationships have been observed between EtO exposure in vivo and
11 hemoglobin adducts (e.g., Walker et al., 1992b), as well as DNA adducts, SCEs, and Hprt
12 mutations (e.g., van Sittert et al., 2000) (see also Sections 3.3 and 3.4). A mutagenic mode of
13 action for EtO carcinogen! city also clearly comports with notions of biological plausibility and
14 coherence because EtO is a direct-acting alkylating agent. Such agents are generally capable of
15 forming DNA adducts, which in turn have the potential to cause genetic damage, including
16 mutations; and mutagenicity, in its turn, is a well-established cause of carcinogenicity. This
17 chain of key events is consistent with current understanding of the biology of cancer.
18 In addition to the clear evidence supporting a mutagenic mode of action in test animals,
19 there are no compelling alternative or additional hypothesized modes of action for EtO
20 carcinogenicity.
21
22 2. Is the hypothesized mode of action relevant to humans?
23 The evidence discussed above demonstrates that EtO is a systemic mutagen in test
24 animals; thus, there is the presumption that it would also be a mutagen in humans. Moreover,
25 there is human evidence directly supporting a mutagenic mode of action for EtO carcinogenicity.
26 Several studies of humans have reported exposure-response relationships between hemoglobin
27 adduct levels and EtO exposure levels (e.g., van Sittert et al., 1993; Schulte et al., 1992; see
28 Section 3.3.2), demonstrating the ability of EtO to bind covalently in systemic human cells, as it
29 does in rodent cells. DNA adducts in EtO-exposed humans have not been well studied, and the
30 evidence of increased DNA adducts is limited.
31 In addition, EtO has yielded positive results in in vitro mutagenicity studies of human
32 cells (see Figure 3-3). Although the studies of point mutations in EtO-exposed humans are few
33 and insensitive and the evidence for mutations is limited, there is clear evidence from a number
34 of human studies that EtO causes chromosomal aberrations, SCEs, and micronucleus formation
35 in peripheral blood lymphocytes (see Section 3.3.3.3 and Table 3-6). At least one study
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-43 DRAFT—DO NOT CITE OR QUOTE
-------
1 suggested an exposure-response relationship for the formation of SCEs in peripheral blood
2 lymphocytes (Major et al., 1996). Another study reported a statistically significant increase in
3 micronuclei in bone marrow cells in EtO-exposed workers (Hogstedt et al., 1983).
4 Finally, there is strong evidence that EtO causes cancer in humans, including cancer
5 types observed in rodent studies (i.e., lymphohematopoietic cancers and breast cancer),
6 providing further weight to the relevance of the aforementioned events to the development of
7 cancer in humans (see Sections 3.1 and 3.5.1).
8 In conclusion, the weight of evidence supports a mutagenic mode of action for EtO
9 carcinogenicity.
10
11 3. Which populations or lifestages can be particularly susceptible to the hypothesized mode of
12 action?
13 The mutagenic mode of action is considered relevant to all populations and lifestages.
14 According to EPA's Supplemental Guidance (U.S. EPA, 2005b), there may be increased
15 susceptibility to early-life exposures to carcinogens with a mutagenic mode of action. Therefore,
16 because the weight of evidence supports a mutagenic mode of action for EtO carcinogenicity,
17 and in the absence of chemical-specific data to evaluate differences in susceptibility, increased
18 early-life susceptibility should be assumed and, if there is early-life exposure, the age-dependent
19 adjustment factors should be applied, in accordance with the Supplemental Guidance (see
20 Section 4.4).
21 In addition, as discussed in Section 3.5.2, people with DNA repair deficiencies or genetic
22 polymorphisms conveying a decreased efficiency in detoxifying enzymes may have increased
23 susceptibility to EtO-induced carcinogenicity.
24
25 3.5. HAZARD CHARACTERIZATION
26 3.5.1. Characterization of Cancer Hazard
27 In studies of humans there is substantial evidence that EtO exposure is causally
28 associated with lymphohematopoietic cancer, but the evidence is not strong enough to be
29 conclusive. There is also evidence that EtO exposure is causally associated with breast cancer,
30 but the database for breast cancer is more limited. Of the eight relevant6 Hill "criteria" (or
31 considerations) for causality (Hill, 1965), temporality, coherence, biological plausibility and
32 analogy are readily satisfied, and the other four criteria (specificity, consistency, biological
33 gradient, and strength of association) are satisfied to varying degrees, as discussed below.
6The ninth consideration is experimental evidence, which is seldom available for human populations and is not
available in the case of human exposures to EtO.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3 -44 DRAFT—DO NOT CITE OR QUOTE
-------
1 Temporality, the sole necessary criterion, is satisfied because the subjects of all the
2 epidemiology studies of EtO were workers who were exposed to EtO before the cancers of
3 interest were observed, i.e., exposure preceded the development of the disease.
4 The related criteria of coherence, biological plausibility and analogy are fulfilled by the
5 well-established knowledge that EtO is mutagenic and genotoxic, which are common
6 mechanistic features of many carcinogens; that EtO is carcinogenic in rodents, with
7 lymphohematopoietic cancers being observed in both rats and mice and mammary carcinomas
8 being observed in female mice; and that EtO is an epoxide and epoxides are capable of directly
9 interacting with DNA and are the active metabolites of many carcinogens.
10 There is some specificity with respect to the lymphohematopoietic system. Most of the
11 studies focus on examining risks associated with subcategories of the lymphohematopoietic
12 system. These cancers include leukemia, Hodgkin lymphoma, NHL, reticulosarcoma, and
13 myeloma. (Note that, with the exception of the Steenland et al. (2004) study, which includes
14 lymphocytic leukemia in a lymphoid cancer category, the studies do not subcategorize leukemia
15 into its distinct myeloid and lymphocytic subtypes.) In most of the studies, an enhanced risk of
16 cancer of the lymphohematopoietic system is evident, and in some studies, it is statistically
17 significant. There also appears to be specificity across the epidemiological database for an
18 increased risk of breast cancer. It should be noted, however, that the specificity criterion is not
19 expected to be strictly satisfied by agents, such as EtO, that are widely distributed in all tissues
20 and are direct-acting chemicals.
21 As just alluded to, there is evidence of consistency between studies with respect to cancer
22 of the lymphohematopoietic system as a whole. Most of the available epidemiologic studies of
23 EtO exposure have reported elevated risks of lymphohematopoietic cancer, and the studies that
24 do not report a significant lymphohematopoietic cancer effect have major limitations, such as
25 small numbers of cases (from small study size and/or insufficient follow-up time), inadequate
26 exposure information, and/or reliance on external analyses (see Table 3-1 and Table A-5 in
27 Appendix A). Overall, about 9 of 11 studies (including only the last follow-up of independent
28 cohorts) with adequate information to determine RR estimates reported an increased risk of
29 lymphohematopoietic cancers or a subgroup thereof, although not all were statistically
30 significant, possibly due to the limitations noted above (see Table 3-1 and Table A-5 in
31 Appendix A). The large, high-quality NIOSH study shows statistically significant
32 exposure-response trends for lymphoid cancers and all lymphohematopoietic cancers (Steenland
33 et al., 2004; see Sections D.3 and D.4 of Appendix D for results for both sexes combined). Four
34 other studies reported statistically significant increases in risk (Swaen et al., 1996; Benson and
35 Teta, 1993; Bisanti et al., 1993; Hogstedt, 1988; Hogstedt et al., 1986), although EtO exposures
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-45 DRAFT—DO NOT CITE OR QUOTE
-------
1 were reportedly low in the Benson and Teta (1993) study and the increased risks may be due to
2 other chemical exposures. Nonsignificant increases in lymphohematopoietic cancer risk were
3 observed in four other studies, based on small numbers of cases (Coggon et al., 2004; Olsen et
4 al., 1997; Hagmar et al., 1995; Norman et al., 1995 [with only 1 case]; Hagmar et al., 1991).
5 Only 2 of the 11 studies showed no evidence of an increase in lymphohematopoietic cancer risk
6 (Swaen et al., 2009; Kiesselbach et al., 1990).
7 Regarding consistency in the breast cancer studies, the large, high-quality NIOSH study
8 shows statistically significant increased risks for both breast cancer mortality (n = 103 deaths;
9 (Steenland et al., 2004)) and breast cancer incidence (n = 319 cases; (Steenland et al., 2003)).
10 Two other studies suggest an increased risk of breast cancer despite their small size ((Norman et
11 al., 1995), n = 12 cases; (Kardos et al., 2003), n = 3 deaths). No elevated risks were seen in the
12 only other two studies reporting breast cancer results; however these studies had few cases,
13 owing to their small size and/or inadequate follow-up time ((Hagmar et al., 1991), n = 5 cases;
14 (Coggon et al., 2004), n = 11 deaths) (see Table 3-2 and Table A-5 in Appendix A).
15 There is also some evidence of dose-response relationships (biological gradient). In the
16 large, high-quality NIOSH study, a statistically significant positive trend was observed in the risk
17 of lymphohematopoietic cancers with increasing (log) cumulative exposure to EtO, although
18 results for this model were reported only for males (Steenland et al., 2004) [the sex difference is
19 not statistically significant, however, and the trend for both sexes combined is also statistically
20 significant (see Tables D-3e and D-4e in Appendix D)]. For only two other cohorts were results
21 for exposure-response analyses reported, probably because most cohorts had too few cases
22 and/or lacked adequate exposure information. In the Swaen et al. (2009) study of the UCC
23 cohort, no statistically significant trends were observed for leukemia or lymphoid cancer using a
24 Cox proportional hazards model with cumulative exposure, a model which notably did not yield
25 statistically significant trends in the NIOSH study, either. In the small study of Hagmar et al.
26 (1995), an SIR for leukemia of 7.14 was reported for subjects with at least 0.14 ppm x years of
27 cumulative exposure and 10 years latency, but this result was based on only two cases and was
28 not statistically significant. For breast cancer, exposure-response analyses were reported only for
29 the NIOSH cohort, again presumably because most cohorts had too few cases and/or lacked
30 adequate exposure information. These analyses yielded clear, statistically significant trends for
31 both breast cancer mortality (Steenland et al., 2004) and breast cancer incidence (Steenland et al.,
32 2003) for a variety of models.
33 Whereas most of the considerations are largely satisfied, as discussed above, there is little
34 strength in the associations, as reflected by the modest magnitude of most of the RR estimates.
35 For example, in the large NIOSH study, the RR estimate for lymphoid cancer mortality in the
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-46 DRAFT—DO NOT CITE OR QUOTE
-------
1 highest exposure quartile is about 3.0 and the RR estimate for breast cancer incidence in the
2 highest exposure quintile in the subcohort with interviews is on the order of 1.9. While large RR
3 estimates increase the confidence that an observed association is not likely due to chance, bias,
4 or confounding, modest RR estimates, such as those observed with EtO, do not preclude a causal
5 association (U.S. EPA, 2005a). With EtO, the modest RR estimates may, in part, reflect the
6 relatively high background rates of these cancers, particularly of breast cancer incidence.
7 In addition to the Hill criteria, other factors such as chance, bias, and confounding are
8 considered in analyzing the weight of epidemiological evidence. Given the consistency of the
9 findings across studies and the exposure-response relationships observed in the largest study,
10 none of these factors is likely to explain the associations between these cancers and EtO
11 exposure. Coexposures to other chemicals are expected to have occurred for workers in the
12 chemical industry cohorts but would have been much less likely in the sterilizer worker cohorts,
13 such as the NIOSH cohort, which reported no evidence of confounding exposures to other
14 occupational carcinogens (Steenland et al., 1991). For breast cancer in the NIOSH subcohort
15 with interviews (Steenland et al., 2003), other risk factors for breast cancer were assessed, and
16 statistically significant factors were included in the exposure-response models.
17 In conclusion, the overall epidemiological evidence for a causal association between EtO
18 exposure and lymphohematopoietic cancer was judged to be strong but less than conclusive. For
19 breast cancer, the existing evidence was strong, but there were few studies and, thus, overall, the
20 epidemiological evidence was judged to be more limited.
21 There is inadequate evidence for other cancer types (e.g., stomach cancer and pancreatic
22 cancer) in the epidemiology studies.
23 The experimental animal evidence for carcinogenicity is concluded to be "sufficient"
24 based on findings of tumors at multiple sites, by both oral and inhalation routes of exposure, and
25 in both sexes of both rats and mice. Tumor types resulting from inhalation exposure included
26 mononuclear cell leukemia in male and female rats and malignant lymphoma and mammary
27 carcinoma in female mice, suggesting some site concordance with the lymphohematopoietic and
28 breast cancers observed in humans, also exposed by inhalation.
29 The evidence of EtO genotoxicity and mutagenicity is unequivocal. EtO is a
30 direct-acting alkylating agent and has invariably tested positive in in vitro mutation assays from
31 bactedophage, bacteria, fungi, yeast, insects, plants, and mammalian cell cultures (including
32 human cells). In mammalian cells (including human cells), EtO-induced genotoxic effects
33 include unscheduled DNA synthesis, gene mutations, SCEs, and chromosomal aberrations. The
34 results of in vivo genotoxicity studies of EtO have also been largely positive, following
35 ingestion, inhalation, or injection. Increases in frequencies of gene mutations have been reported
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-47 DRAFT—DO NOT CITE OR QUOTE
-------
1 in the lung, T-lymphocytes, bone marrow, and testes of EtO-exposed mice. In particular,
2 increases in frequencies of oncogene mutations have been observed in several tumor types from
3 EtO-exposed mice compared to spontaneous mouse tumors of the same types. Several inhalation
4 studies in laboratory animals have demonstrated that EtO exposure levels in the range of those
5 used in the rodent bioassays (i.e., 10-100 ppm, 6-7 hours/day, 5 days/week) induce SCEs.
6 Evidence for micronuclei and chromosomal aberrations from these same exposure levels in
7 short-term studies (4 weeks or less) is less consistent, although concerns have been raised about
8 some of the negative studies. A recent study showed clear, statistically significant increases in
9 chromosomal aberrations with longer durations of exposure (>12 weeks) to the concentration
10 levels used in the rodent bioassays. The studies of point mutations in EtO-exposed humans are
11 few and insensitive and the evidence for mutations is limited; however, there is clear evidence
12 from a number of human studies that EtO causes chromosomal aberrations, SCEs, and
13 micronucleus formation in peripheral blood lymphocytes, and one study has reported increased
14 levels of micronuclei in bone marrow cells in EtO-exposed workers.
15 Under EPA's 2005 Guidelines for Carcinogen Risk Assessment (U.S. EPA, 2005a), the
16 conclusion can be made that EtO is "carcinogenic to humans." In general, the descriptor
17 "carcinogenic to humans" is appropriate when there is convincing epidemiologic evidence of a
18 causal association between human exposure and cancer. This descriptor is also appropriate when
19 there is a lesser weight of epidemiologic evidence that is strengthened by specific lines of
20 evidence set forth in the Guidelines, which are satisfied for EtO. The lines of evidence
21 supporting the characterization of "carcinogenic to humans" include the following: (1) there is
22 strong, although less than conclusive on its own, evidence of cancer in humans associated with
23 EtO exposure via inhalation, specifically, evidence of lymphohematopoietic cancers and female
24 breast cancer in EtO-exposed workers; (2) there is extensive evidence of EtO-induced
25 carcinogenicity in laboratory animals, including lymphohematopoietic cancers in rats and mice
26 and mammary carcinomas in mice following inhalation exposure; (3) EtO is a direct-acting
27 alkylating agent whose mutagenic and genotoxic capabilities have been well established in a
28 variety of experimental systems, and a mutagenic mode of carcinogenic action has been
29 identified in animals involving the key precursor events of DNA adduct formation and
30 subsequent DNA damage, including point mutations and chromosomal effects; and (4) there is
31 strong evidence that the key precursor events are anticipated to occur in humans and progress to
32 tumors, including evidence of chromosome damage, such as chromosomal aberrations, SCEs,
33 and micronuclei in EtO-exposed workers.
34
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-48 DRAFT—DO NOT CITE OR QUOTE
-------
1 3.5.2. Susceptible Life stages and Populations
2 There are no data on the relative susceptibility of children and the elderly when compared
3 with adult workers, in whom the evidence of hazard has been gathered, but because EtO does not
4 have to be metabolized before binding to DNA and proteins, the maturing of enzyme systems in
5 very young children is thought not to be a predominant factor in its hazard, at least for activation.
6 However, the immaturity of detoxifying enzymes in very young children may increase children's
7 susceptibility because children may clear EtO at a slower rate than adults. As discussed in
8 Section 3.3.1, EtO is metabolized (i.e., detoxified) primarily by hydrolysis in humans but also by
9 glutathione conjugation. Both hydrolytic activity and glutathione-S-transferase activity
10 apparently develop after birth (Clewell et al., 2002); thus, very young children might have a
11 decreased capacity to detoxify EtO compared to adults. In the absence of data on the relative
12 susceptibility associated with EtO exposure in early life, increased early-life susceptibility is
13 assumed, in accordance with EPA's Supplemental Guidance (U.S. EPA, 2005b), because the
14 weight of evidence supports the conclusion of a mutagenic mode of action for EtO
15 carcinogenicity (see Section 3.4).
16 Other than the occurrence of sex-specific cancers (e.g., breast cancer in human females,
17 mammary and uterine carcinomas in female mice, and testicular peritoneal mesotheliomas in
18 male rats; see Section 3.2), there is no clear sex difference in EtO-induced carcinogenicity. With
19 the exception of the sex-specific cancers and the observation of malignant lymphomas in female
20 but not male mice, there is no sex difference in EtO-induced cancer types in the rat and mouse
21 bioassays. Cancer potency estimates for females are roughly 50% higher than those for males
22 for both mice and rats (see Table 4-18 in Section 4.2.5). In humans, in the large NIOSH study
23 (Steenland et al., 2004), the association between lymphoid cancers and EtO exposure was seen
24 primarily in males, but the sex difference was not statistically significant (see Appendix D), and
25 the SAB panel that reviewed an earlier draft of this assessment recommended that data from both
26 sexes be combined for the derivation of quantitative risk estimates for the lymphohematopoietic
27 cancers (SAB, 2007).
28 Brown et al. (1996) reported that sex differences in EtO toxicokinetics were observed in
29 mice but not in rats; female mice had a significantly higher steady-state blood EtO concentration
30 after 4 hours of exposure to either 100 or 330 ppm than male mice. As noted above and
31 discussed in Section 3.3.1, EtO is metabolized primarily by hydrolysis in humans. Mertes et al.
32 (1985) reported no sex difference in microsomal or cytosolic epoxide hydrolase activities in
33 human liver in vitro using benzo[a]pyrene 4,5-oxide or trans-stilbene oxide, respectively, as
34 substrates. Using EtO as a substrate, but with far fewer subjects, Fennell and Brown (2001)
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-49 DRAFT—DO NOT CITE OR QUOTE
-------
1 reported similar values for males and females for epoxide hydrolase activity in human liver
2 microsomes and for GSH transferase in human liver cytosolic fractions.
3 Because EtO is detoxified by glutathione conjugation or hydrolysis, people with
4 genotypes conveying deficiencies in glutathione-S-transferase or epoxide hydrolase activities
5 may be at increased risk of cancer from EtO exposure. Yong et al. (2001) measured
6 approximately twofold greater EtO-hemoglobin adduct levels in occupationally exposed persons
7 with a null GSTT1 genotype than in those with positive genotypes. Similarly, in a study of
8 hospital workers, Haufroid et al. (2007) reported increased urinary excretion of a glutathione
9 conjugate of EtO, reflecting increased detoxification of EtO, associated with a nonnull GSTT1
10 genotype, although the increase was not statistically significant in all the regression models
11 tested; associations were less clear for other glutathione-S-transferase or epoxide hydrolase
12 polymorphisms.
13 In addition, people with DNA repair deficiencies such as xeroderma pigmentosum,
14 Bloom's syndrome, Fanconi anemia, and ataxia telangiectasia (Gelehrter et al., 1990) are
15 expected to be especially sensitive to the damaging effects of EtO exposure. Paz-y-Mifio et al.
16 (2002) have recently identified a specific polymorphism in the excision repair pathway gene
17 hMSH2. The polymorphism was present in 7.5% of normal individuals and in 22.7% of NHL
18 patients, suggesting that this polymorphism may be associated with an increased risk of
19 developing NHL.
20
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 3-50 DRAFT—DO NOT CITE OR QUOTE
-------
1 4. CANCER DOSE-RESPONSE ASSESSMENT FOR INHALATION EXPOSURE
2 This chapter presents the derivation of cancer unit risk estimates from human and rodent
3 data. Section 4.1 discusses the derivation of unit risk estimates for lymphohematopoietic
4 cancers, breast cancer, and total cancer from human data, as well as sources of uncertainty in
5 these estimates. Section 4.2 presents the derivation of unit risk estimates from rodent data.
6 Section 4.3 summarizes the unit risk estimates derived from the different data sets. Section 4.4
7 discusses adjustments for assumed increased early-life susceptibility, based on recommendations
8 from EPA's Supplemental Guidance (U.S. EPA, 2005b), because the weight of evidence supports
9 the conclusion of a mutagenic mode of action for EtO carcinogenicity (see Section 3.4).
10 Section 4.5 presents conclusions about the unit risk estimates. Section 4.6 compares the unit risk
11 estimates derived in this EPA assessment to those derived in other assessments. Finally,
12 Section 4.7 provides risk estimates derived for some general occupational exposure scenarios.
13
14 4.1. INHALATION UNIT RISK ESTIMATES DERIVED FROM HUMAN DATA
15 The NIOSH retrospective cohort study of more than 18,000 workers in 13 sterilizing
16 facilities (most recent update by Steenland et al., 2004; Steenland et al., 2003) provides the most
17 appropriate data sets for deriving quantitative cancer risk estimates in humans for several
18 reasons: (1) exposure estimates were derived for the individual workers using a comprehensive
19 exposure assessment, (2) the cohort was large and diverse (e.g., 55% female), and (3) there was
20 little reported exposure to chemicals other than EtO. Exposure estimates, including estimates for
21 early exposures for which no measurements were available, were determined using a regression
22 model that estimated exposures to each individual as a function of facility, exposure category,
23 and time period. The regression model was based on extensive personal monitoring data from
24 18 facilities spanning a number of years as well as information on factors influencing exposure,
25 such as engineering controls (Hornung et al., 1994; see also Section A.2.8 in Appendix A).
26 When evaluated against test data, the model accounted for 85% of the variation in average EtO
27 exposure levels. The investigators were then able to estimate the cumulative exposure
28 (ppm x days) for each individual worker by multiplying the estimated exposure for each job
29 (exposure category) held by the worker by the number of days spent in that job and summing
30 over all the jobs held by the worker. Steenland et al. (2004) present follow-up results for the
31 cohort mortality study previously discussed by Steenland et al. (1991) and Stayner et al. (1993).
32 Positive findings in the current follow-up include increased rates of (lympho)hematopoietic
33 cancer mortality and of breast cancer mortality in females. Steenland et al. (2003) present results
34 of a breast cancer incidence study of a subcohort of 7,576 women from the NIOSH cohort.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-1 DRAFT—DO NOT CITE OR QUOTE
-------
1 The other major occupational study (most recent update by Swaen et al., 2009) described
2 risks to Union Carbide workers exposed to EtO at two chemical plants in West Virginia, but this
3 study is less useful for estimating quantitative cancer risks for a number of reasons. First, the
4 exposure assessment is much less extensive than that used for the NIOSH cohort, with greater
5 likelihood for exposure misclassification, especially in the earlier time periods when no
6 measurements were available (1925-1973). Exposure estimation for the individual workers was
7 based on a relatively crude exposure matrix which cross-classified 3 levels of exposure intensity
8 with 4 time periods. The exposure estimates for 1974-1988 were based on measurements from
9 air sampling at the West Virginia plants since 1976. The exposure estimates for 1957-1973 were
10 based on measurements in a similar plant in Texas. The exposure estimates for 1940-1956 were
11 based loosely on "rough" estimates reported for chlorohydrin-based EtO production in a Swedish
12 facility in the 1940s. The exposure estimates for 1925-1939 were essentially guesses. Thus, for
13 the two earliest time periods (1925-1939 and 1940-1956) at least, the exposure estimates are
14 highly uncertain. (See Section A.2.20 of Appendix A for a more detailed discussion of the
15 exposure assessment for the Union Carbide cohort.) This is in contrast to the NIOSH exposure
16 assessment in which exposure estimates were based on extensive sampling data and regression
17 modeling. In addition, the sterilization processes used by the NIOSH cohort workers were fairly
18 constant back in time, unlike chemical production processes, which likely involved much higher
19 and more variable exposure levels in the past. Furthermore, the Union Carbide cohort is of much
20 smaller size and has far fewer deaths than the NIOSH cohort, it is restricted to males and so
21 cannot be used to investigate breast cancer risk in females, and there are coexposures to other
22 chemicals.
23 A third study (Hagmar et al., 1995; Hagmar et al., 1991) estimated cumulative exposures
24 for individual workers; however, insufficient exposure-response data are presented for the
25 derivation of unit risk estimates. Exposure-response results for specific cancers are provided
26 only in the 1991 paper and then only for two lymphohematopoietic cancers across two
27 categorical exposure groups.
28 Table 4-1 provides a summary of the judgments made in selecting the NIOSH study as
29 the basis for the derivation of unit risk estimates. The NIOSH EtO cohort data can be obtained
30 from the Industrywide Studies Branch of NIOSH.7
31 The derivation of unit risk estimates, defined as the lifetime risk of cancer from chronic
32 inhalation of EtO per unit of air concentration, for lymphohematopoietic cancer mortality and
33
'industrywide Studies Branch; Division of Surveillance, Hazard Evaluations and Field Studies: NIOSH;
Centers for Disease Control and Prevention, 4676 Columbia Parkway MS R-13,Cincinnati, Ohio 45226, telephone:
513-841-4203.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-2 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
Table 4-1. Considerations used in this assessment for selecting epidemiology
studies for quantitative risk estimation
Consideration
Studies Selected
Comments
Availability of
quantitative exposure
estimates
Hagmar et al. (1995) and Hagmar
etal. (1991) [Swedish sterilizer
cohort]
Swaen et al. (2009) [latest
follow-up of Union Carbide
Corporation (UCC) cohort]
Steenland et al. (2004) and
Steenland et al. (2003) [latest
follow-up of NIOSH cohort]
These are the only 3 studies with quantitative
exposure estimates, which is an essential
criterion for quantitative risk estimation.
Availability of exposure-
response information
1. Swaen et al. (2009)
2. Steenland et al. (2004) and
Steenland et al. (2003)
Hagmar et al. (1995) and Hagmar et al. (1991)
did not present sufficient exposure-response
results, presumably because they had a short
follow-up time and thus few cases of specific
cancers (5 breast cancer cases;
6 lymphohematopoietic cancer cases).
Other factors affecting
the utility of
epidemiology studies for
quantitative risk
estimation
Steenland et al. (2004) and Steenland
et al. (2003)
The NIOSH study [Steenland et al. (2004) and
Steenland et al. (2003)] alone was selected for
quantitative risk estimation, as it was judged to
be substantially superior to the UCC study
(Swaen et al., 2009) with respect to a number of
key considerations [in particular, in order of
importance: (1) quality of the exposure
estimates, (2) cohort size, and (3) the absence of
coexposures and the inclusion of women].
4
5
6
7
8
9
10
11
12
13
14
incidence and for breast cancer mortality and incidence in females, based on results of the recent
analyses of the NIOSH cohort, is presented in the following subsections.
The exposure-response models used to fit the epidemiological data are empirical
"curve-fitting" models. Considerations used in the selection of the exposure-response models
upon which to base the derivation of unit risk estimates included statistical fit (as reflected by
^-values), visual fit of the models to the categorical results, and biological plausibility. When
multiple models were deemed to be reasonable candidates for selection based on those
o
considerations, AIC was also considered in selecting the "preferred" model.
Akaike Information Criteria. The AIC is a measure of information loss from a dose-response model that can be
used to compare a specified set of models. The AIC is defined as 2p - 21n(L), where p is the number of estimated
parameters included in the model and L is the maximized value of the likelihood function. Among a set of specified
models, the model with the lowest AIC is the preferred model.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-3 DRAFT—DO NOT CITE OR QUOTE
-------
1 4.1.1. Risk Estimates for Lymphohematopoietic Cancer
2 4.1.1.1. Lymphohematopoietic Cancer Results From the NIOSH Study
3 Steenland et al. (2004) investigated the relationship between (any) EtO exposure and
4 mortality from cancer at a number of sites using life-table analyses with the U.S. population as
5 the comparison population. Categorical SMR analyses were also done by quartiles of cumulative
6 exposure. Then, to further investigate apparent exposure-response relationships observed for
7 (lympho)hematopoietic cancer and breast cancer, internal exposure-response analyses were
8 conducted using Cox proportional hazards models, which have the form
9
10
11 Relative rate (RR) = epx, (4-1)
12
13
14 where P represents the regression coefficient and X is the exposure (or some function of
15 exposure, e.g., the natural log of exposure). Internal analyses were done two ways—with
16 exposure as a categorical variable and with exposure as a continuous variable. A nested
17 case-control approach was used, with age as the time variable used to form the risk sets. Risk
18 sets were constructed with 100 controls randomly selected for each case from the pool of those
19 surviving to at least the age of the index case. According to the authors, use of 100 controls per
20 case has been shown to result in ORs virtually identical to the RR estimates obtained with full
21 cohorts. Cases and controls were matched on race (white/nonwhite), sex, and date of birth
22 (within 5 years). Exposure was the only covariate in the model, so the/>-value for the model also
23 serves as a/?-value for the regression coefficient, P, as well as for a test of exposure-response
24 trend.
25 For lymphohematopoietic cancer mortality, Steenland et al. (2004) analyzed both all
26 lymphohematopoietic cancers combined and a subcategory of lymphohematopoietic cancers that
27 they called "lymphoid" cancers; these included NHL, myeloma, and lymphocytic leukemia.
28 Their exposure-response analyses focused on cumulative exposure and (natural) log cumulative
29 exposure, with various lag periods. Other EtO exposure metrics (duration of exposure, average
30 exposure, and peak exposure) were also examined, but models using these metrics did not
31 generally predict lymphohematopoietic cancer as well as models using cumulative exposure. A
32 lag period defines an interval before death, or end of follow-up, during which any exposure is
33 disregarded because it is not considered relevant to the outcome under investigation. For
34 lymphohematopoietic (and lymphoid) cancer mortality, a 15-year lag provided the best fit to the
35 data, based on the likelihood ratio test. One ppm x day was added to cumulative exposures in
36 lagged analyses to avoid taking the log of 0. For both all lymphohematopoietic and lymphoid
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-4 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
8
9
10
11
12
13
cancers, Steenland et al. (2004) found stronger positive exposure-response trends in males and so
presented the results for some of the regression models separately by sex. The apparent sex
difference was not statistically significant (see Appendix D), however, and results for both sexes
combined were subsequently obtained from Dr. Steenland (see Appendix D; Section 3 for
lymphoid cancer, Section 4 for all lymphohematopoietic cancer). These results are presented in
Table 4-2. For additional details and discussion of the Steenland et al. (2004) study, see
Appendix A.
Table 4-2. Cox regression results for all lymphohematopoietic cancer and
lymphoid cancer mortality in both sexes in the NIOSH cohort, for the models
presented by Steenland et al. (2004)
Exposure variable"
p valueb
Coefficient (SE)
ORs by category0 (95% CI)
All lymphohematopoietic cancerd
Cumulative exposure, 15-yr
lag
Log cumulative exposure, 15-
yr lag
Categorical cumulative
exposure, 15-yr lag
0.40
0.009
0.10
0.00000326
(0.00000349)
0.107(0.0418)
1.00, 2.33 (0.93-5.86), 3.46 (1.33-8.95),
3.02 (1.16-7.89), 2.96 (1.12-7.81)
Lymphoid cancer"
Cumulative exposure, 15-yr
lag
Log cumulative exposure, 15-
yr lag
Categorical cumulative
exposure, 15-yr lag
0.22
0.02
0.21
0.00000474
(0.00000335)
0.112(0.0486)
1.00, 1.75 (0.59-5.25), 3.15 (1.04-9.49),
2.44 (0.80-7.50), 3.00 (1.02-8.45)
14
15
16
17
18
19
20
21
22
23
24
25
26
27
""Cumulative exposure is in ppm x days.
V-values from likelihood ratio test.
"Exposure categories are 0, >0-1,199, 1,200-3,679, 3,680-13,499, >13,500 ppm x days.
d9th revision ICD codes 200-208; results based on 74 cases.
eNHL, myeloma, and lymphocytic leukemia (9th revision ICD codes 200, 202, 203, 204); results based on 53 cases.
Source: Additional analyses performed by Dr. Steenland (see Sections D.3 and D.4 of Appendix D).
4.1.1.2. Prediction of Lifetime Extra Risk of Lymphohematopoietic Cancer Mortality
The exposure-response trends for lymphohematopoietic cancers observed by Steenland et
al. (2004) appear to be driven largely by the lymphoid cancers; therefore, the primary risk
analyses for lymphohematopoietic cancer are based on the lymphoid cancer results.
Lymphohematopoietic cancers are a diverse group of diseases with diverse etiologies, and
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-5 DRAFT—DO NOT CITE OR QUOTE
-------
1 myeloid and lymphoid cells develop from different progenitor cells; thus, there is stronger
2 support for an etiologic role of EtO in the development of lymphoid cancers than in the
3 development of the cancers in the aggregate all lymphohematopoietic cancer category. The
4 consideration of NHL, (plasma cell) myeloma, and lymphocytic leukemia together as
5 "lymphoid" cancers is consistent with the current World Health Organization classification of
6 such cancers based on their derivation from B-cells, T-cells, and NK-cells rather than previous
7 distinctions (Harris et al., 1999). Nonetheless, for comprehensiveness and for the reasons listed
8 below, risk estimates based on the all lymphohematopoietic cancer results are presented for
9 comparison. Judging roughly from the ^-values, the model fits do not appear notably better for
10 lymphoid cancers than for all lymphohematopoietic cancers (see Table 4-2, ^-values for log
11 cumulative exposure models), and the "lymphoid" category did not include Hodgkin lymphoma,
12 which also exhibited evidence of exposure-response trends, although based on few cases
13 (Steenland et al., 2004). In addition, misclassification or nonclassification of tumor type is more
14 likely to occur for subcategories of lymphohematopoietic cancer (e.g., 4 of the 25 leukemias in
15 the analyses were classified as "not specified" and so could not be considered for the lymphoid
16 cancer analysis).
17 The results of internal exposure-response analyses of lymphoid cancer in the NIOSH
18 cohort (Cox regression analyses, summarized in Table 4-2) were used for predicting the extra
19 risks of lymphoid cancer mortality from continuous environmental exposure to EtO. Extra risk
20 is defined as
21
22
23 Extra risk = (Rx - Ro)/(l - Ro), (4-2)
24
25
26 where Rx is the lifetime risk in the exposed population and Ro is the lifetime risk in an
27 unexposed population (i.e., the background risk). These risk estimates were calculated using the
28 p regression coefficients and an actuarial program (life-table analysis) that accounts for
29 competing causes of death.9 An inherent assumption in the Cox regression model and its
30 application in the life-table analyses is that RR is independent of age. (An alternate assumption
31 of increased susceptibility from early-life exposure to EtO, as recommended in EPA's
32 Supplemental Guidance (U.S. EPA, 2005b) for chemicals, such as EtO [see Section 3.4], with a
33 mutagenic mode of action, is considered in Section 4.4. This alternate assumption is the
9This program is an adaptation of the approach previously used by the Committee on the Biological Effects of
Ionizing Radiation (BEIR, 1988). A spreadsheet illustrating the extra risk calculation for the derivation of the LEC0i
for lymphoid cancer incidence (see Section 4.1.1.3) is presented in Appendix E.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-6 DRAFT—DO NOT CITE OR QUOTE
-------
1 prevailing assumption in this assessment, based on the recommendations in the Supplemental
2 Guidance. Risk estimates are first developed under the assumption of age independence,
3 however, because that is the standard approach in the absence of evidence to the contrary or of
4 sufficient evidence of a mutagenic mode of action to invoke the divergent assumption of
5 increased early-life susceptibility.)
6 U.S. age-specific all-cause mortality rates for 2004 for both sexes of all race groups
7 combined (Arias, 2007) were used to specify the all-cause background mortality rates in the
8 actuarial program. For the cause-specific background mortality rates for lymphoid cancers,
9 age-specific mortality rates for the relevant subcategories of lymphohematopoietic cancer (NHL
10 [C82-C85 of 10th revision of the International Classification of Diseases (ICD)], multiple
11 myeloma [C88, C90], and lymphoid leukemia [C91]) for the year 2004 were obtained from the
12 National Center for Health Statistics Data Warehouse website
13 (http://www.cdc.gov/nchs/datawh/statab/unpubd/mortabs.htm). The risks were computed up to
14 age 85 for continuous exposures to EtO beginning at birth.10 Conversions between occupational
15 EtO exposures and continuous environmental exposures were made to account for differences in
16 the number of days exposed per year (240 vs. 365 days) and in the amount of EtO-contaminated
17 air inhaled per day (10 vs. 20 m3; (U.S. EPA, 1994)). An adjustment was also made for the lag
18 period. The reported standard errors for the regression coefficients from Table 4-2 were used to
19 compute the 95% upper confidence limits (UCLs) for the relative rates, based on a normal
20 approximation.
