Вы находитесь на странице: 1из 130

AMERICAN JOURNAL OF EPIDEMIOLOGY

issn 0002-9262 (print)


ISSN 1476-6256 (online)
printed in the u.s.a.

American Journal of
Associate Director for Epidemiology and Genetics Research Program
Division of Cancer Control and Population Sciences
Epidemiology
The National Cancer Institute (NCI), a major research component of the National Institutes of Health (NIH) Volume 169 Number 8 April 15, 2009
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 www.aje.oxfordjournals.org
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.
ORIGINAL CONTRIBUTIONS
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
919 Parkinson’s Disease and Residential Exposure to Maneb and Paraquat From
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 Agricultural Applications in the Central Valley of California. Sadie Costello, Myles Cockburn,
on cancer epidemiology and other population sciences that informs development of creative and cost-effective Jeff Bronstein, Xinbo Zhang, and Beate Ritz
strategies to advance science is essential. 927 Overweight and Obesity Over the Adult Life Course and Incident Mobility Limitation in
The Associate Director will lead the Epidemiology and Genetics Research Program (EGRP), which includes its Older Adults: The Health, Aging and Body Composition Study. Denise K. Houston, Jingzhong
Office of the Associate Director and four branches: Clinical and Translational Epidemiology, Host Susceptibility Ding, Barbara J. Nicklas, Tamara B. Harris, Jung Sun Lee, Michael C. Nevitt, Susan M. Rubin, Frances
Factors, Methods and Technologies, and Modifiable Risk Factors. He/she provides scientific and administrative A. Tylavsky, and Stephen B. Kritchevsky for the Health ABC Study
leadership for the Program, supervises the staff, and represents NCI to a wide variety of professional, academic, 937 Association of Diabetes With Prostate Cancer Risk in the Multiethnic Cohort. Kevin
and advocacy organizations. The Associate Director also develops and facilitates collaborations with funders of M. Waters, Brian E. Henderson, Daniel O. Stram, Peggy Wan, Laurence N. Kolonel, and Christopher
other types of population science, including the NIH Institutes and Centers, Centers for Disease Control and
A. Haiman
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 370 946 Modification of the Effect of Vitamin E Supplementation on the Mortality of Male
research grants and numerous interagency agreements and consortia. Smokers by Age and Dietary Vitamin C. Harri Hemilä and Jaakko Kaprio

Volume 169
Qualifications: The successful applicant will be an experienced epidemiologist (M.D. or Ph.D.-level training re- 954 Dietary Acrylamide Intake and Risk of Premenopausal Breast Cancer. Kathryn M. Wilson,
quired) with leadership experience, excellent communication skills, and a strong record of peer-reviewed publica- Lorelei A. Mucci, Eunyoung Cho, David J. Hunter, Wendy Y. Chen, and Walter C. Willett
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, 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,

Number 8
scientific staff, training programs, interdisciplinary collaborations, and/or funded programs is highly valued.
Leslie Bernstein, Joan Largent, Marinela Capanu, Colin B Begg, Lene Mellemkjær, Charles F. Lynch,
Salary: This position is an excepted service position (Title 42) with a salary range of $160,000-$195,000. Kathleen E. Malone, Anne S. Reiner, Xiaolin Liang, Robert W. Haile, John D. Boice, Jr., WECARE Study
Collaborative Group, and Jonine L. Bernstein.
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: 969 Positive Associations Between Ionizing Radiation and Lymphoma Mortality Among
Men. David B. Richardson, Hiromi Sugiyama, Steve Wing, Ritsu Sakata, Eric Grant, Yukiko Shimizu,
Rachel Ballard-Barbash, M.D., M.P.H., Associate Director Nobuo Nishi, Susan Geyer, Midori Soda, Akihiko Suyama, Fumiyoshi Kasagi, and Kazunori Kodama
Applied Research Program
Division of Cancer Control and Population Sciences 977 Biomarker-calibrated Energy and Protein Consumption and Increased Cancer Risk
National Cancer Institute Among Postmenopausal Women. Ross L. Prentice, Pamela A. Shaw, Sheila A. Bingham, Shirley
Pages 919–1042

6130 Executive Blvd, Room 4005, MSC 7344 A. A. Beresford, Bette Caan, Marian L. Neuhouser, Ruth E. Patterson, Marcia L. Stefanick, Suzanne
Bethesda, MD 20892-7344 Satterfield, Cynthia A. Thomson, Linda Snetselaar, Asha Thomas, and Lesley F. Tinker
Express Mail: Rockville, MD 20852

Continued on Inside Front Cover


For more information about DCCPS and EGRP, see http://cancercontrol.cancer.gov

THE DHHS/NIH/NCI ARE EQUAL OPPORTUNITY EMPLOYERS


Published for the Johns Hopkins Bloomberg School of Public Health by Oxford
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 MARLENE B. GOLDMAN, Lebanon, NH ROBERT C. MILLIKAN, Chapel Hill, NC
ALBERTO ASCHERIO, Boston, MA ELISEO GUALLAR, Baltimore, MD LORENE M. NELSON, Stanford, CA
BRAD ASTOR, Baltimore, MD M. ELIZABETH HALLORAN, Seattle, WA ROBERTA B. NESS, Houston, TX
DONNA DAY BAIRD, Research Triangle BERNARD L. HARLOW, Minneapolis, MN CRAIG J. NEWSCHAFFER, Philadelphia, PA
Park, NC PATRICIA HARTGE, Rockville, MD THOMAS R. O’BRIEN, Bethesda, MD
OLGA BASSO, Research Triangle Park, NC MAUREEN C. HATCH, Bethesda, MD ANDREW OLSHAN, Chapel Hill, NC
EDWARD J. BOYKO, Seattle, WA RICHARD B. HAYES, New York, NY NANCY PADIAN, San Francisco, CA
ROBERT F. BREIMAN, Atlanta, GA MIGUEL A. HERNÁN, Boston, MA JULIE R. PALMER, Brookline, MA
GERMAINE M. BUCK, Rockville, MD LISA HERRINTON, Oakland, CA JULIE PARSONNET, Stanford, CA
KENNETH P. CANTOR, Bethesda, MD IRVA HERTZ-PICCIOTTO, Davis, CA WENDY S. POST, Baltimore, MD
WONG-HO CHOW, Bethesda, MD DONALD HOOVER, Piscataway, NJ NANCY POTISCHMAN, Bethesda, MD
JOHN D. CLEMENS, Bethesda, MD JANE A. HOPPIN, Research Triangle Park, NC ERIC B. RIMM, Boston, MA
STEPHEN COLE, Chapel Hill, NC DAVID J. HUNTER, Boston, MA HARVEY A. RISCH, New Haven, CT
GLINDA S. COOPER, Research Triangle Park, NC JOUNI JAAKKOLA, Birmingham, United Kingdom LESLIE ROBISON, Minneapolis, MN
JOSEF CORESH, Baltimore, MD STEVEN J. JACOBSEN, Pasadena, CA KARIN ROSENBLATT, Champaign, IL
STEVEN S. COUGHLIN, Washington, DC KONRAD JAMROZIK, Nedlands, Australia AUDREY F. SAFTLAS, Iowa City, IA
ROSA M. CRUM, Baltimore, MD STANISLAV V. KASL, New Haven, CT GLEN A. SATTEN, Atlanta, GA
RALPH B. D’AGOSTINO, JR., Winston-Salem, NC FREYA KAMEL, Research Triangle Park, NC ENRIQUE F. SCHISTERMAN, Bethesda, MD
RALPH B. D’AGOSTINO, SR., Boston, MA JAY S. KAUFMAN, Montreal, Canada THOMAS A. SELLERS, Tampa, FL
CONSTANTINE DASKALAKIS, Philadelphia, PA WILLIAM C. KNOWLER, Phoenix, AZ EYAL SHAHAR, Tucson, AZ
ANA V. DIEZ-ROUX, Ann Arbor, MI ALAN R. KRISTAL, Seattle, WA A. RICHEY SHARRETT, Baltimore, MD
BRENDA ESKENAZI, Berkeley, CA MICHAEL F. LEITZMANN, Regensburg, Germany GARY SHAW, Oakland, CA
GUY D. ESLICK, Boston, MA DE-KUN LI, Oakland, CA COLIN L. SOSKOLNE, Edmonton, Canada
EDUARDO FAERSTEIN, Rio de Janeiro, Brazil RONGLING LI, Memphis, TN MEIR STAMPFER, Boston, MA
KATHERINE M. FLEGAL, Hyattsville, MD MARC LIPSITCH, Boston, MA RACHAEL STOLZENBERG-SOLOMON,
BETSY FOXMAN, Ann Arbor, MI JULIAN LITTLE, Aberdeen, United Kingdom Rockville, MD
LIN FRITSCHI, Crawley, Australia STEPHANIE LONDON, Research Triangle SHOLOM WACHOLDER, Bethesda, MD
SANDRO GALEA, Ann Arbor, MI Park, NC DOUGLAS L. WEED, Bethesda, MD
MARILIE D. GAMMON, Chapel Hill, NC MATTHEW P. LONGNECKER, Research Triangle SCOTT T. WEISS, Boston, MA
PETER J. GERGEN, Rockville, MD Park, NC PHYLLIS A. WINGO, Atlanta, GA
RICHARD F. GILLUM, Hyattsville, MD GERALD MCGWIN, JR., Birmingham, AL JUN ZHANG, Bethesda, MD
EDWARD GIOVANNUCCI, Boston, MA JOSEPH K. MCLAUGHLIN, Rockville, MD WEI ZHENG, Nashville, TN
American Journal of Epidemiology
The American Journal of Epidemiology (ISSN 0002-9262) is published please contact: Journals Customer Services, Oxford University Press,
twice a month, 24 times a year in 2 volumes (12 issues per volume), by 4-5-10-8F Shiba, Minato-ku, Tokyo, 108-8386, Japan. Telephone: +81 3
Oxford University Press, 2001 Evans Road, Cary, NC 27513-2009, for the 5444 5858, Fax: +81 3 3454 2929, e-mail: custserv.jp@oxfordjournals.org.
Johns Hopkins Bloomberg School of Public Health and is sponsored by the Advertising: Advertising, inserts, and artwork enquiries should be
Society for Epidemiologic Research. Information for authors appears on the addressed to Advertising and Special Sales, Oxford Journals, Oxford
Journal’s home page at www.aje.oxfordjournals.org. Authors should consult University Press, Great Clarendon Street, Oxford OX2 6DP, UK.
these instructions before submitting manuscripts. Telephone: +44 (0) 1865 353329. Fax: +44 (0) 1865 353774. E-mail:
Oxford University Press is a department of the University of Oxford. It jnlsadvertising@oxfordjournals.org.
furthers the University’s objective of excellence in research, scholarship, Digital Object Identifiers: For information on dois and how to resolve
and education by publishing worldwide. them, please visit www.doi.org.
Subscriptions: A subscription to the American Journal of Epidemiology Oxford Journals Environmental and Ethical Policies: Oxford
comprises 24 issues. Prices include postage; for subscribers outside the Journals is committed to working with the global community to bring the
Americas, issues are sent air freight. highest quality research to the widest possible audience. Oxford Journals
Annual subscription rate (Volumes 169 and 170, 24 issues, 2009): will protect the environment by implementing environmentally friendly
Institutional: Print and online $762/£508/€762, Online only $724/£483/ policies and practices wherever possible. Please see http://www.
€724, and Print only $724/£483/€724; Personal: (Combination: the Journal oxfordjournals.org/ethicalpolicies.html for further information on Oxford
and Epidemiologic Reviews) Online only $355/£237/€355, and Print only Journals’ environmental and ethical policies.
$373/£249/€373). Please note: UK£ rate applies to UK and Rest of World, Requests for permissions, reprints, and photocopies: All rights
except US ($) and Europe (€). There may be other subscription rates available. reserved; no part of this publication may be reproduced, stored in a retrieval
For a complete listing please see www.oxfordjournals.org/our_journals/aje/ system, or transmitted in any form or by any means (electronic, mechanical,
access_purchase/price_list.html. photocopying, recording, or otherwise) without either prior written
Full pre-payment in the correct currency is required for all orders. permission of the publisher (Oxford University Press, Rights and Business
Payment should be in US dollars for orders being delivered to the USA or Development Journals Division, Great Clarendon St., Oxford OX2 6DP,
Canada; Euros for orders being delivered within Europe (excluding the UK. Tel. +44 1865 354490/353695; Fax: +44 1865 353485; E-mail:
UK); GBP sterling for orders being delivered elsewhere (i.e. not being journals.permissions@oxfordjournals.org; Web: www.oxfordjournals.org/
delivered to USA, Canada, or Europe). All orders should be accompanied jnls/permissions) or a license permitting restricted copying, issued in the
by full payment and sent to your nearest Oxford Journals office. US by the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA
Subscriptions are accepted for complete volumes only. Orders are regarded 01923 (Fax: 978-750-4470) or in the UK by the Copyright Licensing
as firm, and payments are not refundable. Our prices include Standard Air Agency Ltd., 90 Tottenham Court Road, London W1P 9HE, UK. Reprints
as postage outside of the UK. Claims must be notified within four months of individual articles are available only from the authors.
of despatch/order date (whichever is later). Subscriptions in the EEC may It is a condition of publication in the Journal that authors grant an
be subject to European VAT. If registered, please supply details to avoid exclusive license to the Johns Hopkins Bloomberg School of Public Health.
unnecessary charges. For subscriptions that include online versions, a pro- However, requests for permission to reprint material found in the Journal
portion of the subscription price may be subject to UK VAT. Subscribers in should be submitted to Oxford University Press. This ensures that requests
Canada, please add GST to the prices quoted. from third parties to reproduce articles are handled efficiently and consis-
Personal rate subscriptions are only available if payment is made by per- tently and will also allow the article to be disseminated as widely as possi-
sonal cheque or credit card, delivery is to a private address, and is for personal ble. As part of the license agreement, authors may use their own material in
use only. other publications provided that the Journal is acknowledged as the origi-
Methods of payment: (i) Check (payable to Oxford University Press, nal place of publication and Oxford University Press is notified in writing
mailed to Oxford University Press, Cashiers Office, Great Clarendon in advance.
Street, Oxford OX2 6DP, UK) in GB£ Sterling (drawn on a UK bank), US$ The author(s) of each article appearing in this Journal is/are solely
Dollars (drawn on a US bank), or EU€ Euros. (ii) Bank transfer to Barclays responsible for the content thereof; the publication of an article shall not
Bank Plc, Oxford Group Office, Oxford (bank sort code 20-65-18) (UK), constitute or be deemed to constitute any representation by the Editors, the
overseas only Swift code BARC GB 22 (GB£ Sterling to account no. Johns Hopkins Bloomberg School of Public Health, or the Society for
70299332, IBAN GB89BARC20651870299332; US$ Dollars to account Epidemiologic Research that the data presented therein are correct or suffi-
no. 66014600, IBAN GB27BARC20651866014600; EU€ Euros to account cient to support the conclusions reached or that the experiment design or
no. 78923655, IBAN GB16BARC20651878923655). (iii) Credit card methodology is adequate.
(Mastercard, Visa, Switch or American Express). The American Journal of Epidemiology is printed on acid-free paper
Back issues: The current year and two previous years’ issues are available that meets the minimum requirements of ANSI Standard Z39.48-1984
from Oxford University Press. Previous volumes can be obtained online (Permanence of Paper).
at http://www.periodicals.com/oxford.html or from the Periodicals Service The Journal is indexed or abstracted in Index Medicus, the Science
Company, 11 Main Street, Germantown, NY 12526, USA. E-mail: Citation Index, Current Contents: Life Sciences, and the BIOSIS database.
psc@periodicals.com. Tel: (518) 537-4700. Fax: (518) 537-5899. Copyright © 2009 by the Johns Hopkins Bloomberg School of Public
For further information, please contact: Journals Customer Service, Health.
Oxford University Press, Great Clarendon Street, Oxford OX2 6DP, UK. The American Journal of Epidemiology (ISSN 0002-9262) is published
Telephone (& answer-phone outside normal working hours): +44 (0) 1865 semimonthly in Jan., Feb., Mar., Apr., May, Jun., Jul., Aug., Sept., Oct., Nov.,
353907, Fax: +44 (0) 1865 353485, E-mail: jnls.cust.serv@oxfordjournals.org. and Dec. by Oxford University Press, 2001 Evans Road, Cary, NC 27513-
In the US, please contact: Journals Customer Service, Oxford University 2009. Periodical Postage paid at Cary, NC and at additional post offices.
Press, 2001 Evans Road, Cary, NC 27513, USA, Telephone (& answer- Postmaster: Send address changes to: American Journal of Epidemiology,
phone outside normal working hours): 800-852-7323 (toll-free in USA/ Journals Customer Service Department, Oxford University Press, 2001
Canada), fax: 919-677-1714, e-mail: jnlorders@oxfordjournals.org. In Japan, Evans Road, Cary, NC 27513-2009.
AMERICAN JOURNAL OF EPIDEMIOLOGY
issn 0002-9262 (print)
ISSN 1476-6256 (online)
printed in the u.s.a.

American Journal of
Associate Director for Epidemiology and Genetics Research Program
Division of Cancer Control and Population Sciences
Epidemiology
The National Cancer Institute (NCI), a major research component of the National Institutes of Health (NIH) Volume 169 Number 8 April 15, 2009
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 www.aje.oxfordjournals.org
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.
ORIGINAL CONTRIBUTIONS
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
919 Parkinson’s Disease and Residential Exposure to Maneb and Paraquat From
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 Agricultural Applications in the Central Valley of California. Sadie Costello, Myles Cockburn,
on cancer epidemiology and other population sciences that informs development of creative and cost-effective Jeff Bronstein, Xinbo Zhang, and Beate Ritz
strategies to advance science is essential. 927 Overweight and Obesity Over the Adult Life Course and Incident Mobility Limitation in
The Associate Director will lead the Epidemiology and Genetics Research Program (EGRP), which includes its Older Adults: The Health, Aging and Body Composition Study. Denise K. Houston, Jingzhong
Office of the Associate Director and four branches: Clinical and Translational Epidemiology, Host Susceptibility Ding, Barbara J. Nicklas, Tamara B. Harris, Jung Sun Lee, Michael C. Nevitt, Susan M. Rubin, Frances
Factors, Methods and Technologies, and Modifiable Risk Factors. He/she provides scientific and administrative A. Tylavsky, and Stephen B. Kritchevsky for the Health ABC Study
leadership for the Program, supervises the staff, and represents NCI to a wide variety of professional, academic, 937 Association of Diabetes With Prostate Cancer Risk in the Multiethnic Cohort. Kevin
and advocacy organizations. The Associate Director also develops and facilitates collaborations with funders of M. Waters, Brian E. Henderson, Daniel O. Stram, Peggy Wan, Laurence N. Kolonel, and Christopher
other types of population science, including the NIH Institutes and Centers, Centers for Disease Control and
A. Haiman
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 370 946 Modification of the Effect of Vitamin E Supplementation on the Mortality of Male
research grants and numerous interagency agreements and consortia. Smokers by Age and Dietary Vitamin C. Harri Hemilä and Jaakko Kaprio

Volume 169
Qualifications: The successful applicant will be an experienced epidemiologist (M.D. or Ph.D.-level training re- 954 Dietary Acrylamide Intake and Risk of Premenopausal Breast Cancer. Kathryn M. Wilson,
quired) with leadership experience, excellent communication skills, and a strong record of peer-reviewed publica- Lorelei A. Mucci, Eunyoung Cho, David J. Hunter, Wendy Y. Chen, and Walter C. Willett
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, 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,

Number 8
scientific staff, training programs, interdisciplinary collaborations, and/or funded programs is highly valued.
Leslie Bernstein, Joan Largent, Marinela Capanu, Colin B Begg, Lene Mellemkjær, Charles F. Lynch,
Salary: This position is an excepted service position (Title 42) with a salary range of $160,000-$195,000. Kathleen E. Malone, Anne S. Reiner, Xiaolin Liang, Robert W. Haile, John D. Boice, Jr., WECARE Study
Collaborative Group, and Jonine L. Bernstein.
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: 969 Positive Associations Between Ionizing Radiation and Lymphoma Mortality Among
Men. David B. Richardson, Hiromi Sugiyama, Steve Wing, Ritsu Sakata, Eric Grant, Yukiko Shimizu,
Rachel Ballard-Barbash, M.D., M.P.H., Associate Director Nobuo Nishi, Susan Geyer, Midori Soda, Akihiko Suyama, Fumiyoshi Kasagi, and Kazunori Kodama
Applied Research Program
Division of Cancer Control and Population Sciences 977 Biomarker-calibrated Energy and Protein Consumption and Increased Cancer Risk
National Cancer Institute Among Postmenopausal Women. Ross L. Prentice, Pamela A. Shaw, Sheila A. Bingham, Shirley
Pages 919–1042

6130 Executive Blvd, Room 4005, MSC 7344 A. A. Beresford, Bette Caan, Marian L. Neuhouser, Ruth E. Patterson, Marcia L. Stefanick, Suzanne
Bethesda, MD 20892-7344 Satterfield, Cynthia A. Thomson, Linda Snetselaar, Asha Thomas, and Lesley F. Tinker
Express Mail: Rockville, MD 20852

