® EPA
EPA/635/R-14/194a
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)
August 2014
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.
ii DRAFT—DO NOT CITE OR QUOTE
-------
NOTE TO REVIEWERS
A draft of this assessment received external peer review by EPA's Science Advisory
Board (SAB) and was being revised according to the peer review recommendations when the
2011 National Research Council (NRC) report was released with general recommendations for
improvements to the IRIS process. As noted in the 2011 report, the NRC encouraged EPA to
proceed with assessments while developing improvements to the IRIS Program. Consistent with
this advice, EPA has indicated that it would not go backwards in the assessment development
process, but would focus on moving forward, while also phasing in improvements.
While the ethylene oxide (EtO) assessment does not incorporate all the revisions to the
IRIS assessment format and methodology recommended in the 2011 NRC recommendations
(and the more recent 2014 NRC Review of the IRIS Process), such as the inclusion of a standard
Preamble, systematic review and standardized approaches for evidence integration, this
assessment is streamlined, and uses tables, figures, and appendices to increase transparency and
clarity. In addition, the assessment is structured to have separate hazard identification and dose-
response sections and the update to the literature search was conducted using systematic
literature search approaches. Additionally, consistent with the goal that assessments should
provide a scientifically sound and transparent evaluation of the relevant scientific literature and
presentation of the analyses performed, this assessment contains an expanded discussion on the
rationales for study evaluation and selection, as well as other key assessment decisions.
Appendix K documents where the recommendations from Chapter 7 of the NRC 2011 report
have been implemented in this assessment.
EPA obtained public comment on an external review draft in 2006 (U.S. EPA, 2006a)
and completed a peer review by a panel of EPA's Science Advisory Board (SAB) in 2007 (SAB,
2007). A summary of the public and peer review comments and EPA's disposition of these
comments is presented in Appendix H of the current revised draft assessment. The consensus
conclusions of the SAB review supported the conclusions in the draft external review assessment
regarding the cancer classification and the selection of the key dataset for quantification of
cancer risk estimates. The SAB recommended that EPA examine several issues further, such as
endogenous EtO production, whether an alternative dataset would add useful exposure-response
information, and whether the primary epidemiology dataset could be modeled using continuous
data across the full exposure range without first converting the exposure and individual
occurrence data into categorical data for the derivation of unit risk estimates. The SAB also
provided comments on both sides of the issue of whether or not a nonlinear low-dose
This document is a draft for review purposes only and does not constitute Agency policy.
iii DRAFT—DO NOT CITE OR QUOTE
-------
extrapolation approach should be included in the analysis with two conflicting views articulated
in the Appendices to the SAB's final report.
Because of the importance of this assessment, the complex issues in modeling
epidemiology data, and the new modeling of epidemiology data done in response to the prior
SAB peer review, EPA decided to seek additional SAB peer review. Prior to the additional
external peer review, EPA released a revised draft for public comment in July 2013. The revised
draft was discussed during an IRIS Bimonthly Public Science meeting in December 2013. A
summary of public comments, and EPA's responses are available in Appendix L.
EPA did not formally record the details of literature search strategies employed in
identifying the relevant EtO literature for the development of the 2006 external peer review draft.
There were no critical studies identified as missing during public comment and peer review, but
the 2007 SAB final report recommended inclusion of additional supporting studies. EPA revised
the assessment according to the recommendations of the peer review panel, including extensive
new modeling of epidemiologic data, and incorporated new studies that were identified through
June 2010 into the revised assessment. As indicated above, EPA decided to seek additional SAB
peer review primarily because of the new modeling of epidemiologic data done in response to
the SAB recommendations. Therefore, in May 2013, in order to ensure that no critical studies
had been missed that would warrant major revisions to the assessment, EPA conducted a well-
documented systematic literature search of literature published from January 2006 through May
2013. The literature search strategy was conducted and documented following the 2011 NRC
recommendations for more formal systematic literature searches. Relevant studies that could
potentially impact the cancer hazard characterization or dose-response assessment were
identified and considered. Appendix J provides documentation of the search methods, the bases
for the judgments of the relevancy of new literature, and the disposition of studies identified in
the 2013 search (Section J.I). Appendix J also includes reviews of two major studies published
after June 2010 that were identified in the search and merited in-depth discussion (Section J.2).
Two additional studies of potential importance to the outcome of the assessment, both published
after the May 2013 search, were noted in public comments in October 2013 and are also
reviewed in Appendix J (Section J.3). None of the additional studies that were identified were
found to have an impact on the final conclusions of the assessment. A list of the 131 references
added after the 2006 external peer review can be found in Appendix I. References considered
and cited in this document, including abstracts, can be found on the Health and Environmental
Research Online (HERO) website.1
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
This document is a draft for review purposes only and does not constitute Agency policy.
iv DRAFT—DO NOT CITE OR QUOTE
-------
Development of the hazard identification and dose-response assessments for EtO has
followed the general guidelines for risk assessment as set forth by the National Research Council
(NRC, 1983). United States Environmental Protection Agency (U.S. EPA) Guidelines and Risk
Assessment Forum Technical Panel Reports that were used in the development of this
assessment include the following: Guidelines for Mutagenicity Risk Assessment (U.S. EPA,
1986), Methods for Derivation of Inhalation Reference Concentrations and Application of
Inhalation Dosimetry (U.S. EPA, 1994), Benchmark Dose Technical Guidance (U.S. EPA,
2012), Science Policy Council Handbook: Risk Characterization (U.S. EPA, 2000), Guidelines
for Carcinogen Risk Assessment fU.S. EPA, 2005aA Supplemental Guidance for Assessing
Susceptibility from Early-Life Exposure to Carcinogens (U.S. EPA, 2005b), and Science Policy
Council Handbook: Peer Review (U.S. EPA, 2006b).
managed in EPA's Office of Research and Development (ORD) by the National Center for Environmental
Assessment (NCEA). The database includes more than 1,000,000 scientific articles from the peer-reviewed
literature. New studies are added continuously to HERO.
This document is a draft for review purposes only and does not constitute Agency policy.
v DRAFT—DO NOT CITE OR QUOTE
-------
CONTENTS
LIST OF TABLES viii
LIST OF FIGURES x
LIST OF ABBREVIATIONS xi
AUTHORS, CONTRIBUTORS, AND REVIEWERS xii
1. EXECUTIVE SUMMARY 1-1
2. INTRODUCTION 2-1
2.1. LITERATURE IDENTIFICATION 2-2
2.2. NRC RECOMMENDATIONS OF 2011 2-3
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 Ethyl ene Oxide Carcinogenicity
under EPA's Mode-of-Action Framework 3-42
3.5. HAZARD CHARACTERIZATION 3-45
3.5.1. Characterization of Cancer Hazard 3-45
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 4-22
4.1.3. Total Cancer Risk Estimates 4-44
4.1.4. Sources of Uncertainty in the Cancer Risk Estimates 4-46
4.1.5. Summary 4-58
4.2. INHALATION UNIT RISK DERIVED FROM EXPERIMENTAL ANIMAL
DATA 4-59
4.2.1. Overall Approach 4-59
4.2.2. Cross-Species Scaling 4-60
4.2.3. Dose-Response Modeling Methods 4-62
4.2.4. Description of Experimental Animal Studies 4-64
4.2.5. Results of Data Analysis of Experimental Animal Studies 4-65
This document is a draft for review purposes only and does not constitute Agency policy.
vi DRAFT—DO NOT CITE OR QUOTE
-------
CONTENTS (continued)
4.3. SUMMARY OF INHALATION UNIT RISK ESTIMATES—NOT
ACCOUNTING FOR ASSUMED INCREASED EARLY-LIFE
SUSCEPTIBILITY 4-67
4.4. ADJUSTMENTS FOR POTENTIAL INCREASED EARLY-LIFE
SUSCEPTIBILITY 4-68
4.5. INHALATION UNIT RISK ESTIMATES—CONCLUSIONS 4-74
4.6. COMPARISON WITH OTHER PUBLISHED RISK ESTIMATES 4-82
4.6.1. Unit Risk Estimates Based on Human Studies 4-82
4.6.2. Unit Risk Estimates Based on Laboratory Animal Studies 4-85
4.7. RISK ESTIMATES FOR SOME OCCUPATIONAL EXPOSURE
SCENARIOS 4-85
REFERENCES R-l
APPENDIX A. CRITICAL REVIEW OF EPIDEMIOLOGIC EVIDENCE A-l
APPENDIX B. REFERENCES FOR FIGURE 3-3 B-l
APPENDIX C. GENOTOXICITY AND MUTAGENICITY OF ETHYLENE OXIDE C-l
APPENDIX D. REANALYSES AND INTERPRETATION OF ETHYLENE OXIDE
EXPOSURE-RESPONSE DATA D-l
APPENDIX E. LIFE-TABLE ANALYSIS E-l
APPENDIX F. EQUATIONS USED FOR WEIGHTED LINEAR REGRESSION F-l
APPENDIX G. MODEL PARAMETERS IN THE ANALYSIS OF ANIMAL TUMOR
INCIDENCE G-l
APPENDIX H. SUMMARY OF 2007 EXTERNAL PEER REVIEW AND PUBLIC
COMMENTS AND DISPOSITION H-l
APPENDIX I. LIST OF REFERENCES ADDED AFTER THE 2006 EXTERNAL
REVIEW DRAFT 1-1
APPENDIX J. SUMMARY OF MAJOR NEW STUDIES SINCE THE LITERATURE
CUTOFF DATE J-l
APPENDIX K. DOCUMENTATION OF IMPLEMENTATIONS OF THE 2011
NATIONAL RESEARCH COUNCIL RECOMMENDATIONS K-1
APPENDIX L. SUMMARY OF PUBLIC COMMENTS RECEIVED ON THE JULY
2013 PUBLIC COMMENT DRAFT AND EPA RESPONSES L-l
This document is a draft for review purposes only and does not constitute Agency policy.
vii DRAFT—DO NOT CITE OR QUOTE
-------
LIST OF TABLES
Table 1-1. Summary of maj or findings 1-7
Table 3-1. Epidemiological studies of ethylene oxide and human
cancer—lymphohematopoietic cancer results51 3-13
Table 3-2. Summary of epidemiological results on EtO and female breast cancer (all
sterilizer workers)51 3-17
Table 3-3. Tumor incidence data in National Toxicology Program Study of B6C3Fi mice
NTP (1987)a and exposure-response modeling results 3-21
Table 3-4. Tumor incidence data in Lynch et al. (1984a; 1984b) study of male F344 rats
and exposure-response modeling results 3-22
Table 3-5. Tumor incidence data in Snellings et al. (1984) and Garman et al. (1985)
reports on F344 ratsa and exposure-response modeling results 3-24
Table 3-6. Cytogenetic effects in humans 3-33
Table 4-1. Considerations used in this assessment for selecting epidemiology studies for
quantitative risk estimation 4-3
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) 4-6
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) 4-12
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 4-14
Table 4-5. ECoi, LECoi, and unit risk estimates for lymphoid cancera 4-16
Table 4-6. ECoi, LECoi, and unit risk estimates for all lymphohematopoietic cancera 4-19
Table 4-7. Cox regression results for breast cancer mortality in females in the NIOSH
cohort,51 for models presented in Steenland et al. (2004) 4-23
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) 4-28
Table 4-9. ECoi, LECoi, and unit risk estimates for breast cancer mortality in females3 4-29
This document is a draft for review purposes only and does not constitute Agency policy.
viii DRAFT—DO NOT CITE OR QUOTE
-------
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 4-31
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) 4-37
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 4-39
Table 4-13. ECoi, LECoi, and unit risk estimates for breast cancer incidence in females—
invasive and in situa 4-41
Table4-14. Calculation ofECoi for total cancer risk 4-45
Table 4-15. Calculation of total cancer unit risk estimate 4-45
Table 4-16. Upper-bound unit risks (per ug/m3) obtained by combining tumor sites 4-64
Table 4-17. Unit risk values from multistage Weibull3 time-to-tumor modeling of mouse
tumor incidence in theNTP (1987) study 4-66
Table 4-18. Summary of unit risk estimates (per ug/m3) in animal bioassays 4-67
Table 4-19. ECoi, LECoi, and unit risk estimates for adult-only exposures* 4-70
Table 4-20. Calculation of ECoi for total cancer (incidence) risk from adult-only
exposure* 4-71
Table 4-21. Calculation of total cancer unit risk estimate from adult-only exposure* 4-71
Table 4-22. Adult-based unit risk estimates for use in ADAF calculations and risk
estimate calculations involving less-than-lifetime exposure scenarios 4-73
Table 4-23. Adult-based extra risk estimates per ppm based on adult-exposure-only
ECoiS (0.01/ECoi estimates)3 4-79
Table 4-24. Summary of key unit risk estimates from this assessment (see Section 4.7 for
risk estimates based on occupational exposure scenarios) 4-81
Table 4-25. Comparison of unit risk estimates3 4-83
Table 4-26. Extra risk estimates for lymphoid cancer in both sexes for various
occupational exposure levels3 4-89
Table 4-27. Extra risk estimates for breast cancer incidence in females for various
occupational exposure levelsa'b 4-92
This document is a draft for review purposes only and does not constitute Agency policy.
ix DRAFT—DO NOT CITE OR QUOTE
-------
LIST OF FIGURES
Figure 3-1. Metabolism of ethylene oxide 3-25
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.) 3-27
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)] 3-29
Figure 4-1. RR estimate for lymphoid cancer vs. mean exposure (with 15-year lag,
unadjusted for continuous exposure) 4-9
Figure 4-2. RR estimate for all lymphohematopoietic cancer vs. mean exposure (with
15-year lag, unadjusted for continuous exposure) 4-18
Figure 4-3. RR estimate for breast cancer mortality vs. mean exposure (with 20-year lag,
unadjusted for continuous exposure) 4-25
Figure 4-4. RR estimate for breast cancer incidence in full cohort vs. mean exposure
(with 15-year lag, unadjusted for continuous exposure) 4-33
Figure 4-5. RR estimate for breast cancer incidence in subcohort with interviews vs.
mean exposure (with 15-year lag, unadjusted for continuous exposure) 4-34
Figure 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 4-40
Figure 4-7. RR estimates for lymphoid cancer from occupational EtO exposures (with
15-year lag) 4-90
Figure 4-8. RR estimates for breast cancer incidence from occupational EtO exposures
(with 15-year lag) 4-95
This document is a draft for review purposes only and does not constitute Agency policy.
x 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
This document is a draft for review purposes only and does not constitute Agency policy.
xi 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
Kyle Steenland (under contract to EPA)
Rollins School of Public Health
Emory University
Atlanta, Georgia
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
This document is a draft for review purposes only and does not constitute Agency policy.
xii DRAFT—DO NOT CITE OR QUOTE
-------
Suryanarayana Vulimiri
National Center for Environmental Assessment
U.S. Environmental Protection Agency
Washington, DC
REVIEWERS
Earlier drafts of this document were provided for review to EPA scientists, interagency
reviewers from other federal agencies and the Executive Office of the President, the public, and
independent scientists external to EPA. Summaries and EPA's dispositions of the comments
received from the independent external peer reviewers and from the public are included in
Appendices H (for the 2006 external review draft) and L (for the 2013 public comment draft).
INTERNAL EPA REVIEWERS
Michele Burgess, OSWER
Deborah Burgin, OPEI
Carol Christensen, OPP/HED
Kerry Dearfield, ORD/OSP (no longer with EPA)
Joyce Donohue, OW
Rebecca Dzubow, AO/OCHP
Michael Firestone, AO/OCHP
Linnea Hansen, OPP
Karen Hogan, ORD/NCEA-IRIS
Ray Kent, OPP/HED
Aparna Koppikar, ORD/NCEA-W (retired)
Tim Leighton, OPP/AD
Tim McMahon, OPP/AD
David Miller, OPP/HED
Deirdre Murphy, OAR/ESD
Steve Nesnow, ORD/NHEERL (retired)
Marian Olsen, Region 2
Brenda Perkovich-Foos, AO/OCHP
Julian Preston, ORD/NHEERL (retired)
Santhini Ramasamy, OPP/HED, OW
Elissa Reaves, OPP/HED
Nancy Rios-Jafolla, Region 3
Tracey Woodruff, AO/NCEE (no longer with EPA)
The authors would like to acknowledge David Bussard, Jane Caldwell, Catherine
Gibbons, Linda Phillips, Julian Preston, Cheryl Scott, Maria Spassova, 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.
xiii DRAFT—DO NOT CITE OR QUOTE
-------
EXTERNAL PEER REVIEWERS
SCIENCE ADVISORY BOARD ETHYLENE OXIDE REVIEW PANEL (2007)
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.
xiv 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.
xv 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.
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 NIOSH study (Steenland et al., 2004;
5 Steenland et al., 2003). This study was selected for the derivation of risk estimates because it is
r\
6 a high-quality study, it is the largest of the available studies, and it has exposure estimates for
7 the individual workers from a high-quality exposure assessment. Multiple modeling approaches
8 were evaluated for the exposure-response data, including modeling the cancer response as a
9 function of either categorical exposures or continuous individual exposure levels. Preferred
10 approaches were defined for each cancer endpoint in consideration of both the statistical
11 properties and biological reasonableness of the resulting model forms (see Tables 4-4 and 4-12
12 for a summary of models investigated in this assessment for lymphoid cancer and breast cancer
13 incidence, respectively, and the considerations used in model selection).
14 Unit risk estimates based on the human data were first derived under the common
15 assumption that relative risk is independent of age. This assumption is later superseded by an
16 assumption of increased early-life susceptibility, and it is the unit risk estimates derived under
17 this latter assumption that are the ultimate estimates proposed in this assessment (presented
18 further below).
19 Under the assumption that relative risk is independent of age, an LECoi (lower 95%
20 confidence limit on the ECoi, the estimated effective concentration associated with 1% extra risk)
21 was calculated using a life-table analysis and linear modeling of the categorical Cox regression
22 analysis results for excess lymphoid cancer mortality [Steenland et al. (2004): additional results
23 for both sexes combined provided in Appendix D] excluding the highest exposure group to
24 mitigate the supralinearity of the exposure-response data.3 Linear low-dose extrapolation below
25 the range of observations is supported by the conclusion that a mutagenic mode of action is
26 operative in EtO carcinogenicity. Linear low-dose extrapolation from the LECoi for lymphoid
27 cancer mortality yielded a lifetime extra cancer unit risk estimate of 2.2 x 10 4 per ug/m3 (4.0
2 The NIOSH study (Steenland et al., 2003, 2004) was judged to be a "high-quality" study based on the attributes
discussed in Section 3.1 and in Section A.2.8 of Appendix A, including availability of individual worker exposure
estimates from a high-quality exposure assessment, cohort study design, large size, inclusion of males and females,
adequate follow-up, absence of any known confounding exposures, and use of internal comparisons. The breast
cancer incidence study using the subcohort of female workers with interviews had the additional attribute of
investigating and controlling for a number of breast cancer risk factors (Steenland et al., 2003).
3 A variety of continuous exposure model forms were considered; however, the exposure-response data for
lymphoid cancer are too supralinear (i.e., the exposure-response relationship is very steep in the lower exposure
range and then attenuates at higher exposures) to obtain useful models from the best-fitting continuous exposure
models (Section 4.1.1.2).
This document is a draft for review purposes only and does not constitute Agency policy.
1 -2 DRAFT—DO NOT CITE OR QUOTE
-------
1 x 10 4 per ppb)4 of continuous EtO exposure. Applying the same linear regression coefficient
2 and life-table analysis to background lymphoid cancer incidence rates and applying linear
3 low-dose extrapolation resulted in a preferred lifetime extra lymphoid cancer unit risk estimate
4 of 4.8 x 10 4 per ug/m3 (8.8 x 10 4 per ppb), as cancer incidence estimates are generally
5 preferred over mortality estimates.
6 Using the same approach,5 a unit risk estimate of 2.8 x 10 4 per ug/m3 (5.1 x 10 4 per
7 ppb) was derived from the breast cancer mortality results of the same epidemiology study
8 (Steenland et al., 2004). Breast cancer incidence risk estimates, on the other hand, were
9 calculated from the data from a breast cancer incidence study of the same occupational cohort
10 (Steenland et al., 2003), and, for these data, a two-piece linear spline model was used for the
11 exposure-response modeling (the two-piece linear spline model was the best-fitting model of the
12 continuous exposure models considered for the breast cancer incidence data; see Section 4.1.2.3).
13 Using the same life-table approach and linear low-dose extrapolation, a unit risk estimate of 9.5
14 x 10 4 per ug/m3 (1.7 x 10 3 per ppb) was obtained for breast cancer incidence. Again, the
15 incidence estimate is preferred over the mortality estimate. Combining the incidence risk
16 estimates for the two cancer types resulted in a total cancer unit risk estimate of 1.2 x io~3 per
17 ug/m3 (2.3 x 1Q~3 per ppb).6
18 Unit risk estimates were also derived from the three chronic rodent bioassays for EtO
19 reported in the literature. These estimates, ranging from 2.2 x 10 5 per ug/m3 to 4.6 x 10 5 per
20 ug/m3, are over an order of magnitude lower than the estimates based on human data. The
21 Agency takes the position that human data, if adequate data are available, provide a more
22 appropriate basis than rodent data for estimating population risks (U.S. EPA, 2005a), primarily
23 because uncertainties in extrapolating quantitative risks from rodents to humans are avoided.
24 Although there is a sizeable difference between the rodent-based and the human-based estimates,
25 the human data are from a large, high-quality study, with EtO exposure estimates for the
26 individual workers and little reported exposure to chemicals other than EtO. Therefore, the
27 estimates based on the human data are the preferred estimates for this assessment.
28 Because the weight of evidence supports a mutagenic mode of action for EtO
29 carcinogenicity, and as there are no chemical-specific data from which to assess early-life
30 susceptibility, increased early-life susceptibility should be assumed, according to EPA's
4Conversion equation: 1 ppm = 1830 ug/m3.
5 Several continuous exposure model forms were considered; however, the exposure-response data for breast cancer
mortality were too supralinear (i.e., the exposure-response relationship is very steep in the lower exposure range and
then attenuates at higher exposures) to obtain useful continuous exposure models (Section 4.1.2.2); thus, as for
lymphoid cancer, a linear regression of the categorical results, excluding the highest exposure group, was used.
6 The method used to derive the total cancer unit risk estimate involves estimating an upper bound on the sum of the
maximum likelihood estimates of risk; see Section 4.1.3.
This document is a draft for review purposes only and does not constitute Agency policy.
1 -3 DRAFT—DO NOT CITE OR QUOTE
-------
1 Supplemental Guidance for Assessing Susceptibility from Early-Life Exposure to
2 Carcinogens-hereinafter referred to as "EPA's Supplemental Guidance" (U.S. EPA, 2005b).
3 This mode-of-action-based assumption of increased early-life susceptibility supersedes the
4 assumption of age independence under which the human-data-based estimates presented above
5 were derived. Thus, using the same approach as for the estimates discussed above but initiating
6 exposure in the life-table analysis at age 16 instead of at birth, adult-exposure-only unit risk
7 estimates were calculated from the human data for selected exposure-response models under an
8 alternate assumption that relative risk is independent of age for adults, which represent the life
9 stage for which the data upon which the exposure-response modeling was conducted pertain.
10 These adult-exposure-only unit risk estimates were then rescaled to a 70-year basis for use in the
11 standard age-dependent adjustment factors (ADAF) calculations and risk estimate calculations
12 involving less-than-lifetime exposure scenarios. The resulting adult-based unit risk estimates
13 were 4.3 x 10 4 per ug/m3 (7.9 x 10~4 per ppb) for lymphoid cancer incidence, 8.2 x 10 4 per
14 ug/m3 (1.5 x 10 3 per ppb) for breast cancer incidence in females, and 1.1 x 10 3 per ug/m3 (2.0
15 x icf3 per ppb) for both cancer types combined. The adult-based unit risk estimates, which were
16 derived under an assumption of increased early-life susceptibility, supersede those presented
17 earlier that were derived under the assumption that RR is independent of age. When using the
18 adult-based unit risk estimates to estimate extra cancer risks for a given exposure scenario, the
19 standard ADAFs should be applied, in accordance with EPA's Supplemental Guidance (U.S.
20 EPA, 2005b). Applying the ADAFs to obtain a full lifetime total cancer unit risk estimate yields
21 1.8 x 1CT3 per ug/m3 (3.3 x 1CT3 per ppb), and the commensurate lifetime chronic (lower-bound)
22 exposure level of EtO corresponding to an increased cancer risk of 10 6 is 0.0006 ug/m3 (0.0003
23 ppb).
24 The major sources of uncertainty in the unit risk estimates derived from the human data
25 include the low-dose extrapolation, the retrospective exposure assessment conducted for the
26 epidemiology study, and the exposure-response modeling of the epidemiological data (see
27 Section 4.1.4 for a discussion of these and other sources of uncertainty in the unit risk estimates).
28 Although there are uncertainties in the unit risk estimate, confidence in the estimate is
29 relatively high. First, there is strong confidence in the hazard characterization of EtO as
30 "carcinogenic to humans," which is based on strong epidemiological evidence supplemented by
31 other lines of evidence. Second, the unit risk estimate is based on human data from a large,
32 high-quality epidemiology study with individual worker exposures estimated using a
33 high-quality regression model. Finally, the use of low-exposure linear extrapolation is strongly
34 supported by the conclusion that EtO carcinogenicity has a mutagenic mode of action.
This document is a draft for review purposes only and does not constitute Agency policy.
1 -4 DRAFT—DO NOT CITE OR QUOTE
-------
1 Confidence in the unit risk estimate is particularly high for the breast cancer component,
2 the largest contributor to the total cancer unit risk estimate, which is based on over 200 incident
3 cases for which the investigators had information on other potential breast cancer risk factors.
4 The selected model for the breast cancer incidence data was the best-fitting model of the models
5 investigated as well as the model which provided the best representation of the categorical
6 results, particularly in the lower exposure range of greatest relevance for the derivation of a unit
7 risk estimate. Alternate estimates calculated from other reasonable models suggest that a unit
8 risk estimate for breast cancer incidence fourfold lower (corresponding to a total cancer unit risk
9 estimate of twofold lower) is plausible; however, unit risk estimates notably lower than that are
10 considered unlikely from the available data (Section 4.1.2.3).
11 There is somewhat less, although still relatively high in general, confidence in the
12 lymphoid cancer component of the unit risk estimate because it is based on fewer events (40
13 lymphoid cancer deaths); incidence risk was estimated from mortality data; and the
14 exposure-response relationship is exceedingly supralinear, such that the best-fitting continuous
15 exposure models yield apparently implausibly steep low-exposure slopes. Although these
16 continuous exposure models provided statistically significant slope coefficients, there was low
17 confidence in such steep slopes, which, particularly for the two-piece spline model, are highly
18 dependent on a small number of cases in the low-exposure range. Thus, a linear regression
19 model of the categorical results for the lowest three quartiles was used to derive the unit risk
20 estimate for lymphoid cancer, and there was greater confidence in the more moderate slope
21 resulting from that model, although it was not statistically significant, because it was based on
22 more data and provided a good representation of the categorical results across this larger data
23 range in the lower exposure region. (The most suitable alternative model of those based on the
24 continuous exposure data yielded a unit risk estimate about twofold higher than the estimate
25 from the linear regression model of the categorical results.)7 So, while there is less confidence in
26 the lymphoid cancer unit risk estimate than in the breast cancer unit risk estimate, the lymphoid
27 cancer estimate is considered a reasonable estimate from the available data, and overall, there is
28 relatively high confidence in the total cancer unit risk estimate.
29 The unit risk estimate is intended to provide a reasonable upper bound on cancer risk.
30 The estimate was developed for environmental exposure levels (it is considered valid for
31 exposures up to 140 ug/m3 [75 ppb]) and is not applicable to higher-level exposures, such as may
7The most suitable alternative model of those based on the continuous exposure data was the two-piece log-linear
spline model with the knot at 1,600 ppm x days, although it was not statistically significant and it was not the two-
piece log-linear model with the maximum likelihood, which had the knot at 100 ppm x days, but rather one based on
a local maximum likelihood with respect to the changes in the value of the knot (see Section 4.1.1.2 and Figure D-3a
of Appendix D).
This document is a draft for review purposes only and does not constitute Agency policy.
1 -5 DRAFT—DO NOT CITE OR QUOTE
-------
1 occur occupationally, which appear to have a different exposure-response relationship.
2 However, occupational exposure levels of EtO are of concern to EPA when EtO is used as a
3 pesticide (e.g., sterilizing agent or fumigant). Therefore, this document also presents extra risk
4 estimates for the two cancer types for a range of occupational exposure scenarios (see
5 Section 4.7). Maximum likelihood estimates of the extra risk of lymphoid cancer and breast
6 cancer combined for the range of occupational exposure scenarios considered (i.e., 0.1 to 1 ppm
7 8-hour TWA for 35 years) ranged from 0.047 to 0.14. The overall uncertainty associated with
8 the extra risk estimates for occupational exposure scenarios is less than that associated with the
9 unit risk estimates for environmental exposures. The extra risk estimates are derived for
10 occupational exposure scenarios that yield cumulative exposures well within the range of the
11 exposures in the NIOSH study. Moreover, the NIOSH study is a study of sterilizer workers who
12 used EtO for the sterilization of medical supplies or spices (Steenland et al., 1991): thus, the
13 results are directly applicable to workers in these occupations, and these are among the
14 occupations of primary concern to EPA.
15 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.
1 -6 DRAFT—DO NOT CITE OR QUOTE
-------
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 datad — lymphoid cancer
incidence and breast cancer incidence in females
1.8 x 10 3
Adult-based unit risk estimates6
Total cancer risk based on human datad — 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 — range from lymphoid
cancer incidence estimates from two different models and
female breast cancer incidence estimates from three models
Total cancer incidence risk estimate from rodent data (female
mouse)
1.1 x 10 3
4.3 x ID'4
8.2 x ID'4
5.6 x lQ-4-1.4x 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 5 yr)f
0.047-0.14
1 aThese unit risk estimates are not intended for use with continuous lifetime exposure levels above about 140 ug/m .
2 See Section 4.7 for risk estimates based on occupational exposure scenarios. Preferred estimates are in bold.
3 bTo convert unit risk estimates to (ppm)"1, multiply the (ug/m3)"1 estimates by 1,830 (ug/m3)/ppm
4 'Because the weight of evidence supports a mutagenic mode of action for EtO carcinogenicity, and because of the
5 lack of chemical-specific data, EPA assumes increased early-life susceptibility and recommends the application of
6 ADAFs, in accordance with EPA's Supplemental Guidance (U.S. EPA 2005b). for exposure scenarios that include
7 early-life exposures. For the full lifetime (upper bound) unit risk estimate presented here, ADAFs have been
8 applied, as described in Section 4.4.
9 technically, this unit risk estimate reflects the total (upper bound) cancer risk to females and not to the general
10 population because the breast cancer risk only applies to females. As a practical matter for regulatory purposes,
11 however, females comprise roughly half the general population and this unit risk estimate enables risk managers to
12 evaluate the individual risk for this substantial population group. For the purposes of estimating numbers of cancer
13 cases attributable to specific exposure levels, e.g., for benefits analyses, it would be more appropriate to use the
14 cancer-specific unit risk estimates (or central tendency estimates), taking sex into account.
This document is a draft for review purposes only and does not constitute Agency policy.
1 -7 DRAFT—DO NOT CITE OR QUOTE
-------
1 Table 1-1. Summary of major findings (continued)
2
3 eThese (upper bound) unit risk estimates are intended for use in ADAF calculations and less-than-lifetime adult
4 exposure scenarios (U.S. EPA. 2005b). Note that these are not the same as the unit risk estimates derived directly
5 from the human data in Section 4.1 under the assumption that RRs are independent of age. Under that assumption,
6 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
7 cancer incidence, and 1.2 x 10~3 per ug/m3 for the combined cancer incidence risk from those two cancers. See
8 Section 4.4 for the derivation of the adult-based unit risk estimates.
9 technically, these sums would reflect the total cancer risk to females and not a mixed-sex workforce because the
10 breast cancer risk only applies to females. As a practical matter for regulatory purposes, however, females typically
11 comprise a substantial proportion of the sterilizer workforce and summing these extra risk estimates enables risk
12 managers to evaluate the individual risk for this substantial workforce group. In a situation in which the workforce
13 of concern is comprised predominantly of males, it might be appropriate to use a sex-weighted sum of the extra risks
14 from the two cancer types (see Section 4.7 for the cancer-specific extra risk estimates).
This document is a draft for review purposes only and does not constitute Agency policy.
1 -8 DRAFT—DO NOT CITE OR QUOTE
-------
1 2. INTRODUCTION
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 certain other items and as a fumigating agent for
5 spices. The largest sources of human exposure are in occupations involving contact with the gas
6 in plants that manufacture or use EtO and in hospitals that sterilize medical equipment. EtO can
7 also be inhaled by residents living near production or sterilizing/fumigating facilities. Based on
8 EPA's 2005 National-scale Air Toxics Assessment (NATA) data, the average environmental
9 exposure concentration from all sources (including concentrations near known sources) in the
10 United States is 0.0062 ug/m3; the average background concentration excluding concentrations
11 near sources of EtO is 0.0044 ug/m3 (NATA 2005 data,
12 http://www.epa.gov/ttn/atw/nata2005/tables.html).
13 EPA offices with an interest in EtO include the Office of Air and Radiation and the
14 Office of Pesticide Programs. The Office of Air and Radiation has an interest because EtO is 1
15 of the 188 hazardous air pollutants listed in the 1990 Clean Air Act Amendments. The Office of
16 Pesticide Programs has an interest in both environmental and occupational exposures resulting
17 from the sterilization uses of EtO because EPA is responsible for pesticide labeling and
18 registration decisions under the Federal Insecticide, Fungicide, and Rodenticide Act (FIFRA).
19 The purpose of this document is to provide scientific support and rationale for the hazard
20 and dose-response assessment in IRIS pertaining to carcinogenicity from chronic inhalation
21 exposure to ethylene oxide (EtO) (CASRN 75-21-8). It is not intended to be a comprehensive
22 treatise on the chemical or toxicological nature of EtO. In general, this IRIS Carcinogenicity
23 Assessment provides information on the carcinogenic hazard potential of EtO and quantitative
24 estimates of risk from inhalation exposure. The information includes a weight-of-evidence
25 judgment of the likelihood that the agent is a human carcinogen and the conditions under which
26 the carcinogenic effects may be expressed. Quantitative risk estimates for inhalation exposure
27 (inhalation unit risks) are derived. The definition of an inhalation unit risk is a plausible upper
28 bound on the estimate of risk per ug/m3 air breathed.
29 Development of the hazard identification and dose-response assessments for EtO has
30 followed the general guidelines for risk assessment as set forth by the National Research Council
31 (NRC, 1983). United States Environmental Protection Agency (U.S. EPA) Guidelines and Risk
32 Assessment Forum Technical Panel Reports that were used in the development of this
3 3 assessment include the following: Guidelines for Mutagenicity Risk Assessment (U.S. EPA,
34 1986), Methods for Derivation of Inhalation Reference Concentrations and Application of
This document is a draft for review purposes only and does not constitute Agency policy.
2-1 DRAFT—DO NOT CITE OR QUOTE
-------
1 Inhalation Dosimetry (U.S. EPA, 1994), Benchmark Dose Technical Guidance (U.S. EPA,
2 2012), Science Policy Council Handbook: Risk Characterization (U.S. EPA, 2000), Guidelines
3 for Carcinogen Risk Assessment U.S. EPA (2005a), Supplemental Guidance for Assessing
4 Susceptibility from Early-Life Exposure to Carcinogens (U.S. EPA, 2005b), and Science Policy
5 Council Handbook: Peer Review (U.S. EPA, 2006b).
6 An earlier external review draft of this carcinogen! city assessment (U.S. EPA, 2006a)
7 was peer reviewed by a panel of EPA's Science Advisory Board (SAB) in 2007 (SAB, 2007).
8 See Appendix H for a summary and EPA's disposition of SAB and public comments on the 2006
9 external review draft. In response to comments from that SAB review, EPA conducted extensive
10 new exposure-response modeling of certain epidemiologic data. In addition, EPA has updated
11 the assessment to reflect new literature through May 2013; this new literature did not
12 substantively impact the conclusions of the assessment (Appendix J). In July 2013, EPA
13 released a revised draft for public comment, and that draft assessment was discussed at EPA's
14 December 2013 IRIS Bimonthly Public Meeting. EPA has also addressed the public comments
15 that were received on the July 2013 draft (Appendix L). This newly revised 2014 external
16 review draft is being released for additional external peer review to receive comments primarily
17 on the expanded exposure-response modeling of the epidemiologic data.
18
19 2.1. LITERATURE IDENTIFICATION
20 The literature search strategy first employed for this assessment was based on the
21 Chemical Abstracts Service Registry Number (CASRN) and at least one common name. Any
22 pertinent scientific information submitted by the public to the IRIS Submission Desk was also
23 considered in the development of this document. References were added after the first external
24 peer review in response to the reviewers' and public comments. In preparation for this second
25 external peer review, a well-documented systematic literature search was conducted for the time
26 frame from January 2006 to May 2013. This systematic literature search is described in Section
27 J. 1 of Appendix J. Based on this search, 56 references were identified as potentially relevant to
28 the EtO assessment. None of the new studies would impact the assessment's major conclusions.
29 Nonetheless, two new studies of high pertinence to the assessment were identified, and these
30 studies are reviewed in Section J.2 of Appendix J for transparency and completeness. Reviews
31 of an additional two new studies of high pertinence were added to Appendix J to address public
32 comments received in October and December of 2013 (Section J.3); these studies similarly do not
33 impact the assessment's major conclusions.
This document is a draft for review purposes only and does not constitute Agency policy.
2-2 DRAFT—DO NOT CITE OR QUOTE
-------
1 See Appendix I for a list of all 131 references added after the first external peer review.
2 The references considered and cited in this document, including abstracts, can be found on the Health
3 and Environmental Research Online (HERO) website.8
4
5 2.2. NRC RECOMMENDATIONS OF 2011
6 On 23 December 2011, the Consolidated Appropriations Act, 2012, was signed into law.9
7 The report language included direction to EPA for the IRIS Program related to recommendations
8 provided by the National Research Council (NRC) in their 2011 review of EPA's draft IRIS
9 assessment of formaldehyde. Consistent with the direction provided by Congress,
10 documentation of how the recommendations from Chapter 7 of the NRC (2011) report have been
11 implemented in this assessment is provided in Appendix K. This documentation also includes an
12 explanation for why certain recommendations were not incorporated.
13 In brief, the EtO assessment was one of a group of chemical assessments that had already
14 completed external peer review at the time the 2011 NRC recommendations were released. For
15 this group of assessments, EPA focused on a subset of the short-term recommendations, such as
16 streamlining documents, increasing transparency and clarity, and using more tables, figures, and
17 appendices to present information and data in assessments. While the EtO assessment does not
18 incorporate recent revisions to the IRIS assessment format recommended in the 2011 NRC
19 recommendations (and the more recent 2014 NRC Review of the IRIS Process), such as the
20 inclusion of a standard Preamble and the revised chapter structure, and does not fully implement
21 the longer-term NRC recommendations, as discussed in Appendix K, the assessment is
22 consistent with the goal that assessments should provide a scientifically sound and transparent
23 evaluation of the relevant scientific literature and presentation of the analyses performed.
24 For general information about this assessment or other questions relating to IRIS, the
25 reader is referred to EPA's IRIS Hotline at (202) 566-1676 (phone), (202) 566-1749 (fax), or
26 hotline.iris@epa.gov (email address).
8HERO 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 1,000,000 scientific articles from the peer-reviewed
literature. New studies are added continuously to HERO.
9Pub. L. No. 112-74, Consolidated Appropriations Act, 2012.
This document is a draft for review purposes only and does not constitute Agency policy.
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.
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 etal. (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.
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: Tetaetal.. 1999; Benson and Teta, 1993; TetaetaL 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 etal.. 1990). through 1988 in the TetaetaL (1993). Benson and Teta (1993). and Teta et al.
15 (1999) studies, and through 2003 in the Swaen et al. (2009) and Valdez-Flores et al. (2010)
16 studies. A large-scale industrial hygiene survey and monitoring of EtO concentrations was
17 carried out in 1976, at which time EtO was in 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 (TetaetaL, 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.