21 The only statistically significant Cox regression model presented by Steenland et al.
22 (2004) for lymphoid cancer mortality in males was for log cumulative exposure with a 15-year
23 lag (p = 0.02). This was similarly true for the analyses of lymphoid cancer using the data for
24 both sexes (see Table 4-2). However, using the log cumulative exposure model to estimate the
25 risks from low environmental exposures is problematic because this model, which is intended to
26 fit the full range of occupational exposures in the study, is inherently supralinear (i.e., risk
27 increases steeply with increasing exposures in the low exposure range and then plateaus), and
28 results are unstable for low exposures (i.e., small changes in exposure correspond to large
29 changes in risk; see Figure 4-1). Some consideration was thus given to the cumulative exposure
30 model, which is typically used and which is stable at low exposures, although the fit to these data
31 was not statistically significant (p = 0.22). However, the Cox regression model with cumulative
10Rates above age 85 years are not included because cause-specific disease rates are less stable for those ages. Note
that 85 years is not employed here as an average lifespan but, rather, as a cutoff point for the life-table analysis,
which uses actual age-specific mortality rates. The average lifespan for males and females combined in a life-table
analysis truncated at age 85 years is about 75 years.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-7 DRAFT—DO NOT CITE OR QUOTE
-------
to
OJ
o
rs
S
1
I
^
'TS
o
o
a
i
EN
**ij
8"
_ . _ e"(p*exp)
----- e"(p*logexp)
• categorical
^— — splinelOO
• spline1600
5000
10000 15000 20000 25000 30000
mean cumulative exposure (ppm* days)
35000
40000
45000
3
o
H
W
o
c
o
H
W
1
Figure 4-1. RR estimate for lymphoid cancer vs. mean exposure (with 15-year lag, unadjusted for continuous
exposure).
*exp): Cox regression results for RR = e x exP°sure; eA
= (p x exP°sure)
*logexp): Cox regression results for RR = e
= (p x ln(exP°sure»
categorical: Cox regression results for RR = e x exposure) with categorical exposures; linear: weighted linear regression
of categorical results, excluding highest exposure group (see text); splinelOO(1600): 2-piece log-linear spline model
with knot at 100 (1,600) ppm x days (see text).
Source: Steenland reanalyses for male and female combined; see Appendix D (except for linear regression of
categorical results, which was done by EPA).
-------
1 exposure is inherently sublinear (i.e., risk increases gradually in the low exposure range and then
2 with increasing steepness as exposure increases) and does not reflect the apparent supralinearity
3 of the data demonstrated by the categorical results and the superior fit of the log cumulative
4 exposure model; thus, this model was not considered further. (Note that all the models discussed
5 in Chapter 4 treat exposure as a continuous variable except for the categorical models and the
6 linear regressions of categorical data, which are specifically described as such.)
7 In a 2006 external review draft of this assessment (U.S. EPA, 2006a), which relied on the
8 original published results of Steenland et al. (2004), EPA proposed that the best way to represent
9 the exposure-response relationship in the lower exposure region, which is the region of interest
10 for low-exposure extrapolation, was through the use of a weighted linear regression of the results
11 from the Cox regression model with categorical cumulative exposure and a 15-year lag (for
12 males only, as this was the significant finding in the published paper of Steenland et al., 2004).
13 In addition, the highest exposure group was not included in the regression to alleviate some of
14 the "plateauing" in the exposure-response relationship at higher exposure levels and to provide a
15 better fit to the lower exposure data. Linear modeling of categorical (i.e., grouped)
16 epidemiologic data and elimination of the highest exposure group(s) under certain circumstances
17 to obtain a better fit of low-exposure data are both standard techniques used in EPA
18 dose-response assessments (U.S. EPA, 2012, 2005a). An established methodology was
19 employed for the weighted linear regression of the categorical epidemiologic data, as described
20 by Rothman (1986) and used by others (e.g., van Wijngaarden and Hertz-Picciotto, 2004).
21 However, the SAB panel that reviewed the draft assessment recommended that EPA employ
22 models using the individual exposure data as an alternative to modeling the published grouped
23 data. The SAB also recommended that both males and females be included in the modeling of
24 lymphohematopoietic cancer mortality (SAB, 2007).
25 In response to these recommendations and in consultation with Dr. Steenland, one of the
26 investigators from the NIOSH cohort studies, EPA determined that, using the full data set, an
27 alternative way to address the supralinearity of the data (while avoiding the extreme
28 low-exposure curvature obtained with the log cumulative exposure model) might be to use a
29 two-piece log-linear spline model. Spline models have been used previously for
30 exposure-response analyses of epidemiological data (Steenland and Deddens, 2004; Steenland et
31 al., 2001). These models are particularly useful for exposure-response data such as the EtO
32 lymphoid cancer data, for which RR initially increases with increasing exposure but then tends to
33 plateau, or level off, at higher exposures. Such plateauing exposure-response relationships have
34 been seen with other occupational carcinogens and may occur for various reasons, including the
35 depletion of susceptible subpopulations at high exposures, mismeasurement of high exposures,
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-9 DRAFT—DO NOT CITE OR QUOTE
-------
1 or a healthy worker survivor effect (Stayner et al., 2003). No other traditional exposure-response
2 models for continuous data which might suitably fit the observed exposure-response pattern were
3 apparent. Dr. Steenland was commissioned to do the spline analyses using the full data set with
4 cumulative exposure as a continuous variable, and his findings are included in Appendix D (see
5 Section D.3 for lymphoid cancer, Section D.4 for all lymphohematopoietic cancer). The results
6 of the spline analyses are presented below.
7 For the two-piece log-linear spline modeling approach, the Cox regression model (eq 4-1)
8 was the underlying basis for the splines which were fit to the lymphoid cancer exposure-response
9 data.u Taking the log of both sides of eq 4-1, log RR is a linear function of exposure
10 (cumulative exposure is used here), and, with the two-piece log-linear spline approach, log RR is
11 a function of two lines which join at a single point of inflection, called a "knot". The shape of
12 the two-piece log-linear spline model, in particular the slope in the low-exposure region, depends
13 on the location of the knot. For this assessment, the knot was generally selected by trying
14 different knots in increments of 1,000 ppm x days, starting at 1,000 ppm x days, and choosing
15 the one that resulted in the largest model likelihood. In some cases, increments of 100 ppm x
16 days were used between the increments of 1,000 ppm x days to fine-tune the knot selection. The
17 model likelihood did not change much across the different trial knots (see Figure D-3a of
18 Appendix D), but it did change slightly; therefore, the largest calculated likelihood was used as a
19 basis for knot selection. For more discussion of the two-piece spline approach, see Appendix D.
20 Using this approach, the largest likelihood was observed with the knot at
21 1,600 ppm x days. However, the graphical results for the two-piece log-linear spline model with
22 a knot at 1,600 ppm x days suggested that the model was underestimating RR in the region
23 where the data were plateauing (see Figure 4-1).12 Therefore, knots below 1,000 ppm x days
24 were also evaluated in increments of 100 ppm x days, and a likelihood was observed with the
25 knot at 100 ppm x days that exceeded the likelihood with the knot at 1,600 ppm x days,
26 although, again, the model likelihood did not actually change much across the different trial
27 knots. See Table 4-3 and Section D.3 of Appendix D for parameter estimates and fit statistics for
28 the two spline models. The graphical results for the two-piece spline model with a knot at
29 100 ppm x days suggested that this model provided a better fit to the region where the data were
30 plateauing (see Figure 4-1). Furthermore, the overall fit of this two-piece spline model was
31 statistically significant (p = 0.048), whereas the/>-value for the two-piece spline model with the
32 knot at 1,600 ppm x days exceeded 0.05, although minimally (p = 0.072). Thus, for the
11 As parameterized in Appendix D, for cumulative exposures less than the value of the knot, RR = e(|31 * exP°sure); for
cumulative exposures greater than the value of the knot, RR = e(pl *exposure + p2 x ^°™e-^\
12The loglinear spline segments appear fairly linear in the plotted range; however, they are not strictly linear.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-10 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
lymphoid cancer mortality data, the optimal two-piece log-linear spline model appeared to be the
one with the knot at 100 ppm x days. This model provided the largest calculated likelihood, was
statistically significant, and presented the best apparent graphical fit to the majority of the range
of the data. However, this model yielded a very steep slope in the exposure range below the knot
of 100 ppm x days (see Figure 4-1), and, as such, there was low confidence in the slope, given
that it is based primarily on a relatively small number of cases in the low-exposure region. Thus,
after examining the new modeling analyses, it was determined that the weighted linear regression
of the categorical data still provided the best available approach for risk estimates for
lymphohematopoietic cancer.
13
Table 4-3. Exposure-response modeling results for all lymphohematopoietic
cancer and lymphoid cancer mortality in both sexes in the NIOSH cohort for
models not presented by Steenland et al. (2004)
Model3
p valueb
Coefficient (SE)
All lymphohematopoietic cancer0
2-piece log-linear spline (knot at 500 ppm *
days)
Linear regression of categorical results,
excluding highest exposure group
0.02
0.08
low-exposure spline segment:
61=0.00201(0.0007731)
0.0003459 (0.0001944)
Lymphoid cancerd
Optimal 2-piece log-linear spline (knot at 100
ppm x days)
Alternate 2-piece log-linear spline (knot at 1,600
ppm x days)
Linear regression of categorical results,
excluding the highest exposure quartile
0.048
0.07
0.18
low-exposure spline segment:
61=0.01010(0.00493)
low-exposure spline segment:
Bl = 0.0004893 (0.0002554)
0.000247 (0.000185)
aAll with cumulative exposure in ppm * days as the exposure variable and with a 15-yr lag.
V-values from likelihood ratio test, except for linear regressions of categorical results, where Wald ^-values are
reported.
"9th revision ICD codes 200-208; results based on 74 cases.
dNHL, myeloma, and lymphocytic leukemia (9th revision ICD codes 200, 202, 203, 204); results based on 53 cases.
Source: Additional analyses performed by Dr. Steenland (see Sections D.3 and D.4 of Appendix D), except for the
linear regression of the categorical results, which was performed by EPA.
When this assessment was near completion, a two-piece linear spline model (with a linear model, i.e., RR = 1 + (3
x exposure, as the underlying basis for the spline pieces) was attempted, using the just-published approach of
Langholz and Richardson (2010) to model the individual data with cumulative exposure as a continuous variable;
however, this model did not alleviate the problem of the excessively steep low-exposure spline segment (see
Figure D-3c in Appendix D) and was not pursued further for the lymphoid cancer data. The Langholz and
Richardson (2010) approach was also employed to model the lymphoid cancer data using linear RR models with
cumulative exposure and log cumulative exposure as continuous variables; however, these linear models similarly
did not alleviate the problems of the corresponding log-linear RR models (see Figure D-3c in Appendix D).
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-11 DRAFT—DO NOT CITE OR QUOTE
-------
1 For the weighted linear regression, the Cox regression results from the model with
2 categorical cumulative exposure and a 15-year lag (see Table 4-2) was used, excluding the
3 highest exposure group, as discussed above.14 The weights used for the ORs were the inverses
4 of the variances, which were calculated from the confidence intervals.15 Mean and median
5 exposures for the cumulative exposure groups were provided by Dr. Steenland (see Table D-3a
6 of Appendix D).16 The mean values were used for the weighted regression analysis because the
7 cancer response is presumed to be a function of cumulative exposure, which is expected to be
8 best represented by mean exposures. If the median values had been used, a slightly larger
9 regression coefficient would have been obtained, resulting in slightly larger risk estimates. See
10 Table 4-3 for the results obtained from the weighted linear regression and Figure 4-1 for a
11 depiction of the resulting model.
12 As the lymphoid cancer data set is the primary data set used for the derivation of unit risk
13 estimates for lymphohematopoietic cancers, a summary of all the models considered for
14 modeling the lymphoid cancer exposure-response data and the judgments made about model
15 selection is provided in Table 4-4. See Figures 4-1 and D-3c in Appendix D for visual
16 representations of the models. See Tables 4-2 and 4-3 and Section D.3 of Appendix D for
17 /7-values and other fit statistics.
18 The linear regression of the categorical results for males and females combined and the
19 actuarial program (life-table analysis) were used to estimate the exposure level (ECX; "effective
20 concentration") and the associated 95% lower confidence limit (LECX) corresponding to an extra
21 risk of 1% (x = 0.01). A 1% extra risk level is commonly used for the determination of the point
22 of departure (POD) for low-exposure extrapolation from epidemiological data; higher extra risk
23 levels, such as 10%, would be an upward extrapolation for these data. Thus, 1% extra risk was
24 selected for determination of the POD, and, consistent with EPA's Guidelines for Carcinogen
25 Risk Assessment (U.S. EPA, 2005a), the LEG value corresponding to that risk level was used as
26 the POD to derive the cancer unit risk estimates.
27
28
"Concerns have been raised that this approach of dropping high-dose data appears arbitrary. It should be noted,
however, that only the highest exposure group was omitted from the linear regression, and the exposure groupings
were derived a priori by the NIOSH investigators and not by U. S. EPA in the course of its analyses.
1 Equations for this weighted linear regression approach are presented in Rothman (1986) and summarized in
Appendix F.
16Mean exposures for both sexes combined with a 15-year lag for the categorical exposure quartiles in Table 4-1
were 446; 2,143; 7,335; and 39,927 ppm * days. Median values were 374; 1,985; 6,755; and 26,373 ppm * days.
These values are for the full cohort, not just the risk sets.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-12 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
Table 4-4. Models considered for modeling the exposure-response data for
lymphoid cancer mortality in both sexes in the NIOSH cohort for the
derivation of unit risk estimates
Model3
Comments
Cox regression (log-linear) model
Inadequate overall statistical fit and poor visual fit in the
low-exposure region
Cox regression model with log cumulative exposure
Good overall statistical fit but too steep in the low-
exposure region
Optimal 2-piece log-linear spline (knot at 100 ppm
days)
Good overall statistical fit but too steep in the low-
exposure region
Alternate 2-piece log-linear spline (knot at 1,600
ppm x days)
Nonsignificant statistical fit and too steep in the low-
exposure region
linear model (RR = 1 + (3 x exposure)
Inadequate overall statistical fit (p = 0.13) and poor visual
fit in the low-exposure region
linear model with log cumulative exposure
Good overall statistical fit (p = 0.02) but too steep in the
low-exposure region
2-piece linear spline model
Good overall statistical fit (p = 0.04) but too steep in the
low-exposure region
5
6
7
8
9
10
11
12
13
14
Linear regression of categorical results,
excluding the highest exposure quartile
SELECTED. The continuous supralinear models (e.g.,
the log-cumulative-exposure models and the optimal 2-
piece log-linear spline model) are statistically significant
for lymphoid cancer mortality; however, they are too
steep in the low-exposure region for the derivation of
stable unit risk estimates. Thus, the linear regression
model of categorical results, excluding the highest
exposure quartile, was used for the derivation of unit risk
estimates, despite the lack of statistical significance, as it
was considered a better representation of the data in the
low-exposure region. Lack of statistical significance is
not critical given the low statistical power with
categorical data and the statistical significance of the
continuous supralinear models, which establishes the
significance of the exposure-response correlation for the
underlying data.
aAll with cumulative exposure as the exposure variable, except where noted, and with a 15-yr lag.
Because EtO is DNA-reactive and has direct mutagenic activity (see Section 3.3.3),
which is one of the cases cited by EPA's Guidelines for Carcinogen Risk Assessment (U.S. EPA,
2005a) for the use of linear low-dose extrapolation, a linear low-exposure extrapolation was
performed. The ECoi, LECoi, and inhalation unit risk estimate calculated for lymphoid cancer
mortality from the linear regression of the categorical results are presented in Table 4-5 (the
incidence results also presented in Table 4-5 are discussed in Section 4.1.1.3 below). The
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-13 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
Table 4-5. ECoi, LECoi, and unit risk estimates for lymphoid cancer"
Model0
Cox regression model, log
cumulative exposure,
15-yrlag
Optimal low-exposure log-linear
spline (knot at 100 ppm x days),6
cumulative exposure,
15-yrlag
Alternate low-exposure log-linear
spline (knot at 1,600 ppm x days)/
cumulative exposure,
15-yrlag
Linear regression of categorical
results, cumulative exposure,
15-yrlagg
Mortality
ECoi
(ppm)
0.00441
0.000982
0.0203
0.0564
LECoi
(ppm)
0.000428
0.000545
0.0109
0.0252
Unit risk
(per ppm)
a
a
f
0.397
Incidence1"
ECoi
(ppm)
0.000288
0.000525
0.0108
0.0254
LECoi
(ppm)
0.0000898
0.000291
0.00583
0.0114
Unit risk
(per ppm)
d
d
f
0.877
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
aFrom lifetime continuous exposure. Unit risk = 0.01/LEC0i.
blncidence estimate s presented here for comparison; they are derived in Section 4.1.1.3.
°From Dr. Steenland's analyses for males and females combined (see Section D.3 of Appendix D), Cox regression
models. Note that the ECOT and LEC0i results presented here will not exactly match those presented in Appendix D
because, although EPA used the regression coefficients reported by Dr. Steenland in Appendix D, the life-table
analyses using 2004 all-cause mortality rates were redone to be more up-to-date and consistent with the
cause-specific mortality rates; the results presented in Appendix D were based on life-table analyses using 2000
all-cause mortality rates.
dUnit risk estimates are not presented for these models because these models were deemed unsuitable for the
derivation of risks from (low) environmental exposure levels (see text).
eUsing regression coefficient from low-exposure segment of optimal two-piece log-linear spline model (largest
likelihood) with knot at 100 ppm x days; see text and Appendix D. Each of the ECOT values is below the value of
0.0013 ppm roughly corresponding to the knot of 100 ppm x days [(100 ppm x days) x (10 m3/20 m3) x
(240 d/365 d) x (365 d/yr)/70 yr = 0.0013 ppm] and, thus, appropriately in the range of the low-exposure segment.
fUsing regression coefficient from low-exposure segment of alternate two-piece log-linear spline model (local
largest likelihood) with a knot at 1,600 ppm x days. Each of the ECM values is below the value of 0.021 ppm
roughly corresponding to the knot of 1,600 ppm x days (see footnote d for calculation) and, thus, appropriately in
the range of the low-exposure segment. Unit risk estimates were not calculated from this model because the fit was
inferior to that of the optimal model (see text).
Degression coefficient derived from linear regression of categorical Cox regression results from Table 4-2, as
described in Section 4.1.1.2. Each of the ECM values is appropriately below the value of 0.090 ppm roughly
corresponding to the value of about 7,000 ppm x days (see footnote d for calculation) above which the linear
regression model of the categorical results does not apply (see Figure 4-1).
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
4-14 DRAFT—DO NOT CITE OR QUOTE
-------
1 resulting unit risk estimate for lymphoid cancer mortality based on the linear regression of the
2 categorical results for both sexes using cumulative exposure with a 15-year lag is 0.397 per ppm.
3 ECoi and LECoi estimates from the other models considered are presented for comparison only,
4 to illustrate the differences in model behavior at the low end of the exposure-response range.
5 Unit risk estimates are not presented for these other models because, as discussed above, these
6 models were deemed unsuitable for the derivation of risks from (low) environmental exposure
7 levels. The log cumulative exposure model, with its extreme supralinearity in the lower
8 exposure region, and the optimal two-piece log-linear spline model, with its very steep
9 low-exposure slope, yield substantially lower ECoi estimates (0.00441 ppm and 0.000982 ppm,
10 respectively). Converting the units, the resulting unit risk estimate of 0.397 per ppm from the
11 linear regression model of the categorical results corresponds to a unit risk estimate of
12 2.17 x 10 4 per ug/m3 for lymphoid cancer mortality.1?
13 As discussed above, risk estimates based on the all lymphohematopoietic cancer results
14 are also derived for comparison. The same methodology presented above for the lymphoid
15 cancer results was used for the all lymphohematopoietic cancer risk estimates. Age-specific
16 background mortality rates for all lymphohematopoietic cancers for the year 2004 were obtained
17 from the NCHS Data Warehouse website (http://www.cdc.gov/nchs/datawh/statab/unpubd/
18 mortabs.htm). The results of Dr. Steenland's reanalyses using the Cox regression models
19 presented in the Steenland et al. (2004) paper with data for males and females combined are
20 presented in Table 4-2. As for lymphoid cancer and for all hematopoietic cancer in males
21 presented in the Steenland et al. (2004) paper, the only statistically significant Cox regression
22 model was for log cumulative exposure with a 15-year lag (p = 0.01). The cumulative exposure
23 model did not provide an adequate fit to the data and is not considered further here (p = 0.35).
24 Because of the problems with the supralinear log cumulative exposure model which are
25 discussed for the lymphoid cancers above, EPA again investigated the use of a two-piece
26 log-linear spline model to attempt to address the supralinearity of the data while avoiding the
27 extreme low-exposure curvature obtained with the log cumulative exposure model. For the all
28 lymphohematopoietic cancer mortality data, the largest calculated likelihood was obtained with a
29 knot of 500 ppm x days (see Figure D-4a of Appendix D). See Table 4-3 and Section D.4 of
30 Appendix D for parameter estimates and fit statistics for the two-piece spline model. As with the
31 lymphoid cancer mortality results, however, this model resulted in an apparently excessively
"Conversion equation: 1 ppm= 1,830 ug/m3.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-15 DRAFT—DO NOT CITE OR QUOTE
-------
1 steep low-exposure spline (see Figure 4-2), so, again, the linear regression model of the
2 categorical results was used to derive the cancer unit risk estimate for this data set.18
3 For the weighted linear regression, the results from the Cox regression model with
4 categorical cumulative exposure and a 15-year lag (see Table 4-2) were used, excluding the
5 highest exposure group, and the approach discussed above for lymphoid cancer mortality. See
6 Table 4-3 for the results obtained from the weighted linear regression and Figure 4-2 for a
7 graphical presentation of the resulting linear regression model. As discussed above, this linear
8 regression model was used to derive the unit risk estimates for all lymphohematopoietic cancer.
9 The ECoi, LECoi, and inhalation unit risk estimate calculated for all
10 lymphohematopoietic cancer mortality from the linear regression model of the categorical results
11 are presented in Table 4-6 (the incidence results also presented in Table 4-6 are discussed in
12 Section 4.1.1.3 below). The resulting unit risk estimate for all lymphohematopoietic cancer
13 mortality based on the linear regression of the categorical results for both sexes using cumulative
14 exposure with a 15-year lag is 0.680 per ppm. ECoi and LECoi estimates from the other models
15 considered are presented for comparison only, to illustrate the differences in model behavior at
16 the low end of the exposure-response range. Unit risk estimates are not presented for these other
17 models because, as discussed above, these models were deemed unsuitable for the derivation of
18 risks from (low) environmental exposure levels. The resulting unit risk estimate for all
19 lymphohematopoietic cancer mortality from the linear regression model of the categorical results
20 is similar to that for lymphoid cancer mortality (70% higher; see Table 4-5). Converting the
21 units, the resulting unit risk estimate of 0.680 per ppm corresponds to a unit risk estimate of
22 3.72 x 10 4 per ug/m3 for all lymphohematopoietic cancer mortality.
23
24 4.1.1.3. Prediction of Lifetime Extra Risk of Lymphohematopoietic Cancer Incidence
25 EPA cancer risk estimates are typically derived to represent an upper bound on increased
26 risk of cancer incidence., as from experimental animal incidence data. Cancer data from
27 epidemiologic studies are more generally mortality data, as is the case in the Steenland et al.
28 (2004) study. For tumor sites with low survival rates, mortality-based estimates are reasonable
18When this assessment was near completion, a two-piece linear spline model (with a linear model, i.e., RR = 1 + (3
x exposure, as the underlying basis for the spline pieces) was attempted, using the just-published approach of
Langholz and Richardson (2010) to model the individual data with cumulative exposure as a continuous variable;
however, this model did not alleviate the problem of the excessively steep low-exposure spline segment (see
Figure D-4c in Appendix D) and was not pursued further for the all lymphohematopoietic cancer data. The
Langholz and Richardson (2010) approach was also employed to model the all lymphohematopoietic cancer data
using linear RR models with cumulative exposure and log cumulative exposure as continuous variables; however,
these linear models similarly did not alleviate the problems of the corresponding log-linear RR models (see
Figure D-4c in Appendix D).
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-16 DRAFT—DO NOT CITE OR QUOTE
-------
to
OJ
o
C5
S
I
o
I
o
a
i
O 5
O I
3
o
H
W
o
c
o
H
W
.
1
1.0
• e»(|3*exp)
• e"(|3*logexp)
categorical
2-piece spline
• linear
5000
10000
15000 20000 25000
cumulative exposure (ppm*days)
30000
35000
40000
Figure 4-2. RR estimate for all lymphohematopoietic cancer vs. mean exposure (with 15-year lag, unadjusted
for continuous exposure).
eA(p*exp): Cox regression results for RR = e(p x exposure); eA(p*logexp): Cox regression results for RR = e(p x ln(exP°sure»;
categorical: Cox regression results for RR = e*x exposure) with categorical exposures; linear: weighted linear regression
of categorical results, excluding highest exposure group (see text); 2-piece spline: 2-piece log-linear spline model with
knot at 500 ppm x days (see text)
Source: Steenland reanalyses for male and female combined; see Appendix D (except for linear regression of the
categorical results, which was done by EPA).
-------
1
2
3
Table 4-6. ECoi, LECoi, and unit risk estimates for all lymphohematopoietic
cancer
Model0
Log cumulative
exposure,
15-yrlag
Low-exposure
log-linear spline;6
cumulative
exposure,
15-yrlag
Linear regression
of categorical
results,
cumulative
exposure,
IS-yrlag1
Mortality
EC01
(ppm)
0.00140
0.00377
0.0283
LEC01
(ppm)
0.000245
0.00231
0.0147
Unit risk
(per ppm)
d
d
0.680
Incidence1"
EC01
(ppm)
0.000190
0.00216
0.0144
LEC01
(ppm)
0.0000753
0.00132
0.00746
Unit risk
(per ppm)
d
d
1.34s
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
"From lifetime continuous exposure. Unit risk = 0.01/LECM.
blncidence estimate s presented here for comparison; they are derived in Section 4.1.1.3.
°From Dr. Steenland's analyses for males and females combined (see Appendix D), Cox regression models. Note
that the ECM and LECOT results presented here will not exactly match those presented in Appendix D because,
although EPA used the regression coefficients reported by Dr. Steenland in Appendix D, the life-table analyses
using 2004 all-cause mortality rates were redone to be more up-to-date and consistent with the cause-specific
mortality rates; the results presented in Appendix D were based on life-table analyses using 2000 all-cause mortality
rates.
dUnit risk estimates are not presented for these models because these models were deemed unsuitable for the
derivation of risks from (low) environmental exposure levels (see text).
eUsing regression coefficient from low-exposure segment of two-piece log-linear spline model with knot at 500 ppm
x days; see text and Appendix D. Each of the EC0i values is below the value of 0.0064 ppm roughly corresponding
to the knot of 500 ppm x days [(500 ppm x days) x (10 m3/20 m3) x (240 d/365 d) x (365 d/yr)/70 yr = 0.0064 ppm]
and, thus, appropriately in the range of the low-exposure segment.
Degression coefficient derived from linear regression of categorical Cox regression results from Table 4-2, as
described in Section 4.1.1.2. Each of the ECM values is appropriately below the value of 0.064 ppm roughly
corresponding to the value of about 5,000 ppm x days (see footnote d for calculation) above which the linear
regression model of the categorical results does not apply (see Figure 4-2).
gFor unit risk estimates below 1, convert to risk per ppb (e.g., 1.34 per ppm = 1.34 x 10~3 per ppb).
approximations of cancer incidence risk; however, for many lymphohematopoietic cancers, the
survival rate is substantial, and incidence-based risks are preferred because EPA endeavors to
protect against cancer occurrence, not just mortality (U.S. EPA, 2005a).
Therefore, another calculation was done using the same regression coefficients presented
above (see Section 4.1.1.2), but with age-specific lymphoid cancer incidence rates for the
relevant subcategories of lymphohematopoietic cancer (NHL, myeloma, and lymphocytic
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-18 DRAFT—DO NOT CITE OR QUOTE
-------
1 leukemia) for 2000-2004 from SEER (Ries et al., 2007); Tables XIX, XVIII, XIII: both sexes,
2 all races) in place of the lymphoid cancer mortality rates in the actuarial program. SEER collects
3 good-quality cancer incidence data from a variety of geographical areas in the United States.
4 The incidence data used here are from "SEER 17," a registry of seventeen states, regions, and
5 cities covering about 26% of the U.S. population.
6 The incidence-based calculation assumes that lymphoid cancer incidence and mortality
7 have the same exposure-response relationship for the relative rate of effect from EtO exposure
8 and that the incidence data are for first occurrences of primary lymphoid cancer or that relapses
9 and secondary lymphoid cancers provide a negligible contribution. (The latter assumption is
10 probably sound; the former assumption is more potentially problematic. Because various
11 lymphoid subtypes with different survival rates are included in the categorization of lymphoid
12 cancers, if the EtO-associated relative rates of the subtypes differ and if the relative
13 rate-weighted survival rates for the lymphoid cancers are different from those for the combined
14 subtypes, a bias could occur, resulting in either an underestimation or overestimation of the extra
15 risk for lymphoid cancer incidence.)19 Potential concern that the incidence estimates might be
16 overestimated would come primarily from the inclusion of multiple myeloma, because that
17 subtype has the lowest incidence:mortality ratios (and, thus, if that subtype were driving the
18 increased mortality observed for the lymphoid cancer grouping, then including the incidence
19 estimates). Multiple myelomas, however, constitute only 25% of the lymphoid cancer cases in
20 the cohort, and there is no evidence that multiple myeloma is driving the EtO-induced rates for
21 the other subtypes, which have higher incidence:mortality ratios, might inflate the not expected
22 to result in an overestimation of the incidence risk estimates; if anything, the incidence risks
23 would likely be diluted with the inclusion of the multiple myeloma rates. The incidence-based
24 calculation also relies on the fact that the lymphoid cancer incidence rates are excess in lymphoid
90
25 cancer mortality. Thus, using the total lymphoid cancer incidence rates is small when
19Sielken and Valdez-Flores (2009) reject the assumption that lymphohematopoietic cancer incidence and mortality
have the same exposure-response relationship, reporting that, except at high exposure levels, the exposure-response
data in the male workers in the NIOSH cohort are consistent with a decreased survival time and suggesting that this
could explain the observed increases in mortality. However, they do not establish that this is what is occurring, and
the mechanistic data support an exposure-related increase in incident cancers. See Appendix A.2.20 for a more
detailed discussion of this issue.
20According to data from SEER (www.seer.cancer.gov), 25% is below the proportion of multiple myeloma deaths
one would expect based on age-adjusted U.S. background mortality rates of multiple myeloma, NHL, and chronic
lymphocytic leukemia, and these 3 subtypes have the same pattern for mortality rates increasing as a function of age
mostly above age 50, so the comparison with lifetime background rates is reasonable. In addition, the low
proportion of multiple myeloma deaths in the lymphoid cancer subgrouping cannot be attributed to an
underrepresentation of blacks, who have incidence rates of multiple myeloma over twice those of whites
(http://seer.cancer.gov/statfacts/html/mulmy.html), in the cohort, because blacks comprise 16% of the cohort versus
12.3% in the U.S. population.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-19 DRAFT—DO NOT CITE OR QUOTE
-------
91
1 compared with the all-cause mortality rates. The resulting ECoi and LECoi estimates for
2 lymphoid cancer incidence from the various models examined are presented in Table 4-5. The
3 unit risk estimate for lymphoid cancer incidence from the selected linear regression model of the
4 categorical results is 0.877 per ppm.
5 The ECoi estimates for cancer incidence range from about 6.5% (log cumulative exposure
6 Cox regression model) to 54% (cumulative exposure Cox regression model) of the corresponding
7 mortality-based estimates. The difference between incidence and mortality rates cannot explain
8 the large discrepancy in ECoi estimates for the log cumulative exposure model. Instead, the
9 discrepancy probably reflects the very different results that can occur from a small shift along the
10 dose-response curve for the log cumulative exposure model, illustrating the low-dose instability
11 of the results from this model. The incidence unit risk estimate from the linear regression model
12 of the categorical results is about 120% higher than (i.e., 2.2 times) the mortality-based estimate.
13 Overall, as discussed above, the preferred estimate for the unit risk for lymphoid cancer is
14 the estimate of 0.877 per ppm (4.79 x 10 4 per ug/m3) derived, using incidence rates for the
15 cause-specific background rates, from the weighted linear regression of the categorical results,
16 dropping the highest exposure group.
17 As discussed in Section 4.1.1.2, risk estimates based on the results of Dr. Steenland's
18 reanalyses of the all lymphohematopoietic cancer data (see Appendix D and Table 4-2) are also
19 derived for comparison. The same methodology presented above for the lymphoid cancer
20 incidence results was used for the all lymphohematopoietic cancer incidence risk estimates, and
21 the same assumptions apply. Age-specific SEER incidence rates for all lymphohematopoietic
22 cancer for the years 2000-2004 were used (Ries et al., 2007); Tables XIX, IX, XVIII, and XIII:
23 both sexes, all races). The ECoi and LECoi estimates for all lymphohematopoietic cancer
24 incidence from the different all lymphohematopoietic cancer mortality models examined are
25 presented in Table 4-6. The resulting unit risk estimate for all lymphohematopoietic cancer
26 incidence from the linear regression of the categorical results is about 2.0-times the
27 mortality-based estimate and about 1.5-times the lymphoid cancer incidence estimate (see
28 Table 4-5).
29
21Sielken and Valdez-Flores (2009) suggest that the methods used by EPA to calculate incidence risk estimates in
the life-table analysis are inappropriate; however, as explained in more detail in Appendix A.2.20, we disagree. For
the situation where the cause-specific incidence rates are small compared to the all-cause mortality rates, as with
lymphoid cancer, there is no problem, as Sielken and Valdez-Flores (2009) themselves demonstrate, and, for the
situation where the cause-specific incidence rates are not negligible compared to the all-cause mortality rates, as
with breast cancer, an adjustment was made in the analysis to remove those with incident cases from the population
at risk, i.e., those "surviving" each interval without a diagnosis of breast cancer (see Section 4.1.2.3). See
Appendix A.2.20 for a more detailed discussion of this issue.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-20 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
4.1.2. Risk Estimates for Breast Cancer
4.1.2.1. .Brazsf Cancer Results From the NIOSH Study
The Steenland et al. (2004) study discussed above in Section 4.1.1.1 also presents results
from exposure-response analyses for breast cancer mortality in female workers. Steenland et al.
(2003) present results of a breast cancer incidence study of a subcohort of the female workers
from the NIOSH cohort. In addition to the analyses presented in the Steenland et al. (2003) and
Steenland et al. (2004) papers, Dr. Steenland did subsequent analyses of the breast cancer
incidence and mortality data sets for EPA; these are discussed below and reported in
Sections D.I and D.2 of Appendix D, respectively.
4.1.2.2. Prediction of Lifetime Extra Risk of Breast Cancer Mortality
Results from the Cox regression models presented by Steenland et al. (2004), with some
reanalyses reported by Dr. Steenland in Appendix D (see Section D.2), are summarized in
Table 4-7. These models were considered for the derivation of unit risk estimates for breast
cancer mortality in females from continuous environmental exposure to EtO, applying the
methodologies described in Section 4.1.1.2.
Table 4-7. Cox regression results for breast cancer mortality in females in
the NIOSH cohort3, for models presented in Steenland et al. (2004)
Exposure variable1"
Cumulative exposure, 20-yr
lage
Log cumulative exposure, 20-
yr lagf
Categorical cumulative
exposure, 20-yr lagf
/7-valuec
0.06
0.01
0.07
Coefficient (SE)
0.0000122
(0.00000641)
0.084 (0.035)
ORs by category"1 (95% CI)
1.00, 1.76 (0.91-3.43), 1.77 (0.88-3.56),
1.97 (0.94-4.06), 3.13 (1.42-6.92)
22
23
24
25
26
27
28
29
30
31
32
33
"Based on 103 breast cancer (ICD-9 174,175) deaths.
bCumulative exposure is in ppm x days.
^-values reported by Steenland et al. (2004).
Exposure categories are 0; >0-646; 647-2,779; 2,780-12,321; >12,322 ppm x days.
eFrom reanalyses in Section D.2 of Appendix D; Steenland et al. (2004) reported the Cox regression results for
cumulative exposure with no lag.
fFrom Table 8 of Steenland et al. (2004).
United States age-specific all-cause mortality rates for 2000 for females of all race groups
combined (Minifio et al., 2002) were used to specify the all-cause background mortality rates in
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
4-21 DRAFT—DO NOT CITE OR QUOTE
-------
1 the actuarial program (life-table analysis). The National Center for Health Statistics 1997-2001
2 cause-specific background mortality rates for invasive breast cancers in females were obtained
3 from a SEER report (Ries et al., 2004). The risks were computed up to age 85 for continuous
4 exposures to EtO, conversions were made between occupational EtO exposures and continuous
5 environmental exposures, and 95% UCLs were calculated for the relative rates, as described
6 above.
