AMERICAN JOURNAL OF EPIDEMIOLOGY
issn 0002-9262 (print)
Associate Director for Epidemiology and Genetics Research Program
Division of Cancer Control and Population Sciences
The National Cancer Institute (NCI), a major research component of the National Institutes of Health (NIH)
within the Department of Health and Human Services (DHHS), seeks a senior scientist to serve as an As-
sociate Director in its Division of Cancer Control and Population Sciences (DCCPS). The Division provides
national scientific leadership and oversight of NCI-funded research in the areas of cancer epidemiology, sur-
veillance, health services, survivorship, and behavioral science. DCCPS also is committed to addressing health
disparities through transdisciplinary research.
This challenging and highly visible position requires broad scientific expertise, a passion for public service,
commitment to collaboration, and an ability to develop effective strategies to identify gaps in research and
overcome barriers to scientific progress. Exploiting scientific opportunities requires visionary leadership and
sound scientific judgment to ensure the greatest payoff from NCI’s investments. Therefore, a broad perspective
on cancer epidemiology and other population sciences that informs development of creative and cost-effective
strategies to advance science is essential.
The Associate Director will lead the Epidemiology and Genetics Research Program (EGRP), which includes its
Office of the Associate Director and four branches: Clinical and Translational Epidemiology, Host Susceptibility
Factors, Methods and Technologies, and Modifiable Risk Factors. He/she provides scientific and administrative
leadership for the Program, supervises the staff, and represents NCI to a wide variety of professional, academic,
and advocacy organizations. The Associate Director also develops and facilitates collaborations with funders of
other types of population science, including the NIH Institutes and Centers, Centers for Disease Control and
Prevention (CDC), and many non-governmental organizations. EGRP’s grants, contracts, interagency agree-
ments, and operating budgets totaled more than $197 million in Fiscal Year 2008. Included were more than 370
research grants and numerous interagency agreements and consortia.
Volume 169
Qualifications: The successful applicant will be an experienced epidemiologist (M.D. or Ph.D.-level training re-
quired) with leadership experience, excellent communication skills, and a strong record of peer-reviewed publica-
tions. Strong leadership skills, an ability to work effectively across disciplinary boundaries, and a commitment to the
highest standards of scientific integrity and quality are required. Experience in managing complex research projects,
Number 8
scientific staff, training programs, interdisciplinary collaborations, and/or funded programs is highly valued.
Salary: This position is an excepted service position (Title 42) with a salary range of $160,000-$195,000.
How to Apply: Applications will be considered until the position is filled. Please submit a letter of interest and
April 15, 2009
CV to the Search Committee Chair:
Rachel Ballard-Barbash, M.D., M.P.H., Associate Director
Applied Research Program
Division of Cancer Control and Population Sciences
National Cancer Institute
Pages 919–1042
6130 Executive Blvd, Room 4005, MSC 7344
Bethesda, MD 20892-7344
Express Mail: Rockville, MD 20852
For more information about DCCPS and EGRP, see http://cancercontrol.cancer.gov
THE DHHS/NIH/NCI ARE EQUAL OPPORTUNITY EMPLOYERS
oxford
ISSN 1476-6256 (online)
printed in the u.s.a.
American Journal of
Epidemiology
Volume 169 Number 8 April 15, 2009
www.aje.oxfordjournals.org
ORIGINAL CONTRIBUTIONS
919 Parkinson’s Disease and Residential Exposure to Maneb and Paraquat From
Agricultural Applications in the Central Valley of California. Sadie Costello, Myles Cockburn,
Jeff Bronstein, Xinbo Zhang, and Beate Ritz
927 Overweight and Obesity Over the Adult Life Course and Incident Mobility Limitation in
Older Adults: The Health, Aging and Body Composition Study. Denise K. Houston, Jingzhong
Ding, Barbara J. Nicklas, Tamara B. Harris, Jung Sun Lee, Michael C. Nevitt, Susan M. Rubin, Frances
A. Tylavsky, and Stephen B. Kritchevsky for the Health ABC Study
937 Association of Diabetes With Prostate Cancer Risk in the Multiethnic Cohort. Kevin
M. Waters, Brian E. Henderson, Daniel O. Stram, Peggy Wan, Laurence N. Kolonel, and Christopher
A. Haiman
946 Modification of the Effect of Vitamin E Supplementation on the Mortality of Male
Smokers by Age and Dietary Vitamin C. Harri Hemilä and Jaakko Kaprio
954 Dietary Acrylamide Intake and Risk of Premenopausal Breast Cancer. Kathryn M. Wilson,
Lorelei A. Mucci, Eunyoung Cho, David J. Hunter, Wendy Y. Chen, and Walter C. Willett
962 Alcohol Intake and Cigarette Smoking and Risk of a Contralateral Breast Cancer: The
Women’s Environmental Cancer and Radiation Epidemiology Study. Julia A. Knight,
Leslie Bernstein, Joan Largent, Marinela Capanu, Colin B Begg, Lene Mellemkjær, Charles F. Lynch,
Kathleen E. Malone, Anne S. Reiner, Xiaolin Liang, Robert W. Haile, John D. Boice, Jr., WECARE Study
Collaborative Group, and Jonine L. Bernstein.
969 Positive Associations Between Ionizing Radiation and Lymphoma Mortality Among
Men. David B. Richardson, Hiromi Sugiyama, Steve Wing, Ritsu Sakata, Eric Grant, Yukiko Shimizu,
Nobuo Nishi, Susan Geyer, Midori Soda, Akihiko Suyama, Fumiyoshi Kasagi, and Kazunori Kodama
977 Biomarker-calibrated Energy and Protein Consumption and Increased Cancer Risk
Among Postmenopausal Women. Ross L. Prentice, Pamela A. Shaw, Sheila A. Bingham, Shirley
A. A. Beresford, Bette Caan, Marian L. Neuhouser, Ruth E. Patterson, Marcia L. Stefanick, Suzanne
Satterfield, Cynthia A. Thomson, Linda Snetselaar, Asha Thomas, and Lesley F. Tinker
Continued on Inside Front Cover
Published for the Johns Hopkins Bloomberg School of Public Health by Oxford
University Press Sponsored by the Society for Epidemiologic Research
Founded in 1920 by W. H. Welch and W. H. Howell as the American Journal of
Hygiene at the Johns Hopkins School of Hygiene and Public Health
 Founded 1920 by
W. H. Welch and
W. H. Howell as the
American Journal of Hygiene

BOARD OF EDITORS POLICY BOARD EDITOR-IN-CHIEF EMERITUS, EDITORIAL PROJECT

MOYSES SZKLO, Editor-in-Chief MICHAEL J. KLAG, Chairman IN MEMORIAM MANAGERS
DONNA K. ARNETT TODD HUMMEL GEORGE W. COMSTOCK DEBORAH J. ANDERSON
TERRI BEATY SHERMAN JAMES SANDRA L. BRYANT
HARVEY CHECKOWAY MALCOLM MACLURE ASSOCIATE EDITOR IN AMY W. REDMON-NORWOOD
AARON R. FOLSOM NOEL R. ROSE RESIDENCE
GARY D. FRIEDMAN SHRUTI H. MEHTA ADMINISTRATOR
ROBERT J. GLYNN HARRIETT TELLJOHANN
ICHIRO KAWACHI
MUIN J. KHOURY ADMINISTRATIVE
MARK KLEBANOFF ASSISTANT
MARTHA S. LINET CYNTHIA A. BARBRE
POLLY MARCHBANKS
POLLY NEWCOMB REVIEW COORDINATOR
JONATHAN M. SAMET GAYLE FINI
DAVID VLAHOV
CLARICE WEINBERG

ASSOCIATE EDITORS

ANTHONY ALBERG, Charleston, SC
ALBERTO ASCHERIO, Boston, MA
BRAD ASTOR, Baltimore, MD
DONNA DAY BAIRD, Research Triangle
Park, NC
OLGA BASSO, Research Triangle Park, NC
EDWARD J. BOYKO, Seattle, WA
ROBERT F. BREIMAN, Atlanta, GA
GERMAINE M. BUCK, Rockville, MD
KENNETH P. CANTOR, Bethesda, MD
WONG-HO CHOW, Bethesda, MD
JOHN D. CLEMENS, Bethesda, MD
STEPHEN COLE, Chapel Hill, NC
GLINDA S. COOPER, Research Triangle Park, NC
JOSEF CORESH, Baltimore, MD
STEVEN S. COUGHLIN, Washington, DC
ROSA M. CRUM, Baltimore, MD
RALPH B. D’AGOSTINO, JR., Winston-Salem, NC
RALPH B. D’AGOSTINO, SR., Boston, MA
CONSTANTINE DASKALAKIS, Philadelphia, PA
ANA V. DIEZ-ROUX, Ann Arbor, MI
BRENDA ESKENAZI, Berkeley, CA
GUY D. ESLICK, Boston, MA
EDUARDO FAERSTEIN, Rio de Janeiro, Brazil
KATHERINE M. FLEGAL, Hyattsville, MD
BETSY FOXMAN, Ann Arbor, MI
LIN FRITSCHI, Crawley, Australia
SANDRO GALEA, Ann Arbor, MI
MARILIE D. GAMMON, Chapel Hill, NC
PETER J. GERGEN, Rockville, MD
RICHARD F. GILLUM, Hyattsville, MD
EDWARD GIOVANNUCCI, Boston, MA
MARLENE B. GOLDMAN, Lebanon, NH
ELISEO GUALLAR, Baltimore, MD
M. ELIZABETH HALLORAN, Seattle, WA
BERNARD L. HARLOW, Minneapolis, MN
PATRICIA HARTGE, Rockville, MD
MAUREEN C. HATCH, Bethesda, MD
RICHARD B. HAYES, New York, NY
MIGUEL A. HERNÁN, Boston, MA
LISA HERRINTON, Oakland, CA
IRVA HERTZ-PICCIOTTO, Davis, CA
DONALD HOOVER, Piscataway, NJ
JANE A. HOPPIN, Research Triangle Park, NC
DAVID J. HUNTER, Boston, MA
JOUNI JAAKKOLA, Birmingham, United Kingdom
STEVEN J. JACOBSEN, Pasadena, CA
KONRAD JAMROZIK, Nedlands, Australia
STANISLAV V. KASL, New Haven, CT
FREYA KAMEL, Research Triangle Park, NC
JAY S. KAUFMAN, Montreal, Canada
WILLIAM C. KNOWLER, Phoenix, AZ
ALAN R. KRISTAL, Seattle, WA
MICHAEL F. LEITZMANN, Regensburg, Germany
DE-KUN LI, Oakland, CA
RONGLING LI, Memphis, TN
MARC LIPSITCH, Boston, MA
JULIAN LITTLE, Aberdeen, United Kingdom
STEPHANIE LONDON, Research Triangle
Park, NC
MATTHEW P. LONGNECKER, Research Triangle
Park, NC
GERALD MCGWIN, JR., Birmingham, AL
JOSEPH K. MCLAUGHLIN, Rockville, MD
ROBERT C. MILLIKAN, Chapel Hill, NC
LORENE M. NELSON, Stanford, CA
ROBERTA B. NESS, Houston, TX
CRAIG J. NEWSCHAFFER, Philadelphia, PA
THOMAS R. O’BRIEN, Bethesda, MD
ANDREW OLSHAN, Chapel Hill, NC
NANCY PADIAN, San Francisco, CA
JULIE R. PALMER, Brookline, MA
JULIE PARSONNET, Stanford, CA
WENDY S. POST, Baltimore, MD
NANCY POTISCHMAN, Bethesda, MD
ERIC B. RIMM, Boston, MA
HARVEY A. RISCH, New Haven, CT
LESLIE ROBISON, Minneapolis, MN
KARIN ROSENBLATT, Champaign, IL
AUDREY F. SAFTLAS, Iowa City, IA
GLEN A. SATTEN, Atlanta, GA
ENRIQUE F. SCHISTERMAN, Bethesda, MD
THOMAS A. SELLERS, Tampa, FL
EYAL SHAHAR, Tucson, AZ
A. RICHEY SHARRETT, Baltimore, MD
GARY SHAW, Oakland, CA
COLIN L. SOSKOLNE, Edmonton, Canada
MEIR STAMPFER, Boston, MA
RACHAEL STOLZENBERG-SOLOMON,
Rockville, MD
SHOLOM WACHOLDER, Bethesda, MD
DOUGLAS L. WEED, Bethesda, MD
SCOTT T. WEISS, Boston, MA
PHYLLIS A. WINGO, Atlanta, GA
JUN ZHANG, Bethesda, MD
WEI ZHENG, Nashville, TN
 American Journal of Epidemiology
The American Journal of Epidemiology (ISSN 0002-9262) is published
twice a month, 24 times a year in 2 volumes (12 issues per volume), by
Oxford University Press, 2001 Evans Road, Cary, NC 27513-2009, for the
Johns Hopkins Bloomberg School of Public Health and is sponsored by the
Society for Epidemiologic Research. Information for authors appears on the
Journal’s home page at www.aje.oxfordjournals.org. Authors should consult
these instructions before submitting manuscripts.
Oxford University Press is a department of the University of Oxford. It
furthers the University’s objective of excellence in research, scholarship,
and education by publishing worldwide.
Subscriptions: A subscription to the American Journal of Epidemiology
comprises 24 issues. Prices include postage; for subscribers outside the
Americas, issues are sent air freight.
Annual subscription rate (Volumes 169 and 170, 24 issues, 2009):
Institutional: Print and online $762/£508/€762, Online only $724/£483/
€724, and Print only $724/£483/€724; Personal: (Combination: the Journal
and Epidemiologic Reviews) Online only $355/£237/€355, and Print only
$373/£249/€373). Please note: UK£ rate applies to UK and Rest of World,
except US ($) and Europe (€). There may be other subscription rates available.
For a complete listing please see www.oxfordjournals.org/our_journals/aje/
access_purchase/price_list.html.
Full pre-payment in the correct currency is required for all orders.
Payment should be in US dollars for orders being delivered to the USA or
Canada; Euros for orders being delivered within Europe (excluding the
UK); GBP sterling for orders being delivered elsewhere (i.e. not being
delivered to USA, Canada, or Europe). All orders should be accompanied
by full payment and sent to your nearest Oxford Journals office.
Subscriptions are accepted for complete volumes only. Orders are regarded
as firm, and payments are not refundable. Our prices include Standard Air
as postage outside of the UK. Claims must be notified within four months
of despatch/order date (whichever is later). Subscriptions in the EEC may
be subject to European VAT. If registered, please supply details to avoid
unnecessary charges. For subscriptions that include online versions, a pro-
portion of the subscription price may be subject to UK VAT. Subscribers in
Canada, please add GST to the prices quoted.
Personal rate subscriptions are only available if payment is made by per-
sonal cheque or credit card, delivery is to a private address, and is for personal
use only.
Methods of payment: (i) Check (payable to Oxford University Press,
mailed to Oxford University Press, Cashiers Office, Great Clarendon
Street, Oxford OX2 6DP, UK) in GB£ Sterling (drawn on a UK bank), US$
Dollars (drawn on a US bank), or EU€ Euros. (ii) Bank transfer to Barclays
Bank Plc, Oxford Group Office, Oxford (bank sort code 20-65-18) (UK),
overseas only Swift code BARC GB 22 (GB£ Sterling to account no.
70299332, IBAN GB89BARC20651870299332; US$ Dollars to account
no. 66014600, IBAN GB27BARC20651866014600; EU€ Euros to account
no. 78923655, IBAN GB16BARC20651878923655). (iii) Credit card
(Mastercard, Visa, Switch or American Express).
Back issues: The current year and two previous years’ issues are available
from Oxford University Press. Previous volumes can be obtained online
at http://www.periodicals.com/oxford.html or from the Periodicals Service
Company, 11 Main Street, Germantown, NY 12526, USA. E-mail:
psc@periodicals.com. Tel: (518) 537-4700. Fax: (518) 537-5899.
For further information, please contact: Journals Customer Service,
Oxford University Press, Great Clarendon Street, Oxford OX2 6DP, UK.
Telephone (& answer-phone outside normal working hours): +44 (0) 1865
353907, Fax: +44 (0) 1865 353485, E-mail: jnls.cust.serv@oxfordjournals.org.
In the US, please contact: Journals Customer Service, Oxford University
Press, 2001 Evans Road, Cary, NC 27513, USA, Telephone (& answer-
phone outside normal working hours): 800-852-7323 (toll-free in USA/
Canada), fax: 919-677-1714, e-mail: jnlorders@oxfordjournals.org. In Japan,
please contact: Journals Customer Services, Oxford University Press,
4-5-10-8F Shiba, Minato-ku, Tokyo, 108-8386, Japan. Telephone: +81 3
5444 5858, Fax: +81 3 3454 2929, e-mail: custserv.jp@oxfordjournals.org.
Advertising: Advertising, inserts, and artwork enquiries should be
addressed to Advertising and Special Sales, Oxford Journals, Oxford
University Press, Great Clarendon Street, Oxford OX2 6DP, UK.
Telephone: +44 (0) 1865 353329. Fax: +44 (0) 1865 353774. E-mail:
jnlsadvertising@oxfordjournals.org.
Digital Object Identifiers: For information on dois and how to resolve
them, please visit www.doi.org.
Oxford Journals Environmental and Ethical Policies: Oxford
Journals is committed to working with the global community to bring the
highest quality research to the widest possible audience. Oxford Journals
will protect the environment by implementing environmentally friendly
policies and practices wherever possible. Please see http://www.
oxfordjournals.org/ethicalpolicies.html for further information on Oxford
Journals’ environmental and ethical policies.
Requests for permissions, reprints, and photocopies: All rights
reserved; no part of this publication may be reproduced, stored in a retrieval
system, or transmitted in any form or by any means (electronic, mechanical,
photocopying, recording, or otherwise) without either prior written
permission of the publisher (Oxford University Press, Rights and Business
Development Journals Division, Great Clarendon St., Oxford OX2 6DP,
UK. Tel. +44 1865 354490/353695; Fax: +44 1865 353485; E-mail:
journals.permissions@oxfordjournals.org; Web: www.oxfordjournals.org/
jnls/permissions) or a license permitting restricted copying, issued in the
US by the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA
01923 (Fax: 978-750-4470) or in the UK by the Copyright Licensing
Agency Ltd., 90 Tottenham Court Road, London W1P 9HE, UK. Reprints
of individual articles are available only from the authors.
It is a condition of publication in the Journal that authors grant an
exclusive license to the Johns Hopkins Bloomberg School of Public Health.
However, requests for permission to reprint material found in the Journal
should be submitted to Oxford University Press. This ensures that requests
from third parties to reproduce articles are handled efficiently and consis-
tently and will also allow the article to be disseminated as widely as possi-
ble. As part of the license agreement, authors may use their own material in
other publications provided that the Journal is acknowledged as the origi-
nal place of publication and Oxford University Press is notified in writing
in advance.
The author(s) of each article appearing in this Journal is/are solely
responsible for the content thereof; the publication of an article shall not
constitute or be deemed to constitute any representation by the Editors, the
Johns Hopkins Bloomberg School of Public Health, or the Society for
Epidemiologic Research that the data presented therein are correct or suffi-
cient to support the conclusions reached or that the experiment design or
methodology is adequate.
The American Journal of Epidemiology is printed on acid-free paper
that meets the minimum requirements of ANSI Standard Z39.48-1984
(Permanence of Paper).
The Journal is indexed or abstracted in Index Medicus, the Science
Citation Index, Current Contents: Life Sciences, and the BIOSIS database.
Copyright © 2009 by the Johns Hopkins Bloomberg School of Public
Health.
The American Journal of Epidemiology (ISSN 0002-9262) is published
semimonthly in Jan., Feb., Mar., Apr., May, Jun., Jul., Aug., Sept., Oct., Nov.,
and Dec. by Oxford University Press, 2001 Evans Road, Cary, NC 27513-
2009. Periodical Postage paid at Cary, NC and at additional post offices.
Postmaster: Send address changes to: American Journal of Epidemiology,
Journals Customer Service Department, Oxford University Press, 2001
Evans Road, Cary, NC 27513-2009.
 AMERICAN JOURNAL OF EPIDEMIOLOGY
Associate Director for Epidemiology and Genetics Research Program
Division of Cancer Control and Population Sciences
The National Cancer Institute (NCI), a major research component of the National Institutes of Health (NIH)
within the Department of Health and Human Services (DHHS), seeks a senior scientist to serve as an As-
sociate Director in its Division of Cancer Control and Population Sciences (DCCPS). The Division provides
national scientific leadership and oversight of NCI-funded research in the areas of cancer epidemiology, sur-
veillance, health services, survivorship, and behavioral science. DCCPS also is committed to addressing health
disparities through transdisciplinary research.
This challenging and highly visible position requires broad scientific expertise, a passion for public service,
commitment to collaboration, and an ability to develop effective strategies to identify gaps in research and
overcome barriers to scientific progress. Exploiting scientific opportunities requires visionary leadership and
sound scientific judgment to ensure the greatest payoff from NCI’s investments. Therefore, a broad perspective
on cancer epidemiology and other population sciences that informs development of creative and cost-effective
strategies to advance science is essential.
The Associate Director will lead the Epidemiology and Genetics Research Program (EGRP), which includes its
Office of the Associate Director and four branches: Clinical and Translational Epidemiology, Host Susceptibility
Factors, Methods and Technologies, and Modifiable Risk Factors. He/she provides scientific and administrative
leadership for the Program, supervises the staff, and represents NCI to a wide variety of professional, academic,
and advocacy organizations. The Associate Director also develops and facilitates collaborations with funders of
other types of population science, including the NIH Institutes and Centers, Centers for Disease Control and
Prevention (CDC), and many non-governmental organizations. EGRP’s grants, contracts, interagency agree-
ments, and operating budgets totaled more than $197 million in Fiscal Year 2008. Included were more than 370
research grants and numerous interagency agreements and consortia.
Volume 169
Qualifications: The successful applicant will be an experienced epidemiologist (M.D. or Ph.D.-level training re-
quired) with leadership experience, excellent communication skills, and a strong record of peer-reviewed publica-
tions. Strong leadership skills, an ability to work effectively across disciplinary boundaries, and a commitment to the
highest standards of scientific integrity and quality are required. Experience in managing complex research projects,
Number 8
scientific staff, training programs, interdisciplinary collaborations, and/or funded programs is highly valued.
Salary: This position is an excepted service position (Title 42) with a salary range of $160,000-$195,000.
How to Apply: Applications will be considered until the position is filled. Please submit a letter of interest and
April 15, 2009
CV to the Search Committee Chair:
Rachel Ballard-Barbash, M.D., M.P.H., Associate Director
Applied Research Program
Division of Cancer Control and Population Sciences
National Cancer Institute
Pages 919–1042
6130 Executive Blvd, Room 4005, MSC 7344
Bethesda, MD 20892-7344
Express Mail: Rockville, MD 20852
For more information about DCCPS and EGRP, see http://cancercontrol.cancer.gov
THE DHHS/NIH/NCI ARE EQUAL OPPORTUNITY EMPLOYERS
oxford
issn 0002-9262 (print)
ISSN 1476-6256 (online)
printed in the u.s.a.
American Journal of
Epidemiology
Volume 169 Number 8 April 15, 2009
www.aje.oxfordjournals.org
ORIGINAL CONTRIBUTIONS
919 Parkinson’s Disease and Residential Exposure to Maneb and Paraquat From
Agricultural Applications in the Central Valley of California. Sadie Costello, Myles Cockburn,
Jeff Bronstein, Xinbo Zhang, and Beate Ritz
927 Overweight and Obesity Over the Adult Life Course and Incident Mobility Limitation in
Older Adults: The Health, Aging and Body Composition Study. Denise K. Houston, Jingzhong
Ding, Barbara J. Nicklas, Tamara B. Harris, Jung Sun Lee, Michael C. Nevitt, Susan M. Rubin, Frances
A. Tylavsky, and Stephen B. Kritchevsky for the Health ABC Study
937 Association of Diabetes With Prostate Cancer Risk in the Multiethnic Cohort. Kevin
M. Waters, Brian E. Henderson, Daniel O. Stram, Peggy Wan, Laurence N. Kolonel, and Christopher
A. Haiman
946 Modification of the Effect of Vitamin E Supplementation on the Mortality of Male
Smokers by Age and Dietary Vitamin C. Harri Hemilä and Jaakko Kaprio
954 Dietary Acrylamide Intake and Risk of Premenopausal Breast Cancer. Kathryn M. Wilson,
Lorelei A. Mucci, Eunyoung Cho, David J. Hunter, Wendy Y. Chen, and Walter C. Willett
962 Alcohol Intake and Cigarette Smoking and Risk of a Contralateral Breast Cancer: The
Women’s Environmental Cancer and Radiation Epidemiology Study. Julia A. Knight,
Leslie Bernstein, Joan Largent, Marinela Capanu, Colin B Begg, Lene Mellemkjær, Charles F. Lynch,
Kathleen E. Malone, Anne S. Reiner, Xiaolin Liang, Robert W. Haile, John D. Boice, Jr., WECARE Study
Collaborative Group, and Jonine L. Bernstein.
969 Positive Associations Between Ionizing Radiation and Lymphoma Mortality Among
Men. David B. Richardson, Hiromi Sugiyama, Steve Wing, Ritsu Sakata, Eric Grant, Yukiko Shimizu,
Nobuo Nishi, Susan Geyer, Midori Soda, Akihiko Suyama, Fumiyoshi Kasagi, and Kazunori Kodama
977 Biomarker-calibrated Energy and Protein Consumption and Increased Cancer Risk
Among Postmenopausal Women. Ross L. Prentice, Pamela A. Shaw, Sheila A. Bingham, Shirley
A. A. Beresford, Bette Caan, Marian L. Neuhouser, Ruth E. Patterson, Marcia L. Stefanick, Suzanne
Satterfield, Cynthia A. Thomson, Linda Snetselaar, Asha Thomas, and Lesley F. Tinker
Continued on Inside Front Cover
Published for the Johns Hopkins Bloomberg School of Public Health by Oxford
University Press Sponsored by the Society for Epidemiologic Research
Founded in 1920 by W. H. Welch and W. H. Howell as the American Journal of
Hygiene at the Johns Hopkins School of Hygiene and Public Health
Contents Continued from Front Cover
Central Arkansas Veterans Healthcare System (CAVHS)
990 Sex-Modified Effect of Hepatitis B Virus Infection on Mortality From Primary Liver Cancer. Na Wang,
Yingjie Zheng, Xinsen Yu, Wenyao Lin, Yue Chen, and Qingwu Jiang
Health Services Researcher
996 Serum Selenium and Peripheral Arterial Disease: Results From the National Health and Nutrition Full-time position
Examination Survey, 2003–2004. Joachim Bleys, Ana Navas-Acien, Martin Laclaustra, Roberto Pastor-Barriuso,
Andy Menke, Jose Ordovas, Saverio Stranges, and Eliseo Guallar
The Geriatric Research Education and Clinical Center (GRECC) at the Central Arkansas
1004 Ambient Air Pollution and Cardiovascular Malformations in Atlanta, Georgia, 1986–2003. Matthew J. Strickland, Veterans Healthcare System (CAVHS) is seeking a full-time health services researcher.
Mitchel Klein, Adolfo Correa, Mark D. Reller, William T. Mahle, Tiffany J. Riehle-Colarusso, Lorenzo D. Botto, W. Dana
Faculty appointment in the Departments of Geriatrics and/or Health Policy and
Flanders, James A. Mulholland, Csaba Siffel, Michele Marcus, and Paige E. Tolbert
Management, University of Arkansas for Medical Sciences at the Assistant/Associate
1015 Maternal Urinary Metabolites of Di-(2-Ethylhexyl) Phthalate in Relation to the Timing of Labor in a Professor level commensurate with experience. Position includes 90% protected time for
US Multicenter Pregnancy Cohort Study. Jennifer J. Adibi, Russ Hauser, Paige L. Williams, Robin M. Whyatt,
research. The incumbent must be an American citizen, hold a doctoral degree (PhD or
Antonia M. Calafat, Heather Nelson, Robert Herrick, and Shanna H. Swan
equivalent), and have: (1) completed an approved residency or postdoctoral training
1025 Longitudinal Trends in Hazardous Alcohol Consumption Among Women With Human Immunodeficiency program in the area of health services research, or (2) have extensive training, research
Virus Infection, 1995–2006. Robert L. Cook, Fang Zhu, Bea Herbeck Belnap, Kathleen Weber, Judith A. Cook,
David Vlahov, Tracey E. Wilson, Nancy A. Hessol, Michael Plankey, Andrea A. Howard, Stephen R. Cole, Gerald B. Sharp,
and dissemination experiences in the area of health services research as documented by
Jean L. Richardson, and Mardge H. Cohen previous research investigations, senior mentor evaluations, presentations at national
meetings, and peer-reviewed publications. The ideal candidate will have expertise and
1033 Suicide Mortality Among Patients Receiving Care in the Veterans Health Administration Health System.
John F. McCarthy, Marcia Valenstein, H. Myra Kim, Mark Ilgen, Kara Zivin, and Frederic C. Blow
interest in the areas of nutrition, exercise, or functional outcomes of older adults.

Women and minorities are encouraged to apply. A nationally competitive salary is

BOOK REVIEWS
available and competitive recruitment packages are possible for superior candidates.
1039 Hyping Health Risks: Environmental Hazards in Daily Life and the Science of Epidemiology By Geoffrey
C. Kabat. David A. Savitz For information and application procedures please contact Brandy Fulmer, Human
Resources Management Specialist, at Brandy.Fulmer@va.gov.
1041 Concepts of Epidemiology: Integrating the Ideas, Theories, Principles and Methods of Epidemiology,
2nd Edition By Raj Bhopal. Jonathan M. Samet

Instructions to Authors can be found on the following website: http://aje.oxfordjournals.org/.

Stay alert to forthcoming contents

of OUP Journals... for free!
We will automatically email you the forthcoming
tables of contents for new issues of journals
published by Oxford University Press.
To register for this free service, go to
www.oupjournals.org/tocmail, select
the journal(s) you are interested in,
enter your email address, and press
‘submit’.

w w w . o u p j o u r n a l s . o r g

1
 American Journal of Epidemiology Vol. 169, No. 8
ª The Author 2009. Published by the Johns Hopkins Bloomberg School of Public Health. DOI: 10.1093/aje/kwp006
All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org. Advance Access publication March 6, 2009

Original Contribution

Parkinson’s Disease and Residential Exposure to Maneb and Paraquat From

Agricultural Applications in the Central Valley of California

Sadie Costello, Myles Cockburn, Jeff Bronstein, Xinbo Zhang, and Beate Ritz

Initially submitted September 12, 2008; accepted for publication January 6, 2009.

Evidence from animal and cell models suggests that pesticides cause a neurodegenerative process leading to
Parkinson’s disease (PD). Human data are insufficient to support this claim for any specific pesticide, largely
because of challenges in exposure assessment. The authors developed and validated an exposure assessment
tool based on geographic information systems that integrated information from California Pesticide Use Reports
and land-use maps to estimate historical exposure to agricultural pesticides in the residential environment. In
1998–2007, the authors enrolled 368 incident PD cases and 341 population controls from the Central Valley of
California in a case-control study. They generated estimates for maneb and paraquat exposures incurred between
1974 and 1999. Exposure to both pesticides within 500 m of the home increased PD risk by 75% (95% confidence
interval (CI): 1.13, 2.73). Persons aged
60 years at the time of diagnosis were at much higher risk when exposed
to either maneb or paraquat alone (odds ratio ¼ 2.27, 95% CI: 0.91, 5.70) or to both pesticides in combination
(odds ratio ¼ 4.17, 95% CI: 1.15, 15.16) in 1974–1989. This study provides evidence that exposure to a combi-
nation of maneb and paraquat increases PD risk, particularly in younger subjects and/or when exposure occurs at
younger ages.

case-control studies; fungicides, industrial; geographic information systems; herbicides; maneb; paraquat;
Parkinson disease; pesticides

Abbreviations: CI, confidence interval; DDE, dichlorodiphenyldichloroethylene; GIS, geographic information system; MPPþ, toxic
metabolite of 1-methyl-4-phenylpyridinium; OR, odds ratio; PD, Parkinson’s disease; PLSS, Public Land Survey System; PUR,
Pesticide Use Reporting.

Parkinson’s disease (PD) has been reported to occur at
high rates among farmers and in rural populations, contrib-
uting to the hypothesis that agricultural pesticides might be
causal agents (1–4). Animal studies have linked certain pes-
ticides to Parkinsonism and dopaminergic cell death. The
pesticide rotenone can produce the behavioral and neuro-
pathologic features of PD in some rodent models through
chronic systemic inhibition of mitochondrial complex I (5,
6). Exposure to a combination of the fungicide maneb and
the herbicide paraquat in mice leads to increased substantia
nigra neuronal pathology (7), age-dependent motor degen-
eration, progressive reductions in dopamine metabolites and
turnover (8), and reduced tyrosine hydroxylase and dopa-
mine transporter immunoreactivity (9, 10).
Human evidence is insufficient to identify any particular
pesticide compound, including those implicated by animal
studies, as being responsible for causing PD (11). Method-
ological limitations have clouded the interpretation of most
epidemiologic studies exploring pesticide exposures and PD
in humans. Past studies have generally relied on self-reports
and recall of chemical usage, making them vulnerable to
information bias and differential recall bias (12).
Because pesticides applied from the air or ground may
drift from their intended treatment sites, with measurable
concentrations subsequently detected in the air, in plants,
and in animals up to several hundred meters from applica-
tion sites (13–15), accurate methods of estimating environ-
mental exposures in rural communities are sorely needed.

Correspondence to Dr. Sadie Costello, Department of Environmental Health Sciences, School of Public Health, University of California, Berkeley,
50 University Hall, #7360, Berkeley, CA 94720-7360 (e-mail: sadie@berkeley.edu).

919 Am J Epidemiol 2009;169:919–926

920 Costello et al.
Geographic information system (GIS)-based methods of as-
sessing exposure to pesticides have become popular in re-
cent years and may prove an effective solution when
pesticide data exist. We developed and employed a validated
GIS-based exposure assessment tool to estimate pesticide
exposure from applications to agricultural crops, relying
on California Pesticide Use Reporting (PUR) data, land-
use maps, and geocoded residential historical locations
(16). We investigated whether exposure to the pesticides
maneb and paraquat, alone and in combination, increased
the risk of incident PD among residents of the Central Valley
of California, an area well-known for its intensive agricul-
ture and potential for pesticide exposure.
MATERIALS AND METHODS
All procedures described have been approved by the
University of California, Los Angeles, institutional review
board for human subjects, and informed consent was ob-
tained from all participants.
Subject recruitment
We used a population-based approach for recruiting cases
and controls from a largely agricultural population in Cali-
fornia. Details are provided elsewhere (17). Briefly, persons
with PD newly diagnosed between January 1998 and January
2007 who resided in 1 of 3 central California counties
(Fresno, Tulare, or Kern county) and had lived in California
for at least 5 years prior to diagnosis were recruited into our
study within 3 years of diagnosis. Altogether, 28 (90%) of
the 31 practicing local neurologists who provided care for
PD patients assisted in recruiting cases for this study. We
solicited collaboration from Kaiser Permanente Medical
Center (Fresno, California), Kern Medical Center (Bakersfield,
California), and Visalia Medical Clinic (Visalia, California)
and from the Veterans Administration, PD support groups,
local newspapers, and local radio stations that broadcast
public service announcements.
Of the 1,167 PD cases who were initially invited, 604
were not eligible: For 397, the case’s diagnosis date fell
outside the 3-year range prior to contact, 51 denied having
received a PD diagnosis, 134 lived outside the tricounty
area, and 22 were too ill to participate. Of the 563 eligible
cases, 473 (84%) were examined by a University of Cali-
fornia, Los Angeles, movement disorder specialist at least
once and were confirmed to have clinically ‘‘probable’’ or
‘‘possible’’ PD; the remaining 90 potential cases could not
be examined or interviewed (54% withdrew, 32% were too
ill or died, and 14% moved out of the area prior to the
examination or did not honor a scheduled appointment).
We examined but excluded another 96 patients because they
had other causes of Parkinsonism. This left us with 377
cases; of these, 368 provided all information needed for
analyses.
Controls aged 65 years or older were identified from
Medicare lists in 2001, but because of implementation of
the Health Insurance Portability and Accountability Act,
which prohibits the use of Medicare enrollees, 70% of our
controls were recruited from randomly selected tax assessor
residential units (parcels) in each of the 3 counties. We
mailed letters of invitation to a random selection of residen-
tial living units and also attempted to identify head-of-
household names and telephone numbers for these parcels,
using the services of marketing companies and Internet
searches.
We contacted 1,212 potential population controls by mail
and/or telephone for eligibility screening. Eligibility criteria
were: 1) not having PD, 2) being at least 35 years of age,
3) currently residing primarily in 1 of the 3 designated coun-
ties, and 4) having lived in California for at least 5 years
prior to the screening. Only 1 person per household was
allowed to enroll. Of the potential controls contacted, 457
were ineligible: 409 were too young, 44 were terminally ill,
and 4 resided primarily outside of the study area. Of the 755
eligible controls, 409 (54%) declined participation, were too
ill to honor an appointment, or moved out of the area prior to
interview; 346 (46%) were enrolled, and 341 provided all
information needed for analyses.
Assessment of environmental pesticide exposure
We conducted telephone interviews to obtain demo-
graphic and exposure information. Detailed residential his-
tory forms were mailed to subjects in advance of their
interview and were reviewed in person or over the phone.
We estimated pesticide exposures in the residential environ-
ment from applications to agricultural crops employing
a validated GIS-based system, which combined PUR data
and land-use maps (16, 18), to produce estimates of residen-
tial ambient pesticide applications within a set distance of
subjects’ homes. We recorded and geocoded lifetime resi-
dential histories and estimated ambient exposures for all
historical addresses at which participants had resided be-
tween 1974 and 1999, the period covered by the PUR data.
A technical discussion of our GIS-based approach is pro-
vided elsewhere (16); here we briefly summarize the data
sources and the exposure modeling process.
Residential addresses. Addresses were automatically
geocoded to TigerLine files (NAVTEQ (Chicago, Illinois),
unpublished data, 2006), and discrepancies were then man-
ually resolved in a multistep process similar to that de-
scribed by McElroy et al. (19). Resulting locations were
recorded, along with the relevant year range of residence,
so they could be matched to the appropriate year-specific
PUR and land-use data (below). For our GIS model, we
relied on addresses in Fresno, Kern, and Tulare counties
(the tricounty area) at which participants had resided be-
tween 1974 and 1999. Out of 9,568 total residential years
contributed by cases (26 years 3 368 cases), 7,593 years
(79%) were spent at addresses within the tricounty area as
compared with 6,757 (76%) of 8,866 years contributed by
controls (26 years 3 341 controls). We geocoded these
tricounty residential addresses for the period 1974–1999
with similar precision for cases and controls; that is, both
had spent 88% of their respective residential years at ad-
dresses we considered to have been mapped with high pre-
cision (i.e., at the level of a residential parcel, street
address, or street intersection rather than a zip code or city
centroid).
Am J Epidemiol 2009;169:919–926
 PD and Residential Maneb and Paraquat Exposure 921

Pesticide use reporting. PUR data are recorded by the to dopaminergic cell damage or possibly PD (organochlo-
California Department of Pesticide Regulation for any com- rines, organophosphates, and dithiocarbamates (23) and pro-
mercial application of restricted-use pesticides (defined as teasome inhibitors (24)).
agents with harmful environmental or toxicologic effects We considered the following demographic variables as
(20)) and, since 1990, for all commercial uses of pesticides potential confounders in all analyses: age (age at diagnosis
regardless of toxicologic profile. The location of each PUR for cases and age at interview for controls), sex, race (white,
record is referenced to the Public Land Survey System nonwhite), education (<12 years, 12 years, >12 years), and
(PLSS), a nationwide grid that parcels land into sections cigarette smoking (current, former, never). We used SAS 9.1
at varying resolutions. Each PUR record includes the name (SAS Institute Inc., Cary, North Carolina) to perform un-
of the pesticide’s active ingredient, the poundage applied, conditional logistic regression analyses.
the crop and acreage of the field, the application method,
and the date of application.
Land-use maps. Because the PUR records link an agri- RESULTS
cultural pesticide application only to a whole PLSS grid
Study participants were predominantly Caucasian, over
section, we added information from land-use maps to more
the age of 60, and without a family history of PD (Table 1).
precisely locate the pesticide application, as described in
Cases were slightly older than controls, were more often
detail elsewhere (18). The California Department of Water
male, and had completed fewer years of education. They
Resources periodically (every 7–10 years) performs county-
were also more likely to have been occupationally exposed
wide large-scale surveys of land use and crop cover, which
to pesticides and to be never or former smokers.
allowed us to identify the locations of specific crops within
We did not find increased risks of PD among subjects
each PLSS grid section. Digital maps from more recent
exposed to paraquat alone during the years 1974–1999
(1996–1999) surveys are available (21), and paper maps
(Table 2). While the rarity of sole maneb exposure (4 subjects)
were manually digitized for earlier periods (1977–1995).
precluded any meaningful interpretation of the maneb-only
The 1977 land-use survey was conducted closest in time
results, combined exposure to both maneb and paraquat in-
to 1974, when PUR data became available. We constructed
creased the risk of PD by 75% (odds ratio (OR) ¼ 1.75,
historical electronic maps of land use and crop type, and
95% confidence interval (CI): 1.13, 2.73), an effect estimate
using the PLSS grid section and the crop type reported in the
which was essentially unchanged after adjustment for
PUR record, we allocated pesticide applications to an agri-
occupational pesticide exposure (OR ¼ 1.74, 95% CI:
cultural site to which we assigned a GIS-based location.
1.11, 2.72).
Deriving estimates of residential pesticide exposure. The
When we examined 2 separate exposure time windows,
time-specific total exposure at each location, by pesticide,
the years 1974–1989 and 1990–1999, the risk increase ob-
was derived through summation of exposures over a fixed
served for the whole period was found to be mainly attribut-
500-m radius (suggested in previous literature (13, 15, 19))
able to exposures incurred during the earlier window
around the home for the relevant years of residence. The
(OR ¼ 2.14, 95% CI: 1.24, 3.68), while being exposed
numbers of pounds of pesticide applied annually per acre
during the later window did not seem to increase PD risk
were summed for each residential buffer and weighted by
(Table 2). Furthermore, for younger (
60 years) subjects,
the proportion of treated acreage in each buffer, resulting in
exposure to both maneb and paraquat in both windows in-
pesticide application rates that could be averaged over spe-
creased PD risk as much as 4- to 6-fold (Table 3). Exposure
cific calendar periods of each subject’s lifetime.
to either maneb or paraquat alone during 1974–1989 also
increased risk of PD in younger subjects (OR ¼ 2.27, 95%
Statistical analysis CI: 0.91, 5.70). When we examined exposure windows
among our older subjects (>60 years), combined exposure
We estimated residential exposures to maneb and para-
to both pesticides in the earlier window only (1974–1989)
quat, alone and in combination, for the following time win-
was also associated with a 2-fold increase in PD risk
dows: 1) 1974–1999, 2) 1974–1989, and 3) 1990–1999, to
(OR ¼ 2.15, 95% CI: 1.15, 4.02), but no increase was found
assess the possibility of an extensive induction period prior
for either the later window (1990–1999) or the combined
to PD onset and the influence of age at exposure. We strat-
exposure periods (Table 3). Stratification by sex suggested
ified models by sex and age (
60 years, >60 years) and, in
no differences in estimates between males and females.
additional sensitivity analyses, controlled for exposure to
some groups of pesticides suspected to increase PD risk.
We controlled for occupational exposure to pesticides DISCUSSION
among subjects who had held jobs in the agricultural sector,
assigning them to categories of ‘‘likely exposed to pesti- In this population-based case-control study, agricultural
cides’’ when they reported pesticide handling and applica- application of both maneb and paraquat within 500 m of
tions or fieldwork and ‘‘possibly exposed to pesticides’’ a residence during the period 1974–1999 greatly increased
when they reported managerial, produce processing, and the risk of developing PD, especially when exposure oc-
other nonfield farm work; all other subjects were considered curred between 1974 and 1989 or when PD was diagnosed
‘‘not occupationally exposed to pesticides’’ (22). In some at a younger age (
60 years). Exposure to both pesticides
models, we also adjusted for residential exposures to groups during the earlier time window (1974–1989) also doubled
of other pesticides that some studies have found to be linked the risk for older cases. Associations were particularly

Am J Epidemiol 2009;169:919–926
922 Costello et al.

Table 1. Odds Ratio for Parkinson’s Disease According to Various Sociodemographic

Characteristics, Central Valley of California, 1998–2008

Cases Controls 95%

(n 5 368) (n 5 341) Odds
Variable Confidence
Ratio
No. or Mean % No. or Mean % Interval

Mean age, 68.1 (34–88) 67.6 (34–92) 1.00 0.99, 1.02

years (range)
Age group, years

40 7 2 6 2
41–50 25 7 26 8
51–60 47 13 55 16
61–70 111 30 95 28
71–80 145 39 121 35
>80 33 9 38 11
Female sex 161 44 165 48 0.83 0.62, 1.12
First-degree relative with 55 15 37 11 1.44 0.93, 2.25
Parkinson’s disease
Race
White 296 80 279 82 1 Reference
Nonwhitea 72 20 62 18 1.09 0.75, 1.60
Asian 4 1 8 2
Black 3 1 13 4
Latino 49 13 31 9
Native American 16 4 10 3
Education, years
<12 68 18 38 11 1.15 0.69, 1.90
12 100 27 64 19 1 Reference
>12 200 54 239 70 0.54 0.37, 0.77
Job exposure matrix
Not occupationally 232 63 240 70 1 Reference
exposed to pesticides
Possibly occupationally 26 7 26 8 1.03 0.58, 1.83
exposed to pesticides
Likely occupationally 110 30 75 22 1.52 1.08, 2.14
exposed to pesticides
Cigarette smoking status
Never smoker 195 53 146 43 1 Reference
Former smoker 151 41 161 47 0.70 0.52, 0.96
Current smoker 22 6 34 10 0.48 0.27, 0.86
Pack-years of
cigarette smoking
0 195 53 146 43 1 Reference
>0–
19 96 26 89 26 0.81 0.56, 1.16
>19 77 21 106 31 0.54 0.38, 0.78
a
The odds ratio was calculated for all nonwhites versus whites.

strong for younger-onset patients (
60 years), who would
have been children, teenagers, and young adults during the
exposure period: Among those exposed in the earlier time
window, risk was increased more than 4-fold with exposure
to both pesticides and more than 2-fold with exposure to just
1 of the pesticides. Consistent with some theories regarding
the progression of PD pathology (25), these data suggest
that the critical window of exposure to toxicants may be
years before the onset of motor symptoms which lead to
diagnosis.
Pesticide and herbicide exposures have previously been
implicated in idiopathic PD. Paraquat is structurally similar
to the toxic metabolite (MPPþ) of the 1-methyl-4-phenyl-
pyridinium ion (a metabolite of 1-methyl-4-phenyl-1,2,3,6-
tetrahydropyridine), an agent known to induce Parkinsonian
symptoms in humans that has been widely used to study

Am J Epidemiol 2009;169:919–926
 PD and Residential Maneb and Paraquat Exposure 923

Table 2. Odds Ratio for Parkinson’s Disease According to Table 3. Odds Ratio for Parkinson’s Disease According to
Residential Ambient Exposure to Maneb and/or Paraquat, Central Residential Ambient Exposure to Maneb and/or Paraquat, by Time
Valley of California, 1974–1999 Window of Exposure and Age Group, Central Valley of California,
1974–1999
Cases Controls 95%
Time Window (n 5 368) (n 5 341) Odds Cases Controls 95%
Confidence
and Exposure Ratioa Age Group Odds
Confidence
No. % No. % Interval and Exposure Ratioa
No. % No. % Interval
1974–1999
1974–1999 Time Window
Missing data 13 4 13 4

60 years
No exposure 115 31 126 37 1 Reference
Missing data 2 3 4 5
Paraquat only 149 40 152 45 1.01 0.71, 1.43
No exposure 18 23 34 39 1 Reference
Maneb only 3 1 1 0 3.04 0.30, 30.86
Paraquat or 38 48 42 48 1.77 0.84, 3.75
Both paraquat 88 24 49 14 1.75 1.13, 2.73 maneb only
and maneb
Both paraquat 21 27 7 8 5.07 1.75, 14.71
1974–1989 and maneb
Missing data 53 14 52 15 >60 years
No exposure 93 25 113 33 1 Reference Missing data 11 4 9 4
Paraquat or 148 40 137 40 1.25 0.85, 1.85 No exposure 97 34 92 36 1 Reference
maneb only
Paraquat or 114 39 111 44 0.90 0.60, 1.34
Both paraquat 74 20 39 11 2.14 1.24, 3.68 maneb only
and maneb
Both paraquat 67 23 42 17 1.36 0.83, 2.23
1990–1999 and maneb
Missing data 15 4 15 4 1974–1989 Time Window
No exposure 215 58 213 62 1 Reference
60 years
Paraquat or 113 31 95 28 0.96 0.64, 1.43 Missing data 16 20 20 23
maneb only
No exposure 13 16 27 31 1 Reference
Both paraquat 25 7 18 5 0.93 0.45, 1.94
and maneb Paraquat or 36 46 34 39 2.27 0.91, 5.70
maneb only
a
Odds ratios were adjusted for age, sex, nonwhite race, education, Both paraquat 14 18 6 7 4.17 1.15, 15.16
and smoking status. Results were mutually adjusted for exposure in and maneb
each time window. >60 years
Missing data 37 13 32 13
Parkinsonism in animal models (26). MPPþ is believed to
cause cell death by interfering with mitochondrial respira- No exposure 80 28 86 34 1 Reference
tion (27), because it concentrates in mitochondria and in- Paraquat or 112 39 103 41 1.18 0.75, 1.84
hibits complex I of the electron transport chain (28). Many maneb only
lines of evidence point to possible mitochondrial dysfunc- Both paraquat 60 21 33 13 2.15 1.15, 4.02
and maneb
tion in PD. Several genes have been identified in familial
forms of PD that are linked to mitochondrial function 1990–1999 Time Window
(PINK1 and DJ1), and in sporadic cases of PD, pathologic
60 years
free radical reactions that damage mitochondria and de- Missing data 2 3 5 6
crease electron transport activity have been described (29). No exposure 43 54 58 67 1 Reference
Impaired electron transport hampers adenosine triphosphate
Paraquat or 27 34 22 25 2.00 0.84, 4.74
production and leads to the diversion of electrons from their maneb only
normal electron transport recipients and, thus, further for- Both paraquat 7 9 2 2 5.74 0.55, 59.62
mation of damaging free radicals (29). and maneb
Although paraquat is also used to induce Parkinsonism in >60 years
some animal models, the mechanism by which it produces
Missing data 13 4 10 4
symptoms is not yet understood (30). Recent mammalian
and yeast-cell experiments suggest that mitochondria take No exposure 172 60 155 61 1 Reference
up paraquat actively across their membranes, where com- Paraquat or 86 30 73 29 0.78 0.49, 1.24
maneb only
plex I reduces it to the paraquat radical cation that subse-
quently produces mitochondria-damaging superoxide (31). Both paraquat 18 6 16 6 0.66 0.29, 1.50
and maneb
It has also been suggested that maneb may inhibit the ubiq-
a
uitin proteasome system, thereby damaging the dopaminer- Age-stratified models with adjustment for sex, nonwhite race,
gic neuron (24, 32). Additionally, maneb has been linked to education, and smoking status. Results were mutually adjusted for
Parkinsonism in mice also exposed to paraquat. In 3 recent exposure in each time window.
studies, investigators reported that only when mice were
exposed to a combination of the fungicide maneb and the

Am J Epidemiol 2009;169:919–926
924 Costello et al.
herbicide paraquat (paraquat þ maneb), not to either pesti-
cide alone, did they exhibit increased neuronal pathology
(7), age-dependent motor degeneration and progressive re-
ductions in dopamine metabolites and dopamine turnover
(8), and reduced tyrosine hydroxylase and dopamine trans-
porter immunoreactivity (9).
The fungicide maneb and the herbicide paraquat are both
used in the Central Valley of California and are often used
on the same crops, including potatoes, dry beans, and toma-
toes. The average amount of maneb applied near the homes
of these study subjects was relatively stable throughout both
time windows; however, annual paraquat exposure in-
creased during the later (1990–1999) time window. Persons
living near fields sprayed with maneb and paraquat may also
be exposed to a host of other agricultural chemicals. When
we controlled for the influence of other groups of pesticides
suspected a priori to be risk factors for PD in our study, the
odds ratios for combined maneb and paraquat exposure and
PD in the younger subjects were still in the 3- to 6-fold range
and statistically significant; however, our precision de-
creased, probably because of correlated exposures. Correla-
tion between pesticides is an inherent problem when
assessing the effects of human exposure. However, since
adjustment for other pesticides did not remove the associa-
tion for maneb and paraquat, our data provide compelling
evidence that these 2 pesticides may in fact affect PD risk in
humans, as has been suggested by animal experiments.
Paraquat and maneb are applied by ground, aerial, and
backpack methods; however, paraquat has a much longer
field half-life of 1,000 days (33), as compared with only
12–36 days for maneb (34). Both chemicals bind strongly
to soil, though, and are not thought to be a threat to ground-
water (35, 36). Such strong binding could result in contam-
inated soil getting blown or tracked into homes by wind,
pets, and shoes, thereby increasing exposure for persons
who live closer to agricultural application sites (3, 37, 38).
In a previous validation study, our prediction model for a
serum measure of dichlorodiphenyldichloroethylene (DDE)
explained 47% of the biomarker’s variance (39). Addition-
ally, our GIS-derived measure of organochlorine exposure
identified persons with high serum DDE levels reasonably
well (specificity of 87%) (39).
Although our GIS model allowed us to calculate the num-
ber of pounds of each active ingredient applied per acre
within a 500-m buffer, these quantities are not comparable
across pesticides. That is, a pound of active ingredient does
not represent the same human neurotoxicity across pesti-
cides, and no information currently exists that would allow
us to standardize these measures. Thus, while we believe
that our model provided us with an accurate indicator of any
pesticide exposure from applications close to a residence,
our exposure measure cannot be considered quantitative
beyond a crude rank ordering of low/medium likelihood
of exposure and high likelihood of exposure. Since we hy-
pothesized that coexposure to 2 pesticides, maneb and para-
quat, would increase the risk of PD, we also lacked the
statistical power to perform extensive categorical analyses
(note that only 3 cases and 1 control were exposed solely to
maneb). We conducted additional analyses after dichoto-
mizing pounds per acre at their median and mean levels
and found that exposure to both pesticides at the highest
level was associated with PD, especially in persons aged

60 years; however, wide confidence intervals surrounding
our point estimates rendered these results generally uninfor-
mative (results not shown).
In only 1 previous analysis, conducted within the Agri-
cultural Health Study cohort (40), did researchers assess the
effects of maneb and paraquat exposures. Statistical power
was limited by the small number (n ¼ 78) of incident cases
identified during follow-up and the very small number (n ¼
4–10) of cases exposed to maneb/mancozeb (OR ¼ 2.1) and
paraquat (OR ¼ 1.4). In a small Taiwanese study, the only
case-control study to date with sufficient statistical power to
examine exposure to the herbicide paraquat, Liou et al. (41)
reported a 4- to 6-fold increase in PD risk among long-term
applicators. In a case-control study from the Mayo Clinic
(Rochester, Minnesota), Brighina et al. (42) presented asso-
ciations between self-reported pesticide exposure and PD in
subjects younger than 60 years only (for all pesticides,
OR ¼ 1.80, 95% CI: 1.12, 2.87; for herbicides, OR ¼ 2.46,
95% CI: 1.34, 4.52).
Our exposure estimates did not depend on the subject’s
recall of pesticide exposure and are therefore unlikely to
have been biased by differential exposure misclassification.
Since all of our PD diagnoses were clinically confirmed, we
expect disease misclassification to have been minimal. Non-
differential exposure misclassification is a possibility in our
study and may have attenuated our effect estimates.
Our results may be biased if cases and controls selected
themselves into our study according to their potential for
pesticide exposure, but our subjects were not asked to self-
report environmental exposures and probably were unaware
of their true historical exposures. There is no reason to
suspect that cases and controls would have chosen to par-
ticipate on the basis of their historical residence near certain
agricultural plots. We saw no difference in estimated effects
when we restricted analyses to only those subjects with
more (12 years) or less (<12 years) education. Similarly,
we saw no difference in our results when we restricted the
sample to persons whose addresses had been mapped with
high precision in the tricounty area during the period 1974–
1999 (363 cases, 336 controls).
Our analysis has confirmed 2 previous observations from
animal studies: 1) exposure to multiple chemicals may po-
tentiate the effect of each chemical (of interest, since hu-
mans are often exposed to more than 1 pesticide in the
environment) and 2) the timing of exposure is important.
To our knowledge, this is the first epidemiologic study to
provide strong evidence that 2 specific pesticides, suggested
by animal research as potentially acting synergistically to
become neurotoxic, strongly increase the risk of PD in
humans, especially given combined exposure and when
encountered earlier in life.
ACKNOWLEDGMENTS
Author affiliations: Department of Environmental Health
Sciences, School of Public Health, University of California,
Am J Epidemiol 2009;169:919–926

Berkeley, Berkeley, California (Sadie Costello); Department
of Preventive Medicine, Keck School of Medicine, Univer-
sity of Southern California, Los Angeles, California (Myles
Cockburn, Xinbo Zhang); Department of Geography, College
of Letters, Arts and Sciences, University of Southern
California, Los Angeles, California (Myles Cockburn,
Xinbo Zhang); Department of Neurology, School of Medicine,
University of California, Los Angeles, Los Angeles, Cali-
fornia (Jeff Bronstein); and Department of Epidemiology,
School of Public Health, University of California, Los
Angeles, Los Angeles, California (Beate Ritz).
This work was supported by the National Institute
of Environmental Health Sciences (grants ES10544,
U54ES12078, and 5P30 ES07048), the National Institute of
Neurological Disorders and Stroke (grant NS 038367), and
the Department of Defense Prostate Cancer Research Pro-
gram (grant 051037). In addition, initial pilot funding was
provided by the American Parkinson’s Disease Association.
The authors thank the participating neurologists and med-
ical centers in Fresno, Kern, and Tulare counties for their
support.
Conflict of interest: none declared.
REFERENCES
1. Ben-Shlomo Y, Finnan F, Allwright S, et al. The epidemiology
of Parkinson’s disease in the Republic of Ireland: observations
from routine data sources. Ir Med J. 1993;86(6):190–191, 194.
2. Burguera JA, Catala J, Taberner P, et al. Mortality from
Parkinson’s disease in Spain (1980–1985). Distribution by age,
sex and geographic areas. Neurologia. 1992;7(3):89–93.
3. Morano A, Jimenez-Jimenez FJ, Molina JA, et al. Risk-factors
for Parkinson’s disease: case-control study in the province of
Cáceres, Spain. Acta Neurol Scand. 1994;89(3):164–170.
4. Svenson LW, Platt GH, Woodhead SE. Geographic variations
in the prevalence rates of Parkinson’s disease in Alberta. Can J
Neurol Sci. 1993;20(4):307–311.
5. Betarbet R, Sherer TB, MacKenzie G, et al. Chronic systemic
pesticide exposure reproduces features of Parkinson’s disease.
Nat Neurosci. 2000;3(12):1301–1306.
6. Sherer TB, Betarbet R, Greenamyre JT. Pesticides and
Parkinson’s disease. Scientific World Journal. 2001;1:207–208.
7. Norris EH, Uryu K, Leight S, et al. Pesticide exposure exac-
erbates alpha-synucleinopathy in an A53T transgenic mouse
model. Am J Pathol. 2007;170(2):658–666.
8. Thiruchelvam M, McCormack A, Richfield EK, et al. Age-
related irreversible progressive nigrostriatal dopaminergic
neurotoxicity in the paraquat and maneb model of the
Parkinson’s disease phenotype. Eur J Neurosci. 2003;18(3):
589–600.
9. Thiruchelvam M, Richfield EK, Baggs RB, et al. The nigro-
striatal dopaminergic system as a preferential target of repeated
exposures to combined paraquat and maneb: implications for
Parkinson’s disease. J Neurosci. 2000;20(24):9207–9214.
10. Thiruchelvam M, Richfield EK, Goodman BM, et al. Devel-
opmental exposure to the pesticides paraquat and maneb and
the Parkinson’s disease phenotype. Neurotoxicology. 2002;
23(4-5):621–633.
11. Brown TP, Rumsby PC, Capleton AC, et al. Pesticides and
Parkinson’s disease—is there a link? Environ Health Perspect.
2006;114(2):156–164.
Am J Epidemiol 2009;169:919–926
PD and Residential Maneb and Paraquat Exposure 925
12. Seidler A, Hellenbrand W, Robra BP, et al. Possible environ-
mental, occupational, and other etiologic factors for
Parkinson’s disease: a case-control study in Germany.
Neurology. 1996;46(5):1275–1284.
13. Chester G, Ward RJ. Occupational exposure and drift hazard
during aerial application of paraquat to cotton. Arch Environ
Contam Toxicol. 1984;13(5):551–563.
14. Currier WW, MacCollom GB, Baumann GL. Drift residues
of air-applied carbaryl in an orchard environment. J Econ
Entomol. 1982;75(6):1062–1068.
15. MacCollom GB, Currier WW, Baumann GL. Drift compari-
sons between aerial and ground orchard application. J Econ
Entomol. 1986;79(2):459–464.
16. Goldberg DW, Wilson JP, Knoblock CA, et al. An effective
and efficient approach for manually improving geocoded data.
Int J Health Geogr. 2008;7:60.
17. Kang GA, Bronstein JM, Masterman DL, et al. Clinical
characteristics in early Parkinson’s disease in a central
California population-based study. Mov Disord. 2005;20(9):
1133–1142.
18. Rull RP, Ritz B. Historical pesticide exposure in California
using pesticide use reports and land-use surveys: an assess-
ment of misclassification error and bias. Environ Health
Perspect. 2003;111(12):1582–1589.
19. McElroy JA, Remington PL, Trentham-Dietz A, et al. Geo-
coding addresses from a large population-based study: lessons
learned. Epidemiology. 2003;14(4):399–407.
20. United States Senate Committee on Agriculture, Nutrition
and Forestry. Federal Insecticide, Fungicide, and
Rotenticide Act [As Amended Through P.L. 110–246,
Effective May 22, 2008]. Section 3(d)(1)(C).
Washington, DC: US Senate, 2008:30. (http://agriculture.
senate.gov/Legislation/Compilations/Fifra/FIFRA.pdf).
(Accessed February 8, 2009).
21. California Department of Water Resources. California Land
and Water Use. Sacramento, CA: California Department of
Water Resources; 2009. (http://www.landwateruse.water.ca.
gov/). (Accessed February 8, 2009).
22. Young HA, Mills PK, Riordan D, et al. Use of a crop and job
specific exposure matrix for estimating cumulative exposure to
triazine herbicides among females in a case-control study in
the Central Valley of California. Occup Environ Med. 2004;
61(11):945–951.
23. Elbaz A, Tranchant C. Epidemiologic studies of environmental
exposures in Parkinson’s disease. J Neurol Sci. 2007;262(1-2):
37–44.
24. Wang XF, Li S, Chou AP, et al. Inhibitory effects of pesticides
on proteasome activity: implication in Parkinson’s disease.
Neurobiol Dis. 2006;23(1):198–205.
25. Braak H, Del Tredici K, Rüb U, et al. Staging of brain pa-
thology related to sporadic Parkinson’s disease. Neurobiol
Aging. 2003;24(2):197–211.
26. Langston JW, Ballard P, Tetrud JW, et al. Chronic Parkinson-
ism in humans due to a product of meperidine-analog syn-
thesis. Science. 1983;219(4587):979–980.
27. Sayre LM, Wang F, Hoppel CL. Tetraphenylborate potentiates
the respiratory inhibition by the dopaminergic neurotoxin
MPPþ in both electron transport particles and intact mito-
chondria. Biochem Biophys Res Commun. 1989;161(2):
809–818.
28. Singer TP, Ramsay RR. Mechanism of the neurotoxicity of
MPTP. An update. FEBS Lett. 1990;274(1-2):1–8.
29. Cassarino DS, Bennett JP Jr. An evaluation of the role of
mitochondria in neurodegenerative diseases: mitochondrial
mutations and oxidative pathology, protective nuclear
926 Costello et al.
responses, and cell death in neurodegeneration. Brain Res
Brain Res Rev. 1999;29(1):1–25.
30. Dinis-Oliveira RJ, Remião F, Carmo H, et al. Paraquat
exposure as an etiological factor of Parkinson’s disease.
Neurotoxicology. 2006;27(6):1110–1122.
31. Cocheme HM, Murphy MP. Complex I is the major site of
mitochondrial superoxide production by paraquat. J Biol
Chem. 2008;283(4):1786–1798.
32. Zhou Y, Shie FS, Piccardo P, et al. Proteasomal inhibition
induced by manganese ethylene-bis-dithiocarbamate: rele-
vance to Parkinson’s disease. Neuroscience. 2004;128(2):
281–291.
33. Extension Toxicology Network. EXTOXNET: Extension Tox-
icology Network. Pesticide Information Profiles. Paraquat.
Corvallis, OR: Oregon State University; 1996. (http://extoxnet.
orst.edu/pips/paraquat.htm). (Accessed June 5, 2008).
34. Extension Toxicology Network. EXTOXNET: Extension
Toxicology Network. Pesticide Information Profiles. Maneb.
Corvallis, OR: Oregon State University; 1996. (http://extoxnet.
orst.edu/pips/maneb.htm). (Accessed June 5, 2008).
35. Environmental Protection Agency. R.E.D. [Reregistration
Eligibility Decision] Facts. Paraquat Dichloride. Washington,
DC: Environmental Protection Agency; 1997. (http://www.
epa.gov/oppsrrd1/REDs/factsheets/0262fact.pdf). (Accessed
June 5, 2008).
36. Environmental Protection Agency. Reregistration Eligibility
Decision (RED) for Maneb. Washington, DC: Environmental
Protection Agency; 2005. (http://www.epa.gov/oppsrrd1/
REDs/maneb_red.pdf). (Accessed June 5, 2008).
37. Jiménez-Jiménez FJ, Mateo D, Giménez-Roldán S. Exposure
to well water and pesticides in Parkinson’s disease: a case-
control study in the Madrid area. Mov Disord. 1992;7(2):
149–152.
38. Rybicki BA, Johnson CC, Uman J, et al. Parkinson’s disease
mortality and the industrial use of heavy metals in Michigan.
Mov Disord. 1993;8(1):87–92.
39. Ritz B, Costello S. Geographic model and biomarker-derived
measures of pesticide exposure and Parkinson’s disease. Ann N
Y Acad Sci. 2006;1076:378–387.
40. Kamel F, Tanner CM, Umbach DM, et al. Pesticide exposure
and self-reported Parkinson’s disease in the Agricultural
Health Study. Am J Epidemiol. 2007;165(4):364–374.
41. Liou HH, Tsai MC, Chen CJ, et al. Environmental risk factors
and Parkinson’s disease: a case-control study in Taiwan.
Neurology 1997;48(6):1583–1588.
42. Brighina L, Frigerio R, Schneider NK, et al. Alpha-synuclein,
pesticides, and Parkinson disease: a case-control study. Neu-
rology. 2008;70(16):1461–1469.
Am J Epidemiol 2009;169:919–926
 American Journal of Epidemiology Vol. 169, No. 8
ª The Author 2009. Published by the Johns Hopkins Bloomberg School of Public Health. DOI: 10.1093/aje/kwp007
All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org. Advance Access publication March 6, 2009

Original Contribution

Overweight and Obesity Over the Adult Life Course and Incident Mobility
Limitation in Older Adults
The Health, Aging and Body Composition Study

Denise K. Houston, Jingzhong Ding, Barbara J. Nicklas, Tamara B. Harris, Jung Sun Lee, Michael
C. Nevitt, Susan M. Rubin, Frances A. Tylavsky, and Stephen B. Kritchevsky for the Health ABC
Study

Initially submitted September 29, 2008; accepted for publication January 6, 2009.

Obesity in middle and old age predicts mobility limitation; however, the cumulative effect of overweight and/or
obesity over the adult life course is unknown. The association between overweight and/or obesity in young, middle,
and late adulthood and its cumulative effect on incident mobility limitation was examined among community-
dwelling US adults aged 70–79 years at baseline (1997–1998) in the Health, Aging and Body Composition Study
(n ¼ 2,845). Body mass index was calculated by using recalled weight at ages 25 and 50 years and measured
weight at ages 70–79 years. Mobility limitation (difficulty walking 1/4 mile (0.4 km) or climbing 10 steps) was
assessed semiannually over 7 years of follow-up and was reported by 43.0% of men and 53.7% of women.
Men and women who were overweight or obese at all 3 time points had an increased risk of mobility limitation
(hazard ratio ¼ 1.61, 95% confidence interval: 1.25, 2.06 and hazard ratio ¼ 2.85, 95% confidence interval: 2.15,
3.78, respectively) compared with those who were normal weight throughout. Furthermore, there was a significant
graded response (P < 0.0001) on risk of mobility limitation for the cumulative effect of obesity in men and over-
weight and/or obesity in women. Onset of overweight and obesity in earlier life contributes to an increased risk of
mobility limitation in old age.

aged; mobility limitation; obesity; overweight

Abbreviations: BMI, body mass index; Health ABC, Health, Aging and Body Composition.

The proportion of older adults in the United States is
expected to grow to 20% of the population by the year
2030 (1). In addition, adults are, on average, increasingly
heavier than earlier generations, with approximately one-
third classified as obese (body mass index (BMI)
30 kg/m2)
(2). Obesity has been consistently associated with increased
risk of cardiovascular disease, diabetes, and several cancers,
as well as other chronic conditions (3, 4). Obesity in middle-
aged and older adults also increases the risk of physical
disability (5–10), possibly as a result of these obesity-
related chronic conditions or by other mechanisms. Thus,
the growing prevalence of obesity, particularly among youn-
ger age groups, could reverse the recent declines in disabil-
ity rates among future generations of older adults (11–13).
Although the association between obesity in middle and
late adulthood and physical disability is established (5–8,
10), few studies have examined the effect of obesity among
younger adults (14, 15) or the cumulative effect of obesity
from young adulthood (9, 16) on disability in older adults.
In the National Health and Nutrition Examination Survey
Epidemiologic Follow-up Study, being obese or becoming
obese over 20 years of follow-up was associated with higher
levels of disability (9). Among Finnish adults aged 55 years
or older, earlier onset of obesity and obesity duration based

Correspondence to Dr. Denise K. Houston, Sticht Center on Aging, Department of Internal Medicine, Section on Gerontology and Geriatric
Medicine, Wake Forest University School of Medicine, Medical Center Boulevard, Winston Salem, NC 27157-1207 (e-mail: dhouston@
wfubmc.edu).