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 Tetaetal. (1993)1. 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 Tetaetal. (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, Tetaetal. (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 Tetaetal. (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.
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 Tetaetal. (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.
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.
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.10 In internal categorical analyses, statistically significant odds ratios (ORs) were observed
5 in 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
10The 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.
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 Tetaet 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.
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 Olsenet 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.
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 subjects with breast cancer 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.
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 cancer,11 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.
"Hungarian 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.
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: Stavner 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 etal., 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.
3-12 DRAFT—DO NOT CITE OR QUOTE
-------
Table 3-1. Epidemiological studies of ethylene oxide and human cancer—lymphohematopoietic cancer results"
Study/Population/
Industry
Hosstedt (1988) and
Hosstedtetal. (1986).
Sterilizers, production
workers, Sweden.
Cossonetal. (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
Greenbers 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; 12.0% of cohort
had died (85 deaths).
Exposure to other chemicals.
Short follow-up; 19.6% of cohort had
died (565 deaths).
Exposure to other chemicals.
Insufficient follow-up; 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
o
a
I
a,
8"
O 5
1
3
o
H
W
o
c
o
H
W
-------
Table 3-1. Epidemiological studies of ethylene oxide and human cancer—lymphohematopoietic cancer results"
(continued)
§•
rj
s
3
§
Study/Population/
Industry
Number of
subjects
Lymphohematopoietic cancer results
Comments
s
•3
o
I
o
a
1
%
s
1
Swaen et al. (2009).
Update of Teta et al.
(1993) rOreenberg 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 (17 deaths)
or leukemia (11 deaths) cancer categories for continuous cumulative
exposure.
Small cohort; long follow-up time
(50.8% had died; 1,048 deaths).
Crude exposure assessment, especially
for the early time periods.
Exposure to other chemicals.
2
O
^
Steenland et al. (2004).
Update of Steenland et
al. (1991). Stavneretal.
(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
V
O
c
o
H
W
-------
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.
Hasmar et al. (1995) and
Hasmaretal. (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.
Number of
subjects
1,971 men
2,170
(861 men,
1,309
women)
1,132
(204 men,
928 women)
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
Comments
Insufficient follow-up; 3.9% of cohort
had died (76 deaths).
Exposure to other chemicals.
Short follow-up period (only 40 cancer
cases).
(See Section J.2.2 and Table J-2 of
Appendix J for a more recent follow-
up of this cohort.)
Short follow-up period and small cohort
(only 28 cancer cases).
§•
§
I
o
a
I
a,
8"
O 5
1
3
o
H
W
o
c
o
H
W
-------
o
c
o
H
W
Table 3-1. Epidemiological studies of ethylene oxide and human cancer—lymphohematopoietic cancer results"
(continued)
Study/Population/
Industry
Swaenetal. (1996).
Nested case-control
study; cases and controls
from a large chemical
production plant,
Belgium.
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
10 cases of
Hodgkin
lymphoma (3
exposed; 7
confirmed)
and 200
controls; all
male
1,361 men
299 women
Lymphohematopoietic cancer results
cancer OR (95% CI)
Hodgkin lymphoma (ICD 201) 8.5 (1.4, 40)
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; expected number not reported.
Comments
Hypothesis-generating study to
investigate a cluster of Hodgkin
lymphomas observed at a chemical plant.
Exposure to other chemicals.
Short follow-up and small cohort; 22.0%
had died; 300 deaths.
Exposure to other chemicals.
Short follow-up period and small cohort
(11 cancer deaths).
Possible exposure to natural radium,
which permeates the region.
§•
§
I
ON
o
a
I
a,
8"
O 5
3
o
H
W
1
"Extracted from Table A-5 of Appendix A, with addition of some summary results (e.g., SMRs); see Table A-5 and Appendix A for more study details.
bCalculated by EPA assuming Poisson distribution.
ICONS: ICD codes not specified; NR: not reported
-------
Table 3-2. Summary of epidemiological results on EtO and female breast cancer (all sterilizer workers)"
This document is a draft for review purposes only and does not constitute Agency policy.
3-17 DRAFT— DO NOT CITE OR QUOTE
Study
Hogstedt et al.
(1986) and
Hoestedt(1988)
Swedish
incidence and
mortality study
Coeeon 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; nested
within Steenland
et al. (2004)
cohort
Number of
Women
170
1,011 women
hospital
workers
9,908
7,576
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)
(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 .74 (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);/? = 0.03 with log cumulative exposure (15-yr lag).
Comments
8 deaths (7 from cancer) had
occurred among the women;
breakdown by cancer type not
reported.
1 1 breast cancer deaths.
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.
-------
Table 3-2. Summary of epidemiological results on EtO and female breast cancer (all sterilizer workers)3
(continued)
§•
§
I
Study
Number of
Women
Breast Cancer Results
Comments
Hagmar et al.
(1995) and
Hagmar et al.
(1991)
Swedish cancer
incidence study
1,309
5 cases vs. 10.8 expected SIR = 0.46 (95% CI: 0.15, 1.08).
5 cases.
(See Section J.2.2 and Table J-2 of
Appendix J for a more recent
follow-up of this cohort.)
Norman et al.
(1995)
U.S. cancer
incidence study
928
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.
12 cases.
00
o
a
I
a,
8"
Kardos et al.
299
(2003)
Hungarian
mortality study
11 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 #5 in
Section 3.1).
3 breast cancer deaths.
O 5
1
3
o
H
W
"Extracted from Table A-5 of Appendix A; see Table A-5 and Appendix A for more study details, and also Table 3-1 above.
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 large NIOSH studies, which found
19 positive exposure-response relationships for both breast cancer incidence and mortality (319
20 incident breast cancer cases; 103 breast cancer deaths). Of the five other studies that included
21 females, none approached the size of the NIOSH studies in terms of breast cancer data - the
22 study with the next largest breast cancer database had only 12 cases. Nonetheless, two of the
23 five other studies were supportive of an increased risk of breast cancer.12
24
25 3.2. EVIDENCE OF CANCER IN LABORATORY ANIMALS
26 The International Agency for Research on Cancer (IARC) monograph (IARC, 1994b) has
27 summarized the rodent studies of carcinogenicity, and Health Canada (2001) has used this
28 information to derive levels of concern for human exposure. EPA concludes that the IARC
29 summary of the key studies is valid and is not aware of any animal cancer bioassays that have
30 been published since 1994. The Ethylene Oxide Industry Council (EPIC, 2001) also reviewed
31 the same studies and did not cite additional studies. The qualitative results are described here
32 and the incidence data are tabulated in the unit risk derivation section of this document.
12 A more recent follow-up of a third study of the five other studies that was published after the cutoff date for
literature inclusion in the main text of this assessment is also supportive of an increased risk of breast cancer (see
Section J.2.2 and Table J-2 of Appendix J).
This document is a draft for review purposes only and does not constitute Agency policy.
3-19 DRAFT—DO NOT CITE OR QUOTE
-------
1 One study of oral administration in rats has been published; there are no oral studies in
2 mice. Dunkelberg (1982) administered EtO in vegetable oil to groups of 50 female
3 Sprague-Dawley rats by gastric intubation twice weekly for 150 weeks. There were two control
4 groups (untreated and oil gavage) and two treated groups (7.5 and 30 mg/kg-day). A
5 dose-dependent increase in the incidence of malignant tumors in the forestomach was observed
6 in the treated groups (8/50 and 31/50 in the low- and high-dose groups, respectively). Of the
7 39 tumors, 37 were squamous cell carcinomas, and metastases to other organs were common in
8 these animals. This study was not evaluated quantitatively because oral risk estimates are
9 beyond the scope of this document.
10 One inhalation assay was reported in mice (NTP, 1987) and two inhalation assays were
11 reported in rats [(Lynch et al.. 1984a: Lynch etal.. 1984b) in males; (Garman et al.. 1986. 1985:
12 Snellings et al., 1984) in both males and femalesl. In the National Toxicology Program (NTP)
13 mouse bioassay (NTP, 1987), groups of 50 male and 50 female B6C3Fi mice were exposed to
14 EtO via inhalation at concentrations of 0, 50, and 100 ppm for 6 hours per day, 5 days per week,
15 for 102 weeks. Mean body weights were similar for treated and control animals, and there was
16 no decrease in survival associated with treatment. A concentration-dependent increase in the
17 incidence of tumors at several sites was observed in both sexes. These data are summarized in
18 Table 3-3. Males had carcinomas and adenomas in the lung. Females had carcinomas and
19 adenomas in the lung, malignant lymphomas, adenocarcinomas in the uterus, and
20 adenocarcinomas in the mammary glands. The NTP also reports that both sexes had dose-related
21 increased incidences of cystadenomas of the Harderian glands, but these are benign lesions and
22 are not considered further here.
23 In the Lynch et al. (Lynch et al., 1984a; Lynch et al., 1984b) bioassay in male Fischer
24 344 (F344) rats, groups of 80 animals were exposed to EtO via inhalation at concentrations of 0,
25 50, and 100 ppm for 7 hours per day, 5 days per week, for 2 years. Mean body weights were
26 statistically significantly decreased in both treated groups compared with controls (p < 0.05).
27 Increased mortality was observed in the treated groups, and the increase was statistically
28 significant in the 100-ppm exposure group (p < 0.01). Lynch et al. (1984a) suggest that survival
29 was affected by a pulmonary infection alone and in combination with EtO exposure.
30 Concentration-dependent increases in the incidence of mononuclear cell leukemia in the spleen,
31 peritoneal mesothelioma in the testes, and glioma in the brain were observed (see Table 3-4).
32 The fact that the increased incidence of mononuclear cell leukemia was statistically significant in
33 the low-exposure group, but not in the high-exposure group, is probably attributable to the
34 increased mortality in the high-exposure group. The increased incidence in just the terminal kill
35 rats in the 100-ppm group was statistically significant compared with controls.
36
This document is a draft for review purposes only and does not constitute Agency policy.
3-20 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
Table 3-3. Tumor incidence data in National Toxicology Program Study of
B6C3Fi mice NTP (1987)a and exposure-response modeling results'5
Gender/tumor type
EtO concentration
(time-weighted average)0
0 ppm
50 ppm
(16.3 mg/m3)
100 ppm
(32.7 mg/m3)
EC10
(LEC10)d,
(mg/m3)
Unit risk
(0.1/LEdo)
(per mg/m3)
Males
Lung adenomas plus
Carcinomas
11/49
19/49
26/49e
6.94
(4.51)
2.22 x 10'2
Females
Lung adenomas plus
Carcinomas
Malignant
Lymphoma
Uterine
Carcinoma
Mammary
carcinoma1
2/44
9/44
0/44
1/44
5/44
6/44
1/44
8/44g
22/49f
22/498
5/49h
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
17
Incidence data were adjusted by EPA by eliminating the animals that died prior to the occurrence of the first tumor
or prior to 52 weeks, whichever was earlier.
bStatistical analyses and exposure-response modeling were conducted by EPA.
0 Adjusted to continuous exposure from experimental exposure conditions of 6 hr/d, 5 d/wk; 1 ppm = 1.83 mg/m3.
dCalculated by EPA using Tox_Risk program.
ep < 0.01 (pairwise Fisher's exact test).
{p < 0.001 (pairwise Fisher's exact test).
sp < 0.05 (pairwise Fisher's exact test).
hp = 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 NIH-07 diet.
'Highest dose was deleted in order to fit a model to the dose-response data.
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 and exposure-response modeling results
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 by EPA to continuous exposure from experimental exposure conditions of 7 hr/d, 5 d/wk; 1 ppm =1.83
mg/m3.
bCalculated by EPA 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).
In the bioassay conducted by Snellings et al. (1984), 120 male and 120 female F344 rats
in each sex and dose group were exposed to EtO via inhalation at concentrations of 0 (2 control
groups of 120 rats of each sex were used), 10, 33, and 100 ppm for 6 hours per day, 5 days per
week, for 2 years, with scheduled kills at 6 (10 rats per group), 12 (10 rats per group), and
18 (20 rats per group) months. Significant decreases in mean body weight were observed in the
100-ppm exposure group in males and in the 100-ppm and 33-ppm exposure groups in females.
During the 15* month of exposure, an outbreak of viral sialodacryoadenitis occurred, resulting in
the deaths of 1-5 animals per group. Snellings etal. (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 22" or 23r 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) and Lynch et al. (1984b). Snellings
This document is a draft for review purposes only and does not constitute Agency policy.
3 -22 DRAFT—DO NOT CITE OR QUOTE
-------
1 et al. (1984) only report incidences (of incidental and nonincidental primary tumors for all
2 exposure groups) for the 24-month (terminal) kill. However, in their paper they state that
3 significant findings for the mononuclear cell leukemias were also obtained when all rats were
4 included and that a mortality-adjusted trend analysis yielded positive findings for the
5 EtO-exposed females (p < 0.005) and males (p < 0.05). Similarly, Snellings et al. (1984) report
6 that when male rats with unscheduled deaths were included in the analysis of peritoneal
7 mesotheliomas, it appeared that EtO exposure was associated with earlier tumor occurrence, and
8 a mortality-adjusted trend analysis yielded a significant positive trend (p < 0.005). In later
9 publications describing brain tumors (Garman et al., 1986, 1985), both males and females had a
10 concentration-dependent increased incidence of brain tumors (see Table 3-5). Garman et al.
11 (1986) and Garman et al. (1985) report incidences including all rats from the 18- and 24-month
12 kills and all rats found dead or killed moribund. The earliest brain tumors were observed in rats
13 killed at 18 months.
14
15 3.2.1. Conclusions Regarding the Evidence of Cancer in Laboratory Animals
16 In conclusion, EtO causes cancer in laboratory animals. After inhalation exposure to
17 EtO, statistically significant increased incidences of cancer have been observed in both rats and
18 mice, in both males and females, and in multiple tissues (lung, mammary gland, uterus,
19 lymphoid cells, brain, tunica vaginalis testis). In addition, one oral study in rats has been
20 conducted, and a significant dose-dependent increase in carcinomas of the forestomach was
21 reported.
22
23 3.3. SUPPORTING EVIDENCE
24 3.3.1. Metabolism and Kinetics
25 Information on the kinetics and metabolism of EtO has been derived primarily from
26 studies conducted with laboratory animals exposed via inhalation, although some limited data
27 from humans have been identified. Details are available in several reviews (Fennell and Brown,
28 2001: Csanadv et al.. 2000: Brown etal.. 1998: Brown etal.. 1996).
29 Following inhalation, EtO is absorbed efficiently into the blood and rapidly distributed to
30 all organs and tissues. EtO is metabolized primarily by two pathways (see Figure 3-1):
31 (1) hydrolysis to ethylene glycol (1,2-ethanediol), with subsequent conversion to oxalic acid,
32 formic acid, and carbon dioxide; and (2) glutathione conjugation and the formation of
33 ,S'-(2-hydroxyethyl)cysteine and N-acetylated derivatives (WHO, 2003). From the available data,
34 the route involving conjugation with glutathione appears to predominate in mice; in larger
35 species (including humans), the conversion of EtO is primarily via hydrolysis through ethylene
This document is a draft for review purposes only and does not constitute Agency policy.
3-23 DRAFT—DO NOT CITE OR QUOTE
-------
Table 3-5. Tumor incidence data in Snellings et al. (1984) and Garman et al. (1985) reports on F344 rats" and
exposure-response modeling results'*
Gender/tumor type
Concentration (time-weighted average)0
Oppmd
lOppm
(3.27 mg/m3)
33 ppm
(10.8 mg/m3)
100 ppm
(32.7 mg/m3)
EC10
(LEC10)e
(mg/m3)
Unit risk (0.1/LEC10)
(per mg/m3)
Males
Splenic mononuclear cell
leukemia
Testicular peritoneal
mesothelioma
Primary brain tumors
13/97
(13%)f
2/97
(2.1%)
1/181
(0.55%)
9/51
(18%)
2/51
(3.9%)
1/92
(1.1%)
12/39g
(32%)
4/39
(10%)
5/85g
(5.9%)
9/30g
(30%)
4/30g
(13%)
7/87h
(8.1%)
12.3
(6.43)
22.3
(11.6)
36.1
(22.3)
1.56 x 10'2
8.66 x 10'3
4.5 x 1(T3
Females
Splenic mononuclear cell
leukemia
Primary brain tumors
11/116
(9.5%)
1/188
(0.53%)
ll/54g
(21%)
1/94
(1.1%)
14/48h
(30%)
3/92
(3.3%)
15/261
(58%)
4/808
(5%)
4.46
(3.1)
63.8
(32.6)
3.23 x ID'2
3.07 x ID'3
§•
§
I
o
a
I
a,
8"
O 5
1
3
o
H
W
Denominators refer to the number of animals for which histopathological diagnosis was performed. For brain tumors, Garman etal. (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.
bStatistical analyses and exposure-response modeling were conducted by EPA.
°Adjusted by EPA to continuous exposure from experimental exposure conditions of 6 hr/d, 5 d/wk; 1 ppm =1.83 mg/m3.
dResults for both control groups combined.
Calculated by EPA using Tox_Risk program.
fNumbers in parentheses indicate percentage incidence values.
sp < 0.05 (pairwise Fisher's exact test).
hp < 0.01 (pairwise Fisher's exact test).
lp < 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
+
CO,
CO2HCO2H
Oxalic acid
1
2
3
4
5
6
7
Figure 3-1. Metabolism of ethylene oxide.
glycol. Because EtO is an epoxide capable of reacting directly with cellular macromolecules,
both pathways are considered to be detoxifying.
Among rodent species, there are clear quantitative differences in metabolic rates. The
rate of clearance of EtO from the blood, brain, muscle, and testes was measured by Brown et al.
(1998) and Brown et al. (1996). Clearance rates were nearly identical across blood and other
tissues. Following a 4-hour inhalation exposure to 100 ppm EtO in mice and rats, the average
This document is a draft for review purposes only and does not constitute Agency policy.
3-25 DRAFT—DO NOT CITE OR QUOTE
-------
1 blood elimination half-lives ranged from 2.4 to 3.2 minutes in mice and 11 to 14 minutes in rats.
2 The elimination half-life in humans is 42 minutes (Filser et al., 1992), and the half-life in salt
3 water is 4 days (IARC, 1994b).
4 In a more detailed study in mice, Brown et al. (1998) measured EtO concentrations in
5 mice after 4-hour inhalation exposures at 0, 50, 100, 200, 300, or 400 ppm. They found that
6 blood EtO concentration increased linearly with inhaled concentrations of less than 200 ppm, but
7 above 200 ppm the blood concentration increased more rapidly. In addition, glutathione levels in
8 liver, lung, kidney, and testes decreased as exposures increased above 200 ppm. The
9 investigators interpreted this, along with other information, to mean that at low concentrations
10 the metabolism and disappearance of EtO is primarily a result of glutathione conjugation, but at
11 higher concentrations, when tissue glutathione begins to be depleted, the elimination occurs via a
12 slower nonenzymatic hydrolysis process, leading to a greater-than-linear increase in blood EtO
13 concentration.
14 Fennell and Brown (2001) constructed physiologically based pharmacokinetic (PBPK)
15 models of uptake and metabolism in mice, rats, and humans, based on previous studies. They
16 reported that the models adequately predicted blood and tissue EtO concentrations in rats and
17 mice, with the exception of the testes, and blood EtO concentrations in humans. Modeling
18 6-hour inhalation exposures yielded simulated blood peak concentrations and areas under the
19 curve (AUCs) that are similar for mice, rats, and humans (human levels are within about 15% of
20 rat and mouse levels; see Figure 3-2). In other words, exposure to a given EtO concentration in
21 air results in similar predicted blood EtO AUCs for mice, rats, and humans.
22 These studies show that tissue concentrations in mice, rats, and humans exposed to a
23 particular air concentration of EtO are approximately equal and that they are linearly related to
24 inhalation concentration, at least in the range of exposures used in the rodent cancer bioassays
25 (i.e., 100 ppm and below).
26
27 3.3.2. Protein Adducts
28 EtO forms DNA (see Section 3.3.3.1) and hemoglobin adducts within tissues throughout
29 the body (Walker et al., 1992a: Walker et al., 1992b). Formation of hemoglobin adducts has
30 been used as a measure of exposure to EtO. The main sites of alkylation are cysteine, histidine,
31 and the N-terminal valine; however, for analytical reasons, the N-(2-hydroxyethyl)valine adduct
32 is generally preferred for measurements (Walker et al., 1990). Walker et al. (1992b) reported
33 measurements of this hemoglobin adduct and showed how the concentration of the adducts
34 changes according to the dynamics of red blood cell turnover. Walker et al. (1992b) measured
35 hemoglobin adduct formation in mice and rats exposed to 0, 3, 10, 33, 100, and 300 (rats only)
This document is a draft for review purposes only and does not constitute Agency policy.
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
18
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.)
ppm of EtO (6 hours/day, 5 days/week, for 4 weeks). Response was linear in both species up to
33 ppm, after which the slope significantly increased. The exposure-related decrease in
glutathione concentration in liver, lung, and other tissues observed by Brown et al. (1998) in
mice is a plausible explanation for the increasing rate of hemoglobin adduct formation at higher
exposures.
In humans, hemoglobin adducts can be used as biomarkers of recent exposure to EtO
(IARC, 2008: Boogaard, 2002; IARC, 1994b), and several studies have reported
exposure-response relationships between hemoglobin adduct levels and EtO exposure levels [e.g.
(van Sittert et al., 1993; Schulte et al., 1992)]. Hemoglobin adducts are good general indicators
of exposure because they are stable (DNA adducts, on the other hand, may be repaired
This document is a draft for review purposes only and does not constitute Agency policy.
3-27 DRAFT—DO NOT CITE OR QUOTE
-------
1 or fixed as mutations and hence are less reliable measures of exposure). However, Fost et al.
2 (1991) noted that human erythrocytes showed marked interindividual differences in the amounts
3 of EtO bound to hemoglobin, and Yong et al. (2001) reported that levels of
4 N-(2-hydroxyethyl)valine were approximately twofold greater in persons with a G<5Tr7-null
5 genotype than in those with positive genotypes. Endogenous ethylene oxide (see
6 Section 3.3.3.1) also contributes to hemoglobin adduct levels, making it more difficult to detect
7 the impacts of low levels of exogenous EtO exposure. In addition, Walker etal. (1993) reported
8 that hemoglobin adducts in mice and rats were lost at a greater rate than would be predicted by
9 the erythrocyte life span.
10
11 3.3.3. Genotoxicity
12 Since the first report of EtO induction of sex-linked recessive lethals in Drosophila
13 (Rapoport, 1948), numerous papers have been published on the positive genotoxic activity in
14 biological systems, spanning a wide range of assay systems, from bacteriophage to higher plants
15 and animals. Figure 3-3 shows the 203 test entries in the EPA Genetic Activity Profile database
16 in 2001. In prokaryotes and lower eukaryotes, EtO induced DNA damage and gene mutations in
17 bacteria, yeast, and fungi and gene conversions in yeast. In mammalian cells (from in vitro
18 and/or in vivo exposures), EtO-induced effects include unscheduled DNA synthesis, gene
19 mutations, sister chromatid exchanges (SCEs), micronuclei, and chromosomal aberrations.
20 Genotoxicity, in particular increased levels of SCEs and chromosomal aberrations, has also been
21 observed in blood cells of workers occupationally exposed to EtO. Several publications contain
22 details of earlier genetic toxicity studies [e.g. (IARC. 2008: Kolman et al.. 2002: Thier and Bolt
23 2000: Nataraian et al.. 1995: Preston et al.. 1995: IARC. 1994b: Dellarco et al.. 1990: Ehrenberg
24 and Hussain, 1981)1. This review briefly summarizes the evidence of the genotoxic potential of
25 EtO, focusing primarily on recently published studies that provide information on the mode of
26 action of EtO (see Appendix C for more details from some individual studies).
27
28 3.3.3.1. DNA Adducts
29 EtO is a direct-acting 8^2 (substitution-nucleophilic-bimolecular)-type monofunctional
30 alkylating agent that forms adducts with cellular macromolecules such as proteins (e.g.,
31 hemoglobin, see Section 3.3.2) and DNA (Pauwels and Veulemans, 1998). Alkylating agents
32 may produce a variety of different DNA alkylation products (Beranek, 1990) in varying
33 proportions, depending primarily on the electrophilic properties of the agent. Reactivity of an
34 alkylating agent is estimated by its Swain-Scott substrate constant (s-value), which ranges from
35 0 to 1, and EtO has a high s-value of 0.96 (Beranek, 1990: Golberg, 1986: Warwick, 1963).
This document is a draft for review purposes only and does not constitute Agency policy.
3-28 DRAFT—DO NOT CITE OR QUOTE
-------
ETHYLENE OXIDE
75-21-8
GENETIC ACTIVITY PROFILE
NCEA-01
2001
2
3
4
5
6
7
8
9
10
11
1 2
13
14
1 5
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
s
A
0
_ S
A
5
E E
B S
R A
D H
F%
-^tfl
J
H
c
M
S
z N
F C
S S
c c
G R
ss
-1- AA
B 79
P
F
PROKARY
K
M-
C
F
1
f±
p
L F
M
_ C
r
s
D
M
X
°" M
M H
f_I
m
99
"" PI
ft
C
t
T
7
S
c
M
~ D
i U
1 u
HH G
_L 1
-H
S
£
C
B
A
S C
L L
H H
D
V M
HU-™
D _
L
R
D v VV. C M
J
F C
H
: T
' A^J,, H
ft -jyiTJ: °T
=c
-
s
G
R
"*"
A
N
G
LOW EUK
PLNT
INS
MAMM VITRO
HUMAN
F
F==
-
1 |
1
_
S
L
P
5
.
" C
EH-
- -~
"P
--
==
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.
Acting by the SN2 mechanism and having a high substrate constant both favor alkylation at the
N7 position of guanine in the DNA (Walker et al., 1990). The predominant DNA adduct formed
by EtO and other SN2-type alkylating agents is N7-(2-hydroxyethyl)guanine (N7-HEG). After in
vitro treatment of DNA with EtO, Segerback (1990) identified three adducts, N7-HEG,
N3-hydroxyethyladenine, and O6- hydroxy ethyl guanine (O6-HEG), in the ratios 200:8.8: 1; two
other peaks, suspected of representing other adenine adducts, were also observed at levels well
below that of N7-HEG.
Ethyl ene, an endogenous precursor of EtO, is produced during normal physiological
processes. Such processes reportedly include oxidation of methionine and hemoglobin, lipid
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
This document is a draft for review purposes only and does not constitute Agency policy.
3-29 DRAFT— DO NOT CITE OR QUOTE
-------
1 P450-mediated conversion of ethylene (Tornqvist 1996). This endogenous production of EtO
2 contributes significantly to background levels of DNA adducts, making it difficult to detect the
3 impacts of low levels of exogenous EtO exposure on DNA adduct levels. For example, in DNA
4 extracted from the lymphocytes of unexposed individuals, mean background levels of N7-HEG
5 ranged from 2 to 8.5 pmol/mg DNA (Bolt, 1996). Using sensitive detection techniques and an
6 approach designed to separately quantify both endogenous N7-HEG adducts and "exogenous"
7 N7-HEG adducts induced by EtO treatment in rats, Marsden et al. (2009) reported increases in
8 exogenous adducts in DNA of spleen and liver at the lowest dose administered (0.0001 mg/kg
9 injected i.p. daily for 3 days) and statistically significant linear dose-response relationships
10 (p< 0.05) for exogenous adducts in all three tissues examined (spleen, liver, and stomach),
11 although the authors caution that some of the adduct levels induced at low EtO concentrations
12 are below the limit of accurate quantitation. Note that the whole range of doses studied by
13 Marsden et al. (2009) lies well below the dose corresponding to the lowest LOAEL from an EtO
14 cancer bioassay (see Section C.7 of Appendix C). Marsden et al. (2009) also observed increases
15 in endogenous N7-HEG adduct formation at the two highest doses (0.05 and 0.1 mg/kg),
16 suggesting that, in addition to direct adduct formation via alkylation, EtO can induce N7-HEG
17 adduct production indirectly. Marsden et al. (2009) hypothesized that this indirect adduct
18 formation by EtO results from the induction of ethyl ene generation under conditions of oxidative
19 stress.
20 In experiments with rats and mice exposed to EtO at concentrations of 0, 3, 10, 33, 100,
21 or 300 (rats only) ppm for 6 hours per day, 5 days per week, for 4 weeks, Walker et al. (1992a)
22 measured N7-HEG adducts in the DNA of lung, brain, kidney, spleen, liver, and testes. At
23 100 ppm, the adduct levels for all tissues except testis were similar (within a factor of 3), despite
24 the fact that not all of these tissues are targets for toxicity. The study's data on the persistence of
25 the DNA adducts indicate that DNA repair rates differ in different tissues. Although Walker et
26 al. (1992a) suggested that N7-HEG adducts are likely to be removed by depurination forming
27 apurinic/apyrimidinic (AP) sites in DNA, a later study from the same group showed that
28 EtO-induced DNA damage is repaired without accumulation of AP sites or involving base
29 excision repair (Rusvn et al., 2005). Rats exposed to high doses of EtO (300 ppm) by inhalation
30 showed steady-state levels of O6-HEG adducts that are -250-300 times lower than the N7-HEG
31 levels (Walker et al., 1992a). Even though low levels of O6-HEG adducts were detected, they
32 are more mutagenic in nature and may contribute to the tumors observed in target organs.
33 Two studies provide evidence of N7-HEG DNA adduct formation in human populations
34 occupationally exposed to EtO, one reporting a modest increase in white blood cells (van Delft et
35 al., 1994) and the other a four- to fivefold increase in granulocytes (Yong et al., 2007) compared
This document is a draft for review purposes only and does not constitute Agency policy.
3-30 DRAFT—DO NOT CITE OR QUOTE
-------
1 to unexposed controls. However, these differences were not statistically significant due to high
2 interindividual variation in adduct levels.
O
4 3.3.3.2. Point Mutations
5 EtO has consistently yielded positive results in in vitro mutation assays from
6 b acted ophage, bacteria, fungi, yeast, insects, plants, and mammalian cell cultures (including
7 human cells). For example, EtO induces single base pair deletions and base substitutions in the
8 HPRT gene in human diploid fibroblasts (Kolman and Chovanec, 2000; Lambert et al., 1994;
9 Bastlova et al., 1993) in vitro. The results of in vivo studies on the mutagenicity of EtO have
10 also been consistently positive following ingestion, inhalation, or injection [e.g.Tates et al.
11 (1999)1. Increases in the frequency of gene mutations in T-lymphocytes (Hprt locus) (Walker et
12 al., 1997) and in bone marrow and testes (Lad locus) (Recio et al., 2004) have been observed in
13 transgenic mice exposed to EtO via inhalation at concentrations similar to those in
14 carcinogenesis bioassays with this species (NTP, 1987). At somewhat higher concentrations
15 than those used in the carcinogenesis bioassays (200 ppm, but for only 4 weeks), increases in the
16 frequency of gene mutations have also been observed in the lungs of transgenic mice (Lad
17 locus) (Sisk et al., 1997) and in T-lymphocytes of rats (Hprt locus) (van Sittert et al., 2000; Tates
18 et al., 1999). In in vivo studies with male mice, EtO also causes heritable mutations and other
19 effects in germ cells (Generoso et al., 1990; Lewis et al., 1986).
20 In a study of mammary gland carcinomas in EtO-exposed B6C3Fi mice from the 1987
21 NTP bioassay (NTP, 1987) and 19 mammary gland carcinomas from concurrent controls in the
22 1987 NTP EtO bioassay and a 1986 NTP benzene bioassay, Houle et al. (2006) measured
23 mutation frequencies in exons 5-8 of thep53 tumor suppressor gene and in codon 61 of the Hras
24 oncogene. Mutation frequencies in the mammary carcinomas of EtO-exposed mice were only
25 slightly increased over frequencies in spontaneous mammary carcinomas (33% of the
26 carcinomas in the EtO-exposed mice had Hras mutations versus 26% of spontaneous tumors;
27 67% of the carcinomas in the EtO-exposed mice hadp53 mutations versus 58% of spontaneous
28 tumors); however, the EtO-induced tumors exhibited a distinct shift in the mutational spectra of
29 thep53 and Hras genes and more commonly displayed concurrent mutations of the two genes
30 (Houle et al., 2006). Furthermore, Houle et al. (2006) detected about sixfold higher levels of p53
31 protein expression in the mammary carcinomas of EtO-exposed mice than in spontaneous
32 mammary carcinomas, and there was an apparent dose-response relationship between EtO
33 exposure level and both p53 protein expression andp53 gene mutation (three of the seven tumors
34 in the 50-ppm exposure group and all five tumors in the 100-ppm group had increased protein
35 expression; also, threep53 gene mutations were found in the seven tumors in the 50-ppm
This document is a draft for review purposes only and does not constitute Agency policy.
3 -31 DRAFT—DO NOT CITE OR QUOTE
-------
1 exposure group and nine were found in the five tumors in the 100-ppm group). Some of the
2 same investigators conducted a similar study ofKras mutations in lung, Harderian gland, and
3 uterine tumors (Hong et al., 2007). Substantial increases were observed in Kras mutation
4 frequencies in the tumors from the EtO-exposed mice. Kras mutations were reported in 100% of
5 the lung tumors from EtO-exposed mice versus 25% of spontaneous lung tumors (108 NTP
6 control animal tumors, including 8 from the EtO bioassay), in 86% of Harderian gland tumors
7 from EtO-exposed mice versus 7% of spontaneous Harderian gland tumors (27 NTP control
8 animal tumors, including 2 from the EtO bioassay), and in 83% of uterine tumors from
9 EtO-exposed mice (there were no uterine tumors in control mice in the 1986 NTP bioassay and
10 none were examined from other control animals). Furthermore, a specific Kras mutation, a
11 G —> T transversion in codon 12, was nearly universal in lung tumors from EtO-exposed mice
12 (21/23) but rare in lung tumors from control animals (1/108). Other specific mutations were also
13 predominant in the Harderian gland and uterine tumors, but too few Kras mutations were
14 available in spontaneous Harderian gland tumors, and no spontaneous uterine tumors were
15 examined; thus, meaningful comparisons could not be made for these sites. Overall, these data
16 strongly suggest that EtO-induced mutations in oncogenes and tumor-suppressor genes play a
17 role in EtO-induced carcinogenesis in multiple tissues.
18 Only a few studies have investigated gene mutations in people occupationally exposed to
19 EtO. In one study, HPRTmutant frequency in peripheral blood lymphocytes was measured in a
20 group of 9 EtO-exposed hospital workers, a group of 15 EtO-exposed factory workers, and their
21 respective controls (Tates etal., 1991). EtO exposure scenarios suggest higher exposures in the
22 factory workers, and this is supported by the measurement of higher hemoglobin adduct levels in
23 those workers. HPRTmutant frequencies were 55% increased in the hospital workers, but the
24 increase was not statistically significant. In the factory workers, a statistically significant
25 increase of 60% was reported. In a study of workers in an EtO production facility (Tates et al.,
26 1995), //Permutations were measured in three exposed groups and one unexposed group (seven
27 workers per group). No significant differences in mutant frequencies were observed between the
28 groups; however, the authors stated that about 50 subjects per group would have been needed to
29 detect a 50% increase.
30 Major et al. (2001) measured //Permutations in female nurses employed in hospitals in
31 Eger and Budapest, Hungary. This study and an earlier study measuring effects on chromosomes
32 (see Table 3-6) were conducted to examine a possible causal relationship between EtO exposure
33 and a cluster of cancers (mostly breast) in nurses exposed to EtO in the Eger hospital. The
34 Budapest hospital was chosen because there was no apparent increase in cancer among nurses
35 exposed to EtO. Controls were female hospital workers in the respective cities, and nurses in
This document is a draft for review purposes only and does not constitute Agency policy.
3-32 DRAFT—DO NOT CITE OR QUOTE
-------
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 S 8
0.5-8
0 S ID
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)
Stollevetal. (1984),
Galloway etal. (1986)
Garry etal. Q 979)
Hansen etal. (1984)
Hosstedt etal. (1983)
Laurent etal. (1984)
Lerda and Rizzi (1992)
Maior etal. (1996)
Mayer etal. (1991)
Poppetal. (1994)
Ribeiro etal. (1994)
Richmond et al. (1985)
§•
§
I
o
a
I
a,
8"
O 5
31
o ^
H t
W ^S'
o
c
o
H
W
-------
Table 3-6. Cytogenetic effects in humans (continued)
§•
§
I
Number exposed
(number of controls)
22 (22)
19(19)
10(10)
9
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
OS 19
0.1-4
4-12
2-6
3-27
Accidental
<5
>15
i ^
6-14
Mean
3
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
-------
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
a
S
O
a
I
a,
8"
O 5
31
o ^
H t
W ^S'
o
c
o
H
W
-------
1 Eger with known cancers were excluded. Mean peak levels of EtO were 5 mg/m3 (2.7 ppm) in
2 Budapest and 10 mg/m3 (5.4 ppm) in Eger. HPRT variant frequencies in both controls and
3 EtO-exposed workers in the Eger hospital were higher than either group in the Budapest hospital,
4 but there was no significant increase among the EtO-exposed workers in either hospital when
5 compared with the respective controls. The authors noted that the HPRT variant frequencies
6 among smoking EtO-exposed nurses in Eger were significantly higher than among smokers in
7 the Eger controls; however, the fact that the HPRT variant frequency was almost three times
8 higher in nonsmokers than in smokers in the Eger hospital control group raises questions about
9 the basis of the claimed EtO effect.
10
11 3.3.3.3. Chromosomal Effects
12 As discussed by Preston (1999) in an extensive review of the cytogenetic effects of EtO,
13 a variety of cytogenetic assays can be used to measure induced chromosome damage. However,
14 most of the assays commonly employed measure events that are detectable only in the first (or in
15 some cases the second) metaphase after exposure and require DNA synthesis to convert DNA
16 damage into a chromosomal aberration. In addition, DNA repair is operating in peripheral
17 lymphocytes to repair induced DNA damage. Thus, for acute exposures, the timing of sampling
18 is of great importance. For chronic studies, the endpoints measure only the most recent
19 exposures, and if the time between last exposure and sampling is long, any induced DNA
20 damage not converted to a stable genotoxic alteration is certain to be missed. The events
21 measured include all types of chromosomal aberrations, micronuclei, SCE, and numerical
22 chromosomal changes. Stable chromosomal aberrations include reciprocal translocations,
23 inversions, and some fraction of insertions and deletions as well as some numerical changes.
24 However, until the development of fluorescent in situ hybridization (FISH), chromosome
25 banding techniques were needed to detect these types of aberrations.
26 In in vitro assays, EtO has consistently tested positive in studies for multiple types of
27 chromosomal effects, including DNA strand breaks, SCEs, micronuclei, and chromosomal
28 aberrations [e.g. see Table 11 of IARC (2008)1. Of note, Adam et al. (2005) measured the
29 sensitivity of different human cell types to EtO-induced DNA damage using the comet assay,
30 which measures direct strand breaks and/or DNA damage converted to strand breaks during
31 alkaline treatment. Adam et al. (2005) reported dose-dependent increases in DNA damage in the
32 concentration range 0-100 uM in each of the cell types examined with no notable cytotoxicity.
33 At the lowest concentration reported (20 uM), significant increases in DNA damage were
34 observed in lymphoblasts, lymphocytes, and breast epithelial cells, but not in keratinocytes or
35 cervical epithelial cells, suggesting that breast epithelial cells may have increased sensitivity to
This document is a draft for review purposes only and does not constitute Agency policy.
3-36 DRAFT—DO NOT CITE OR QUOTE
-------
1 EtO-induced genotoxicity compared to other nonlymphohematopoietic cell types. In addition,
2 Godderis et al. (2006) investigated the effects of genetic polymorphisms on DNA damage
3 induced by EtO in peripheral blood lymphocytes of 20 nonsmoking university students. No
4 significant increases in micronuclei were observed following EtO treatment; however,
5 dose-related increases in DNA strand breaks were seen in the comet assay. GST polymorphisms
6 did not have a significant impact on the EtO-induced effects; however, significant increases in
7 DNA strand breaks were associated with low-activity alleles of two DNA repair enzymes
8 compared to wild-type alleles.