7 The only statistically significant Cox regression model presented by Steenland et al.
8 (2004) for breast cancer mortality in females was for log cumulative exposure with a 20-year lag
9 (p = 0.01). However, as for the lymphohematopoietic cancers in Section 4.1.1, using the log
10 cumulative exposure model to estimate the risks from low environmental exposures is
11 problematic because this model is highly supralinear and results are unstable for low exposures
12 (see Figure 4-3). The cumulative exposure model, which is typically used and which is stable at
13 low exposures, was nearly statistically significant (p = 0.06 with a 20-year lag; see Section D.2
14 of Appendix D) in terms of the global fit to the data; however, at low exposures, the Cox
15 regression model with cumulative exposure is sublinear and does not reflect the apparent
16 supralinearity of the breast cancer mortality data (see Figure 4-3).
17 In a 2006 external review draft of this assessment (U.S. EPA, 2006a), which relied on the
18 original published results of Steenland et al. (2004), EPA proposed that the best way to reflect
19 the exposure-response relationship in the lower exposure region, which is the region of interest
20 for low-exposure extrapolation, was to do a weighted linear regression of the results from the
21 Cox regression model with categorical cumulative exposure and a 20-year lag. In addition, the
22 highest exposure group was not included in the regression to alleviate some of the "plateauing"
23 in the exposure-response relationship at higher exposure levels and to provide a better fit to the
24 lower exposure data. Linear modeling of categorical epidemiologic data and elimination of the
25 highest exposure group(s) in certain circumstances to obtain a better fit of low-exposure data are
26 both standard techniques used in EPA dose-response assessments (U.S. EPA, 2005a). However,
27 as discussed in Section 4.1.1.2 for the similarly supralinear lymphohematopoietic cancer data,
28 the SAB panel that reviewed the draft assessment recommended that EPA employ models using
29 the individual exposure data as an alternative to modeling the published grouped data (SAB,
30 2007). Consequently, it was determined that, using the full data set, an alternative way to
31 address the supralinearity of the data (while avoiding the extreme low-exposure curvature
32 obtained with the log cumulative exposure model) might be to use a two-piece spline model, and
33 Dr. Steenland was commissioned to do the spline analyses using the full data set with cumulative
34 exposure as a continuous variable. His findings are reported in Section D.2 of Appendix D, and
35 the results for the breast cancer mortality analyses are summarized below.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-22 DRAFT—DO NOT CITE OR QUOTE
-------
to
o
Th
ocu
en
draft for
ew
rposes on
-23
and does not constitute Age
DRAFT— DO NO
n
HH
H
W
O
V
C>
C
o
H
W
RR estimate
----- eA(B*logexp)
• categorical
— — eA(B*exp)
-- spline700
5000
10000 15000 20000
mean cumulative exposure(ppm*days)
25000
30000
35000
Figure 4-3. RR estimate for breast cancer mortality vs. mean exposure (with 20-year lag, unadjusted for
continuous exposure).
eA(B*exp): Cox regression results for RR = e(p x exP°sure); eA(B*logexp): Cox regression results for RR = e(p x ln(exP°sure»;
categorical: Cox regression results for RR = e(p x exposure:) with categorical exposures; linear: weighted linear regression
of categorical results, excluding highest exposure group (see text); spline700(13000): 2-piece log-linear spline model
with knot at 700 (13,000) ppm x days (see text).
Source: Steenland reanalyses with 20-year lag; see Section D.2 of Appendix D (except for linear regression of the
categorical results, which was done by EPA).
-------
1 For the two-piece log-linear spline modeling approach, as described in Section 4.1.1.2
2 and discussed more fully in Section D.2 of Appendix D, the Cox regression model was the
3 underlying basis for the splines which were fit to the breast cancer mortality exposure-response
4 data (cumulative exposure is used here, with a 20-year lag), and thus, log RR is a function of two
5 lines which join at a single point of inflection, called a "knot." The shape of the two-piece
6 log-linear spline model, in particular the slope in the low-exposure region, depends on the
7 location of the knot. For this assessment, knot selection was first attempted by trying different
8 knots in increments of 1,000 ppm x days, starting at 1,000 ppm x days, and choosing the one that
9 resulted in the largest model likelihood. The model likelihood did not actually change much
10 across the different trial knots (see Figure D-2a of Appendix D), but it did change slightly, and
11 this approach indicated that a knot of 13,000 ppm x days for the breast cancer mortality data
99
12 yielded the largest likelihood. However, a visual inspection of the model fit suggested that the
13 two-piece log-linear spline model with a knot at 13,000 ppm x days underestimates the
14 low-exposure results (see Figure 4-3). Thus, knots below 1,000 ppm x days in increments of
15 100 ppm x days were investigated, and it was revealed that a knot at 700 ppm x days yielded a
16 model with a likelihood that exceeded that for the model with the knot at 13,000 ppm x days (see
9^
17 Figures D-2a and D-2a' of Appendix D). The model with the knot at 700 ppm x days,
18 however, has a seemingly implausibly steep low-exposure slope, as was the case with the largest
19 likelihood models for the lymphohematopoietic cancers above. Moreover, neither the model
20 with the knot at 700 ppm x days nor the one with the knot at 13,000 ppm x days was statistically
21 significant overall, although both were nearly so (p = 0.067 and 0.074, respectively). See
22 Table 4-8 and Section D.2 of Appendix D for parameter estimates and fit statistics for the two
23 spline models. Because there was low confidence in the steep low-exposure slope from the
24 two-piece spline model with the largest likelihood, which is based on a relatively small number
25 of cases in that exposure range, and because the model with the knot at 13,000 ppm x days,
26 which had a local largest likelihood, appeared to have a poor fit to the low-exposure data, it was
27 determined that the weighted linear regression of the categorical results was more appropriate as
28 the basis for the unit risk estimates. For more discussion of the breast cancer mortality
29 exposure-response modeling using the continuous data, see Section D.2 of Appendix D.
30 For the weighted linear regression, the results from the Cox regression model with
31 categorical cumulative exposure (and a 20-year lag) presented in Table 4-7 were used, excluding
Using the log-linear spline model with the knot at 13,000 ppm x days, a regression coefficient of 0.0000607 per
ppm x day (SE = 0.0000309 per ppm x day) was obtained for the low-exposure spline segment (see Appendix D).
23Using the optimal two-piece log-linear spline model with the knot at 700 ppm x days, a regression coefficient of
0.0006877 per ppm x day (SE = 0.0004171 per ppm x day) was obtained for the low-exposure spline segment (see
Appendix D).
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-24 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
the highest exposure group, and the approach discussed above for the lymphoid cancers (see
Section 4.1.1.2).24 Mean and median exposures for the cumulative exposure groups were
9S
provided by Dr. Steenland (see Appendix D). See Table 4-8 for the results obtained from the
weighted linear regression of the categorical results and mean exposures and Figure 4-3 for a
depiction of the resulting linear regression model.
The linear regression of the categorical results and the actuarial program (life-table
analysis) were used to estimate the exposure level (ECX) and the associated 95% lower
confidence limit (LECX) corresponding to an extra risk of 1% (x = 0.01). As discussed in
Section 4.1.1.2, a 1% extra risk level is a more reasonable response level for defining the POD
for these epidemiologic data than 10%.
Table 4-8. Exposure-response modeling results for breast cancer mortality
in females in the NIOSH cohort for models not presented by Steenland et al.
(2004)
Model3
Optimal 2-piece log-linear spline (knot at 700
ppm x days)
Alternate 2-piece log-linear spline (knot at
13,000 ppm x days)
Linear regression of categorical results,
excluding the highest exposure quartile
p value
0.067
0.074
0.09
Coefficient (SE)
low-exposure spline segment:
61=0.000688(0.000417)
low-exposure spline segment:
Bl = 0.0000607 (0.0000309)
0.000201 (0.000120)
17
18
19
20
21
22
23
24
25
26
27
28
29
aAll with cumulative exposure in ppm x days as the exposure variable and with a 20-yr lag; based on 103 breast
cancer deaths.
V-values from likelihood ratio test, except for linear regression of categorical results, where Wald ^-values are
reported.
Source: Additional analyses performed by Dr. Steenland (see Section D.2 of Appendix D), except for the linear
regression of the categorical results, which was performed by EPA.
Because EtO is DNA-reactive and has direct mutagenic activity (see Section 3.3.3),
which is one of the cases cited by EPA's Guidelines for Carcinogen Risk Assessment (U.S. EPA,
2005a) for the use of linear low-dose extrapolation, a linear low-exposure extrapolation was
Equations for this weighted linear regression approach are presented in Rothman (1986) and summarized in
Appendix F.
25Mean exposures for females with a 20-year lag for the categorical exposure quartiles in Table 8 of Steenland et al.
(2004) were 276; 1,453; 5,869; and 26,391 ppm x days. Median values were 250; 1,340; 5,300; and
26,676 ppm x days. These values are for the risk sets but should provide a good approximation to the full cohort
values.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-25 DRAFT—DO NOT CITE OR QUOTE
-------
1 performed. The ECoi, LECoi, and inhalation unit risk estimate calculated for breast cancer
2 mortality from the linear regression model of the categorical results are presented in Table 4-9.
3 The resulting unit risk estimate for breast cancer mortality based on the linear regression of the
4 categorical results using cumulative exposure with a 20-year lag is 0.513 per ppm. ECoi and
5 LECoi estimates from the other models considered are presented for comparison only, to
6 illustrate the differences in model behavior at the low end of the exposure-response range. Unit
7 risk estimates are not presented for these other models because, as discussed above, these models
8 were deemed unsuitable for the derivation of risks from (low) environmental exposure levels.
9 As one can see, the standard Cox regression cumulative exposure model, with its extreme
10 sublinearity in the lower exposure region, yields a substantially higher ECoi estimate
11 (0.530 ppm) than the ECoi estimate of 0.0387 ppm from the linear regression of the categorical
12 results, while the log cumulative exposure Cox regression model, with its extreme supralinearity
13 in the lower exposure region, yields a substantially lower ECoi estimates (0.00112 ppm). The
14 estimates from the two-piece log-linear spline models flank the result from the linear regression
15 of the categorical results more closely. The steep low-exposure segment of the two-piece
16 log-linear spline model with the optimal knot at 700 ppm x days yields an ECoi estimate of
17 0.00941 ppm, whereas the shallower low-exposure slope from the two-piece log-linear spline
18 model with the local maximum likelihood suggesting a knot at 13,000 ppm x days yields an ECoi
19 estimate of 0.107 ppm. Converting the units, the unit risk estimate of 0.513 per ppm for breast
20 cancer mortality from the linear regression model of the categorical results corresponds to a unit
21 risk estimate of 2.80 x 10~4 per ug/m3.
22
23 4.1.2.3. Prediction of Lifetime Extra Risk of Breast Cancer Incidence
24 As discussed in Section 4.1.1.3, risk estimates for cancer incidence are preferred to
25 estimates for cancer mortality, especially for cancer types with good survival rates, such as breast
26 cancer. In the case of female breast cancer in the NIOSH cohort, there is a corresponding
27 incidence study (Steenland et al., 2003) with exposure-response results for breast cancer
28 incidence, so one can estimate cancer incidence risks directly rather than estimate them from
29 mortality data. The incidence study used a (sub)cohort of 7,576 (76%) of the female workers
30 from the original cohort. Cohort eligibility for the incidence study was restricted to the female
31 workers who had been employed at 1 of the 14 plants for at least 1 year, owing to cost
32 considerations and the greater difficulties in locating workers with short-term employment.
33 Interviews were sought from all the women in the incidence study cohort or their next-of-kin
34 (18% of the cohort had died). Completed interviews were obtained for 5,139 (68%) of the
35
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-26 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
Table 4-9. ECoi, LECoi, and unit risk estimates for breast cancer mortality
in females"
Model
Log cumulative exposure, 20-
yrlagb
Cumulative exposure, 20-yr
lagd
Low-exposure log-linear
spline, cumulative exposure
with knot at 700 ppm x days,
20-yr lage
Low-exposure log-linear
spline, cumulative exposure
with knot at 13,000 ppm x
days, 20-yr lagf
Categorical; cumulative
exposure, 20-yr lag8
EC01
(ppm)
0.00112
0.530
0.00941
0.107
0.0387
LECoi
(ppm)
0.000219
0.285
0.00471
0.0580
0.0195
Unit risk
(per ppm)
c
c
c
c
0.513
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
aFrom lifetime continuous exposure. Unit risk = 0.01/LECoi.
bFrom Table 8 of Steenland et al. (2004), Cox regression model.
°Unit risk estimates are not presented for these models because these models were deemed unsuitable for the
derivation of risks from (low) environmental exposure levels (see text).
dFrom Dr. Steenland's reanalyses (see Table D-2d of Appendix D), Cox regression model.
eFrom low-exposure segment of two-piece log-linear spline model with largest model likelihood and a knot at
700 ppm x days; see text and Table D-2c of Appendix D. The ECM value is below the value of 0.009 ppm roughly
corresponding to the knot of 700 ppm x days [(700 ppm x days) x (10 m3/20 m3) x (240 d/365 d) x (365 d/yr)/70 yr
= 0.0013 ppm] and, thus, appropriately in the range of the low-exposure segment.
fFrom low-exposure segment of two-piece log-linear spline model with a local largest likelihood for knot at
13,000 ppm x days; see text and Table D-2f of Appendix D. The ECm value is below the value of 0.17 ppm roughly
corresponding to the knot of 13,000 ppm x days (see calculation in footnote e) and, thus, appropriately in the range
of the low-exposure segment.
Degression coefficient derived from linear regression of categorical Cox regression results from Table 8 of
Steenland et al. (2004), as described in Section 4.1.2.2. The ECM value is appropriately below the value of 0.064
ppm roughly corresponding to the value of about 5,000 ppm x days (see footnote e for calculation) above which the
linear regression model of the categorical results does not apply (see Figure 4-3).
7,576 women in the cohort. The investigators also attempted to acquire breast cancer incidence
data for the cohort from cancer registries (available for 9 of the 11 states in which the plants were
located) and death certificates; thus, results are presented for both the full cohort (n = 7,576) and
the subcohort of women with interviews (n = 5,139). For additional details and discussion of the
Steenland et al. (2003) study, see Section A.2.16 of Appendix A.
Steenland et al. (2003) identified 319 incident cases of breast cancer in the cohort through
1998. Interview (questionnaire) data were available for 73% (233 cases). Six percent were
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
4-27 DRAFT—DO NOT CITE OR QUOTE
-------
1 carcinoma in situ (20 cases). Steenland et al. (2003) performed internal exposure-response
2 analyses similar to those described in their 2004 paper and in Section 4.1.1.1 above. Controls for
3 each case were selected from the cohort members without breast cancer at the age of diagnosis of
4 the case. Cases and controls were matched on race. Of the potential confounders evaluated for
5 those with interviews, only parity and breast cancer in a first-degree relative were important
6 predictors of breast cancer, and only these variables were included in the final models for the
7 subcohort analyses. In situ cases were included with invasive breast cancer cases in the analyses;
8 however, the in situ cases represent just 6% of the total, and excluding them reportedly did not
9 greatly affect the results.
10 From the Steenland et al. (2003) internal analyses (Cox regression) using the full cohort,
11 the best-fitting model with exposure as a continuous variable was for (natural) log cumulative
12 exposure, lagged 15 years (p = 0.05). Duration of exposure, lagged 15 years, provided a slightly
13 better fitting model. Models using maximum or average exposure did not fit as well. In
14 addition, use of a threshold model did not provide a statistically significant improvement in fit.
15 For internal analyses using the subcohort with interviews, the cumulative exposure and log
16 cumulative exposure models, both lagged 15 years, and the log cumulative exposure model with
17 no lag all fit almost equally well, and the duration of exposure (also lagged 15 years) model fit
18 slightly better. Results of the Cox regression analyses for the cumulative and log cumulative
19 exposure models, with 15-year lags, are shown in Table 4-10, and these are the results
20 considered for the unit risk calculations. The models using duration of exposure are less useful
21 for estimating exposure-related risks, duration of exposure and cumulative exposure are
22 correlated, and the fits for these models are only marginally better than those with cumulative
23 exposure. The log cumulative exposure model with no lag was considered less biologically
24 realistic than the corresponding model with a 15-year lag because some lag period would be
25 expected for the development of breast cancer. Furthermore, although initial risk estimates
26 based on the full cohort results are calculated for comparison, the preferred estimates are those
27 based on the subcohort with interviews because the subcohort should have more complete case
28 ascertainment and has additional information available on potential breast cancer confounders.
29 For the actuarial program (life-table analysis), U.S. age-specific all-cause mortality rates
30 for 2004 for females of all race groups combined (Arias, 2007) were used to specify the all-cause
31 background mortality rates. Because breast cancer incidence rates are not negligible compared
32 to all-cause mortality rates, the all-cause mortality rates in the life-table analysis were adjusted to
33 reflect women dying or being diagnosed with breast cancer in a given age interval. All-cause
34 mortality rates and breast cancer incidence rates were summed, and breast cancer mortality rates
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-28 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
Table 4-10. Cox regression results for breast cancer incidence in females
from the NIOSH cohort, for the models presented by Steenland et al.
(2003)a'b
Cohort
Full incidence
study cohort
n = 7,576
319 cases
Subcohort with
interviews
n = 5,139
233 cases
Exposure variable0
Cumulative exposure, 15-yr
lag
Log cumulative exposure, 15-
yr lag
Categorical cumulative
exposure, 15-yr lag
Cumulative exposure, 15-yr
lag
Log cumulative exposure, 15-
yr lag
Categorical cumulative
exposure, 15-yr lag
Coefficient (SE),
^-valued
0.0000054
(0.0000035),
p = O.U
0.037 (0.019),
^ = 0.05
0.0000095
(0.0000041),
p = 0.02
0.050 (0.023),
^ = 0.03
f
ORs by category6 (95% CI)
1.00, 1.07 (0.72-1.59), 1.00
(0.67-1.50), 1.24 (0.85-1.90), 1.17
(0.78-1.78), 1.74(1.16-2.65)
1.00, 1.06(0.66-1.71), 0.99
(0.61-1.60), 1.24 (0.76-2.00), 1.42
(0.88-2.29), 1.87(1.12-3.10)
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
"Invasive breast cancer (ICD-9 174) and carcinoma in situ (ICD-9 233.0).
bCases and controls matched on age and race (white/nonwhite). Full cohort models include cumulative exposure
and categorical variable for yr of birth (quartiles). Subcohort models include cumulative exposure, categorical
variables for yr of birth (quartiles), breast cancer in first-degree relative, and parity.
'Cumulative exposure is in ppm x days.
d/>-values for exposure variable from Wald test, as reported by Steenland et al. (2003).
"Exposure categories are 0, >0-647, 647-2,026, 2,026-4,919, 4,919-14,620, >14,620 ppm x days.
f/>-value for the addition of the categorical exposure variables = 0.11 (email dated 5 March 2010 from Kyle
Steenland, Emory University, to Jennifer Jinot, EPA).
Source: Tables 4 and 5 of Steenland et al. (2003).
were subtracted so that those dying of breast cancer were not counted twice (i.e., as deaths and as
incident cases of breast cancer). The National Center for Health Statistics 2002-2006 mortality
rates for invasive breast cancer in females were obtained from a SEER report (Horner et al.,
2009). The SEER report also provided SEER-17 incidence rates for invasive and in situ breast
cancer. The Cox regression results reported by Steenland et al. (2003) are for invasive and in
situ breast cancers combined. It is consistent with EPA's Guidelines for Carcinogen Risk
Assessment (U.S. EPA, 2005a) to combine these two tumor types because the in situ tumors can
progress to invasive tumors. Thus, the primary risk calculations in this assessment use the sum
of invasive and in situ breast cancer incidence rates for the cause-specific background rates.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-29 DRAFT—DO NOT CITE OR QUOTE
-------
1 Comparison calculations were performed using just the invasive breast cancer incidence rates for
2 the cause-specific rates; this issue is further discussed in Section 4.1.3 on sources of uncertainty.
3 The risks were computed up to age 85 for continuous exposures to EtO, conversions were made
4 between occupational EtO exposures and continuous environmental exposures, and 95% UCLs
5 were calculated for the relative rates, as described in Section 4.1.1.2 above.
6 For breast cancer incidence in both the full cohort (see Figure 4-4) and the subcohort with
7 interviews (see Figure 4-5), the low-exposure categorical results suggest a more linear
8 low-exposure exposure-response relationship than that obtained with either the continuous
9 variable log cumulative exposure (supralinear) or cumulative exposure (sublinear) Cox
10 regression models, the two of which lie on opposite sides of the low-exposure categorical results.
11 Thus, as with the lymphohematopoietic cancer and the breast cancer mortality results above,
12 EPA proposed in the 2006 Draft Assessment (U.S. EPA, 2006a), which relied on the original
13 published results of Steenland et al. (2003), that the best way to reflect the data in the lower
14 exposure region, which is the region of interest for low-exposure extrapolation, was to do a
15 weighted linear regression of the results from the model with categorical cumulative exposure
16 (with a 15-year lag). In addition, the highest exposure group was not included in the regression
17 to provide a better fit to the lower-exposure data (The RR estimates for the highest exposure
18 quintiles suggest somewhat supralinear exposure-response relationships for both the full cohort
19 and the subcohort with interviews and supralinearity is evidenced in the subcohort with
20 interviews by the strong influence of the top 5% of cumulative exposures on dampening the
21 slope of the [cumulative exposure] Cox regression model [see Section D.I and Figure D-ld of
22 Appendix D]. Moreover, there is more uncertainty in using the mean cumulative exposure to
23 represent the range of exposures in a highest exposure categorical group because such groups
24 contain a wider range of exposures; for example, for the subcohort with interviews, the highest
25 exposure quintile contains exposures ranging from about 14,500 ppm x days to over
26 250,000 ppm x days). Linear modeling of categorical (i.e., grouped) epidemiologic data and
27 elimination of the highest exposure group(s) under certain circumstances to obtain a better fit of
28 low-exposure data are both standard techniques used in EPA dose-response assessments (U.S.
29 EPA, 2012, 2005a). However, as discussed in Section 4.1.1.2 for the lymphohematopoietic
30 cancer data, the SAB panel that reviewed the draft assessment recommended that EPA not rely
31 on the published grouped data but, rather, do additional analyses using the individual data (SAB,
32 2007).
33 Consequently, it was determined that using the individual data, a better way to address
34 the apparent supralinearity of the data (while avoiding the extreme low-exposure curvature
35 obtained with the log cumulative exposure Cox regression model) might be to use a two-piece
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-30 DRAFT—DO NOT CITE OR QUOTE
-------
to
o
§•
I
<>>
o
i
§
!
O
H
H
W
O
/O
O
• eA(p*exp)
• eA(p*logexp)
• linear
categorical
5000
10000
15000 20000 25000 30000
mean cumulative exposu re (ppm* days)
35000
40000
45000
Figure 4-4. RR estimate for breast cancer incidence in full cohort vs. mean exposure (with 15-year lag,
unadjusted for continuous exposure).
A*
exp): Cox regression results for RR =
exposure)A*
logexp): Cox regression results for RR = e
= (p x ln(exposure))
categorical: Cox regression results for RR = e^x exposure) with categorical exposures; linear: weighted linear regression
of categorical results, excluding highest exposure group (see text).
Source: Steenland et al. (2003) (except for linear regression of the categorical results, which was done by EPA).
-------
to
o
§•
K)
§
i
a,
rs
I
§
o
H
H
W
o
&
o
c
o
— eA([3*exp)
eA((3*logexp)
— linear regression
categorical
— loglinearspline
- linearspline
eA(B*sqrtexp)
- lincar(l+(3*exp)
1+ (3*logexp
5QQQ 10000 15000 20000 25000 30000 35000 40000
mean cumulative exposure (ppm * days)
Figure 4-5. RR estimate for breast cancer incidence in subcohort with interviews vs. mean exposure (with
15-year lag, unadjusted for continuous exposure).
eA(p*exp): Cox regression results for RR = e(p x exP°sure); eA(p*logexp): Cox regression results for RR = e(p x ln(exP°sure»;
categorical: Cox regression results for RR = e*x exposure) with categorical exposures; eA(P*sqrtexp): Cox regression
results for RR = e(px sirt(exPosure)); linear regression: weighted linear regression of categorical results, excluding highest
exposure group (see text); log-linear and linear spline: 2-piece spline models, both with knots at 5,800 ppm x days (see
text); linear: RR = 1 + p x exposure, with exposure as a continuous variable; 1 + P*logexp: RR = 1 + p x
In(exposure), exposure continuous.
Sources: Steenland et al. (2003) except for Steenland 2-piece spline models (see Appendix D) and linear regression of
the categorical results, which was done by EPA.
-------
1 spline model, and Dr. Steenland was commissioned to do the spline analyses. His findings are
2 reported in Appendix D (see Section D. 1), and the results for the breast cancer incidence
3 analyses are summarized below. Note that, for the two-piece spline analyses, only the data from
4 the subcohort with interviews and for the invasive and in situ breast cancers combined were
5 analyzed, because this was the preferred data set, as discussed above. (Dr. Steenland also
6 employed a cubic spline model as a semiparametric approach to visualize the underlying
7 exposure-response relationship; however, this approach produces an overly complicated function
8 for an empirical model, as opposed to a biologically based model, and was not used for risk
9 assessment purposes. In addition, Dr. Steenland investigated the use of a Cox regression model
10 with a square-root transformation of cumulative exposure; however, this approach, though less
11 extreme than using the log transformation of cumulative exposure, also yields a notably
12 supralinear model [see Figure 4-5], which can result in unstable low-exposure risk estimates.
13 The model results for both the cubic spline and square-root transformation models are included
14 in Appendix D, Section D.I, but are not considered further here. EPA chose to pursue the
15 development of two-piece spline models, which avoid the problem of unstable risk estimates
16 from supralinear curvature in the low-exposure region and provide a more general approach to
17 modeling supralinear exposure-response data, as opposed to using random, arbitrary
18 power-transformations of the exposure variable.)
19 For the two-piece log-linear spline modeling approach, as described in Section 4.1.1.2
20 and discussed more fully in Appendix D, the Cox regression model was the underlying basis for
21 the splines which were fit to the breast cancer incidence exposure-response data (cumulative
22 exposure is used here, with a 15-year lag), and, thus, log RR is a function of two lines which join
23 at a single point of inflection, called a "knot." The shape of the two-piece spline model, in
24 particular the slope in the low-exposure region, depends on the location of the knot. For this
25 assessment, the knot was generally selected by trying different knots in increments of 1,000 ppm
26 x days, starting at 1,000 ppm x days, and choosing the one that resulted in the largest model
27 likelihood. In some cases, increments of 100 ppm x days were used between the increments of
28 1,000 ppm x days to fine-tune the knot selection. The model likelihood did not actually change
29 much across the different trial knots (see Figure D-la of Appendix D), but it did change slightly,
30 and a knot of 5,800 ppm x days for the breast cancer incidence data based on the largest
31 likelihood was chosen. The two-piece log-linear spline model with this knot provided a
32 statistically significant fit to the data (p = 0.01 for the addition of the exposure terms), as well as
33 a good visual fit (see Figure 4-5).
34 A two-piece linear spline model was also fitted, using the just-published approach of
35 Langholz and Richardson (2010). This model is similar to the log-linear spline model discussed
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-33 DRAFT—DO NOT CITE OR QUOTE
-------
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
above; however, for the linear spline model, the underlying basis for the splines is a linear model
(i.e., RR = 1 + P x z, where z represents the covariate data, including exposure, and P are the
parameters being estimated). The knot was selected as for the log-linear spline model, and the
same knot of 5,800 ppm x days yielded the largest likelihood (see Figure D-lh of Appendix D)
and was also chosen for the two-piece linear spline model. The two-piece linear spline model
with this knot provided a statistically significant fit to the data (p = 0.002 for the addition of the
exposure terms), as well as a good visual fit (see Figure 4-5). Because this model provided a
better fit than the log-linear spline model, i.e., it had a lower AIC, the two-piece linear spline
model was selected as the preferred model for the unit risk estimates for breast cancer incidence.
See Table 4-11 and Section D.I of Appendix D for parameter estimates and fit statistics for the
two spline models.
Table 4-11. Exposure-response modeling results for breast cancer incidence
in females from the NIOSH cohort for models not presented by Steenland et
al. (2003)
Model3
p valueb
Coefficient (SE)
Full incidence study cohort0
Linear regression of categorical results,
excluding the highest exposure quintile
0.33
0.0000264 (0.0000269)
Subcohort with interviews'1
2-piece log-linear spline (knot at 5,800 ppm *
days)
2-piece linear spline (knot at 5,800 ppm * days)
linear
linear with log cumulative exposure
Linear regression of categorical results,
excluding the highest exposure quintile
0.01
0.002
0.003
0.01
0.16
low-exposure spline segment:
61=0.0000770(0.0000317)
low-exposure spline segment:
61=0.000119(0.0000677)
0.0000304 (0.0000175)
0.0713 (0.0392)
0.0000517(0.0000369)
aAll with cumulative exposure in ppm * days as the exposure variable and with a 15-yr lag.
V-value for addition of exposure variables from likelihood ratio test, except for the linear regressions of categorical
results, where Wald ^-values are reported.
°319 breast cancer cases.
d233 breast cancer cases.
Source: Additional analyses performed by Dr. Steenland (see Section D.2 of Appendix D), except for the linear
regressions of categorical results, which were performed by EPA using the equations of Rothman (1986) presented
in Appendix F.
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
4-34 DRAFT—DO NOT CITE OR QUOTE
-------
1 Linear RR models with cumulative exposure and log cumulative exposure as continuous
2 variables were also investigated using the approach of Langholz and Richardson (2010), and
3 these models fit better than the corresponding log RR models (see Table 4-11 and Section D. 1 of
4 Appendix D) although not as well as the two-piece linear spline model, which had the lowest
5 AIC. Risk estimates based on the linear model with cumulative exposure are developed for
6 comparison, but the linear model with log cumulative exposure is too steep in the low-exposure
7 region (see Figure 4-5) and is not considered further. For more details of the breast cancer
8 incidence exposure-response modeling, see Section D.I of Appendix D.
9 Risk estimates based on the original linear regression analyses of the categorical results
10 are also presented for comparison. For the approach of using a weighted linear regression of the
11 results from the Cox regression model with categorical cumulative exposure (and a 15-year lag),
12 excluding the highest exposure group, the weights used for the ORs were the inverses of the
13 variances, which were calculated from the confidence intervals.26 Mean and median exposures
14 for the cumulative exposure groups for the full cohort were kindly provided by Dr. Steenland
15 (email dated April 21, 2004, from Kyle Steenland, Emory University, to Jennifer Jinot, EPA).27
16 The mean values were used for the weighted regression analysis because the (arithmetic) mean
17 exposures best represent the model's linear relationship between exposure and cancer response.
18 Differences between means and medians were not large for the females, especially for the lower
19 four quintiles. If the median values had been used, a slightly larger regression coefficient would
20 have been obtained, resulting in slightly larger risk estimates. Although the exposure values are
21 for risk sets from the full cohort, they should be reasonably close to the values for the subcohort
22 with interviews. See Table 4-11 for the results from the weighted linear regressions of the
23 categorical results and Figures 4-4 and 4-5 for a depiction of the resulting linear regression
24 models.
25 As the subcohort with interviews from the NIOSH incidence study cohort provides the
26 preferred data set for the derivation of unit risk estimates for breast cancer, a summary of all the
27 models considered for modeling the breast cancer exposure-response data from the subcohort
28 and the judgments made about model selection is provided in Table 4-12. See Figure 4-5 for
29 visual representations of the models. See Tables 4-10 and 4-11 and Section D.I of Appendix D
30 for parameter estimates, ^-values, and other fit statistics. Three of the models presented in
31
26Equations for this weighted linear regression approach are presented in Rothman (1986) and summarized in
Appendix F.
27Mean exposures for females with a 15-year lag for the exposure categories in Table 3 of Steenland et al. (2003)
were 280; 1,241; 3,304; 8,423; and 36,022 ppm x days. Median values were 253; 1,193; 3,241; 7,741; and 26,597
ppm x days. These values are for the risk sets but should provide a good approximation to the full cohort values.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-35 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
Table 4-12. Models considered for modeling the exposure-response data for
breast cancer incidence in females in the subcohort with interviews from the
NIOSH incidence study cohort for the derivation of unit risk estimates
Model3
Cox regression (log-linear)
model
Cox regression model with log
cumulative exposure
Cox regression model with
square-root transformation of
exposure
Linear regression of categorical
results, excluding the highest
exposure quintile
2-piece log-linear spline model
(knot at 5,800 ppm x days)
linear model (RR = 1 + (3 x
exposure)
linear model with log
cumulative exposure
2-piece linear spline model
(knot at 5,800 ppm x days)
AICb
1956.675
1956.176
1953.028
C
1954.485
1952.260
1954.267
1950.935
Comments
Good overall statistical fit but poor visual fit (too
shallow) in the low-exposure region.
Good overall statistical fit but too steep in the low-
exposure region.
Good overall statistical fit but still notably supralinear
(steep) in the low-exposure region, though less so than
with the log transformation; also preference was given to
the two-piece spline models as providing a more general
approach to modeling supralinear data.
Not statistically significant, though that is unsurprising
since the approach, which is based on categorical data,
has low statistical power; preference given to models that
treated exposure as a continuous variable, as
recommended by the SAB, and that also provided
reasonable representations of the low-exposure region.
Good overall statistical fit and good visual fit; preference
given to the 2-piece linear spline model because it had a
better statistical fit (lower AIC) and better apparent fit to
the lower-exposure data.
Good overall statistical fit and good visual fit; preference
given to the 2-piece linear spline model because it had a
better statistical fit (lower AIC) and better apparent fit to
the lower-exposure data.
Good overall statistical fit but too steep in the low-
exposure region.
SELECTED. Good overall statistical fit and good visual
fit; lower AIC than 2-piece log-linear spline and linear
model and better apparent fit to the lower-exposure data.
5
6
7
8
9
10
11
aAll with cumulative exposure as the exposure variable, except where noted, and with a 15-yr lag.
bAIC = 2p-2LL, where/) = # of parameters and LL = In(likelihood), assuming two exposure parameters for the
two-piece spline models.
°Not calculated.
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
4-36 DRAFT—DO NOT CITE OR QUOTE
-------
1 Table 4-12 had a good overall statistical fit, a good visual fit, and a credible low-exposure slope
2 (the linear and log-linear two-piece spline models and the [continuous] linear RR model). To
3 better compare these models, they are plotted again in Figure 4-6 this time against the categorical
4 data in deciles. Earlier categorical results in this assessment were based on the (log-linear) Cox
5 regression model; however, the deciles in Figure 4-6 are based on a linear RR categorical
6 model—this model had a lower AIC than the log-linear decile model (1963.94 vs. 1966.91), and
7 it provides a statistically significant fit to the data (p = 0.004), so the deciles should provide a
8 good representation of the data for the purposes of comparing the models (the decile results from
9 the log-linear and linear RR categorical models and the mean cumulative exposure estimates for
10 the deciles are presented in Section D.I of Appendix D). As can be seen in Figure 4-6, the
11 two-piece linear spline model, in addition to having the lowest AIC (see Table 4-12), appears to
12 have a better fit to the lower-exposure data, which are of the greatest interest in estimating
13 low-exposure risk. It also appears from Figure 4-6 that the linear model has a poorer fit to the
14 lower-exposure data than either of the two-piece spline models. This is consistent with the
15 analysis presented in Section D. 1 of Appendix D showing the strong influence of the upper tail
16 of cumulative exposures on the results of the cumulative exposure Cox regression model. The
17 responses in the upper tail of exposures are relatively dampened, such that when the highest 5%
18 of exposures are excluded, the slope of the Cox regression model is substantially increased (e.g.,
19 at 10,000 ppm x days, the RR estimate increases from about 1.1 to almost 1.5; see Figure D-ld
20 in Appendix D). This strong influence of the upper tail of exposures would similarly attenuate
21 the slope of the (continuous) linear model. The two-piece spline models, on the other hand, are
22 more flexible, and the influence of the upper tail of exposures would be primarily on the upper
23 spline segment; thus, the two-piece models are able to provide a better fit to the lower-exposure
24 data.
25 The exposure level (ECX) and the associated 95% lower confidence limit (LECX)
26 corresponding to an extra risk of 1% (x = 0.01) for breast cancer incidence in females (based on
27 invasive + in situ tumors in the subcohort with interviews) for the models discussed above were
28 estimated using the actuarial program (life-table analysis). As noted in Section 4.1.1.2, a 1%
29 extra risk level is a more reasonable response level for defining the POD for these epidemiologic
30 data than a 10% level. The results are presented in Table 4-13.
31 Because EtO is DNA-reactive and has direct mutagenic activity (see Section 3.3.3),
32 which is one of the cases cited by EPA's Guidelines for Carcinogen Risk Assessment (U.S. EPA,
33 2005a) for the use of linear low-dose extrapolation, a linear low-exposure extrapolation was
34 performed. The inhalation unit risk estimates for the different breast cancer incidence models
35 considered suitable for low-exposure extrapolation are presented in Table 4-13. As discussed
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-37 DRAFT—DO NOT CITE OR QUOTE
-------
This document is a draft for review purposes only and does not constitute Agency policy.