Continued on Inside Front Cover


For more information about DCCPS and EGRP, see http://cancercontrol.cancer.gov

THE DHHS/NIH/NCI ARE EQUAL OPPORTUNITY EMPLOYERS


Published for the Johns Hopkins Bloomberg School of Public Health by Oxford
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 To register for this free service, go to
www.oupjournals.org/tocmail, select
tables of contents for new issues of journals
the journal(s) you are interested in,
published by Oxford University Press. 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 Human evidence is insufficient to identify any particular
high rates among farmers and in rural populations, contrib- pesticide compound, including those implicated by animal
uting to the hypothesis that agricultural pesticides might be studies, as being responsible for causing PD (11). Method-
causal agents (1–4). Animal studies have linked certain pes- ological limitations have clouded the interpretation of most
ticides to Parkinsonism and dopaminergic cell death. The epidemiologic studies exploring pesticide exposures and PD
pesticide rotenone can produce the behavioral and neuro- in humans. Past studies have generally relied on self-reports
pathologic features of PD in some rodent models through and recall of chemical usage, making them vulnerable to
chronic systemic inhibition of mitochondrial complex I (5, information bias and differential recall bias (12).
6). Exposure to a combination of the fungicide maneb and Because pesticides applied from the air or ground may
the herbicide paraquat in mice leads to increased substantia drift from their intended treatment sites, with measurable
nigra neuronal pathology (7), age-dependent motor degen- concentrations subsequently detected in the air, in plants,
eration, progressive reductions in dopamine metabolites and and in animals up to several hundred meters from applica-
turnover (8), and reduced tyrosine hydroxylase and dopa- tion sites (13–15), accurate methods of estimating environ-
mine transporter immunoreactivity (9, 10). 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- residential units (parcels) in each of the 3 counties. We
sessing exposure to pesticides have become popular in re- mailed letters of invitation to a random selection of residen-
cent years and may prove an effective solution when tial living units and also attempted to identify head-of-
pesticide data exist. We developed and employed a validated household names and telephone numbers for these parcels,
GIS-based exposure assessment tool to estimate pesticide using the services of marketing companies and Internet
exposure from applications to agricultural crops, relying searches.
on California Pesticide Use Reporting (PUR) data, land- We contacted 1,212 potential population controls by mail
use maps, and geocoded residential historical locations and/or telephone for eligibility screening. Eligibility criteria
(16). We investigated whether exposure to the pesticides were: 1) not having PD, 2) being at least 35 years of age,
maneb and paraquat, alone and in combination, increased 3) currently residing primarily in 1 of the 3 designated coun-
the risk of incident PD among residents of the Central Valley ties, and 4) having lived in California for at least 5 years
of California, an area well-known for its intensive agricul- prior to the screening. Only 1 person per household was
ture and potential for pesticide exposure. allowed to enroll. Of the potential controls contacted, 457
were ineligible: 409 were too young, 44 were terminally ill,
MATERIALS AND METHODS
and 4 resided primarily outside of the study area. Of the 755
eligible controls, 409 (54%) declined participation, were too
All procedures described have been approved by the ill to honor an appointment, or moved out of the area prior to
University of California, Los Angeles, institutional review interview; 346 (46%) were enrolled, and 341 provided all
board for human subjects, and informed consent was ob- information needed for analyses.
tained from all participants.
Assessment of environmental pesticide exposure
Subject recruitment
We conducted telephone interviews to obtain demo-
We used a population-based approach for recruiting cases graphic and exposure information. Detailed residential his-
and controls from a largely agricultural population in Cali- tory forms were mailed to subjects in advance of their
fornia. Details are provided elsewhere (17). Briefly, persons interview and were reviewed in person or over the phone.
with PD newly diagnosed between January 1998 and January We estimated pesticide exposures in the residential environ-
2007 who resided in 1 of 3 central California counties ment from applications to agricultural crops employing
(Fresno, Tulare, or Kern county) and had lived in California a validated GIS-based system, which combined PUR data
for at least 5 years prior to diagnosis were recruited into our and land-use maps (16, 18), to produce estimates of residen-
study within 3 years of diagnosis. Altogether, 28 (90%) of tial ambient pesticide applications within a set distance of
the 31 practicing local neurologists who provided care for subjects’ homes. We recorded and geocoded lifetime resi-
PD patients assisted in recruiting cases for this study. We dential histories and estimated ambient exposures for all
solicited collaboration from Kaiser Permanente Medical historical addresses at which participants had resided be-
Center (Fresno, California), Kern Medical Center (Bakersfield, tween 1974 and 1999, the period covered by the PUR data.
California), and Visalia Medical Clinic (Visalia, California) A technical discussion of our GIS-based approach is pro-
and from the Veterans Administration, PD support groups, vided elsewhere (16); here we briefly summarize the data
local newspapers, and local radio stations that broadcast sources and the exposure modeling process.
public service announcements. Residential addresses. Addresses were automatically
Of the 1,167 PD cases who were initially invited, 604 geocoded to TigerLine files (NAVTEQ (Chicago, Illinois),
were not eligible: For 397, the case’s diagnosis date fell unpublished data, 2006), and discrepancies were then man-
outside the 3-year range prior to contact, 51 denied having ually resolved in a multistep process similar to that de-
received a PD diagnosis, 134 lived outside the tricounty scribed by McElroy et al. (19). Resulting locations were
area, and 22 were too ill to participate. Of the 563 eligible recorded, along with the relevant year range of residence,
cases, 473 (84%) were examined by a University of Cali- so they could be matched to the appropriate year-specific
fornia, Los Angeles, movement disorder specialist at least PUR and land-use data (below). For our GIS model, we
once and were confirmed to have clinically ‘‘probable’’ or relied on addresses in Fresno, Kern, and Tulare counties
‘‘possible’’ PD; the remaining 90 potential cases could not (the tricounty area) at which participants had resided be-
be examined or interviewed (54% withdrew, 32% were too tween 1974 and 1999. Out of 9,568 total residential years
ill or died, and 14% moved out of the area prior to the contributed by cases (26 years 3 368 cases), 7,593 years
examination or did not honor a scheduled appointment). (79%) were spent at addresses within the tricounty area as
We examined but excluded another 96 patients because they compared with 6,757 (76%) of 8,866 years contributed by
had other causes of Parkinsonism. This left us with 377 controls (26 years 3 341 controls). We geocoded these
cases; of these, 368 provided all information needed for tricounty residential addresses for the period 1974–1999
analyses. with similar precision for cases and controls; that is, both
Controls aged 65 years or older were identified from had spent 88% of their respective residential years at ad-
Medicare lists in 2001, but because of implementation of dresses we considered to have been mapped with high pre-
the Health Insurance Portability and Accountability Act, cision (i.e., at the level of a residential parcel, street
which prohibits the use of Medicare enrollees, 70% of our address, or street intersection rather than a zip code or city
controls were recruited from randomly selected tax assessor 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 years before the onset of motor symptoms which lead to
have been children, teenagers, and young adults during the diagnosis.
exposure period: Among those exposed in the earlier time Pesticide and herbicide exposures have previously been
window, risk was increased more than 4-fold with exposure implicated in idiopathic PD. Paraquat is structurally similar
to both pesticides and more than 2-fold with exposure to just to the toxic metabolite (MPPþ) of the 1-methyl-4-phenyl-
1 of the pesticides. Consistent with some theories regarding pyridinium ion (a metabolite of 1-methyl-4-phenyl-1,2,3,6-
the progression of PD pathology (25), these data suggest tetrahydropyridine), an agent known to induce Parkinsonian
that the critical window of exposure to toxicants may be 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- and found that exposure to both pesticides at the highest
cide alone, did they exhibit increased neuronal pathology level was associated with PD, especially in persons aged
(7), age-dependent motor degeneration and progressive re- 60 years; however, wide confidence intervals surrounding
ductions in dopamine metabolites and dopamine turnover our point estimates rendered these results generally uninfor-
(8), and reduced tyrosine hydroxylase and dopamine trans- mative (results not shown).
porter immunoreactivity (9). In only 1 previous analysis, conducted within the Agri-
The fungicide maneb and the herbicide paraquat are both cultural Health Study cohort (40), did researchers assess the
used in the Central Valley of California and are often used effects of maneb and paraquat exposures. Statistical power
on the same crops, including potatoes, dry beans, and toma- was limited by the small number (n ¼ 78) of incident cases
toes. The average amount of maneb applied near the homes identified during follow-up and the very small number (n ¼
of these study subjects was relatively stable throughout both 4–10) of cases exposed to maneb/mancozeb (OR ¼ 2.1) and
time windows; however, annual paraquat exposure in- paraquat (OR ¼ 1.4). In a small Taiwanese study, the only
creased during the later (1990–1999) time window. Persons case-control study to date with sufficient statistical power to
living near fields sprayed with maneb and paraquat may also examine exposure to the herbicide paraquat, Liou et al. (41)
be exposed to a host of other agricultural chemicals. When reported a 4- to 6-fold increase in PD risk among long-term
we controlled for the influence of other groups of pesticides applicators. In a case-control study from the Mayo Clinic
suspected a priori to be risk factors for PD in our study, the (Rochester, Minnesota), Brighina et al. (42) presented asso-
odds ratios for combined maneb and paraquat exposure and ciations between self-reported pesticide exposure and PD in
PD in the younger subjects were still in the 3- to 6-fold range subjects younger than 60 years only (for all pesticides,
and statistically significant; however, our precision de- OR ¼ 1.80, 95% CI: 1.12, 2.87; for herbicides, OR ¼ 2.46,
creased, probably because of correlated exposures. Correla- 95% CI: 1.34, 4.52).
tion between pesticides is an inherent problem when Our exposure estimates did not depend on the subject’s
assessing the effects of human exposure. However, since recall of pesticide exposure and are therefore unlikely to
adjustment for other pesticides did not remove the associa- have been biased by differential exposure misclassification.
tion for maneb and paraquat, our data provide compelling Since all of our PD diagnoses were clinically confirmed, we
evidence that these 2 pesticides may in fact affect PD risk in expect disease misclassification to have been minimal. Non-
humans, as has been suggested by animal experiments. differential exposure misclassification is a possibility in our
Paraquat and maneb are applied by ground, aerial, and study and may have attenuated our effect estimates.
backpack methods; however, paraquat has a much longer Our results may be biased if cases and controls selected
field half-life of 1,000 days (33), as compared with only themselves into our study according to their potential for
12–36 days for maneb (34). Both chemicals bind strongly pesticide exposure, but our subjects were not asked to self-
to soil, though, and are not thought to be a threat to ground- report environmental exposures and probably were unaware
water (35, 36). Such strong binding could result in contam- of their true historical exposures. There is no reason to
inated soil getting blown or tracked into homes by wind, suspect that cases and controls would have chosen to par-
pets, and shoes, thereby increasing exposure for persons ticipate on the basis of their historical residence near certain
who live closer to agricultural application sites (3, 37, 38). agricultural plots. We saw no difference in estimated effects
In a previous validation study, our prediction model for a when we restricted analyses to only those subjects with
serum measure of dichlorodiphenyldichloroethylene (DDE) more (12 years) or less (<12 years) education. Similarly,
explained 47% of the biomarker’s variance (39). Addition- we saw no difference in our results when we restricted the
ally, our GIS-derived measure of organochlorine exposure sample to persons whose addresses had been mapped with
identified persons with high serum DDE levels reasonably high precision in the tricounty area during the period 1974–
well (specificity of 87%) (39). 1999 (363 cases, 336 controls).
Although our GIS model allowed us to calculate the num- Our analysis has confirmed 2 previous observations from
ber of pounds of each active ingredient applied per acre animal studies: 1) exposure to multiple chemicals may po-
within a 500-m buffer, these quantities are not comparable tentiate the effect of each chemical (of interest, since hu-
across pesticides. That is, a pound of active ingredient does mans are often exposed to more than 1 pesticide in the
not represent the same human neurotoxicity across pesti- environment) and 2) the timing of exposure is important.
cides, and no information currently exists that would allow To our knowledge, this is the first epidemiologic study to
us to standardize these measures. Thus, while we believe provide strong evidence that 2 specific pesticides, suggested
that our model provided us with an accurate indicator of any by animal research as potentially acting synergistically to
pesticide exposure from applications close to a residence, become neurotoxic, strongly increase the risk of PD in
our exposure measure cannot be considered quantitative humans, especially given combined exposure and when
beyond a crude rank ordering of low/medium likelihood encountered earlier in life.
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 ACKNOWLEDGMENTS
(note that only 3 cases and 1 control were exposed solely to
maneb). We conducted additional analyses after dichoto- Author affiliations: Department of Environmental Health
mizing pounds per acre at their median and mean levels Sciences, School of Public Health, University of California,

Am J Epidemiol 2009;169:919–926
PD and Residential Maneb and Paraquat Exposure 925

Berkeley, Berkeley, California (Sadie Costello); Department 12. Seidler A, Hellenbrand W, Robra BP, et al. Possible environ-
of Preventive Medicine, Keck School of Medicine, Univer- mental, occupational, and other etiologic factors for
sity of Southern California, Los Angeles, California (Myles Parkinson’s disease: a case-control study in Germany.
Cockburn, Xinbo Zhang); Department of Geography, College Neurology. 1996;46(5):1275–1284.
of Letters, Arts and Sciences, University of Southern 13. Chester G, Ward RJ. Occupational exposure and drift hazard
during aerial application of paraquat to cotton. Arch Environ
California, Los Angeles, California (Myles Cockburn,
Contam Toxicol. 1984;13(5):551–563.
Xinbo Zhang); Department of Neurology, School of Medicine, 14. Currier WW, MacCollom GB, Baumann GL. Drift residues
University of California, Los Angeles, Los Angeles, Cali- of air-applied carbaryl in an orchard environment. J Econ
fornia (Jeff Bronstein); and Department of Epidemiology, Entomol. 1982;75(6):1062–1068.
School of Public Health, University of California, Los 15. MacCollom GB, Currier WW, Baumann GL. Drift compari-
Angeles, Los Angeles, California (Beate Ritz). sons between aerial and ground orchard application. J Econ
This work was supported by the National Institute Entomol. 1986;79(2):459–464.
of Environmental Health Sciences (grants ES10544, 16. Goldberg DW, Wilson JP, Knoblock CA, et al. An effective
U54ES12078, and 5P30 ES07048), the National Institute of and efficient approach for manually improving geocoded data.
Neurological Disorders and Stroke (grant NS 038367), and Int J Health Geogr. 2008;7:60.
the Department of Defense Prostate Cancer Research Pro- 17. Kang GA, Bronstein JM, Masterman DL, et al. Clinical
gram (grant 051037). In addition, initial pilot funding was characteristics in early Parkinson’s disease in a central
California population-based study. Mov Disord. 2005;20(9):
provided by the American Parkinson’s Disease Association.
1133–1142.
The authors thank the participating neurologists and med- 18. Rull RP, Ritz B. Historical pesticide exposure in California
ical centers in Fresno, Kern, and Tulare counties for their using pesticide use reports and land-use surveys: an assess-
support. ment of misclassification error and bias. Environ Health
Conflict of interest: none declared. 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
REFERENCES and Forestry. Federal Insecticide, Fungicide, and
Rotenticide Act [As Amended Through P.L. 110–246,
1. Ben-Shlomo Y, Finnan F, Allwright S, et al. The epidemiology
Effective May 22, 2008]. Section 3(d)(1)(C).
of Parkinson’s disease in the Republic of Ireland: observations
Washington, DC: US Senate, 2008:30. (http://agriculture.
from routine data sources. Ir Med J. 1993;86(6):190–191, 194.
senate.gov/Legislation/Compilations/Fifra/FIFRA.pdf).
2. Burguera JA, Catala J, Taberner P, et al. Mortality from
(Accessed February 8, 2009).
Parkinson’s disease in Spain (1980–1985). Distribution by age,
21. California Department of Water Resources. California Land
sex and geographic areas. Neurologia. 1992;7(3):89–93.
3. Morano A, Jimenez-Jimenez FJ, Molina JA, et al. Risk-factors and Water Use. Sacramento, CA: California Department of
for Parkinson’s disease: case-control study in the province of Water Resources; 2009. (http://www.landwateruse.water.ca.
Cáceres, Spain. Acta Neurol Scand. 1994;89(3):164–170. gov/). (Accessed February 8, 2009).
4. Svenson LW, Platt GH, Woodhead SE. Geographic variations 22. Young HA, Mills PK, Riordan D, et al. Use of a crop and job
in the prevalence rates of Parkinson’s disease in Alberta. Can J specific exposure matrix for estimating cumulative exposure to
Neurol Sci. 1993;20(4):307–311. triazine herbicides among females in a case-control study in
5. Betarbet R, Sherer TB, MacKenzie G, et al. Chronic systemic the Central Valley of California. Occup Environ Med. 2004;
pesticide exposure reproduces features of Parkinson’s disease. 61(11):945–951.
Nat Neurosci. 2000;3(12):1301–1306. 23. Elbaz A, Tranchant C. Epidemiologic studies of environmental
6. Sherer TB, Betarbet R, Greenamyre JT. Pesticides and exposures in Parkinson’s disease. J Neurol Sci. 2007;262(1-2):
Parkinson’s disease. Scientific World Journal. 2001;1:207–208. 37–44.
7. Norris EH, Uryu K, Leight S, et al. Pesticide exposure exac- 24. Wang XF, Li S, Chou AP, et al. Inhibitory effects of pesticides
erbates alpha-synucleinopathy in an A53T transgenic mouse on proteasome activity: implication in Parkinson’s disease.
model. Am J Pathol. 2007;170(2):658–666. Neurobiol Dis. 2006;23(1):198–205.
8. Thiruchelvam M, McCormack A, Richfield EK, et al. Age- 25. Braak H, Del Tredici K, Rüb U, et al. Staging of brain pa-
related irreversible progressive nigrostriatal dopaminergic thology related to sporadic Parkinson’s disease. Neurobiol
neurotoxicity in the paraquat and maneb model of the Aging. 2003;24(2):197–211.
Parkinson’s disease phenotype. Eur J Neurosci. 2003;18(3): 26. Langston JW, Ballard P, Tetrud JW, et al. Chronic Parkinson-
589–600. ism in humans due to a product of meperidine-analog syn-
9. Thiruchelvam M, Richfield EK, Baggs RB, et al. The nigro- thesis. Science. 1983;219(4587):979–980.
striatal dopaminergic system as a preferential target of repeated 27. Sayre LM, Wang F, Hoppel CL. Tetraphenylborate potentiates
exposures to combined paraquat and maneb: implications for the respiratory inhibition by the dopaminergic neurotoxin
Parkinson’s disease. J Neurosci. 2000;20(24):9207–9214. MPPþ in both electron transport particles and intact mito-
10. Thiruchelvam M, Richfield EK, Goodman BM, et al. Devel- chondria. Biochem Biophys Res Commun. 1989;161(2):
opmental exposure to the pesticides paraquat and maneb and 809–818.
the Parkinson’s disease phenotype. Neurotoxicology. 2002; 28. Singer TP, Ramsay RR. Mechanism of the neurotoxicity of
23(4-5):621–633. MPTP. An update. FEBS Lett. 1990;274(1-2):1–8.
11. Brown TP, Rumsby PC, Capleton AC, et al. Pesticides and 29. Cassarino DS, Bennett JP Jr. An evaluation of the role of
Parkinson’s disease—is there a link? Environ Health Perspect. mitochondria in neurodegenerative diseases: mitochondrial
2006;114(2):156–164. mutations and oxidative pathology, protective nuclear