9 In vivo, several inhalation studies in laboratory animals have demonstrated that EtO
10 exposure levels in the range of those used in the rodent bioassays induce SCEs [see Table 11 of
11 IARC (2008)]; however, evidence for micronuclei and chromosomal aberrations from these same
12 exposure levels is less consistent. In particular, studies by van Sittert et al. (2000) and Lorenti
13 Garcia etal. (2001) observed increases in micronuclei and chromosomal aberrations in splenic
14 lymphocytes of rats exposed to 50, 100, or 200 ppm EtO for 6 hours/day, 5 days/week, for
15 4 weeks compared to levels from control rats, but the increases were not statistically significant.
16 IARC (2008) noted, however, that "strong conclusions cannot be drawn" from these two studies
17 because the cytogenetic analyses "were initiated 5 days after the final day of exposure, a
18 suboptimal time, and the power of the (FISH) studies were limited by analysis of only a single
19 chromosome and the small numbers of rats per group examined," which was 3 per exposure
20 group in both of the studies, although numerous cells/rat were examined. Moreover, a recent
21 study by Donner etal. (2010) showed clear, statistically significant increases in chromosomal
22 aberrations with longer durations of exposure (>12 weeks) to the concentration levels used in the
23 rodent bioassays.
24 In humans, various studies of occupationally exposed workers have reported SCEs and
25 other chromosomal effects associated with EtO exposure, including micronuclei and
26 chromosomal aberrations. The genotoxicity of EtO was demonstrated in humans as early as
27 1979. Table 3-6 summarizes the cytogenetic effects of EtO on human exposures (see also
28 Appendix C for more details on some of the studies).
29 As illustrated in Table 3-6, numerous studies observed increased SCEs in occupationally
30 exposed workers, especially for workers with the highest exposures [e.g., (Major et al., 1996;
31 SartoetaL 1991: TatesetaL 1991: SartoetaL 1987)1. Several studies of occupationally
32 exposed workers have also reported increased micronucleus formation in lymphocytes (Ribeiro
33 et al., 1994; Tates et al., 1991), in nasal mucosal cells (Sarto et al., 1990), and in bone marrow
34 cells (Hogstedt et al., 1983), although this endpoint seems to be less sensitive than SCEs. An
35 association between increased micronucleus frequency and cancer risk has been reported in at
This document is a draft for review purposes only and does not constitute Agency policy.
3-37 DRAFT—DO NOT CITE OR QUOTE
-------
1 least one large prospective general population study (Bonassi et al., 2007). In addition,
2 chromosomal aberrations have been reported in multiple studies of workers occupationally
3 exposed to EtO (Ribeiro et al.. 1994: Tates et al.. 1991: Sarto et al.. 1987). Chromosomal
4 aberrations have been linked to an increased risk of cancer in several large prospective general
5 population studies [e.g., (Boffetta et al., 2007; Rossner et al., 2005; Hagmar et al., 2004; Liou et
6 al.. 1999)1.
7
8 3.3.3.4. Summary
9 The available data from in vitro studies, laboratory animal models, and epidemiological
10 studies establish that EtO is a mutagenic and genotoxic agent that causes a variety of types of
11 genetic damage.
12
13 3.4. MODE OF ACTION
14 EtO is an alkylating agent that has consistently been found to produce numerous
15 genotoxic effects in a variety of biological systems ranging from bacteriophage to occupationally
16 exposed humans. It is carcinogenic in mice and rats, inducing tumors of the
17 lymphohematopoietic system, brain, lung, connective tissues, uterus, and mammary gland. In
18 addition, epidemiological studies have shown an increased risk of various types of human
19 cancers (see Table A-5 in Appendix A), in particular lymphohematopoietic and breast cancers.
20 Target tissues for EtO carcinogenicity in laboratory animals are varied, and the cancers are not
21 clearly attributable to any specific type of genetic alteration. Although the precise mechanisms
22 by which the multisite carcinogenicity in mice, rats, and humans occurs are unknown, EtO is
23 clearly a mutagenic and genotoxic agent, as discussed in Section 3.3.3, and mutagenicity and
24 genotoxicity are well established as playing a key role in carcinogenicity.
25 Exposure of cells to DNA-reactive agents results in the formation of carcinogen-DNA
26 adducts. The formation of DNA adducts results from a sequence of events involving absorption
27 of the agent, distribution to different tissues, and accessibility of the molecular target (Swenberg
28 et al., 1990). Alkylating agents may induce several different DNA alkylation products (Beranek,
29 1990) with varying proportions, depending primarily on the electrophilic properties of the agent.
30 The predominant DNA adduct formed by EtO is N7-HEG, although other adducts, such as
31 N3-hydroxyethyladenine and O6-HEG, have also been observed, in much lesser amounts
32 (Segerback, 1990). In addition to direct DNA adduct formation via alkylation, Marsden et al.
33 (2009) observed an indirect effect of EtO exposure on endogenous N7-HEG adduct formation
34 and hypothesized that EtO could also indirectly cause adduct formation via oxidative stress (see
35 also Section 3.3.3.1 and Appendix C). The various adducts are processed by different repair
This document is a draft for review purposes only and does not constitute Agency policy.
3-38 DRAFT—DO NOT CITE OR QUOTE
-------
1 pathways, and the subsequent genotoxic responses elicited by unrepaired DNA adducts are
2 dependent on a wide range of variables. The specific adduct(s) responsible for EtO-induced
3 genotoxicity and the mechanism(s) by which this adduct(s) induces the genotoxic damage are
4 unknown.
5 It had been postulated that the predominant EtO-DNA adduct, N7-HEG, although
6 unlikely to be directly promutagenic, could be subject to depurination, resulting in an apurinic
7 site which could be vulnerable to miscoding during cell replication [e.g., Walker and Skopek
8 (1993)1. However, in a study designed to test this hypothesis, Rusyn et al. (2005) failed to detect
9 an accumulation of abasic sites in brain, spleen, and liver tissues of rats exposed to EtO. Rusvn
10 et al. (2005) conclude that the accumulation of abasic sites is unlikely to be a primary
11 mechanism for EtO mutagenicity, although they note that it is also possible that their assay was
12 not sufficiently sensitive to detect small increases in abasic sites or that abasic sites are only
13 mutagenic under conditions of rapid cell turnover, when cell replication may occur before repair
14 of the abasic site (the tissues examined in their study were relatively quiescent). Another
15 potential mechanism for EtO-induced mutagenicity is the direct mutagenicity of the
16 promutagenic adducts such as O6-HEG, although these adducts are generally considered to occur
17 at levels too low to explain all of the observed mutagenicity (IARC, 2008). In an in vitro study,
18 Tompkins et al. (2009) exposed plasmid DNA to a range of EtO concentrations in water and
19 reported that only the N7-HEG adduct was detectable after exposure to EtO concentrations up to
20 2,000 |iM; at higher EtO concentrations (>10 mM), Nl-hydroxyethyladenine and O6-HEG
21 adducts were also quantifiable but at much lower levels than the N7-HEG adducts. Tompkins et
22 al. (2009) then examined the mutagenicity of these adducts in a supF forward mutation assay and
23 reported that the relative mutation frequencies were statistically significantly elevated only for
24 plasmids exposed to these higher EtO concentrations (increases in relative mutation frequency
25 were observed for N7-HEG adduct levels corresponding to lower EtO concentrations, however,
26 and biologically relevant EtO-related increases in mutation frequency at these lower
27 concentrations cannot be ruled out given the variability of the data and the limitations of the
28 study) (see Appendix C, Sections C.I.2 and C.2.2, for a more detailed discussion of this study).
29 An additional mechanism that has been suggested for EtO-induced mutagenicity is the imidazole
30 ring-opening of N7-HEG, which can result in stable, potentially mutagenic lesions; however,
31 EtO-induced N7-HEG ring-opening has not been corroborated in vivo (IARC, 2008).
32 The events involved in the formation of chromosomal damage by EtO are similarly
33 unknown. N-alklylated bases are removed from DNA by base excision repair pathways. A
34 review by Memisoglu and Samson (2000) notes that the action of DNA glycosylase and apurinic
35 endonuclease creates a DNA single-strand break, which can in turn lead to DNA double-strand
This document is a draft for review purposes only and does not constitute Agency policy.
3-39 DRAFT—DO NOT CITE OR QUOTE
-------
1 breaks (DSBs). DSBs can also be produced by normal cellular functions, such as during V(D)J
2 recombination in the development of lymphoid cells or topoisomerase II-mediated cleavage at
3 defined sites. A review of mechanisms of DSB repair indicates that the molecular mechanisms
4 are not fully understood (Pfeiffer et al., 2000). This review provides a thorough discussion of
5 both sources (endogenous and exogenous) of DSBs and the variety of repair pathways that have
6 evolved to process the breaks. Although homology-directed repair generally restores the original
7 sequence, during nonhomologous end-joining, the ends of the breaks are frequently modified by
8 addition or deletion of nucleotides. The lack of accumulation of abasic sites observed in the
9 Rusvn et al. (2005) study discussed above argues against a mechanism involving abasic sites as
10 hot spots for strand breaks, although it is possible that abasic sites accumulate more readily in
11 replicating lymphocytes, which were not examined in the study of Rusyn et al. (2005). Another
12 postulated mechanism for EtO-induced strand breaks is via the formation of hydroxyethyl
13 adducts on the phosphate backbone of the DNA, but this mechanism requires further study
14 CIARC. 2008).
15 Lymphohematopoietic malignancies, like all other cancers, are considered to be a
16 consequence of an accumulation of genetic and epigenetic changes involving multiple genes and
17 chromosomal alterations. Although it is clear that chromosome translocations are common
18 features of some hematopoietic cancers, there is evidence that mutations in p53 or NRAS are
19 involved in certain types of leukemia (U.S. EPA, 1997). It should also be noted that
20 therapy-related leukemias exhibiting reciprocal translocations are generally only seen in patients
21 who have previously been treated with chemotherapeutic agents that act as topoisomerase II
22 inhibitors (U.S. EPA, 1997). In NHL, the BCL6 gene is frequently activated by translocations
23 (Chaganti etal., 1998) as well as by mutations within the gene coding sequence (Lossos and
24 Levy, 2000). Preudhomme et al. (2000) observed point mutations in the AML1 gene in 9 of
25 22 patients with the MO type (minimally differentiated acute myeloblastic leukemia) of acute
26 myeloid leukemia (AML), and Harada et al. (2003) identified AML1 point mutations in cases of
27 radiation-associated and therapy-related myelodysplastic syndrome (MDS)/AML. In both
28 reports, point mutations within the coding sequence were found in patients with normal
29 karyotypes as well as some with translocations or other chromosomal abnormalities.
30 Zharlyganova et al. (2008) identified AML1 mutations in 7 of 18 radiation-exposed MDS/AML
31 patients but in none of 13 unexposed MDS/AML cases. Other point mutations have also been
32 identified in therapy-related MDS/AML patients, includingp53 gene mutations after exposure to
33 alkylating agents (Christiansen et al., 2001) and mutations in RAS and other genes in the receptor
34 tyrosine kinase signal transduction pathway (Christiansen et al., 2005). Several models have
35 been developed to integrate these various types of genetic alterations. One recent model
This document is a draft for review purposes only and does not constitute Agency policy.
3-40 DRAFT—DO NOT CITE OR QUOTE
-------
1 suggests that the pathogenesis of MDS/AML can be subdivided into at least eight genetic
2 pathways that have different etiologies and different biologic characteristics (Pedersen-Bjergaard
3 et al.. 2006).
4 A mode-of-action-motivated modeling approach based solely on chromosome
5 translocations has been proposed by Kirman et al. (2004). The authors suggested a nonlinear
6 dose-response relationship for EtO and leukemia, based on a consideration that "chromosomal
7 aberrations are the characteristic initiating events in chemically induced acute leukemia and gene
8 mutations are not characteristic initiating events." They proposed that EtO must be responsible
9 for two nearly simultaneous DNA adducts, yielding a dose-squared (quadratic) relationship
10 between EtO exposure and leukemia risk. However, as discussed above, there is evidence that
11 does not support the assumption that chromosomal aberrations represent the sole initiating event.
12 In fact, these aberrations or translocations could be a downstream event resulting from genomic
13 instability. In addition, it is not clear that acute leukemia is the lymphohematopoietic cancer
14 subtype associated with EtO exposure; in the large NIOSH study, increases in
15 lymphohematopoietic cancer risk were driven by increases in lymphoid cancer subtypes.
16 Furthermore, even if two reactions with DNA resulting in chromosomal aberrations or
17 translocations are early-occurring events in some EtO-induced lymphohematopoietic cancers, it
18 is not necessary that both events be associated with EtO exposure (e.g., background error repair
19 rates or exposure to other alkylating agents may be the cause). Moreover, EtO could also
20 produce translocations indirectly by forming DNA or protein adducts that affect the normally
21 occurring recombination activities of lymphocytes or the repair of spontaneous double-strand
22 breaks. Thus, broader mode-of-action considerations were not regarded as supportive of the
23 hypothesis that the exposure-response relationship is purely quadratic.
24 Breast cancer is similarly considered to be a consequence of an accumulation of genetic
25 and epigenetic changes involving multiple genes and chromosomal alterations (Ingvarsson,
26 1999). Again, the precise mechanisms by which EtO induces breast cancer are unknown. As
27 discussed in Section 3.3.3.2, in a study of mammary gland carcinomas in EtO-exposed mice,
28 Houle et al. (2006) noted that the EtO-induced tumors exhibited a distinct shift in the mutational
29 spectra of thep53 and Hras genes and more commonly displayed concurrent mutations of the
30 two genes. The comet assay results of Adam et al. (2005) suggest that human breast epithelial
31 cells may have increased sensitivity to EtO-induced genotoxicity compared to other
32 nonlymphohematopoietic cell types (see Section 3.3.3.3); however, the basis for any increased
33 sensitivity of breast epithelial cells is similarly unknown.
34 In summary, EtO induces a variety of types of genetic damage. It directly interacts with
35 DNA, resulting in DNA adducts, gene mutations, and chromosome damage. Depending on a
This document is a draft for review purposes only and does not constitute Agency policy.
3 -41 DRAFT—DO NOT CITE OR QUOTE
-------
1 number of variables, EtO-induced DNA adducts (1) may be repaired, (2) may result in a
2 base-pair mutation during replication, or (3) may be converted to a DSB, which also may be
3 repaired or result in unstable (micronuclei) or stable (translocation) cytogenetic damage. The
4 available data are strongly supportive of a mutagenic mode of action involving gene mutations
5 and chromosomal aberrations (translocations, deletions, or inversions) that critically alter the
6 function of oncogenes or tumor suppressor genes. Although it is clear that chromosome
7 translocations are common features of many hematopoietic cancers, there is evidence that
8 mutations inp53, AML1, or Nras are also involved in some leukemias. The above scientific
9 evidence provides support for a mutagenic mode of action.
10
11 3.4.1. Analysis of the Mode of Action for Ethylene Oxide Carcinogenicity under EPA's
12 Mode-of-Action Framework
13 In this section, the mode of action evidence for EtO carcinogenicity is analyzed under the
14 mode of action framework in EPA's 2005 Guidelines for Carcinogen Risk Assessment (U.S.
15 EPA (2005aX Section 2.4.3).
16 The hypothesis is that EtO carcinogenicity has a mutagenic mode of action. This
17 hypothesized mode of action is presumed to apply to all of the tumor types.
18 The key events in the hypothesized mutagenic mode of action are (1) DNA adduct
19 formation by EtO, which is a direct-acting alkylating agent; (2) the resulting genetic damage,
20 including the formation and fixation of DNA mutations, particularly in oncogenes and tumor
21 suppressor genes, as well as chromosomal alterations; and (3) clonal expansion of mutated cells
22 during later stages of cancer development. Mutagenicity is a well-established cause of
23 carcinogenicity.
24
25 1. Is the hypothesized mode of action sufficiently supported in the test animals?
26 Numerous studies have demonstrated that EtO forms protein and DNA adducts, in mice
27 and rats (see Sections 3.3.1 and 3.4 and Figure 3-2). For example, Walker et al. (1992b) and
28 Walker et al. (1992a) demonstrated that EtO forms protein adducts with hemoglobin in the blood
29 and DNA adducts with tissues throughout the body, including in the lung, brain, kidney, spleen,
30 liver, and testes.
31 In addition, there is incontrovertible evidence that EtO is mutagenic (see Section 3.3.3).
32 The evidence is strong and consistent; EtO has invariably yielded positive results in in vitro
33 mutation assays from bacteriophage, bacteria, fungi, yeast, insects, plants, and mammalian cell
34 cultures. The results of in vivo studies on the mutagenicity and genotoxicity of EtO have also
35 been consistently positive following ingestion, inhalation, or injection. Increases in the
This document is a draft for review purposes only and does not constitute Agency policy.
3 -42 DRAFT—DO NOT CITE OR QUOTE
-------
1 frequency of gene mutations in the lung, in T-lymphocytes, in bone marrow, and in testes have
2 been observed in transgenic mice exposed to EtO via inhalation at concentrations similar to those
3 in the mouse carcinogenesis bioassays. Furthermore, in a study ofp53 (tumor suppressor gene)
4 and Hras (oncogene) mutations in mammary gland carcinomas of EtO-exposed and control
5 mice, Houle et al. (2006) noted that the EtO-induced tumors exhibited a distinct shift in the
6 mutational spectra of thep53 and Hras genes and more commonly displayed concurrent
7 mutations of the two genes, and in a similar study ofKras (oncogene) mutations in lung,
8 Harderian gland, and uterine tumors, substantial increases were observed in Kras mutation
9 frequencies in the tumors from the EtO-exposed mice (Hong et al., 2007).
10 Several inhalation studies in laboratory animals have demonstrated that EtO exposure
11 levels in the range of those used in the rodent bioassays induce SCEs. Evidence for micronuclei
12 and chromosomal aberrations from these same exposure levels has been less consistent;
13 however, IARC (2008) has noted analytical limitations with some of these analyses (see
14 Section 3.3.3.3). Moreover, a recent study by Donner et al. (2010) showed clear, statistically
15 significant increases in chromosomal aberrations with exposure durations of >12 weeks to the
16 concentration levels used in the rodent bioassays.
17 Ethylene oxide induces a variety of mutagenic and genotoxic effects, including
18 chromosome breaks, micronuclei, SCEs, and gene mutations; however, the more general effect
19 of mutagenicity/genotoxicity is specific and occurs in the absence of cytotoxicity or other overt
20 toxicity. A temporal relationship is also clearly evident, with adducts and mutagenicity
21 observed in subchronic assays.
22 Dose-response relationships have been observed between EtO exposure in vivo and
23 hemoglobin adducts [e.g., Walker et al. (1992b)], as well as DNA adducts, SCEs, and Hprt
24 mutations [e.g., van Sittert et al. (2000)1 (see also Sections 3.3 and 3.4). A mutagenic mode of
25 action for EtO carcinogenicity also clearly comports with notions of biological plausibility and
26 coherence because EtO is a direct-acting alkylating agent. Such agents are generally capable of
27 forming DNA adducts, which in turn have the potential to cause genetic damage, including
28 mutations; and mutagenicity, in its turn, is a well-established cause of carcinogenicity. This
29 chain of key events is consistent with current understanding of the biology of cancer.
30 In addition to the clear evidence supporting a mutagenic mode of action in test animals,
31 there are no compelling alternative or additional hypothesized modes of action for EtO
32 carcinogenicity. For example, there is no cytotoxicity or other toxicity indicative of regenerative
33 proliferation or some other toxicity-related mode of action. Oxidative stress has been
34 hypothesized as a mode of action, but there is little evidentiary support for this hypothesis and
This document is a draft for review purposes only and does not constitute Agency policy.
3-43 DRAFT—DO NOT CITE OR QUOTE
-------
1 the role of oxidative stress in EtO-induced carcinogen!city is speculative at this time (see Section
2 J.3.2 of Appendix J and the response to comment #7 in Appendix L).
O
4 2. Is the hypothesized mode of action relevant to humans?
5 The evidence discussed above demonstrates that EtO is a systemic mutagen in test
6 animals; thus, there is the presumption that it would also be a mutagen in humans. Moreover,
7 there is human evidence directly supporting a mutagenic mode of action for EtO carcinogenicity.
8 Several studies of humans have reported exposure-response relationships between hemoglobin
9 adduct levels and EtO exposure levels [e.g., van Sittert et al. (1993); Schulte et al. (1992); see
10 Section 3.3.2], demonstrating the ability of EtO to bind covalently in systemic human cells, as it
11 does in rodent cells. DNA adducts in EtO-exposed humans have not been well studied, and the
12 evidence of increased DNA adducts is limited.
13 In addition, EtO has yielded positive results in in vitro mutagenicity studies of human
14 cells (see Figure 3-3). Although the studies of point mutations in EtO-exposed humans are few
15 and insensitive and the evidence for mutations is limited, there is clear evidence from a number
16 of human studies that EtO causes chromosomal aberrations, SCEs, and micronucleus formation
17 in peripheral blood lymphocytes (see Section 3.3.3.3 and Table 3-6). At least one study
18 suggested an exposure-response relationship for the formation of SCEs in peripheral blood
19 lymphocytes (Major et al., 1996). Another study reported a statistically significant increase in
20 micronuclei in bone marrow cells in EtO-exposed workers (Hogstedt et al., 1983).
21 Finally, there is strong evidence that EtO causes cancer in humans, including cancer
22 types observed in rodent studies (i.e., lymphohematopoietic cancers and breast cancer),
23 providing further weight to the relevance of the aforementioned events to the development of
24 cancer in humans (see Sections 3.1 and 3.5.1).
25 In conclusion, the weight of evidence supports a mutagenic mode of action for EtO
26 carcinogenicity. Although oxidative stress or other processes might contribute to the
27 development of EtO-induced cancers, the available evidence best supports a mutagenic mode of
28 action as the primary mode for EtO carcinogenicity.
29
30 3. Which populations or lifestages can be particularly susceptible to the hypothesized mode of
31 action?
32 The mutagenic mode of action is considered relevant to all populations and lifestages.
33 According to EPA's Supplemental Guidance (U.S. EPA, 2005b), there may be increased
34 susceptibility to early-life exposures to carcinogens with a mutagenic mode of action. Therefore,
35 because the weight of evidence supports a mutagenic mode of action for EtO carcinogenicity,
This document is a draft for review purposes only and does not constitute Agency policy.
3 -44 DRAFT—DO NOT CITE OR QUOTE
-------
1 and in the absence of chemical-specific data to evaluate differences in susceptibility, increased
2 early-life susceptibility should be assumed and, if there is early-life exposure, the age-dependent
3 adjustment factors should be applied, in accordance with the Supplemental Guidance (see
4 Section 4.4).
5 In addition, as discussed in Section 3.5.2, people with DNA repair deficiencies or genetic
6 polymorphisms conveying a decreased efficiency in detoxifying enzymes may have increased
7 susceptibility to EtO-induced carcinogenicity.
8
9 3.5. HAZARD CHARACTERIZATION
10 3.5.1. Characterization of Cancer Hazard
11 In studies of humans there is substantial evidence that EtO exposure is causally
12 associated with lymphohematopoietic cancer, but the evidence is not strong enough to be
13 conclusive. There is also evidence that EtO exposure is causally associated with breast cancer,
14 but the database for breast cancer is more limited. Of the eight relevant13 Hill "criteria" (or
15 considerations) for causality (Hill, 1965), temporality, coherence, biological plausibility and
16 analogy are readily satisfied, and the other four criteria (specificity, consistency, biological
17 gradient, and strength of association) are satisfied to varying degrees, as discussed below.
18 Temporality, the sole necessary criterion, is satisfied because the subjects of all the
19 epidemiology studies of EtO were workers who were exposed to EtO before the cancers of
20 interest were observed, i.e., exposure preceded the development of the disease.
21 The related criteria of coherence, biological plausibility and analogy are fulfilled by the
22 well-established knowledge that EtO is mutagenic and genotoxic, which are common
23 mechanistic features of many carcinogens; that EtO is carcinogenic in rodents, with
24 lymphohematopoietic cancers being observed in both rats and mice and mammary carcinomas
25 being observed in female mice; and that EtO is an epoxide and epoxides are capable of directly
26 interacting with DNA and are the active metabolites of many carcinogens.
27 There is some specificity with respect to the lymphohematopoietic system. Most of the
28 studies focus on examining risks associated with subcategories of the lymphohematopoietic
29 system. These cancers include leukemia, Hodgkin lymphoma, NHL, reticulosarcoma, and
30 myeloma. (Note that with the exception of the Steenland et al. (2004) study, which includes
31 lymphocytic leukemia in a lymphoid cancer category, the studies do not subcategorize leukemia
32 into its distinct myeloid and lymphocytic subtypes.) In most of the studies, an enhanced risk of
33 cancer of the lymphohematopoietic system is evident, and in some studies, it is statistically
13The 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.
3-45 DRAFT—DO NOT CITE OR QUOTE
-------
1 significant. There also appears to be specificity across the epidemiological database for an
2 increased risk of breast cancer. It should be noted, however, that the specificity criterion is not
3 expected to be strictly satisfied by agents, such as EtO, that are widely distributed in all tissues
4 and are direct-acting chemicals.
5 As just alluded to, there is evidence of consistency between studies with respect to cancer
6 of the lymphohematopoietic system as a whole. Most of the available epidemiologic studies of
7 EtO exposure have reported elevated risks of lymphohematopoietic cancer, and the studies that
8 do not report a significant lymphohematopoietic cancer effect have major limitations, such as
9 small numbers of cases (from small study size and/or insufficient follow-up time), inadequate
10 exposure information, and/or reliance on external analyses (see Table 3-1 and Table A-5 in
11 Appendix A). Overall, about 9 of 11 studies (including only the last follow-up of independent
12 cohorts) with adequate information to determine RR estimates reported an increased risk of
13 lymphohematopoietic cancers or a subgroup thereof, although not all were statistically
14 significant, possibly due to the limitations noted above (see Table 3-1 and Table A-5 in
15 Appendix A). The large, high-quality14 NIOSH study shows statistically significant
16 exposure-response trends for lymphoid cancers and all lymphohematopoietic cancers [ Steenland
17 et al. (2004): see Sections D.3 and D.4 of Appendix D for results for both sexes combined]. Four
18 other studies reported statistically significant increases in risk (Swaen etal., 1996; Benson and
19 Teta, 1993; Bisanti et al., 1993; Hogstedt et al., 1986), although EtO exposures were reportedly
20 low in the Benson and Teta (1993) study and the increased risks may be due to other chemical
21 exposures. Nonsignificant increases in lymphohematopoietic cancer risk were observed in four
22 other studies, based on small numbers of cases (Coggon et al., 2004; Olsen et al., 1997; Hagmar
23 etal.. 1995: Norman etal.. 1995: Hagmar etal.. 1991) [e.g., Norman etal. (1995) had only 1
24 case]. Only 2 of the 11 studies showed no evidence of an increase in lymphohematopoietic
25 cancer risk (Swaen et al.. 2009: Kiesselbach et al.. 1990).
26 Regarding consistency in the breast cancer studies, the large, high-quality NIOSH study
27 shows statistically significant increased risks for both breast cancer mortality [n = 103 deaths;
28 Steenland et al. (2004)1 and breast cancer incidence [n = 319 cases; Steenland et al. (2003)1.
29 Two other studies suggest an increased risk of breast cancer despite their small size [Norman et
14The NIOSH study ( ENREF 125: ENREF 127) was judged to be a "high-quality" study based on the
attributes discussed (Section 3.1 and Section A.2.8 of Appendix A), including availability of individual worker
exposure estimates from a high-quality exposure assessment (which used a comprehensive job-exposure matrix with
exposure-level estimates from a regression model that incorporated a variety of plant and production variables and
that explained 85% of the variability in an independent data set of EtO measurements), cohort study design, large
size, inclusion of males and females, adequate follow-up, absence of any known confounding exposures, and use of
internal comparisons. The breast cancer incidence study using the subcohort with interviews had the additional
attribute of investigating and controlling for several breast cancer risk factors (ENREF 127).
This document is a draft for review purposes only and does not constitute Agency policy.
3-46 DRAFT—DO NOT CITE OR QUOTE
-------
1 al. (1995), n = 12 cases; Kardos et al. (2003), n = 3 deaths]. No elevated risks were seen in the
2 only other two studies reporting breast cancer results; however these studies had few cases,
3 owing to their small size and/or inadequate follow-up time [Hagmar et al. (1991), n = 5 cases;
4 Coggon et al. (2004), n = 11 deaths] (see Table 3-2 and Table A-5 in Appendix A).
5 There is also some evidence of dose-response relationships (biologicalgradient). In the
6 large, high-quality NIOSH study, a statistically significant positive trend was observed in the risk
7 of lymphohematopoietic cancers with increasing (log) cumulative exposure to EtO, although
8 results for this model were reported only for males (Steenland et al., 2004) (the sex difference is
9 not statistically significant, however, and the trend for both sexes combined is also statistically
10 significant; see Tables D-3e and D-4e in Appendix D). For only two other cohorts were results
11 for exposure-response analyses reported, probably because most cohorts had too few cases
12 and/or lacked adequate exposure information. In the Swaen et al. (2009) study of the UCC
13 cohort, no statistically significant trends were observed for leukemia or lymphoid cancer using a
14 Cox proportional hazards model with cumulative exposure, a model which notably did not yield
15 statistically significant trends in the NIOSH study, either. In the small study of Hagmar et al.
16 (1995), an SIR for leukemia of 7.14 was reported for subjects with at least 0.14 ppm x years of
17 cumulative exposure and 10 years latency, but this result was based on only two cases and was
18 not statistically significant. For breast cancer, exposure-response analyses were reported only for
19 the NIOSH cohort, again presumably because most cohorts had too few cases and/or lacked
20 adequate exposure information. These analyses yielded clear, statistically significant trends for
21 both breast cancer mortality (Steenland et al., 2004) and breast cancer incidence (Steenland et al.,
22 2003) for a variety of models.
23 Whereas most of the considerations are largely satisfied, as discussed above, there is little
24 strength in the associations, as reflected by the modest magnitude of most of the RR estimates.
25 For example, in the large NIOSH study, the RR estimate for lymphoid cancer mortality in the
26 highest exposure quartile is about 3.0 and the RR estimate for breast cancer incidence in the
27 highest exposure quintile in the subcohort with interviews is on the order of 1.9. While large RR
28 estimates increase the confidence that an observed association is not likely due to chance, bias,
29 or confounding, modest RR estimates, such as those observed with EtO, do not preclude a causal
30 association (U.S. EPA, 2005a). With EtO, the modest RR estimates may, in part, reflect the
31 relatively high background rates of these cancers, particularly of breast cancer incidence.
32 In addition to the Hill criteria, other factors such as chance, bias, and confounding are
33 considered in analyzing the weight of epidemiological evidence. Given the consistency of the
34 findings across studies and the exposure-response relationships observed in the largest study,
35 none of these factors is likely to explain the associations between these cancers and EtO
This document is a draft for review purposes only and does not constitute Agency policy.
3-47 DRAFT—DO NOT CITE OR QUOTE
-------
1 exposure. Coexposures to other chemicals are expected to have occurred for workers in the
2 chemical industry cohorts but would have been much less likely in the sterilizer worker cohorts,
3 such as the NIOSH cohort, which reported no evidence of confounding exposures to other
4 occupational carcinogens (Steenland et al., 1991). For breast cancer in the NIOSH subcohort
5 with interviews (Steenland et al., 2003), other risk factors for breast cancer were assessed, and
6 statistically significant factors were included in the exposure-response models.
7 In conclusion, the overall epidemiological evidence for a causal association between EtO
8 exposure and lymphohematopoietic cancer was judged to be strong but less than conclusive. For
9 breast cancer, the existing evidence was strong based on the large, high-quality NIOSH study,
10 but other studies lacked sufficient breast cancer data and, thus, overall, the epidemiological
11 evidence was judged to be more limited.
12 There is inadequate evidence for other cancer types (e.g., stomach cancer and pancreatic
13 cancer) in the epidemiology studies.
14 The experimental animal evidence for carcinogenicity is concluded to be "sufficient"
15 based on findings of tumors at multiple sites, by both oral and inhalation routes of exposure, and
16 in both sexes of both rats and mice. Tumor types resulting from inhalation exposure included
17 mononuclear cell leukemia in male and female rats and malignant lymphoma and mammary
18 carcinoma in female mice, suggesting some site concordance with the lymphohematopoietic and
19 breast cancers observed in humans, also exposed by inhalation.
20 The evidence of EtO genotoxicity and mutagenicity is unequivocal. EtO is a
21 direct-acting alkylating agent and has invariably tested positive in in vitro mutation assays from
22 bactedophage, bacteria, fungi, yeast, insects, plants, and mammalian cell cultures (including
23 human cells). In mammalian cells (including human cells), EtO-induced genotoxic effects
24 include unscheduled DNA synthesis, gene mutations, SCEs, and chromosomal aberrations. The
25 results of in vivo genotoxicity studies of EtO have also been largely positive, following
26 ingestion, inhalation, or injection. Increases in frequencies of gene mutations have been reported
27 in the lung, T-lymphocytes, bone marrow, and testes of EtO-exposed mice. In particular,
28 increases in frequencies of oncogene mutations have been observed in several tumor types from
29 EtO-exposed mice compared to spontaneous mouse tumors of the same types. Several inhalation
30 studies in laboratory animals have demonstrated that EtO exposure levels in the range of those
31 used in the rodent bioassays (i.e., 10-100 ppm, 6-7 hours/day, 5 days/week) induce SCEs.
32 Evidence for micronuclei and chromosomal aberrations from these same exposure levels in
33 short-term studies (4 weeks or less) is less consistent, although concerns have been raised about
34 some of the negative studies. A recent study showed clear, statistically significant increases in
35 chromosomal aberrations with longer durations of exposure (>12 weeks) to the concentration
This document is a draft for review purposes only and does not constitute Agency policy.
3-48 DRAFT—DO NOT CITE OR QUOTE
-------
1 levels used in the rodent bioassays. The studies of point mutations in EtO-exposed humans are
2 few and insensitive and the evidence for mutations is limited; however, there is clear evidence
3 from a number of human studies that EtO causes chromosomal aberrations, SCEs, and
4 micronucleus formation in peripheral blood lymphocytes, and one study has reported increased
5 levels of micronuclei in bone marrow cells in EtO-exposed workers.
6 Under EPA's 2005 Guidelines for Carcinogen Risk Assessment (U.S. EPA. 2005a). the
7 conclusion can be made that EtO is "carcinogenic to humans." In general, the descriptor
8 "carcinogenic to humans" is appropriate when there is convincing epidemiologic evidence of a
9 causal association between human exposure and cancer. This descriptor is also appropriate when
10 there is a lesser weight of epidemiologic evidence that is strengthened by specific lines of
11 evidence set forth in the Guidelines (U.S. EPA, 2005a), which are satisfied for EtO. The lines of
12 evidence supporting the characterization of "carcinogenic to humans" include the following:
13 (1) there is strong, although less than conclusive on its own, evidence of cancer in humans
14 associated with EtO exposure via inhalation, specifically, evidence of lymphohematopoietic
15 cancers and female breast cancer in EtO-exposed workers; (2) there is extensive evidence of
16 EtO-induced carcinogenicity in laboratory animals, including lymphohematopoietic cancers in
17 rats and mice and mammary carcinomas in mice following inhalation exposure; (3) EtO is a
18 direct-acting alkylating agent whose mutagenic and genotoxic capabilities have been well
19 established in a variety of experimental systems, and a mutagenic mode of carcinogenic action
20 has been identified in animals involving the key precursor events of DNA adduct formation and
21 subsequent DNA damage, including point mutations and chromosomal effects; and (4) there is
22 strong evidence that the key precursor events are anticipated to occur in humans and progress to
23 tumors, including evidence of chromosome damage, such as chromosomal aberrations, SCEs,
24 and micronuclei in EtO-exposed workers.
25
26 3.5.2. Susceptible Life stages and Populations
27 There are no data on the relative susceptibility of children and the elderly when compared
28 with adult workers, in whom the evidence of hazard has been gathered, but because EtO does not
29 have to be metabolized before binding to DNA and proteins, the maturing of enzyme systems in
30 very young children is thought not to be a predominant factor in its hazard, at least for activation.
31 However, the immaturity of detoxifying enzymes in very young children may increase children's
32 susceptibility because children may clear EtO at a slower rate than adults. As discussed in
33 Section 3.3.1, EtO is metabolized (i.e., detoxified) primarily by hydrolysis in humans but also by
34 glutathione conjugation. Both hydrolytic activity and glutathione-S-transferase activity
35 apparently develop after birth (Clewell et al., 2002); thus, very young children might have a
This document is a draft for review purposes only and does not constitute Agency policy.
3-49 DRAFT—DO NOT CITE OR QUOTE
-------
1 decreased capacity to detoxify EtO compared to adults. In the absence of data on the relative
2 susceptibility associated with EtO exposure in early life, increased early-life susceptibility is
3 assumed, in accordance with EPA's Supplemental Guidance (U.S. EPA, 2005b), because the
4 weight of evidence supports the conclusion of a mutagenic mode of action for EtO
5 carcinogenicity (see Section 3.4).
6 Other than the occurrence of sex-specific cancers (e.g., breast cancer in human females,
7 mammary and uterine carcinomas in female mice, and testicular peritoneal mesotheliomas in
8 male rats; see Section 3.2), there is no clear sex difference in EtO-induced carcinogenicity. With
9 the exception of the sex-specific cancers and the observation of malignant lymphomas in female
10 but not male mice, there is no sex difference in EtO-induced cancer types in the rat and mouse
11 bioassays. Cancer potency estimates for females are roughly 50% higher than those for males
12 for both mice and rats (see Table 4-18 in Section 4.2.5). In humans, in the large NIOSH study
13 (Steenland et al., 2004), the association between lymphoid cancers and EtO exposure was seen
14 primarily in males, but the sex difference was not statistically significant (see Section D.3.b in
15 Appendix D), and the SAB panel that reviewed an earlier draft of this assessment recommended
16 that data from both sexes be combined for the derivation of quantitative risk estimates for the
17 Ivmphohematopoietic cancers (SAB, 2007).
18 Brown et al. (1996) reported that sex differences in EtO toxicokinetics were observed in
19 mice but not in rats; female mice had a significantly higher steady-state blood EtO concentration
20 after 4 hours of exposure to either 100 or 330 ppm than male mice. As noted above and
21 discussed in Section 3.3.1, EtO is metabolized primarily by hydrolysis in humans. Mertes et al.
22 (1985) reported no sex difference in microsomal or cytosolic epoxide hydrolase activities in
23 human liver in vitro using benzo[a]pyrene 4,5-oxide or trans-stilbene oxide, respectively, as
24 substrates. Using EtO as a substrate, but with far fewer subjects, Fennell and Brown (2001)
25 reported similar values for males and females for epoxide hydrolase activity in human liver
26 microsomes and for GSH transferase in human liver cytosolic fractions.
27 Because EtO is detoxified by glutathione conjugation or hydrolysis, people with
28 genotypes conveying deficiencies in glutathione-S-transferase or epoxide hydrolase activities
29 may be at increased risk of cancer from EtO exposure. Yongetal. (2001) measured
30 approximately twofold greater EtO-hemoglobin adduct levels in occupationally exposed persons
31 with a null GSTT1 genotype than in those with positive genotypes. Similarly, in a study of
32 hospital workers, Haufroid et al. (2007) reported increased urinary excretion of a glutathione
33 conjugate of EtO, reflecting increased detoxification of EtO, associated with a nonnull GSTT1
34 genotype, although the increase was not statistically significant in all the regression models
This document is a draft for review purposes only and does not constitute Agency policy.
3-50 DRAFT—DO NOT CITE OR QUOTE
-------
1 tested; associations were less clear for other glutathione-S-transferase or epoxide hydrolase
2 polymorphisms.
3 In addition, people with DNA repair deficiencies such as xeroderma pigmentosum,
4 Bloom's syndrome, Fanconi anemia, and ataxia telangiectasia (Gelehrter et al., 1990) are
5 expected to be especially sensitive to the damaging effects of EtO exposure. Paz-y-Mifio et al.
6 (2002) have recently identified a specific polymorphism in the excision repair pathway gene
7 hMSH2. The polymorphism was present in 7.5% of normal individuals and in 22.7% of NHL
8 patients, suggesting that this polymorphism may be associated with an increased risk of
9 developing NHL.
This document is a draft for review purposes only and does not constitute Agency policy.