7/20 13 4-38 DRAFT— DO NOT CITE OR QUOTE
2.6
2.4
HI
c 2.2
OJ
1 2
01
S 1.8
ITS
-o 1.6
o
M—
fjl.4
-t-»
in
0)
cc 1.2
oe
1 ^
o.s
c
/
^xl^-
^^7^
* ^^^/ -----
*r^^ -^~ ""
1r_ '+/
41 /
— — loglinearspline
linear spline
linear {l+p*exp}
•* categorical
i/
+
) 10000 20000 30000 40000
mean cumulative exposure {ppm * days)
50000
figure 4-6. RR estimate for breast cancer incidence in subcohort with interviews vs. mean exposure (with
[5-year lag, unadjusted for continuous exposure); select models compared to deciles.
Categorical: linear model (RR = 1 + P x exposure) with categorical exposures; log-linear and linear spline: 2-piece
spline models, both with knots at 5,800 ppm x days (see text); linear: RR = 1 + P x exposure, with exposure as a
continuous variable.
Sources: Steenland analyses in Appendix D.
-------
1
2
3
Table 4-13. ECoi, LECoi, and unit risk estimates for breast cancer incidence
in females—invasive and in situa
Model
Cox regression,
cumulative
exposure,
15-yrlagb
Cox regression, log
cumulative
exposure,
15-yrlagb
Linear regression of
categorical results,
excluding highest
exposure quintile;
cumulative
exposure,
15-yrlagb'd
Low-exposure log-
linear spline,
cumulative
exposure,
15-yrlage
Linear model with
continuous
cumulative
exposure, 15-yrlag11
Low-exposure
linear spline,
cumulative
exposure,
15-yrlage
With interviews
EC01
(ppm)
0.135
0.0000765
0.0257
0.0166
0.0437
0.0112
LEC01
(ppm)
0.0788
0.0000422
0.0118
0.00991
0.0224
0.00576
Unit risk
(per ppm)
c
c
0.847
1.01f
0.4461
1.74fj
Full cohort
EC01
(ppm)
0.237
0.000124
0.0503
LEC01
(ppm)
0.115
0.0000529
0.0188
Unit risk
(per ppm)
C
c
0.532
__g
__g
__g
4
5
6
7
"All-cause mortality adjusted (to dying of something other than breast cancer or developing breast cancer). Unit
risk = 0.01/LECoi. Note that the ECOT and LEC0i results presented here will not exactly match those presented in
Appendix D because, although the regression coefficients reported by Dr. Steenland in Appendix D were used, the
life-table analyses using 2004 all-cause mortality and 2002-2006 cause-specific mortality and incidence rates were
redone to be more up-to-date; the results presented in Appendix D were based on life-table analyses using 2000
all-cause mortality rates and comparable cause-specific rates.
bFrom Tables 4 and 5 of Steenland et al. (2003), Cox regression models.
°Unit risk estimates are not presented for these models because these models were deemed unsuitable for the
derivation of risks from (low) environmental exposure levels (see text).
Degression coefficient derived from linear regression of categorical results, as described in Section 4.1.2.3.
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
4-39 DRAFT—DO NOT CITE OR QUOTE
-------
1 Table 4-13. ECoi, LECoi, and unit risk estimates for breast cancer incidence
2 in females—invasive and in situa (continued)
eFrom low-exposure segment of two-piece spline analysis; see text and Table D-lc of Appendix D for log-linear
model or Table D-li for linear model; two-piece spline analyses not performed for the full cohort. The ECM value is
below the value of 0.075 ppm roughly corresponding to the knot of 5,800 ppm x days [(5,800 ppm x days) x
(10 m3/20 m3) x (240 d/365 d) x (365 d/yr)/70 yr = 0.075 ppm] and, thus, appropriately in the range of the
low-exposure segment.
fFor unit risk estimates above 1, convert to risk per ppb (e.g., 1.74 per ppm = 1.74 x 10~3 per ppb).
8Not estimated.
hFrom linear analyses in Section D.l.b.2 and Table D-li of Appendix D.
'Confidence intervals used in deriving the LECMs were estimated employing the Wald approach. Confidence
intervals for linear RR models, however, in contrast to those for the log-linear RR models, may not be symmetrical.
EPA also evaluated application of a profile likelihood approach for the linear RR models (Langholz and Richardson,
2010), which allows for asymmetric CIs, for comparison with the Wald approach. The MLE for the regression
coefficient of the linear model is 0.0000304 per ppm x day. Using the profile likelihood method, the (95%
one-sided) upper bound on the regression coefficient is 0.0000745 per ppm x day and the (95% one-sided) lower
bound on the regression coefficient is 0.00000975 per ppm x day. Based on these profile likelihood estimates, the
LECoi estimate is 0.0174 ppm, the UEC0i estimate is 0.133 ppm, and the unit risk estimate for breast cancer
incidence from the linear model would have been 0.575 per ppm, slightly higher (29%) than the value of 0.446 per
ppm obtained using the Wald approach.
JConfidence intervals used in deriving the LECMs were estimated employing the Wald approach. Confidence
intervals for linear RR models, however, in contrast to those for the log-linear RR models, may not be symmetrical.
EPA also evaluated application of a profile likelihood approach for the linear RR models (Langholz and Richardson,
2010), which allows for asymmetric CIs, for comparison with the Wald approach. The MLE for the regression
coefficient of the first spline segment is 0.000119 per ppm x day. Using the profile likelihood method, the (95%
one-sided) upper bound on the regression coefficient is 0.000309 per ppm x day and the (95% one-sided) lower
bound on the regression coefficient is 0.000032 per ppm x day. Based on these profile likelihood estimates, the
LECoi estimate is 0.00430 ppm, the UECW estimate is 0.0415 ppm, and the unit risk estimate for breast cancer
incidence from the low-exposure linear spline would have been 2.33 per ppm, slightly higher (34%) than the value
of 1.74 per ppm obtained using the Wald approach.
O
4
5 above, the unit risk estimate based on the two-piece linear spline model using cumulative
6 exposure with a 15-year lag (i.e., 1.74 per ppm, or 1.74 x 10 3 per ppb) is the preferred estimate.
7 The two-piece log-linear spline model resulted in a unit risk estimate of 1.01 per ppm, while the
8 linear regression of categorical results yielded a unit risk estimate of 0.847 per ppm and the
9 continuous linear model produced a unit risk estimate of 0.446 per ppm; these alternate estimates
10 are about 60%, 50%, and 25%, respectively, of the estimate based on the preferred two-piece
11 linear spline model. ECoi and LECoi estimates from the other models examined are presented
12 for comparison only, to illustrate the differences in model behavior at the low end of the
13 exposure-response range. Unit risk estimates are not presented for these other models because,
14 as discussed above, the log cumulative exposure Cox regression model was considered overly
15 supralinear and the cumulative exposure Cox regression model was considered overly sublinear
16 for the data in the lower exposure range (e.g., first 4 quintiles of exposure). As one can see from
17 the results for the subcohort with interviews, the standard Cox regression cumulative exposure
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-40 DRAFT—DO NOT CITE OR QUOTE
-------
1 model, with its extreme sublinearity in the lower exposure region, yields a notably higher ECoi
2 estimate (0.135 ppm) than that from the two-piece linear spline model (0.0112 ppm), while the
3 log cumulative exposure model, with its extreme supralinearity in the lower exposure region,
4 yields a substantially lower ECoi estimate (0.0000765 ppm). Converting the units, the preferred
5 unit risk estimate of 1.74 per ppm corresponds to an estimate of 9.51 x 10~4 per ug/m3 for breast
6 cancer incidence.
7 As discussed above, the primary risk calculations for breast cancer incidence were based
8 on invasive and in situ tumors in the subcohort of women with interviews, and the primary
9 model was the two-piece linear spline model. For this assessment, the two-piece spline analyses
10 were not performed with the full cohort and the life-table analyses were not replicated for the
11 invasive cancers only. In the 2006 Draft Assessment (U.S. EPA, 2006a), however, comparison
12 analyses were done. Using the linear regression of the categorical results, the comparable unit
13 risk estimate for the full cohort was about 40% lower than the estimate based on the subcohort
14 with interviews. The corresponding unit risk estimate derived based on the subcohort results but
15 using invasive breast cancer only for the background incidence rates was about 17% lower than
16 the estimate based on invasive and in situ tumors, reflecting the difference between incidence
17 rates for invasive breast cancer only and for combined in situ and invasive breast cancer.
18 The unit risk estimate of 1.74 per ppm (1.74 x 10~3 per ppb) is the preferred estimate for
19 female breast cancer risk because it is based on incidence data versus mortality data, it is based
20 on more cases (n = 233) than the mortality estimate (n = 103), and information on personal
21 breast cancer risk factors obtained from the interviews is taken into account. Furthermore, the
22 two-piece linear spline model, which uses the complete data set with exposure as a continuous
23 variable, was statistically significant and had the lowest AIC and the best apparent visual fit to
24 the lower-exposure data of the models considered. Converting the units, 1.74 per ppm
25 corresponds to a unit risk of 9.51 x 10 4 per ug/m3.
26
27 4.1.3. Total Cancer Risk Estimates
28 According to EPA's Guidelines for Carcinogen Risk Assessment (U.S. EPA, 2005a),
29 cancer risk estimates are intended to reflect total cancer risk, not site-specific cancer risk;
30 therefore, an additional calculation was made to estimate the combined risk for (incident)
31 lymphoid and breast cancers, because females would be at risk for both cancer types. Assuming
32 that the cancer types are independent and that the risk estimates are approximately normally
33 distributed, one can estimate the 95% UCL (one-sided) on the total risk as the 95% UCL on the
34 sum of the maximum likelihood estimates (MLEs) of the risk estimates according to the formula
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-41 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
8
9
10
11
12
95% UCL = MLE + 1.645(SE),
(4-3)
13
14
15
16
17
18
19
20
21
22
23
24
where MLE is the MLE of total cancer risk (i.e., the sum of the individual MLEs) and the SE of
the sum of the MLEs is the square root of the sum of the individual variances (i.e., the variance
of the sum is the sum of the variances, and the SE is the square root of the variance). First, an
ECoi of 0.0078 ppm for the total cancer risk (i.e., lymphoid cancer incidence + breast cancer
incidence) was estimated, as summarized in Table 4-14.
Table 4-14. Calculation of ECoi for total cancer risk
Cancer type
Lymphoid
Breast
Total3
ECoi
(ppm)
0.0254
0.0112
-
0.01/ECoi
(per ppm)
0.394
0.893
1.29
ECoi for total
cancer risk
(ppm)
-
-
0.00775
aThe total 0.01/ECM value equals the sum of the individual 0.01/ECM values; the ECM for the total
cancer risk then equals 0.017(0.01/EC0i).
Then, a unit risk estimate of 2.3 per ppm for the total cancer risk (i.e., lymphoid cancer
incidence + breast cancer incidence) was derived, as shown in Table 4-15. An LECoi estimate of
0.00441 ppm for the total cancer risk can be calculated as 0.017(2.27 per ppm).
Table 4-15. Calculation of total cancer unit risk estimate
Cancer type
Lymphoid
Breast
Total
Unit risk
estimate
(per ppm)
0.877
1.74
-
0.01/ECoi
(per ppm)
0.394
0.893
1.29
SEa
(per ppm)
0.294
0.515
(0.593)b
Variance
0.0864
0.265
0.351
Total cancer
unit risk
estimate
(per ppm)
-
-
2.27C
LECoi for total
cancer riskd
(ppm)
-
-
0.00441
25
26
27
28
29
30
aSE = (unit risk - 0.01/EC0i)/1.645.
bThe SE of the total cancer risk is calculated as the square root of the sum of the variances (next column), not as the
sum of the SEs.
"Total cancer unit risk = 1.29 + 1.645 x 0.593.
dThe LECoi for the total cancer risk equals 0.01/(total cancer unit risk estimate).
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
4-42 DRAFT—DO NOT CITE OR QUOTE
-------
1 Thus, the total cancer unit risk estimate is 2.3 per ppm (or 2.3 x 10 3 per ppb;
2 1.2 x 1CT3 per ug/m3). Recall that this is the unit risk estimate derived under the assumption that
3 RR is independent of age (see Section 4.1.1.2). The preferred assumption of increased early-life
4 susceptibility, in accordance with EPA's Supplemental Guidance (U.S. EPA, 2005b), is
5 considered in Section 4.4. While there are uncertainties regarding the assumption of a normal
6 distribution of risk estimates, the resulting unit risk estimate is appropriately bounded in the
7 roughly twofold range between estimates based on the sum of the individual MLEs (i.e., 1.29)
8 and the sum of the individual 95% UCLs (i.e., unit risk estimates, 2.6), or more precisely in this
9 case, between the largest individual unit risk estimate (1.74) and the sum of the unit risk
10 estimates (2.6). Thus, any inaccuracy in the total cancer risk estimate resulting from the
11 approach used to combine risk estimates across cancer types is relatively minor.
12
13 4.1.4. Sources of Uncertainty in the Cancer Risk Estimates
14 The two major sources of uncertainty in quantitative cancer risk estimates are generally
15 interspecies extrapolation and high-dose to low-dose extrapolation. The risk estimates derived
16 from the Steenland et al. (2003) and Steenland et al. (2004) and additional Steenland (see
17 Appendix D) analyses are not subject to interspecies uncertainty because they are based on
18 human data. Furthermore, the human-based estimates are less affected by high-dose to low-dose
19 extrapolation than are rodent-based estimates and, thus, uncertainty from that source is reduced
20 somewhat. For example, the average exposure in the NIOSH cohort was more than 10 times
21 lower than the lowest exposure level in a rodent bioassay after adjustment to continuous lifetime
22 exposure. Nonetheless, uncertainty remains in the extrapolation from occupational exposures to
23 lower environmental exposures. Although the actual exposure-response relationship at low
24 exposure levels is unknown, the clear evidence of EtO mutagenicity supports the linear
25 low-exposure extrapolation that was used (U.S. EPA, 2005a).
26 Because of the existence of endogenous EtO (see Section 3.3.3.1), several members of
27 the SAB panel that reviewed EPA's external review draft assessment felt that the
28 exposure-response relationship for cancer at low exposures would be nonlinear and suggested
29 that it would be consistent with EPA's 2005 Guidelines for Carcinogen Risk Assessment (U.S.
30 EPA, 2005a) to present a nonlinear approach for "extrapolation" to lower exposures (SAB,
31 2007). EPA considered this suggestion but judged that the support for a nonlinear approach was
32 inadequate. In brief, as discussed in Sections 3.1 through 3.3.3, EtO is a DNA-reactive,
33 mutagenic, multisite carcinogen in humans and experimental species; as such, it has the
34 hallmarks of a compound for which low-dose linear extrapolation is strongly supported under
35 EPA's Guidelines for Carcinogen Risk Assessment (U.S. EPA, 2005a). EPA's Guidelines for
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-43 DRAFT—DO NOT CITE OR QUOTE
-------
1 Carcinogen Risk Assessment (U.S. EPA, 2005a) do provide for presenting alternate approaches
2 when those alternatives have significant biological support; however, EPA's analysis of the
3 arguments for using a nonlinear approach presented on page 23 and in Appendix C of the SAB
4 report (SAB, 2007) did not find these arguments to be persuasive. The arguments posited by the
5 SAB panel members who supported using a nonlinear approach were largely that (1) DNA
6 adducts may show a nonlinear response when identical adducts are formed endogenously and
7 (2) mutations do not have linear relationships with exposure but exhibit an "inflection point."
8 However, as discussed in Section 3.3.3.1, recent data from Marsden et al. (2009) support a linear
9 exposure-response relationship for EtO exposure and DNA adducts (p < 0.05) and demonstrate
10 increases of DNA adducts from exogenous EtO exposure above those from endogenous EtO for
11 very low exposures to exogenous EtO, providing direct evidence against argument (1).
12 Moreover, Appendix C of the SAB report (SAB, 2007) presents two EtO-specific mutation data
13 sets in support of argument (2); however, EPA's analysis of these data sets finds that they are in
14 fact consistent with low-dose linearity. See the response to this comment under charge question
15 2.b in Appendix H for a more comprehensive discussion of EPA's consideration and rejection of
16 a nonlinear approach and for the details of EPA's analysis of the two EtO mutation data sets.
17 Other sources of uncertainty emanate from the epidemiologic studies and their analyses
18 (Steenland et al., 2004; Steenland et al., 2003; Steenland analyses in Appendix D), including the
19 retrospective estimation of EtO exposures in the cohort, the modeling of the epidemiologic
20 exposure-response data, the proper dose metric for exposure-response analysis, and potential
21 confounding or modifying factors. Although these are common areas of uncertainty in
22 epidemiologic studies, they were generally well addressed in the NIOSH studies.
23 Regarding exposure estimation, the NIOSH investigators conducted a detailed
24 retrospective exposure assessment to estimate the individual worker exposures. They used
25 extensive data from 18 facilities, spanning a number of years, to develop a regression model
26 (Hornung et al., 1994; Greife et al., 1988) [see also Section A.2.8 for more details about the
27 development and evaluation of the regression model]. The model accounted for 85% of the
28 variation in average EtO exposure levels in an independent set of test data. In addition, the
29 modeled estimates were not highly biased nor biased in one direction when compared to the
30 predictions of a panel of 11 industrial hygiene experts familiar with EtO levels in the sterilization
31 industry. Detailed work history data for the individual workers were collected for the 1987
32 follow-up (Steenland et al., 1991). For the extended follow-up (Steenland et al., 2004; Steenland
33 et al., 2003), additional information on the date last employed was obtained for those workers
34 still employed and exposed at the time of the original work history collection for the plants still
35 using EtO (25% of the cohort). It was then assumed that exposure for these workers continued
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-44 DRAFT—DO NOT CITE OR QUOTE
-------
1 until the date of last employment and that their exposure level stayed the same as that in their last
2 job held at the time of the original data collection. Thus, there would be more exposure
3 misclassification in the extended follow-up. However, when the investigators compared
4 cumulative exposures estimated with and without the extended work histories, they found little
5 difference because exposure levels were very low by the mid-1980s and, therefore, had little
6 impact on cumulative exposure (Steenland et al., 2004; Steenland et al., 2003). While the
7 NIOSH regression model performed well in estimating exposures in validation tests (Hornung et
8 al., 1994), there is, nonetheless, uncertainty associated with any retrospective exposure
9 assessment, and this can affect the ability to discriminate among exposure-response models.
10 With respect to the lymphohematopoietic cancer response, it is not clear exactly which
11 lymphohematopoietic cancer subtypes are related to EtO exposure, so analyses were done for
12 both lymphoid cancers and all lymphohematopoietic cancers (Steenland et al., 2004). The
13 associations observed for all lymphohematopoietic cancers was largely driven by the lymphoid
14 cancer responses, and biologically, there is stronger support for an etiologic role for EtO in the
15 development of the more closely related lymphoid cancers than in the development of the more
16 diverse cancers in the aggregate all lymphohematopoietic cancer grouping; thus, the lymphoid
17 cancer analysis is the preferred analysis for the lymphohematopoietic cancers. Nonetheless, the
18 preferred unit risk estimate for all lymphohematopoietic cancers was similar to (about 50%
19 greater than) that for the lymphoid cancers.
20 For the lymphoid cancer response (Steenland et al., 2004), modeling the
21 exposure-response relationship is limited by the small number of cases (n = 53). The Cox
22 proportional hazards model used by Steenland et al. (2004) is commonly used for this type of
23 analysis because exposure can be modeled as a continuous variable, competing causes of
24 mortality can be taken into account, and potential confounding factors can be controlled for in
25 the regression. Normally, model dependence should be minimized by the practice, under EPA's
26 2005 Guidelines for Carcinogen Risk Assessment (U.S. EPA, 2005a), of modeling only in the
27 observable range and then performing a linear extrapolation from the "POD" (in this case the
28 LECoi). However, the log cumulative exposure Cox regression model with 15-year lag, which
29 provides the best fit to the overall data, is too steep in the low-exposure region and then plateaus
30 rapidly at higher exposures, making it difficult to derive stable risk estimates (i.e., estimates that
31 are not highly dependent on the POD). And the alternative cumulative exposure model, though
32 typically used for epidemiologic data, is too sublinear in the low-exposure region for these data,
33 which exhibit supralinearity. EPA attempted to fit two-piece log-linear and linear spline models
34 to the individual continuous data to address the supralinearity of the data while avoiding the
35 extreme low-exposure curvature of the log cumulative exposure model; however, these models
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-45 DRAFT—DO NOT CITE OR QUOTE
-------
1 resulted in low-exposure slopes that appeared to be implausibly steep (i.e., they suggested
2 excessively large changes in risk from small changes in exposure). The steep low-exposure
3 slopes are a manifestation of apparently high risks in workers with relatively low exposures;
4 however, this elevation is based on small numbers of cancer cases in that exposure range, and
5 EPA has low confidence in the low-exposure slopes. The two-piece spline model with the knot
6 at a higher exposure level could have been used, but without model likelihood as a basis for knot
7 selection, such selection becomes arbitrary, and with the knot at the higher exposure level which
8 had an apparent local maximum for the log-linear model (1,600 ppm x days rather than 100 ppm
9 x days), the visual fit was poor (see Figure 4-1). Thus, EPA opted for a weighted linear
10 regression model based on the Cox regression categorical results, excluding the highest exposure
11 group, to reflect the exposure-response relationship in the exposure region below the "plateau."
12 The all lymphohematopoietic cancer data set had more cases (n = 74) but was heavily dominated
13 by the lymphoid cancer response and conveyed the same problems for exposure-response
14 modeling; thus, a linear regression model, excluding the highest exposure group, was used for
15 this data set as well.
16 The linear model is a parsimonious choice that assumes neither a sublinear nor a
17 supralinear exposure-response relationship and acknowledges the inherent imprecision in the
18 epidemiological data. The highest exposure group was excluded because it is less relevant to the
19 low-exposure risks of interest for low-exposure extrapolation and its inclusion would have overly
20 influenced the linear regression, resulting in a slope that would have substantially underestimated
21 the apparent low-exposure risks. Excluding data can appear arbitrary, but EPA aimed to avoid
22 an arbitrary selection by using the a priori exposure groups presented by Steenland et al. (2004)
23 and excluding only the highest exposure group, with the exposures least relevant to low
24 environmental exposure levels. The linear regression has its own limitations (e.g., it is based on
25 categorical rather than continuous data and the slopes were not statistically significant);
26 nonetheless, it was judged to be the most reasonable approach for deriving low-exposure risk
27 estimates from the available lymphohematopoietic cancer data.
28 Although the linear regression model of the categorical results seems to be a reasonable
29 approach for best reflecting the exposure-response results at the lower end of the exposure range,
30 clearly there is uncertainty regarding the exposure-response model. The log cumulative
31 exposure Cox regression model, which was the best-fitting model overall of the models
32 investigated, yields lower ECoi and LECoi estimates than the linear regression model of the
33 categorical results (see Table 4-6), but the estimates based on the linear regression model are
34 preferred because the linear regression model is more stable.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-46 DRAFT—DO NOT CITE OR QUOTE
-------
1 Another, more minor area of uncertainty related to the exposure-response modeling is the
2 lag period. The best-fitting models presented by Steenland et al. (2004) for
3 lymphohematopoietic cancer mortality had a 15-year lag (lag periods of 0, 5, 10, 15, and
4 20 years were considered). A 15-year lag period means that exposures in the 15 years prior to
5 death or the end of follow-up are not taken into account. In other words, in the best-fitting
6 models, relevant exposures for the development of the lymphohematopoietic cancers occurred
7 over 15 years before death. For the best-fitting continuous model for lymphoid cancer reported
8 by Steenland et al. (2004), the log cumulative exposure Cox regression model, the actual
9 difference between the regression coefficients from the 15-year-lagged and the unlagged models
10 was negligible (the regression coefficient from the unlagged model was about 8% lower than that
11 from the 15-year-lagged model; however, it should be noted that the unlagged model did not
12 provide a statistically significant fit to the data (p = 0.17) (the results for the unlagged model are
13 presented in Section D.3.e of Appendix D).
14 In addition, the analyses of the NIOSH investigators indicate that the regression
15 coefficient for cumulative exposure might have decreased with increasing follow-up, suggesting
16 that the higher exposure levels encountered by the workers in the more distant past are having
17 less of an impact on more recent risk. The regression coefficient for lymphoid cancers was
18 1.2 x 10"5 per ppm x day, for both sexes with a 10-year lag, in the 1987 follow-up (Stayner et al.,
19 1993) versus 4.7 x 10 6 per ppm x day, for both sexes with a 15-year lag, in the 1998 follow-up
20 (see Steenland reanalyses in Appendix D). A similar decrease was found in the regression
21 coefficient for cumulative exposure for all lymphohematopoietic cancers. The life-table analysis
22 used in this dose-response assessment assumes exposure accrues over the full lifetime for the
23 cumulative exposure metric. If, in fact, exposures in the distant past cease to have a meaningful
24 impact on the risk of lymphohematopoietic cancers, this approach would tend to overestimate the
25 unit risk. Thus, a comparison analysis was conducted to evaluate the impact of ignoring
26 exposures over 55 years in the past in the life-table analysis. The actual value of such a cut
27 point, if warranted, is unknown. A value less than 55 years might not be appropriate because
28 exposures for some of the workers began in 1943, so any diminution of potency for past
29 exposures occurring since 1943 is already reflected in the regression coefficient with follow-up
30 through 1998, at least for those workers, although it is unknown what proportion of workers had
31 such early exposures and how long they survived. The comparison analysis for lymphoid cancer
32 yielded an LECoi of 0.0156 ppm and a unit risk estimate of 0.64 per ppm, which is about
33 27% less than the estimate obtained from the unrestricted life-table analysis. Because the
34 appropriate cut point for excluding past exposures is unknown and the unit risk estimate from the
35 linear regression model of the categorical results is already substantially less than that obtained
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-47 DRAFT—DO NOT CITE OR QUOTE
-------
1 from the best-fitting log cumulative exposure Cox regression model, the estimate from the full
2 life-table analysis is preferred. In any event, the preferred estimate is not appreciably different
3 from the estimate from the analysis which considered only the most recent 55 years of exposure
4 in the life-table analysis.
5 Several dose metrics (cumulative exposure, duration of exposure, maximum [8-hour
6 TWA] exposure, and average exposure) were analyzed by Steenland et al. (2004), and
7 cumulative exposure was the best predictor of mortality from lymphohematopoietic cancers.
8 Cumulative exposure is considered a good measure of total exposure because it integrates
9 exposure (levels) over time.
10 Also, the important potential modifying/confounding factors of age, sex, race, and
11 calendar time were taken into account in the analysis, and the plants included in this cohort were
12 specifically selected for the absence of any known confounding exposures (Stayner et al., 1993).
13 With respect to the breast cancer mortality response (Steenland et al., 2004), the
14 exposure-response modeling was based on 103 deaths. As for the lymphohematopoietic cancer
15 responses, the exposure-response data for breast cancer mortality are fairly supralinear,
16 especially for the low-exposure groups. An attempt was again made to fit two-piece log-linear
17 and linear spline models to the individual continuous data to address the supralinearity of the
18 data while avoiding the extreme low-exposure curvature of the log cumulative exposure Cox
19 regression model; however, these models resulted in low-exposure slopes that appeared to be
20 implausibly steep and the model fits were not convincing (i.e., they were neither statistically
21 significant nor visually compelling; see Figure 4-3). Thus, the same linear regression approach,
22 excluding the highest exposure group, was taken to obtain a regression coefficient for the
23 life-table analysis. As discussed above, the linear regression has its own limitations (e.g., it is
24 based on categorical rather than continuous data and the slope is not statistically significant);
25 nonetheless, it was judged to be the most reasonable approach for deriving low-exposure risk
26 estimates from the available breast cancer mortality data.
27 For the lag period, the best-fitting model had a lag of 20 years, which was the longest lag
28 period investigated. This is a commonly used lag period for solid tumors, which typically have
29 longer latency periods than lymphohematopoietic cancers. It is unknown whether a lag period
30 longer than 20 years would have provided a better model fit. The Steenland et al. (2004)
31 analysis took into account age, race, and calendar time. Other risk factors for breast cancer could
32 not be included in the mortality analysis, but many of these factors were considered in the breast
33 cancer incidence study (Steenland et al., 2003), as discussed below, and the preferred breast
34 cancer risk estimates are based on the breast cancer incidence data.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-48 DRAFT—DO NOT CITE OR QUOTE
-------
1 Steenland et al. (2003) conducted an incidence study for breast cancer; therefore, it was
2 not necessary to calculate unit risk estimates for breast cancer incidence indirectly from the
3 mortality data as was done for lymphohematopoietic cancer. Further advantages to using the
4 results from the incidence study are that more cases were available for the exposure-response
5 modeling (319 cases) and that the investigators were able to include data on potential
6 confounders in the modeling for the subcohort with interviews (233 cases). Because the
7 subcohort with interviews had complete case ascertainment and provided data on potential
8 confounders, it was the preferred breast cancer incidence data set, although some results based
9 on the full cohort are presented for comparison. For the full cohort, the continuous exposure Cox
10 regression model providing the best fit to the data was again the log cumulative exposure model.
11 With breast cancer incidence, a 15-year lag provided the best model fits. For the subcohort, the
12 cumulative exposure and log cumulative exposure Cox regression models fit nearly equally well.
13 For both groups, the categorical Cox regression results suggest that a linear model lying between
14 the supralinear log cumulative exposure model and the sublinear cumulative exposure model
15 would better represent the low-exposure data than either of the two presented
16 continuous-variable models (see Figures 4-4 and 4-5). Thus, for both groups, in EPA's original
17 draft analyses based on the published summary data, a linear regression was fitted to the
18 categorical results, dropping the highest exposure group to provide a better fit to the
19 lower-exposure data (U.S. EPA, 2006a). In addition, in subsequent analyses by Dr. Steenland
20 (see Appendix D) of the individual data using exposure as a continuous variable, two-piece
21 log-linear and linear spline models and other linear RR models were used to model the subcohort
22 data; the two-piece linear spline model was the best-fitting of these models and provided the
23 preferred breast cancer incidence risk estimates.
24 Confidence intervals were determined using the Wald approach. Confidence intervals for
25 linear RR models, however, in contrast to those for the log-linear RR models, may not be
26 symmetrical. EPA also evaluated application of a profile likelihood approach for the linear RR
27 models (Langholz and Richardson, 2010), which allows for asymmetric CIs, for comparison with
28 the Wald approach. Using the profile likelihood method and the two-piece linear spline model,
29 the resulting unit risk estimate for breast cancer incidence would have been 2.33 per ppm,
30 slightly higher (34%) than the value of 1.74 per ppm obtained as the unit risk estimate for breast
31 cancer incidence in this assessment. These results suggest that if the profile likelihood method
32 had been used for the linear RR models in this assessment, the total cancer risk estimate, which
33 incorporates the breast cancer incidence estimate as a component, would be less than 34% higher
34 than the total cancer risk estimate presented here.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-49 DRAFT—DO NOT CITE OR QUOTE
-------
1 With respect to the two-piece spline models, the use of this model form is not intended to
2 imply that an abrupt change in biological response occurs at the knot but, rather, to allow
3 description of an exposure-response relationship in which the slope of the relationship differs
4 notably in the low-exposure versus high-exposure regions. The two-piece model is used here
5 primarily for its representation of the low-exposure data. The main uncertainty in the two-piece
6 spline models is in the selection of the knot, and the location of the knot is critical in defining the
7 low-exposure slope. The model likelihood was used to provide a statistical basis for knot
8 selection; although, as shown in Appendix D (see Figure D-la), the likelihood did not generally
9 change appreciably over a range of possible knots. Thus, because of the importance of knot
10 selection, a sensitivity analysis was done to examine the impacts of selecting different knots (see
11 Section D.6 of Appendix D). For the sensitivity analysis, the two-piece log-linear model was run
12 with knots roughly one increment (1,000 ppm x days) below and one increment above the
13 selected knot. For breast cancer incidence, this sensitivity analysis yielded ECoi estimates of
14 0.0133 ppm and 0.0176 ppm, respectively (i.e., about 14% lower and 14% higher, respectively,
15 than the ECoi of 0.0154 ppm obtained with the originally selected knot of 6,000 ppm x days).28
16 As can be seen in Table 4-13, there is substantial variation in the ECoi estimates obtained
17 from the different models. Although some plateauing is apparent with the highest exposure
18 group and is evidenced in the subcohort with interviews by the strong influence of the top 5% of
19 cumulative exposures on dampening the slope of the (cumulative exposure) Cox regression
20 model (see Section D.I and Figure D-ld of Appendix D), the categorical data for breast cancer
21 incidence do not display the supralinearity in the lower exposure groups seen in the cases
22 discussed above (i.e., lymphohematopoietic cancers and breast cancer mortality). Thus, for the
23 subcohort with interviews, the difference between the ECoi estimates from the standard
24 cumulative exposure Cox regression model and the two-piece spline models or the linear
25 regression of the categorical results or continuous linear models are not as dramatic as seen in
26 those cases (the ECoi estimates from the latter four approaches are nearly within an order of
27 magnitude of that of the cumulative exposure model). For the subcohort with interviews, the
28 two-piece spline models, the continuous linear model, and the linear regression of the categorical
29 results gave similar results—the unit risk estimates spanned less than a fourfold range. This
30 range is bounded by the two best-fitting (based on AIC) continuous models—the two-piece
31 linear spline model and the continuous linear model. If the continuous linear model had been
32 selected rather than the two-piece linear spline model, which had a slightly lower AIC value and
28 About 12% lower and 17% higher, respectively, than the ECM of 0.0151 ppm obtained with the more finely tuned
knot of 5,800 ppm * days (see Appendix D). The ECM value of 0.0166 presented in this assessment (see
Table 4-13) is not directly comparable to the values in the sensitivity analysis because more recent background
incidence and mortality rates were used in the lifetable analyses upon which the assessment estimates were based.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-50 DRAFT—DO NOT CITE OR QUOTE
-------
1 a better apparent visual fit of the lower-exposure data, the breast cancer incidence unit risk
2 estimate would have been 0.446 per ppm rather than 1.74 per ppm, and the total cancer unit risk
3 estimate would have been 1.15 per ppm rather than 2.27 per ppm. In other words, of the models
4 investigated, the total cancer unit risk estimate from the best-fitting alternate model (based on
5 AIC) is about 50% lower than that of the best-fitting model. However, data in the lower
6 exposure range of greatest relevance for the derivation of a unit risk estimate support a steeper
7 slope in the lower exposure range; thus, although the lower estimate obtained from the
8 continuous linear model is plausible, unit risk estimates notably lower than that are considered
9 unlikely from the available data.
10 The best-fitting models presented by Steenland et al. (2003) for breast cancer incidence
11 generally had a 15-year lag (lag periods of 0, 5, 10, 15, and 20 years were considered). A
12 15-year lag period means that exposures in the 15 years prior to diagnosis or the end of
13 follow-up are not taken into account. For the various continuous models for breast cancer
14 incidence in the full cohort and the subcohort with interviews reported by Steenland et al. (2003,
15 Tables 4 and 5), none of the unlagged models provided a statistically significant fit to the data,
16 with the exception of the log cumulative exposure Cox regression model for the subcohort,
17 where the unlagged model fit marginally better than the 15-year-lagged model. However, as
18 noted in Section 4.1.2.3, the log cumulative exposure model with no lag was considered less
19 biologically realistic than the corresponding model with a 15-year lag because some lag period
20 would be expected for the development of breast cancer; thus, the 15-year-lagged model was
21 used in this assessment. The regression coefficient from the unlagged log cumulative exposure
22 Cox regression model was about 90% higher than that from the 15-year-lagged model.
23 With respect to dose metrics for breast cancer incidence, models using duration provided
24 better model fits than those using cumulative exposure (Steenland et al., 2003); however,
25 duration is less useful for estimating unit risks and the cumulative exposure models also provided
26 statistically significant fits to the data, thus the cumulative exposure metric was used for the
27 quantitative risk estimates. Models using peak or average exposure did not fit as well.
28 Regarding potential confounders/modifying factors, analyses for the full cohort were
29 adjusted for age, race, and calendar time, and exposures to other chemicals in these plants were
30 reportedly minimal. For the subcohort with interviews, a number of specific breast cancer risk
31 factors were investigated, including body mass index, breast cancer in a first-degree relative,
32 parity, age at menopause, age at menarche, socioeconomic status, and diet; however, only parity
33 and breast cancer in a first-degree relative were determined to be important predictors of breast
34 cancer and were included in the final models.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-51 DRAFT—DO NOT CITE OR QUOTE
-------
1 An area of uncertainty in the life-table analysis for breast cancer incidence pertains to the
2 rates used for the cause-specific background rate. The regression coefficients presented by
3 Steenland et al. (2003) represent invasive and in situ cases combined, where 6% of the cases are
4 in situ, and the preferred unit risk estimates in this assessment are calculated similarly using
5 background rates for invasive and in situ cases combined. The regression coefficients for
6 invasive and in situ cases combined should be good approximations for regression coefficients
7 for invasive cases alone; however, it is uncertain how well they reflect the exposure-response
8 relationships for in situ cases alone. Diagnosed cases of in situ breast cancer would presumably
9 be remedied and not progress to invasive breast cancer, so double-counting is unlikely to be a
10 significant problem. Carcinoma in situ is a risk factor for invasive breast cancer; however, this
11 observation is most likely explained by the fact that these two types of breast cancer have other
12 breast cancer risk factors in common, some of which have been considered in the subcohort
13 analysis. One might hypothesize that EtO exposure could cause a more rapid progression to
14 invasive tumors; however, there is no specific evidence that this occurs. On the other hand, there
15 is some indication that in situ cases in the incidence study might have been diagnosed at
16 relatively low rates in comparison to the invasive cases. Steenland et al. (2003) reported that 6%
17 of the cases in their study are in situ; according to the National Cancer Institute, however, ductal
18 carcinoma in situ accounted for about 18% of newly diagnosed cases of breast cancer in 1998
19 (NCI, 2004).