927 Am J Epidemiol 2009;169:927–936

928 Houston et al.
on recalled weights at ages 20, 30, 40, and 50 years in-
creased the likelihood of reporting walking limitations
(16). However, neither of these studies included a measure
of incident disability.
In the Health, Aging and Body Composition (Health
ABC) Study, we previously showed that, compared with
those of normal weight, men and women who reported be-
ing overweight or obese at ages 25, 50, and 70–79 years had
significantly worse physical performance at study baseline
(ages 70–79 years) based on objective measures (17), pre-
dictive of subsequent mobility disability (18). The primary
objective of the present study was to examine the associa-
tion of overweight and/or obesity in young, middle, and late
adulthood, as well as the cumulative effect of overweight
and/or obesity across all 3 time points, with incident mobil-
ity limitation in men and women in their 70s and 80s.
MATERIALS AND METHODS
Study population
Data for this analysis were derived from the Health ABC
Study, a prospective cohort study investigating the associa-
tions among body composition, weight-related health con-
ditions, and incident functional limitations in older adults.
The Health ABC Study enrolled 3,075 community-dwelling
male and female blacks and whites aged 70–79 years be-
tween April 1997 and June 1998. Participants were recruited
from a random sample of white and all black Medicare-
eligible residents in the Pittsburgh, Pennsylvania, and
Memphis, Tennessee, metropolitan areas. Participants were
eligible if they 1) reported no difficulty walking one-fourth
of a mile (0.4 km), climbing up 10 steps, or performing
activities of daily living; 2) were free of life-threatening
illness; 3) planned to remain in the geographic area for at
least 3 years; and 4) were not enrolled in lifestyle interven-
tion trials. All participants provided written informed con-
sent, and all protocols were approved by the institutional
review boards at both study sites.
Participants who were missing data on recalled weight at
age 25 years (n ¼ 75) or age 50 years (n ¼ 68) or data on
recalled height at age 25 years (n ¼ 39) were excluded.
Participants with an extremely low body weight (BMI
<15 kg/m2) at ages 25, 50, or 70–79 years (n ¼ 15) or an
absolute weight change of more than 100 pounds (45.4 kg)
from age 50 to ages 70–79 years (n ¼ 8) were also excluded.
Participants who were missing information on mobility lim-
itation during follow-up (n ¼ 2) as well as other pertinent
baseline covariates (n ¼ 23) were also excluded. The final
analysis sample was 2,845 participants.
Mobility limitation
Occurrence of mobility limitation during follow-up was
assessed during annual clinic visits alternating with tele-
phone interviews every 6 months. Persistent mobility limi-
tation was defined as 2 consecutive reports of having any
difficulty walking one-fourth of a mile (0.4 km) or climbing
up 10 steps without resting because of health or a physical
problem. Follow-up included incident cases ascertained
through 6.9 years of follow-up, with a mean follow-up of
4.2 (standard deviation, 2.4) years.
BMI history
At the baseline visit, participants were asked to recall
their usual body weight (women were instructed to answer
for a time when not pregnant) and height (without shoes) at
age 25 years and to recall their usual body weight at age 50
years. Body weight at study baseline (ages 70–79 years) was
measured in kilograms by using a standard balance beam
scale, and height was measured in millimeters by using
a Harpenden stadiometer (Holtain Ltd., Crosswell, United
Kingdom). BMI at ages 25 and 50 years was calculated by
using recalled weight at age 25 or 50 years and recalled
height at age 25 years, while BMI at ages 70–79 years
was calculated by using measured weight and height. BMI
was categorized as normal weight (<25.0 kg/m2), over-
weight (25.0–29.9 kg/m2), and obese (
30.0 kg/m2). Addi-
tional analyses that excluded underweight (BMI <18.5 kg/m2)
participants at ages 25, 50, and 70–79 years were similar
to those that included them with the normal-weight group
(data not shown).
To assess the cumulative effect of overweight and/or
obesity across all 3 time points, the following categories
were created: normal weight/nonobese at ages 25, 50, and
70–79 years; overweight and/or obese at ages 70–79 years
but not at ages 25 or 50 years; overweight and/or obese at
ages 70–79 and 50 years but not at age 25 years; over-
weight and/or obese at all 3 time points; and overweight
and/or obese at age 50 years but not at ages 70–79 years.
Other BMI history patterns were categorized as other. For
these analyses, recalled height at age 25 years was used to
calculate BMI at ages 70–79 years in addition to BMI at
ages 25 and 50 years to minimize possible misclassifica-
tion bias due to systematic differences between recalled
height at age 25 years and measured height at study base-
line. Results were similar when measured height at study
baseline was used to calculate BMI at ages 25, 50, and
70–79 years (data not shown).
Potential confounders
Demographic characteristics (age, gender, race, and
education), smoking status, alcohol consumption, and
physical activity were ascertained by an interviewer-
administered questionnaire at study baseline. The time
and intensity of self-reported physical activities performed
in the past 7 days, including walking for exercise, other
walking, climbing stairs, aerobic dance, weight training,
and other high- and medium-intensity activities, were
summed (kcal/week). Because BMI history may affect mo-
bility as a consequence of weight-related health condi-
tions, prevalent health conditions at study baseline were
examined as potential mediators of the associations. The
prevalence of diabetes, coronary heart disease, congestive
heart failure, stroke, chronic obstructive pulmonary dis-
ease, and knee pain were determined by using algorithms
based on self-report and medication use at study baseline;
Am J Epidemiol 2009;169:927–936

participants with definitive health conditions were coded as
‘‘yes.’’ The 20-item Center for Epidemiologic Studies
Depression Scale was used as an indicator of depressed
mood, and persons scoring 16 or higher were classified
as depressed (19). The Modified Mini-Mental State Exam-
ination was used as an indicator of general cognitive status,
with a minimum score of zero and maximum score of
100 (best) (20). Because mean scores on the Modified
Mini-Mental State Examination vary by education, a cut-
point of <75 for individuals with less than a high school
education and a cutpoint of <80 for individuals with a high
school education were used to classify participants as cog-
nitively impaired.
Statistical analyses
Cox proportional hazards regression models were used to
examine the associations between BMI status at ages 25, 50,
and 70–79 years and risk of incident mobility limitation
with SAS version 9.1 software (SAS Institute, Inc., Cary,
North Carolina). The cumulative effect of overweight and/or
obesity across all 3 time points and incident mobility limi-
tation was also examined. Participants who survived with no
evidence of incident mobility limitation were censored at
their next-to-last 6-month contact. Participants who died
and had no evidence of incident mobility limitation were
censored at their time of death, and those who were lost to
follow-up were censored at their last visit. Two-way inter-
actions between gender and BMI status at ages 25, 50, and
70–79 years were tested, and interactions were found at the
significance level of a ¼ 0.10. Interactions between race and
BMI status at ages 25, 50, and 70–79 years within gender
group were also tested but were not significant. Thus, in this
paper, all analyses are presented for men and women sepa-
rately, with race groups combined. Models were adjusted for
age, race, field center, education, smoking, alcohol con-
sumption, and physical activity at study baseline. Additional
models were also adjusted for prevalent health conditions at
study baseline. Proportional hazards assumptions were as-
sessed by examining log(log S(t)) plots as well as testing
interactions of each variable with time in the model. The
proportional hazards assumptions were not violated. All
P values were 2 sided.
RESULTS
The mean age of the study population was 73.6 years,
50.6% were women, and 39.7% were black. Participants
excluded from the present analysis (n ¼ 230, 7.5%) were
more likely to be female, black, and older; have less than
a high school education; have a higher BMI at ages 70–79
years; and report incident mobility limitation during follow-
up (P < 0.01). The descriptive characteristics of the study
population by gender and incident mobility limitation are
shown in Table 1. Approximately 43.0% of men and 53.7%
of women reported becoming limited in mobility over the
7 years of follow-up. Participants who reported incident
mobility limitation were more likely to be black, have less
than a high school education, report being a current smoker
Am J Epidemiol 2009;169:927–936
Weight History and Mobility Limitation 929
and a former drinker, engage in less physical activity, and
report prevalent chronic conditions than those who did not
report incident mobility limitation. Overweight and obesity
at ages 25, 50, and 70–79 years were more prevalent among
those who reported incident mobility limitation. Participants
who reported mobility limitation were less likely to be nor-
mal weight or nonobese at all 3 time points.
The hazard ratios and 95% confidence intervals of inci-
dent mobility limitation by BMI status at ages 25, 50, and
70–79 years for men and women are shown in Table 2. Men
who were obese at age 25 years and overweight or obese at
ages 50 and 70–79 years were at significantly increased risk
of incident mobility limitation compared with men who
were normal weight at ages 25, 50, and 70–79 years, re-
spectively. For women, those who were overweight at age
25 years and overweight or obese at ages 50 and 70–79 years
were at significantly increased risk of incident mobility lim-
itation compared with women who were normal weight at
each of the 3 time points. Adjusting for prevalent health
conditions at study baseline attenuated the associations be-
tween BMI status at ages 25, 50, and 70–79 years and in-
cident mobility limitation slightly, but, in general, the
associations remained significant.
Figures 1 and 2 show the hazard ratios and 95% confi-
dence intervals for incident mobility limitation associated
with overweight and/or obesity from age 25 to ages 70–79
years. Compared with that for those of normal weight at all
3 time points, the risk of incident mobility limitation ap-
peared to increase by duration of overweight or obesity
(BMI
25 kg/m2) for those who were overweight or obese
at ages 70–79 years (Figure 1). Although fewer participants
reported being obese at age 50 years or earlier, similar asso-
ciations were seen for duration of obesity (BMI
30 kg/m2),
particularly among men (Figure 2). After excluding partic-
ipants who were overweight and/or obese at age 50 years but
not at ages 70–79 years and those who had a BMI history
pattern of ‘‘other,’’ we found a significant graded response
for the cumulative effect of obesity in men (P for trend <
0.0001) and overweight and/or obesity in women (P for
trend < 0.0001) on risk of mobility limitation. Women
who reported being overweight and/or obese at age 50 years
but not at ages 70–79 years were at increased risk of incident
mobility limitation compared with those who were normal
weight or nonobese at all 3 time points. Men who reported
being obese at age 50 years but not at ages 70–79 years were
also at increased risk of incident mobility limitation. Further
adjustment for prevalent health conditions at study baseline
attenuated the associations slightly, but, in general, the as-
sociations remained significant.
The cumulative effect of overweight and/or obesity on
incident mobility limitation may partly be explained by at-
tained BMI at ages 70–79 years. Participants who reported
being overweight or obese at age 50 years or earlier had
higher BMIs in late adulthood compared with those not
classified as being overweight or obese until ages 70–79
years (30.8 kg/m2 vs. 28.0 kg/m2, P < 0.0001). Thus, we
examined the cumulative effect of overweight or obesity in
analyses limited to those who were overweight or obese at
ages 70–79 years, adjusting for age, race, education, field
center, smoking, alcohol consumption, and physical activity
930 Houston et al.

Table 1. Participant Characteristics at Study Baseline (Ages 70–79 Years) by Incident Mobility
Limitation in Men and Women, the Health, Aging and Body Composition Study, 1997–1998a

Men Women
No Mobility Mobility No Mobility Mobility
Limitation Limitation Limitation Limitation
(n 5 800) (n 5 604) (n 5 667) (n 5 774)

Age, yearsb 73.5 (2.9) 74.0 (2.9) 73.1 (2.8) 73.7 (2.9)
Black race 30.5 43.0 35.1 50.5
<High school education 22.1 32.0 13.9 27.3
Smoking status
Never 32.5 25.5 60.4 54.4
Current 8.8 13.2 8.2 10.2
Former 58.8 61.3 31.3 35.4
Alcohol consumption
Never 19.8 12.8 34.2 41.5
Former 21.4 31.3 14.2 23.0
Low (<1 drink/week) 19.5 19.7 24.3 19.8
Moderate (1–7 drinks/week) 26.2 26.2 22.6 12.9
Heavy (>7 drinks/week) 13.1 10.1 4.6 2.8
Physical activity
<500 kcal/week 35.6 50.7 50.8 68.0
500–<1,500 kcal/week 28.4 27.6 32.7 22.5

1,500 kcal/week 36.0 21.7 16.5 9.6
Prevalent chronic conditions
Diabetes 15.9 28.2 9.6 19.0
Stroke 4.5 9.9 5.6 9.2
Coronary heart disease 18.1 27.8 7.5 14.5
Congestive heart failure 2.4 6.3 0.6 3.2
COPD 8.0 17.2 6.4 16.0
Knee pain 9.1 20.9 8.7 25.8
Depression 1.9 6.1 2.6 8.0
Cognitive impairment 5.9 11.9 2.0 8.4
BMI at age 25 years
Normal weight 78.6 72.5 95.4 88.5
Overweight 19.9 24.5 3.3 8.7
Obese 1.5 3.0 1.4 2.8
Table continues

at study baseline. Among men, the hazard ratios of incident
mobility limitation were 1.22 (95% confidence interval:
0.96, 1.54) for those who were overweight or obese at ages
25, 50, and 70–79 years and 1.15 (95% confidence interval:
0.92, 1.44) for those who were overweight or obese at ages
50 and 70–79 years compared with those who were over-
weight or obese at ages 70–79 years but not at ages 50 or 25
years (P for trend ¼ 0.15). Among women, the hazard
ratios of incident mobility limitation were 1.71 (95% con-
fidence interval: 1.30, 2.23) for those who were overweight
or obese at all 3 time points and 1.23 (95% confidence
interval: 1.02, 1.47) for those who were overweight or
obese at ages 50 and 70–79 years compared with those
who were overweight or obese at ages 70–79 years but
not earlier (P for trend ¼ 0.0002). Further adjustment for
prevalent health conditions at study baseline attenuated the
associations slightly; however, the trend remained signifi-
cant for women.
DISCUSSION
For both men and women, overweight and/or obesity in
young, middle, and late adulthood were associated with an
increased risk of incident mobility limitation in late adult-
hood compared with normal weight. The risk of incident
mobility limitation was approximately 2.8-fold higher for
women and 1.6-fold higher for men who were overweight or
obese at ages 25, 50, and 70–79 years compared with being
normal weight at all 3 time points. For men, the risk of
incident mobility limitation was approximately 2.4-fold

Am J Epidemiol 2009;169:927–936
 Weight History and Mobility Limitation 931

Table 1. Continued

Men Women
No Mobility Mobility No Mobility Mobility
Limitation Limitation Limitation Limitation
(n 5 800) (n 5 604) (n 5 667) (n 5 774)

BMI at age 50 years

Normal weight 52.5 44.4 74.2 54.3
Overweight 41.9 43.4 21.3 31.8
Obese 5.6 12.2 4.5 14.0
BMI at ages 70–79 years
Normal weight 34.1 26.3 45.1 25.6
Overweight 49.1 46.7 38.5 35.8
Obese 16.8 27.0 16.3 38.6
Life-course overweight or obesity
Normal weight all 26.4 19.4 43.0 22.2
Overweight or obese at 24.2 23.0 29.8 31.3
ages 70–79 years
Overweight or obese at 23.5 27.5 21.6 32.7
ages 50 and 70–79 years
Overweight or obese 17.2 22.7 3.0 10.0
at ages 25, 50, and
70–79 years
Overweight or obese at 6.8 5.5 1.2 3.1
age 50 years but normal
weight at ages 70–79 years
Other 1.9 2.0 1.4 0.8
Life-course obesity
Nonobese all 80.5 67.6 82.6 58.0
Obese at ages 70–79 years 13.2 19.4 12.6 27.5
Obese at ages 50 and 3.1 6.1 2.7 9.3
70–79 years
Obese at ages 25, 50, 0.2 1.2 0.9 1.8
and 70–79 years
Obese at age 50 years but 2.2 5.0 0.9 2.8
nonobese at ages
70–79 years
Other 0.6 0.8 0.3 0.5

Abbreviations: BMI, body mass index; COPD, chronic obstructive pulmonary disease.
a
Except for age, all values are expressed as frequency (%).
b
Mean (standard deviation).

higher among those who were obese compared with non-
obese at all 3 time points. Although not significant, there
was a trend for women who were obese at all 3 time points
to have a 1.7-fold higher risk of mobility limitation com-
pared with those who were nonobese throughout. Further-
more, the risk of incident mobility limitation appeared to
increase with duration of overweight and/or obesity. Among
those who were normal weight at ages 70–79 years, having
a history of overweight and/or obesity in midlife was asso-
ciated with an increased risk of incident mobility limitation
compared with those who were normal weight or nonobese
at all 3 time points. Thus, being overweight and/or obese at
any time during adulthood increased the risk of mobility
limitation later in life, and, the longer the duration of over-
weight and/or obesity, the greater the risk. This finding ex-
tends prior work showing that overweight and obesity in
young and middle-aged adults predicts physical disability
in late life (14–16).
Overweight and obesity may lead to joint wear and tear,
reduced exercise capacity, and a higher rate of chronic dis-
ease such as cardiovascular disease, diabetes, and arthritis,
thus resulting in physical disability. The onset of overweight
or obesity in young and middle adulthood may result in
lower physical activity levels, contributing to decreased
muscle strength and cardiovascular fitness and greater de-
clines in physical function because of longer duration of
excess body weight and earlier onset of chronic disease.
In the National Health and Nutrition Examination Survey
Epidemiologic Follow-up Study, being obese or becoming
obese over 20 years of follow-up was associated with higher
levels of upper- and lower-body disability among adults
aged 25–74 years at baseline (9). Among Finnish adults

Am J Epidemiol 2009;169:927–936
932 Houston et al.

Table 2. Incident Mobility Limitation Among Men and Women by Body Mass Index at Ages 25, 50, and 70–79 Years, the Health, Aging and Body
Composition Study, 7 Years of Follow-upa

Men Women
Model 1b Model 2c Model 1b Model 2c
No. No.
HR 95% CI HR 95% CI HR 95% CI HR 95% CI

Age 25 years
Normal weight 1,067 1.00 1.00 1,321 1.00 1.00
Overweight 307 1.20 1.00, 1.45 1.09 0.90, 1.32 89 1.60 1.24, 2.06 1.70 1.31, 2.20
Obese 30 1.62 1.00, 2.60 1.47 0.91, 2.38 31 1.38 0.90, 2.12 1.18 0.76, 1.81
Age 50 years
Normal weight 688 1.00 1.00 915 1.00 1.00
Overweight 597 1.19 1.00, 1.41 1.16 0.98, 1.38 388 1.50 1.27, 1.76 1.40 1.18, 1.65
Obese 119 1.83 1.41, 2.37 1.49 1.13, 1.95 138 2.22 1.77, 2.80 1.76 1.39, 2.23
Ages 70–79 years
Normal weight 425 1.00 1.00 474 1.00 1.00
Overweight 675 1.23 1.01, 1.50 1.24 1.02, 1.52 534 1.50 1.24, 1.81 1.40 1.16, 1.70
Obese 297 1.77 1.41, 2.22 1.61 1.28, 2.02 408 2.46 2.03, 2.99 2.14 1.75, 2.62

Abbreviations: CI, confidence interval; HR, hazard ratio.

a
Separate Cox proportional hazards regression models for ages 25, 50, and 70–79 years.
b
Model 1 was adjusted for age, race, field center, education (<high school,
high school), smoking status (former, current, never), alcohol
consumption (former, never, <1 drink/week, 1–7 drinks/week, >7 drinks/week), and physical activity (<500 kcal/week, 500–<1,500 kcal/week,

1,500 kcal/week) at study baseline.
c
Model 2 was adjusted for model 1 variables plus prevalent diabetes, coronary heart disease, congestive heart failure, stroke, chronic
obstructive pulmonary disease, knee pain, depression, and cognitive impairment at study baseline.

aged 55 years or older, those who reported being obese at
age 30, 40, or 50 years had a 4-fold or higher increased
likelihood of walking limitations, and there was a significant
linear trend between duration of obesity and walking limi-
tations (16). However, reverse causality cannot be ruled out
because neither of these studies included a measure of in-
cident disability. In the Health ABC cohort, participants
who had a history of being overweight and/or obese at ages
25, 50, and/or 70–79 years had an increased risk of incident
mobility limitation compared with those who were normal
weight at all 3 time points. Adjustment for prevalent di-
abetes, cardiovascular disease, pulmonary disease, and
knee pain attenuated the associations slightly; however,
the associations between BMI history and incident mobil-
ity remained.
Whether overweight and/or obesity in young and middle
adulthood are independent of overweight and/or obesity in
late adulthood is difficult to untangle. In the Health ABC
cohort, recalled BMI at ages 25 and 50 years and measured
BMI at ages 70–79 years were moderately correlated (r ¼
0.33 and r ¼ 0.64, respectively). This finding raises the issue
of whether BMI in earlier life or attained BMI at ages 70–79
years was the driving force behind the associations seen
between recalled BMI at ages 25 and 50 years and mobility
limitation in late adulthood. Thus, caution must be exercised
when interpreting the results. Participants who reported be-
ing overweight or obese in midlife or earlier were heavier in
late adulthood than those who became overweight or obese
later in life. Nonetheless, there appeared to be a graded re-
sponse between age at onset of overweight and/or obesity
and increased risk of mobility limitation, particularly for
women. Furthermore, a history of overweight and/or obesity
in midlife or earlier was associated with increased risk of
incident mobility limitation among those who were normal
weight at ages 70–79 years.
A limitation of the work presented here is the use of
recalled weight and height from the distant past. However,
previous studies have shown high correlations (r
0.80)
between recalled and measured weight in young adulthood
among middle-aged and older men and women (21–23). In
the Health ABC cohort, self-reported and measured weight
at study baseline (ages 70–79 years) was highly correlated
(r ¼ 0.98). The correlation between recalled height at age
25 years and measured height at study baseline was also
high (r ¼ 0.93; mean difference, 3 cm). Other studies have
also shown high correlations between self-reported and
measured height among older adults (r
0.77), with a mean
difference of approximately 2–3 cm (24, 25). However, the
overreporting of height by older adults may be due, in part,
to loss of height associated with aging (26) and may more
accurately reflect height in younger adulthood. Results were
similar when measured height at study baseline was used in
the analysis in place of recalled height at age 25 years to
calculate BMI.
Important characteristics of the Health ABC cohort limit
generalization of these findings. First, participants were re-
cruited to be well functioning and free of mobility limitation
at study baseline. Thus, selection bias may have weakened
the association between BMI history and incident mobility
limitation because persons with a history of overweight and/
or obesity may have developed mobility limitation prior to
study entry and were excluded from the study. Another

Am J Epidemiol 2009;169:927–936
 Weight History and Mobility Limitation 933

Figure 1. Hazard ratios and 95% confidence intervals for incident mobility limitation among men (A) and women (B) by history of overweight or
obesity (body mass index
25 kg/m2), the Health, Aging and Body Composition Study, 7 years of follow-up. Models were adjusted for age, race,
field center, education (<high school,
high school), smoking status (former, current, never), alcohol consumption (former, never, <1 drink/week,
1–7 drinks/week, >7 drinks/week), and physical activity (<500 kcal/week, 500–<1,500 kcal/week,
1,500 kcal/week) at study baseline.

limitation is the use of self-reported mobility limitation as shown that self-reported mobility limitation is valid and
the primary endpoint. However, previous studies have has clinical significance (27). Furthermore, the use of

Am J Epidemiol 2009;169:927–936
934 Houston et al.

Figure 2. Hazard ratios and 95% confidence intervals for incident mobility limitation among men (A) and women (B) by history of obesity (body
mass index
30 kg/m2), the Health, Aging and Body Composition Study, 7 years of follow-up. Models were adjusted for age, race, field center,
education (<high school,
high school), smoking status (former, current, never), alcohol consumption (former, never, <1 drink/week, 1–7 drinks/
week, >7 drinks/week), and physical activity (<500 kcal/week, 500–<1,500 kcal/week,
1,500 kcal/week) at study baseline.

2 consecutive reports of mobility limitation reduces the in- rogate of body fatness is also a limitation of these analyses.
fluence of transient mobility limitation. Use of BMI as a sur- Although BMI is correlated with body fatness in young and

Am J Epidemiol 2009;169:927–936

middle-aged adults, changes in body composition that ac-
company aging alter the relation between BMI and body
fatness in older adults such that older adults have more body
fat for a given BMI compared with younger adults (28).
Finally, the observational nature of our study did not enable
us to evaluate a causal association between BMI history and
mobility limitation. It is biologically plausible that a history
of overweight and/or obesity may result in mobility limita-
tion. However, although the analyses were adjusted for be-
havioral characteristics such as smoking and physical
activity and for prevalent health conditions at study base-
line, overweight and/or obesity may serve as a proxy mea-
sure for other relevant participant characteristics that may
lead to both overweight and obesity and mobility limitation.
In conclusion, overweight and obesity in young, middle,
and late adulthood was associated with increased risk of
incident mobility limitation in this well-functioning cohort
of older adults. Those with a history of overweight and/or
obesity in midlife or earlier but not in late adulthood also
tended to have an increased risk of incident mobility limi-
tation compared with those who were normal weight
throughout. Furthermore, there appeared to be a graded re-
sponse between age at onset of overweight and/or obesity
and risk of mobility limitation. These data suggest that in-
terventions targeting prevention of overweight and obesity
in young and middle-aged adults may be useful in prevent-
ing or delaying the onset of mobility limitation later in life.
ACKNOWLEDGMENTS
Author affiliations: Sticht Center on Aging, Wake Forest
University School of Medicine, Winston-Salem, North Car-
olina (Denise K. Houston, Jingzhong Ding, Barbara J.
Nicklas, Stephen B. Kritchevsky); Laboratory of Epidemi-
ology, Demography, and Biometry, National Institute on
Aging, Bethesda, Maryland (Tamara B. Harris); Department
of Foods and Nutrition, University of Georgia, Athens,
Georgia (Jung Sun Lee); Department of Epidemiology and
Biostatistics, University of California, San Francisco, Cali-
fornia (Michael C. Nevitt, Susan M. Rubin); and Department
of Preventive Medicine, University of Tennessee Health Sci-
ence Center, Memphis, Tennessee (Frances A. Tylavsky).
This work was supported in part by the Intramural
Research Program of the National Institute on Aging,
National Institutes of Health; National Institute on Aging,
National Institutes of Health (contracts N01-AG-6-2101,
N01-AG-6-2103, and N01-AG-6-2106); and the Wake
Forest University Claude D. Pepper Older Americans
Independence Center (grant P30-AG21332 to D. K. H.).
Conflict of interest: none declared.
REFERENCES
1. Federal Interagency Forum on Aging-Related Statistics. Older
Americans 2004: Key Indicators of Well-Being. Washington,
DC: US Government Printing Office; 2004.
Am J Epidemiol 2009;169:927–936
Weight History and Mobility Limitation 935
2. Ogden CL, Carroll MD, Curtin LR, et al. Prevalence of over-
weight and obesity in the United States, 1999–2004. JAMA.
2006;295(13):1549–1555.
3. Willett WC, Dietz WH, Colditz GA. Guidelines for healthy
weight. N Engl J Med. 1999;341(6):427–434.
4. Overweight, obesity, and health risk. National Task Force on
the Prevention and Treatment of Obesity. Arch Intern Med.
2000;160(7):898–904.
5. LaCroix AZ, Guralnik JM, Berkman LF, et al. Maintaining
mobility in late life. II, Smoking, alcohol consumption,
physical activity, and body mass index. Am J Epidemiol.
1993;137(8):858–869.
6. Launer LJ, Harris T, Rumpel C, et al. Body mass index, weight
change, and risk of mobility disability in middle-aged and
older women: the epidemiologic follow-up study of NHANES
I. JAMA. 1994;271(14):1093–1098.
7. Jensen GL, Friedmann JM. Obesity is associated with func-
tional decline in community-dwelling rural older persons.
J Am Geriatr Soc. 2002;50(5):918–923.
8. Ostbye T, Taylor DH Jr, Krause KM, et al. The role of smoking
and other modifiable lifestyle risk factors in maintaining and
restoring lower body mobility in middle-aged and older
Americans: results from the HRS and AHEAD. Health and
Retirement Study. Asset and Health Dynamics Among the
Oldest Old. J Am Geriatr Soc. 2002;50(4):691–699.
9. Ferraro KF, Su YP, Gretebeck RJ, et al. Body mass index and
disability in adulthood: a 20-year panel study. Am J Public
Health. 2002;92(5):834–840.
10. Daviglus ML, Liu K, Yan LL, et al. Body mass index in middle
age and health-related quality of life in older age. Arch Intern
Med. 2003;163(20):2448–2455.
11. Sturm R, Ringel JS, Andreyeva T. Increasing obesity rates
and disability trends. Health Aff (Millwood). 2004;23(2):
199–205.
12. Lakdawalla DN, Bhattacharya J, Goldman DP. Are the young
becoming more disabled? Health Aff (Millwood). 2004;23(1):
168–176.
13. Alley DE, Chang VW. The changing relationship of obesity
and disability, 1988–2004. JAMA. 2007;298(17):2020–2027.
14. Hubert HB, Bloch DA, Fries JF. Risk factors for physical
disability in an aging cohort: the NHANES I Epidemiologic
Followup Study. J Rheumatol. 1993;20(3):480–488.
15. Houston DK, Stevens J, Cai J, et al. Role of weight history on
functional limitations and disability in late adulthood: the
ARIC study. Obes Res. 2005;13(10):1793–1802.
16. Stenholm S, Rantanen T, Alanen E, et al. Obesity history as
a predictor of walking limitation at old age. Obesity (Silver
Spring). 2007;15(4):929–938.
17. Houston DK, Ding J, Nicklas BJ, et al. The association be-
tween weight history and physical performance in the Health,
Aging and Body Composition study. Int J Obes (Lond). 2007;
31(11):1680–1687.
18. Guralnik JM, Ferrucci L, Pieper CF, et al. Lower extremity
function and subsequent disability: consistency across studies,
predictive models, and value of gait speed alone compared
with the short physical performance battery. J Gerontol A Biol
Sci Med Sci. 2000;55(4):M221–M231.
19. Radloff L. The CES-D Scale: a self-report depression scale for
research in the general population. Appl Psychol Meas. 1977;
1:385–401.
20. Teng EL, Chui HC. The Modified Mini-Mental State (3MS)
examination. J Clin Psychiatry. 1987;48(8):314–318.
21. Stevens J, Keil JE, Waid LR, et al. Accuracy of current, 4-year,
and 28-year self-reported body weights in an elderly popula-
tion. Am J Epidemiol. 1990;132(6):1156–1163.
936 Houston et al.
22. Rhoads GG, Kagan A. The relation of coronary disease,
stroke, and mortality to weight in youth and in middle age.
Lancet. 1983;1(8323):492–495.
23. Casey VA, Dwyer JT, Berkey CS, et al. Long-term memory of
body weight and past weight satisfaction: a longitudinal
follow-up study. Am J Clin Nutr. 1991;53(6):1493–1498.
24. Kuczmarski MF, Kuczmarski RJ, Najjar M. Effects of age on
validity of self-reported height, weight, and body mass index:
findings from the Third National Health and Nutrition Exam-
ination Survey, 1988–1994. J Am Diet Assoc. 2001;101(1):
28–34.
25. Gunnell D, Berney L, Holland P, et al. How accurately are
height, weight and leg length reported by the elderly, and how
closely are they related to measurements recorded in child-
hood? Int J Epidemiol. 2000;29(3):456–464.
26. Sorkin JD, Muller DC, Andres R. Longitudinal change in the
heights of men and women: consequential effects on body
mass index. Epidemiol Rev. 1999;21(2):247–260.
27. Fried LP, Young Y, Rubin G, et al. Self-reported preclinical
disability identifies older women with early declines in per-
formance and early disease. J Clin Epidemiol. 2001;54(9):
889–901.
28. Villareal DT, Apovian CM, Kushner RF, et al. Obesity in older
adults: technical review and position statement of the Ameri-
can Society for Nutrition and NAASO, The Obesity Society.
Am J Clin Nutr. 2005;82(5):923–934.
Am J Epidemiol 2009;169:927–936
 American Journal of Epidemiology Vol. 169, No. 8
ª The Author 2009. Published by the Johns Hopkins Bloomberg School of Public Health. DOI: 10.1093/aje/kwp003
All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org. Advance Access publication February 24, 2009

Original Contribution

Association of Diabetes With Prostate Cancer Risk in the Multiethnic Cohort

Kevin M. Waters, Brian E. Henderson, Daniel O. Stram, Peggy Wan, Laurence N. Kolonel, and
Christopher A. Haiman

Initially submitted August 25, 2008; accepted for publication January 6, 2009.

Among men of European ancestry, diabetics have a lower risk of prostate cancer than do nondiabetics. The
biologic basis of this association is unknown. The authors have examined whether the association is robust across
populations in a population-based prospective study. The analysis included 5,941 prostate cancer cases identified
over a 12-year period (1993–2005) among 86,303 European-American, African-American, Latino, Japanese-
American, and Native Hawaiian men from the Multiethnic Cohort. The association between diabetes and prostate-
specific antigen (PSA) levels (n ¼ 2,874) and PSA screening frequencies (n ¼ 46,970) was also examined.
Diabetics had significantly lower risk of prostate cancer than did nondiabetics (relative risk ¼ 0.81, 95% confidence
interval (CI): 0.74, 0.87; P < 0.001), with relative risks ranging from 0.65 (95% CI: 0.50, 0.84; P ¼ 0.001) among
European Americans to 0.89 (95% CI: 0.77, 1.03; P ¼ 0.13) among African Americans. Mean PSA levels were
significantly lower in diabetics than in nondiabetics (mean PSA levels, 1.07 and 1.28, respectively; P ¼ 0.003) as
were PSA screening frequencies (44.7% vs. 48.6%; P < 0.001); however, this difference could explain only a small
portion (~20%) of the inverse association between these diseases. Diabetes is a protective factor for prostate
cancer across populations, suggesting shared risk factors that influence a common mechanism.

cohort studies; diabetes mellitus, type 2; ethnology; prostate-specific antigen; prostatic neoplasms

Abbreviations: CI, confidence interval; PSA, prostate-specific antigen; RR, relative risk.

Prostate cancer and type 2 diabetes are 2 of the most
common chronic diseases that afflict the aging male popu-
lation. Epidemiologic studies conducted primarily in popu-
lations of European ancestry have provided evidence of an
inverse relation between these diseases, with diabetics hav-
ing ~20% lower risk of developing prostate cancer than do
nondiabetics (1–8). However, considerable effect heteroge-
neity has been noted among studies, highlighting the need
for additional prospective analyses of these endpoints in
large representative population-based studies. The biologic
basis of this suspected relation is currently unknown and,
aside from age and perhaps obesity (4, 9, 10), these 2 dis-
eases share no known nongenetic risk factors.
The inverse relation between these endpoints may be due
to direct effects on prostate cancer growth and development,
as men with type 2 diabetes have been found to have lower
prostate-specific antigen (PSA) levels, on average, than do
men without type 2 diabetes (11, 12). The reported protec-
tive effects of type 2 diabetes may also be attributed to
differences in prostate cancer screening in diabetic and non-
diabetic patients. Differences in health maintenance, access
to medical care, and the presence of serious medical con-
ditions may result in more (or less) medical attention and
preventive measures (13). Thus, examining the association
between type 2 diabetes status and prostate cancer screening
frequencies is important to quantify the degree to which
detection bias may explain the apparent relation.
The incidence rates of type 2 diabetes and prostate cancer
vary widely across populations. However, the ethnic dispar-
ities for these common diseases are not correlated; that is,
not all populations with high rates of diabetes are at low risk
of prostate cancer (14–20). The extent to which these dis-
eases are linked in non-European populations is not clear.
To confirm the previously reported association in a large,

Correspondence to Dr. Christopher A. Haiman, Harlyne Norris Research Tower, 1450 Biggy Street, Room 1504A, Mail Code LG591 MC9601,
Los Angeles, CA 90033 (e-mail: haiman@usc.edu).

937 Am J Epidemiol 2009;169:937–945

938 Waters et al.
representative population-based prospective study, as well
as to examine the consistency of the association across
racial/ethnic populations with differing rates of prostate
cancer, we evaluated prostate cancer incidence by type
2 diabetes status in a multiethnic sample of 86,303 men
from the Multiethnic Cohort (21). We also assessed the pre-
sumed effect of diabetes status in the etiology of prostate
cancer by examining PSA levels in a multiethnic sample of
men with and without type 2 diabetes. We also evaluated
PSA screening frequencies in diabetics and nondiabetics to
define the role of screening bias in explaining the observed
association between these common diseases.
MATERIALS AND METHODS
Study population
The Multiethnic Cohort is a prospective cohort study that
includes 215,251 men and women, the majority from 5
racial/ethnic groups in Hawaii and Los Angeles, California
(African Americans, European Americans, Native Hawaiians,
Japanese Americans, and Latinos) (21). Between 1993 and
1996, participants entered the cohort by completing a
26-page, self-administered questionnaire that asked about
diet and demographic factors, personal behaviors (e.g.,
physical activity), history of prior medical conditions
(e.g., diabetes), and family history of common cancers.
Potential cohort members were identified primarily through
Department of Motor Vehicles drivers’ license files and,
additionally for African Americans, Health Care Financing
Administration data files. Participants were between the ages
of 45 and 75 years at the time of recruitment.
In the cohort, incident prostate cancer cases are identified
annually through cohort linkage to the population-based
Surveillance, Epidemiology, and End Results (SEER) can-
cer registries in Hawaii and Los Angeles County, as well as
the California Cancer Registry. Information on stage and
grade of disease is also obtained through these registries.
Linkage with these registries is complete through December
31, 2004, in Hawaii and December 31, 2005, in California.
Over this period, 5,941 incident cases of invasive prostate
cancer were identified. Deaths within the cohort are deter-
mined from linkages to the death certificate files in Hawaii
and California, supplemented with linkages to the National
Death Index. In the Multiethnic Cohort, diabetes status is
defined on the basis of self-report on the baseline question-
naire. This question did not differentiate between type 1
diabetes and type 2 diabetes, and thus we expect a small
fraction (<10%) of the respondents to have type 1 diabetes
and to be potentially misclassified (22).
In addition to self-reported race/ethnicity, the following
risk factors were included in the analysis: body mass index
(weight (kg)/height (m)2), educational level (
12 years,
some college or vocational, and college graduate), first-
degree family history of prostate cancer, and amount of
vigorous physical activity (0, >0–1.5, >1.5–5, and >5 hours/
week). Vigorous activity includes both vigorous sports and
vigorous work.
We limited our analysis to 91,018 men in the Multiethnic
Cohort from the 5 major racial/ethnic groups. We excluded
men with a prevalent report of prostate cancer (n ¼ 3,004)
based on self-report or from the Surveillance, Epidemiol-
ogy, and End Results registries. We excluded men with
missing information for body mass index (n ¼ 838), educa-
tional level (n ¼ 872), and diabetes status (n ¼ 1). The
prospective analysis of the association between diabetes
status and prostate cancer incidence in this study includes
86,303 men.
PSA levels were previously measured on 4,623 men in the
Multiethnic Cohort (23). These men were randomly selected
from the cohort to evaluate the distribution of PSA levels
across ethnic groups. We excluded 194 men with prevalent
prostate cancer at baseline. We also excluded 1,527 men
with incident prostate cancer during the follow-up period,
to ensure that elevated PSA levels among undiagnosed cases
did not influence the results. Another 28 men with missing
body mass index data were excluded from the analysis,
leaving 2,874 men who are included in the final analysis
of the effect of type 2 diabetes on PSA levels.
In 2001, we sent a short follow-up questionnaire to cohort
members. On this questionnaire, we also asked about PSA
screening prior to 1999. Of the 86,303 men included in the
primary analysis of type 2 diabetes and prostate cancer,
23,768 (27.5%) did not complete the follow-up question-
naire. We also excluded 4,649 men with incident prostate
cancer. Finally, we excluded men under the age of 50 years
(n ¼ 10,916) because annual PSA screening is recommen-
ded to begin at age 50 (24). This leaves 46,970 men included
in the analysis of the association between type 2 diabetes
and PSA screening.
The informed consent and study protocol were approved
by the institutional review boards at the University of South-
ern California and the University of Hawaii.
Statistical analysis
Cox regression was used to estimate hazard ratios (re-
ported as relative risks) for the effect of type 2 diabetes on
prostate cancer incidence (STATA, version 8, software; Sta-
taCorp LP, College Station, Texas). We adjusted for age,
body mass index, educational level, and race/ethnicity (in
pooled analyses). Neither body mass index nor educational
level was associated with prostate cancer risk, but both re-
mained in the model as the former was found to be associ-
ated with PSA levels and the latter was found to be a highly
significant predictor of PSA screening. Physical activity and
family history of prostate cancer were left out of the final
model because neither had an effect on the association be-
tween type 2 diabetes and prostate cancer. Stratified analy-
ses were performed in older age groups to assess whether
type 2 diabetes duration and long-term exposure to declin-
ing insulin levels may be important in prostate cancer de-
velopment. Because men may be at increased risk of
prostate cancer within the first few years following a diabetes
diagnosis as a result of higher insulin levels, and since the
date of type 2 diabetes diagnosis is unknown for cohort
members, we also performed a sensitivity analysis to exam-
ine whether the association might be attenuated in recently
diagnosed diabetics. In this analysis, we censored follow-up
of incident prostate cancer cases incrementally by year from
Am J Epidemiol 2009;169:937–945
 Type 2 Diabetes and Prostate Cancer Risk 939

Table 1. Descriptive Characteristics by Race/Ethnicity and Diabetes Status (Yes/No) in the Multiethnic Cohort (n ¼ 86,303), Los Angeles,
California, and Hawaii, 1993–2005

European African Native Japanese

Latinos
Americans Americans Hawaiians Americans Total
Yes No Yes No Yes No Yes No Yes No

No. of men 1,440 20,606 1,791 9,475 988 5,088 3,160 22,927 3,446 17,382 86,303
Mean age, years (SD) 62.8 (8.1) 58.5 (9.0) 63.4 (8.0) 60.9 (8.9) 59.2 (8.0) 56.1 (8.6) 63.6 (8.3) 60.6 (9.2) 61.7 (7.0) 59.7 (7.8)
No. of prostate 60 1,193 215 1,295 35 213 160 1,302 198 1,270 5,941
cancer cases
Prostate cancer 242.8 413.9 686.2 845.2 257.9 373.8 270.1 354.9 304.2 433.7
incidence ratesa
Family history of 7.0 7.9 7.9 8.8 5.6 5.7 6.1 6.2 5.1 5.8
prostate cancer, %b
Body mass index
(kg/m2), %b
<23 13.1 21.3 11.1 17.5 7.6 13.8 21.0 26.3 12.3 16.7
23–24.99 8.1 17.9 9.6 14.6 4.7 11.1 17.3 25.1 8.1 10.8
25–29.99 42.7 45.7 44.5 47.8 40.0 45.6 45.1 41.4 49.1 53.8
30–34.99 23.7 11.9 24.4 15.9 28.1 20.3 12.9 6.2 22.8 15.1
35 12.3 3.1 10.0 4.3 19.6 9.1 3.7 1.0 7.7 3.5
Educational level, %b

12 years 34.6 23.2 41.8 40.0 63.0 53.8 36.6 34.8 67.4 64.1
Some college or 29.0 29.1 36.9 37.0 25.8 28.5 34.3 30.5 22.5 23.5
vocational
College graduate 36.4 47.7 20.8 22.9 11.2 17.7 29.1 34.6 10.2 12.3
Physical activity
(hours/week), %b,c
0 39.5 27.9 43.7 34.1 29.7 22.1 39.3 33.0 37.7 28.9
>0–1.5 16.0 14.4 17.5 15.6 13.9 13.4 19.6 18.2 14.3 14.3
>1.5–5 20.8 23.5 15.7 22.0 23.4 25.7 20.9 23.4 18.0 20.9
>5 20.1 31.6 17.2 23.9 29.3 35.7 17.3 23.0 24.4 31.1

Abbreviation: SD, standard deviation.

a
Adjusted to the 1970 US standard population.
b
Age standardized (5-year age groups) to the total population included in the study.
c
Percentages do not add up to 100% because of missing values.

1 to 5 years after cohort entry. We also examined the asso-
ciation in analyses stratified by body mass index (25 kg/
m2 and <25 kg/m2). Analyses stratified by Gleason score to
determine the effect of type 2 diabetes status on prostate
cancer severity were also conducted. This latter analysis
excludes 370 prostate cancer cases with missing information
on the Gleason score.
In the analysis of PSA levels, generalized linear models
were used to estimate least-squared mean PSA levels by
type 2 diabetes status (SAS, version 9.1, software; SAS In-
stitute, Inc., Cary, North Carolina). Models were adjusted
for the putative confounders of age, body mass index, and
race/ethnicity. We calculated PSA screening frequencies
adjusted for both age and educational level by type 2 di-
abetes status, and we tested for a difference using logistic
regression; body mass index was not found to influence the
effect of type 2 diabetes on PSA screening. The fraction of
the association between type 2 diabetes and prostate cancer
that may be attributable to PSA screening was estimated.
Assuming that prostate cancer incidence roughly doubled
since the initiation of PSA screening (25), with about 50%
of men being screened, we estimate that incidence rates
have increased by 0.02 per 1% of the population screened.
We then used this slope to estimate the relative impact of
screening on prostate cancer incidence in diabetic and non-
diabetic men as follows: relative risk (RR)PSA ¼ (1 þ 0.02 3
screening frequency in nondiabetics)/(1 þ 0.02 3 screening
frequency in diabetics), with (1  RRPSA/1  RRT2D) being
an estimate of the fraction of the association between type 2
diabetes (T2D) and prostate cancer incidence attributable to
PSA screening.
RESULTS
The mean age of the men (n ¼ 86,303) in this study was
59.9 (standard deviation, 8.8) years and ranged from 56.6
for Native Hawaiians to 61.3 for African Americans (Table 1).
The age-standardized prostate cancer incidence rate of
830.2 (per 100,000) was nearly 2 times greater in African

Am J Epidemiol 2009;169:937–945
940 Waters et al.
Americans than in the other populations (Table 1). The age-
adjusted prevalence of type 2 diabetes also varied widely
across populations, from 6.9% in European Americans to
18.0% in Native Hawaiians. The mean age of the diabetic
men in our study at baseline was slightly higher than that of
the nondiabetic men for each racial/ethnic group, ranging
from 59.2 years (vs. 56.1 in nondiabetics) in Native Hawai-
ians to 63.6 years (vs. 60.6 in nondiabetics) in Japanese
Americans. The age-standardized prostate cancer incidence
rates were lower in diabetic men than in nondiabetic men for
each racial/ethnic group, ranging from 242.8 (per 100,000)
in European-American diabetics (vs. 413.9 in nondiabetics)
to 686.2 in African-American diabetics (vs. 845.2 in non-
diabetics). First-degree family history of prostate cancer
was also less common in diabetic men than in nondiabetic
men for each racial/ethnic group, ranging from 5.1% in
Latino diabetic men (vs. 5.8% in nondiabetics) to 7.9% in
African-American diabetic men (vs. 8.8% in nondiabetics).
As expected, in each population, diabetic men were more
likely to be overweight and less physically active than men
without type 2 diabetes (Table 1).
In multivariate analyses, men with type 2 diabetes had
significantly lower risk of prostate cancer than did men
without type 2 diabetes (RR ¼ 0.81, 95% confidence in-
terval (CI): 0.74, 0.87; P < 0.001) (Table 2). The inverse
association was observed consistently in all 5 populations
and ranged from 0.65 (95% CI: 0.50, 0.84) in European
Americans to 0.89 (95% CI: 0.77, 1.03) in African Ameri-
cans (Pheterogeneity ¼ 0.32). We also examined the effect of
type 2 diabetes status on prostate cancer incidence by age at
entry into the cohort as a surrogate for duration of type 2
diabetes, as the progressive decline of insulin levels with age
among type 2 diabetics has been suggested to be protective
for prostate cancer (1, 3, 6, 8). We found no evidence that
the inverse association was strengthened among older men
(Table 2). However, we did observe a slight, yet consistent,
decrease in the relative risk when censoring the follow-up of
incident cases, by year, within the first 5 years of follow-up
(Appendix Figure 1). We observed no significant difference
in the association when stratified by body mass index (25
kg/m2: RR ¼ 0.78, 95% CI: 0.71, 0.86; <25 kg/m2: RR ¼
0.86, 95% CI: 0.74, 1.00; Pinteraction ¼ 0.24). We also ob-
served consistent effects by disease severity (Gleason score

7, n ¼ 3,853: RR ¼ 0.81, 95% CI: 0.73, 0.90; Gleason
score >7, n ¼ 1,703: RR ¼ 0.76, 95% CI: 0.65, 0.89).
Association of PSA levels with type 2 diabetes status
and body mass index
In the subset of 2,874 men with PSA measurements, di-
abetic men (n ¼ 344) were found to have significantly lower
least square geometric mean PSA levels than did nondia-
betic men (n ¼ 2,530; 1.04 vs. 1.29 ng/mL; P < 0.001).
Adjusting for body mass index had little effect on this as-
sociation (1.07 vs. 1.28 ng/mL; P ¼ 0.003) (Table 3). This
association was noted in all populations except Native
Hawaiians and was statistically significant in European
Americans (0.62 vs. 1.21 ng/mL; P ¼ 0.003) and Latinos
(0.99 vs. 1.27 ng/mL; P ¼ 0.02). Consistent with previous
reports (11, 12), this report also shows an inverse relation
between body mass index and PSA levels (Table 3). In
ethnicity-pooled analyses, compared with men with a body
mass index of <25 kg/m2, men with a body mass index of
30 had 13.8% lower mean PSA levels (P ¼ 0.009). In
multivariate generalized linear models, adjusted for type 2
diabetes status and age at blood draw, we estimated a 1-unit
increase in body mass index to be associated with a 1.6%
decrease in mean PSA level (P < 0.001).
Association of type 2 diabetes status and education
with PSA screening frequencies
In the sample of 46,970 men over 50 years of age with
information on PSA screening, 48.2% reported a PSA
screening test prior to 1999. European-American men were
more likely to report having been screened (55.8%), while
Native Hawaiians (34.3%) were the least likely to have had
PSA testing (Pheterogeneity < 0.001). We observed a modest,
yet highly statistically significant difference in age and ed-
ucational level regarding standardized PSA screening fre-
quencies between diabetics (44.7%) and nondiabetics
(48.6%; P < 0.001) (Table 4). The lower PSA screening
frequencies among diabetics were noted in all populations
except in African Americans and were statistically signifi-
cant in Japanese Americans (P < 0.001) and Latinos (P ¼
0.02). Men with higher educational levels were much more
likely to have had a PSA test than were men with
12 years
of schooling (Table 4). We adjusted for educational level as
a surrogate for PSA screening in the primary cohort analyses
discussed above, yet it had little impact on the association.
Next, we examined the potential impact of detection bias
on the observed association between type 2 diabetes and
prostate cancer. On the basis of the 3.9% difference in
PSA screening frequencies observed between diabetics
and nondiabetics, we estimated that detection bias is likely
to account for only ~20% of the inverse association between
type 2 diabetes and prostate cancer risk.
DISCUSSION
In this prospective analysis of 5 racial/ethnic populations,
we found a highly significant association between type 2
diabetes status and prostate cancer incidence, with diabetics
having ~20% lower risk of developing prostate cancer. This
inverse association was observed in all populations, with the
magnitude of the effect being consistent with that of the
majority of other studies conducted in men of European
ancestry (1–8).
In this study, type 2 diabetes status was based on self-
report, which may have led to misclassification. Previous
studies, however, have shown that self-reported responses
for many common chronic diseases such as diabetes are
reliable when compared with medical records (26–28).
The analysis does not account for incident cases of type 2
diabetes over the 8-year follow-up period. However, inci-
dent cases of diabetes in the nondiabetes group would make
the 2 groups more similar and create an underestimation of
the inverse association. Another limitation of our study is
that we cannot differentiate between cases of type 1 diabetes
Am J Epidemiol 2009;169:937–945
Am J Epidemiol 2009;169:937–945

Table 2. Relative Risk of Prostate Cancer Associated With Diabetes Status by Age and Gleason Score in the Multiethnic Cohort, Los Angeles, California, and Hawaii (n ¼ 86,303),
1993–2005

European African Native Japanese

Latinos All
Americans Americans Hawaiians Americans
95% 95% 95% 95% 95% 95%
No. of Relative No. of Relative No. of Relative No. of Relative No. of Relative No. of Relative
Confidence Confidence Confidence Confidence Confidence Confidence
Cases Riska Cases Riska Cases Riska Cases Riska Cases Riska Cases Riska
Interval Interval Interval Interval Interval Interval

All men 1,253 0.65 0.50, 0.84 1,510 0.89 0.77, 1.03 248 0.73 0.51, 1.05 1,462 0.81 0.69, 0.96 1,468 0.78 0.67, 0.91 5,941 0.81 0.74, 0.87
P value 0.001 0.13 0.09 0.01 0.001 <0.001
Pheterogeneity 0.32
Age 50 years 1,192 0.66 0.51, 0.86 1,436 0.87 0.75, 1.01 233 0.72 0.49, 1.04 1,422 0.80 0.68, 0.95 1,434 0.79 0.68, 0.91 5,717 0.80 0.74, 0.87
P value 0.002 0.06 0.08 0.01 0.002 <0.001
Pheterogeneity 0.47
Age 60 years 932 0.66 0.49, 0.87 1,109 0.90 0.77, 1.06 170 0.76 0.50, 1.14 1,216 0.81 0.68, 0.97 1,069 0.77 0.65, 0.92 4,496 0.81 0.74, 0.89
P value 0.004 0.22 0.19 0.02 0.003 <0.001
Pheterogeneity 0.25
Age 70 years 346 1.03 0.70, 1.49 371 0.92 0.70, 1.21 47 0.71 0.31, 1.60 521 0.80 0.62, 1.05 293 0.69 0.49, 0.97 1,578 0.84 0.72, 0.97
P value 0.90 0.55 0.40 0.11 0.03 0.02

Type 2 Diabetes and Prostate Cancer Risk

Pheterogeneity 0.57
Gleason score
7 753 0.66 0.47, 0.92 1,075 0.87 0.73, 1.03 139 0.86 0.54, 1.35 830 0.78 0.63, 0.98 1,056 0.80 0.67, 0.95 3,853 0.81 0.73, 0.90
P value 0.02 0.11 0.51 0.03 0.01 <0.001
Pheterogeneity 0.60
Gleason score >7 384 0.68 0.43, 1.07 340 0.95 0.71, 1.29 95 0.61 0.32, 1.14 558 0.76 0.58, 1.00 341 0.68 0.49, 0.94 1,718 0.76 0.65, 0.89
P value 0.09 0.76 0.12 0.05 0.02 <0.001
Pheterogeneity 0.54
a
Adjusted for age, body mass index, and educational level. Adjusted for race in pooled analysis.

941
 942 Waters et al.

and type 2 diabetes. Although they have similar phenotypes,

Pheterogeneity
Table 3. Geometric Mean Prostate-specific Antigen Levels by Ethnicity, Diabetes Status, and Body Mass Index in the Multiethnic Cohort (n ¼ 2,874), Los Angeles, California, and Hawaii,

<0.001
type 1 diabetes and type 2 diabetes have distinct mecha-

0.11

0.49
0.93
nisms of pathogenesis and may have dissimilar associations
with prostate cancer incidence. However, we expect this
differential misclassification to be minimal as the preva-

935 Referent
0.003

0.009
<0.001
P Value

0.70
lence of type 1 diabetes is comparatively low in these
Alla

populations (22).
Diabetes and prostate cancer are both traditionally under-
2,874

2,530

1,384
344

555
No.

diagnosed diseases. In this study, undiagnosed type 2 dia-

1.6
betes would result in prostate cancer incidence rates being

Percent change in geometric prostate-specific antigen levels with increase of 1 body mass index unit adjusted for diabetes status and age at blood draw.
ng/mL
Mean,

more similar between the diabetic and nondiabetic groups.

1.26

1.07
1.28

1.30
1.28
1.12
Undiagnosed cases of prostate cancer would result in lower
rates of prostate cancer among both diabetics and nondia-
196 Referent
P Value

0.02

0.97
0.16
0.06
betics. As a result, we would expect these simultaneous
events of disease misclassification to counter the inverse
Latinos

association that we noted in this study toward the null. It

93
714

621

365
153
No.

is also possible that men who do not receive frequent med-

1.6

ical care would be underdiagnosed and misclassified for

ng/mL
Mean,

Adjusted for diabetes status, body mass index, and age at blood draw. African Americans versus each ethnic group, P < 0.001.
1.26

0.99
1.27

1.27
1.27
1.09

both diseases. As a result, the underdiagnosis of prostate

cancer and the lower risk of prostate cancer among the
240 Referent
P Value

misclassified diabetics would also result in a bias toward

0.53

0.03
0.18
0.06

the null of the underlying association. Although we expect

Americans
Japanese

that the underdiagnosis of these diseases is unable to explain

56

29
485

429

216
No.

their inverse association, future studies demanding regular

2.4

blood glucose and PSA screening will be needed to quantify

ng/mL
Mean,

1.22

1.15
1.26

1.37
1.14
1.07

the impact of this bias.

In this study, we also found that men with diabetes are
less likely to report PSA screening than are men without
59 Referent
P Value

0.86

0.26
0.18
0.06

diabetes. These findings are contrary to those of a previous

Hawaiians

study that reported that men with diabetes are more likely to
Native

undergo screening for prostate cancer (3). PSA screening

44
313

269

148
106
No.

frequencies were lower among diabetics in all populations

1.9

except in African Americans. Education, which is a surro-

ng/mL
Mean,

1.14

1.00
0.97

1.12
0.96
0.92

Adjusted for body mass index, diabetes status, ethnicity, and age at blood draw.

gate for socioeconomic status and access to health care, was

significantly associated with both PSA screening frequen-
266 Referent
P Value

cies and diabetes status. However, further adjustment for

0.29

0.84
0.07
0.03

education in the main cohort analyses did not change the

Americans
African

results. We estimated that the potential bias incurred by

916

126
790

457
193
No.

differential PSA screening (~4%) in diabetics and nondia-

1.8

betics explained only ~20% of the protective effect of type

ng/mL
Mean,

1.50

1.46
1.64

1.66
1.69
1.38

2 diabetes on prostate cancer risk. In addition, if the asso-

ciation between these conditions was influenced by detec-
Adjusted for body mass index and age at blood draw.
Adjusted for diabetes status and age at blood draw.
174 Referent

tion bias, then one would expect the inverse association to

P Value

0.003

0.18
0.84
0.65

diminish among severe cases of prostate cancer, because

Americans
European

they are likely to have been diagnosed without the use of

PSA screening. However, we observed only a minimal
25

74
446

421

198
No.

change in the association between diabetes status and pros-

0.6
ng/mL
Mean,

tate cancer incidence when stratified by disease severity.

1.19

0.62
1.21

1.09
1.27
1.12

Thus, detection bias associated with lower PSA levels

and/or lower PSA screening frequencies in diabetics is un-
Body mass index, % changee

likely to explain the strong and highly significant inverse

Body mass index (kg/m2)d

association between type 2 diabetes and prostate cancer in

this study.
Studies have suggested that the protective effect of di-
Diabetes statusc

abetes on prostate cancer incidence may be greater among

men with longstanding type 2 diabetes (3, 6, 8). One theory
25–29.99
1993–2005

is that hyperinsulinemia, which is observed at onset, is

All menb

<25

30
Yes

associated with increased levels of growth factors (e.g.,

No

insulin-like growth factor-I) that may induce prostate cancer

c
a
b

d
e

during the first few years of type 2 diabetes. Subsequently,

Am J Epidemiol 2009;169:937–945
 Type 2 Diabetes and Prostate Cancer Risk 943

prostate cancer rates would decrease in the later stages of

Pheterogeneity
Table 4. Prostate-specific Antigen Screening Frequencies by Level of Education and Diabetes Status in the Multiethnic Cohort (n ¼ 46,970), Los Angeles, California, and Hawaii, 1993–2005

<0.001

<0.001
type 2 diabetes when insulin levels decrease and men be-

0.45
come hypoinsulinemic. We do not have data on the date of
diagnosis for type 2 diabetes, but we did analyze the asso-
ciation between type 2 diabetes status and prostate cancer

6,856 41.1 Referent 19,353 40.0 Referent

<0.001
<0.001

<0.001
P Value

incidence by age at entry to the cohort as a surrogate for

All

longstanding type 2 diabetes. With both of these theories,

one would expect the magnitude of the inverse association
46,970 48.2

13,588 49.1
14,029 58.6

5,765 44.7
41,205 48.6
%

between diabetes and prostate cancer to be greater among

older men. However, we found no difference in the associ-
ation in older men. We did, however, notice a modest
No.

increase in the magnitude of the inverse association when

removing incident cases within the first 5 years of follow-up,
<0.001
<0.001
P Value

0.02
which supports the hypothesis that men with newly diag-
nosed diabetes may have an increased risk of prostate
Latinos

cancer.
10,973 44.7

2,702 50.8
1,415 50.5

1,793 44.3
9,180 47.4
%

Most, but not all, studies have shown that, on average,

men with diabetes have lower PSA levels than do those
P values are from a logistic regression model and are adjusted for age, educational level, and race/ethnicity (pooled analysis).

without diabetes (11, 12, 29). In our multiethnic sample,

No.

PSA levels were lower in diabetic men. However, what this

indicates is not clear. Lower PSA levels in diabetics may
6,247 37.1 Referent
<0.001
<0.001

<0.001
P Value

signal a lower prevalence of prostate cancer and an indica-

tion of a biologic effect of type 2 diabetes status on prostate
Americans
Japanese

growth and development. At the same time, the effect

16,021 46.3

4,846 46.0
4,928 56.4

1,947 42.1
14,074 46.0
%

P values are from a logistic regression model and are adjusted for age and race/ethnicity (pooled analysis).

of diabetes status on PSA levels could result in decreased

Percentages are age (5-year age groups) and education standardized to the total population in the study.

follow-up for prostate cancer diagnosis among diabetics,

No.

which may partially account for the inverse relation between

Percentages are age standardized (5-year age groups) to the total population included in the study.

type 2 diabetes and prostate cancer risk. Additional work

will be needed to understand whether type 2 diabetes status
2,747 44.5 Referent 1,954 46.3 Referent 1,549 30.1 Referent
0.001
<0.001
P Value

0.37

influences the accuracy of PSA screening or directly con-

tributes to prostate cancer risk. Consistent with previous
Hawaiians
Native

studies (11, 12, 30), our analysis also suggests an inverse

2,995 34.3

874 35.6
572 45.7

487 35.1
2,508 36.4
%

relation between body mass index and PSA levels. Further

studies of this association are necessary to determine if
No.

obese men should have lower PSA thresholds to indicate

further work-up for prostate cancer.
<0.001
<0.001
P Value

Only a small number of studies have investigated the re-

0.84

lation between type 2 diabetes and prostate cancer risk in

Americans

non-European populations (2, 31–33). Most of these studies

African

have observed nonsignificant inverse associations; however,

4,946 52.1

1,819 53.3
1,173 59.8

742 52.5
4,204 52.1
%

small sample sizes have limited interpretation of the find-

No. of men with prostate-specific antigen screening data.

ings. Our results, from a population-based prospective study

No.

of over 5,900 prostate cancer cases from 5 racial/ethnic

populations, provide strong support for the pan-ethnic na-
<0.001
<0.001
P Value

0.46

ture of the association between these common diseases.

Recently, a common variant in the hepatocyte nuclear
Americans
European

factor-1 b gene (HNF1b) was found to be associated with

12,035 55.8

3,347 53.6
5,941 63.4

796 51.4
11,239 53.0
b

an increased risk of prostate cancer. This same variant was

%

also found to be associated with a decreased risk of type 2

a

diabetes (34). Common genetic variation in another gene,

No.

JAZF1, has also been associated with risks of both prostate

cancer and type 2 diabetes (35, 36). These findings, along
College or vocational

with findings from other studies that have shown that di-
Percent screened.
College graduate
c,d

abetes is inversely associated with a family history of pros-

e,f
Educational level

Diabetes status

tate cancer (5, 37), which we also noted, point to both

12 years

shared genetic risk and common molecular and/or meta-

bolic pathways that are important in the etiology of these
All menc

Yes

diseases.
No

In summary, in this large, multiethnic prospective study,

a
b

d
e
c

we observed consistent inverse associations between type 2

Am J Epidemiol 2009;169:937–945
944 Waters et al.
diabetes and prostate cancer risk across multiple racial/
ethnic populations. These findings provide strong support
for the hypothesis that type 2 diabetes is a protective factor
for prostate cancer. We also confirmed results from previous
studies, showing that PSA levels are decreased in diabetic
men. Our findings that diabetic men are less likely to be
screened for prostate cancer could not account for these
results. Future studies aimed at determining the biologic
link between diabetes and prostate cancer are warranted
and should focus on common environmental and genetic
factors that are shared across populations.
ACKNOWLEDGMENTS
Author affiliations: Department of Preventive Medicine,
Keck School of Medicine, University of Southern California/
Norris Comprehensive Cancer Center, Los Angeles, Cali-
fornia (Kevin M. Waters, Brian E. Henderson, Daniel O.
Stram, Peggy Wan, Christopher A. Haiman); and Epidemi-
ology Program, Cancer Research Center, University of
Hawaii, Honolulu, Hawaii (Laurence N. Kolonel).
This work was supported by the National Cancer Institute
(grant CA54281).
Conflict of interest: none declared.
REFERENCES
1. Kasper JS, Giovannucci E. A meta-analysis of diabetes mel-
litus and the risk of prostate cancer. Cancer Epidemiol Bio-
markers Prev. 2006;15(11):2056–2062.
2. Calton BA, Chang SC, Wright ME, et al. History of diabetes
mellitus and subsequent prostate cancer risk in the NIH-AARP
Diet and Health Study. Cancer Causes Control. 2007;18(5):
493–503.
3. Giovannucci E, Rimm EB, Stampfer MJ, et al. Diabetes mel-
litus and risk of prostate cancer. Cancer Causes Control.
1998;9(1):3–9.
4. Gong Z, Neuhouser ML, Goodman PJ, et al. Obesity, diabetes,
and risk of prostate cancer: results from the prostate cancer
prevention trial. Cancer Epidemiol Biomarkers Prev. 2006;
15(10):1977–1983.
5. Velicer CM, Dublin S, White E. Diabetes and the risk of
prostate cancer: the role of diabetes treatment and complica-
tions. Prostate Cancer Prostatic Dis. 2007;10(1):46–51.
6. Rodriguez C, Patel AV, Mondul AM, et al. Diabetes and risk of
prostate cancer in a prospective cohort of US men. Am J
Epidemiol. 2005;161(2):147–152.
7. Pierce BL, Plymate S, Ostrander EA, et al. Diabetes mellitus
and prostate cancer risk. Prostate. 2008;68(10):1126–1132.
8. Tavani A, Gallus S, Bosetti C, et al. Diabetes and the risk of
prostate cancer. Eur J Cancer Prev. 2002;11(2):125–128.
9. Giovannucci E, Rimm EB, Stampfer MJ, et al. Height, body
weight, and risk of prostate cancer. Cancer Epidemiol Bio-
markers Prev. 1997;6(8):557–563.
10. Huxley R, James WP, Barzi F, et al. Ethnic comparisons of the
cross-sectional relationships between measures of body size
with diabetes and hypertension. Obes Rev. 2008;9(suppl 1):
53–61.
11. Fukui M, Tanaka M, Kadano M, et al. Serum prostate-specific
antigen level in men with type 2 diabetes. Diabetes Care.
2008;31(5):930–931.
12. Werny DM, Saraiya M, Gregg EW. Prostate-specific antigen
values in diabetic and nondiabetic US men, 2001–2002. Am J
Epidemiol. 2006;164(10):978–983.
13. Ross LE, Berkowitz Z, Ekwueme DU. Use of the prostate-
specific antigen test among U.S. men: findings from the 2005
National Health Interview Survey. Cancer Epidemiol Bio-
markers Prev. 2008;17(3):636–644.
14. Abate N, Chandalia M. The impact of ethnicity on type 2
diabetes. J Diabetes Complications. 2003;17(1):39–58.
15. Borrell LN, Crawford ND, Dailo FJ. Race/ethnicity and self-
reported diabetes among adults in the National Health Interview
Survey: 2000–2003. Public Health Rep. 2007;122(5):616–625.
16. Mau MK, Glanz K, Severino R, et al. Mediators of lifestyle
behavior change in Native Hawaiians: initial findings from the
Native Hawaiian Diabetes Intervention Program. Diabetes
Care. 2001;24(10):1770–1775.
17. Aluli N. Prevalence of obesity in a Native Hawaiian popula-
tion. Am J Clin Nutr. 1991;53(6 suppl):1556S–1560S.
18. Brawley OW, Jani AB, Master V. Prostate cancer and race.
Curr Probl Cancer. 2007;31(3):211–225.
19. Jemal A, Siegel R, Ward E, et al. Cancer statistics, 2008. CA
Cancer J Clin. 2008;58(2):71–96.
20. Kolonel LN, Altshuler D, Henderson BE. The Multiethnic
Cohort Study: exploring genes, lifestyle, and cancer risk. Nat
Rev Cancer. 2004;4(7):519–527.
21. Kolonel LN, Henderson BE, Hankin JH, et al. A multiethnic
cohort in Hawaii and Los Angeles: baseline characteristics.
Am J Epidemiol. 2000;151(4):346–357.
22. National Center for Chronic Disease Prevention and Health
Promotion. National diabetes fact sheet. Atlanta, GA: Centers
for Disease Control and Prevention; 2005. (http://www.cdc.gov/
diabetes/pubs/general.htm#what).
23. Cheng I, Yu MC, Koh WP, et al. Comparison of prostate-
specific antigen and hormone levels among men in Singapore
and the United States. Cancer Epidemiol Biomarkers Prev.
2005;14(7):1692–1695.
24. Smith RA, Cokkinides V, Eyre HJ, et al. American Cancer
Society guidelines for the early detection of cancer, 2003. CA
Cancer J Clin. 2003;53(1):27–43.
25. Collin SM, Martin RM, Metcalfe C, et al. Prostate-cancer
mortality in the USA and UK in 1975–2004: an ecological
study. Lancet Oncol. 2008;9(5):445–452.
26. Midthjell K, Holmen J, Biøndal A, et al. Is questionnaire in-
formation valid in the study of a chronic disease such as di-
abetes? The Nord-Trøndelag Diabetes Study. J Epidemiol
Community Health. 1992;46(5):537–542.
27. Okura Y, Urban LH, Mahoney DW, et al. Agreement between
self-report questionnaires and medical record data was sub-
stantial for diabetes, hypertension, myocardial infarction and
stroke but not heart failure. J Clin Epidemiol. 2004;57(10):
1096–1103.
28. Walitt BT, Constantinescu F, Katz JD, et al. Validation of self-
report of rheumatoid arthritis and systemic lupus erythematosus:
the Women’s Health Initiative. J Rheumatol. 2008;35(5):811–818.
29. Chan JM, Latini DM, Cowan J, et al. History of diabetes,
clinical features of prostate cancer, and prostate cancer
recurrence-data for CaPSURETM (United States). Cancer
Causes Control. 2005;16(7):789–797.
30. Fowke JH, Signorello LB, Chang SS, et al. Effects of obesity
and height on prostate-specific antigen (PSA) and percentage
of free PSA levels among African-American and Caucasian
men. Cancer. 2006;107(10):2361–2367.
Am J Epidemiol 2009;169:937–945
 Type 2 Diabetes and Prostate Cancer Risk 945

31. Coker AL, Sanderson M, Zheng W, et al. Diabetes mellitus 1.2

and prostate cancer risk among older men: population based 1.1
case-control study. Br J Cancer. 2004;90(11):2171–2175.
32. Mishina T, Watanabe H, Araki H, et al. Epidemiological study 1
of prostatic cancer by matched-pair analysis. Prostate. 1985;
0.9
6(4):423–436.

Relative Risk
33. Rosenberg DJ, Neugut AI, Ahsan H, et al. Diabetes mellitus
0.8
and the risk of prostate cancer. Cancer Invest. 2002;20(2):
157–165.
34. Gudmundsson J, Sulem P, Steinthorsdottir V, et al. Two var- 0.7
iants on chromosome 17 confer prostate cancer risk, and the
one in TCF2 protects against type 2 diabetes. Nat Genet. 0.6
2007;39(8):977–983.
35. Thomas G, Jacobs KB, Yeager M, et al. Multiple loci identified
in a genome-wide association study of prostate cancer. Nat 0.5
Genet. 2008;40(3):310–315. 0 1 2 3 4 5
36. Zeggini E, Scott LJ, Saxena R, et al. Meta-analysis of genome- Cox Analysis, years after cohort entry
wide association data and large-scale replication identifies
additional susceptibility loci for type 2 diabetes. Nat Genet. Appendix Figure 1. Ethnically pooled association between type 2
2008;40(5):638–645. diabetes and prostate cancer risk by analysis start point, Los Angeles,
37. Meyer P, Zuern C, Hermanns N, et al. The association between California, and Hawaii, 1993–2005. There were 5,941, 5,441, 4,933,
paternal prostate cancer and type 2 diabetes [electronic article]. 4,456, 3,941, and 3,373 cases at 0, 1, 2, 3, 4, and 5 years after cohort
J Carcinog. 2007;6:14. entry, respectively. Bars, 95% confidence interval.

Am J Epidemiol 2009;169:937–945
 American Journal of Epidemiology Vol. 169, No. 8
ª 2009 The Authors DOI: 10.1093/aje/kwn413
This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial
License (http://creativecommons.org/licenses/by-nc/2.0/uk/) which permits unrestricted non-commercial use, Advance Access publication February 13, 2009
distribution, and reproduction in any medium, provided the original work is properly cited.

Original Contribution

Modification of the Effect of Vitamin E Supplementation on the Mortality of Male

Smokers by Age and Dietary Vitamin C

Harri Hemilä and Jaakko Kaprio

Initially submitted September 1, 2008; accepted for publication December 15, 2008.

The Alpha-Tocopherol, Beta-Carotene Cancer Prevention (ATBC) Study (1985–1993) recruited 29,133 Finnish
male cigarette smokers, finding that vitamin E supplementation had no overall effect on mortality. The authors of
this paper found that the effect of vitamin E on respiratory infections in ATBC Study participants was modified by
age, smoking, and dietary vitamin C intake; therefore, they examined whether the effect of vitamin E supplemen-
tation on mortality is modified by the same variables. During a median follow-up time of 6.1 years, 3,571 deaths
occurred. Age and dietary vitamin C intake had a second-order interaction with vitamin E supplementation of
50 mg/day. Among participants with a dietary vitamin C intake above the median of 90 mg/day, vitamin E increased
mortality among those aged 50–62 years by 19% (95% confidence interval: 5, 35), whereas vitamin E decreased
mortality among those aged 66–69 years by 41% (95% CI:
56,
21). Vitamin E had no effect on participants who
had a dietary vitamin C intake below the median. Smoking quantity did not modify the effect of vitamin E. This study
provides strong evidence that the effect of vitamin E supplementation on mortality varies between different
population groups. Further study is needed to confirm this heterogeneity.

aging; antioxidants; oxidative stress; population characteristics; randomized controlled trial; smoking; survival rate

Abbreviations: ATBC, Alpha-Tocopherol, Beta-Carotene Cancer Prevention; CI, confidence interval; RR, risk ratio.

Taking vitamin E supplements is a common practice,
particularly among older people. In the United States, about
a quarter of adults aged 60 years or older take supplements
containing vitamin E (1). Such a common habit makes the
health effects of this practice an important public health
issue: does vitamin E supplementation improve health or
not?
The rationale behind taking the lipid-soluble antioxidant
vitamin E is to protect against oxidative stress, which con-
tributes to the aging processes and may affect longevity
(2, 3). However, 3 meta-analyses of randomized trials found
that vitamin E supplementation did not reduce mortality,
implying that vitamin E does not lead to universal systemic
benefits against the processes that lead to chronic disease
(4–6).
Vitamin E is a major lipid-soluble antioxidant, whereas
vitamin C is a major water-soluble antioxidant. They inter-
act in vitro and in vivo (7–11). Smoking increases the
plasma a-tocopherol disappearance rate, which is normal-
ized by vitamin C supplementation (10). Thus, since smok-
ing seems to modify the interaction of these 2 antioxidants,
intake of vitamin C is particularly important when examin-
ing the effect of vitamin E on smokers.
The Alpha-Tocopherol, Beta-Carotene Cancer Prevention
(ATBC) Study was a randomized, double-blind, placebo-
controlled trial that examined the effects of vitamin E and
b-carotene on lung cancer in male smokers (12, 13). Pre-
viously, we found significant heterogeneity in the effects of
vitamin E supplementation in the ATBC Study; vitamin E
increased the risk of tuberculosis for heavy smokers with
high dietary vitamin C intake but had no effect on other
participants (14). The effect of vitamin E on common cold
and pneumonia risk was modified by age and smoking (11,
15–19).
In the ATBC Study, vitamin E supplementation had no
overall effect on mortality (12). In this study, we tested the

Correspondence to Dr. Harri Hemilä, Department of Public Health, University of Helsinki, POB 41, Helsinki, FIN-00014, Finland (e-mail:
harri.hemila@helsinki.fi).