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

responses, and cell death in neurodegeneration. Brain Res 36. Environmental Protection Agency. Reregistration Eligibility
Brain Res Rev. 1999;29(1):1–25. Decision (RED) for Maneb. Washington, DC: Environmental
30. Dinis-Oliveira RJ, Remião F, Carmo H, et al. Paraquat Protection Agency; 2005. (http://www.epa.gov/oppsrrd1/
exposure as an etiological factor of Parkinson’s disease. REDs/maneb_red.pdf). (Accessed June 5, 2008).
Neurotoxicology. 2006;27(6):1110–1122. 37. Jiménez-Jiménez FJ, Mateo D, Giménez-Roldán S. Exposure
31. Cocheme HM, Murphy MP. Complex I is the major site of
to well water and pesticides in Parkinson’s disease: a case-
mitochondrial superoxide production by paraquat. J Biol
Chem. 2008;283(4):1786–1798. control study in the Madrid area. Mov Disord. 1992;7(2):
32. Zhou Y, Shie FS, Piccardo P, et al. Proteasomal inhibition 149–152.
induced by manganese ethylene-bis-dithiocarbamate: rele- 38. Rybicki BA, Johnson CC, Uman J, et al. Parkinson’s disease
vance to Parkinson’s disease. Neuroscience. 2004;128(2): mortality and the industrial use of heavy metals in Michigan.
281–291. Mov Disord. 1993;8(1):87–92.
33. Extension Toxicology Network. EXTOXNET: Extension Tox- 39. Ritz B, Costello S. Geographic model and biomarker-derived
icology Network. Pesticide Information Profiles. Paraquat. measures of pesticide exposure and Parkinson’s disease. Ann N
Corvallis, OR: Oregon State University; 1996. (http://extoxnet. Y Acad Sci. 2006;1076:378–387.
orst.edu/pips/paraquat.htm). (Accessed June 5, 2008). 40. Kamel F, Tanner CM, Umbach DM, et al. Pesticide exposure
34. Extension Toxicology Network. EXTOXNET: Extension and self-reported Parkinson’s disease in the Agricultural
Toxicology Network. Pesticide Information Profiles. Maneb.
Health Study. Am J Epidemiol. 2007;165(4):364–374.
Corvallis, OR: Oregon State University; 1996. (http://extoxnet.
41. Liou HH, Tsai MC, Chen CJ, et al. Environmental risk factors
orst.edu/pips/maneb.htm). (Accessed June 5, 2008).
35. Environmental Protection Agency. R.E.D. [Reregistration and Parkinson’s disease: a case-control study in Taiwan.
Eligibility Decision] Facts. Paraquat Dichloride. Washington, Neurology 1997;48(6):1583–1588.
DC: Environmental Protection Agency; 1997. (http://www. 42. Brighina L, Frigerio R, Schneider NK, et al. Alpha-synuclein,
epa.gov/oppsrrd1/REDs/factsheets/0262fact.pdf). (Accessed pesticides, and Parkinson disease: a case-control study. Neu-
June 5, 2008). 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 ger age groups, could reverse the recent declines in disabil-
expected to grow to 20% of the population by the year ity rates among future generations of older adults (11–13).
2030 (1). In addition, adults are, on average, increasingly Although the association between obesity in middle and
heavier than earlier generations, with approximately one- late adulthood and physical disability is established (5–8,
third classified as obese (body mass index (BMI) 30 kg/m2) 10), few studies have examined the effect of obesity among
(2). Obesity has been consistently associated with increased younger adults (14, 15) or the cumulative effect of obesity
risk of cardiovascular disease, diabetes, and several cancers, from young adulthood (9, 16) on disability in older adults.
as well as other chronic conditions (3, 4). Obesity in middle- In the National Health and Nutrition Examination Survey
aged and older adults also increases the risk of physical Epidemiologic Follow-up Study, being obese or becoming
disability (5–10), possibly as a result of these obesity- obese over 20 years of follow-up was associated with higher
related chronic conditions or by other mechanisms. Thus, levels of disability (9). Among Finnish adults aged 55 years
the growing prevalence of obesity, particularly among youn- 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- problem. Follow-up included incident cases ascertained
creased the likelihood of reporting walking limitations through 6.9 years of follow-up, with a mean follow-up of
(16). However, neither of these studies included a measure 4.2 (standard deviation, 2.4) years.
of incident disability.
In the Health, Aging and Body Composition (Health BMI history
ABC) Study, we previously showed that, compared with
those of normal weight, men and women who reported be- At the baseline visit, participants were asked to recall
ing overweight or obese at ages 25, 50, and 70–79 years had their usual body weight (women were instructed to answer
significantly worse physical performance at study baseline for a time when not pregnant) and height (without shoes) at
(ages 70–79 years) based on objective measures (17), pre- age 25 years and to recall their usual body weight at age 50
dictive of subsequent mobility disability (18). The primary years. Body weight at study baseline (ages 70–79 years) was
objective of the present study was to examine the associa- measured in kilograms by using a standard balance beam
tion of overweight and/or obesity in young, middle, and late scale, and height was measured in millimeters by using
adulthood, as well as the cumulative effect of overweight a Harpenden stadiometer (Holtain Ltd., Crosswell, United
and/or obesity across all 3 time points, with incident mobil- Kingdom). BMI at ages 25 and 50 years was calculated by
ity limitation in men and women in their 70s and 80s. 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
MATERIALS AND METHODS was categorized as normal weight (<25.0 kg/m2), over-
weight (25.0–29.9 kg/m2), and obese (30.0 kg/m2). Addi-
Study population tional analyses that excluded underweight (BMI <18.5 kg/m2)
participants at ages 25, 50, and 70–79 years were similar
Data for this analysis were derived from the Health ABC to those that included them with the normal-weight group
Study, a prospective cohort study investigating the associa- (data not shown).
tions among body composition, weight-related health con- To assess the cumulative effect of overweight and/or
ditions, and incident functional limitations in older adults. obesity across all 3 time points, the following categories
The Health ABC Study enrolled 3,075 community-dwelling were created: normal weight/nonobese at ages 25, 50, and
male and female blacks and whites aged 70–79 years be- 70–79 years; overweight and/or obese at ages 70–79 years
tween April 1997 and June 1998. Participants were recruited but not at ages 25 or 50 years; overweight and/or obese at
from a random sample of white and all black Medicare- ages 70–79 and 50 years but not at age 25 years; over-
eligible residents in the Pittsburgh, Pennsylvania, and weight and/or obese at all 3 time points; and overweight
Memphis, Tennessee, metropolitan areas. Participants were and/or obese at age 50 years but not at ages 70–79 years.
eligible if they 1) reported no difficulty walking one-fourth Other BMI history patterns were categorized as other. For
of a mile (0.4 km), climbing up 10 steps, or performing these analyses, recalled height at age 25 years was used to
activities of daily living; 2) were free of life-threatening calculate BMI at ages 70–79 years in addition to BMI at
illness; 3) planned to remain in the geographic area for at ages 25 and 50 years to minimize possible misclassifica-
least 3 years; and 4) were not enrolled in lifestyle interven- tion bias due to systematic differences between recalled
tion trials. All participants provided written informed con- height at age 25 years and measured height at study base-
sent, and all protocols were approved by the institutional line. Results were similar when measured height at study
review boards at both study sites. baseline was used to calculate BMI at ages 25, 50, and
Participants who were missing data on recalled weight at 70–79 years (data not shown).
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 Potential confounders
<15 kg/m2) at ages 25, 50, or 70–79 years (n ¼ 15) or an Demographic characteristics (age, gender, race, and
absolute weight change of more than 100 pounds (45.4 kg) education), smoking status, alcohol consumption, and
from age 50 to ages 70–79 years (n ¼ 8) were also excluded. physical activity were ascertained by an interviewer-
Participants who were missing information on mobility lim- administered questionnaire at study baseline. The time
itation during follow-up (n ¼ 2) as well as other pertinent and intensity of self-reported physical activities performed
baseline covariates (n ¼ 23) were also excluded. The final in the past 7 days, including walking for exercise, other
analysis sample was 2,845 participants. walking, climbing stairs, aerobic dance, weight training,
and other high- and medium-intensity activities, were
Mobility limitation summed (kcal/week). Because BMI history may affect mo-
bility as a consequence of weight-related health condi-
Occurrence of mobility limitation during follow-up was tions, prevalent health conditions at study baseline were
assessed during annual clinic visits alternating with tele- examined as potential mediators of the associations. The
phone interviews every 6 months. Persistent mobility limi- prevalence of diabetes, coronary heart disease, congestive
tation was defined as 2 consecutive reports of having any heart failure, stroke, chronic obstructive pulmonary dis-
difficulty walking one-fourth of a mile (0.4 km) or climbing ease, and knee pain were determined by using algorithms
up 10 steps without resting because of health or a physical based on self-report and medication use at study baseline;

Am J Epidemiol 2009;169:927–936
Weight History and Mobility Limitation 929

participants with definitive health conditions were coded as and a former drinker, engage in less physical activity, and
‘‘yes.’’ The 20-item Center for Epidemiologic Studies report prevalent chronic conditions than those who did not
Depression Scale was used as an indicator of depressed report incident mobility limitation. Overweight and obesity
mood, and persons scoring 16 or higher were classified at ages 25, 50, and 70–79 years were more prevalent among
as depressed (19). The Modified Mini-Mental State Exam- those who reported incident mobility limitation. Participants
ination was used as an indicator of general cognitive status, who reported mobility limitation were less likely to be nor-
with a minimum score of zero and maximum score of mal weight or nonobese at all 3 time points.
100 (best) (20). Because mean scores on the Modified The hazard ratios and 95% confidence intervals of inci-
Mini-Mental State Examination vary by education, a cut- dent mobility limitation by BMI status at ages 25, 50, and
point of <75 for individuals with less than a high school 70–79 years for men and women are shown in Table 2. Men
education and a cutpoint of <80 for individuals with a high who were obese at age 25 years and overweight or obese at
school education were used to classify participants as cog- ages 50 and 70–79 years were at significantly increased risk
nitively impaired. of incident mobility limitation compared with men who
were normal weight at ages 25, 50, and 70–79 years, re-
Statistical analyses
spectively. For women, those who were overweight at age
25 years and overweight or obese at ages 50 and 70–79 years
Cox proportional hazards regression models were used to were at significantly increased risk of incident mobility lim-
examine the associations between BMI status at ages 25, 50, itation compared with women who were normal weight at
and 70–79 years and risk of incident mobility limitation each of the 3 time points. Adjusting for prevalent health
with SAS version 9.1 software (SAS Institute, Inc., Cary, conditions at study baseline attenuated the associations be-
North Carolina). The cumulative effect of overweight and/or tween BMI status at ages 25, 50, and 70–79 years and in-
obesity across all 3 time points and incident mobility limi- cident mobility limitation slightly, but, in general, the
tation was also examined. Participants who survived with no associations remained significant.
evidence of incident mobility limitation were censored at Figures 1 and 2 show the hazard ratios and 95% confi-
their next-to-last 6-month contact. Participants who died dence intervals for incident mobility limitation associated
and had no evidence of incident mobility limitation were with overweight and/or obesity from age 25 to ages 70–79
censored at their time of death, and those who were lost to years. Compared with that for those of normal weight at all
follow-up were censored at their last visit. Two-way inter- 3 time points, the risk of incident mobility limitation ap-
actions between gender and BMI status at ages 25, 50, and peared to increase by duration of overweight or obesity
70–79 years were tested, and interactions were found at the (BMI 25 kg/m2) for those who were overweight or obese
significance level of a ¼ 0.10. Interactions between race and at ages 70–79 years (Figure 1). Although fewer participants
BMI status at ages 25, 50, and 70–79 years within gender reported being obese at age 50 years or earlier, similar asso-
group were also tested but were not significant. Thus, in this ciations were seen for duration of obesity (BMI 30 kg/m2),
paper, all analyses are presented for men and women sepa- particularly among men (Figure 2). After excluding partic-
rately, with race groups combined. Models were adjusted for ipants who were overweight and/or obese at age 50 years but
age, race, field center, education, smoking, alcohol con- not at ages 70–79 years and those who had a BMI history
sumption, and physical activity at study baseline. Additional pattern of ‘‘other,’’ we found a significant graded response
models were also adjusted for prevalent health conditions at for the cumulative effect of obesity in men (P for trend <
study baseline. Proportional hazards assumptions were as- 0.0001) and overweight and/or obesity in women (P for
sessed by examining log(log S(t)) plots as well as testing trend < 0.0001) on risk of mobility limitation. Women
interactions of each variable with time in the model. The who reported being overweight and/or obese at age 50 years
proportional hazards assumptions were not violated. All but not at ages 70–79 years were at increased risk of incident
P values were 2 sided. 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
RESULTS also at increased risk of incident mobility limitation. Further
adjustment for prevalent health conditions at study baseline
The mean age of the study population was 73.6 years, attenuated the associations slightly, but, in general, the as-
50.6% were women, and 39.7% were black. Participants sociations remained significant.
excluded from the present analysis (n ¼ 230, 7.5%) were The cumulative effect of overweight and/or obesity on
more likely to be female, black, and older; have less than incident mobility limitation may partly be explained by at-
a high school education; have a higher BMI at ages 70–79 tained BMI at ages 70–79 years. Participants who reported
years; and report incident mobility limitation during follow- being overweight or obese at age 50 years or earlier had
up (P < 0.01). The descriptive characteristics of the study higher BMIs in late adulthood compared with those not
population by gender and incident mobility limitation are classified as being overweight or obese until ages 70–79
shown in Table 1. Approximately 43.0% of men and 53.7% years (30.8 kg/m2 vs. 28.0 kg/m2, P < 0.0001). Thus, we
of women reported becoming limited in mobility over the examined the cumulative effect of overweight or obesity in
7 years of follow-up. Participants who reported incident analyses limited to those who were overweight or obese at
mobility limitation were more likely to be black, have less ages 70–79 years, adjusting for age, race, education, field
than a high school education, report being a current smoker center, smoking, alcohol consumption, and physical activity

Am J Epidemiol 2009;169:927–936
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 prevalent health conditions at study baseline attenuated the
mobility limitation were 1.22 (95% confidence interval: associations slightly; however, the trend remained signifi-
0.96, 1.54) for those who were overweight or obese at ages cant for women.
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 DISCUSSION
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 For both men and women, overweight and/or obesity in
years (P for trend ¼ 0.15). Among women, the hazard young, middle, and late adulthood were associated with an
ratios of incident mobility limitation were 1.71 (95% con- increased risk of incident mobility limitation in late adult-
fidence interval: 1.30, 2.23) for those who were overweight hood compared with normal weight. The risk of incident
or obese at all 3 time points and 1.23 (95% confidence mobility limitation was approximately 2.8-fold higher for
interval: 1.02, 1.47) for those who were overweight or women and 1.6-fold higher for men who were overweight or
obese at ages 50 and 70–79 years compared with those obese at ages 25, 50, and 70–79 years compared with being
who were overweight or obese at ages 70–79 years but normal weight at all 3 time points. For men, the risk of
not earlier (P for trend ¼ 0.0002). Further adjustment for 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- young and middle-aged adults predicts physical disability
obese at all 3 time points. Although not significant, there in late life (14–16).
was a trend for women who were obese at all 3 time points Overweight and obesity may lead to joint wear and tear,
to have a 1.7-fold higher risk of mobility limitation com- reduced exercise capacity, and a higher rate of chronic dis-
pared with those who were nonobese throughout. Further- ease such as cardiovascular disease, diabetes, and arthritis,
more, the risk of incident mobility limitation appeared to thus resulting in physical disability. The onset of overweight
increase with duration of overweight and/or obesity. Among or obesity in young and middle adulthood may result in
those who were normal weight at ages 70–79 years, having lower physical activity levels, contributing to decreased
a history of overweight and/or obesity in midlife was asso- muscle strength and cardiovascular fitness and greater de-
ciated with an increased risk of incident mobility limitation clines in physical function because of longer duration of
compared with those who were normal weight or nonobese excess body weight and earlier onset of chronic disease.
at all 3 time points. Thus, being overweight and/or obese at In the National Health and Nutrition Examination Survey
any time during adulthood increased the risk of mobility Epidemiologic Follow-up Study, being obese or becoming
limitation later in life, and, the longer the duration of over- obese over 20 years of follow-up was associated with higher
weight and/or obesity, the greater the risk. This finding ex- levels of upper- and lower-body disability among adults
tends prior work showing that overweight and obesity in 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 women. Furthermore, a history of overweight and/or obesity
age 30, 40, or 50 years had a 4-fold or higher increased in midlife or earlier was associated with increased risk of
likelihood of walking limitations, and there was a significant incident mobility limitation among those who were normal
linear trend between duration of obesity and walking limi- weight at ages 70–79 years.
tations (16). However, reverse causality cannot be ruled out A limitation of the work presented here is the use of
because neither of these studies included a measure of in- recalled weight and height from the distant past. However,
cident disability. In the Health ABC cohort, participants previous studies have shown high correlations (r 0.80)
who had a history of being overweight and/or obese at ages between recalled and measured weight in young adulthood
25, 50, and/or 70–79 years had an increased risk of incident among middle-aged and older men and women (21–23). In
mobility limitation compared with those who were normal the Health ABC cohort, self-reported and measured weight
weight at all 3 time points. Adjustment for prevalent di- at study baseline (ages 70–79 years) was highly correlated
abetes, cardiovascular disease, pulmonary disease, and (r ¼ 0.98). The correlation between recalled height at age
knee pain attenuated the associations slightly; however, 25 years and measured height at study baseline was also
the associations between BMI history and incident mobil- high (r ¼ 0.93; mean difference, 3 cm). Other studies have
ity remained. also shown high correlations between self-reported and
Whether overweight and/or obesity in young and middle measured height among older adults (r 0.77), with a mean
adulthood are independent of overweight and/or obesity in difference of approximately 2–3 cm (24, 25). However, the
late adulthood is difficult to untangle. In the Health ABC overreporting of height by older adults may be due, in part,
cohort, recalled BMI at ages 25 and 50 years and measured to loss of height associated with aging (26) and may more
BMI at ages 70–79 years were moderately correlated (r ¼ accurately reflect height in younger adulthood. Results were
0.33 and r ¼ 0.64, respectively). This finding raises the issue similar when measured height at study baseline was used in
of whether BMI in earlier life or attained BMI at ages 70–79 the analysis in place of recalled height at age 25 years to
years was the driving force behind the associations seen calculate BMI.
between recalled BMI at ages 25 and 50 years and mobility Important characteristics of the Health ABC cohort limit
limitation in late adulthood. Thus, caution must be exercised generalization of these findings. First, participants were re-
when interpreting the results. Participants who reported be- cruited to be well functioning and free of mobility limitation
ing overweight or obese in midlife or earlier were heavier in at study baseline. Thus, selection bias may have weakened
late adulthood than those who became overweight or obese the association between BMI history and incident mobility
later in life. Nonetheless, there appeared to be a graded re- limitation because persons with a history of overweight and/
sponse between age at onset of overweight and/or obesity or obesity may have developed mobility limitation prior to
and increased risk of mobility limitation, particularly for 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
Weight History and Mobility Limitation 935

middle-aged adults, changes in body composition that ac- 2. Ogden CL, Carroll MD, Curtin LR, et al. Prevalence of over-
company aging alter the relation between BMI and body weight and obesity in the United States, 1999–2004. JAMA.
fatness in older adults such that older adults have more body 2006;295(13):1549–1555.
fat for a given BMI compared with younger adults (28). 3. Willett WC, Dietz WH, Colditz GA. Guidelines for healthy
Finally, the observational nature of our study did not enable weight. N Engl J Med. 1999;341(6):427–434.
4. Overweight, obesity, and health risk. National Task Force on
us to evaluate a causal association between BMI history and
the Prevention and Treatment of Obesity. Arch Intern Med.
mobility limitation. It is biologically plausible that a history 2000;160(7):898–904.
of overweight and/or obesity may result in mobility limita- 5. LaCroix AZ, Guralnik JM, Berkman LF, et al. Maintaining
tion. However, although the analyses were adjusted for be- mobility in late life. II, Smoking, alcohol consumption,
havioral characteristics such as smoking and physical physical activity, and body mass index. Am J Epidemiol.
activity and for prevalent health conditions at study base- 1993;137(8):858–869.
line, overweight and/or obesity may serve as a proxy mea- 6. Launer LJ, Harris T, Rumpel C, et al. Body mass index, weight
sure for other relevant participant characteristics that may change, and risk of mobility disability in middle-aged and
lead to both overweight and obesity and mobility limitation. older women: the epidemiologic follow-up study of NHANES
In conclusion, overweight and obesity in young, middle, I. JAMA. 1994;271(14):1093–1098.
and late adulthood was associated with increased risk of 7. Jensen GL, Friedmann JM. Obesity is associated with func-
incident mobility limitation in this well-functioning cohort tional decline in community-dwelling rural older persons.
J Am Geriatr Soc. 2002;50(5):918–923.
of older adults. Those with a history of overweight and/or
8. Ostbye T, Taylor DH Jr, Krause KM, et al. The role of smoking
obesity in midlife or earlier but not in late adulthood also and other modifiable lifestyle risk factors in maintaining and
tended to have an increased risk of incident mobility limi- restoring lower body mobility in middle-aged and older
tation compared with those who were normal weight Americans: results from the HRS and AHEAD. Health and
throughout. Furthermore, there appeared to be a graded re- Retirement Study. Asset and Health Dynamics Among the
sponse between age at onset of overweight and/or obesity Oldest Old. J Am Geriatr Soc. 2002;50(4):691–699.
and risk of mobility limitation. These data suggest that in- 9. Ferraro KF, Su YP, Gretebeck RJ, et al. Body mass index and
terventions targeting prevention of overweight and obesity disability in adulthood: a 20-year panel study. Am J Public
in young and middle-aged adults may be useful in prevent- Health. 2002;92(5):834–840.
ing or delaying the onset of mobility limitation later in life. 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):
ACKNOWLEDGMENTS 199–205.
12. Lakdawalla DN, Bhattacharya J, Goldman DP. Are the young
Author affiliations: Sticht Center on Aging, Wake Forest becoming more disabled? Health Aff (Millwood). 2004;23(1):
University School of Medicine, Winston-Salem, North Car- 168–176.
olina (Denise K. Houston, Jingzhong Ding, Barbara J. 13. Alley DE, Chang VW. The changing relationship of obesity
Nicklas, Stephen B. Kritchevsky); Laboratory of Epidemi- and disability, 1988–2004. JAMA. 2007;298(17):2020–2027.
ology, Demography, and Biometry, National Institute on 14. Hubert HB, Bloch DA, Fries JF. Risk factors for physical
Aging, Bethesda, Maryland (Tamara B. Harris); Department disability in an aging cohort: the NHANES I Epidemiologic
of Foods and Nutrition, University of Georgia, Athens, Followup Study. J Rheumatol. 1993;20(3):480–488.
Georgia (Jung Sun Lee); Department of Epidemiology and 15. Houston DK, Stevens J, Cai J, et al. Role of weight history on
Biostatistics, University of California, San Francisco, Cali- functional limitations and disability in late adulthood: the
fornia (Michael C. Nevitt, Susan M. Rubin); and Department ARIC study. Obes Res. 2005;13(10):1793–1802.
16. Stenholm S, Rantanen T, Alanen E, et al. Obesity history as
of Preventive Medicine, University of Tennessee Health Sci-
a predictor of walking limitation at old age. Obesity (Silver
ence Center, Memphis, Tennessee (Frances A. Tylavsky). Spring). 2007;15(4):929–938.
This work was supported in part by the Intramural 17. Houston DK, Ding J, Nicklas BJ, et al. The association be-
Research Program of the National Institute on Aging, tween weight history and physical performance in the Health,
National Institutes of Health; National Institute on Aging, Aging and Body Composition study. Int J Obes (Lond). 2007;
National Institutes of Health (contracts N01-AG-6-2101, 31(11):1680–1687.
N01-AG-6-2103, and N01-AG-6-2106); and the Wake 18. Guralnik JM, Ferrucci L, Pieper CF, et al. Lower extremity
Forest University Claude D. Pepper Older Americans function and subsequent disability: consistency across studies,
Independence Center (grant P30-AG21332 to D. K. H.). predictive models, and value of gait speed alone compared
Conflict of interest: none declared. 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.
REFERENCES 20. Teng EL, Chui HC. The Modified Mini-Mental State (3MS)
examination. J Clin Psychiatry. 1987;48(8):314–318.
1. Federal Interagency Forum on Aging-Related Statistics. Older 21. Stevens J, Keil JE, Waid LR, et al. Accuracy of current, 4-year,
Americans 2004: Key Indicators of Well-Being. Washington, and 28-year self-reported body weights in an elderly popula-
DC: US Government Printing Office; 2004. tion. Am J Epidemiol. 1990;132(6):1156–1163.