3-51 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. (Note that the estimates in Section 4.1 were derived under the common
6 assumption that relative risk is independent of age. This assumption is later superseded by an
7 assumption of increased early-life susceptibility, and it is the unit risk estimates derived under
8 this latter assumption, which are developed in Section 4.4, that are the ultimate estimates
9 proposed in this assessment.) Section 4.2 presents the derivation of unit risk estimates from
10 rodent data. Section 4.3 summarizes the unit risk estimates derived from the different data sets.
11 Section 4.4 discusses adjustments for assumed increased early-life susceptibility, based on
12 recommendations from EPA's Supplemental Guidance (U.S. EPA, 2005b), because the weight of
13 evidence supports the conclusion of a mutagenic mode of action for EtO carcinogenicity (see
14 Section 3.4). Section 4.5 presents conclusions about the unit risk estimates. Section 4.6
15 compares the unit risk estimates derived in this EPA assessment to those derived in other
16 assessments. Finally, Section 4.7 provides risk estimates derived for some general occupational
17 exposure scenarios.
18
19 4.1. INHALATION UNIT RISK ESTIMATES DERIVED FROM HUMAN DATA
20 The NIOSH retrospective cohort study of more than 18,000 workers in 13 sterilizing
21 facilities [most recent update by Steenland et al. (2004) and Steenland et al. (2003)] provides the
22 most appropriate data sets for deriving quantitative cancer risk estimates in humans for several
23 reasons: (1) exposure estimates were derived for the individual workers using a comprehensive
24 exposure assessment, (2) the cohort was large and diverse (e.g., 55% female), and (3) there was
25 little reported exposure to chemicals other than EtO. Exposure estimates, including estimates for
26 early exposures for which no measurements were available, were determined using a regression
27 model that estimated exposures to each individual as a function of facility, exposure category,
28 and time period. The regression model was based on extensive personal monitoring data from
29 18 facilities spanning a number of years as well as information on factors influencing exposure,
30 such as engineering controls [Hornung et al. (1994): see also Section A.2.8 in Appendix A].
31 When evaluated against test data, the model accounted for 85% of the variation in average EtO
32 exposure levels. The investigators were then able to estimate the cumulative exposure
33 (ppm x days) for each individual worker by multiplying the estimated exposure for each job
34 (exposure category) held by the worker by the number of days spent in that j ob and summing
This document is a draft for review purposes only and does not constitute Agency policy.
4-1 DRAFT—DO NOT CITE OR QUOTE
-------
1 over all the jobs held by the worker. Steenland et al. (2004) present follow-up results for the
2 cohort mortality study previously discussed by Steenland et al. (1991) and Stayner et al. (1993).
3 Positive findings in the current follow-up include increased rates of (lympho)hematopoietic
4 cancer mortality and of breast cancer mortality in females. Steenland et al. (2003) present results
5 of a breast cancer incidence study of a subcohort of 7,576 women from the NIOSH cohort.
6 The other major occupational study [most recent update by Swaen et al. (2009)]
7 described risks to Union Carbide workers exposed to EtO at two chemical plants in West
8 Virginia, but this study is less useful for estimating quantitative cancer risks for a number of
9 reasons. First, the exposure assessment is much less extensive than that used for the NIOSH
10 cohort, with greater likelihood for exposure misclassification, especially in the earlier time
11 periods when no measurements were available (1925-1973). Exposure estimation for the
12 individual workers was based on a relatively crude exposure matrix which cross-classified 3
13 levels of exposure intensity with 4 time periods. The exposure estimates for 1974-1988 were
14 based on measurements from air sampling at the West Virginia plants since 1976. The exposure
15 estimates for 1957-1973 were based on measurements in a similar plant in Texas. The exposure
16 estimates for 1940-1956 were based loosely on "rough" estimates reported for
17 chlorohydrin-based EtO production in a Swedish facility in the 1940s. The exposure estimates
18 for 1925-1939 were essentially guesses. Thus, for the two earliest time periods (1925-1939 and
19 1940-1956) at least, the exposure estimates are highly uncertain. (See Section A.2.20 of
20 Appendix A for a more detailed discussion of the exposure assessment for the Union Carbide
21 cohort.) This is in contrast to the NIOSH exposure assessment in which exposure estimates were
22 based on extensive sampling data and regression modeling. In addition, the sterilization
23 processes used by the NIOSH cohort workers were fairly constant back in time, unlike chemical
24 production processes, which likely involved much higher and more variable exposure levels in
25 the past. Furthermore, the Union Carbide cohort is of much smaller size and has far fewer deaths
26 than the NIOSH cohort, it is restricted to males and so cannot be used to investigate breast cancer
27 risk in females, and there are coexposures to other chemicals.
28 A third study (Hagmar et al., 1995; Hagmaretal., 1991) estimated cumulative exposures
29 for individual workers; however, insufficient exposure-response data are presented for the
30 derivation of unit risk estimates. Exposure-response results for specific cancers are provided
31 only in the 1991 paper and then only for two lymphohematopoietic cancers across two
32 categorical exposure groups.
This document is a draft for review purposes only and does not constitute Agency policy.
4-2 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
Table 4-1 provides a summary of the considerations taken into account in selecting the
NIOSH study as the basis for the derivation of unit risk estimates. The NIOSH EtO cohort
mortality data can be obtained from the Industrywide Studies Branch of NIOSH.
15
Table 4-1. Considerations used in this assessment for selecting epidemiology
studies for quantitative risk estimation
Consideration
Studies
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)1 alone was selected for
quantitative risk estimation, as it was judged to
be substantially superior to the UCC study
(Swaen etal.. 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].
9
10
11
12
13
14
15
16
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.
4-3 DRAFT—DO NOT CITE OR QUOTE
-------
1 The derivation of unit risk estimates, defined as the lifetime risk of cancer from chronic
2 inhalation of EtO per unit of air concentration, for lymphohematopoietic cancer mortality and
3 incidence and for breast cancer mortality and incidence in females, based on results of the recent
4 analyses of the NIOSH cohort, is presented in the following subsections.
5 The exposure-response models used to fit the epidemiological data are empirical
6 "curve-fitting" models. Considerations used in the selection of the exposure-response models
7 upon which to base the derivation of unit risk estimates included statistical fit (as reflected by
8 ^-values),16 visual fit of the model shape to the categorical results, and biological plausibility.
9 When multiple models were deemed to be reasonable candidates for selection based on those
10 considerations, AIC17 was also considered in selecting the "preferred" model. (Note that all the
11 models discussed in Chapter 4 are fitted treating exposure as a continuous variable except for the
12 categorical models and the linear regressions of categorical results, which are explicitly
13 described as such.)
14
15 4.1.1. Risk Estimates for Lymphohematopoietic Cancer
16 4.1.1.1. Lymphohematopoietic Cancer Results From the NIOSH Study
17 Steenland et al. (2004) investigated the relationship between (any) EtO exposure and
18 mortality from cancer at a number of sites using life-table analyses with the U.S. population as
19 the comparison population. Categorical SMR analyses were also done by quartiles of cumulative
20 exposure. Then, to further investigate apparent exposure-response relationships observed for
21 (lympho)hematopoietic cancer and breast cancer, internal exposure-response analyses were
22 conducted using Cox proportional hazards models, which have the form
23
24
25 Relative rate (RR) = epx, (4-1)
26
27
28 where P represents the regression coefficient and X is the exposure (or some function of
29 exposure, e.g., the natural log of exposure). Internal analyses were done two ways—with
30 exposure as a categorical variable and with exposure as a continuous variable. A nested
16/>-values generally were obtained from the likelihood ratio test. Smalls-values (conventionally, p < 0.05) indicate
a statistically significant fit, e.g., rejection of the null hypothesis that the regression coefficient on the exposure term
(P) is 0.
17Akaike 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.
4-4 DRAFT—DO NOT CITE OR QUOTE
-------
1 case-control approach was used, with age as the time variable used to form the risk sets. Risk
2 sets were constructed with 100 controls randomly selected for each case from the pool of those
3 surviving to at least the age of the index case. According to the authors, use of 100 controls per
4 case has been shown to result in ORs virtually identical to the RR estimates obtained with full
5 cohorts. Cases and controls were matched on race (white/nonwhite), sex, and date of birth
6 (within 5 years). Exposure was the only covariate in the model, so the/?-value for the model also
7 serves as ap-va\ue for the regression coefficient, P, as well as for a test of exposure-response
8 trend.
9 For lymphohematopoietic cancer mortality, Steenland et al. (2004) analyzed both all
10 lymphohematopoietic cancers combined and a subcategory of lymphohematopoietic cancers that
11 they called "lymphoid" cancers; these included NHL, myeloma, and lymphocytic leukemia.
12 Their exposure-response analyses focused on cumulative exposure and (natural) log cumulative
13 exposure, with various lag periods. Other EtO exposure metrics (duration of exposure, average
14 exposure, and peak exposure) were also examined, but models using these metrics did not
15 generally predict lymphohematopoietic cancer as well as models using cumulative exposure. A
16 lag period defines an interval before death, or end of follow-up, during which any exposure is
17 disregarded because it is not considered relevant to the outcome under investigation. For
18 lymphohematopoietic (and lymphoid) cancer mortality, a 15-year lag provided the best fit to the
19 data, based on the likelihood ratio test. One ppm x day was added to cumulative exposures in
20 lagged analyses to avoid taking the log of 0. For both all lymphohematopoietic and lymphoid
21 cancers, Steenland et al. (2004) found stronger positive exposure-response trends in males and so
22 presented the results for some of the regression models separately by sex. The apparent sex
23 difference was not statistically significant (see Appendix D), however, and results for both sexes
24 combined were subsequently obtained from Dr. Steenland (see Appendix D; Section 3 for
25 lymphoid cancer, Section 4 for all lymphohematopoietic cancer). These results are presented in
26 Table 4-2. For additional details and discussion of the Steenland et al. (2004) study, see
27 Appendix A.
28
This document is a draft for review purposes only and does not constitute Agency policy.
4-5 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
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"
/7-valueb
Coefficient (SE)
(per ppm x day)
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)
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
""Cumulative exposure is in ppm * 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
myeloid and lymphoid cells develop from different progenitor cells; thus, there is stronger
support for an etiologic role of EtO in the development of lymphoid cancers than in the
development of the cancers in the aggregate all lymphohematopoietic cancer category. The
consideration of NHL, (plasma cell) myeloma, and lymphocytic leukemia together as
"lymphoid" cancers is consistent with the current World Health Organization classification of
such cancers based on their derivation from B-cells, T-cells, and NK-cells rather than previous
distinctions (Harris et al., 1999). Nonetheless, for comprehensiveness and for the reasons listed
below, risk estimates based on the all lymphohematopoietic cancer results are presented for
This document is a draft for review purposes only and does not constitute Agency policy.
4-6 DRAFT—DO NOT CITE OR QUOTE
-------
1 comparison. Judging roughly from the ^-values, the model fits do not appear notably better for
2 lymphoid cancers than for all lymphohematopoietic cancers (see Table 4-2, ^-values for log
3 cumulative exposure models), and the "lymphoid" category did not include Hodgkin lymphoma,
4 which also exhibited evidence of exposure-response trends, although based on few cases
5 (Steenland et al., 2004). In addition, misclassification or nonclassification of tumor type is more
6 likely to occur for subcategories of lymphohematopoietic cancer (e.g., 4 of the 25 leukemias in
7 the analyses were classified as "not specified" and so could not be considered for the lymphoid
8 cancer analysis).
9 The results of internal exposure-response analyses of lymphoid cancer in the NIOSH
10 cohort (Cox regression analyses, summarized in Table 4-2) were used for predicting the extra
11 risks of lymphoid cancer mortality from continuous environmental exposure to EtO. Extra risk
12 is defined as
13
14
15 Extra risk = (Rx - Ro)/(l - Ro), (4-2)
16
17
18 where Rx is the lifetime risk in the exposed population and Ro is the lifetime risk in an
19 unexposed population (i.e., the background risk). These risk estimates were calculated using the
20 p regression coefficients and an actuarial program (life-table analysis) that accounts for
1 &
21 competing causes of death. An inherent assumption in the Cox regression model and its
22 application in the life-table analyses is that RR is independent of age. (An alternate assumption
23 of increased susceptibility from early-life exposure to EtO, as recommended in EPA's
24 Supplemental Guidance (U.S. EPA, 2005b) for chemicals, such as EtO [see Section 3.4], with a
25 mutagenic mode of action, is considered in Section 4.4. This alternate assumption is the
26 prevailing assumption in this assessment, based on the recommendations in the Supplemental
27 Guidance. Risk estimates are first developed under the assumption of age independence,
28 however, because that is the standard approach in the absence of evidence to the contrary or of
29 sufficient evidence of a mutagenic mode of action to invoke the divergent assumption of
30 increased early-life susceptibility.)
31 U. S. age-specific all-cause mortality rates for 2004 for both sexes of all race groups
32 combined (Arias, 2007) were used to specify the all-cause background mortality rates in the
33 actuarial program. For the cause-specific background mortality rates for lymphoid cancers,
18This 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 LECOT
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.
4-7 DRAFT—DO NOT CITE OR QUOTE
-------
1 age-specific mortality rates for the relevant subcategories of lymphohematopoietic cancer (NHL
2 [C82-C85 of 10th revision of the International Classification of Diseases (ICD)], multiple
3 myeloma [C88, C90], and lymphoid leukemia [C91]) for the year 2004 were obtained from the
4 National Center for Health Statistics Data Warehouse website
5 (http://www.cdc.gov/nchs/datawh/statab/unpubd/mortabs.htm). The risks were computed up to
6 age 85 for continuous exposures to EtO beginning at birth.19 Conversions between occupational
7 EtO exposures and continuous environmental exposures were made to account for differences in
8 the number of days exposed per year (240 vs. 365 days) and in the amount of EtO-contaminated
9 air inhaled per day [10 vs. 20 m3; U.S. EPA (1994)1. An adjustment was also made for the lag
10 period. The reported standard errors for the regression coefficients from Table 4-2 were used to
11 compute the 95% upper confidence limits (UCLs) for the relative rates, based on a normal
12 approximation.
13 The only statistically significant Cox regression model presented by Steenland et al.
14 (2004) for lymphoid cancer mortality was for log cumulative exposure with a 15-year lag in
15 males (p = 0.02). This was similarly true for the analyses of lymphoid cancer using the data for
16 both sexes (see Table 4-2). However, using the log cumulative exposure model to estimate the
17 risks from low environmental exposures is problematic because this model, which is intended to
18 fit the full range of occupational exposures in the study, is inherently supralinear (i.e., risk
19 increases steeply with increasing exposures in the low exposure range and then plateaus), and
20 results are unstable for low exposures (i.e., small changes in exposure correspond to large
21 changes in risk; see Figure 4-1). Some consideration was thus given to the cumulative exposure
22 model, which is typically used and which is stable at low exposures, although the fit to these data
23 was not statistically significant (p = 0.22). However, the Cox regression model with cumulative
24 exposure is inherently sublinear (i.e., risk increases gradually in the low exposure range and then
25 with increasing steepness as exposure increases) and does not reflect the apparent supralinearity
26 of the data demonstrated by the categorical results and the superior fit of the log cumulative
27 exposure model; thus, this model was not considered further.
19Rates 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 lifespanbut, 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.
4-8 DRAFT—DO NOT CITE OR QUOTE
-------
o
C5
S
I
VO
O
a
i
. . eA(p*exp)
eA(p*logexp)
• categorical
^^^^^— linear
splinelOO
spline1600
15000 20000 25000
mean cumulative exposure (ppm * days)
O 5
O I
2 ^
31
o ^
H t
W ^S'
o
c
o
H
W
Figure 4-1. RR estimate for lymphoid cancer 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 ta(exP°sure»; categorical: Cox
regression results for RR = e(|3 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). (Note that, with
the exception of the categorical results and the linear regression of the categorical results, the different models have different implicitly
estimated baseline risks; thus, they are not strictly comparable to each other in terms of RR values, i.e., along the y-axis. They are, however,
comparable in terms of general shape.)
Source: Steenland reanalyses for males and females combined; see Appendix D (except for linear regression of categorical results, which was
done by EPA).
-------
1 In a 2006 external review draft of this assessment (U.S. EPA, 2006a), which relied on the
2 original published results of Steenland et al. (2004), EPA proposed that the best way to represent
3 the exposure-response relationship in the lower exposure region, which is the region of interest
4 for low-exposure extrapolation, was through the use of a weighted linear regression of the results
5 from the Cox regression model with categorical cumulative exposure and a 15-year lag [for
6 males only, as this was the significant finding in the published paper of Steenland et al. (2004)].
7 In addition, the highest exposure group was not included in the regression to alleviate some of
8 the "plateauing" in the exposure-response relationship at higher exposure levels and to provide a
9 better fit to the lower exposure data. Linear modeling of categorical (i.e., grouped)
10 epidemiologic data and elimination of the highest exposure group(s) under certain circumstances
11 to obtain a better fit of low-exposure data are both standard techniques used in EPA
12 dose-response assessments (U.S. EPA, 2012, 2005a). An established methodology was
13 employed for the weighted linear regression of the categorical epidemiologic data, as described
14 by Rothman (1986) and used by others [e.g., van Wijngaarden and Hertz-Picciotto (2004)].
15 However, the SAB panel that reviewed the draft assessment recommended that EPA employ
16 models using the individual continuous exposure data as an alternative to modeling the published
17 grouped data. The SAB also recommended that both males and females be included in the
18 modeling of lymphohematopoietic cancer mortality (SAB, 2007).
19 In response to these recommendations and in consultation with Dr. Steenland, one of the
20 investigators from the NIOSH cohort studies, EPA determined that, using the full continuous
21 exposure data set, an alternative way to address the supralinearity of the data (while avoiding the
22 extreme low-exposure curvature obtained with the log cumulative exposure model) might be to
23 use a two-piece log-linear spline model. Spline models have been used previously for
24 exposure-response analyses of epidemiological data (Steenland and Deddens, 2004; Steenland et
25 al., 2001). These models are particularly useful for exposure-response data such as the EtO
26 lymphoid cancer data, for which RR initially increases with increasing exposure but then tends to
27 plateau, or attenuate, at higher exposures. Such plateauing exposure-response relationships have
28 been seen with other occupational carcinogens and may occur for various reasons, including the
29 depletion of susceptible subpopulations at high exposures, mismeasurement of high exposures,
30 or a healthy worker survivor effect (Stayner et al., 2003). No other traditional exposure-response
31 models for continuous exposure data that might suitably fit the observed exposure-response
32 pattern were apparent. Dr. Steenland was commissioned to do the spline analyses using the full
33 data set with cumulative exposure as a continuous variable, and his findings are included in
34 Appendix D (see Section D.3 for lymphoid cancer, Section D.4 for all lymphohematopoietic
35 cancer). The results of the spline analyses are presented below.
This document is a draft for review purposes only and does not constitute Agency policy.
4-10 DRAFT—DO NOT CITE OR QUOTE
-------
1 For the two-piece log-linear spline modeling approach, the Cox regression model
2 (Equation 4-1) was the underlying basis for the splines which were fit to the lymphoid cancer
90
3 exposure-response data. Taking the log of both sides of Equation 4-1, log RR is a linear
4 function of exposure (cumulative exposure is used here), and, with the two-piece log-linear
5 spline approach, log RR is a function of two lines which join at a single point of inflection, called
6 a "knot." The shape of the two-piece log-linear spline model, in particular the slope in the
7 low-exposure region, depends on the location of the knot. For this assessment, the knot was
8 generally selected by evaluating different knots in increments of 100 ppm x days over some
9 range of cumulative exposures starting at 0 and then choosing the one that resulted in the best
10 (largest) model likelihood. The model likelihood did not change much across the different trial
11 knots for any of the data sets, but it did change slightly, and the largest calculated likelihood was
12 used as the basis for knot selection. For more discussion of the two-piece spline approach, see
13 Appendix D.
14 For the lymphoid cancer data, the range examined for knot selection was from 0 to
15 15,000 ppm x days, and the largest model likelihood was observed with the knot at 100 ppm x
16 days, although, as noted above, the model likelihood did not actually change much across the
17 different trial knots (see Figure D-3a of Appendix D). The overall fit of this two-piece spline
18 model was statistically significant (p = 0.048); however, this model yielded a very steep slope in
91
19 the exposure range below the knot of 100 ppm x days (see Figure 4-1), and there was low
20 confidence in the slope, given that it is based primarily on a relatively small number of cases in
21 the low-exposure region. Consideration was also given to the two-piece log-linear spline model
22 with a knot at 1,600 ppm x days, which provided a local maximum for the likelihood (see
23 Figure D-3a of Appendix D; see Table 4-3 and Section D.3 of Appendix D for parameter
24 estimates and fit statistics for the two spline models). However, the two-piece spline model with
25 the knot at 1,600 ppm x days was not the model with the maximum likelihood; the/?-value
26 exceeded 0.05, although minimally (p = 0.072); and the slope in the low-exposure range (i.e.,
27 below the knot), although not as steep as the low-exposure slope from the two-piece log-linear
28 spline model with the knot at 100 ppm x days, was still based primarily on a relatively small
29 number of cases in the low-exposure region and, thus, there was less confidence in this slope
30 than in that of the linear regression of the categorical results discussed above. Therefore, after
31 examining the new modeling analyses, it was determined that the weighted linear regression of
32 the categorical results still provided the best available approach for risk estimates for lymphoid
20As 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 (-p—1-0*.
21 Although the log-linear spline segments appear fairly linear in the plotted range, they are not strictly linear.
This document is a draft for review purposes only and does not constitute Agency policy.
4-11 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
cancer. Nonetheless, the two-piece log-linear spline model with the knot at 1,600 ppm x days
might be considered a reasonable alternative based on the continuous exposure data, and thus,
unit risk calculations using this model are included in the assessment for comparison purposes.
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
/7-valueb
Coefficient (SE)
(per ppm x day)
All lymphohematopoietic cancer0
2-piece log-linear spline (knot at 500 ppm x
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
2-piece log-linear spline with maximum
likelihood (knot at 100 ppm x days)
Alternative 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:
61=0.0004893(0.0002554)
0.000247 (0.000185)
10
11
12
13
14
15
16
17
18
19
20
21
aAll with cumulative exposure in ppm x 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.
C9^ 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.
For the weighted linear regression, the Cox regression results from the model with
categorical cumulative exposure and a 15-year lag (see Table 4-2) was used, excluding the
When this assessment was largely complete, 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 then 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.
4-12 DRAFT—DO NOT CITE OR QUOTE
-------
9^
1 highest exposure group, as discussed above. The weights used for the ORs were the inverses
2 of the variances, which were calculated from the confidence intervals.24 Mean and median
3 exposures for the cumulative exposure groups were provided by Dr. Steenland (see Table D-3a
9S
4 of Appendix D). The mean values were used for the weighted regression analysis because the
5 cancer response is presumed to be a function of cumulative exposure, which is expected to be
6 best represented by mean exposures. If the median values had been used, a slightly larger
7 regression coefficient would have been obtained, resulting in slightly larger risk estimates. See
8 Table 4-3 for the results obtained from the weighted linear regression and Figure 4-1 for a
9 depiction of the resulting model.
10 As the lymphoid cancer data set is the primary data set used for the derivation of unit risk
11 estimates for lymphohematopoietic cancers, a summary of all the models considered for
12 modeling the lymphoid cancer exposure-response data and the judgments made about model
13 selection is provided in Table 4-4. See Figures 4-1 and D-3c in Appendix D for visual
14 representations of the models. See Tables 4-2 and 4-3 and Section D.3 of Appendix D for other
15 information about the models.
23Concerns 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.
24Equations for this weighted linear regression approach are presented in Rothman(1986) and summarized in
Appendix F.
25Mean 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 x days. Median values were 374; 1,985; 6,755; and 26,373 ppm x 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.
4-13 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
Cox regression (log-linear) model
Cox regression model with log cumulative
exposure
2-piece log-linear spline with maximum
likelihood (knot at 100 ppm x days)
Alternative 2-piece log-linear spline (knot at
1,600 ppm x days)
linear model (RR = 1 + (3 x exposure)13
linear model with log cumulative exposure
2-piece linear spline model (knot at 100 ppm
x days)
Linear regression of categorical results,
excluding the highest exposure quartile
Comments
Inadequate overall statistical fit (p = 0.22) and poor visual fit
Good overall statistical fit (p = 0.02) but too steep in the low-
exposure region
Good overall statistical fit (p = 0.048) but too steep in the low-
exposure region
ALTERNATIVE. Not 2-piece spline model with maximum
likelihood and not statistically significant fit (p = 0.07) but
reasonable alternative model (see text above)
Inadequate overall statistical fit (p = 0.13) and poor visual fit
Good overall statistical fit (p = 0.02) but too steep in the low-
exposure region
Good overall statistical fit (p = 0.04) but too steep in the low-
exposure region
SELECTED. The continuous exposure supralinear models (e.g.,
the log-cumulative-exposure models and the 2-piece log-linear
spline model with the maximum likelihood) 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 (p = 0. 18), 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 exposure supralinear models, which establishes the
significance of the exposure-response correlation for the
underlying data.
5
6
7
aAll with cumulative exposure as the exposure variable, except where noted, and with a 15-yr lag.
bThe linear models are discussed in footnote #14 above.
9
10
11
12
13
14
15
The linear regression of the categorical results for males and females combined and the
actuarial program (life-table analysis) were used to estimate the exposure level (ECX; "effective
concentration") and the associated 95% lower confidence limit (LECX) corresponding to an extra
risk of 1% (x = 0.01). A 1% extra risk level is commonly used for the determination of the point
of departure (POD) for low-exposure extrapolation from epidemiological cancer data (except for
rare cancers); higher extra risk levels, such as 10%, would be an upward extrapolation for these
This document is a draft for review purposes only and does not constitute Agency policy.
4-14 DRAFT—DO NOT CITE OR QUOTE
-------
1 data. Thus, 1% extra risk was selected for determination of the POD, and, consistent with EPA's
2 Guidelines for Carcinogen Risk Assessment (U.S. EPA, 2005a), the LEG value corresponding to
3 that risk level was used as the POD to derive the cancer unit risk estimates.
4 Because EtO is DNA-reactive and has direct mutagenic activity (see Section 3.3.3),
5 which is one of the cases cited by EPA's Guidelines for Carcinogen Risk Assessment (U.S. EPA,
6 2005a) for the use of linear low-dose extrapolation, a linear low-exposure extrapolation was
7 performed. The ECoi, LECoi, and inhalation unit risk estimate calculated for lymphoid cancer
8 mortality from the linear regression of the categorical results are presented in Table 4-5 (the
9 incidence results also presented in Table 4-5 are discussed in Section 4.1.1.3 below). The
10 resulting unit risk estimate for lymphoid cancer mortality based on the linear regression of the
11 categorical results for both sexes using cumulative exposure with a 15-year lag is 0.397 per ppm.
12 The unit risk estimate from the alternative model, the two-piece log-linear spline model with the
13 knot at 1,600 ppm x days, is 0.917 per ppm, about 2.3-times higher than the estimate from the
14 preferred model. ECoi and LECoi estimates from the other models considered are presented for
15 comparison only, to illustrate the differences in model behavior at the low end of the
16 exposure-response range. Unit risk estimates are not presented for these other models because,
17 as discussed above, these models were deemed unsuitable for the derivation of risks from (low)
18 environmental exposure levels. The log cumulative exposure model, with its extreme
19 supralinearity in the lower exposure region, and the two-piece log-linear spline model with the
20 maximum likelihood (knot at 100 ppm x days), with its very steep low-exposure slope, yield
21 substantially lower ECoi estimates (0.00441 ppm and 0.000982 ppm, respectively). Converting
22 the units, the resulting unit risk estimate of 0.397 per ppm from the linear regression model of
23 the categorical results corresponds to a unit risk estimate of 2.17 x ICT4 per ug/m3 for lymphoid
24 cancer mortality.26
25 As discussed above, risk estimates based on the all lymphohematopoietic cancer results
26 are also derived for comparison. The same methodology presented above for the lymphoid
27 cancer results was used for the all lymphohematopoietic cancer risk estimates. Age-specific
28 background mortality rates for all lymphohematopoietic cancers for the year 2004 were obtained
29 from the NCHS Data Warehouse website (http://www.cdc.gov/nchs/datawh/statab/unpubd/
30 mortabs.htm). The results of Dr. Steenland's reanalyses using the Cox regression models
31 presented in the Steenland et al. (2004) paper with data for males and females combined are
32 presented in Table 4-2. As for lymphoid cancer and for all hematopoietic cancer in males
33 presented in the Steenland et al. (2004) paper, the only statistically significant Cox regression
Conversion equation: 1 ppm= 1,830 ug/m
This document is a draft for review purposes only and does not constitute Agency policy.
4-15 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
low-exposure log-linear spline from
2-piece spline model with
maximum likelihood (knot at 100
ppm x days),6
cumulative exposure,
15-yrlag
Alternative low-exposure log-linear
spline (knot at 1,600 ppm x days)/
cumulative exposure,
15-yrlag
Linear regression of categorical
results, cumulative exposure,
15-yrlag11
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
0.917
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
1.72g
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 estimates are 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 two-piece log-linear spline model with largest
likelihood (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 alternative two-piece log-linear spline model (local
maximum 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 e for calculation) and, thus, appropriately in
the range of the low-exposure segment.
8To obtain unit risk estimates less than 1, convert to risk per ppb (e.g., 1.72 per ppm = 1.72 x l(T3 per ppb).
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 e for calculation) above which the linear
regression model of the categorical results does not apply (see Figure 4-1).
This document is a draft for review purposes only and does not constitute Agency policy.
4-16 DRAFT—DO NOT CITE OR QUOTE
-------
1 model was for log cumulative exposure with a 15-year lag (p = 0.01). The cumulative exposure
2 model did not provide an adequate fit to the data and is not considered further here (p = 0.35).
3 Because of the problems with the supralinear log cumulative exposure model which are
4 discussed for the lymphoid cancers above, EPA again investigated the use of a two-piece
5 log-linear spline model to attempt to address the supralinearity of the data while avoiding the
6 extreme low-exposure curvature obtained with the log cumulative exposure model. For the all
7 lymphohematopoietic cancer mortality data, the range examined for knot selection was from 0 to
8 7,000 ppm x days, and the largest model likelihood was obtained with the knot at 500 ppm x
9 days (see Figure D-4a of Appendix D). See Table 4-3 and Section D.4 of Appendix D for
10 parameter estimates and fit statistics for the two-piece spline model. As with the lymphoid
11 cancer mortality results, however, this model resulted in an apparently excessively steep
12 low-exposure spline (see Figure 4-2), so, again, the linear regression model of the categorical
13 results was used to derive the cancer unit risk estimate for this data set.27
14 For the weighted linear regression, the results from the Cox regression model with
15 categorical cumulative exposure and a 15-year lag (see Table 4-2) were used, excluding the
16 highest exposure group, and the approach discussed above for lymphoid cancer mortality. See
17 Table 4-3 for the results obtained from the weighted linear regression and Figure 4-2 for a
18 graphical presentation of the resulting linear regression model. As discussed above, this linear
19 regression model was used to derive the unit risk estimates for all lymphohematopoietic cancer.
20
27When this assessment was largely complete, 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 then 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.
4-17 DRAFT—DO NOT CITE OR QUOTE
-------
o
C5
S
I
00 §
O
a
i
(-\
**ij
8"
e-(p'exp)
eA(p*logexp)
• categorical
2-piece spline
^^—^^— linear
15000 20000
cumulative exposure (ppm*days)
Figure 4-2. RR estimate for all lymphohematopoietic cancer vs. mean exposure (with 15-year lag, unadjusted
for continuous exposure).
3
o
H
W
/o
c
o
H
W
1
eA(p*exp): Cox regression results for RR = e(p x exposure); eA(p*logexp): Cox regression results for RR = e(p x ta(exP°sure»; categorical: Cox
regression results for RR = e(|3 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). (Note that, with the
exception of the categorical results and the linear regression of the categorical results, the different models have different implicitly estimated
baseline risks; thus, they are not strictly comparable to each other in terms of RR values, i.e., along the y-axis. They are, however, comparable
in terms of general shape.)
Source: Steenland reanalyses for males and females combined; see Appendix D (except for linear regression of the categorical results, which
was done by EPA).
-------
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
The ECoi, LECoi, and inhalation unit risk estimate calculated for all
lymphohematopoietic cancer mortality from the linear regression model of the categorical results
are presented in Table 4-6 (the incidence results also presented in Table 4-6 are discussed in
Section 4.1.1.3 below). The resulting unit risk estimate for all lymphohematopoietic cancer
mortality based on the linear regression of the categorical results for both sexes using cumulative
exposure with a 15-year lag is 0.680 per ppm. ECoi and LECoi estimates from the other models
considered are presented for comparison only, to illustrate the differences in model behavior at
the low end of the exposure-response range. Unit risk estimates are not presented for these other
models because, as discussed above, these models were deemed unsuitable for the derivation of
risks from (low) environmental exposure levels. The resulting unit risk estimate for all
lymphohematopoietic cancer mortality from the linear regression model of the categorical results
is similar to that for lymphoid cancer mortality (70% higher; see Table 4-5). Converting the
units, the resulting unit risk estimate of 0.680 per ppm corresponds to a unit risk estimate of
3.72 x 10 4 per ug/m3 for all lymphohematopoietic cancer mortality.
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,
15-yrlagf
Mortality
EC0i
(ppm)
0.00140
0.00377
0.0283
LECoi
(ppm)
0.000245
0.00231
0.0147
Unit risk
(per ppm)
d
d
0.680
Incidence1"
EC0i
(ppm)
0.000190
0.00216
0.0144
LECoi
(ppm)
0.0000753
0.00132
0.00746
Unit risk
(per ppm)
d
d
1.34s
20
21
22
23
aFrom lifetime continuous exposure. Unit risk = 0.01/LECM.
Incidence estimates presented here for comparison; they are derived in Section 4.1.1.3.
This document is a draft for review purposes only and does not constitute Agency policy.
4-19 DRAFT—DO NOT CITE OR QUOTE
-------
1 Table 4 6. EC01, LEC01, and unit risk estimates for all
2 lymphohematopoietic cancer" (continued)
3
4
5 °From Dr. Steenland's analyses for males and females combined (see Appendix D), Cox regression models. Note
6 that the EC0i and LECM results presented here will not exactly match those presented in Appendix D because,
7 although EPA used the regression coefficients reported by Dr. Steenland in Appendix D, the life-table analyses
8 using 2004 all-cause mortality rates were redone to be more up-to-date and consistent with the cause-specific
9 mortality rates; the results presented in Appendix D were based on life-table analyses using 2000 all-cause mortality
10 rates.
11 dUnit risk estimates are not presented for these models because these models were deemed unsuitable for the
12 derivation of risks from (low) environmental exposure levels (see text).
13 eUsing regression coefficient from low-exposure segment of two-piece log-linear spline model with knot at 500 ppm
14 x days; see text and Appendix D. Each of the ECOT values is below the value of 0.0064 ppm roughly corresponding
15 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]
16 and, thus, appropriately in the range of the low-exposure segment.
17 Degression coefficient derived from linear regression of categorical Cox regression results from Table 4-2, as
18 described in Section 4.1.1.2. Each of the ECOT values is appropriately below the value of 0.064 ppm roughly
19 corresponding to the value of about 5,000 ppm x days (see footnote d for calculation) above which the linear
20 regression model of the categorical results does not apply (see Figure 4-2).
21 8For unit risk estimates below 1, convert to risk per ppb (e.g., 1.34perppm= 1.34 x 10~3perppb).
22
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 commonly mortality data, as is the case in the Steenland et al. (2004)
28 study. For tumor sites with low survival rates, mortality-based estimates are reasonable
29 approximations of cancer incidence risk; however, for many lymphohematopoietic cancers, the
30 survival rate is substantial, and incidence risk estimates are preferred by EPA (U.S. EPA, 2005a).
31 Therefore, another calculation was done using the same regression coefficients presented
32 above (see Section 4.1.1.2), but with age-specific lymphoid cancer incidence rates for the
33 relevant subcategories of lymphohematopoietic cancer (NHL, myeloma, and lymphocytic
34 leukemia) for 2000-2004 from SEER (Ries et al.. 2007): Tables XIX, XVIII, XIII: both sexes,
35 all races) in place of the lymphoid cancer mortality rates in the actuarial program. SEER collects
36 good-quality cancer incidence data from a variety of geographical areas in the United States.
37 The incidence data used here are from "SEER 17," a registry of seventeen states, regions, and
38 cities covering about 26% of the U.S. population.
39 The incidence risk calculation assumes that (1) lymphoid cancer incidence and mortality
40 have the same exposure-response relationship for the relative rate of effect from EtO exposure
41 and that (2) the incidence data are for first occurrences of primary lymphoid cancer or that
42 relapses and secondary lymphoid cancers provide a negligible contribution. (The latter
43 assumption is probably sound; the former assumption is more potentially problematic. Because
This document is a draft for review purposes only and does not constitute Agency policy.
4-20 DRAFT—DO NOT CITE OR QUOTE
-------
1 various lymphoid cancer subtypes with different survival rates are included in the categorization
2 of lymphoid cancers, if the EtO-associated relative rates of the subtypes differ, a bias could
3 occur, resulting in either an underestimation or overestimation of the extra risk for lymphoid
9R
4 cancer incidence.) Potential concern that the incidence risk estimates might be overestimated
5 would come primarily from the inclusion of multiple myeloma, because that subtype has the
6 lowest incidence:mortality ratios (and, thus, if that subtype were driving the increased mortality
7 observed for the lymphoid cancer grouping, then including the incidence rates for the other
8 subtypes, which have higher incidence:mortality ratios, might inflate the incidence risk
9 estimates). Multiple myelomas, however, constitute only 25% of the lymphoid cancer cases in
10 the cohort, and there is no evidence that multiple myeloma is driving the EtO-induced excess in
9Q
11 lymphoid cancer mortality. Thus, using the total lymphoid cancer incidence rates is not
12 expected to result in an overestimation of the incidence risk estimates; if anything, the incidence
13 risks would likely be diluted with the inclusion of the multiple myeloma rates. The incidence
14 risk calculation also relies on the fact that the lymphoid cancer incidence rates (more
15 specifically, the differential rates obtained by subtracting the mortality rates from the incidence
16 rates) are small when compared with the all-cause mortality rates.30 The resulting ECoi and
17 LECoi estimates for lymphoid cancer incidence from the various models examined are presented
18 in Table 4-5. The unit risk estimate for lymphoid cancer incidence from the selected linear
19 regression model of the categorical results is 0.877 per ppm. The unit risk estimate for lymphoid
20 cancer incidence from the alternative model, the two-piece log-linear spline model with the knot
28Sielken 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.
29According 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 more than 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.
30Sielken 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.
4-21 DRAFT—DO NOT CITE OR QUOTE
-------
1 at 1,600 ppm x days, is 1.72 per ppm, about 2.0 times higher than the estimate from the preferred
2 model.
3 Overall, as discussed above, the preferred estimate for the unit risk for lymphoid cancer is
4 the estimate of 0.877 per ppm (4.79 x 10 4 per ug/m3) derived, using incidence rates for the
5 cause-specific background rates, from the weighted linear regression of the categorical results,
6 dropping the highest exposure group. The incidence unit risk estimate from the linear regression
7 model of the categorical results is about 120% higher than (i.e., 2.2 times) the mortality-based
8 estimate from the same model.
9 As discussed in Section 4.1.1.2, risk estimates based on the results of Dr. Steenland's
10 reanalyses of the all lymphohematopoietic cancer data (see Appendix D and Table 4-2) are also
11 derived for comparison. The same methodology presented above for the lymphoid cancer
12 incidence results was used for the all lymphohematopoietic cancer incidence risk estimates, and
13 the same assumptions apply. Age-specific SEER incidence rates for all lymphohematopoietic
14 cancer for the years 2000-2004 were used (Ries et al.. 2007): Tables XIX, IX, XVIII, and XIII:
15 both sexes, all races). The ECoi and LECoi estimates for all lymphohematopoietic cancer
16 incidence from the different all lymphohematopoietic cancer mortality models examined are
17 presented in Table 4-6. The resulting unit risk estimate for all lymphohematopoietic cancer
18 incidence from the linear regression of the categorical results is about 2.0-times the
19 mortality-based estimate and about 1.5-times the lymphoid cancer incidence estimate (see
20 Table 4-5).
21
22 4.1.2. Risk Estimates for Breast Cancer
23 4.1.2.1. Breast Cancer Results From the NIOSH Study
24 The Steenland et al. (2004) study discussed above in Section 4.1.1.1 also presents results
25 from exposure-response analyses for breast cancer mortality in female workers. Steenland et al.