946 Am J Epidemiol 2009;169:946–953

 Vitamin E Supplementation and Heterogeneity in Mortality
hypothesis that the variables that modify the effect of vita-
min E supplementation on respiratory infections would also
modify the effect of vitamin E on mortality. We did not
presume that changes in the incidence of respiratory infec-
tions would significantly affect overall mortality, but that
changes in respiratory infections may reflect nonspecific
systemic effects of vitamin E that might also affect
mortality.
MATERIALS AND METHODS
Participants
The design and methods of the ATBC Study examining
the effects of vitamin E (dl-a-tocopheryl acetate, 50 mg/day)
and b-carotene (20 mg/day) on the incidence of lung cancer
and other cancers have been described earlier (12, 13).
The ATBC Study is registered at the website www.
ClinicalTrials.gov under the identifier NCT00342992.
In brief, male participants aged 50–69 years had to smoke
5 or more cigarettes per day at entry to be eligible, and those
enrolled in the trial (N ¼ 29,133) were randomized to 1 of 4
intervention arms and were administered placebo, vitamin
E, b-carotene, or vitamin E þ b-carotene; a 2 3 2 factorial
design was used. Compliance with supplementation was
high: some 90% of the participants took more than 90%
of their prescribed capsules during their active participation
in the trial; there were no differences in capsule consump-
tion among the intervention groups (13). Supplementation
increased the serum level of a-tocopherol by 50% compared
with baseline levels (12, 13). The intervention continued
until April 30, 1993. The trial was approved by the institu-
tional review boards, and all participants gave written in-
formed consent.
Baseline characteristics
Before randomization at baseline, the men completed
questionnaires on their medical and smoking histories and
general background characteristics, and their weight was
measured. A detailed dietary history questionnaire provided
data regarding vitamin C, vitamin E, fruit, vegetable, and
berry consumption (20, 21). The validity of the dietary his-
tory questionnaire was assessed by comparing it with the
food consumption records of 190 participants for 12 sepa-
rate 2-day periods distributed evenly over 6 months. Clas-
sified by the dietary history questionnaire, 74%–76% of
participants categorized by food consumption records were
in the same vitamin C intake quintile or in the within-one-
quintile category (20). Dietary data were not available for
2,022 of the 29,133 participants.
Outcome and follow-up time
Deaths were identified by using the National Death Reg-
istry, as previously described (12). In the database we ana-
lyzed was one death additional to those in the 1994 report.
Follow-up time for each participant began from the day of
randomization and continued until death or the end of the
trial. The median follow-up time for the participants in the
Am J Epidemiol 2009;169:946–953
947
present analysis was 6.1 years, and there was a total of
169,731 person-years of observation.
Statistical models
We estimated the effect of vitamin E supplementation on
mortality through proportional hazards regression models.
We calculated the risk ratio and the 95% confidence interval
of the risk ratio by using the PROC PHREG procedure in
SAS software (release 8.2; SAS Institute, Inc., Cary, North
Carolina). The 2 3 2 factorial design of the trial permitted
assessment of the effect of vitamin E independent of
b-carotene after confirming no statistical interaction be-
tween the agents. Thus, we compared the trial participants
administered vitamin E with those not receiving vitamin E
(no-vitamin-E participants). We did not analyze the effects
of b-carotene in this study. Regarding supplementation, we
carried out the analyses following the intention-to-treat
principle. Because deaths were identified in the National
Death Registry, which registers all deaths occurring in
Finland, loss-to-follow-up is insignificant.
To test the statistical significance of interaction between
vitamin E supplementation and potential modifying factors,
we first added the supplementation and modifying factor to the
regression model. The statistical significance of the interaction
was thereafter calculated from the change in
2 3 log(likeli-
hood) when the interaction term of vitamin E supplementation
and the modifying factor was added to the model.
In our subgroup analyses, we split dietary vitamin E and C
levels, and the residual of fruit, vegetable, and berry consump-
tion, at rounded levels close to the medians. Dietary vitamin C
was also used as a continuous variable because interaction
with a continuous variable refutes the possibility that dichot-
omizing dietary vitamin C intake might cause a spurious in-
teraction; to decrease the skewness of the distribution, the
statistical model included the logarithm of dietary vitamin C
intake.
Nelson-Aalen cumulative hazard functions were con-
structed by using the STATA sts program (release 9.1; Stata
Corporation, College Station, Texas). Two-tailed P values
were used.
Examination of the specificity of effect modification by
vitamin C
The major sources of vitamin C in the diet of study par-
ticipants were fruit, vegetables, and berries, on average 58%
of dietary vitamin C originating from these foods. Total in-
take of fruit, vegetables, and berries was strongly correlated
with the calculated vitamin C intake (r ¼ 0.88). Thus, it is
possible that an association with dietary vitamin C is a sta-
tistical artifact reflecting other substances in these foods or
the lifestyle related to eating these foods. To examine the
possible role of dietary compounds other than vitamin C in
these foods, we calculated the residual of fruit, vegetables,
and berries intake by using linear regression to model fruit,
vegetables, and berries as a function of dietary vitamin C, as
previously (22). As designed, the residual of fruit, vegeta-
bles, and berries intake has no correlation with dietary vita-
min C. We assumed that any other putative compound that
948 Hemilä and Kaprio

might interact with vitamin E supplementation has no perfect Table 1. Effect of Vitamin E Supplementation on Mortality by Age
correlation with vitamin C and, therefore, that variation in the and Dietary Vitamin C Intake, Alpha-Tocopherol, Beta-Carotene
other compound remains as variation in the residual of fruit, Cancer Prevention Study, 1985–1993
vegetables, and berries intake, which was split at 0 g/day, Vitamin C Test for
close to the median. High residual of fruit, vegetables, and Age at Baseline
Vitamin C
<90 mg/day ‡90 mg/day Interaction
berries intake (above zero) indicates that the participant with (n 5 13,567)a (n 5 13,544)a (P Value)
a given vitamin C level consumes more than the average
quantity of fruit, vegetables, and berries, whereas low resid- 50–62 years
(n ¼ 22,413)
ual of fruit, vegetables, and berries intake (below zero) in-
dicates less-than-average intake of these food classes. Risk ratiob 1.00 1.19 0.048
95% confidence 0.90, 1.13 1.05, 1.35
interval
Deathsc 614/616 552/469
RESULTS
63–65 years
During the 169,731 person-years of follow-up of the (n ¼ 2,761)
ATBC Study participants, 3,571 deaths occurred, equivalent Risk ratiob 0.95 0.89 0.7
to 21.0 deaths per 1,000 person-years. The deaths were 95% confidence 0.75, 1.20 0.68, 1.17
equally divided between the vitamin E and no-vitamin-E interval
groups: 1,801 vs. 1,770, corresponding to a risk ratio of Deathsc 142/139 106/110
1.02 (95% confidence interval (CI): 0.95, 1.09). 66–69 years
Dietary vitamin C intake did not modify the effect of (n ¼ 1,937)
vitamin E supplementation. For participants with a vitamin Risk ratiob 1.07 0.59 0.002
C intake of less than 90 mg/day, the effect of vitamin E on 95% confidence 0.84, 1.36 0.44, 0.79
mortality was a risk ratio of 1.00 (95% CI: 0.92, 1.11); for interval
those with a higher vitamin C intake, the effect was a risk Deathsc 139/137 71/124
ratio of 1.04 (95% CI: 0.94, 1.16). Age did not modify the Test for interaction; 0.4 0.0003
effect of vitamin E (P ¼ 0.06; interaction between vitamin E age as a continuous
and age as a continuous variable). variable (P value)d
We found that vitamin E supplementation had a second- a
Information on dietary vitamin C intake was missing for 2,022
order interaction with dietary vitamin C and age (Table 1). participants, with 177 deaths of vitamin-E and 175 deaths of no-
Vitamin E did not affect mortality for participants with low vitamin-E participants.
b
vitamin C intake. However, for participants with high dietary Proportional hazards regression model comparing participants
vitamin C intake, the effect of vitamin E depended on age, so who received vitamin E with those who did not.
c
that it increased mortality in the young (aged <63 years) Number of deaths of vitamin-E participants/number of deaths of
participants by 19% but reduced mortality in the old (aged no-vitamin-E participants.
d
66 years) participants by 41% (Table 1, Figures 1 and 2). The second-order interaction term between vitamin E supplemen-
tation, dietary vitamin C, and age improved the regression model by
Division of participants into the age categories shown in
v2(1 df) ¼ 10.1, P ¼ 0.0015. The uniformity of the vitamin E effect
Table 1 was based on the inspection of data, but the findings was also tested by adding a dummy variable for vitamin E effect in
were not sensitive to the cutpoints of age. For young partic- 5 groups of the table, allowing each of the 6 groups its own vitamin E
ipants with high vitamin C intake, the point estimate of the supplementation effect. The regression model was improved by
vitamin E effect was very similar in narrower age catego- v2(5 df) ¼ 22.2, P ¼ 0.0005 compared with the model with a uniform
ries: risk ratio ¼ 1.17 (95% CI: 0.93, 1.48; 297 deaths), risk vitamin E effect. Adding the vitamin E effect to only those groups aged
ratio ¼ 1.25 (95% CI: 1.02, 1.53; 371 deaths), and risk 50–62 and 66–69 years with high vitamin C intake led to similar im-
ratio ¼ 1.13 (95% CI: 0.92, 1.40; 353 deaths) in the groups provement in the regression model, by v2(2 df) ¼ 21.0 compared with
aged 50–54, 55–58, and 59–62 years, respectively, which the model with a uniform vitamin E effect.
justified combining these age groups. In participants aged
66–69 years with high dietary vitamin C intake, vitamin E
had the same effect in 2-year categories: risk ratio ¼ 0.58 The main food sources of vitamin C are fruit, vegetables,
(95% CI: 0.39, 0.86; 113 deaths) and risk ratio ¼ 0.58 (95% and berries. Thus, modification of the vitamin E effect by
CI: 0.37, 0.91; 82 deaths) in the groups aged 66–67 and dietary vitamin C might be explained by high levels of other
68–69 years, respectively. substances in these foods. Residual fruit, vegetables, and
There seemed to be an approximately 3-year lag period berry intake (refer to the Materials and Methods section)
before vitamin E started to increase the mortality of partic- did not modify the effect of vitamin E in participants aged
ipants aged 50–62 years with high vitamin C intake (Figure 1). 50–62 years (Table 2), indicating that other substances in
During the first 3 years, vitamin E had no effect on mor- these foods do not explain the effect modification by vitamin C.
tality, but thereafter it increased mortality by 38%. Inclu- The vitamin E effect was modified by vitamin C as a con-
sion of the lag period in the regression model led to tinuous variable, indicating that the cutpoint for dichotomi-
significant improvement in the model. The survival curves zation was not crucial to the finding. Dietary vitamin E
for participants aged 66–69 years suggested a 2-year lag intake and b-carotene supplementation did not modify the
period (Figure 2). effect of vitamin E supplementation. In the participants aged

Am J Epidemiol 2009;169:946–953
 Vitamin E Supplementation and Heterogeneity in Mortality 949

0.15 No Vitamin E 0.50 No Vitamin E

Vitamin E Vitamin E

0.40
Cumulative Hazard of Death

Cumulative Hazard of Death

0.10
0.30

0.20
0.05

0.10

0.00
0.00
0 2 4 6 8
0 2 4 6 8
Years of Follow-up
Years of Follow-up

Figure 1. Effect of vitamin E supplementation on mortality among

Figure 2. Effect of vitamin E supplementation on mortality among
participants aged 50–62 years with a dietary vitamin C intake of
participants aged 66–69 years with a dietary vitamin C intake of >90
>90 mg/day (n ¼ 11,448 with 1,021 deaths), Alpha-Tocopherol,
mg/day (n ¼ 872 with 195 deaths), Alpha-Tocopherol, Beta-Carotene
Beta-Carotene Cancer Prevention Study, 1985–1993. Nelson-Aalen
Cancer Prevention Study, 1985–1993. Nelson-Aalen cumulative haz-
cumulative hazard functions for the vitamin-E and no-vitamin-E
ard functions for the vitamin-E and no-vitamin-E groups are shown.
groups are shown. Each step indicates 1 death. For the difference
Each step indicates 1 death. For the difference between the 2 groups,
between the 2 groups, log-rank-test P ¼ 0.006. The number of par-
log-rank-test P ¼ 0.0003. The number of participants with follow-up
ticipants with follow-up time of 7 years was 2,316; the curves are cut
time of 7 years was 128; the curves are cut at 7.8 years because the
at 7.8 years because the number of participants declined abruptly
number of participants declined abruptly thereafter. The possibility of
thereafter. The possibility of a lag period was examined by adding
a lag period was examined by adding 2 different risk ratio terms for
a different risk ratio term for vitamin E effect starting at variable time
vitamin E effect starting at variable time points because the 2 curves
points. The best improvement in the regression model was achieved
diverge at the initiation of supplementation and at about 2 years. The
by adding the second vitamin E effect starting at 3.3 person-years,
best improvement in the regression model was achieved by adding
which improved the regression model by v2(1 df) ¼ 7.1, P ¼ 0.007.
the second vitamin E effect starting at 0.3 person-years and the third
This model gives risk ratios of 0.99 (95% confidence interval: 0.82,
risk ratio starting from 1.9 years, which improved the regression model
1.19) during the first 3.3 years and 1.38 (95% confidence interval:
by v2(2 df) ¼ 5.2, P ¼ 0.073. This model gives risk ratios of 0.15
1.17, 1.63) thereafter.
(95% confidence interval: 0.02, 1.2) during the first 0.3 years, 1.04
(95% confidence interval: 0.53, 2.04) during the period 0.3–1.9 years,
and 0.54 (95% confidence interval: 0.39, 0.76) thereafter. During the
first 0.3 years of follow-up, there were 5 deaths in the b-carotene arm,
66–69 years, residual fruit, vegetables, and berry intake; 2 deaths in the placebo arm, 1 death in the vitamin E arm, and no
deaths in the vitamin E þ b-carotene arm.
dietary vitamin E intake; and b-carotene supplementation
did not modify the effect of vitamin E (Table 3).
Smoking had a marginally significant modification effect on
the participants aged 66–69 years with high vitamin C intake
(P ¼ 0.051). The decrease in mortality with vitamin E
subgroup. On the other hand, the cumulative hazard esti-
supplementation was more evident for those who smoked
mates over the 8-year follow-up (Figure 2) indicate mor-
less than a pack of cigarettes per day (risk ratio (RR) ¼
tality of 28% in the vitamin-E group and 49% in the
0.44, 95% CI: 0.28, 0.68; 89 deaths) compared with
no-vitamin-E group, corresponding to the number needed
a pack or more (RR ¼ 0.78, 95% CI: 0.53, 1.15; 106
to treat of 4.8 over an 8-year supplementation period.
deaths). For the participants aged 50–62 years with high
vitamin C intake, the effect was more apparent for those
who smoked a pack of cigarettes or more per day (RR ¼ DISCUSSION
1.22, 95% CI: 1.05, 1.42; 689 deaths) than for those who
smoked less (RR ¼ 1.12, 95% CI: 0.90, 1.39; 332 deaths), Vitamin E supplementation had no overall effect on mor-
but the confidence intervals overlapped broadly and the tality among the ATBC Study participants, yet we found
difference was not statistically significant (P ¼ 0.5 in strong evidence of heterogeneity between subgroups.
heterogeneity test). Vitamin E increased, decreased, or had no effect on mortal-
Of the participants aged 66–69 years with high vitamin C ity depending on age and vitamin C intake. The numerical
intake, 16.7% of the vitamin-E participants died during estimates calculated for the subgroups are less essential than
follow-up, whereas 27.7% of the no-vitamin-E participants the strong evidence of heterogeneity. When the effect of
died. Thus, based on the baseline cohort, the number needed vitamin E depends on the characteristics of people, it seems
to treat to prevent one death during follow-up was 9.1 in this obvious that the estimates of intervention effect obtained

Am J Epidemiol 2009;169:946–953
950 Hemilä and Kaprio

Table 2. Specificity of Vitamin C in Modifying the Effect of Vitamin E Supplementation on the

Mortality of Participants Aged 50–62 Years, Alpha-Tocopherol, Beta-Carotene Cancer
Prevention Study, 1985–1993

Vitamin E No Vitamin E 95% Test of

No. of Risk
Subgroup No. of No. of a Confidence Interaction
Participants Rateb Rateb Ratio
Deaths Deaths Interval (P Value)

All 24,000 1,292 18.4 1,202 17.0 1.08 1.00, 1.17

Dietary vitamin Cc
<90 mg/day 10,965 614 19.2 616 19.1 1.00 0.90, 1.13 0.048d
90 mg/day 11,448 552 16.3 469 13.7 1.19 1.05, 1.35
Residual of fruit,
vegetables,
and berriese
<0 g/day 11,839 638 18.5 575 16.5 1.11 1.00, 1.25 0.5
0 g/day 10,574 528 16.9 510 16.1 1.05 0.93, 1.19
Dietary vitamin Ec
<10 mg/day 9,295 516 19.1 499 18.2 1.05 0.92, 1.19 0.5
10 mg/day 13,118 650 16.8 586 15.0 1.11 1.00, 1.25
b-Carotene
No 12,041 617 17.5 567 15.9 1.10 0.98, 1.23 0.7
Yes 11,959 675 19.3 635 18.1 1.06 0.95, 1.19
a
Proportional hazards regression model comparing participants who received vitamin E with
those who did not.
b
Number of deaths per 1,000 person-years.
c
Information on dietary vitamins C and E intake was missing for 1,587 participants, with 126
deaths of vitamin-E and 117 deaths of no-vitamin-E participants.
d
Dietary vitamin C as a continuous variable: test for vitamin E interaction, P ¼ 0.011.
e
Difference between an individual’s intake and the mean intake with a given dietary vitamin C
intake; refer to the Materials and Methods section of the text. Information on fruit, vegetables, and
berries intake was missing for 1,587 participants, with 126 deaths of vitamin-E and 117 deaths of
no-vitamin-E participants.

from one study cannot be confidently generalized to other
population groups.
Although we divided the participants into several sub-
groups, the multiple comparison problem did not seem to be
a concern in our study. First, the subgroup heterogeneity in the
6 groups defined by age and dietary vitamin C intake was
significant even when we allowed each of the 6 subgroups to
have its own vitamin E effect (Table 1). Second, our current
subgroup analyses tested our earlier subgroup findings for res-
piratory infection outcomes in the ATBC Study (11, 14–19).
Our findings have several important implications. Sub-
group analysis has been discouraged because it can lead to
false-positive findings due to the multiple comparison prob-
lem (23–25). It has even been argued that ‘‘believing that
a treatment effect exists in one stratum of patients, even
though no overall significant treatment effect exists, is
a common error’’ (24, p. 15). Furthermore, large trials are
set up after years of deliberation by experts, and it is as-
sumed that all relevant knowledge is therefore incorporated
into the study plans (25). However, biology is complex, and
it seems unlikely that all important biologic knowledge
could ever be taken into account properly when setting up
a pragmatic controlled trial. A single estimate of treatment
effect calculated for a large population can be misleading if
it is thought to be valid for all participants, as shown in our
work. The overall estimate for all ATBC Study participants,
risk ratio ¼ 1.02, is inconsistent with our subgroup findings
for nearly half of all participants, that is, the young and old
with high dietary vitamin C intake.
Our current study and earlier subgroup analyses of the
ATBC Study suggest that subgroup analyses of large trial
databases should be encouraged even if there is no overall
effect in the study population, keeping in mind the possibil-
ity of spurious findings from multiple testing. Given the
long-term commitment of participants and the resources
invested, it might even be considered a moral imperative
to analyze the large trial databases as extensively as possible
rather than simply to calculate an overall effect. Evidently,
subgroup findings should be considered cautiously, and the
interpretation of P values must be related to the number of
subgroup analyses being carried out. Nevertheless, our sub-
group findings in the ATBC Study point out the need for
further research on vitamin E, whereas the overall estimate
of effect suggests that no further studies would have been
worthwhile. We concur with Feinstein’s concern (26) that
anti-subgroup doctrines have become so entrenched that
they often hamper investigation of socially and biologically
sound subgroups with public health relevance.

Am J Epidemiol 2009;169:946–953
 Vitamin E Supplementation and Heterogeneity in Mortality 951

Table 3. Specificity of Vitamin C in Modifying the Effect of Vitamin E Supplementation on the

Mortality of Participants Aged 66–69 Years, Alpha-Tocopherol, Beta-Carotene Cancer
Prevention Study, 1985–1993

Vitamin E No Vitamin E 95% Test of

No. of Risk
Subgroup No. of No. of a Confidence Interaction
Participants Rateb Rateb Ratio
Deaths Deaths Interval (P Value)

All 2,140 237 41.6 291 48.6 0.87 0.73, 1.03

Dietary vitamin Cc
<90 mg/day 1,065 139 49.6 137 46.4 1.07 0.84, 1.36 0.002d
90 mg/day 872 71 29.6 124 50.8 0.59 0.44, 0.79
Residual of fruit,
vegetables,
and berriese
<0 g/day 992 111 41.3 127 47.9 0.88 0.68, 1.13 0.7
0 g/day 945 99 39.3 134 48.9 0.81 0.62, 1.05
Dietary vitamin Ec
<10 mg/day 1,057 123 43.8 144 49.7 0.89 0.69, 1.13 0.5
10 mg/day 880 87 36.4 117 46.9 0.78 0.59, 1.03
b-Carotene
No 1,070 122 41.5 143 49.6 0.85 0.66, 1.08 0.8
Yes 1,070 115 41.7 148 47.6 0.89 0.69, 1.13
a
Proportional hazards regression model comparing participants who received vitamin E with
those who did not.
b
Number of deaths per 1,000 person-years.
c
Information on dietary vitamins C and E intake was missing for 203 participants, with 27
deaths of vitamin-E and 30 deaths of no-vitamin-E participants.
d
Dietary vitamin C as a continuous variable: test for vitamin E interaction, P ¼ 0.002.
e
Difference between an individual’s intake and the mean intake with a given dietary vitamin C
intake; refer to the Materials and Methods section of the text. Information on fruit, vegetables, and
berries intake was missing for 203 participants, with 27 deaths of vitamin-E and 30 deaths of no-
vitamin-E participants.

Three meta-analyses of controlled trials found no effect of
vitamin E supplementation on mortality (4–6). Combining
several large trials leads to estimates with narrow confidence
intervals; however, pooling is based on the assumption of the
same universal effect in all populations of the combined
studies. Our strong evidence of heterogeneity in the ATBC
Study refutes the assumption of a uniform effect and casts
doubt on the validity of the pooled estimates of the meta-
analyses. Although the pooled estimates reject the proposal
that vitamin E supplementation would be beneficial for wide-
ranging population groups, our study suggests that important
subgroups being harmed or benefiting from vitamin E may be
hidden in the misleadingly narrow confidence intervals.
The recently published Physicians’ Health Study II found
no overall effect of vitamin E and C supplementation on
mortality (27). However, the lack of average effect does
not conflict with our findings of heterogeneity because vi-
tamin E had no overall effect in the ATBC Study either. The
number of deaths in the ATBC Study (n ¼ 3,571) was twice
that in the Physicians’ Health Study II (n ¼ 1,661); conse-
quently, there is more statistical power to analyze potential
subgroup differences in the ATBC Study. Furthermore,
ATBC Study participants were recruited from the general
population, so there is greater potential for analyzing het-
erogeneity compared with the Physicians’ Health Study II
focused on a single upper-class profession.
Our strongest findings concern vitamin E supplementation
because the division of participants into vitamin-E and no-
vitamin-E groups was random. The evidence of heterogeneity
of the vitamin E effect is strong irrespective of the specificity
of vitamin C as the modifier. Nevertheless, by using the re-
sidual of fruit, vegetables, and berries, which has no corre-
lation with dietary vitamin C, we were able to show that
modification of the vitamin E effect is not explained by other
substances in these foods, suggesting that vitamin C is the
specific substance explaining the modification. Furthermore,
the subgroup divisions in our current study were based on the
effects of vitamin E on respiratory infections, and vitamin C
has also affected respiratory infections in certain controlled
trials, showing that its physiologic effects are not limited to
preventing scurvy (28, 29). Finally, the interaction between
vitamins E and C is well established (7–11). These argu-
ments support the possibility that vitamin C specifically
causes the modification of the vitamin E effect, although our
vitamin C intake levels were based on observational data.
All ATBC Study participants were smokers, and no direct
conclusions can be drawn for nonsmokers. The benefit of
vitamin E supplementation on respiratory infections was

Am J Epidemiol 2009;169:946–953
952 Hemilä and Kaprio
previously more evident in ATBC Study participants who
smoked less, and the harm was more apparent in those who
smoked more (14, 15, 18, 19). We saw similar trends in the
effect of vitamin E on mortality, but the modification of the
vitamin E effect caused by smoking quantity was not statis-
tically significant. Nevertheless, the greater reduction in mor-
tality among old participants with high vitamin C intake who
smoked less suggests that vitamin E supplementation might
affect mortality among older male nonsmokers as well.
Since vitamin E is a fat-soluble substance and it may take
months before tissue levels are substantially elevated (30,
31), short trials may be uninformative. In the current study,
there seemed to be 2- and 3-year lag periods before vitamin
E started to decrease or increase mortality (Figures 1 and 2).
These lag periods are consistent with a delay in the effects of
the fat-soluble vitamin. Nevertheless, the maximum of
8 years of follow-up in the ATBC Study was long enough
to observe that vitamin E supplementation does have effects.
On the other hand, the risk of tuberculosis increased signif-
icantly within a year of supplementation being initiated (14)
and the risk of pneumonia in participants who exercised in
their leisure time decreased without a lag (16), so not all
effects of vitamin E are delayed.
It has been suggested that the vitamin E doses in the ran-
domized trials were too low to show any effect (32), and high
doses in some trials increased mortality (5). Given our ob-
servations that 50 mg/day of vitamin E caused significant
increase and decrease in mortality in the ATBC Study pop-
ulation, depending on the characteristics of participants, no
justification exists for claiming that the vitamin E doses in
randomized trials have been too low. Our study also suggests
that it may be primarily subject characteristics and not dose
of vitamin E that determines whether vitamin E causes harm.
Many people take vitamin E supplements because they
believe that the vitamin protects them against oxidative
stress, which has a role in the aging processes (2, 3). Our
findings among the older ATBC Study participants are con-
sistent with the concept that the antioxidant vitamins C and
E can counteract oxidative stress processes in old people
under some conditions. Assuming that the benefit of vitamin
E supplementation in old people is conditional on a high
intake of vitamin C, it seems possible that old people with
low dietary vitamin C might benefit from a combination of
the 2. Therefore, the most informative type of further study
would be a factorial design with vitamins E and C on old
people with low dietary vitamin C intake.
Finally, the US nutritional recommendations and a recent
review consider that vitamin E is safe in doses up to 1,000
mg/day (33, 34). Our finding that 50 mg/day of vitamin E
significantly increased mortality among half of the male
smokers in the ATBC Study aged 50–62 years contradicts
the universal safety of 1,000 mg/day of vitamin E. Previously,
we found that 50 mg/day of vitamin E significantly increased
the risk of the common cold, pneumonia, and tuberculosis in
subgroups of the ATBC Study, also indicating that such a low
dose can cause harmful effects in some people (14, 18, 19).
In conclusion, our subgroup analyses of the ATBC Study
cohort support the conclusions of meta-analysis of con-
trolled trials (4–6), in that vitamin E supplementation seems
ineffective or harmful for middle-aged male smokers. Evi-
dently, in people younger than age 65 years, taking vitamin
E supplements should be strongly discouraged until clear
evidence emerges that some population groups of younger
or middle-aged people benefit. On the other hand, our study
indicates that vitamin E supplementation may lead to ben-
eficial effects in some subgroups of old people, and this
possibility should be investigated by using a factorial design
with vitamin C supplementation. Finally, the substantial de-
crease in mortality with vitamin E supplementation among
the older participants with high dietary vitamin C intake
raises the question of whether the decrease in overall mor-
tality is attributable to a single cause of death or a few
causes, or whether it suggests a general decrease in frailty
reflected in lessened mortality from diverse causes.
ACKNOWLEDGMENTS
Authors affiliation: Department of Public Health, Univer-
sity of Helsinki, Helsinki, Finland (Harri Hemilä, Jaakko
Kaprio).
The authors thank the ATBC Study (the National Public
Health Institute, Finland, and the National Cancer Institute,
United States) for access to the data.
H. H. planned the study and wrote the draft of the manu-
script. J. K. participated in planning the analyses and the
critical revision of the manuscript. H. H. had full access to
all data in this study and takes responsibility for the integrity
of the data and the accuracy of the data analysis.
Conflict of interest: none declared.
REFERENCES
1. Radimer K, Bindewald B, Hughes J, et al. Dietary supplement
use by US adults: data from the National Health and Nutrition
Examination Survey, 1999–2000. Am J Epidemiol. 2004;
160(4):339–349.
2. Beckman KB, Ames BN. The free radical theory of aging
matures. Physiol Rev. 1998;78(2):547–581.
3. Finkel T, Holbrook NJ. Oxidants, oxidative stress and the bi-
ology of ageing. Nature. 2000;408(6809):239–247.
4. Vivekananthan DP, Penn MS, Sapp SK, et al. Use of antioxi-
dant vitamins for the prevention of cardiovascular disease:
meta-analysis of randomised trials. Lancet. 2003;361(9374):
2017–2023.
5. Miller ER, Pastor-Barriuso R, Dalal D, et al. Meta-analysis:
high-dosage vitamin E supplementation may increase
all-cause mortality. Ann Intern Med. 2005;142(1):37–46.
(Comments: Ann Intern Med. 2005;143(2):150–158).
6. Bjelakovic G, Nikolova D, Gluud LL, et al. Mortality in ran-
domized trials of antioxidant supplements for primary and
secondary prevention: systematic review and meta-analysis.
JAMA. 2007;297(8):842–857. (Comments: JAMA. 2007;
298(4):400–403).
7. Packer JE, Slater TF, Willson RL. Direct observation of a free
radical interaction between vitamin E and vitamin C. Nature.
1979;278(5706):737–738.
8. Tanaka K, Hashimoto T, Tokumaru S, et al. Interactions be-
tween vitamin C and vitamin E are observed in tissues of in-
herently scorbutic rats. J Nutr. 1997;127(10):2060–2064.
Am J Epidemiol 2009;169:946–953
 Vitamin E Supplementation and Heterogeneity in Mortality
9. Hill KE, Montine TJ, Motley AK, et al. Combined deficiency
of vitamins E and C causes paralysis and death in guinea pigs.
Am J Clin Nutr. 2003;77(6):1484–1488.
10. Bruno RS, Leonard SW, Atkinson J, et al. Faster plasma
vitamin E disappearance in smokers is normalized by vitamin C
supplementation. Free Radic Biol Med. 2006;40(4):689–697.
(Comments: Free Radic Biol Med. 2007;42(4):578–580).
11. Hemilä H. Do Vitamins C and E Affect Respiratory Infections?
[dissertation]. Helsinki, Finland: University of Helsinki;
2006:10–11, 56–57. (Available at http://ethesis.helsinki.fi/
julkaisut/laa/kansa/vk/hemila/). (Accessed December 15, 2008).
12. The effect of vitamin E and beta-carotene on the incidence of
lung cancer and other cancers in male smokers. The Alpha-
Tocopherol, Beta-Carotene Cancer Prevention Study Group.
N Engl J Med. 1994;330(15):1029–1035.
13. The alpha-tocopherol, beta-carotene lung cancer prevention
study: design, methods, participant characteristics, and com-
pliance. The ATBC Cancer Prevention Study Group. Ann
Epidemiol. 1994;4(1):1–10.
14. Hemilä H, Kaprio J. Vitamin E supplementation may tran-
siently increase tuberculosis risk in males who smoke heavily
and have high dietary vitamin C intake. Br J Nutr. 2008;
100(4):896–902. (Comments: Br J Nutr. 2009;101(1):145–147).
15. Hemilä H, Virtamo J, Albanes D, et al. Vitamin E and beta-
carotene supplementation and hospital-treated pneumonia in-
cidence in male smokers. Chest. 2004;125(2):557–565.
16. Hemilä H, Kaprio J, Albanes D, et al. Physical activity and the
risk of pneumonia in male smokers administered vitamin E
and b-carotene. Int J Sports Med. 2006;27(4):336–341.
17. Hemilä H, Kaprio J, Albanes D, et al. Vitamin C, vitamin E,
and beta-carotene in relation to common cold incidence in
male smokers. Epidemiology. 2002;13(1):32–37.
18. Hemilä H, Virtamo J, Albanes D, et al. The effect of vitamin E on
common cold incidence is modified by age, smoking and resi-
dential neighborhood. J Am Coll Nutr. 2006;25(4):332–339.
19. Hemilä H, Kaprio J. Vitamin E supplementation and pneu-
monia risk in males who initiated smoking at an early age:
effect modification by body weight and vitamin C [electronic
article]. Nutr J. 2008;7(1):33.
20. Pietinen P, Hartman AM, Haapa E, et al. Reproducibility
and validity of dietary assessment instruments. I. A self-
administered food use questionnaire with a portion size picture
booklet. Am J Epidemiol. 1988;128(3):655–666.
21. Ovaskainen ML, Valsta M, Lauronen J. The compilation of
food analysis values as a database for dietary studies—the
Finnish experience. Food Chem. 1996;57(1):133–136.
Am J Epidemiol 2009;169:946–953
953
22. Hemilä H, Kaprio J, Pietinen P, et al. Vitamin C and other
compounds in vitamin C rich food in relation to risk of tu-
berculosis in male smokers. Am J Epidemiol. 1999;150(6):
632–641.
23. Assmann SF, Pocock SJ, Enos LE, et al. Subgroup analysis and
other (mis)uses of baseline data in clinical trials. Lancet.
2000;355(9209):1064–1069.
24. Peto R, Pike MC, Armitage P, et al. Design and analysis of
randomized clinical trials requiring prolonged observation of
each patient. II. Analysis and examples. Br J Cancer. 1977;
35(1):1–39.
25. Vandenbroucke JP. Observational research, randomised trials,
and two views of medical science [electronic article]. PLoS
Med. 2008;5(3):e67.
26. Feinstein AR. The problem of cogent subgroups: a clinico-
statistical tragedy. J Clin Epidemiol. 1998;51(4):297–299.
27. Sesso HD, Buring JE, Christen WG, et al. Vitamins E and C in
the prevention of cardiovascular disease in men: the Physi-
cians’ Health Study II randomized controlled trial. JAMA.
2008;300(18):2123–2133.
28. Douglas RM, Hemilä H. Vitamin C for preventing and treating
the common cold [electronic article]. PLoS Med. 2005;2(6):
e168.
29. Hemilä H, Louhiala P. Vitamin C may affect lung infections.
J R Soc Med. 2007;100(11):495–498.
30. Kitagawa M, Mino M. Effects of elevated d-alpha(RRR)-
tocopherol dosage in man. J Nutr Sci Vitaminol (Tokyo). 1989;
35(2):133–142.
31. Handelman GJ, Epstein WL, Peerson J, et al. Human adipose
alpha-tocopherol and gamma-tocopherol kinetics during and
after 1 y of alpha-tocopherol supplementation. Am J Clin Nutr.
1994;59(5):1025–1032.
32. Blumberg JB, Frei B. Why clinical trials of vitamin E and
cardiovascular diseases may be fatally flawed. Commentary on
‘‘The relationship between dose of vitamin E and suppression
of oxidative stress in humans’’. Free Radic Biol Med. 2007;
43(10):1374–1376.
33. Food and Nutrition Board, Institute of Medicine. Dietary
Reference Intakes for Vitamin C, Vitamin E, Selenium, and
Carotenoids. Washington, DC: The National Academies Press;
2000: 249–260. (Available at http://www.nap.edu/books/
0309069351/html/249.html). (Accessed December 15, 2008).
34. Hathcock JN, Azzi A, Blumberg J, et al. Vitamins E and C are
safe across a broad range of intakes. Am J Clin Nutr. 2005;
81(4):736–745. (Comments: Am J Clin Nutr. 2005;82(5):
1141–1143).
 American Journal of Epidemiology Vol. 169, No. 8
ª The Author 2009. Published by the Johns Hopkins Bloomberg School of Public Health. DOI: 10.1093/aje/kwn421
All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org. Advance Access publication February 18, 2009

Original Contribution

Dietary Acrylamide Intake and Risk of Premenopausal Breast Cancer

Kathryn M. Wilson, Lorelei A. Mucci, Eunyoung Cho, David J. Hunter, Wendy Y. Chen, and Walter
C. Willett

Initially submitted July 29, 2008; accepted for publication December 22, 2008.

Acrylamide, a probable human carcinogen, is formed during high-temperature cooking of many commonly
consumed foods. It is widespread; approximately 30% of calories consumed in the United States are from foods
containing acrylamide. In animal studies, acrylamide causes mammary tumors, but it is unknown whether the level
of acrylamide in foods affects human breast cancer risk. The authors studied the association between acrylamide
intake and breast cancer risk among 90,628 premenopausal women in the Nurses’ Health Study II. They calculated
acrylamide intake from food frequency questionnaires in 1991, 1995, 1999, and 2003. From 1991 through 2005,
they documented 1,179 cases of invasive breast cancer. They used Cox proportional hazards models to assess
the association between acrylamide and breast cancer risk. The multivariable-adjusted relative risk of premeno-
pausal breast cancer was 0.92 (95% confidence interval: 0.76, 1.11) for the highest versus the lowest quintile of
acrylamide intake (Ptrend ¼ 0.61). Results were similar regardless of smoking status or estrogen and progesterone
receptor status of the tumors. The authors found no associations between intakes of foods high in acrylamide,
including French fries, coffee, cereal, potato chips, potatoes, and baked goods, and breast cancer risk. They found
no evidence that acrylamide intake, within the range of US diets, is associated with increased risk of premeno-
pausal breast cancer.

acrylamide; breast neoplasms; diet

Abbreviations: CI, confidence interval; ER, estrogen receptor; FFQ, food frequency questionnaire; PR, progesterone receptor; RR,
relative risk.

Acrylamide is classified as a probable human carcino-
gen (1), and it is formed during high-temperature pro-
cessing of many commonly consumed foods (2). The
discovery of acrylamide in foods in 2002 caused consider-
able concern worldwide. Acrylamide is widespread in the
food supply, with approximately 38% of calories con-
sumed in the United States coming from foods that
contain acrylamide (3). Potatoes, cold breakfast cereal,
coffee, and baked goods are major sources of acrylamide
intake in the United States (4). Prior to the discovery of
acrylamide in foods, industrial use and tobacco use were
thought to be the major sources of acrylamide exposure in
humans (1).
In animal tests, acrylamide administered in high levels in
drinking water causes several types of hormone-sensitive
cancers, including mammary tumors in female rats (5, 6).
Given the burden of breast cancer, it is of interest to study
the association between acrylamide intake in humans and
the risk of breast cancer. Epidemiologic studies have had
mixed results. Two prospective studies of dietary acrylam-
ide exposure in humans found no association with pre- or
postmenopausal breast cancer risk (7, 8). Both of these
reports used food frequency questionnaires (FFQs) to as-
sess acrylamide intake. A prospective study in the Danish
Diet, Cancer, and Health Cohort used a biomarker of ac-
rylamide exposure, acrylamide adducts to hemoglobin, and
found an increased risk of breast cancer among postmen-
opausal women with higher adducts. The increased risk
appeared limited to smokers and to estrogen receptor-positive
cancers (9).

Correspondence to Dr. Kathryn M. Wilson, Channing Laboratory, Brigham and Women’s Hospital, 181 Longwood Avenue, Third Floor, Boston,
MA 02115 (e-mail: kwilson@hsph.harvard.edu).

954 Am J Epidemiol 2009;169:954–961


We used data from the Nurses’ Health Study II to assess
the association between acrylamide intake and premeno-
pausal breast cancer risk. This cohort has repeated measures
of diet, which allows us to study acrylamide intake over an
extended time. We previously reported on the creation of an
acrylamide food composition database for this cohort (10);
we found a moderate association between calculated acryl-
amide intake and hemoglobin adducts of acrylamide in
a subset of the cohort.
MATERIALS AND METHODS
Study population
The Nurses’ Health Study II is a prospective cohort
study of 116,671 female registered nurses aged 25–42
years at the start of the study in 1989. Follow-up ques-
tionnaires have been sent biennially to update information
on lifestyle and health. Beginning in 1991, and every
4 years thereafter, a semiquantitative FFQ was sent to
participants to assess their usual dietary intake over the
previous year. Women who completed the first FFQ in
1991 (n ¼ 97,807) form the study population for this
analysis.
We excluded women who had an implausible energy in-
take (<800 or >4,200 kcal/day) or who left more than 70
food items blank (n ¼ 2,361). We also excluded women
who reported a diagnosis of cancer (excluding nonmela-
noma skin cancer) before baseline in 1991 (n ¼ 1,308).
The analysis was limited to premenopausal women, so
women who were postmenopausal at baseline were ex-
cluded (n ¼ 3,462), and women were censored after they
reached natural or surgical menopause. Women who had
a hysterectomy without a bilateral oophorectomy were ex-
cluded (n ¼ 48) or censored at the time of surgery because
their menopausal status was unknown. This left a total of
90,628 premenopausal women with baseline diet informa-
tion for the analysis. The response rate was approximately
90% among these women through the end of follow-up on
June 1, 2005. This study was approved by the human re-
search committees at the Harvard School of Public Health
and Brigham and Women’s Hospital.
Assessment of acrylamide intake
FFQs with over 130 food items were completed in 1991,
1995, 1999, and 2003. Participants were asked how fre-
quently they had consumed a specified portion size of each
item over the previous year with 9 possible responses,
ranging from never or less than once a month to 6 or more
times per day. The FFQ includes the major acrylamide-
contributing foods according to US Food and Drug Admin-
istration surveys: French fries, cold breakfast cereal, potato
chips, cookies, coffee, breads, baked goods, and snack
foods (4).
We previously reported on the creation and validation of
an acrylamide food composition database for the FFQ (10).
Briefly, 42 food items on the FFQ were assigned acrylam-
ide contents based on published data from the US Food and
Drug Administration and additional analyses of US foods
Am J Epidemiol 2009;169:954–961
Acrylamide and Premenopausal Breast Cancer 955
by the Swedish National Food Administration. We calcu-
lated daily acrylamide intake for each participant by mul-
tiplying the acrylamide content of 1 serving of food by the
frequency of consumption of that food and summing across
all food items on the questionnaire. Acrylamide intake
from cold breakfast cereal was based on participants’ re-
porting of which brand they use most often. The correla-
tion between 1999 acrylamide intake and a biomarker of
acrylamide exposure, the sum of hemoglobin adducts
of acrylamide and its metabolite glycidamide, was 0.34
(P < 0.0001) among 296 nonsmoking women from the
Nurses’ Health Study II cohort. The accuracy of reporting
for individual food items on a similar FFQ was measured
by comparing FFQ responses and 28 days of diet records in
a subset of women in the Nurses’ Health Study (11). The
correlation between FFQ and diet records for the top
acrylamide-contributing foods was 0.73 for French fries,
0.78 for coffee, 0.60 for potato chips, and 0.79 for cold
breakfast cereal.
Because acrylamide may have an effect on carcinogenesis
over an extended period of time, we used the cumulative
average intake of acrylamide to represent long-term dietary
intake. That is, 1991 intake was used for the 1991–1995
follow-up period, the average of 1991 and 1995 intakes
was used for the 1995–1999 follow-up period, the average
of 1991, 1995, and 1999 intakes was used for the 1999–2003
follow-up period, and the average of all 4 questionnaires
was used for the 2003–2005 follow-up period. Data from
the previous FFQ were carried forward to the next time
period for participants with incomplete FFQ information
after baseline. In secondary analyses, we examined the as-
sociation between baseline acrylamide intake and breast
cancer risk.
Ascertainment of breast cancer cases
Biennial follow-up questionnaires were used to identify
newly diagnosed cases of breast cancer. Deaths were
documented by responses to questionnaires by family
members, by the postal service, or through the National
Death Index. Cause of death was confirmed by medical
record review, information from relatives, or review of
death certificates.
When participants reported breast cancer, we asked the
participant for confirmation of the diagnosis and permis-
sion to obtain relevant medical records. Pathology reports
confirmed 98% of the self-reported breast cancers. Infor-
mation on estrogen and progesterone receptor status was
obtained from pathology reports and was available for
78% of cases. A recent validation study in the Nurses’
Health Study I cohort demonstrated that pathology reports
provide accurate information on estrogen receptor status
(12). Cases of carcinoma-in-situ were not included in the
analysis.
Statistical analysis
Each participant contributed person-time from the date of
return of the 1991 questionnaire until the time of breast can-
cer diagnosis, menopause, death, or June 1, 2005, whichever
956 Wilson et al.

Table 1. Age-standardized Characteristics of the Nurses’ Health Study II Cohort in 1991a

Calorie-adjusted Acrylamide Intake

Quintile 1, Low Quintile 2 Quintile 3 Quintile 4 Quintile 5, High
(n 5 20,934) (n 5 17,416) (n 5 16,768) (n 5 16,331) (n 5 19,179)
Acrylamide intake, lg/day 10.8 16.6 20.2 24.6 37.8
Acrylamide by body weight, 0.17 0.26 0.32 0.38 0.58
lg/kg/day
Age, years 36 36 36 36 36
2
Body mass index, kg/m 25 25 24 24 25
Current smokers, % 9 10 11 13 17
Physical activity, METs/week 24 22 20 20 17
Age at menarche <12 years, % 25 24 24 24 25
Nulliparous, % 32 28 26 26 27
Current oral contraceptive users, % 11 10 11 11 11
Family history of breast cancer, % 6 6 6 6 6
History of benign breast disease, % 33 32 33 34 33
Nutrient intakes
Energy intake, kcal/day 1,796 1,854 1,805 1,724 1,772
Alcohol, g/day 3.0 3.1 3.3 3.3 2.9
Animal fat, g/dayb 35 35 35 35 35
b,c
Glycemic load 123 122 121 120 120
Intakes of high acrylamide
foods, servings/day
French fries 0.03 0.1 0.1 0.1 0.2
Coffee 0.7 1.2 1.6 2.1 2.3
Breakfast cereal 0.3 0.4 0.4 0.4 0.4
Potato chips 0.1 0.1 0.2 0.2 0.3
Potatoes (baked, roasted, mashed) 0.3 0.3 0.3 0.3 0.3

Abbreviation: MET, metabolic equivalent.

a
All data (except for mean age) are standardized to the age distribution of the cohort in 1991. Means or percent-
ages are shown.
b
Animal fat and glycemic load are adjusted for total energy intake.
c
Each unit of dietary glycemic load represents the glycemic equivalent of 1 g of carbohydrate from white bread.
Intakes shown are per day.

came first. Participants were divided into quintiles based on
their acrylamide intake and their consumption of acrylamide-
rich foods. Acrylamide intake was adjusted for total energy
intake by using the residual method. Relative risks of breast
cancer were calculated as the incidence rate for a given
quintile of consumption divided by the rate in the lowest
quintile.
We used Cox proportional hazards regression to adjust
for potential confounding by other breast cancer risk fac-
tors. To control as finely as possible for confounding by
age, calendar time, and any possible 2-way interactions
between these 2 time scales, we stratified the analysis
jointly by age in months at the start of each follow-up
period and calendar year of the current questionnaire cycle.
We used multivariable models to adjust for the following
factors: body mass index (<18.5, 18.5–19.9, 20.0–22.4,
22.5–24.9, 25.0–29.9, and
30 kg/m2), height (<62,
62–<65, 65–<68, and
68 inches; 1 inch ¼ 2.54 cm), oral
contraceptive use (never, former use <4 years, former use

4 years, current use <8 years, and current use
8 years),
parity and age at first birth (nulliparous, 1–2 children and
age at first birth <25 years, 1–2 children and age at first
birth 25–<30 years, 1–2 children and age at first birth
30
years, 3 or more children and age at first birth <25 years,
3 or more children and age at first birth
25 years), age at
menarche (<12, 12, 13, or
14 years), family history of
breast cancer (yes/no), history of benign breast disease
(yes/no), smoking (never, former smoker of <25 ciga-
rettes/day, former smoker of
25 cigarettes/day, current
smoker of <25 cigarettes/day, and current smoker of
25
cigarettes/day), physical activity (18 and >18 metabolic
equivalent (MET)-hours/week), animal fat (quintiles), glyc-
emic load (quintiles), alcohol intake (continuous g/day), and
total energy intake (continuous kcal/day). We adjusted for
animal fat and glycemic load as they have previously been
associated with breast cancer risk in this cohort (13, 14).
We also considered adjustment for quintile of vegetable
fat intake, trans fat intake, and glycemic index, as these
dietary factors were most correlated with acrylamide in-
take, but they were not included in final models because

Am J Epidemiol 2009;169:954–961
 Acrylamide and Premenopausal Breast Cancer 957

Table 2. Relative Risk (95% Confidence Intervals) of Breast Cancer by Quintile of Calorie-adjusted Acrylamide Intake, Nurses’ Health Study II,
1991–2005

Calorie-adjusted Acrylamide Intakea

Quintile 1, Low Quintile 2 Quintile 3 Quintile 4 Quintile 5, High Ptrendb
(12 mg/day) (17 mg/day) (20 mg/day) (24 mg/day) (33 mg/day)

All premenopausal breast cancer

No. of cases 237 236 232 264 210
Age-adjusted relative risk 1.00 0.96 (0.80, 1.15) 0.95 (0.79, 1.14) 1.04 (0.87, 1.24) 0.92 (0.76, 1.10) 0.58
Multivariable relative riskc 1.00 0.95 (0.79, 1.14) 0.94 (0.78, 1.13) 1.03 (0.87, 1.24) 0.92 (0.76, 1.11) 0.61
By smoking status
Never smokers
No. of cases 165 149 148 165 111
Age-adjusted relative risk 1.00 0.91 (0.72, 1.13) 0.93 (0.75, 1.17) 1.06 (0.85, 1.32) 0.81 (0.64, 1.04) 0.28
Multivariable relative riskc 1.00 0.91 (0.73, 1.14) 0.94 (0.75, 1.18) 1.08 (0.86, 1.34) 0.82 (0.64, 1.05) 0.33
Former smokers
No. of cases 56 64 63 74 68
Age-adjusted relative risk 1.00 0.98 (0.68, 1.41) 0.91 (0.63, 1.32) 0.99 (0.69, 1.40) 1.05 (0.73, 1.50) 0.70
Multivariable relative riskc 1.00 1.00 (0.69, 1.44) 0.92 (0.64, 1.34) 1.01 (0.70, 1.43) 1.09 (0.75, 1.56) 0.57
Current smokers
No. of cases 16 23 23 25 31
Age-adjusted relative risk 1.00 1.16 (0.60, 2.25) 1.14 (0.59, 2.21) 0.88 (0.46, 1.69) 0.97 (0.52, 1.81) 0.61
Multivariable relative riskc 1.00 1.09 (0.55, 2.17) 1.16 (0.58, 2.30) 0.82 (0.41, 1.62) 1.05 (0.55, 2.02) 0.89
By ER and PR status
ERþ/PRþ breast cancer
No. of cases 105 129 111 138 114
Age-adjusted relative risk 1.00 1.16 (0.90, 1.50) 1.00 (0.77, 1.31) 1.19 (0.93, 1.54) 1.13 (0.87, 1.48) 0.38
Multivariable relative riskc 1.00 1.14 (0.88, 1.48) 0.98 (0.75, 1.28) 1.16 (0.90, 1.50) 1.11 (0.85, 1.46) 0.45
ER/PR breast cancer
No. of cases 39 45 34 43 35
Age-adjusted relative risk 1.00 1.10 (0.72, 1.70) 0.86 (0.54, 1.36) 1.06 (0.69, 1.65) 0.93 (0.59, 1.47) 0.73
Multivariable relative riskc 1.00 1.09 (0.70, 1.68) 0.85 (0.53, 1.35) 1.04 (0.67, 1.62) 0.90 (0.57, 1.43) 0.62

Abbreviations: ER, estrogen receptor; MET, metabolic equivalent; PR, progesterone receptor; þ, positive; , negative.
a
Median intake.
b
Test for trend calculated by using the median intake in each quintile as a continuous variable.
c
Multivariable models are stratified by age in months and calendar year and adjusted for the following: body mass index (<18.5, 18.5–19.9,
20.0–22.4, 22.5–24.9, 25.0–29.9, and
30 kg/m2), height (<62, 62–<65, 65–<68, and
68 inches; 1 inch ¼ 2.54 cm), oral contraceptive use
(never, former use <4 years, former use
4 years, current use <8 years, and current use
8 years), parity and age at first birth (nulliparous, parity
1–2 and age at first birth <25 years, parity 1–2 and age at first birth 25–<30 years, parity 1–2 and age at first birth
30 years, parity
3 and age at
first birth <25 years, parity
3 and age at first birth
25 years), age at menarche (<12, 12, 13, or
14 years), family history of breast cancer (yes/
no), history of benign breast disease (yes/no), smoking (never, former smoker <25 cigarettes/day, former smoker
25 cigarettes/day, current
smoker <25 cigarettes/day, and current smoker
25 cigarettes/day), physical activity (18 and >18 MET-hours/week), animal fat (quintiles),
glycemic load (quintiles), alcohol intake (continuous), and total energy intake (continuous).

they had no substantial effect on the relative risk or stan-
dard error estimates for acrylamide. All covariates except
height and age at menarche were updated in each question-
naire cycle. The SAS Proc PHREG procedure (SAS Insti-
tute, Inc., Cary, North Carolina) was used for all analyses,
and the Anderson-Gill data structure was used to handle
time-varying covariates efficiently. To test for a linear
trend across quintiles of intake, we modeled acrylamide
intake as a continuous variable using the median value for
each quintile.
We examined whether the association between acryl-
amide intake and breast cancer risk was modified by in-
dividual characteristics, including age, smoking status,
body mass index, alcohol intake, and glycemic load, by
modeling the association separately in each group. We
tested the significance of interactions by adding cross-
product terms between acrylamide intake and the variable
of interest to the model and comparing this model to the
model without the cross-product term using the likelihood
ratio test.

Am J Epidemiol 2009;169:954–961
958 Wilson et al.

Table 3. Relative Risk (95% Confidence Intervals) of Breast Cancer by Intake of High-Acrylamide Foods, Nurses’ Health Study II, 1991–2005

Intake of High-Acrylamide Foods

Ptrenda
Quintile 1, Low Quintile 2 Quintile 3 Quintile 4 Quintile 5, High

French fries
Median intake, servings/week 0 0.2 0.5 0.7 1.0
No. of cases 255 211 255 195 263
b
Multivariable relative risk 1.00 1.08 (0.89, 1.31) 0.93 (0.77, 1.11) 0.93 (0.76, 1.13) 0.96 (0.80, 1.16) 0.35
Coffee
Median intake, servings/day 0 0.2 1 2.5 3.5
No. of cases 270 155 230 266 258
b
Multivariable relative risk 1.00 1.11 (0.91, 1.36) 0.97 (0.81, 1.16) 1.01 (0.85, 1.21) 0.92 (0.77, 1.11) 0.28
Breakfast cereal
Median intake, servings/week 0 0.7 2.0 3.0 6.0
No. of cases 207 254 226 272 220
b
Multivariable relative risk 1.00 1.11 (0.92, 1.33) 1.07 (0.88, 1.30) 1.13 (0.94, 1.37) 1.10 (0.89, 1.34) 0.55
Potato chips
Median intake, servings/week 0 0.5 0.6 1.0 3.0
No. of cases 219 313 204 216 227
b
Multivariable relative risk 1.00 1.01 (0.85, 1.20) 1.00 (0.82, 1.22) 1.04 (0.86, 1.26) 0.98 (0.80, 1.19) 0.76
Potatoes (baked,
roasted, mashed)
Median intake, servings/week 0.5 1.0 1.5 2.0 3.0
No. of cases 221 302 173 174 309
b
Multivariable relative risk 1.00 1.04 (0.87, 1.24) 1.01 (0.82, 1.24) 0.96 (0.78, 1.19) 0.97 (0.80, 1.17) 0.48
Popcorn
Median intake, servings/week 0 0.5 0.7 1.0 3.0
No. of cases 256 220 247 273 183
Multivariable relative riskb 1.00 1.02 (0.85, 1.23) 1.09 (0.91, 1.31) 0.99 (0.83, 1.18) 0.78 (0.64, 0.95) 0.002
Muffins
Median intake, servings/week 0 0.2 0.5 1.0 2.0
No. of cases 216 155 324 212 272
Multivariable relative riskb 1.00 1.01 (0.81, 1.24) 1.09 (0.92, 1.30) 0.98 (0.80, 1.19) 1.18 (0.98, 1.43) 0.10
Crackers
Median intake, servings/week 0 0.5 0.7 1.2 3.0
No. of cases 244 195 204 285 251
Multivariable relative riskb 1.00 0.94 (0.77, 1.14) 0.96 (0.79, 1.16) 0.96 (0.81, 1.15) 1.10 (0.92, 1.33) 0.13
Table continues

RESULTS acrylamide intake was 10.8 lg/day in the lowest quintile

During 14 years (945,764 person-years) of follow-up, we
documented 1,179 cases of invasive breast cancer among
90,628 premenopausal women in the cohort. The age range
of women in 1991 was 26–46 years. Ages at breast cancer
diagnosis ranged from 26 to 56 years. We had information
on estrogen receptor (ER)/progesterone receptor (PR) status
for 916 (78%) cases. Of these, 597 were ER and PR positive
(ERþ/PRþ), and 196 were ER and PR negative (ER/
PR). Because of the small number of mixed ER/PR status
tumors, we did not include these cases in our analysis by
ER/PR status.
Table 1 shows the characteristics of the cohort in 1991 by
quintile of energy-adjusted acrylamide intake. The mean
and 37.8 lg/day in the highest quintile. The major food
contributors to acrylamide intake were French fries (23%),
coffee (15%), cold breakfast cereal (12%), potato chips
(9%), and other potatoes (baked, roasted, mashed; 5%).
Those in the highest quintile of acrylamide consumption
were more likely to be current smokers and were less likely
to exercise than those in the lowest quintile.
Intake of acrylamide was not associated with risk of pre-
menopausal breast cancer (Table 2). The multivariable rel-
ative risk of breast cancer was 0.92 (95% confidence interval
(CI): 0.76, 1.11) in the highest quintile of intake compared
with the lowest quintile. The P value for a linear trend across
quintiles was 0.61. No association was found for ERþ/PRþ
or ER/PR cancers.

Am J Epidemiol 2009;169:954–961
 Acrylamide and Premenopausal Breast Cancer 959

Table 3. Continued

Intake of High-Acrylamide Foods

Ptrenda
Quintile 1, Low Quintile 2 Quintile 3 Quintile 4 Quintile 5, High

Dark bread
Median intake, servings/week 0 1.0 3.0 4.7 10.2
No. of cases 219 267 280 199 214
Multivariable relative riskb 1.00 1.30 (1.08, 1.55) 1.11 (0.93, 1.33) 0.96 (0.79, 1.17) 0.96 (0.79, 1.17) 0.05
English muffins, bagels, rolls
Median intake, servings/week 0 0.5 1.0 3.0 5.0
No. of cases 149 279 270 268 213
Multivariable relative riskb 1.00 1.04 (0.85, 1.28) 1.02 (0.83, 1.25) 0.93 (0.76, 1.15) 0.98 (0.78, 1.22) 0.38
Pizza
Median intake, servings/week 0.2 0.5 0.7 1.0 2.0
No. of cases 185 268 324 271 131
Multivariable relative riskb 1.00 0.91 (0.75, 1.11) 1.10 (0.91, 1.33) 0.94 (0.76, 1.15) 0.93 (0.73, 1.18) 0.55
All potatoesc
Median intake, servings/week 1.0 2.0 3.0 4.5 7.0
No. of cases 246 225 250 239 219
Multivariable relative riskb 1.00 0.90 (0.74, 1.08) 0.89 (0.74, 1.07) 0.95 (0.78, 1.15) 0.90 (0.73, 1.11) 0.59
Breads/starchesc
Median intake, servings/day 0.7 1.2 1.6 2.2 3.4
No. of cases 253 223 228 258 217
Multivariable relative riskb 1.00 0.86 (0.71, 1.03) 0.86 (0.71, 1.04) 0.95 (0.78, 1.16) 0.87 (0.70, 1.09) 0.59
Baked goodsc
Median intake, servings/week 0.9 2.0 3.3 5.2 9.9
No. of cases 257 223 220 237 242
Multivariable relative riskb 1.00 0.94 (0.78, 1.13) 0.85 (0.70, 1.02) 0.88 (0.73, 1.07) 0.91 (0.74, 1.11) 0.55

Abbreviation: MET, metabolic equivalent.

a
Test for trend calculated by using the median intake in each quintile as a continuous variable.
b
Multivariable models are stratified by age in months and calendar year and adjusted for the following: body mass index (<18.5, 18.5–19.9,
20.0–22.4, 22.5–24.9, 25.0–29.9, and
30 kg/m2), height (<62, 62–<65, 65–<68, and
68 inches; 1 inch ¼ 2.54 cm), oral contraceptive use
(never, former use <4 years, former use
4 years, current use <8 years, and current use
8 years), parity and age at first birth (nulliparous, parity
1–2 and age at first birth <25 years, parity 1–2 and age at first birth 25–<30 years, parity 1–2 and age at first birth
30 years, parity
3 and age at
first birth <25 years, parity
3 and age at first birth
25 years), age at menarche (<12, 12, 13, or
14 years), family history of breast cancer (yes/
no), history of benign breast disease (yes/no), smoking (never, former smoker <25 cigarettes/day, former smoker
25 cigarettes/day, current
smoker <25 cigarettes/day, and current smoker
25 cigarettes/day), physical activity (18 and >18 MET-hours/week), animal fat (quintiles),
glycemic load (quintiles), alcohol intake (continuous), and total energy intake (continuous).
c
All potatoes include French fries, potato chips, and potatoes (baked, roasted, and mashed). Breads/starches include white bread, dark bread,
English muffins/rolls/bagels, muffins, tortillas, pancakes, crackers, and pizza. Baked goods include cookies, brownies, donuts, cake, pie, and
sweet rolls.

Because tobacco use is a major source of acrylamide
exposure, we examined the association among never
smokers, former smokers, and current smokers separately
(Table 2). There was no indication of increased risk of
breast cancer for higher acrylamide intakes in any of these
groups.
We found no significant differences in the association
between dietary acrylamide intake and breast cancer risk
when we stratified the population by age, body mass index,
alcohol intake, glycemic index, or glycemic load (data not
shown).
We repeated the analysis measuring acrylamide exposure
relative to body weight (i.e., lg/kg of body weight/day), as
this is the exposure measurement used in toxicology studies,
and again found similar results. The relative risk for the
highest versus the lowest quintile of acrylamide by body
weight was 1.00 (95% CI: 0.82, 1.22), with a P value for
linear trend of 0.95. Baseline acrylamide intake through diet
was also not associated with breast cancer risk. The relative
risks relative to the lowest quintile of baseline acrylamide
intake were 0.96 (95% CI: 0.79, 1.15) for quintile 2,
1.05 (95% CI: 0.88, 1.26) for quintile 3, 1.01 (95% CI:
0.84, 1.21) for quintile 4, and 1.03 (95% CI: 0.86, 1.24)
for quintile 5 (Ptrend ¼ 0.62).
Table 3 shows the association between consumption
of the major acrylamide-contributing foods and

Am J Epidemiol 2009;169:954–961
960 Wilson et al.
premenopausal breast cancer risk. We examined all individ-
ual foods that contributed at least 2% to the total estimated
acrylamide intake in our population, and none was posi-
tively associated with breast cancer risk. We also examined
several food groups that are major sources of acrylamide: all
potatoes (French fries, potato chips, and baked/mashed/
roasted potatoes), breads (white and dark bread, English
muffins/bagels/rolls, tortillas, pancakes, pizza, and crack-
ers), and baked goods (cookies, brownies, donuts, cake,
pie, and sweet rolls). None of these food groups was asso-
ciated with breast cancer risk.
DISCUSSION
We found no association between acrylamide intake and
premenopausal breast cancer risk in this cohort. There was
no association for ER/PR-positive or -negative cancers or by
smoking status. In addition, there was no association be-
tween intake of any major acrylamide-contributing foods
and breast cancer risk.
These findings are in line with both previous prospective
studies of acrylamide intake and breast cancer risk using
FFQs. Mucci et al. (7) found no significant association be-
tween acrylamide intake and breast cancer risk among
mostly premenopausal Swedish women. Hogervorst et al.
(8) found no significant association among postmenopausal
Dutch women. The mean acrylamide intake and the contrast
in intakes between high and low quintiles were quite similar
across the 3 studies. However, the main sources of acrylam-
ide vary between populations. French fries, coffee, cold
breakfast cereal, and potato chips contribute the most to
intake in our US population. Coffee is a larger contributor
to acrylamide intake in both the Swedish and Dutch cohorts
(7, 8). Fried potatoes and crisp bread are other major con-
tributors in the Swedish women (7), and Dutch spice cake
and cookies are major contributors in the Dutch women (8).
In addition, Pelucchi et al. found no association between
acrylamide intake (15) or fried potato intake (16) and breast
cancer risk in a hospital-based, case-control study in Italy
and Switzerland.
In a nested case-control study, Olesen et al. (9) studied the
association between acrylamide adducts to hemoglobin,
a biomarker of acrylamide exposure, and postmenopausal
breast cancer risk. Current smokers with higher levels of
acrylamide adducts to hemoglobin at baseline had a signifi-
cantly increased risk of breast cancer (for a 10-fold increase
in adducts, relative risk (RR) ¼ 3.1, 95% CI: 1.0, 9.7). To-
bacco use is an important source of acrylamide exposure,
and smokers have acrylamide adduct levels 3–5 times higher
than do nonsmokers. Therefore, adduct levels among smok-
ers reflect both tobacco use and dietary intake of acrylam-
ide. Among nonsmoking women, whose adduct levels are
thought primarily to represent dietary acrylamide exposure,
Olesen et al. found no statistically significant association
between adduct levels and breast cancer risk (for a 10-fold
increase in adducts, RR ¼ 1.5, 95% CI: 0.6, 3.6). The
meaning of the positive association with adduct levels
among smokers is not clear; however, their results among
nonsmokers seem in line with our finding that acrylamide
exposure in the range of dietary intakes is not clearly asso-
ciated with breast cancer risk. Olesen et al. also found a sig-
nificantly increased risk of ERþ cancers among those with
higher adduct levels (for a 10-fold increase in adducts, RR ¼
2.7, 95% CI: 1.1, 6.6). This result was among smokers
and nonsmokers combined, with multivariable adjustment
for smoking behavior; however, given the relative contribu-
tions of smoking and diet to adduct levels, it is not clear that
such adjustment provides adequate control of confounding by
smoking. In our analysis of food frequency questionnaire-
assessed acrylamide and ERþ cancers, we found no signifi-
cant association.
Strengths of our study include its prospective design,
large number of premenopausal cases, and high rates of
follow-up. In addition, we are the first to study acrylamide
intake and cancer risk using multiple FFQs administered to
collect updated dietary data throughout the follow-up pe-
riod, rather than a single point in time. This improves our
assessment of long-term diet and reduces measurement error
(17). Our study also uses an extensive acrylamide database
with 42 acrylamide-contributing foods, approximately twice
as many as used in previous studies.
We have previously found that FFQ acrylamide intake is
significantly correlated with hemoglobin adducts of acryl-
amide and its metabolite glycidamide in this population
(10). However, misclassification of acrylamide intake re-
mains a limitation of our study, as acrylamide poses unique
challenges for FFQ assessment. Acrylamide formation is
affected by many parameters (for review, refer to Stadler
and Scholz (18)), such as cooking temperature and even
the length and temperature of storage of ingredients such
as potatoes, which implies that acrylamide content varies
widely among different brands of prepared foods, and de-
pends on the cooking methods used at home. Therefore, the
assignment of a single acrylamide value for each food likely
results in nondifferential misclassification of acrylamide in-
take, which would bias our observed relative risks toward
the null. As a result, we may have missed modest associa-
tions between acrylamide intake and cancer risk. We also
found no association between breast cancer risk and intake
of any of the top 11 acrylamide-contributing foods, which
are well measured by the FFQ within the similar Nurses’
Health Study cohort (11).
Residual confounding is also a concern in observational
studies. Adjustment for known breast cancer risk factors had
very little effect on the relative risk estimates, suggesting that
it is unlikely that confounding was a source of substantial
bias. Finally, it is not clear that adult diet is the most rel-
evant period of exposure. It is possible that high acrylam-
ide intake in childhood or adolescence may increase breast
cancer risk later in life, and our study cannot address this
question.
In conclusion, we found no association between acrylam-
ide intake from the diet and risk of premenopausal breast
cancer risk. Combined with the results of other prospective
cohort studies, this suggests that intake of foods high in
acrylamide is not a major risk factor for breast cancer. How-
ever, a modest association could have been missed, because
the substantial variation in the acrylamide content of foods
makes measurement of intake difficult.
Am J Epidemiol 2009;169:954–961

ACKNOWLEDGMENTS
Author affiliations: Department of Epidemiology,
Harvard School of Public Health, Boston, Massachusetts
(Kathryn M. Wilson, Lorelei A. Mucci, David J. Hunter,
Walter C. Willett); Department of Nutrition, Harvard
School of Public Health, Boston, Massachusetts (Kathryn
M. Wilson, Eunyoung Cho, David J. Hunter, Walter C.
Willett); Channing Laboratory, Department of Medicine,
Harvard Medical School and Brigham and Women’s
Hospital, Boston, Massachusetts (Lorelei A. Mucci,
Eunyoung Cho, David J. Hunter, Wendy Y. Chen, Walter
C. Willett); and Department of Medical Oncology, Dana-
Farber Cancer Institute, Boston, Massachusetts (Wendy Y.
Chen).
This work was supported by a grant from the National
Cancer Institute (CA050385) and by a National Cancer
Institute/National Institutes of Health training grant (T32
CA09001 to K. M. W.)
Conflict of interest: none declared.
REFERENCES
1. International Agency for Research on Cancer. Monographs on
the Evaluation of Carcinogenic Risks to Humans. Vol 60.
Lyon, France: International Agency for Research on Cancer;
1994.
2. Tareke E, Rydberg P, Karlsson P, et al. Analysis of acrylamide,
a carcinogen formed in heated foodstuffs. J Agric Food Chem.
2002;50(17):4998–5006.
3. Petersen BJ, Tran N. Exposure to acrylamide: placing
exposure in context. Adv Exp Med Biol. 2005;561:
63–76.
4. DiNovi M. The 2006 Exposure Assessment for Acrylamide.
Rockville, MD: Center for Food Safety and Applied Nutrition,
Food and Drug Administration; 2006. (http://www.cfsan.fda.
gov/~dms/acryexpo.html).
5. Johnson KA, Gorzinski SJ, Bodner KM, et al. Chronic toxicity
and oncogenicity study on acrylamide incorporated in the
drinking water of Fischer 344 rats. Toxicol Appl Pharmacol.
1986;85(2):154–168.
Am J Epidemiol 2009;169:954–961
Acrylamide and Premenopausal Breast Cancer 961
6. Friedman MA, Dulak LH, Stedham MA. A lifetime oncogenicity
study in rats with acrylamide. Fundam Appl Toxicol. 1995;
27(1):95–105.
7. Mucci LA, Sandin S, Bälter K, et al. Acrylamide intake and
breast cancer risk in Swedish women. JAMA. 2005;293(11):
1326–1327.
8. Hogervorst JG, Schouten LJ, Konings EJ, et al. A prospective
study of dietary acrylamide intake and the risk of endometrial,
ovarian, and breast cancer. Cancer Epidemiol Biomarkers
Prev. 2007;16(11):2304–2313.
9. Olesen PT, Olsen A, Frandsen H, et al. Acrylamide exposure
and incidence of breast cancer among postmenopausal women
in the Danish Diet, Cancer and Health Study. Int J Cancer.
2008;122(9):2094–2100.
10. Wilson KM, Vesper HW, Tocco P, et al. Validation of a food
frequency questionnaire measurement of dietary acrylamide in-
take using hemoglobin adducts of acrylamide and glycidamide.
Cancer Causes Control. 2008. (doi: 10.1007/s10552-008-9241-7).
(http://www.springerlink.com/content/5470507675718735/
fulltext.pdf).
11. Salvini S, Hunter DJ, Sampson L, et al. Food-based validation
of a dietary questionnaire: the effects of week-to-week varia-
tion in food consumption. Int J Epidemiol. 1989;18(4):
858–867.
12. Collins LC, Marotti JD, Baer HJ, et al. Comparison of estrogen
receptor results from pathology reports with results from
central laboratory testing. J Natl Cancer Inst. 2008;100(3):
218–221.
13. Cho E, Spiegelman D, Hunter DJ, et al. Premenopausal fat
intake and risk of breast cancer. J Natl Cancer Inst. 2003;
95(14):1079–1085.
14. Cho E, Spiegelman D, Hunter DJ, et al. Premenopausal dietary
carbohydrate, glycemic index, glycemic load, and fiber in re-
lation to risk of breast cancer. Cancer Epidemiol Biomarkers
Prev. 2003;12(11 pt 1):1153–1158.
15. Pelucchi C, Galeone C, Levi F, et al. Dietary acrylamide and
human cancer. Int J Cancer. 2006;118(2):467–471.
16. Pelucchi C, Franceschi S, Levi F, et al. Fried potatoes and
human cancer. Int J Cancer. 2003;105(4):558–560.
17. Willett WC. Nutritional Epidemiology. New York, NY: Oxford
University Press; 1998.
18. Stadler RH, Scholz G. Acrylamide: an update on current
knowledge in analysis, levels in food, mechanisms of forma-
tion, and potential strategies of control. Nutr Rev. 2004;62(12):
449–467.
 American Journal of Epidemiology Vol. 169, No. 8
ª The Author 2009. Published by the Johns Hopkins Bloomberg School of Public Health. DOI: 10.1093/aje/kwn422
All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org. Advance Access publication February 11, 2009

Original Contribution

Alcohol Intake and Cigarette Smoking and Risk of a Contralateral Breast Cancer
The Women’s Environmental Cancer and Radiation Epidemiology Study

Julia A. Knight, Leslie Bernstein, Joan Largent, Marinela Capanu, Colin B. Begg, Lene Mellemkjær,
Charles F. Lynch, Kathleen E. Malone, Anne S. Reiner, Xiaolin Liang, Robert W. Haile, John
D. Boice, Jr., WECARE Study Collaborative Group, and Jonine L. Bernstein

Initially submitted September 12, 2008; accepted for publication December 19, 2008.

Women with primary breast cancer are at increased risk of developing second primary breast cancer. Few
studies have evaluated risk factors for the development of asynchronous contralateral breast cancer in women
with breast cancer. In the Women’s Environmental Cancer and Radiation Epidemiology Study (1985–2001), the
roles of alcohol and smoking were examined in 708 women with asynchronous contralateral breast cancer (cases)
compared with 1,399 women with unilateral breast cancer (controls). Cases and controls aged less than 55 years at
first breast cancer diagnosis were identified from 5 population-based cancer registries in the United States and
Denmark. Controls were matched to cases on birth year, diagnosis year, registry region, and race and counter-
matched on radiation treatment. Risk factor information was collected by telephone interview. Rate ratios and 95%
confidence intervals were estimated by using conditional logistic regression. Ever regular drinking was associated
with an increased risk of asynchronous contralateral breast cancer (rate ratio ¼ 1.3, 95% confidence interval: 1.0,
1.6), and the risk increased with increasing duration (P ¼ 0.03). Smoking was not related to asynchronous con-
tralateral breast cancer. In this, the largest study of asynchronous contralateral breast cancer to date, alcohol is
a risk factor for the disease, as it is for a first primary breast cancer.

alcohol drinking; breast neoplasms; neoplasms, second primary; smoking

Abbreviations: CI, confidence interval; RR, rate ratio; WECARE, Women’s Environmental Cancer and Radiation Epidemiology.

The risk of developing asynchronous contralateral breast
cancer, a primary breast cancer occurring in the opposite
breast subsequent to a first breast cancer diagnosis in 1 breast,
in female survivors of breast cancer is considerably higher
than the risk of developing a first primary breast cancer in
unaffected women (1). The population of women with breast
cancer have a higher prevalence of all breast cancer risk
factors, both genetic and nongenetic, than women without
breast cancer. It would be expected that women who go on
to develop asynchronous contralateral breast cancer would
have an even higher prevalence of risk factors than those
who develop only 1 primary, although there may be mitigat-
ing factors such as the treatment received for the first primary
and changes in behavior. Few studies have examined the role
of alcohol intake or cigarette smoking, 2 potentially modifi-
able risk factors, in the development of asynchronous contra-
lateral breast cancer. Given the high level of risk in these
women, clarifying the role of modifiable risk factors in asyn-
chronous contralateral breast cancer in women with breast
cancer is important.
A number of meta-analyses confirm that alcohol is a risk
factor for both pre- and postmenopausal first primary breast
cancer, although the magnitude of increased risk associated
with consuming 1 or more drinks per day is moderate (2–4).
Recent studies confirm this finding (5, 6). Thus far, a signif-
icant association between alcohol intake and second pri-
mary breast cancer has not been observed (7–9), although
2 of the studies observed elevated risk estimates of 1.09 and

Correspondence to Dr. Julia Knight, Samuel Lunenfeld Research Institute, Mount Sinai Hospital, 60 Murray Street, Room 5-237, Box 18, Toronto,
Ontario, Canada M5T 3L9 (e-mail: knight@lunenfeld.ca).