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

22. Rhoads GG, Kagan A. The relation of coronary disease, closely are they related to measurements recorded in child-
stroke, and mortality to weight in youth and in middle age. hood? Int J Epidemiol. 2000;29(3):456–464.
Lancet. 1983;1(8323):492–495. 26. Sorkin JD, Muller DC, Andres R. Longitudinal change in the
23. Casey VA, Dwyer JT, Berkey CS, et al. Long-term memory of heights of men and women: consequential effects on body
body weight and past weight satisfaction: a longitudinal mass index. Epidemiol Rev. 1999;21(2):247–260.
follow-up study. Am J Clin Nutr. 1991;53(6):1493–1498. 27. Fried LP, Young Y, Rubin G, et al. Self-reported preclinical
24. Kuczmarski MF, Kuczmarski RJ, Najjar M. Effects of age on disability identifies older women with early declines in per-
validity of self-reported height, weight, and body mass index: formance and early disease. J Clin Epidemiol. 2001;54(9):
findings from the Third National Health and Nutrition Exam- 889–901.
ination Survey, 1988–1994. J Am Diet Assoc. 2001;101(1): 28. Villareal DT, Apovian CM, Kushner RF, et al. Obesity in older
28–34. adults: technical review and position statement of the Ameri-
25. Gunnell D, Berney L, Holland P, et al. How accurately are can Society for Nutrition and NAASO, The Obesity Society.
height, weight and leg length reported by the elderly, and how 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 men without type 2 diabetes (11, 12). The reported protec-
common chronic diseases that afflict the aging male popu- tive effects of type 2 diabetes may also be attributed to
lation. Epidemiologic studies conducted primarily in popu- differences in prostate cancer screening in diabetic and non-
lations of European ancestry have provided evidence of an diabetic patients. Differences in health maintenance, access
inverse relation between these diseases, with diabetics hav- to medical care, and the presence of serious medical con-
ing ~20% lower risk of developing prostate cancer than do ditions may result in more (or less) medical attention and
nondiabetics (1–8). However, considerable effect heteroge- preventive measures (13). Thus, examining the association
neity has been noted among studies, highlighting the need between type 2 diabetes status and prostate cancer screening
for additional prospective analyses of these endpoints in frequencies is important to quantify the degree to which
large representative population-based studies. The biologic detection bias may explain the apparent relation.
basis of this suspected relation is currently unknown and, The incidence rates of type 2 diabetes and prostate cancer
aside from age and perhaps obesity (4, 9, 10), these 2 dis- vary widely across populations. However, the ethnic dispar-
eases share no known nongenetic risk factors. ities for these common diseases are not correlated; that is,
The inverse relation between these endpoints may be due not all populations with high rates of diabetes are at low risk
to direct effects on prostate cancer growth and development, of prostate cancer (14–20). The extent to which these dis-
as men with type 2 diabetes have been found to have lower eases are linked in non-European populations is not clear.
prostate-specific antigen (PSA) levels, on average, than do 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 men with a prevalent report of prostate cancer (n ¼ 3,004)
as to examine the consistency of the association across based on self-report or from the Surveillance, Epidemiol-
racial/ethnic populations with differing rates of prostate ogy, and End Results registries. We excluded men with
cancer, we evaluated prostate cancer incidence by type missing information for body mass index (n ¼ 838), educa-
2 diabetes status in a multiethnic sample of 86,303 men tional level (n ¼ 872), and diabetes status (n ¼ 1). The
from the Multiethnic Cohort (21). We also assessed the pre- prospective analysis of the association between diabetes
sumed effect of diabetes status in the etiology of prostate status and prostate cancer incidence in this study includes
cancer by examining PSA levels in a multiethnic sample of 86,303 men.
men with and without type 2 diabetes. We also evaluated PSA levels were previously measured on 4,623 men in the
PSA screening frequencies in diabetics and nondiabetics to Multiethnic Cohort (23). These men were randomly selected
define the role of screening bias in explaining the observed from the cohort to evaluate the distribution of PSA levels
association between these common diseases. 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,
MATERIALS AND METHODS to ensure that elevated PSA levels among undiagnosed cases
Study population did not influence the results. Another 28 men with missing
body mass index data were excluded from the analysis,
The Multiethnic Cohort is a prospective cohort study that leaving 2,874 men who are included in the final analysis
includes 215,251 men and women, the majority from 5 of the effect of type 2 diabetes on PSA levels.
racial/ethnic groups in Hawaii and Los Angeles, California In 2001, we sent a short follow-up questionnaire to cohort
(African Americans, European Americans, Native Hawaiians, members. On this questionnaire, we also asked about PSA
Japanese Americans, and Latinos) (21). Between 1993 and screening prior to 1999. Of the 86,303 men included in the
1996, participants entered the cohort by completing a primary analysis of type 2 diabetes and prostate cancer,
26-page, self-administered questionnaire that asked about 23,768 (27.5%) did not complete the follow-up question-
diet and demographic factors, personal behaviors (e.g., naire. We also excluded 4,649 men with incident prostate
physical activity), history of prior medical conditions cancer. Finally, we excluded men under the age of 50 years
(e.g., diabetes), and family history of common cancers. (n ¼ 10,916) because annual PSA screening is recommen-
Potential cohort members were identified primarily through ded to begin at age 50 (24). This leaves 46,970 men included
Department of Motor Vehicles drivers’ license files and, in the analysis of the association between type 2 diabetes
additionally for African Americans, Health Care Financing and PSA screening.
Administration data files. Participants were between the ages The informed consent and study protocol were approved
of 45 and 75 years at the time of recruitment. by the institutional review boards at the University of South-
In the cohort, incident prostate cancer cases are identified ern California and the University of Hawaii.
annually through cohort linkage to the population-based
Surveillance, Epidemiology, and End Results (SEER) can- Statistical analysis
cer registries in Hawaii and Los Angeles County, as well as
the California Cancer Registry. Information on stage and Cox regression was used to estimate hazard ratios (re-
grade of disease is also obtained through these registries. ported as relative risks) for the effect of type 2 diabetes on
Linkage with these registries is complete through December prostate cancer incidence (STATA, version 8, software; Sta-
31, 2004, in Hawaii and December 31, 2005, in California. taCorp LP, College Station, Texas). We adjusted for age,
Over this period, 5,941 incident cases of invasive prostate body mass index, educational level, and race/ethnicity (in
cancer were identified. Deaths within the cohort are deter- pooled analyses). Neither body mass index nor educational
mined from linkages to the death certificate files in Hawaii level was associated with prostate cancer risk, but both re-
and California, supplemented with linkages to the National mained in the model as the former was found to be associ-
Death Index. In the Multiethnic Cohort, diabetes status is ated with PSA levels and the latter was found to be a highly
defined on the basis of self-report on the baseline question- significant predictor of PSA screening. Physical activity and
naire. This question did not differentiate between type 1 family history of prostate cancer were left out of the final
diabetes and type 2 diabetes, and thus we expect a small model because neither had an effect on the association be-
fraction (<10%) of the respondents to have type 1 diabetes tween type 2 diabetes and prostate cancer. Stratified analy-
and to be potentially misclassified (22). ses were performed in older age groups to assess whether
In addition to self-reported race/ethnicity, the following type 2 diabetes duration and long-term exposure to declin-
risk factors were included in the analysis: body mass index ing insulin levels may be important in prostate cancer de-
(weight (kg)/height (m)2), educational level (12 years, velopment. Because men may be at increased risk of
some college or vocational, and college graduate), first- prostate cancer within the first few years following a diabetes
degree family history of prostate cancer, and amount of diagnosis as a result of higher insulin levels, and since the
vigorous physical activity (0, >0–1.5, >1.5–5, and >5 hours/ date of type 2 diabetes diagnosis is unknown for cohort
week). Vigorous activity includes both vigorous sports and members, we also performed a sensitivity analysis to exam-
vigorous work. ine whether the association might be attenuated in recently
We limited our analysis to 91,018 men in the Multiethnic diagnosed diabetics. In this analysis, we censored follow-up
Cohort from the 5 major racial/ethnic groups. We excluded 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- since the initiation of PSA screening (25), with about 50%
ciation in analyses stratified by body mass index (25 kg/ of men being screened, we estimate that incidence rates
m2 and <25 kg/m2). Analyses stratified by Gleason score to have increased by 0.02 per 1% of the population screened.
determine the effect of type 2 diabetes status on prostate We then used this slope to estimate the relative impact of
cancer severity were also conducted. This latter analysis screening on prostate cancer incidence in diabetic and non-
excludes 370 prostate cancer cases with missing information diabetic men as follows: relative risk (RR)PSA ¼ (1 þ 0.02 3
on the Gleason score. screening frequency in nondiabetics)/(1 þ 0.02 3 screening
In the analysis of PSA levels, generalized linear models frequency in diabetics), with (1  RRPSA/1  RRT2D) being
were used to estimate least-squared mean PSA levels by an estimate of the fraction of the association between type 2
type 2 diabetes status (SAS, version 9.1, software; SAS In- diabetes (T2D) and prostate cancer incidence attributable to
stitute, Inc., Cary, North Carolina). Models were adjusted PSA screening.
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- RESULTS
abetes status, and we tested for a difference using logistic
regression; body mass index was not found to influence the The mean age of the men (n ¼ 86,303) in this study was
effect of type 2 diabetes on PSA screening. The fraction of 59.9 (standard deviation, 8.8) years and ranged from 56.6
the association between type 2 diabetes and prostate cancer for Native Hawaiians to 61.3 for African Americans (Table 1).
that may be attributable to PSA screening was estimated. The age-standardized prostate cancer incidence rate of
Assuming that prostate cancer incidence roughly doubled 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- between body mass index and PSA levels (Table 3). In
adjusted prevalence of type 2 diabetes also varied widely ethnicity-pooled analyses, compared with men with a body
across populations, from 6.9% in European Americans to mass index of <25 kg/m2, men with a body mass index of
18.0% in Native Hawaiians. The mean age of the diabetic 30 had 13.8% lower mean PSA levels (P ¼ 0.009). In
men in our study at baseline was slightly higher than that of multivariate generalized linear models, adjusted for type 2
the nondiabetic men for each racial/ethnic group, ranging diabetes status and age at blood draw, we estimated a 1-unit
from 59.2 years (vs. 56.1 in nondiabetics) in Native Hawai- increase in body mass index to be associated with a 1.6%
ians to 63.6 years (vs. 60.6 in nondiabetics) in Japanese decrease in mean PSA level (P < 0.001).
Americans. The age-standardized prostate cancer incidence
rates were lower in diabetic men than in nondiabetic men for
Association of type 2 diabetes status and education
each racial/ethnic group, ranging from 242.8 (per 100,000)
with PSA screening frequencies
in European-American diabetics (vs. 413.9 in nondiabetics)
to 686.2 in African-American diabetics (vs. 845.2 in non- In the sample of 46,970 men over 50 years of age with
diabetics). First-degree family history of prostate cancer information on PSA screening, 48.2% reported a PSA
was also less common in diabetic men than in nondiabetic screening test prior to 1999. European-American men were
men for each racial/ethnic group, ranging from 5.1% in more likely to report having been screened (55.8%), while
Latino diabetic men (vs. 5.8% in nondiabetics) to 7.9% in Native Hawaiians (34.3%) were the least likely to have had
African-American diabetic men (vs. 8.8% in nondiabetics). PSA testing (Pheterogeneity < 0.001). We observed a modest,
As expected, in each population, diabetic men were more yet highly statistically significant difference in age and ed-
likely to be overweight and less physically active than men ucational level regarding standardized PSA screening fre-
without type 2 diabetes (Table 1). quencies between diabetics (44.7%) and nondiabetics
In multivariate analyses, men with type 2 diabetes had (48.6%; P < 0.001) (Table 4). The lower PSA screening
significantly lower risk of prostate cancer than did men frequencies among diabetics were noted in all populations
without type 2 diabetes (RR ¼ 0.81, 95% confidence in- except in African Americans and were statistically signifi-
terval (CI): 0.74, 0.87; P < 0.001) (Table 2). The inverse cant in Japanese Americans (P < 0.001) and Latinos (P ¼
association was observed consistently in all 5 populations 0.02). Men with higher educational levels were much more
and ranged from 0.65 (95% CI: 0.50, 0.84) in European likely to have had a PSA test than were men with 12 years
Americans to 0.89 (95% CI: 0.77, 1.03) in African Ameri- of schooling (Table 4). We adjusted for educational level as
cans (Pheterogeneity ¼ 0.32). We also examined the effect of a surrogate for PSA screening in the primary cohort analyses
type 2 diabetes status on prostate cancer incidence by age at discussed above, yet it had little impact on the association.
entry into the cohort as a surrogate for duration of type 2 Next, we examined the potential impact of detection bias
diabetes, as the progressive decline of insulin levels with age on the observed association between type 2 diabetes and
among type 2 diabetics has been suggested to be protective prostate cancer. On the basis of the 3.9% difference in
for prostate cancer (1, 3, 6, 8). We found no evidence that PSA screening frequencies observed between diabetics
the inverse association was strengthened among older men and nondiabetics, we estimated that detection bias is likely
(Table 2). However, we did observe a slight, yet consistent, to account for only ~20% of the inverse association between
decrease in the relative risk when censoring the follow-up of type 2 diabetes and prostate cancer risk.
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 DISCUSSION
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- In this prospective analysis of 5 racial/ethnic populations,
served consistent effects by disease severity (Gleason score we found a highly significant association between type 2
7, n ¼ 3,853: RR ¼ 0.81, 95% CI: 0.73, 0.90; Gleason diabetes status and prostate cancer incidence, with diabetics
score >7, n ¼ 1,703: RR ¼ 0.76, 95% CI: 0.65, 0.89). having ~20% lower risk of developing prostate cancer. This
inverse association was observed in all populations, with the
Association of PSA levels with type 2 diabetes status magnitude of the effect being consistent with that of the
and body mass index majority of other studies conducted in men of European
ancestry (1–8).
In the subset of 2,874 men with PSA measurements, di- In this study, type 2 diabetes status was based on self-
abetic men (n ¼ 344) were found to have significantly lower report, which may have led to misclassification. Previous
least square geometric mean PSA levels than did nondia- studies, however, have shown that self-reported responses
betic men (n ¼ 2,530; 1.04 vs. 1.29 ng/mL; P < 0.001). for many common chronic diseases such as diabetes are
Adjusting for body mass index had little effect on this as- reliable when compared with medical records (26–28).
sociation (1.07 vs. 1.28 ng/mL; P ¼ 0.003) (Table 3). This The analysis does not account for incident cases of type 2
association was noted in all populations except Native diabetes over the 8-year follow-up period. However, inci-
Hawaiians and was statistically significant in European dent cases of diabetes in the nondiabetes group would make
Americans (0.62 vs. 1.21 ng/mL; P ¼ 0.003) and Latinos the 2 groups more similar and create an underestimation of
(0.99 vs. 1.27 ng/mL; P ¼ 0.02). Consistent with previous the inverse association. Another limitation of our study is
reports (11, 12), this report also shows an inverse relation 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/ 11. Fukui M, Tanaka M, Kadano M, et al. Serum prostate-specific
ethnic populations. These findings provide strong support antigen level in men with type 2 diabetes. Diabetes Care.
for the hypothesis that type 2 diabetes is a protective factor 2008;31(5):930–931.
for prostate cancer. We also confirmed results from previous 12. Werny DM, Saraiya M, Gregg EW. Prostate-specific antigen
studies, showing that PSA levels are decreased in diabetic values in diabetic and nondiabetic US men, 2001–2002. Am J
Epidemiol. 2006;164(10):978–983.
men. Our findings that diabetic men are less likely to be
13. Ross LE, Berkowitz Z, Ekwueme DU. Use of the prostate-
screened for prostate cancer could not account for these specific antigen test among U.S. men: findings from the 2005
results. Future studies aimed at determining the biologic National Health Interview Survey. Cancer Epidemiol Bio-
link between diabetes and prostate cancer are warranted markers Prev. 2008;17(3):636–644.
and should focus on common environmental and genetic 14. Abate N, Chandalia M. The impact of ethnicity on type 2
factors that are shared across populations. 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
ACKNOWLEDGMENTS behavior change in Native Hawaiians: initial findings from the
Native Hawaiian Diabetes Intervention Program. Diabetes
Author affiliations: Department of Preventive Medicine, Care. 2001;24(10):1770–1775.
Keck School of Medicine, University of Southern California/ 17. Aluli N. Prevalence of obesity in a Native Hawaiian popula-
Norris Comprehensive Cancer Center, Los Angeles, Cali- tion. Am J Clin Nutr. 1991;53(6 suppl):1556S–1560S.
fornia (Kevin M. Waters, Brian E. Henderson, Daniel O. 18. Brawley OW, Jani AB, Master V. Prostate cancer and race.
Stram, Peggy Wan, Christopher A. Haiman); and Epidemi- Curr Probl Cancer. 2007;31(3):211–225.
ology Program, Cancer Research Center, University of 19. Jemal A, Siegel R, Ward E, et al. Cancer statistics, 2008. CA
Hawaii, Honolulu, Hawaii (Laurence N. Kolonel). Cancer J Clin. 2008;58(2):71–96.
This work was supported by the National Cancer Institute 20. Kolonel LN, Altshuler D, Henderson BE. The Multiethnic
(grant CA54281). Cohort Study: exploring genes, lifestyle, and cancer risk. Nat
Conflict of interest: none declared. 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
REFERENCES Promotion. National diabetes fact sheet. Atlanta, GA: Centers
for Disease Control and Prevention; 2005. (http://www.cdc.gov/
1. Kasper JS, Giovannucci E. A meta-analysis of diabetes mel- diabetes/pubs/general.htm#what).
litus and the risk of prostate cancer. Cancer Epidemiol Bio- 23. Cheng I, Yu MC, Koh WP, et al. Comparison of prostate-
markers Prev. 2006;15(11):2056–2062. specific antigen and hormone levels among men in Singapore
2. Calton BA, Chang SC, Wright ME, et al. History of diabetes and the United States. Cancer Epidemiol Biomarkers Prev.
mellitus and subsequent prostate cancer risk in the NIH-AARP 2005;14(7):1692–1695.
Diet and Health Study. Cancer Causes Control. 2007;18(5): 24. Smith RA, Cokkinides V, Eyre HJ, et al. American Cancer
493–503. Society guidelines for the early detection of cancer, 2003. CA
3. Giovannucci E, Rimm EB, Stampfer MJ, et al. Diabetes mel- Cancer J Clin. 2003;53(1):27–43.
litus and risk of prostate cancer. Cancer Causes Control. 25. Collin SM, Martin RM, Metcalfe C, et al. Prostate-cancer
1998;9(1):3–9. mortality in the USA and UK in 1975–2004: an ecological
4. Gong Z, Neuhouser ML, Goodman PJ, et al. Obesity, diabetes, study. Lancet Oncol. 2008;9(5):445–452.
and risk of prostate cancer: results from the prostate cancer 26. Midthjell K, Holmen J, Biøndal A, et al. Is questionnaire in-
prevention trial. Cancer Epidemiol Biomarkers Prev. 2006; formation valid in the study of a chronic disease such as di-
15(10):1977–1983. abetes? The Nord-Trøndelag Diabetes Study. J Epidemiol
5. Velicer CM, Dublin S, White E. Diabetes and the risk of Community Health. 1992;46(5):537–542.
prostate cancer: the role of diabetes treatment and complica- 27. Okura Y, Urban LH, Mahoney DW, et al. Agreement between
tions. Prostate Cancer Prostatic Dis. 2007;10(1):46–51. self-report questionnaires and medical record data was sub-
6. Rodriguez C, Patel AV, Mondul AM, et al. Diabetes and risk of stantial for diabetes, hypertension, myocardial infarction and
prostate cancer in a prospective cohort of US men. Am J stroke but not heart failure. J Clin Epidemiol. 2004;57(10):
Epidemiol. 2005;161(2):147–152. 1096–1103.
7. Pierce BL, Plymate S, Ostrander EA, et al. Diabetes mellitus 28. Walitt BT, Constantinescu F, Katz JD, et al. Validation of self-
and prostate cancer risk. Prostate. 2008;68(10):1126–1132. report of rheumatoid arthritis and systemic lupus erythematosus:
8. Tavani A, Gallus S, Bosetti C, et al. Diabetes and the risk of the Women’s Health Initiative. J Rheumatol. 2008;35(5):811–818.
prostate cancer. Eur J Cancer Prev. 2002;11(2):125–128. 29. Chan JM, Latini DM, Cowan J, et al. History of diabetes,
9. Giovannucci E, Rimm EB, Stampfer MJ, et al. Height, body clinical features of prostate cancer, and prostate cancer
weight, and risk of prostate cancer. Cancer Epidemiol Bio- recurrence-data for CaPSURETM (United States). Cancer
markers Prev. 1997;6(8):557–563. Causes Control. 2005;16(7):789–797.
10. Huxley R, James WP, Barzi F, et al. Ethnic comparisons of the 30. Fowke JH, Signorello LB, Chang SS, et al. Effects of obesity
cross-sectional relationships between measures of body size and height on prostate-specific antigen (PSA) and percentage
with diabetes and hypertension. Obes Rev. 2008;9(suppl 1): of free PSA levels among African-American and Caucasian
53–61. 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, plasma a-tocopherol disappearance rate, which is normal-
particularly among older people. In the United States, about ized by vitamin C supplementation (10). Thus, since smok-
a quarter of adults aged 60 years or older take supplements ing seems to modify the interaction of these 2 antioxidants,
containing vitamin E (1). Such a common habit makes the intake of vitamin C is particularly important when examin-
health effects of this practice an important public health ing the effect of vitamin E on smokers.
issue: does vitamin E supplementation improve health or The Alpha-Tocopherol, Beta-Carotene Cancer Prevention
not? (ATBC) Study was a randomized, double-blind, placebo-
The rationale behind taking the lipid-soluble antioxidant controlled trial that examined the effects of vitamin E and
vitamin E is to protect against oxidative stress, which con- b-carotene on lung cancer in male smokers (12, 13). Pre-
tributes to the aging processes and may affect longevity viously, we found significant heterogeneity in the effects of
(2, 3). However, 3 meta-analyses of randomized trials found vitamin E supplementation in the ATBC Study; vitamin E
that vitamin E supplementation did not reduce mortality, increased the risk of tuberculosis for heavy smokers with
implying that vitamin E does not lead to universal systemic high dietary vitamin C intake but had no effect on other
benefits against the processes that lead to chronic disease participants (14). The effect of vitamin E on common cold
(4–6). and pneumonia risk was modified by age and smoking (11,
Vitamin E is a major lipid-soluble antioxidant, whereas 15–19).
vitamin C is a major water-soluble antioxidant. They inter- In the ATBC Study, vitamin E supplementation had no
act in vitro and in vivo (7–11). Smoking increases the 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 947