26 (2003) present results of a breast cancer incidence study of a subcohort of the female workers
27 from the NIOSH cohort. In addition to the analyses presented in the Steenland et al. (2003) and
28 Steenland et al. (2004) papers, Dr. Steenland did subsequent analyses of the breast cancer
29 incidence and mortality data sets for EPA; these are discussed below and reported in
30 Sections D. 1 and D.2 of Appendix D, respectively.
31
32 4.1.2.2. Prediction of Lifetime Extra Risk of Breast Cancer Mortality
33 Results from the Cox regression models presented by Steenland et al. (2004), with some
34 reanalyses reported by Dr. Steenland in Appendix D (see Section D.2), are summarized in
35 Table 4-7. These models were considered for the derivation of unit risk estimates for breast
This document is a draft for review purposes only and does not constitute Agency policy.
4-22 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
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 cohort," 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)
(per ppm x day)
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)
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
"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).
U.S. 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
the actuarial program (life-table analysis). The National Center for Health Statistics 1997-2001
cause-specific background mortality rates for invasive breast cancers in females were obtained
from a SEER report (Ries et al., 2004). The risks were computed up to age 85 for continuous
exposures to EtO, conversions were made between occupational EtO exposures and continuous
environmental exposures, and 95% UCLs were calculated for the relative rates, as described
above.
The only statistically significant Cox regression model presented by Steenland et al.
(2004) for breast cancer mortality in females was for log cumulative exposure with a 20-year lag
(p = 0.01). However, as for the lymphohematopoietic cancers in Section 4.1.1, using the log
cumulative exposure model to estimate the risks from low environmental exposures is
problematic because this model is highly supralinear and results are unstable for low exposures
(see Figure 4-3). The cumulative exposure model, which is typically used and which is stable at
low exposures, was nearly statistically significant (p = 0.06 with a 20-year lag; see Section D.2
of Appendix D) in terms of the global fit to the data; however, the Cox regression model with
This document is a draft for review purposes only and does not constitute Agency policy.
4-23 DRAFT—DO NOT CITE OR QUOTE
-------
1 cumulative exposure is sublinear and does not reflect the apparent supralinearity of the breast
2 cancer mortality data (see Figure 4-3).
3 In a 2006 external review draft of this assessment (U.S. EPA, 2006a), which relied on the
4 original published results of Steenland et al. (2004), EPA proposed that the best way to reflect
5 the exposure-response relationship in the lower exposure region, which is the region of interest
6 for low-exposure extrapolation, was to do a weighted linear regression of the results from the
7 Cox regression model with categorical cumulative exposure and a 20-year lag. In addition, the
8 highest exposure group was not included in the regression to alleviate some of the "plateauing"
9 in the exposure-response relationship at higher exposure levels and to provide a better fit to the
10 lower exposure data. Linear modeling of categorical epidemiologic data and elimination of the
11 highest exposure group(s) in certain circumstances to obtain a better fit of low-exposure data are
12 both standard techniques used in EPA dose-response assessments (U.S. EPA, 2005a). However,
13 as discussed in Section 4.1.1.2 for the similarly supralinear lymphohematopoietic cancer data,
14 the SAB panel that reviewed the draft assessment recommended that EPA employ models using
15 the individual exposure data as an alternative to modeling the published grouped data (SAB,
16 2007). Consequently, it was determined that, using the full data set, an alternative way to
17 address the supralinearity of the data (while avoiding the extreme low-exposure curvature
18 obtained with the log cumulative exposure model) might be to use a two-piece spline model, and
19 Dr. Steenland was commissioned to do the spline analyses using the full data set with cumulative
20 exposure as a continuous variable. His findings are reported in Section D.2 of Appendix D, and
21 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.
4-24 DRAFT—DO NOT CITE OR QUOTE
-------
o
C5
S
I
O
a
i
------- e"(B*logexp)
^^^^^^ linear
• categorical
eA(B*exp)
^ ^ — spline/00
10000 15000 20000
mean cumulative exposure (ppm*days)
O 5
O I
3
o
H
W
o
c
o
H
W
1
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(|3 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 * days (see text). (Note that, with
the exception of the categorical results and the linear regression of the categorical results, the different models have different implicitly
estimated baseline risks; thus, they are not strictly comparable to each other in terms of RR values, i.e., along the y-axis. They are, however,
comparable in terms of general shape.)
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, with a 20-year lag), and thus, log RR is a function of two lines that
5 join at a single point of inflection, called a "knot." The shape of the two-piece log-linear spline
6 model, in particular the slope in the low-exposure region, depends on the location of the knot.
7 Knot selection was made by trying different knots over a reasonable range and choosing the one
8 that resulted in the largest model likelihood. For the breast cancer mortality data, the range
9 examined for knot selection was from 0 to 25,000 ppm x days, using increments of
10 100 ppm x days to 7,000 ppm x days and increments of 1,000 ppm x days above
11 7,000 ppm x days. The largest model likelihood was observed with the knot at 700 ppm x days,
12 although, as noted above, the model likelihood did not change much across the various trial
13 knots (see Figure D-2a of Appendix D).31 The overall fit of this two-piece spline model was not
14 statistically significant (p = 0.067). Of greater concern, this model yielded a very steep slope in
15 the exposure range below the knot of 700 ppm x days (see Figure 4-3), and confidence in the
16 slope was low, given that it is based primarily on relatively few cases in the low-exposure region.
17 Consideration was also given to the two-piece log-linear spline model with a knot at
18 13,000 ppm x days, which provided a local maximum for the likelihood (see Figure D-2a' of
19 Appendix D; see Table 4-8 and Section D.2 of Appendix D for parameter estimates and fit
^9
20 statistics for the two spline models). However, the two-piece spline model with the knot at
21 13,000 ppm x days was not the model with the maximum likelihood; its/>-value similarly
22 exceeded 0.05, although minimally (p = 0.074); and the large difference between the knots in the
23 maximum-likelihood and alternative two-piece spline models decreased confidence in this
24 alternative two-piece spline model. Therefore, after examining the new modeling analyses, it
25 was determined that the weighted linear regression of the categorical results still provided the
26 best available approach for risk estimates for lymphoid cancer.33 Nonetheless, the two-piece
27 log-linear spline model with the knot at 13,000 ppm x days might be considered a reasonable
31Using the two-piece log-linear spline model with the maximum likelihood (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).
32Using the alternative two-piece 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).
33When this assessment was largely complete, some linear models (i.e., RR = 1 + (3 x exposure) were investigated,
using the then just-published approach of Langholz and Richardson (2010). to model the individual data with
cumulative exposure as a continuous variable; however, these linear models did not alleviate the problems of the
corresponding log-linear RR models and the results are not presented here (see Section D.2.c and Figure D-2d in
Appendix D).
This document is a draft for review purposes only and does not constitute Agency policy.
4-26 DRAFT—DO NOT CITE OR QUOTE
-------
1 alternative based on the continuous exposure data, and thus, unit risk calculations using this
2 model are included in the assessment for comparison purposes. (For more discussion of the
3 breast cancer mortality exposure-response modeling using the continuous exposure data, see
4 Section D.2 of Appendix D.)
5 For the weighted linear regression, the results from the Cox regression model with
6 categorical cumulative exposure (and a 20-year lag) presented in Table 4-7 were used, excluding
7 the highest exposure group, and the approach discussed above for the lymphoid cancers (see
8 Section 4.1.1.2).34 Mean and median exposures for the cumulative exposure groups were
9 provided by Dr. Steenland (see Appendix D).35 See Table 4-8 for the results obtained from the
10 weighted linear regression of the categorical results and mean exposures and Figure 4-3 for a
11 depiction of the resulting linear regression model.
12 The linear regression of the categorical results and the actuarial program (life-table
13 analysis) were used to estimate the exposure level (ECX) and the associated 95% lower
14 confidence limit (LECX) corresponding to an extra risk of 1% (x = 0.01). As discussed in
15 Section 4.1.1.2, a 1% extra risk level is a more reasonable response level for defining the POD
16 for these epidemiologic data than 10%.
34Equations for this weighted linear regression approach are presented in Rothman(1986) and summarized in
Appendix F.
35Mean 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 * 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.
4-27 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
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)
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
Model3
2-piece log-linear spline with maximum
likelihood (knot at 700 ppm x days)
Alternative 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)
(per ppm x day)
low-exposure spline segment:
61=0.000688(0.000417)
low-exposure spline segment:
61=0.0000607(0.0000309)
0.000201 (0.000120)
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
performed. The ECoi, LECoi, and inhalation unit risk estimate calculated for breast cancer
mortality from the linear regression model of the categorical results are presented in Table 4-9.
The resulting unit risk estimate for breast cancer mortality based on the linear regression of the
categorical results using cumulative exposure with a 20-year lag is 0.513 per ppm. The unit risk
estimate from the alternative model, the two-piece log-linear spline model with the knot at
13,000 ppm x days, is 0.172 per ppm, about 34% of the estimate from the preferred model. ECoi
and LECoi estimates from the other models considered are presented for comparison only, to
illustrate the differences in model behavior at the low end of the exposure-response range. Unit
risk estimates are not presented for these other models because, as discussed above, these models
were deemed unsuitable for the derivation of risks from (low) environmental exposure levels.
As one can see, the standard Cox regression cumulative exposure model, with its extreme
sublinearity in the lower exposure region, yields a substantially higher ECoi estimate
(0.530 ppm) than the ECoi estimate of 0.0387 ppm from the linear regression of the categorical
results, while the log cumulative exposure Cox regression model, with its extreme supralinearity
in the lower exposure region, yields a substantially lower ECoi estimates (0.00112 ppm). The
estimates from the two-piece log-linear spline models flank the result from the linear regression
of the categorical results more closely. The steep low-exposure segment of the two-piece
This document is a draft for review purposes only and does not constitute Agency policy.
4-28 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
8
9
10
log-linear spline model with the maximum likelihood (knot at 700 ppm x days) yields an ECoi
estimate of 0.00941 ppm, whereas the shallower low-exposure slope from the two-piece
log-linear spline model with the local maximum likelihood suggesting a knot at 13,000 ppm x
days yields an ECoi estimate of 0.107 ppm. Converting the units, the unit risk estimate of 0.513
per ppm for breast cancer mortality from the linear regression model of the categorical results
corresponds to a unit risk estimate of 2.80 x 10 4 per ug/m3.
Table 4-9. ECoi, LECoi, and unit risk estimates for breast cancer mortality
in females11
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
EC0i
(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
0.172
0.513
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
"From lifetime continuous exposure. Unit risk = 0.01/LEC0i.
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 EC0i 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 ECW 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).
This document is a draft for review purposes only and does not constitute Agency policy.
4-29 DRAFT—DO NOT CITE OR QUOTE
-------
1 4.1.2.3. Prediction of Lifetime Extra Risk of Breast Cancer Incidence
2 As discussed in Section 4.1.1.3, risk estimates for cancer incidence are preferred to
3 estimates for cancer mortality, especially for cancer types with good survival rates, such as breast
4 cancer. In the case of female breast cancer in the NIOSH cohort, there is a corresponding
5 incidence study (Steenland et al., 2003) with exposure-response results for breast cancer
6 incidence, so one can estimate cancer incidence risks directly rather than estimate them from
7 mortality data. The incidence study used a (sub)cohort of 7,576 (76%) of the female workers
8 from the original cohort. Cohort eligibility for the incidence study was restricted to the female
9 workers who had been employed at 1 of the 14 plants for at least 1 year, owing to cost
10 considerations and the greater difficulties in locating workers with short-term employment.
11 Interviews were sought from all the women in the incidence study cohort or their next-of-kin
12 (18% of the cohort had died). Completed interviews were obtained for 5,139 (68%) of the
13 7,576 women in the cohort. The investigators also attempted to acquire breast cancer incidence
14 data for the cohort from cancer registries (available for 9 of the 11 states in which the plants were
15 located) and death certificates; thus, results are presented for both the full cohort (n = 7,576) and
16 the subcohort of women with interviews (n = 5,139). For additional details and discussion of the
17 Steenland et al. (2003) study, see Section A.2.16 of Appendix A.
18 Steenland et al. (2003) identified 319 incident cases of breast cancer in the cohort through
19 1998. Interview (questionnaire) data were available for 73% (233 cases). Six percent were
20 carcinoma in situ (20 cases). Steenland et al. (2003) performed internal exposure-response
21 analyses similar to those described in their 2004 paper and in Section 4.1.1.1 above. Controls for
22 each case were selected from the cohort members without breast cancer at the age of diagnosis of
23 the case. Cases and controls were matched on race. Of the potential confounders evaluated for
24 those with interviews, only parity and breast cancer in a first-degree relative were important
25 predictors of breast cancer, and only these variables were included in the final models for the
26 subcohort analyses. In situ cases were included with invasive breast cancer cases in the analyses;
27 however, the in situ cases represent just 6% of the total, and excluding them reportedly did not
28 greatly affect the results.
29 From the Steenland et al. (2003) internal analyses (Cox regression) using the full cohort,
30 the best-fitting model with exposure as a continuous variable was for (natural) log cumulative
31 exposure, lagged 15 years (p = 0.05). Duration of exposure, lagged 15 years, provided a slightly
32 better fitting model. Models using maximum or average exposure did not fit as well. In
33 addition, use of a threshold model did not provide a statistically significant improvement in fit.
34 For internal analyses using the subcohort with interviews, the cumulative exposure and log
35 cumulative exposure models, both lagged 15 years, and the log cumulative exposure model with
This document is a draft for review purposes only and does not constitute Agency policy.
4-30 DRAFT—DO NOT CITE OR QUOTE
-------
1 no lag all fit almost equally well, and the duration of exposure (also lagged 15 years) model fit
2 slightly better. Results of the Cox regression analyses for the cumulative and log cumulative
3 exposure models, with 15-year lags, are shown in Table 4-10, and these are the results
4 considered for the unit risk calculations. The models using duration of exposure are less useful
5 for estimating exposure-related risks, duration of exposure and cumulative exposure are
6 correlated, and the fits for these models are only marginally better than those with cumulative
7 exposure. The log cumulative exposure model with no lag was considered less biologically
9
10
11
12
Table 4-10. Cox regression results for breast cancer incidence in females
from the NIOSH cohort, for the models presented by Steenland et al.
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)
(per ppm x day),
/7-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)
13
14
15
16
17
18
19
20
21
22
23
24
25
26
alnvasive 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 * 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).
This document is a draft for review purposes only and does not constitute Agency policy.
4-31 DRAFT—DO NOT CITE OR QUOTE
-------
1 realistic than the corresponding model with a 15-year lag because some lag period would be
2 expected for the development of breast cancer. Furthermore, although initial risk estimates
3 based on the full cohort results are calculated for comparison, the preferred estimates are those
4 based on the subcohort with interviews because the subcohort should have more complete case
5 ascertainment and has additional information available on potential breast cancer confounders.
6 For the actuarial program (life-table analysis), U.S. age-specific all-cause mortality rates
7 for 2004 for females of all race groups combined (Arias, 2007) were used to specify the all-cause
8 background mortality rates. Because breast cancer incidence rates (more specifically, the
9 differential rates obtained by subtracting the mortality rates from the incidence rates) are not
10 negligible compared to all-cause mortality rates, the all-cause mortality rates in the life-table
11 analysis were adjusted to reflect women dying or being diagnosed with breast cancer in a given
12 age interval. All-cause mortality rates and breast cancer incidence rates were summed, and
13 breast cancer mortality rates were subtracted so that those dying of breast cancer were not
14 counted twice (i.e., as deaths and as incident cases of breast cancer). The National Center for
15 Health Statistics 2002-2006 mortality rates for invasive breast cancer in females were obtained
16 from a SEER report (Horner et al., 2009). The SEER report also provided SEER-17 incidence
17 rates for invasive and in situ breast cancer. The Cox regression results reported by Steenland et
18 al. (2003) are for invasive and in situ breast cancers combined. It is consistent with EPA's
19 Guidelines for Carcinogen Risk Assessment (U.S. EPA, 2005a) to combine these two tumor
20 types because the in situ tumors can progress to invasive tumors. Thus, the primary risk
21 calculations in this assessment use the sum of invasive and in situ breast cancer incidence rates
22 for the cause-specific background rates. Comparison calculations were performed using just the
23 invasive breast cancer incidence rates for the cause-specific rates; this issue is further discussed
24 in Section 4.1.3 on sources of uncertainty. The risks were computed up to age 85 for continuous
25 exposures to EtO, conversions were made between occupational EtO exposures and continuous
26 environmental exposures, and 95% UCLs were calculated for the relative rates, as described in
27 Section 4.1.1.2 above.
28 For breast cancer incidence in both the full cohort (see Figure 4-4) and the subcohort with
29 interviews (see Figure 4-5), the low-exposure categorical results suggest a more linear
30 low-exposure exposure-response relationship than that obtained with either the continuous
31 variable log cumulative exposure (supralinear) or cumulative exposure (sublinear) Cox
32 regression models. Thus, as with the lymphohematopoietic cancer and the breast cancer
33 mortality results above, EPA proposed in the 2006 Draft Assessment (U.S. EPA, 2006a), which
34 relied on the original published results of Steenland et al. (2003), that the best way to reflect the
35 data in the lower exposure region, which is the region of interest for low-exposure extrapolation,
This document is a draft for review purposes only and does not constitute Agency policy.
4-32 DRAFT—DO NOT CITE OR QUOTE
-------
cf
rs
§
^
.>!
§
i
a.
1.5
1
1
n: 1.4
/
eA(p*exp)
• — eA(p*logexp)
• categorical
15000 20000 25000
mean cumulative exposure (ppm * days)
o
2
H
W
O
&
O
c
o
H
W
Figure 4-4. RR estimate for breast cancer incidence in full cohort 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 ta(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). (Note that, with the exception of the categorical results and the linear regression of the categorical results, the
various models have different implicitly estimated baseline risks; thus, they are not strictly comparable to each other in terms of RR values, i.e..
along the y-axis. They are, however, comparable in terms of general shape.)
Source: Steenland et al. (2003) (except for linear regression of the categorical results, which was done by EPA).
-------
§•
— eA([3*exp)
cA((5*logexp)
— linear regression
categorical
— loglinearspline
- linearspline
eA(B*sqrtexp)
- linear(l+(3*exp)
1+ [3*logexp
5000 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 exposure).
eA(P*logexp): Cox regression results for RR = e'
(p x ln(exposure)).
; categorical: Cox
regression results for RR = e(p x exP°sure) with categorical exposures; eA(P*sqrtexp): Cox regression results for RR = e(p x s
-------
1 was to do a weighted linear regression of the results from the model with categorical cumulative
2 exposure (with a 15-year lag). In addition, the highest exposure group was not included in the
3 regression to provide a better fit to the lower-exposure data (The RR estimates for the highest
4 exposure quintiles suggest somewhat supralinear exposure-response relationships for both the
5 full cohort and the subcohort with interviews, and supralinearity is evidenced in the subcohort
6 with interviews by the strong influence of the top 5% of cumulative exposures on dampening the
7 slope of the [cumulative exposure] Cox regression model [see Section D.I and Figure D-ld of
8 Appendix D]. Moreover, there is more uncertainty in using the mean cumulative exposure to
9 represent the range of exposures in a highest exposure categorical group because such groups
10 contain a wider range of exposures; for example, for the subcohort with interviews, the highest
11 exposure quintile contains exposures ranging from about 14,500 ppm x days to over
12 250,000 ppm x days). Linear modeling of categorical (i.e., grouped) epidemiologic data and
13 elimination of the highest exposure group(s) under certain circumstances to obtain a better fit of
14 low-exposure data are both standard techniques used in EPA dose-response assessments (U.S.
15 EPA, 2012, 2005a). However, as discussed in Section 4.1.1.2 for the lymphohematopoietic
16 cancer data, the SAB panel that reviewed the draft assessment recommended that EPA not rely
17 on the published grouped data but, rather, do additional analyses using the individual data (SAB,
18 2007).
19 Consequently, it was determined that using the individual data, a better way to address
20 the apparent supralinearity of the data (while avoiding the extreme low-exposure curvature
21 obtained with the log cumulative exposure Cox regression model) might be to use a two-piece
22 spline model, and Dr. Steenland was commissioned to do the spline analyses. His findings are
23 reported in Appendix D (see Section D. 1), and the results for the breast cancer incidence
24 analyses are summarized below. Note that, for the two-piece spline analyses, only the data from
25 the subcohort with interviews and for the invasive and in situ breast cancers combined were
26 analyzed, because this was the preferred data set, as discussed above. (Dr. Steenland also
27 employed a cubic spline model as a semiparametric approach to visualize the underlying
28 exposure-response relationship; however, this approach produces an overly complicated function
29 for an empirical model, as opposed to a biologically based model, and was not used for risk
30 assessment purposes. In addition, Dr. Steenland investigated the use of a Cox regression model
31 with a square-root transformation of cumulative exposure; however, this approach, though less
32 extreme than using the log transformation of cumulative exposure, also yields a notably
33 supralinear model [see Figure 4-5], which can result in unstable low-exposure risk estimates.
34 The model results for both the cubic spline and square-root transformation models are included
35 in Appendix D, Section D.I, but are not considered further here. EPA chose to pursue the
This document is a draft for review purposes only and does not constitute Agency policy.
4-3 5 DRAFT—DO NOT CITE OR QUOTE
-------
1 development of two-piece spline models to attempt to avoid the problem of unstable risk
2 estimates from supralinear curvature in the low-exposure region because these models provide a
3 more general and systematic approach to modeling supralinear exposure-response data, as
4 opposed to using random, arbitrary power-transformations of the exposure variable.)
5 For the two-piece log-linear spline modeling approach, as described in Section 4.1.1.2
6 and discussed more fully in Appendix D, the Cox regression model was the underlying basis for
7 the splines which were fit to the breast cancer incidence exposure-response data (cumulative
8 exposure is used here, with a 15-year lag), and, thus, log RR is a function of two lines which join
9 at a single point of inflection, called a "knot." The shape of the two-piece spline model, in
10 particular the slope in the low-exposure region, depends on the location of the knot. For breast
11 cancer incidence, the knot was selected by evaluating different knots from 100 to
12 15,000 ppm x days in increments of 100 ppm x days and then choosing the one that resulted in
13 the best (largest) model likelihood. The model likelihood did not actually change much across
14 the different trial knots (see Figure D-la of Appendix D), but it did change slightly, and a knot of
15 5,800 ppm x days was chosen for the breast cancer incidence data based on the largest
16 likelihood. The two-piece log-linear spline model with this knot provided a statistically
17 significant fit to the data (p = 0.01 for the addition of the exposure terms; see Table D-lg in
18 Appendix D), as well as a good visual fit (see Figure 4-5).
19 A two-piece linear spline model was also fitted, using the just-published approach of
20 Langholz and Richardson (2010). This model is similar to the log-linear spline model discussed
21 above; however, for the linear spline model, the underlying basis for the splines is a linear model
22 (i.e., RR = 1 + P x z, where z represents the covariate data, including exposure, and P are the
23 parameters being estimated). The knot was selected as for the log-linear spline model, and the
24 same knot of 5,800 ppm x days yielded the largest likelihood (see Figure D-lh of Appendix D)
25 and was also chosen for the two-piece linear spline model. The two-piece linear spline model
26 with this knot provided a statistically significant fit to the data (p = 0.002 for the addition of the
27 exposure terms), as well as a good visual fit (see Figure 4-5). Because this model provided a
28 better fit than the log-linear spline model, i.e., it had a lower AIC, the two-piece linear spline
29 model was selected as the preferred model for the unit risk estimates for breast cancer incidence.
30 See Table 4-11 and Section D.I of Appendix D for parameter estimates and fit statistics for the
31 two spline models.
32 Linear RR models with cumulative exposure and log cumulative exposure as continuous
33 variables were also investigated using the approach of Langholz and Richardson (2010), and
34 these models fit better than the corresponding log RR models (see Table 4-11 and Section D. 1 of
35 Appendix D) although not as well as the two-piece linear spline model, which had the lowest
This document is a draft for review purposes only and does not constitute Agency policy.
4-36 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
AIC. Risk estimates based on the linear model with cumulative exposure are developed for
comparison, but the linear model with log cumulative exposure is too steep in the low-exposure
region (see Figure 4-5) and is not considered further. For more details of the breast cancer
incidence exposure-response modeling, see Section D.I of Appendix D.
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
/7-valueb
Coefficient (SE)
(per ppm x day)
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 x
days)
2-piece linear spline (knot at 5,800 ppm x 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:
Bl = 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.
Risk estimates based on the original linear regression analyses of the categorical results
are also presented for comparison. For the approach of using a weighted linear regression of the
results from the Cox regression model with categorical cumulative exposure (and a 15-year lag),
excluding the highest exposure group, the weights used for the ORs were the inverses of the
variances, which were calculated from the confidence intervals.36 Mean and median exposures
36Equations for this weighted linear regression approach are presented in Rothman (1986) and summarized in
Appendix F.
This document is a draft for review purposes only and does not constitute Agency policy.
4-37 DRAFT—DO NOT CITE OR QUOTE
-------
1 for the cumulative exposure groups for the full cohort were kindly provided by Dr. Steenland
2 (email dated April 21, 2004, from Kyle Steenland, Emory University, to Jennifer Jinot, EPA).37
3 The mean values were used for the weighted regression analysis because the (arithmetic) mean
4 exposures best represent the model's linear relationship between exposure and cancer response.
5 Differences between means and medians were not large for the females, especially for the lower
6 four quintiles. If the median values had been used, a slightly larger regression coefficient would
7 have been obtained, resulting in slightly larger risk estimates. Although the exposure values are
8 for risk sets from the full cohort, they should be reasonably close to the values for the subcohort
9 with interviews. See Table 4-11 for the results from the weighted linear regressions of the
10 categorical results and Figures 4-4 and 4-5 for a depiction of the resulting linear regression
11 models.
12 As the subcohort with interviews from the NIOSH incidence study cohort provides the
13 preferred data set for the derivation of unit risk estimates for breast cancer, a summary of all the
14 models considered for modeling the breast cancer exposure-response data from the subcohort
15 and the judgments made about model selection is provided in Table 4-12. See Figure 4-5 for
16 visual representations of the models. See Tables 4-10 and 4-11 and Section D.I of Appendix D
17 for parameter estimates, ^-values, and other fit statistics. Three of the models presented in
18 Table 4-12 had a good overall statistical fit, a good visual fit, and a credible low-exposure slope
19 (the linear and log-linear two-piece spline models and the [continuous exposure] linear RR
20 model). To better compare these models, they are plotted again in Figure 4-6 this time against
21 the categorical data in deciles. Earlier categorical results in this assessment were based on the
22 (log-linear) Cox regression model; however, the deciles in Figure 4-6 are based on a linear RR
23 categorical model—this model had a lower AIC than the log-linear decile model (1963.94 vs.
24 1966.91), and it provides a statistically significant fit to the data (p = 0.004), so the deciles
25 should provide a good representation of the data for the purposes of comparing the models (the
26 decile results from the log-linear and linear RR categorical models and the mean cumulative
27 exposure estimates for the deciles are presented in Section D. 1 of Appendix D). As can be seen
28 in Figure 4-6, the two-piece linear spline model, in addition to having the lowest AIC (see
29 Table 4-12), appears to have a better fit to the lower-exposure data, which are of the greatest
30 interest in estimating low-exposure risk. It also appears from Figure 4-6 that the linear model
31 has a poorer fit (too shallow) to the lower-exposure data than either of the two-piece spline
32 models. This is consistent with the analysis presented in Section D.I of Appendix D
33
37Mean 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.
4-3 8 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
systematic 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
aAll with cumulative exposure as the exposure variable, except where noted, and with a 15-yr lag.
bAIC = 2p-2LL, where p = #of parameters and LL = In(likelihood), assuming two exposure parameters for the
two-piece spline models.
°Not calculated.
This document is a draft for review purposes only and does not constitute Agency policy.
4-39 DRAFT—DO NOT CITE OR QUOTE
-------
o
C5
S
I
TO
'TS
.>!
§
i
-
§•
fe I
f 1
o I
o I
o ^
H »
O ^
HH ^3
I i o
W ,^
/o
c
o
H
W
2.6
loglinearspline
• linear spline
linear {l+p*exp)
categorical
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
15-year lag, unadjusted for continuous exposure); select models compared to deciles.
Categorical: linear model (RR = 1 + (3 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 + (3 x exposure, with exposure as a continuous variable. (Note that, with the exception of
the categorical results and the linear regression of the categorical results, the various models have different implicitly estimated baseline risks;
thus, they are not strictly comparable to each other in terms of RR values, i.e., along the y-axis. They are, however, comparable in terms of
general shape.)
Source: Steenland analyses in Appendix D.
-------
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
showing the strong influence of the upper tail of cumulative exposures on the results of the
cumulative exposure Cox regression model. The responses in the upper tail of exposures are
relatively dampened, such that when the highest 5% of exposures are excluded, the slope of the
Cox regression model is substantially increased (e.g., at 10,000 ppm x days, the RR estimate
increases from about 1.1 to almost 1.5; see Figure D-ld in Appendix D). This strong influence
of the upper tail of exposures would similarly attenuate the slope of the linear model. The
two-piece spline models, on the other hand, are more flexible, and the influence of the upper tail
of exposures would be primarily on the upper spline segment; thus, the two-piece models are
able to provide a better fit to the lower-exposure data.
The exposure level (ECX) and the associated 95% lower confidence limit (LECX)
corresponding to an extra risk of 1% (x = 0.01) for breast cancer incidence in females (based on
invasive + in situ tumors in the subcohort with interviews) for the models discussed above
(except for the square-root model and the linear model with log cumulative exposure, which
were not considered further) were estimated using the actuarial program (life-table analysis). As
noted 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 a 10% level. The results are presented in Table 4-13.
Table 4-13. ECoi, LECoi, and unit risk estimates for breast cancer incidence
in females—invasive and in situ"
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
With interviews
ECoi
(ppm)
0.135
0.0000765
0.0257
0.0166
LECoi
(ppm)
0.0788
0.0000422
0.0118
0.00991
Unit risk
(per ppm)
c
c
0.847
1.01f
Full cohort
ECoi
(ppm)
0.237
0.000124
0.0503
LECoi
(ppm)
0.115
0.0000529
0.0188
Unit risk
(per ppm)
C
c
0.532
__g
21
This document is a draft for review purposes only and does not constitute Agency policy.
4-41 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
Table 4-13. ECoi, LECoi, and unit risk estimates for breast cancer incidence
in females—invasive and in situa (continued)
4
5
6
7
Linear model with
continuous
cumulative exposure,
15-yrlag11
Low-exposure linear
spline, cumulative
exposure,
15-yrlage
0.0437
0.0112
0.0224
0.00576
0.4461
1.74fj
__g
__g
aAll-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.
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; two-piece spline and linear RR models not developed for the full cohort.
hFrom linear analyses in Section D.l.b.2 and Table D-li of Appendix D.
'Confidence intervals used in deriving the LECoi s 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.
J 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 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 UECM 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.
10
This document is a draft for review purposes only and does not constitute Agency policy.
4-42 DRAFT—DO NOT CITE OR QUOTE
-------
1 Because EtO is DNA-reactive and has direct mutagenic activity (see Section 3.3.3),
2 which is one of the cases cited by EPA's Guidelines for Carcinogen Risk Assessment (U.S. EPA,
3 2005a) for the use of linear low-dose extrapolation, a linear low-exposure extrapolation was
4 performed. The inhalation unit risk estimates for the different breast cancer incidence models
5 considered suitable for low-exposure extrapolation are presented in Table 4-13. As discussed
6 above, the unit risk estimate based on the two-piece linear spline model using cumulative
7 exposure with a 15-year lag (i.e., 1.74 per ppm, or 1.74 x 1CT3 per ppb) is the preferred estimate.
8 The two-piece log-linear spline model resulted in a unit risk estimate of 1.01 per ppm, while the
9 linear regression of categorical results yielded a unit risk estimate of 0.847 per ppm and the
10 continuous exposure linear model produced a unit risk estimate of 0.446 per ppm; these alternate
11 estimates are about 60%, 50%, and 25%, respectively, of the estimate based on the preferred
12 two-piece linear spline model. ECoi and LECoi estimates from the other models examined are
13 presented for comparison only, to illustrate the differences in model behavior at the low end of
14 the exposure-response range. Unit risk estimates are not presented for these other models
15 because, as discussed above, the log cumulative exposure Cox regression model was considered
16 overly supralinear and the cumulative exposure Cox regression model was considered overly
17 sublinear for the data in the lower exposure range (e.g., first 4 quintiles of exposure). As one can
18 see from the results for the subcohort with interviews, the standard Cox regression cumulative
19 exposure model, with its extreme sublinearity in the lower exposure region, yields a notably
20 higher ECoi estimate (0.135 ppm) than that from the two-piece linear spline model (0.0112 ppm),
21 while the log cumulative exposure model, with its extreme supralinearity in the lower exposure
22 region, yields a substantially lower ECoi estimate (0.0000765 ppm). Converting the units, the
23 preferred unit risk estimate of 1.74 per ppm corresponds to an estimate of 9.51 x 1CT4 per ug/m3
24 for breast cancer incidence.
25 As discussed above, the primary risk calculations for breast cancer incidence were based
26 on invasive and in situ tumors in the subcohort of women with interviews, and the primary
27 model was the two-piece linear spline model. For this assessment, the two-piece spline analyses
28 were not performed with the full cohort and the life-table analyses were not replicated for the
29 invasive cancers only. In the 2006 Draft Assessment (U.S. EPA, 2006a), however, comparison
30 analyses were done. Using the linear regression of the categorical results, the comparable unit
31 risk estimate for the full cohort was about 40% lower than the estimate based on the subcohort
32 with interviews. The corresponding unit risk estimate derived based on the subcohort results but
33 using invasive breast cancer only for the background incidence rates was about 17% lower than
34 the estimate based on invasive and in situ tumors, reflecting the difference between incidence
35 rates for invasive breast cancer only and for combined in situ and invasive breast cancer.
This document is a draft for review purposes only and does not constitute Agency policy.
4-43 DRAFT—DO NOT CITE OR QUOTE
-------
1 The unit risk estimate of 1.74 per ppm (1.74 x 10 3 per ppb) is the preferred estimate for
2 female breast cancer risk because it is based on incidence data versus mortality data, it is based
3 on more cases (n = 233) than the mortality estimate (n = 103), and information on personal
4 breast cancer risk factors obtained from the interviews is taken into account. Furthermore, the
5 two-piece linear spline model, which uses the complete data set with exposure as a continuous
6 variable, was statistically significant and had the lowest AIC and the best apparent visual fit to
7 the lower-exposure data of the models considered. Converting the units, 1.74 per ppm
8 corresponds to a unit risk estimate of 9.51 x 1CT4 per ug/m3.
9
10 4.1.3. Total Cancer Risk Estimates
11 According to EPA's Guidelines for Carcinogen Risk Assessment (U.S. EPA, 2005a),
12 cancer risk estimates are intended to reflect total cancer risk, not site-specific cancer risk;
13 therefore, an additional calculation was made to estimate the combined risk for (incident)
14 lymphoid and breast cancers, because females would be at risk for both cancer types. Assuming
15 that the cancer types are independent and that the risk estimates are approximately normally
16 distributed, one can estimate the 95% UCL (one-sided) on the total risk as the 95% UCL on the
17 sum of the maximum likelihood estimates (MLEs) of the risk estimates according to the formula
18
19
20 95% UCL = MLE + 1.645(SE), (4-3)
21
22
23 where MLE is the MLE of total cancer risk (i.e., the sum of the individual MLEs) and the SE of
24 the sum of the MLEs is the square root of the sum of the individual variances (i.e., the variance
25 of the sum is the sum of the variances, and the SE is the square root of the variance). Because
26 both models are linear in the range around the PODs, the combining-risk calculations can be
27 done directly from the unit risk and 0.01/ECoi estimates rather than having to do the calculations
28 at a common exposure level near where the ECoi and LECoi for the combined risk would be.
29 First, an ECoi of 0.0078 ppm for the total cancer risk (i.e., lymphoid cancer incidence + breast
30 cancer incidence) was estimated, as summarized in Table 4-14.
31
32
33
This document is a draft for review purposes only and does not constitute Agency policy.
4-44 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
Table 4-14. Calculation of ECoi for total cancer risk
4
5
6
7
8
9
10
11
12
13
14
Cancer type
Lymphoid
Breast
Total3
EC01
(ppm)
0.0254
0.0112
-
0.01/ECoi
(per ppm)
0.394
0.893
1.29
EC0i for total
cancer risk
(ppm)
-
-
0.00775
"The 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/ECM).
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
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
aSE = (unit risk - 0.017EC0i)/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).
Thus, the total cancer unit risk estimate is 2.3 per ppm (or 2.3 x 10 3 per ppb;
1.2 x 10 3 per ug/m3). Recall that this is the unit risk estimate derived under the assumption that
RR is independent of age (see Section 4.1.1.2). The preferred assumption of increased early-life
susceptibility, in accordance with EPA's Supplemental Guidance (U.S. EPA, 2005b), is
considered in Section 4.4. 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 individual MLEs (i.e., 1.29)
and the sum of the individual 95% UCLs (i.e., unit risk estimates, 2.6), or more precisely in this
This document is a draft for review purposes only and does not constitute Agency policy.
4-45 DRAFT—DO NOT CITE OR QUOTE
-------
1 case, between the largest individual unit risk estimate (1.74) and the sum of the unit risk
2 estimates (2.6). Thus, any inaccuracy in the total cancer risk estimate resulting from the
3 approach used to combine risk estimates across cancer types is relatively minor.
4
5 4.1.4. Sources of Uncertainty in the Cancer Risk Estimates
6 The two major sources of uncertainty in quantitative cancer risk estimates are generally
7 interspecies extrapolation and high-dose to low-dose extrapolation. The risk estimates derived
8 from the Steenland et al. (2003). Steenland et al. (2004). and additional Steenland (see
9 Appendix D) analyses are not subject to interspecies uncertainty because they are based on
10 human data. Furthermore, the human-based estimates are less affected by high-dose to low-dose
11 extrapolation than are rodent-based estimates and, thus, uncertainty from that source is reduced
12 somewhat. For example, the average exposure in the NIOSH cohort was more than 10 times
13 lower than the lowest exposure level in a rodent bioassay after adjustment to continuous lifetime
14 exposure. Nonetheless, uncertainty remains in the extrapolation from occupational exposures to
15 lower environmental exposures. Although the actual exposure-response relationship at low
16 exposure levels is unknown, the clear evidence of EtO mutagenicity supports the linear
17 low-exposure extrapolation that was used (U.S. EPA, 2005a).
18 Because of the existence of endogenous EtO (see Section 3.3.3.1), several members of
19 the SAB panel that reviewed EPA's external review draft assessment felt that the
20 exposure-response relationship for cancer at low exposures would be nonlinear and suggested
21 that it would be consistent with EPA's 2005 Guidelines for Carcinogen Risk Assessment (U.S.
22 EPA, 2005a) to present a nonlinear approach for "extrapolation" to lower exposures (SAB,
23 2007). EPA considered this suggestion but judged that the support for a nonlinear approach was
24 inadequate. In brief, as discussed in Sections 3.1 through 3.3.3, EtO is a DNA-reactive,
25 mutagenic, multisite carcinogen in humans and experimental species; as such, it has the
26 hallmarks of a compound for which low-dose linear extrapolation is strongly supported under
27 EPA's Guidelines for Carcinogen Risk Assessment (U.S. EPA, 2005a). EPA's Guidelines for
28 Carcinogen Risk Assessment (U.S. EPA, 2005a) do provide for presenting alternate approaches
29 when those alternatives have significant biological support; however, EPA's analysis of the
30 arguments for using a nonlinear approach presented on page 23 and in Appendix C of the SAB
31 report (SAB, 2007) did not find these arguments to be persuasive. The arguments posited by the
32 SAB panel members who supported using a nonlinear approach were largely that (1) DNA
33 adducts may show a nonlinear response when identical adducts are formed endogenously and
34 (2) mutations do not have linear relationships with exposure but exhibit an "inflection point."