962 Am J Epidemiol 2009;169:962–968


1.11 associated with recent or ever drinking prior to the first
diagnosis, respectively (7, 9). A variety of mechanisms,
which may all contribute to the relation, have been proposed
to explain the association between alcohol and breast cancer
including estrogen metabolism, acetaldehyde mutagenesis,
oxidation and free radicals, and 1-carbon metabolism (10).
Cigarette smoking has been more controversial as a possi-
ble risk factor for breast cancer with inconsistent results in the
literature (11, 12). Conflicting results may be due to the com-
peting effects of smoking at different ages. In some studies,
increased risk has been associated specifically with smoking
at an early age during breast development, although observa-
tions regarding the effect of early smoking vary (13–15). The
relation between smoking and second primary breast cancer
has also been inconsistent, with some studies observing
evidence of an association (16, 17) and others not (7–9).
In this study, we evaluate the evidence for an association
between alcohol intake and cigarette smoking and the devel-
opment of asynchronous contralateral breast cancer among
women with a first diagnosis of breast cancer from the
Women’s Environmental Cancer and Radiation Epidemiol-
ogy (WECARE) Study.
MATERIALS AND METHODS
The WECARE Study is a multicenter, population-based,
nested case-control study where women with asynchronous
contralateral breast cancer serve as cases and women with
unilateral breast cancer serve as matched controls (18). Ad-
ditional detail on data collection has been reported previously
(19). All participants were identified through 5 population-
based tumor registries, 4 in the United States (Los Angeles
County Cancer Surveillance Program, Cancer Surveillance
System of the Fred Hutchinson Cancer Research Center,
State Health Registry of Iowa, Cancer Surveillance Program
of Orange County/San Diego-Imperial Organization for
Cancer Control) and 1 in Denmark (the Danish Breast Cancer
Cooperative Group Registry supplemented by data from the
Danish Cancer Registry).
Study population
Women were eligible as cases if they were diagnosed be-
tween January 1, 1985, and December 31, 2000, and were
aged less than 55 years with a first primary invasive breast
cancer that did not spread beyond the regional lymph nodes at
diagnosis and a second primary in situ or invasive breast
cancer diagnosed in the contralateral breast at least 1 year
after the first breast cancer diagnosis. The asynchronous con-
tralateral breast cancer had to have been diagnosed no later
than December 31, 2001. Case patients were required to have
resided in the same reporting area at the time of diagnosis of
both cancers, to have had no prior or intervening cancer di-
agnosis between their first and second primary breast cancers,
and to be alive at the time of contact. Two control subjects
were individually matched to each case on year of birth, year
of diagnosis, registry region, and race and were 1:2 counter-
matched on registry-reported radiation exposure, so that each
triplet consisted of 1 radiation-unexposed and 2 radiation-
exposed subjects. In addition, controls had to meet the
Am J Epidemiol 2009;169:962–968
Alcohol and Smoking and Contralateral Breast Cancer 963
following criteria: 1) diagnosed since January 1, 1985, with
a first primary invasive breast cancer while residing in 1 of
the study reporting areas; 2) residing on the reference date
(the date of first diagnosis plus ‘‘at-risk interval,’’ the time
between the first and second diagnoses, of the matched case)
in the same registry reporting area where they were diagnosed
with their breast cancer; 3) never diagnosed (by reference
date) with a second primary breast cancer or any other cancer;
4) alive at the time of contact; and 5) without prophylactic
mastectomy of the contralateral breast following diagnosis of
their first primary. The control sampling is accounted for in
the analysis by the inclusion of sampling weights. The design
has been discussed in detail previously (18).
A total of 998 women with asynchronous contralateral
breast cancer were eligible and approached for inclusion
in the study as cases, and 2,112 women with unilateral
breast cancer were eligible as controls. Of these potential
participants, 708 cases (71%) and 1,399 controls (66%)
completed the interview and had a blood sample drawn.
Data collection
All participants in the WECARE Study were interviewed
over the telephone to obtain information on known and
suspected risk factors for breast cancer including the follow-
ing: personal demographics; age at menarche, first birth, and
menopause; parity and lactation; body size; and family his-
tory of cancer. All questions were asked in reference to the
period before the reference date (defined above). The
following descriptions reflect the actual wording used in
the questionnaire or by the interviewer. The specific infor-
mation collected on alcohol and smoking consisted of
whether the women had ever smoked cigarettes or drunk
any alcoholic beverages regularly (at least 1 cigarette
a day for 6 months or longer or at least 1 drink per month,
respectively) before the reference date, how old they were
when they first started smoking cigarettes or drinking alco-
holic beverages regularly, whether they had stopped smok-
ing or drinking regularly, at what age they last stopped, how
many total years they smoked or drank regularly, and, dur-
ing periods when they smoked regularly, on average, how
many cigarettes they usually smoked per day, week, or
month. With respect to alcohol consumption, they were told
that 1 drink is equal to 1 bottle or can of beer, 1 glass of wine
or bottle of wine cooler, or 1 cocktail, shot, or mixed drink
of liquor and then asked to describe their average alcohol
consumption before the reference date in categories (never
or less than 1 drink each month, 1–3 drinks each month, 1
drink each week, 2–4 drinks each week, 5 or 6 drinks each
week, 1 drink each day, 2 or 3 drinks each day, or 4 or more
drinks each day). Medical records, pathology reports, and
hospital charts were used to collect detailed information on
treatment and tumor characteristics. The study protocol was
approved by the institutional review boards at each study
site and by the ethical committee system in Denmark.
Statistical analysis
We used conditional logistic regression analysis with the
inclusion of a log weight covariate in the model where the
964 Knight et al.

Table 1. Characteristics of Women With Unilateral and Table 1. Continued

Asynchronous Contralateral Breast Cancer, the Women’s
Environmental Cancer and Radiation Epidemiology Study, Breast Cancer
1985–2001 Asynchronous
Unilateral
Contralateral
(n 5 1,399)
Breast Cancer (n 5 708)

Asynchronous No. %a No. %

Unilateral
Contralateral c
(n 5 1,399) Family history of breast
(n 5 708)
a
cancer
No. % No. %
No 1,088 77.9 472 66.7
Matched characteristics
Yes 285 20.4 225 31.8
Registry
Unknown 26 1.8 11 1.6
Iowa 222 15.9 113 16.0
Histology of first breast
Orange County/San 231 16.5 118 16.7 cancer
Diego, California
Lobular 131 8.7 90 12.7
Los Angeles County, 390 27.9 199 28.1
California Medullar 51 3.4 33 4.7

Seattle, Washington 198 14.2 99 14.0 Ductal and other 1,213 87.9 584 82.6

Denmark 358 25.6 179 25.3 Stage of first breast cancer

Race Localized 916 64.3 506 71.5

Non-Hispanic white 1,288 92.1 649 91.7 Regional 483 35.7 202 28.5

Hispanic white 48 3.4 24 3.4 Chemotherapy

Black 39 2.8 21 3.0 No 629 42.5 386 54.5

Other 24 1.7 14 2.0 Yes 770 57.5 322 45.5

Mean age at first diagnosis, 45 (23–55) 46 (24–55) Hormone therapy

years (range) No 909 66.3 511 72.2
Mean age at reference date, 51 (27–69) 51 (27–71) Yes 488 33.7 197 27.8
yearsb (range)
a
Mean at-risk period, years 5 (1–16) 5 (1–16) Proportions are weighted for the countermatching with the excep-
(range) tion of factors contributing to the study matching schema.
b
Countermatched characteristic
‘‘Reference date’’ is the date of diagnosis for asynchronous con-
tralateral breast cancer and the corresponding date for unilateral
Radiation treatment breast cancer.
c
No 266 50.2 362 51.1 First-degree family history.
Yes 1,133 49.8 346 48.9
Consumption during the at-risk period was defined as start-
Other characteristics
ing prior to or during the period between first diagnosis and
Year of second the reference date and stopping during or after the period
diagnosis
between the first diagnosis and the reference date. The av-
1986–1988 21 3.0 erage smoking amount was defined as half a pack per day or
1989–1991 75 10.6 less and greater than half a pack per day versus never
1992–1994 137 19.4 smoked and also as pack-years in tertiles versus never
1995–1997 202 28.5
smoked. Average alcohol consumption, which was collected
as categories described above, was defined as less than 1
1998–2001 273 38.6
drink per day and 1 drink per day or more compared with
Table continues never drinking. After adjusting for age in models examining
ever versus never drinking and smoking, we tested the fol-
coefficient of this log weight is fixed at 1 (i.e., an offset in lowing potential confounders: education, first-degree family
the model). The weights used the numbers of registry- history of breast cancer, body mass index (both at first di-
reported, radiation-exposed and -unexposed women in the agnosis and at the reference date), age at menarche, age at
risk set to account for the countermatched sampling design menopause, stage of first primary, histology of first primary,
(18, 20). Further, because controls were independently sam- exposure to chemotherapy, exposure to radiation treatment,
pled from the failure time risk sets, the estimated parameters use of tamoxifen, number of full-term pregnancies, age at
are rate ratios in the proportional hazards model for cohort first full-term pregnancy, and ever breastfeeding. In addi-
data (21), and standard likelihood methods apply (22). tion, to test for mutual confounding, ever regular smoking
Smoking and alcohol consumption were each examined as was added to models of alcohol drinking, and ever regular
ever regular use (yes/no), regular use during the at-risk pe- drinking was added to models of smoking. As there was no
riod (yes/no), lifetime duration in tertiles defined in controls indication of confounding, defined as a change in the rate
versus never use, and age at starting drinking or smoking. ratio estimate of 10% or more over the rate ratio from the

Am J Epidemiol 2009;169:962–968
 Alcohol and Smoking and Contralateral Breast Cancer 965

Table 2. The Age-adjusted Association Between Alcohol Drinking and the Risk of Developing Asynchronous
Contralateral Breast Cancer, the Women’s Environmental Cancer and Radiation Epidemiology Study, 1985–2001

Breast Cancer
Asynchronous 95% Confidence
Unilateral Rate Ratioa
Contralateral Interval
No. %b No. %

Ever drank regularly

No 550 42.4 275 39.0 1.0
Yes 846 57.6 431 61.0 1.3 1.0, 1.6
c
Ever drank regularly during at-risk period
No 672 52.0 348 49.3 1.0
Yes 722 48.0 358 50.7 1.2 0.9, 1.5
Lifetime duration of drinking
Never 550 42.4 275 39.2 1.0
<20 years 279 19.4 137 19.5 1.2 0.9, 1.7
20–<30 years 286 19.3 143 20.4 1.3 0.9, 1.7

30 years 276 18.9 147 20.9 1.4 1.0, 1.9
Ptrendd 0.03
Average drinking amount
Never 550 43.1 275 39.5 1.0
<1 drink/day 659 45.2 338 48.5 1.3 1.0, 1.7

1 drink/day 171 11.7 84 12.1 1.2 0.8, 1.7
d
Ptrend 0.16
Starting age
Never 550 42.4 275 39.0 1.0

20 years of age 526 34.5 282 40.0 1.4 1.1, 1.8
<20 years of age 315 23.1 148 21.0 1.1 0.8, 1.5
a
Accounting for countermatching and adjusted for age at first diagnosis.
b
Proportions are weighted for the countermatching.
c
Consumption during the at-risk period was defined as starting prior to or during the period between the first
diagnosis and the reference date and stopping during or after the period between the first diagnosis and the reference
date.
d
Test for trend across categories.

age-adjusted model, only the age-adjusted models are pre-
sented. We performed tests for trend across categories of
duration and consumption. All analyses were conducted using
SAS, version 9.1, software (SAS Institute, Inc., Cary, North
Carolina), and a 2-sided P < 0.05 was considered significant.
RESULTS
Table 1 shows selected characteristics of the WECARE
Study population. The 2 groups were similar on all matched
characteristics. Women with asynchronous contralateral
breast cancer were more likely to have a family history of
breast cancer. The majority (67%) of second diagnoses oc-
curred in 1995 or later. In addition, 56% of cases and 49% of
controls were interviewed within 5 years of the reference
date, and 88% of cases and 86% of controls were inter-
viewed within 10 years. Table 2 shows that ever regular
drinking was associated with an elevated risk of developing
asynchronous contralateral breast cancer (rate ratio (RR) ¼
1.3, 95% confidence interval (CI): 1.0, 1.6), although the
increased risk associated with drinking specifically after
the first diagnosis (the at-risk period) did not achieve statis-
tical significance (RR ¼ 1.2, 95% CI: 0.9, 1.5). The risk
increased with increasing lifetime duration of drinking
(Ptrend ¼ 0.03). We observed no apparent trend associated
with the average amount of alcohol consumed. The risk of
asynchronous contralateral breast cancer increased with
later initiation of drinking (at age 20 or more years) but
not with early initiation (before age 20 years). We did not
observe any differences in the risks associated with the du-
ration of drinking prior to the first pregnancy, adjusted for
the age at first full-term pregnancy: began drinking after first
pregnancy (RR ¼ 1.5, 95% CI: 1.1, 2.1); drank <7 years
before first pregnancy (RR ¼ 1.4, 95% CI: 1.0, 1.9); drank

7 years before first pregnancy (RR ¼ 1.2, 95% CI: 0.8,
1.7), versus parous and never drank. There was no evidence
for any association with smoking (Table 3) nor any consis-
tent pattern with smoking before the first pregnancy, ad-
justed for age at first full-term pregnancy: began smoking
after first pregnancy (RR ¼ 1.6, 95% CI: 1.0, 2.5); smoked

Am J Epidemiol 2009;169:962–968
966 Knight et al.

Table 3. The Age-adjusted Association Between Cigarette Smoking and the Risk of Developing Asynchronous
Contralateral Breast Cancer, the Women’s Environmental Cancer and Radiation Epidemiology Study, 1985–2001

Breast Cancer
Asynchronous 95% Confidence
Unilateral Rate Ratioa
Contralateral Interval
No. %b No. %

Ever regularly smoked

No 701 49.9 349 49.3 1.0
Yes 698 50.1 359 50.7 1.1 0.9, 1.6
c
Ever smoked regularly during at-risk period
No 1,108 79.6 542 76.9 1.0
Yes 290 20.4 163 23.1 1.2 0.9, 1.5
Lifetime duration of smoking
Never 701 49.9 349 49.4 1.0
<20 years 323 23.4 150 21.2 1.0 0.8, 1.3

20 years 374 26.7 207 29.3 1.1 0.9, 1.5
Ptrendd 0.36
Average smoking amount
Never 701 49.9 349 49.7 1.0

pack/day 358 25.6 180 25.6 1.0 0.8, 1.3
>
pack/day 337 24.5 173 24.6 1.1 0.8, 1.4
Ptrendd 0.57
Pack-years of smoking
Never 701 50.0 349 49.9 1.0
<6 235 17.2 94 13.4 0.8 0.6, 1.1
6–<19 232 16.4 132 18.9 1.3 1.0, 1.8

19 227 16.4 125 17.9 1.1 0.8, 1.4
Ptrendd 0.27
Starting age
Never 701 49.9 349 49.3 1.0

20 years of age 225 16.0 104 14.7 1.0 0.8, 1.4
<20 years of age 473 34.1 255 36.0 1.1 0.9, 1.4
a
Accounting for countermatching and adjusted for age at first diagnosis.
b
Proportions are weighted for the countermatching.
c
Smoking during the at-risk period was defined as starting prior to or during the period between the first diagnosis
and the reference date and stopping during or after the period between the first diagnosis and the reference date.
d
Test for trend across categories.

<6 years before first pregnancy (RR ¼ 1.0, 95% CI: 0.7,
1.4); smoked
6 years before first pregnancy (RR ¼ 1.3,
95% CI: 1.0, 1.8), versus parous and never smoked. We did
not observe any difference in the effect of smoking or alco-
hol by categories of radiation exposure or any difference in
the effect of alcohol by tamoxifen use (data not shown). We
also did not observe consistent differences in results across
categories when we conducted the analysis by the time be-
tween reference date and interview in 3 categories, 0–5
years, 6–10 years, and >10 years (data not shown).
DISCUSSION
In the WECARE Study, we found that consuming alco-
hol, particularly over longer periods of time, was associated
with an increased risk of asynchronous contralateral breast
cancer. We did not find evidence that smoking cigarettes
increased the risk of this disease. Our results differ from
previous studies that did not observe an increased risk of
asynchronous contralateral breast cancer associated with
alcohol (7–9). However, as with a first primary breast can-
cer, the effect of alcohol is modest and could be missed if the
sample size was insufficient. Previous studies included 488,
77, or 136 cases (7–9) compared with the 708 cases in our
study. Varying definitions and prevalences of drinking may
also have contributed to the inconsistency among studies. In
the largest previous study, only information on recent drink-
ing prior to the first diagnosis was available, and the authors
were unable to evaluate lifetime duration of drinking (7).
We did not find a relation with reported average amount

Am J Epidemiol 2009;169:962–968

consumed, but the women were asked to average their
lifetime consumption, and changes in intake were not cap-
tured, including changes occurring around the time of the
first diagnosis. Relatively few women in this population
(12%) reported consuming 1 drink per day or more on
average.
Previous studies of smoking and asynchronous contralat-
eral breast cancer have yielded inconsistent results (7–9, 16,
17), similar to studies of first primary breast cancer (11, 12).
As with alcohol, this inconsistency may be due to issues of
study design, such as variation in sample size, smoking
definitions, and smoking prevalences. Our results from the
WECARE Study do not support the hypothesis that smoking
is a risk factor for asynchronous contralateral breast cancer,
although a small increased risk associated with smoking
cannot be ruled out. Previous studies have not evaluated
changes in drinking and smoking behavior. In the WECARE
Study, we found that only a small proportion of women
changed their drinking status after their first breast cancer
diagnosis (10% of cases and 9% of controls), but the pro-
portion who stopped smoking was greater (28% of cases and
29% of controls).
The WECARE Study included women who were aged
less than 55 years at first diagnosis (mean age, 45 years
among controls and 46 years among cases), although they
were somewhat older at the reference date (mean age in both
groups, 51 years). These women were younger than those in
some other studies (7) but not others (8). However, there is
no evidence that the effect of smoking or alcohol on the risk
of first primary breast cancer differs by menopausal status
(2–4, 10).
An important strength of the WECARE Study is that it is
the largest case-control study conducted of asynchronous con-
tralateral breast cancer to date that includes direct patient
interview. We have also been able to detect other associa-
tions in their expected directions for fewer full-term preg-
nancies and early menarche (19), radiation treatment (23),
and treatment of the first primary breast cancer with chemo-
therapy or tamoxifen (24). However, although we did cap-
ture some information on lifetime duration of smoking and
alcohol consumption, the level of detail of the information
collected was limited. We did not capture the changes in the
patterns of consumption at various times of life. Although
recall bias is a problem in case-control studies in general,
whether or not it is an issue in our study is unclear, as both
the cases and controls have been affected by breast cancer.
In this study, cases and controls were matched on time since
the first diagnosis, and the majority (56%) of cases were
interviewed within 5 years and most (88%) within 10 years
of the second diagnosis. It is likely that women can recall
accurately whether they drank or smoked, and the broad
categories used in the analysis of duration of drinking and
smoking minimize the potential for misclassification errors
in recall of when drinking and smoking began and ended. As
with amount, we were unable to examine duration in greater
detail. We were also unable to consider results by estrogen
and progesterone status, as a considerable proportion of the
cases were missing information on receptor status. Results
from studies relating alcohol to first breast cancer according
to hormone receptor status are inconsistent (10, 25). Further,
Am J Epidemiol 2009;169:962–968
Alcohol and Smoking and Contralateral Breast Cancer 967
if drinking and/or smoking adversely affects survival, and if
fewer women who drink and/or smoke survive to be diag-
nosed with a second primary or to be included in the study
after the second diagnosis, this could affect the relative risks
reported here.
Women with a first primary unilateral breast cancer have
an elevated risk of developing cancer in the contralateral
breast. Although we did not observe an increased risk of
asynchronous contralateral breast cancer associated with
cigarette smoking, there are many other reasons to quit
smoking including reducing the risk of smoking-related
cancer or heart disease. Alcohol appears to be associated
with an increased risk of asynchronous contralateral breast
cancer.
ACKNOWLEDGMENTS
Author affiliations: Samuel Lunenfeld Research Institute,
Mount Sinai Hospital, Toronto, Canada (Julia A. Knight);
City of Hope, Duarte, California (Leslie Bernstein); Univer-
sity of California at Irvine, Irvine, California (Joan Largent);
Memorial Sloan Kettering Cancer Center, New York, New
York (Marinela Capanu, Colin Begg, Anne S. Reiner,
Xiaolin Liang, Jonine L. Bernstein); Danish Cancer Society,
Copenhagen, Denmark (Lene Mellemkjær); University of
Iowa, Iowa City, Iowa (Charles F. Lynch); Fred Hutchinson
Cancer Research Center, Seattle, Washington (Kathleen E.
Malone); University of Southern California, Los Angeles,
California (Robert W. Haile); and International Epidemiol-
ogy Institute, Rockville, Maryland, and Vanderbilt Univer-
sity, Nashville, Tennessee (John D. Boice, Jr.).
This work was supported by the National Institutes of
Health (grants U01 CA83178 and R01 CA97397).
WECARE Study Collaborative Group: Memorial Sloan
Kettering Cancer Center (New York, New York): Jonine L.
Bernstein (WECARE Study Principal Investigator), Colin
Begg, Marinela Capanu, Xiaolin Liang, Anne S. Reiner,
Irene Orlow, Tracy Layne; City of Hope (Duarte, Califor-
nia): Leslie Bernstein (subcontract Principal Investigator),
Laura Donnelly-Allen (some work performed at University
of Southern California, Los Angeles, California); Danish
Cancer Society (Copenhagen, Denmark): Jørgen H. Olsen
(subcontract Principal Investigator), Michael Andersson,
Lisbeth Bertelsen, Per Guldberg, Lene Mellemkjær; Fred
Hutchinson Cancer Research Center (Seattle, Washington):
Kathleen E. Malone (subcontract Principal Investigator),
Noemi Epstein; International Epidemiology Institute
(Rockville, Maryland) and Vanderbilt University (Nashville,
Tennessee): John D. Boice, Jr. (subcontract Principal Investi-
gator), David P. Atencio; National Cancer Institute (Bethesda,
Maryland): Daniela Seminara; New York University (New
York, New York): Roy E. Shore (subcontract Principal Inves-
tigator); University of California at Irvine (Irvine, California):
Hoda Anton-Culver (subcontract Principal Investigator),
Joan Largent (Coinvestigator); University of California at
Los Angeles (Los Angeles, California): Richard A. Gatti
(Coinvestigator); University of Iowa (Iowa City, Iowa):
Charles F. Lynch (subcontract Principal Investigator), Jeanne
968 Knight et al.
DeWall; University of Southern California (Los Angeles, Cal-
ifornia): Robert W. Haile (subcontract Principal Investigator),
Bryan M. Langholz (Coinvestigator), Duncan C. Thomas
(Coinvestigator), Anh T. Diep (Coinvestigator), Shanyan
Xue, Nianmin Zhou, Yong Liu, Evgenia Ter-Karapetova;
University of Southern Maine (Portland, Maine): W. Douglas
Thompson (subcontract Principal Investigator); University of
Texas, M. D. Anderson Cancer Center (Houston, Texas):
Marilyn Stovall (subcontract Principal Investigator), Susan
Smith (Coinvestigator); and University of Virginia (Charlot-
tesville, Virginia): Patrick Concannon (subcontract Principal
Investigator), Sharon Teraoka (Coinvestigator), Eric R. Olson
(some work performed at Benaroya Research Institute at Vir-
ginia Mason, Seattle, Washington), V. Anne Morrison, Karen
M. Cerosaletti, Lemuel Navarro, Jocyndra Wright.
Conflict of interest: none declared.
REFERENCES
1. Chen Y, Thompson W, Semenciw R, et al. Epidemiology of
contralateral breast cancer. Cancer Epidemiol Biomarkers
Prev. 1999;8(10):855–861.
2. Ellison RC, Zhang Y, McLennan CE, et al. Exploring the re-
lation of alcohol consumption to risk of breast cancer. Am J
Epidemiol. 2001;154(8):740–747.
3. Smith-Warner SA, Spiegelman D, Yaun SS, et al. Alcohol and
breast cancer in women: a pooled analysis of cohort studies.
JAMA. 1998;297(7):535–540.
4. Collaborative Group on Hormonal Factors in Breast Cancer.
Alcohol, tobacco and breast cancer—collaborative reanalysis
of individual data from 53 epidemiological studies, including
58515 women with breast cancer and 95067 women without
the disease. Br J Cancer. 2002;87(11):1234–1245.
5. Berstad P, Ma H, Bernstein L, et al. Alcohol intake and breast
cancer among young women. Breast Cancer Res Treat. 2007;
108(1):113–120.
6. Tjønneland A, Christensen J, Olsen A, et al. Alcohol intake
and breast cancer risk: the European Prospective Investigation
into Cancer and Nutrition (EPIC). Cancer Causes Control.
2007;18(4):361–373.
7. Trentham-Dietz A, Newcomb PA, Nichols H, et al. Breast
cancer risk factors and second primary malignancies among
women with breast cancer. Breast Cancer Res Treat. 2007;
105(2):195–207.
8. Li CI, Malone KE, Porter PL, et al. Epidemiologic and
molecular risk factors for contralateral breast cancer
among young women. Br J Cancer. 2003;89(3):513–518.
9. Bernstein JL, Thompson WD, Risch N, et al. Risk factors
predicting the incidence of second primary breast cancer
among women diagnosed with a first primary breast cancer.
Am J Epidemiol. 1992;136(8):925–936.
10. Dumitrescu RG, Shields PG. The etiology of alcohol-induced
breast cancer. Alcohol. 2005;35(3):213–225.
11. Terry PD, Rohan TE. Cigarette smoking and the risk of breast
cancer in women: a review of the literature. Cancer Epidemiol
Biomarkers Prev. 2002;11(10 pt 1):953–971.
12. Morabia A. Smoking (active and passive) and breast cancer:
epidemiologic evidence up to June 2001. Environ Mol Mutagen.
2002;39(2-3):89–95.
13. Reynolds P, Hurley S, Goldberg DE, et al. Active smoking,
household passive smoking, and breast cancer: evidence from
the California Teachers Study. J Natl Cancer Inst. 2004;96(1):
29–37.
14. Ha M, Mabuchi K, Sigurdson AJ, et al. Smoking cigarettes
before first childbirth and risk of breast cancer. Am J Epidemiol.
2007;166(1):55–61.
15. Prescott J, Ma H, Bernstein L, et al. Cigarette smoking is not
associated with breast cancer risk in young women. Cancer
Epidemiol Biomarkers Prev. 2007;16(3):620–622.
16. Fowble B, Hanlon A, Freedman G, et al. Second cancers
after conservative surgery and radiation for stages I–II
breast cancer: identifying a subset of women at increased
risk. Int J Radiat Oncol Biol Phys. 2001;51(3):679–690.
17. Horn PL, Thompson WD. Risk of contralateral breast cancer:
associations with factors related to initial breast cancer. Am J
Epidemiol. 1988;128(2):309–323.
18. Bernstein JL, Langholz B, Haile RW, et al. Study design:
evaluating gene–environment interactions in the etiology of
breast cancer—the WECARE Study. Breast Cancer Res. 2004;
6(3):R199–R214.
19. Largent J, Capanu M, Bernstein L, et al. Reproductive history
and risk of second primary breast cancer: the WECARE Study.
Cancer Epidemiol Biomarkers Prev. 2007;16(5):906–911.
20. Langholz B, Borgan Ø. Counter-matching: a stratified nested
case-control sampling method. Biometrika. 1995;82(1):69–79.
21. Cox DR. Regression models and life tables. J R Stat Soc (B).
1972;34:187–202.
22. Borgan Ø, Goldstein L, Langholz B. Methods for the analysis
of sampled cohort data in the Cox proportional hazards model.
Ann Statist. 1995;23(5):1749–1778.
23. Stovall M, Smith SA, Langholz BM, et al. Dose to the con-
tralateral breast from radiotherapy and risk of second primary
breast cancer in the WECARE Study. Int J Radiat Oncol Biol
Phys. 2008;72(4):1021–1030.
24. Bertelsen L, Bernstein L, Olsen JH, et al. Effect of systemic
adjuvant treatment on risk for contralateral breast cancer in the
Women’s Environment, Cancer and Radiation Epidemiology
Study. J Natl Cancer Inst. 2008;100(1):32–40.
25. Althuis MD, Fergenbaum JH, Garcia-Closas M, et al. Etiology
of hormone receptor-defined breast cancer: a systematic re-
view of the literature. Cancer Epidemiol Biomarkers Prev.
2004;13(10):1558–1568.
Am J Epidemiol 2009;169:962–968
 American Journal of Epidemiology Vol. 169, No. 8
ª The Author 2009. Published by the Johns Hopkins Bloomberg School of Public Health. DOI: 10.1093/aje/kwp018
All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org. Advance Access publication March 6, 2009

Original Contribution

Positive Associations Between Ionizing Radiation and Lymphoma Mortality

Among Men

David B. Richardson, Hiromi Sugiyama, Steve Wing, Ritsu Sakata, Eric Grant, Yukiko Shimizu,
Nobuo Nishi, Susan Geyer, Midori Soda, Akihiko Suyama, Fumiyoshi Kasagi, and
Kazunori Kodama

Initially submitted August 7, 2008; accepted for publication January 8, 2009.

The authors investigated the relation between ionizing radiation and lymphoma mortality in 2 cohorts: 1) 20,940
men in the Life Span Study, a study of Japanese atomic bomb survivors who were aged 15–64 years at the time of
the bombings of Hiroshima and Nagasaki, and 2) 15,264 male nuclear weapons workers who were hired at the
Savannah River Site in South Carolina between 1950 and 1986. Radiation dose-mortality trends were evaluated
for all malignant lymphomas and for non-Hodgkin’s lymphoma. Positive associations between lymphoma mortality
and radiation dose under a 5-year lag assumption were observed in both cohorts (excess relative rates per sievert
were 0.79 (90% confidence interval: 0.10, 1.88) and 6.99 (90% confidence interval: 0.96, 18.39), respectively).
Exclusion of deaths due to Hodgkin’s disease led to small changes in the estimates of association. In each cohort,
evidence of a dose-response association was primarily observed more than 35 years after irradiation. These
findings suggest a protracted induction and latency period for radiation-induced lymphoma mortality.

lymphoma; mortality; nuclear weapons; radiation, ionizing

Abbreviations: CI, confidence interval; ERR, excess relative rate; ICD, International Classification of Diseases; LRT, likelihood
ratio test; LSS, Life Span Study; ND, not determined; NHL, non-Hodgkin’s lymphoma; SRS, Savannah River Site.

Ionizing radiation has been considered as a cause of lym-
phoma by a number of investigators. In a review of this
literature, Boice (1) concluded that the evidence of associ-
ation between ionizing radiation and non-Hodgkin’s lym-
phoma (NHL) is extremely weak and that there is no
evidence of association between radiation and Hodgkin’s
disease. The United Nations Scientific Committee on the
Effects of Atomic Radiation noted that studies of NHL fol-
lowing external exposure to ionizing radiation have yielded
mixed results and concluded that overall there is little evi-
dence of an association between NHL and external exposure
to ionizing radiation (2). Ron (3) reached a similar conclu-
sion, noting that evidence of association between radiation
and NHL has been inconsistent and Hodgkin’s disease has
rarely been related to radiation exposure; and Melbye and
Trichopoulos (4) stated that there is no evidence that ioniz-
ing radiation causes NHL. However, this conclusion is not
universally shared. Hartge et al. argued that the evidence
suggests that ionizing radiation probably causes lymphoma
(5) and observed that high doses of ionizing radiation appear
to be associated with lymphoma risk in some studies of
radiotherapy (6).
Lack of a consistent association between ionizing radia-
tion and lymphoma could mean that there is no causal re-
lation or that a causal relation is obscured by bias or
deficiencies in exposure measurement, case classification,
duration of follow-up, or some combination of these factors.
Given that lymphoma is often an indolent disease, long-term
studies of radiation-exposed populations may be needed to
observe an effect. The development of nuclear weapons in
the early 1940s led to 2 types of epidemiologic studies that
can now provide evidence regarding the radiation-
lymphoma association: studies based on follow-up of work-
ers exposed to ionizing radiation during nuclear weapons

Correspondence to Dr. David B. Richardson, Department of Epidemiology, School of Public Health, University of North Carolina at Chapel Hill,
Chapel Hill, NC 27599 (e-mail: david.richardson@unc.edu).

969 Am J Epidemiol 2009;169:969–976

970 Richardson et al.
production and studies based on follow-up of people ex-
posed to ionizing radiation from the use of nuclear weapons.
Most prominent among the latter is the Life Span Study
(LSS), a study of Japanese survivors of the atomic bombings
of Hiroshima and Nagasaki. Radiation risk estimates from
studies of nuclear workers are often compared with esti-
mates from the LSS in order to evaluate the consistency of
risk estimates in a population that includes people exposed
to acute high doses with estimates from populations that are
chronically exposed to low doses (7–9).
We examined the association between ionizing radiation
and lymphoma mortality in a US occupational cohort and in
a sample of LSS atomic bomb survivors and compared find-
ings from the 2 populations. Follow-up of each cohort com-
menced in 1950 and spanned approximately 5 decades. To
the extent possible, we conducted these analyses as parallel
analyses employing comparable methods. We focused, in
particular, on variation in the associations between radiation
dose and lymphoma mortality by time since exposure.
MATERIALS AND METHODS
The LSS cohort includes 86,611 people who were alive at
the time of the 1950 Japanese census, reported being in
Hiroshima or Nagasaki at the time of the bombings (August
1945), and had dose estimates based on the DS02 dosimetry
system (10). Follow-up for ascertainment of vital status and
cause-of-death information started on October 1, 1950, and
continued until December 31, 2000.
The Savannah River Site (SRS) was constructed near
Aiken, South Carolina, in 1950 as a facility to produce
materials for the US nuclear weapons program. A cohort
of 18,883 workers who were hired at the SRS prior to
1987, who worked there for at least 90 days, who were
not known to have been employed at another US Depart-
ment of Energy facility, and who had complete information
on name, Social Security number, sex, date of birth, and date
of hire was enumerated (11). Vital status and cause-of-death
information were ascertained through December 31, 2002.
Cohort restrictions for comparability
Since over 95% of the collective dose at SRS was in-
curred by males, there was little ability to estimate risk
due to radiation exposure among female SRS workers. We
therefore restricted the analyses to males in both cohorts.
Since the youngest age at hire at SRS was 15 years and most
SRS workers terminated their employment by age 65 years,
LSS analyses were restricted to people who were aged 15–
64 years at the time of the bombings. This resulted in a co-
hort of 15,264 male SRS workers and a cohort of 20,940
male LSS subjects who were aged 15–64 years at the time of
the bombings.
Dosimetry data
For the LSS, we used DS02 revised colon dose estimates
adjusted for dosimetry errors, with shielded kerma estimates
above 4 Gy truncated to 4 Gy (12). For consistency with
analyses of the SRS cohort, dose estimates calculated as the
sum of the c-radiation dose plus 10 times the neutron dose
are expressed in sieverts; some recent reports on LSS ana-
lyses refer to this quantity as the weighted dose in grays (13,
14). Interactions between radiation and lymphocytes may
occur in the lymphatic or circulatory system at a variety
of anatomic sites; the choice of target organ for dose esti-
mation may depend on the characteristics of the lymphoma,
including anatomic location (15, 16). The colon dose has
been taken as a representative dose to the organs involved at
a variety of anatomic locations, similar to the approach
employed in prior analyses of solid cancers (17). The colon
dose estimate has been used by previous investigators as an
estimate comparable to the quantity estimated by the radi-
ation dosimeters worn by nuclear industry workers (i.e., the
‘‘deep dose’’).
For SRS workers, the exposure of interest was defined as
cumulative whole-body radiation dose equivalent from ex-
ternal sources and tritium received during employment at
the site, expressed in sieverts; neutron doses were multiplied
by a factor of 10. Personal radiation monitoring data were
available for the period 1950–1999. Whole-body doses were
estimated for work-years with missing dose data using dose
estimates from adjacent time periods and average values for
similar workers; estimated annual doses constituted 4% of
employment years for male workers (18).
Outcome definitions
In the LSS, underlying cause of death was coded accord-
ing to the International Classification of Diseases, Ninth
Revision (ICD-9), which was issued in 1977. In the SRS
study, underlying cause of death was coded according to the
Eighth Revision of the ICD (ICD-8) for deaths occurring
prior to 1979 and according to the ICD revision in effect at
the time of death for deaths occurring in 1979 or later. (The
Tenth Revision of the ICD (ICD-10) was issued in 1992.)
As in prior analyses (17, 19), we examined the broad
category of malignant lymphoma (ICD-8 and ICD-9 codes
200–202; ICD-10 codes C81–C85). In addition, we exam-
ined the subcategory of NHL (ICD-8 and ICD-9 codes 200
and 202; ICD-10 codes C82–C85). There were too few
deaths due to Hodgkin’s disease to support separate analyses
of that outcome in these cohorts.
Statistical methods
Poisson regression methods were used. The analytical
data file for the LSS cohort consisted of a tabulation of
person-time and numbers of deaths by city, age at exposure
(in 5-year intervals), attained age (in 5-year intervals), cal-
endar time (1950–1952, 1953–1955, and then 5-year inter-
vals up to 1995, 1996–1997, and 1998–2000), and dose
(<0.005, 0.005–<0.02, 0.02–<0.04, 0.04–<0.06, 0.06–
<0.08, 0.08–<0.1, 0.1–<0.125, 0.125–<0.150, 0.150–
<0.175, 0.175–<0.2, 0.2–<0.25, 0.25–<0.3, 0.3–<0.5,
0.5–<0.75, 0.75–<1, 1–<1.25, 1.25–<1.5, 1.5–<1.75,
1.75–<2, 2–<2.5, 2.5–<3, and
3 Sv). The analytical data
Am J Epidemiol 2009;169:969–976

file for the SRS cohort consisted of a tabulation of person-
time and events by attained age (in 5-year intervals), race
(black vs. other), year of birth (before 1915, 1915–1924,
1925–1929, 1930–1934, 1935–1949, or 1950 or later), pay
code (paid monthly, weekly, or hourly), employment status
(employed, terminated within the last 2 years, or terminated
more than 2 years prior, classified separately for risk
ages <62 years and
62 years) (20–22), and dose
(0, >0–<0.005, 0.005–<0.02, 0.02–<0.04, 0.04–<0.06, 0.06–
<0.08, 0.08–<0.1, 0.1–<0.125, 0.125–<0.150, 0.150–
<0.175, 0.175–<0.2, 0.2–<0.25, 0.25–<0.3, and
0.3 Sv).
Covariate control was achieved through background strat-
ification of regression models. In analyses of the LSS co-
hort, the stratifying factors were attained age, age at
exposure, and city; in analyses of the SRS cohort, the strat-
ifying factors were attained age, birth cohort, race, pay code,
and employment status. Radiation dose-mortality associa-
tions were estimated via a regression model of the form
rate ¼ eai ð1 þ b xÞ;
where ai indexes the stratum-specific mortality rate in the
absence of radiation exposure and b̂ provides an estimate of
the excess relative rate (ERR) per sievert (23, 24).
In analyses of the LSS cohort, x represents the estimated
radiation dose delivered at the time of the bombings in
August 1945. Since follow-up of the LSS cohort began in
October 1950, this implies a minimal lag of approximately
5 years between exposure and its effect. We also present re-
sults from analyses in which we assumed that there was no
excess risk during the period 1950–1955; that is, a minimum
latency period of approximately 10 years was assumed. A
10-year lag assumption has been used in previous nuclear
worker studies that examined lymphoma mortality (25, 26).
We refer to analyses of LSS data that examine excess mor-
tality risk since 1950 and since 1956 as analyses carried out
under 5- and 10-year lag assumptions, respectively. In anal-
yses of the SRS cohort, x represents the cumulative radiation
dose under a 5- or 10-year lag assumption. Lagging dose
assignment by L years means that an increment of dose was
included in the calculation of cumulative dose at time t if it
had been received at or before time t  L years; person-time
and events at time t were then classified according to that
category of lagged cumulative dose.
The dose range in the LSS, 0–4 Sv, was wider than the dose
range in the SRS study (0–<0.5 Sv). In order to evaluate
dose-response associations over a comparable range of doses,
we also conducted analyses based upon LSS data limited to
the 19,183 survivors with doses in the range of 0–<0.5 Sv.
In analyses of the LSS cohort, we assessed variation in
radiation risk with time since exposure via a regression
model of the form
rate ¼ eai ð1 þ b1 xPeriod1 þ b2 xPeriod2
þ b3 xPeriod3 þ b4 xPeriod4Þ;
where Period1–Period4 are indicator variables for the cal-
endar time periods 1950–1970, 1971–1980, 1981–1990, and
1991–2000, respectively. The values b̂1 ; b̂2 ; b̂3 ; and b̂4 pro-
Am J Epidemiol 2009;169:969–976
Ionizing Radiation and Lymphoma Mortality 971
vide estimates of the ERR per 1-Sv dose during the periods
5–25, 26–35, 36–45, and 46–55 years after the bombings. In
analyses of the SRS cohort, we fitted a model of the form
rate ¼ edi ð1 þ /1 d1 þ /2 d2 þ /3 d3 Þ;
where d1–d3 represent the cumulative radiation doses ac-
crued in the exposure time windows 5–25, 26–35, and

36 years prior to observation of a person-year or
event and /̂1 ; /̂2 ; and /̂3 provide associated estimates of
the ERR per 1-Sv dose.
We estimated parameters using the EPICURE statistical
package (Hirosoft International Corporation, Seattle,
Washington); for consistency with recent reports (2, 26),
we generated 90% confidence intervals for estimated param-
eters via the likelihood method (27). In some analyses, con-
fidence bounds could not be determined (designated ‘‘not
determined’’ (ND)). In order to aid interpretation of model
fittings, we report the 1-sided P value derived via a likeli-
hood ratio test (LRT) for each reported point estimate.
Tabulations of observed versus expected numbers of deaths
by category of cumulative dose are reported; we calculated
expected counts for each cell of the person-time table using
a regression model that included all variables except the
dose term.
RESULTS
With follow-up through 2000, 90 malignant lymphoma
deaths were observed among the male atomic bomb sur-
vivors exposed at ages 15–64 years, including 6 deaths from
Hodgkin’s disease (Table 1). Sixty-three malignant lym-
phoma deaths occurred among residents of Hiroshima (58
due to NHL) and 27 malignant lymphoma deaths occurred
among residents of Nagasaki (26 due to NHL). No deaths
due to malignant lymphoma occurred among survivors at
attained ages less than 30 years. In the SRS cohort, 56
lymphoma deaths were observed; 5 of these deaths were
due to Hodgkin’s disease. One death due to malignant lym-
phoma was observed among black males (it was a case of
NHL), and 18, 14, and 24 deaths due to malignant lym-
phoma were observed among workers paid monthly,
weekly, and hourly, respectively. Three deaths due to ma-
lignant lymphoma occurred among actively employed SRS
workers (all were cases of NHL) and 6 deaths occurred
within 2 years of termination of employment (all were cases
of NHL), while the remaining 47 deaths due to malignant
lymphoma occurred 2 or more years after termination of
employment at SRS (42 due to NHL).
In the LSS, the estimated ERR of malignant lymphoma
per sievert, under a 5-year lag assumption, was 0.79 (90%
confidence interval (CI): 0.10, 1.88). The goodness of model
fit was slightly improved, and the magnitude of association
was slightly increased, upon exclusion of deaths due to
Hodgkin’s disease (Table 2). Under a 10-year lag assump-
tion, these estimated associations were slightly larger in
magnitude. In the SRS study, the estimated ERRs of malig-
nant lymphoma per sievert under 5- and 10-year lag assump-
tions were 6.99 (90% CI: 0.96, 18.39) and 8.18 (90% CI:
1.44, 21.16), respectively. Upon exclusion of deaths due to
972 Richardson et al.

Table 1. Observed Numbers of Deaths Due to Malignant Lymphoma Among Male Atomic
Bomb Survivors (1950–2000) and Male Workers at the Savannah River Site (1950–2002),
by Age Group, Japan and South Carolinaa

Atomic Bomb Savannah River

Survivorsb Site Workers
Attained
No. of Deaths No. of Deaths
Age, years Person-Years Person-Years
of Follow-Up Malignant Non-Hodgkin’s of Follow-Up Malignant Non-Hodgkin’s
Lymphoma Lymphoma Lymphoma Lymphoma

<35 50,103 1 1 119,174 2 2

35–39 31,253 2 1 66,573 2 2
40–44 39,991 3 2 66,937 2 2
45–49 50,727 3 3 61,141 0 0
50–54 63,495 6 4 53,782 3 3
55–59 73,109 4 4 47,115 4 4
60–64 76,830 9 9 41,019 4 3
65–69 74,314 14 13 33,865 15 11
70–74 58,446 19 19 21,880 15 15
75–79 37,956 17 16 9,712 5 5

80 35,138 12 12 4,494 4 4
Total 591,359 90 84 525,691 56 51
a
Because of rounding, column totals for person-time differ slightly from the sums of rows.
b
Japanese males who were aged 15–64 years and present in Hiroshima or Nagasaki at the
time of the bombings.

Hodgkin’s disease, these estimated associations were ciation between dose and malignant lymphoma mortality
slightly smaller in magnitude. The SRS cohort included during the periods 5–35 years postexposure and 36–55 years
a single death due to malignant lymphoma among black postexposure. There was no evidence of association 5–35
workers; upon restriction to nonblack workers, the esti- years after exposure (ERR/Sv ¼ 0.03, 90% CI: ND, 1.15;
mated ERRs of malignant lymphoma per sievert under LRT ¼ 0.00, P ¼ 0.96); however, there was a positive
5- and 10-year lag assumptions were 7.10 (90% CI: 1.00,
18.66) and 8.18 (90% CI: 1.44, 21.16), respectively.
When the LSS data were limited to survivors with doses Table 2. Estimated Association Between Lymphoma Mortality and
in the range of 0–<0.5 Sv, estimates of radiation-lymphoma Ionizing Radiation Dose Under 5- and 10-Year Exposure Lags
mortality associations were of greater magnitude than esti- Among Male Atomic Bomb Survivors (1950–2000) and Male
mates obtained from model fittings over the entire dose Workers at the Savannah River Site (1950–2002), Japan and South
Carolina
range. Under a 5-year lag assumption, the estimated ERRs
of malignant lymphoma and NHL per sievert were 3.02 Atomic Bomb Savannah River
(90% CI: 0.33, 7.22) and 2.86 (90% CI: 0.10, 7.24), respec- Exposure Survivorsa Site Workers
Lag and
tively. While this suggests nonlinearity in the dose-response ERR Malignant Non-Hodgkin’s Malignant Non-Hodgkin’s
association, comparison of a linear-quadratic dose-response Lymphoma Lymphoma Lymphoma Lymphoma
function with a purely linear dose-response function indi- 5 years
cated that inclusion of a quadratic term resulted in very little ERR per 0.79 0.86 6.99 6.45
improvement in model fit (LRT ¼ 0.07, 1 df; P ¼ 0.79). Sv
Under a 10-year lag assumption, the estimated ERRs of 90% CI 0.10, 1.88 0.13, 2.03 0.96, 18.39 0.48, 17.95
malignant lymphoma and NHL per sievert were 4.54 P value b
0.05 0.04 0.04 0.07
(90% CI: 1.16, 9.93) and 4.24 (90% CI: 0.83, 9.76),
10 years
respectively.
In the LSS, there was no evidence of an association be- ERR per 1.06 1.12 8.18 7.62
Sv
tween radiation dose and lymphoma mortality during the
periods 5–25 years or 26–35 years after irradiation (Table 3). 90% CI 0.24, 2.38 0.26, 2.51 1.44, 21.16 0.93, 20.77
Positive associations between lymphoma mortality and P value 0.02 0.02 0.03 0.05
dose were observed during the periods 36–45 years and 46– Abbreviations: CI, confidence interval; ERR, excess relative rate.
55 years after irradiation. Analyses of associations between a
Japanese males who were aged 15–64 years and present in
radiation dose and NHL led to risk estimates similar to those Hiroshima or Nagasaki at the time of the bombings.
obtained via analyses of all malignant lymphoma (Table 3). b
P value from a likelihood ratio test that the reported parameter for
In a nested model, defined post hoc, we evaluated the asso- the estimated ERR was equal to 0.

Am J Epidemiol 2009;169:969–976
 Ionizing Radiation and Lymphoma Mortality 973

Table 3. Estimated Association Between Radiation Dose and Lymphoma Mortality Among
Male Atomic Bomb Survivors,a by Time Since Exposure, Hiroshima and Nagasaki, Japan,
1950–2000

Time Since Exposure, years

Lymphoma Type (Calendar Period)
and ERR 5–25 26–35 36–45 46–55
(1950–1970) (1971–1980) (1981–1990) (1991–2000)

Malignant lymphoma
ERR per Sv 0.08 0.10 2.23 1.70
90% CI ND, ND ND, ND 0.09, 6.91 0.16, 5.36
P valueb 0.89 0.91 0.08 0.05
No. of deaths 31 20 16 23
Non-Hodgkin’s lymphoma
ERR per Sv 0.17 0.10 2.23 1.70
90% CI ND, ND ND, ND 0.09, 6.91 0.16, 5.36
P value 0.79 0.91 0.08 0.05
No. of deaths 25 20 16 23

Abbreviations: CI, confidence interval; ERR, excess relative rate; ND, not determined.
a
Japanese males who were aged 15–64 years and present in Hiroshima or Nagasaki at the
time of the bombings.
b
P value from a likelihood ratio test that the reported parameter for the estimated ERR was
equal to 0.

association between dose and lymphoma mortality
36 Table 5 shows observed and expected numbers of malig-
years after exposure (ERR/Sv ¼ 1.93, 90% CI: 0.48, 4.66; nant lymphoma deaths by dose category under 5- and
LRT ¼ 6.83, P < 0.01). 10-year lag assumptions. The distribution of events among
In analyses of the SRS cohort, there was a highly impre- SRS workers with respect to dose was relatively narrow in
cise positive association between lymphoma mortality and comparison with the LSS data. Over the dose range at which
doses accrued during the periods 5–25 and 26–35 years the ratio of observed to expected numbers of malignant
prior. The association with doses accrued
36 years prior lymphoma deaths could be compared in these 2 cohorts
was of the largest magnitude and contributed most to the (i.e., 0–<0.5 Sv), these ratios were similar in magnitude
goodness of model fit. The estimated dose-response associ- for analyses of the 2 cohorts, although values tended to be
ation within each exposure time window was based on the slightly greater for the SRS cohort than for the LSS cohort.
total number of lymphoma deaths. Similar estimates were Ratios of observed to expected numbers of deaths were
obtained in analyses restricted to NHL (Table 4).
When the LSS data were limited to those survivors with
doses in the range of 0–<0.5 Sv, there were positive, albeit
Table 4. Estimated Association Between Radiation Dose and
imprecise, estimates of association between radiation dose
Lymphoma Mortality Among Male Workers at the Savannah River
and malignant lymphoma mortality during the periods 5–25 Site, by Time Since Exposure, South Carolina, 1950–2002
years after irradiation (ERR/Sv ¼ 0.64, 90% CI: 1.69,
5.94; LRT ¼ 0.1, P ¼ 0.75), 26–35 years after irradiation Lymphoma Type Time Since Exposure, years
(ERR/Sv ¼ 2.52, 90% CI: 1.48, 11.71; LRT ¼ 0.7, P ¼ and ERR 5–25 26–35 36–52
0.40), 36–45 years (ERR/Sv ¼ 7.08, 90% CI: 0.08, 22.86;
Malignant lymphoma
LRT ¼ 2.6, P ¼ 0.11), and 46–55 years after irradiation
(ERR/Sv ¼ 6.42, 90% CI: 0.22, 23.11; LRT ¼ 2.4, P ¼ ERR per Sv 1.18 4.06 33.28
0.12). Results for analyses of NHL were similar to those for 90% CI ND, ND ND, 25.34 4.83, 107.9
all lymphoma mortality. There was a negative association P valuea 0.85 0.64 0.03
between radiation dose and NHL mortality during the period Non-Hodgkin’s lymphoma
5–25 years after irradiation (ERR/Sv ¼ 0.41, 90% CI: ND, ERR per Sv 1.51 0.58 38.35
5.00; LRT ¼ 0.03, P ¼ 0.85) and positive associations be-
90% CI ND, 16.02 ND, 22.83 7.02, 121.57
tween radiation dose and mortality during the periods 26–
35 years after irradiation (ERR/Sv ¼ 2.46, 90% CI: 1.50, P value 0.80 0.95 0.02
11.55; LRT ¼ 0.68, P ¼ 0.41), 36–45 years after irradia- Abbreviations: CI, confidence interval; ERR, excess relative rate;
tion (ERR/Sv ¼ 7.07, 90% CI: 0.08, 22.83; LRT ¼ 2.61, ND, not determined.
P ¼ 0.11), and 46–55 years after irradiation (ERR/Sv ¼ a
P value from a likelihood ratio test that the reported parameter for
6.42, 90% CI: 0.23, 23.11; LRT ¼ 2.41, P ¼ 0.12). the estimated ERR was equal to 0.

Am J Epidemiol 2009;169:969–976
974 Richardson et al.

Table 5. Observed and Expected Numbers of Deaths Due to Malignant Lymphoma Among Male Atomic Bomb
Survivors (1950–2000) and Male Workers at the Savannah River Site (1950–2002), by Radiation Dose, Japan and
South Carolinaa

Radiation Dose, Sv
Assumed Lag and Cohort
<0.005 0.005–<0.10 0.10–<0.20 0.20–<0.50 0.50–<1 1–<2 ‡2

5-year lag
Atomic bomb survivorsb
No. of deaths observed 32 29 8 11 3 5 2
c
Obs/Exp ratio 0.80 0.97 1.33 1.61 0.72 2.04 2.60
Mean dose, Sv 0.001 0.032 0.141 0.322 0.721 1.340 2.392
Person-years of follow-up 260,641 195,354 38,255 45,932 28,566 16,674 5,937
Savannah River Site workers
No. of deaths observed 20 24 7 5 0 0 0
Obs/Exp ratio 0.77 1.01 1.78 2.14
Mean dose, Sv 0.001 0.028 0.142 0.266
Person-years of follow-up 305,131 181,767 25,961 12,830 0 0 0
10-year lag
Atomic bomb survivors
No. of deaths observed 27 27 8 11 3 5 2
Obs/Exp ratio 0.73 0.97 1.44 1.73 0.78 2.19 2.73
Mean dose, Sv 0.001 0.032 0.141 0.322 0.722 1.338 2.392
Person-years of follow-up 213,808 160,274 31,330 37,840 23,545 13,827 4,926
Savannah River Site workers
No. of deaths observed 21 24 6 5 0 0 0
Obs/Exp ratio 0.77 1.05 1.60 2.35
Mean dose, Sv 0.001 0.028 0.141 0.264
Person-years of follow-up 344,948 149,706 21,197 9,840 0 0 0

Abbreviations: Exp, expected; Obs, observed.

a
Because of rounding, some column totals for person-time differ slightly from the sums of rows.
b
Japanese males who were aged 15–64 years and present in Hiroshima or Nagasaki at the time of the bombings.
c
Ratio of the number of deaths observed to the number of deaths expected.

minimally affected by exclusion of deaths due to Hodgkin’s
disease (results not shown).
DISCUSSION
In a previous analysis of lymphoma mortality among sur-
vivors in the LSS, Pierce et al. (17) reported evidence of
a nonsignificant positive association with radiation dose
among males (ERR/Sv ¼ 0.27, 90% CI: ND, 1.49) and
a nonsignificant negative association among females
(ERR/Sv ¼ 0.17, 90% CI: ND, 0.30). In those analyses,
a time-constant ERR model was fitted to mortality follow-
up through 1990. In the present paper, time-window ana-
lyses helped to explain the observation of a significant positive
association between radiation dose and lymphoma mortality
among male atomic bomb survivors with more recent fol-
low-up, showing that positive associations have been ob-
served only since 1980. Such findings suggest a protracted
induction and latency period. If considered within the
framework of a multistage model of carcinogenesis, the
relatively long empirical induction period for lymphoma
following radiation exposure may be consistent with action
at an early stage of a multistage process.
The point estimates for the radiation dose-lymphoma
mortality association under 5- and 10-year lag assumptions
derived from analysis of the SRS cohort are larger than the
estimates derived from analysis of the LSS cohort (Table 2).
Differences in the magnitude and rate of exposure may in-
fluence the comparability of dose-response estimates. These
cohorts also differ with regard to potential biases from con-
founding, selection, and exposure measurement error. While
it is not an established cause of NHL, benzene is suspected
to be related to NHL (28). However, benzene was not used
in the production process at SRS, nor was it routinely used
as a degreaser. Plutonium-239 is a radiologic hazard at SRS.
While a recent study suggested that the contribution of plu-
tonium doses to total dose estimates for these workers was
relatively small (29), we did not directly assess confounding
by plutonium exposure. Selection bias could have influ-
enced these estimates of association—for example, via the
‘‘healthy worker’’ survivor effect (20). Although we ad-
justed for employment status, such an approach is sub-
optimal if employment status is an intermediate variable

Am J Epidemiol 2009;169:969–976

as well as a confounder of the association of interest. How-
ever, in studies of chronic diseases with long latency peri-
ods, cumulative exposure will typically not appreciably
influence employment termination rates; under such condi-
tions, employment status will play a minor role as an
intermediate variable but could have a strong role as a con-
founder of the association (22). Frequent reading of dosim-
eters could have led to dose underestimation if dosimeters
were not sufficiently exposed to reach a minimum detect-
able dose. However, prior work suggests that the impact of
this source of measurement error on estimates of radiation
dose-response trends is modest (30–32).
Problems of bias could also influence estimates of
radiation-mortality associations among atomic bomb survi-
vors. DS02 estimates account for the initial radiation
released from the detonation of the weapons but not radia-
tion from fallout or neutron activation of the ground and
structures (33). The available data suggest that most people
in Hiroshima and Nagasaki had low cumulative external
doses from fallout, with maximum estimates in the range
of 0.2–0.4 Sv for several hundred people who were in an
area of Nagasaki approximately 3 km from the hypocenter
(33, 34). Selective survival in the LSS cohort is another
concern and is a generic consideration when trying to un-
derstand the temporal evolution of exposure-related risk
(35). A relation between short-term survival after the bomb-
ings and later risk of lymphoma could lead to bias in
dose-response estimates. Evidence of selection has been
suggested by some empirical analyses (36, 37); however,
values for the magnitude of dose-related selective survival
assumed in a recent study suggested a modest potential for
bias in dose-response estimates (38).
These analyses provide evidence of a positive association
between ionizing radiation dose and malignant lymphoma
mortality among male Japanese atomic bomb survivors and
SRS workers. We did not address risk estimates for females,
for whom there was no evidence of a positive association
between radiation dose and lymphoma mortality in follow-
up through 1990 (17). The radiation-NHL mortality associ-
ations among these male atomic bomb survivors and SRS
workers are of larger magnitude than the estimate reported
in a 15-country study of nuclear workers (under a 10-year
lag assumption, ERR/Sv ¼ 0.44, 90% CI: <0, 4.78) (7);
however, in the current analyses, positive dose-response as-
sociations were primarily observed more than 35 years after
irradiation. These findings underscore the importance of
continued follow-up of the LSS cohort and nuclear worker
cohorts.
ACKNOWLEDGMENTS
Author affiliations: Department of Epidemiology, School
of Public Health, University of North Carolina at Chapel
Hill, Chapel Hill, North Carolina (David Richardson, Steve
Wing); Department of Epidemiology, Radiation Effects Re-
search Foundation, Hiroshima, Japan (Hiromi Sugiyama,
Ritsu Sakata, Eric J. Grant, Yukiko Shimizu, Nobuo Nishi,
Fumiyoshi Kasagi); Department of Statistics, Radiation
Am J Epidemiol 2009;169:969–976
Ionizing Radiation and Lymphoma Mortality 975
Effects Research Foundation, Hiroshima, Japan (Susan Geyer);
Department of Epidemiology, Radiation Effects Research
Foundation, Nagasaki, Japan (Midori Soda, Akihiko Suyama);
and Radiation Effects Research Foundation, Hiroshima, Japan
(Kazunori Kodama).
This project was supported in part by grant R01
OH007871 from the US National Institute for Occupational
Safety and Health. Support was also provided by the Radi-
ation Effects Research Foundation (Hiroshima and Naga-
saki, Japan), which is funded by the Japan Ministry of
Health, Labour and Welfare and the US Department of En-
ergy, the latter partly through the National Academy of
Sciences.
Conflict of interest: none declared.
REFERENCES
1. Boice JD Jr. Radiation and non-Hodgkin’s lymphoma. Cancer
Res. 1992;52(19 suppl):5489s–5491s.
2. United Nations Scientific Committee on the Effects of Atomic
Radiation. Sources and Effects of Ionizing Radiation. New
York, NY: United Nations; 2000.
3. Ron E. Ionizing radiation and cancer risk: evidence from
epidemiology. Radiat Res. 1998;150(5 suppl):S30–S41.
4. Melbye M, Trichopoulos D. Non-Hodgkin’s lymphomas. In:
Adami H-O, Hunter D, Trichopoulos D, eds. Textbook of
Cancer Epidemiology. New York, NY: Oxford University
Press; 2002:535–555.
5. Hartge P, Smith MT. Environmental and behavioral factors and
the risk of non-Hodgkin lymphoma. Cancer Epidemiol Bio-
markers Prev. 2007;16(3):367–368.
6. Hartge P, Wang SS, Bracci PM, et al. Non-Hodgkin lym-
phoma. In: Schottenfeld D, Fraumeni JF, eds. Cancer Epide-
miology and Prevention. New York, NY: Oxford University
Press; 2006:898–918.
7. Cardis E, Vrijheid M, Blettner M, et al. Risk of cancer after
low doses of ionising radiation: retrospective cohort study in
15 countries. BMJ. 2005;331(7508):77.
8. Muirhead CR, Goodill AA, Haylock RG, et al. Occupational
radiation exposure and mortality: second analysis of the Na-
tional Registry for Radiation Workers. J Radiol Prot. 1999;
19(1):3–26.
9. Schubauer-Berigan M, Daniels RD, Fleming DA, et al. Risk of
chronic myeloid and acute leukemia mortality after exposure
to ionizing radiation among workers at four U.S. nuclear
weapons facilities and a nuclear naval shipyard. Radiat Res.
2007;167(2):222–232.
10. Preston DL, Pierce DA, Shimizu Y, et al. Effect of recent
changes in atomic bomb survivor dosimetry on cancer
mortality risk estimates. Radiat Res. 2004;162(4):
377–389.
11. Richardson DB, Wing S. Leukemia mortality among workers
at the Savannah River Site. Am J Epidemiol. 2007;166(9):
1015–1022.
12. Pierce DA, Stram DO, Vaeth M. Allowing for random errors in
radiation dose estimates for the atomic bomb survivor data.
Radiat Res. 1990;123(3):275–284.
13. Preston DL, Ron E, Tokuoka S, et al. Solid cancer incidence in
atomic bomb survivors: 1958–1998. Radiat Res. 2007;
168(1):1–64.
14. International Commission on Radiological Protection. 1990
Recommendations of the International Commission on
976 Richardson et al.
Radiological Protection. Oxford, United Kingdom: Pergamon
Press; 1991.
15. National Institute for Occupational Safety and Health.
Changes to the dose reconstruction target organ selection for
lymphoma under the Energy Employees Occupational Illness
Compensation Program Act of 2000. Fed Regist. 2006;71(31):
7969–7970.
16. Office of Compensation Analysis and Support, National In-
stitute for Occupational Safety and Health. Selection for In-
ternal and External Dosimetry Target Organs for Lymphatic/
Hematopoietic Cancers. (Technical Information Bulletin).
Cincinnati, OH: National Institute for Occupational Safety and
Health; 2006. (Document no. OCAS-TIB-0012).
17. Pierce DA, Shimizu Y, Preston DL, et al. Studies of the mor-
tality of atomic bomb survivors. Report 12, part I. Cancer:
1950–1990. Radiat Res. 1996;146(1):1–27.
18. Richardson DB, Wing S, Daniels RD. Evaluation of external
radiation dosimetry records at the Savannah River Site,
1951–1989. J Expo Sci Environ Epidemiol. 2007;17(1):
13–24.
19. Shimizu Y, Kato H, Schull WJ. Studies of the mortality of
A-bomb survivors. 9. Mortality, 1950–1985: part 2. Cancer
mortality based on the recently revised doses (DS86). Radiat
Res. 1990;121(2):120–141.
20. Arrighi HM, Hertz-Picciotto I. The evolving concept of the
healthy worker survivor effect. Epidemiology. 1994;5(2):
189–196.
21. Steenland K, Stayner L. The importance of employment status
in occupational cohort mortality studies. Epidemiology. 1991;
2(6):418–423.
22. Steenland K, Deddens J, Salvan A, et al. Negative bias in
exposure-response trends in occupational studies: modeling
the healthy workers survivor effect. Am J Epidemiol. 1996;
143(2):202–210.
23. Thomas D. General relative-risk models for survival time and
matched case-control analysis. Biometrics. 1981;37(4):
673–686.
24. Greenland S. Introduction to regression models. In:
Rothman K, Greenland S, eds. Modern Epidemiology. 2nd ed.
Philadelphia, PA: Lippincott, Williams & Wilkins; 1998:
359–399.
25. Gilbert ES, Cragle DL, Wiggs LD. Updated analyses of
combined mortality data for workers at the Hanford Site, Oak
Ridge National Laboratory, and Rocky Flats Weapons Plant.
Radiat Res. 1993;136(3):408–421.
26. Cardis E, Vrijheid M, Blettner M, et al. The 15-Country
Collaborative Study of Cancer Risk among Radiation Workers
in the Nuclear Industry: estimates of radiation related cancer
risks. Radiat Res. 2007;167(4):396–416.
27. Preston DL, Lubin JH, Pierce DA, et al. Epicure: User’s
Guide. Seattle, WA: Hirosoft International Corporation; 1993.
28. Smith MT, Jones RM, Smith AH. Benzene exposure and risk
of non-Hodgkin lymphoma. Cancer Epidemiol Biomarkers
Prev. 2007;16(3):385–391.
29. Daniels RD, Lodwick CJ, Schubauer-Berigan MK, et al.
Assessment of plutonium exposures for an epidemiological
study of US nuclear workers. Radiat Prot Dosimetry. 2006;
118(1):43–55.
30. Daniels RD, Yiin JH. A comparison of statistical methods for
estimation of less than detectable ionising radiation exposures.
Radiat Prot Dosimetry. 2006;121(3):240–51.
31. Daniels RD, Schubauer-Berigan MK. Bias and uncertainty of
penetrating photon dose measured by film dosimeters in an
epidemiological study of US nuclear workers. Radiat Prot
Dosimetry. 2005;113(3):275–289.
32. Shin H, Ramsay T, Krewski D, et al. The effect of censoring on
cancer risk estimates based on the Canadian National Dose
Registry of occupational radiation exposure. J Expo Anal
Environ Epidemiol. 2005;15(5):398–406.
33. Committee for the Compilation of Materials on Damage
Caused by the Atomic Bombs in Hiroshima and Nagasaki.
Hiroshima and Nagasaki: The Physical, Medical, and Social
Effects of the Atomic Bombings. Tokyo, Japan: Iwanami
Shoten, Publishers; 1981.
34. Radiation Effects Research Foundation. Frequently Asked
Questions. Hiroshima, Japan: Radiation Effects Research
Foundation; 2007. (http://www.rerf.or.jp/general/qa_e/
qa12.html). (Accessed June 30, 2008).
35. Flanders WD, Klein M. Properties of 2 counterfactual effect
definitions of a point exposure. Epidemiology. 2007;18(4):
453–460.
36. Stewart AM, Kneale GW. A-bomb survivors: factors that may
lead to a re-assessment of the radiation hazard. Int J Epide-
miol. 2000;29(4):708–714.
37. Shimizu Y, Pierce DA, Preston DL, et al. Studies of the mor-
tality of atomic bomb survivors. Report 12, part II. Noncancer
mortality: 1950–1990. Radiat Res. 1999;152(4):374–389.
38. Pierce DA, Vaeth M, Shimizu Y. Selection bias in cancer risk
estimation from A-bomb survivors. Radiat Res. 2007;167(6):
735–741.
Am J Epidemiol 2009;169:969–976
 American Journal of Epidemiology Vol. 169, No. 8
ª The Author 2009. Published by the Johns Hopkins Bloomberg School of Public Health. DOI: 10.1093/aje/kwp008
All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org. Advance Access publication March 3, 2009

Original Contribution

Biomarker-calibrated Energy and Protein Consumption and Increased Cancer

Risk Among Postmenopausal Women

Ross L. Prentice, Pamela A. Shaw, Sheila A. Bingham, Shirley A. A. Beresford, Bette Caan, Marian
L. Neuhouser, Ruth E. Patterson, Marcia L. Stefanick, Suzanne Satterfield, Cynthia A. Thomson,
Linda Snetselaar, Asha Thomas, and Lesley F. Tinker

Initially submitted October 3, 2008; accepted for publication January 6, 2009.

The authors previously reported equations, derived from the Nutrient Biomarker Study within the Women’s Health
Initiative, that produce calibrated estimates of energy, protein, and percentage of energy from protein consumption
from corresponding food frequency questionnaire estimates and data on other factors, such as body mass index,
age, and ethnicity. Here, these equations were applied to yield calibrated consumption estimates for 21,711 women
enrolled in the Women’s Health Initiative dietary modification trial comparison group and 59,105 women enrolled in
the observational study. These estimates were related prospectively to total and site-specific invasive cancer
incidence (1993–2005). In combined cohort analyses that do not control for body mass, uncalibrated energy was not
associated with total cancer incidence or site-specific cancer incidence for most sites, whereas biomarker-calibrated
energy was positively associated with total cancer (hazard ratio ¼ 1.18, 95% confidence interval: 1.10, 1.27, for 20%
consumption increase), as well as with breast, colon, endometrial, and kidney cancer (respective hazard ratios of
1.24, 1.35, 1.83, and 1.47). Calibrated protein was weakly associated, and calibrated percentage of energy from
protein was inversely associated, with total cancer. Calibrated energy and body mass index associations were highly
interdependent. Implications for the interpretation of nutritional epidemiology studies are described.

bias (epidemiology); biological markers; diet; energy intake; epidemiologic methods; neoplasms; nutrition assess-
ment; proteins

Abbreviations: CI, confidence interval; DM, dietary modification; FFQ, food frequency questionnaire; HR, hazard ratio; WHI,
Women’s Health Initiative.

Early international correlation studies reported a positive
association between energy consumption and the incidence
and mortality from cancer. Among women, associations
were reported for breast, colon, rectal, endometrial, ovarian,
and kidney cancer (1). Rodent feeding experiments indicate
that underfeeding typically inhibits the development of site-
specific and overall cancer (2, 3).
Analytical epidemiologic studies of diet, nutrition, and
cancer date to the 1970s. Initial case-control studies used
a range of dietary assessment procedures, including food
records, recalls, and frequencies. Concern about dietary re-
call bias subsequently led to cohort studies as the predom-
inant design for dietary association studies. Because these
studies typically involve tens of thousands of enrollees,
a self-administered, machine-readable food frequency ques-
tionnaire (FFQ) has been the principal dietary assessment
tool in cohort studies.
However, like other dietary assessment methods, the mea-
surement properties of FFQs remain substantially unknown.
Comparison of FFQ assessments with food records reveals
noteworthy differences (4) that imply an important error
component to self-reported nutrient intake. Small-scale
studies using a doubly labeled water biomarker (5) of en-
ergy consumption suggest important systematic biases also,
as obese persons may systematically underreport energy
consumption (6) in some populations. Measurement error,

Correspondence to Dr. Ross L. Prentice, Division of Public Health Sciences, Fred Hutchinson Cancer Research Center, 1100 Fairview Avenue
North, P.O. Box 19024, Seattle, WA 98109-1024 (e-mail: rprentic@fhcrc.org).

977 Am J Epidemiol 2009;169:977–989

978 Prentice et al.

Table 1. Subject Characteristics for Women in the Women’s Health Initiative Dietary
Modification Trial Comparison Group and Observational Study, 1993–1998

Dietary Modification Observational

Trial Comparison Study
Characteristic Group (n 5 21,711)a (n 5 59,105)a
% No. % No.

Age, yearsb
50–59 30 6,421 19 11,135
60–69 48 10,495 43 25,257
70–79 21 4,667 35 20,555
80–89 1 128 4 2,158
2
Body mass index, kg/m
Normal (<25.0) 26 5,704 42 24,938
Overweight (25.0–29.9) 36 7,767 34 20,361
Obese (
30) 38 8,239 23 13,806
Race
White 82 17,889 86 51,028
Black 10 2,161 6 3,661
Hispanic 3 725 3 1,736
b
Other 4 936 5 2,680
Income (total yearly)
<$20,000 15 3,218 14 8,159
$20,000–$34,999 25 5,335 23 13,605
$35,000–$49,999 21 4,593 21 12,214
$50,000–$74,999 21 4,546 21 12,407

$75,000 18 4,009 22 12,720
Education
Less than high school diploma 4 893 4 2,196
High school diploma or equivalent 18 3,803 16 9,379
School after high school 40 8,593 36 21,421
College degree or higher 39 8,422 44 26,109
Smoking
Current 6 1,392 6 3,307
Past 52 11,373 51 30,232
Never 41 8,946 43 25,566
Table continues

especially systematic biases, may substantially distort diet
and cancer associations. It is important to examine nutrient
and disease associations in a manner that appropriately ac-
commodates FFQ measurement errors.
The accumulated data on diet and cancer were reviewed
by an international panel of experts in 1997 (7). Rather few
‘‘definite’’ or ‘‘probable’’ dietary associations emerged. The
authors wrote, ‘‘The significance of the data on energy in-
take and cancer risk in humans remains unclear’’ (7, p. 371),
and ‘‘In the view of the panel, the effect of energy intake on
cancer is best assessed by examining the data on related
factors: rate of growth, body mass, and physical activity’’
(7, p. 371). This state of affairs has evidently not changed in
the intervening decade (8) and reflects considerable uncer-
tainty about energy consumption estimates and related as-
sociation study findings. The 1997 panel also assessed that
protein consumption was not ‘‘probably or convincingly’’
related to the risk of any cancer (7, p. 394).
Good-quality biomarkers of both total energy consump-
tion (5) and protein consumption (9) have been developed
but, for cost and logistics reasons, have received little use in
epidemiologic research. These biomarkers involve urinary
recovery of metabolites produced when these nutrients are
expended. In weight-stable persons, they provide objective
estimates of short-term energy and protein consumption.
The associated measurement error plausibly adheres to
a simple classical measurement model,
W ¼ Z þ e; ð1Þ
where Z is the targeted (log-transformed) nutrient consump-
tion, W is the (log-transformed) biomarker-measured

Am J Epidemiol 2009;169:977–989
 Biomarker-calibrated Energy and Cancer 979

Table 1. Continued

Dietary Modification Observational

Trial Comparison Study
Characteristic Group (n 5 21,711)a (n 5 59,105)a
% No. % No.