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

Am J Epidemiol 2009;169:946–953
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 it is thought to be valid for all participants, as shown in our
population groups. work. The overall estimate for all ATBC Study participants,
Although we divided the participants into several sub- risk ratio ¼ 1.02, is inconsistent with our subgroup findings
groups, the multiple comparison problem did not seem to be for nearly half of all participants, that is, the young and old
a concern in our study. First, the subgroup heterogeneity in the with high dietary vitamin C intake.
6 groups defined by age and dietary vitamin C intake was Our current study and earlier subgroup analyses of the
significant even when we allowed each of the 6 subgroups to ATBC Study suggest that subgroup analyses of large trial
have its own vitamin E effect (Table 1). Second, our current databases should be encouraged even if there is no overall
subgroup analyses tested our earlier subgroup findings for res- effect in the study population, keeping in mind the possibil-
piratory infection outcomes in the ATBC Study (11, 14–19). ity of spurious findings from multiple testing. Given the
Our findings have several important implications. Sub- long-term commitment of participants and the resources
group analysis has been discouraged because it can lead to invested, it might even be considered a moral imperative
false-positive findings due to the multiple comparison prob- to analyze the large trial databases as extensively as possible
lem (23–25). It has even been argued that ‘‘believing that rather than simply to calculate an overall effect. Evidently,
a treatment effect exists in one stratum of patients, even subgroup findings should be considered cautiously, and the
though no overall significant treatment effect exists, is interpretation of P values must be related to the number of
a common error’’ (24, p. 15). Furthermore, large trials are subgroup analyses being carried out. Nevertheless, our sub-
set up after years of deliberation by experts, and it is as- group findings in the ATBC Study point out the need for
sumed that all relevant knowledge is therefore incorporated further research on vitamin E, whereas the overall estimate
into the study plans (25). However, biology is complex, and of effect suggests that no further studies would have been
it seems unlikely that all important biologic knowledge worthwhile. We concur with Feinstein’s concern (26) that
could ever be taken into account properly when setting up anti-subgroup doctrines have become so entrenched that
a pragmatic controlled trial. A single estimate of treatment they often hamper investigation of socially and biologically
effect calculated for a large population can be misleading if 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 erogeneity compared with the Physicians’ Health Study II
vitamin E supplementation on mortality (4–6). Combining focused on a single upper-class profession.
several large trials leads to estimates with narrow confidence Our strongest findings concern vitamin E supplementation
intervals; however, pooling is based on the assumption of the because the division of participants into vitamin-E and no-
same universal effect in all populations of the combined vitamin-E groups was random. The evidence of heterogeneity
studies. Our strong evidence of heterogeneity in the ATBC of the vitamin E effect is strong irrespective of the specificity
Study refutes the assumption of a uniform effect and casts of vitamin C as the modifier. Nevertheless, by using the re-
doubt on the validity of the pooled estimates of the meta- sidual of fruit, vegetables, and berries, which has no corre-
analyses. Although the pooled estimates reject the proposal lation with dietary vitamin C, we were able to show that
that vitamin E supplementation would be beneficial for wide- modification of the vitamin E effect is not explained by other
ranging population groups, our study suggests that important substances in these foods, suggesting that vitamin C is the
subgroups being harmed or benefiting from vitamin E may be specific substance explaining the modification. Furthermore,
hidden in the misleadingly narrow confidence intervals. the subgroup divisions in our current study were based on the
The recently published Physicians’ Health Study II found effects of vitamin E on respiratory infections, and vitamin C
no overall effect of vitamin E and C supplementation on has also affected respiratory infections in certain controlled
mortality (27). However, the lack of average effect does trials, showing that its physiologic effects are not limited to
not conflict with our findings of heterogeneity because vi- preventing scurvy (28, 29). Finally, the interaction between
tamin E had no overall effect in the ATBC Study either. The vitamins E and C is well established (7–11). These argu-
number of deaths in the ATBC Study (n ¼ 3,571) was twice ments support the possibility that vitamin C specifically
that in the Physicians’ Health Study II (n ¼ 1,661); conse- causes the modification of the vitamin E effect, although our
quently, there is more statistical power to analyze potential vitamin C intake levels were based on observational data.
subgroup differences in the ATBC Study. Furthermore, All ATBC Study participants were smokers, and no direct
ATBC Study participants were recruited from the general conclusions can be drawn for nonsmokers. The benefit of
population, so there is greater potential for analyzing het- 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 dently, in people younger than age 65 years, taking vitamin
smoked less, and the harm was more apparent in those who E supplements should be strongly discouraged until clear
smoked more (14, 15, 18, 19). We saw similar trends in the evidence emerges that some population groups of younger
effect of vitamin E on mortality, but the modification of the or middle-aged people benefit. On the other hand, our study
vitamin E effect caused by smoking quantity was not statis- indicates that vitamin E supplementation may lead to ben-
tically significant. Nevertheless, the greater reduction in mor- eficial effects in some subgroups of old people, and this
tality among old participants with high vitamin C intake who possibility should be investigated by using a factorial design
smoked less suggests that vitamin E supplementation might with vitamin C supplementation. Finally, the substantial de-
affect mortality among older male nonsmokers as well. crease in mortality with vitamin E supplementation among
Since vitamin E is a fat-soluble substance and it may take the older participants with high dietary vitamin C intake
months before tissue levels are substantially elevated (30, raises the question of whether the decrease in overall mor-
31), short trials may be uninformative. In the current study, tality is attributable to a single cause of death or a few
there seemed to be 2- and 3-year lag periods before vitamin causes, or whether it suggests a general decrease in frailty
E started to decrease or increase mortality (Figures 1 and 2). reflected in lessened mortality from diverse causes.
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. ACKNOWLEDGMENTS
On the other hand, the risk of tuberculosis increased signif-
icantly within a year of supplementation being initiated (14) Authors affiliation: Department of Public Health, Univer-
and the risk of pneumonia in participants who exercised in sity of Helsinki, Helsinki, Finland (Harri Hemilä, Jaakko
their leisure time decreased without a lag (16), so not all Kaprio).
effects of vitamin E are delayed. The authors thank the ATBC Study (the National Public
It has been suggested that the vitamin E doses in the ran- Health Institute, Finland, and the National Cancer Institute,
domized trials were too low to show any effect (32), and high United States) for access to the data.
doses in some trials increased mortality (5). Given our ob- H. H. planned the study and wrote the draft of the manu-
servations that 50 mg/day of vitamin E caused significant script. J. K. participated in planning the analyses and the
increase and decrease in mortality in the ATBC Study pop- critical revision of the manuscript. H. H. had full access to
ulation, depending on the characteristics of participants, no all data in this study and takes responsibility for the integrity
justification exists for claiming that the vitamin E doses in of the data and the accuracy of the data analysis.
randomized trials have been too low. Our study also suggests Conflict of interest: none declared.
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 REFERENCES
stress, which has a role in the aging processes (2, 3). Our 1. Radimer K, Bindewald B, Hughes J, et al. Dietary supplement
findings among the older ATBC Study participants are con- use by US adults: data from the National Health and Nutrition
sistent with the concept that the antioxidant vitamins C and Examination Survey, 1999–2000. Am J Epidemiol. 2004;
E can counteract oxidative stress processes in old people 160(4):339–349.
under some conditions. Assuming that the benefit of vitamin 2. Beckman KB, Ames BN. The free radical theory of aging
E supplementation in old people is conditional on a high matures. Physiol Rev. 1998;78(2):547–581.
intake of vitamin C, it seems possible that old people with 3. Finkel T, Holbrook NJ. Oxidants, oxidative stress and the bi-
low dietary vitamin C might benefit from a combination of ology of ageing. Nature. 2000;408(6809):239–247.
the 2. Therefore, the most informative type of further study 4. Vivekananthan DP, Penn MS, Sapp SK, et al. Use of antioxi-
dant vitamins for the prevention of cardiovascular disease:
would be a factorial design with vitamins E and C on old meta-analysis of randomised trials. Lancet. 2003;361(9374):
people with low dietary vitamin C intake. 2017–2023.
Finally, the US nutritional recommendations and a recent 5. Miller ER, Pastor-Barriuso R, Dalal D, et al. Meta-analysis:
review consider that vitamin E is safe in doses up to 1,000 high-dosage vitamin E supplementation may increase
mg/day (33, 34). Our finding that 50 mg/day of vitamin E all-cause mortality. Ann Intern Med. 2005;142(1):37–46.
significantly increased mortality among half of the male (Comments: Ann Intern Med. 2005;143(2):150–158).
smokers in the ATBC Study aged 50–62 years contradicts 6. Bjelakovic G, Nikolova D, Gluud LL, et al. Mortality in ran-
the universal safety of 1,000 mg/day of vitamin E. Previously, domized trials of antioxidant supplements for primary and
we found that 50 mg/day of vitamin E significantly increased secondary prevention: systematic review and meta-analysis.
the risk of the common cold, pneumonia, and tuberculosis in JAMA. 2007;297(8):842–857. (Comments: JAMA. 2007;
298(4):400–403).
subgroups of the ATBC Study, also indicating that such a low 7. Packer JE, Slater TF, Willson RL. Direct observation of a free
dose can cause harmful effects in some people (14, 18, 19). radical interaction between vitamin E and vitamin C. Nature.
In conclusion, our subgroup analyses of the ATBC Study 1979;278(5706):737–738.
cohort support the conclusions of meta-analysis of con- 8. Tanaka K, Hashimoto T, Tokumaru S, et al. Interactions be-
trolled trials (4–6), in that vitamin E supplementation seems tween vitamin C and vitamin E are observed in tissues of in-
ineffective or harmful for middle-aged male smokers. Evi- herently scorbutic rats. J Nutr. 1997;127(10):2060–2064.

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

9. Hill KE, Montine TJ, Motley AK, et al. Combined deficiency 22. Hemilä H, Kaprio J, Pietinen P, et al. Vitamin C and other
of vitamins E and C causes paralysis and death in guinea pigs. compounds in vitamin C rich food in relation to risk of tu-
Am J Clin Nutr. 2003;77(6):1484–1488. berculosis in male smokers. Am J Epidemiol. 1999;150(6):
10. Bruno RS, Leonard SW, Atkinson J, et al. Faster plasma 632–641.
vitamin E disappearance in smokers is normalized by vitamin C 23. Assmann SF, Pocock SJ, Enos LE, et al. Subgroup analysis and
supplementation. Free Radic Biol Med. 2006;40(4):689–697. other (mis)uses of baseline data in clinical trials. Lancet.
(Comments: Free Radic Biol Med. 2007;42(4):578–580). 2000;355(9209):1064–1069.
11. Hemilä H. Do Vitamins C and E Affect Respiratory Infections? 24. Peto R, Pike MC, Armitage P, et al. Design and analysis of
[dissertation]. Helsinki, Finland: University of Helsinki; randomized clinical trials requiring prolonged observation of
2006:10–11, 56–57. (Available at http://ethesis.helsinki.fi/ each patient. II. Analysis and examples. Br J Cancer. 1977;
julkaisut/laa/kansa/vk/hemila/). (Accessed December 15, 2008). 35(1):1–39.
12. The effect of vitamin E and beta-carotene on the incidence of 25. Vandenbroucke JP. Observational research, randomised trials,
lung cancer and other cancers in male smokers. The Alpha- and two views of medical science [electronic article]. PLoS
Tocopherol, Beta-Carotene Cancer Prevention Study Group. Med. 2008;5(3):e67.
N Engl J Med. 1994;330(15):1029–1035. 26. Feinstein AR. The problem of cogent subgroups: a clinico-
13. The alpha-tocopherol, beta-carotene lung cancer prevention statistical tragedy. J Clin Epidemiol. 1998;51(4):297–299.
study: design, methods, participant characteristics, and com- 27. Sesso HD, Buring JE, Christen WG, et al. Vitamins E and C in
pliance. The ATBC Cancer Prevention Study Group. Ann the prevention of cardiovascular disease in men: the Physi-
Epidemiol. 1994;4(1):1–10. cians’ Health Study II randomized controlled trial. JAMA.
14. Hemilä H, Kaprio J. Vitamin E supplementation may tran- 2008;300(18):2123–2133.
siently increase tuberculosis risk in males who smoke heavily 28. Douglas RM, Hemilä H. Vitamin C for preventing and treating
and have high dietary vitamin C intake. Br J Nutr. 2008; the common cold [electronic article]. PLoS Med. 2005;2(6):
100(4):896–902. (Comments: Br J Nutr. 2009;101(1):145–147). e168.
15. Hemilä H, Virtamo J, Albanes D, et al. Vitamin E and beta- 29. Hemilä H, Louhiala P. Vitamin C may affect lung infections.
carotene supplementation and hospital-treated pneumonia in- J R Soc Med. 2007;100(11):495–498.
cidence in male smokers. Chest. 2004;125(2):557–565. 30. Kitagawa M, Mino M. Effects of elevated d-alpha(RRR)-
16. Hemilä H, Kaprio J, Albanes D, et al. Physical activity and the tocopherol dosage in man. J Nutr Sci Vitaminol (Tokyo). 1989;
risk of pneumonia in male smokers administered vitamin E 35(2):133–142.
and b-carotene. Int J Sports Med. 2006;27(4):336–341. 31. Handelman GJ, Epstein WL, Peerson J, et al. Human adipose
17. Hemilä H, Kaprio J, Albanes D, et al. Vitamin C, vitamin E, alpha-tocopherol and gamma-tocopherol kinetics during and
and beta-carotene in relation to common cold incidence in after 1 y of alpha-tocopherol supplementation. Am J Clin Nutr.
male smokers. Epidemiology. 2002;13(1):32–37. 1994;59(5):1025–1032.
18. Hemilä H, Virtamo J, Albanes D, et al. The effect of vitamin E on 32. Blumberg JB, Frei B. Why clinical trials of vitamin E and
common cold incidence is modified by age, smoking and resi- cardiovascular diseases may be fatally flawed. Commentary on
dential neighborhood. J Am Coll Nutr. 2006;25(4):332–339. ‘‘The relationship between dose of vitamin E and suppression
19. Hemilä H, Kaprio J. Vitamin E supplementation and pneu- of oxidative stress in humans’’. Free Radic Biol Med. 2007;
monia risk in males who initiated smoking at an early age: 43(10):1374–1376.
effect modification by body weight and vitamin C [electronic 33. Food and Nutrition Board, Institute of Medicine. Dietary
article]. Nutr J. 2008;7(1):33. Reference Intakes for Vitamin C, Vitamin E, Selenium, and
20. Pietinen P, Hartman AM, Haapa E, et al. Reproducibility Carotenoids. Washington, DC: The National Academies Press;
and validity of dietary assessment instruments. I. A self- 2000: 249–260. (Available at http://www.nap.edu/books/
administered food use questionnaire with a portion size picture 0309069351/html/249.html). (Accessed December 15, 2008).
booklet. Am J Epidemiol. 1988;128(3):655–666. 34. Hathcock JN, Azzi A, Blumberg J, et al. Vitamins E and C are
21. Ovaskainen ML, Valsta M, Lauronen J. The compilation of safe across a broad range of intakes. Am J Clin Nutr. 2005;
food analysis values as a database for dietary studies—the 81(4):736–745. (Comments: Am J Clin Nutr. 2005;82(5):
Finnish experience. Food Chem. 1996;57(1):133–136. 1141–1143).

Am J Epidemiol 2009;169:946–953
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- cancers, including mammary tumors in female rats (5, 6).
gen (1), and it is formed during high-temperature pro- Given the burden of breast cancer, it is of interest to study
cessing of many commonly consumed foods (2). The the association between acrylamide intake in humans and
discovery of acrylamide in foods in 2002 caused consider- the risk of breast cancer. Epidemiologic studies have had
able concern worldwide. Acrylamide is widespread in the mixed results. Two prospective studies of dietary acrylam-
food supply, with approximately 38% of calories con- ide exposure in humans found no association with pre- or
sumed in the United States coming from foods that postmenopausal breast cancer risk (7, 8). Both of these
contain acrylamide (3). Potatoes, cold breakfast cereal, reports used food frequency questionnaires (FFQs) to as-
coffee, and baked goods are major sources of acrylamide sess acrylamide intake. A prospective study in the Danish
intake in the United States (4). Prior to the discovery of Diet, Cancer, and Health Cohort used a biomarker of ac-
acrylamide in foods, industrial use and tobacco use were rylamide exposure, acrylamide adducts to hemoglobin, and
thought to be the major sources of acrylamide exposure in found an increased risk of breast cancer among postmen-
humans (1). opausal women with higher adducts. The increased risk
In animal tests, acrylamide administered in high levels in appeared limited to smokers and to estrogen receptor-positive
drinking water causes several types of hormone-sensitive 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


Acrylamide and Premenopausal Breast Cancer 955

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

Am J Epidemiol 2009;169:954–961
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 parity and age at first birth (nulliparous, 1–2 children and
their acrylamide intake and their consumption of acrylamide- age at first birth <25 years, 1–2 children and age at first
rich foods. Acrylamide intake was adjusted for total energy birth 25–<30 years, 1–2 children and age at first birth 30
intake by using the residual method. Relative risks of breast years, 3 or more children and age at first birth <25 years,
cancer were calculated as the incidence rate for a given 3 or more children and age at first birth 25 years), age at
quintile of consumption divided by the rate in the lowest menarche (<12, 12, 13, or 14 years), family history of
quintile. breast cancer (yes/no), history of benign breast disease
We used Cox proportional hazards regression to adjust (yes/no), smoking (never, former smoker of <25 ciga-
for potential confounding by other breast cancer risk fac- rettes/day, former smoker of 25 cigarettes/day, current
tors. To control as finely as possible for confounding by smoker of <25 cigarettes/day, and current smoker of 25
age, calendar time, and any possible 2-way interactions cigarettes/day), physical activity (18 and >18 metabolic
between these 2 time scales, we stratified the analysis equivalent (MET)-hours/week), animal fat (quintiles), glyc-
jointly by age in months at the start of each follow-up emic load (quintiles), alcohol intake (continuous g/day), and
period and calendar year of the current questionnaire cycle. total energy intake (continuous kcal/day). We adjusted for
We used multivariable models to adjust for the following animal fat and glycemic load as they have previously been
factors: body mass index (<18.5, 18.5–19.9, 20.0–22.4, associated with breast cancer risk in this cohort (13, 14).
22.5–24.9, 25.0–29.9, and 30 kg/m2), height (<62, We also considered adjustment for quintile of vegetable
62–<65, 65–<68, and 68 inches; 1 inch ¼ 2.54 cm), oral fat intake, trans fat intake, and glycemic index, as these
contraceptive use (never, former use <4 years, former use dietary factors were most correlated with acrylamide in-
4 years, current use <8 years, and current use 8 years), 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- We examined whether the association between acryl-
dard error estimates for acrylamide. All covariates except amide intake and breast cancer risk was modified by in-
height and age at menarche were updated in each question- dividual characteristics, including age, smoking status,
naire cycle. The SAS Proc PHREG procedure (SAS Insti- body mass index, alcohol intake, and glycemic load, by
tute, Inc., Cary, North Carolina) was used for all analyses, modeling the association separately in each group. We
and the Anderson-Gill data structure was used to handle tested the significance of interactions by adding cross-
time-varying covariates efficiently. To test for a linear product terms between acrylamide intake and the variable
trend across quintiles of intake, we modeled acrylamide of interest to the model and comparing this model to the
intake as a continuous variable using the median value for model without the cross-product term using the likelihood
each quintile. 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