35 However, as discussed in Section 3.3.3.1, recent data from Marsden et al. (2009) support a linear
This document is a draft for review purposes only and does not constitute Agency policy.
4-46 DRAFT—DO NOT CITE OR QUOTE
-------
1 exposure-response relationship for EtO exposure and DNA adducts (p < 0.05) and demonstrate
2 increases of DNA adducts from exogenous EtO exposure above those from endogenous EtO for
3 very low exposures to exogenous EtO, providing evidence against argument (1). Moreover,
4 Appendix C of the SAB report (SAB, 2007) presents two EtO-specific mutation data sets in
5 support of argument (2); however, EPA's analysis of these data sets finds that they are in fact
6 consistent with low-dose linearity. See the response to this comment under charge question 2.b
7 in Appendix H for a more comprehensive discussion of EPA's consideration and rejection of a
8 nonlinear approach and for the details of EPA's analysis of the two EtO mutation data sets. EPA
9 also considered more recent (2013) public comments proposing nonlinear modes of action for
10 EtO carcinogenicity; however, EPA found these hypothetical modes of action to be speculative
11 at this time (see Appendix L and Section J.3.2 of Appendix J).
12 Other sources of uncertainty emanate from the epidemiologic studies and their analyses
13 [Steenland et al. (2004):Steenland et al. (2003): Steenland analyses in Appendix D], including
14 the retrospective estimation of EtO exposures in the cohort, the modeling of the epidemiologic
15 exposure-response data, the proper dose metric for exposure-response analysis, and potential
16 confounding or modifying factors. Although these are common areas of uncertainty in
17 epidemiologic studies, they were generally well addressed in the NIOSH studies.
18 Regarding exposure estimation, the NIOSH investigators conducted a detailed
19 retrospective exposure assessment to estimate the individual worker exposures. They used
20 extensive data from 18 facilities, spanning a number of years, to develop a regression model for
21 estimating EtO exposure levels associated with different jobs (exposure categories), facilities,
22 and time periods (Hornung et al., 1994; Greife et al., 1988) [see also Section A.2.8 for more
23 details about the development and evaluation of the regression model]. The model accounted for
24 85% of the variation in average EtO exposure levels in an independent set of test data. In
25 addition, the modeled estimates were not highly biased nor biased in one direction when
26 compared to the predictions of a panel of 11 industrial hygiene experts familiar with EtO levels
27 in the sterilization industry. The regression model was used to develop an exposure matrix
28 stratified by time and exposure category for each facility in the cohort study. Detailed work
29 history data for the individual workers were collected for the 1987 follow-up and used in
30 conjunction with the facility-specific exposure matrices to derive cumulative exposure estimates
31 for the individual workers (Steenland et al., 1991). For the extended follow-up (Steenland et al.,
32 2004; Steenland et al., 2003), additional information on the date last employed was obtained for
33 those workers still employed and exposed at the time of the original work history collection for
34 the plants still using EtO (25% of the cohort). It was then assumed that exposure for these
35 workers continued until the date of last employment and that their exposure level stayed the
This document is a draft for review purposes only and does not constitute Agency policy.
4-47 DRAFT—DO NOT CITE OR QUOTE
-------
1 same as that in their last job held at the time of the original data collection. Thus, there would be
2 more exposure misclassification in the extended follow-up. However, when the investigators
3 compared cumulative exposures estimated with and without the extended work histories, they
4 found little difference because exposure levels were very low by the mid-1980s and, therefore,
5 had little impact on cumulative exposure (Steenland et al., 2004; Steenland et al., 2003). While
6 the NIOSH regression model performed well in estimating exposures in validation tests
7 (Hornung et al., 1994), there is, nonetheless, uncertainty associated with any retrospective
8 exposure assessment, and this can affect the ability to discriminate among exposure-response
9 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 (shallow) in the low-exposure region for
33 these data, which exhibit supralinearity. EPA attempted to fit two-piece log-linear and linear
34 spline models to the individual continuous exposure data to address the supralinearity of the data
35 while avoiding the extreme low-exposure curvature of the log cumulative exposure model;
This document is a draft for review purposes only and does not constitute Agency policy.
4-48 DRAFT—DO NOT CITE OR QUOTE
-------
1 however, these models resulted in low-exposure slopes that appeared to be implausibly steep
2 (i.e., they suggested excessively large changes in risk from small changes in exposure). The
3 steep low-exposure slopes are a manifestation of apparently high risks in workers with relatively
4 low exposures; however, this elevation is based on small numbers of cancer cases in that
5 exposure range, and EPA has low confidence in the low-exposure slopes. A two-piece spline
6 model with the knot at a higher exposure level could have been used, but without model
7 likelihood as a basis for knot selection, such selection becomes arbitrary, and with the knot at the
8 higher exposure level which had an apparent local maximum for the log-linear model
9 (1,600 ppm x days rather than 100 ppm x days), the model was not statistically significant
10 (p = 0.07). Thus, EPA opted for a weighted linear regression model based on the Cox regression
11 categorical results, excluding the highest exposure group, to reflect the exposure-response
12 relationship in the exposure region below the "plateau." The all lymphohematopoietic cancer
13 data set had more cases (n = 74) but was heavily dominated by the lymphoid cancer response and
14 conveyed the same problems for exposure-response modeling; thus, a linear regression model,
15 excluding the highest exposure group, was used for 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 exposure 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. The most suitable alternative
35 model based on the continuous exposure data is the two-piece log-linear spline model with the
This document is a draft for review purposes only and does not constitute Agency policy.
4-49 DRAFT—DO NOT CITE OR QUOTE
-------
1 knot at 1,600 ppm x days, and thus, unit risk calculations using this model are included in the
2 assessment for comparison purposes. The unit risk estimate from this alternative model is about
3 2.0 times the estimate from the linear regression model of the categorical results. If the
4 alternative model had been selected for lymphoid cancer rather than the linear regression model
5 of the categorical results, the total cancer unit risk estimate would have been 2.90 per ppm rather
6 than 2.27 per ppm, that is, about 28% higher than the total cancer risk estimate with the preferred
7 lymphoid cancer model.38
8 Another area of uncertainty related to the exposure-response modeling is the lag period.
9 The best-fitting models presented by Steenland et al. (2004) for lymphohematopoietic cancer
10 mortality had a 15-year lag (lag periods of 0, 5, 10, 15, and 20 years were considered). A
11 15-year lag period means that exposures in the 15 years prior to death or the end of follow-up are
12 not taken into account. In other words, in the best-fitting models, relevant exposures for the
13 development of the lymphohematopoietic cancers occurred over 15 years before death. For the
14 best-fitting continuous exposure model for lymphoid cancer reported by Steenland et al. (2004),
15 the log cumulative exposure Cox regression model, the actual difference between the regression
16 coefficients from the 15-year-lagged and the unlagged models was negligible (the regression
17 coefficient from the unlagged model was about 8% lower than that from the 15-year-lagged
18 model; however, it should be noted that the unlagged model did not provide a statistically
19 significant fit to the data (p = 0.17) (the results for the unlagged model are presented in
20 S ecti on D. 3. e of Appendix D).
21 In addition, the analyses of the NIOSH investigators indicate that the regression
22 coefficient for cumulative exposure might have decreased with increasing follow-up, suggesting
23 that the higher exposure levels encountered by the workers in the more distant past are having
24 less of an impact on more recent risk. The regression coefficient for lymphoid cancers was
25 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.,
26 1993) versus 4.7 x 10~6 per ppm x day, for both sexes with a 15-year lag, in the 1998 follow-up
27 (see Steenland reanalyses in Appendix D). A similar decrease was found in the regression
28 coefficient for cumulative exposure for all lymphohematopoietic cancers. The life-table analysis
29 used in this dose-response assessment assumes exposure accrues over the full lifetime for the
30 cumulative exposure metric. If, in fact, exposures in the distant past cease to have a meaningful
31 impact on the risk of lymphohematopoietic cancers, this approach would tend to overestimate the
32 unit risk. Thus, a comparison analysis was conducted to evaluate the impact of ignoring
33 exposures over 55 years in the past in the life-table analysis. The actual value of such a cut
38The total cancer risk calculations were conducted at 0.003 ppm, an exposure level expected to be near the LEC0i
estimate for the total cancer risk.
This document is a draft for review purposes only and does not constitute Agency policy.
4-50 DRAFT—DO NOT CITE OR QUOTE
-------
1 point, if warranted, is unknown. A value less than 55 years might not be appropriate because
2 exposures for some of the workers began in 1943, so any diminution of potency for past
3 exposures occurring since 1943 is already reflected in the regression coefficient with follow-up
4 through 1998, at least for those workers, although it is unknown what proportion of workers had
5 such early exposures and how long they survived. The comparison analysis for lymphoid cancer
6 yielded an LECoi of 0.0156 ppm and a unit risk estimate of 0.64 per ppm, which is about
7 27% less than the estimate obtained from the unrestricted life-table analysis. Because the
8 appropriate cut point for excluding past exposures is unknown and the unit risk estimate from the
9 linear regression model of the categorical results is already substantially less than that obtained
10 from the best-fitting log cumulative exposure Cox regression model or even that from the
11 alternative two-piece log-linear spline model with the knot at 1,600 ppm x days, the estimate
12 from the full life-table analysis is preferred. In any event, the preferred estimate is not
13 appreciably different from the estimate from the analysis which considered only the most recent
14 55 years of exposure in the life-table analysis.
15 Several dose metrics (cumulative exposure, duration of exposure, maximum [8-hour
16 TWA] exposure, and average exposure) were analyzed by Steenland et al. (2004), and
17 cumulative exposure was the best predictor of mortality from lymphohematopoietic cancers.
18 Cumulative exposure is considered a good measure of total exposure because it integrates
19 exposure (levels) over time.
20 Also, the important potential modifying/confounding factors of age, sex, race, and
21 calendar time were taken into account in the analysis, and the plants included in this cohort were
22 specifically selected for the absence of any known confounding exposures (Stayner et al., 1993).
23 With respect to the breast cancer mortality response (Steenland et al., 2004), the
24 exposure-response modeling was based on 103 deaths. As for the lymphohematopoietic cancer
25 responses, the exposure-response data for breast cancer mortality are fairly supralinear,
26 especially for the low-exposure groups. An attempt was again made to fit two-piece log-linear
27 and linear spline models to the individual continuous exposure data to address the supralinearity
28 of the data while avoiding the extreme low-exposure curvature of the log cumulative exposure
29 Cox regression model; however, these models resulted in low-exposure slopes that appeared to
30 be implausibly steep and the model fits were not statistically significant. Thus, the same linear
31 regression approach, excluding the highest exposure group, was taken to obtain a regression
32 coefficient for the life-table analysis. As discussed above, the linear regression has its own
33 limitations (e.g., it is based on categorical rather than continuous exposure data and the slope is
34 not statistically significant); nonetheless, it was judged to be the most reasonable approach for
35 deriving low-exposure risk estimates from the available breast cancer mortality data. The most
This document is a draft for review purposes only and does not constitute Agency policy.
4-51 DRAFT—DO NOT CITE OR QUOTE
-------
1 suitable alternative model based on the continuous exposure data is the two-piece log-linear
2 spline model with the knot at 13,000 ppm x days (not the knot with the largest likelihood, but a
3 local maximum), and thus, unit risk calculations using this model are included in the assessment
4 for comparison purposes. The unit risk estimate from this alternative model is about 34% of the
5 estimate from the linear regression model of the categorical results.
6 For the lag period, the best-fitting model had a lag of 20 years, which was the longest lag
7 period investigated. This is a commonly used lag period for solid tumors, which typically have
8 longer latency periods than lymphohematopoietic cancers. It is unknown whether a lag period
9 longer than 20 years would have provided a better model fit. The Steenland et al. (2004)
10 analysis took into account age, race, and calendar time. Other risk factors for breast cancer could
11 not be included in the mortality analysis, but many of these factors were considered in the breast
12 cancer incidence study (Steenland et al., 2003), as discussed below, and the preferred breast
13 cancer risk estimates are based on the breast cancer incidence data.
14 Steenland et al. (2003) conducted an incidence study for breast cancer; therefore, it was
15 not necessary to calculate unit risk estimates for breast cancer incidence indirectly from the
16 mortality data as was done for lymphohematopoietic cancer. Further advantages to using the
17 results from the incidence study are that more cases were available for the exposure-response
18 modeling (319 cases) and that the investigators were able to include data on potential
19 confounders in the modeling for the subcohort with interviews (233 cases). Because the
20 subcohort with interviews had complete case ascertainment and provided data on potential
21 confounders, it was the preferred breast cancer incidence data set, although some results based
22 on the full cohort are presented for comparison. For the full cohort, the continuous exposure Cox
23 regression model providing the best fit to the data was again the log cumulative exposure model.
24 With breast cancer incidence, a 15-year lag provided the best model fits. For the subcohort, the
25 cumulative exposure and log cumulative exposure Cox regression models fit nearly equally well
26 (e.g., similar AICs). For both groups, the categorical Cox regression results suggest that a linear
27 model, providing a compromise between the supralinear log cumulative exposure model and the
28 sublinear cumulative exposure model, would better represent the low-exposure data than either
29 of the two presented continuous exposure models (see Figures 4-4 and 4-5). Thus, for both
30 groups, in EPA's original draft analyses based on the published summary data, a linear
31 regression was fitted to the categorical results, dropping the highest exposure group to provide a
32 better fit to the lower-exposure data (U.S. EPA, 2006a). In subsequent analyses by Dr. Steenland
33 (see Appendix D) of the individual data using exposure as a continuous variable, two-piece
34 log-linear and linear spline models and other linear RR models were used to model the subcohort
This document is a draft for review purposes only and does not constitute Agency policy.
4-52 DRAFT—DO NOT CITE OR QUOTE
-------
1 data; the two-piece linear spline model was the best-fitting of these models and provided the
2 preferred breast cancer incidence risk estimates.
3 Confidence intervals were determined using the Wald approach. Confidence intervals for
4 linear RR models, however, in contrast to those for the log-linear RR models, may not be
5 symmetrical. EPA also evaluated application of a profile likelihood approach for the linear RR
6 models (Langholz and Richardson, 2010), which allows for asymmetric CIs, for comparison with
7 the Wald approach. Using the profile likelihood method and the two-piece linear spline model,
8 the resulting unit risk estimate for breast cancer incidence would have been 2.33 per ppm,
9 slightly higher (34%) than the value of 1.74 per ppm obtained as the unit risk estimate for breast
10 cancer incidence in this assessment. Thus, if the profile likelihood method had been used for the
11 linear RR models in this assessment, the total cancer risk estimate, which incorporates the breast
12 cancer incidence estimate as a component, would be less than 34% higher than the total cancer
13 risk estimate presented here.
14 With respect to the two-piece spline models, the use of this model form is not intended to
15 imply that an abrupt change in biological response occurs at the knot but, rather, to allow
16 description of an exposure-response relationship in which the slope of the relationship differs
17 notably in the low-exposure versus high-exposure regions. The two-piece model is used here
18 primarily for its representation of the low-exposure data. The main uncertainty in the two-piece
19 spline models is in the selection of the knot, and the location of the knot is critical in defining the
20 low-exposure slope. The model likelihood was used to provide a statistical basis for knot
21 selection; although, as shown in Appendix D (see Figure D-la), the likelihood did not generally
22 change appreciably over a range of possible knots. Thus, because of the importance of knot
23 selection, a sensitivity analysis was done to examine the impacts of selecting different knots (see
24 Section D.6 of Appendix D). For the sensitivity analysis, the two-piece log-linear model was run
25 with knots roughly one increment (1,000 ppm x days) below and one increment above the
26 selected knot. For breast cancer incidence, this sensitivity analysis yielded ECoi estimates of
27 0.0133 ppm and 0.0176 ppm, respectively (i.e., about 14% lower and 14% higher, respectively,
28 than the ECoi of 0.0154 ppm obtained with the originally selected knot of 6,000 ppm x days).39
29 As can be seen in Table 4-13, there is substantial variation in the ECoi estimates obtained
30 from the different models. Although some plateauing is apparent with the highest exposure
31 group and is evidenced in the subcohort with interviews by the strong influence of the top 5% of
32 cumulative exposures on dampening the slope of the (cumulative exposure) Cox regression
39 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.
4-53 DRAFT—DO NOT CITE OR QUOTE
-------
1 model (see Section D.I and Figure D-ld of Appendix D), the categorical data for breast cancer
2 incidence do not display the strong supralinearity (steep rise) in the lower exposure groups seen
3 in the cases discussed above (i.e., lymphohematopoietic cancers and breast cancer mortality).
4 Thus, for the subcohort with interviews, the difference between the ECoi estimates from the
5 standard cumulative exposure Cox regression model and the two-piece spline models or the
6 linear regression of the categorical results or continuous exposure linear models are not as
7 dramatic as seen in those cases (the ECoi estimates from the latter four approaches are nearly
8 within an order of magnitude of that of the cumulative exposure model). For the subcohort with
9 interviews, the two-piece spline models, the continuous exposure linear model, and the linear
10 regression of the categorical results gave similar results—the unit risk estimates spanned less
11 than a fourfold range. This range is bounded by the two best-fitting (based on AIC) continuous
12 exposure models—the two-piece linear spline model and the linear model. If the continuous
13 exposure linear model had been selected rather than the two-piece linear spline model, which had
14 a slightly lower AIC value and a better apparent visual fit of the lower-exposure data, the breast
15 cancer incidence unit risk estimate would have been 0.446 per ppm rather than 1.74 per ppm, and
16 the total cancer unit risk estimate would have been 1.15 per ppm rather than 2.27 per ppm. In
17 other words, of the models investigated, the total cancer unit risk estimate with the best-fitting
18 alternate breast cancer incidence model (based on AIC) is about 50% lower than that of the
19 best-fitting model. However, data in the lower-exposure range of greatest relevance for the
20 derivation of a unit risk estimate support a steeper slope in the lower-exposure range; thus,
21 although the lower estimate obtained from the continuous exposure linear model is plausible,
22 unit risk estimates notably lower than that are considered unlikely from the available data.
23 The best-fitting models presented by Steenland et al. (2003) for breast cancer incidence
24 generally had a 15-year lag (lag periods of 0, 5, 10, 15, and 20 years were considered). A
25 15-year lag period means that exposures in the 15 years prior to diagnosis or the end of
26 follow-up are not taken into account. For the various continuous exposure models for breast
27 cancer incidence in the full cohort and the subcohort with interviews reported by Steenland et al.
28 (2003),Tables 4 and 5, none of the unlagged models provided a statistically significant fit to the
29 data, with the exception of the log cumulative exposure Cox regression model for the subcohort,
30 where the unlagged model fit marginally better than the 15-year-lagged model. However, as
31 noted in Section 4.1.2.3, the log cumulative exposure model with no lag was considered less
32 biologically realistic than the corresponding model with a 15-year lag because some lag period
33 would be expected for the development of breast cancer; thus, the 15-year-lagged model was
34 used in this assessment. The regression coefficient from the unlagged log cumulative exposure
35 Cox regression model was about 90% higher than that from the 15-year-lagged model.
This document is a draft for review purposes only and does not constitute Agency policy.
4-54 DRAFT—DO NOT CITE OR QUOTE
-------
1 With respect to dose metrics for breast cancer incidence, models using duration of
2 exposure provided better model fits than those using cumulative exposure (Steenland et al.,
3 2003): however, duration is less useful for estimating unit risks and the cumulative exposure
4 models also provided statistically significant fits to the data, thus the cumulative exposure metric
5 was used for the quantitative risk estimates. Models using peak (highest one-time exposure) or
6 average exposure (cumulative exposure divided by duration) did not fit as well.
7 Regarding potential confounders/modifying factors, analyses for the full cohort were
8 adjusted for age, race, and calendar time, and exposures to other chemicals in these plants were
9 reportedly minimal. For the subcohort with interviews, a number of specific breast cancer risk
10 factors were investigated, including body mass index, breast cancer in a first-degree relative,
11 parity, age at menopause, age at menarche, socioeconomic status, and diet; however, only parity
12 and breast cancer in a first-degree relative were determined to be important predictors of breast
13 cancer and were included in the final models.
14 An area of uncertainty in the life-table analysis for breast cancer incidence pertains to the
15 rates used for the cause-specific background rate. The regression coefficients presented by
16 Steenland et al. (2003) represent invasive and in situ cases combined, where 6% of the cases are
17 in situ, and the preferred unit risk estimates in this assessment are calculated similarly using
18 background rates for invasive and in situ cases combined. The regression coefficients for
19 invasive and in situ cases combined should be good approximations for regression coefficients
20 for invasive cases alone; however, it is uncertain how well they reflect the exposure-response
21 relationships for in situ cases alone. Diagnosed cases of in situ breast cancer would presumably
22 be remedied and not progress to invasive breast cancer, so double-counting is unlikely to be a
23 significant problem. Carcinoma in situ is a risk factor for invasive breast cancer; however, this
24 observation is most likely explained by the fact that these two types of breast cancer have other
25 breast cancer risk factors in common, some of which have been considered in the subcohort
26 analysis. One might hypothesize that EtO exposure could cause a more rapid progression to
27 invasive tumors; however, there is no specific evidence that this occurs. On the other hand, there
28 is some indication that in situ cases in the incidence study might have been diagnosed at
29 relatively low rates in comparison to the invasive cases. Steenland et al. (2003) reported that 6%
30 of the cases in their study are in situ; according to the National Cancer Institute, however, ductal
31 carcinoma in situ accounted for about 18% of newly diagnosed cases of breast cancer in 1998
32 (NCI, 2004).
33 There are several possible explanations for this difference. One is that it reflects
34 differences in diagnosis with calendar time because the rate of diagnosis of carcinoma in situ has
35 increased over time with increased use of mammography. Another is that the difference is
This document is a draft for review purposes only and does not constitute Agency policy.
4-55 DRAFT—DO NOT CITE OR QUOTE
-------
1 partially a reflection of the age distribution in the cohort because the proportion of new cases
2 diagnosed as carcinoma in situ varies by age. A third possible explanation is that the low
3 proportion of in situ cases is at least partially a consequence of underascertainment of cases
4 because in situ cases will not be reported on death certificates, although, even if all 20 in situ
5 cases were in the subcohort with interviews, that would still be only 8.6% of the cases. In any
6 event, this is a relatively minor source of uncertainty, and a comparison of the unit risk estimates
7 using invasive + in situ breast cancer background rates and invasive-only background rates,
8 using EPA's original analyses in the 2006 Draft Assessment, found that the estimate based on the
9 invasive + in situ background rates was less than 20% higher than the corresponding estimate
10 using only invasive breast cancer background rates (U.S. EPA, 2006a).
11 The results for the subcohort with interviews are used for the primary breast cancer unit
12 risk calculations because, in addition to including the data on potential confounders, the
13 subcohort is considered to have full ascertainment of the breast cancer cases, whereas the full
14 cohort for the incidence study has incomplete case ascertainment, as illustrated by the fact that
15 death certificates were the only source of case ascertainment for 14% of the cases. Complete
16 interviews were available for only 68% of the 7,576 women in the full incidence cohort, and
17 thus, some potential exists for participation selection bias in the subcohort. There is, however,
18 no basis for considering participation to be associated with breast cancer or EtO exposure, and
19 the major reason for nonparticipation was a failure to locate (22% of full incidence cohort) and
20 not lack of response (3% of cohort) or refusal to participate (7% of cohort). Risk estimates based
21 on the full cohort were calculated for comparison with the subcohort estimates using the original
22 linear regression analyses of the categorical results. The unit risk estimate based on the
23 subcohort was about 60% higher than the corresponding estimate from the full cohort (U.S. EPA,
24 2006a).
25 Some additional sources of uncertainty are not so much inherent in the exposure-response
26 modeling or in the epidemiologic data themselves but, rather, arise in the process of obtaining
27 more general Agency risk estimates from the epidemiologic results. EPA cancer risk estimates
28 are typically derived to represent an upper bound on increased risk of cancer incidence for all
29 sites affected by an agent for the general population. From experimental animal studies, this is
30 accomplished by using tumor incidence data and summing across all the tumor sites that
31 demonstrate significantly increased incidences, customarily for the most sensitive sex and
32 species, to be protective of the general human population. However, in estimating comparable
33 risks from the NIOSH epidemiologic data, certain limitations are encountered. First, the study
34 reported by Steenland et al. (2004) is a retrospective mortality study, and cancer incidence data
35 are not available for lymphohematopoietic cancer [for breast cancer, a separate incidence study
This document is a draft for review purposes only and does not constitute Agency policy.
4-56 DRAFT—DO NOT CITE OR QUOTE
-------
1 (Steenland et al., 2003) was available]. Second, these occupational epidemiology data represent
2 a healthy worker cohort. Third, the epidemiologic study may not have sufficient statistical
3 power and follow-up time to observe associations for all the tumor sites that may be affected by
4 EtO.
5 The first limitation was addressed quantitatively in the life-table analysis for the
6 lymphohematopoietic cancer risk estimates. Although assumptions are made in using incidence
7 rates for the cause-specific background rates, as discussed in Section 4.1.1.3, the resulting
8 incidence-based estimates are believed to be better estimates of cancer incidence risk than are the
9 mortality-based estimates. The incidence unit risk estimate is about 120% higher than (i.e.,
10 2.2 times) the mortality-based estimate, which seems reasonable given the relatively high
11 survival rates for lymphoid cancers (according to SEER data [www.seer.cancer.gov], 5-year
12 survival rates are 65% for NHL; 78% for chronic lymphocytic leukemia, which are the vast
13 majority of the lymphocytic leukemias in adults; and 40% for multiple myeloma).
14 The healthy-worker effect is often an issue in occupational epidemiology studies, but the
15 internal exposure-response analyses conducted by these investigators help address this concern,
16 at least partially. In terms of representing the general population, the NIOSH study cohort was
17 relatively diverse. It contained both female (55%) and male workers, and the workers were 79%
18 white, 16% black, and 5% "other." Furthermore, because of EtO's mutagenic mode of action,
19 increased early-life susceptibility is assumed and ADAFs are applied for exposure scenarios
20 involving early life (see Section 4.4).
21 With respect to other possible tumor sites of concern, the rodent data suggest that
22 lymphohematopoietic cancers are a major tumor type associated with EtO exposure in female
23 mice and in male and female rats. Thus, it is reasonable that this might be a cancer type of
24 concern in humans also. Likewise, the mouse data suggest an increased risk of mammary gland
25 tumors from EtO exposure, and evidence of that can be seen in the Steenland et al. (2004) and
26 Steenland et al. (2003) studies. However, the rodent data suggest associations between EtO
27 exposure and other tumor types as well, and although site concordance across species is not
28 generally assumed, it is possible that the NIOSH study, despite its relatively large size and long
29 follow-up (mean length of follow-up was 26.8 years), had insufficient power to observe small
30 increases in risk in certain other sites. For example, the tumor site with the highest potency
31 estimate in both male and female mice was the lung. In the NIOSH study, one cannot rule out a
32 small increase in the risk of lung cancer, which has a high background rate, thus making small
33 increases difficult to detect.
34 To obtain the risk estimate for total cancer risk (2.3 per ppm, or 2.3 x 10~3 per ppb), the
35 preferred estimates for lymphoid cancer incidence and breast cancer incidence were combined.
This document is a draft for review purposes only and does not constitute Agency policy.
4-57 DRAFT—DO NOT CITE OR QUOTE
-------
1 While there are uncertainties in the approach used to combine the individual estimates, the
2 resulting unit risk estimate is appropriately bounded in the roughly twofold range between
3 estimates based on the sum of the individual MLEs of risk and the sum of the individual 95%
4 UCLs, and thus, any inaccuracy in the total cancer unit risk estimate resulting from the approach
5 used is relatively minor. Because the breast cancer component of the total cancer risk estimate
6 applies only to females, the total cancer risk estimate is expected to overestimate the cancer risk
7 to males somewhat (the preferred unit risk estimate for lymphoid cancer alone was 0.877 per
8 ppm [or 8.77 x 1CT4 per ppb], which is about 40% of the total cancer risk estimate).
9 Despite these uncertainties, the inhalation cancer unit risk estimate of 2.3 per ppm (or
10 2.3 x 1Q~3 per ppb) for the total cancer risk from lymphoid cancer incidence and female breast
11 cancer incidence has the advantages of being based on human data from a large, high-quality
12 epidemiologic study with individual exposure estimates for each worker. Furthermore, the breast
13 cancer component of the risk estimate, which contributes approximately 60% of the total cancer
14 risk, is based on a substantial number of incident cases (233 total, the vast majority of which
15 were in the exposure range below the knot of 5,800 ppm x days [see Table D.I a of
16 Appendix D]).
17 A further area of uncertainty pertains to the assumption that RR is independent of age,
18 which is a common assumption in the dose-response modeling of epidemiological data and is an
19 underlying assumption in the Cox regression model. For the NIOSH worker cohort, the
20 proportional hazards model assumption of RR being independent of age was tested by checking
21 the significance of an interaction between age and cumulative exposure, and none of the models
22 had a significant interaction term. This suggests that, for adults at least, the assumption that RR
23 is independent of age is valid. However, the worker cohort contains no children and is
24 uninformative on the issue of early-life susceptibility. In the absence of data on early-life
25 susceptibility, EPA's Supplemental Guidance (U.S. EPA, 2005b) recommends that increased
26 early-life susceptibility be assumed for carcinogens with a mutagenic mode of action, and the
27 conclusion was made in Section 3.4 that the weight of evidence supports a mutagenic mode of
28 action for EtO. Thus, in accordance with the Supplemental Guidance, the alternate assumption
29 of increased early-life susceptibility is preferred as the basis for risk estimates in this assessment,
30 and risk estimates derived under this preferred assumption are presented in Section 4.4.
31
32 4.1.5. Summary
33 Under the common assumption that RR is independent of age, an inhalation unit risk
34 estimate for lymphoid cancer incidence of 0.877 per ppm (or 8.77 x 10~4 per ppb; 4.79 x 10~4 per
35 |^g/m3) was calculated using a life-table analysis and a weighted linear regression of the
This document is a draft for review purposes only and does not constitute Agency policy.
4-58 DRAFT—DO NOT CITE OR QUOTE
-------
1 categorical Cox regression results, excluding the highest exposure group, for excess lymphoid
2 cancer mortality from a high-quality occupational epidemiology study. Similarly, an inhalation
3 unit risk estimate for female breast cancer incidence of 1.74 per ppm (or 1.74 x 10 3 per ppb;
4 9.51 x 10 4 per ug/m3) was calculated using a life-table analysis and two-piece linear spline
5 modeling of the continuous exposure data for excess breast cancer incidence from the same
6 high-quality occupational epidemiology study. The linear regression of the categorical results
7 with the exclusion of the highest exposure group for the lymphoid cancer results and the
8 two-piece linear spline analysis for the breast cancer incidence data were different modeling
9 approaches used to address the supralinearity of the exposure-response data in the two data sets.
10 Low-dose linear extrapolation was used, as warranted by the clear mutagenicity of EtO. An
11 ECoi estimate of 0.0078 ppm, a LECoi estimate of 0.0044 ppm, and a unit risk estimate of 2.3 per
12 ppm (or 2.3 x 10 3 per ppb; 1.2 x 10 3 per ug/m3) were obtained for the total cancer risk
13 combined across both cancer types. Despite the uncertainties discussed above, this inhalation
14 unit risk estimate has the advantages of being based on human data from a high-quality
15 epidemiologic study with individual exposure estimates for each worker.
16 In the absence of data on early-life susceptibility, EPA's Supplemental Guidance (U.S.
17 EPA, 2005a) recommends that increased early-life susceptibility be assumed for carcinogens
18 with a mutagenic mode of action, and the conclusion was made in Section 3.4 that the weight of
19 evidence supports a mutagenic mode of action for EtO. Thus, in accordance with the
20 Supplemental Guidance, the alternate assumption of increased early-life susceptibility is
21 preferred as the basis for risk estimates in this assessment, and risk estimates derived under this
22 preferred assumption are presented in Section 4.4. Other than the use of the alternate assumption
23 about early-life susceptibility, the approach used to derive the estimates presented in Section 4.4
24 is identical to the approach used for the estimates derived here in Section 4.1, and the
25 comparisons made between various options and the issues and uncertainties discussed here in
26 Section 4.1 are applicable to the estimates derived in Section 4.4.
27
28 4.2. INHALATION UNIT RISK DERIVED FROM EXPERIMENTAL ANIMAL DATA
29 4.2.1. Overall Approach
30 Lifetime animal cancer bioassays of inhaled EtO have been carried out in three
31 laboratories, as described in Section 3.2. The data from these reports are presented in Tables 3-3
32 through 3-5. These studies have also been reviewed by the IARC (1994b) and Health Canada
33 (2001). Health Canada calculated the EDos for each data set using the benchmark dose
34 methodology. The EOIC report (EPIC, 2001) tabulated only lymphatic tumors because they
35 constituted the predominant risk.
This document is a draft for review purposes only and does not constitute Agency policy.
4-59 DRAFT—DO NOT CITE OR QUOTE
-------
1 The overall approach in this derivation is to find a unit risk for each of the
2 bioassays—keeping data on males and females separate—from data on the incidence of all tumor
3 types and then to use the maximum of these values as the summary measure of the unit risk from
4 animal studies (i.e., the unit risk represents the most sensitive species and sex). The unit risk for
5 the animals in these bioassays is converted to a unit risk in humans by first determining the
6 continuous exposures in humans that are equivalent to the rodent bioassay exposures and then by
7 assuming that the lifetime incidence in humans is equivalent to lifetime incidence in rodents, as
8 is commonly accepted in interspecies risk extrapolations. For cross-species scaling of exposure
9 levels (see Section 4.2.2 below), an assumption of ppm equivalence is used; thus, no interspecies
10 conversion is needed for the exposure concentrations. Bioassay exposure levels are adjusted to
11 equivalent continuous exposures by multiplying by (hours of exposure/24 hours) and by (5/7) for
12 the number of days exposed per week. The unit risk in humans (risk per unit air concentration)
13 is then assumed to be numerically equal to that in rodents (after adjustment to continuous
14 exposures); the calculations from the rodent bioassay data are shown in Tables 3-3 through 3-5.
15
16 4.2.2. Cross-Species Scaling
17 In the absence of chemical-specific information, EPA's 1994 inhalation dosimetry
18 methods (U.S. EPA, 1994) provide standard methods and default scaling factors for
19 cross-species scaling. Under EPA's methodology, EtO would be considered a Category 2 gas
20 because it is reactive and water soluble and has clear systemic distribution and effects.
21 Dosimetry equations for Category 2 gases are undergoing EPA re-evaluation and are not being
22 used at this time. For cross-species scaling of extrarespiratory effects, current practice is to treat
23 Category 2 gases as Category 3 gases. For Category 3 gases, ppm equivalence is assumed (i.e.,
24 responses across species are equivalent on a ppm exposure basis), unless the airblood partition
25 coefficient for the experimental species is less than the coefficient for humans [U.S. EPA (1994),
26 p. 4-61]. In the case of EtO, measured airblood partition coefficients are 78 in the mouse
27 (Fennell and Brown, 2001), 64 in the rat (Krishnan et al., 1992), and 61 in the human (Csanady
28 et al., 2000); thus, ppm equivalence for cross-species scaling to humans can be assumed for
29 extrarespiratory effects observed in mice and rats. The assumption of ppm equivalence is further
30 supported by the PBPK modeling of Fennell and Brown (2001), who reported that simulated
31 blood AUCs for EtO after 6 hours of exposure to concentrations between 1 ppm and 100 ppm
32 were similar for mice, rats, and humans and were linearly related to the exposure concentration
33 (see Section 3.3.1 and Figure 3-2). This modeling was validated against measured blood EtO
34 concentrations for rodents and humans. For Category 2 gases with respiratory effects, there is no
35 clear guidance on an interim approach. One suggested approach is to do cross-species scaling
This document is a draft for review purposes only and does not constitute Agency policy.
4-60 DRAFT—DO NOT CITE OR QUOTE
-------
1 using both Category 1 and Category 3 gas equations and then decide which is most appropriate.
2 In this document, the preferred approach was to assume ppm equivalence was also valid for the
3 lung tumors in mice because of the clear systemic distribution of EtO (e.g., see Section 3.1).
4 Treating EtO as a Category 1 gas for cross-species scaling of the lung tumors would presume
5 that the lung tumors are arising only from the immediate and direct action of EtO as it comes into
6 first contact with the lung. In fact, some of the EtO dose contributing to lung tumors is likely
7 attributable to recirculation of systemic EtO through the lung.
8 If one were to treat EtO as a Category 1 gas for the cross-species scaling of the lung
9 tumor response as a bounding exercise, EPA's 1994 inhalation dosimetry methods present
10 equations for estimating the RGDRpu, i.e., the regional gas dose ratio for the pulmonary region,
11 which acts as an adjustment factor for estimating human equivalent exposure concentrations
12 from experimental animal exposure concentrations (adjusted for continuous exposure) [U.S. EPA
13 (1994), pp. 4-49 to 4-51]. These equations rely on parameters describing mass transport of the
14 gas (EtO) in the extrathoracic and tracheobronchial regions for both the experimental animal
15 species (mouse) and humans. Without experimental data for these parameters, it seems
16 reasonable to estimate RGDRPU using a simplified equation and the adjusted alveolar ventilation
17 rates of Fennell and Brown (2001). Fennell and Brown adjusted the alveolar ventilation rates to
18 reflect limited pulmonary uptake of EtO, a phenomenon commonly observed for highly
19 water-soluble gases (Johanson and Filser, 1992). The adjusted ventilation rates were then used
20 by Fennell and Brown in their PBPK modeling simulations, and good fits to blood concentration
21 data were reported for both the mouse and human models. In this document, the adjusted
22 alveolar ventilation rates were used to estimate the RGDRpu as follows:
23
24
25 RGDRpu = (RGDpu)m/(RGDpu)h = (Qaiv/SAPU)m/(Qaiv/SApu)h, (4-4)
26 where:
27 RGDpu = regional gas dose to the pulmonary region,
28 Qaiv = (adjusted) alveolar ventilation rate,
29 SApu = surface area of the pulmonary region, and
30 the subscripts "m" and "h" denote mouse and human values.
31
32
33 Then, using adjusted alveolar ventilation rates from Fennell and Brown (2001) and surface area
34 values from EPA [U.S. EPA (19941 p. 4-26],
35
36
37 RGDRPU = ((0.78L/h)/(0.05m2))/((255L/h)/(54.0m2)) = 3.3. (4-5)
This document is a draft for review purposes only and does not constitute Agency policy.
4-61 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3 Using this value for the RGDRPU would increase the human equivalent concentration about
4 threefold, resulting in a decreased risk for lung tumors of about threefold, as a lower bound. The
5 true value of the RGDRpu is expected to be between 1 and 3, and any adjustment to the lung
6 tumor risks would still be expected to result in unit risk estimates roughly within the range of the
7 rodent unit risk estimates derived later in Section 4.2 under the assumption of ppm equivalence.
8
9 4.2.3. Dose-Response Modeling Methods
10 In this document the following steps were used:
11 1. Extract the incidence data presented in the original studies. In order to crudely adjust
12 for early mortality in the analysis of the NTP (1987) data, the incidence data have been corrected
13 for a specific tumor type by eliminating the animals that died prior to the occurrence of the first
14 tumor or prior to 52 weeks, whichever was earlier. It was not possible to make this adjustment
15 with the other studies where data on individual animals were not available. With these
16 exceptions, the tumor incidence data in Tables 3-3 through 3-5 match the original data.
17 2. Fit the multistage model to the dose-response data using the Tox Risk program.
18 The likelihood-ratio test was used to determine the lowest value of the multistage
19 polynomial degree that provided the best fit to the data while requiring selection of the most
20 parsimonious model. In this procedure, if a good fit to the data in the neighborhood of the POD
21 is not obtained with the multistage model because of a nonmonotonic reduction in risk at the
22 highest dose tested (as sometimes occurs when there is early mortality from other causes), that
23 data point is eliminated and the model is fit again to the remaining data. Such a deletion was
24 found necessary in two cases (mammary tumors in the NTP study and mononuclear cell
25 leukemia in the Lynch study). The goodness-of-fit measures for the dose-response curves and
26 the parameters derived from them are shown in Appendix G.