Recreational physical activity,

metabolic equivalents/week
<1.5 25 5,335 16 9,508
1.5–6.2 25 5,378 20 11,715
6.3–14.7 26 5,541 27 15,708

14.8 25 5,457 38 22,174
Breast cancer family history, yes 18 3,729 19 10,631
Gail 5-year risk score, %
<1.00 15 3,355 11 6,778
1.00–1.99 62 13,368 62 36,531
2.00–2.99 14 3,073 16 9,623

3.00 9 1,915 10 6,173
Colon cancer family history, yes 16 3,257 17 9,006
History of polyps, yes 8 1,780 9 5,275
Unopposed estrogen use ever, yes 37 8,084 38 22,736
Estrogen þ progesterone 28 6,054 31 18,395
use ever, yes
Diabetes, yes 6 1,313 5 2,664
Hypertension, yes 41 8,909 37 22,029
Alcohol use
Nondrinker 10 2,086 10 6,063
Current drinker
<1 drink/week 36 7,708 32 18,768
1–7 drinks/week 27 5,923 27 16,048
>7 drinks/week 10 2,116 14 8,010
Past drinker 18 3,878 17 10,216
a
Number of subjects for whom there were no missing values for the energy regression cali-
bration or for total cancer hazard ratio analysis.
b
Age at food frequency questionnaire measurement (year 1 dietary modification trial compar-
ison group and year 3 observational study).

consumption, and e is measurement error that is assumed to
be independent of Z and of all other study subject character-
istics. The cost to ascertain these biomarkers for each par-
ticipant in a cohort study would be excessive. Instead,
a substudy that includes both the biomarker and FFQ can
be used to produce calibrated consumption estimates for all
cohort members.
The measurement model for the self-reported data typi-
cally needs to be more complex than the classical measure-
ment model (equation 1). Other factors, such as body mass,
ethnicity, and age, may affect the assessment, and measure-
ment errors may be correlated if the assessment is repeated
for specific study subjects. Hence, we consider the measure-
ment model (10, 11),
Q ¼ S0 þ S1 Z þ S2 V þ S3 VZ þ r þ u; ð2Þ
for the (log-transformed) self-reported nutrient assessment
Q, where V is a set of characteristics that may relate to
systematic bias in the assessment, r is a person-specific error
variable that will be present in each self-reported assessment
for a study subject, and u is an independent measurement
error term. In addition, S0, S1, S2, and S3 are constants to be
estimated, and all variables on the right sides of equations 1
and 2 are assumed to be independent, given V.
We have recently reported (12) FFQ measurement error
findings from the Nutritional Biomarker Study among 544
women enrolled in the Women’s Health Initiative (WHI)
dietary modification (DM) trial. FFQ estimates of energy,
protein, and percentage of energy from protein were each
found to incorporate important systematic bias, and corre-
sponding calibration equations were developed. Here, we
use these equations to produce calibrated estimates of en-
ergy, protein, and percentage of energy from protein for
women in the DM trial comparison (control) group and
for women in the WHI observational study. The 2 cohorts

Am J Epidemiol 2009;169:977–989
980 Prentice et al.
will be used, separately and combined, to assess associa-
tions between calibrated nutrient consumption and cancer
incidence as observed during WHI follow-up. Cancer risk
among DM intervention group women may depend in
a complex manner on baseline and follow-up dietary pat-
terns, so that intervention group women were excluded from
the present analyses.
MATERIALS AND METHODS
Study cohorts
Detailed accounts of the design of the WHI Clinical Trial
and Observational Study and of the DM trial findings have
been presented (13–18). This paper uses a subset of women
assigned to the DM trial comparison group (n ¼ 29,294) and
a subset of the observational study cohort (n ¼ 93,676).
Both cohorts included only women who were 50–79 years
of age at recruitment (1993–1998), were postmenopausal,
and had no medical condition associated with less than 3
years’ predicted survival. Both provided common core ques-
tionnaires at baseline on medical history, reproductive history,
family history, personal habits, psychosocial attributes, and
food frequency (19, 20).
DM trial women, who could be assigned to overlapping
trials of postmenopausal hormone therapy and of calcium
and vitamin D supplementation, also satisfied additional
exclusionary criteria. To maximize commonality with the
DM cohort, the 76,987 observational study women consid-
ered here were those remaining after imposing additional
DM trial baseline exclusionary criteria as follows: prior
history of breast cancer, colorectal cancer, or other cancer
(except nonmelanoma skin cancer) within the preceding
10 years; a stroke or myocardial infarction in the preceding
6 months; severe hypertension (systolic blood pressure
>200 mm or diastolic blood pressure >105 mm); already fol-
lowing a low-fat diet; underweight (body mass index <18); or
FFQ-reported daily energy of <600 kcal or >5,000 kcal.
WHI food frequency questionnaire
All DM trial and observational study women completed
FFQs at baseline. DM trial women repeated the FFQ at 1
year following enrollment and approximately every 3 years
thereafter, while observational study women repeated the
FFQ at 3 years following enrollment. FFQs were provided
in connection with visits to the 40 participating clinical
centers, where completeness and quality control checks
were applied. The self-administered FFQ included 122 line
items for individual foods/food groups and 19 adjustment
items regarding fat intake, as well as summary questions.
Nutrition Data System for Research, version 2005, software
(University of Minnesota, Minneapolis, Minnesota) was
used to compute daily average nutrient consumption esti-
mates (21, 22).
Nutrient Biomarker Study
The WHI Nutrient Biomarker Study was conducted in
2004–2005 to assess measurement properties of this FFQ
and to produce calibrated consumption estimates for energy
and protein. The eligibility and recruitment methods for the
Nutrient Biomarker Study have been described (12); 544
representative women from the DM trial cohort were en-
rolled (276 comparison group, 268 intervention group).
These weight-stable women participated in a doubly labeled
water protocol to estimate daily total energy expenditure
over a 2-week period, as well as a urinary nitrogen protocol
to estimate daily protein consumption over a 24-hour period,
and also provided a concurrent FFQ and other questionnaire
data. Twenty percent (n ¼ 111) repeated the entire Nutrient
Biomarker Study protocol an average of 6 months later to
provide reliability data for measurement error component
estimation (12). FFQ total energy and protein were found to
be underestimated, while the percentage of energy from
protein was overestimated. Women having high body mass
index (weight (kg)/height (m)2) and younger women under-
estimated energy consumption to a comparatively greater
extent. Calibration equations were developed for each of
energy, protein, and percentage of energy from protein by
linear regression of log-biomarker estimates on correspond-
ing log-FFQ estimates, body mass index, age, ethnicity,
and other factors (12). For example, the calibrated log-
energy consumption is given by 7.61 þ 0.062 (log-FFQ
energy  7.27) þ 0.013 (body mass index  28.2) 
0.005 (age  70.9 years), plus some less influential terms
involving ethnicity, family income, and physical activity.
DM intervention group assignment did not meet inclusion
criteria for any of the 3 calibration equations.
Nutrient Biomarker Study application to WHI cohorts
Here, we apply these calibration equations to FFQ data
that were collected earlier in the WHI and relate the cali-
brated consumption estimates to subsequent cancer inci-
dence. Doing so is complicated by the use of the FFQ in
participant screening for the DM trial. The exclusion of
about 50% of the women who had baseline FFQ percent-
age of energy from fat of less than 32, in conjunction with
FFQ measurement error, implies that the baseline FFQ
percentage of energy from fat is overestimated in the
DM trial (by about 3% on average), with corresponding
estimates of energy likewise distorted. Observational
study baseline estimates are distorted in the opposite di-
rection because many women screened out from the DM
trial enrolled in the observational study. In terms of equa-
tion 2, these distortions arise because women tend to meet
the FFQ inclusion criteria when the independent random
error term (u) that attends a particular FFQ application is
positive. Later FFQs for a woman, following a sufficient
period of time (e.g., 6 months) to avoid carry-over effects
on this measurement component, can be expected to be
free of this measurement effect. Hence, our analyses rely
on FFQs obtained at year 1 in the DM trial comparison
group and at year 3 in the observational study, and only
cancer diagnoses that follow these FFQ collections are
included in analyses. These FFQs were collected an aver-
age 6.5 years (DM trial comparison group) and 4 years
(observational study) prior to the Nutrient Biomarker
Study data collection.
Am J Epidemiol 2009;169:977–989
 Biomarker-calibrated Energy and Cancer 981

Table 2. Incidence of Invasive Cancer in the Women’s Health Initiative Following Year 1
(Dietary Modification Trial Comparison Group) and Year 3 (Observational Study) Food
Frequency Data Collection, 1993–2005

Dietary Modification
Observational Study Total
Trial Comparison
(n 5 59,105)a (N 5 80,816)a
Cancer Group (n 5 21,711)a
Incidence/1,000 No. of Incidence/1,000 No. of Incidence/1,000 No. of
Person-Years Cases Person-Years Cases Person-Years Cases

Total cancerb 12.34 1,807 11.06 3,234 11.48 5,041

Breast 4.98 685 4.73 1,018 4.83 1,703
Colon 0.89 123 0.87 240 0.88 363
Rectum 0.33 47 0.14 40 0.21 87
Ovary 0.63 72 0.57 131 0.59 203
Endometrium 1.32 115 1.21 220 1.25 335
Bladder 0.25 39 0.20 60 0.22 99
Kidney 0.28 42 0.27 81 0.27 123
Pancreas 0.26 40 0.23 71 0.24 111
Lung 0.95 146 0.91 275 0.92 421
Lymphoma 0.57 88 0.57 175 0.57 263
Leukemia 0.32 49 0.20 60 0.24 109
a
The number of subjects in the cohort for whom there were no missing values for the energy
calibration or for total cancer hazard ratio analysis. The number of subjects with no missing values
varied slightly by cancer site and nutrient.
b
Exclusive of nonmelanoma skin cancer.

Dietary consumption and disease risk associations were
estimated for total invasive cancer, as well as for invasive
cancers of the breast, colon, rectum, ovary, endometrium,
bladder, kidney, pancreas, and lung, and for lymphoma and
leukemia. The ovarian cancer analyses were restricted to
women without bilateral oophorectomy at baseline, and
the endometrial cancer analysis was restricted to women
with a uterus at baseline.
DM comparison group women were queried twice per
year, and observational study women annually, concerning
diagnosis of any cancer other than nonmelanoma skin can-
cer. Cancer reports were verified by medical record and
pathology report review by centrally trained physician ad-
judicators at participating clinical centers (23).
Statistical analyses
Log-consumption estimates were calibrated directly from
the biomarker assessments (equation 1) for the few women
included in the Nutrient Biomarker Study and for other
women by using the calibration equations previously devel-
oped (12).
Hazard ratio estimates were based on Cox regression
(24). Follow-up times extended from year 1 (DM trial com-
parison group) or year 3 (observational study) to the earliest
of cancer occurrence, death, lost to follow-up, or March 31,
2005, when the intervention phase of WHI ended. To min-
imize mammographic screening influences on results, the
breast cancer analyses censored the follow-up time for
a woman the first time she exceeded 2 years without a mam-
mogram. The Cox model baseline hazard rates for each
cancer outcome were stratified on baseline age in 5-year
categories and, for the DM trial comparison group, also
on hormone therapy trial participation (active estrogen; es-
trogen placebo; active estrogen plus progestin; estrogen plus
progestin placebo; not randomized). Analyses that combine
the 2 cohorts stratify also on cohort. Analysis for specific
cancer outcomes included standard risk factors in the
Cox regression model to control confounding, as shown in
Appendix Table 1. Women having missing confounding
factors were excluded from analysis.
Principal analyses modeled the log-hazard ratio linearly
on log-nutrient consumption, so that the hazard ratio for
a fractional increase in the nutrient is independent of the
consumption. For display purposes, we present hazard ratios
for a 20% increase in consumption. For a woman with me-
dian consumption, a 20% increment corresponds to about
413 kcal of energy, 15 g of protein, or 2.9 units in percentage
of energy from protein.
Usual Cox model standard error estimates were calcu-
lated for uncalibrated consumption regression coefficients.
A more complex standard error estimation procedure is
needed for the calibrated consumption coefficients to ac-
knowledge uncertainty in the calibration parameter esti-
mates, as well as in the ‘‘regression calibration’’ hazard
ratio estimation procedure (11), which has been shown to
be free of practically important biases in extensive simula-
tion studies. A bootstrap procedure (500 bootstrap samples),
with bootstrap sampling stratified on cohort and member-
ship in the Nutrient Biomarker Study and in the Nutrient
Biomarker Study reliability subset, was applied for cali-
brated standard error estimation. A bootstrap procedure

Am J Epidemiol 2009;169:977–989
982 Prentice et al.

(500 samples) was also used to test the equality of hazard

95% Confidence
Table 3. Geometric Mean Consumption and 95% Confidence Intervals for Uncalibrated Dietary Consumption, as Estimated by the Women’s Health Initiative Food Frequency Questionnaire,
ratios in the DM trial comparison group and observational

11.9, 17.3

11.8, 17.6
and for Calibrated Consumption Using Nutritional Biomarker Data, in the Women’s Health Initiative Dietary Modification Trial Comparison Group and Observational Study, 1993–2005

Interval
study cohorts.

Calibrated
Calibrated energy turns out to be strongly positively cor-
related with body mass index. The data analyzed here do not

Percentage of Energy
allow one to determine whether a high body mass should be

Mean
From Protein
regarded as a consequence of a high-energy diet, in which

14.4

14.4
case body mass index should be excluded from the set of
potential confounding factors to avoid overcorrection, or

95% Confidence
whether a high body mass may arise for other reasons

11.5, 24.0

11.5, 25.0
Interval
(e.g., sedentary lifestyle), in which case energy consumption

Uncalibrated
may be high as a result of related energy requirements, and
body mass index control would be needed in regression
analyses. Hence, we present hazard ratio estimates for
energy and for body mass index separately and jointly.

Mean

16.6

16.9
Two-sided P values are used throughout.

95% Confidence

58.4, 104.4

54.8, 100.5

Calibrated by using the measurement model (equation 1) for women in the Nutrition Biomarker Study and equation 2 otherwise.
Interval
RESULTS

Calibrated
A total of 26,531 (91%) DM trial comparison group
women and 66,788 (87%) observational study women pro-

Protein (g/day)
vided FFQs (year 1 DM, year 3 observational study) and

Mean

78.1

74.2
were without a prior cancer diagnosis during WHI follow-
up. Of these, 21,711 (82%) DM trial comparison group and

95% Confidence
59,105 (88%) observational study women had all the data

26.3, 142.1

24.8, 138.1
Interval
needed for energy calibration and for confounding control

Uncalibrated
for total cancer. Table 1 shows some demographic and life-
style characteristics for these women. Analyses of other
cancer outcomes or other nutrients involve a slightly differ-
ent set of women, because of different confounding factors

Mean

61.2

58.6
and, hence, missing data exclusions.
Table 2 shows incidence rates and the number of invasive

1,786.9, 2,564.2

1,722.3, 2,453.9
95% Confidence
cancers through March 31, 2005, for energy analyses for
each cancer site. Incidence rates are similar between the 2
Interval

cohorts. A total of 5,041 invasive cancers contribute to the

Calibrateda

total cancer analyses, but the number of incident cancers is

<300 for specific cancers other than breast, colon, endome-
trial, and lung.
2,140.6

2,055.8
Energy (kcal/day)

Table 3 shows the geometric mean consumption and 95%

Mean

confidence interval for the consumption of energy, protein,

and percentage of energy from protein for both cohorts, with
95% Confidence

and without calibration. The distribution of calibrated con-

676.6, 3,224.9

641.0, 2,989.4

sumption estimates is similar in the 2 cohorts. The narrower

Interval

confidence intervals for the calibrated versus uncalibrated

Uncalibrated

estimates reflect, in part, smaller variations in actual con-

sumption compared with that assessed by the FFQ.
Table 4 shows hazard ratio estimates for a 20% increase
1,477.2

1,384.3

in total energy consumption under a linear log-hazard ratio

Mean

model that excludes body mass index. A 20% increase cor-

responds to about 2 standard deviations for calibrated en-
ergy and percentage of energy from protein and about 1.3
trial comparison group

standard deviations for calibrated protein. For comparison,

Observational study
Dietary modification

extreme quartile medians differ by about 2.3 standard devi-

(n ¼ 21,711)

ations, and extreme tertile medians differ by about 1.9 stan-

(n ¼ 59,105)

dard deviations, for normally distributed exposures.

Separate hazard ratio estimates are given for the DM trial
comparison group and observational study cohorts, without
and with biomarker calibration of consumption estimates.
a

Biomarker calibration clearly has a major impact on hazard

Am J Epidemiol 2009;169:977–989
 Biomarker-calibrated Energy and Cancer 983

Table 4. Hazard Ratio Estimates for a 20% Increase in Energy (kcal/day) Consumption in the Women’s Health Initiative Dietary Modification Trial
Comparison Group and Observational Study, Without and With Biomarker Calibration, 1993–2005

Dietary Modification Trial Observational Test of Equality of

Comparison Group Study Hazard Ratios
Cancer Uncalibrated Calibrated Uncalibrated Calibrated
Uncalibrated Calibrated
Hazard 95% Confidence Hazard 95% Confidence Hazard 95% Confidence Hazard 95% Confidence P Valuea P Valuea
Ratiob Interval Ratiob Intervalc Ratiob Interval Ratiob Inervalc

Total cancer 1.00 0.98, 1.02 1.13 1.02, 1.26 1.01 0.99, 1.03 1.21 1.11, 1.32 0.52 0.30
Breast 0.99 0.95, 1.02 1.25 1.07, 1.47 1.02 0.99, 1.05 1.23 1.06, 1.41 0.20 0.85
Colon 0.93 0.86, 1.00 1.11 0.75, 1.66 0.96 0.91, 1.02 1.47 1.11, 1.94 0.44 0.26
Rectum 1.10 0.96, 1.26 1.00 0.49, 2.02 1.00 0.87, 1.14 1.52 0.94, 2.47 0.30 0.34
Ovary 0.98 0.89, 1.09 1.00 0.61, 1.63 1.04 0.96, 1.12 1.09 0.71, 1.65 0.42 0.80
Endometrium 1.00 0.92, 1.09 1.73 1.21, 2.49 1.07 1.00, 1.14 1.88 1.48, 2.39 0.21 0.69
Bladder 0.99 0.87, 1.13 1.07 0.58, 1.97 1.10 0.98, 1.23 1.27 0.82, 1.97 0.26 0.70
Kidney 1.14 1.00, 1.30 1.87 0.95, 3.68 1.00 0.90, 1.11 1.28 0.81, 2.05 0.11 0.42
Pancreas 1.02 0.90, 1.16 1.72 1.09, 2.73 1.01 0.91, 1.12 1.02 0.49, 2.10 0.88 0.22
Lung 0.99 0.93, 1.06 1.01 0.72, 1.42 0.97 0.93, 1.03 0.76 0.55, 1.06 0.73 0.26
Lymphoma 0.96 0.88, 1.04 0.75 0.47, 1.23 0.98 0.92, 1.05 0.75 0.53, 1.08 0.69 0.97
Leukemia 0.97 0.86, 1.10 0.90 0.52, 1.56 1.14 1.01, 1.28 1.93 1.15, 3.21 0.07 0.05
a
P value based on the difference between log-hazard ratios from the dietary modification trial comparison group and observational study
cohorts, with a bootstrap estimate of standard deviation for the difference between the calibrated log-hazard ratios.
b
Hazard ratio associated with a 20% increase in daily consumption by considering the hazard ratio for log(1.2x) compared with log(x):
exp(beta)log 1.2, where beta is the estimated coefficient in Cox regression.
c
The 95% confidence intervals for calibrated hazard ratios are based on the log-estimated hazard ratio 6 1.96 3 the bootstrap standard error.

ratio estimates, with evidence for positive associations be- 1.01, 1.19), endometrial cancer (HR ¼ 1.37, 95% CI: 1.16,
tween calibrated energy and total cancer, as well as certain 1.61), and leukemia (HR ¼ 1.39, 95% CI: 1.05, 1.83). These
site-specific cancers, in both the DM trial comparison group positive associations may be substantially attributable to
and observational study cohorts, but with little evidence of correlation between protein and energy consumption, since
association for uncalibrated energy. There is also little evi- the hazard ratio estimates for percentage of energy from
dence of difference in hazard ratios between the 2 cohorts, protein are less than 1 for total and most specific cancers,
with or without calibration, with the possible exception of and the inverse association is significant for total cancer
leukemia.
Figure 1 shows corresponding hazard ratio estimates and
95% confidence intervals from the analysis of the 2 cohorts
combined. Calibrated energy is positively related to total
(hazard ratio (HR) ¼ 1.18, 95% confidence interval (CI):
1.10, 1.27), breast (HR ¼ 1.24, 95% CI: 1.11, 1.38), colon
(HR ¼ 1.35, 95% CI: 1.06, 1.71), endometrial (HR ¼ 1.83,
95% CI: 1.49, 2.25), and kidney (HR ¼ 1.47, 95% CI: 1.00,
2.16) cancer, while uncalibrated energy was not signifi-
cantly related to total cancer or to any specific cancer, with
the exception of an inverse association with colon cancer.
The wider confidence intervals for calibrated versus uncali-
brated energy hazard ratios reflect both uncertainty in the
coefficients of the calibration equations and deattenuation
that arises from acknowledging dietary assessment mea-
surement error in the hazard ratio estimation procedure.
Analyses of calibrated protein and percentage of energy
from protein similarly yielded little evidence of hazard ratio
differences between the 2 cohorts (each P > 0.05). Figures
2 and 3 show corresponding combined cohort hazard ratios Figure 1. Estimated hazard ratios and 95% confidence intervals for
and 95% confidence intervals for a 20% increase in these a 20% increase in energy consumption (kcal/day), from combined
analysis of data from the Women’s Health Initiative dietary modifica-
nutritional factors. The hazard ratios for a 20% increase in tion trial comparison group and observational study, without and with
calibrated protein are above 1 for total cancer (HR ¼ 1.06, biomarker calibration of consumption, 1993–2005. Unfilled square,
95% CI: 1.01, 1.12), breast cancer (HR ¼ 1.09, 95% CI: uncalibrated; filled circle, calibrated.

Am J Epidemiol 2009;169:977–989
984 Prentice et al.
Figure 2. Estimated hazard ratios and 95% confidence intervals for
a 20% increase in protein consumption (g/day), from combined anal-
ysis of data from the Women’s Health Initiative dietary modification
trial comparison group and observational study, without and with bio-
marker calibration of consumption, 1993–2005. Unfilled square, un-
calibrated; filled circle, calibrated.
(HR ¼ 0.92, 95% CI: 0.85, 0.99 for a 20% increase in
percentage of energy from protein). Results corresponding
to Figures 1–3 by quartile of calibrated consumption are
given in Appendix Tables 2–4.
The correlation coefficients for body mass index with log-
transformed energy, protein, and percentage of energy from
protein in the combined cohorts were, respectively, 0.07,
0.10, and 0.07 without calibration and 0.81, 0.46, and
0.12 following calibration. Hence, it may be difficult to
distinguish between total energy and body mass index asso-
ciations, with total or site-specific cancer. Table 5 examines
Figure 3. Estimated hazard ratios and 95% confidence intervals for
a 20% increase in percentage of energy from protein, from combined
analysis of data from the Women’s Health Initiative dietary modifica-
tion trial comparison group and observational study, without and with
biomarker calibration of consumption, 1993–2005. Unfilled square,
uncalibrated; filled circle, calibrated.
the effect of including body mass index in the log-hazard
ratio model on the calibrated energy hazard ratios shown in
Figure 1 and also shows the effect of including calibrated
energy on the hazard ratio for body mass index. Hazard
ratios for both energy and body mass index are not signif-
icant for most cancer sites and may be unstable in the pres-
ence of the other variable, and confidence intervals are wide.
DISCUSSION
This report has both methodological and substantive im-
plications. On the methodology side, it provides a first ap-
plication of the use of urinary recovery markers to correct
for systematic bias in dietary self-reported data, in an epi-
demiologic cohort setting. In analyses that control for stan-
dard confounding factors but not body mass index, FFQ
estimates of energy, protein, or percentage of energy from
protein were not significantly associated with total invasive
cancer incidence. In contrast, following biomarker calibra-
tion, the associations with total cancer incidence were
strong for energy (P < 0.0001), moderate for protein (P ¼
0.01), and inverse for percentage of energy from protein
(P ¼ 0.03), suggesting that macronutrients other than pro-
tein drive the positive energy association. Likewise, calibrated
energy consumption was found to be positively associated
with the risk of breast, colon, endometrial, and kidney cancer,
whereas uncalibrated energy was not.
These comparisons suggest that systematic bias in dietary
assessment could have a profound effect on nutritional ep-
idemiology findings. Total energy assessment is a recog-
nized weak aspect of FFQs. Uncalibrated FFQs are
generally believed to be more reliable for nutrient density
than for absolute consumption estimates. However, bio-
marker calibration also qualitatively affected the findings
for protein density in relation to total cancer (Figure 3).
Measurement error has typically been acknowledged in
epidemiology reporting through a simple deattenuation fac-
tor, as befits measurement equation 2 in the absence of
systematic bias (i.e., S2 ¼ S3 ¼ 0). Such deattenuation typ-
ically has little effect on significance levels. The presence of
systematic bias changes this feature, however, because re-
gression coefficients are corrected for distortions beyond
simple attenuation, possibly leading to substantially altered
P values.
To help interpret the calibrated energy variable defined
here, we note that calibrated energy can be viewed as esti-
mated actual short-term energy consumption, as determined
by FFQ energy, body mass index, age, and other factors. The
correlations of calibrated energy, in our combined cohorts,
with log-FFQ energy, body mass index, and age are, respec-
tively, 0.35, 0.81, and 0.44. The strong associations with
age and especially with body mass index imply that log-FFQ
energy does not adhere to a simple classical measurement
model. A linear regression of body mass index on log-
calibrated energy gives a projected body mass index increase
of 9.2 units corresponding to a 20% increase in calibrated
energy, suggesting, in conjunction with Table 5, that much of
the observed dependence of cancer incidence rates on total
energy can be explained by body mass associations with
Am J Epidemiol 2009;169:977–989
 Biomarker-calibrated Energy and Cancer 985

Table 5. Hazard Ratio Estimates for a 20% Increase in Calibrated Energy (kcal/day) Consumption and for a 10-Unit Increase in Body Mass
Index, in Analyses That Either Exclude (Unadjusted) or Include (Adjusted) the Other Variable, Using Data from the Women’s Health Initiative
Dietary Modification Trial Comparison Group and Observational Study, 1993–2005

Calibrated Energy Body Mass Index

Body Mass Index Unadjusted Body Mass Index Adjusted Energy Unadjusted Energy Adjusted
Cancer
Hazard 95% Confidence Hazard 95% Confidence Hazard 95% Confidence Hazard 95% Confidence
Ratio Intervala Ratio Intervala Ratio Interval Ratio Intervala

Total cancer 1.18 1.10, 1.27 0.90 0.76, 1.06 1.17 1.12, 1.23 1.27 1.11, 1.44
Breast 1.24 1.11, 1.38 1.11 0.81, 1.53 1.20 1.10, 1.30 1.10 0.86, 1.40
Colon 1.35 1.06, 1.71 0.70 0.41, 1.18 1.36 1.16, 1.61 1.81 1.19, 2.76
Rectum 1.23 0.79, 1.91 2.09 0.67, 6.50 1.15 0.81, 1.63 0.62 0.26, 1.52
Ovary 1.05 0.76, 1.45 1.12 0.61, 2.04 1.00 0.79, 1.28 0.95 0.58, 1.56
Endometrium 1.83 1.49, 2.25 1.40 0.83, 2.35 1.60 1.37, 1.87 1.26 0.84, 1.88
Bladder 1.18 0.83, 1.68 1.64 0.63, 4.25 1.08 0.77, 1.51 0.74 0.34, 1.60
Kidney 1.47 1.00, 2.16 1.27 0.52, 3.11 1.41 1.08, 1.83 1.14 0.59, 2.20
Pancreas 1.26 0.78, 2.03 0.88 0.41, 1.91 1.17 0.86, 1.59 1.37 0.76, 2.50
Lung 0.85 0.67, 1.08 0.58 0.37, 0.92 0.98 0.83, 1.16 1.44 0.99, 2.11
Lymphoma 0.75 0.56, 1.02 0.60 0.34, 1.04 0.87 0.70, 1.08 1.26 0.83, 1.90
Leukemia 1.41 0.93, 2.14 1.88 0.76, 4.60 1.22 0.90, 1.65 0.78 0.40, 1.51
a
The 95% confidence intervals for analyses that include calibrated energy are based on the log-estimated hazard ratio 61.96 3 the bootstrap
standard error.

these diseases. Table 5 likewise suggests that much of the
dependence of cancer incidence rates on body mass index
can be explained by energy consumption associations with
these diseases.
Our analyses yielded similar results when calibration
equations were applied in the DM cohort where they were
derived and when exported to the observational study. How-
ever, this extrapolation is under near-optimal conditions as
the 2 cohorts were drawn from essentially the same popu-
lations, with much commonality in eligibility and exclusion-
ary criteria. Comparison with calibration equations from
nutritional biomarker studies in other populations (25, 26)
could be informative.
As noted above, the Nutrient Biomarker Study was con-
ducted in 2004–2005, an average of about 6.5 years after the
1-year FFQ data collection for the DM trial comparison
group women and about 4 years on average after the 3-year
FFQ data collection for observational study women. Our
application assumes that the calibration equations devel-
oped from Nutrient Biomarker Study data apply to FFQs
at these earlier time points. Moreover, the biomarker data
provide consumption estimates over a rather short period of
time (e.g., 6 months between initial and repeat applications
in the 20% subsample). However, dietary patterns are ex-
pected to track over longer time periods for most women in
these cohorts.
On the substantive side, we observe strong positive asso-
ciations between calibrated energy consumption and the risk
of total and certain site-specific cancers. There are also sug-
gestions of a positive association between protein consump-
tion and leukemia and an inverse association between
percentage of energy from protein and bladder cancer
(Figures 2 and 3) that would be worth examining in other
settings. More comprehensive temporal data on the interplay
between a high-energy diet and body fat accumulation will
be needed to understand the mechanisms leading to elevated
cancer risk with a high level of energy consumption. How-
ever, regardless of whether body fat accumulation results
from a history of high-energy consumption, or whether
a high body mass leads to increased energy requirements,
or both, it is evident that a high body mass index is an
important aspect of total and site-specific cancer risk, and
efforts to prevent obesity deserve a continued high priority
in national cancer control efforts.
ACKNOWLEDGMENTS
Author affiliations: Division of Public Health Sciences, Fred
Hutchinson Cancer Research Center, Seattle, Washington
(Ross L. Prentice, Marian L. Neuhouser, Ruth E. Patterson,
Lesley F. Tinker); Biostatistics Research Branch, National
Institute of Allergy and Infectious Diseases, Bethesda,
Maryland (Pamela A. Shaw); Medical Research Council
Dunn Human Nutrition Unit, University of Cambridge,
Cambridge, United Kingdom (Sheila A. Bingham); Depart-
ment of Epidemiology, University of Washington, Seattle,
Washington (Shirley A. A. Beresford); Kaiser Permanente
Division of Research, Oakland, California (Bette Caan);
Stanford Prevention Research Center, Palo Alto, California
(Marcia L. Stefanick); University of Tennessee Health Sciences
Center, Memphis, Tennessee (Suzanne Satterfield); Depart-
ment of Nutritional Sciences, University of Arizona, Tucson,
Arizona (Cynthia A. Thomson); Department of Community
and Behavioral Health, University of Iowa, Iowa City, Iowa
(Linda Snetselaar); Medstar Research Institute, Washington,
District of Columbia (Asha Thomas); and Johns Hopkins

Am J Epidemiol 2009;169:977–989
986 Prentice et al.
University/Sinai Hospital, Baltimore, Maryland (Asha
Thomas).
This work was supported by the National Heart, Lung, and
Blood Institute, National Institutes of Health, US Department
of Health and Human Services (contracts N01WH22110,
24152, 32100-2, 32105-6, 32108-9, 32111-13, 32115, 32118-
19, 32122, 42107-26, 42129-32, and 44221). Clinical Trials
Registration: ClinicalTrials.gov identifier: NCT00000611. Dr.
Prentice’s work was partially supported by grant CA53996 from
the National Cancer Institute.
The authors thank the WHI investigators and staff for
their outstanding dedication and commitment. A full listing
of WHI investigators can be found at the following website:
http://www.whi.org.
A list of key WHI investigators involved in this research
follows. Program Office—National Heart, Lung, and Blood
Institute, Bethesda, Maryland: Elizabeth Nabel, Jacques
Rossouw, Shari Ludlam, Linda Pottern, Joan McGowan,
Leslie Ford, Nancy Geller. Clinical Coordinating Cen-
ters—Fred Hutchinson Cancer Research Center, Seattle,
Washington: Ross Prentice, Garnet Anderson, Andrea
LaCroix, Charles L. Kooperberg, Ruth E. Patterson, Anne
McTiernan; Wake Forest University School of Medicine,
Winston-Salem, North Carolina: Sally Shumaker; Medical
Research Labs, Highland Heights, Kentucky: Evan Stein;
University of California at San Francisco, San Francisco,
California: Steven Cummings. Clinical Centers—Albert
Einstein College of Medicine, Bronx, New York: Sylvia
Wassertheil-Smoller; Baylor College of Medicine, Houston,
Texas: Aleksandar Rajkovic; Brigham and Women’s Hospi-
tal, Harvard Medical School, Boston, Massachusetts: JoAnn
Manson; Brown University, Providence, Rhode Island:
Annlouise R. Assaf; Emory University, Atlanta, Georgia:
Lawrence Phillips; Fred Hutchinson Cancer Research Cen-
ter, Seattle, Washington: Shirley Beresford; George Wash-
ington University Medical Center, Washington, District of
Columbia: Judith Hsia; Los Angeles Biomedical Research In-
stitute at Harbor-UCLA Medical Center, Torrance, California:
Rowan Chlebowski; Kaiser Permanente Center for Health
Research, Portland, Oregon: Evelyn Whitlock; Kaiser Per-
manente Division of Research, Oakland, California: Bette
Caan; Medical College of Wisconsin, Milwaukee, Wisconsin:
Jane Morley Kotchen; MedStar Research Institute/Howard
University, Washington, District of Columbia: Barbara V.
Howard; Northwestern University, Chicago/Evanston,
Illinois: Linda Van Horn; Rush Medical Center, Chicago,
Illinois: Henry Black; Stanford Prevention Research Center,
Stanford, California: Marcia L. Stefanick; State University of
New York at Stony Brook, Stony Brook, New York: Dorothy
Lane; The Ohio State University, Columbus, Ohio: Rebecca
Jackson; University of Alabama at Birmingham, Birming-
ham, Alabama: Cora E. Lewis; University of Arizona,
Tucson/Phoenix, Arizona: Tamsen Bassford; University at
Buffalo, Buffalo, New York: Jean Wactawski-Wende; Univer-
sity of California at Davis, Sacramento, California: John
Robbins; University of California at Irvine, Irvine, California:
F. Allan Hubbell; University of California at Los Angeles,
Los Angeles, California: Lauren Nathan; University of
California at San Diego, LaJolla/Chula Vista, California:
Robert D. Langer; University of Cincinnati, Cincinnati,
Ohio: Margery Gass; University of Florida, Gainesville/
Jacksonville, Florida: Marian Limacher; University of
Hawaii, Honolulu, Hawaii: David Curb; University of Iowa,
Iowa City/Davenport, Iowa: Robert Wallace; University of
Massachusetts/Fallon Clinic, Worcester, Massachusetts:
Judith Ockene; University of Medicine and Dentistry of New
Jersey, Newark, New Jersey: Norman Lasser; University of
Miami, Miami, Florida: Mary Jo O’Sullivan; University of
Minnesota, Minneapolis, Minnesota: Karen Margolis; Univer-
sity of Nevada, Reno, Nevada: Robert Brunner; University of
North Carolina, Chapel Hill, North Carolina: Gerardo Heiss;
University of Pittsburgh, Pittsburgh, Pennsylvania: Lewis
Kuller; University of Tennessee, Memphis, Tennessee: Karen
C. Johnson; University of Texas Health Science Center, San
Antonio, Texas: Robert Brzyski; University of Wisconsin,
Madison, Wisconsin: Gloria E. Sarto; Wake Forest University
School of Medicine, Winston-Salem, North Carolina: Mara
Vitolins; and Wayne State University School of Medicine/
Hutzel Hospital, Detroit, Michigan: Susan Hendrix.
Decisions concerning study design, data collection and
analysis, interpretation of results, preparation of the manu-
script, or submission of the manuscript for publication resided
with committees comprising WHI investigators that included
National Heart, Lung, and Blood Institute representatives.
Conflict of interest: none declared.
REFERENCES
1. Armstrong B, Doll R. Environmental factors and cancer in-
cidence and mortality in different countries, with special ref-
erence to dietary practices. Int J Cancer. 1975;15(4):617–631.
2. Kritchevsky D, Klurfeld DM. Influence of caloric intake on
experimental carcinogenesis: a review. Adv Exp Med Biol.
1986;206:55–68.
3. Freedman LS, Clifford C, Messina M. Analysis of dietary fat,
calories, body weight, and the development of mammary tu-
mors in rats and mice: a review. Cancer Res. 1990;50(18):
5710–5719.
4. Willett WC, Sampson L, Stampfer MJ, et al. Reproducibility
and validity of a semiquantitative food frequency question-
naire. Am J Epidemiol. 1985;122(1):51–65.
5. Schoeller DA. Recent advances from application of doubly-
labeled water to measurement of human energy expenditure.
J Nutr. 1999;129(10):1765–1768.
6. Heitmann BL, Lissner L. Dietary underreporting by obese
individuals—is it specific or non-specific? BMJ. 1995;
311(7011):986–989.
7. World Cancer Research Fund/American Institute for Cancer
Research. Food, Nutrition and the Prevention of Cancer: A
Global Perspective. Washington, DC: American Institute for
Cancer Research; 1997.
8. World Cancer Research Fund/American Institute for Cancer
Research. Food, Nutrition, Physical Activity, and the Preven-
tion of Cancer: A Global Perspective. Washington, DC:
American Institute for Cancer Research; 2007.
9. Bingham SA. Urine nitrogen as a biomarker for the validation
of dietary protein intake. J Nutr. 2003;133(suppl 3):921S–
924S.
10. Prentice RL, Sugar E, Wang CY, et al. Research strategies and
the use of nutrient biomarkers in studies of diet and chronic
disease. Public Health Nutr. 2002;5(6A):977–984.
Am J Epidemiol 2009;169:977–989
 Biomarker-calibrated Energy and Cancer 987

11. Sugar EA, Wang CY, Prentice RL. Logistic regression with
exposure biomarkers and flexible measurement error. Bio-
metrics. 2007;63(1):143–151.
12. Neuhouser ML, Tinker L, Shaw PA, et al. Use of recovery
biomarkers to calibrate nutrient consumption self-reports in
the Women’s Health Initiative. Am J Epidemiol. 2008;
167(10):1247–1259.
13. Women’s Health Initiative Study Group. Design of the
Women’s Health Initiative clinical trial and observational
study. Control Clin Trials. 1998;19(1):61–109.
14. Women’s Health Initiative Study Group. Dietary adherence in
the Women’s Health Initiative dietary modification trial. J Am
Diet Assoc. 2004;104(4):654–658.
15. Prentice RL, Caan B, Chlebowski RT, et al. Low-fat dietary
pattern and risk of invasive breast cancer: the Women’s Health
Initiative randomized controlled dietary modification trial.
JAMA. 2006;295(6):629–642.
16. Beresford SA, Johnson KC, Ritenbaugh C, et al. Low-fat di-
etary pattern and risk of colorectal cancer: the Women’s
Health Initiative randomized controlled dietary modification
trial. JAMA. 2006;295(6):643–654.
17. Howard BV, Van Horn L, Hsia J, et al. Low-fat dietary pattern
and risk of cardiovascular disease: the Women’s Health Ini-
tiative randomized controlled dietary modification trial.
JAMA. 2006;295(6):655–666.
18. Prentice RL, Thomson CA, Caan B, et al. Low-fat dietary
pattern and cancer incidence in the Women’ s Health Initiative
dietary modification randomized controlled trial. J Natl Can-
cer Inst. 2007;99(20):1534–1543.
19. Hays J, Hunt JR, Hubbell FA, et al. The Women’s Health
Initiative recruitment methods and results. Ann Epidemiol.
2003;13(9 suppl):S18–S77.
20. Langer RD, White E, Lewis CE, et al. The Women’s Health
Initiative Observational Study: baseline characteristics of
participants and reliability of baseline measures. Ann Epide-
miol. 2003;13(9 suppl):S107–S121.
21. Kristal AK, Shattuck AL, Williams ME. Food frequency
questionnaires for diet intervention research. In: 17th National
Nutrient Databank Conference Proceedings. Baltimore, MD:
International Life Sciences Institute; 1992:110–125.
22. Schakel SF, Buzzard IM, Gebhardt SE. Procedures for esti-
mating nutrient values for food composition databases. J Food
Compost Anal. 1997;10(2):102–114.
23. Curb JD, McTiernan A, Heckbert SR, et al. Outcomes ascer-
tainment and adjudication methods in the Women’s Health
Initiative. Ann Epidemiol. 2003;13(9 suppl):S122–S128.
24. Cox DR. Regression analysis and life tables (with discussion).
J R Stat Soc (B). 1972;34:187–220.
25. Subar AF, Kipnis V, Troiano RP, et al. Using intake biomarkers
to evaluate the extent of dietary misreporting in a large sample
of adults: the OPEN study. Am J Epidemiol. 2003;158(1):1–13.
26. Kipnis V, Subar AF, Midthune D, et al. Structure of dietary
measurement error: results of the OPEN biomarker study. Am
J Epidemiol. 2003;158(1):14–21; discussion 22–26.

APPENDIX

Appendix Table 1. Baseline Factors Included in Cox Model Hazard Ratio Analyses to Control Confounding, in the Dietary Modification Trial
Comparison Group and Observational Study Components of the Women’s Health Initiative, 1993–2005a

Cancer Site
Total Lymphoma,
Cancer Colon, Bladder, Kidney, Leukemia
Breast Ovary Endometrium
Rectum Pancreas, Lung
Raceb (white/other, black, Hispanic) x x xc x xd
Education (high school or less, beyond x x
high school, college degree)
Exercise (METs/week) x x x
Smokingb (never, past, current) x x x x xe
b
Alcohol (never, past, <1/week, 1–7/week, x x x x
>7/week)
Breast cancer family history x x
(no, yes)
Gail 5-year risk (5-year absolute risk %) x
Unopposed estrogen use ever x x xc x x
(no, yes)
Estrogen plus progesterone use ever x x xc x x
(no, yes)
Colon cancer family history (no, yes) x
History of colorectal polyps (no, yes) xc
History of diabetes (no, yes) x
Hypertension (no, yes) x x xf

Abbreviation: MET, metabolic equivalent.

a
The same factors were used for the dietary modification comparison group and observational study cohorts.
b
For rare cancers: race: black/Hispanic (yes/no); smoking: ever (yes/no); alcohol: nondrinker (past/never), light drinker (<1 drink/week), moderate/heavy
(
1 drinks/week).
c
Colon cancer only.
d
Lung only.
e
Leukemia only.
f
Kidney only.

Am J Epidemiol 2009;169:977–989
988 Prentice et al.

Appendix Table 2. Hazard Ratios by Quartile of Biomarker-calibrated Energy Consumption

From the Analyses of Combined Data From the Women’s Health Initiative Dietary Modification
Trial Comparison Group and Observational Study, 19932005a

Energy (kcal/day)
Quartile 2 Quartile 3 Quartile 4
Cancer
Hazard 95% Confidence Hazard 95% Confidence Hazard 95% Confidence
Ratio Interval Ratio Interval Ratio Interval

Total cancer 1.07 0.97, 1.17 1.07 0.97, 1.19 1.18 1.07, 1.31
Breast 1.07 0.90, 1.28 1.17 0.98, 1.40 1.33 1.12, 1.58
Colon 1.27 0.88, 1.85 1.12 0.78, 1.60 1.51 1.03, 2.21
Rectum 1.82 0.81, 4.08 2.34 1.04, 5.26 1.51 0.64, 3.58
Ovary 1.23 0.78, 1.93 1.19 0.75, 1.89 0.91 0.58, 1.43
Endometrium 1.02 0.66, 1.57 1.26 0.81, 1.96 2.03 1.38, 3.00
Bladder 1.76 0.89, 3.46 2.14 1.02, 4.51 1.05 0.47, 2.39
Kidney 1.42 0.77, 2.62 1.31 0.71, 2.43 1.44 0.80, 2.61
Pancreas 0.94 0.49, 1.79 1.24 0.67, 2.32 1.33 0.68, 2.60
Lung 0.90 0.68, 1.19 0.75 0.53, 1.07 0.79 0.58, 1.08
Lymphoma 0.99 0.70, 1.42 0.81 0.54, 1.21 0.66 0.42, 1.03
Leukemia 1.46 0.71, 3.03 1.63 0.84, 3.18 1.46 0.69, 3.10
a
Estimated hazard ratios and 95% confidence intervals for the second, third, and fourth quar-
tiles relative to the first quartile of biomarker-calibrated energy consumption. Confidence intervals
for log-hazard ratios derive from the log-hazard ratio estimate 61.96 times the corresponding
bootstrapped standard deviation estimate.

Appendix Table 3. Hazard Ratios by Quartile of Biomarker-calibrated Protein Consumption

From the Analyses of Combined Data From the Women’s Health Initiative Dietary Modification
Trial Comparison Group and Observational Study, 19932005a

Protein (g/day)
Quartile 2 Quartile 3 Quartile 4
Cancer
Hazard 95% Confidence Hazard 95% Confidence Hazard 95% Confidence
Ratio Interval Ratio Interval Ratio Interval

Total cancer 1.07 0.97, 1.18 1.10 0.99, 1.22 1.09 0.98, 1.22
Breast 1.15 0.97, 1.36 1.07 0.90, 1.28 1.22 0.99, 1.49
Colon 0.97 0.70, 1.34 1.11 0.76, 1.61 0.96 0.64, 1.44
Rectum 1.22 0.56, 2.65 1.57 0.72, 3.41 1.08 0.48, 2.41
Ovary 0.68 0.42, 1.10 1.10 0.73, 1.66 0.85 0.55, 1.31
Endometrium 1.36 0.91, 2.04 1.59 1.08, 2.35 1.85 1.26, 2.70
Bladder 1.11 0.54, 2.28 1.25 0.64, 2.45 0.96 0.45, 2.05
Kidney 1.12 0.62, 2.03 0.98 0.53, 1.80 1.31 0.73, 2.36
Pancreas 1.41 0.82, 2.41 0.95 0.47, 1.92 1.19 0.60, 2.37
Lung 1.00 0.75, 1.34 1.05 0.76, 1.43 0.78 0.55, 1.10
Lymphoma 0.88 0.59, 1.30 0.96 0.63, 1.44 0.68 0.41, 1.11
Leukemia 1.38 0.64, 2.99 2.05 1.02, 4.09 1.77 0.82, 3.81
a
Estimated hazard ratios and 95% confidence intervals for the second, third, and fourth quar-
tiles relative to the first quartile of biomarker-calibrated protein consumption. Confidence intervals
for log-hazard ratios derive from the log-hazard ratio estimate 61.96 times the corresponding
bootstrapped standard deviation estimate.

Am J Epidemiol 2009;169:977–989
 Biomarker-calibrated Energy and Cancer 989

Appendix Table 4. Hazard Ratios by Quartile of Biomarker-calibrated Percentage of Energy

From Protein Consumption From the Analyses of Combined Data From the Women’s Health
Initiative Dietary Modification Trial Comparison Group and Observational Study, 19932005a

Percentage of Energy From Protein

Quartile 2 Quartile 3 Quartile 4
Cancer
Hazard 95% Confidence Hazard 95% Confidence Hazard 95% Confidence
Ratio Interval Ratio Interval Ratio Interval

Total cancer 0.94 0.85, 1.04 0.94 0.85, 1.04 0.92 0.82, 1.04
Breast 0.97 0.83, 1.13 0.92 0.78, 1.09 0.94 0.78, 1.12
Colon 0.83 0.58, 1.2 0.98 0.71, 1.35 1.06 0.74, 1.51
Rectum 0.85 0.43, 1.67 1.24 0.63, 2.44 1.01 0.52, 1.96
Ovary 1.16 0.76, 1.77 1.13 0.72, 1.8 1.08 0.69, 1.69
Endometrium 0.92 0.65, 1.29 0.99 0.71, 1.39 0.92 0.63, 1.35
Bladder 0.72 0.39, 1.33 0.84 0.44, 1.58 0.58 0.28, 1.22
Kidney 0.80 0.44, 1.48 1.10 0.63, 1.92 0.86 0.48, 1.53
Pancreas 0.89 0.54, 1.44 0.65 0.36, 1.17 0.92 0.56, 1.53
Lung 0.99 0.73, 1.33 0.90 0.66, 1.23 0.92 0.67, 1.26
Lymphoma 1.11 0.77, 1.59 0.89 0.61, 1.28 0.93 0.61, 1.40
Leukemia 1.29 0.74, 2.24 1.28 0.73, 2.25 1.39 0.76, 2.54
a
Estimated hazard ratios and 95% confidence intervals for the second, third, and fourth quar-
tiles relative to the first quartile of biomarker-calibrated percentage of energy from protein con-
sumption. Confidence intervals for log-hazard ratios derive from the log-hazard ratio estimate
61.96 times the corresponding bootstrapped standard deviation estimate.

Am J Epidemiol 2009;169:977–989
 American Journal of Epidemiology Vol. 169, No. 8
ª The Author 2009. Published by the Johns Hopkins Bloomberg School of Public Health. DOI: 10.1093/aje/kwn418
All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org. Advance Access publication February 18, 2009

Original Contribution

Sex-Modified Effect of Hepatitis B Virus Infection on Mortality From Primary Liver

Cancer

Na Wang, Yingjie Zheng, Xinsen Yu, Wenyao Lin, Yue Chen, and Qingwu Jiang

Initially submitted April 3, 2008; accepted for publication December 19, 2008.

Sex and hepatitis B virus (HBV) infection are both important risk factors for primary liver cancer. However, their
possible biologic interaction has not been well studied. The authors examined data from 89,789 subjects aged
25–69 years who participated in a 14-year cohort study (1992–2006) conducted in Haimen, China. An age-
stratified Cox proportional hazards model was used for multivariate analysis. The authors assessed the combined
effect of sex and HBV infection on liver cancer mortality by calculating 3 interaction measures: the relative risk due
to interaction, the attributable proportion of interaction, and the synergy index. There was a greater risk difference
between hepatitis B surface antigen carriers and noncarriers among men than among women. After adjustment for
potential confounders, the relative risk due to interaction, the attributable proportion of interaction, and the synergy
index were 33.27 (95% confidence interval (CI): 22.54, 43.99), 0.59 (95% CI: 0.55, 0.63), and 2.49 (95% CI: 2.13,
2.90), respectively, suggesting a significant synergistic effect of the interaction between sex and HBV infection on
liver cancer mortality. HBV infection had a larger impact on liver cancer mortality in men than in women, which may
explain at least part of the sex difference in liver cancer risk.

China; cohort studies; hepatitis B virus; liver neoplasms; mortality; risk factors; sex

Abbreviations: CI, confidence interval; HBsAg, hepatitis B surface antigen; HBV, hepatitis B virus; RERI, relative risk due to
interaction.

Hepatocellular carcinoma represents the majority of cases
of primary liver cancer (hereafter called liver cancer); it is
the fifth most common cancer and the third leading cause of
cancer death worldwide (1, 2). China is one of the countries
with the highest incidence of hepatocellular carcinoma in
the world—approximately 35 cases per 100,000 population
for males and 13 per 100,000 for females (2). Hepatocellular
carcinoma is a male-predominant cancer, with a male:female
ratio ranging from 2:1 to 4:1 (2). Hepatitis B virus (HBV)
infection, which is endemic in many Asian countries, in-
cluding China, is a leading risk factor for hepatocellular
carcinoma and accounts for approximately 53% of hepato-
cellular carcinoma cases worldwide (3). Other important
risk factors for hepatocellular carcinoma include hepatitis
C virus infection, alcoholism, aflatoxin B1 ingestion, and
family history of liver cancer (4). So far, it has been unclear
whether HBV infection can explain the sex difference in the
occurrence of liver cancer. In this study, we examined effect
modification by sex of the association between HBV infec-
tion and liver cancer mortality, based on data from a 14-year
longitudinal study conducted in Haimen, China.
MATERIALS AND METHODS
Study population
Study methods have been described in detail previously
(5). Briefly, the study was started in February 1992 in
Haimen, China, and enrollment continued until December
1993. A total of 90,236 residents (60,306 men and 29,930
women) voluntarily participated. The subjects were 22–80
years of age. All of the participants completed a questionnaire

Correspondence to Prof. Yue Chen, Department of Epidemiology and Community Medicine, Faculty of Medicine, University of Ottawa, 451
Smyth Road, Ottawa, Ontario K1H 8M5, Canada (e-mail: ychen@uottawa.ca) or Prof. Qingwu Jiang, Department of Epidemiology, School of
Public Health, Fudan University, 130 Dong’an Road, Shanghai 200032, People’s Republic of China (e-mail: jiangqu@fudan.edu.cn).

990 Am J Epidemiol 2009;169:990–995

 Sex-Modified Effect of HBV on Liver Cancer Mortality 991

and provided a blood specimen at entry. The questionnaire Table 1. Characteristics of Participants at Study Entry, Haimen
collected information about name, residence, date of birth, Cohort Study, China, 1992–2006
occupation (peasant, nonpeasant), regular cigarette smoking Men Women Total
(daily), regular alcohol drinking (
4 drinks per week), regu- Characteristic
No. % No. % No. %
lar tea drinking (
4 times per week), past pesticide exposure,
drinking water source (ditch, pond, shallow well, deep well) Age, years
in every decade from the 1960s to the 1990s, staple food type <40 23,081 38.4 14,179 47.7 37,260 41.5
(corn, rice, wheat) in every decade from the 1960s to the 40–54 27,511 45.8 12,362 41.6 39,873 44.4
1990s, history of acute hepatitis, history of jaundice, history
55 9,484 15.8 3,172 10.7 12,656 14.1
of cirrhosis, and family history of liver cancer. The subjects
Hepatitis B surface
were tested for hepatitis B surface antigen (HBsAg) by antigen status
radioimmunoassay.
Positive 9,171 15.3 5,015 16.9 14,186 15.8
The study protocol was approved by the Institutional
Review Board of Fox Chase Cancer Center (Philadelphia, Negative 50,905 84.7 24,698 83.1 75,603 84.2
Pennsylvania), the Medical Ethics Review Group of Haimen Occupation
City (Haimen, China), and the Ethics Review Committee of Peasant 43,307 72.1 24,792 83.4 68,099 75.8
Shanghai Medical University (Shanghai, China). Question- Nonpeasant 16,769 27.9 4,921 16.6 21,690 24.2
naire data were collected by trained interviewers employed Regular cigarette
at the Haimen City Anti-Epidemic Station. In this analysis, smoking (daily)
we included all 89,789 participants (60,076 men and 29,713 Yes 38,864 64.7 1,014 3.4 39,878 44.4
women) who were aged 25–69 years at entry.
No 21,212 35.3 28,699 96.6 49,911 55.6
All of the subjects were followed up every year for vital
status and liver cancer mortality until December 31, 2006. Regular alcohol
drinking (
4
Death certification information was reported to the death drinks/week)
register system at Haimen City Anti-Epidemic Station Yes 35,985 59.9 3,812 12.8 39,797 44.3
monthly by village or township doctors. Information on
No 24,091 40.1 25,901 87.2 49,992 55.7
cause of death for persons who had temporarily moved out
of Haimen was also collected. Persons who had a permanent History of acute
hepatitis
change of residency were identified and no longer followed.
Yes 12,042 20.0 4,499 15.1 16,541 18.4
No 48,034 80.0 25,214 84.9 73,248 81.6
Statistical analysis
Family history of
Person-years of follow-up were calculated for each par- liver cancer
ticipant from the date of the baseline survey to the date of Yes 1,992 3.3 1,394 4.7 3,386 3.8
death, the date of loss to follow-up, or, if neither of those No 58,084 96.7 28,319 95.3 86,403 96.2
events occurred, the end of the study period (December 31,
2006). Subjects who were lost to follow-up were censored
on the last confirmed date of their presence in Haimen. lated as RRAþBþ  RRAþB  RRABþ þ 1), the attributable
A Cox proportional hazards model was used to estimate proportion of interaction (calculated as RERI/RRAþBþ),
the sex-specific hazard ratio and 95% confidence interval and the synergy index (calculated as (RRAþBþ 1)/
for liver cancer mortality associated with HBsAg status. [(RRAþB  1) þ (RRABþ  1)])—and their 95% confidence
Time-dependent covariates were used to assess proportion- intervals to assess the combined effect of male sex and HBV
ality. The baseline hazard in the proportional hazards re- infection as measured by HBsAg status. Subjects were grouped
gression models was stratified by 1-year age category, and into 4 categories: females who were HBsAg-negative (AB),
then the proportional hazards assumption for the other var- females who were HBsAg-positive (AþB), males who were
iables was satisfied. Estimated hazard ratios were adjusted HBsAg-negative (ABþ), and males who were HBsAg-
for other potential confounders, including occupation, cig- positive (AþBþ). If there is no interaction, RERI and the at-
arette smoking, alcohol consumption, history of hepatitis, tributable proportion of interaction are both 0 and the synergy
and family history of liver cancer. index is 1 (9).
The modifying effect of sex on the relation between Statistical analysis was performed using the Statistical
HBsAg and liver cancer mortality was examined on a mul- Analysis System, version 9.1 (SAS Institute, Inc., Cary,
tiplicative scale; it was not statistically significant in a pre- North Carolina). The SAS program provided by Li and
vious analysis (5) and also was not significant in this study. Chambless (11) was used to calculate the 3 measures of
It has been argued that interaction on an additive scale is additive interaction and to test for significance.
more meaningful for assessing public health and biologic
significance (6, 7) and also more indicative of the underly-
ing causal mechanism (8). In this analysis, we assessed the RESULTS
effect of the additive interaction of HBsAg status and sex on
risk of liver cancer mortality. We calculated 3 measures The 89,789 subjects had a total of 1,147,713 person-years
(9, 10)—the relative excess risk of interaction (RERI) (calcu- of follow-up. Table 1 shows the demographic characteristics

Am J Epidemiol 2009;169:990–995
992 Wang et al.

Table 2. Primary Liver Cancer Mortality and Risk Difference According to Hepatitis B Surface
Antigen Status, by Sex and Age Group, Haimen Cohort Study, China, 1992–2006

Mortality per
Sex or Age Group (years) % HBsAg- No. of Person-Years Risk
100,000
and HBsAg Status Positive Deaths of Follow-Up Difference
Person-Years

Total
Sex
Male
Negative 423 660,324.1 64.1 971.1
Positive 15.3 1,113 107,510.8 1,035.2
Female
Negative 51 317,463.7 16.1 330.0
Positive 16.9 216 62,414.8 346.1
Age group, years
<40
Men
Negative 74 246,731.0 30.0 729.5
Positive 18.5 397 52,271.1 759.5
Women
Negative 7 147,818.2 4.7 223.7
Positive 17.8 72 31,518.6 228.4
40–54
Men
Negative 210 309,071.4 67.9 1,181.2
Positive 14.3 566 45,311.6 1,249.1
Women
Negative 27 134,954.8 20.0 394.6
Positive 16.4 105 25,325.1 414.6

55
Men
Negative 139 104,521.7 133.0 1,377.9
Positive 10.0 150 9,928.1 1,510.9
Women
Negative 17 34,690.7 49.0 651.0
Positive 14.8 39 5,571.2 700.0

Abbreviation: HBsAg, hepatitis B surface antigen.

of the participants at baseline. Among the subjects, 15.8%
were HBsAg-positive (men: 15.3%; women: 16.9%).
A total of 13,529 deaths or losses to follow-up were iden-
tified, among which 1,803 participants died from liver cancer,
with 73.7% of them being HBsAg-positive. The median age
at death for liver cancer patients was 51.7 years for men and
52.4 years for women. Mortality from liver cancer varied in
the 4 groups defined by sex and HBsAg status (Table 2). Male
HBsAg carriers had the highest liver cancer mortality
(1,035.2 deaths per 100,000 person-years), while female
non-HBsAg carriers had the lowest (16.1 deaths per
100,000 person-years). The mortality difference for liver can-
cer between HBsAg carriers and non-HBsAg carriers was
much greater in men than in women among all age groups.
HBsAg carriers had an increased risk of mortality from
liver cancer, with a hazard ratio of 16 as compared with non-
HBsAg carriers. Male sex and older age were also important
predictors of liver cancer mortality. Compared with women
who were HBsAg-negative, mortality was significantly in-
creased in males and persons who were HBsAg-positive,
and the combined effect of male sex and HBsAg positivity
was synergistic (Figure 1). After adjustment for potential
confounders, the 3 interaction measures on the additive
scale all showed significantly higher values than the null
(RERI: 33.27 (95% confidence interval (CI): 22.54,
43.99); attributable proportion of interaction: 0.59 (95%
CI: 0.55, 0.63); synergy index: 2.49 (95% CI: 2.13, 2.90)),
suggesting a synergistic effect of male sex and HBsAg pos-
itivity on liver cancer mortality. We further examined this
joint effect according to age (Table 3). The RERI decreased
with increasing age, suggesting that the sex-HBsAg inter-
action was stronger in younger people.

Am J Epidemiol 2009;169:990–995
 Sex-Modified Effect of HBV on Liver Cancer Mortality 993

measure of additivity in the hazards model (11), indicated

Figure 1. Individual and joint effects of sex and hepatitis B surface
antigen (HBsAg) status on mortality from primary liver cancer, Haimen
Cohort Study, China, 1992–2006. HBV, hepatitis B virus.
DISCUSSION
It has been well documented that men are more likely
than women to develop liver cancer (2), and HBV infection
is a major risk factor (3). In this analysis, we focused on the
joint effect of male sex and HBV infection on liver cancer
mortality. The proportion of HBsAg-positive men was com-
parable to that of women in our study. It has been suggested
that the higher rate of liver cancer in males than in females is
due to higher levels of exposure to risk factors (2). However,
after adjustment for covariates, including HBsAg status,
age, history of hepatitis, occupation, regular alcohol drink-
ing, regular cigarette smoking, and family history of liver
cancer, men still had higher mortality from liver cancer than
women did. This may suggest that some susceptibility other
than exposure level is more important.
We assessed additive interaction by calculating 3 mea-
sures, including RERI, the attributable proportion of inter-
action, and the synergy index, all of which indicated that
there was a significant synergistic effect of HBsAg positivity
and male sex on risk of liver cancer mortality. HBsAg pos-
itivity had an increased effect on liver cancer mortality in
men compared with women. RERI, considered the best
that the excess risk for male HBsAg carriers due to interac-
tion was 33 times higher than the risk for female non-
HBsAg carriers. A high proportion (59%) of the mortality
among male HBsAg carriers was attributable to the interac-
tion. Because of the additive interaction, the excess risk of
liver cancer mortality resulting from combined exposure to
male sex and HBsAg positivity was 2.49 times higher than
the sum of their independent effects. These results suggest
that men are more sensitive to the effect of HBV infection on
liver cancer mortality than women. An increased impact of
HBV infection may explain, at least in part, the elevated risk
of liver cancer in men.
Many studies have explored potential mechanisms of
liver cancer development, as well as the sex difference in
incidence. Sex hormones may play an important role in sex
disparity in the process of oncogenesis. In a multicenter
case-control study, Yu et al. (12) observed an inverse rela-
tion between exposure to estrogen and liver cancer, suggest-
ing that estrogen may provide a protective effect against
liver cancer. Naugler et al. (13) found that estrogen-
mediated inhibition of interleukin-6 production by Kupffer
cells reduced the risk of liver cancer in females. An intra-
cellular signaling protein called MyD88 and the transcrip-
tion factor nuclear factor jB were also found to be involved
in this signaling pathway (13, 14). Wang et al. (15) inte-
grated the HBsAg gene into the mouse p21neomycin locus and
found that estrogen receptor b was extremely up-regulated,
which indicates that estrogen receptor b may play an im-
portant role in the development of liver cancer related to
HBsAg. In a case-control study, Yu et al. (16) and Sung et al.
(17) found that there was a positive association between
testosterone levels and liver cancer risk in HBV carriers
and that genes involved in the regulation of testosterone,
such as SRD5A2 and V89L, may also play a role in the
etiology of liver cancer.
Liver cancer arises most frequently in the setting of
chronic liver inflammation (18). After infection by the hep-
atitis virus, the host’s inflammatory immune response to the
viral antigens induces hepatocyte damage and is followed
by the pathogenesis of liver cancer (19). Sex hormones may
interact with HBV infection in the process and lead to a dom-
inant sex disparity in liver cancer risk.

Table 3. Effects of 3 Measures of Additive Interaction Between Sex and Hepatitis B Virus Infection on Mortality From Primary Liver Cancer, by
Age Group, Haimen Cohort Study, China, 1992–2006

Age Group, Unadjusted Adjusteda

years SI 95% CI RERI 95% CI AP 95% CI SI 95% CI RERI 95% CI AP 95% CI

Total 2.69 2.34, 3.10 41.69 29.34, 54.04 0.62 0.58, 0.66 2.49 2.13, 2.90 33.27 22.54, 43.99 0.59 0.55, 0.63
<40 3.06 2.41, 3.88 107.72 26.29, 189.16 0.67 0.58, 0.76 2.85 2.19, 3.72 90.03 19.38, 160.67 0.64 0.55, 0.74
40–54 2.75 2.25, 3.36 39.35 23.44, 55.26 0.63 0.58, 0.68 2.53 2.02, 3.16 30.24 16.83, 43.65 0.59 0.54, 0.65

55 1.96 1.39, 2.76 15.36 5.37, 25.35 0.47 0.39, 0.56 1.86 1.28, 2.72 13.25 3.33, 23.17 0.45 0.35, 0.54

Abbreviations: AP, attributable proportion of interaction; CI, confidence interval; RERI, relative excess risk of interaction; SI, synergy index.
a
Adjusted for age, occupation (peasant vs. nonpeasant), family history of liver cancer (yes vs. no), regular cigarette smoking (daily; yes vs. no),
regular alcohol drinking (
4 drinks per week; yes vs. no), and history of hepatitis (yes vs. no).

Am J Epidemiol 2009;169:990–995
994 Wang et al.
In our analysis, the sex-HBsAg interaction tended to de-
crease with age. Previous study has demonstrated that liver
cancer tends to occur in men and postmenopausal women,
which may result from lower production of estradiol and
a reduced response to the action of estradiol (20). That
may contribute to a lower sex disparity in the HBV-related
risk of liver cancer in older people. Another possibility is
potentially competing causes of death, which may be more
prominent in older subjects than in younger ones and is also
sex-related.
One limitation of our study is that HBsAg status was used
as the sole indicator of HBV infection. Occult HBV infec-
tion, a type of chronic HBV infection that is characterized
by the absence of detectable HBsAg in the blood and very
low levels of HBV DNA in the blood and liver (21, 22), can
also lead to liver cancer, with a mechanism similar to that of
overt cases (23). Therefore, part of the mortality among
HBsAg noncarriers might be attributable to the effect of
occult hepatitis B infection, resulting in underestimation
of the effect of HBV infection as well as its interaction with
sex. Additional data on HBV infection, such as information
on hepatitis B e antigen, HBV virus load, and HBV geno-
type, would consolidate the finding (3, 23).
In summary, we found a synergistic effect of male sex and
HBV infection on risk of liver cancer mortality, suggesting
that HBV infection may have a greater impact on liver can-
cer risk in men than in women, at least for persons living in
high-risk areas. Possible mechanisms for the interaction be-
tween sex and HBV infection in liver cancer mortality need
to be further explored. If the finding is confirmed by other
studies, a specific HBV vaccination program may be devel-
oped to reduce liver cancer risk. It has been well demon-
strated that HBV infection frequently occurs as a result of
vertical (mother-to-child) or horizontal (child-to-child)
transmission in endemic areas (24). It has also been reported
that HBV infection can be acquired in adulthood through
invasive medical procedures or household contact with
HBV carriers (25). In Haimen, where the current study
was conducted, an HBV vaccination program for all new-
borns was implemented in 1992. Adults, however, are not
included in the program. Because of the increased risk of
liver cancer in HBsAg-positive males, immunization strat-
egies might be expanded to also include male adults not yet
infected with HBV, to reduce the risk of developing liver
cancer.
ACKNOWLEDGMENTS
Author affiliations: Department of Epidemiology, School
of Public Health, Fudan University, Shanghai, People’s Re-
public of China (Na Wang, Yingjie Zheng, Qingwu Jiang);
Haimen City Center for Disease Control and Prevention,
Haimen City, People’s Republic of China (Xinsen Yu,
Wenyao Lin); and Department of Epidemiology and Com-
munity Medicine, Faculty of Medicine, University of Ot-
tawa, Ottawa, Ontario, Canada (Yue Chen).
This work was supported by a grant from the Ministry of
Science and Technology of the People’s Republic of China
(2006BAI02A03) and the Shanghai Leading Academic Dis-
cipline Project (B118). Na Wang received a scholarship
from the China Scholarship Council for her training at the
University of Ottawa.
Conflict of interest: none declared.
REFERENCES
1. Parkin DM. Global cancer statistics in the year 2000. Lancet
Oncol. 2001;2(9):533–543.
2. El-Serag HB, Rudolph KL. Hepatocellular carcinoma: epide-
miology and molecular carcinogenesis. Gastroenterology.
2007;132(7):2557–2576.
3. Lupberger J, Hildt E. Hepatitis B virus-induced oncogenesis.
World J Gastroenterol. 2007;13(1):74–81.
4. Bosch FX, Ribes J, Dı́az M, et al. Primary liver cancer:
worldwide incidence and trends. Gastroenterology. 2004;
127(5 suppl 1):S5–S16.
5. Evans AA, Chen G, Ross EA, et al. Eight-year follow-up of the
90,000-person Haimen City cohort: I. Hepatocellular carci-
noma mortality, risk factors, and gender differences. Cancer
Epidemiol Biomarkers Prev. 2002;11(4):369–376.
6. Rothman KJ. Epidemiology: An Introduction. Oxford, NY:
Oxford University Press; 2002.
7. Rothman KJ, Greenland S. Modern Epidemiology. Philadel-
phia, PA: Lippincott-Raven; 1998.
8. Kalilani L, Atashili J. Measuring additive interaction using
odds ratios [electronic article]. Epidemiol Perspect Innov.
2006;3:5.
9. Andersson T, Alfredsson L, Källberg H, et al. Calculating
measures of biological interaction. Eur J Epidemiol. 2005;
20(7):575–579.
10. Assmann SF, Hosmer DW, Lemeshow S, et al. Confidence
intervals for measures of interaction. Epidemiology. 1996;7(3):
286–290.
11. Li R, Chambless L. Test for additive interaction in propor-
tional hazards models. Ann Epidemiol. 2007;17(3):227–236.
12. Yu MW, Chang HC, Chang SC, et al. Role of reproductive
factors in hepatocellular carcinoma: impact on hepatitis B- and
C-related risk. Hepatology. 2003;38(6):1393–1400.
13. Naugler WE, Sakurai T, Kim S, et al. Gender disparity in liver
cancer due to sex differences in MyD88-dependent IL-6
production. Science. 2007;317(5834):121–124.
14. Lawrence T, Hageman T, Balkwill F. Cancer: sex, cytokines,
and cancer. Science. 2007;317(5834):51–52.
15. Wang Y, Cui F, Lv Y, et al. HBsAg and HBx knocked into
the p21 locus causes hepatocellular carcinoma in mice.
Hepatology. 2004;39(2):318–324.
16. Yu MW, Yang YC, Yang SY, et al. Hormonal markers and
hepatitis B virus-related hepatocellular carcinoma risk:
a nested case-control study among men. J Natl Cancer Inst.
2001;93(21):1644–1651.
17. Sung YM, Tang NL, Lai PB, et al. Re: Hormonal markers and
hepatitis B virus-related hepatocellular carcinoma risk:
a nested case-control study among men [letter]. J Natl Cancer
Inst. 2003;95(7):559–560.
18. Prieto J. Inflammation, hepatocellular carcinoma and sex:
IL-6 in the centre of the triangle. J Hepatol. 2008;48(2):
380–381.
19. Wands J. Hepatocellular carcinoma and sex. N Engl J Med.
2007;357(19):1974–1976.
20. Shimizu I, Kohno N, Tamaki K, et al. Female hepatology:
favorable role of estrogen in chronic liver disease with
Am J Epidemiol 2009;169:990–995

hepatitis B virus infection. World J Gastroenterol. 2007;13(32):
4295–4305.
21. Vivekanandan P, Kannangai R, Ray SC, et al. Comprehensive
genetic and epigenetic analysis of occult hepatitis B from liver
tissue samples. Clin Infect Dis. 2008;46(8):1227–1236.
22. Torbenson M, Thomas DL. Occult hepatitis B. Lancet Infect
Dis. 2002;2(8):479–486.
Am J Epidemiol 2009;169:990–995
Sex-Modified Effect of HBV on Liver Cancer Mortality 995
23. Chemin I, Trépo C. Clinical impact of occult HBV infections.
J Clin Virol. 2005;34(suppl 1):S15–S21.
24. Marrero CR, Marrero JA. Viral hepatitis and hepatocellular
carcinoma. Arch Med Res. 2007;38(6):612–620.
25. Zhang HW, Yin JH, Li YT, et al. Risk factors for acute hep-
atitis B and its progression to chronic hepatitis in Shanghai,
China. Gut. 2008;57(12):1713–20.
 American Journal of Epidemiology Vol. 169, No. 8
ª The Author 2009. Published by the Johns Hopkins Bloomberg School of Public Health. DOI: 10.1093/aje/kwn414
All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org. Advance Access publication February 16, 2009

Original Contribution

Serum Selenium and Peripheral Arterial Disease: Results From the National
Health and Nutrition Examination Survey, 2003–2004

Joachim Bleys, Ana Navas-Acien, Martin Laclaustra, Roberto Pastor-Barriuso, Andy Menke,
Jose Ordovas, Saverio Stranges, and Eliseo Guallar

Initially submitted October 12, 2008; accepted for publication December 15, 2008.

The authors conducted a cross-sectional study of the association of serum selenium with the prevalence of
peripheral arterial disease among 2,062 US men and women 40 years of age or older participating in the National
Health and Nutrition Examination Survey, 2003–2004. Serum selenium was measured by using inductively cou-
pled plasma-dynamic reaction cell-mass spectrometry. Peripheral arterial disease was defined as an ankle-
brachial blood pressure index <0.90. The age-, sex-, and race-adjusted prevalence of peripheral arterial disease
decreased with increasing serum selenium (P for linear trend ¼ 0.02), but there was an indication of an upturn in
risk in the highest quartile of serum selenium. The fully adjusted odds ratios for peripheral arterial disease com-
paring selenium quartiles 2, 3, and 4 with the lowest quartile were 0.75 (95% confidence interval: 0.37, 1.52), 0.58
(95% confidence interval: 0.28, 1.19), and 0.67 (95% confidence interval: 0.34, 1.31), respectively. In spline
regression models, peripheral arterial disease prevalence decreased with increasing serum selenium levels up
to 150–160 ng/mL, followed by a gradual increase at higher selenium levels. The association between serum
selenium levels and the prevalence of peripheral arterial disease was not statistically significant, although
a U-shaped relation was suggested.

antioxidants; cardiovascular diseases; cross-sectional studies; nutrition surveys; peripheral vascular diseases;
selenium

Abbreviations: ABI, ankle-brachial blood pressure index; NHANES, National Health and Nutrition Examination Survey.

Selenium, an essential micronutrient involved in anti-
oxidant selenoenzymes such as glutathione peroxidases,
has been hypothesized to prevent atherosclerotic disease
(1–3). Glutathione peroxidase synthesis and activity,
however, plateau at selenium levels above 70–90 ng/mL
(4). In the United States compared with other countries,
selenium intake is substantially higher (5), and most
adults have serum selenium levels above 95 ng/mL (4).
At these high concentrations, selenium is incorporated
nonspecifically as selenomethionine in the synthesis of
other plasma proteins, with unknown health effects (4).
It is thus unclear whether increased selenium levels con-
fer additional benefit for atherosclerosis prevention in the
United States.
A recent meta-analysis of 14 prospective cohort studies
found a modest, but statistically significant inverse associa-
tion between selenium levels and coronary heart disease (3),
but the 2 US studies in this meta-analysis found no associ-
ation (3, 6, 7). Moreover, serum selenium levels were not
associated with cardiovascular disease mortality in a pro-
spective study in a representative US sample, although
a U-shaped relation was suggested (8). Given the current
interest in selenium supplements for chemoprevention of
cancer (5), it is important to understand the overall impact
of increased selenium intake on other health endpoints, in-
cluding vascular disease.
Peripheral arterial disease, which affects about 8 million
Americans, is characterized by flow-limiting atherosclerosis

Correspondence to Dr. Eliseo Guallar, Departments of Epidemiology and Medicine and the Welch Center for Prevention, Epidemiology and
Clinical Research, Johns Hopkins Bloomberg School of Public Health, 2024 East Monument Street, Room 2-639, Baltimore, MD 21205-2223
(e-mail: eguallar@jhsph.edu).