20 There are several possible explanations for this difference. One is that it reflects
21 differences in diagnosis with calendar time because the rate of diagnosis of carcinoma in situ has
22 increased over time with increased use of mammography. Another is that the difference is
23 partially a reflection of the age distribution in the cohort because the proportion of new cases
24 diagnosed as carcinoma in situ varies by age. A third possible explanation is that the low
25 proportion of in situ cases is at least partially a consequence of underascertainment of cases
26 because in situ cases will not be reported on death certificates, although, even if all 20 in situ
27 cases were in the subcohort with interviews, that would still be only 8.6% of the cases. In any
28 event, this is a relatively minor source of uncertainty, and a comparison of the unit risk estimates
29 using invasive + in situ breast cancer background rates and invasive-only background rates,
30 using EPA's original analyses in the 2006 Draft Assessment, found that the estimate based on the
31 invasive + in situ background rates was less than 20% higher than the corresponding estimate
32 using only invasive breast cancer background rates (U.S. EPA, 2006a).
33 The results for the subcohort with interviews are used for the primary breast cancer unit
34 risk calculations because, in addition to including the data on potential confounders, the
35 subcohort is considered to have full ascertainment of the breast cancer cases, whereas the full
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-52 DRAFT—DO NOT CITE OR QUOTE
-------
1 cohort for the incidence study has incomplete case ascertainment, as illustrated by the fact that
2 death certificates were the only source of case ascertainment for 14% of the cases. Complete
3 interviews were available for only 68% of the 7,576 women in the full incidence cohort, and
4 thus, some potential exists for participation selection bias in the subcohort. There is, however,
5 no basis for considering participation to be associated with breast cancer or EtO exposure, and
6 the major reason for nonparticipation was a failure to locate (22% of full incidence cohort) and
7 not lack of response (3% of cohort) or refusal to participate (7% of cohort). Risk estimates based
8 on the full cohort were calculated for comparison with the subcohort estimates using the original
9 linear regression analyses of the categorical results. The unit risk estimate based on the
10 subcohort was about 60% higher than the corresponding estimate from the full cohort (U.S. EPA,
11 2006a).
12 Some additional sources of uncertainty are not so much inherent in the exposure-response
13 modeling or in the epidemiologic data themselves but, rather, arise in the process of obtaining
14 more general Agency risk estimates from the epidemiologic results. EPA cancer risk estimates
15 are typically derived to represent an upper bound on increased risk of cancer incidence for all
16 sites affected by an agent for the general population. From experimental animal studies, this is
17 accomplished by using tumor incidence data and summing across all the tumor sites that
18 demonstrate significantly increased incidences, customarily for the most sensitive sex and
19 species, to be protective of the general human population. However, in estimating comparable
20 risks from the NIOSH epidemiologic data, certain limitations are encountered. First, the study
21 reported by Steenland et al. (2004) is a retrospective mortality study, and cancer incidence data
22 are not available for lymphohematopoietic cancer (for breast cancer, a separate incidence study
23 [Steenland et al., 2003] was available). Second, these occupational epidemiology data represent
24 a healthy-worker cohort. Third, the epidemiologic study may not have sufficient statistical
25 power and follow-up time to observe associations for all the tumor sites that may be affected by
26 EtO.
27 The first limitation was addressed quantitatively in the life-table analysis for the
28 lymphohematopoietic cancer risk estimates. Although assumptions are made in using incidence
29 rates for the cause-specific background rates, as discussed in Section 4.1.1.3, the resulting
30 incidence-based estimates are believed to be better estimates of cancer incidence risk than are the
31 mortality-based estimates. The incidence unit risk estimate is about 120% higher than (i.e.,
32 2.2 times) the mortality-based estimate, which seems reasonable given the relatively high
33 survival rates for lymphoid cancers (according to SEER data [www.seer.cancer.gov], 5-year
34 survival rates are 65% for NHL; 78% for chronic lymphocytic leukemia, which are the vast
35 majority of the lymphocytic leukemias in adults; and 40% for multiple myeloma).
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-53 DRAFT—DO NOT CITE OR QUOTE
-------
1 The healthy-worker effect is often an issue in occupational epidemiology studies, but the
2 internal exposure-response analyses conducted by these investigators help address this concern,
3 at least partially. In terms of representing the general population, the NIOSH study cohort was
4 relatively diverse. It contained both female (55%) and male workers, and the workers were 79%
5 white, 16% black, and 5% "other." Furthermore, because of EtO's mutagenic mode of action,
6 increased early-life susceptibility is assumed and ADAFs are applied for exposure scenarios
7 involving early life (see Section 4.4).
8 With respect to other possible tumor sites of concern, the rodent data suggest that
9 lymphohematopoietic cancers are a major tumor type associated with EtO exposure in female
10 mice and in male and female rats. Thus, it is reasonable that this might be a cancer type of
11 concern in humans also. Likewise, the mouse data suggest an increased risk of mammary gland
12 tumors from EtO exposure, and evidence of that can be seen in the Steenland et al. (2004) and
13 Steenland et al. (2003) studies. However, the rodent data suggest associations between EtO
14 exposure and other tumor types as well, and although site concordance across species is not
15 generally assumed, it is possible that the NIOSH study, despite its relatively large size and long
16 follow-up (mean length of follow-up was 26.8 years), had insufficient power to observe small
17 increases in risk in certain other sites. For example, the tumor site with the highest potency
18 estimate in both male and female mice was the lung. In the NIOSH study, one cannot rule out a
19 small increase in the risk of lung cancer, which has a high background rate.
20 To obtain the risk estimate for total cancer risk (2.3 per ppm, or 2.3 x 10 3 per ppb), the
21 preferred estimates for lymphoid cancer incidence and breast cancer incidence were combined.
22 While there are uncertainties in the approach used to combine the individual estimates, the
23 resulting unit risk estimate is appropriately bounded in the roughly twofold range between
24 estimates based on the sum of the individual MLEs of risk and the sum of the individual 95%
25 UCLs, and thus, any inaccuracy in the total cancer unit risk estimate resulting from the approach
26 used is relatively minor. Because the breast cancer component of the total cancer risk estimate
27 applies only to females, the total cancer risk estimate is expected to overestimate the cancer risk
28 to males somewhat (the preferred unit risk estimate for lymphoid cancer alone was 0.877 per
29 ppm [or 8.77 x 10~4 per ppb], which is about 40% of the total cancer risk estimate).
30 Despite these uncertainties, the inhalation cancer unit risk estimate of 2.3 per ppm (or
31 2.3 x 1Q~3 per ppb) for the total cancer risk from lymphoid cancer incidence and female breast
32 cancer incidence has the advantages of being based on human data from a large, high-quality
33 epidemiologic study with individual exposure estimates for each worker. Furthermore, the breast
34 cancer component of the risk estimate, which contributes approximately 60% of the total cancer
35 risk, is based on a substantial number of incident cases (233 total, the vast majority of which
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-54 DRAFT—DO NOT CITE OR QUOTE
-------
1 were in the exposure range below the knot of 5,800 ppm x days [see Table D.la of
2 Appendix D]).
3 A further area of uncertainty pertains to the assumption that RR is independent of age,
4 which is a common assumption in the dose-response modeling of epidemiological data and is an
5 underlying assumption in the Cox regression model. For the NIOSH worker cohort, the
6 proportional hazards model assumption of RR being independent of age was tested by checking
7 the significance of an interaction between age and cumulative exposure, and none of the models
8 had a significant interaction term. This suggests that, for adults at least, the assumption that RR
9 is independent of age is valid. However, the worker cohort contains no children and is
10 uninformative on the issue of early-life susceptibility. In the absence of data on early-life
11 susceptibility, EPA's Supplemental Guidance (U.S. EPA, 2005b) recommends that increased
12 early-life susceptibility be assumed for carcinogens with a mutagenic mode of action, and the
13 conclusion was made in Section 3.4 that the weight of evidence supports a mutagenic mode of
14 action for EtO. Thus, in accordance with the Supplemental Guidance, the alternate assumption
15 of increased early-life susceptibility is preferred as the basis for risk estimates in this assessment,
16 and risk estimates derived under this preferred assumption are presented in Section 4.4.
17
18 4.1.5. Summary
19 Under the common assumption that RR is independent of age, an inhalation unit risk
20 estimate for lymphoid cancer incidence of 0.877 per ppm (or 8.77 x 10 4 per ppb; 4.79 x 10 4 per
21 ug/m3) was calculated using a life-table analysis and a weighted linear regression of the
22 categorical Cox regression results, excluding the highest exposure group, for excess lymphoid
23 cancer mortality from a high-quality occupational epidemiology study. Similarly, an inhalation
24 unit risk estimate for female breast cancer incidence of 1.74 per ppm (or 1.74 x 10~3 per ppb;
25 9.51 x 10 4 per ug/m3) was calculated using a life-table analysis and two-piece linear spline
26 modeling of the continuous data for excess breast cancer incidence from the same high-quality
27 occupational epidemiology study. The linear regression of the categorical results with the
28 exclusion of the highest exposure group for the lymphoid cancer results and the two-piece linear
29 spline analysis for the breast cancer incidence data were different modeling approaches used to
30 address the supralinearity of the exposure-response data in the two data sets. Low-dose linear
31 extrapolation was used, as warranted by the clear mutagenicity of EtO. An ECoi estimate of
32 0.0078 ppm, a LECoi estimate of 0.0044 ppm, and a unit risk estimate of 2.3 per ppm (or
33 2.3 x 10 3 per ppb; 1.2 x 10 3 per ug/m3) were obtained for the total cancer risk combined across
34 both cancer types. Despite the uncertainties discussed above, this inhalation unit risk estimate
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-55 DRAFT—DO NOT CITE OR QUOTE
-------
1 has the advantages of being based on human data from a high-quality epidemiologic study with
2 individual exposure estimates for each worker.
3 In the absence of data on early-life susceptibility, EPA's Supplemental Guidance (U.S.
4 EPA, 2005a) recommends that increased early-life susceptibility be assumed for carcinogens
5 with a mutagenic mode of action, and the conclusion was made in Section 3.4 that the weight of
6 evidence supports a mutagenic mode of action for EtO. Thus, in accordance with the
7 Supplemental Guidance., the alternate assumption of increased early-life susceptibility is
8 preferred as the basis for risk estimates in this assessment, and risk estimates derived under this
9 preferred assumption are presented in Section 4.4. Other than the use of the alternate assumption
10 about early-life susceptibility, the approach used to derive the estimates presented in Section 4.4
11 is identical to the approach used for the estimates derived here in Section 4.1, and the
12 comparisons made between various options and the issues and uncertainties discussed here in
13 Section 4.1 are applicable to the estimates derived in Section 4.4.
14
15 4.2. INHALATION UNIT RISK DERIVED FROM EXPERIMENTAL ANIMAL DATA
16 4.2.1. Overall Approach
17 Lifetime animal cancer bioassays of inhaled EtO have been carried out in three
18 laboratories, as described in Section 3.2. The data from these reports are presented in Tables 3-3
19 through 3-5. These studies have also been reviewed by the IARC (1994b) and Health Canada
20 (2001). Health Canada calculated the EDos for each data set using the benchmark dose
21 methodology. The EOIC report (EOIC, 2001) tabulated only lymphatic tumors because they
22 constituted the predominant risk.
23 The overall approach in this derivation is to find a unit risk for each of the
24 bioassays—keeping data on males and females separate—from data on the incidence of all tumor
25 types and then to use the maximum of these values as the summary measure of the unit risk from
26 animal studies (i.e., the unit risk represents the most sensitive species and sex). The unit risk for
27 the animals in these bioassays is converted to a unit risk in humans by first determining the
28 continuous exposures in humans that are equivalent to the rodent bioassay exposures and then by
29 assuming that the lifetime incidence in humans is equivalent to lifetime incidence in rodents, as
30 is commonly accepted in interspecies risk extrapolations. For cross-species scaling of exposure
31 levels (see Section 4.2.2 below), an assumption of ppm equivalence is used; thus, no interspecies
32 conversion is needed for the exposure concentrations. Bioassay exposure levels are adjusted to
33 equivalent continuous exposures by multiplying by (hours of exposure/24 hours) and by (5/7) for
34 the number of days exposed per week. The unit risk in humans (risk per unit air concentration)
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-56 DRAFT—DO NOT CITE OR QUOTE
-------
1 is then assumed to be numerically equal to that in rodents (after adjustment to continuous
2 exposures); the calculations from the rodent bioassay data are shown in Tables 3-3 through 3-5.
O
4 4.2.2. Cross-Species Scaling
5 In the absence of chemical-specific information, EPA's 1994 inhalation dosimetry
6 methods (U.S. EPA, 1994) provide standard methods and default scaling factors for
7 cross-species scaling. Under EPA's methodology, EtO would be considered a Category 2 gas
8 because it is reactive and water soluble and has clear systemic distribution and effects.
9 Dosimetry equations for Category 2 gases are undergoing EPA re-evaluation and are not being
10 used at this time. For cross-species scaling of extrarespiratory effects, current practice is to treat
11 Category 2 gases as Category 3 gases. For Category 3 gases, ppm equivalence is assumed (i.e.,
12 responses across species are equivalent on a ppm exposure basis), unless the airblood partition
13 coefficient for the experimental species is less than the coefficient for humans (U.S. EPA, 1994,
14 p. 4-61). In the case of EtO, measured airblood partition coefficients are 78 in the mouse
15 (Fennell and Brown, 2001), 64 in the rat (Krishnan et al., 1992), and 61 in the human (Csanady
16 et al., 2000); thus, ppm equivalence for cross-species scaling to humans can be assumed for
17 extrarespiratory effects observed in mice and rats. The assumption of ppm equivalence is further
18 supported by the PBPK modeling of Fennell and Brown (2001), who reported that simulated
19 blood AUCs for EtO after 6 hours of exposure to concentrations between 1 ppm and 100 ppm
20 were similar for mice, rats, and humans and were linearly related to the exposure concentration
21 (see Section 3.3.1 and Figure 3-2). This modeling was validated against measured blood EtO
22 concentrations for rodents and humans. For Category 2 gases with respiratory effects, there is no
23 clear guidance on an interim approach. One suggested approach is to do cross-species scaling
24 using both Category 1 and Category 3 gas equations and then decide which is most appropriate.
25 In this document, the preferred approach was to assume ppm equivalence was also valid for the
26 lung tumors in mice because of the clear systemic distribution of EtO (e.g., see Section 3.1).
27 Treating EtO as a Category 1 gas for cross-species scaling of the lung tumors would presume
28 that the lung tumors are arising only from the immediate and direct action of EtO as it comes into
29 first contact with the lung. In fact, some of the EtO dose contributing to lung tumors is likely
30 attributable to recirculation of systemic EtO through the lung.
31 If one were to treat EtO as a Category 1 gas for the cross-species scaling of the lung
32 tumor response as a bounding exercise, EPA's 1994 inhalation dosimetry methods present
33 equations for estimating the RGDRpu, i.e., the regional gas dose ratio for the pulmonary region,
34 which acts as an adjustment factor for estimating human equivalent exposure concentrations
35 from experimental animal exposure concentrations (adjusted for continuous exposure) (U.S.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-57 DRAFT—DO NOT CITE OR QUOTE
-------
1 EPA, 1994, pp. 4-49 to 4-51). These equations rely on parameters describing mass transport of
2 the gas (EtO) in the extrathoracic and tracheobronchial regions for both the experimental animal
3 species (mouse) and humans. Without experimental data for these parameters, it seems
4 reasonable to estimate RGDRpu using a simplified equation and the adjusted alveolar ventilation
5 rates of Fennell and Brown (2001). Fennell and Brown adjusted the alveolar ventilation rates to
6 reflect limited pulmonary uptake of EtO, a phenomenon commonly observed for highly
7 water-soluble gases (Johanson and Filser, 1992). The adjusted ventilation rates were then used
8 by Fennell and Brown in their PBPK modeling simulations, and good fits to blood concentration
9 data were reported for both the mouse and human models. In this document, the adjusted
10 alveolar ventilation rates were used to estimate the RGDRpu as follows:
11
12
13 RGDRpu = (RGDpu)m/(RGDpu)h = (Qaiv/S APU)m/(Qaiv/S APU)h, (4-4)
14 where:
15 RGDpu = regional gas dose to the pulmonary region,
16 Qaiv = (adjusted) alveolar ventilation rate,
17 SApu = surface area of the pulmonary region, and
18 the subscripts "m" and "h" denote mouse and human values.
19
20
21 Then, using adjusted alveolar ventilation rates from Fennell and Brown (2001) and surface area
22 values from EPA (U.S. EPA, 1994, p. 4-26),
23
24
25 RGDRpu = ((0.78L/h)/(0.05m2))/((255L/h)/(54.0m2)) = 3.3. (4-5)
26
27
28 Using this value for the RGDRpu would increase the human equivalent concentration about
29 threefold, resulting in a decreased risk for lung tumors of about threefold, as a lower bound. The
30 true value of the RGDRpu is expected to be between 1 and 3, and any adjustment to the lung
31 tumor risks would still be expected to result in unit risk estimates roughly within the range of the
32 rodent unit risk estimates derived later in Section 4.2 under the assumption of ppm equivalence.
33
34 4.2.3. Dose-Response Modeling Methods
35 In this document the following steps were used:
36 1. Extract the incidence data presented in the original studies. In order to crudely adjust
37 for early mortality in the analysis of the NTP (1987) data, the incidence data have been corrected
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-58 DRAFT—DO NOT CITE OR QUOTE
-------
1 for a specific tumor type by eliminating the animals that died prior to the occurrence of the first
2 tumor or prior to 52 weeks, whichever was earlier. It was not possible to make this adjustment
3 with the other studies where data on individual animals were not available. With these
4 exceptions, the tumor incidence data in Tables 3-3 through 3-5 match the original data.
5 2. Fit the multistage model to the dose-response data using the Tox Risk program.
6 The likelihood-ratio test was used to determine the lowest value of the multistage
7 polynomial degree that provided the best fit to the data while requiring selection of the most
8 parsimonious model. In this procedure, if a good fit to the data in the neighborhood of the POD
9 is not obtained with the multistage model because of a nonmonotonic reduction in risk at the
10 highest dose tested (as sometimes occurs when there is early mortality from other causes), that
11 data point is eliminated and the model is fit again to the remaining data. Such a deletion was
12 found necessary in two cases (mammary tumors in the NTP study and mononuclear cell
13 leukemia in the Lynch study). The goodness-of-fit measures for the dose-response curves and
14 the parameters derived from them are shown in Appendix G.
15 In the NTP bioassay, where the individual animal data were available, a time-to-tumor
16 analysis was undertaken to account for early mortality. The general model used in this analysis
17 is the multistage Weibull model:
18
19
20 P(d,t)=l-exp[-(q0 + qid + q2d2 + ... + qkdk)x(t-to)z], (4-6)
21
22
23 where P(d,t) represents the probability of a tumor by age t (in bioassay weeks) for dose d (i.e.,
24 human equivalent exposure), and the parameter ranges are restricted as follows: z > 1, to > 0,
25 and q; > 0 for i = 0, 1, ..., k. The parameter to represents the time between when a potentially
26 fatal tumor becomes observable and when it causes death. The analyses were conducted using
27 the computer software Tox_Risk version 3.5, which is based on methods developed by Krewski
28 et al. (1983). Parameters are estimated in Tox_Risk using the method of maximum likelihood.
29 Tumor types can be categorized by tumor context as either fatal or incidental. Incidental
30 tumors are those tumors thought not to have caused the death of an animal, whereas fatal tumors
31 are thought to have resulted in animal death. Tumors at all sites were treated as incidental
32 (although it was recognized that this may not have been the case, the experimental data are not
33 detailed enough to conclude otherwise). The parameter to was set equal to 0 because there were
34 insufficient data to reliably estimate it.
35 The likelihood-ratio test was used to determine the lowest value of the multistage
36 polynomial degree k that provided the best fit to the data while requiring selection of the most
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-59 DRAFT—DO NOT CITE OR QUOTE
-------
1 parsimonious model. The one-stage Weibull (i.e., k = 1) was determined to be the most optimal
2 value for all the tumor types analyzed.
3 3. Select the POD and calculate the unit risk for each tumor site. The effective
4 concentration that causes a 10% extra risk for tumor incidence, ECio, and the 95% lower bound
5 of that concentration, LECio, are derived from the dose-response model. The LECio is then used
6 as the POD for a linear low-dose extrapolation, and the unit risk is calculated as 0.1/LECio. This
7 is the procedure specified in the EPA's Guidelines for Carcinogen Risk Assessment (U.S. EPA,
8 2005a) for agents such as EtO that have direct mutagenic activity. See Section 3.4 for a
9 discussion of the mode of action for EtO. Tables 3-3 through 3-5 present the unit risk estimates
10 for the individual tumor sites in each bioassay.
11 4. Develop a unit risk estimate based on the incidence of all tumors combined. This
12 method assumes that occurrences of tumors at multiple sites are independent and, further, that
13 the risk estimate for each tumor type is normally distributed. Then, at a given exposure level, the
14 MLEs of extra risk due to each tumor type are added to obtain the MLE of total cancer risk. The
15 variances corresponding to each tumor type are added to give the variance associated with the
16 sum of the MLEs. The one-sided 95% UCL of the MLE for the combined risk is then calculated
17 as:
18
19
20 95% UCL = MLE + 1.645(SE), (4-7)
21
22
23 where SE is the standard error and is the square root of the summed variance. (Note that as a
24 precursor to this step, when Tox _Risk is used to fit the incidence of a single tumor type, it
25 provides the MLE and 95% UCL of extra risk at a specific dose. The standard error in the MLE
26 is determined using the above formula). The calculation is repeated for a few exposure levels,
27 and the exposure yielding a value of 0.1 for the upper bound on extra risk is determined by
28 interpolation. The unit risk is then the slope of the linear extrapolation from this POD. The
29 results are given in Table 4-16.
30
31 4.2.4. Description of Experimental Animal Studies
32 NTP (1987) exposed male and female B6C3Fi mice to concentrations of 0, 50, and
33 100 ppm for 6 hours per day, 5 days per week, for 102 weeks. An elevated incidence of lung
34 carcinomas was found in males, and elevated lung carcinomas, malignant lymphomas, uterine
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-60 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
Table 4-16. Upper-bound unit risks (per ug/m3) obtained by combining
tumor sites
Combination method"
UCL on sum of risks0
Sum of unit risksd
Time-to-tumor analysis and
u.c.b on sum of risks0
NTP (1987)
female mouse
2.71 x 10'5
4.12 x 10'5
4.55 x 10'5
Lynch et al. (1984a);
Lynch et al. (1984b)
male rat
4.17 x 1(T5
3.66 x 1(T5
-
Snellings et al. (1984)b
Male rat
2.19 x 1(T5
2.88 x 1(T5
-
Female rat
3.37 x lO'5
3.54 x 10'5
-
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
aUnit risk in these methods is the slope of the straight line extrapolation from a point of departure at the dose
corresponding to a value of 0.1 for the 95% upper confidence bound on total extra risk.
Includes data on brain tumors from the analysis by Garman et al. (1985). See Table 3-3.
°UCL = 95% upper confidence bound. At a given dose, the MLE of the combined extra risk was determined by
summing the MLE of risk due to each tumor type. The variance associated with this value was determined by
summing over the variances due to each tumor type.
dSum of values in last column of Tables 3-1 through 3-3.
adenocarcinomas, and mammary carcinomas were found in females. These data are shown in
Table 3-3.
Lynch et al. [Lynch et al. (1984a); Lynch et al. (1984b)] exposed male F344 rats to 0, 50,
and 100 ppm for 7 hours per day, 5 days per week, for 2 years. They found excess incidence of
tumors at three sites: mononuclear cell leukemia in the spleen, testicular peritoneal
mesothelioma, and brain glioma. In this study the survival in the high-dose group (19%) was
less than that of controls (49%), which reduced the incidence of leukemias. In the animals in the
high-dose group that survived to the termination of the experiment, the incidence of leukemias
was statistically significantly higher than for controls (p < 0.01). The incidence data are shown
in Table 3-4, uncorrected for the high-dose-group mortality. If the individual animal data were
available to perform the correction, the incidence would be higher. Therefore, using these data
results in an underestimate of risk.
Snellings et al. (1984) exposed male and female F344 rats to 0, 10, 33, and 100 ppm for
6 hours per day, 5 days per week, for 2 years and described their results for all sites except the
brain. In two subsequent publications for the same study, Garman et al. (1986, 1985) described
the development of brain tumors in a different set of F344 rats. The Snellings et al. (1984)
publication reported an elevated incidence of splenic mononuclear cell leukemia and peritoneal
mesothelioma in males and an elevated incidence of splenic mononuclear cell leukemia in
females. The mortality was higher in the 100-ppm groups than the other three groups for both
males and females. The incidences in the animals killed after 24 months in Snellings et al.
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
4-61 DRAFT—DO NOT CITE OR QUOTE
-------
1 (1984) are shown in Table 3-5. Table 3-5 also presents the brain tumor incidence data for male
2 and female rats from the Garman et al. (1986, 1985) publications. The brain tumor incidence
3 was lower than that of the other tumors, particularly the splenic mononuclear cell leukemias.
4
5 4.2.5. Results of Data Analysis of Experimental Animal Studies
6 The unit risks calculated from the individual site-sex-bioassay data sets are presented in
7 Tables 3-3 through 3-5. The highest unit risk of any individual site is 3.23 x 10 5 per ug/m3,
8 which is for mononuclear cell leukemia in the female rats of the Snellings et al. (1984) study.
9 Table 4-17 presents the results of the time-to-tumor method applied to the individual
10 animals in the NTP bioassay, compared with the results from the dose group incidence data in
11 Table 3-3. This comparison was done for each tumor type separately. The time-to-tumor
12 method of analyzing the individual animals results in generally higher unit risk estimates than
13 does the analysis of dose group data, as shown in Table 4-17. The ratio is not large (less than
14 2.2) across the tumor types. (In the case of mammary tumors this ratio is actually less than 1. It
15 must be noted that the incidence at the highest dose [where the incidence was substantially less
16 than at the intermediate dose] was deleted from the analysis of grouped data, whereas it was
17 retained in the time-to-tumor analysis. Therefore, the comparison for the mammary tumors is
18 not a strictly valid comparison of methods.) The results also show the extent to which a time-to-
19 tumor analysis of individual animal data increases the risk estimated from data on dose groups.
20 It is expected that if individual animal data were available for the Lynch et al. (1984a); Lynch et
21 al. (1984b) and the Snellings et al. (1984) bioassays, then the time-to-tumor analysis would also
22 result in higher estimates because both those studies also showed early mortality in the highest
23 dose group.
24 The results of combining tumor types are summarized in Table 4-16. The sums of the
25 individual unit risks tabulated in Tables 3-3 to 3-5 are given in the second row of Table 4-16.
26 Note that as expected they are greater than the unit risks computed from the upper bound on the
27 sum of risks for all data sets except for the Lynch et al. [Lynch et al. (1984a); Lynch et al.
28 (1984b)] data. The reason for this exception is not known, but the differences are small. It is
29 likely that the problem arises from the methodology used to combine the risks across tumor sites.
30 In an attempt to be consistent with the new two-step methodology (i.e., modeling in the
31 observable range to a POD and then doing a linear extrapolation to zero extra risk at zero
32 exposure), the exposure concentration at which the sum of the independent tumor site risks
33 yielded a 95% upper bound on 10% extra risk was estimated and used as the POD. Summing
34 risks in this way results in a POD for the combined tumor risk that is different (lower) than the
35 points of departure for each individual tumor site risk. Thus, the risk estimate for the sum is not
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-62 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
Table 4-17. Unit risk values from multistage Weibulla time-to-tumor
modeling of mouse tumor incidence in the NTP (1987) study
Tumor type
Unit risk, 0.1/LEC10
(per jig/m3)
from time-to-tumor
analysis
Unit risk,
0.1/LEdo
(per jig/m3)
(Table 3-3)b
Ratio of unit risks
time-to-
tumor/grouped data
Males
Lung: alveolar/bronchiolar
adenoma and carcinoma
3.01 x 10'5
2.22 x 10'5
1.4
Females
Lung: alveolar/bronchiolar
adenoma and carcinoma
Malignant lymphoma
Uterine carcinoma
Mammary carcinoma
2.40 x 10'5
1.43 x 10'5
6.69 x 10'6
8.69 x 1Q-6
1.10 x 10'5
7.18 x 10'6
4.33 x 10'6
1.87 x 1Q-5
2.2
2.0
1.5
0.5
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
aP(d,t) D1 ~ exp[-(q0 + qid + q2d2 + ... + qkdk) x (t - t0)z], where d is inhaled ethylene oxide concentration in ppm, t
is weeks until death with tumor. In all cases, k = I provided the optimal model.
blncidence data modeled using multistage model without taking time to tumor into account.
strictly comparable to the individual risks that constitute it. These tumor-site-specific risks were
based on points of departure individually calculated to correspond with a 10% extra risk. In any
event, adding the upper bound risks of individual tumor sites should overestimate the upper
bound of the sum, and the latter is the preferred measure of the total cancer risk because it avoids
the overestimate. However, for the exceptional Lynch et al. [Lynch et al. (1984a); Lynch et al.
(1984b)] data, the sum of upper bounds, 3.66 x 10 5 per ug/m3, is already an overestimate of the
total risk, and this value is preferred over the anomalously high value of 4.17 x 10~5 per ug/m3
corresponding to the upper bound on the sum of risks. The latter value is considered to be an
excessive overestimate and is therefore not carried over into the summary Table 4-18. For the
Snellings et al. (1984) data sets, the upper confidence bound on the sum of risks is used in the
summary Table 4-18. The results of the sum-of-risks calculations on the NTP bioassay time-to-
tumor data are included in the third row of Table 4-16. The estimate for the NTP female mice is
4.55 x 1Q~5 per ug/m3, which is higher than the other two measures of total tumor risk in that
bioassay. This value is preferable to the other measures because it utilizes the individual animal
data available for that bioassay.
Summary of results. The summary of unit risks from the five data sets is shown in
Table 4-18. The data set giving the highest risk (4.55 x io~5 per ug/m3) is the NTP (1987) data
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
4-63 DRAFT—DO NOT CITE OR QUOTE
-------
1 Table 4-18. Summary of unit risk estimates (per ug/m3) in animal bioassays
2
Assay
NTP (1987), B6C3FJ mice
Lynch et al. (1984a); Lynch et al. (1984b),
F344 rats
Snellings et al. (1984), F344 rats
Males
3.01 x lQ-5a
3.66x 1Q-5C
2.19 x lQ-5d
Females
4.55 x lQ-5b
_
3.37 x lQ-5d
4 "From time-to-tumor analysis of lung adenomas and carcinomas, Table 4-17.
5 bUpper bound on sum of risks from the time-to-tumor analysis of the NTP data, Table 4-16.
6 °Sum of (upper bound) unit risks (see text for explanation), Table 4-16.
7 dUpper bound on sum of risks, Table 4-16.
8
9
10 on combined tumors in female mice. The other values are within about a factor of 2 of the
11 highest value.
12
13 4.3. SUMMARY OF INHALATION UNIT RISK ESTIMATES—NOT ACCOUNTING
14 FOR ASSUMED INCREASED EARLY-LIFE SUSCEPTIBILITY
15 For both humans and laboratory animals, tumors occur at multiple sites. In humans, there
16 was a combination of tumors having lymphohematopoietic, in particular lymphoid, origins in
17 both sexes and breast cancer in females, and, in rodents, lymphohematopoietic tumors, mammary
18 carcinomas, and tumors of other sites were observed. From human data, an extra cancer unit risk
19 estimate of 4.79 x 10 4 per ug/m3 (8.77 x 10 4 per ppb) was calculated for lymphoid cancer
20 incidence, and a unit risk estimate of 9.31 x 10 4 per ug/m3 (1.74 x 10 3 per ppb) was calculated
21 for breast cancer incidence in females. The total extra cancer unit risk estimate was 1.2 x io~3
22 per ug/m3 (2.3 x 10 3 per ppb) for both cancer types combined (ECoi = 0.0078 ppm;
23 LECoi = 0.0043 ppm). Unit risk estimates derived from the three chronic rodent bioassays for
24 EtO ranged from 2.2 x 10"5 per ug/m3 to 4.6 x 10~5 per ug/m3, over an order of magnitude lower
25 than the estimates based on human data.
26 Adequate human data, if available, are considered to provide a more appropriate basis
27 than rodent data for estimating human risks (U.S. EPA, 2005a), primarily because uncertainties
28 in extrapolating quantitative risks from rodents to humans are avoided. Although there is a
29 sizeable difference between the rodent-based and the human-based estimates, the human data are
30 from a large, high-quality study, with EtO exposure estimates for the individual workers and
31 little reported exposure to chemicals other than EtO. Therefore, the total extra cancer unit risk
32 estimate of 1.2 x 10 3 per ug/m3 (2.3 x 10 3 per ppb) calculated for lymphoid cancers and breast
33 cancer combined is the preferred estimate of those estimates not taking assumed increased early-
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-64 DRAFT—DO NOT CITE OR QUOTE
-------
1 life susceptibility into account (estimates accounting for assumed increased early-life
2 susceptibility are presented in Section 4.4). The unit risk estimate is intended to be an upper
3 bound on cancer risk for use with exposures below the POD (i.e., the LECoi). The unit risk
4 estimate should not generally be used above the POD; however, in the case of this total extra
5 cancer unit risk, which is based on cancer type-specific unit risk estimates from two linear
6 models, the estimate should be valid for exposures up to about 0.075 ppm (140 ug/m3), which is
7 the minimum of the limits for the lymphoid cancer unit risk estimate (0.090 ppm; see
8 Section 4.1.1.2) and the breast cancer unit risk estimate (0.075 ppm; see Section 4.1.2.3).
9 Because a mutagenic mode of action for EtO carcinogenicity (see Section 3.3.2) is
10 "sufficiently supported in (laboratory) animals" and "relevant to humans", and as there are no
11 chemical-specific data to evaluate the differences between adults and children, increased
12 early-life susceptibility should be assumed and, if there is early-life exposure, the age-dependent
13 adjustment factors (ADAFs) should be applied, as appropriate, in accordance with EPA's
14 Supplemental Guidance (U.S. EPA, 2005b; see Section 4.4 below for more details on the
15 application of ADAFs).
16
17 4.4. ADJUSTMENTS FOR POTENTIAL INCREASED EARLY-LIFE
18 SUSCEPTIBILITY
19 There are no chemical-specific data on age-specific susceptibility to EtO-induced
20 carcinogenesis. However, there is sufficient weight of evidence to conclude that EtO operates
21 through a mutagenic mode of action (see Section 3.4.1). In such circumstances (i.e., the absence
22 of chemical-specific data on age-specific susceptibility but sufficient evidence of a mutagenic
23 mode of action), EPA's Supplemental Guidance for Assessing Susceptibility from Early-Life
24 Exposure to Carcinogens (U.S. EPA, 2005b) recommends the assumption of increased early-life
25 susceptibility and the application of default age-dependent adjustment factors (ADAFs) to adjust
26 for this potential increased susceptibility from early-life exposure. See the Supplemental
27 Guidance for detailed information on the general application of these adjustment factors. In
28 brief, the Supplemental Guidance establishes ADAFs for three specific age groups. The current
29 ADAFs and their age groupings are 10 for <2 years, 3 for 2 to <16 years, and 1 for 16 years and
30 above (U.S. EPA, 2005b). For risk assessments based on specific exposure assessments, the
31 10-fold and 3-fold adjustments to the unit risk estimates are to be combined with age-specific
32 exposure estimates when estimating cancer risks from early-life (<16 years of age) exposure.
33 These ADAFs, however, were formulated based on comparisons of the ratios of cancer
34 potency estimates from juvenile-only exposures to cancer potency estimates from adult-only
35 exposures from rodent bioassay data sets with appropriate exposure scenarios, and they are
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-65 DRAFT—DO NOT CITE OR QUOTE
-------
1 designed to be applied to cancer potency estimates derived from adult-only exposures. Thus,
2 alternate life-table analyses were conducted to derive comparable adult-exposure-only unit risk
3 estimates to which ADAFs would be applied to account for early-life exposure. For these
4 alternate life-table analyses, it was assumed that RR is independent of age for adults, which
5 represent the life stage for which the exposure-response data and the Cox regression modeling
6 results from the NIOSH cohort study specifically pertain, but that there is increased early-life
7 susceptibility, based on the weight-of-evidence-based conclusion that EtO carcinogenicity has a
8 mutagenic mode of action (see Section 3.4), which supersedes the assumption that RR is
9 independent of age for all ages including children.