27 In the NTP bioassay, where the individual animal data were available, a time-to-tumor
28 analysis was undertaken to account for early mortality. The general model used in this analysis
29 is the multistage Weibull model:
30
31
32 P(d,t)=l-exp[-(qo + qid + q2d2 + ... + qkdk)x(t-to)z], (4-6)
33
34
35 where P(d,t) represents the probability of a tumor by age t (in bioassay weeks) for dose d (i.e.,
36 human equivalent exposure), and the parameter ranges are restricted as follows: z > 1, to > 0,
37 and q; > 0 for i = 0, 1, ..., k. The parameter to represents the time between when a potentially
This document is a draft for review purposes only and does not constitute Agency policy.
4-62 DRAFT—DO NOT CITE OR QUOTE
-------
1 fatal tumor becomes observable and when it causes death. The analyses were conducted using
2 the computer software Tox_Risk version 3.5, which is based on methods developed by Krewski
3 et al. (1983). Parameters are estimated in Tox_Risk using the method of maximum likelihood.
4 Tumor types can be categorized by tumor context as either fatal or incidental. Incidental
5 tumors are those tumors thought not to have caused the death of an animal, whereas fatal tumors
6 are thought to have resulted in animal death. Tumors at all sites were treated as incidental
7 (although it was recognized that this may not have been the case, the experimental data are not
8 detailed enough to conclude otherwise). The parameter to was set equal to 0 because there were
9 insufficient data to reliably estimate it.
10 The likelihood-ratio test was used to determine the lowest value of the multistage
11 polynomial degree k that provided the best fit to the data while requiring selection of the most
12 parsimonious model. The one-stage Weibull (i.e., k = 1) was determined to be the most optimal
13 value for all the tumor types analyzed.
14 3. Select the POD and calculate the unit risk for each tumor site. The effective
15 concentration that causes a 10% extra risk for tumor incidence, ECio, and the 95% lower bound
16 of that concentration, LECio, are derived from the dose-response model. The LECio is then used
17 as the POD for a linear low-dose extrapolation, and the unit risk is calculated as 0.1/LECio. This
18 is the procedure specified in the EPA's Guidelines for Carcinogen Risk Assessment (U.S. EPA,
19 2005a) for agents such as EtO that have direct mutagenic activity. See Section 3.4 for a
20 discussion of the mode of action for EtO. Tables 3-3 through 3-5 present the unit risk estimates
21 for the individual tumor sites in each bioassay.
22 4. Develop a unit risk estimate based on the incidence of all tumors combined. This
23 method assumes that occurrences of tumors at multiple sites are independent and, further, that
24 the risk estimate for each tumor type is normally distributed. Then, at a given exposure level, the
25 MLEs of extra risk due to each tumor type are added to obtain the MLE of total cancer risk. The
26 variances corresponding to each tumor type are added to give the variance associated with the
27 sum of the MLEs. The one-sided 95% UCL of the MLE for the combined risk is then calculated
28 as:
29
30
31 95% UCL = MLE + 1.645(SE), (4-7)
32
33
34 where SE is the standard error and is the square root of the summed variance. (Note that as a
35 precursor to this step, when Tox_Risk is used to fit the incidence of a single tumor type, it
36 provides the MLE and 95% UCL of extra risk at a specific dose. The standard error in the MLE
This document is a draft for review purposes only and does not constitute Agency policy.
4-63 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
is determined using the above formula). The calculation is repeated for a few exposure levels,
and the exposure yielding a value of 0.1 for the upper bound on extra risk is determined by
interpolation. The unit risk is then the slope of the linear extrapolation from this POD. The
results are given in Table 4-16.
Table 4-16. Upper-bound unit risks (per ug/m ) 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
Lvnch et al. (1984b).
Lynch et al. (1984a)
male rat
4.17 x 10'5
3.66 x lO'5
-
Snellings et al. (1984)b
Male rat
2.19 x 10'5
2.88 x lO'5
-
Female rat
3.37 x 10'5
3.54 x lO'5
-
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
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.
blncludes 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.
4.2.4. Description of Experimental Animal Studies
NTP (1987) exposed male and female B6C3Fi mice to concentrations of 0, 50, and
100 ppm for 6 hours per day, 5 days per week, for 102 weeks. An elevated incidence of lung
carcinomas was found in males, and elevated lung carcinomas, malignant lymphomas, uterine
adenocarcinomas, and mammary carcinomas were found in females. These data are shown in
Table 3-3.
Lynch et al. (Lynch etal., 1984a: Lynch etal., 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
This document is a draft for review purposes only and does not constitute Agency policy.
4-64 DRAFT—DO NOT CITE OR QUOTE
-------
1 available to perform the correction, the incidence would be higher. Therefore, using these data
2 results in an underestimate of risk.
3 Snellings et al. (1984) exposed male and female F344 rats to 0, 10, 33, and 100 ppm for
4 6 hours per day, 5 days per week, for 2 years and described their results for all sites except the
5 brain. In two subsequent publications for the same study, (Garman etal., 1986, 1985) described
6 the development of brain tumors in a different set of F344 rats. The Snellings et al. (1984)
7 publication reported an elevated incidence of splenic mononuclear cell leukemia and peritoneal
8 mesothelioma in males and an elevated incidence of splenic mononuclear cell leukemia in
9 females. The mortality was higher in the 100-ppm groups than the other three groups for both
10 males and females. The incidences in the animals killed after 24 months in Snellings et al.
11 (1984) are shown in Table 3-5. Table 3-5 also presents the brain tumor incidence data for male
12 and female rats from the (Garman et al., 1986, 1985) publications. The brain tumor incidence
13 was lower than that of the other tumors, particularly the splenic mononuclear cell leukemias.
14
15 4.2.5. Results of Data Analysis of Experimental Animal Studies
16 The unit risks calculated from the individual site-sex-bioassay data sets are presented in
17 Tables 3-3 through 3-5. The highest unit risk of any individual site is 3.23 x 10 5 per ug/m3,
18 which is for mononuclear cell leukemia in the female rats of the Snellings et al. (1984) study.
19 Table 4-17 presents the results of the time-to-tumor method applied to the individual
20 animals in the NTP bioassay, compared with the results from the dose group incidence data in
21 Table 3-3. This comparison was done for each tumor type separately. The time-to-tumor
22 method of analyzing the individual animals results in generally higher unit risk estimates than
23 does the analysis of dose group data, as shown in Table 4-17. The ratio is not large (less than
24 2.2) across the tumor types. (In the case of mammary tumors this ratio is actually less than 1. It
25 must be noted that the incidence at the highest dose [where the incidence was substantially less
26 than at the intermediate dose] was deleted from the analysis of grouped data, whereas it was
27 retained in the time-to-tumor analysis. Therefore, the comparison for the mammary tumors is
28 not a strictly valid comparison of methods.) The results also show the extent to which a time-to-
29 tumor analysis of individual animal data increases the risk estimated from data on dose groups.
30 It is expected that if individual animal data were available for the Lynch et al. (1984a) and Lynch
31 et al., 1984b) and the Snellings et al. (1984) bioassays, then the time-to-tumor analysis would
32 also result in higher estimates because both those studies also showed early mortality in the
33 highest dose group.
34 The results of combining tumor types are summarized in Table 4-16. The sums of the
35 individual unit risks tabulated in Tables 3-3 to 3-5 are given in the second row of Table 4-16.
This document is a draft for review purposes only and does not constitute Agency policy.
4-65 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
27
aP(d,t) = 1 - 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.
Note that as expected they are greater than the unit risks computed from the upper bound on the
sum of risks for all data sets except for the Lynch et al. (Lynch etal., 1984a: Lynch etal., 1984b)
data. The reason for this exception is not known, but the differences are small. It is likely that
the problem arises from the methodology used to combine the risks across tumor sites. In an
attempt to be consistent with the new two-step methodology (i.e., modeling in the observable
range to a POD and then doing a linear extrapolation to zero extra risk at zero exposure), the
exposure concentration at which the sum of the independent tumor site risks yielded a 95% upper
bound on 10% extra risk was estimated and used as the POD. Summing risks in this way results
in a POD for the combined tumor risk that is different (lower) than the points of departure for
each individual tumor site risk. Thus, the risk estimate for the sum is not 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 etal., 1984a: Lynch et al., 1984b) data, the
sum of upper bounds, 3.66 x io~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
This document is a draft for review purposes only and does not constitute Agency policy.
4-66 DRAFT—DO NOT CITE OR QUOTE
-------
1 and is therefore not carried over into the summary Table 4-18. For the Snellings et al. (1984)
2 data sets, the upper confidence bound on the sum of risks is used in the summary Table 4-18.
3 The results of the sum-of-risks calculations on the NTP bioassay time-to-tumor data are included
4 in the third row of Table 4-16. The estimate for the NTP female mice is 4.55 x 10 5 per ug/m3,
5 which is higher than the other two measures of total tumor risk in that bioassay. This value is
6 preferable to the other measures because it utilizes the individual animal data available for that
7 bioassay.
8 Summary of results. The summary of unit risks from the five data sets is shown in
9 Table 4-18. The data set giving the highest risk (4.55 x io~5 per ug/m3) is the NTP (1987) data
10 on combined tumors in female mice. The other values are within about a factor of 2 of the
11 highest value.
12
13 Table 4-18. Summary of unit risk estimates (per ug/m3) in animal bioassays
14
Assay
NTP ( 1987). B6C3F, mice
Lvnch et al. <1984b). Lvnch et al. <1984a).
F344 rats
Snellinss et al. (1984), F344 rats
Males
3.01 x lQ-5a
3.66 x lQ-5c
2.19 x 10'5d
Females
4.55 x lQ-5b
_
3.37 x 10'5d
15
16 aFrom time-to-tumor analysis of lung adenomas and carcinomas, Table 4-17.
17 bUpper bound on sum of risks from the time-to-tumor analysis of the NTP data, Table 4-16.
18 °Sum of (upper bound) unit risks (see text for explanation), Table 4-16.
19 dUpper bound on sum of risks, Table 4-16.
20
21
22 4.3. SUMMARY OF INHALATION UNIT RISK ESTIMATES—NOT ACCOUNTING
23 FOR ASSUMED INCREASED EARLY-LIFE SUSCEPTIBILITY
24 For both humans and laboratory animals, tumors occur at multiple sites. In humans, there
25 was a combination of tumors having lymphohematopoietic, in particular lymphoid, origins in
26 both sexes and breast cancer in females, and, in rodents, lymphohematopoietic tumors, mammary
27 carcinomas, and tumors of other sites were observed. From human data, an extra cancer unit risk
28 estimate of 4.79 x 10 4 per ug/m3 (8.77 x 10 4 per ppb) was calculated for lymphoid cancer
29 incidence, and a unit risk estimate of 9.31 x icr4 per ug/m3 (1.74 x io~3 per ppb) was calculated
30 for breast cancer incidence in females. The total extra cancer unit risk estimate was 1.2 x 10 3
31 per ug/m3 (2.3 x 10~3 per ppb) for both cancer types combined (ECoi = 0.0078 ppm;
32 LECoi = 0.0044 ppm). Unit risk estimates derived from the three chronic rodent bioassays for
This document is a draft for review purposes only and does not constitute Agency policy.
4-67 DRAFT—DO NOT CITE OR QUOTE
-------
1 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
2 than the estimates based on human data.
3 Adequate human data, if available, are considered to provide a more appropriate basis
4 than rodent data for estimating human risks (U.S. EPA, 2005a), primarily because uncertainties
5 in extrapolating quantitative risks from rodents to humans are avoided. Although there is a
6 sizeable difference between the rodent-based and the human-based estimates, the human data are
7 from a large, high-quality study, with EtO exposure estimates for the individual workers and
8 little reported exposure to chemicals other than EtO. Therefore, the total extra cancer unit risk
9 estimate of 1.2 x icf3 per ug/m3 (2.3 x icf3 per ppb) calculated for lymphoid cancers and breast
10 cancer combined is the preferred estimate of those estimates not taking assumed increased early-
11 life susceptibility into account (estimates accounting for assumed increased early-life
12 susceptibility are presented in Section 4.4). The unit risk estimate is intended to be an upper
13 bound on cancer risk for use with exposures below the POD (i.e., the LECoi). The unit risk
14 estimate should not generally be used above the POD; however, in the case of this total extra
15 cancer unit risk, which is based on cancer type-specific unit risk estimates from two linear
16 models, the estimate should be valid for exposures up to about 0.075 ppm (140 ug/m3), which is
17 the minimum of the limits for the lymphoid cancer unit risk estimate (0.090 ppm; see
18 Section 4.1.1.2) and the breast cancer unit risk estimate (0.075 ppm; see Section 4.1.2.3).
19 Because a mutagenic mode of action for EtO carcinogenicity (see Section 3.3.2) is
20 "sufficiently supported in (laboratory) animals" and "relevant to humans," and as there are no
21 chemical-specific data to evaluate the differences between adults and children, increased
22 early-life susceptibility should be assumed and, if there is early-life exposure, the age-dependent
23 adjustment factors (ADAFs) should be applied, as appropriate, in accordance with EPA's
24 Supplemental Guidance (U.S. EPA, 2005b) see Section 4.4 below for more details on the
25 application of ADAFs).
26
27 4.4. ADJUSTMENTS FOR POTENTIAL INCREASED EARLY-LIFE
28 SUSCEPTIBILITY
29 There are no chemical-specific data on age-specific susceptibility to EtO-induced
30 carcinogenesis. However, there is sufficient weight of evidence to conclude that EtO operates
31 through a mutagenic mode of action (see Section 3.4.1). In such circumstances (i.e., the absence
32 of chemical-specific data on age-specific susceptibility but sufficient evidence of a mutagenic
33 mode of action), EPA's Supplemental Guidance for Assessing Susceptibility from Early-Life
34 Exposure to Carcinogens (U.S. EPA, 2005b) recommends the assumption of increased early-life
35 susceptibility and the application of default age-dependent adjustment factors (ADAFs) to adjust
This document is a draft for review purposes only and does not constitute Agency policy.
4-68 DRAFT—DO NOT CITE OR QUOTE
-------
1 for this potential increased susceptibility from early-life exposure. See the Supplemental
2 Guidance for detailed information on the general application of these adjustment factors. In
3 brief, the Supplemental Guidance establishes ADAFs for three specific age groups. The current
4 ADAFs and their age groupings are 10 for <2 years, 3 for 2 to <16 years, and 1 for 16 years and
5 above (U.S. EPA, 2005b). For risk assessments based on specific exposure assessments, the
6 10-fold and 3-fold adjustments to the unit risk estimates are to be combined with age-specific
7 exposure estimates when estimating cancer risks from early-life (<16 years of age) exposure.
8 These ADAFs, however, were formulated based on comparisons of the ratios of cancer
9 potency estimates from juvenile-only exposures to cancer potency estimates from adult-only
10 exposures from rodent bioassay data sets with appropriate exposure scenarios, and they are
11 designed to be applied to cancer potency estimates derived from adult-only exposures. Thus,
12 alternate life-table analyses were conducted to derive comparable adult-exposure-only unit risk
13 estimates to which ADAFs would be applied to account for early-life exposure.40 For these
14 alternate life-table analyses, it was assumed that RR is independent of age for adults, which
15 represent the life stage for which the exposure-response data and the Cox regression modeling
16 results from the NIOSH cohort study specifically pertain, but that there is increased early-life
17 susceptibility, based on the weight-of-evidence-based conclusion that EtO carcinogenicity has a
18 mutagenic mode of action (see Section 3.4), which supersedes the assumption that RR is
19 independent of age for all ages including children.
20 In the alternate analyses, exposure in the life table was taken to start at age 16 years, the
21 age cut point that was established in EPA's Supplemental Guidance (U.S. EPA, 2005b), to derive
22 an adult-exposure-only unit risk estimate to which ADAFs would be applied to account for
23 early-life exposure. Other than the age at which exposure was initiated, the life-table analyses
24 are identical to those conducted for the results presented in Section 4.1. Adult-exposure-only
25 unit risk estimates were derived for both cancer incidence and mortality for both lymphoid and
26 breast cancers. Alternate estimates were not derived for all lymphohematopoietic cancers
27 because lymphoid cancer was the preferred endpoint (see Section 4.1.1.2). Incidence estimates
28 are preferred over mortality estimates, but both are calculated here for comparison and because
29 mortality estimates are sometimes used in addition to incidence estimates in benefit-cost
40
In this assessment, adult-exposure-only unit risk estimates refer to estimates derived from the life-table analysis
assuming exposure only for ages > 16 years. The adult-exposure-only unit risk estimates are merely intermediate
values in the calculation of adult-based unit risk estimates and should not be used in any risk calculations. Adult-
based unit risk estimates refer to estimates derived after rescaling the adult-exposure-only unit risk estimates to a
(70-year) lifetime, as described later in Section 4.4. The adult-based unit risk estimates are intended to be used in
ADAF calculations (U.S, EPA, 2005b) for the computation of extra risk estimates for specific exposure scenarios.
Note that the unit risk estimates in this section, which are derived under an assumption of increased early-life
susceptibility, supersede those that were derived in Section 4.1 under the assumption that RR is independent of age.
This document is a draft for review purposes only and does not constitute Agency policy.
4-69 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
8
9
10
11
analyses. For each cancer endpoint, the same exposure-response model was used as that which
was selected for the unit risk estimates in Section 4.1 (i.e., linear regression of the categorical
results, excluding the highest exposure category, for lymphoid cancer and breast cancer mortality
and two-piece linear spline model for breast cancer incidence). The results are presented in
Table 4-19 along with the unit risk estimates derived assuming that RR was independent of age
for all ages (see Section 4.1) for comparison. As can be seen in Table 4-19, the unit risk
estimates for adult-only exposures range from about 66% to about 72% of the unit risk estimates
derived under the assumption of age independence across all ages.
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
LECoi
(ppm)
0.0352
0.0163
0.0297
0.00863
Adult-exposure-
only unit risk
estimate"
(per ppm)
0.284
0.613
0.337
1.16°
Lifetime-exposure unit risk estimate
under assumption of age
independence1"
(per ppm)
0.397
0.877
0.513
1.74C
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
*These are intermediate values. See Table 4-22 below for the final adult-based cancer-type-specific unit risk
estimates.
"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).
According to EPA's Guidelines for Carcinogen Risk Assessment (U.S. EPA, 2005a)
cancer risk estimates are intended to reflect total cancer risk, not site-specific cancer risk;
therefore, an additional calculation was made to estimate the combined risk for (incident)
lymphoid and breast cancers from adult-only exposures, because females would be at risk for
both cancer types. Assuming that the tumor types are independent and that the risk estimates are
approximately normally distributed, this calculation can be made as described in Section 4.1.3.
First, an ECoi of 0.0114 ppm for the total cancer risk (i.e., lymphoid cancer incidence + breast
cancer incidence) from adult-only exposure was estimated, as summarized in Table 4-20.
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
This document is a draft for review purposes only and does not constitute Agency policy.
4-70 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
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).
Table 4-20. Calculation of ECoi for total cancer (incidence) 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
EC0i for total risk
(ppm)
-
-
0.0114
*These are intermediate values. See Table 4-23 below for the final adult-based cancer-type-
specific 0.01/ECoi estimates.
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/ECOT).
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
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
*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.
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
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
This document is a draft for review purposes only and does not constitute Agency policy.
4-71 DRAFT—DO NOT CITE OR QUOTE
-------
1 the sum of the unit risk estimates (1.77), and thus, any inaccuracy in the total cancer risk estimate
2 resulting from the approach used to combine risk estimates across cancer types is relatively
3 minor.
4 When EPA derives unit risk estimates from rodent bioassay data, there is a blurring of the
5 distinction between lifetime and adult-only exposures because the relative amount of time that a
6 rodent spends as a juvenile is negligible (<8%) compared to its lifespan. [According to EPA's
7 Supplemental Guidance., puberty begins around 5-7 weeks of age in rats and around 4-6 weeks
8 in mice (U.S. EPA, 2005b)1. Thus, when exposure in a rodent is initiated at 5-8 weeks, as in the
9 typical rodent bioassay, and the bioassay is terminated after 104 weeks of exposure, the unit risk
10 estimate derived from the resulting cancer incidence data is considered a unit risk estimate from
11 lifetime exposure, except when the ADAFs were formulated and are applied, in which case the
12 same estimate is considered to apply to adult-only exposure. Yet, when adult exposures are
13 considered in the application of ADAFs, the adult-exposure-only unit risk estimate is pro-rated
14 over the full default (average) human lifespan of 70 years, presumably because that is how adult
15 exposures are treated when a unit risk estimate calculated in the same manner from the same
16 bioassay exposure paradigm is taken as a lifetime unit risk estimate.
17 However, in humans, a greater proportion of time is spent in childhood (e.g., 16 of
18 70 years = 23%), and the distinction between lifetime exposure and adult-only exposure cannot
19 be ignored when human data are used as the basis for the unit risk estimates. Thus, as described
20 above, adult-exposure-only unit risk estimates were calculated distinct from the lifetime
21 estimates that were derived in Section 4.1 under the assumption of age independence for all ages.
22 In addition, the adult-exposure-only unit risk estimates need to be rescaled to a 70-year lifespan
23 in order to be used in the ADAF calculations and risk estimate calculations involving less-than-
24 lifetime exposure scenarios in the standard manner, which includes prorating even adult-based
25 unit risk estimates over 70 years. Thus, the adult-exposure-only unit risk estimates are
26 multiplied by 70/54 to rescale the 54-year adult period of the 70-year default lifespan to 70 years.
27 Then, for example, if a risk estimate were calculated for a less-than-lifetime exposure scenario
28 involving exposure only for the full adult period of 54 years, the rescaled unit risk estimate
29 would be multiplied by 54/70 in the standard calculation and the adult-exposure-only unit risk
30 estimate would be appropriately reproduced. Without reseating the adult-exposure-only unit risk
31 estimates, the example calculation just described for exposure only for the full adult period of
32 54 years would result in a risk estimate 77% (i.e., 54/70) of that obtained directly from the
33 adult-exposure-only unit risk estimates, which would be illogical. The rescaled adult-based unit
34 risk estimates for use in ADAF calculations and risk estimate calculations involving less-than-
This document is a draft for review purposes only and does not constitute Agency policy.
4-72 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
lifetime exposure scenarios are presented in Table 4-22. Rescaled LECoi and ECoi estimates for
adult-based total cancer risk are 5.0 x ICT3 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
Adult-based unit risk estimate (per
ppm)
Adult-based unit risk estimate (per
Ug/m3)
Preferred Models3
Lymphoid cancer mortality
Lymphoid cancer incidence
Breast cancer mortality
Breast cancer incidence
Total cancer incidence
0.368
0.795
0.436
1.50b
1.98b
2.01 x 1Q-4
4.35 x 1Q-4
2.39 x 1Q-4
8.21 x 1Q-4
1.08 x 10 3
Alternative Models
Breast cancer incidence0
Total cancer incidence"1
Lymphoid cancer incidence6
Total cancer incidence'
0.386
1.03b
1.56b
2.56b
2.11 x 10'4
5.64 x 10'4
8.53 x 10'4
1.40 x 10'3
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
aLinear regression of categorical results for lymphoid cancer and breast cancer mortality; two-piece linear spline
model for breast cancer incidence.
bFor unit risk estimates above 1, convert to risk per ppb (e.g., 1.16perppm= 1.16 x 10"3perppb).
°Linear RR model for breast cancer incidence, representing the other (low) end of the range of reasonable
alternatives to the two-piece linear spline model, which is at the high end of the range.
dWith preferred lymphoid cancer incidence model but alternative linear RR model for breast cancer incidence,
representing the low end of the range of reasonable estimates for total cancer risk.
Two-piece log-linear spline model with knot at 1,600 ppm x days, an alternative model for lymphoid cancer based
on continuous exposure data.
fWith preferred breast cancer incidence model but alternative two-piece log-linear spline model with knot at
1,600 ppm x days for lymphoid cancer incidence.
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
size, the ADAF calculation is fairly straightforward. Thus, the ADAF-adjusted lifetime total-
cancer unit risk estimate is calculated as follows:
This document is a draft for review purposes only and does not constitute Agency policy.
4-73 DRAFT—DO NOT CITE OR QUOTE
-------
1 total cancer risk from exposure to constant EtO exposure level of 1 ug/m3 from ages 0-70 years:
2
3 unit risk exposure duration partial
4 Age group ADAF (per ug/m3) cone (fjg/m3) adjustment risk
5 0 to <2 years 10 1.08 x 10~3 1 2 years/70 years 3.09 x 10~4
6 2to<16years 3 1.08 x 1Q~3 1 14 years/70 years 6.48 x 1Q~4
7 >16 years 1 1.08 x 10~3 1 54 years/70 years 8.33 x 1Q'4
8 total lifetime risk = 1.80 x 10~3
9
10 The partial risk for each age group is the product of the values in columns 2-5 [e.g.,
11 10 x (1.08 x 1Q~3) x 1 x 2/70 = 3.09 x 10~4], and the total risk is the sum of the partial risks.
12
13
14 This 70-year risk estimate for a constant exposure of 1 ug/m3 is equivalent to a lifetime
15 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
16 potential increased early-life susceptibility, assuming a 70-year lifetime and constant exposure
17 across age groups. Note that because of the use of the rescaled adult-based unit risk estimate, the
18 partial risk for the >16 years age group is the same as would be obtained for a 1 ug/m3 constant
19 exposure directly from the total cancer adult-exposure-only unit risk estimate of 8.36 x 10~4 per
20 ug/m3 that was presented above, as it should be (the small difference in the second decimal place
21 i s due to round-off error).
22 In addition to the uncertainties discussed above for the inhalation unit risk estimate, there
23 are uncertainties in the application of ADAFs to adjust for potential increased early-life
24 susceptibility. The ADAFs reflect an expectation of increased risk from early-life exposure to
25 carcinogens with a mutagenic mode of action (U.S. EPA, 2005b), but they are general
26 adjustment factors and are not specific to EtO. With respect to the breast cancer estimates, for
27 example, evidence suggests that puberty/early adulthood is a particularly susceptible life stage
28 for breast cancer induction (U.S. EPA, 2005b: Russo and Russo, 1999): however, EPA has not,
29 at this time, developed alternate ADAFs to reflect such a pattern of increased early-life
30 susceptibility, and there is currently no EPA guidance on an alternate approach for adjusting for
31 early-life susceptibility to potential breast carcinogens.
32
33 4.5. INHALATION UNIT RISK ESTIMATES—CONCLUSIONS
34 For both humans and laboratory animals, tumors occur at multiple sites. In humans, there
35 was a combination of tumors having lymphohematopoietic, in particular lymphoid, origins in
36 both sexes and breast cancer in females, and, in rodents, lymphohematopoietic tumors, mammary
37 carcinomas, and tumors of other sites were observed. From human data, an extra cancer unit risk
This document is a draft for review purposes only and does not constitute Agency policy.
4-74 DRAFT—DO NOT CITE OR QUOTE
-------
1 estimate of 4.79 x 10 4 per ug/m3 (8.77 x 10 4 per ppb) was calculated for lymphoid cancer
2 incidence, and a unit risk estimate of 9.49 x icf4 per ug/m3 (1.74 x icf3 per ppb) was calculated
3 for breast cancer incidence in females, under the assumption that RR is independent of age for all
4 ages (see Section 4.1). The total extra cancer unit risk estimate was 1.24 x 10 3 per ug/m3
5 (2.27 x 1CT3 per ppb) for both cancer types combined (ECoi = 0.00775 ppm; LECoi =
6 0.00441 ppm). Unit risk estimates derived from the three chronic rodent bioassays for EtO
7 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
8 the estimates based on human data.
9 Because a mutagenic mode of action for EtO carcinogenicity (see Section 3.4.1) 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, in accordance with EPA's Supplemental Guidance
13 (U.S. EPA, 2005b). This assumption of increased early-life susceptibility supersedes the
14 assumption of age independence under which the human-data-based estimates presented in the
15 previous paragraph were derived. Thus, as described in Section 4.4, adult-exposure-only (i.e.,
16 ages > 16 years) unit risk estimates were calculated from the human data under an alternate
17 assumption that RR is independent of age for adults, which represent the life stage for which the
18 data upon which the exposure-response modeling was conducted pertain. These adult-exposure-
19 only unit risk estimates were then rescaled to a 70-year basis to derive adult-based unit risk
20 estimates for use in the standard ADAF calculations and risk estimate calculations involving
21 less-than-lifetime exposure scenarios. The resulting adult-based unit risk estimates were 4.35 x
22 10~4 per ug/m3 (7.95 x 10~4 per ppb) for lymphoid cancer incidence and 8.21 x io~4 per ug/m3
23 (1.50 x io~3 per ppb) for breast cancer incidence in females. The adult-based total extra cancer
24 unit risk estimate for use in ADAF calculations and risk estimate calculations involving
25 less-than-lifetime exposure scenarios was 1.08 x 10~3 per ug/m3 (1.98 x 10~3 per ppb) for both
26 cancer types combined.
27 For exposure scenarios involving early-life exposure, the ADAFs should be applied, in
28 accordance with EPA's Supplemental Guidance (U.S. EPA, 2005b). Applying the ADAFs to
29 obtain a full lifetime unit risk estimate yields
30
31
32 1.98/ppm x ((10 x 2 years/70 years) + (3 x 14/70) + (1 x 54/70)) (4-8)
33 = 3.29/ppm = 1.80 x 10~3/(ug/m3).
34
35
This document is a draft for review purposes only and does not constitute Agency policy.
4-75 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.41 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, the major ones related to
16 exposure misclassification, model uncertainty, and low-dose extrapolation, as discussed in
17 Section 4.1.4, confidence in the unit risk estimate is relatively high. First, there is strong
18 confidence in the hazard characterization of EtO as "carcinogenic to humans," which is based on
19 strong epidemiological evidence supplemented by other lines of evidence, such as genotoxicity
20 in both rodents and humans (see Section 3.5.1). Second, the unit risk estimate is based on human
21 data from a large, high-quality epidemiology study with individual worker exposures estimated
22 using a high-quality regression model (see Section 4.1 and Section A.2.8 of Appendix A).
23 Finally, the use of low-exposure linear extrapolation is strongly supported by the conclusion that
24 EtO 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
41 Technically, this unit risk estimate reflects the total (upper bound) cancer risk to females and not to the general
population because the breast cancer risk applies only to females. As a practical matter for regulatory purposes,
however, females comprise roughly half the general population, and this unit risk estimate enables risk managers to
evaluate the individual risk for this substantial population group. For the purposes of estimating numbers of cancer
cases attributable to specific exposure levels, e.g., for benefits analyses, it would be more appropriate to use the
cancer-specific unit risk estimates (or central tendency estimates), taking sex into account.
This document is a draft for review purposes only and does not constitute Agency policy.
4-76 DRAFT—DO NOT CITE OR QUOTE
-------
1 relevance for the derivation of a unit risk estimate. Alternate estimates calculated from other
2 reasonable models suggest that a unit risk estimate for breast cancer incidence that is up to
3 fourfold lower (corresponding to a total cancer unit risk estimate up to twofold lower) is
4 plausible; however, unit risk estimates notably lower than that are considered unlikely from the
5 available data.
6 There is lower confidence in the lymphoid cancer component of the unit risk estimate
7 because it is based on fewer events (40 lymphoid cancer deaths); incidence risk was estimated
8 from mortality data; and the exposure-response relationship is exceedingly supralinear, such that
9 continuous exposure models yield apparently implausibly steep low-exposure slopes (see
10 Figure 4-1). Although these continuous exposure models provided statistically significant slope
11 coefficients, there was low confidence in such steep slopes, which, particularly for the two-piece
12 spline model with the maximum likelihood, are highly dependent on a small number of cases in
13 the low-exposure range. Thus, a linear regression model of the categorical results for the lowest
14 three quartiles was used to derive the unit risk estimate for lymphoid cancer, and there was
15 greater confidence in the more moderate slope resulting from that model, although it was not
16 statistically significant, because it was based on more data and provided a good representation of
17 the categorical results across this larger data range in the lower-exposure region (see
18 Section 4.1.1.2). So, while there is lower confidence in the lymphoid cancer unit risk estimate
19 than in the breast cancer unit risk estimate, the lymphoid cancer estimate is considered a
20 reasonable estimate from the available data, and overall, there is relatively high confidence in the
21 total cancer unit risk estimate. An alternative lymphoid cancer model (two-piece log-linear
22 spline model with the knot defined by a local maximum likelihood rather than the largest
23 likelihood), judged to be the most suitable model based on the continuous exposure data, was
24 also considered for comparison. This alternative model yielded a unit risk estimate for lymphoid
25 cancer incidence two times that from the linear regression of the categorical results, resulting in a
26 total cancer unit risk estimate about 28% higher than the total cancer risk estimate with the
27 preferred lymphoid cancer model.
28 The unit risk estimate is intended to be an upper bound on cancer risk for use with
29 exposures below the POD (i.e., the LECoi). The unit risk estimate should not generally be used
30 above the POD; however, in the case of this total extra cancer unit risk, which is based on cancer
31 type-specific unit risk estimates from two linear models, the estimate should be valid for
32 exposures up to about 0.075 ppm (140 ug/m3), which is the minimum of the limits for the
33 lymphoid cancer unit risk estimate (0.090 ppm: see Section 4.1.1.2) and the breast cancer unit
34 risk estimate (0.075 ppm; see Section 4.1.2.3). (See Section 4.7 for risk estimates based on
35 occupational exposure scenarios.)
This document is a draft for review purposes only and does not constitute Agency policy.
4-77 DRAFT—DO NOT CITE OR QUOTE
-------
1 Using the above full lifetime unit risk estimate of 3.3 x 10 3 per ppb (1.8 x 10 3 per
2 ug/m3), the (lower bound) lifetime chronic exposure level of EtO corresponding to an increased
3 cancer risk of 10 6 can be estimated as follows:
4
5
6 (10~6)/(3.3/ppm) = 3.0 x 10~7 ppm = 0.00030 ppb = 0.0006 ug/m3. (4-9)
7
8
9 The inhalation unit risk estimate presented above, which is calculated based on a linear
10 extrapolation from the POD (LECoi), is expected to provide an upper bound on the risk of cancer
11 incidence. For some applications, however, estimates of "central tendency" for the risk below
12 the POD are desired. Thus, adult-based extra risk estimates per ppm for some of the cancer
13 responses based on linear extrapolation from the adult-exposure-only ECoi (i.e., 0.01/ECoi), and
14 rescaled to a 70-year basis for use in ADAF calculations and risk estimate calculations involving
15 less-than-lifetime exposure scenarios (see Section 4.4), are reported in Table 4-23. The
16 adult-exposure-only ECoiS were from the linear regression models of the categorical results for
17 lymphoid cancers and breast cancer mortality and from the two-piece linear spline model
18 (low-dose segment) for breast cancer incidence. (Note that, for each of these models, the
19 low-exposure extrapolated estimates are a straight linear continuation of the linear models used
20 above the PODs, and thus, the statistical properties of the models are preserved.) These estimates
21 are dependent on the suitability of the ECoi estimates as well as on the applicability of the linear
22 low-dose extrapolation. The assumption of low-dose linearity is supported by the mutagenicity
23 of EtO (see Section 3.4). If these estimates are to be used, ADAFs should be applied if early-life
24 exposure occurs, in accordance with EPA's Supplemental Guidance (U.S. EPA, 2005b).
25 As can be seen by comparing the adult-based 0.01/ECoi estimates in Table 4-23 with the
26 adult-based unit risk estimates (i.e., 0.01/LECoi estimates) in Table 4-22, the 0.01/ECoi estimates
27 are about 45% of the unit risk estimates for the lymphoid cancer responses and about 50% of the
28 unit risk estimates for the breast cancer responses.
29
30
This document is a draft for review purposes only and does not constitute Agency policy.
4-78 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
Table 4-23. Adult-based extra risk estimates per ppm based on
adult-exposure-only ECoiS (0.01/ECoi estimates)"
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
Cancer response
Lymphoid cancer mortality (both sexes)
Lymphoid cancer incidence (both sexes)
Breast cancer mortality (females)
Breast cancer incidence (females)
Total cancer incidence
Adult-exposure-only
ECoi (ppm)
0.0787
0.0364
0.0590
0.0167
0.0114
Adult-based
0.01/ECoi (per ppm)b'c
0.165
0.356
0.219
0.776
1.14d
"ADAFs should be applied to the adult-based 0.01/ECM estimates 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 conversion to per ug/m3, divide by 1830.
dFor unit risk estimates above 1, convert to risk per ppb (e.g., 1.14perppm= 1.14 x 10~3perppb).
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
make it hard to observe statistically significant increases in risk from low levels of exogenous
exposure. However, there is clear evidence of carcinogenic hazard from the rodent bioassays
and strong evidence from human studies (see Section 3.5), and the genotoxicity/mutagenicity of
EtO (see Section 3.4) supports low-dose linear extrapolation of risk estimates from those studies
(U.S. EPA, 2005a). In fact, as discussed in Section 3.3.3.1, Marsden et al. (2009), using
sensitive detection techniques and an approach designed to separately quantify endogenous
N7-HEG adducts and "exogenous" N7-HEG adducts induced by EtO treatment in rats, reported
increases in exogenous adducts in DNA of spleen and liver at the lowest dose administered
(0.0001 mg/kg injected i.p. daily for 3 days, which is a very low dose compared to the LOAELs
in the carcinogenicity bioassays; see Section C.7 of Appendix C). Marsden et al. (2009) also
reported statistically significant linear dose-response relationships (p < 0.05) for exogenous
adducts in all three tissues examined (spleen, liver, and stomach), although they caution that
some of the adduct levels induced at low EtO concentrations are below the limit of accurate
quantitation. Furthermore, while the contributions to DNA damage from low exogenous EtO
exposures may be relatively small compared to those from endogenous EtO exposure, low levels
of exogenous EtO may nonetheless be responsible for additional risk (above background risk).
This document is a draft for review purposes only and does not constitute Agency policy.
4-79 DRAFT—DO NOT CITE OR QUOTE
-------
1 This is not inconsistent with the much higher levels of background cancer risk, to which
2 endogenous EtO may contribute, for the two cancer types observed in the human
3 studies—lymphoid cancers have a background lifetime incidence risk on the order of 3%, while
4 the background lifetime incidence risk for breast cancer is on the order of 15%.42
5 Also related to the issue of endogenous EtO, Starr and Swenberg (2013) have proposed
6 an approach for bounding the cancer risk from low levels of exogenous exposure to chemicals
7 that also exist endogenously. In brief, Starr and Swenberg (2013) assume that all background
8 cancer risk (Po) for a specific cancer type is attributable to background levels (Co) of some
9 endogenous adduct (as a marker of exposure) of the chemical of interest [Starr and Swenberg
10 (2013) use formaldehyde as an example] in that tissue; they then use the ratio Po/Co (actually the
11 lower bound on Co) to estimate a linear slope for risk as a function of endogenous adduct level
12 down to zero adducts, which they claim is a conservative (upper) bound on cancer risk from low
13 levels of exogenous exposure (similarly expressed in terms of adduct levels). EPA disagrees that
14 this approach necessarily yields a conservative bound on risk, however, because, even //the
15 adduct level is an appropriate dose metric for comparing cancer risks, the approach relies on the
16 assumption that the dose-response relationship over the dose range of endogenous adducts is
17 linear down to zero adducts. In contrast to this assumption, EPA considers it highly plausible
18 that the dose-response relationship over the endogenous range is sublinear (e.g., that the baseline
19 levels of DNA repair enzymes and other protective systems evolved to deal with endogenous
20 DNA damage would work more effectively for lower levels of endogenous adducts), that is, that
21 the slope of the dose-response relationship for risk per adduct would increase as the level of
22 endogenous adducts increases. If the dose-response relationship over the endogenous range is
23 sublinear, rather than linear as assumed by Starr and Swenberg (2013), then the approach
24 proposed by Starr and Swenberg (2013) does not necessarily produce a conservative bound on
25 risk.
26 See Table 4-24 for a summary of key unit risk estimates derived in this assessment. See
27 Section 4.7 for risk estimates based on occupational exposure scenarios.
28
29
42These 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 of Ro 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.