996 Am J Epidemiol 2009;169:996–1003


in the muscular arteries of the lower extremities and is an
important marker of generalized atherosclerosis (9–11).
Data on the association of selenium levels with peripheral
arterial disease are very limited (12, 13). The objective of
the present study was to assess the association between
serum levels of selenium and reduced ankle-brachial
blood pressure index (ABI), a specific subclinical marker
for peripheral arterial disease, in the National Health and
Nutrition Examination Survey (NHANES), 2003–2004. ABI
values below 0.90 are considered diagnostic of peripheral
arterial disease. Furthermore, a low ABI is an independent
predictor of cardiovascular risk after adjusting for traditional
cardiovascular risk factors (14, 15).
MATERIALS AND METHODS
NHANES is conducted by the National Center for
Health Statistics (Hyattsville, Maryland) by using a com-
plex multistage sampling design to obtain a representative
sample of the civilian, noninstitutionalized US popula-
tion. We used data from NHANES 2003–2004 (16) be-
cause it was the first NHANES survey to measure
selenium and ABI levels simultaneously. In NHANES
2003–2004 interviews and physical examinations, the
overall response rate was 76%. Serum selenium and
ABI measurements were restricted to participants aged
40 years or older (N ¼ 3,086). We excluded 2 pregnant
women, 183 participants without selenium measure-
ments, 514 participants without ABI measurements in
both legs, and 314 participants with missing information
on any adjustment covariate. We finally excluded 11 par-
ticipants with left or right ABI measurements of more
than 1.5, usually due to vessel stiffness. The final sample
size was 2,062. The 2003–2004 NHANES study protocols
were approved by the National Center for Health Statis-
tics institutional review board. Oral and written informed
consent was obtained from all participants.
Serum selenium
Collection materials were screened for potential selenium
contamination. After blood collection, serum aliquots were
obtained, frozen at
20
C, and shipped to the Trace Ele-
ments Laboratory at the Wadsworth Center of the New York
State Department of Health for analysis. Serum selenium
levels were measured by using inductively coupled
plasma-dynamic reaction cell-mass spectrometry. The lab-
oratory procedures and quality control methods for serum
selenium measurement have been described in detail else-
where (17). The between-assay coefficients of variation for
quality-control pooled samples analyzed throughout the du-
ration of the survey ranged from 2.5% to 2.9%.
Peripheral arterial disease
A specific protocol was used to measure ABI in
NHANES 2003–2004 (18). The measurements of blood
pressure used for ABI were additional to and different from
other measurements of blood pressure used to evaluate hy-
pertension. Systolic blood pressure was measured on the
Am J Epidemiol 2009;169:996–1003
Selenium and Peripheral Arterial Disease 997
right arm (brachial artery) and both ankles (posterior tibial
arteries) with a Doppler device, the Parks Mini-Lab IV, model
3100 (Parks Medical Electronics, Inc., Aloha, Oregon). If the
participant had a condition that would interfere with blood
pressure reading in the right arm, the left arm was used.
Systolic blood pressure was measured twice at each site for
participants aged 40–59 years and once at each site for
participants aged 60 years or older. The left and right
ABI measurements were obtained by dividing the mean
systolic blood pressure in each ankle by the mean systolic
blood pressure in the arm. Peripheral arterial disease was
defined as an ABI value of less than 0.90 in at least one leg
(14, 15).
Other variables
Information about age, sex, race-ethnicity, education,
family income, menopausal status for women, cigarette
smoking, alcohol consumption, use of dietary supplements,
and use of cholesterol- and blood-pressure-lowering medi-
cations was based on self-report. Body mass index was cal-
culated by dividing measured weight in kilograms by
measured height in meters squared. Three to 4 systolic blood
pressure measurements were taken and were averaged by
using standardized protocols. Diabetes was defined as a fast-
ing serum glucose concentration of 126 mg/dL or higher,
a nonfasting serum glucose concentration of 200 mg/dL or
higher, a self-reported physician diagnosis, or current med-
ication use. Glomerular filtration rate was estimated by us-
ing the Modification of Diet in Renal Disease Study
equation with serum creatinine values (19).
Statistical methods
Participants were grouped in quartiles of serum sele-
nium levels based on the weighted population distribu-
tion. Odds ratios and 95% confidence intervals for
peripheral arterial disease prevalence comparing the 3
highest quartiles of serum selenium with the lowest quar-
tile were estimated by using logistic regression. Tests for
linear risk trend across serum selenium quartiles were
performed by including an ordinal variable with the
median selenium level of each quartile in the logistic re-
gression models. To further explore the shape of the dose-
response relation between serum selenium levels and
peripheral arterial disease prevalence, we used restricted
quadratic splines with knots at the 5th, 50th, and 95th
percentiles of serum selenium distribution. These spline
models require the same number of parameters as the
quartile analysis, but they can accommodate a wide vari-
ety of smooth risk trends (20). Sensitivity analyses using
different numbers and locations of the knots, with cubic
instead of quadratic splines, and log-transforming serum
selenium levels gave similar results (not shown). Statisti-
cal analyses were performed with the survey package in R
software (to account for the complex sampling design in
NHANES 2003–2004) (21, 22). Strata, primary sampling
units, and examination sample weights were used to ob-
tain unbiased point estimates and robust linearized stan-
dard errors.
998 Bleys et al.

Table 1. Characteristics of the Study Population by the Presence or Absence of Peripheral

Arterial Disease, National Health and Nutrition Examination Survey, 2003–2004a

Peripheral No Peripheral
Arterial Disease Arterial Disease P value
(n 5 169) (n 5 1,893)

Age, years 67.2 (1.2) 55.6 (0.4) <0.001

Sex: men 48.3 (3.6) 49.6 (1.4) 0.75
Race-ethnicity 0.001
White 76.2 (4.8) 80.4 (3.3)
Black 17.0 (3.8) 8.5 (1.3)
Mexican American 4.4 (2.9) 4.7 (1.7)
Education <high school 24.9 (4.5) 15.8 (1.6) 0.01
Family income <$20,000 33.3 (4.6) 19.6 (2.0) 0.006
Postmenopausal women 44.1 (4.6) 32.7 (1.4) 0.02
Cigarette smoking 0.02
Former 48.0 (6.6) 33.8 (1.2)
Current 23.6 (3.7) 20.7 (1.1)
Serum cotinine, ng/mL 0.45 (1.5) 0.40 (1.2) 0.74
Current alcohol drinking 21.1 (3.6) 32.8 (2.7) 0.03
Body mass index, kg/m2 29.9 (0.5) 28.3 (0.2) 0.01
Dietary supplement use 61.9 (3.7) 63.1 (1.6) 0.79
C-reactive protein, mg/L 3.4 (1.1) 2.0 (1.0) 0.001
Total cholesterol, mg/dL 201.9 (4.5) 209.3 (1.4) 0.13
High density lipoprotein 51.8 (1.1) 54.8 (0.5) 0.03
cholesterol, mg/dL
Cholesterol-lowering-medication use 39.3 (4.7) 19.2 (1.3) <0.001
Systolic blood pressure, mm Hg 134.5 (1.4) 127.1 (0.8) <0.001
Blood-pressure-lowering-medication use 55.5 (5.2) 29.2 (1.8) <0.001
Diabetes 31.5 (6.8) 11.0 (1.1) 0.003
Glomerular filtration rate 26.1 (4.5) 10.3 (0.8) <0.001
<60 mL/minute per 1.73 m2
Serum selenium, ng/mL 134.5 (2.0) 137.7 (1.2) 0.11
a
Values are expressed as percentages (standard errors) for categorical variables or means
(standard errors) for continuous variables; for serum cotinine and C-reactive protein, geometric
means (geometric standard errors) are reported.

RESULTS cholesterol, and systolic blood pressure and were inversely

The weighted prevalence of peripheral arterial disease in
the study population was 4.9%. Compared with participants
without peripheral arterial disease, those with disease were
more likely to be older, black, ever smokers, nondrinkers, and
diabetic; to have a lower educational level and family in-
come; and to use cholesterol- and blood-pressure-lowering
medications (Table 1). Participants with peripheral arterial
disease, compared with those without disease, also had higher
average levels of body mass index, systolic blood pressure,
and C-reactive protein and lower high density lipoprotein
cholesterol levels and glomerular filtration rate.
Participants in the highest quartile of serum selenium
levels, compared with those in the lowest quartile, were
more likely to be older, men, white, and nonsmokers and
to use dietary supplements and cholesterol-lowering med-
ications (Table 2). Serum selenium levels were also positive-
ly associated with total cholesterol, high density lipoprotein
associated with body mass index and serum cotinine levels.
The age-, sex-, and race-adjusted prevalence of peripheral
arterial disease decreased with increasing serum selenium
levels (P for linear trend ¼ 0.02) (Table 3). However, there
was an indication of an upturn in risk trend in the highest
quartile of serum selenium, particularly after adjusting for
cardiovascular risk factors. The fully adjusted odds ratios
for peripheral arterial disease comparing selenium quartiles
2, 3, and 4 with the lowest quartile were 0.75 (95% confi-
dence interval: 0.37, 1.52), 0.58 (95% confidence interval:
0.28, 1.19), and 0.67 (95% confidence interval: 0.34, 1.31),
respectively. In spline regression models, peripheral arterial
disease prevalence decreased with increasing serum sele-
nium levels up to 150–160 ng/mL (80th–91st percentiles
of the serum selenium distribution in the study population),
followed by a gradual increase at higher selenium levels
(Figure 1). Consistently, there was a marginally significant
U-shaped dose-response relation between serum selenium

Am J Epidemiol 2009;169:996–1003
 Selenium and Peripheral Arterial Disease 999

Table 2. Characteristics of the Study Population by Quartile of Serum Selenium Level, National Health and
Nutrition Examination Survey, 2003–2004a

Quartile of Serum Selenium (ng/mL)

P Value for
Quartile 1 Quartile 2 Quartile 3 Quartile 4 Linear Trendb
(<125) (125–134) (135–147) (‡148)

Median serum selenium, ng/mL 118 131 142 157

Age, years 55.9 55.6 56.1 57.1 0.04
Sex: men 36.0 47.1 56.9 56.0 <0.001
Race-ethnicity
White 77.3 77.9 82.4 83.6 0.02
Black 14.2 8.9 7.7 5.1 <0.001
Mexican American 3.2 4.8 4.4 5.3 0.05
Education <high school 15.8 13.3 12.3 13.2 0.54
Family income <$20,000 25.5 14.3 16.8 19.8 0.33
Postmenopausal women 36.8 30.3 30.3 28.9 0.05
Cigarette smoking
Former 28.0 35.4 31.9 38.7 0.03
Current 29.4 19.8 15.9 12.6 <0.001
Serum cotinine, ng/mL 0.96 0.44 0.33 0.21 <0.001
Current alcohol drinking 27.0 32.2 33.5 31.9 0.47
Body mass index, kg/m2 28.8 28.2 28.8 27.6 0.01
Dietary supplement use 54.0 63.0 67.0 71.6 0.001
C-reactive protein, mg/L 2.4 2.0 2.0 1.9 0.09
Total cholesterol, mg/dL 199.0 204.9 209.9 220.8 <0.001
High density lipoprotein cholesterol, mg/dL 52.5 54.9 54.6 56.3 0.01
Cholesterol-lowering-medication use 16.7 15.3 19.7 20.4 0.04
Systolic blood pressure, mm Hg 125.3 126.2 129.2 128.8 0.02
Blood-pressure-lowering-medication use 27.8 25.6 29.7 31.2 0.18
Diabetes 10.6 8.8 10.0 13.9 0.25
Glomerular filtration rate 5.6 6.0 6.3 7.3 0.24
<60 mL/minute per 1.73 m2
Peripheral arterial disease 5.0 3.3 2.5 2.5 0.02
a
Values are expressed as percentages for categorical variables or means for continuous variables adjusted for
age (years), sex, and race-ethnicity; for serum cotinine and C-reactive protein, adjusted geometric means are
reported.
b
P value for linear trend in percentages or means across quartiles of serum selenium adjusted for age (years), sex,
and race-ethnicity.

and peripheral arterial disease prevalence (P for quadratic
spline terms ¼ 0.06 and 0.12 after adjustment for age, sex,
and race and after full adjustment, respectively).
DISCUSSION
In this cross-sectional study, conducted in a representative
sample of the US population, the association between serum
selenium levels and the prevalence of peripheral arterial
disease was not statistically significant, although a U-shaped
relation was suggested: the prevalence of peripheral arterial
disease decreased with increasing serum selenium levels up
to 150 ng/mL but increased with increasing selenium levels
above 160 ng/mL. Selenium intake varies around the world
primarily because of geographic variation in the amount of
selenium in the soil (1, 23). In the United States, estimated
selenium intake ranges from 60 lg/day to 220 lg/day (23),
higher than the recommended dietary allowance for healthy
adults (55 lg/day) (4). As a consequence, serum selenium
levels in the United States are high: in NHANES 2003–
2004, the median selenium level was 134 ng/mL, and 99%
of study participants had serum selenium levels above 95
ng/mL. These concentrations are considerably higher than
in other countries. In Europe, for instance, average serum
selenium levels range from 50 ng/mL to 90 ng/mL (23, 24).
Very limited data are available on the association of
selenium with peripheral arterial disease. Two small studies
found similar selenium levels in patients with peripheral
arterial disease compared with controls, but the dose-
response relation was not evaluated (12, 13). For other car-
diovascular outcomes, most prospective studies have been

Am J Epidemiol 2009;169:996–1003
1000 Bleys et al.

Table 3. Odds Ratios and 95% Confidence Intervals for Peripheral Arterial Disease by Quartile of Serum Selenium Level, National Health and
Nutrition Examination Survey, 2003–2004

Quartile of Serum Selenium (ng/mL)

P Value for
Quartile 1 (<125) Quartile 2 (125–134) Quartile 3 (135–147) Quartile 4 (‡148)
Linear Trenda
OR 95% CI OR 95% CI OR 95% CI OR 95% CI

Median serum selenium, ng/mL 118 131 142 157

Cases/noncases 50/440 43/453 38/515 38/485
Model 1b 1.00 Reference 0.65 0.40, 1.07 0.48 0.26, 0.90 0.49 0.28, 0.85 0.02
c
Model 2 1.00 Reference 0.73 0.41, 1.30 0.57 0.29, 1.11 0.64 0.33, 1.23 0.16
Model 3d 1.00 Reference 0.75 0.37, 1.52 0.58 0.28, 1.19 0.67 0.34, 1.31 0.18

Abbreviations: CI, confidence interval; OR, odds ratio.

a
P value for linear risk trend across quartiles of serum selenium.
b
Adjusted for age (years), sex (men, women), and race-ethnicity (white, black, Mexican American, other).
c
Further adjusted for education (<high school, high school), family income (<$20,000, $20,000), postmenopausal status for women (yes,
no), cigarette smoking (never, former, current), serum cotinine (log-transformed), alcohol consumption (yes, no), body mass index (kg/m2), and
dietary supplement use (yes, no).
d
Further adjusted for C-reactive protein (log-transformed), total cholesterol (mg/dL), high density lipoprotein cholesterol (mg/dL), cholesterol-
lowering-medication use (yes, no), systolic blood pressure (mm Hg), blood-pressure-lowering-medication use (yes, no), diabetes (yes, no), and
glomerular filtration rate (<60, 60–<90, 90 mL/minute per 1.73 m2).

conducted in populations with suboptimal selenium levels in Findings from the only 2 prospective studies of serum
Europe or China (3, 25–36). These studies tended to report selenium levels and coronary heart disease conducted in
inverse associations between serum selenium levels and cor- the United States, however, are consistent with a U-shaped
onary heart disease incidence, but their sample sizes were relation (6, 8). In the Physicians’ Health Study, the relative
too small for detailed dose-response analyses. risks for incident myocardial infarction comparing quintiles
2–5 of plasma selenium with the lowest quintile were 0.87,
0.82, 0.60, and 1.53, respectively (6). The cutoff levels for
quintiles 1 and 5 of serum selenium in this study were 92
Odds Ratio for Peripheral Arterial Disease

2 25 ng/mL and 134 ng/mL, respectively. In the NHANES III

Mortality Study, the relative risks for cardiovascular dis-
20 ease mortality comparing tertiles 2 and 3 of serum sele-
nium with the lowest tertile were 0.90 and 0.98,
Weighted Percentage

1 respectively; for stroke mortality, the corresponding rel-

0.5
0.25
100 120 140 160
Serum Selenium, ng/ml
Figure 1. Odds ratios for peripheral arterial disease by serum sele-
nium levels, National Health and Nutrition Examination Survey, 2003–
2004. Curves represent adjusted odds ratios (solid line) and their 95%
confidence intervals (dashed lines) based on restricted quadratic
splines for serum selenium levels with knots at the 5th, 50th, and
95th percentiles. The reference value (odds ratio ¼ 1) was set at the
10th percentile of serum selenium distribution (116 ng/mL). Odd ratios
were adjusted for age, sex, race-ethnicity, education, family income,
postmenopausal status, smoking, serum cotinine, alcohol consump-
tion, body mass index, dietary supplement use, C-reactive protein, total
cholesterol, high density lipoprotein cholesterol, cholesterol-lowering
medication use, systolic blood pressure, blood-pressure-lowering
medication use, diabetes, and glomerular filtration rate. Bars represent
the weighted histogram of serum selenium distribution.
15
ative risks were 0.73 and 1.23 (8). The cutoff levels for
serum selenium tertiles in NHANES III were 117.3 ng/mL and
10 130.4 ng/mL, respectively. In this study, a dose-response anal-
ysis showed that cardiovascular and coronary heart disease
mortality decreased with increasing serum selenium levels
5
up to 120 ng/mL followed by an increase at higher levels,
although the U-shaped relation was not statistically significant
0 (8). Finally, in the Health Professionals Follow-up Study, the
180 200 odds ratios for incident coronary heart disease comparing
quintiles 2–5 of toenail selenium levels with the first quintile
were 1.03, 0.99, 1.32, and 0.86, respectively, with no clear
dose-response relation (7). Both serum and toenail selenium
levels reflect selenium status, although toenails reflect longer-
term exposure. It is unclear, however, whether both biomarkers
are comparable in their ability to capture the different types of
selenium compounds.
Few randomized trials have evaluated the effect of selenium
supplementation on cardiovascular outcomes or atherosclero-
sis progression, and most of these studies combined selenium
with other vitamins and minerals (3, 37). Only 2 of these trials
were conducted in the United States, both reporting null results
(38, 39). In the Nutritional Prevention of Cancer trial, the
relative risk for cardiovascular disease incidence comparing

Am J Epidemiol 2009;169:996–1003

200 lg/day of selenium supplementation with placebo was
1.03 (95% confidence interval: 0.78, 1.37) (38). In the HDL-
Atherosclerosis Treatment Study, the progression of athero-
sclerosis measured by coronary angiography in patients with
coronary artery disease was similar among participants ran-
domized to an antioxidant supplement containing 100 lg/day
of selenium, 800 IU/day of vitamin E, 1 g/day of vitamin C,
and 25 mg/day of b-carotene and participants randomized to
placebo (39). Overall, limited evidence from randomized trials
has not shown a protective effect of selenium supplementation
in US studies. With respect to observational studies, those in
the United States have not been able to detect a significant
linear association between serum selenium and cardiovascular
outcomes, but the dose-response associations in these studies
were U-shaped.
The biologic mechanisms underlying a potential effect of
selenium on cardiovascular disease are likely complex, but
they may be related to the dual role of selenium as an es-
sential and toxic element. Selenium is an essential micro-
nutrient that is incorporated into glutathione peroxidases
and other selenoproteins (4). Increasing serum selenium
levels increase the concentration and activity of glutathione
peroxidases, but this dose-response relation reaches a pla-
teau at serum selenium levels of 70–90 ng/mL (4). As a con-
sequence, higher selenium levels could potentially prevent
atherosclerosis development and progression in populations
whose selenium exposure is below the levels needed to
maximize glutathione peroxidases (1–3). In selenium-
replete populations such as in the United States, in which
virtually all participants have serum selenium levels above
70–90 ng/mL, the mechanisms underlying a potential ben-
eficial effect of increased selenium levels are unclear. Since
selenium supplementation is actively promoted in the
United States, and large randomized controlled trials testing
the efficacy of selenium supplementation in prostate cancer
prevention are under way (40, 41), mechanistic studies are
urgently needed to establish the biologic basis for a protec-
tive effect of selenium in populations whose selenium status
is already high.
Selenium, however, has a narrow therapeutic range (4),
and it may even be harmful at intake levels below the current
tolerable Upper Intake Level of 400 lg/day (2). In fact,
some selenium compounds have been documented to gen-
erate reactive oxygen species (42–44), and the upturn in
peripheral arterial disease prevalence that we observed at
selenium levels above 160 ng/mL could be associated with
selenium-induced increased oxidative stress. This upturn in
risk is also consistent with recent reports showing increased
risk of diabetes (45, 46) and elevated lipid levels (47) with
high selenium levels in US populations. For instance, the
Nutritional Prevention of Cancer trial showed an increased
risk of diabetes for participants receiving 200 lg/day of
selenium compared with placebo (hazard ratio ¼ 1.50,
95% confidence interval: 1.03, 2.33) (46). Interestingly,
the excess risk was limited to participants in the upper tertile
of the serum selenium distribution (>121.6 ng/mL), who
had a hazard ratio for diabetes of 2.70 (95% confidence
interval: 1.30, 5.61). Further research is needed to establish
the mechanisms underlying the association of high-normal
selenium levels with peripheral arterial disease and with
Am J Epidemiol 2009;169:996–1003
Selenium and Peripheral Arterial Disease 1001
metabolic abnormalities, and to determine whether the
change point in risk associated with elevated selenium levels
depends on genetic polymorphisms in candidate genes for
selenium metabolism (48).
Several limitations of our study need to be considered.
The use of a cross-sectional design and of prevalent cases of
peripheral arterial disease limited our ability to determine
the direction and the causality of the observed association. It
is possible that the pathophysiologic changes of atheroscle-
rosis could modify serum selenium levels or that partici-
pants with peripheral arterial disease change their health
behaviors, including selenium intake through diet and di-
etary supplements. As a consequence, our findings must be
confirmed in prospective studies with incident cases of pe-
ripheral arterial disease. Another limitation of our study is
the use of a single measurement of serum selenium, which
reflects short-term selenium intake and may be subject to
high within-person variability (49). Furthermore, our study
measured only total serum selenium, and we did not have
information on selenoprotein levels or activity or about non-
specific incorporation of selenium as selenomethionine in
other plasma proteins. More detailed analysis of different
compartments of serum selenium will be needed to better
understand the association of selenium with peripheral
arterial disease. The strengths of our study come from
the rigorous sampling design and the quality of the study
measurements used in NHANES; the representativeness of
the NHANES sample; and the use of ABI, a noninvasive
measure of subclinical atherosclerosis.
In summary, the association between serum selenium lev-
els and the prevalence of peripheral arterial disease in
NHANES 2003–2004 was not statistically significant,
although a U-shaped relation was suggested. Other sources
of evidence (6, 8) also suggest a U-shaped relation between
serum selenium levels and cardiovascular outcomes in the
United States, a selenium-replete population. In many pop-
ulations worldwide, selenium intake is lower than in the
United States (1, 23). At these lower levels of selenium
intake, the association of selenium with peripheral arterial
disease remains unknown. Prospective studies of selenium
status across populations with different levels of selenium
intake and randomized trials stratified by baseline sele-
nium status are needed to establish the optimal selenium
levels to minimize the risk of cardiovascular and other
chronic diseases.
ACKNOWLEDGMENTS
Author affiliations: Department of Epidemiology, Johns
Hopkins Bloomberg School of Public Health, Baltimore,
Maryland (Joachim Bleys, Ana Navas-Acien, Martin
Laclaustra, Andy Menke, Eliseo Guallar); Department of
Medicine, Johns Hopkins School of Medicine, Baltimore,
Maryland (Eliseo Guallar); Welch Center for Prevention,
Epidemiology and Clinical Research, Johns Hopkins
Bloomberg School of Public Health and Johns Hopkins
School of Medicine, Baltimore, Maryland (Joachim Bleys,
Ana Navas-Acien, Martin Laclaustra, Andy Menke, Eliseo
1002 Bleys et al.
Guallar); Department of Environmental Health Sciences,
Johns Hopkins Bloomberg School of Public Health,
Baltimore, Maryland (Ana Navas-Acien); Department of
Cardiovascular Epidemiology and Population Genetics,
National Center for Cardiovascular Research (CNIC),
Madrid, Spain (Martin Laclaustra, Eliseo Guallar); National
Center for Epidemiology, Instituto de Salud Carlos III, and
the CIBER in Epidemiology and Public Health (CIBERESP),
Madrid, Spain (Roberto Pastor-Barriuso); Friedman School
of Nutrition Science and Policy at Tufts University, Boston,
Massachusetts (Jose M. Ordovas); and Clinical Sciences
Research Institute at Warwick Medical School, Coventry,
United Kingdom (Saverio Stranges).
Supported by grants 1 R01 ES012673 from the National
Institute of Environmental Health Sciences and 0230232N
from the American Heart Association.
Conflict of interest: none declared.
REFERENCES
1. Rayman MP. Food-chain selenium and human health:
emphasis on intake. Br J Nutr. 2008;100(2):254–268.
2. Navas-Acien A, Bleys J, Guallar E. Selenium intake and
cardiovascular risk: what is new? Curr Opin Lipidol. 2008;
19(1):43–49.
3. Flores-Mateo G, Navas-Acien A, Pastor-Barriuso R, et al.
Selenium and coronary heart disease: a meta-analysis. Am J
Clin Nutr. 2006;84(4):762–773.
4. Food and Nutrition Board, Institute of Medicine. Dietary
Reference Intakes for Vitamin C, Vitamin E, Selenium, and
Carotenoids. A report of the Panel on Dietary Antioxidants
and Related Compounds, Subcommittees on Upper Reference
Levels of Nutrients and Interpretation and Uses of Dietary
Reference Intakes, and the Standing Committee on the Sci-
entific Evaluation of Dietary Reference Intakes. Washington,
DC: The National Academies Press; 2000.
5. Rayman MP. Selenium in cancer prevention: a review of the
evidence and mechanism of action. Proc Nutr Soc. 2005;
64(4):527–542.
6. Salvini S, Hennekens CH, Morris JS, et al. Plasma levels of
the antioxidant selenium and risk of myocardial infarction
among U.S. physicians. Am J Cardiol. 1995;76(17):
1218–1221.
7. Yoshizawa K, Ascherio A, Morris JS, et al. Prospective study
of selenium levels in toenails and risk of coronary heart dis-
ease in men. Am J Epidemiol. 2003;158(9):852–860.
8. Bleys J, Navas-Acien A, Guallar E. Serum selenium levels and
all-cause, cancer, and cardiovascular mortality among US
adults. Arch Intern Med. 2008;168(4):404–410.
9. McDermott MM, Liu K, Criqui MH, et al. Ankle-brachial
index and subclinical cardiac and carotid disease: the
Multi-Ethnic Study of Atherosclerosis. Am J Epidemiol. 2005;
162(1):33–41.
10. Murabito JM, Evans JC, Larson MG, et al. The ankle-brachial
index in the elderly and risk of stroke, coronary disease, and
death: the Framingham Study. Arch Intern Med. 2003;163(16):
1939–1942.
11. Newman AB, Shemanski L, Manolio TA, et al. Ankle-arm
index as a predictor of cardiovascular disease and mortality in
the Cardiovascular Health Study. The Cardiovascular Health
Study Group. Arterioscler Thromb Vasc Biol. 1999;19(3):
538–545.
12. Ondrus P, Alberty R, Vassanyiova Z. Importance of lipid per-
oxidation, protective enzymes and trace elements in chronic
leg ischaemia. Eur J Clin Chem Clin Biochem. 1996;34(6):
471–475.
13. Mansoor MA, Bergmark C, Haswell SJ, et al. Correlation
between plasma total homocysteine and copper in patients
with peripheral vascular disease. Clin Chem. 2000;46(3):
385–391.
14. Hirsch AT, Haskal ZJ, Hertzer NR, et al. ACC/AHA 2005
Practice Guidelines for the management of patients with
peripheral arterial disease (lower extremity, renal, mesen-
teric, and abdominal aortic): a collaborative report from the
American Association for Vascular Surgery/Society for
Vascular Surgery, Society for Cardiovascular Angiography
and Interventions, Society for Vascular Medicine and
Biology, Society of Interventional Radiology, and the ACC/
AHA Task Force on Practice Guidelines (Writing Committee
to Develop Guidelines for the Management of Patients With
Peripheral Arterial Disease): endorsed by the American
Association of Cardiovascular and Pulmonary Rehabilitation;
National Heart, Lung, and Blood Institute; Society for Vas-
cular Nursing; TransAtlantic Inter-Society Consensus; and
Vascular Disease Foundation. Circulation. 2006;113(11):
e463–e654.
15. Fowkes FG, Murray GD, Butcher I, et al. Ankle brachial index
combined with Framingham Risk Score to predict cardiovas-
cular events and mortality: a meta-analysis. JAMA. 2008;
300(2):197–208.
16. National Center for Health Statistics, Centers for Disease
Control and Prevention. National Health and Nutrition
Examination Survey, 2003–2004. 2007. (http://www.cdc.gov/
nchs/about/major/nhanes/nhanes2003-2004/nhanes03_04.htm).
(Accessed August 24, 2008).
17. National Center for Health Statistics, Centers for Disease
Control and Prevention. NHANES 2003–2004 Data Docu-
mentation. Laboratory Assessment: Lab 39—Erythrocyte
Protoporphyrin and Selenium. 2007. (http://www.cdc.gov/
nchs/data/nhanes/nhanes_03_04/l39epp_c.pdf). (Accessed
August 24, 2008).
18. National Center for Health Statistics, Centers for Disease
Control and Prevention. NHANES 2003–2004 Data Docu-
mentation. Exam Component: Ankle-Brachial Blood Pressure
Index (ABPI) Section of the Lower Extremity Disease Ex-
amination (LEXAB_C). 2005. (http://www.cdc.gov/nchs/
data/nhanes/nhanes_03_04/lexab_c.pdf). (Accessed August
24, 2008).
19. Levey AS, Coresh J, Greene T, et al. Using standardized serum
creatinine values in the Modification of Diet in Renal Disease
Study equation for estimating glomerular filtration rate. Ann
Intern Med. 2006;145(4):247–254.
20. Greenland S. Dose-response and trend analysis in epidemiol-
ogy: alternatives to categorical analysis. Epidemiology. 1995;
6(4):356–365.
21. R Development Core Team. R: A Language and Environment
for Statistical Computing. Vienna, Austria: R Foundation for
Statistical Computing; 2008. (ISBN 3-900051-07-0). (http://
www.R-project.org).
22. Lumley T. The Survey Package. (http://cran.fhcrc.org/web/
packages/survey/survey.pdf). (Accessed August 24, 2008).
23. Combs GF Jr. Selenium in global food systems. Br J Nutr.
2001;85(5):517–547.
24. Rayman MP. The importance of selenium to human health.
Lancet. 2000;356(9225):233–241.
25. Salonen JT, Alfthan G, Huttunen JK, et al. Association
between cardiovascular death and myocardial infarction and
Am J Epidemiol 2009;169:996–1003

serum selenium in a matched-pair longitudinal study. Lancet.
1982;2(8291):175–179.
26. Miettinen TA, Alfthan G, Huttunen JK, et al. Serum selenium
concentration related to myocardial infarction and fatty acid
content of serum lipids. Br Med J (Clin Res Ed). 1983;
287(6391):517–519.
27. Salonen JT, Salonen R, Penttila I, et al. Serum fatty acids,
apolipoproteins, selenium and vitamin antioxidants and the
risk of death from coronary artery disease. Am J Cardiol.
1985;56(4):226–231.
28. Virtamo J, Valkeila E, Alfthan G, et al. Serum selenium and
the risk of coronary heart disease and stroke. Am J Epidemiol.
1985;122(2):276–282.
29. Ringstad J, Thelle D. Risk of myocardial infarction in relation
to serum concentrations of selenium. Acta Pharmacol Toxicol
(Copenh). 1986;59(suppl 7):336–339.
30. Kok FJ, de Bruijn AM, Hofman A, et al. Selenium status and
chronic disease mortality: Dutch epidemiological findings. Int
J Epidemiol. 1987;16(2):329–332.
31. Ringstad J, Jacobsen BK, Thomassen Y, et al. The Tromsø
Heart Study: serum selenium and risk of myocardial infarction
a nested case-control study. J Epidemiol Community Health.
1987;41(4):329–332.
32. Suadicani P, Hein HO, Gyntelberg F. Serum selenium con-
centration and risk of ischaemic heart disease in a prospective
cohort study of 3000 males. Atherosclerosis. 1992;96(1):
33–42.
33. Marniemi J, Jarvisalo J, Toikka T, et al. Blood vitamins,
mineral elements and inflammation markers as risk factors of
vascular and non-vascular disease mortality in an elderly
population. Int J Epidemiol. 1998;27(5):799–807.
34. Kilander L, Berglund L, Boberg M, et al. Education, lifestyle
factors and mortality from cardiovascular disease and cancer.
A 25-year follow-up of Swedish 50-year-old men. Int J Epi-
demiol. 2001;30(5):1119–1126.
35. Wei WQ, Abnet CC, Qiao YL, et al. Prospective study of
serum selenium concentrations and esophageal and gastric
cardia cancer, heart disease, stroke, and total death. Am J Clin
Nutr. 2004;79(1):80–85.
36. Akbaraly NT, Arnaud J, Hininger-Favier I, et al. Selenium and
mortality in the elderly: results from the EVA study. Clin
Chem. 2005;51(11):2117–2123.
37. Bleys J, Miller ER III, Pastor-Barriuso R, et al. Vitamin-
mineral supplementation and the progression of atherosclerosis:
Am J Epidemiol 2009;169:996–1003
Selenium and Peripheral Arterial Disease 1003
a meta-analysis of randomized controlled trials. Am J Clin Nutr.
2006;84(4):880–887.
38. Stranges S, Marshall JR, Trevisan M, et al. Effects of selenium
supplementation on cardiovascular disease incidence and
mortality: secondary analyses in a randomized clinical trial.
Am J Epidemiol. 2006;163(8):694–699.
39. Brown BG, Zhao XQ, Chait A, et al. Simvastatin and niacin,
antioxidant vitamins, or the combination for the prevention of
coronary disease. N Engl J Med. 2001;345(22):1583–1592.
40. Lippman SM, Goodman PJ, Klein EA, et al. Designing the
Selenium and Vitamin E Cancer Prevention Trial (SELECT).
J Natl Cancer Inst. 2005;97(2):94–102.
41. Marshall JR, Sakr W, Wood D, et al. Design and progress of
a trial of selenium to prevent prostate cancer among men with
high-grade prostatic intraepithelial neoplasia. Cancer
Epidemiol Biomarkers Prev. 2006;15(8):1479–1484.
42. Spallholz JE. On the nature of selenium toxicity and carcino-
static activity. Free Radic Biol Med. 1994;17(1):45–64.
43. Spallholz JE, Palace VP, Reid TW. Methioninase and
selenomethionine but not Se-methylselenocysteine generate
methylselenol and superoxide in an in vitro chemiluminescent
assay: implications for the nutritional carcinostatic activity
of selenoamino acids. Biochem Pharmacol. 2004;67(3):
547–554.
44. Drake EN. Cancer chemoprevention: selenium as a prooxi-
dant, not an antioxidant. Med Hypotheses. 2006;67(2):
318–322.
45. Bleys J, Navas-Acien A, Guallar E. Serum selenium and
diabetes in U.S. adults. Diabetes Care. 2007;30(4):829–834.
46. Stranges S, Marshall JR, Natarajan R, et al. Effects of long-
term selenium supplementation on the incidence of type 2
diabetes: a randomized trial. Ann Intern Med. 2007;147(4):
217–223.
47. Bleys J, Navas-Acien A, Stranges S, et al. Serum selenium and
serum lipids in US adults. Am J Clin Nutr. 2008;88(2):
416–423.
48. Méplan C, Crosley LK, Nicol F, et al. Genetic polymorphisms
in the human selenoprotein P gene determine the response of
selenoprotein markers to selenium supplementation in a gender-
specific manner (the SELGEN study). FASEB J. 2007;21(12):
3063–3074.
49. Longnecker MP, Stampfer MJ, Morris JS, et al. A 1-y trial of
the effect of high-selenium bread on selenium concentrations
in blood and toenails. Am J Clin Nutr. 1993;57(3):408–413.
 American Journal of Epidemiology Vol. 169, No. 8
Published by the Johns Hopkins Bloomberg School of Public Health 2009. DOI: 10.1093/aje/kwp011
Advance Access publication March 3, 2009

Original Contribution

Ambient Air Pollution and Cardiovascular Malformations in Atlanta, Georgia,

1986–2003

Matthew J. Strickland, Mitchel Klein, Adolfo Correa, Mark D. Reller, William T. Mahle, Tiffany
J. Riehle-Colarusso, Lorenzo D. Botto, W. Dana Flanders, James A. Mulholland, Csaba Siffel,
Michele Marcus, and Paige E. Tolbert

Initially submitted February 22, 2008; accepted for publication January 8, 2009.

Associations between ambient air pollution levels during weeks 3–7 of pregnancy and risks of cardiovascular
malformations were investigated among the cohort of pregnancies reaching at least 20 weeks’ gestation that were
conceived during January 1, 1986–March 12, 2003, in Atlanta, Georgia. Surveillance records obtained from the
Metropolitan Atlanta Congenital Defects Program, which conducts active, population-based surveillance on this
cohort, were reviewed to classify cardiovascular malformations. Ambient 8-hour maximum ozone and 24-hour
average carbon monoxide, nitrogen dioxide, particulate matter with an average aerodynamic diameter of <10 lm
(PM10), and sulfur dioxide measurements were obtained from centrally located stationary monitors. Temporal
associations between these pollutants and daily risks of secundum atrial septal defect, aortic coarctation, hypo-
plastic left heart syndrome, patent ductus arteriosus, valvar pulmonary stenosis, tetralogy of Fallot, transposition of
the great arteries, muscular ventricular septal defect, perimembranous ventricular septal defect, conotruncal
defects, left ventricular outflow tract defect, and right ventricular outflow defect were modeled by using Poisson
generalized linear models. A statistically significant association was observed between PM10 and patent ductus
arteriosus (for an interquartile range increase in PM10 levels, risk ratio ¼ 1.60, 95% confidence interval: 1.11, 2.31).
Of the 60 associations examined in the primary analysis, no other significant associations were observed.

air pollution; heart defects, congenital

Abbreviations: CI, confidence interval; MACDP, Metropolitan Atlanta Congenital Defects Program; PM10,, particulate matter with
an average aerodynamic diameter of <10 lm; RR, risk ratio.

A growing body of epidemiologic evidence suggests as-
sociations between ambient air pollution and adverse preg-
nancy outcomes (1–6). Associations between air pollution
levels during pregnancy and risks of cardiovascular malfor-
mations among the offspring were investigated in 2 previous
population-based case-control studies (7, 8). In these stud-
ies, cardiovascular malformations were classified by using
preexisting surveillance database codes, and analyses were
based on contrasts in ambient pollution levels over space
and time.
In the first study, conducted in southern California, inves-
tigators reported an association between ambient carbon
monoxide levels and risk of ventricular septal defects
(fourth quartile vs. first quartile, odds ratio ¼ 2.95, 95%
confidence interval (CI): 1.44, 6.05). Elevated risks of aortic
artery and valve defects, pulmonary artery and valve de-
fects, and conotruncal defects with increasing ambient
ozone levels were also reported (7). The second investiga-
tion, conducted in Texas, did not corroborate the southern
California findings, although a suggestive association be-
tween ozone and pulmonary artery and valve defects was
observed. The Texas investigators reported positive associ-
ations for carbon monoxide and tetralogy of Fallot, particu-
late matter with an average aerodynamic diameter of <10 lm
(PM10) and atrial septal defects, and sulfur dioxide and ven-
tricular septal defects (8).

Correspondence to Dr. Matthew Strickland, Department of Environmental and Occupational Health, Emory University, 1518 Clifton Road NE,
Atlanta, GA 30322 (e-mail: mjstric@sph.emory.edu).

1004 Am J Epidemiol 2009;169:1004–1014


Table 1. Cardiovascular Malformation Outcome Groups, Number of Cases,a and Outcome Group Definitions
Outcome Group No. of Cases
Atrial septal defect, secundum 379
Coarctation of the aorta 275
Hypoplastic left heart syndrome 175
Patent ductus arteriosus 219
Pulmonary stenosis, valvar 312
Tetralogy of Fallot 299
Transposition of the great arteries 165
Ventricular septal defect, muscular 1,108
Ventricular septal defect, perimembranous 546
Conotruncal defectb 661
Left ventricular outflow tract defectb 558
Right ventricular outflow tract defectb 421
a
Number of cases identified among the cohort of pregnancies reaching at least 20 weeks’ gestation in Atlanta, Georgia, with an estimated date
of conception during January 1, 1986–March 12, 2003.
b
Aggregate grouping of cardiovascular malformations.
We conducted a retrospective cohort study in Atlanta,
Georgia, to explore temporal associations between ambient
air pollution levels during pregnancy and risks of cardiovas-
cular malformations. We did not closely replicate the meth-
odologies of the previous studies. In addition to the
retrospective cohort design and the temporal analytical ap-
proach, we reviewed and reclassified each surveillance rec-
ord using a modified version of the International Pediatric
and Congenital Cardiac Code implemented in the Society of
Thoracic Surgeons Congenital Heart Surgery Database (9–13).
This activity permitted classification of cardiovascular
malformations according to embryologic (rather than only
anatomic) considerations.
MATERIALS AND METHODS
Study population
Vital records for the cohort of liveborn and stillborn in-
fants of at least 20 weeks’ gestation whose mothers resided
in 1 of 5 central Atlanta counties at delivery were obtained
from the Office of Health Information and Policy, Georgia
Division of Public Health. The records of infants with an
Am J Epidemiol 2009;169:1004–1014
Ambient Air Pollution and Cardiovascular Malformations 1005
Definition
Includes secundum-type atrial septal defect.
Includes coarctation of the aorta, aortic arch hypoplasia, and interrupted aortic
arch type A.
Includes hypoplastic left heart syndrome with or without ventricular septal defect.
Includes term infants (
36 weeks’ gestation) with patent ductus arteriosus
persisting for
6 weeks following delivery. Infants were excluded if the
patent ductus arteriosus was an obligatory shunt lesion or if patency
was maintained by prostaglandin infusion.
Includes valvar and unspecified pulmonary stenosis, as well as dysplastic
pulmonary valve.
Includes typical tetralogy of Fallot, tetralogy of Fallot with absent pulmonary valve,
pulmonary atresia with ventricular septal defect, pulmonary atresia with major
aortopulmonary collateral arteries, and tetralogy of Fallot-type double-outlet
right ventricle.
Includes all types of transposition with concordant atrioventricular connections and
discordant ventricular arterial connections, with or without ventricular septal
defect or left ventricular outflow tract obstruction. Also includes double-outlet
right ventricle with malpositioned great arteries.
Includes muscular-type ventricular septal defect.
Includes perimembranous-type ventricular septal defect.
Includes all cardiovascular malformations in the ‘‘Tetralogy of Fallot’’ and
‘‘Transposition of the great arteries’’ outcome groups. Also includes
aortopulmonary window defect, all other double-outlet right-ventricle variants,
interrupted aortic arch type B, unspecified interrupted aortic arch, vascular
rings, and subarterial-type ventricular septal defect.
Includes all cardiovascular malformations in the ‘‘Coarctation of the aorta’’ and
‘‘Hypoplastic left heart syndrome’’ outcome groups. Also includes stenosis and
atresia of the aortic valve and isolated bicuspid aortic valve.
Includes all cardiovascular malformations in the ‘‘Pulmonary stenosis, valvar’’
outcome group. Also includes pulmonary valve atresia with intact ventricular
septum, tricuspid valve atresia, double-chambered right ventricle, and isolated
supravalvar pulmonary artery stenosis.
indication of a cardiovascular malformation were obtained
from the Metropolitan Atlanta Congenital Defects Program
(MACDP), which conducts active, population-based birth
defects surveillance on the cohort of pregnancies reaching
at least 20 weeks’ gestation to mothers residing in 1 of 5
central Atlanta counties at delivery/termination (14).
MACDP ascertains livebirths, stillbirths, and elective termi-
nations with major structural defects, chromosomal abnor-
malities, and clinical syndromes diagnosed by the age of
6 years. When available, details from echocardiography,
catheterization, and surgical reports are abstracted; general
pregnancy information, such as gestational age and birth
weight, is also collected.
Pregnancies with an estimated date of conception during
January 1, 1986–March 12, 2003, were included in the anal-
ysis. For each conception date, we estimated the number of
pregnancies with a cardiovascular malformation (numera-
tor) and total pregnancies (denominator). Approximately
0.3% of pregnancies with cardiovascular malformations
and 1.9% of total pregnancies were excluded because of
missing or implausible gestational age information (defined
as <20 weeks or >44 weeks). For each pregnancy, we es-
timated the date of conception (assuming that conception
1006 Strickland et al.

Table 2. Prevalencea of Cardiovascular Malformations, by Season and Year of Conception, for the Cohort of Pregnancies Reaching at Least 20
Weeks’ Gestation in Atlanta, Georgia, With an Estimated Date of Conception During January 1, 1986–March 12, 2003

Prevalence,a by Year of
Prevalence,a by Season of Conception
No. of Conception Overall
Cases March– June– September– December– 1986– 1992– 1998– Prevalencea
May August November February 1991 1997 2003

Atrial septal defect, secundum 379 4.2 5.4 5.9 5.6 2.8 4.7 7.9 5.3
Coarctation of the aorta 275 3.6 4.0 3.1 4.6 3.9 3.8 3.8 3.8
Hypoplastic left heart syndrome 175 2.0 2.2 3.1 2.4 2.9 2.3 2.3 2.5
Patent ductus arteriosus 219 2.8 3.1 2.9 3.5 3.6 3.1 2.7 3.1
Pulmonary stenosis, valvar 312 3.6 4.6 4.8 4.4 2.7 4.6 5.4 4.4
Tetralogy of Fallot 299 4.4 3.8 4.0 4.5 4.2 4.0 4.3 4.2
Transposition of the great arteries 165 2.3 2.2 2.2 2.5 2.2 2.2 2.5 2.3
Ventricular septal defect, muscular 1,108 14.1 14.8 17.1 15.8 5.3 14.3 24.9 15.5
Ventricular septal defect, 546 6.6 7.8 8.4 7.7 4.8 8.1 9.5 7.6
perimembranous
Conotruncal defectb 661 9.1 8.3 9.1 10.4 8.4 9.0 10.2 9.2
Left ventricular outflow tract defectb 558 6.4 7.5 8.6 8.7 7.9 7.6 7.9 7.8
Right ventricular outflow tract defectb 421 5.0 6.2 6.1 6.2 3.9 6.9 6.5 5.9
a
Prevalence per 10,000 pregnancies reaching at least 20 weeks’ gestation.
b
Aggregate grouping of cardiovascular malformations.

occurred 14 days after the last menstrual period date) using
vital records data. Unfortunately, these estimates were un-
reliable; for 30% of pregnancies, the last menstrual period
date was on the 15th of the month. When we subtracted the
clinical gestational age estimate from the birth date, too
many last menstrual period dates fell between the 12th
and 18th of the month.
To compensate for this data quality limitation, we used
gestational age estimates from MACDP surveillance rec-
ords (which are obtained from medical chart review and
are frequently based on ultrasound measurements) to es-
timate conception dates for pregnancies with cardiovascu-
lar malformations, because day-of-month patterns were
not evident among these estimates. For the denominators,
we modeled the daily count of conceptions. We calculated
the average daily number of conceptions for each month
using vital records information. We then created a daily
time-series data set, in which each day was assigned
its monthly average, and fit a cubic spline with 6 knots
per year to it. The predicted values from this model were
used as the daily estimates of conceptions (n ¼ 715,500
pregnancies).
Cardiovascular malformation outcome groups
Each MACDP surveillance record with a diagnosis of
a cardiovascular malformation was reviewed by a pediatric
cardiologist and classified by using the Society of Thoracic
Surgeons Congenital Heart Surgery Database, version 2.30,
nomenclature (11–13, 15). This nomenclature is specific to
cardiovascular malformations and is more detailed than
what is typically used in birth defects surveillance (16–
18). We classified infants with isolated transient newborn
cardiac conditions (e.g., patent foramen ovale) as physio-
logically normal, and we placed infants with more than 1
cardiovascular malformation in multiple outcome groups
only when these malformations were thought to be embry-
ologically independent; otherwise, we coded only the major
cardiovascular malformation. Further details about this ac-
tivity are available (12).
For our analyses, we excluded infants who had normal
cardiac physiology and those with identified trisomies, evi-
dence of heterotaxy syndrome, and abnormal cardiac loop-
ing. Results are presented for 12 outcome groups; 3 of these
are aggregate groupings of cardiovascular malformations
(Table 1).
Ambient air quality data
Ambient air quality measurements of daily 8-hour max-
imum ozone and 24-hour average carbon monoxide, nitro-
gen dioxide, PM10, and sulfur dioxide were obtained from
the US Environmental Protection Agency Air Quality Sys-
tem, Georgia Department of Natural Resources, and the
Metro Atlanta Index. For each pollutant, we selected 1 cen-
tral monitoring station for use in analyses. When central
station measurements of carbon monoxide, nitrogen diox-
ide, and sulfur dioxide were missing, pollution levels at the
central station were modeled by using measurements from
other stations. The central station for ozone did not operate
during November–February. During 1993–2003, wintertime
ozone levels were modeled by using ozone measurements
from a nearby monitor, with the model developed by use of
data during the later time period when wintertime ozone
levels were available. During 1986–1992, wintertime ozone
levels were modeled by using maximum temperature and
1-hour maximum nitrogen dioxide measurements from
a nearby monitor. Measurements of PM10 were available

Am J Epidemiol 2009;169:1004–1014
 Ambient Air Pollution and Cardiovascular Malformations 1007

Table 3. Interquartile Range and Mean Values, by Season and Year of Conception, for the
Weighted 5-Week Air Pollution Metrica Assigned to the Cohort of Pregnancies Reaching at Least
20 Weeks’ Gestation in Atlanta, Georgia, With an Estimated Date of Conception During January
1, 1986–March 12, 2003

24-Hour 24-Hour 24-Hour

8-Hour 24-Hour
Nitrogen Carbon Sulfur
Ozone, PM10,
Dioxide, Monoxide, Dioxide,
ppbb,c mg/m3 c,d
ppbc ppmc ppbc

Interquartile range 29.9 14.2 5.7 0.3 4.0

Mean value, by season of
conception
March–May 54.6 36.0 24.2 0.6 5.4
June–August 56.5 38.7 22.6 0.8 5.4
September–November 25.4 31.2 26.9 0.9 6.9
December–February 29.2 27.3 26.5 0.7 7.1
Mean value, by year of
conception
1986–1991 43.3 43.2 28.0 0.7 8.7
1992–1997 39.8 30.0 24.3 0.8 5.5
1998–2003 41.2 25.8 22.5 0.7 4.0

Abbreviation: PM10, particulate matter with an average aerodynamic diameter of <10 lm.
a
The air pollution metric is a 5-week weighted average of daily ambient air pollution levels
measured at a central monitor during weeks 3–7 of pregnancy. Relative weights are 0.7, 0.9, and
1.0 for pollution levels during the first and last week, the second and fourth week, and the middle
week of the window, respectively.
b
The central station for ozone did not operate during winter months (November–February).
c
Daily central monitoring station measurements available: 67% (4,251 of 6,315 days) for
ozone, 41% (2,563 of 6,315 days) for PM10, 90% (5,670 of 6,315 days) for nitrogen dioxide,
92% (5,804 of 6,315 days) for carbon monoxide, and 94% (5,966 of 6,315 days) for sulfur dioxide.
When feasible, missing daily measurements were modeled: 33% (2,064 of 6,315 days) for ozone,
59% (3,735 of 6,315 days) for PM10, 10% (609 of 6,315 days) for nitrogen dioxide, 6% (388 of
6,315 days) for carbon monoxide, and 5% (311 of 6,315 days) for sulfur dioxide.
d
PM10 was measured every sixth day during 1986–1992, Sunday–Thursday during 1993–
1995, and daily during 1996–2003. The location of the PM10 central monitoring station changed
on January 1, 1993, and on January 1, 1998; on January 1, 1998, the measurement method
changed from the federal reference method to tapered element oscillating microbalance.

every sixth day during 1986–1992, Sunday–Thursday during
1993–1995, and daily during 1996–2003; linear interpolation
between measurements was used to estimate missing PM10
levels. The location of the PM10 central monitoring station
changed on January 1, 1993, and on January 1, 1998; on
January 1, 1998, the measurement method changed from
the federal reference method to the tapered element oscillat-
ing microbalance method.
Statistical analyses
For each conception date (January 1, 1986–March 12,
2003), we estimated the number of pregnancies with a par-
ticular cardiovascular malformation (numerator) and total
pregnancies (denominator). All pregnancies on a particular
conception date were assigned the same pollutant metric,
which was a weighted average of the 35 daily ambient air
pollution measurements during weeks 3–7 of pregnancy
(a period when the 4 chambers, inflow tract, and outflow tract
of the heart develop (19)). Relative weights were 0.7 for
measurements during weeks 3 and 7, 0.9 for measurements
during weeks 4 and 6, and 1.0 for measurements during
week 5. We chose this a priori weighting scheme, which
emphasizes pollution levels during the center of the window,
because of uncertainty in the date of conception estimates.
We then created 52 strata representing the week-of-year
as follows: across all calendar years we grouped January 1–
January 7 as the first week of the year, January 8–January 14
as the second week of the year, and so on. We included
February 29, when present, in the ninth week of the year
(February 26–March 4). The 52nd week of the year was
8 days long (December 24–December 31).
We modeled temporal associations between ambient air
pollution and daily risks of cardiovascular malformations
using Poisson generalized linear models with a log link
and scaled variance estimates. We modeled the pollution
metric as a continuous variable and used the natural loga-
rithm of total conceptions as the offset. We included indi-
cator variables for the 52 strata representing week-of-year to
control for potential confounding by factors with seasonal
variation, and we included a cubic spline for day of follow-
up with 1 knot per year to control for long-term trends. All
risk ratios and confidence intervals corresponded to an in-
terquartile range increase in the ambient pollutant metric.

Am J Epidemiol 2009;169:1004–1014
1008 Strickland et al.

Table 4. Spearman’s Partial Correlation Coefficients for the Weighted 5-Week Air Pollution
Metric Assigned to the Cohort of Pregnancies Reaching at Least 20 Weeks’ Gestation in Atlanta,
Georgia, With an Estimated Date of Conception During January 1, 1986–March 12, 2003a

24-Hour 24-Hour 24-Hour

8-Hour 24-Hour
Nitrogen Carbon Sulfur
Ozone, PM10,
Dioxide, Monoxide, Dioxide,
ppb mg/m3
ppb ppm ppb

8-Hour ozone, ppb 1.00

24-Hour PM10, lg/m3 0.49*** 1.00
24-Hour nitrogen 0.45*** 0.46*** 1.00
dioxide, ppb
24-Hour carbon 0.07*** 0.32*** 0.41*** 1.00
monoxide, ppm
24-Hour sulfur 0.30*** 0.41*** 0.39*** 0.23*** 1.00
dioxide, ppb

Abbreviation: PM10, particulate matter with an average aerodynamic diameter of <10 lm.
*** P < 0.001.
a
Linear regression models containing week-of-year indicator variables and a cubic spline for
day of follow-up with 1 knot per year were used to predict the daily pollution metrics. The pairwise
Spearman partial correlation coefficients were estimated by using the residuals from these mod-
els. The air pollution metric is a weighted average of the 35 daily ambient air pollution levels
measured at a central monitor during weeks 3–7 of pregnancy.

Models were created by using R statistical software, version poral controls by replacing the week-of-year indicator
2.5.0 (20). variables with a cubic spline for day-of-year that had 3
We performed several sensitivity analyses. In one sensi- knots; instead of including yearly knots in the cubic spline
tivity analysis, we relaxed the seasonal and long-term tem- for day of follow-up, we placed knots once every 3 years.

Table 5. Risk Ratios and 95% Confidence Intervals for Associations Between the Weighted 5-Week Air Pollution Metric and Cardiovascular
Malformations Among the Cohort of Pregnancies Reaching at Least 20 Weeks’ Gestation in Atlanta, Georgia, With an Estimated Date of
Conception During January 1, 1986–March 12, 2003a

8-Hour 24-Hour 24-Hour Nitrogen 24-Hour Carbon 24-Hour Sulfur

Ozone, ppb PM10, mg/m3 Dioxide, ppb Monoxide, ppm Dioxide, ppb
No. of
Cases 95% 95% 95% 95% 95%
Risk Risk Risk Risk Risk
Confidence Confidence Confidence Confidence Confidence
Ratio Ratio Ratio Ratio Ratio
Interval Interval Interval Interval Interval

Atrial septal defect, secundum 379 1.16 0.67, 2.00 1.12 0.82, 1.53 1.15 0.92, 1.43 0.92 0.74, 1.15 1.00 0.72, 1.38
Coarctation of the aorta 275 1.15 0.65, 2.06 1.15 0.84, 1.58 1.11 0.87, 1.41 0.99 0.79, 1.24 1.04 0.75, 1.43
Hypoplastic left heart 175 0.82 0.37, 1.84 0.89 0.60, 1.31 0.91 0.66, 1.24 0.80 0.60, 1.06 0.77 0.50, 1.18
syndrome
Patent ductus arteriosus 219 1.39 0.72, 2.68 1.60 1.11, 2.31 1.27 0.96, 1.70 1.18 0.92, 1.51 1.22 0.86, 1.74
Pulmonary stenosis, valvar 312 0.97 0.53, 1.75 0.87 0.63, 1.21 1.01 0.80, 1.28 1.06 0.86, 1.32 0.70 0.49, 1.00
Tetralogy of Fallot 299 1.09 0.59, 2.00 0.88 0.65, 1.20 0.94 0.74, 1.20 1.13 0.91, 1.40 0.85 0.61, 1.17
Transposition of the great 165 1.29 0.58, 2.85 1.12 0.74, 1.72 0.80 0.57, 1.11 0.94 0.68, 1.30 1.13 0.75, 1.71
arteries
Ventricular septal defect, 1,108 1.08 0.77, 1.50 1.01 0.83, 1.23 1.09 0.96, 1.24 0.99 0.85, 1.14 0.95 0.77, 1.17
muscular
Ventricular septal defect, 546 1.06 0.67, 1.68 0.94 0.73, 1.22 1.12 0.94, 1.33 0.96 0.81, 1.14 0.99 0.76, 1.28
perimembranous
Conotruncal defect 661 1.22 0.81, 1.85 0.99 0.80, 1.22 0.95 0.81, 1.12 1.04 0.89, 1.21 1.06 0.86, 1.31
Left ventricular outflow 558 1.09 0.70, 1.68 1.03 0.83, 1.29 1.01 0.85, 1.20 0.97 0.82, 1.13 0.97 0.76, 1.22
tract defect
Right ventricular outflow 421 0.73 0.44, 1.22 0.85 0.64, 1.12 1.02 0.84, 1.25 1.16 0.96, 1.40 0.74 0.55, 1.00
tract defect

Abbreviation: PM10, particulate matter with an average aerodynamic diameter of <10 lm.
a
Risk ratios and 95% confidence intervals correspond to an increase in the interquartile range of the 5-week air pollutant metric (a weighted
average of the 35 daily ambient air pollution levels measured at a central monitor during weeks 3–7 of pregnancy). The interquartile ranges were
29.9 ppb for ozone, 14.2 lg/m3 for PM10, 5.7 ppb for nitrogen dioxide, 0.30 ppm for carbon monoxide, and 4.0 ppb for sulfur dioxide.

Am J Epidemiol 2009;169:1004–1014
 Ambient Air Pollution and Cardiovascular Malformations 1009

Table 6. Sensitivity Analysis With Less Stringent Control of Seasonal and Long-Term Temporal Variation, With Risk Ratios and 95% Confidence
Intervals for Associations Between the Weighted 5-Week Air Pollution Metric and Cardiovascular Malformations Among the Cohort of Pregnancies
Reaching at Least 20 Weeks’ Gestation in Atlanta, Georgia, With an Estimated Date of Conception During January 1, 1986–March 12, 2003a

8-Hour 24-Hour 24-Hour Nitrogen 24-Hour Carbon 24-Hour Sulfur

Ozone, ppb PM10, mg/m3 Dioxide, ppb Monoxide, ppm Dioxide, ppb
No. of
Cases 95% 95% 95% 95% 95%
Risk Risk Risk Risk Risk
Confidence Confidence Confidence Confidence Confidence
Ratio Ratio Ratio Ratio Ratio
Interval Interval Interval Interval Interval

Atrial septal defect, secundum 379 1.25 0.74, 2.09 1.12 0.82, 1.52 1.16 0.94, 1.43 0.97 0.80, 1.18 1.05 0.77, 1.43
Coarctation of the aorta 275 1.21 0.68, 2.16 1.21 0.90, 1.63 1.03 0.82, 1.30 1.00 0.82, 1.22 1.02 0.75, 1.40
Hypoplastic left heart 175 1.37 0.63, 3.00 1.12 0.75, 1.66 0.92 0.68, 1.25 0.93 0.73, 1.20 0.84 0.55, 1.27
syndrome
Patent ductus arteriosus 219 1.26 0.65, 2.43 1.40 1.01, 1.95 1.40 1.07, 1.83 1.19 0.96, 1.47 1.27 0.90, 1.79
Pulmonary stenosis, valvar 312 1.23 0.70, 2.15 1.00 0.73, 1.37 1.07 0.86, 1.33 1.09 0.90, 1.32 0.80 0.57, 1.13
Tetralogy of Fallot 299 1.24 0.72, 2.14 0.98 0.73, 1.30 1.02 0.82, 1.26 1.10 0.92, 1.30 0.82 0.61, 1.10
Transposition of the great 165 1.70 0.83, 3.48 1.01 0.69, 1.49 0.77 0.58, 1.03 0.94 0.72, 1.23 0.98 0.67, 1.44
arteries
Ventricular septal defect, 1,108 1.17 0.87, 1.56 1.02 0.85, 1.23 1.05 0.94, 1.18 1.01 0.90, 1.13 0.93 0.77, 1.13
muscular
Ventricular septal defect, 546 1.11 0.73, 1.69 0.92 0.72, 1.17 1.05 0.89, 1.23 1.00 0.86, 1.17 0.98 0.76, 1.25
perimembranous
Conotruncal defect 661 1.34 0.93, 1.93 1.01 0.73, 1.30 1.00 0.87, 1.15 1.05 0.93, 1.19 0.96 0.79, 1.17
Left ventricular outflow 558 1.37 0.91, 2.06 1.17 0.95, 1.44 0.99 0.84, 1.16 1.02 0.90, 1.17 0.99 0.79, 1.23
tract defect
Right ventricular outflow 421 0.94 0.58, 1.52 0.94 0.72, 1.23 1.00 0.83, 1.20 1.16 0.98, 1.36 0.85 0.65, 1.13
tract defect

Abbreviation: PM10, particulate matter with an average aerodynamic diameter of <10 lm.
a
Risk ratios and 95% confidence intervals correspond to an increase in the interquartile range of the 5-week air pollutant metric (a weighted
average of the 35 daily ambient air pollution levels measured at a central monitor during weeks 3–7 of pregnancy). The interquartile ranges were
29.9 ppb for ozone, 14.2 lg/m3 for PM10, 5.7 ppb for nitrogen dioxide, 0.30 ppm for carbon monoxide, and 4.0 ppb for sulfur dioxide.

We also investigated the effect of limiting analyses to single
gestation pregnancies, limiting analyses to infants with only
1 cardiovascular malformation, using unweighted 5-week
pollution metrics, and using unweighted 9-week pollution
metrics (the average of measurements during the first 63
days of pregnancy).
RESULTS
Surveillance records of pregnancies reaching 20 weeks’
gestation with an estimated date of conception during
January 1, 1986–March 12, 2003, indication of a cardiovas-
cular malformation, and no evidence of trisomy were iden-
tified (n ¼ 7,102). Of these, only 4,639 (65%) were deemed
to have a cardiovascular malformation following review and
classification. Infants with complex malformations involving
abnormal cardiac looping (n ¼ 86) or heterotaxy syndrome
(n ¼ 96) were excluded from analysis, irrespective of the
other specific cardiovascular lesions that might have been
present. From the remaining 4,457 infants, 3,338 (75%) were
included in 1 or more of the outcome groups shown in Table 1.
Overall, there were 715,500 liveborn and stillborn infants with
an estimated date of conception during this period.
The prevalence of cardiovascular malformations, shown
in Table 2 by season and year of conception, suggested some
seasonal variation across outcome groups. Whereas the ob-
served prevalence of the relatively severe lesions (hypoplas-
tic left heart syndrome, transposition of the great arteries,
and tetralogy of Fallot) remained stable over time, the ob-
served prevalence of the less severe lesions (secundum atrial
septal defect, valvar pulmonary stenosis, muscular ventric-
ular septal defect, and perimembranous ventricular septal
defect) increased markedly.
Descriptive statistics for the air pollution metrics, which
are weighted averages of daily ambient pollution measure-
ments during weeks 3–7 of pregnancy, indicated that all
pollutants had seasonal variation (Table 3). Levels of
PM10, nitrogen dioxide, and sulfur dioxide declined over
time. Pairwise Spearman partial correlation coefficients
for the pollution metrics are presented in Table 4.
As shown in Table 5, we observed a statistically signifi-
cant positive association between PM10 and patent ductus
arteriosus (for an interquartile range increase in PM10 levels,
risk ratio (RR) ¼ 1.60, 95% CI: 1.11, 2.31). The 95%
confidence intervals for all other associations, presented in
Table 5, included the null value.
Results from 5 sensitivity analyses are presented in Tables
6–10. We consistently observed positive associations be-
tween PM10 and patent ductus arteriosus (less stringent sea-
sonal and long-term temporal control, RR ¼ 1.40, 95% CI:
1.01, 1.95; limited to single gestation pregnancies, RR ¼
1.57, 95% CI: 1.07, 2.28; limited to infants with only 1
cardiovascular malformation, RR ¼ 1.70, 95% CI: 1.12,
2.56; using an unweighted 5-week metric, RR ¼ 1.53, 95%

Am J Epidemiol 2009;169:1004–1014
1010 Strickland et al.

Table 7. Sensitivity Analysis Limited to Single Gestation Pregnancies, With Risk Ratios and 95% Confidence Intervals for Associations Between
the Weighted 5-Week Air Pollution Metric and Cardiovascular Malformations Among the Cohort of Pregnancies Reaching at Least 20 Weeks’
Gestation in Atlanta, Georgia, With an Estimated Date of Conception During January 1, 1986–March 12, 2003a

8-Hour 24-Hour 24-Hour Nitrogen 24-Hour Carbon 24-Hour Sulfur

Ozone, ppb PM10, mg/m3 Dioxide, ppb Monoxide, ppm Dioxide, ppb
No. of
Cases 95% 95% 95% 95% 95%
Risk Risk Risk Risk Risk
Confidence Confidence Confidence Confidence Confidence
Ratio Ratio Ratio Ratio Ratio
Interval Interval Interval Interval Interval

Atrial septal defect, secundum 347 1.19 0.67, 2.11 1.10 0.79, 1.51 1.12 0.89, 1.41 0.86 0.68, 1.09 1.02 0.73, 1.43
Coarctation of the aorta 254 1.11 0.61, 2.01 1.24 0.89, 1.72 1.14 0.89, 1.46 1.02 0.81, 1.29 1.09 0.78, 1.53
Hypoplastic left heart 166 0.92 0.40, 2.11 0.90 0.60, 1.35 0.89 0.65, 1.23 0.79 0.59, 1.06 0.78 0.50, 1.21
syndrome
Patent ductus arteriosus 215 1.25 0.64, 2.46 1.57 1.07, 2.28 1.23 0.92, 1.65 1.19 0.92, 1.52 1.17 0.81, 1.67
Pulmonary stenosis, valvar 286 0.95 0.51, 1.78 0.87 0.62, 1.23 1.00 0.78, 1.27 1.07 0.85, 1.34 0.73 0.51, 1.05
Tetralogy of Fallot 284 1.04 0.55, 1.96 0.87 0.63, 1.20 0.94 0.73, 1.21 1.15 0.93, 1.44 0.85 0.61, 1.18
Transposition of the great 160 1.15 0.51, 2.60 1.13 0.74, 1.72 0.79 0.56, 1.10 0.91 0.65, 1.27 1.09 0.72, 1.66
arteries
Ventricular septal defect, 1,027 1.08 0.76, 1.51 0.99 0.81, 1.22 1.13 0.99, 1.29 1.00 0.86, 1.16 0.94 0.75, 1.17
muscular
Ventricular septal defect, 514 1.06 0.66, 1.70 0.95 0.72, 1.23 1.11 0.93, 1.33 0.96 0.81, 1.14 0.94 0.72, 1.23
perimembranous
Conotruncal defect 629 1.16 0.76, 1.78 0.98 0.63, 1.20 0.93 0.79, 1.10 1.04 0.89, 1.21 1.03 0.83, 1.28
Left ventricular outflow 523 1.09 0.70, 1.71 1.08 0.85, 1.35 1.00 0.84, 1.20 0.97 0.82, 1.14 1.00 0.79, 1.27
tract defect
Right ventricular outflow 389 0.68 0.40, 1.16 0.83 0.62, 1.12 1.00 0.81, 1.23 1.16 0.96, 1.41 0.76 0.56, 1.03
tract defect

Abbreviation: PM10, particulate matter with an average aerodynamic diameter of <10 lm.
a
Risk ratios and 95% confidence intervals correspond to an increase in the interquartile range of the 5-week air pollutant metric (a weighted
average of the 35 daily ambient air pollution levels measured at a central monitor during weeks 3–7 of pregnancy). The interquartile ranges were
29.9 ppb for ozone, 14.2 lg/m3 for PM10, 5.7 ppb for nitrogen dioxide, 0.30 ppm for carbon monoxide, and 4.0 ppb for sulfur dioxide.

CI: 1.02, 2.29; using an unweighted 9-week metric, RR ¼
1.71, 95% CI: 0.98, 3.00). We observed a positive, statisti-
cally significant association between nitrogen dioxide and
patent ductus arteriosus in the analysis with less stringent
seasonal and long-term temporal control (RR ¼ 1.40, 95%
CI: 1.07, 1.83). Using unweighted 9-week pollution metrics,
we observed positive, statistically significant associations
between nitrogen dioxide and both secundum atrial septal
defect (RR ¼ 1.58, 95% CI: 1.02, 2.29) and muscular ven-
tricular septal defect (RR ¼ 1.20, 95% CI: 1.02, 1.41). We
also observed negative, statistically significant associations
between ozone and right ventricular outflow tract defects in
the analysis limited to infants with only 1 cardiovascular
malformation (RR ¼ 0.52, 95% CI: 0.29, 0.93) and between
sulfur dioxide and right ventricular outflow tract defects in
the analysis based on an unweighted 5-week metric (RR ¼
0.73, 95% CI: 0.54, 0.98).
DISCUSSION
We investigated temporal associations between ambient air
pollution levels during weeks 3–7 of pregnancy and risks of
cardiovascular malformations. Except for the association be-
tween PM10 and patent ductus arteriosus, all 95% confidence
intervals from the primary analysis were consistent with little
or no association. Many confidence intervals, particularly
those for ozone, were wide and were therefore compatible
with both no effect and a harmful effect of air pollution.
Patent ductus arteriosus was not investigated in the 2 pre-
vious studies (7, 8), likely because identification of infants
with congenital patent ductus arteriosus is difficult. In our
study population, we used restrictive criteria to exclude patent
ductus arteriosus in premature or newborn infants and when it
occurred as an obligate shunt lesion in the presence of other
cardiovascular malformations. Of the 2,273 surveillance rec-
ords we reviewed that contained a code for patent ductus
arteriosus, the records for only 219 infants met these criteria.
We are unaware of experimental animal evidence support-
ing an association between PM10 and patent ductus arterio-
sus. Analogous epidemiologic evidence comes from
Swedish registry data, wherein a positive association was
observed between first trimester maternal smoking and risk
of patent ductus arteriosus among term infants (21). One
plausible biologic mechanism for our result relates to fetal
growth and development. Relative to term infants, premature
infants have a much higher incidence of patent ductus (22,
23). Associations between reduced fetal growth and mater-
nal smoking (24), environmental tobacco smoke (25), and
ambient particulate matter (26, 27) have been observed in
epidemiologic studies. If high PM10 levels in utero restrict
fetal development, this could explain an association between
PM10 and patent ductus arteriosus among term infants.

Am J Epidemiol 2009;169:1004–1014
 Ambient Air Pollution and Cardiovascular Malformations 1011

Table 8. Sensitivity Analysis Limited to Infants and Fetuses With Only 1 Cardiovascular Malformation, With Risk Ratios and 95% Confidence
Intervals for Associations Between the Weighted 5-Week Air Pollution Metric and Cardiovascular Malformations Among the Cohort of Pregnancies
Reaching at Least 20 Weeks’ Gestation in Atlanta, Georgia, With an Estimated Date of Conception During January 1, 1986–March 12, 2003a

8-Hour 24-Hour 24-Hour Nitrogen 24-Hour Carbon 24-Hour Sulfur

Ozone, ppb PM10, mg/m3 Dioxide, ppb Monoxide, ppm Dioxide, ppb
No. of
Cases 95% 95% 95% 95% 95%
Risk Risk Risk Risk Risk
Confidence Confidence Confidence Confidence Confidence
Ratio Ratio Ratio Ratio Ratio
Interval Interval Interval Interval Interval

Atrial septal defect, secundum 202 0.80 0.36, 1.73 1.03 0.67, 1.57 1.04 0.77, 1.41 1.08 0.80, 1.46 1.16 0.75, 1.79
Coarctation of the aorta 145 1.09 0.51, 2.32 1.16 0.75, 1.79 1.08 0.78, 1.50 1.04 0.76, 1.43 0.89 0.56, 1.39
Hypoplastic left heart 167 0.80 0.35, 1.82 0.90 0.61, 1.34 0.91 0.66, 1.24 0.83 0.62, 1.10 0.82 0.53, 1.28
syndrome
Patent ductus arteriosus 171 1.20 0.56, 2.59 1.70 1.12, 2.56 1.31 0.95, 1.81 1.25 0.95, 1.64 1.37 0.92, 2.04
Pulmonary stenosis, valvar 225 0.76 0.37, 1.56 0.88 0.61, 1.27 1.01 0.77, 1.33 1.13 0.88, 1.45 0.74 0.49, 1.11
Tetralogy of Fallot 279 0.96 0.50, 1.81 0.85 0.62, 1.18 0.89 0.69, 1.15 1.17 0.94, 1.46 0.87 0.62, 1.22
Transposition of the great 140 1.28 0.54, 3.03 1.12 0.71, 1.78 0.79 0.55, 1.13 0.93 0.65, 1.31 1.19 0.76, 1.87
arteries
Ventricular septal defect, 976 0.90 0.64, 1.28 0.98 0.79, 1.20 1.06 0.92, 1.21 1.00 0.85, 1.16 0.91 0.73, 1.15
muscular
Ventricular septal defect, 388 1.22 0.71, 2.11 0.93 0.68, 1.26 1.19 0.97, 1.46 0.97 0.79, 1.18 1.08 0.80, 1.47
perimembranous
Conotruncal defect 571 1.12 0.72, 1.76 0.96 0.76, 1.21 0.91 0.76, 1.09 1.07 0.91, 1.27 1.05 0.83, 1.32
Left ventricular outflow 406 1.00 0.61, 1.66 0.99 0.76, 1.28 0.94 0.77, 1.14 0.97 0.80, 1.16 0.89 0.67, 1.18
tract defect
Right ventricular outflow 331 0.52 0.29, 0.93 0.81 0.59, 1.10 1.02 0.81, 1.28 1.23 1.00, 1.51 0.77 0.56, 1.07
tract defect

Abbreviation: PM10, particulate matter with an average aerodynamic diameter of <10 lm.
a
Risk ratios and 95% confidence intervals correspond to an increase in the interquartile range of the 5-week air pollutant metric (a weighted
average of the 35 daily ambient air pollution levels measured at a central monitor during weeks 3–7 of pregnancy). The interquartile ranges were
29.9 ppb for ozone, 14.2 lg/m3 for PM10, 5.7 ppb for nitrogen dioxide, 0.30 ppm for carbon monoxide, and 4.0 ppb for sulfur dioxide.