and 37.8 lg/day in the highest quintile. The major food
During 14 years (945,764 person-years) of follow-up, we contributors to acrylamide intake were French fries (23%),
documented 1,179 cases of invasive breast cancer among coffee (15%), cold breakfast cereal (12%), potato chips
90,628 premenopausal women in the cohort. The age range (9%), and other potatoes (baked, roasted, mashed; 5%).
of women in 1991 was 26–46 years. Ages at breast cancer Those in the highest quintile of acrylamide consumption
diagnosis ranged from 26 to 56 years. We had information were more likely to be current smokers and were less likely
on estrogen receptor (ER)/progesterone receptor (PR) status to exercise than those in the lowest quintile.
for 916 (78%) cases. Of these, 597 were ER and PR positive Intake of acrylamide was not associated with risk of pre-
(ERþ/PRþ), and 196 were ER and PR negative (ER/ menopausal breast cancer (Table 2). The multivariable rel-
PR). Because of the small number of mixed ER/PR status ative risk of breast cancer was 0.92 (95% confidence interval
tumors, we did not include these cases in our analysis by (CI): 0.76, 1.11) in the highest quintile of intake compared
ER/PR status. with the lowest quintile. The P value for a linear trend across
Table 1 shows the characteristics of the cohort in 1991 by quintiles was 0.61. No association was found for ERþ/PRþ
quintile of energy-adjusted acrylamide intake. The mean 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 this is the exposure measurement used in toxicology studies,
exposure, we examined the association among never and again found similar results. The relative risk for the
smokers, former smokers, and current smokers separately highest versus the lowest quintile of acrylamide by body
(Table 2). There was no indication of increased risk of weight was 1.00 (95% CI: 0.82, 1.22), with a P value for
breast cancer for higher acrylamide intakes in any of these linear trend of 0.95. Baseline acrylamide intake through diet
groups. was also not associated with breast cancer risk. The relative
We found no significant differences in the association risks relative to the lowest quintile of baseline acrylamide
between dietary acrylamide intake and breast cancer risk intake were 0.96 (95% CI: 0.79, 1.15) for quintile 2,
when we stratified the population by age, body mass index, 1.05 (95% CI: 0.88, 1.26) for quintile 3, 1.01 (95% CI:
alcohol intake, glycemic index, or glycemic load (data not 0.84, 1.21) for quintile 4, and 1.03 (95% CI: 0.86, 1.24)
shown). for quintile 5 (Ptrend ¼ 0.62).
We repeated the analysis measuring acrylamide exposure Table 3 shows the association between consumption
relative to body weight (i.e., lg/kg of body weight/day), as 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- exposure in the range of dietary intakes is not clearly asso-
ual foods that contributed at least 2% to the total estimated ciated with breast cancer risk. Olesen et al. also found a sig-
acrylamide intake in our population, and none was posi- nificantly increased risk of ERþ cancers among those with
tively associated with breast cancer risk. We also examined higher adduct levels (for a 10-fold increase in adducts, RR ¼
several food groups that are major sources of acrylamide: all 2.7, 95% CI: 1.1, 6.6). This result was among smokers
potatoes (French fries, potato chips, and baked/mashed/ and nonsmokers combined, with multivariable adjustment
roasted potatoes), breads (white and dark bread, English for smoking behavior; however, given the relative contribu-
muffins/bagels/rolls, tortillas, pancakes, pizza, and crack- tions of smoking and diet to adduct levels, it is not clear that
ers), and baked goods (cookies, brownies, donuts, cake, such adjustment provides adequate control of confounding by
pie, and sweet rolls). None of these food groups was asso- smoking. In our analysis of food frequency questionnaire-
ciated with breast cancer risk. assessed acrylamide and ERþ cancers, we found no signifi-
cant association.
Strengths of our study include its prospective design,
DISCUSSION large number of premenopausal cases, and high rates of
follow-up. In addition, we are the first to study acrylamide
We found no association between acrylamide intake and intake and cancer risk using multiple FFQs administered to
premenopausal breast cancer risk in this cohort. There was collect updated dietary data throughout the follow-up pe-
no association for ER/PR-positive or -negative cancers or by riod, rather than a single point in time. This improves our
smoking status. In addition, there was no association be- assessment of long-term diet and reduces measurement error
tween intake of any major acrylamide-contributing foods (17). Our study also uses an extensive acrylamide database
and breast cancer risk. with 42 acrylamide-contributing foods, approximately twice
These findings are in line with both previous prospective as many as used in previous studies.
studies of acrylamide intake and breast cancer risk using We have previously found that FFQ acrylamide intake is
FFQs. Mucci et al. (7) found no significant association be- significantly correlated with hemoglobin adducts of acryl-
tween acrylamide intake and breast cancer risk among amide and its metabolite glycidamide in this population
mostly premenopausal Swedish women. Hogervorst et al. (10). However, misclassification of acrylamide intake re-
(8) found no significant association among postmenopausal mains a limitation of our study, as acrylamide poses unique
Dutch women. The mean acrylamide intake and the contrast challenges for FFQ assessment. Acrylamide formation is
in intakes between high and low quintiles were quite similar affected by many parameters (for review, refer to Stadler
across the 3 studies. However, the main sources of acrylam- and Scholz (18)), such as cooking temperature and even
ide vary between populations. French fries, coffee, cold the length and temperature of storage of ingredients such
breakfast cereal, and potato chips contribute the most to as potatoes, which implies that acrylamide content varies
intake in our US population. Coffee is a larger contributor widely among different brands of prepared foods, and de-
to acrylamide intake in both the Swedish and Dutch cohorts pends on the cooking methods used at home. Therefore, the
(7, 8). Fried potatoes and crisp bread are other major con- assignment of a single acrylamide value for each food likely
tributors in the Swedish women (7), and Dutch spice cake results in nondifferential misclassification of acrylamide in-
and cookies are major contributors in the Dutch women (8). take, which would bias our observed relative risks toward
In addition, Pelucchi et al. found no association between the null. As a result, we may have missed modest associa-
acrylamide intake (15) or fried potato intake (16) and breast tions between acrylamide intake and cancer risk. We also
cancer risk in a hospital-based, case-control study in Italy found no association between breast cancer risk and intake
and Switzerland. of any of the top 11 acrylamide-contributing foods, which
In a nested case-control study, Olesen et al. (9) studied the are well measured by the FFQ within the similar Nurses’
association between acrylamide adducts to hemoglobin, Health Study cohort (11).
a biomarker of acrylamide exposure, and postmenopausal Residual confounding is also a concern in observational
breast cancer risk. Current smokers with higher levels of studies. Adjustment for known breast cancer risk factors had
acrylamide adducts to hemoglobin at baseline had a signifi- very little effect on the relative risk estimates, suggesting that
cantly increased risk of breast cancer (for a 10-fold increase it is unlikely that confounding was a source of substantial
in adducts, relative risk (RR) ¼ 3.1, 95% CI: 1.0, 9.7). To- bias. Finally, it is not clear that adult diet is the most rel-
bacco use is an important source of acrylamide exposure, evant period of exposure. It is possible that high acrylam-
and smokers have acrylamide adduct levels 3–5 times higher ide intake in childhood or adolescence may increase breast
than do nonsmokers. Therefore, adduct levels among smok- cancer risk later in life, and our study cannot address this
ers reflect both tobacco use and dietary intake of acrylam- question.
ide. Among nonsmoking women, whose adduct levels are In conclusion, we found no association between acrylam-
thought primarily to represent dietary acrylamide exposure, ide intake from the diet and risk of premenopausal breast
Olesen et al. found no statistically significant association cancer risk. Combined with the results of other prospective
between adduct levels and breast cancer risk (for a 10-fold cohort studies, this suggests that intake of foods high in
increase in adducts, RR ¼ 1.5, 95% CI: 0.6, 3.6). The acrylamide is not a major risk factor for breast cancer. How-
meaning of the positive association with adduct levels ever, a modest association could have been missed, because
among smokers is not clear; however, their results among the substantial variation in the acrylamide content of foods
nonsmokers seem in line with our finding that acrylamide makes measurement of intake difficult.

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

ACKNOWLEDGMENTS 6. Friedman MA, Dulak LH, Stedham MA. A lifetime oncogenicity


study in rats with acrylamide. Fundam Appl Toxicol. 1995;
Author affiliations: Department of Epidemiology, 27(1):95–105.
Harvard School of Public Health, Boston, Massachusetts 7. Mucci LA, Sandin S, Bälter K, et al. Acrylamide intake and
(Kathryn M. Wilson, Lorelei A. Mucci, David J. Hunter, breast cancer risk in Swedish women. JAMA. 2005;293(11):
Walter C. Willett); Department of Nutrition, Harvard 1326–1327.
School of Public Health, Boston, Massachusetts (Kathryn 8. Hogervorst JG, Schouten LJ, Konings EJ, et al. A prospective
study of dietary acrylamide intake and the risk of endometrial,
M. Wilson, Eunyoung Cho, David J. Hunter, Walter C.
ovarian, and breast cancer. Cancer Epidemiol Biomarkers
Willett); Channing Laboratory, Department of Medicine, Prev. 2007;16(11):2304–2313.
Harvard Medical School and Brigham and Women’s 9. Olesen PT, Olsen A, Frandsen H, et al. Acrylamide exposure
Hospital, Boston, Massachusetts (Lorelei A. Mucci, and incidence of breast cancer among postmenopausal women
Eunyoung Cho, David J. Hunter, Wendy Y. Chen, Walter in the Danish Diet, Cancer and Health Study. Int J Cancer.
C. Willett); and Department of Medical Oncology, Dana- 2008;122(9):2094–2100.
Farber Cancer Institute, Boston, Massachusetts (Wendy Y. 10. Wilson KM, Vesper HW, Tocco P, et al. Validation of a food
Chen). frequency questionnaire measurement of dietary acrylamide in-
This work was supported by a grant from the National take using hemoglobin adducts of acrylamide and glycidamide.
Cancer Institute (CA050385) and by a National Cancer Cancer Causes Control. 2008. (doi: 10.1007/s10552-008-9241-7).
(http://www.springerlink.com/content/5470507675718735/
Institute/National Institutes of Health training grant (T32
fulltext.pdf).
CA09001 to K. M. W.) 11. Salvini S, Hunter DJ, Sampson L, et al. Food-based validation
Conflict of interest: none declared. 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
REFERENCES receptor results from pathology reports with results from
central laboratory testing. J Natl Cancer Inst. 2008;100(3):
1. International Agency for Research on Cancer. Monographs on 218–221.
the Evaluation of Carcinogenic Risks to Humans. Vol 60. 13. Cho E, Spiegelman D, Hunter DJ, et al. Premenopausal fat
Lyon, France: International Agency for Research on Cancer; intake and risk of breast cancer. J Natl Cancer Inst. 2003;
1994. 95(14):1079–1085.
2. Tareke E, Rydberg P, Karlsson P, et al. Analysis of acrylamide, 14. Cho E, Spiegelman D, Hunter DJ, et al. Premenopausal dietary
a carcinogen formed in heated foodstuffs. J Agric Food Chem. carbohydrate, glycemic index, glycemic load, and fiber in re-
2002;50(17):4998–5006. lation to risk of breast cancer. Cancer Epidemiol Biomarkers
3. Petersen BJ, Tran N. Exposure to acrylamide: placing Prev. 2003;12(11 pt 1):1153–1158.
exposure in context. Adv Exp Med Biol. 2005;561: 15. Pelucchi C, Galeone C, Levi F, et al. Dietary acrylamide and
63–76. human cancer. Int J Cancer. 2006;118(2):467–471.
4. DiNovi M. The 2006 Exposure Assessment for Acrylamide. 16. Pelucchi C, Franceschi S, Levi F, et al. Fried potatoes and
Rockville, MD: Center for Food Safety and Applied Nutrition, human cancer. Int J Cancer. 2003;105(4):558–560.
Food and Drug Administration; 2006. (http://www.cfsan.fda. 17. Willett WC. Nutritional Epidemiology. New York, NY: Oxford
gov/~dms/acryexpo.html). University Press; 1998.
5. Johnson KA, Gorzinski SJ, Bodner KM, et al. Chronic toxicity 18. Stadler RH, Scholz G. Acrylamide: an update on current
and oncogenicity study on acrylamide incorporated in the knowledge in analysis, levels in food, mechanisms of forma-
drinking water of Fischer 344 rats. Toxicol Appl Pharmacol. tion, and potential strategies of control. Nutr Rev. 2004;62(12):
1986;85(2):154–168. 449–467.

Am J Epidemiol 2009;169:954–961
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 of alcohol intake or cigarette smoking, 2 potentially modifi-
cancer, a primary breast cancer occurring in the opposite able risk factors, in the development of asynchronous contra-
breast subsequent to a first breast cancer diagnosis in 1 breast, lateral breast cancer. Given the high level of risk in these
in female survivors of breast cancer is considerably higher women, clarifying the role of modifiable risk factors in asyn-
than the risk of developing a first primary breast cancer in chronous contralateral breast cancer in women with breast
unaffected women (1). The population of women with breast cancer is important.
cancer have a higher prevalence of all breast cancer risk A number of meta-analyses confirm that alcohol is a risk
factors, both genetic and nongenetic, than women without factor for both pre- and postmenopausal first primary breast
breast cancer. It would be expected that women who go on cancer, although the magnitude of increased risk associated
to develop asynchronous contralateral breast cancer would with consuming 1 or more drinks per day is moderate (2–4).
have an even higher prevalence of risk factors than those Recent studies confirm this finding (5, 6). Thus far, a signif-
who develop only 1 primary, although there may be mitigat- icant association between alcohol intake and second pri-
ing factors such as the treatment received for the first primary mary breast cancer has not been observed (7–9), although
and changes in behavior. Few studies have examined the role 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


Alcohol and Smoking and Contralateral Breast Cancer 963

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

Am J Epidemiol 2009;169:962–968
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- increased risk associated with drinking specifically after
sented. We performed tests for trend across categories of the first diagnosis (the at-risk period) did not achieve statis-
duration and consumption. All analyses were conducted using tical significance (RR ¼ 1.2, 95% CI: 0.9, 1.5). The risk
SAS, version 9.1, software (SAS Institute, Inc., Cary, North increased with increasing lifetime duration of drinking
Carolina), and a 2-sided P < 0.05 was considered significant. (Ptrend ¼ 0.03). We observed no apparent trend associated
with the average amount of alcohol consumed. The risk of
RESULTS asynchronous contralateral breast cancer increased with
later initiation of drinking (at age 20 or more years) but
Table 1 shows selected characteristics of the WECARE not with early initiation (before age 20 years). We did not
Study population. The 2 groups were similar on all matched observe any differences in the risks associated with the du-
characteristics. Women with asynchronous contralateral ration of drinking prior to the first pregnancy, adjusted for
breast cancer were more likely to have a family history of the age at first full-term pregnancy: began drinking after first
breast cancer. The majority (67%) of second diagnoses oc- pregnancy (RR ¼ 1.5, 95% CI: 1.1, 2.1); drank <7 years
curred in 1995 or later. In addition, 56% of cases and 49% of before first pregnancy (RR ¼ 1.4, 95% CI: 1.0, 1.9); drank
controls were interviewed within 5 years of the reference 7 years before first pregnancy (RR ¼ 1.2, 95% CI: 0.8,
date, and 88% of cases and 86% of controls were inter- 1.7), versus parous and never drank. There was no evidence
viewed within 10 years. Table 2 shows that ever regular for any association with smoking (Table 3) nor any consis-
drinking was associated with an elevated risk of developing tent pattern with smoking before the first pregnancy, ad-
asynchronous contralateral breast cancer (rate ratio (RR) ¼ justed for age at first full-term pregnancy: began smoking
1.3, 95% confidence interval (CI): 1.0, 1.6), although the 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, with an increased risk of asynchronous contralateral breast
1.4); smoked 6 years before first pregnancy (RR ¼ 1.3, cancer. We did not find evidence that smoking cigarettes
95% CI: 1.0, 1.8), versus parous and never smoked. We did increased the risk of this disease. Our results differ from
not observe any difference in the effect of smoking or alco- previous studies that did not observe an increased risk of
hol by categories of radiation exposure or any difference in asynchronous contralateral breast cancer associated with
the effect of alcohol by tamoxifen use (data not shown). We alcohol (7–9). However, as with a first primary breast can-
also did not observe consistent differences in results across cer, the effect of alcohol is modest and could be missed if the
categories when we conducted the analysis by the time be- sample size was insufficient. Previous studies included 488,
tween reference date and interview in 3 categories, 0–5 77, or 136 cases (7–9) compared with the 708 cases in our
years, 6–10 years, and >10 years (data not shown). study. Varying definitions and prevalences of drinking may
also have contributed to the inconsistency among studies. In
DISCUSSION the largest previous study, only information on recent drink-
ing prior to the first diagnosis was available, and the authors
In the WECARE Study, we found that consuming alco- were unable to evaluate lifetime duration of drinking (7).
hol, particularly over longer periods of time, was associated We did not find a relation with reported average amount

Am J Epidemiol 2009;169:962–968
Alcohol and Smoking and Contralateral Breast Cancer 967

consumed, but the women were asked to average their if drinking and/or smoking adversely affects survival, and if
lifetime consumption, and changes in intake were not cap- fewer women who drink and/or smoke survive to be diag-
tured, including changes occurring around the time of the nosed with a second primary or to be included in the study
first diagnosis. Relatively few women in this population after the second diagnosis, this could affect the relative risks
(12%) reported consuming 1 drink per day or more on reported here.
average. Women with a first primary unilateral breast cancer have
Previous studies of smoking and asynchronous contralat- an elevated risk of developing cancer in the contralateral
eral breast cancer have yielded inconsistent results (7–9, 16, breast. Although we did not observe an increased risk of
17), similar to studies of first primary breast cancer (11, 12). asynchronous contralateral breast cancer associated with
As with alcohol, this inconsistency may be due to issues of cigarette smoking, there are many other reasons to quit
study design, such as variation in sample size, smoking smoking including reducing the risk of smoking-related
definitions, and smoking prevalences. Our results from the cancer or heart disease. Alcohol appears to be associated
WECARE Study do not support the hypothesis that smoking with an increased risk of asynchronous contralateral breast
is a risk factor for asynchronous contralateral breast cancer, 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 ACKNOWLEDGMENTS
changed their drinking status after their first breast cancer
diagnosis (10% of cases and 9% of controls), but the pro- Author affiliations: Samuel Lunenfeld Research Institute,
portion who stopped smoking was greater (28% of cases and Mount Sinai Hospital, Toronto, Canada (Julia A. Knight);
29% of controls). City of Hope, Duarte, California (Leslie Bernstein); Univer-
The WECARE Study included women who were aged sity of California at Irvine, Irvine, California (Joan Largent);
less than 55 years at first diagnosis (mean age, 45 years Memorial Sloan Kettering Cancer Center, New York, New
among controls and 46 years among cases), although they York (Marinela Capanu, Colin Begg, Anne S. Reiner,
were somewhat older at the reference date (mean age in both Xiaolin Liang, Jonine L. Bernstein); Danish Cancer Society,
groups, 51 years). These women were younger than those in Copenhagen, Denmark (Lene Mellemkjær); University of
some other studies (7) but not others (8). However, there is Iowa, Iowa City, Iowa (Charles F. Lynch); Fred Hutchinson
no evidence that the effect of smoking or alcohol on the risk Cancer Research Center, Seattle, Washington (Kathleen E.
of first primary breast cancer differs by menopausal status Malone); University of Southern California, Los Angeles,
(2–4, 10). California (Robert W. Haile); and International Epidemiol-
An important strength of the WECARE Study is that it is ogy Institute, Rockville, Maryland, and Vanderbilt Univer-
the largest case-control study conducted of asynchronous con- sity, Nashville, Tennessee (John D. Boice, Jr.).
tralateral breast cancer to date that includes direct patient This work was supported by the National Institutes of
interview. We have also been able to detect other associa- Health (grants U01 CA83178 and R01 CA97397).
tions in their expected directions for fewer full-term preg- WECARE Study Collaborative Group: Memorial Sloan
nancies and early menarche (19), radiation treatment (23), Kettering Cancer Center (New York, New York): Jonine L.
and treatment of the first primary breast cancer with chemo- Bernstein (WECARE Study Principal Investigator), Colin
therapy or tamoxifen (24). However, although we did cap- Begg, Marinela Capanu, Xiaolin Liang, Anne S. Reiner,
ture some information on lifetime duration of smoking and Irene Orlow, Tracy Layne; City of Hope (Duarte, Califor-
alcohol consumption, the level of detail of the information nia): Leslie Bernstein (subcontract Principal Investigator),
collected was limited. We did not capture the changes in the Laura Donnelly-Allen (some work performed at University
patterns of consumption at various times of life. Although of Southern California, Los Angeles, California); Danish
recall bias is a problem in case-control studies in general, Cancer Society (Copenhagen, Denmark): Jørgen H. Olsen
whether or not it is an issue in our study is unclear, as both (subcontract Principal Investigator), Michael Andersson,
the cases and controls have been affected by breast cancer. Lisbeth Bertelsen, Per Guldberg, Lene Mellemkjær; Fred
In this study, cases and controls were matched on time since Hutchinson Cancer Research Center (Seattle, Washington):
the first diagnosis, and the majority (56%) of cases were Kathleen E. Malone (subcontract Principal Investigator),
interviewed within 5 years and most (88%) within 10 years Noemi Epstein; International Epidemiology Institute
of the second diagnosis. It is likely that women can recall (Rockville, Maryland) and Vanderbilt University (Nashville,
accurately whether they drank or smoked, and the broad Tennessee): John D. Boice, Jr. (subcontract Principal Investi-
categories used in the analysis of duration of drinking and gator), David P. Atencio; National Cancer Institute (Bethesda,
smoking minimize the potential for misclassification errors Maryland): Daniela Seminara; New York University (New
in recall of when drinking and smoking began and ended. As York, New York): Roy E. Shore (subcontract Principal Inves-
with amount, we were unable to examine duration in greater tigator); University of California at Irvine (Irvine, California):
detail. We were also unable to consider results by estrogen Hoda Anton-Culver (subcontract Principal Investigator),
and progesterone status, as a considerable proportion of the Joan Largent (Coinvestigator); University of California at
cases were missing information on receptor status. Results Los Angeles (Los Angeles, California): Richard A. Gatti
from studies relating alcohol to first breast cancer according (Coinvestigator); University of Iowa (Iowa City, Iowa):
to hormone receptor status are inconsistent (10, 25). Further, Charles F. Lynch (subcontract Principal Investigator), Jeanne