10 In the alternate analyses, exposure in the life table was taken to start at age 16 years, the
11 age cut point that was established in EPA's Supplemental Guidance (U.S. EPA, 2005b), to derive
12 an adult-exposure-only unit risk estimate to which ADAFs would be applied to account for
13 early-life exposure. Other than the age at which exposure was initiated, the life-table analyses
14 are identical to those conducted for the results presented in Section 4.1. Adult-exposure-only
15 unit risk estimates were derived for both cancer incidence and mortality for both lymphoid and
16 breast cancers. Alternate estimates were not derived for all lymphohematopoietic cancers
17 because lymphoid cancer was the preferred endpoint (see Section 4.1.1.2). Incidence estimates
18 are preferred over mortality estimates, but both are calculated here for comparison and because
19 mortality estimates are sometimes used in addition to incidence estimates in benefit-cost
20 analyses. For each cancer endpoint, the same exposure-response model was used as that which
21 was selected for the unit risk estimates in Section 4.1 (i.e., linear regression of the categorical
22 results, excluding the highest exposure category, for lymphoid cancer and breast cancer mortality
23 and two-piece linear spline model for breast cancer incidence). The results are presented in
24 Table 4-19 along with the unit risk estimates derived assuming that RR was independent of age
25 for all ages (see Section 4.1) for comparison. As can be seen in Table 4-19, the unit risk
26 estimates for adult-only exposures range from about 66% to about 72% of the unit risk estimates
27 derived under the assumption of age independence across all ages.
28 According to EPA's Guidelines for Carcinogen Risk Assessment (U.S. EPA, 2005a)
29 cancer risk estimates are intended to reflect total cancer risk, not site-specific cancer risk;
30 therefore, an additional calculation was made to estimate the combined risk for (incident)
31 lymphoid and breast cancers from adult-only exposures, because females would be at risk for
32 both cancer types. Assuming that the tumor types are independent and that the risk estimates are
33 approximately normally distributed, this calculation can be made as described in Section 4.1.3.
34 First, an ECoi of 0.0114 ppm for the total cancer risk (i.e., lymphoid cancer incidence + breast
35 cancer incidence) from adult-only exposure was estimated, as summarized in Table 4-20.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-66 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
Table 4-19. ECoi, LECoi, and unit risk estimates for adult-only exposures*
Cancer response
Lymphoid cancer mortality
(both sexes)
Lymphoid cancer
incidence (both sexes)
Breast cancer mortality
(females)
Breast cancer incidence
(females)
ECoi (ppm)
0.0787
0.0364
0.0590
0.0167
LEC01
(ppm)
0.0352
0.0163
0.0297
0.00863
Adult-exposure-
only unit risk
estimate"
(per ppm)
0.284
0.613
0.337
1.16C
Lifetime-exposure unit risk estimate
under assumption of age
independence1"
(per ppm)
0.397
0.877
0.513
1.74C
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
"Unit risk estimate = 0.01/LECm.
bFrom Tables 4-5, 4-9, and 4-13 of Section 4.1.
Tor unit risk estimates above 1, convert to risk per ppb (e.g., 1.16 per ppm = 1.16 x 10"3 per ppb).
*These are intermediate values. See Table 4-22 below for the final adult-based cancer-type-specific unit risk
estimates.
Table 4-20. Calculation of ECoi for total cancer risk from adult-only
exposure
Cancer type
Lymphoid
Breast
Total3
EC01
(ppm)
0.0364
0.0167
-
0.01/ECoi
(per ppm)
0.275
0.599
0.874
ECoi for total risk
(ppm)
~
~
0.0114
aThe total 0.01/ECM value equals the sum of the individual 0.01/ECM values; the ECM for the
total cancer risk then equals 0.01/(0.01/EQ>i).
Then, a unit risk estimate of 1.5 per ppm for the total cancer risk (i.e., lymphoid cancer
incidence + breast cancer incidence) from adult-only exposure was derived, as shown in
Table 4-21. An LECoi estimate of 0.00654 ppm for the total cancer risk can be calculated as
0.017(1.53 per ppm).
Thus, the total cancer unit risk estimate from adult-only exposure is 1.53 per ppm (or
1.53 x 10 3 per ppb; 8.36 x 10 4 per ug/m3). While there are uncertainties regarding the
assumption of a normal distribution of risk estimates, the resulting unit risk estimate is
appropriately bounded in the roughly twofold range between estimates based on the sum of the
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
4-67 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
Table 4-21. Calculation of total cancer unit risk estimate from adult-only
exposure*
Cancer type
Lymphoid
Breast
Total
Adult-exposure-
only unit risk
estimate
(per ppm)
0.613
1.16
~
0.01/ECoi
(per ppm)
0.275
0.599
0.874
SEa
(per ppm)
0.205
0.340
(0.397)b
Variance
0.0422
0.115
0.158
Adult-exposure-only
total cancer unit risk
estimate
(per ppm)
~
~
1.53C
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
*These are intermediate values. See Table 4-22 below for the final adult-based cancer-type-specific unit risk
estimates.
aSE = (unit risk - 0.01/EC0i)/1.645.
bThe SE of the total cancer risk is calculated as the square root of the sum of the variances (next column), not as the
sum of the SEs.
"Total cancer unit risk = 0.874 + 1.645 x 0.397.
individual MLEs (i.e., 0.874) and the sum of the individual 95% UCLs (i.e., unit risk estimates,
1.77), or more precisely in this case, between the largest individual unit risk estimate (1.16) and
the sum of the unit risk estimates (1.77), and thus, any inaccuracy in the total cancer risk estimate
resulting from the approach used to combine risk estimates across cancer types is relatively
minor.
When EPA derives unit risk estimates from rodent bioassay data, there is a blurring of the
distinction between lifetime and adult-only exposures because the relative amount of time that a
rodent spends as a juvenile is negligible (<8%) compared to its lifespan. [According to EPA's
Supplemental Guidance, puberty begins around 5-7 weeks of age in rats and around 4-6 weeks
in mice (U.S. EPA, 2005b)]. Thus, when exposure in a rodent is initiated at 5-8 weeks, as in the
typical rodent bioassay, and the bioassay is terminated after 104 weeks of exposure, the unit risk
estimate derived from the resulting cancer incidence data is considered a unit risk estimate from
lifetime exposure, except when the ADAFs were formulated and are applied, in which case the
same estimate is considered to apply to adult-only exposure. Yet, when adult exposures are
considered in the application of ADAFs, the adult-exposure-only unit risk estimate is pro-rated
over the full default human lifespan of 70 years, presumably because that is how adult exposures
are treated when a unit risk estimate calculated in the same manner from the same bioassay
exposure paradigm is taken as a lifetime unit risk estimate.
However, in humans, a greater proportion of time is spent in childhood (e.g., 16 of
70 years = 23%), and the distinction between lifetime exposure and adult-only exposure cannot
be ignored when human data are used as the basis for the unit risk estimates. Thus, as described
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-68 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
above, adult-exposure-only unit risk estimates were calculated distinct from the lifetime
estimates that were derived in Section 4.1 under the assumption of age independence for all ages.
In addition, the adult-exposure-only unit risk estimates need to be rescaled to a 70-year lifespan
in order to be used in the ADAF calculations and risk estimate calculations involving less-than-
lifetime exposure scenarios in the standard manner, which includes prorating even adult-based
unit risk estimates over 70 years. Thus, the adult-exposure-only unit risk estimates are
multiplied by 70/54 to rescale the 54-year adult period of the 70-year default lifespan to 70 years.
Then, for example, if a risk estimate were calculated for a less-than-lifetime exposure scenario
involving exposure only for the full adult period of 54 years, the rescaled unit risk estimate
would be multiplied by 54/70 in the standard calculation and the adult-exposure-only unit risk
estimate would be appropriately reproduced. Without reseating the adult-exposure-only unit risk
estimates, the example calculation just described for exposure only for the full adult period of
54 years would result in a risk estimate 77% (i.e., 54/70) of that obtained directly from the
adult-exposure-only unit risk estimates, which would be illogical. The rescaled adult-based unit
risk estimates for use in ADAF calculations and risk estimate calculations involving less-than-
lifetime exposure scenarios are presented in Table 4-22. Rescaled LECoi and ECoi estimates for
adult-based total cancer risk are 5.0 x 10~3 ppm (9.2 ug/m3) and 8.8 x 10~3 ppm (16 ug/m3).
Table 4-22. Adult-based unit risk estimates for use in ADAF calculations
and risk estimate calculations involving less-than-lifetime exposure scenarios
Cancer response
Lymphoid cancer mortality
Lymphoid cancer incidence
Breast cancer mortality
Breast cancer incidence
Total cancer incidence
Adult-based unit risk estimate (per
ppm)
0.368
0.795
0.436
1.50a
1.98a
Adult-based unit risk estimate
Ug/m3)
(per
2.01 x 10'4
4.35 x 10'4
2.39 x 10'4
8.21 x 10'4
1.08 x 10'3
23
24
25
26
27
28
29
30
Tor unit risk estimates above 1, convert to risk per ppb (e.g., 1.16perppm= 1.16 x 10"3perppb).
An example calculation illustrating the application of the ADAFs to the human-data-
derived adult-based (rescaled as discussed above) unit risk estimate for EtO for a lifetime
exposure scenario is presented below. For inhalation exposures, assuming ppm equivalence
across age groups, i.e., equivalent risk from equivalent exposure levels, independent of body
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
4-69 DRAFT—DO NOT CITE OR QUOTE
-------
1 size, the ADAF calculation is fairly straightforward. Thus, the ADAF-adjusted lifetime total-
2 cancer unit risk estimate is calculated as follows:
3
4 total cancer risk from exposure to constant EtO exposure level of 1 ug/m3 from ages 0-70 years:
5
6 unit risk exposure duration partial
7 Age group ADAF (per ug/m3) cone (ug/m3) adjustment risk
8 0 to <2 years 10 1.08 x 10~3 1 2 years/70 years 3.09 x 10~4
9 2to<16years 3 1.08 x 1(T3 1 14 years/70 years 6.48 x 1(T4
10 >16 years 1 1.08 x 10~3 1 54 years/70 years 8.33 x 1(T4
11 total lifetime risk = 1.80 x 10~3
12
13 The partial risk for each age group is the product of the values in columns 2-5 [e.g.,
14 10 x (1.08 x io~3) x 1 x 2/70 = 3.09 x 10~4], and the total risk is the sum of the partial risks.
15
16 This 70-year risk estimate for a constant exposure of 1 ug/m3 is equivalent to a lifetime
17 unit risk estimate of 1.8 x 10~3 per ug/m3 (3.3 per ppm, or 3.3 x 10 3 per ppb), adjusted for
18 potential increased early-life susceptibility, assuming a 70-year lifetime and constant exposure
19 across age groups. Note that because of the use of the rescaled adult-based unit risk estimate, the
20 partial risk for the >16 years age group is the same as would be obtained for a 1 ug/m3 constant
21 exposure directly from the total cancer adult-exposure-only unit risk estimate of 8.36 x 10~4 per
22 ug/m3 that was presented above, as it should be (the small difference in the second decimal place
23 is due to round-off error).
24 In addition to the uncertainties discussed above for the inhalation unit risk estimate, there
25 are uncertainties in the application of ADAFs to adjust for potential increased early-life
26 susceptibility. The ADAFs reflect an expectation of increased risk from early-life exposure to
27 carcinogens with a mutagenic mode of action (U.S. EPA, 2005b), but they are general
28 adjustment factors and are not specific to EtO. With respect to the breast cancer estimates, for
29 example, evidence suggests that puberty/early adulthood is a particularly susceptible life stage
30 for breast cancer induction (U.S. EPA, 2005b; Russo and Russo, 1999); however, EPA has not,
31 at this time, developed alternate ADAFs to reflect such a pattern of increased early-life
32 susceptibility, and there is currently no EPA guidance on an alternate approach for adjusting for
33 early-life susceptibility to potential breast carcinogens.
34
35 4.5. INHALATION UNIT RISK ESTIMATES—CONCLUSIONS
36 For both humans and laboratory animals, tumors occur at multiple sites. In humans, there
37 was a combination of tumors having lymphohematopoietic, in particular lymphoid, origins in
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-70 DRAFT—DO NOT CITE OR QUOTE
-------
1 both sexes and breast cancer in females, and, in rodents, lymphohematopoietic tumors, mammary
2 carcinomas, and tumors of other sites were observed. From human data, an extra cancer unit risk
3 estimate of 4.79 x 10 4 per ug/m3 (8.77 x 10 4 per ppb) was calculated for lymphoid cancer
4 incidence, and a unit risk estimate of 9.49 x 10 4 per ug/m3 (1.74 x 10 3 per ppb) was calculated
5 for breast cancer incidence in females, under the assumption that RR is independent of age for all
6 ages (see Section 4.1). The total extra cancer unit risk estimate was 1.24 x 10 3 per ug/m3
7 (2.27 x 1Q~3 per ppb) for both cancer types combined (ECoi = 0.00775 ppm; LECoi =
8 0.00441 ppm). Unit risk estimates derived from the three chronic rodent bioassays for EtO
9 ranged from 2.2 x 10 5 per ug/m3 to 4.6 x 10 5 per ug/m3, over an order of magnitude lower than
10 the estimates based on human data.
11 Because a mutagenic mode of action for EtO carcinogenicity (see Section 3.4.1) is
12 "sufficiently supported in (laboratory) animals" and "relevant to humans," and as there are no
13 chemical-specific data to evaluate the differences between adults and children, increased
14 early-life susceptibility should be assumed, in accordance with EPA's Supplemental Guidance
15 (U.S. EPA, 2005b). This assumption of increased early-life susceptibility supersedes the
16 assumption of age independence under which the human-data-based estimates presented in the
17 previous paragraph were derived. Thus, as described in Section 4.4, adult-exposure-only unit
18 risk estimates were calculated from the human data under an alternate assumption that RR is
19 independent of age for adults, which represent the life stage for which the data upon which the
20 exposure-response modeling was conducted pertain. These adult-exposure-only unit risk
21 estimates were then rescaled to a 70-year basis for use in the standard ADAF calculations and
22 risk estimate calculations involving less-than-lifetime exposure scenarios. The resulting
23 adult-based unit risk estimates were 4.35 x 10~4 per ug/m3 (7.95 x 10~4 per ppb) for lymphoid
24 cancer incidence and 8.21 x 10 4 per ug/m3 (1.50 x 10 3 per ppb) for breast cancer incidence in
25 females. The adult-based total extra cancer unit risk estimate for use in ADAF calculations and
26 risk estimate calculations involving less-than-lifetime exposure scenarios was 1.08 x 10 3 per
27 ug/m3 (1.98 x 10 3 per ppb) for both cancer types combined.
28 For exposure scenarios involving early-life exposure, the age-dependent adjustment
29 factors (ADAFs) should be applied, in accordance with EPA's Supplemental Guidance (U.S.
30 EPA, 2005b). Applying the ADAFs to obtain a full lifetime unit risk estimate yields
31
32
33 1.98/ppm x ((10 x 2 years/70 years) + (3 x 14/70) + (1 x 54/70)) (4-8)
34 = 3.29/ppm = 1.80 x 10~3/(ug/m3).
35
36
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-71 DRAFT—DO NOT CITE OR QUOTE
-------
1 Applying the ADAFs to the unit risk estimates derived from the three chronic rodent bioassays
2 for EtO yields estimates ranging from 3.7 x 10~5 per ug/m3 to 7.6 x 10 5 per ug/m3, still over an
3 order of magnitude lower than the estimate based on human data.
4 Adequate human data, if available, are considered to provide a more appropriate basis
5 than rodent data for estimating human risks (U.S. EPA, 2005a), primarily because uncertainties
6 in extrapolating quantitative risks from rodents to humans are avoided. Although there is a
7 sizeable difference between the rodent-based and the human-based estimates, the human data are
8 from a large, high-quality study, with EtO exposure estimates for the individual workers and
9 little reported exposure to chemicals other than EtO. Therefore, the human-based full lifetime
10 total extra cancer unit risk estimate of 1.8 x io~3 per ug/m3 (3.3 x io~3 per ppb) calculated
11 for lymphoid cancers and breast cancer combined and applying the ADAFs is the preferred
12 lifetime unit risk estimate. For less-than-lifetime exposure scenarios, the human-data-derived
13 (rescaled) adult-based unit risk estimate of 1.1 x 10 3 per ug/m3 (2.0 x 10 3 per ppb) should be
14 used, in conjunction with the ADAFs if early-life exposures occur.
15 Although there are uncertainties in this unit risk estimate, primarily related to exposure
16 misclassification, model uncertainty, and low-dose extrapolation, as discussed in Section 4.1.4,
17 confidence in the unit risk estimate is relatively high. First, there is strong confidence in the
18 hazard characterization of EtO as "carcinogenic to humans," which is based on strong
19 epidemiological evidence supplemented by other lines of evidence, such as genotoxicity in both
20 rodents and humans (see Section 3.5.1). Second, the unit risk estimate is based on human data
21 from a large, high-quality epidemiology study with individual worker exposures estimated using
22 a high-quality regression model (see Section 4.1 and Section A.2.8 of Appendix A). Finally, the
23 use of low-exposure linear extrapolation is strongly supported by the conclusion that EtO
24 carcinogenicity has a mutagenic mode of action (see Section 3.4.1).
25 Confidence in the unit risk estimate is particularly high for the breast cancer component,
26 the largest contributor to the total cancer unit risk estimate, which is based on over 200 incident
27 cases for which the investigators had information on other potential breast cancer risk factors
28 (see Section 4.1.2.3). The selected model for the breast cancer incidence data was the
29 best-fitting model of the models investigated as well as the model that provided the best
30 representation of the categorical results, particularly in the lower exposure range of greatest
31 relevance for the derivation of a unit risk estimate. Alternate estimates calculated from other
32 reasonable models suggest that a unit risk estimate for breast cancer incidence that is fourfold
33 lower (corresponding to a total cancer unit risk estimate of twofold lower) is plausible; however,
34 unit risk estimates notably lower than that are considered unlikely from the available data.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-72 DRAFT—DO NOT CITE OR QUOTE
-------
1 There is lower confidence in the lymphoid cancer component of the unit risk estimate
2 because it is based on fewer events (40 lymphoid cancer deaths); incidence risk was estimated
3 from mortality data; and the exposure-response relationship is exceedingly supralinear, such that
4 continuous models yield apparently implausibly steep low-exposure slopes (see Figure 4-1).
5 Although these continuous models provided statistically significant slope coefficients, there was
6 low confidence in such steep slopes, which, particularly for the two-piece spline models, are
7 highly dependent on a small number of cases in the low-exposure range. Thus, a linear
8 regression model of the categorical results for the lowest three quartiles was used to derive the
9 unit risk estimate for lymphoid cancer, and there was greater confidence in the more moderate
10 slope resulting from that model, although it was not statistically significant, because it was based
11 on more data and provided a good representation of the categorical results across this larger data
12 range in the lower-exposure region (see Section 4.1.1.2). So, while there is lower confidence in
13 the lymphoid cancer unit risk estimate than in the breast cancer unit risk estimate, the lymphoid
14 cancer estimate is considered a reasonable estimate from the available data, and overall, there is
15 relatively high confidence in the total cancer unit risk estimate.
16 The unit risk estimate is intended to be an upper bound on cancer risk for use with
17 exposures below the POD (i.e., the LECoi). The unit risk estimate should not generally be used
18 above the POD; however, in the case of this total extra cancer unit risk, which is based on cancer
19 type-specific unit risk estimates from two linear models, the estimate should be valid for
20 exposures up to about 0.075 ppm (140 ug/m3), which is the minimum of the limits for the
21 lymphoid cancer unit risk estimate (0.090 ppm: see Section 4.1.1.2) and the breast cancer unit
22 risk estimate (0.075 ppm; see Section 4.1.2.3). (See Section 4.7 for risk estimates based on
23 occupational exposure scenarios.)
24 Using the above full lifetime unit risk estimate of 3.3 x 10~3 per ppb (1.8 x 10~3 per
25 ug/m3), the lifetime chronic exposure level of EtO corresponding to an increased cancer risk of
26 10~6 can be estimated as follows:
27
28
29 (10~6)/(3.3/ppm) = 3.0 x 10~7 ppm = 0.00030 ppb = 0.0006 ug/m3. (4-9)
30
31
32 The inhalation unit risk estimate presented above, which is calculated based on a linear
33 extrapolation from the POD (LECoi), is expected to provide an upper bound on the risk of cancer
34 incidence. However, estimates of "central tendency" for the risk below the POD are also
35 presented. Adult-based extra risk estimates per ppm for some of the cancer responses, based on
36 linear extrapolation from the adult-exposure-only ECoi (i.e., 0.01/ECoi) and reseating to a
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-73 DRAFT—DO NOT CITE OR QUOTE
-------
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
70-year basis for use in ADAF calculations and risk estimate calculations involving less-than-
lifetime exposure scenarios (see Section 4.4), are reported in Table 4-23. The adult-exposure-
only ECoiS were from the linear regression models of the categorical results for lymphoid
cancers and breast cancer mortality and from the two-piece linear spline model (low-dose
segment) for breast cancer incidence. (Note that, for each of these models, the low-exposure
extrapolated estimates are a straight linear continuation of the linear models used above the
PODs, and thus, the statistical properties of the models are preserved.) These estimates are
dependent on the suitability of the ECoi estimates as well as on the applicability of the linear
low-dose extrapolation. The assumption of low-dose linearity is supported by the mutagenicity
of EtO (see Section 3.4). If these estimates are to be used, ADAFs should be applied if early-life
exposure occurs, in accordance with EPA's Supplemental Guidance (U.S. EPA, 2005b).
As can be seen by comparing the adult-based rescaled 0.01/ECoiestimates in Table 4-23
with the adult-based unit risk estimates in Table 4-22, the 0.01/ECoi estimates are about 45% of
the unit risk estimates for the lymphoid cancer responses and about 50% of the unit risk
estimates for the breast cancer responses.
Table 4-23. Adult-based extra risk estimates per ppm based on
adult-exposure-only ECoiSa
Cancer response
Lymphoid cancer mortality (both sexes)
Lymphoid cancer incidence (both sexes)
Breast cancer mortality (females)
Breast cancer incidence (females)
Total cancer incidence
ECoi (ppm)
0.0787
0.0364
0.0590
0.0167
0.0114
Adult-based
0.01/ECoi (per ppm)b
0.165
0.356
0.219
0.776
1.14C
"ADAFs should be applied if early-life exposure occurs, in accordance with EPA's Supplemental
Guidance.
bThese estimates are calculated as 0.01/ECM for the adult-exposure-only extra risk estimate per
ppm rescaled to a 70-yr basis by multiplying by 70/54 (see Section 4.4).
Tor unit risk estimates above 1, convert to risk per ppb (e.g., 1.14 per ppm = 1.14 x 10~3 per ppb).
Finally, it should be noted that some investigators have posited that the high and variable
background levels of endogenous EtO-induced DNA damage in the body (see Section 3.3.3.1)
may overwhelm any contribution from low levels of exogenous EtO exposure (Marsden et al.,
2009; SAB, 2007). It is true that the existence of these high and variable background levels may
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
4-74 DRAFT—DO NOT CITE OR QUOTE
-------
1 make it hard to observe statistically significant increases in risk from low levels of exogenous
2 exposure. However, there is clear evidence of carcinogenic hazard from the rodent bioassays
3 and strong evidence from human studies (see Section 3.5), and the genotoxicity/mutagenicity of
4 EtO (see Section 3.4) supports low-dose linear extrapolation of risk estimates from those studies
5 (U.S. EPA, 2005a). In fact, as noted in Section 3.3.3.1, Marsden et al. (2009), using sensitive
6 detection techniques and an approach designed to separately quantify both endogenous N7-HEG
7 adducts and "exogenous" N7-HEG adducts induced by EtO treatment in rats, reported increases
8 in exogenous adducts in DNA of spleen and liver consistent with a linear dose-response
9 relationship (p < 0.05), down to the lowest dose administered (0.0001 mg/kg injected i.p. daily
10 for 3 days, which is a very low dose compared to the LOAELs in the carcinogenicity bioassays;
11 see Section C.7 of Appendix C). Furthermore, while the contributions to DNA damage from low
12 exogenous EtO exposures may be relatively small compared to those from endogenous EtO
13 exposure, low levels of exogenous EtO may nonetheless be responsible for levels of risk (above
14 background risk). This is not inconsistent with the much higher levels of background cancer
15 risk, to which endogenous EtO may contribute, for the two cancer types observed in the human
16 studies—lymphoid cancers have a background lifetime incidence risk on the order of 3%, while
17 the background lifetime incidence risk for breast cancer is on the order of 15%.29
18 See Table 4-24 for a summary of key unit risk estimates derived in this assessment. See
19 Section 4.7 for risk estimates based on occupational exposure scenarios.
20
21 4.6. COMPARISON WITH OTHER PUBLISHED RISK ESTIMATES
22 The unit risk values derived in this document are compared with other recent risk
23 estimates presented in the published literature (see Table 4-25).
24
25 4.6.1. Unit Risk Estimates Based on Human Studies
26 Kirman et al. (2004) used leukemia data only and pooled data from both the Stayner et al.
27 (1993) and the UCC studies (Teta et al., 1999; Teta et al., 1993). Based on the assumption that
28 leukemias are due to chromosome translocations, requiring two independent events
29 (chromosome breaks), the Kirman et al. (2004) proposed that two independent EtO-induced
30 events are required for EtO-induced leukemias and used a dose-squared model, yielding a unit
31 risk value of 4.5 x io~8 (ug/m3)"1 as their preferred estimate.
32
29These background lifetime incidence values were obtained from the lifetable analysis, based on SEER rates, as
discussed in Sections 4.1.1.3 and 4.1.2.3. For lymphoid cancer, for example, see the value ofRo at the bottom of
the lifetable analysis in Appendix E.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-75 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
Table 4-24. Summary of key unit risk estimates from this assessment (see
Section 4.7 for risk estimates based on occupational exposure scenarios)
Basis
Inhalation unit risk estimate"
(per jig/m3)b
Full lifetime unit risk estimate0
Total cancer risk based on human data (NIOSH cohort of
sterilizer workers) — lymphoid cancer incidence (linear
regression of categorical results) and breast cancer incidence in
females (2 -piece linear spline model)
1.80 x 10 3
Adult-based unit risk estimates'1
Total cancer risk based on human data (NIOSH
cohort) — lymphoid cancer incidence (linear regression of
categorical results) and breast cancer incidence in females (2-
piece linear spline model)
Lymphoid cancer incidence based on human data
(NIOSH) — linear regression of categorical results
Breast cancer incidence in females based on human data
(NIOSH) — stimate based on best-fitting model: the 2-piece
linear spline model
Breast cancer incidence in females based on human data
(NIOSH) — range based on 3 reasonable statistically significant
continuous models: 2-piece linear spline model, 2-piece log-
linear spline model, and linear model
Total cancer risk based on human data (NIOSH
cohort) — lymphoid cancer incidence (linear regression of
categorical results) and range of female breast cancer incidence
estimates (2-piece linear spline model, 2-piece log-linear spline
model, and linear model)
Lymphoid cancer mortality based on human data
(NIOSH) — linear regression of categorical results
Breast cancer mortality in females based on human data
(NIOSH) — linear regression of categorical results
Preferred total cancer incidence risk estimate from rodent data
(female mouse)
Range of total cancer incidence risk estimates from rodent data
(mouse and rat)
1.08 x 10 3
4.35 x 1(T4
8.21 x 1(T4
2.10 x 10'4 to 8.21
5.64 x 10'4to 1.08
x 10-4
x 10"3
2.01 x 10'4
2.39 x 10'4
4.6 x 10'5
2.2 x 10'5 to 4.6 x
io-5
0.01/ECoiestimatese
Lymphoid cancer incidence based on human data
(NIOSH) — linear regression of categorical results
Breast cancer incidence in females based on human data
(NIOSH) — estimate based on best-fitting model: the 2-piece
linear spline model
Lymphoid cancer mortality based on human data
(NIOSH) — linear regression of categorical results
1.9 x 10'4
4.2 x 10'4
9.0 x 10'5
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-76 DRAFT—DO NOT CITE OR QUOTE
-------
Table 4 24. Summary of key unit risk estimates from this assessment (see
Section 4.7 for risk estimates based on occupational exposure scenarios)
(continued)
Basis
Breast cancer mortality in females based on human data
(NIOSH) — linear regression of categorical results
Total cancer incidence based on human data (NIOSH)
Inhalation unit risk estimate"
(per jig/m3)b
1.2 x 1(T4
6.2 x 1(T4
3 "Technically, the values listed in this table are not all unit risk estimates as defined by EPA, but they are all potency
4 estimates that, when multiplied by an exposure value, give an estimate of extra cancer risk. These potency estimates
5 are not intended for use with continuous lifetime exposure levels above 140 ug/m3. See Section 4.7 for risk
6 estimates based on occupational exposure scenarios. Preferred estimates are in bold.
7 bTo convert unit risk estimates to (ppm)"1, multiply the (ug/m3)"1 estimates by 1,830 (ug/m3)/ppm.
8 ° Because the weight of evidence supports a mutagenic mode of action for EtO carcinogenicity, and because of the
9 lack of chemical-specific data, EPA assumes increased early-life susceptibility and recommends the application of
10 ADAFs, in accordance with EPA's Supplemental Guidance (U.S. EPA, 2005b), for exposure scenarios that include
11 early-life exposures. For the full lifetime (upper bound) unit risk estimate presented here, ADAFs have been
12 applied, as described in Section 4.4.
13 dThese (upper bound) unit risk estimates are intended for use in AD AF calculations and less-than-lifetime adult
14 exposure scenarios (U.S. EPA, 2005b). Note that these are not the same as the unit risk estimates derived directly
15 from the human data in Section 4.1 under the assumption that RRs are independent of age. Under that assumption,
16 the key unit risk estimates were 4.8 x 10"4 per ug/m3 for lymphoid cancer incidence, 9.5 x 10"4 per ug/m3 for breast
17 cancer incidence from the best-fitting 2-piece linear spline model, and 1.2 x 10"3 per ug/m3 for the combined cancer
18 incidence risk from those two cancers. See Section 4.4 for the derivation of the adult-based unit risk estimates.
19 eThese are not upper-bound risk estimates but, rather, estimates based on linear extrapolation from the ECM.
20 ADAFs should be applied if early-life exposure occurs, in accordance with EPA's Supplemental Guidance (U.S.
21 EPA, 2005b).
22
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-77 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
Table 4-25. Comparison of unit risk estimates"
Assessments
Data source
Inhalation unit risk estimate1"
(per jig/m3)
Based on human data
EPA
(this document)0
Kirman et al. (2004)
Valdez-Flores et al. (2010)
Lymphoid cancer incidence in sterilizer
workers (NIOSH)d
Breast cancer incidence in female
sterilizer workers (NIOSH)e
Total cancer risk based on the NIOSH
data
Leukemia mortality in combined NIOSH
and UCC cohorts (earlier follow-ups)
multiple individual cancer endpoints,
including all lymphohematopoietic,
lymphoid, and breast cancers, in
combined updated NIOSH and updated
UCC cohorts
7.2 x 1Q-4
1.4 x 1Q-3
1.8 x KT3
4.5 x 1Q-8
Range of 1.4 x 10-8tol.4x KT7 f
5.5 x KT7 to 1.6 x lQ-6g
Based on rodent data
EPA (this document)'
Kirman et al. (2004)
Female mouse tumors
Mononuclear cell leukemia in
rats and lymphomas in mice
7.6 x 1Q-5
2.6 x KT8 to 1.5 x lQ-5h
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
aUpper-bound estimates except where footnoted to indicate that estimates are based on EC values (i.e., estimates
with footnotes f and g).
bBecause the weight of evidence supports a mutagenic mode of action for EtO carcinogenicity, and in the absence of
chemical-specific data, EPA assumes increased early-life susceptibility, in accordance with EPA's Supplemental
Guidance (U.S. EPA, 2005b), and for the EPA lifetime unit risk estimates presented in this table, ADAFs have been
applied, as described in Section 4.4. The corresponding adult-based unit risk estimates are 4.4 x 10"4 (ug/m3)"1 for
human-based lymphoid cancer incidence, 8.2 x 10"4 (ug/m3)-1 for human-based breast cancer incidence,
1.1 x 10"3 (ug/m3)-1 for human-based total cancer incidence, and 4.6 x 10"5 (ug/m3)-1 for rodent-based total cancer
incidence. The non-EPA estimates in the table are shown as reported and do not account for potential increased
early-life susceptibility for lifetime exposures that include childhood, with the exception of the Valdez-Flores et al.
(2010) estimates, which are purported to include the ADAFs, but the ADAFs were in fact misapplied and have
essentially no impact (see Appendix A.2.20).
°See Table 4-24 in Section 4.5 for a more complete summary of estimates from this assessment. See Section 4.7 for
risk estimates for occupational exposure scenarios.
dFor lymphoid cancer mortality, the ADAF-adjusted lifetime unit risk estimate is 3.3 x 10"4 (ug/m3)-1 and the adult-
based unit risk estimate is 2.0 x 10"4 (ug/m3)-1.
Tor breast cancer mortality, the ADAF-adjusted lifetime unit risk estimate is 4.0 x 10"4 (ug/m3)-1 and the adult-
based unit risk estimate is 2.4 x 10"4 (ug/m3)-1.
Estimates based on linear extrapolation from EC0001-EC000001 obtained from the quadratic model.
Estimates based on range of EC(l/million)s of 0.001-0.003 ppm obtained from the model RR = e(P * exposure) for
relevant cancer endpoints.
hEstimates based on quadratic extrapolation model below the observable range of the data (i.e., below the LECio or
LECoi obtained using multistage model) with various points of departure (LEC0i-LEC0ooooi) for final linear
extrapolation (see Section 4.4.2).
7/2013
This document is a draft for review purposes only and does not constitute Agency policy.
4-78 DRAFT—DO NOT CITE OR QUOTE
-------
1 The Kirman et al. (2004) values are different from those in the current document because
2 of the different assumptions inherent in the Kirman et al. (2004) approach and because the study
3 used unpublished data from earlier follow-ups of the two cohorts. A key difference is that EPA
4 uses a linear model rather than a quadratic (dose-squared) model in the range of observation.
5 Then, EPA uses a higher extra risk level (1%) for establishing the POD, whereas Kirman et al.
6 (2004) used a risk level of 10"5 for their best estimate and a risk range of 10"4 to 10"6 for their
7 range of values. The extra risk level and the corresponding POD are not critical with the linear
8 model; however, with the quadratic model used by Kirman et al. (2004), the lower the risk level
9 (and hence the POD), the greater the impact of the quadratic model and the lower the resulting
10 unit risk estimates.
11 In addition, EPA (1) uses data for lymphoid cancers (and female breast cancers) rather
12 than leukemias, (2) includes ages up to 85 years in the life-table analysis rather than stopping at
13 70 years, (3) calculates unit risk estimates for cancer incidence as well as mortality, (4) uses a
14 lower bound as the POD rather than the maximum likelihood estimate, (5) uses the results of
15 lagged analyses rather than unlagged analyses, and (6) uses adult-based unit risk estimates in
16 cojunction with ADAFs (see Section 4.4) to derive the lifetime unit risk estimates.
17 Another key difference is that Kirman et al. (2004) relied on earlier NIOSH results
18 (Stayner et al., 1993), whereas EPA uses the results of NIOSH's more recent follow-up of the
19 cohort (Steenland et al., 2004). Kirman et al. (2004) claim that a quadratic dose-response model
20 provided the best fit to the data in the observable range and that this provides support for their
21 assumed mode of action. However, the 2004 NIOSH data for lymphohematopoietic cancers
22 suggest a supralinear exposure-response relationship (see Section 4.1.1.2 and Figures 4-1 and
23 4-2), which is inconsistent with a dose-squared model. Furthermore, EPA's review of the mode
24 of action evidence does not support the mode of action assumed by Kirman et al. (2004) (see
25 Section 3.4).
26 The Valdez-Flores et al. (2010) unit risk estimates (see Table 4-25) are similarly much
27 lower than those in the current document because of the different assumptions used. A key
28 difference is that EPA uses a linear model or a two-piece linear spline model in the range of
29 observation rather than an exponential model (RR = ep x exposure); which was used by Valdez-
30 Flores et al. (2010) despite its lack of fit. Then, EPA uses a higher extra risk level (1%) for
31 establishing the POD for linear extrapolation, whereas Valdez-Flores et al. (2010) used a risk
32 level of 10"6. In addition, EPA (1) includes ages up to 85 years in the life-table analysis rather
33 than stopping at 70 years, (2) calculates unit risk estimates for cancer incidence as well as
34 mortality, (3) uses a lower bound as the POD rather than the maximum likelihood estimate, and
35 (4) uses the results of lagged analyses rather than unlagged analyses. See Appendix A.2.20 for a
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-79 DRAFT—DO NOT CITE OR QUOTE
-------
1 more detailed discussion of the differences between the EPA and Valdez-Flores et al. (2010)
2 analyses.