Similar to our study, most results reported in the 2 previous
studies were consistent with no association (7, 8). In southern
California, investigators reported 4 significant positive associ-
ations from 144 models. Twenty-four of these models aver-
aged pollution levels over weeks 5–8 of pregnancy; the 4
positive associations were observed during this window (7).
The Texas study investigators observed 3 positive associations
from 75 models; air pollution levels were averaged during
weeks 3–8 of pregnancy (8). No significant association from
the southern California study was replicated in the Texas study.
Because of our review and classification of surveillance
records, comparison of results from our study with those
from previous studies was difficult. Through our review,
we excluded infants with structurally normal hearts and tran-
sient newborn conditions; 35% of the records we reviewed
were reclassified as ‘‘structurally normal.’’ Outcome groups
were based on embryologic considerations, and infants with
multiple congenital heart defect codes were included in mul-
tiple outcome groups only when the malformations were
thought to be embryologically independent. Consequently,
our outcome groups differed from those used in previous
studies (7, 8). For example, in the southern California study
(7), an association was observed between carbon monoxide
and isolated ventricular septal defects (fourth quartile vs.
first quartile, odds ratio ¼ 2.95, 95% CI: 1.44, 6.05). This
outcome group incorporated the 4 major types of ventricular
septal defects, as well as pulmonary atresia with ventricular
septal defect. In our study, we distinguished among the 4
types of ventricular septal defects, because each is thought to
develop through unique embryologic mechanisms (28, 29).
We analyzed perimembranous and muscular ventricular sep-
tal defects as distinct outcome groups, and we included sub-
arterial ventricular septal defects in the conotruncal defects
outcome group (Table 1). There were too few inlet ventricular
septal defects to permit analysis. We grouped pulmonary
atresia with ventricular septal defect, which is the extreme
end of the anatomic spectrum of tetralogy of Fallot (30), in the
tetralogy of Fallot outcome group. To provide a more direct
comparison with the southern California result, we performed
a secondary analysis in which we ignored our classifications
and defined isolated ventricular septal defect according to the
preexisting codes in the MACDP database. Using our primary
analytical approach, we observed no evidence for an associ-
ation between carbon monoxide and isolated ventricular sep-
tal defects (RR ¼ 0.98, 95% CI: 0.82, 1.16).
Our analytical approach and exposure assignment meth-
ods were also different. Whereas the previous study inves-
tigators conducted spatial-temporal analyses using pollution
measurements assigned to women on the basis of residential
location, we opted for a temporal approach using measure-
ments from centrally located monitors. We implemented
a temporal analysis because of our desire to preclude con-
cerns about our study results based on arguments of spatial
confounding. We characterized pollution levels using

Am J Epidemiol 2009;169:1004–1014
1012 Strickland et al.

Table 9. Sensitivity Analysis Based on Unweighted 5-Week Pollution Metrics, With Risk Ratios and 95% Confidence Intervals for Associations
Between Air Pollution and Cardiovascular Malformations Among the Cohort of Pregnancies Reaching at Least 20 Weeks’ Gestation in Atlanta,
Georgia, With an Estimated Date of Conception During January 1, 1986–March 12, 2003a

8-Hour 24-Hour 24-Hour Nitrogen 24-Hour Carbon 24-Hour Sulfur

Ozone, ppb PM10, mg/m3 Dioxide, ppb Monoxide, ppm Dioxide, ppb
No. of
Cases 95% 95% 95% 95% 95%
Risk Risk Risk Risk Risk
Confidence Confidence Confidence Confidence Confidence
Ratio Ratio Ratio Ratio Ratio
Interval Interval Interval Interval Interval

Atrial septal defect, secundum 379 1.09 0.62, 1.90 1.09 0.77, 1.54 1.14 0.92, 1.42 0.93 0.75, 1.16 1.00 0.72, 1.38
Coarctation of the aorta 275 1.18 0.66, 2.12 1.19 0.84, 1.68 1.11 0.87, 1.41 0.98 0.79, 1.22 1.04 0.75, 1.44
Hypoplastic left heart 175 0.78 0.34, 1.77 0.85 0.55, 1.31 0.89 0.65, 1.21 0.82 0.63, 1.08 0.76 0.49, 1.18
syndrome
Patent ductus arteriosus 219 1.35 0.69, 2.64 1.53 1.02, 2.29 1.26 0.94, 1.69 1.18 0.93, 1.50 1.19 0.84, 1.70
Pulmonary stenosis, valvar 312 0.92 0.50, 1.68 0.83 0.57, 1.19 0.99 0.78, 1.25 1.03 0.84, 1.28 0.71 0.50, 1.01
Tetralogy of Fallot 299 1.08 0.58, 2.01 0.89 0.63, 1.26 0.94 0.73, 1.21 1.15 0.93, 1.41 0.81 0.59, 1.13
Transposition of the great 165 1.26 0.57, 2.82 1.20 0.76, 1.91 0.80 0.57, 1.11 0.99 0.73, 1.35 1.13 0.75, 1.71
arteries
Ventricular septal defect, 1,108 1.08 0.77, 1.50 1.05 0.84, 1.30 1.08 0.95, 1.23 0.99 0.86, 1.14 0.95 0.77, 1.17
muscular
Ventricular septal defect, 546 0.99 0.62, 1.57 0.93 0.70, 1.24 1.11 0.93, 1.32 0.93 0.78, 1.09 0.98 0.76, 1.26
perimembranous
Conotruncal defect 661 1.21 0.79, 1.84 1.01 0.80, 1.28 0.95 0.81, 1.12 1.07 0.93, 1.24 1.05 0.85, 1.30
Left ventricular outflow 558 1.07 0.69, 1.67 1.04 0.81, 1.33 1.00 0.84, 1.19 0.98 0.84, 1.14 0.96 0.76, 1.22
tract defect
Right ventricular outflow 421 0.71 0.43, 1.19 0.80 0.58, 1.09 1.01 0.83, 1.24 1.16 0.97, 1.39 0.73 0.54, 0.98
tract defect

Abbreviation: PM10, particulate matter with an average aerodynamic diameter of <10 lm.
a
Risk ratios and 95% confidence intervals correspond to an increase in the interquartile range of the 5-week air pollutant metric (an unweighted
average of the 35 daily ambient air pollution levels measured at a central monitor during weeks 3–7 of pregnancy). The interquartile ranges were
29.7 ppb for ozone, 15.4 lg/m3 for PM10, 5.7 ppb for nitrogen dioxide, 0.30 ppm for carbon monoxide, and 4.0 ppb for sulfur dioxide.

a centrally located monitor because of practical limitations.
Namely, geocoded data were available dating back only to
1994, whereas the central monitor approach allowed us to
utilize data back to 1986. Throughout much of follow-up,
there were only 2 or 3 monitors for each primary pollutant,
and we determined that some monitors were unduly im-
pacted by local sources. We were hesitant to use these mea-
surements, which likely better reflect local conditions than
population exposure. The use of 1 central monitor for each
pollutant also provided some consistency throughout
follow-up; otherwise, modeling assumptions would have
been needed to account for changes in the number and lo-
cation of monitoring stations. For the PM10 analyses, we
relied on such assumptions, and cubic spline knots were
placed on the dates when the monitor location and measure-
ment method changed. To examine the effect of these as-
sumptions, we stratified the data according to measurement
method (federal reference method: 1986–1997; tapered el-
ement oscillating microbalance: 1998–2003). For PM10 and
patent ductus arteriosus, we observed a strong association
during 1986–1997 (RR ¼ 1.89, 95% CI: 1.26, 2.84) and no
association during 1998–2003 (RR ¼ 0.78, 95% CI: 0.28,
2.04). Perhaps the strong association observed during 1986–
1997 was attributable to average PM10 levels being 42%
higher during this period of follow-up (Table 3).
A limitation of our study, also a limitation in previous
studies (7, 8), was that results were based on the cohort of
pregnancies reaching at least 20 weeks’ gestation. Given that
our gestational window of interest spanned weeks 3–7 of
pregnancy, we would have preferred our cohort to consist of
all pregnancies at week 3. The consequence of this limitation
might depend on the causal effect of air pollution on cardio-
vascular malformations. For example, atrioventricular septal
defect, Ebstein’s anomaly, and tricuspid valve dysplasia can
all cause intrauterine congestive heart failure, increasing the
risk of intrauterine fetal death (31). If air pollution were to
increase the risk of these malformations, in turn increasing the
risk of fetal loss before week 20, then our study would have
been unable to detect this harmful effect of pollution.
Measurement error or its discrete counterpart, misclassifi-
cation, which was present in the air quality, vital records, and
surveillance data, was another limitation in our study as in
previous studies (7, 8). Our use of ambient air pollution mea-
surements from stationary monitors as proxies for personal
exposure was likely the largest component of measurement
error. If this measurement error was nondifferential, a bias to-
ward the null could explain some null results (32). If the mea-
surement error was differential, for example, if the magnitude
of measurement error varied according to meteorologic con-
ditions, then risk ratio estimates may have been biased toward,
away from, or across the null. We did not observe broad pat-
terns of consistently positive or negative risk ratios suggesting
the possibility of differential measurement error, although dif-
ferential measurement error may have biased some results.

Am J Epidemiol 2009;169:1004–1014
 Ambient Air Pollution and Cardiovascular Malformations 1013

Table 10. Sensitivity Analysis Based on Unweighted 9-Week Pollution Metrics, With Risk Ratios and 95% Confidence Intervals for Associations
Between Air Pollution and Cardiovascular Malformations Among the Cohort of Pregnancies Reaching at Least 20 Weeks’ Gestation in Atlanta,
Georgia, With an Estimated Date of Conception During January 1, 1986–March 12, 2003a

8-Hour 24-Hour 24-Hour Nitrogen 24-Hour Carbon 24-Hour Sulfur

Ozone, ppb PM10, mg/m3 Dioxide, ppb Monoxide, ppm Dioxide, ppb
No. of
Cases 95% 95% 95% 95% 95%
Risk Risk Risk Risk Risk
Confidence Confidence Confidence Confidence Confidence
Ratio Ratio Ratio Ratio Ratio
Interval Interval Interval Interval Interval

Atrial septal defect, secundum 379 1.36 0.70, 2.65 1.19 0.73, 1.94 1.58 1.19, 2.09 1.08 0.85, 1.37 1.04 0.70, 1.56
Coarctation of the aorta 275 1.29 0.64, 2.57 1.32 0.82, 2.12 1.19 0.88, 1.61 0.93 0.73, 1.20 1.00 0.65, 1.52
Hypoplastic left heart 175 1.16 0.44, 3.03 1.03 0.58, 1.83 1.00 0.68, 1.47 0.80 0.59, 1.09 0.80 0.47, 1.36
syndrome
Patent ductus arteriosus 219 1.38 0.61, 3.10 1.71 0.98, 3.00 1.21 0.84, 1.74 1.09 0.83, 1.43 1.20 0.76, 1.89
Pulmonary stenosis, valvar 312 1.01 0.50, 2.04 0.99 0.59, 1.65 1.04 0.78, 1.39 1.07 0.84, 1.37 0.73 0.47, 1.14
Tetralogy of Fallot 299 1.03 0.50, 2.11 1.01 0.64, 1.59 0.98 0.72, 1.33 1.22 0.97, 1.54 0.78 0.51, 1.17
Transposition of the great 165 1.88 0.72, 4.94 1.68 0.90, 3.17 0.77 0.51, 1.15 0.93 0.65, 1.32 1.24 0.74, 2.09
arteries
Ventricular septal defect, 1,108 1.11 0.74, 1.65 1.18 0.88, 1.59 1.20 1.02, 1.41 0.98 0.83, 1.15 1.00 0.76, 1.30
muscular
Ventricular septal defect, 546 1.00 0.58, 1.73 0.82 0.55, 1.21 1.12 0.90, 1.40 0.94 0.78, 1.13 0.88 0.64, 1.22
perimembranous
Conotruncal defect 661 1.22 0.75, 2.00 1.19 0.87, 1.64 0.97 0.79, 1.19 1.10 0.93, 1.30 1.16 0.89, 1.52
Left ventricular outflow 558 1.19 0.71, 1.99 1.21 0.87, 1.70 1.06 0.85, 1.31 0.94 0.79, 1.12 0.98 0.73, 1.33
tract defect
Right ventricular outflow 421 0.88 0.48, 1.61 1.00 0.65, 1.53 1.02 0.80, 1.31 1.17 0.95, 1.45 0.82 0.57, 1.18
tract defect

Abbreviation: PM10, particulate matter with an average aerodynamic diameter of <10 lm.
a
Risk ratios and 95% confidence intervals correspond to an increase in the interquartile range of the 9-week air pollutant metric (an unweighted
average of the 63 daily ambient air pollution levels measured at a central monitor during weeks 1–9 of pregnancy). The interquartile ranges were
27.4 ppb for ozone, 15.1 lg/m3 for PM10, 5.3 ppb for nitrogen dioxide, 0.28 ppm for carbon monoxide, and 3.9 ppb for sulfur dioxide.

Misclassification in the vital records data was evident by
the strong day-of-month pattern observed in the date-of-
conception estimates. Although a statistical model was cre-
ated to remove this day-of-month pattern, uncertainty in the
estimates remained. The date-of-conception estimates from
MACDP surveillance records, although lacking an obvious
day-of-month pattern, likewise had uncertainty. Although
we selected our exposure window to coincide with the pe-
riod of cardiac morphogenesis, it is possible that exposures
earlier or later in pregnancy could have affected the devel-
opment of certain malformations (33). Some results were
sensitive to the window definition; for pollution levels dur-
ing the first 9 weeks of pregnancy, we observed positive,
statistically significant associations between nitrogen diox-
ide and both secundum atrial septal defect and muscular
ventricular septal defect. Although the point estimates for
these 2 associations were elevated in our primary analysis,
neither was statistically significant (Table 5).
Ours is the third epidemiologic study of air pollution
and cardiovascular malformations reported to date (7, 8).
Results from these 3 studies do not seem consistent. This
inconsistency could be due to the absence of true associa-
tions between ambient air pollution and risks of cardiovas-
cular malformations; it also could be due to differences in
populations, pollution levels, outcome definitions, or analyt-
ical approaches. If ambient air pollution levels do cause
cardiovascular malformations, then the lack of consistency
of results might be due to issues relating to statistical power
and measurement error. Although our study population was
large, we nevertheless would have struggled to detect very
small increases in risk.
ACKNOWLEDGMENTS
Author affiliations: Department of Environmental and
Occupational Health, Rollins School of Public Health,
Emory University, Atlanta, Georgia (Matthew J. Strickland,
Mitchel Klein, Paige E. Tolbert); Department of Epidemi-
ology, Rollins School of Public Health, Emory University,
Atlanta, Georgia (Matthew J. Strickland, W. Dana Flanders,
Michele M. Marcus); National Center on Birth Defects and
Developmental Disabilities, Centers for Disease Control
and Prevention, Atlanta, Georgia (Matthew J. Strickland,
Adolfo Correa, Tiffany Riehle-Colarusso, Csaba Siffel);
Oregon Health and Science University, Portland, Oregon
(Mark D. Reller); Children’s Healthcare of Atlanta, Emory
University, Atlanta, Georgia (William T. Mahle); Depart-
ment of Pediatrics, University of Utah, Salt Lake City, Utah
(Lorenzo D. Botto); School of Civil and Environmental En-
gineering, Georgia Institute of Technology, Atlanta, Georgia
(James A. Mulholland); and Computer Sciences Corpora-
tion, Atlanta, Georgia (Csaba Siffel).

Am J Epidemiol 2009;169:1004–1014
1014 Strickland et al.
Supported by the National Institute of Environmental
Health Sciences (R01-ES012967-01A1), the Health Resour-
ces and Services Administration (T03MC07651), and Envi-
ronmental Public Health Tracking at the Centers for Disease
Control and Prevention.
The findings and conclusions in this report are those of
the author(s) and do not necessarily represent the official
position of the Centers for Disease Control and Prevention.
Conflict of interest: none declared.
REFERENCES
1. Glinianaia SV, Rankin J, Bell R, et al. Particulate air pollution
and fetal health: a systematic review of the epidemiologic
evidence. Epidemiology. 2004;15(1):36–45.
2. Maisonet M, Correa A, Misra D, et al. A review of the liter-
ature on the effects of ambient air pollution on fetal growth.
Environ Res. 2004;95(1):106–115.
3. Sram RJ, Binkova B, Dejmek J, et al. Ambient air pollution
and pregnancy outcomes: a review of the literature. Environ
Health Perspect. 2005;113(4):375–382.
4. Lacasana M, Esplugues A, Ballester F. Exposure to ambient
air pollution and prenatal and early childhood health effects.
Eur J Epidemiol. 2005;20(2):183–199.
5. Perera FP, Rauh V, Whyatt RM, et al. A summary of recent
findings on birth outcomes and developmental effects of pre-
natal ETS, PAH, and pesticide exposures. Neurotoxicology.
2005;26(4):573–587.
6. Choi H, Jedrychowski W, Spengler J, et al. International studies of
prenatal exposure to polycyclic aromatic hydrocarbons and fetal
growth. Environ Health Perspect. 2006;114(11):1744–1750.
7. Ritz B, Yu F, Fruin S, et al. Ambient air pollution and risk of
birth defects in southern California. Am J Epidemiol. 2002;
155(1):17–25.
8. Gilboa SM, Mendola P, Olshan AF, et al. Relation between
ambient air quality and selected birth defects, seven county study,
Texas, 1997–2000. Am J Epidemiol. 2005;162(3):238–252.
9. Mavroudis C, Jacobs JP. Congenital Heart Surgery Nomen-
clature and Database Project: overview and minimum dataset.
Ann Thorac Surg. 2000;69(4 suppl):S2–S17.
10. Beland MJ, Franklin RC, Jacobs JP, et al. Update from the
International Working Group for Mapping and Coding of
Nomenclatures for Paediatric and Congenital Heart Disease.
Cardiol Young. 2004;14(2):225–229.
11. Jacobs JP, Mavroudis C, Jacobs ML, et al. Lessons learned
from the data analysis of the second harvest 1998–2001 of the
Society of Thoracic Surgeons (STS) Congenital Heart Surgery
Database. Eur J Cardiothorac Surg. 2004;26(1):18–37.
12. Riehle-Colarusso T, Strickland MJ, Reller MD, et al. Improv-
ing the quality of surveillance data on congenital heart defects
in the Metropolitan Atlanta Congenital Defects Program. Birth
Defects Res A Clin Mol Teratol. 2007;79(11):743–753.
13. The International Working Group for Mapping and Coding of
Nomenclatures for Pediatric and Congenital Heart Disease.
International pediatric and congenital cardiac code. Montreal,
Quebec, Canada: International Society for Nomenclature of
Congenital and Paediatric Heart Disease; 2004. (http://
www.ipccc.net).
14. Correa A, Cragan JD, Kucik JE, et al. Reporting birth defects
surveillance data 1968–2003. Birth Defects Res A Clin Mol
Teratol. 2007;79(2):65–186.
15. Society of Thoracic Surgeons. STS Congenital Heart Surgery
Database, version 2.30. Chicago, IL: Society of Thoracic Sur-
geons; 2004. (http://www.sts.org/sections/stsnationaldatabase/
datamanagers/congenitalheartsurgerydb/datacollection/index.
html).
16. Centers for Disease Control and Prevention. Metropolitan
Atlanta Congenital Defects Program defect code list. Atlanta,
GA: Centers for Disease Control and Prevention; 2004. (http://
www.cdc.gov/ncbddd/bd/macdp_resources.htm).
17. Jacobs JP, Mavroudis C, Jacobs ML, et al. Nomenclature and
databases—the past, the present, and the future: a primer for
the congenital heart surgeon. Pediatr Cardiol. 2007;28(2):
105–115.
18. National Center on Birth Defects and Developmental Dis-
abilities, Centers for Disease Control and Prevention. State
birth defects surveillance program directory. Birth Defects Res
A Clin Mol Teratol. 2006;76(12):837–893.
19. Sadler TW. Langman’s Medical Embryology. Philadelphia,
PA: Lippincott Williams & Wilkins; 2006.
20. R Development Core Team. R: A Language and Environment
for Statistical Computing v 2.5.0. Vienna, Austria: R Foun-
dation for Statistical Computing; 2007.
21. Kallen K. Maternal smoking and congenital heart defects. Eur
J Epidemiol. 1999;15(8):731–737.
22. Schneider DJ, Moore JW. Patent ductus arteriosus. Circula-
tion. 2006;114(17):1873–1882.
23. Kitterman JA, Edmunds LH Jr, Gregory GA, et al. Patent
ducts arteriosus in premature infants. Incidence, relation to
pulmonary disease and management. N Engl J Med. 1972;
287(10):473–477.
24. US Department of Health and Human Services. Women
and Smoking: A Report of the Surgeon General. Atlanta,
GA: Centers for Disease Control and Prevention; 2001.
25. US Department of Health and Human Services. The
Health Consequences of Involuntary Exposure to Tobacco
Smoke: A Report of the Surgeon General. Atlanta, GA:
Centers for Disease Control and Prevention; 2006.
26. Wilhelm M, Ritz B. Local variations in CO and particulate air
pollution and adverse birth outcomes in Los Angeles County,
California, USA. Environ Health Perspect. 2005;113(9):
1212–1221.
27. Parker JD, Woodruff TJ, Basu R, et al. Air pollution and birth
weight among term infants in California. Pediatrics. 2005;
115(1):121–128.
28. Clark EB. Pathogenetic mechanisms of congenital cardiovas-
cular malformations revisited. Semin Perinatol. 1996;20(6):
465–472.
29. Jacobs JP, Burke RP, Quintessenza JA, et al. Congenital
Heart Surgery Nomenclature and Database Project: ventricular
septal defect. Ann Thorac Surg. 2000;69(4 suppl):S25–S35.
30. Jacobs ML. Congenital Heart Surgery Nomenclature and
Database Project: tetralogy of Fallot. Ann Thorac Surg. 2000;
69(4 suppl):S77–S82.
31. Hoffman JIE. Epidemiology of congenital heart disease: eti-
ology, pathogenesis and incidence. In: Yagel S, Silverman NH,
Gembruch U, eds. Fetal Cardiology: Embryology, Genetics,
Physiology, Echocardiographic Evaluation, Diagnosis and
Perinatal Management of Cardiac Diseases. London, United
Kingdom: Martin Dunitz; 2003:79–88.
32. Zeger SL, Thomas D, Dominici F, et al. Exposure measurement
error in time-series studies of air pollution: concepts and con-
sequences. Environ Health Perspect. 2000;108(5):419–426.
33. Trines J, Hornberger LK. Evolution of heart disease in utero.
Pediatr Cardiol. 2004;25(3):287–298.
Am J Epidemiol 2009;169:1004–1014
 American Journal of Epidemiology Vol. 169, No. 8
ª The Author 2009. Published by the Johns Hopkins Bloomberg School of Public Health. DOI: 10.1093/aje/kwp001
All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org. Advance Access publication February 27, 2009

Original Contribution

Maternal Urinary Metabolites of Di-(2-Ethylhexyl) Phthalate in Relation to the

Timing of Labor in a US Multicenter Pregnancy Cohort Study

Jennifer J. Adibi, Russ Hauser, Paige L. Williams, Robin M. Whyatt, Antonia M. Calafat,
Heather Nelson, Robert Herrick, and Shanna H. Swan

Initially submitted March 10, 2008; accepted for publication January 6, 2009.

Di-(2-ethylhexyl) phthalate (DEHP) is a plasticizer used in consumer and medical products that can cross the
placenta, disrupt steroid hormone synthesis, and activate peroxisome proliferator-activated receptor c. The authors
examined DEHP exposure in relation to the timing of labor in a pregnancy cohort study of 283 women recruited in 4
US states (California, Iowa, Minnesota, and Missouri) between 2000 and 2004. The authors estimated associations
between concentrations of DEHP metabolites and gestational age at delivery using linear regression models and
associations between DEHP metabolites and clinical outcomes using logistic regression models. After covariate
adjustment, women at the 75th percentile of DEHP metabolite concentrations had a 2-day-longer mean length of
gestation than women at the 25th percentile (95% confidence interval: 1.4, 3.3). Log-unit increases in mono-
2-ethylhexyl phthalate and mono-2-ethyl-5-oxohexyl phthalate concentrations were associated with increased
odds of cesarean section delivery (30% and 50% increased odds, respectively), increased odds of delivering at
41 weeks or later (100% and 120% increased odds), and reduced odds of preterm delivery (50% and 60%
decreased odds). These data suggest that DEHP may interfere with signaling related to the timing of parturition.

creatinine; diethylhexyl phthalate; endocrine disruptors; gestational age; parturition; placenta; PPAR c; pregnancy

Abbreviations: CI, confidence interval; DEHP, di-(2-ethylhexyl) phthalate; MEHHP, mono-2-ethylhexyl phthalate; MEHP, mono-
2-ethylhexyl phthalate; MEOHP, mono-2-ethyl-5-oxohexyl phthalate; NHANES, National Health and Nutrition Examination
Survey; PPARc, peroxisome proliferator-activated receptor c; SFF, Study for Future Families.

Di-(2-ethylhexyl) phthalate (DEHP) is a plasticizer that is
widely used in consumer products, such as polyvinyl chlo-
ride flooring, carpeting, roofing, vinyl, upholstery, clothing,
and packaging (1). DEHP exposure occurs through inges-
tion of food and water (2, 3) and inhalation of household
dust (4), as well as parenterally from medical devices (1).
DEHP metabolites have been detected in 95% of the US
population aged
6 years (5).
The hydrolytic metabolite of DEHP, mono-(2-ethylhexyl)
phthalate (MEHP), can cross the placenta and enter fetal
circulation (6, 7). Metabolites of DEHP have been measured
in umbilical cord blood (8), amniotic fluid (9), maternal
urine (10), placental tissue (7, 11), and neonatal urine and
meconium (12, 13). MEHP is further metabolized into oxi-
dative metabolites, including mono-2-ethyl-5-hydroxyhexyl
phthalate (MEHHP) and mono-2-ethyl-5-oxohexyl phthal-
ate (MEOHP). MEHP and its oxidative metabolites can
be glucuronidated and excreted in urine and feces. These
3 metabolites account for approximately 50% of the received
dose of DEHP (14).
Among rodents, DEHP suppresses fetal testosterone syn-
thesis in males (15, 16) and inhibits ovarian aromatase tran-
scription in adult females (17, 18). Aromatase, which is
responsible for the conversion of androgens to estradiol, is
expressed in several tissues, including the human placenta
(19). Inhibition of steroid hormone synthesis in the ovary of
the adult rodent, fetal testis, and fetal brain has been linked
to binding and activation of the transcription factor peroxi-
some proliferator-activated receptor c (PPARc) (20–22).
MEHP can activate PPARc in primate fibroblasts, in rodent

Correspondence to Dr. Jennifer J. Adibi, Center for Reproductive Sciences, Department of Obstetrics/Gynecology and Reproductive Sciences,
University of California, San Francisco, 513 Parnassus Avenue, Box 0556, San Francisco, CA 94143 (e-mail: adibij@obgyn.ucsf.edu).

1015 Am J Epidemiol 2009;169:1015–1024

1016 Adibi et al.
granulosa and trophoblast cells, and in human cervical can-
cer cells (21–24).
PPARc activation increases early in pregnancy to pro-
mote zygote implantation and placental development
(25–27). During pregnancy, high levels of placental PPARc
contribute to uterine quiescence by down-regulating cyclo-
oxygenase 2 and inhibiting prostaglandin synthesis. By con-
trast, at parturition, PPARc activation decreases in order to
allow for increased prostaglandin production, which is es-
sential in the stimulation of uterine contractions (26, 28). It
has been hypothesized that exposure to PPARc ligands dur-
ing pregnancy reduces the risk of preterm labor by suppress-
ing the inflammatory response in fetal tissues (27).
We tested the hypothesis that exposure to DEHP alters the
timing of labor by measuring the associations between ma-
ternal urinary concentrations of DEHP metabolites during
pregnancy and gestational age at delivery, as well as be-
tween these metabolites and risk of cesarean delivery,
preterm birth, and delivering at greater than 41 weeks’
gestation.
MATERIALS AND METHODS
Women included in this analysis were recruited into the
Study for Future Families (SFF) between March 2000 and
August 2004 at 4 US sites. Details on the SFF are available
elsewhere (29). Briefly, women were recruited at prenatal
clinics associated with university hospitals in Los Angeles,
California (Harbor-UCLA and Cedars-Sinai); Minneapolis,
Minnesota (University of Minnesota Health Center); Co-
lumbia, Missouri (University of Missouri School of Medi-
cine); and Iowa City, Iowa (University of Iowa School of
Medicine). Women were eligible if they were at least 18
years of age, conception had been natural, and the preg-
nancy was not medically threatened.
Of the 783 pregnant women enrolled in the SFF between
2000 and 2004, only those who were fully enrolled in
a follow-up study for postnatal evaluation of their babies
were eligible for inclusion in the phthalate study (n ¼ 441).
For a woman to be eligible for follow-up, the pregnancy
had to end in a livebirth, the baby had to be 2–36 months of
age, the mother and baby had to live within 50 miles (80 km)
of the clinic, and the mother had to attend at least 1 study
visit. Reasons for exclusion were unlikely to be related
to either phthalate exposure or gestational age at delivery
(except possibly a nonviable birth outcome, which was ex-
tremely rare). The current analysis (n ¼ 283) was restricted
to mothers for whom we had urinary phthalate concentra-
tions (n ¼ 304) and complete medical record data (n ¼ 298)
after excluding twin births (12 babies), 2 babies with miss-
ing data, and 1 baby born at 30 weeks with eclampsia.
Women provided 1 urine sample at the time of recruitment,
which was on average 12.2 weeks (standard deviation,
7.6 weeks) before delivery or the beginning of the third
trimester.
Human subject committees at each of the participating
centers approved study procedures, and all participants gave
signed informed consent. Data were stripped of identifying
information before analysis.
Gestational age at delivery was calculated in 2 ways. First,
the date of the last menstrual period as reported by the woman
at study entry was subtracted from the date of delivery to
obtain an estimate in days. Second, the clinical estimate re-
corded by the obstetrician was abstracted directly from the
birth record. The clinical estimate was based on ultrasound
data, examination of the newborn, and dates reported by
the mother and was rounded up to the next-highest
week. There was a high correlation between the 2 estimates
(r ¼ 0.92, n ¼ 283) after we substituted the clinical estimate
for cases for which there was a discrepancy of 14 or more days
(30). We used the clinical estimate in the current analysis.
Urinary phthalate metabolite concentrations were mea-
sured using analytical methods described in detail else-
where (31). Briefly, the phthalate metabolites were first
enzymatically deconjugated and then extracted from the
urine using automated on-line solid phase extraction, sep-
arated by high-performance liquid chromatography, and
detected by isotope-dilution tandem mass spectrometry.
Each analytical run included calibration standards, reagent
blanks, and quality control materials of high and low con-
centration to monitor for accuracy and precision. The lim-
its of detection were 0.98 ng/mL (MEHP), 0.95 ng/mL
(MEHHP), and 1.07 ng/mL (MEOHP).
Because of the high correlation and pharmacokinetic sim-
ilarities between MEOHP and MEHHP (r ¼ 0.99), only
estimates for MEOHP are described below.
Statistical analysis
Geometric mean values for urinary phthalates and their
95% confidence intervals, unadjusted and adjusted for cre-
atinine concentration, were calculated and compared with
US population estimates for women of reproductive age
(i.e., 18–40 years) derived from 1999–2000 and 2001–
2002 National Health and Nutrition Examination Survey
(NHANES) data (32, 33). We also calculated geometric
mean values and 95% confidence intervals for US pregnant
women using the NHANES data (n ¼ 209 for MEHP and
n ¼ 104 for MEOHP, MEHHP, and %MEHP (defined be-
low)). Because the NHANES sample was nonrandom, we
used the recommended weights to correctly estimate varian-
ces (34). Phthalate metabolite concentrations, which were
all right-skewed, were transformed using the natural loga-
rithm. For concentrations below the limit of detection, we
assigned a value equal to the limit of detection divided by
the square root of 2 (35). A potential phenotypic marker of
DEHP metabolism, %MEHP, was calculated as the ratio of
MEHP concentration to the sum of the 3 DEHP metabolite
concentrations (in nanomoles) and transformed using the
natural logarithm (36).
We adjusted for urinary dilution in 2 ways. For compar-
isons with NHANES data, we used metabolite concentra-
tions divided by creatinine concentrations (lg/g creatinine).
However, because creatinine is associated with demo-
graphic and physiologic parameters that may be on the
causal pathway from exposure to outcome (37, 38), we in-
cluded a square root transformation of creatinine (mg/dL) as
a covariate in all regression models rather than dividing the
concentration by creatinine.
Am J Epidemiol 2009;169:1015–1024

Spearman correlation coefficients were used to estimate
pairwise associations among phthalate concentrations. Mul-
tivariate linear regression models were used to evaluate as-
sociations between urinary phthalate concentrations and
gestational age at delivery. Although the distribution of ges-
tational ages was slightly skewed, a sensitivity analysis us-
ing generalized estimating equations, which is valid even
under departure from normality, produced similar results.
Logistic regression was used to calculate associations of
phthalate concentrations with binary outcomes, such as ce-
sarean section delivery.
Covariates, including demographic characteristics (race,
geographic site, mother’s education, mother’s age, mother’s
employment status during pregnancy), sample characteris-
tics (timing of urine sample, creatinine), previous pregnancy
history (parity, history of miscarriage), sex of the baby, ma-
ternal and paternal smoking, job-related stress, and mother’s
prepregnancy health (nongestational diabetes, thyroid dis-
orders, fibroids, high blood pressure, and respiratory condi-
tions (asthma and chronic obstructive pulmonary disease))
were considered as potential confounders (Table 1) and
were retained in the model if they were significant at P 
0.15. Potential effect modifiers evaluated included sex of the
baby, %MEHP status (low/high), geographic location, tim-
ing of the urine sample, and parity. We had no information
on mother’s weight prior to pregnancy, body mass index, or
weight gain during pregnancy.
Statistical significance was defined by a (2-sided) P value
of 0.05 or lower and was expressed as a 95% confidence
interval. SAS 9.1 software (SAS Institute Inc., Cary, North
Carolina) was used to conduct all analyses.
RESULTS
Demographic and other sample characteristics varied by
geographic site (Table 1), with significant differences in
parity, race, education, history of fibroids, and gestational
age. After creatinine adjustment, DEHP metabolite concen-
trations varied across geographic locations, with signifi-
cantly higher concentrations of MEHP among women
from Missouri and Minnesota than among women from
California and Iowa. White non-Hispanic mothers, who
made up 84% of the study population, had marginally higher
concentrations of MEHP and MEOHP than Hispanics.
Women with some college education had somewhat lower
concentrations of MEOHP than women with a high school
education or less. Higher levels of job-related stress were
slightly associated with higher MEHP concentrations.
Women with a previous miscarriage had significantly lower
urinary concentrations of MEHP. Metabolite and creatinine
concentrations were not related to the timing of the urine
sample.
The unadjusted mean concentrations of all 3 DEHP me-
tabolites in the SFF women were somewhat lower than the
NHANES US population estimates for pregnant women
and women of reproductive age (Table 2), but only the
mean MEHHP concentration in SFF women was signifi-
cantly lower than that for nonpregnant women (11.9 ng/mL
vs. 19.8 ng/mL). Neither the unadjusted nor the creatinine-
Am J Epidemiol 2009;169:1015–1024
Maternal DEHP Metabolites and Timing of Labor 1017
adjusted mean values for pregnant women differed signif-
icantly between the SFF and the NHANES.
Associations of the DEHP metabolites with gestational
age, unadjusted and adjusted for covariates, are presented
in Table 3 and Figures 1 and 2. Overall, a log-unit increase
in the urinary concentrations of metabolites was associated
with a 1.3- to 1.5-day increase in gestational age. The effect
size was slightly greater for MEOHP than for MEHP, and
we saw no association between %MEHP and gestational
age. Effect estimates were somewhat attenuated after adjust-
ment for potential confounders. Each tertile increase in
MEHP concentration was associated with a 1.8-day (95%
confidence interval (CI): 0.2, 3.4) increase in gestational
age, and each tertile increase in MEOHP concentration
was associated with a 1.9-day increase (95% CI: 0.1, 3.8).
We compared estimates after removing 1 outlier for urinary
MEHP concentration (1,600 ng/mL) and 1 outlier for ges-
tational age (32 weeks) (Figure 1). The effect estimates did
not change appreciably. Since we did not see any reasons,
medical or otherwise, for excluding these subjects from the
analysis, they were included.
The odds of giving birth at greater than 41 weeks in-
creased by 100% and 120% with each log-unit increase in
MEHP and MEOHP concentrations, respectively, after ad-
justment for covariates (Table 4). Conversely, we observed
50% and 60% reductions in the odds of preterm delivery
with a log-unit increase in MEHP and MEOHP concentra-
tions, respectively. The frequency of premature delivery
(6%) was slightly lower than the NHANES US population
estimate of 11.5% for whites, because we excluded twins
and an eclampsia case, which would bring the overall prev-
alence to 9.5%. The higher socioeconomic status of this
cohort also contributed to the slightly lower prevalence.
The odds of a cesarean section delivery increased by 50%
with each log-unit increase in MEOHP concentration and by
30% with each log-unit increase in MEHP concentration.
We explored potential explanations for the association of
cesarean section with DEHP metabolite concentrations.
Women who had ‘‘failure of labor to progress’’ listed as
a reason for cesarean section delivery had a mean MEOHP
concentration that was 0.33 log units higher (95% CI: 0.08,
0.57) than that of all other women, after controlling for other
reasons for cesarean section. We found no associations of
DEHP metabolite concentrations with breech presentation
(n ¼ 8), repeat cesarean section (n ¼ 15), or medically in-
duced (n ¼ 85) or augmented (n ¼ 55) labor or with duration
of labor measured in minutes.
We conducted stratified analyses by geographic site and
%MEHP status (low/high). In Missouri, Iowa, and Califor-
nia, gestational age at delivery tended to increase with
higher MEHP and MEOHP concentrations, while in Min-
nesota gestational age tended to decrease (though not
significantly). After adjustment for creatinine, mother’s
education, parity, and job-related stress, the difference in
slopes for the 2 groups (California, Iowa, and Missouri vs.
Minnesota) was significant for MEOHP (P ¼ 0.02) and
marginally significant for MEHP (P ¼ 0.06).
The association of MEOHP concentrations with gesta-
tional age among women in the low-%MEHP category
was approximately 3 times stronger in magnitude than in
1018 Adibi et al.

Table 1. Characteristics of Participants (n ¼ 283) in the Study for Future Families, a US Multicenter Pregnancy Cohort Study, 2000–2004

Study Center
Overall
Characteristic California Iowa Minnesota Missouri
No. or Mean % No. or Mean % No. or Mean % No. or Mean % No. or Mean %
No. of subjects 283 100 48 17 80 28 87 31 68 24
Time between urine collection and delivery, weeksa
Mean 12 (8)b 14 (8) 11 (8) 12 (8) 12 (6)
First trimester 5 2 1 2 0 0 4 5 0 0
Second trimester 119 42 24 51 34 43 34 39 27 40
Third trimester 157 56 22 47 45 57 49 56 41 60
Sex of baby
Male 143 51 24 50 38 47 45 52 36 53
Female 140 49 24 50 42 53 42 48 32 47
Parity (no. of prior livebirths)***
0 145 51 30 63 32 40 52 60 31 46
1 87 31 11 23 25 31 27 31 24 35

2 51 18 7 15 23 29 8 9 13 19
Racec,***
White 238 84 22 46 72 91 84 97 60 88
Hispanic 26 9 20 42 2 3 0 0 4 6
Other 18 6 6 13 5 6 3 3 4 6
Education*
High school or less 25 9 9 19 6 8 3 3 7 10
Some college 63 22 19 40 14 18 20 23 10 15
College graduation 106 38 10 21 34 43 33 38 29 43
Graduate school 89 31 10 21 26 33 31 36 22 32
History of miscarriage (yes/no) 60 21 9 19 20 25 14 16 17 25
Employed during pregnancy (yes/no) 231 82 33 69 63 79 79 91 56 82
Job-related stress
No work-related stress 52 18 15 31 17 21 8 9 12 18
Not at all stressful 17 6 2 4 4 5 6 7 5 7
Not too stressful 80 28 14 29 21 26 29 33 16 24
Somewhat stressful 105 37 12 25 32 40 35 40 26 38
Very stressful 29 10 5 10 6 8 9 10 9 13
Maternal health
Diabetes 14 5 4 8 6 8 1 1 3 4
Thyroid disorders 17 6 5 10 5 6 2 2 5 7
Fibroids* 16 6 1 2 8 10 2 2 5 7
High blood pressure 19 7 4 8 5 6 5 6 5 7
Respiratory conditions 26 9 4 8 10 13 5 6 7 10
Mean creatinine level, mg/dL 92 (61) 99 (63) 93 (64) 82 (51) 97 (66)
Mean maternal age, years* 30.2 (6.0) 28.3 (6.0) 30.6 (4.6) 30.8 (5.3) 30.3 (4.6)
Gestational age at delivery (clinical estimate), weeksd
Mean** 39.2 (1.5) 38.9 (1.3) 38.9 (1.3) 39.7 (1.3) 39.1 (1.8)
<37 14 5 3 6 4 5 2 2 5 7
37–41 262 93 45 94 76 95 81 93 60 88
>41 7 2 0 0 0 0 4 5 3 4

* P < 0.05; **P < 0.01; ***P < 0.001 (P for difference between at least 2 of the 4 study centers).
a
Information on timing of urine sample collection was missing for 2 participants.
b
Numbers in parentheses, standard deviation.
c
Information on race was missing for 1 participant.
d
Clinical estimate of gestational age abstracted from the birth medical record; values were reported rounded to the next-highest week.
Information was missing for 1 participant.

Am J Epidemiol 2009;169:1015–1024
 Maternal DEHP Metabolites and Timing of Labor 1019

Table 2. Urinary Concentrations (ng/mL) of Di-(2-Ethylhexyl) Phthalate Metabolites Among Participants (n ¼ 283) in the Study for Future
Families (2000–2004) as Compared With Participants in the National Health and Nutrition Examination Survey (1999–2002)

NHANES
Di-(2-Ethylhexyl) % of Samples With Study for NHANES Pregnant
Percentile Reproductive-
Future Families Womena
Phthalate Values Greater Than Age Womenb
Metabolite Limit of Detection
5th 25th 50th 75th 95th GM 95% CI GM 95% CI GM 95% CI

MEHP 76 0.6 1.1 3.5 8.2 40.2 3.6 3.1, 4.3 4.8 3.8, 6.0 4.5 4.0, 5.0
MEHHP 97 1.1 5.6 11.2 25.5 99.4 11.9 10.1, 13.9 19.0 13.5, 26.7 19.8 14.9, 26.2
MEOHP 96 1.2 5.1 9.9 24.6 68.4 10.9 9.3, 12.6 15.4 11.3, 21.1 13.8 10.4, 18.3
%MEHPc NA 4 8 14 23 48 14 12, 15 17 13, 20 13 11, 14

Abbreviations: CI, confidence interval; GM, geometric mean; MEHHP, mono-2-ethyl-5-hydrohexyl phthalate; MEHP, mono-2-ethylhexyl phthal-
ate; MEOHP, mono-2-ethyl-5-oxohexyl phthalate; NA, not applicable; NHANES, National Health and Nutrition Examination Survey.
a
NHANES data on pregnant women, 1999–2000 (MEHP) and 2001–2002 (MEHP, MEOHP, and MEHHP) (n ¼ 209 for MEHP and n ¼ 104 for
MEHHP, MEOHP, and %MEHP).
b
NHANES data on women aged 18–40 years, 1999–2000 (MEHP) and 2001–2002 (MEHP, MEOHP, and MEHHP) (n ¼ 853 for MEHP and
n ¼ 437 for MEHHP, MEOHP, and %MEHP).
c
MEHP/R(MEHP, MEOHP, MEHHP) 3 100.

the high-%MEHP group after adjustment for creatinine,
geographic site, mother’s education, and job-related stress.
When modeled as a multiplicative interaction term, the
P value was 0.10 for differences in slope between the
2 groups (Figure 3). There was no difference in MEHP slope
between the low- and high-%MEHP categories.
DISCUSSION
In approximately 300 pregnant women from 4 US loca-
tions, we found that gestation was 1.1 days and 1.3 days
longer with each log-unit increase in urinary concentrations
of the DEHP metabolites MEHP and MEOHP, respectively.
Women at the 75th percentile of urinary MEHP concen-
tration had a duration of gestation that was 2.3 days
(95% CI: 1.4, 3.3) longer than that of women at the 25th
percentile of exposure, after we controlled for urinary
creatinine concentration, demographic factors, maternal
health, stress, and parity. DEHP exposure in this cohort
was similar to the NHANES US population estimate for
pregnant females.
The clinical or population significance of an exposure-
related shift of 2–3 days in gestational length is difficult to
evaluate. For this reason, we also estimated associations
with clinical outcomes and saw increased odds of delivering
after 41 weeks, decreased odds of preterm delivery, and in-
creased odds of delivering by cesarean section. Delivery
after 41 weeks is associated with an increase in perinatal
mortality due to meconium aspiration, fetal distress, as-
phyxia, pneumonia, malformations, shoulder dystocia, and
traumatic injuries (39, 40). Women who undergo cesarean
section are at increased risk in subsequent pregnancies of
malpresentation, abnormal placentation, antepartum hemor-
rhage, placenta accreta, prolonged labor, uterine rupture,
preterm birth, low birth weight, and stillbirth (41). The

Table 3. Relations of Urinary Concentrations of Di-(2-Ethylhexyl) Phthalate Metabolites in Pregnant Women to

Gestational Age at Delivery in Linear Regression Models (n ¼ 283), Study for Future Families, 2000–2004a,*

MEHP MEOHP MEHHP

Model
Estimate 95% CI Estimate 95% CI Estimate 95% CI

Adjusted for creatinine 0.18 0.04, 0.32 0.21 0.05, 0.38 0.19 0.05, 0.34
Adjusted for creatinine þ 0.17 0.03, 0.32 0.20 0.03, 0.36 0.18 0.02, 0.33
demographic factorsb
Adjusted for above factors þ 0.17 0.03, 0.31 0.19 0.03, 0.35 0.17 0.02, 0.32
maternal healthc
Adjusted for above factors þ 0.16 0.02, 0.30 0.19 0.03, 0.35 0.16 0.01, 0.31
parity

Abbreviations: CI, confidence interval; MEHHP, mono-2-ethyl-5-hydrohexyl phthalate; MEHP, mono-2-ethylhexyl

phthalate; MEOHP, mono-2-ethyl-5-oxohexyl phthalate.
* P < 0.05 for all results shown in table.
a
Change in gestational age at delivery (weeks) per log-unit increase in urinary di-(2-ethylhexyl) phthalate metab-
olite concentration.
b
Geographic center, mother’s educational level, and job-related stress.
c
Nongestational diabetes, thyroid disorders, and fibroids.

Am J Epidemiol 2009;169:1015–1024
1020 Adibi et al.

Figure 1. Clinically estimated gestational age as a function of log Figure 2. Clinically estimated gestational age as a function of log
urinary mono-2-ethylhexyl phthalate (MEHP) concentration (ng/mL) urinary mono-2-ethyl-5-oxohexyl phthalate (MEOHP) concentration
(n ¼ 283) after adjustment for creatinine, Study for Future Families, (ng/mL) (n ¼ 283) after adjustment for creatinine, Study for Future
2000–2004. Families, 2000–2004.

observed decrease in the risk of preterm delivery may be A prior study using MEHP concentration in umbilical
protective or may be indicative of abnormal function of the cord blood from 84 Italian mother/newborn pairs found an
placenta (26); in this study, we could not distinguish be- association that pointed in the opposite direction than was
tween these possibilities. observed here (8). Latini et al. (8) reported that gestational

Table 4. Relations of Urinary Concentrations of Di-(2-Ethylhexyl) Phthalate Metabolites in Pregnant Women to

Clinical Outcomes Related to Parturition and Labor in Logistic Regression Models (n ¼ 283), Study for Future
Families, 2000–2004a

No. of MEHP MEOHP MEHHP

Outcome %
Subjects OR 95% CI OR 95% CI OR 95% CI

Gestational age >41 weeks 7 2

Adjusted for creatinine 1.8* 1.1, 3.0 2.1** 1.2, 3.4 1.9* 1.1, 3.2
Adjusted for creatinine þ covariatesb 2.0* 1.1, 3.5 2.2** 1.3, 4.0 2.1* 1.3, 3.7
Preterm delivery (<37 weeks) 14 5
Adjusted for creatinine 0.5* 0.3, 0.9 0.5* 0.2, 0.9 0.5* 0.3, 0.9
c
Adjusted for creatinine þ covariates 0.5* 0.3, 0.9 0.4* 0.2, 0.9 0.5* 0.3, 0.9
Cesarean section delivery 62 22
Adjusted for creatinine 1.3 1.0, 1.6 1.4* 1.1, 1.9 1.4* 1.1, 1.8
Adjusted for creatinine þ covariatesd 1.3* 1.0, 1.6 1.5** 1.1, 1.9 1.4* 1.1, 1.8
Failure to progress as reason 18 6
for cesarean section delivery
Adjusted for creatinine 1.3 0.9, 1.8 1.6* 1.1, 2.4 1.5* 1.1, 2.2
Adjusted for creatinine þ covariatese 1.2 0.9, 1.7 1.6* 1.1, 2.3 1.5* 1.0, 2.1

Abbreviations: CI, confidence interval; MEHHP, mono-2-ethylhexyl phthalate; MEHP, mono-2-ethylhexyl phthal-
ate; MEOHP, mono-2-ethyl-5-oxohexyl phthalate; OR, odds ratio.
* P < 0.05; **P < 0.01.
a
Change in log odds per log-unit increase in urinary di-(2-ethylhexyl) phthalate metabolite concentration.
b
Adjusted for geographic site (Minnesota vs. California, Iowa, and Missouri) and respiratory conditions.
c
Adjusted for high blood pressure and nongestational diabetes.
d
Adjusted for mother’s age (
35 years vs. <35 years), geographic site (Minnesota vs. California, Iowa, and
Missouri), nongestational diabetes, and fibroids.
e
Adjusted for parity, thyroid conditions, and high blood pressure. Information on the reason for cesarean section
was missing for 2 subjects.

Am J Epidemiol 2009;169:1015–1024

Figure 3. Clinically estimated gestational age as a function of log
urinary mono-2-ethyl-5-oxohexyl phthalate (MEOHP) concentration
(ng/mL), by percent mono-2-ethylhexyl phthalate (%MEHP) status
(low/high) (n ¼ 283), after adjustment for creatinine and covariates,
Study for Future Families, 2000–2004. %MEHP was calculated as the
ratio of MEHP concentration to the sum of the concentrations of
3 di-(2-ethylhexyl) phthalate metabolites (MEHP, MEOHP, and mono-
2-ethylhexyl phthalate) (in nanomoles) and transformed using the
natural logarithm.
age was shorter by 1.2 weeks in the MEHP-positive pairs
than in the MEHP-negative subjects. In that study, they
measured MEHP and DEHP in umbilical cord blood, which
may have been subject to contamination by DEHP in the
sampling and analytic equipment (42). Measuring MEHP in
blood is not recommended because of its short half-life (14).
In that analysis, exposure status was dichotomized as ex-
posed and nonexposed. Given that more than 95% of the
general US population has detectable urinary metabolites of
DEHP (5), this approach could have resulted in misclassifi-
cation among the nonexposed subjects. The discrepant re-
sults between these 2 studies could be due to differences in
study design, exposure assessment, and/or the underlying
characteristics of the populations.
When we stratified SFF subjects by geographic loca-
tion, gestational age tended to increase with increasing
phthalate metabolite concentrations for all sites except
Minnesota, where it tended to remain flat or decrease
slightly. The women from Minnesota tended to have
higher MEOHP concentrations and %MEHP, higher ges-
tational age, higher maternal age, and more education than
women from other study centers, most markedly relative
to California, and they were predominantly white (97%).
We can speculate on 2 possible explanations. It is possible
that the dose-response curve was nonlinear and essentially
reversed at the higher doses among the Minnesota sub-
jects. It is also possible that the site-specific populations
differed in ways that modified the relation between DEHP
exposure and placental function. The significant differen-
ces in race, education, maternal age, and parity between
study centers could be proxies for other unmeasured effect
Am J Epidemiol 2009;169:1015–1024
Maternal DEHP Metabolites and Timing of Labor 1021
modifiers, such as nutritional factors, coexposures, and/or
lifestyle factors.
Other known causes of prolonged gestation include fish
oil consumption during pregnancy (43, 44), deficiency in
placental sulfatase, which is another cause of decreased
estrogen synthesis (45), and living in a highly polluted area
(46). Fish oil contains n-3 long-chain polyunsaturated fatty
acids, which are also ligands of PPARc and may contribute
to a suppressed inflammatory response late in pregnancy
(47, 48). If this were the case, it is possible that a competitive
interaction between DEHP and fatty acids in the diet exists.
We could not test this hypothesis, since fish consumption
among our subjects was generally low (88% of those who
consumed fish had 2 or fewer servings per week). We did not
have information on the type of fish consumed or on how it
was prepared.
We hypothesized that the association between DEHP ex-
posure and timing of labor could also differ depending on
a woman’s ability to metabolize and excrete DEHP. To test
this, we dichotomized %MEHP values at the median and
compared metabolite associations within the low and high
strata. The association of MEOHP concentrations with
longer gestation was 3-fold stronger in the low-%MEHP
group than in the high-%MEHP group. The association of
MEOHP concentration with the timing of labor might be
due to disruption in parturition and signaling by DEHP me-
tabolites, but it might also be due to differences in a woman’s
ability to metabolize and excrete DEHP. In a previous
report, we found %MEHP to be approximately twice as re-
producible within a woman over the last 6 weeks of preg-
nancy as DEHP metabolites (10), suggesting that %MEHP
may reflect stable interindividual differences that could
be relevant to the metabolism and excretion of these com-
pounds in pregnancy.
Concern exists about the potential for systematic error in
estimating gestational age using the last-menstrual-period
approach (49, 50). We addressed this by also using the
clinical estimate, which takes into account ultrasound data
and examination of the newborn in cases where there are
inconsistencies in last-menstrual-period dating and clinical
presentation, but the clinical estimate may have been less
precise, given that it was rounded up. We found results to
be consistent using both measures when including all ges-
tational ages and less consistent when modeling associa-
tions with preterm and postterm delivery. Misclassification
of gestational age by last menstrual period is most prob-
lematic among preterm and postterm births, with the de-
gree of misclassification being associated with maternal
race, age, education, parity, month that prenatal care began
(51), and regularity of the menstrual period (52). Given
that some of these factors are also related to phthalate
exposure and pregnancy outcomes, we relied on the clini-
cal estimate to model associations with DEHP metabolite
concentrations.
Some misclassification of DEHP exposure was present in
our data, given that we had a single spot urine sample for
characterizing exposure. In a previous analysis, we showed
that DEHP metabolite concentrations are not highly repro-
ducible in pregnant women over the last 6 weeks of preg-
nancy, probably because of physiologic changes occurring
1022 Adibi et al.
in the third trimester (10). Of the SFF subjects, 55% were
sampled in the third trimester. In addition to misclassifica-
tion, there may have been other exposures and risk factors
associated with urinary DEHP metabolite concentrations
and birth outcomes that we were not able to adequately
control for, resulting in residual confounding.
In conclusion, we observed an association between in-
creased concentrations of DEHP metabolites in maternal
urine measured during pregnancy and gestational age in a
US multicenter pregnancy cohort study. The direction and
size of the effect appeared to differ depending in part on
geographic location and a woman’s ability to metabolize
and eliminate DEHP. Our results support the hypothesis that
DEHP exposure may alter the dialogue between the maternal
and fetal compartments that is essential for normal labor.
Consistent with this hypothesis, urinary DEHP metabolite
concentrations were associated with an increased risk of ce-
sarean section delivery and of delivering at more than 41
weeks. The binding affinity of the metabolite MEHP for
PPARc and the central role PPARc plays in regulating pla-
cental function may provide an explanation for this associa-
tion, but this was not directly explored in the present study.
These results need to be replicated in other populations.
There is likewise a need for more in vitro and in vivo re-
search to better understand molecular mechanisms by which
DEHP may alter placental development and/or function.
ACKNOWLEDGMENTS
Author affiliations: Department of Environmental Health,
Harvard School of Public Health, Boston, Massachusetts
(Jennifer J. Adibi, Russ Hauser, Heather Nelson, Robert
Herrick); Department of Biostatistics, Harvard School of
Public Health, Boston, Massachusetts (Paige L. Williams);
Department of Environmental Health Sciences, Mailman
School of Public Health, Columbia University, New York,
New York (Robin M. Whyatt); National Center for Environ-
mental Health, Centers for Disease Control and Prevention,
Atlanta, Georgia (Antonia M. Calafat); and Department of
Obstetrics and Gynecology, University of Rochester Medi-
cal Center, Rochester, New York (Shanna H. Swan).
This research was funded by the Environmental Protec-
tion Agency (Star Grant R-82943601-0) and the National
Institute of Environmental Health Sciences (grant R01
ES013543). Jennifer Adibi’s doctoral training was funded
by the Harvard Education and Research Center for Occupa-
tional Safety and Health (grant T42 OH008416).
The authors thank Maureen Nealon and Fan Liu for the
processing and transfer of data, Dr. Sonia Hernandez-Diaz
for her mentoring with regard to epidemiologic methods,
Dr. Shruthi Mahalingaiah for her help in interpreting clinical
data, and Dr. Rick Stalhut for his assistance in accessing
and analyzing data from the National Health and Nutrition
Examination Survey. The authors also thank Dr. Manori
Silva, Dr. Jack Reidy, Jim Preau, and Ella Samandar of
the Centers for Disease Control and Prevention for the
phthalate metabolite measurements.
The findings and conclusions presented in this report are
those of the authors and do not necessarily represent the
views of the Centers for Disease Control and Prevention.
Conflict of interest: none declared.
REFERENCES
1. Schettler T. Human exposure to phthalates via consumer
products. Int J Androl. 2006;29(1):134–139.
2. Kavlock R, Barr D, Boekelheide K, et al. NTP-CERHR expert
panel update on the reproductive and developmental toxicity
of di(2-ethylhexyl) phthalate. Reprod Toxicol. 2006;22(3):
291–399.
3. Agency for Toxic Substances and Disease Registry, Centers
for Disease Control and Prevention. Toxicological Profile for
Di(2-ethylhexyl)phthalate (DEHP). Atlanta, GA: Centers for
Disease Control and Prevention; 2002. (http://www.atsdr.cdc.
gov/toxprofiles/tp9.html#9). (Accessed December 15, 2007).
4. Rudel RA, Camann DE, Spengler JD, et al. Phthalates,
alkylphenols, pesticides, polybrominated diphenyl
ethers, and other endocrine-disrupting compounds in
indoor air and dust. Environ Sci Technol. 2003;37(20):
4543–4553.
5. Kato K, Silva MJ, Reidy JA, et al. Mono(2-ethyl-5-
hydroxyhexyl) phthalate and mono-(2-ethyl-5-oxohexyl)
phthalate as biomarkers for human exposure assessment
to di-(2-ethylhexyl) phthalate. Environ Health Perspect.
2004;112(3):327–330.
6. Singh AR, Lawrence WH, Autian J. Maternal-fetal transfer of
14C-di-2-ethylhexyl phthalate and 14C-diethyl phthalate in
rats. J Pharm Sci. 1975;64(8):1347–1350.
7. Mose T, Mortensen GK, Hedegaard M, et al. Phthalate
monoesters in perfusate from a dual placenta perfusion system,
the placenta tissue and umbilical cord blood. Reprod Toxicol.
2007;23(1):83–91.
8. Latini G, De Felice C, Presta G, et al. In utero exposure
to di-(2-ethylhexyl)phthalate and duration of human
pregnancy. Environ Health Perspect. 2003;111(14):
1783–1785.
9. Silva MJ, Reidy JA, Herbert AR, et al. Detection of phthalate
metabolites in human amniotic fluid. Bull Environ Contam
Toxicol. 2004;72(6):1226–1231.
10. Adibi JJ, Whyatt RM, Williams PL, et al. Characterization of
phthalate exposure among pregnant women assessed by repeat
air and urine samples. Environ Health Perspect. 2008;116(4):
467–473.
11. Poole CF, Wibberley DG. Determination of di-(2-ethylhexyl)
phthalate in human placenta. J Chromatogr. 1977;132(3):
511–518.
12. Kato K, Silva MJ, Needham LL, et al. Quantifying phthalate
metabolites in human meconium and semen using automated
off-line solid-phase extraction coupled with on-line SPE and
isotope-dilution high-performance liquid chromatography–
tandem mass spectrometry. Anal Chem. 2006;78(18):
6651–6655.
13. Weuve J, Sanchez BN, Calafat AM, et al. Exposure to phthal-
ates in neonatal intensive care unit infants: urinary concen-
trations of monoesters and oxidative metabolites. Environ
Health Perspect. 2006;114(9):1424–1431.
14. Koch HM, Bolt HM, Preuss R, et al. New metabolites of di(2-
ethylhexyl)phthalate (DEHP) in human urine and serum after
single oral doses of deuterium-labelled DEHP. Arch Toxicol.
2005;79(7):367–376.
Am J Epidemiol 2009;169:1015–1024

15. Parks LG, Ostby JS, Lambright CR, et al. The plasticizer
diethylhexyl phthalate induces malformations by decreasing
fetal testosterone synthesis during sexual differentiation in the
male rat. Toxicol Sci. 2000;58(2):339–349.
16. Akingbemi BT, Youker RT, Sottas CM, et al. Modulation of rat
Leydig cell steroidogenic function by di(2-ethylhexyl)phthal-
ate. Biol Reprod. 2001;65(4):1252–1259.
17. Davis BJ, Maronpot RR, Heindel JJ. Di-(2-ethylhexyl)
phthalate suppresses estradiol and ovulation in cycling rats.
Toxicol Appl Pharmacol. 1994;128(2):216–223.
18. Lovekamp TN, Davis BJ. Mono-(2-ethylhexyl) phthalate
suppresses aromatase transcript levels and estradiol production
in cultured rat granulosa cells. Toxicol Appl Pharmacol.
2001;172(3):217–224.
19. Fournet-Dulguerov N, MacLusky NJ, Leranth CZ, et al. Im-
munohistochemical localization of aromatase cytochrome
P-450 and estradiol dehydrogenase in the syncytiotrophoblast
of the human placenta. J Clin Endocrinol Metab. 1987;65(4):
757–764.
20. Borch J, Metzdorff SB, Vinggaard AM, et al. Mechanisms
underlying the anti-androgenic effects of diethylhexyl
phthalate in fetal rat testis. Toxicology. 2006;223(1-2):
144–155.
21. Lovekamp-Swan T, Jetten AM, Davis BJ. Dual activation of
PPARa and PPARc by mono-(2-ethylhexyl) phthalate in rat
ovarian granulosa cells. Mol Cell Endocrinol. 2003;201(1-2):
133–141.
22. Xu Y, Cook TJ, Knipp GT. Effects of di-(2-ethylhexyl)-
phthalate (DEHP) and its metabolites on fatty acid homeo-
stasis regulating proteins in rat placental HRP-1 trophoblast
cells. Toxicol Sci. 2005;84(2):287–300.
23. Feige JN, Gelman L, Rossi D, et al. The endocrine disruptor
monoethyl-hexyl-phthalate is a selective peroxisome
proliferator-activated receptor c modulator that promotes
adipogenesis. J Biol Chem. 2007;282(26):19152–19166.
24. Hurst CH, Waxman DJ. Activation of PPARa and PPARc by
environmental phthalate monoesters. Toxicol Sci. 2003;74(2):
297–308.
25. Fournier T, Tsatsaris V, Handschuh K, et al. PPARs and the
placenta. Placenta. 2007;28(2-3):65–76.
26. Froment P, Gizard F, Defever D, et al. Peroxisome proliferator-
activated receptors in reproductive tissues: from game-
togenesis to parturition. J Endocrinol. 2006;189(2):
199–209.
27. Schaiff WT, Barak Y, Sadovsky Y. The pleiotropic function of
PPARc in the placenta. Mol Cell Endocrinol. 2006;249(1-2):
10–15.
28. Dunn-Albanese LR, Ackerman WE IV, Xie Y, et al. Reciprocal
expression of peroxisome proliferator-activated receptor-
gamma and cyclooxygenase-2 in human term parturition. Am J
Obstet Gynecol. 2004;190(3):809–816.
29. Swan SH, Main KM, Liu F, et al. Decrease in anogenital
distance among male infants with prenatal phthalate exposure.
Environ Health Perspect. 2005;113(8):1056–1061.
30. Qin C, Dietz PM, England LJ, et al. Effects of different data-
editing methods on trends in race-specific preterm delivery
rates, United States, 1990–2002. Paediatr Perinat Epidemiol.
2007;21(suppl 2):41–49.
31. Silva MJ, Slakman AR, Reidy JA, et al. Analysis of human
urine for fifteen phthalate metabolites using automated solid-
phase extraction. J Chromatogr B Analyt Technol Biomed Life
Sci. 2004;805(1):161–167.
32. National Center for Health Statistics. National Health and
Nutrition Examination Survey. NHANES 1999–2000. Hyatts-
ville, MD: National Center for Health Statistics; 2006. (http://
Am J Epidemiol 2009;169:1015–1024
Maternal DEHP Metabolites and Timing of Labor 1023
www.cdc.gov/nchs/about/major/nhanes/nhanes99_00.htm).
(Accessed July 11, 2006).
33. National Center for Health Statistics. National Health and Nu-
trition Examination Survey. NHANES 2001–2002. Hyattsville,
MD: National Center for Health Statistics; 2006. (http://
www.cdc.gov/nchs/about/major/nhanes/nhanes01-02.htm).
(Accessed July 11, 2006).
34. National Center for Health Statistics. Continuous NHANES
Web Tutorial. Hyattsville, MD: National Center for Health
Statistics; 2007. (http://www.cdc.gov/nchs/tutorials/NHANES/
index_current.htm). (Accessed June 18, 2007).
35. Hornung RW, Reed LD. Estimation of average concentration
in the presence of nondetectable values. Appl Occup Environ
Hyg. 1990;5(1):46–51.
36. Hauser R, Meeker JD, Singh NP, et al. DNA damage in human
sperm is related to urinary levels of phthalate monoester and
oxidative metabolites. Hum Reprod. 2007;22(3):688–695.
37. Barr DB, Wilder LC, Caudill SP, et al. Urinary creatinine
concentrations in the US population: implications for urinary
biologic monitoring measurements. Environ Health Perspect.
2005;113(2):192–200.
38. Schisterman EF, Whitcomb BW, Louis GM, et al. Lipid ad-
justment in the analysis of environmental contaminants and
human health risks. Environ Health Perspect. 2005;113(7):
853–857.
39. Olesen AW, Westergaard JG, Olsen J. Perinatal and maternal
complications related to postterm delivery: a national register-
based study, 1978–1993. Am J Obstet Gynecol. 2003;189(1):
222–227.
40. Hilder L, Costeloe K, Thilaganathan B. Prolonged pregnancy:
evaluating gestation-specific risks of fetal and infant mortality.
Br J Obstet Gynaecol. 1998;105(2):169–173.
41. Kennare R, Tucker G, Heard A, et al. Risks of adverse out-
comes in the next birth after a first cesarean delivery. Obstet
Gynecol. 2007;109(2):270–276.
42. Calafat AM, Needham LL. Factors affecting the evaluation of
biomonitoring data for human exposure assessment. Int J An-
drol. 2008;31(2):139–143.
43. Makrides M, Duley L, Olsen SF. Marine oil, and other pros-
taglandin precursor, supplementation for pregnancy uncom-
plicated by pre-eclampsia or intrauterine growth restriction.
Cochrane Database Syst Rev. 2006;3:CD003402.
44. Szajewska H, Horvath A, Koletzko B. Effect of n-3 long-chain
polyunsaturated fatty acid supplementation of women with
low-risk pregnancies on pregnancy outcomes and growth
measures at birth: a meta-analysis of randomized controlled
trials. Am J Clin Nutr. 2006;83(6):1337–1344.
45. Cunningham FG, Williams JW. Williams Obstetrics. New
York, NY: McGraw-Hill Professional; 2005.
46. Shea KM, Wilcox AJ, Little RE. Postterm delivery: a challenge
for epidemiologic research. Epidemiology. 1998;9(2):199–204.
47. Olsen SF, Østerdal ML, Salvig JD, et al. Duration of preg-
nancy in relation to fish oil supplementation and habitual fish
intake: a randomised clinical trial with fish oil. Eur J Clin Nutr.
2007;61(8):976–985.
48. Schaiff WT, Knapp FF Jr, Barak Y, et al. Ligand-activated
peroxisome proliferator activated receptor gamma alters pla-
cental morphology and placental fatty acid uptake in mice.
Endocrinology. 2007;148(8):3625–3634.
49. Hediger M, Kiely J. Foreword. Paediatr Perinat Epidemiol.
2007;21(suppl 2):1–3.
50. Savitz DA, Terry JW Jr, Dole N, et al. Comparison of preg-
nancy dating by last menstrual period, ultrasound scanning,
and their combination. Am J Obstet Gynecol. 2002;187(6):
1660–1666.
1024 Adibi et al.

51. Dietz PM, England LJ, Callaghan WM, et al. A comparison of 52. Ananth CV. Menstrual versus clinical estimate of gestational
LMP-based and ultrasound-based estimates of gestational age age dating in the United States: temporal trends and variability
using linked California livebirth and prenatal screening rec- in indices of perinatal outcomes. Paediatr Perinat Epidemiol.
ords. Paediatr Perinat Epidemiol. 2007;21(suppl 2):62–71. 2007;21(suppl 2):22–30.

Am J Epidemiol 2009;169:1015–1024
 American Journal of Epidemiology Vol. 169, No. 8
ª The Author 2009. Published by the Johns Hopkins Bloomberg School of Public Health. DOI: 10.1093/aje/kwp004
All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org. Advance Access publication March 6, 2009

Original Contribution

Longitudinal Trends in Hazardous Alcohol Consumption Among Women With

Human Immunodeficiency Virus Infection, 1995–2006

Robert L. Cook, Fang Zhu, Bea Herbeck Belnap, Kathleen Weber, Judith A. Cook, David Vlahov,
Tracey E. Wilson, Nancy A. Hessol, Michael Plankey, Andrea A. Howard, Stephen R. Cole, Gerald
B. Sharp, Jean L. Richardson, and Mardge H. Cohen

Initially submitted August 28, 2008; accepted for publication January 6, 2009.

Hazardous alcohol consumption among women with human immunodeficiency virus (HIV) infection is associ-
ated with several adverse health and behavioral outcomes, but the proportion of HIV-positive women who engage
in hazardous drinking over time is unclear. The authors sought to determine rates of hazardous alcohol consump-
tion among these women over time and to identify factors associated with this behavior. Subjects were 2,770 HIV-
positive women recruited from 6 US cities who participated in semiannual follow-up visits in the Women’s
Interagency HIV Study from 1995 to 2006. Hazardous alcohol consumption was defined as exceeding daily
(4 drinks) or weekly (>7 drinks) consumption recommendations. Over the 11-year follow-up period, 14%–24%
of the women reported past-year hazardous drinking, with a slight decrease in hazardous drinking over time.
Women were significantly more likely to report hazardous drinking if they were unemployed, were not high school
graduates, had been enrolled in the original cohort (1994–1995), had a CD4 cell count of 200–500 cells/mL, were
hepatitis C-seropositive, or had symptoms of depression. Approximately 1 in 5 of the women met criteria for
hazardous drinking. Interventions to identify and address hazardous drinking among HIV-positive women are
urgently needed.

alcohol drinking; HIV; longitudinal studies; women

Abbreviations: HAART, highly active antiretroviral therapy; HCV, hepatitis C virus; HIV, human immunodeficiency virus; WIHS,
Women’s Interagency HIV Study.

Between 1990 and 2000, the US human immunodefi-
ciency virus (HIV) epidemic expanded to increasingly
include women. Today, women account for more than
one-quarter of all new diagnoses of HIV infection and ac-
quired immunodeficiency syndrome in the United States (1).
As a growing number of women become infected with HIV,
the need to identify and address modifiable factors that af-
fect disease progression or survival becomes more urgent.
Hazardous alcohol use is one such modifiable behavior.
Hazardous alcohol use is defined as a pattern of alcohol
consumption that is associated with an increase in alcohol-
related or other health problems but may not necessarily be
classified as alcoholism (2). The National Institute on Alco-
hol Abuse and Alcoholism defines hazardous alcohol use for
women as an average consumption of >7 drinks per week or
4 drinks at 1 sitting at least weekly. It is important to
distinguish hazardous alcohol use from nonhazardous use,
since low-to-moderate levels of alcohol consumption are
often associated with improved health outcomes and sur-
vival, whereas hazardous consumption is associated with
poorer outcomes (3, 4).
Although studies carried out before the era of highly ac-
tive antiretroviral therapy (HAART) did not find alcohol
consumption to be associated with disease progression (5),
more recent studies have demonstrated a significant associ-
ation between hazardous alcohol consumption and several
adverse health outcomes and behaviors in HIV-positive
women. These include increased HIV viral load, lower

Correspondence to Dr. Robert L. Cook, Department of Epidemiology and Biostatistics, College of Public Health and Health Professions,
University of Florida, P.O. Box 100231, Gainesville, FL 32607 (e-mail: cookrl@phhp.ufl.edu).