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

DeWall; University of Southern California (Los Angeles, Cal- among women diagnosed with a first primary breast cancer.
ifornia): Robert W. Haile (subcontract Principal Investigator), Am J Epidemiol. 1992;136(8):925–936.
Bryan M. Langholz (Coinvestigator), Duncan C. Thomas 10. Dumitrescu RG, Shields PG. The etiology of alcohol-induced
(Coinvestigator), Anh T. Diep (Coinvestigator), Shanyan breast cancer. Alcohol. 2005;35(3):213–225.
11. Terry PD, Rohan TE. Cigarette smoking and the risk of breast
Xue, Nianmin Zhou, Yong Liu, Evgenia Ter-Karapetova;
cancer in women: a review of the literature. Cancer Epidemiol
University of Southern Maine (Portland, Maine): W. Douglas Biomarkers Prev. 2002;11(10 pt 1):953–971.
Thompson (subcontract Principal Investigator); University of 12. Morabia A. Smoking (active and passive) and breast cancer:
Texas, M. D. Anderson Cancer Center (Houston, Texas): epidemiologic evidence up to June 2001. Environ Mol Mutagen.
Marilyn Stovall (subcontract Principal Investigator), Susan 2002;39(2-3):89–95.
Smith (Coinvestigator); and University of Virginia (Charlot- 13. Reynolds P, Hurley S, Goldberg DE, et al. Active smoking,
tesville, Virginia): Patrick Concannon (subcontract Principal household passive smoking, and breast cancer: evidence from
Investigator), Sharon Teraoka (Coinvestigator), Eric R. Olson the California Teachers Study. J Natl Cancer Inst. 2004;96(1):
(some work performed at Benaroya Research Institute at Vir- 29–37.
ginia Mason, Seattle, Washington), V. Anne Morrison, Karen 14. Ha M, Mabuchi K, Sigurdson AJ, et al. Smoking cigarettes
before first childbirth and risk of breast cancer. Am J Epidemiol.
M. Cerosaletti, Lemuel Navarro, Jocyndra Wright.
2007;166(1):55–61.
Conflict of interest: none declared. 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
REFERENCES after conservative surgery and radiation for stages I–II
breast cancer: identifying a subset of women at increased
1. Chen Y, Thompson W, Semenciw R, et al. Epidemiology of risk. Int J Radiat Oncol Biol Phys. 2001;51(3):679–690.
contralateral breast cancer. Cancer Epidemiol Biomarkers 17. Horn PL, Thompson WD. Risk of contralateral breast cancer:
Prev. 1999;8(10):855–861. associations with factors related to initial breast cancer. Am J
2. Ellison RC, Zhang Y, McLennan CE, et al. Exploring the re- Epidemiol. 1988;128(2):309–323.
lation of alcohol consumption to risk of breast cancer. Am J 18. Bernstein JL, Langholz B, Haile RW, et al. Study design:
Epidemiol. 2001;154(8):740–747. evaluating gene–environment interactions in the etiology of
3. Smith-Warner SA, Spiegelman D, Yaun SS, et al. Alcohol and breast cancer—the WECARE Study. Breast Cancer Res. 2004;
breast cancer in women: a pooled analysis of cohort studies. 6(3):R199–R214.
JAMA. 1998;297(7):535–540. 19. Largent J, Capanu M, Bernstein L, et al. Reproductive history
4. Collaborative Group on Hormonal Factors in Breast Cancer. and risk of second primary breast cancer: the WECARE Study.
Alcohol, tobacco and breast cancer—collaborative reanalysis Cancer Epidemiol Biomarkers Prev. 2007;16(5):906–911.
of individual data from 53 epidemiological studies, including 20. Langholz B, Borgan Ø. Counter-matching: a stratified nested
58515 women with breast cancer and 95067 women without case-control sampling method. Biometrika. 1995;82(1):69–79.
the disease. Br J Cancer. 2002;87(11):1234–1245. 21. Cox DR. Regression models and life tables. J R Stat Soc (B).
5. Berstad P, Ma H, Bernstein L, et al. Alcohol intake and breast 1972;34:187–202.
cancer among young women. Breast Cancer Res Treat. 2007; 22. Borgan Ø, Goldstein L, Langholz B. Methods for the analysis
108(1):113–120. of sampled cohort data in the Cox proportional hazards model.
6. Tjønneland A, Christensen J, Olsen A, et al. Alcohol intake Ann Statist. 1995;23(5):1749–1778.
and breast cancer risk: the European Prospective Investigation 23. Stovall M, Smith SA, Langholz BM, et al. Dose to the con-
into Cancer and Nutrition (EPIC). Cancer Causes Control. tralateral breast from radiotherapy and risk of second primary
2007;18(4):361–373. breast cancer in the WECARE Study. Int J Radiat Oncol Biol
7. Trentham-Dietz A, Newcomb PA, Nichols H, et al. Breast Phys. 2008;72(4):1021–1030.
cancer risk factors and second primary malignancies among 24. Bertelsen L, Bernstein L, Olsen JH, et al. Effect of systemic
women with breast cancer. Breast Cancer Res Treat. 2007; adjuvant treatment on risk for contralateral breast cancer in the
105(2):195–207. Women’s Environment, Cancer and Radiation Epidemiology
8. Li CI, Malone KE, Porter PL, et al. Epidemiologic and Study. J Natl Cancer Inst. 2008;100(1):32–40.
molecular risk factors for contralateral breast cancer 25. Althuis MD, Fergenbaum JH, Garcia-Closas M, et al. Etiology
among young women. Br J Cancer. 2003;89(3):513–518. of hormone receptor-defined breast cancer: a systematic re-
9. Bernstein JL, Thompson WD, Risch N, et al. Risk factors view of the literature. Cancer Epidemiol Biomarkers Prev.
predicting the incidence of second primary breast cancer 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- universally shared. Hartge et al. argued that the evidence
phoma by a number of investigators. In a review of this suggests that ionizing radiation probably causes lymphoma
literature, Boice (1) concluded that the evidence of associ- (5) and observed that high doses of ionizing radiation appear
ation between ionizing radiation and non-Hodgkin’s lym- to be associated with lymphoma risk in some studies of
phoma (NHL) is extremely weak and that there is no radiotherapy (6).
evidence of association between radiation and Hodgkin’s Lack of a consistent association between ionizing radia-
disease. The United Nations Scientific Committee on the tion and lymphoma could mean that there is no causal re-
Effects of Atomic Radiation noted that studies of NHL fol- lation or that a causal relation is obscured by bias or
lowing external exposure to ionizing radiation have yielded deficiencies in exposure measurement, case classification,
mixed results and concluded that overall there is little evi- duration of follow-up, or some combination of these factors.
dence of an association between NHL and external exposure Given that lymphoma is often an indolent disease, long-term
to ionizing radiation (2). Ron (3) reached a similar conclu- studies of radiation-exposed populations may be needed to
sion, noting that evidence of association between radiation observe an effect. The development of nuclear weapons in
and NHL has been inconsistent and Hodgkin’s disease has the early 1940s led to 2 types of epidemiologic studies that
rarely been related to radiation exposure; and Melbye and can now provide evidence regarding the radiation-
Trichopoulos (4) stated that there is no evidence that ioniz- lymphoma association: studies based on follow-up of work-
ing radiation causes NHL. However, this conclusion is not 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- above 4 Gy truncated to 4 Gy (12). For consistency with
posed to ionizing radiation from the use of nuclear weapons. analyses of the SRS cohort, dose estimates calculated as the
Most prominent among the latter is the Life Span Study sum of the c-radiation dose plus 10 times the neutron dose
(LSS), a study of Japanese survivors of the atomic bombings are expressed in sieverts; some recent reports on LSS ana-
of Hiroshima and Nagasaki. Radiation risk estimates from lyses refer to this quantity as the weighted dose in grays (13,
studies of nuclear workers are often compared with esti- 14). Interactions between radiation and lymphocytes may
mates from the LSS in order to evaluate the consistency of occur in the lymphatic or circulatory system at a variety
risk estimates in a population that includes people exposed of anatomic sites; the choice of target organ for dose esti-
to acute high doses with estimates from populations that are mation may depend on the characteristics of the lymphoma,
chronically exposed to low doses (7–9). including anatomic location (15, 16). The colon dose has
We examined the association between ionizing radiation been taken as a representative dose to the organs involved at
and lymphoma mortality in a US occupational cohort and in a variety of anatomic locations, similar to the approach
a sample of LSS atomic bomb survivors and compared find- employed in prior analyses of solid cancers (17). The colon
ings from the 2 populations. Follow-up of each cohort com- dose estimate has been used by previous investigators as an
menced in 1950 and spanned approximately 5 decades. To estimate comparable to the quantity estimated by the radi-
the extent possible, we conducted these analyses as parallel ation dosimeters worn by nuclear industry workers (i.e., the
analyses employing comparable methods. We focused, in ‘‘deep dose’’).
particular, on variation in the associations between radiation For SRS workers, the exposure of interest was defined as
dose and lymphoma mortality by time since exposure. 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
MATERIALS AND METHODS by a factor of 10. Personal radiation monitoring data were
available for the period 1950–1999. Whole-body doses were
The LSS cohort includes 86,611 people who were alive at estimated for work-years with missing dose data using dose
the time of the 1950 Japanese census, reported being in estimates from adjacent time periods and average values for
Hiroshima or Nagasaki at the time of the bombings (August similar workers; estimated annual doses constituted 4% of
1945), and had dose estimates based on the DS02 dosimetry employment years for male workers (18).
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. Outcome definitions
The Savannah River Site (SRS) was constructed near In the LSS, underlying cause of death was coded accord-
Aiken, South Carolina, in 1950 as a facility to produce ing to the International Classification of Diseases, Ninth
materials for the US nuclear weapons program. A cohort Revision (ICD-9), which was issued in 1977. In the SRS
of 18,883 workers who were hired at the SRS prior to study, underlying cause of death was coded according to the
1987, who worked there for at least 90 days, who were Eighth Revision of the ICD (ICD-8) for deaths occurring
not known to have been employed at another US Depart- prior to 1979 and according to the ICD revision in effect at
ment of Energy facility, and who had complete information the time of death for deaths occurring in 1979 or later. (The
on name, Social Security number, sex, date of birth, and date Tenth Revision of the ICD (ICD-10) was issued in 1992.)
of hire was enumerated (11). Vital status and cause-of-death As in prior analyses (17, 19), we examined the broad
information were ascertained through December 31, 2002. category of malignant lymphoma (ICD-8 and ICD-9 codes
200–202; ICD-10 codes C81–C85). In addition, we exam-
Cohort restrictions for comparability ined the subcategory of NHL (ICD-8 and ICD-9 codes 200
and 202; ICD-10 codes C82–C85). There were too few
Since over 95% of the collective dose at SRS was in- deaths due to Hodgkin’s disease to support separate analyses
curred by males, there was little ability to estimate risk of that outcome in these cohorts.
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 Statistical methods
SRS workers terminated their employment by age 65 years,
Poisson regression methods were used. The analytical
LSS analyses were restricted to people who were aged 15–
data file for the LSS cohort consisted of a tabulation of
64 years at the time of the bombings. This resulted in a co-
person-time and numbers of deaths by city, age at exposure
hort of 15,264 male SRS workers and a cohort of 20,940
(in 5-year intervals), attained age (in 5-year intervals), cal-
male LSS subjects who were aged 15–64 years at the time of
endar time (1950–1952, 1953–1955, and then 5-year inter-
the bombings.
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–
Dosimetry data <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,
For the LSS, we used DS02 revised colon dose estimates 0.5–<0.75, 0.75–<1, 1–<1.25, 1.25–<1.5, 1.5–<1.75,
adjusted for dosimetry errors, with shielded kerma estimates 1.75–<2, 2–<2.5, 2.5–<3, and 3 Sv). The analytical data

Am J Epidemiol 2009;169:969–976
Ionizing Radiation and Lymphoma Mortality 971

file for the SRS cohort consisted of a tabulation of person- vide estimates of the ERR per 1-Sv dose during the periods
time and events by attained age (in 5-year intervals), race 5–25, 26–35, 36–45, and 46–55 years after the bombings. In
(black vs. other), year of birth (before 1915, 1915–1924, analyses of the SRS cohort, we fitted a model of the form
1925–1929, 1930–1934, 1935–1949, or 1950 or later), pay
code (paid monthly, weekly, or hourly), employment status rate ¼ edi ð1 þ /1 d1 þ /2 d2 þ /3 d3 Þ;
(employed, terminated within the last 2 years, or terminated
more than 2 years prior, classified separately for risk where d1–d3 represent the cumulative radiation doses ac-
ages <62 years and 62 years) (20–22), and dose crued in the exposure time windows 5–25, 26–35, and
(0, >0–<0.005, 0.005–<0.02, 0.02–<0.04, 0.04–<0.06, 0.06– 36 years prior to observation of a person-year or
<0.08, 0.08–<0.1, 0.1–<0.125, 0.125–<0.150, 0.150– event and /̂1 ; /̂2 ; and /̂3 provide associated estimates of
<0.175, 0.175–<0.2, 0.2–<0.25, 0.25–<0.3, and 0.3 Sv). the ERR per 1-Sv dose.
Covariate control was achieved through background strat- We estimated parameters using the EPICURE statistical
ification of regression models. In analyses of the LSS co- package (Hirosoft International Corporation, Seattle,
hort, the stratifying factors were attained age, age at Washington); for consistency with recent reports (2, 26),
exposure, and city; in analyses of the SRS cohort, the strat- we generated 90% confidence intervals for estimated param-
ifying factors were attained age, birth cohort, race, pay code, eters via the likelihood method (27). In some analyses, con-
and employment status. Radiation dose-mortality associa- fidence bounds could not be determined (designated ‘‘not
tions were estimated via a regression model of the form determined’’ (ND)). In order to aid interpretation of model
fittings, we report the 1-sided P value derived via a likeli-
rate ¼ eai ð1 þ b xÞ; hood ratio test (LRT) for each reported point estimate.
Tabulations of observed versus expected numbers of deaths
where ai indexes the stratum-specific mortality rate in the by category of cumulative dose are reported; we calculated
absence of radiation exposure and b̂ provides an estimate of expected counts for each cell of the person-time table using
the excess relative rate (ERR) per sievert (23, 24). a regression model that included all variables except the
In analyses of the LSS cohort, x represents the estimated dose term.
radiation dose delivered at the time of the bombings in
August 1945. Since follow-up of the LSS cohort began in RESULTS
October 1950, this implies a minimal lag of approximately
5 years between exposure and its effect. We also present re- With follow-up through 2000, 90 malignant lymphoma
sults from analyses in which we assumed that there was no deaths were observed among the male atomic bomb sur-
excess risk during the period 1950–1955; that is, a minimum vivors exposed at ages 15–64 years, including 6 deaths from
latency period of approximately 10 years was assumed. A Hodgkin’s disease (Table 1). Sixty-three malignant lym-
10-year lag assumption has been used in previous nuclear phoma deaths occurred among residents of Hiroshima (58
worker studies that examined lymphoma mortality (25, 26). due to NHL) and 27 malignant lymphoma deaths occurred
We refer to analyses of LSS data that examine excess mor- among residents of Nagasaki (26 due to NHL). No deaths
tality risk since 1950 and since 1956 as analyses carried out due to malignant lymphoma occurred among survivors at
under 5- and 10-year lag assumptions, respectively. In anal- attained ages less than 30 years. In the SRS cohort, 56
yses of the SRS cohort, x represents the cumulative radiation lymphoma deaths were observed; 5 of these deaths were
dose under a 5- or 10-year lag assumption. Lagging dose due to Hodgkin’s disease. One death due to malignant lym-
assignment by L years means that an increment of dose was phoma was observed among black males (it was a case of
included in the calculation of cumulative dose at time t if it NHL), and 18, 14, and 24 deaths due to malignant lym-
had been received at or before time t  L years; person-time phoma were observed among workers paid monthly,
and events at time t were then classified according to that weekly, and hourly, respectively. Three deaths due to ma-
category of lagged cumulative dose. lignant lymphoma occurred among actively employed SRS
The dose range in the LSS, 0–4 Sv, was wider than the dose workers (all were cases of NHL) and 6 deaths occurred
range in the SRS study (0–<0.5 Sv). In order to evaluate within 2 years of termination of employment (all were cases
dose-response associations over a comparable range of doses, of NHL), while the remaining 47 deaths due to malignant
we also conducted analyses based upon LSS data limited to lymphoma occurred 2 or more years after termination of
the 19,183 survivors with doses in the range of 0–<0.5 Sv. employment at SRS (42 due to NHL).
In analyses of the LSS cohort, we assessed variation in In the LSS, the estimated ERR of malignant lymphoma
radiation risk with time since exposure via a regression per sievert, under a 5-year lag assumption, was 0.79 (90%
model of the form 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
rate ¼ eai ð1 þ b1 xPeriod1 þ b2 xPeriod2 Hodgkin’s disease (Table 2). Under a 10-year lag assump-
þ b3 xPeriod3 þ b4 xPeriod4Þ; tion, these estimated associations were slightly larger in
magnitude. In the SRS study, the estimated ERRs of malig-
where Period1–Period4 are indicator variables for the cal- nant lymphoma per sievert under 5- and 10-year lag assump-
endar time periods 1950–1970, 1971–1980, 1981–1990, and tions were 6.99 (90% CI: 0.96, 18.39) and 8.18 (90% CI:
1991–2000, respectively. The values b̂1 ; b̂2 ; b̂3 ; and b̂4 pro- 1.44, 21.16), respectively. Upon exclusion of deaths due to

Am J Epidemiol 2009;169:969–976
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 following radiation exposure may be consistent with action
disease (results not shown). 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
DISCUSSION derived from analysis of the SRS cohort are larger than the
estimates derived from analysis of the LSS cohort (Table 2).
In a previous analysis of lymphoma mortality among sur- Differences in the magnitude and rate of exposure may in-
vivors in the LSS, Pierce et al. (17) reported evidence of fluence the comparability of dose-response estimates. These
a nonsignificant positive association with radiation dose cohorts also differ with regard to potential biases from con-
among males (ERR/Sv ¼ 0.27, 90% CI: ND, 1.49) and founding, selection, and exposure measurement error. While
a nonsignificant negative association among females it is not an established cause of NHL, benzene is suspected
(ERR/Sv ¼ 0.17, 90% CI: ND, 0.30). In those analyses, to be related to NHL (28). However, benzene was not used
a time-constant ERR model was fitted to mortality follow- in the production process at SRS, nor was it routinely used
up through 1990. In the present paper, time-window ana- as a degreaser. Plutonium-239 is a radiologic hazard at SRS.
lyses helped to explain the observation of a significant positive While a recent study suggested that the contribution of plu-
association between radiation dose and lymphoma mortality tonium doses to total dose estimates for these workers was
among male atomic bomb survivors with more recent fol- relatively small (29), we did not directly assess confounding
low-up, showing that positive associations have been ob- by plutonium exposure. Selection bias could have influ-
served only since 1980. Such findings suggest a protracted enced these estimates of association—for example, via the
induction and latency period. If considered within the ‘‘healthy worker’’ survivor effect (20). Although we ad-
framework of a multistage model of carcinogenesis, the justed for employment status, such an approach is sub-
relatively long empirical induction period for lymphoma optimal if employment status is an intermediate variable

Am J Epidemiol 2009;169:969–976
Ionizing Radiation and Lymphoma Mortality 975

as well as a confounder of the association of interest. How- Effects Research Foundation, Hiroshima, Japan (Susan Geyer);
ever, in studies of chronic diseases with long latency peri- Department of Epidemiology, Radiation Effects Research
ods, cumulative exposure will typically not appreciably Foundation, Nagasaki, Japan (Midori Soda, Akihiko Suyama);
influence employment termination rates; under such condi- and Radiation Effects Research Foundation, Hiroshima, Japan
tions, employment status will play a minor role as an (Kazunori Kodama).
intermediate variable but could have a strong role as a con- This project was supported in part by grant R01
founder of the association (22). Frequent reading of dosim- OH007871 from the US National Institute for Occupational
eters could have led to dose underestimation if dosimeters Safety and Health. Support was also provided by the Radi-
were not sufficiently exposed to reach a minimum detect- ation Effects Research Foundation (Hiroshima and Naga-
able dose. However, prior work suggests that the impact of saki, Japan), which is funded by the Japan Ministry of
this source of measurement error on estimates of radiation Health, Labour and Welfare and the US Department of En-
dose-response trends is modest (30–32). ergy, the latter partly through the National Academy of
Problems of bias could also influence estimates of Sciences.
radiation-mortality associations among atomic bomb survi- Conflict of interest: none declared.
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 REFERENCES
in Hiroshima and Nagasaki had low cumulative external
doses from fallout, with maximum estimates in the range 1. Boice JD Jr. Radiation and non-Hodgkin’s lymphoma. Cancer
of 0.2–0.4 Sv for several hundred people who were in an Res. 1992;52(19 suppl):5489s–5491s.
area of Nagasaki approximately 3 km from the hypocenter 2. United Nations Scientific Committee on the Effects of Atomic
(33, 34). Selective survival in the LSS cohort is another Radiation. Sources and Effects of Ionizing Radiation. New
concern and is a generic consideration when trying to un- York, NY: United Nations; 2000.
3. Ron E. Ionizing radiation and cancer risk: evidence from
derstand the temporal evolution of exposure-related risk
epidemiology. Radiat Res. 1998;150(5 suppl):S30–S41.
(35). A relation between short-term survival after the bomb- 4. Melbye M, Trichopoulos D. Non-Hodgkin’s lymphomas. In:
ings and later risk of lymphoma could lead to bias in Adami H-O, Hunter D, Trichopoulos D, eds. Textbook of
dose-response estimates. Evidence of selection has been Cancer Epidemiology. New York, NY: Oxford University
suggested by some empirical analyses (36, 37); however, Press; 2002:535–555.
values for the magnitude of dose-related selective survival 5. Hartge P, Smith MT. Environmental and behavioral factors and
assumed in a recent study suggested a modest potential for the risk of non-Hodgkin lymphoma. Cancer Epidemiol Bio-
bias in dose-response estimates (38). markers Prev. 2007;16(3):367–368.
These analyses provide evidence of a positive association 6. Hartge P, Wang SS, Bracci PM, et al. Non-Hodgkin lym-
between ionizing radiation dose and malignant lymphoma phoma. In: Schottenfeld D, Fraumeni JF, eds. Cancer Epide-
miology and Prevention. New York, NY: Oxford University
mortality among male Japanese atomic bomb survivors and Press; 2006:898–918.
SRS workers. We did not address risk estimates for females, 7. Cardis E, Vrijheid M, Blettner M, et al. Risk of cancer after
for whom there was no evidence of a positive association low doses of ionising radiation: retrospective cohort study in
between radiation dose and lymphoma mortality in follow- 15 countries. BMJ. 2005;331(7508):77.
up through 1990 (17). The radiation-NHL mortality associ- 8. Muirhead CR, Goodill AA, Haylock RG, et al. Occupational
ations among these male atomic bomb survivors and SRS radiation exposure and mortality: second analysis of the Na-
workers are of larger magnitude than the estimate reported tional Registry for Radiation Workers. J Radiol Prot. 1999;
in a 15-country study of nuclear workers (under a 10-year 19(1):3–26.
lag assumption, ERR/Sv ¼ 0.44, 90% CI: <0, 4.78) (7); 9. Schubauer-Berigan M, Daniels RD, Fleming DA, et al. Risk of
however, in the current analyses, positive dose-response as- chronic myeloid and acute leukemia mortality after exposure
to ionizing radiation among workers at four U.S. nuclear
sociations were primarily observed more than 35 years after weapons facilities and a nuclear naval shipyard. Radiat Res.
irradiation. These findings underscore the importance of 2007;167(2):222–232.
continued follow-up of the LSS cohort and nuclear worker 10. Preston DL, Pierce DA, Shimizu Y, et al. Effect of recent
cohorts. 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):
ACKNOWLEDGMENTS
1015–1022.
Author affiliations: Department of Epidemiology, School 12. Pierce DA, Stram DO, Vaeth M. Allowing for random errors in
radiation dose estimates for the atomic bomb survivor data.
of Public Health, University of North Carolina at Chapel Radiat Res. 1990;123(3):275–284.
Hill, Chapel Hill, North Carolina (David Richardson, Steve 13. Preston DL, Ron E, Tokuoka S, et al. Solid cancer incidence in
Wing); Department of Epidemiology, Radiation Effects Re- atomic bomb survivors: 1958–1998. Radiat Res. 2007;
search Foundation, Hiroshima, Japan (Hiromi Sugiyama, 168(1):1–64.
Ritsu Sakata, Eric J. Grant, Yukiko Shimizu, Nobuo Nishi, 14. International Commission on Radiological Protection. 1990
Fumiyoshi Kasagi); Department of Statistics, Radiation Recommendations of the International Commission on