O
4 4.6.2. Unit Risk Estimates Based on Laboratory Animal Studies
5 Kirman et al. (2004) also used linear and dose-squared extrapolation models to derive
6 unit risk estimates based on the rat mononuclear cell leukemia data and the mouse lymphoma
7 data. First, they used the multistage model to calculate the LECio (LECoi for the male mouse
8 lymphoma data) for the POD from the observable range. Then, using these PODs for linear
9 extrapolation, Kirman et al. (2004) obtained a unit risk range of 3.9 x 10~6 (ug/m3)"1 to
10 1.5 x 10~5 (ug/m3)"1. Alternatively, Kirman et al. (2004) used a quadratic extrapolation model
11 below the observable range to estimate secondary points of departure (LECoi-LECoooooi;
12 LECooi-LECoooooi for the male mouse) for final linear low-dose extrapolation, yielding unit risks
13 ranging from 2.6 x 10 8 (ug/m3)"1 to 4.9 x 10~6 (ug/m3)"1. These values are all smaller than the
14 unit risks derived from the rodent data in this document.
15
16 4.7. RISK ESTIMATES FOR SOME OCCUPATIONAL EXPOSURE SCENARIOS
17 The unit risk estimates derived in the preceding sections were developed for
18 environmental exposure levels, where maximum modeled levels are on the order of 1-2 ug/m3
19 (email dated October 3, 2005, from Mark Morris, EPA, to Jennifer Jinot, EPA), i.e., roughly
20 0.5-1 ppb, and are not applicable to higher exposures, including some occupational exposure
21 levels. However, occupational exposure levels of EtO are of concern to EPA when EtO is used
22 as a pesticide (e.g., sterilizing agent or fumigant). The occupational exposure scenarios of
23 interest to EPA include some cumulative exposures corresponding to exposure levels in the
24 nonlinear range of some of the models (i.e., above the maximum exposure level at which the
25 low-dose-linear unit risk estimates apply). Therefore, extra risk estimates were calculated for a
26 number of occupational exposure scenarios of possible concern. Extra risk estimates are
27 estimates of the extra cancer risk above background and are the same type of estimate that one
28 gets from multiplying a unit risk estimate by an exposure level. In this case, the exposure level is
29 used directly in the exposure-response model, thus accounting for any nonlinearities in the model
30 above the range of exposure levels for which the linear unit risk estimate is applicable. For these
31 occupational exposure scenarios, exposure-response models based on data from the NIOSH
32 cohort were used in conjunction with the life-table program, as previously discussed in
33 Section 4.1. A 35-year exposure occurring between ages 20 and 55 years was assumed, and
34 exposure levels ranging from 0.1 to 1 ppm 8-hour TWA were examined (i.e., ranging from about
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-80 DRAFT—DO NOT CITE OR QUOTE
-------
1 1,300 to 13,000 ppm x days). (Note that the current Occupational Safety and Health
2 Administration Permissible Exposure Limit is 1 ppm [8-hour TWA].)
3 For lymphoid cancer mortality in both sexes, the best-fitting (natural) log cumulative
4 exposure Cox regression model (see Steenland reanalyses in Appendix D; see also
5 Section 4.1.1.2), lagged 15 years, was used. The log cumulative exposure Cox regression model
6 was the best-fitting model for lymphoid cancer in males in the Steenland et al. (2004) study, and
7 the same model form is used here but with the data from both sexes. Although this model was
8 deemed too steep in the low-exposure region to be useful for the derivation of unit risk estimates
9 for lower (environmental) exposures, the model is well suited for the occupational exposure
10 scenarios of interest in this assessment because the corresponding cumulative exposures are well
11 within the range of the cumulative exposures in the NIOSH cohort. The model was statistically
12 significant (p = 0.02) and provided a better fit, based on AIC, than the two-piece spline models
13 that were considered as alternative models in Section 4.1.1.2 (the AICs were 460.426, 461.847,
14 and 461.48 for the log cumulative exposure Cox regression, log-linear two-piece spline and
15 linear two-piece spline models, respectively [as reported in Section D.3 of Appendix D]; a lower
16 AIC indicates a better fit), as well as a more plausible, smoothly curved exposure-response
17 relationship than the two-piece spline models, both of which, with knots at 100 ppm x days, had
18 a very steep rise and then a very sharp change in slope at the knot. In addition, the log
19 cumulative exposure Cox regression model had a slightly lower AIC (460.426 versus 460.54)
20 than the log cumulative exposure linear model (see Section D.3.c of Appendix D) and has the
21 advantage of being a standard epidemiological model for continuous exposure data (the Cox
22 regression model, albeit with log cumulative exposure to accommodate the supralinearity of the
23 exposure-response data). The log cumulative exposure linear model yields slightly higher RR
24 estimates than the log cumulative exposure Cox regression model, as can be seen by comparing
25 the log cumulative exposure models in Figures D-3b and D-3c in Appendix D, and would thus
26 result in slightly higher extra risk estimates than the log cumulative exposure Cox regression
27 model. For example, the MLEs of extra risk from the log cumulative exposure linear model
28 would range from about 23% higher for the 0.1 ppm 8-hour TWA to 6% higher for the 1 ppm
29 8-hour TWA.
30 The extra risk results for lymphoid cancer mortality and incidence in both sexes for the
31 log cumulative exposure Cox regression model are presented in Table 4-26. For lymphoid
32 cancer incidence, the exposure-response relationship was assumed to be the same as for mortality
33 (see Section 4.1.1.3). As can be seen in Table 4-26, the extra risks for these occupational
34 exposure levels are in the "plateau" region of the exposure-response relationships and increase
35 less than proportionately with exposure. For occupational exposures less than about 1,000 ppm
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-81 DRAFT—DO NOT CITE OR QUOTE
-------
1 x days, or about 0.08 ppm 8-hour TWA for 35 years, risk estimates are no longer in the plateau
2 region (see Figure 4-1) but rather in the steep low-exposure region, which is a region of greater
3 uncertainty for the log cumulative exposure model, and one might want to use the linear
4 regression of the categorical results that was used for lower exposures (see Section 4.1.1.2;
5 Appendix D). Furthermore, if one is using the linear regression model in this range and also
6 estimating risks for exposure levels in the range between about 0.08 and 0.6 ppm (near where the
7 linear regression and log cumulative exposure Cox regression models meet) 8-hour TWA, then
8 one might want to use the linear regression model for the entire range up to 0.6 ppm 8-hour
9 TWA to avoid a discontinuity between the two models; thus, results for the linear regression
10 model for exposure levels up to 0.6 ppm 8-hour TWA are also presented in Table 4-26. While
11 the best-fitting continuous exposure model, the log cumulative exposure Cox regression model,
12 would generally be preferred in the exposure range between 0.08 and 0.6 ppm 8-hour TWA,
13 there is model uncertainty, so the use of either model could be justified. For exposures higher
14 than where the linear regression and log cumulative exposure Cox regression models meet, the
15 log cumulative exposure model exclusively is recommended. The models used to derive the
16 extra risk estimates presented in Table 4-26 for lymphoid cancer for the occupational exposure
17 scenarios are displayed in Figure 4-7 over the range of occupational cumulative exposures of
18 interest; the categorical results are included for comparison.
19 For breast cancer, incidence data were available from the NIOSH incidence study; thus,
20 only incidence estimates were calculated. In addition to being the preferred type of cancer risk
21 estimate, the breast cancer incidence risk estimates are based on more cases than were available
22 in the mortality study and the incidence data (for the subcohort with interviews) are adjusted for
23 a number of breast cancer risk factors (see Section 4.1.2.3). In terms of the incidence data, the
24 subcohort data are preferred to the full cohort data because the subcohort data are adjusted for
25 these potential confounders and also because the full cohort data have incomplete ascertainment
26 of breast cancer cases.
27 For breast cancer incidence in the subcohort with interviews, a number of Cox regression
28 exposure-response models from the Steenland et al. (2003) breast cancer incidence study fit
29 almost equally well (see Section 4.1.2.3). These include a log cumulative exposure model and a
30 cumulative exposure model, both with a 15-year lag, and a log cumulative exposure model with
31 no lag. The latter model was omitted from the calculations because the inclusion of a 15-year lag
32 for the development of breast cancer was considered more biologically realistic than not
33 including a lag. Steenland et al. (2003) also provide a duration-of-exposure Cox regression
34 model with a marginally better fit; however, models using duration of exposure are less useful
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-82 DRAFT—DO NOT CITE OR QUOTE
-------
to
OJ
Table 4-26. Extra risk estimates for lymphoid cancer in both sexes for various occupational exposure levels"
8-hr TWA
(ppm)
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Lymphoid cancer mortality
Log cumulative exposure Cox
regression model0
MLE
0.014
0.016
0.017
0.018
0.018
0.019
0.019
0.020
0.020
0.021
95% UCL
0.032
0.038
0.042
0.045
0.047
0.049
0.051
0.052
0.054
0.055
Linear regression modeld
MLE
0.003
0.007
0.010
0.013
0.016
0.019
-
-
-
-
95% UCL
0.007
0.014
0.022
0.029
0.036
0.042
0.049
-
-
-
Lymphoid cancer incidence1"
Log cumulative exposure Cox
regression model0
MLE
0.031
0.035
0.038
0.040
0.042
0.043
0.044
0.045
0.046
0.047
95% UCL
0.071
0.084
0.093
0.099
0.10
0.11
0.11
0.12
0.12
0.12
Linear regression modeld
MLE
0.007
0.014
0.021
0.028
0.035
0.042
-
-
-
-
95% UCL
0.016
0.031
0.047
0.062
0.076
0.090
-
-
-
-
§•
§
I
^
'TS
o
a
I
a,
8"
O 5
31
o *
H 1
W K-
""Assuming a 35-yr exposure between ages 20 and 55 years (see Section 4.7).
bAssumes same exposure-response relationship as for lymphoid cancer mortality.
°From the best-fitting log cumulative exposure Cox regression model for lymphoid cancer mortality in both sexes; 15-yr lag (see Appendix D; see also
Section 4.1.1.2).
dLinear regression of categorical results for both sexes (see Appendix D; 15-yr lag), excluding the highest exposure group (see Section 4.1.1.2); extra risk
estimates from the linear model are provided only up to the exposure level where the linear model meets the log cumulative Cox regression model.
MLE: maximum likelihood estimate; UCL: (one-sided) upper confidence limit estimate.
o
c
o
H
W
-------
to
OJ
o
rs
S
1
I
o
a
i
O 5
O I
^ ^
31
o ^
H t
W ??•
o •
V
o
c
o
H
W
t
O.I ppm x35 years
6OOO 8OOO
Cumulative Exposure (ppm x days)
• log exposure Cox
regression model
linear regression
of categorical
results
lymphoid cancer
quartiles
t
ppm x 35 years
Figure 4-7 RR estimates for lymphoid cancer from occupational EtO exposures (with 15-year lag).
Lymphoid cancer models (see Section 4.1.1.2): log cumulative exposure Cox regression model; categorical results from Cox
regression model; linear regression of categorical results, excluding highest exposure group.
-------
1 for estimating exposure-related risks, and duration of exposure and cumulative exposure are
2 correlated. Thus, only the lagged cumulative exposure models are considered here.
3 The extra risk estimates for breast cancer incidence in females from the lagged
4 cumulative exposure and log cumulative exposure Cox regression models listed above are
5 presented in Table 4-27. As can be seen in Table 4-27, the extra risk estimates for the lagged log
6 cumulative exposure and cumulative exposure models differ substantially. Furthermore, the
7 categorical Cox regression results for breast cancer incidence in the subcohort with interviews
8 suggest that, for the lowest four exposure quintiles, the log cumulative exposure model
9 overestimates the RR, while the cumulative exposure model generally underestimates the RR,
10 with the categorical results largely falling between the RR estimates of those two models (see
11 Figure 4-5). (The lowest four exposure quintiles represent individual worker exposures ranging
12 from 0 to about 15,000 ppm x days, which covers the range of cumulative exposures for the
13 occupational exposure scenarios of interest in this assessment, the maximum of which is
14 12,775 ppm x days.) Therefore, the two-piece linear spline model (with a 15-year lag) (see
15 Section 4.1.2.3) was also used to calculate the extra risk estimates. The two-piece linear spline
16 model provides a better fit to the data than the log cumulative exposure or cumulative exposure
17 Cox regression models, as indicated by a lower AIC value (1,950.9 for two-piece linear spline
18 model vs. 1,956.2 for the log cumulative exposure Cox regression model and 1,956.8 for the
19 cumulative exposure Cox regression model; Table 4-12 and Appendix D). In fact, the two-piece
20 linear spline model provided the best fit to the breast cancer incidence data of all the models
21 investigated in Section 4.1.2.3, and it provides the best representation of the categorical RR
22 results, particularly for the range of cumulative exposures for the occupational exposure
23 scenarios of interest (see Figures 4-5, 4-6, and 4-8). The extra risk estimates calculated using the
24 two-piece linear spline model are also presented in Table 4-27 and are the preferred estimates
25 because they are derived from the best-fitting model.
26 In addition, extra risk estimates for breast cancer incidence in females from the
27 continuous linear model (with a 15-year lag) (see Section 4.1.2.3) are presented in Table 4-27.
28 This model, with an AIC of 1,952.3 (see Table 4-12), was the second-best-fitting model and also
29 provided a good visual fit to the categorical data (see Figure 4-6). Moreover, the two best-fitting
30 models (i.e., the continuous linear model and the two-piece linear spline model) span the range
31 of RR estimates from the three best-fitting models investigated in Section 4.1.2.3 (the third
32 being the two-piece log-linear spline model) over the range of cumulative exposures for the
33 occupational exposure scenarios of interest in this assessment (see Figure 4-6). Comparing the
34 results of the two best-fitting models shows that the extra risk estimates differ by just under
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-85 DRAFT—DO NOT CITE OR QUOTE
-------
to
OJ
Table 4-27. Extra risk estimates for breast cancer incidence in females for various occupational exposure
levelsa'b
8-hr TWA
(ppm)
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Log cumulative exposure Cox
regression model0
MLE
0.055
0.061
0.065
0.068
0.070
0.072
0.073
0.074
0.076
0.077
95% UCL
0.11
0.12
0.13
0.14
0.14
0.14
0.15
0.15
0.15
0.16
Cumulative exposure Cox
regression model0
MLE
0.0013
0.0026
0.0040
0.0053
0.0067
0.0081
0.0095
0.011
0.012
0.014
95% UCL
0.0023
0.0046
0.0069
0.0092
0.012
0.014
0.017
0.019
0.022
0.024
Continuous linear modeld
MLE
0.0042
0.0084
0.012
0.017
0.021
0.025
0.029
0.033
0.037
0.041
95% UCLf
0.0081
0.016
0.024
0.032
0.040
0.048
0.055
0.063
0.070
0.078
Two-piece linear spline model"
MLE
0.016
0.032
0.048
0.063
0.075
0.081
0.086
0.089
0.093
0.095
95% UCLg
0.031
0.061
0.090
0.118
0.139
0.150
0.157
0.162
0.167
0.171
§
I
§
I
a,
O 3
31
o *
H 1
W £•
o
c
o
H
W
""Assuming a 35-yr exposure between ages 20 and 55 years.
bFrom incidence data for subcohort with interviews; invasive and in situ tumors (Steenland et al, 2003).
°Cox regression models from Steenland et al. (2003, Table 5), with 15-yr lag.
dLinear model with cumulative exposure as a continuous variable (see Section 4.1.2.3 and Section D.2 of Appendix D).
eTwo-piece linear spline model results for occupational exposures use both spline segments (see Section D.2 of Appendix D), knot at 5,800 ppm x days; with
15-yr lag. For the 95% UCL, for exposures below the knot, RR = 1 + ((31 + 1.645 x SE1) x exposure; for exposures above the knot, RR = 1 + ((31 x exp + (32 x
(exp-knot) + 1.645 x sqrt(exp2 x varl + (exp-knot)2 x var2 + 2 x exp x (exp-knot) x covar)), where exp = cumulative exposure, var = variance, covar =
covariance (see Section D.2 of Appendix D for the parameter values).
Confidence intervals used in deriving the 95% UCLs were estimated employing the Wald approach. Confidence intervals for linear RR models, however, in
contrast to those for the log-linear RR models, may not be symmetrical. EPA also evaluated application of a profile likelihood approach for the linear RR models
(Langholz and Richardson, 2010), which allows for asymmetric CIs, for comparison with the Wald approach. Using the profile likelihood method, the resulting
extra risk estimates for breast cancer incidence for the continuous linear model would have been about 29% higher than those obtained using the Wald approach.
-------
Table 4-27. Extra risk estimates for breast cancer incidence in females for various occupational exposure
levels"'"5 (continued)
Confidence intervals used in deriving the 95% UCLs were estimated employing the Wald approach. Confidence intervals for linear RR models, however, in
^ contrast to those for the log-linear RR models, may not be symmetrical. EPA also evaluated application of a profile likelihood approach for the linear RR models
S^ (Langholz and Richardson, 2010), which allows for asymmetric CIs, for comparison with the Wald approach. Using the profile likelihood method, the resulting
§- extra risk estimates for breast cancer incidence for the low-exposure linear spline segment (i.e., below 0.4 ppm 8-hr TWA) would have been about 34% higher
s than those obtained using the Wald approach. Calculating the profile likelihood CIs in the region of the second spline segment is computationally difficult and
1 was not pursued here.
I
§
I
a,
O 3
31
o ^
H t
W K-
o
c
o
H
W
MLE: maximum likelihood estimates; UCL: (one-sided) upper confidence limit estimate.
-------
1 fourfold at the lowest exposure level (0.1 ppm 8-hour TWA) and the difference tapers to just
2 over twofold at the highest exposure level (1 ppm 8-hour TWA), with the estimates of the
3 best-fitting model, the two-piece linear spline model, yielding the higher extra risk estimates
4 across the range.
5 Finally, for comparison, maximum likelihood estimates (MLEs) of extra risk for the log
6 cumulative exposure Cox regression model with a 10-year lag were calculated. The model with
7 a 10-year lag also provided a statistically significant fit to the data (p = 0.03), whereas, the
8 models with 5- and 15-year lags did not. These estimates ranged from 0.067 for 0.1 ppm
9 exposure to 0.094 for 1.0 ppm exposure. Thus, the MLEs of extra risk with a 10-year lag were
10 about 20% higher than those with the 15-year lag.
11 The continuous models (with a 15-year lag) considered for deriving the extra risk
12 estimates for breast cancer incidence in females for the occupational exposure scenarios are
13 displayed in Figure 4-8 over the range of occupational cumulative exposures of interest.
14 Categorical results are also presented for comparison (deciles from the categorical linear model
15 are presented because it had a better fit than the log-linear categorical model, as indicated by the
16 AICs, which were 1,963.9 and 1,966.9, respectively; Appendix D). The recommended model is
17 the two-piece linear spline model; this was the best-fitting continuous model of those evaluated
18 in this assessment, and it provides the best visual fit in comparison to the categorical results in
19 the range of the occupational exposure scenarios of interest. As shown in Figure 4-8, the log
20 cumulative exposure Cox regression model is too flat across the range of exposures of interest,
21 apparently overestimating the risks at lower exposures and underestimating those at higher
22 exposures. It also appears from Figure 4-8 that the cumulative exposure Cox regression model
23 and the linear model both underestimate risks across the range of exposures of interest. This is
24 consistent with the analysis presented in Section D.I of Appendix D showing the strong
25 influence of the upper tail of cumulative exposures on the results of the cumulative exposure Cox
26 regression model. The responses in the upper tail of exposures are relatively dampened, such
27 that when the highest 5% of exposures (exposures > 27,500 ppm x days, which are well in
28 excess of the exposures of corresponding to the occupational exposure scenarios considered
29 here) are excluded, the slope of the Cox regression model is substantially increased (e.g., at
30 10,000 ppm x days, the RR estimate increases from about 1.1 to almost 1.5; see Figure D-ld in
31 Appendix D). This strong influence of the upper tail of exposures would similarly attenuate the
32 slope of the (continuous) linear model, resulting in underestimation of the lower-exposure risks.
33 The two-piece linear spline model, on the other hand, is more flexible, and the influence of the
34 upper tail of exposures would be primarily on the upper spline segment; thus, the two-piece
35 model is able to provide a better fit to the lower-exposure data.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-88 DRAFT—DO NOT CITE OR QUOTE
-------
o
D
>J
5'
rs
o-
o
s
1
<--k
>!'
&
I
^
'TS
i. S
0 'o
o
a
^T
|
3 1
3 1
3 »•
?!
3 <$
H a
"> i2
_^ ^3
_j O
n ^
\^-
^H
n
2
1.8
1.6
.2£
he
-------
1 For the total cancer risk combined across the two cancer types, the MLE can be obtained
2 directly by summing the MLEs for the individual cancer types. An upper bound can be
3 approximated by summing the 95% UCL estimates for the individual cancer types; however, this
4 will overestimate the corresponding 95% UCL on total cancer risk.
5 Although there is model uncertainty, as discussed above, there is less overall uncertainty
6 associated with the extra risk estimates for occupational exposure scenarios than with the unit
7 risk estimates for environmental exposures. The extra risk estimates are derived for occupational
8 exposure scenarios that yield cumulative exposures well within the range of the exposures in the
9 NIOSH study. Moreover, the NIOSH study is a study of sterilizer workers who used EtO for the
10 sterilization of medical supplies or spices (Steenland et al., 1991); thus, the results are directly
11 applicable to workers in these occupations, and these are among the occupations of primary
12 concern for current occupational EtO exposures.
13
14 Calculation of Extra Risk Estimates for Other Occupational Exposure Scenarios:
15
16 Some detailed guidance is provided here for calculating extra risk estimates outside of the
17 range of occupational scenarios considered above. Note that for 35-year exposures to exposure
18 levels between the exposure levels presented in Tables 4-26 and 4-27, e.g., 0.15 ppm, one could
19 interpolate between the extra risk estimates presented for the closest exposure levels on either
20 side.
21
22 For occupational exposures with durations other than 35 years:
23
24 Extra risk estimates for a 45-year exposure to the same exposure levels were nearly
25 identical to those from the 35-year exposure for both lymphoid cancer in both sexes and breast
26 cancer in females (results not shown). With the 15-year lag, the assumption of an additional
27 10 years of exposure only negligibly affects the risks above age 70 and has little impact on
28 lifetime risk. For exposure scenarios of 35-45 years but with 8-hour TWAs falling between
29 those presented in the tables, one can estimate the extra risk by interpolation. For exposure
30 scenarios with durations of exposure less than 30-35 years, one could roughly estimate extra
31 risks by calculating the cumulative exposure and finding the extra risks for a similar cumulative
32 exposure in Tables 4-26 and 4-27. For a more precise estimation, or for exposure scenarios of
33 much shorter duration or for specific age groups, one should do the calculations using a life-table
34 analysis, as presented in Appendix E but modified for the specific exposure scenarios.
35
36
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-90 DRAFT—DO NOT CITE OR QUOTE
-------
1 For occupational exposures below 0.1 ppm:
2
3 For lymphoid cancer, use of the log cumulative exposure Cox regression model is not
4 advised below 0.1 ppm (x 35 years). Instead, the low-exposure continuation of the linear
5 regression model presented in Table 4-26 of the assessment is recommended. For 35-year
6 exposures, the following formulae would apply:
7
8
9 95% UCL on extra risk for lymphoid cancer incidence ~ (8-h TWA occ exp [in ppm]) x
10 (0.016/0.1 ppm) = (8-h TWA occ exp [in ppm]) x (0.16/ppm)
11
12 MLE of extra risk for lymphoid cancer incidence = (8-h TWA occ exp [in ppm]) x
13 (0.007/0.1 ppm) = (8-h TWA occ exp [in ppm]) x (0. 07/ppm)
14
15
16 If one is considering occupational exposure scenarios using a range of 8-h TWA
17 exposure levels on both sides of 0.1 ppm, one might want to use the linear regression model for
18 all the exposure levels up to about 0.6 ppm 8-h TWA (approximately where the linear regression
19 model intersects the log cumulative exposure Cox regression model) to avoid the discontinuity
20 between the two models below where they intersect. Note that the extra risk estimates from the
21 different models differ by at most about 4.5-fold (at 0.1 ppm) and that there is model uncertainty
22 in this range, so the use of either model could be justified. Above where the models intersect,
23 only the log cumulative exposure Cox regression model should be used.
24
25 For breast cancer, the low-exposure continuation of the two-piece linear spline model
26 presented in Table 4-27 of the assessment is recommended. For 35-year exposures, the
27 following formulae would apply:
28
29
30 95% UCL on extra risk for breast cancer incidence ~ (8-h TWA occ exp [in ppm]) x
31 (0.031/0.1 ppm) = (8-h TWA occ exp [in ppm]) x (0. 31/ppm)
32
33 MLE of extra risk for breast cancer incidence = (8-h TWA occ exp [in ppm]) x
34 (0.016/0. Ippm) = (8-h TWA occ exp [in ppm]) x (0. 16/ppm)
35
36
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-91 DRAFT—DO NOT CITE OR QUOTE
-------
1 Above 0.1 ppm 8-h TWA, use the two-piece linear spline model results in Table 4-27 of the
2 assessment.
O
4 Alternatively, for exposures below 0.1 ppm, one could use the formulae presented below,
5 which are based on the unit risk estimates presented in Tables 4-22 (for 95% UCLs) and
6 Table 4-23 (for MLEs), with conversions for adjusting occupational to environmental exposures.
7 Note, however, that the extra risk results for 35 years of exposure based on these unit risk values
8 do not exactly match the values in Tables 4-26 and 4-27 for the linear models (the formulae
9 below yield extra risk estimates that are 15-20% lower than the values in Tables 4-26 and 4-27
10 for the low end of the exposure range [e.g., 0.1-0.4 ppm] where the comparison with the
11 unit-risk-based estimates is appropriate). This is because the results in Tables 4-26 and 4-27 are
12 based on life-table analyses, which take into account age-specific background rates of the
13 cancers and ages of exposure (assumed to be from 20 to 55 years of age in these occupational
14 exposure scenarios), whereas the formulae below are approximations that do not take
15 age-specific considerations into account. The advantage of the formulae based on the unit risk
16 values is that they can incorporate durations other than -35 years.
17
18
19 8-h TWA occ exp [in ppm] x (10 m3/day/20 m3/day) x (240 days/year/365 days/year) x
20 (3 5 years/70 years) = (continuous lifetime) env exp [in ppm]
21 (Note that for exposure durations other than 35 years, replace 35 years with the alternate
22 duration in the formula above.)
23
24 95% UCL on extra risk for lymphoid cancer incidence ~ 0.795/ppm x env exp [in ppm]
25
26 MLE of extra risk for lymphoid cancer incidence = 0.356/ppm x env exp [in ppm]
27
28 95% UCL on extra risk for breast cancer incidence (in females) = 1.50/ppm x env exp [in
29 ppm]
30
31 MLE of extra risk for breast cancer incidence (in females) = 0.776/ppm x env exp [in
32
33
34
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 4-92 DRAFT—DO NOT CITE OR QUOTE
-------
REFERENCES
This reference list includes all the references cited in the document except for
Appendix B, which is a reference list pertaining to Figure 3-3, and Appendix J, which includes
some more recent references. References added after the 2007 external peer review are also
listed separately in Appendix I. References identified in a May 2013 literature search but
appearing after the 30 June 2010 cutoff date for literature inclusion into this carcinogenicity
assessment are cited and discussed in Appendix J.
Adam. B: Bardos. H: Adany. R. (2005). Increased genotoxic susceptibility of breast epithelial cells to ethylene
oxide. MutatRes 585: 120-126. http://dx.doi.0rg/10.1016/j.mrgentox.2005.04.009
Arias. E. (2007). United States life tables, 2004. Atlanta, GA: Centers of Disease Control and Prevention; National
Center for Health Statistics, http://www.cdc.gov/nchs/data/nvsr/nvsr56/nvsr56 09.pdf
Bastlova. T: Andersson. B: Lambert. B: Kolman. A. (1993). Molecular analysis of ethylene oxide-induced mutations
at the HPRT locus in human diploid fibroblasts. Mutat Res 287: 283-292.
BEIR (Committee on the Biological Effects of Ionizing Radiation). (1988). Health risks of radon and other
internally deposited alpha-emitters. Washington, DC: National Academy Press.
Benson. LO: Teta. MJ. (1993). Mortality due to pancreatic and lymphopoietic cancers in chlorohydrin production
workers. Br J Ind Med 50: 710-716. http://dx.doi.0rg/10.1136/oem.50.8.710
Beranek. DT. (1990). Distribution of methyl and ethyl adducts following alkylation with monofunctional alkylating
agents [Review]. MutatRes 231: 11-30.
Bisanti L: Magginj M: Raschetti R: Alegianj SS: Ippolito. FM: Caffari. B: Segnan. N: Ponti. A. (1993). Cancer
mortality in ethylene oxide workers. Br J Ind Med 50: 317-324.
Boffetta. P: vanderHel. O: Norppa. H: Fabianova. E: Fucic. A: Gundy. S: Lazutka. J: Cebulska-Wasilewska. A:
Puskailerova. D: Znaor. A: Kelecsenyj Z: Kurtinaitis. J: Rachtan. J: Fornj A: Vermeulen. R: Bonassi. S. (2007).
Chromosomal aberrations and cancer risk: results of a cohort study from Central Europe. Am J Epidemiol 165: 36-
43. http://dx.doi.org/10.1093/aje/kwj367
Bolt. HM. (1996). Quantification of endogenous carcinogens. The ethylene oxide paradox [Review]. Biochem
Pharmacol 52: 1-5.
Bonassi. S: Znaor. A: Ceppi M: Lando. C: Chang. WP: Holland. N: Kirsch-Volders. M: Zeiger. E: Ban. S: Barale.
R: Bigatti MP: Bolognesi C: Cebulska-Wasilewska. A: Fabianova. E: Fucic. A: Hagmar. L: Joksic. G: Martelli A:
Migliore. L: Mirkova. E: Scarfi MR: Zijno. A: Norppa. H: Fenech. M. (2007). An increased micronucleus
frequency in peripheral blood lymphocytes predicts the risk of cancer in humans. Carcinogenesis 28: 625-631.
http://dx.doi.org/10.1093/carcin/bgll77
Boogaard. PJ. (2002). Use of haemoglobin adducts in exposure monitoring and risk assessment [Review]. J
ChromatogrB Analyt Technol Biomed Life Sci 778: 309-322. http://dx.doi.org/10.1016/S0378-4347(01)00445-5
Brown. CD: Bahman. A: Turner. MJ: Fennell. TR. (1998). Ethylene oxide dosimetry in the mouse. Toxicol Appl
Pharmacol 148: 215-222. http://dx.doi.org/10.1006/taap.1997.8349
Brown. CD: Wong. BA: Fennell. TR. (1996). In vivo and in vitro kinetics of ethylene oxide metabolism in rats and
mice. Toxicol Appl Pharmacol 136: 8-19. http://dx.doi.org/10.1006/taap.1996.0002
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 R-1 DRAFT—DO NOT CITE OR QUOTE
-------
Chaganti SR: Chen. W: Parsa. N: Offit. K: Louie. DC: Dalla-Favera. R: Chaganti RS. (1998). Involvement of
BCL6 in chromosomal aberrations affecting band 3q27 in B-cell non-Hodgkin lymphoma. Genes Chromosomes
Cancer 23: 323-327.
Christiansen. DH: Andersen. MK: Desta. F: Pedersen-Bjergaard. J. (2005). Mutations of genes in the receptor
tyrosine kinase (RTK)TRAS-BRAF signal transduction pathway in therapy-related myelodysplasia and acute
myeloid leukemia. Leukemia 19: 2232-2240. http://dx.doi.org/10.1038/sj.leu.2404009
Christiansea DH: Andersen. MK: Pedersen-Bjergaard. J. (2001). Mutations with loss of heterozygosity of p53 are
common in therapy-related myelodysplasia and acute myeloid leukemia after exposure to alkylating agents and
significantly associated with deletion or loss of 5q, a complex karyotype, and a poor prognosis. J Clin Oncol 19:
1405-1413.
Clare. MG: Dean. BJ: De Jong. G: Van Sittert NJ. (1985). Chromosome analysis of lymphocytes from workers at
an ethylene oxide plant. DNA Repair 156: 109-116.
Clewell HJ: Teeguarden. J: McDonald. T: Sarangapanj R: Lawrence. G: Covington. T: Gentry. R: Shipp. A.
(2002). Review and evaluation of the potential impact of age- and gender-specific pharmacokinetic differences on
tissue dosimetry [Review]. CritRev Toxicol 32: 329-389. http://dx.doi.org/10.1080/20024091064264
Coggon. D: Harris. EC: Poole. J: Palmer. KT. (2004). Mortality of workers exposed to ethylene oxide: extended
follow up of a British cohort. Occup Environ Med 61: 358-362.
Csanadv. GA: Denk. B: Putz. C: Kreuzer. PE: Kessler. W: Baur. C: Gargas. ML: Filser. JG. (2000). A physiological
toxicokinetic model for exogenous and endogenous ethylene and ethylene oxide in rat, mouse, and human:
formation of 2-hydroxyethyl adducts with hemoglobin and DNA. Toxicol Appl Pharmacol 165: 1-26.
http://dx.doi.org/10.1006/taap.2000.8918
Dellarco. VU Generoso. WM: Sega. GA: Fowle. JR. Ill: Jacobson-Kram. D. (1990). Review of the mutagenicity of
ethylene oxide [Review]. EnvironMol Mutagen 16: 85-103.
Dormer. EM: Wong. BA: James. RA: Preston. RJ. (2010). Reciprocal translocations in somatic and germ cells of
mice chronically exposed by inhalation to ethylene oxide: implications for risk assessment. Mutagenesis 25: 49-55.
http://dx.doi.org/10.1093/mutage/gep042
Dunkelberg. H. (1982). Carcinogenicity of ethylene oxide and 1,2-propylene oxide upon intragastric administration
to rats. Br J Cancer 46: 924-933.
Ehrenberg. L: Hussain. S. (1981). Genetic toxicity of some important epoxides [Review]. DNA Repair 86: 1-113.
EPIC (Ethylene Oxide Industry Council). (2001). Toxicological review of ethylene oxide in support of summary
information on the integrated risk information system. Arlington, VA.
Fennell. TR: Brown. CD. (2001). A physiologically based pharmacokinetic model for ethylene oxide in mouse, rat,
and human. Toxicol Appl Pharmacol 173: 161-175. http://dx.doi.org/10.1006/taap.2001.9184
Filser. JG: Denk. B: Tornqvist. M: Kessler. W: Ehrenberg. L. (1992). Pharmacokinetics of ethylene in man: body
burden with ethylene oxide and hydroxyethylation of hemoglobin due to endogenous and environmental ethylene.
Arch Toxicol 66: 157-163.
Post. U: Hallier. E: Ottenwalder. H: Bolt. HM: Peter. H. (1991). Distribution of ethylene oxide in human blood and
its implications for biomonitoring. HumExp Toxicol 10: 25-31.
Galloway. SM: Berry. PK: Nichols. WW: Wolman. SR: Soper. KA: Stolley. PD: Archer. P. (1986). Chromosome
aberrations in individuals occupationally exposed to ethylene oxide, and in a large control population. Mutat Res
Genet Toxicol 170: 55-74. http://dx.doi.org/10.1016/0165-1218(86)90082-0
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 R-2 DRAFT—DO NOT CITE OR QUOTE
-------
Gardner. MJ: Coggon. D: Pannett B: Harris. EC. (1989). Workers exposed to ethylene oxide: a follow up study.
Occup Environ Med 46: 860-665.
Garman. RH: Snellings. WM: Maronpot RR. (1985). Brain tumors in F344 rats associated with chronic inhalation
exposure to ethylene oxide. Neurotoxicology 6: 117-137.
Garman. RH: Snellings. WM: Maronpot RR. (1986). Frequency, size and location of brain tumours in F-344 rats
chronically exposed to ethylene oxide. Food Chem Toxicol 24: 145-153. http://dx.doi.org/10.1016/0278-
6915(86)90349-2
Garry. VF: Hozier. J: Jacobs. D: Wade. RU Gray. DG. (1979). Ethylene oxide: evidence of human chromosomal
effects. Environ Mutagen 1: 375-382. http://dx.doi.org/10.1002/em.2860010410
Gelehrter. TD: Collins. FS: Ginsburg. D. (1990). Principles of medical genetics. Baltimore, MD: Williams and
Wilkins.
Generoso. WM: Cain. KT: Cornett CV: Cacheiro. NLA: Hughes. LA. (1990). Concentration-response curves for
ethylene-oxide-induced heritable translocations and dominant lethal mutations. Environ Mol Mutagen 16: 126-131.
Godderis. L: Aka. P: Matecuca. R: Kirsch-Volders. M: Lisoa D: Veulemans. H. (2006). Dose-dependent influence
of genetic polymorphisms on DNA damage induced by styrene oxide, ethylene oxide and gamma-radiation.
Toxicology 219: 220-229. http://dx.doi.0rg/10.1016/i.tox.2005.ll.021
Golberg. L. (1986). Chemical and physical properties. In Hazard assessment of ethylene oxide. Boca Raton, FL:
CRC Press.
Greenberg. HL: Ott. MG: Shore. RE. (1990). Men assigned to ethylene oxide production or other ethylene oxide
related chemical manufacturing: A mortality study. Br J Ind Med 47: 221-230.
http://dx.doi.0rg/10.1136/oem.47.4.221
Greife. AU Hornung. RW: Stayner. LG: Steenland. KN. (1988). Development of a model for use in estimating
exposure to ethylene oxide in a retrospective cohort mortality study. Scand J Work Environ Health 1: 29-30.