4-80 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)
1.8 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
cohort) — preferred model: linear regression of categorical
results
Breast cancer incidence in females based on human data
(NIOSH cohort) — preferred model: 2-piece linear spline
model
Lymphoid cancer incidence based on human data (NIOSH
cohort) — range based on 2 models: the preferred model (linear
regression of categorical results) and an alternative continuous
exposure model (2-piece log-linear spline model)
Breast cancer incidence in females based on human data
(NIOSH cohort) — range based on three reasonable statistically
significant continuous exposure models: 2-piece linear spline
model, 2-piece log-linear spline model, and linear model
Total cancer risk based on human data (NIOSH cohort) — range
based on total estimates from range of lymphoid cancer
incidence estimates (linear regression of categorical results and
2-piece log-linear spline model) 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
cohort) — linear regression of categorical results
Breast cancer mortality in females based on human data
(NIOSH cohort) — 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)
1.1 x 10 3
4.3 x 1(T4
8.2 x 1(T4
4.3 x 10'4 to 8.5 x 10'4
2.1 x lQ-4to8.2 x 1(T4
5.6 x 10'4to 1.4 x 1(T3
2.0 x 1(T4
2.4 x 1(T4
4.6 x 1(T5
2.2 x 10'5 to 4.6 x 10'5
This document is a draft for review purposes only and does not constitute Agency policy.
4-81 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
Inhalation unit risk estimate"
(per jig/m3)b
0.01/ECoiestimatese
Lymphoid cancer incidence based on human data (NIOSH
cohort) — linear regression of categorical results
Breast cancer incidence in females based on human data
(NIOSH cohort) — estimate based on best-fitting model: the 2-
piece linear spline model
Lymphoid cancer mortality based on human data (NIOSH
cohort) — linear regression of categorical results
Breast cancer mortality in females based on human data
(NIOSH cohort) — linear regression of categorical results
Total cancer incidence based on human data (NIOSH cohort)
1.9 x 1(T4
4.2 x 1(T4
9.0 x 1(T5
1.2 x 1(T4
6.2 x 1(T4
1 ""Technically, the values listed in this table are not all unit risk estimates as defined by EPA, but they are all potency
2 estimates that, when multiplied by an exposure value, give an estimate of extra cancer risk. These potency estimates
3 are not intended for use with continuous lifetime exposure levels above 140 ug/m3. See Section 4.7 for risk
4 estimates based on occupational exposure scenarios. Preferred estimates are in bold.
5 bTo convert unit risk estimates to (ppm)"1, multiply the (ug/m3)"1 estimates by 1,830 (ug/m3)/ppm.
6 ° Because the weight of evidence supports a mutagenic mode of action for EtO carcinogenicity, and because of the
7 lack of chemical-specific data, EPA assumes increased early-life susceptibility and recommends the application of
8 ADAFs, in accordance with EPA's Supplemental Guidance (U.S. EPA. 2005b). for exposure scenarios that include
9 early-life exposures. For the full lifetime (upper-bound) unit risk estimate presented here, ADAFs have been
10 applied, as described in Section 4.4.
11 dThese (upper-bound) unit risk estimates are intended for use in ADAF calculations and less-than-lifetime adult
12 exposure scenarios (U.S. EPA. 2005b). Note that these are not the same as the unit risk estimates derived directly
13 from the human data in Section 4.1 under the assumption that RRs are independent of age. Under that assumption,
14 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
15 cancer incidence, and 1.2 x 10"3 per ug/m3 for the combined cancer incidence risk from those two cancers. See
16 Section 4.4 for the derivation of the adult-based unit risk estimates.
17 eThese are not upper-bound risk estimates but, rather, estimates based on linear extrapolation from the ECM.
18 ADAFs should be applied if early-life exposure occurs, in accordance with EPA's Supplemental Guidance (U.S.
19 EPA. 2005b).
20 4.6. COMPARISON WITH OTHER PUBLISHED RISK ESTIMATES
21 The unit risk values derived in this document are compared with other recent risk
22 estimates presented in the published literature (see Table 4-25).
23
24 4.6.1. Unit Risk Estimates Based on Human Studies
25 Kirman et al. (2004) used leukemia data only and pooled data from both the Stayner et al.
26 (1993) and the UCC studies (Tetaetal.. 1999: Tetaetal.. 1993). Based on the assumption that
This document is a draft for review purposes only and does not constitute Agency policy.
4-82 DRAFT—DO NOT CITE OR QUOTE
-------
1
2
3
4
5
6
7
leukemias are due to chromosome translocations, requiring two independent events
(chromosome breaks), Kirman et al. (2004) proposed that two independent EtO-induced events
are required for EtO-induced leukemias and used a dose-squared model, yielding a unit risk
3\-l
value of 4.5 x 10 (ug/m ) as their preferred estimate.
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 cohort)d
Breast cancer incidence in female
sterilizer workers (NIOSH cohort)6
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 1Q-3
4.5 x 1Q-8
Range of 1.4 x 10-8tol.4 x 10'7 f
5.5 x 10-7tol.6x 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 lQ-8to 1.5 x 10'5h
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
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.
This document is a draft for review purposes only and does not constitute Agency policy.
4-83 DRAFT—DO NOT CITE OR QUOTE
-------
1 Table 4-25. Comparison of unit risk estimates" (continued)
2
3
4 Tor breast cancer mortality, the ADAF-adjusted lifetime unit risk estimate is 4.0 x 10~4 (ug/m3)-1 and the adult-
5 based unit risk estimate is 2.4 x 10~4 (ug/m3)-1.
6 Estimates based on linear extrapolation from ECOOO1-ECOOOOO1 obtained from the quadratic model.
7 Estimates based on range of EC(l/million)s of 0.001-0.003 ppm obtained from the model RR = e(P * exposure) for
8 relevant cancer endpoints.
9 Estimates based on quadratic extrapolation model below the observable range of the data (i.e., below the LECio or
10 LECoi obtained using multistage model) with various points of departure (LEC0i-LEC0ooooi) for final linear
11 extrapolation (see Section 4.4.2).
12
13
14 The Kirman et al. (2004) values are different from those in the current document because
15 of the different assumptions inherent in the Kirman et al. (2004) approach and because the study
16 used unpublished data from earlier follow-ups of the two cohorts. A key difference is that EPA
17 uses a linear model or a two-piece linear spline model rather than a quadratic (dose-squared)
18 model in the range of observation. Then, EPA uses a higher extra risk level (1%) for
19 establishing the POD, whereas Kirman et al. (2004) used a risk level of 10"5 for their best
20 estimate and a risk range of 10"4 to 10"6 for their range of values. The extra risk level and the
21 corresponding POD are not critical with the linear model; however, with the quadratic model
22 used by Kirman et al. (2004), the lower the risk level (and hence the POD), the greater the
23 impact of the quadratic model and the lower the resulting unit risk estimates.
24 In addition, EPA (1) uses data for lymphoid cancers (and female breast cancers) rather
25 than leukemias, (2) includes ages up to 85 years in the life-table analysis rather than stopping at
26 70 years, (3) calculates unit risk estimates for cancer incidence as well as mortality, (4) uses a
27 lower bound as the POD rather than the maximum likelihood estimate, (5) uses the results of
28 lagged analyses rather than unlagged analyses, and (6) uses adult-based unit risk estimates in
29 conjunction with ADAFs (see Section 4.4) to derive the lifetime unit risk estimates.
30 Another key difference is that Kirman et al. (2004) relied on earlier NIOSH results
31 (Stayner et al., 1993), whereas EPA uses the results of NIOSH's more recent follow-up of the
32 cohort (Steenland et al., 2004). Kirman et al. (2004) claim that a quadratic dose-response model
33 provided the best fit to the data in the observable range and that this provides support for their
34 assumed mode of action. However, the 2004 NIOSH data for lymphohematopoietic cancers
35 suggest a supralinear exposure-response relationship (see Section 4.1.1.2 and Figures 4-1 and
36 4-2), which is inconsistent with a dose-squared model. Furthermore, EPA's review of the mode
37 of action evidence does not support the mode of action assumed by Kirman et al. (2004) (see
38 Section 3.4).
This document is a draft for review purposes only and does not constitute Agency policy.
4-84 DRAFT—DO NOT CITE OR QUOTE
-------
1 The Valdez-Flores et al. (2010) unit risk estimates (see Table 4-25) are similarly much
2 lower than those in the current document because of the different assumptions used. A key
3 difference is that EPA uses a linear model or a two-piece linear spline model in the range of
4 observation rather than an exponential model (RR = e^x exposure)3 which was used by Valdez-
5 Flores et al. (2010) despite its lack of fit. Then, EPA uses a 1% extra risk level for establishing
6 the POD for linear extrapolation, whereas Valdez-Flores et al. (2010) used a risk level of 10'6. In
7 addition, EPA (1) includes ages up to 85 years in the life-table analysis rather than stopping at
8 70 years, (2) calculates unit risk estimates for cancer incidence as well as mortality, (3) uses a
9 lower bound as the POD rather than the maximum likelihood estimate, and (4) uses the results of
10 lagged analyses rather than unlagged analyses. See Appendix A.2.20 for a more detailed
11 discussion of the differences between the EPA and Valdez-Flores et al. (2010) analyses.
12
13 4.6.2. Unit Risk Estimates Based on Laboratory Animal Studies
14 Kirman et al. (2004) also used linear and dose-squared extrapolation models to derive
15 unit risk estimates based on the rat mononuclear cell leukemia data and the mouse lymphoma
16 data. First, they used the multistage model to calculate the LECio (LECoi for the male mouse
17 lymphoma data) for the POD from the observable range. Then, using these PODs for linear
18 extrapolation, Kirman et al. (2004) obtained a unit risk range of 3.9 x 10~6 (ug/m3)"1 to
19 1.5 x 10~5 (ug/m3)'1. Alternatively, Kirman et al. (2004) used a quadratic extrapolation model
20 below the observable range to estimate secondary points of departure (LECoi-LECoooooi;
21 LECooi-LECoooooi for the male mouse) for final linear low-dose extrapolation, yielding unit risks
22 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
23 unit risks derived from the rodent data in this document.
24
25 4.7. RISK ESTIMATES FOR SOME OCCUPATIONAL EXPOSURE SCENARIOS
26 The unit risk estimates derived in the preceding sections were developed for
27 environmental exposure levels, where maximum modeled levels are on the order of 1-2 ug/m3
28 (email dated October 3, 2005, from Mark Morris, EPA, to Jennifer Jinot, EPA), i.e., roughly
29 0.5-1 ppb, and are not applicable to higher exposures, including some occupational exposure
30 levels. However, occupational exposure levels of EtO are of concern to EPA when EtO is used
31 as a pesticide (e.g., sterilizing agent or fumigant). The occupational exposure scenarios of
32 interest to EPA include some cumulative exposures corresponding to exposure levels in the
33 nonlinear range of some of the models (i.e., above the maximum exposure level at which the
34 low-dose-linear unit risk estimates apply). Therefore, extra risk estimates were calculated for a
35 number of occupational exposure scenarios of possible concern. Extra risk estimates are
This document is a draft for review purposes only and does not constitute Agency policy.
4-85 DRAFT—DO NOT CITE OR QUOTE
-------
1 estimates of the extra cancer risk above background and are the same type of estimate that one
2 gets from multiplying a unit risk estimate by an exposure level. In this case, the exposure level is
3 used directly in the exposure-response model, thus accounting for any nonlinearities in the model
4 above the range of exposure levels for which the linear unit risk estimate is applicable. For these
5 occupational exposure scenarios, exposure-response models based on data from the NIOSH
6 cohort were used in conjunction with the life-table program, as previously discussed in
7 Section 4.1. A 35-year exposure occurring between ages 20 and 55 years was assumed, and
8 exposure levels ranging from 0.1 to 1 ppm 8-hour TWA were examined (i.e., ranging from about
9 1,300 to 13,000 ppm x days). (Note that the current Occupational Safety and Health
10 Administration Permissible Exposure Limit is 1 ppm [8-hour TWA].)
11 For lymphoid cancer mortality in both sexes, the best-fitting (natural) log cumulative
12 exposure Cox regression model (see Steenland reanalyses in Appendix D; see also
13 Section 4.1.1.2), lagged 15 years, was used. The log cumulative exposure Cox regression model
14 was the best-fitting model for lymphoid cancer in males in the Steenland et al. (2004) study, and
15 the same model form is used here but with the data from both sexes. Although this model was
16 deemed too steep in the low-exposure region to be useful for the derivation of unit risk estimates
17 for lower (environmental) exposures, the model is well suited for the occupational exposure
18 scenarios of interest in this assessment because the corresponding cumulative exposures are well
19 within the range of the cumulative exposures in the NIOSH cohort. The model was statistically
20 significant (p = 0.02) and provided a better fit, based on AIC, than the two-piece spline models
21 with maximum likelihoods that were considered in Section 4.1.1.2 (the AICs were 460.426,
22 461.847, and 461.48 for the log cumulative exposure Cox regression, log-linear two-piece spline
23 and linear two-piece spline models, respectively [as reported in Section D.3 of Appendix D]; a
24 lower AIC indicates a better fit), as well as a more plausible, smoothly curved exposure-response
25 relationship than the two-piece spline models, both of which, with knots at 100 ppm x days, had
26 a very steep rise and then a very sharp change in slope at the knot. In addition, the log
27 cumulative exposure Cox regression model had a slightly lower AIC (460.426 versus 460.54)
28 than the log cumulative exposure linear model (see Section D.3.c of Appendix D) and has the
29 advantage of being a standard epidemiological model for continuous exposure data (the Cox
30 regression model, albeit with log cumulative exposure to accommodate the supralinearity of the
31 exposure-response data). The log cumulative exposure linear model yields slightly higher RR
32 estimates than the log cumulative exposure Cox regression model, as can be seen by comparing
33 the log cumulative exposure models in Figures D-3b and D-3c in Appendix D, and would thus
34 result in slightly higher extra risk estimates than the log cumulative exposure Cox regression
35 model. For example, the MLEs of extra risk from the log cumulative exposure linear model
This document is a draft for review purposes only and does not constitute Agency policy.
4-86 DRAFT—DO NOT CITE OR QUOTE
-------
1 would range from about 23% higher for the 0.1 ppm 8-hour TWA to 6% higher for the 1 ppm
2 8-hour TWA. (The two-piece log-linear spline model with the knot at 1,600 ppm x days [for
3 which there was a local maximum likelihood], which was considered as an alternative model for
4 the derivation of a unit risk estimate for lymphoid cancer, was not considered for the
5 occupational exposure extra risk estimates because there are better fitting models, discussed
6 above, that are suitable for the range of occupational exposures of interest.)
7 The extra risk results for lymphoid cancer mortality and incidence in both sexes for the
8 log cumulative exposure Cox regression model are presented in Table 4-26. For lymphoid
9 cancer incidence, the exposure-response relationship was assumed to be the same as for mortality
10 (see Section 4.1.1.3). As can be seen in Table 4-26, the extra risks for these occupational
11 exposure levels are in the "plateau" region of the exposure-response relationships and increase
12 less than proportionately with exposure. For occupational exposures less than about 1,000 ppm
13 x days, or about 0.08 ppm 8-hour TWA for 35 years, risk estimates are no longer in the plateau
14 region (see Figure 4-1) but rather in the steep low-exposure region, which is a region of greater
15 uncertainty for the log cumulative exposure model, and one might want to use the linear
16 regression of the categorical results that was used for lower exposures (see Section 4.1.1.2;
17 Appendix D). Furthermore, if one is using the linear regression model in this range and also
18 estimating risks for exposure levels in the range between about 0.08 and 0.6 ppm 8-hour TWA
19 (near where the linear regression and log cumulative exposure Cox regression models meet),
20 then one might want to use the linear regression model for the entire range up to 0.6 ppm 8-hour
21 TWA to avoid a discontinuity between the two models; thus, results for the linear regression
22 model for exposure levels up to 0.6 ppm 8-hour TWA are also presented in Table 4-26. While
23 the best-fitting continuous exposure model, the log cumulative exposure Cox regression model,
24 would generally be preferred in the exposure range between 0.08 and 0.6 ppm 8-hour TWA,
25 there is model uncertainty, so the use of either model could be justified. For exposures higher
26 than where the linear regression and log cumulative exposure Cox regression models meet, the
27 log cumulative exposure model exclusively is recommended. The models used to derive the
28 extra risk estimates presented in Table 4-26 for lymphoid cancer for the occupational exposure
29 scenarios are displayed in Figure 4-7 over the range of occupational cumulative exposures of
30 interest; the categorical results are included for comparison.
31 For breast cancer., incidence data were available from the NIOSH incidence study; thus,
32 only incidence estimates were calculated. In addition to being the preferred type of cancer risk
33 estimate, the breast cancer incidence risk estimates are based on more cases than were available
34 in the mortality study and the incidence data (for the subcohort with interviews) are adjusted for
35 a number of breast cancer risk factors (see Section 4.1.2.3). In terms of the incidence data, the
This document is a draft for review purposes only and does not constitute Agency policy.
4-87 DRAFT—DO NOT CITE OR QUOTE
-------
1 subcohort data are preferred to the full cohort data because the subcohort data are adjusted for
2 these potential confounders and also because the full cohort data have incomplete ascertainment
3 of breast cancer cases.
4 For breast cancer incidence in the subcohort with interviews, a number of Cox regression
5 exposure-response models from the Steenland et al. (2003) breast cancer incidence study fit
6 almost equally well (see Section 4.1.2.3). These include a log cumulative exposure model and a
7 cumulative exposure model, both with a 15-year lag, and a log cumulative exposure model with
8 no lag. The latter model was omitted from the calculations because the inclusion of a 15-year lag
9 for the development of breast cancer was considered more biologically realistic than not
10 including a lag. Steenland et al. (2003) also provide a duration-of-exposure Cox regression
11 model with a marginally better fit; however, models using duration of exposure are less useful
12 for estimating exposure-related risks, and duration of exposure and cumulative exposure are
13 correlated. Thus, only the lagged cumulative exposure models are considered here.
14
This document is a draft for review purposes only and does not constitute Agency policy.
4-88 DRAFT—DO NOT CITE OR QUOTE
-------
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 model of
categorical results'1
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 model of
categorical results'1
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
VO
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
-------
o
C5
S
I
O
a
i
(-\
**ij
8"
O 5
O I
3
o
H
W
o
c
o
H
W
1
O.I ppm x 35 years
t
I ppm x 35 years
Cumulative Exposure (ppm x days)
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. (Note that, with the exception of
the categorical results and the linear regression of the categorical results, the various models have different implicitly estimated
baseline risks; thus, they are not strictly comparable to each other in terms of RR values, i.e., along the y-axis. They are,
however, comparable in terms of general shape.)
-------
1 The extra risk estimates for breast cancer incidence in females from the lagged
2 cumulative exposure and log cumulative exposure Cox regression models listed above are
3 presented in Table 4-27. As can be seen in Table 4-27, the extra risk estimates for the lagged log
4 cumulative exposure and cumulative exposure models differ substantially. Furthermore, the
5 categorical Cox regression results for breast cancer incidence in the subcohort with interviews
6 suggest a more linear low-exposure exposure-response relationship than that obtained with either
7 the continuous variable log cumulative exposure (supralinear) or cumulative exposure (sublinear)
8 Cox regression models (see Figure 4-5). (The lowest four exposure quintiles represent individual
9 worker exposures ranging from 0 to about 15,000 ppm x days, which covers the range of
10 cumulative exposures for the occupational exposure scenarios of interest in this assessment, the
11 maximum of which is 12,775 ppm x days.) Therefore, the two-piece linear spline model (with a
12 15-year lag) (see Section 4.1.2.3) was also used to calculate the extra risk estimates. The
13 two-piece linear spline model provides a better fit to the data than the log cumulative exposure or
14 cumulative exposure Cox regression models, as indicated by a lower AIC value (1,950.9 for
15 two-piece linear spline model vs. 1,956.2 for the log cumulative exposure Cox regression model
16 and 1,956.8 for the cumulative exposure Cox regression model; Table 4-12 and Appendix D). In
17 fact, the two-piece linear spline model provided the best fit to the breast cancer incidence data of
18 all the models investigated in Section 4.1.2.3, and it provides the best representation of the
19 categorical RR results, particularly for the range of cumulative exposures for the occupational
20 exposure scenarios of interest (see Figures 4-5, 4-6, and 4-8). The extra risk estimates calculated
21 using the two-piece linear spline model are also presented in Table 4-27 and are the preferred
22 estimates because they are derived from the best-fitting model.
23 In addition, extra risk estimates for breast cancer incidence in females from the
24 continuous exposure linear model (with a 15-year lag) (see Section 4.1.2.3) are presented in
25 Table 4-27 for comparison. This model, with an AIC of 1,952.3 (see Table 4-12), was the
26 second-best-fitting model and also provided an adequate visual fit to the categorical data (see
27 Figure 4-6). Moreover, the two best-fitting models (i.e., the continuous exposure linear model
28 and two-piece linear spline model) span the range of RR estimates from the three best-fitting
29 models investigated in Section 4.1.2.3 (the third being the two-piece log-linear spline model)
30 over the range of cumulative exposures for the occupational exposure scenarios of interest in this
31 assessment (see Figure 4-6). Comparing the results of the two best-fitting models shows that the
32 extra risk estimates differ by just under fourfold at the lowest exposure level (0.1 ppm 8-hour
This document is a draft for review purposes only and does not constitute Agency policy.
4-91 DRAFT—DO NOT CITE OR QUOTE
-------
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
o
a
I
a.
O <$
H 1
O
V
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 Table 5 of Steenland et al. (2003). 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 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
levelsa'b (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.
<-t.
a MLE: maximum likelihood estimates; UCL: (one-sided) upper confidence limit estimate.
I
-
o
a
I
a,
O 5
31
o ^
H t
W ??•
o
c
o
H
W
-------
1 TWA) and the difference tapers to just over twofold at the highest exposure level (1 ppm 8-hour
2 TWA), with the estimates of the best-fitting model, the two-piece linear spline model, yielding
3 the higher extra risk estimates across the range.
4 Finally, also for comparison, maximum likelihood estimates (MLEs) of extra risk for the
5 log cumulative exposure Cox regression model with a 10-year lag were calculated. The model
6 with a 10-year lag also provided a statistically significant fit to the data (p = 0.03), whereas, the
7 models with 5- and 20-year lags did not. These estimates ranged from 0.067 for 0.1 ppm
8 exposure to 0.094 for 1.0 ppm exposure. Thus, the MLEs of extra risk with a 10-year lag were
9 about 20% higher than those with the 15-year lag.
10 The continuous exposure models (with a 15-year lag) considered for deriving the extra
11 risk estimates for breast cancer incidence in females for the occupational exposure scenarios are
12 displayed in Figure 4-8 over the range of occupational cumulative exposures of interest.
13 Categorical results are also presented for comparison (deciles from the categorical linear model
14 are presented because it had a better fit than the log-linear categorical model, as indicated by the
15 AICs, which were 1,963.9 and 1,966.9, respectively; Appendix D). The recommended model is
16 the two-piece linear spline model; this was the best-fitting continuous exposure model of those
17 evaluated in this assessment, and it provides the best visual fit in comparison to the categorical
18 results in the range of the occupational exposure scenarios of interest. As shown in Figure 4-8,
19 the log cumulative exposure Cox regression model is too flat across the range of exposures of
20 interest. It also appears from Figure 4-8 that the slopes of the cumulative exposure Cox
21 regression model and the (continuous exposure) linear model are both insufficiently steep across
22 the range of exposures of interest. This is consistent with the analysis presented in Section D. 1
23 of Appendix D showing the strong influence of the upper tail of cumulative exposures on the
24 results of the cumulative exposure Cox regression model. The responses in the upper tail of
25 exposures are relatively dampened, such that when the highest 5% of exposures (exposures >
26 27,500 ppm x days, which are well in excess of the exposures corresponding to the occupational
27 exposure scenarios considered here) are excluded, the slope of the Cox regression model is
28 substantially increased (e.g., at 10,000 ppm x days, the RR estimate increases from about 1.1 to
29 almost 1.5; see Figure D-ld in Appendix D). This strong influence of the upper tail of exposures
30 would similarly attenuate the slope of the linear model, resulting in underestimation of the
31 lower-exposure risks. The two-piece linear spline model, on the other hand, is more flexible, and
32 the influence of the upper tail of exposures would be primarily on the upper spline segment; thus,
33 the two-piece 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.
4-94 DRAFT—DO NOT CITE OR QUOTE
-------
o
C5
S
I
I
^
o
a
i
(-\
**ij
8"
O 5
O I
3
o
H
W
o
c
o
H
W
1
1.8
1.6
01 „ .
.> 1.4
_
01
QC
1.2
0.8
/
linear spline model
log cumulative
exposure Cox
regression model
— cumulative
exposure Cox
regression model
» breast cancer
deciles
— • linear model
0.1 ppm x 35 years
ppm x 35 years
4000 6000 8000
Cumulative Exposure (ppm x days)
Figure 4-8. RR estimates for breast cancer incidence from occupational EtO exposures (with 15-year lag).
Breast cancer models (see Section 4.1.2.3): linear 2-piece spline model, with knot at 5,800 ppm x days; log cumulative exposure Cox
regression model; (cumulative exposure) Cox regression model; (continuous exposure) linear model; categorical results (deciles) from
continuous exposure linear model. (Note that the various models have different implicitly estimated baseline risks; thus, they are not
strictly comparable to each other in terms of RR values, i.e., along the y-axis. They are, however, comparable in terms of general shape.)
-------
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.43
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 to EPA.
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
43Technically, these sums would reflect the total cancer risk to females and not a mixed-sex workforce because the
breast cancer risk applies only to females. As a practical matter for regulatory purposes, females typically comprise
a substantial proportion of the sterilizer workforce and summing the extra risk estimates enables risk managers to
evaluate the individual risk for this substantial workforce group. In a situation in which the workforce of concern is
comprised predominantly of males, it might be appropriate to use a sex-weighted sum of the extra risks from the two
cancer types. For the purposes of estimating numbers of cancer cases attributable to specific exposure levels, e.g.,
for benefits analyses, it would be most suitable to use the cancer-specific unit risk estimates (or central tendency
estimates), taking sex into account.
This document is a draft for review purposes only and does not constitute Agency policy.
4-96 DRAFT—DO NOT CITE OR QUOTE
-------
1 scenarios with durations of exposure less than 30-35 years, one could roughly estimate extra
2 risks by calculating the cumulative exposure and finding the extra risks for a similar cumulative
3 exposure in Tables 4-26 and 4-27. For a more precise estimation, or for exposure scenarios of
4 much shorter duration or for specific age groups, one should do the calculations using a life-table
5 analysis, as presented in Appendix E but modified for the specific exposure scenarios.
6
7 For occupational exposures below 0.1 ppm:
8
9 For lymphoid cancer, use of the log cumulative exposure Cox regression model is not
10 advised below 0.1 ppm (x 35 years). Instead, the low-exposure continuation of the linear
11 regression model presented in Table 4-26 of the assessment is recommended. For 35-year
12 exposures, the following formulae would apply:
13
14
15 95% UCL on extra risk for lymphoid cancer incidence ~ (8-h TWA occ exp [in ppm]) x
16 (0.016/0.1 ppm) = (8-h TWA occ exp [in ppm]) x (0.16/ppm)
17
18 MLE of extra risk for lymphoid cancer incidence = (8-h TWA occ exp [in ppm]) x
19 (0.007/0.1 ppm) = (8-h TWA occ exp [in ppm]) x (0. 07/ppm)
20
21
22 If one is considering occupational exposure scenarios using a range of 8-h TWA
23 exposure levels on both sides of 0.1 ppm, one might want to use the linear regression model for
24 all the exposure levels up to about 0.6 ppm 8-h TWA (approximately where the linear regression
25 model intersects the log cumulative exposure Cox regression model) to avoid the discontinuity
26 between the two models below where they intersect. Note that the extra risk estimates from the
27 different models differ by at most about 4.5-fold (at 0.1 ppm) and that there is model uncertainty
28 in this range, so the use of either model could be justified. Above where the models intersect,
29 only the log cumulative exposure Cox regression model should be used.
30
31 For breast cancer, the low-exposure continuation of the two-piece linear spline model
32 presented in Table 4-27 of the assessment is recommended. For 35-year exposures, the
33 following formulae would apply:
34
35
36 95% UCL on extra risk for breast cancer incidence ~ (8-h TWA occ exp [in ppm]) x
37 (0.031/0.1 ppm) = (8-h TWA occ exp [in ppm]) x (0. 31/ppm)
This document is a draft for review purposes only and does not constitute Agency policy.
4-97 DRAFT—DO NOT CITE OR QUOTE
-------
2 MLE of extra risk for breast cancer incidence = (8-h TWA occ exp [in ppm]) x
3 (0.016/0. Ippm) = (8-h TWA occ exp [in ppm]) x (0. 16/ppm)
4
5
6 Alternatively, for exposures below 0.1 ppm, one could use the formulae presented below,
7 which are based on the unit risk estimates presented in Tables 4-22 (for 95% UCLs) and
8 Table 4-23 (for MLEs), with conversions for adjusting occupational to environmental exposures.
9 Note, however, that the extra risk results for 35 years of exposure based on these unit risk values
10 do not exactly match the values in Tables 4-26 and 4-27 for the linear models (the formulae
11 below yield extra risk estimates that are 15-20% lower than the values in Tables 4-26 and 4-27
12 for the low end of the exposure range [e.g., 0.1-0.4 ppm] where the comparison with the
13 unit-risk-based estimates is appropriate). This is because the results in Tables 4-26 and 4-27 are
14 based on life-table analyses, which take into account age-specific background rates of the
15 cancers and ages of exposure (assumed to be from 20 to 55 years of age in these occupational
16 exposure scenarios), whereas the formulae below are approximations that do not take
17 age-specific considerations into account. The advantage of the formulae based on the unit risk
18 values is that they can incorporate durations other than -35 years.
19
20
21 8-h TWA occ exp [in ppm] x (10 m3/day/20 m3/day) x (240 days/year/365 days/year) x
22 (35 years/70 years) = (continuous lifetime) env exp [in ppm]
23 (Note that for exposure durations other than 35 years, replace 35 years with the
24 alternative duration in the formula above.)
25
26 95% UCL on extra risk for lymphoid cancer incidence ~ 0.795/ppm x env exp [in ppm]
27
28 MLE of extra risk for lymphoid cancer incidence = 0.356/ppm x env exp [in ppm]
29
30 95% UCL on extra risk for breast cancer incidence (in females) = 1.50/ppm x env exp [in
31 PPm]
32
33 MLE of extra risk for breast cancer incidence (in females) = 0.776/ppm x env exp [in
34 PPm]
This document is a draft for review purposes only and does not constitute Agency policy.
4-98 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. 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. Mutat Res 585: 120-126.
http://dx.doi.0rg/10.1016/i.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]. Mutat Res 231: 11-30.
Bisanti, L; Maggini, M; Raschetti, R: Alegiani, 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; van der Hel, O; Norppa, H; Fabianova, E; Fucic, A; Gundy, S; Lazutka, J; Cebulska-
Wasilewska, A; Puskailerova, D: Znaor, A; Kelecsenyi, Z; Kurtinaitis, J: Rachtan, J:
Forni, 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/aie/kwi367
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 Chromatogr B 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
This document is a draft for review purposes only and does not constitute Agency policy.
R-1 DRAFT—DO NOT CITE OR QUOTE
-------
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
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)/RAS-BRAF signal transduction pathway in therapy-
related myelodysplasia and acute myeloid leukemia. Leukemia 19: 2232-2240.
http://dx.doi.orR/10.1038/si.leu.2404009
Christiansen, 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; Sarangapani, 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]. Crit Rev 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, VL: Generoso, WM: Sega, GA: Fowle, JR, III: Jacob son-Kram, D. (1990). Review of
the mutagenicity of ethylene oxide [Review]. Environ Mol Mutagen 16: 85-103.
Donner, 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.
This document is a draft for review purposes only and does not constitute Agency policy.
R-2 DRAFT—DO NOT CITE OR QUOTE
-------
Post U: Hallien E: Ottenwalder. H: Bolt HM: Peter. H. (1991). Distribution of ethylene oxide in
human blood and its implications for biomonitoring. Hum Exp 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
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, RL: 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: Cornell 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; Lison, 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, AL: 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: Ill-Ill.
This document is a draft for review purposes only and does not constitute Agency policy.
R-3 DRAFT—DO NOT CITE OR QUOTE
-------
Hansen. JP: Allen. J: Brock. K: Falconer. J: Helms. MJ: Shaver. GC: Strohtn. 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. NL: Jaffe. ES: Diebold, J: Flandrin, G: Muller-Hermelink, HK: Vardiman, J: Listen 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.
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). Methods for detecting DNA damaging agents in humans: Applications in
cancer epidemiology and prevention Epidemiologic studies on ethylene oxide and cancer:
An updating. In IARC Sci Publ. 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: Ton, 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/l975_2006/
Hornung, RW: Greife, AL; Stavner, LT; Steenland, NK; Herrick, RF; Elliott U: 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 from
B6C3F1 mice. Toxicol Pathol 34: 752-762.
http://dx.doi.org/10.1080/01926230600935912
This document is a draft for review purposes only and does not constitute Agency policy.
R-4 DRAFT—DO NOT CITE OR QUOTE
-------
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). Ethyl ene oxide. In IARC
Monographs on the Evaluation of Carcinogenic Risks to Humans (pp. 73-159). Lyon,
France. http://monographs.iarc.fr/ENG/Monographs/vol60/volume60.pdf
IARC (International Agency for Research on Cancer). (2008). 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
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.
Kardos, L; Szeles, G; Gombkoto, G; Szeremi, 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/j.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.
Krewski, D: Crump. KS: Farmer. J: Gavlor, DW: Howe. R: Portier. C: Salsburg, D: Sielken. RL:
Van Ryzin, 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.
This document is a draft for review purposes only and does not constitute Agency policy.
R-5 DRAFT—DO NOT CITE OR QUOTE
-------
Lerda, D: Rizzi, R. (1992). Cytogenetic study of persons occupationally exposed to ethylene
oxide. MutatRes 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 ethyl ene 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 ethyl ene oxide. Mutat Res Genet Toxicol Environ
Mutagen 492: 59-67. http://dx.doi.org/10.1016/S1383-5718(01)00149-8
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.
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-
2280Q996)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,
1C (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.
This document is a draft for review purposes only and does not constitute Agency policy.
R-6 DRAFT—DO NOT CITE OR QUOTE
-------
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. In National Vital Statistics Reports. Hyattsville, MD: National Center for Health
Statistics, http://www.cdc.gov/nchs/data/nvsr/nvsr50/nvsr50_l 5.pdf
Natarajan, 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.
http://dx.doi.Org/10.1093/iie/24.2.276
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
NRC (National Research Council). (2011). Review of the Environmental Protection Agency's
draft IRIS assessment of formaldehyde. Washington, DC: National Academies Press.
http://www.nap.edu/catalog/13142.html
NTP (National Toxicology Program). (1987). Toxicology and carcinogenesis studies of ethylene
oxide (CAS no 75-21-8) in B6C3F1 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. MutatRes 418: 21-33.
Paz-y-Mifio. 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.
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/si.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,
1C (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.
This document is a draft for review purposes only and does not constitute Agency policy.
R-7 DRAFT—DO NOT CITE OR QUOTE
-------
Preston, RJ. (1999). Cytogenetic effects of ethylene oxide, with an emphasis on population
monitoring [Review]. CritRev 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 ethyl ene oxide exposures. Environ Mol Mutagen 26: 189-202.
http://dx.doi.org/10.1002/em.2850260303
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 ethyl ene oxide, glycide and glycol on genetic mutations.
Dokl Biochem Biophys 60: 469-472.
Recio, L; Donner, 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 ethyl ene oxide.
Mutagenesis 19: 215-222.
Ribeiro, LR; Salvador!, DM; Rios, AC: Costa, SL; Tates, AD; Tornqvist M; Natarajan, AT.
(1994). Biological monitoring of workers occupationally exposed to ethylene oxide.
MutatRes313: 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; Hankey, 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/l 975_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, 1975-2004. Bethesda, MD: National Cancer Institute, U.S.
Department of Health, Education, and Welfare, National Institutes of Health.
http://seer.cancer.gOv/csr/l 975_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
Rothman, KJ. (1986). Modern epidemiology. Boston, MA: Little Brown & Co.
Russo, J: Russo, IH. (1999). Cellular basis of breast cancer susceptibility [Review]. Oncol Res
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.Org/10.1016/j.dnarep.2005.05.009
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.
This document is a draft for review purposes only and does not constitute Agency policy.
R-8 DRAFT—DO NOT CITE OR QUOTE
-------
http://vosetnite.epa.gov/sab/sabproduct.nsf/368203f97al5308a852574ba005bbd01/5D66
lBC118B527A3852573B80068C97B/$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 andN-(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/i.vrtph.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.
Starr, TB; Swenberg, JA. (2013). A novel bottom-up approach to bounding low-dose human
cancer risks from chemical exposures. Regul Toxicol Pharmacol 65: 311-315.
http://dx.doi.0rg/10.1016/j.vrtph.2013.01.004
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. Am J Epidemiol 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 Environ Med 61:
2-7.
This document is a draft for review purposes only and does not constitute Agency policy.
R-9 DRAFT—DO NOT CITE OR QUOTE
-------
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
Steenland, K; Whelan, E; Deddens, J; Stavner, 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.
MutatRes 129: 89-102. http://dx.doi.org/10.1016/0027-5107(84)90127-1
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, GMH; Slangen, JMM; 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.
http://dx.doi.org/10.1007/BF00381432
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.), (pp. 161-184). New York, NY: Elsevier.
Tates, AD: Boogaard, PJ: Darroudi, F: Nataraian, AT: Caubo, ME: Van Sitteit 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: van Mossel, H: Schoemaker,
HM: Osterman-Golkar, S: Uebel, C: Tang, YS: Zwinderman, AH; Natarajan, AT;
Ehrenberg, L. (1991). Biological and chemical monitoring of occupational exposure to
ethylene oxide. MutatRes 250: 483-497. http://dx.doi.org/10.1016/0027-5107(91)90205-
3
Tates, AD; Van Dam, FJ; Natarajan, AT; Van Tevlingen, 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: Benson, 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]. CritRev 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: Brown. K. (2009). Mutagenicity of DNA
adducts derived from ethylene oxide exposure in the pSP189 shuttle vector replicated in
This document is a draft for review purposes only and does not constitute Agency policy.
R-10 DRAFT—DO NOT CITE OR QUOTE
-------
human Ad293 cells. Mutat Res 678: 129-137.
http://dx.doi.0rg/10.1016/i.tnrgentox.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.
U.S. EPA (U.S. Environmental Protection Agency). (1994). Methods for derivation of inhalation
reference concentrations and application of inhalation dosimetry. (EPA/600/8-90/066F).
Research Triangle Park, NC: U.S. Environmental Protection Agency, Environmental
Criteria and Assessment Office.
http://cfpub.epa.gov/ncea/cfm/recordisplay.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/recordisplay.cfm?deid=12181
U.S. EPA (U.S. Environmental Protection Agency). (2000). Science policy council handbook:
risk characterization. (EPA/100/B-00/002). Washington, D.C.: U.S. Environmental
Protection Agency, Office of Science Policy.
http://www.epa.gov/osa/spc/pdfs/rchandbk.pdf
U.S. EPA (U.S. Environmental Protection Agency). (2005a). Guidelines for carcinogen risk
assessment. (EPA/630/P-03/001F). Washington, DC: U.S. Environmental Protection
Agency, 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/recordisplay.cfm?deid=l57664
U.S. EPA (U.S. Environmental Protection Agency). (2006b). Science policy council peer review
handbook 3rd edition. (EPA/100/B-06/002). U.S. Environmental Protection Agency,
Science Policy Council. http://cfpub.epa.gov/ncea/cfm/recordisplay.cfm?deid=l57664
U.S. EPA (U.S. Environmental Protection Agency). (2012). Benchmark dose technical guidance.
(EPA/1 OO/R-12/001). Washington, DC: Risk Assessment Forum.
http://www.epa.gov/raf/publicati ons/pdfs/benchmark_dose_guidance.pdf
Valdez-Flores, C: Sielken, 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.Org/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.
This document is a draft for review purposes only and does not constitute Agency policy.
R-11 DRAFT—DO NOT CITE OR QUOTE
-------
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
Perspect99: 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/S0027-
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.Org/10.1016/j.scitotenv.2004.04.005
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/aiim.20443
This document is a draft for review purposes only and does not constitute Agency policy.
R-12 DRAFT—DO NOT CITE OR QUOTE
-------
Yong. LC: Schulte. PA: Wiencke. JK: Boeniger. MF: Connallv. 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.
R-13 DRAFT—DO NOT CITE OR QUOTE
------- |