1025 Am J Epidemiol 2009;169:1025–1032

1026 Cook et al.
medication adherence, increased risky sexual behavior, and
more rapid disease progression (6–12). Women who are
coinfected with HIV and hepatitis C virus (HCV) are at even
greater risk of alcohol-associated health problems.
Previous studies have suggested that 6%–54% of HIV-
positive women meet criteria for hazardous drinking in dif-
ferent settings, depending on the specific measure of haz-
ardous drinking used (6, 13–16). These studies were limited
by assessment at only 1 time point or recruitment from
a single setting. Longitudinal assessment of alcohol con-
sumption is important, because drinking behavior may
change with increasing age or with HIV disease progression.
Drinking patterns may also be influenced by period effects
(factors common to all age groups at a particular time point)
or cohort effects (factors common to all women enrolled in
a study in a given year) (17). Furthermore, many drinkers do
not show consistent behavior, varying their drinking pattern
and volume over time (17).
Here we describe longitudinal patterns of alcohol con-
sumption in HIV-positive women. Identification of demo-
graphic, clinical, or behavioral characteristics of women at
increased risk for hazardous drinking might help clinicians
recognize those women, so that increased screening, pre-
vention, and therapeutic services could be provided. Our
objectives in this study were to determine the proportion
of HIV-positive women with hazardous alcohol consump-
tion over 11 years of follow-up and to identify factors asso-
ciated with hazardous drinking in these women.
MATERIALS AND METHODS
Study population
Data were obtained from women participating in the
Women’s Interagency HIV Study (WIHS), the largest ongo-
ing longitudinal cohort study of HIV-positive women in the
United States. The WIHS is a multicenter, prospective study
that was established in 1994 to carry out comprehensive
investigations of the impact of HIV infection among women
aged 18 years. In a second recruitment phase, new mem-
bers were enrolled in the study during 2000–2001. Women
are seen at semiannual study visits. Our analyses were lim-
ited to 2,770 women who were confirmed by serologic
analysis to be HIV-positive at study entry.
Participating study sites are centered in the Bronx and
Brooklyn, New York; Washington, DC; Los Angeles and
San Francisco, California; and Chicago, Illinois. WIHS par-
ticipants were recruited from a variety of sources, including
HIV primary care clinics, hospital-based programs, research
programs, community outreach sites, women’s support
groups, drug rehabilitation programs, HIV testing sites,
and referrals from enrolled participants. The study design
has been described previously (18, 19), and additional infor-
mation is available at the study’s Web site (https://statepiaps.
jhsph.edu/wihs/).
The WIHS participants undergo semiannual physical ex-
aminations, complete study questionnaires, and provide
blood for biomarker measurements. The semiannual ques-
tionnaire data are obtained face to face by interviewers who
receive extensive standardized training regarding how to ask
questions on sensitive issues and how to minimize social
desirability bias. Participants who report heavy alcohol con-
sumption or drug use during the interviews are referred to
substance abuse treatment programs in the community with
established linkages to the WIHS sites. The number of
women who pursue such referrals is not known.
Institutional review boards at each of the study centers
and their community affiliates approved the WIHS proto-
cols, and informed written consent was obtained from all
participants.
Measures
Alcohol consumption. The WIHS has collected data on
quantity and frequency of alcohol consumption since its
inception. At each visit, WIHS participants are asked to
report the average number of days per week on which they
have consumed an alcoholic drink since their previous as-
sessment 6 months earlier. A drink is defined as ‘‘1 can,
bottle, or glass of beer, a glass of wine, a shot of liquor,
a mixed drink with that amount of liquor, or any other kind
of alcoholic beverage.’’ Response options are every day, 5–6
days per week, 3–4 days per week, 1–2 days per week, less
than once per week, and none. Next, participants are asked
about the usual number of drinks they have consumed per
day since their previous assessment. Open-ended responses
were elicited between 1996 and 2002, and choices were
categorized for visits taking place after October 2002
(0, 1–2, 3–4, 5–6, or 7 drinks per day). Open-ended responses
such as ‘‘a pint of vodka’’ (0.5 L) are converted to numbers
of standard drinks. At baseline, women also provided in-
formation on whether they had ever received treatment for
problematic alcohol consumption, and if so, which type of
treatment (e.g., inpatient alcohol detoxification, outpatient al-
cohol detoxification, attendance at Alcoholics Anonymous).
Using the WIHS data from visits 1–23 (1995–2006), we
created 2 types of alcohol consumption measures based on
the 2 questions relating to hazardous drinking and based the
cutpoint for these analyses on guidelines issued by the
National Institute on Alcohol Abuse and Alcoholism (2). First,
we calculated the average quantity of drinks per week by
multiplying quantity by frequency. For response items that
involved a range, we used the midpoint of the range. Weekly
consumption was also categorized as >7 drinks per week
(excessive weekly consumption) or 7 drinks per week.
Second, we used the average number of drinks consumed
per day to categorize women into those who consumed 4
drinks per day (excessive daily consumption) and those who
consumed <4 drinks per day. Then, using these cutpoints at
each visit, we categorized a woman’s alcohol consumption
into 1 of 3 groups: hazardous drinking (>7 drinks per week
or 4 drinks per day), moderate drinking (any drinking that
did not qualify as hazardous drinking), or nondrinking. We
also created a dichotomous variable for any hazardous
drinking in the past year (i.e., current visit or previous visit).
Additional measures. Women provided information
about their age, race, educational attainment, marital status,
and employment status at the time of study enrollment. We
categorized the women according to their geographic site of
enrollment and study enrollment time (original cohort,
Am J Epidemiol 2009;169:1025–1032

recruited in 1994–1995, or second cohort, recruited in
2001–2002). Illicit drug use was assessed at each visit by
asking participants to indicate the quantity and frequency of
use of the following substances since their previous visit:
tobacco, marijuana, cocaine, ‘‘crack’’ or freebase cocaine,
heroin, and methadone. Current drug use was defined as
taking the drug at least once per month. Participants also
reported whether they had ever injected any drugs and
whether they were currently injecting drugs.
At each visit, a CD4 cell count was determined, with the
results categorized as <200 cells/mL, 200–500 cells/mL,
and >500 cells/mL. Hepatitis C infection was determined
by antibody testing from blood collected at study enrollment
(actual testing was done several years after enrollment).
Symptoms of depression were measured at each visit with
the 20-item Center for Epidemiologic Studies Depression
Scale (20); women were classified as having symptoms of
depression if they had a score of 16.
Statistical analyses
We compared characteristics of HIV-positive women with
and without hazardous drinking behavior at baseline or the
first follow-up visit. We used the t test to compare mean
values and the chi-squared test to compare categorical var-
iables (SAS, version 9.1; SAS Institute Inc., Cary, North
Carolina). Next, we used multivariable logistic regression
to assess the effect of sociodemographic and clinical vari-
ables on the risk of hazardous drinking at each visit. We
used a generalized estimating equations approach, allowing
participants to contribute data from each time point, with
adjustment for repeated measures within subjects. All socio-
demographic factors and clinical variables were selected for
the initial model, with the exception of drug use, which was
left out of the model because of significant collinearity with
the alcohol measure. Using stepwise elimination, the least
significant variable was dropped from the current model
until all of the variables left in the model were significant
(P < 0.05). The corresponding estimates were compared
with the initial estimates from the full model, and the
changes were negligible. These analyses were repeated in
a stratified manner according to WIHS enrollment cohort
date (1995–1996 or 2001–2002); since the results were sim-
ilar, we present the combined data as representative of the
entire sample.
We plotted the proportion of women with hazardous
drinking or moderate drinking in the past year, overall and
stratified by HCV status at baseline. To assess whether the
proportion of women who were hazardous drinkers might be
affected by study loss to follow-up or differential survival
rates, we repeated these analyses using only data from
women who completed 1 of the last 2 visits. The observed
trends were similar; therefore, we present results from the
entire cohort.
RESULTS
Baseline characteristics of the 2,770 HIV-infected
women, overall and stratified by baseline hazardous drink-
ing behavior pattern, are shown in Table 1. Women report-
Am J Epidemiol 2009;169:1025–1032
Alcohol Consumption in HIV-Positive Women 1027
ing hazardous drinking at baseline were significantly more
likely to be older than 40 years, to be African-American, to
be unemployed, to have not completed high school, or to be
part of the original (1994–1995) cohort. Women with haz-
ardous drinking patterns were also significantly more likely
to be HCV-seropositive and to have high levels of depressive
symptoms. They were also more likely to smoke cigarettes,
to use other drugs, and to report current or past injection
drug use.
Among women who did not engage in hazardous drinking
at baseline, 981 (45%) were drinking at moderate levels, 380
(17%) were current nondrinkers with a history of hazardous
drinking, and 825 (37%) were current nondrinkers with no
history of hazardous drinking. Among women who did en-
gage in hazardous drinking at baseline, 138 (25%) exceeded
weekly consumption levels only, 142 (25%) exceeded daily
consumption levels only, and 282 (50%) exceeded both
weekly and daily consumption levels. Overall, 17% of the
women had received 1 or more types of treatment for alco-
hol problems in the past, with the most common being
inpatient alcohol detoxification (13%), Alcoholics Anony-
mous attendance (12%), and outpatient alcohol detoxifica-
tion (8%).
Figure 1 shows the proportion of women who had con-
sumed alcohol in the previous year during the period 1995–
2006. Over this interval, 14%–24% of women met the
criteria for hazardous drinking, and 32%–48% of women
met the criteria for moderate drinking. There was a gradual
decline in the proportion of women who met the criteria for
hazardous drinking over time, with a corresponding increase
in the proportion of women who met the criteria for mod-
erate drinking during follow-up.
Figure 2 shows the proportion of HIV-positive women
who reported hazardous drinking in the previous year ac-
cording to their HCV serostatus at study enrollment. In the
mid-to-late 1990s, HCV-seropositive women were much
more likely to be hazardous drinkers than HCV-seronegative
women. However, this difference decreased over time, and
there was no significant difference in hazardous drinking
behavior according to HCV status at the most recent time
points.
In multivariable analysis, women were slightly but statis-
tically significantly less likely to engage in hazardous drink-
ing with each follow-up visit, a finding that was independent
of age (Table 2). Women were also significantly less likely
to be hazardous drinkers if they were employed at baseline
or if they had completed high school. In contrast, women
were significantly more likely to be hazardous drinkers if
they had CD4 cell counts of 200–500 cells/mL, were HCV-
seropositive, or had high levels of depressive symptoms.
There was no significant association of hazardous drinking
with age, race, or enrollment cohort.
DISCUSSION
In this paper, we report patterns of alcohol consumption
between 1995 and 2006 over time in a geographically di-
verse sample of 2,770 HIV-positive women from the WIHS
cohort. We found that at any time point, 14%–24% of
1028 Cook et al.

Table 1. Baseline Characteristics of HIV-Positive Women According to Self-reported Alcohol Consumption Status During the Past Year,
Women’s Interagency HIV Study, 1995–2006

Alcohol Consumption Statusa

Nonhazardous
Total (n 5 2,770) Hazardous Drinking
P Value
Drinking (n 5 562) (Moderate/None)
(n 5 2,208)
No. % No. % No. %

Sociodemographic factors
Mean age,b years 36 (7.8)c 37 (7.3) 36 (7.9) 0.06
Age group,b years
<30 642 23.2 102 18.2 540 24.5
30–40 1,334 48.2 257 45.7 1,077 48.8
>40 791 28.6 203 36.1 588 26.7 <0.0001
Race
White (non-Hispanic) 422 15.2 79 14.1 343 15.5
African-American (non-Hispanic) 1,606 58.0 366 65.1 1,240 56.2
Hispanic 654 23.6 104 18.5 550 24.9
Other 88 3.2 13 2.3 75 3.4 0.001
Marital statusb
Single (never married) 910 33.4 200 36.0 710 32.7
Married 1,010 37.0 192 34.6 818 37.7
Separated/divorced/widowed 807 29.6 163 29.4 644 29.7 0.27
Employedb 671 24.3 84 15.0 587 26.6 <0.0001
More than a high school educationb 878 31.8 141 25.1 737 33.5 0.0002
Study site
Bronx, New York 545 19.7 103 18.3 442 20.0
Brooklyn, New York 454 16.4 77 13.7 377 17.1
Washington, DC 411 14.8 79 14.1 332 15.0
Los Angeles, California 569 20.5 99 17.6 470 21.3
San Francisco, California 419 15.1 109 19.4 310 14.0
Chicago, Illinois 372 13.4 95 16.9 277 12.5 0.0006
Enrollment cohort
1994–1995 2,041 73.7 493 87.7 1,548 70.1
2001–2002 729 26.3 69 12.3 660 29.9 <0.0001
Clinical characteristics
CD4 cell count,b cells/mL
<200 660 24.6 119 22.0 541 25.2
200–500 1,164 43.4 254 47.0 910 42.4
>500 860 32.0 167 30.9 693 32.3 0.13
Hepatitis C-seropositiveb 948 35.3 290 53.0 658 30.7 <0.0001
Depressive symptoms (CES-D score 16)b 1,467 54.1 366 66.7 1,101 50.9 <0.0001
Drug useb
Current use
Cocaine 321 11.6 170 30.2 151 6.8 <0.0001
Heroin 282 10.2 121 21.5 161 7.3 <0.0001
Crack/freebase cocaine 437 15.8 222 39.5 215 9.7 <0.0001
(Illicit) methadone 38 1.4 19 3.4 19 0.9 <0.0001
Marijuana/hashish 591 21.4 236 42.1 355 16.1 <0.0001
Tobacco 1,415 51.1 437 77.8 978 44.4 <0.0001
Ever use of injected drugs 914 33.0 292 52.0 622 28.2 <0.0001
Abbreviations: CES-D, Center for Epidemiologic Studies Depression [Scale]; HIV, human immunodeficiency virus.
a
Hazardous drinking was defined as excessive weekly consumption (>7 drinks per week) or excessive daily consumption (4 drinks per
occasion). Information on drinking status for the first 2 visits was missing for 45 women.
b
Information on this measure was missing for some women.
c
Numbers in parentheses, standard deviation.

Am J Epidemiol 2009;169:1025–1032
 Alcohol Consumption in HIV-Positive Women 1029

Figure 1. Proportions of 2,770 human immunodeficiency virus (HIV)-positive women reporting having engaged in hazardous drinking and
nonhazardous drinking during the past year, Women’s Interagency HIV Study, 1995–2006. The alcohol measurement items were modified slightly
in 2002.

women reported consumption of alcohol at hazardous levels
in the past year, while over 50% of women had consumed
any alcohol.
The proportion of women with hazardous drinking at any
time point in our study was similar to or greater than the
proportion in other cross-sectional studies that evaluated
hazardous alcohol consumption among HIV-positive
women (6%–54%) (6–12). However, it is difficult to directly
compare proportions of women with hazardous drinking
across these studies because of differences in the measure-
ment of hazardous drinking. The proportions in our sample
were also similar to a recent (2001–2002) general popula-
tion estimate for US women aged 18 years or older, which
found that 22% of US adult women engage in hazardous
drinking (21).
Our analysis differs from previous studies of hazardous
drinking in HIV-positive women by documenting trends
over time. This allowed us to see, for example, that the
proportion of women with hazardous drinking behavior de-
clined gradually during the time period of 1995–2006,
whereas the overall proportion of women with any alcohol
consumption remained stable. Several other studies have
demonstrated a gradual decrease in hazardous drinking with
increasing age (17, 22). However, there is also evidence of

Figure 2. Proportion of 2,770 human immunodeficiency virus (HIV)-positive women reporting having engaged in hazardous drinking during the
past year, according to their hepatitis C virus status at baseline, Women’s Interagency HIV Study, 1995–2006.

Am J Epidemiol 2009;169:1025–1032
1030 Cook et al.

Table 2. Significant Predictors of Hazardous Alcohol Consumptiona with more rapid disease progression (9, 10, 12). One possi-
at Each Time Point in HIV-Positive Women (Multivariate Analyses), ble hypothesis regarding this shift is that in the HAART era,
Women’s Interagency HIV Study, 1995–2006b people with HIV infection are feeling better and therefore
Odds 95% Confidence continuing or increasing their use of recreational substances
Parameter P Value
Ratio Interval such as alcohol. Alternatively, hazardous drinking could di-
Current visit (vs. previous visit) 0.96 0.96, 0.97 <0.0001 rectly shorten survival through its effects on the immune
system, adverse health behaviors such as nonadherence to
Employment (yes vs. no) 0.80 0.72, 0.90 0.0002
medication protocols, or HIV-related comorbid conditions
More than a high school 0.72 0.59, 0.87 0.0009 such as hepatitis.
education (vs. high school
or less) The fact that women with HCV infection drank as much
Study site as or more than women without HCV is concerning, al-
though the steeper decline in rates of hazardous drinking
Brooklyn, New York 1 Referent
among women infected with both HIV and HCV is reassur-
Bronx, New York 1.15 0.87, 1.52 0.33 ing. Liver disease, mostly related to HCV, is now the leading
Washington, DC 1.65 1.23, 2.21 0.0009 cause of non-HIV-related death in women with HIV/
Los Angeles, California 1.18 0.87, 1.59 0.28 acquired immunodeficiency syndrome (26). Hazardous al-
San Francisco, California 1.49 1.12, 1.99 0.007 cohol consumption is also a significant barrier to receiving
Chicago, Illinois 1.31 0.96, 1.77 0.09
treatment for HCV infection (27). The more rapid decline
among women dually infected with both HIV and HCV
CD4 cell count, cells/mL
could be due to increased awareness and clinical interven-
>500 1 Referent tion, to decreased survival, or to some combination of these
200–500 1.10 1.00, 1.21 0.05 factors. In this study, baseline HCV antibody testing was not
<200 1.06 0.95, 1.24 0.22 conducted for several years. Therefore, many women were
Hepatitis C-seropositive 1.62 1.35, 1.95 <0.0001 not initially aware of their HCV status, and they may have
(yes vs. no) reduced their drinking only after learning they were HCV-
Depressive symptoms 1.31 1.21, 1.43 <0.0001 seropositive (28).
(CES-D score 16) The overlap of hazardous alcohol consumption with drug
(yes vs. no) use is also important to note. At baseline, over one-third of
Abbreviations: CES-D, Center for Epidemiologic Studies Depres- the women with hazardous drinking also had a history of
sion [Scale]; HIV, human immunodeficiency virus. injection or noninjection drug use. The relative contribution
a
Hazardous drinking was defined as excessive weekly consump- of hazardous alcohol consumption versus other drugs to ma-
tion (>7 drinks per week) or excessive daily consumption (4 drinks jor health outcomes in HIV infection can be difficult to dis-
per occasion). cern. Noninjection drug use has been associated with HIV
b
Other variables in the model that were not statistically significant progression and all-cause mortality in this cohort (29). In
included age, race, and enrollment cohort (1994–1995 or 2001– addition, nearly all of the women with HCV infection in this
2002).
sample had a history of current or past injection drug use,
which could party explain the increased prevalence of haz-
ardous drinking among HCV-seropositive women.
mixed longitudinal patterns, with a significant proportion of Several issues related to self-reported alcohol consump-
older women initiating hazardous drinking in middle to late tion measures warrant mention. Only a few alcohol con-
life (22). Our finding that the decrease in hazardous drinking sumption measurements were obtained at each semiannual
over time was independent of age or study cohort suggests study visit (i.e., quantity and frequency), limiting our ability
that the trend may represent a period effect, which refers to to examine participants for additional drinking patterns such
general changes in the population related to year of data as periodic binge drinking. Measures of quantity and fre-
collection (23). We also considered whether the gradual de- quency also occasionally lead to underestimates in compar-
crease in hazardous drinking could be related to differential ison with other drinking assessment methods (30). In
loss to follow-up or shorter survival among women with addition, the definition of a standard drink of alcohol may
hazardous drinking, but additional analyses did not support have been inconsistent across women. Early in the study,
this explanation. Although researchers in prior analyses of women could indicate specific types of beverages and
retention rates among women enrolled in the original cohort amounts that did not easily translate into numbers of stan-
did not find a difference between heavier drinkers and those dard drinks (e.g., a pint (0.5 L) of vodka). However, later in
who drank less (24), further investigation of the relation the study, we did not obtain this open-ended information,
between hazardous drinking and survival is warranted. which could account for part of the observed shift from
To date, conclusions about the relative impact of hazard- hazardous drinking to moderate drinking during the last
ous drinking on major health outcomes such as survival or few years of follow-up. Hazardous drinking was estimated
disease progression in HIV-positive persons have been for some women because of response items that included
mixed. In the pre-HAART era, most studies showed no ev- ranges, and because drinking patterns during the past month
idence of an effect of alcohol consumption on survival (5, could have varied. Social desirability can also result in un-
25). However, in more recent studies, investigators have derestimation of heavy drinking in some women. Although
concluded that hazardous alcohol consumption is correlated these measurement biases could point in either direction,

Am J Epidemiol 2009;169:1025–1032

they are probably more likely to underestimate drinking,
and thus the true rate of hazardous drinking could be higher.
In summary, 14%–24% of HIV-positive women in the
WIHS met criteria for past-year hazardous drinking over
a decade of follow-up. Although more than half of the
HIV-positive women in this study consumed at least some
alcohol, interventions should attempt to target those women
who are drinking at hazardous or unhealthy levels. In the
current era of HAART, hazardous drinking could have
a greater impact on health outcomes because of its associ-
ation with nonadherence to medication protocols. Data col-
lection for the WIHS started at the beginning of the HAART
era, and it is possible that some women cut down on their
drinking in order to improve their chances of survival.
Women with additional risk factors such as depression or
HCV coinfection may need even more intensive screening
and targeted intervention. Advice on how to screen and in-
tervene is available (2, 31, 32); yet despite the consistent
evidence of harm associated with hazardous drinking, in
many clinical settings hazardous alcohol consumption is
often neither detected nor addressed (33). Although the
study design does not allow for formal cause-effect conclu-
sions, these findings suggest that women with identified risk
factors such as depression or HCV infection may need ad-
ditional or more aggressive screening for hazardous drink-
ing. Further research and guidance are needed to translate
effective alcohol-related interventions into clinical practice
and to determine the impact of such interventions on the
health outcomes of women with HIV infection.
ACKNOWLEDGMENTS
Author affiliations: Department of Epidemiology and
Biostatistics, College of Public Health and Health Profes-
sions, University of Florida, Gainesville, Florida (Robert L.
Cook); Department of Medicine, School of Medicine, Uni-
versity of Pittsburgh, Pittsburgh, Pennsylvania (Fang Zuy,
Bea Herbeck Belnap); The CORE Center, Cook County
Bureau of Health Services, Chicago, Illinois (Kathleen
Weber, Mardge H. Cohen); Department of Internal Medicine,
Rush School of Medicine, Rush University, Chicago,
Illinois (Kathleen Weber, Mardge H. Cohen); Department
of Psychiatry, College of Medicine at Chicago, University of
Illinois, Chicago, Illinois (Judith A. Cook); New York Acad-
emy of Medicine, New York, New York (David Vlahov);
Department of Preventive Medicine and Community
Health, SUNY Downstate Medical Center, Brooklyn, New
York (Tracey E. Wilson); Departments of Clinical Pharmacy
and Medicine, Schools of Pharmacy and Medicine, Univer-
sity of California, San Francisco, San Francisco, California
(Nancy A. Hessol); Department of Medicine, Georgetown
University Medical Center, Washington, DC (Michael
Plankey); Department of Epidemiology and Population
Health, Montefiore Medical Center–Albert Einstein College of
Medicine, Bronx, New York (Andrea A. Howard); Depart-
ment of Epidemiology, School of Public Health, University
of North Carolina, Chapel Hill, North Carolina (Stephen R.
Cole); National Institute of Allergy and Infectious Diseases,
Am J Epidemiol 2009;169:1025–1032
Alcohol Consumption in HIV-Positive Women 1031
Bethesda, Maryland (Gerald B. Sharp); and Department of
Preventive Medicine, Keck School of Medicine, University
of Southern California, Los Angeles, California (Jean L.
Richardson).
The Women’s Interagency HIV Study (WIHS) is funded
by the National Institute of Allergy and Infectious Diseases
(grants UO1-AI-35004, UO1-AI-31834, UO1-AI-34994,
UO1-AI-34989, UO1-AI-34993, and UO1-AI-42590) and
the National Institute of Child Health and Human Develop-
ment (grant UO1-HD-32632). The study is cofunded by the
National Cancer Institute, the National Institute on Drug
Abuse, and the National Institute on Deafness and Other
Communication Disorders. Funding is also provided by
the National Center for Research Resources (University of
California, San Francisco–Clinical and Translational
Science Institute grant UL1 RR024131).
Data for this analysis were collected by the WIHS
Collaborative Study Group at the following centers: New York
City/Bronx Consortium (Kathryn Anastos, Principal Inves-
tigator (PI)); Brooklyn, New York (Howard Minkoff, PI);
Washington, DC, Metropolitan Consortium (Mary Young,
PI); The Connie Wofsy Study Consortium of Northern
California (Ruth Greenblatt, PI); Los Angeles County/
Southern California Consortium (Alexandra Levine, PI);
Chicago Consortium (Mardge Cohen, PI); and Data Coor-
dinating Center (Baltimore, Maryland) (Stephen Gange, PI).
The contents of this publication are solely the responsi-
bility of the authors and do not necessarily represent the
official views of the National Institutes of Health.
Conflict of interest: none declared.
REFERENCES
1. Centers for Disease Control and Prevention. HIV/AIDS Among
Women. (CDC HIV/AIDS Fact Sheet). Atlanta, GA: Centers
for Disease Control and Prevention; 2008. (http://www.cdc.
gov/hiv/topics/women/resources/factsheets/pdf/women.pdf).
(Accessed January 4, 2009).
2. National Institute on Alcohol Abuse and Alcoholism. Helping
Patients Who Drink Too Much: A Clinician’s Guide. Bethesda,
MD: National Institute on Alcohol Abuse and Alcoholism;
2005. (NIH publication no. 07-3769). (http://pubs.niaaa.nih.
gov/publications/Practitioner/CliniciansGuide2005/guide.pdf).
(Accessed January 4, 2009).
3. Caetano R, Ramisetty-Mikler S, Floyd LR, et al. The epide-
miology of drinking among women of child-bearing age. Al-
cohol Clin Exp Res. 2006;30(6):1023–1030.
4. Mukamal KJ, Conigrave KM, Mittleman MA, et al. Roles of
drinking pattern and type of alcohol consumed in coronary
heart disease in men. N Engl J Med. 2003;348(2):109–118.
5. Kaslow RA, Blackwelder WC, Ostrow DG, et al. No evidence
for a role of alcohol or other psychoactive drugs in acceler-
ating immunodeficiency in HIV-1-positive individuals. A re-
port from the Multicenter AIDS Cohort Study. JAMA. 1989;
261(23):3424–3429.
6. Cook RL, Sereika SM, Hunt SC, et al. Problem drinking and
medication adherence among persons with HIV infection.
J Gen Intern Med. 2001;16(2):83–88.
7. National Institute on Alcohol Abuse and Alcoholism. Alcohol
and HIV/AIDS. (Alcohol Alert no. 57). Bethesda, MD:
1032 Cook et al.
National Institute on Alcohol Abuse and Alcoholism; 2002.
(http://pubs.niaaa.nih.gov/publications/aa57.htm). (Accessed
January 4, 2009).
8. Wilson TE, Massad LS, Riester KA, et al. Sexual contracep-
tive, and drug use behaviors of women with HIV and those at
high risk for infection: results from the Women’s Interagency
HIV Study. AIDS. 1999;13(5):591–598.
9. Samet JH, Horton NJ, Traphagen ET, et al. Alcohol con-
sumption and HIV disease progression: are they related?
Alcohol Clin Exp Res. 2003;27(5):862–867.
10. Theall KP, Clark RA, Powell A, et al. Alcohol consumption,
ART usage and high-risk sex among women infected with
HIV. AIDS Behav. 2007;11(2):205–215.
11. Braithwaite RS, McGinnis KA, Conigliaro J, et al. A temporal
and dose-response association between alcohol consumption
and medication adherence among veterans in care. Alcohol
Clin Exp Res. 2005;29(7):1190–1197.
12. Chander G, Lau B, Moore RD. Hazardous alcohol use: a risk
factor for non-adherence and lack of suppression in HIV in-
fection. J Acquir Immune Defic Syndr. 2006;43(4):411–417.
13. Galvan FH, Bing EG, Fleishman JA, et al. The prevalence of
alcohol consumption and heavy drinking among people with
HIV in the United States: results from the HIV Cost and Serv-
ices Utilization Study. J Stud Alcohol. 2002;63(2):179–186.
14. Seth P, Wingood GM, Diclemente RJ. Exposure to alcohol
problems and its association with sexual behavior and
biologically-confirmed Trichomonas vaginalis among women
living with HIV. Sex Transm Infect. 2008;84(5):390–392.
15. Theall KP, Amedee A, Clark RA, et al. Alcohol consumption
and HIV-1 vaginal RNA shedding among women. J Stud
Alcohol Drugs. 2008;69(3):454–458.
16. Metsch LR, Pereyra M, Colfax G, et al. HIV-positive patients’
discussion of alcohol use with their HIV primary care pro-
viders. Drug Alcohol Depend. 2008;95(1-2):37–44.
17. Moore AA, Gould R, Reuben DB, et al. Longitudinal patterns
and predictors of alcohol consumption in the United States.
Am J Public Health. 2005;95(3):458–465.
18. Barkan SE, Melnick SL, Preston-Martin S, et al. The Women’s
Interagency HIV Study. WIHS Collaborative Study Group.
Epidemiology. 1998;9(2):117–124.
19. Bacon MC, von Wyl V, Alden C, et al. The Women’s Inter-
agency HIV Study: an observational cohort brings clinical
sciences to the bench. Clin Diagn Lab Immunol. 2005;12(9):
1013–1019.
20. Radloff LS. The CES-D Scale: a self-report depression scale
for research in the general population. Appl Psychol Meas.
1977;1(3):385–401.
21. National Institute on Alcohol Abuse and Alcoholism. Alcohol
Use and Alcohol Use Disorders in the United States: Main
Findings from the 2001–2002 National Epidemiologic Survey
on Alcohol and Related Conditions (NESARC). (U.S. Alcohol
Epidemiologic Data Reference Manual, vol 8, no. 1).
Bethesda, MD: National Institute on Alcohol Abuse and
Alcoholism; 2006. (NIH publication no. 05-5737). (http://
defeataddictions.org/files/2001_2002_National_Epidemiologic_
Survey.pdf). (Accessed January 4, 2009).
22. Liberto JG, Oslin DW, Ruskin PE. Alcoholism in older per-
sons: a review of the literature. Hosp Community Psychiatry.
1992;43(10):975–984.
23. Wilsnack RW, Kristjanson AF, Wilsnack SC, et al. Are U.S.
women drinking less (or more)? Historical and aging trends,
1981–2001. J Stud Alcohol. 2006;67(3):341–348.
24. Hessol NA, Schneider M, Greenblatt RM, et al. Retention of
women enrolled in a prospective study of human immunode-
ficiency virus infection: impact of race, unstable housing, and
use of human immunodeficiency virus therapy. Am J Epide-
miol. 2001;154(6):563–573.
25. Chandiwana SK, Sebit MB, Latif AS, et al. Alcohol con-
sumption in HIV-I infected persons: a study of immunological
markers, Harare, Zimbabwe. Cent Afr J Med. 1999;45(11):
303–308.
26. Cohen MH, French AL, Benning L, et al. Causes of death
among women with human immunodeficiency virus infection
in the era of combination antiretroviral therapy. Am J Med.
2002;113(2):91–98.
27. Nunes D, Saitz R, Libman H, et al. Barriers to treatment of
hepatitis C in HIV/HCV-coinfected adults with alcohol prob-
lems. Alcohol Clin Exp Res. 2006;30(9):1520–1526.
28. Tsui JI, Saitz R, Cheng DM, et al. Awareness of hepatitis C
diagnosis is associated with less alcohol use among persons
co-infected with HIV. J Gen Intern Med. 2007;22(6):822–825.
29. Kapadia F, Cook JA, Cohen MH, et al. The relationship be-
tween non-injection drug use behaviors on progression to
AIDS and death in a cohort of HIV seropositive women in the
era of highly active antiretroviral therapy use. Addiction.
2005;100(7):990–1002.
30. Allen JP, Wilson VB, eds. Assessing Alcohol Problems: A
Guide for Clinicians and Researchers. 2nd ed. Bethesda, MD:
National Institute on Alcohol Abuse and Alcoholism; 2003.
(NIH publication no. 03-3745). Available at URL: (http://
pubs.niaaa.nih.gov/publications/Assesing%20Alcohol/index.htm).
(Accessed January 4, 2009).
31. Saitz R. Unhealthy alcohol use. N Engl J Med. 2005;352(6):
596–607.
32. Aharonovich E, Hatzenbuehler ML, Johnston B, et al.
A low-cost, sustainable intervention for drinking reduction
in the HIV primary care setting. AIDS Care. 2006;18(6):
561–568.
33. Conigliaro J, Gordon AJ, McGinnis KA, et al. How harmful is
hazardous alcohol use and abuse in HIV infection: do health
care providers know who is at risk? J Acquir Immune Defic
Syndr. 2003;33(4):521–525.
Am J Epidemiol 2009;169:1025–1032
 American Journal of Epidemiology Vol. 169, No. 8
Published by the Johns Hopkins Bloomberg School of Public Health 2009. DOI: 10.1093/aje/kwp010
Advance Access publication February 27, 2009

Original Contribution

Suicide Mortality Among Patients Receiving Care in the Veterans Health

Administration Health System

John F. McCarthy, Marcia Valenstein, H. Myra Kim, Mark Ilgen, Kara Zivin, and Frederic C. Blow

Initially submitted February 20, 2008; accepted for publication January 12, 2009.

Understanding and reducing mortality from suicide among veterans is a national priority, particularly for individ-
uals receiving care from the US Veterans Health Administration (VHA). This report examines suicide rates among
VHA patients and compares them with rates in the general population. Suicide mortality was assessed in fiscal year
2001 for patients alive at the start of that fiscal year and with VHA use in fiscal years 2000–2001 (n ¼ 4,692,034).
Deaths from suicide were identified by using National Death Index data. General population rates were identified by
use of the Web-based Injury Statistics Query and Reporting System. VHA rates were 43.13/100,000 person-years
for men and 10.41/100,000 person-years for women. For male patients, the age-adjusted standardized mortality
ratio was 1.66; for females, it was 1.87. Male patients aged 30–79 years had increased risks relative to men in the
general population; standardized mortality ratios ranged from 2.56 (ages 30–39 years) to 1.33 (ages 70–79 years).
Female patients aged 40–59 years had greater risks than did women in the general population, with standardized
mortality ratios of 2.15 (ages 40–49 years) and 2.36 (ages 50–59 years). Findings offer heretofore unavailable
comparison points for health systems. Prior to the conflicts in Afghanistan and Iraq and before recent VHA
initiatives, rates were higher among VHA patients than in the general population. Female patients had particularly
high relative risks.

military medicine; mortality; suicide; veterans

Abbreviations: CI, confidence interval; VHA, Veterans Health Administration.

Understanding and reducing suicide risks among veterans
is a national concern and a major priority for the Department
of Veterans Affairs Veterans Health Administration (VHA)
(1–4). Each year, the VHA provides health services to approx-
imately 5 million veterans—roughly one-fifth of all veterans
living in the United States. The characteristics of the VHA
patient population include important suicide risk factors. The
VHA serves a predominantly male older population, with
substantial medical morbidities, high levels of substance
abuse and mental illness, and increased knowledge of and
access to firearms (5–7). However, there have been few com-
prehensive assessments of suicide mortality among veterans
generally (8) and none to our knowledge among veterans re-
ceiving services in the VHA health system.
In 2007, Congress required the Department of Veterans
Affairs to implement a comprehensive suicide prevention
program (9). This legislation reflects concerns regarding
the high mental health needs observed among veterans of
Operations Enduring Freedom and Iraqi Freedom (10–12),
the influx of these veterans now eligible for VHA health
services (13, 14), and indications that suicide risks may be
elevated among veterans compared with nonveterans. Nota-
bly, a highly publicized study observed that, in the period
1986–1997, male respondents to the US National Health
Interview Survey (NHIS) who indicated veteran status had
greater suicide risks than those who did not (adjusted hazard
ratio ¼ 2.04, 95% confidence interval (CI): 1.10, 3.80) (8).
Even before passage of this legislation, however, the
VHA had implemented key provisions and system-level
mental health enhancements. Based on its Mental Health
Strategic Plan (3), the VHA has worked to ensure access
to quality mental health services and to develop programs

Correspondence to Dr. John F. McCarthy, US Department of Veterans Affairs Serious Mental Illness Treatment Research and Evaluation Center,
P.O. Box 13017, Ann Arbor, MI 48113-0170 (e-mail: john.mccarthy2@va.gov).

1033 Am J Epidemiol 2009;169:1033–1038

1034 McCarthy et al.
specifically designed to prevent suicides. For example, since
2005, the VHA has added more than 3,600 mental health
staff, and VHA mental health service expenditures in-
creased from $2 billion in fiscal year 2001 to nearly
$4 billion in fiscal year 2008 (15). Evaluation of the impact
of ongoing changes in the VHA patient population and of
recent VHA mental health and suicide prevention initiatives
will require assessment of baseline suicide rates among
VHA patients.
Much of the existing research regarding suicide mortality
among veterans has focused on combat theater veteran co-
horts or veteran subgroups with service-related injuries.
These have not found increased rates among combat theater
cohorts (16–18), but risks have been shown to be higher in
association with combat trauma and post-traumatic stress
disorder (19, 20).
The few studies that have examined suicide mortality
among VHA patients have been limited to high-risk patient
subgroups. Increased risks have been demonstrated in pa-
tients receiving depression treatment (21), those receiving
outpatient mental health services (22), and those who have
received psychiatric discharge (23). However, no study has
examined rates for the entire VHA patient population. In-
deed, to date there have been no comprehensive assessments
of suicide mortality for all patients of any health system or
managed care organization.
Patterns of suicide risk across age groups may differ
among VHA patients compared with risks in the general
population. Older persons are conventionally understood
to have higher risks for suicide than younger persons (24).
However, among individuals receiving VHA treatment for
depression, Zivin et al. (21) found that suicide risks were
highest among younger patients (age, <45 years).
It is also important to consider gender differentials in
suicide mortality among patients receiving VHA care.
Women veterans are one of the fastest growing segments
of the population of veterans eligible for VHA care (25).
In the general population, men are 4 times as likely to die
from suicide, although women are 3 times as likely to at-
tempt suicide (15, 26). However, little is known about gen-
der differences in suicide risks among individuals receiving
VHA services.
This paper presents new information regarding suicide
mortality among all patients receiving VHA services. We
examined suicide mortality during fiscal year 2001 (October 1,
2000–September 30, 2001) among all individuals who
were alive at the start of fiscal year 2001 and who received
VHA services during either fiscal year 2000 or fiscal year
2001. We assessed suicide mortality overall and by age and
gender categories. VHA suicide rates were compared with
those of the general US population.
MATERIALS AND METHODS
Data sources
This study used data from the VHA’s National Patient Care
Database to identify all veterans with inpatient or outpatient
services utilization in any VHA facility during fiscal year
2000 or fiscal year 2001. Measures of vital status and cause
of death were based on information from the National Death
Index, which is considered the ‘‘gold standard’’ for mortality
assessment (27). It includes national data regarding dates and
causes of death for all US residents, derived from death cer-
tificates filed in state vital statistics offices. National Death
Index searches were performed for cohorts of VHA patients
who received any VHA services use during fiscal year 2000 or
fiscal year 2001 and had no subsequent use through June
2006. This cost-efficient method for conducting National
Death Index searches enables comprehensive assessment of
vital status and cause of death among all individuals who
received VHA services in fiscal year 2000 or fiscal year
2001. National Death Index search results often include mul-
tiple records that are potential matches. ‘‘True matches’’ were
identified on the basis of established procedures (28).
Measures
Patients’ age and gender were identified from administra-
tive files included in the National Patient Care Database.
The age at the start of fiscal year 2001 was categorized as
being 18–29, 30–39, 40–49, 50–59, 60–69, 70–79, or
80
years. Information regarding race/ethnicity was not consis-
tently available in the National Patient Care Database for all
VHA patients; therefore, it was not possible to calculate
race-adjusted rates.
Using National Death Index data, we identified dates and
causes of death. Suicide deaths were identified by using
International Classification of Diseases, Tenth Revision,
codes X60–X84, Y87.0, and U03 (29). We calculated the
total suicide deaths observed in fiscal year 2001 divided by
the person-years at risk for an observed suicide death. This
quotient was then multiplied by 100,000 to generate suicide
rates per 100,000 person-years.
Person-years at risk for an observed suicide death were
made operational as follows. For those veterans who received
VHA services in fiscal year 2000 (and survived that year), the
risk period began on the first day of fiscal year 2001. For
veterans with no VHA services in fiscal year 2000, the risk
period began on the day of their first VHA use in fiscal year
2001. Risk periods ended at either the date of death (from any
cause) or the end of the year, whichever came first.
Cohort
The VHA patient population at risk for observed suicide
deaths during fiscal year 2001 included all 4,692,034 indi-
viduals with VHA inpatient or outpatient services use in
fiscal year 2001 or fiscal year 2000 and who were alive at
the start of fiscal year 2001. Of these patients, 21,066 (0.4%)
were excluded because of missing or out-of-range values for
age or gender. Among the excluded patients, 2 suicides were
observed. The final cohort consisted of 4,670,968 patients.
This project was reviewed and approved by a VHA human
subjects committee.
Rates in the US general population
Suicide rates in the general population were identified
from the Web-based Injury Statistics Query and Reporting
Am J Epidemiol 2009;169:1033–1038
 Suicide Mortality Among VHA Health System Patients 1035

System (WISQARS) (26). These data are derived from the Table 1. Characteristics of US VHA Patients With Risk Time in
National Vital Statistics System. Suicide rates in the general Fiscal Year 2001a
population were assessed overall and by age and gender All Patients Female Male
categories. These are available by calendar year. Therefore,
rates in the general population were calculated for calendar No. 4,670,968 470,057 4,200,911
year 2001. % 100.0 10.1 89.9
Age, years
Analyses Mean (SD) 59.8 (15.7) 48.4 (16.4) 61.1 (15.1)
Median 61.1 46.6 63.1
Standardized mortality ratios for suicide risks among
Age distribution, %b
VHA patients were calculated for age and gender subgroups
compared with those in the general US population (30). We 18–29 4.3 13.9 3.3
calculated 95% confidence intervals for the standardized 30–39 7.5 18.7 6.2
mortality ratios using the exact method based on Poisson 40–49 15.0 26.3 13.8
distribution (31). 50–59 21.6 18.0 22.1
60–69 19.8 10.5 20.9
RESULTS 70–79 24.1 9.1 25.8

80 7.6 3.5 8.1
Table 1 presents characteristics of the VHA patients in-
Total 99.9 100.0 100.2
cluded in the risk cohort. The mean age was 59.8 years
(standard deviation, 15.7) at the start of fiscal year 2001. With VHA use
The mean number of risk days in fiscal year 2001 was 316.7 In fiscal year 2001 28.0 35.6 27.3
but not in fiscal
(standard deviation, 97.0). There were 1,613 deaths by year 2000, %
suicide.
In fiscal year 2000 17.6 33.1 15.9
Table 2 presents suicide rates among VHA patients com- but not in fiscal
pared with the general US population. Overall, for men and year 2001, %
women combined, suicide risks among VHA patients were Risk days, no.
66% higher than those observed in the general population Mean (SD) 316.7 (97.0) 320.6 (92.0) 316.3 (97.5)
(age- and gender-adjusted standardized mortality ratio ¼
Median 365.0 365.0 365.0
1.66, 95% CI: 1.58, 1.75). The crude suicide rate among
male VHA patients was 43.1/100,000 person-years, com- 25% quartile 342.0 346.0 342.0
pared with 23.2/100,000 person-years among males in the Total suicide 1,613 43 1,570
deaths, no.
general population. The standardized mortality ratio for
male VHA patients relative to males in the general popula- Abbreviation: SD, standard deviation; VHA, Veterans Health
tion, with adjustment for differences in age distributions, Administration.
a
was 1.66 (95% CI: 1.58, 1.74). That is, suicide risks among Patients who had some VHA services utilization in fiscal year
male patients were 66% greater than for males in the gen- 2000 or fiscal year 2001 and who were alive at the start of fiscal year
eral population. Analyses by age subgroups indicated that 2001.
b
male VHA patients between 30 and 79 years of age had Does not sum to 100% because of rounding.
greater suicide risks than did similarly aged men in the
general population. Standardized mortality ratios ranged
heightened when the government is also responsible for
from 1.33 among VHA patients aged 70–79 (95% CI:
their health care. This study presents new information re-
1.20, 1.47) years to 2.56 among patients aged 30–39
garding suicide mortality among the entire patient popula-
(95% CI: 2.13, 3.03) years.
tion of the VHA, representing approximately 5 million
Among female patients, the crude suicide rate was 10.41/
veterans. The VHA is the only large health-care organiza-
100,000 person-years, compared with 5.22/100,000 person-
tion to assess suicide mortality among its entire patient pop-
years in the general population, with a standardized mortal-
ulation. Thus, comparable data regarding suicide mortality
ity ratio of 1.87 (95% CI: 1.35, 2.47). Compared with
in other health system populations are not available—even
similarly aged women in the general population, female
though health systems are likely to be important venues for
patients had significantly higher risks: For those aged
suicide prevention efforts. This report from the nation’s
40–49 years, the standardized mortality ratio was 2.15
largest integrated health system provides a heretofore un-
(95% CI: 1.25, 3.29), and for those aged 50–59 years, it
available comparison point for other health systems that
was 2.36 (95% CI: 1.22, 3.88).
may be initiating surveillance of suicide mortality in their
populations. Study findings provide a baseline for evalua-
DISCUSSION tions of the most comprehensive suicide prevention program
ever put into place in a health-care system.
Government agencies and the public acknowledge a re- Suicide rates among VHA patients were significantly
sponsibility to monitor adverse outcomes among individuals higher than those in the general US population. However,
discharged from military service. This responsibility is the differentials between suicide rates for VHA patients and

Am J Epidemiol 2009;169:1033–1038
1036 McCarthy et al.

Table 2. Suicide Rates in Fiscal Year 2001 for US VHA Patients and in Calendar Year 2001 for the General US Population, by Gender and Age
Group

General US Populationa VA Patients

Standardized 95% Sex- and
General Crude Rate/ Person- Crude Rate/ Mortality Confidence Age-adjusted
Suicides, Suicides,
US Population, 100,000 Years 100,000 Ratio Interval Rate
no. no.
no. Person-Years at Riskb Person-Years

Males, years 23,850 102,872,290 23.18 1,570 3,639,873 43.13 1.66c 1.58, 1.74 38.46
18–29 4,830 23,983,737 20.14 31 111,234 27.87 1.38 0.94, 1.91
30–39 4,702 21,583,911 21.78 126 225,768 55.81 2.56c 2.13, 3.03
40–49 5,146 21,538,269 23.89 287 514,375 55.80 2.34c 2.07, 2.61
50–59 3,554 15,875,192 22.39 381 820,818 46.46 2.08c 1.87, 2.29
c
60–69 2,030 9,692,852 20.94 247 756,268 32.66 1.56 1.37, 1.76
70–79 2,072 6,961,055 29.77 366 924,752 39.58 1.33c 1.20, 1.47

80 1,516 3,237,274 46.83 132 287,458 45.92 0.98 0.82, 1.15
Females, years 5,730 109,719,004 5.22 43 412,873 10.41 1.87c 1.35, 2.47 9.76
18–29 782 22,972,302 3.40 1 52,972 1.89 0.55 0.01, 2.05
30–39 1,155 21,449,920 5.38 7 76,959 9.10 1.69 0.68, 3.15
40–49 1,573 22,024,321 7.14 17 110,730 15.35 2.15c 1.25, 3.29
50–59 1,113 16,725,057 6.65 12 76,254 15.74 2.36c 1.22, 3.88
60–69 514 10,945,634 4.70 3 44,241 6.78 1.44 0.30, 3.48
70–79 367 9,273,850 3.96 3 37,412 8.02 2.03 0.42, 4.88

80 226 6,327,920 3.57 0 14,306 0.00 0.00 NA
Total 29,580 212,591,294 13.91 1,613 4,052,746 39.80 1.66c 1.58, 1.75 23.15

Abbreviations: ICD-10, International Statistical Classification of Diseases and Related Health Problems, 10th Revision, 2nd ed; NA, not avail-
able; VA, Department of Veterans Affairs; VHA, Veterans Health Administration; WISQARS, Web-based Injury Statistics Query and Reporting
System.
a
Source: WISQARS, Centers for Disease Control and Prevention (http://www.cdc.gov/ncipc/wisqars/). Suicide identified by ICD-10 diagnosis
codes X60–X84, Y87.0, and U03 (29).
b
Person-years at risk for being observed with a suicide death during fiscal year 2001. Person-years of risk time are presented here as rounded
numbers.
c
Standardized mortality ratios where the confidence interval did not contain 1.0.

for the general US population were not as large as the differ-
entials previously estimated for veterans compared with
nonveterans in the community (8). Whereas Kaplan et al.
(8) reported that risks for male veterans were twice those
of male nonveterans in the community (adjusted hazard
ratio ¼ 2.04, 95% CI: 1.10, 3.80), we observed an overall
standardized mortality ratio of 1.66 (95% CI: 1.58, 1.74) for
male VHA patients compared with men in the general pop-
ulation. However, it is difficult to make direct comparisons
across studies because of differences in timeframes, popu-
lations, and methodologies. Given that VHA patients had
sought health services and likely had greater medical and
psychiatric morbidity than did veterans in the community,
one could expect higher relative risks in this treatment-
seeking patient population. Alternatively, treatment popula-
tions may have lower suicide risks as a function of increased
mental health and substance use care access and treatment.
This study provides new information on age-specific
differentials in suicide risks among VHA patients relative
to age- and gender-matched comparisons in the general
population. Among male patients, rates were highest
among those aged 30–49 years and lowest among patients
aged 18–29 and 60–69 years. The differential in suicide
rates for male patients, relative to men in the general pop-
ulation, was highest for patients aged 30–39 years, and it
fell among older patient groups. Further research is needed
to evaluate whether these patterns reflect cohort effects or
age-patterned selection of VHA services among veterans.
Among female patients, rates were highest among those
aged 50–59 years.
Consistent with the literature, crude rates among female
VHA patients were substantially lower than those among
male patients. However, compared with women in the gen-
eral population, female VHA patients had high suicide mor-
tality, with a standardized mortality ratio of 1.87. This is
a larger differential than the standardized mortality ratio of
1.66 for male patients.
Increased relative risks among female VHA patients may
result from not only the characteristics of treatment-seeking
populations generally (e.g., greater morbidity) but also the
factors unique to women who have served in the military.
Female VHA patients may have greater experience with and
access to firearms than do women in the general population,
and they may be more likely to have experienced traumatic
episodes, increasing risks for both post-traumatic stress dis-
order and suicide (32–37). In designing and implementing
suicide prevention initiatives, VHA providers must address
suicide risks among both male and female patients.

Am J Epidemiol 2009;169:1033–1038

Several issues should be considered when interpreting
study findings. First, there are longstanding concerns re-
garding potential undercounting of suicide deaths. Research
is needed to assess potential differential recording of suicide
deaths by veteran status. Second, although we compared
rates among VHA patients with those among the general
US population, we note that VHA patients are included in
the general population. Third, we note that the VHA patient
population includes some nonveterans (e.g., widows or de-
pendents of veterans), whose utilization accounts for ap-
proximately 1% of VHA inpatient admissions and 2% of
outpatient visits (38). This study examined suicide mortality
among all VHA patients, of whom the vast majority are
veterans. Finally, it was not possible to adjust rates for race.
In the general population, suicide rates are lower among
minority populations, and minorities may be overrepre-
sented among VHA patients.
Study findings indicate that VHA patients were at increased
risks for suicide, yet the differential in risks was less than what
might be expected given previous comparisons (8). Future re-
search is necessary to assess patients’ sociodemographic and
clinical factors that may be associated with suicide mortality,
to assess the impact of specific health-care services on suicide
risks among veterans, and to evaluate trends in suicide mor-
tality among VHA patients and among veterans who do not
receive VHA care. For health-care providers, and particularly
the Veterans Health Administration, research on suicide rates
and risk factors among veterans may directly inform treat-
ment and suicide prevention interventions.
ACKNOWLEDGMENTS
Author affiliations: US Department of Veterans Affairs
Serious Mental Illness Treatment Research and Evaluation
Center, Ann Arbor, Michigan (John F. McCarthy, Marcia
Valenstein, Mark Ilgen, Kara Zivin, Frederic C. Blow);
US Department of Veterans Affairs Center for Clinical
Management Research, Ann Arbor, Michigan (John F.
McCarthy, Marcia Valenstein, H. Myra Kim, Mark Ilgen,
Kara Zivin, Frederic C. Blow); and University of Michigan
Department of Psychiatry, Ann Arbor, Michigan (John F.
McCarthy, Marcia Valenstein, H. Myra Kim, Mark Ilgen,
Kara Zivin, Frederic C. Blow).
The authors thank Robert Bilgrad and Michelle Goodier
(National Center for Health Statistics) for assistance with
searching the National Death Index and Dr. Ira R. Katz
(Veterans Health Administration) and Dr. Katherine J. Hoggatt
(University of Michigan) for comments on an earlier version
of this manuscript.
Conflict of interest: none declared.
REFERENCES
1. Institute of Medicine. Reducing Suicide: A National Impera-
tive. Washington, DC: National Academies Press; 2004.
2. US Department of Health and Human Services. The Surgeon
General’s Call to Action to Prevent Suicide. Washington, DC:
National Institutes of Health; 1999.
Am J Epidemiol 2009;169:1033–1038
Suicide Mortality Among VHA Health System Patients 1037
3. Department of Veterans Affairs. The Comprehensive VHA
Mental Health Strategic Plan. Washington, DC: Department
of Veterans Affairs; 2005.
4. Office of Inspector General, Department of Veterans Affairs.
Healthcare inspection: implementing VHA’s Mental Health
Strategic Plan initiatives for suicide prevention. Report 06-
03706-126. Washington, DC: VA Office of Inspector General;
2007. (http://www.va.gov/oig/54/reports/VAOIG-06-03706-
126.pdf). (Accessed December 4, 2007).
5. Agha Z, Lofgren RP, VanRuiswyk JV, et al. Are patients at
Veterans Affairs medical centers sicker? A comparative analy-
sis of health status and medical resource use. Arch Intern Med.
2000;160(21):3252–3257.
6. Lambert MT, Fowler DR. Suicide risk factors among veterans:
risk management in the changing culture of the Department of
Veterans Affairs. J Ment Health Adm. 1997;24(3):350–358.
7. Hankin CS, Spiro A, Miller DR, et al. Mental disorders and
mental health treatment among US Department of Veterans
Affairs outpatients: the Veterans Health Study. Am J Psychi-
atry. 1999;156(12):1924–1930.
8. Kaplan MS, Huguet N, McFarland BH, et al. Suicide among
male veterans: a prospective population-based study. J Epi-
demiol Community Health. 2007;61(7):619–624.
9. The Joshua Omvig Veterans Suicide Prevention Act, Pub L
No. 110-110, 121 Stat 1031-1034. (http://npl.ly.gov.tw/pdf/
6173.pdf).
10. Kang HK, Hyams KC. Mental health care needs among recent
war veterans. N Engl J Med. 2005;352(13):1289.
11. Milliken CS, Auchterloni JL, Hoge CW. Longitudinal assess-
ment of mental health problems among active and reserve
component soldiers returning from the Iraq war. JAMA.
2007;298(18):2141–2148.
12. Seal KH, Bertentha D, Miner CR, et al. Bringing the war back
home: mental health disorders among 103,788 US veterans
returning from Iraq and Afghanistan seen at Department of
Veterans Affairs facilities. Arch Intern Med. 2007;167(5):
476–482.
13. Kang HK. Committee on Evaluation of the VA’s Presumptive
Disability Decision-Making Process. VA health care utiliza-
tion among OIF/OEF veterans. Washington, DC: Institute of
Medicine; 2006. (http://www.iom.edu/Object.File/Master/36/
181/Kang1.pdf). (Accessed November 17, 2007).
14. Department of Veterans Affairs. Statement of Ira Katz, MD,
Deputy Chief Patient Care Services Officer for Mental Health,
Department of Veterans Affairs, before House Committee on
Veterans Affairs July 25, 2007. Washington, DC: Department
of Veterans Affairs; 2007. (http://www.va.gov/OCA/
testimony/hvac/070725IK.asp). (Accessed December 4, 2007).
15. Office of the Inspector General, Department of Veterans Af-
fairs. Statement of Michael Shepherd, MD, Physician, Office of
Healthcare Inspections, Office of Inspector General, Depart-
ment of Veterans Affairs, before Special Committee on Aging,
United States Senate, on the Department of Veterans Affairs
implementation of suicide prevention initiatives from the
Mental Health Strategic Plan, October 3, 2007. Washington,
DC: Department of Veterans Affairs; 2007. (http://www.va.
gov/oig/pubs/VAOIG-statement-20071003-shepherd.pdf).
(Accessed December 6, 2007).
16. Boehmer TK, Flanders WD, McGeehin MA, et al. Postservice
mortality in Vietnam veterans: 30-year follow-up. Arch Intern
Med. 2004;164(17):1908–1916.
17. Thomas TL, Kang HK, Dalager NA. Mortality among women
Vietnam veterans, 1973–1987. Am J Epidemiol. 1991;134(9):
973–980.
1038 McCarthy et al.
18. Kang HK, Bullman TA. Mortality among US veterans of the
Persian Gulf War: 7-year follow-up. Am J Epidemiol. 2001;
154(5):399–405.
19. Bullman TA, Kang HK. The risk of suicide among wounded
Vietnam veterans. Am J Public Health. 1996;86(5):
662–667.
20. Bullman TA, Kang HK. Posttraumatic stress disorder and the
risk of traumatic deaths among Vietnam veterans. J Nerv Ment
Dis. 1994;182(11):604–610.
21. Zivin K, Kim HM, McCarthy JF, et al. Suicide mortality
among individuals receiving treatment for depression in the
Veterans Affairs Health System: associations with patient and
treatment setting characteristics. Am J Public Health. 2007;
97(12):2193–2198.
22. Desai MM, Rosenheck RA, Desai RA. Time trends and pre-
dictors of suicide among mental health outpatients in the
Department of Veterans Affairs. J Behav Health Serv Res.
2008;35(1):115–124.
23. Desai RA, Dausey DJ, Rosenheck RA. Mental health service
delivery and suicide risk: the role of individual patient and
facility factors. Am J Psychiatry. 2005;162(2):311–318.
24. National Institute of Mental Health. Older Adults Depression
and Suicide Facts. Washington, DC: National Institute of
Mental Health; 2003.
25. Meehan S. Improving health care for women veterans. J Gen
Intern Med. 2006;21(suppl 3):S1–S2.
26. National Center for Injury Prevention and Control. WISQARS
fatal injuries: mortality reports. Atlanta, GA: Centers for
Disease Control and Prevention; 2007. (http://webapp.cdc.
gov/sasweb/ncipc/mortrate.html). (Accessed November 26,
2007).
27. Cowper DC, Kubal JD, Maynard C, et al. A primer and com-
parative review of major US mortality databases. Ann Epide-
miol. 2002;12(7):462–468.
28. Sohn MW, Arnold N, Maynard C, et al. Accuracy and
completeness of mortality data in the Department of Veterans
Affairs [electronic article]. Popul Health Metr. 2006;4:2.
29. World Health Organization. International Statistical
Classification of Diseases and Related Health Problems. 10th
Revision. 2nd ed. (ICD-10). Geneva, Switzerland: World
Health Organization; 2004.
30. Hennekens CH, Buring JE. Epidemiology in Medicine.
Boston, MA: Little, Brown and Company; 1987:82–85.
31. Ulm K. A simple method to calculate the confidence interval
of a standardized mortality ratio. Am J Epidemiol. 1990;
131(2):373–375.
32. Fontana A, Schwartz LS, Rosenheck R. Posttraumatic stress
disorder among female Vietnam veterans: a causal model of
etiology. Am J Public Health. 1997;87(2):169–175.
33. Frayne SM, Seaver MR, Loveland S, et al. Burden of
medical illness in women with depression and post-
traumatic stress disorder. Arch Intern Med. 2004;164(12):
1306–1312.
34. Himmelfarb N, Yaeger D, Mintz J. Posttraumatic stress dis-
order in female veterans with military and civilian sexual
trauma. J Trauma Stress. 2006;19(6):837–846.
35. Kang H, Dalager N, Mahan C, et al. The role of sexual assault
on the risk of PTSD among Gulf War veterans. Ann Epidemiol.
2005;15(3):191–195.
36. Lehmann LL, McCormick RA, McCracken L. Suicidal be-
havior among patients in the VA Health Care System. Psy-
chiatr Serv. 1995;46(10):1069–1071.
37. Qin P, Agerbo E, Mortensen PB. Suicide risk in relation to
socioeconomic, demographic, psychiatric, and familial
factors: a national register-based study of all suicides in
Denmark, 1981–1997. Am J Psychiatry. 2003;160(4):
765–772.
38. King S, Shane A, Smith MW. A Guide to Identifying Non-
Veteran Records in the Inpatient and Outpatient Databases.
(Technical report #20). Menlo Park, CA: VHA Health
Economics Resource Center (HERC); 2006.
Am J Epidemiol 2009;169:1033–1038
 American Journal of Epidemiology
ª The Author 2009. Published by the Johns Hopkins Bloomberg School of Public Health.
All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org.
Book Reviews
Hyping Health Risks: Environmental Hazards in Daily Life and the
Science of Epidemiology
By Geoffrey C. Kabat
ISBN: 978-0-231-14148-2, Columbia University Press, Irvington, New York (Telephone: 914-591-9111, Fax: 1-800-944-
1844, E-mail: cup_book@columbia.edu, World Wide Web: http://cup.columbia.edu), 2008, 272 pp., $27.95 Hardcover
For a scholarly book in epidemiology, Hyping Health
Risks by Geoffrey C. Kabat is more reflective and opinion-
ated than most and even has somewhat of a plot. Few books
about epidemiology generate a visceral in addition to
a solely cerebral response. On reading the book, I found
passages causing offense and others eliciting support for
Kabat’s thoughtful and articulate laments. For the most part,
the story of truth and misrepresentation of evidence on
health risks was engaging, including personal portraits of
good and evil intent. However, I did not always agree with
the author on who belonged to which category and often
found myself thinking, ‘‘that’s not quite the whole story;
it’s more complicated than that,’’ or ‘‘I can see why you’d
interpret it that way, but I don’t.’’ Perhaps, being provocative
in that way is a real credit to the author.
Kabat has written a fairly succinct treatise (272 pages, but
small print) making the case that environmental health con-
cerns have been given undue attention and too much funding
and have been subjected to insufficient critical evaluation
for a variety of reasons that run counter to good science. The
bulk of the book consists of 4 lengthy, self-contained
vignettes on the pursuit of environmental contaminants as
a cause of breast cancer and the evaluation and interpreta-
tion of health effects of exposures to electromagnetic fields,
indoor radon, and environmental tobacco smoke. The exag-
gerations by advocates to generate support for these issues
and claimed censoring of critics are portrayed as the product
of self-serving, funding-hungry researchers; well-intentioned,
but misguided political activists; exploiters from the media
or real politics; and, more broadly, societal tenacity in pur-
suing the wrong things in the wrong way rather than the
right things in the right way. While the author does allude
in passing to the countervailing arguments—that remote
possibilities of harm from ubiquitous environmental agents
need careful scrutiny, that the public pays the bills and has
some right to influence funding priorities, that researchers
believe they are serving public health—these modest at-
tempts at balance fall short of a comprehensive apprecia-
tion of the complexity and ambiguity of the issues under
discussion.
The stories of the rise and fall of the controversies are
very nicely written, capturing a time in our recent history
that the author (and many of the potential readers of the
book) lived through. The dynamics and personalities come
across vividly, with clarity and accuracy regarding the tech-
1039
Vol. 169, No. 8
nical issues behind the controversies and a thoughtful, if
highly subjective, answer to the question, What happened?
The distance of at most a couple of decades and the benefits
of hindsight regarding the science and politics allow the
author to tell a story with a beginning, middle, and end.
Though I suspect that Dr. Kabat will be flattered and
offended by the comparison, it reminds me very much of
Paul Brodeur, who wrote extensively with great influ-
ence for the New Yorker, and, to some extent, Gary Taubes,
who also weaves science and personality into a compelling
story.
The book provokes reflection and reexamination of our
roles and motives as epidemiologists (my direct experience
concerned the electromagnetic field saga, for which I was
interviewed by Dr. Kabat) and, more broadly, of the inter-
section of epidemiology with social, political, and ideolog-
ical issues. The urge to reflect on our place in society has
tended to be one-sided, coming largely from the political
left, focusing on the moral obligation to use our science to
promote social justice and encouraging us to be activists in
order to change and not just study the world. This book
serves as a powerful counterbalance in reminding us that
some of the causes for which epidemiologists and their ad-
vocates join forces may not be grounded in good science and
therefore not likely to provide public health benefit. To be
useful in advancing public health, epidemiologic findings
have to be valid; when that is not the case, then the enthu-
siastic support of the public and regulatory agencies can be
harmful to public health by diverting resources and attention
from more important issues. This book forcefully examines
that question—What goes wrong when the good intentions
of scientists and activists are based on weak epidemiologic
findings?
The specific ‘‘problem’’ varies across the 4 topics consid-
ered. Regarding the first one, breast cancer activists gener-
ated interest from politicians and epidemiologists about the
potential role of environmental pollutants in the causation of
elevated breast cancer rates, particularly on Long Island,
New York. In fact, there was no elevation beyond that ex-
pected based on the demographics of the affected commu-
nities. In the case of health effects of electromagnetic fields,
epidemiologists persisted despite the forceful arguments
from physicists that biologic or health effects were implau-
sible. Indoor radon and environmental tobacco smoke are
both characterized by Kabat as real, but small hazards that
Am J Epidemiol 2009;169:1039–1042
1040 Book Reviews
were exaggerated by regulatory agencies (particularly the
US Environmental Protection Agency) and the antitobacco
lobby, respectively. In fact, the National Academy of Sci-
ences estimated that more than 10% of lung cancer deaths
in the United States are attributable to indoor radon, but the
author notes (correctly) that most of these deaths occur
among smokers, such that if smoking were eliminated, the
effect of radon would be markedly reduced. In the case of
both indoor radon and environmental tobacco smoke, the
author expresses deep concern about the distraction from
addressing active smoking, the real culprit in his view. Also
common to the latter 2 issues is the contention that certainty
of evidence and the magnitude of possible effects were ex-
aggerated for political reasons. In the case of environmental
tobacco smoke, there is a perceived conspiracy to suppress
criticism of the research or those who generate contradictory
data regarding the impact of environmental tobacco smoke
on risk of lung cancer, unfairly lumping such researchers
with the long lineage of tobacco industry shills.
To the best of my knowledge, the author was not incor-
rect about any of the important technical points, yet the
way he bundled the facts and drew broader inferences was
incomplete and oversimplified to make his point clearer. To
suggest that topics such as environmental contributors to
breast cancer or lung cancer risks associated with indoor
radon should not have been addressed at all or as persis-
tently as they were, or that incomplete data should not have
been used to promote regulatory action on radon and en-
vironmental tobacco smoke, fails to capture the complexity
of policy decisions. Even within the purely scientific arena,
how does one combine limited direct epidemiologic evi-
dence on low-level indoor radon with compelling evidence
from high exposures among uranium miners and a strong
theoretical and empirical basis for extrapolation to the lev-
els encountered in residences? Is the combination suffi-
cient for drawing firm conclusions about indoor radon
even with more uncertain evidence directly from the stud-
ies of indoor radon?
Determining which topics deserve initial and continuing
attention and funding is ultimately a policy decision, a blend
of science—considering a range of disciplinary perspectives—
and political and societal concerns. Early in the evolution of
new topics with environmental relevance, there is often
a set-aside of research funding (generated by advocates),
after which the scientific merit and promise either moves
the topic forward or allows it to atrophy. Evidence on envi-
ronmental pollutants as a major contributor to breast cancer
has not taken hold, and any proposed new studies on that
topic need to stand or fall based on their merits, not because
there continues to be advocacy for pursuing this possibility.
Similarly, electromagnetic fields continue to be of limited
interest to epidemiologists after a period of extensive fund-
ing, but those advocating new proposed research must ac-
knowledge the findings and shortcomings of the studies that
have come before and make the case that the new studies
will be better. The question of when the funding and atten-
tion to topics exceed their objective, public health merit is
important, but we need to be wary of a stopping rule that
precludes low-plausibility topics from being investigated at
all, particularly when there are forceful advocates for having
a look. Topics not considered in the book have taken other
courses, such as the evolution of information on particulate
air pollution at levels many once thought too low to be of
biologic interest. In hindsight, of course we would do well to
invest exclusively in topics that will improve public health,
but those cannot be predicted in advance. The question of
how much to invest in novel, unexplored areas versus those
of proven public health importance is complex. Some effort
needs to be devoted to topics other than those representing
compelling global epidemics—tobacco, HIV, and obesity,
for example.
I completely concur with Dr. Kabat in his effort to stim-
ulate more explicit examination of the issues involved in the
evolution of these topics, assessing how the perspectives of
epidemiologists blend and conflict with those of other sci-
entists and how the views of scientists are integrated by
advocates who call for funding and use research findings
for their particular purposes. We should be more aware of
the epidemiologist’s important but circumscribed role in
setting research and regulatory policy. Through such under-
standing, epidemiologists could become more effective con-
tributors to the broader debates and more accepting when
our personal perspective on the epidemiology, or even the
consensus of epidemiologic thinking, does not carry the
day. If the argument of activists overrides that of the epi-
demiologists, or the empirical epidemiologic evidence is
more persuasive than the theoretical arguments of physi-
cists or toxicologists (or vice versa), or public health acti-
vists override the views of epidemiologists, it is not
necessarily an injustice. In the policy arena, where funding
and regulatory decisions are made, it is presumptuous to
suggest that in the face of ambiguity (common to each of
the issues in the book), the current evidence from epide-
miology (or physics, or toxicology, or molecular biology)
should always prevail.
Those of us who have chosen to operate in the realm of
research value the relative clarity of the terms of battle,
drawing on the tools of evidence, new studies, and refine-
ment of hypotheses. Science necessarily isolates and ab-
stracts from the broader world, and, within the realm of
science, epidemiology represents a further narrowing of
the frame of reference. Narrower still is the direct evidence
on a specific question, for example, indoor radon and lung
cancer, versus the broader epidemiologic evidence on ra-
don and lung cancer generally. By those standards, intru-
sions and distractions from scientific principles are
harmful. On the other hand, if suggestive evidence regard-
ing environmental tobacco smoke and lung cancer is a valu-
able tool for antismoking activists and those who generate
scientific challenges weaken their efforts, the critics of that
evidence are naturally seen as acting counter to public
health even if the criticisms are scientifically valid. The
ethical issues in slanting or selectively attending to the
science to make wise policy are applicable more broadly
than to epidemiology.
Reading and reflecting on the thesis of this book can only
help epidemiologists be more aware of our place in society
and thus be more effective contributors as we venture be-
yond the technical aspects of epidemiology into the broader,
messier world. Many of us recognize that regulatory policy
Am J Epidemiol 2009;169:1039–1042

requires a framework encompassing considerations other
than epidemiologic evidence, and this book helps to extend
that perspective to include research funding policy. Kabat
identifies real issues that need closer examination and more
open debate than has take place up to now. In the bipartisan
spirit of the era, there is a need to examine the points of
contention (between advocates and scientists, epidemiolo-
gists and physicists, regulators and researchers) to find com-
mon ground and enable good science and sound policy to
move forward, if not always hand in hand.
Am J Epidemiol 2009;169:1039–1042
Book Reviews 1041
ACKNOWLEDGMENTS
Conflict of interest: none declared.
David A. Savitz (e-mail: david.savitz@mssm.edu)
Mount Sinai School of Medicine, Center of Excellence in
Epidemiology, Biostatistics, and Disease Prevention,
New York, NY 10029-6574
DOI: 10.1093/aje/kwp013; Advance Access publication February 10, 2009

Concepts of Epidemiology: Integrating the Ideas, Theories, Principles and
Methods of Epidemiology, 2nd Edition
By Raj Bhopal
ISBN-10: 0-1995-4314-3, ISBN-13: 978-0-1995-4314-4, Oxford University Press, Inc., New York, New York (Telephone:
212-726-6000, Website: http://www.oup.com/us/), 2008, 472 pp., $55.95 Softcover
Several years ago, as the Department of Epidemiology of
the Johns Hopkins Bloomberg School of Public Health re-
vised its core methods courses, the faculty scanned then
available texts, seeking to find books that could be used over
the 1-year sequence. We found an unexpectedly large num-
ber of introductory texts, including the first edition of Raj
Bhopal’s Concepts of Epidemiology (1); of these, most were
intended for a broad introductory course and for the diverse
student groups that might take such courses. Bhopal had
thoughtfully reviewed 25 such introductory texts in 1997
(2), and 2 years later he commented on the role that such
books play in defining the conceptual domain of the field
(3). His review of these books, along with extensive teach-
ing experience, contributed to his writing an introductory
text that differed from many long-used standards.
The first edition of Concepts of Epidemiology, befitting its
title, emphasized the conceptual basis of epidemiology and its
links to population health. It was intended primarily for the
broad audience of postgraduate students throughout the world
who take an introductory course. The second edition (4) builds
on the foundation set by the first, with some expansion and the
addition of questions and answers at the end of each chapter.
The didactic approach of this edition is a strength. The book
has numerous embedded exercises that challenge readers to
consider the implications and applications of the concepts that
are discussed. Case studies are abundant, and the questions
and answers that end the chapters should be particularly useful
for those using the book outside of a course setting. Figures
have been used creatively to explain key concepts.
In its organization, the book follows a structure distinct
from that of most texts, which generally begin with broad
concepts and the indicators of population health, followed
by chapters on various study designs and measures of asso-
ciation, and ending with discussions of bias, causal infer-
ence, and application. Concepts of Epidemiology, by
Am J Epidemiol 2009;169:1039–1042
Book Reviews 1041
contrast, begins after a framing introductory chapter with
consideration of the ‘‘epidemiological concept of popula-
tion’’ (4, p. 17) followed by a chapter on disease variation by
time, place, and person that does not introduce either in-
cidence and mortality rates or prevalence. These measures
are covered in the seventh chapter. The book’s fourth chap-
ter covers bias and effect modification, while the fifth ad-
dresses causation; this sequence appears out of order, as
causal association is the target of much epidemiologic
research and bias is an unwanted source of association.
Additionally, these 2 chapters come before study design
and measures of association that are reviewed in chapters
8 and 9, respectively. The concluding 10th chapter offers
a perspective on the future of epidemiology, considers ethics
and practice, and provides historical vignettes.
We also faced the challenge of how best to order the pre-
sentation of concepts and methods when we revised the core
sequence of epidemiology courses at Johns Hopkins a few
years ago. There are different ways to structure an introduc-
tory course: The course can be anchored on association and
causation and on ‘‘descriptive’’ and ‘‘analytical’’ methods,
for example. However, the rationale for the sequencing in
Concepts of Epidemiology is not transparent and leads to
potential difficulties in using the book for a course. For ex-
ample, variation in disease by person, place, and time is
covered without illustration by incidence or mortality rates.
Causation and bias are considered prior to the introduction of
measures of association. Bhopal’s intent is likely to ensure
that readers give emphasis to concept rather than calculations,
but in my view a more integrated approach would be prefer-
able. In the Johns Hopkins’ courses, we settled on concepts
and models of causation as the starting point, followed by
measures of occurrence, risk, and rates per person-time.
Measures of association and study designs followed. Only
then were bias and effect modification covered. The sequence
1042 Book Reviews
for the Hopkins courses is still unsettled as opportunities for
improvement have been identified with each iteration.
Beyond approach and organization, the utility of an intro-
ductory text is determined by its content. Bhopal’s extensive
experience in population health is an asset to the book, and he
repeatedly emphasizes the linkages of epidemiology to pop-
ulation health and draws in examples from his work. The
range of topics is broad, updated, and appropriate for an in-
troductory text. On the other hand, I found laxity in the def-
inition of some key concepts and measures, perhaps reflecting
a tendency to largely offer descriptions, rather than sharp
definitions or formulae. For example, person-time is not ex-
plicitly defined nor sufficiently covered, and a lengthy discus-
sion still does not clarify the relation between incidence rate
and cumulative incidence. I have found the same deficiency in
other introductory texts. Selection bias is vaguely defined as
follows: ‘‘Bias can result from the choice of populations to be
studied. This is known as selection bias’’ (4, p. 90). The sub-
sequent discussion emphasizes external validity rather than
distortion of measures of association as a consequence of
selection processes. The Dictionary of Epidemiology (5)
would have been a useful resource for ensuring that definitions
were better aligned with the most up-to-date formulations.
Chapter 9, covering study designs, introduces the concept of
the base population but then deviates from offering a unifying
formulation of study design; the various major designs are
covered as though not related, even though the chapter’s title
spoke to ‘‘an integrated suite of methods.’’ These are only
some of the examples of the book’s content that caught my
attention; of course, opinions vary among epidemiologists on
the field’s methods, and others may find Bhopal’s presentation
to be better aligned with their views.
The bottom line for any review of an introductory text is
whether the book can be recommended for use in an intro-
ductory course. Concepts of Epidemiology is sufficiently
distinct in its approach that anyone considering its use
should read it closely to determine if it will fit the needs
of his/her particular course and students. I am certain that,
when used by Dr. Bhopal, it works well.
ACKNOWLEDGMENTS
Conflict of interest: none declared.
REFERENCES
1. Bhopal R. Concepts of Epidemiology: An Integrated
Introduction to the Ideas, Theories, Principles and Methods of
Epidemiology. New York, NY: Oxford University Press, Inc,
2002.
2. Bhopal R. Which book? A comparative review of 25
introductory epidemiology textbooks. J Epidemiol Community
Health. 1997;51(6):612–622.
3. Bhopal R. Paradigms in epidemiology textbooks: in the
footsteps of Thomas Kuhn. Am J Public Health. 1999;89(8):
1162–1165.
4. Bhopal R. Concepts of Epidemiology: Integrating the Ideas,
Theories, Principles and Methods of Epidemiology. Second
Edition. New York, NY: Oxford University Press, Inc, 2008.
5. Last JM, ed. A Dictionary of Epidemiology. 4th ed. New York,
NY: Oxford University Press, Inc, 2001.
Jonathan M. Samet (e-mail: jsamet@usc.edu)
Department of Preventive Medicine and Institute for Global
Health, Keck School of Medicine, University of Southern
California, Los Angeles, CA 90089
DOI: 10.1093/aje/kwp009; Advance Access publication February 10, 2009
Am J Epidemiol 2009;169:1039–1042