Hagmar. L: Mikoczy. Z: Welinder. H. (1995). Cancer incidence in Swedish sterilant workers exposed to ethylene
oxide. Occup Environ Med 52: 154-156. http://dx.doi.0rg/10.1136/oem.52.3.154
Hagmar. L: Stromberg. U: Bonassi. S: Hansteen. IL: Knudsen. LE: Lindholm. C: Norppa. H. (2004). Impact of
types of lymphocyte chromosomal aberrations on human cancer risk: results from Nordic and Italian cohorts. Cancer
Res 64: 2258-2263. http://dx.doi.org/10.1158/0008-5472.CAN-03-3360
Hagmar. L: Welinder. H: Linden. K: Attewell R: Osterman-Golkar. S: Tornqvist M. (1991). An epidemiological
study of cancer risk among workers exposed to ethylene oxide using hemoglobin adducts to validate environmental
exposure assessments. Int Arch Occup Environ Health 63: 271-277.
Hansea JP: Allea J: Brock. K: Falconer. J: Helms. MJ: Shaver. GC Strohm. B. (1984). Normal sister chromatid
exchange levels in hospital sterilization employees exposed to ethylene oxide. J Occup Med 26: 29-32.
Harada. H: Harada. Y: Tanaka. H: Kimura. A: Inaba. T. (2003). Implications of somatic mutations in the AML1
gene in radiation-associated and therapy-related myelodysplastic syndrome/acute myeloid leukemia. Blood 101:
673-680. http://dx.doi.org/10.1182/blood-2002-04-1010
Harris. NU Jaffe. ES: Diebold. J: Flandrin. G: Muller-Hermelink. HK: Vardiman. J: Lister. TA: Bloomfield. CD.
(1999). World Health Organization classification of neoplastic diseases of the hematopoietic and lymphoid tissues:
report of the Clinical Advisory Committee meeting-Airlie House, Virginia, November 1997. J Clin Oncol 17: 3835-
3849.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 R-3 DRAFT—DO NOT CITE OR QUOTE
-------
Haufroid. V: Merz. B: Hofmann. A: Tschopp. A: Lison. D: Hotz. P. (2007). Exposure to ethylene oxide in hospitals:
biological monitoring and influence of glutathione S-transferase and epoxide hydrolase polymorphisms. Cancer
Epidemiol Biomarkers Prev 16: 796-802. http://dx.doi.org/10.1158/1055-9965.EPI-06-0915
Health Canada. (2001). Priority substances list assessment report: Ethylene oxide. (Cat. no. En40-215/51E). Ottawa,
Ontario, http://www.hc-sc.gc.ca/ewh-semt/pubs/contaminants/psl2-lsp2/ethylene_oxide/index-eng.php
Hill AB. (1965). The environment and disease: Association or causation? Proc R Soc Med 58: 295-300.
Hogstedt B: Gullberg. B: Hedner. K: Kolnig. AM: Mitelman. F: Skerfving. S: Widegren. B. (1983). Chromosome
aberrations and micronuclei in bone marrow cells and peripheral blood lymphocytes in humans exposed to ethylene
oxide. Hereditas 98: 105-113.
Hogstedt. C. (1988). Epidemiologic studies on ethylene oxide and cancer: An updating. In Methods for detecting
DNA damaging agents in humans: Applications in cancer epidemiology and prevention. Lyon, France: International
Agency for Research on Cancer.
Hogstedt C: Aringer. L: Gustavsson. A. (1986). Epidemiologic support for ethylene oxide as a cancer-causing
agent. JAMA 255: 1575-1578.
Hong. HH: Houle. CD: Toa TV: Sills. RC. (2007). K-ras mutations in lung tumors and tumors from other organs
are consistent with a common mechanism of ethylene oxide tumorigenesis in the B6C3F1 mouse. Toxicol Pathol 35:
81-85. http://dx.doi.org/10.1080/01926230601063839
Horner. MJ: Ries. LAG: Krapcho. M: Neyman. N: Aminou. R: Howlader. N: Altekruse. SF: Feuer. EJ: Huang. L:
Mariotto. A: Miller. BA: Lewis. PR: Eisner. MP: Stinchcomb. DG: Edwards. BK. (2009). SEER cancer statistics
review, 1975-2006. Bethesda, MD: National Cancer Institute. http://seer.cancer.gov/csr/1975_2006/
Hornung. RW: Greife. AL: Stayner. LT: Steenland. NK: Herrick. RF: Elliott LJ: Ringenburg. VL: Morawetz. J.
(1994). Statistical model for prediction of retrospective exposure to ethylene oxide in an occupational mortality
study. Am J Ind Med 25: 825-836.
Houle. CD: Ton. TV: Clayton. N: Huff. J: Hong. HH: Sills. RC. (2006). Frequent p53 and H-ras mutations in
benzene- and ethylene oxide-induced mammary gland carcinomas fromB6C3Fl mice. Toxicol Pathol 34: 752-762.
http://dx.doi.org/10.1080/01926230600935912
IARC (International Agency for Research on Cancer). (1994a). Ethylene. In IARC monographs on the evaluation of
carcinogenic risks to humans Some industrial chemicals (pp. 45-71). Lyon, France.
http://monographs.iarc.fr/ENG/Monographs/vol60/volume60.pdf
IARC (International Agency for Research on Cancer). (1994b). Ethylene oxide. In IARC monographs on the
evaluation of carcinogenic risks to humans Some industrial chemicals (pp. 73-159). Lyon, France.
http://monographs.iarc.fr/ENG/Monographs/vol60/volume60.pdf
IARC (International Agency for Research on Cancer). (2008a). 1,3-butadiene, ethylene oxide and vinyl halides
(vinyl fluoride, vinyl chloride and vinyl bromide). In IARC monographs on the evaluation of carcinogenic risks to
humans (pp. 3-471). Lyon, France. http://monographs.iarc.fr/ENG/Monographs/vol97/mono97.pdf
IARC (International Agency for Research on Cancer). (2008b). IARC monographs on the evaluation of carcinogenic
risks to humans. Volume 97. 1,3-butadiene, ethylene oxide and vinyl halides (vinyl fluoride, vinyl chloride and vinyl
bromide). 97: 3-471.
Ingvarsson. S. (1999). Molecular genetics of breast cancer progression [Review]. Semin Cancer Biol 9: 277-288.
http://dx.doi.org/10.1006/scbi.1999.0124
Johanson. G: Filser. JG. (1992). Experimental data from closed chamber gas uptake studies in rodents suggest lower
uptake rate of chemical than calculated from literature values on alveolar ventilation. Arch Toxicol 66: 291-295.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 R-4 DRAFT—DO NOT CITE OR QUOTE
-------
Kardos. L: Szeles. G: Gombkoto. G: Szeremj M: Tompa. A: Adany. R. (2003). Cancer deaths among hospital staff
potentially exposed to ethylene oxide: an epidemiological analysis. Environ Mol Mutagen 42: 59-60.
http://dx.doi.org/10.1002/em. 10167
Kiesselbach. N: Ulm. K: Lange. HJ: Korallus. U. (1990). A multicentre mortality study of workers exposed to
ethylene oxide. Br J Ind Med 47: 182-188.
Kirman. CR: Sweeney. LM: Teta. MJ: Sielken. RL: Valdez-Flores. C: Albertini. RJ: Gargas. ML. (2004).
Addressing nonlinearity in the exposure-response relationship for a genotoxic carcinogen: cancer potency estimates
for ethylene oxide. Risk Anal 24: 1165-1183. http://dx.doi.org/10.1111/i.0272-4332.2004.00517.x
Kolman. A: Chovanec. M. (2000). Combined effects of gamma-radiation and ethylene oxide in human diploid
fibroblasts. Mutagenesis 15: 99-104. http://dx.doi.0rg/10.1093/mutage/15.2.99
Kolman. A: Chovanec. M: Osterman-Golkar. S. (2002). Genotoxic effects of ethylene oxide, propylene oxide and
epichlorohydrin in humans: update review (1990-2001) [Review]. DNA Repair 512: 173-194.
Krewskj D: Crump. KS: Farmer. J: Gaylor. DW: Howe. R: Portier. C: Salsburg. D: Sielken. RL: VanRyzin. J.
(1983). A comparison of statistical methods for low dose extrapolation utilizing time-to-tumour data. Fundam Appl
Toxicol 3: 140-160. http://dx.doi.org/10.1016/S0272-0590(83)80075-X
Krishnan. K: Gargas. ML: Fennell TR: Andersen. ME. (1992). A physiologically based description of ethylene
oxide dosimetry in the rat. Toxicol Ind Health 8: 121-140.
Lambert. B: Andersson. B: Bastlova. T: Hou. SM: Hellgren. D: Kolman. A. (1994). Mutations induced in the
hypoxanthine phosphoribosyl transferase gene by three urban air pollutants: Acetaldehyde, benzo[a]pyrene
diolepoxide, and ethylene oxide. Environ Health Perspect Suppl 102: 135-138.
Langholz. B: Richardson. DB. (2010). Fitting general relative risk models for survival time and matched case-
control analysis. Am J Epidemiol 171: 377-383. http://dx.doi.org/10.1093/aje/kwp403
Laurent. C: Frederic. J: Leonard. AY. (1984). Sister chromatid exchange frequency in workers exposed to high
levels of ethylene oxide, in a hospital sterilization service. Int Arch Occup Environ Health 54:33-43.
Lerda. D: Rizzi. R. (1992). Cytogenetic study of persons occupationally exposed to ethylene oxide. Mutat Res 281:
31-37.
Lewis. SE: Barnett. LB: Felton. C: Johnson. FM: Skow. LC: Cacheiro. N: Shelby. MD. (1986). Dominant visible
and electrophoretically expressed mutations induced in male mice exposed to ethylene oxide by inhalation. Environ
Mol Mutagen 8: 867-872.
Liou. SH: Lung. JC: Chen. YH: Yang. T: Hsieh. LL: Chen. CJ: Wu. TN. (1999). Increased chromosome-type
chromosome aberration frequencies as biomarkers of cancer risk in a blackfoot endemic area. Cancer Res 59: 1481-
1484.
Lorenti Garcia. C: Darroudi. F: Tates. AD: Natarajan. AT. (2001). Induction and persistence of micronuclei, sister-
chromatid exchanges and chromosomal aberrations in splenocytes and bone-marrow cells of rats exposed to
ethylene oxide. Mutat Res Genet Toxicol Environ Mutagen 492: 59-67. http://dx.doi.org/10.1016/S1383-
5718(01100149-8
Lossos. IS: Levy. R. (2000). Mutation analysis of the 5' noncoding regulatory region of the BCL-6 gene in non-
Hodgkin lymphoma: evidence for recurrent mutations and intraclonal heterogeneity. Blood 95: 1400-1405.
Lynch. DW: Lewis. TR: Moorman. WJ: Burg. JR: Groth. DH: Khan. A: Ackerman. LJ: Cockrell BY. (1984a).
Carcinogenic and toxicologic effects of inhaled ethylene oxide and propylene oxide in F344 rats. Toxicol Appl
Pharmacol 76: 69-84.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 R-5 DRAFT—DO NOT CITE OR QUOTE
-------
Lynch. DW: Lewis. TR: Moorman. WJ: Burg. JR: Lai. JB: Setzer. JV: Groth. DH: Gulati. DK: Zavos. PM:
Sabharwal. PS: Ackerman. LJ: Cockrell. BY: Sprinz. H. (1984b). Effects on monkeys and rats of long-term
inhalation exposure to ethylene oxide: Major findings of the NIOSH study. In Inhospital ethylene oxide sterilization:
Current issues in ethylene oxide toxicity and occupational exposure (pp. 7-10). (AAMI Technology Assessment
Report No. 8-84). Arlington, VA: Association for the Advancement of Medical Instrumentation.
Major. J: Jakab. MG: Tompa. A. (1996). Genotoxicological investigation of hospital nurses occupationally exposed
to ethylene-oxide: I Chromosome aberrations, sister-chromatid exchanges, cell cycle kinetics, and UV-induced DNA
synthesis in peripheral blood lymphocytes. Environ Mol Mutagen 27: 84-92. http://dx.doi.org/10.1002/(SICI) 1098-
2280(1996)27:2<:84::AID-EM2>:3.0.CO:2-E
Major. J: Jakab. MG: Tompa. A. (2001). Genotoxicological investigation of hospital nurses occupationally exposed
to ethylene oxide. II. HPRT mutation frequencies. Central Eur J Occup Env Med 7: 195-208.
Marsden. DA: Jones. DJ: Britton. RG: Ognibene. T: Ubick. E: Johnson. GE: Farmer. PB: Brown. K. (2009). Dose-
response relationships for N7-(2-hydroxyethyl)guanine induced by low-dose [14C]ethylene oxide: evidence for a
novel mechanism of endogenous adduct formation. Cancer Res 69: 3052-3059. http://dx.doi.org/10.1158/0008-
5472.CAN-08-4233
Mayer. J: Warburton. D: Jeffrey. AM: Pero. R: Walles. S: Andrews. L: Toor. M: Latriano. L: Wazneh. L: Tang. D:
Tsai. WY: Kuroda. M: Perera. F. (1991). Biologic markers in ethylene oxide exposed workers and controls. Mutat
Res 248: 163-176. http://dx.doi.org/10.1016/0027-5107(91)90098-9
Memisoglu. A: Samson. L. (2000). Base excision repair in yeast and mammals [Review]. Mutat Res 451: 39-51.
Mertes. I: Fleischmann. R: Glatt H: Oesch. F. (1985). Interindividual variations in the activities of cytosolic and
microsomal epoxide hydrolase in human liver. Carcinogenesis 6: 219-223.
Minifio. AM: Arias. E: Kochanek. KD: Murphy. SL: Smith. BL. (2002). Deaths: Final Data for 2000. Hyattsville,
MD: National Center for Health Statistics, http://www.cdc.gov/nchs/data/nvsr/nvsr50/nvsr50_15.pdf
Nataraian. AT: Preston. RJ: Dellarco. V: Ehrenberg. L: Generoso. W: Lewis. S: Tates. AD. (1995). Ethylene oxide:
evaluation of genotoxicity data and an exploratory assessment of genetic risk [Review]. DNA Repair 330: 55-70.
NCI (National Cancer Institute). (2004). Breast cancer screening (PDQ): Health professional version. Washington,
DC: U.S. Department of Health, Education, and Welfare, National Institutes of Health.
http://www.cancer.gov/cancertopics/pdq/screening/breast/HealthProfessional/
Norman. SA: Berlin. JA: Soper. KA: Middendorf. BF: Stolley. PP. (1995). Cancer incidence in a group of workers
potentially exposed to ethylene oxide. Int J Epidemiol 24: 276-284.
NRC (National Research Council). (1983). Risk assessment in the federal government: Managing the process.
Washington, DC: National Academies Press. http://www.nap.edu/openbook.php?record_id=366&page=Rl
NTP (National Toxicology Program). (1987). Toxicology and carcinogenesis studies of ethylene oxide (CAS no 75-
21-8) inB6C3Fl mice (inhalation studies).
Olsen. GW: Lacy. SE: Bodner. KM: Chau. M: Arceneaux. TG: Cartmill. JB: Ramlow. JM: Boswell. JM. (1997).
Mortality from pancreatic and lymphopoietic cancer among workers in ethylene and propylene chlorohydrin
production. Occup Environ Med 54: 592-598. http://dx.doi.0rg/10.1136/oem.54.8.592
Pauwels. W: Veulemans. H. (1998). Comparison of ethylene, propylene and styrene 7,8-oxide in vitro adduct
formation on N-terminal valine in human haemoglobin and on N-7-guanine in human DNA. Mutat Res 418: 21-33.
Paz-v-Mino. C: Perez. JC: Fiallo. BF: Leone. PE. (2002). A polymorphism in the hMSH2 gene (gIVS12-6T>C)
associated with non-Hodgkin lymphomas. Cancer Genet Cytogenet 133: 29-33.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 R-6 DRAFT—DO NOT CITE OR QUOTE
-------
Pedersen-Bjergaard. J: Christiansen. DH: Desta. F: Andersen. MK. (2006). Alternative genetic pathways and
cooperating genetic abnormalities in the pathogenesis of therapy-related myelodysplasia and acute myeloid
leukemia [Review]. Leukemia 20: 1943-1949. http://dx.doi.org/10.1038/sj.leu.2404381
Pfeiffer. P: Goedecke. W: Obe. G. (2000). Mechanisms of DNA double-strand break repair and their potential to
induce chromosomal aberrations [Review]. Mutagenesis 15: 289-302.
Popp. W: Vahrenholz. C: Przygoda. H: Brauksiepe. A: Goch. S: Muller. G: Schell C: Norpoth. K. (1994). DNA-
protein cross-links and sister chromatid exchange frequencies in lymphocytes and hydroxyethyl mercapturic acid in
urine of ethylene oxide-exposed hospital workers. Int Arch Occup Environ Health 66: 325-332.
Preston. RJ. (1999). Cytogenetic effects of ethylene oxide, with an emphasis on population monitoring [Review].
Crit Rev Toxicol 29: 263-282. http://dx.doi.org/10.1080/10408449991349212
Preston. RJ: Fennell. TR: Leber. AP: Sielken. RL: Swenberg. JA. (1995). Reconsideration of the genetic risk
assessment for ethylene oxide exposures. Environ Mol Mutagen 26: 189-202.
Preudhomme. C: Warot-Loze. D: Roumier. C: Grardel-Duflos. N: Garand. R: Lai. JL: Dastugue. N: Macintyre. E:
Denis. C: Bauters. F: Kerckaert. JP: Cosson. A: Fenaux. P. (2000). High incidence of biallelic point mutations in the
Runt domain of the AML1/PEBP2 alpha B gene in Mo acute myeloid leukemia and in myeloid malignancies with
acquired trisomy 21. Blood 96: 2862-2869.
Rapoport IA. (1948). The effect of ethylene oxide, glycide and glycol on genetic mutations. Dokl Biochem Biophys
60: 469-472.
Recio. L: Dormer. M: Abernethy. D: Pluta. L: Steen. AM: Wong. BA: James. A: Preston. RJ. (2004). In vivo
mutagenicity and mutation spectrum in the bone marrow and testes of B6C3F1 lacl transgenic mice following
inhalation exposure to ethylene oxide. Mutagenesis 19: 215-222.
Ribeiro. LR: Salvadori. DM: Rios. AC: Costa. SL: Tates. AD: Tornqvist M: Natarajan. AT. (1994). Biological
monitoring of workers occupationally exposed to ethylene oxide. Mutat Res 313: 81-87.
Richmond. GW: Abrahams. RH: Nemenzo. JH: Hine. CH. (1985). An evaluation of possible effects on health
following exposure to ethylene oxide. Arch Environ Occup Health 40: 20-25.
Ries. LAG: Eisner. MP: Kosary. CL: Hankev. BF: Miller. BA: Clegg. L: Mariotto. A: Feuer. EF: Edwards. BK.
(2004). SEER (Surveillance Epidemiology and End Results) cancer statistics review, 1975-2001. Bethesda, MD:
National Cancer Institute, U.S. Department of Health, Education, and Welfare, National Institutes of Health.
http://seer.cancer.gov/csr/1975 2001
Ries. LAG: Melbert. D: Krapcho. M: Mariotto. A: Miller. BA: Feuer. EJ: Clegg. L: Horner. MJ: Howlader. N:
Eisner. MP. (2007). SEER (Surveillance Epidemiology and End Results) cancer statistics review, 19752004.
Bethesda, MD: National Cancer Institute, U.S. Department of Health, Education, and Welfare, National Institutes of
Health. http://seer.cancer.gov/csr/1975_2004
Rossner. P: Boffetta. P: Ceppi. M: Bonassi. S: Smerhovsky. Z: Landa. K: Juzova. D: Sram. RJ. (2005).
Chromosomal aberrations in lymphocytes of healthy subjects and risk of cancer. Environ Health Perspect 113: 517-
520. http://dx.doi.org/10.1289/ehp.6925
Rothmaa KJ. (1986). Modern epidemiology. Boston, MA: Little Brown & Co.
Russo. J: Russo. IH. (1999). Cellular basis of breast cancer susceptibility [Review]. OncolRes 11: 169-178.
Rusyn. I: Asakura. S: Li. Y: Kosyk. O: Koc. H: Nakamura. J: Upton. PB: Swenberg. JA. (2005). Effects of ethylene
oxide and ethylene inhalation on DNA adducts, apurinic/apyrimidinic sites and expression of base excision DNA
repair genes in rat brain, spleen, and liver. DNA Repair 4: 1099-1110.
http://dx.doi.0rg/10.1016/j.dnarep.2005.05.009
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 R-7 DRAFT—DO NOT CITE OR QUOTE
-------
SAB (Science Advisory Board). (2007). Review of Office of Research and Development (ORD) draft assessment
entitled "Evaluation of the carcinogenicity of ethylene oxide". Washington, DC: Science Advisory Board, U.S.
Environmental Protection Agency.
http://vosemite.epa.gov/sab/sabproduct.nsf/368203f97al5308a852574ba005bbd01/5D661BC118B527A3852573B8
0068C97B/$File/EP A-SAB-08-004-unsigned.pdf
Sarto. F: Clonfero. E: Bartolucci. GB: Franceschi. C: Chiricolo. M: Levis. AG. (1987). Sister chromatid exchanges
and DNA repair capability in sanitary workers exposed to ethylene oxide: Evaluation of the dose-effect relationship.
Am J Ind Med 12: 625-637. http://dx.doi.org/10.1002/aiim.4700120515
Sarto. F: Cominato. I: Pinton. AM: Brovedani. PG: Faccioli. CM: Bianchi. V: Levis. AG. (1984). Cytogenetic
damage in workers exposed to ethylene oxide. Mutat Res 138: 185-195.
Sarto. F: Tomanin. R: Giacomelli. L: lannini. G: Cupiraggi. AR. (1990). The micronucleus assay in human
exfoliated cells of the nose and mouth: application to occupational exposures to chromic acid and ethylene oxide.
Mutat Res 244: 345-351.
Sarto. F: Tornqvist MA: Tomanin. R: Bartolucci. GB: Osterman-Golkar. SM: Ehrenberg. L. (1991). Studies of
biological and chemical monitoring of low-level exposure to ethylene oxide. Scand J Work Environ Health 17: 60-
64.
Schulte. PA: Boeniger. M: Walker. JT: Schober. SE: Pereira. MA: Gulati. DK: Wojciechowski. JP: Garza. A:
Froelich. R: Strauss. G. (1992). Biologic markers in hospital workers exposed to low levels of ethylene oxide. Mutat
Res 278: 237-251.
Segerback. D. (1990). Reaction products in hemoglobin and DNA after in vitro treatment with ethylene oxide and
N-(2-hydroxyethyl)-N-nitrosourea. Carcinogenesis 11: 307-312.
Sielken. RL: Valdez-Flores. C. (2009). Life-table calculations of excess risk for incidence versus mortality: ethylene
oxide case study. Regul Toxicol Pharmacol 55: 82-89. http://dx.doi.0rg/10.1016/j.yrtph.2009.06.003
Sisk. SC: Pluta. LJ: Meyer. KG: Wong. BC: Recio. L. (1997). Assessment of the in vivo mutagenicity of ethylene
oxide in the tissues of B6C3F1 lacl transgenic mice following inhalation exposure. Mutat Res 391: 153-164.
Snellings. WM: Weil. CS: Maronpot. RR. (1984). A two-year inhalation study of the carcinogenic potential of
ethylene oxide in Fischer 344 rats. Toxicol Appl Pharmacol 75: 105-117.
Stayner. L: Steenland. K: Dosemeci M: Hertz-Picciotto. I. (2003). Attenuation of exposure-response curves in
occupational cohort studies at high exposure levels. Scand J Work Environ Health 29: 317-324.
Stayner. L: Steenland. K: Greife. A: Hornung. R: Hayes. RB: Nowlin. S: Morawetz. J: Ringenburg. V: Elliot. L:
Halperin. W. (1993). Exposure-response analysis of cancer mortality in a cohort of workers exposed to ethylene
oxide. AmJEpidemiol 138: 787-798.
Steenland. K: Deddens. J: Piacitelli. L. (2001). Risk assessment for 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD)
based on an epidemiologic study. Am J Epidemiol 154: 451-458.
Steenland. K: Deddens. JA. (2004). A practical guide to dose-response analyses and risk assessment in occupational
epidemiology [Review]. Epidemiology 15: 63-70. http://dx.doi.org/10.1097/01.ede.0000100287.45004.e7
Steenland. K: Stayner. L: Deddens. J. (2004). Mortality analyses in a cohort of 18 235 ethylene oxide exposed
workers: follow up extended from 1987 to 1998. Occup EnvironMed 61: 2-7.
Steenland. K: Stayner. L: Greife. A: Halperin. W: Haves. R: Hornung. R: Nowlin. S. (1991). Mortality among
workers exposed to ethylene oxide. N Engl J Med 324: 1402-1407.
http://dx.doi.org/10.1056/NEJM199105163242004
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 R-8 DRAFT—DO NOT CITE OR QUOTE
-------
Steenland. K: Whelan. E: Deddens. J: Stayner. L: Ward. E. (2003). Ethylene oxide and breast cancer incidence in a
cohort study of 7576 women (United States). Cancer Causes Control 14: 531-539.
Stolley. PD: Soper. KA: Galloway. SM: Nichols. WW: Norman. SA: Wolman. SR. (1984). Sister-chromatid
exchanges in association with occupational exposure to ethylene oxide. Mutat Res 129: 89-102.
http://dx.doi.org/10.1016/0027-5107(84)90127-l
Swaen. GM: Burns. C: Teta. JM: Bodner. K: Keenan. D: Bodnar. CM. (2009). Mortality study update of ethylene
oxide workers in chemical manufacturing: a 15 year update. J Occup Environ Med 51: 714-723.
http://dx.doi.org/10.1097/JOM.Ob013e3181a2ca20
Swaen. GM: Slangen. JM: Ott MG: Kusters. E: Van Den Langenbergh. G: Arends. JW: Zober. A. (1996).
Investigation of a cluster often cases of Hodgkin's disease in an occupational setting. Int Arch Occup Environ
Health 68: 224-228.
Swenberg. JA: Fedtke. N: Fennell TR: Walker. VE. (1990). Relationships between carcinogen exposure, DNA
adducts and carcinogenesis. In DB Clayson; 1C Munro; P Shubik; JA Swenberg (Eds.), Progress in predictive
toxicology (pp. 161-184). New York, NY: Elsevier.
Tates. AD: Boogaard. PJ: Darroudi. F: Natarajaa AT: Caubo. ME: Van Sittert. NJ. (1995). Biological effect
monitoring in industrial workers following incidental exposure to high concentrations of ethylene oxide. DNA
Repair 329: 63-77.
Tates. AD: Grummt. T: Tornqvist. M: Farmer. PB: van Dam. FJ: vanMossel. H: Schoemaker. HM: Osterman-
Golkar. S: Uebel. C: Tang. YS: Zwinderman. AH: Nataraian. AT: Ehrenberg. L. (1991). Biological and chemical
monitoring of occupational exposure to ethylene oxide. Mutat Res 250: 483-497. http://dx.doi.org/10.1016/0027-
5107(91190205-3
Tates. AD: Van Dam. FJ: Natarajan. AT: Van Teylingen. CMM: De Zwart. FA: Zwinderman. AH: Van Sittert. NJ:
Nilsen. A: Nilsen. OG: Zahlsen. K: Magnusson. AL: Tornqvist. M. (1999). Measurement of HPRT mutations in
splenic lymphocytes and haemoglobin adducts in erythrocytes of Lewis rats exposed to ethylene oxide. DNA Repair
431:397-415.
Teta. MJ: Bensoa LO: Vitale. JN. (1993). Mortality study of ethylene oxide workers in chemical manufacturing: a
10 year update. Br J Ind Med 50: 704-709.
Teta. MJ: Sielken. RL: Valdez-Flores. C. (1999). Ethylene oxide cancer risk assessment based on epidemiological
data: application of revised regulatory guidelines. Risk Anal 19: 1135-1155.
Thier. R: Bolt. HM. (2000). Carcinogenicity and genotoxicity of ethylene oxide: new aspects and recent advances
[Review]. Crit Rev Toxicol 30: 595-608. http://dx.doi.org/10.1080/10408440008951121
Tompa. A: Major. J: Jakab. MG. (1999). Is breast cancer cluster influenced by environmental and occupational
factors among hospital nurses in Hungary. Pathol Oncol Res 5: 117-121.
Tompkins. EM: McLuckie. KI: Jones. DJ: Farmer. PB: Browa K. (2009). Mutagenicity of DNA adducts derived
from ethylene oxide exposure in the pSP189 shuttle vector replicated in human Ad293 cells. Mutat Res 678: 129-
137. http://dx.doi.0rg/10.1016/i.mrgentox.2009.05.011
Tornqvist M. (1996). Ethylene oxide as a biological reactive intermediate of endogenous origin [Review]. Adv Exp
Med Biol 387: 275-283.
U.S. EPA (U.S. Environmental Protection Agency). (1986). Guidelines for carcinogen risk assessment. Fed Reg 51:
33993-34003.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 R-9 DRAFT—DO NOT CITE OR QUOTE
-------
U.S. EPA (U.S. Environmental Protection Agency). (1994). Methods for derivation of inhalation reference
concentrations and application of inhalation dosimetry [EPA Report]. (EPA/600/8-90/066F). Research Triangle
Park, NC. http://cfpub.epa.gov/ncea/cfm/recordisplav.cfm?deid=71993
U.S. EPA (U.S. Environmental Protection Agency). (1997). Chemical and radiation leukemogenesis in humans and
rodents and the value of rodent models for assessing risks of lymphohematopoietic cancers [EPA Report].
(EPA/600/R-97/090). Washington, DC: National Center for Environmental Assessment, Office of Research and
Development, http://cfpub.epa. gov/ncea/cfm/recordisplav.cfm?deid= 12181
U.S. EPA (U.S. Environmental Protection Agency). (2000). Science policy council handbook: Risk characterization
[EPA Report]. (EPA 100-B-00-002). Washington, D.C.: Office of Science Policy, Office of Research and
Development, http://www.epa.gov/osa/spc/pdfs/rchandbk.pdf
U.S. EPA (U.S. Environmental Protection Agency). (2005a). Guidelines for carcinogen risk assessment [EPA
Report]. (EPA/630/P-03/001F). Washington, DC: Risk Assessment Forum, http://www.epa.gov/cancerguidelines/
U.S. EPA (U.S. Environmental Protection Agency). (2005b). Supplemental guidance for assessing susceptibility
from early-life exposure to carcinogens. (EPA/630/R-03/003F). Washington, DC: U.S. Environmental Protection
Agency, Risk Assessment Forum. http://www.epa.gov/raf/publications/pdfs/childrens_supplement_fmal.pdf.
U.S. EPA (U.S. Environmental Protection Agency). (2006a). Evaluation of the carcinogenicity of ethylene oxide:
external review draft [EPA Report]. (EPA/635/R-06/003). Washington, DC: National Center for Environmental
Assessment, Office of Research and Development, http://cfpub.epa. gov/ncea/cfm/recordisplav.cfm?deid= 157664
U.S. EPA (U.S. Environmental Protection Agency). (2006b). Science policy council peer review handbook 3rd
edition [EPA Report]. (100B06002). Science Policy Council.
http://cfpub.epa. gov/ncea/cfm/recordisplay.cfm?deid= 157664
U.S. EPA (U.S. Environmental Protection Agency). (2012). Benchmark dose technical guidance. (EPA/100/R-
12/001). Washington, DC: Risk Assessment Forum.
http://www.epa. gov/raf/publications/pdfs/benchmark_dose_guidance.pdf
Valdez-Flores. C: Sielkea RL: Teta. MJ. (2010). Quantitative cancer risk assessment based on NIOSH and UCC
epidemiological data for workers exposed to ethylene oxide. Regul Toxicol Pharmacol 56: 312-320.
http://dx.doi.0rg/10.1016/i.vrtph.2009.10.001
van Delft JH: van Winden. MJ: Luiten-Schuite. A: Ribeiro. LR: Baan. RA. (1994). Comparison of various
immunochemical assays for the detection of ethylene oxide-DNA adducts with monoclonal antibodies against
imidazole ring-opened N7-(2-hydroxyethyl)guanosine: application in a biological monitoring study. Carcinogenesis
15: 1867-1873.
van Sittert. NJ: Beulink. GD: van Vliet. EW: van der Waal. H. (1993). Monitoring occupational exposure to
ethylene oxide by the determination of hemoglobin adducts. Environ Health Perspect 99: 217-220.
van Sittert. NJ: Boogaard. PJ: Natarajan. AT: Tates. AD: Ehrenberg. LG: Tornqvist MA. (2000). Formation of
DNA adducts and induction of mutagenic effects in rats following 4 weeks inhalation exposure to ethylene oxide as
a basis for cancer risk assessment. Mutat Res-Fundam Mol Mech Mutagen 447: 27-48.
http://dx.doi.org/10.1016/80027-5107(99)00208-0
Van Sittert. NJ: de Jong. G. (1985). Biomonitoring of exposure to potential mutagens and carcinogens in industrial
populations. Food Chem Toxicol 23: 23-31.
van Wijngaarden. E: Hertz-Picciotto. I. (2004). A simple approach to performing quantitative cancer risk assessment
using published results from occupational epidemiology studies. Sci Total Environ 332: 81-87.
http://dx.doi.0rg/10.1016/i.scitotenv.2004.04.005
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 R-10 DRAFT—DO NOT CITE OR QUOTE
-------
Walker. VE: Fennell TR: Boucheron. JA: Fedtke. N: Ciroussel F: Swenberg. JA. (1990). Macromolecular adducts
of ethylene oxide: a literature review and a time-course study on the formation of 7-(2-hydroxyethyl)guanine
following exposures of rats by inhalation [Review]. DNA Repair 233: 151-164.
Walker. VE: Fennell TR: Upton. PB: MacNeela. JP: Swenberg. JA. (1993). Molecular dosimetry of DNA and
hemoglobin adducts in mice and rats exposed to ethylene oxide. Environ Health Perspect 99: 11-17.
Walker. VE: Fennell TR: Upton. PB: Skopek. TR: Prevost V: Shuker. PEG: Swenberg. JA. (1992a). Molecular
dosimetry of ethylene oxide: formation and persistence of 7-(2-hydroxyethyl)guanine in DNA following repeated
exposures of rats and mice. Cancer Res 52: 4328-4334.
Walker. VE: MacNeela. JP: Swenberg. JA: Turner. MJ. Jr: Fennell. TR. (1992b). Molecular dosimetry of ethylene
oxide: formation and persistence of N-(2-hydroxyethyl)valine in hemoglobin following repeated exposures of rats
and mice. Cancer Res 52: 4320-4327.
Walker. VE: Sisk. SC: Upton. PB: Wong. BA: Recio. L. (1997). In vivo mutagenicity of ethylene oxide at the hprt
locus in T-lymphocytes of B6C3F1 lacl transgenic mice following inhalation exposure. Mutat Res 392: 211-222.
Walker. VE: Skopek. TR. (1993). A mouse model for the study of in vivo mutational spectra: sequence specificity
of ethylene oxide at the hprt locus. Mutat Res 288: 151-162. http://dx.doi.org/10.1016/0027-5107(93)90216-3
Warwick. GP. (1963). The mechanism of action of alkylating agents [Review]. Cancer Res 23: 1315-1333.
WHO (World Health Organization). (2003). Concise International Chemical Assessment Document: Ethylene oxide.
Geneva: International Programme on Chemical Safety (IPCS).
http://www.inchem.org/documents/cicads/cicad54.htm
Wong. O: Trent. LS. (1993). An epidemiological study of workers potentially exposed to ethylene oxide. Occup
Environ Med 50: 308-316.
Yager. JW: Hines. CJ: Spear. RC. (1983). Exposure to ethylene oxide at work increases sister chromatid exchanges
in human peripheral lymphocytes. Science 219: 1221-1223. http://dx.doi.org/10.1126/science.6828851
Yong. LC: Schulte. PA: Kao. CY: Giese. RW: Boeniger. MF: Strauss. GH: Petersen. MR: Wiencke. JK. (2007).
DNA adducts in granulocytes of hospital workers exposed to ethylene oxide. Am J Ind Med 50: 293-302.
http://dx.doi.org/10.1002/ajim.20443
Yong. LC: Schulte. PA: Wiencke. JK: Boeniger. MF: Connally. LB: Walker. JT: Whelan. EA: Ward. EM. (2001).
Hemoglobin adducts and sister chromatid exchanges in hospital workers exposed to ethylene oxide: effects of
glutathione S-transferase Tl and Ml genotypes. Cancer Epidemiol Biomarkers Prev 10: 539-550.
Zharlyganova. D: Harada. H: Harada. Y: Shinkarev. S: Zhumadilov. Z: Zhunusova. A: Tchaizhunusova. NJ:
Apsalikov. KN: Kemaikin. V: Zhumadilov. K: Kawano. N: Kimura. A: Hoshi. M. (2008). High frequency of
AML1/RUNX1 point mutations in radiation-associated myelodysplastic syndrome around Semipalatinsk nuclear
test site. J Radiat Res (Tokyo) 49: 549-555.
This document is a draft for review purposes only and does not constitute Agency policy.
7/2013 R-11 DRAFT—DO NOT CITE OR QUOTE
------- |