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

Radiological Protection. Oxford, United Kingdom: Pergamon 26. Cardis E, Vrijheid M, Blettner M, et al. The 15-Country
Press; 1991. Collaborative Study of Cancer Risk among Radiation Workers
15. National Institute for Occupational Safety and Health. in the Nuclear Industry: estimates of radiation related cancer
Changes to the dose reconstruction target organ selection for risks. Radiat Res. 2007;167(4):396–416.
lymphoma under the Energy Employees Occupational Illness 27. Preston DL, Lubin JH, Pierce DA, et al. Epicure: User’s
Compensation Program Act of 2000. Fed Regist. 2006;71(31): Guide. Seattle, WA: Hirosoft International Corporation; 1993.
7969–7970. 28. Smith MT, Jones RM, Smith AH. Benzene exposure and risk
16. Office of Compensation Analysis and Support, National In- of non-Hodgkin lymphoma. Cancer Epidemiol Biomarkers
stitute for Occupational Safety and Health. Selection for In- Prev. 2007;16(3):385–391.
ternal and External Dosimetry Target Organs for Lymphatic/ 29. Daniels RD, Lodwick CJ, Schubauer-Berigan MK, et al.
Hematopoietic Cancers. (Technical Information Bulletin). Assessment of plutonium exposures for an epidemiological
Cincinnati, OH: National Institute for Occupational Safety and study of US nuclear workers. Radiat Prot Dosimetry. 2006;
Health; 2006. (Document no. OCAS-TIB-0012). 118(1):43–55.
17. Pierce DA, Shimizu Y, Preston DL, et al. Studies of the mor- 30. Daniels RD, Yiin JH. A comparison of statistical methods for
tality of atomic bomb survivors. Report 12, part I. Cancer: estimation of less than detectable ionising radiation exposures.
1950–1990. Radiat Res. 1996;146(1):1–27. Radiat Prot Dosimetry. 2006;121(3):240–51.
18. Richardson DB, Wing S, Daniels RD. Evaluation of external 31. Daniels RD, Schubauer-Berigan MK. Bias and uncertainty of
radiation dosimetry records at the Savannah River Site, penetrating photon dose measured by film dosimeters in an
1951–1989. J Expo Sci Environ Epidemiol. 2007;17(1): epidemiological study of US nuclear workers. Radiat Prot
13–24. Dosimetry. 2005;113(3):275–289.
19. Shimizu Y, Kato H, Schull WJ. Studies of the mortality of 32. Shin H, Ramsay T, Krewski D, et al. The effect of censoring on
A-bomb survivors. 9. Mortality, 1950–1985: part 2. Cancer cancer risk estimates based on the Canadian National Dose
mortality based on the recently revised doses (DS86). Radiat Registry of occupational radiation exposure. J Expo Anal
Res. 1990;121(2):120–141. Environ Epidemiol. 2005;15(5):398–406.
20. Arrighi HM, Hertz-Picciotto I. The evolving concept of the 33. Committee for the Compilation of Materials on Damage
healthy worker survivor effect. Epidemiology. 1994;5(2): Caused by the Atomic Bombs in Hiroshima and Nagasaki.
189–196. Hiroshima and Nagasaki: The Physical, Medical, and Social
21. Steenland K, Stayner L. The importance of employment status Effects of the Atomic Bombings. Tokyo, Japan: Iwanami
in occupational cohort mortality studies. Epidemiology. 1991; Shoten, Publishers; 1981.
2(6):418–423. 34. Radiation Effects Research Foundation. Frequently Asked
22. Steenland K, Deddens J, Salvan A, et al. Negative bias in Questions. Hiroshima, Japan: Radiation Effects Research
exposure-response trends in occupational studies: modeling Foundation; 2007. (http://www.rerf.or.jp/general/qa_e/
the healthy workers survivor effect. Am J Epidemiol. 1996; qa12.html). (Accessed June 30, 2008).
143(2):202–210. 35. Flanders WD, Klein M. Properties of 2 counterfactual effect
23. Thomas D. General relative-risk models for survival time and definitions of a point exposure. Epidemiology. 2007;18(4):
matched case-control analysis. Biometrics. 1981;37(4): 453–460.
673–686. 36. Stewart AM, Kneale GW. A-bomb survivors: factors that may
24. Greenland S. Introduction to regression models. In: lead to a re-assessment of the radiation hazard. Int J Epide-
Rothman K, Greenland S, eds. Modern Epidemiology. 2nd ed. miol. 2000;29(4):708–714.
Philadelphia, PA: Lippincott, Williams & Wilkins; 1998: 37. Shimizu Y, Pierce DA, Preston DL, et al. Studies of the mor-
359–399. tality of atomic bomb survivors. Report 12, part II. Noncancer
25. Gilbert ES, Cragle DL, Wiggs LD. Updated analyses of mortality: 1950–1990. Radiat Res. 1999;152(4):374–389.
combined mortality data for workers at the Hanford Site, Oak 38. Pierce DA, Vaeth M, Shimizu Y. Selection bias in cancer risk
Ridge National Laboratory, and Rocky Flats Weapons Plant. estimation from A-bomb survivors. Radiat Res. 2007;167(6):
Radiat Res. 1993;136(3):408–421. 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 studies typically involve tens of thousands of enrollees,
association between energy consumption and the incidence a self-administered, machine-readable food frequency ques-
and mortality from cancer. Among women, associations tionnaire (FFQ) has been the principal dietary assessment
were reported for breast, colon, rectal, endometrial, ovarian, tool in cohort studies.
and kidney cancer (1). Rodent feeding experiments indicate However, like other dietary assessment methods, the mea-
that underfeeding typically inhibits the development of site- surement properties of FFQs remain substantially unknown.
specific and overall cancer (2, 3). Comparison of FFQ assessments with food records reveals
Analytical epidemiologic studies of diet, nutrition, and noteworthy differences (4) that imply an important error
cancer date to the 1970s. Initial case-control studies used component to self-reported nutrient intake. Small-scale
a range of dietary assessment procedures, including food studies using a doubly labeled water biomarker (5) of en-
records, recalls, and frequencies. Concern about dietary re- ergy consumption suggest important systematic biases also,
call bias subsequently led to cohort studies as the predom- as obese persons may systematically underreport energy
inant design for dietary association studies. Because these 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 protein consumption was not ‘‘probably or convincingly’’
and cancer associations. It is important to examine nutrient related to the risk of any cancer (7, p. 394).
and disease associations in a manner that appropriately ac- Good-quality biomarkers of both total energy consump-
commodates FFQ measurement errors. tion (5) and protein consumption (9) have been developed
The accumulated data on diet and cancer were reviewed but, for cost and logistics reasons, have received little use in
by an international panel of experts in 1997 (7). Rather few epidemiologic research. These biomarkers involve urinary
‘‘definite’’ or ‘‘probable’’ dietary associations emerged. The recovery of metabolites produced when these nutrients are
authors wrote, ‘‘The significance of the data on energy in- expended. In weight-stable persons, they provide objective
take and cancer risk in humans remains unclear’’ (7, p. 371), estimates of short-term energy and protein consumption.
and ‘‘In the view of the panel, the effect of energy intake on The associated measurement error plausibly adheres to
cancer is best assessed by examining the data on related a simple classical measurement model,
factors: rate of growth, body mass, and physical activity’’
(7, p. 371). This state of affairs has evidently not changed in W ¼ Z þ e; ð1Þ
the intervening decade (8) and reflects considerable uncer-
tainty about energy consumption estimates and related as- where Z is the targeted (log-transformed) nutrient consump-
sociation study findings. The 1997 panel also assessed that 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 Q, where V is a set of characteristics that may relate to
be independent of Z and of all other study subject character- systematic bias in the assessment, r is a person-specific error
istics. The cost to ascertain these biomarkers for each par- variable that will be present in each self-reported assessment
ticipant in a cohort study would be excessive. Instead, for a study subject, and u is an independent measurement
a substudy that includes both the biomarker and FFQ can error term. In addition, S0, S1, S2, and S3 are constants to be
be used to produce calibrated consumption estimates for all estimated, and all variables on the right sides of equations 1
cohort members. and 2 are assumed to be independent, given V.
The measurement model for the self-reported data typi- We have recently reported (12) FFQ measurement error
cally needs to be more complex than the classical measure- findings from the Nutritional Biomarker Study among 544
ment model (equation 1). Other factors, such as body mass, women enrolled in the Women’s Health Initiative (WHI)
ethnicity, and age, may affect the assessment, and measure- dietary modification (DM) trial. FFQ estimates of energy,
ment errors may be correlated if the assessment is repeated protein, and percentage of energy from protein were each
for specific study subjects. Hence, we consider the measure- found to incorporate important systematic bias, and corre-
ment model (10, 11), sponding calibration equations were developed. Here, we
use these equations to produce calibrated estimates of en-
Q ¼ S0 þ S1 Z þ S2 V þ S3 VZ þ r þ u; ð2Þ ergy, protein, and percentage of energy from protein for
women in the DM trial comparison (control) group and
for the (log-transformed) self-reported nutrient assessment 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- and to produce calibrated consumption estimates for energy
tions between calibrated nutrient consumption and cancer and protein. The eligibility and recruitment methods for the
incidence as observed during WHI follow-up. Cancer risk Nutrient Biomarker Study have been described (12); 544
among DM intervention group women may depend in representative women from the DM trial cohort were en-
a complex manner on baseline and follow-up dietary pat- rolled (276 comparison group, 268 intervention group).
terns, so that intervention group women were excluded from These weight-stable women participated in a doubly labeled
the present analyses. 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,
MATERIALS AND METHODS and also provided a concurrent FFQ and other questionnaire
Study cohorts data. Twenty percent (n ¼ 111) repeated the entire Nutrient
Biomarker Study protocol an average of 6 months later to
Detailed accounts of the design of the WHI Clinical Trial provide reliability data for measurement error component
and Observational Study and of the DM trial findings have estimation (12). FFQ total energy and protein were found to
been presented (13–18). This paper uses a subset of women be underestimated, while the percentage of energy from
assigned to the DM trial comparison group (n ¼ 29,294) and protein was overestimated. Women having high body mass
a subset of the observational study cohort (n ¼ 93,676). index (weight (kg)/height (m)2) and younger women under-
Both cohorts included only women who were 50–79 years estimated energy consumption to a comparatively greater
of age at recruitment (1993–1998), were postmenopausal, extent. Calibration equations were developed for each of
and had no medical condition associated with less than 3 energy, protein, and percentage of energy from protein by
years’ predicted survival. Both provided common core ques- linear regression of log-biomarker estimates on correspond-
tionnaires at baseline on medical history, reproductive history, ing log-FFQ estimates, body mass index, age, ethnicity,
family history, personal habits, psychosocial attributes, and and other factors (12). For example, the calibrated log-
food frequency (19, 20). energy consumption is given by 7.61 þ 0.062 (log-FFQ
DM trial women, who could be assigned to overlapping energy  7.27) þ 0.013 (body mass index  28.2) 
trials of postmenopausal hormone therapy and of calcium 0.005 (age  70.9 years), plus some less influential terms
and vitamin D supplementation, also satisfied additional involving ethnicity, family income, and physical activity.
exclusionary criteria. To maximize commonality with the DM intervention group assignment did not meet inclusion
DM cohort, the 76,987 observational study women consid- criteria for any of the 3 calibration equations.
ered here were those remaining after imposing additional
DM trial baseline exclusionary criteria as follows: prior Nutrient Biomarker Study application to WHI cohorts
history of breast cancer, colorectal cancer, or other cancer
(except nonmelanoma skin cancer) within the preceding Here, we apply these calibration equations to FFQ data
10 years; a stroke or myocardial infarction in the preceding that were collected earlier in the WHI and relate the cali-
6 months; severe hypertension (systolic blood pressure brated consumption estimates to subsequent cancer inci-
>200 mm or diastolic blood pressure >105 mm); already fol- dence. Doing so is complicated by the use of the FFQ in
lowing a low-fat diet; underweight (body mass index <18); or participant screening for the DM trial. The exclusion of
FFQ-reported daily energy of <600 kcal or >5,000 kcal. about 50% of the women who had baseline FFQ percent-
age of energy from fat of less than 32, in conjunction with
WHI food frequency questionnaire FFQ measurement error, implies that the baseline FFQ
percentage of energy from fat is overestimated in the
All DM trial and observational study women completed DM trial (by about 3% on average), with corresponding
FFQs at baseline. DM trial women repeated the FFQ at 1 estimates of energy likewise distorted. Observational
year following enrollment and approximately every 3 years study baseline estimates are distorted in the opposite di-
thereafter, while observational study women repeated the rection because many women screened out from the DM
FFQ at 3 years following enrollment. FFQs were provided trial enrolled in the observational study. In terms of equa-
in connection with visits to the 40 participating clinical tion 2, these distortions arise because women tend to meet
centers, where completeness and quality control checks the FFQ inclusion criteria when the independent random
were applied. The self-administered FFQ included 122 line error term (u) that attends a particular FFQ application is
items for individual foods/food groups and 19 adjustment positive. Later FFQs for a woman, following a sufficient
items regarding fat intake, as well as summary questions. period of time (e.g., 6 months) to avoid carry-over effects
Nutrition Data System for Research, version 2005, software on this measurement component, can be expected to be
(University of Minnesota, Minneapolis, Minnesota) was free of this measurement effect. Hence, our analyses rely
used to compute daily average nutrient consumption esti- on FFQs obtained at year 1 in the DM trial comparison
mates (21, 22). group and at year 3 in the observational study, and only
cancer diagnoses that follow these FFQ collections are
Nutrient Biomarker Study included in analyses. These FFQs were collected an aver-
age 6.5 years (DM trial comparison group) and 4 years
The WHI Nutrient Biomarker Study was conducted in (observational study) prior to the Nutrient Biomarker
2004–2005 to assess measurement properties of this FFQ 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 cancer outcome were stratified on baseline age in 5-year
estimated for total invasive cancer, as well as for invasive categories and, for the DM trial comparison group, also
cancers of the breast, colon, rectum, ovary, endometrium, on hormone therapy trial participation (active estrogen; es-
bladder, kidney, pancreas, and lung, and for lymphoma and trogen placebo; active estrogen plus progestin; estrogen plus
leukemia. The ovarian cancer analyses were restricted to progestin placebo; not randomized). Analyses that combine
women without bilateral oophorectomy at baseline, and the 2 cohorts stratify also on cohort. Analysis for specific
the endometrial cancer analysis was restricted to women cancer outcomes included standard risk factors in the
with a uterus at baseline. Cox regression model to control confounding, as shown in
DM comparison group women were queried twice per Appendix Table 1. Women having missing confounding
year, and observational study women annually, concerning factors were excluded from analysis.
diagnosis of any cancer other than nonmelanoma skin can- Principal analyses modeled the log-hazard ratio linearly
cer. Cancer reports were verified by medical record and on log-nutrient consumption, so that the hazard ratio for
pathology report review by centrally trained physician ad- a fractional increase in the nutrient is independent of the
judicators at participating clinical centers (23). consumption. For display purposes, we present hazard ratios
for a 20% increase in consumption. For a woman with me-
Statistical analyses dian consumption, a 20% increment corresponds to about
413 kcal of energy, 15 g of protein, or 2.9 units in percentage
Log-consumption estimates were calibrated directly from of energy from protein.
the biomarker assessments (equation 1) for the few women Usual Cox model standard error estimates were calcu-
included in the Nutrient Biomarker Study and for other lated for uncalibrated consumption regression coefficients.
women by using the calibration equations previously devel- A more complex standard error estimation procedure is
oped (12). needed for the calibrated consumption coefficients to ac-
Hazard ratio estimates were based on Cox regression knowledge uncertainty in the calibration parameter esti-
(24). Follow-up times extended from year 1 (DM trial com- mates, as well as in the ‘‘regression calibration’’ hazard
parison group) or year 3 (observational study) to the earliest ratio estimation procedure (11), which has been shown to
of cancer occurrence, death, lost to follow-up, or March 31, be free of practically important biases in extensive simula-
2005, when the intervention phase of WHI ended. To min- tion studies. A bootstrap procedure (500 bootstrap samples),
imize mammographic screening influences on results, the with bootstrap sampling stratified on cohort and member-
breast cancer analyses censored the follow-up time for ship in the Nutrient Biomarker Study and in the Nutrient
a woman the first time she exceeded 2 years without a mam- Biomarker Study reliability subset, was applied for cali-
mogram. The Cox model baseline hazard rates for each 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.

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
Figure 2. Estimated hazard ratios and 95% confidence intervals for estimates of energy, protein, or percentage of energy from
a 20% increase in protein consumption (g/day), from combined anal-
ysis of data from the Women’s Health Initiative dietary modification protein were not significantly associated with total invasive
trial comparison group and observational study, without and with bio- cancer incidence. In contrast, following biomarker calibra-
marker calibration of consumption, 1993–2005. Unfilled square, un- tion, the associations with total cancer incidence were
calibrated; filled circle, calibrated. 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
(HR ¼ 0.92, 95% CI: 0.85, 0.99 for a 20% increase in energy consumption was found to be positively associated
percentage of energy from protein). Results corresponding with the risk of breast, colon, endometrial, and kidney cancer,
to Figures 1–3 by quartile of calibrated consumption are whereas uncalibrated energy was not.
given in Appendix Tables 2–4. These comparisons suggest that systematic bias in dietary
The correlation coefficients for body mass index with log- assessment could have a profound effect on nutritional ep-
transformed energy, protein, and percentage of energy from idemiology findings. Total energy assessment is a recog-
protein in the combined cohorts were, respectively, 0.07, nized weak aspect of FFQs. Uncalibrated FFQs are
0.10, and 0.07 without calibration and 0.81, 0.46, and generally believed to be more reliable for nutrient density
0.12 following calibration. Hence, it may be difficult to than for absolute consumption estimates. However, bio-
distinguish between total energy and body mass index asso- marker calibration also qualitatively affected the findings
ciations, with total or site-specific cancer. Table 5 examines 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-
Figure 3. Estimated hazard ratios and 95% confidence intervals for calibrated energy gives a projected body mass index increase
a 20% increase in percentage of energy from protein, from combined of 9.2 units corresponding to a 20% increase in calibrated
analysis of data from the Women’s Health Initiative dietary modifica-
tion trial comparison group and observational study, without and with energy, suggesting, in conjunction with Table 5, that much of
biomarker calibration of consumption, 1993–2005. Unfilled square, the observed dependence of cancer incidence rates on total
uncalibrated; filled circle, calibrated. 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 between a high-energy diet and body fat accumulation will
dependence of cancer incidence rates on body mass index be needed to understand the mechanisms leading to elevated
can be explained by energy consumption associations with cancer risk with a high level of energy consumption. How-
these diseases. ever, regardless of whether body fat accumulation results
Our analyses yielded similar results when calibration from a history of high-energy consumption, or whether
equations were applied in the DM cohort where they were a high body mass leads to increased energy requirements,
derived and when exported to the observational study. How- or both, it is evident that a high body mass index is an
ever, this extrapolation is under near-optimal conditions as important aspect of total and site-specific cancer risk, and
the 2 cohorts were drawn from essentially the same popu- efforts to prevent obesity deserve a continued high priority
lations, with much commonality in eligibility and exclusion- in national cancer control efforts.
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- ACKNOWLEDGMENTS
ducted in 2004–2005, an average of about 6.5 years after the
1-year FFQ data collection for the DM trial comparison Author affiliations: Division of Public Health Sciences, Fred
group women and about 4 years on average after the 3-year Hutchinson Cancer Research Center, Seattle, Washington
FFQ data collection for observational study women. Our (Ross L. Prentice, Marian L. Neuhouser, Ruth E. Patterson,
application assumes that the calibration equations devel- Lesley F. Tinker); Biostatistics Research Branch, National
oped from Nutrient Biomarker Study data apply to FFQs Institute of Allergy and Infectious Diseases, Bethesda,
at these earlier time points. Moreover, the biomarker data Maryland (Pamela A. Shaw); Medical Research Council
provide consumption estimates over a rather short period of Dunn Human Nutrition Unit, University of Cambridge,
time (e.g., 6 months between initial and repeat applications Cambridge, United Kingdom (Sheila A. Bingham); Depart-
in the 20% subsample). However, dietary patterns are ex- ment of Epidemiology, University of Washington, Seattle,
pected to track over longer time periods for most women in Washington (Shirley A. A. Beresford); Kaiser Permanente
these cohorts. Division of Research, Oakland, California (Bette Caan);
On the substantive side, we observe strong positive asso- Stanford Prevention Research Center, Palo Alto, California
ciations between calibrated energy consumption and the risk (Marcia L. Stefanick); University of Tennessee Health Sciences
of total and certain site-specific cancers. There are also sug- Center, Memphis, Tennessee (Suzanne Satterfield); Depart-
gestions of a positive association between protein consump- ment of Nutritional Sciences, University of Arizona, Tucson,
tion and leukemia and an inverse association between Arizona (Cynthia A. Thomson); Department of Community
percentage of energy from protein and bladder cancer and Behavioral Health, University of Iowa, Iowa City, Iowa
(Figures 2 and 3) that would be worth examining in other (Linda Snetselaar); Medstar Research Institute, Washington,
settings. More comprehensive temporal data on the interplay 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 Ohio: Margery Gass; University of Florida, Gainesville/
Thomas). Jacksonville, Florida: Marian Limacher; University of
This work was supported by the National Heart, Lung, and Hawaii, Honolulu, Hawaii: David Curb; University of Iowa,
Blood Institute, National Institutes of Health, US Department Iowa City/Davenport, Iowa: Robert Wallace; University of
of Health and Human Services (contracts N01WH22110, Massachusetts/Fallon Clinic, Worcester, Massachusetts:
24152, 32100-2, 32105-6, 32108-9, 32111-13, 32115, 32118- Judith Ockene; University of Medicine and Dentistry of New
19, 32122, 42107-26, 42129-32, and 44221). Clinical Trials Jersey, Newark, New Jersey: Norman Lasser; University of
Registration: ClinicalTrials.gov identifier: NCT00000611. Dr. Miami, Miami, Florida: Mary Jo O’Sullivan; University of
Prentice’s work was partially supported by grant CA53996 from Minnesota, Minneapolis, Minnesota: Karen Margolis; Univer-
the National Cancer Institute. sity of Nevada, Reno, Nevada: Robert Brunner; University of
The authors thank the WHI investigators and staff for North Carolina, Chapel Hill, North Carolina: Gerardo Heiss;
their outstanding dedication and commitment. A full listing University of Pittsburgh, Pittsburgh, Pennsylvania: Lewis
of WHI investigators can be found at the following website: Kuller; University of Tennessee, Memphis, Tennessee: Karen
http://www.whi.org. C. Johnson; University of Texas Health Science Center, San
A list of key WHI investigators involved in this research Antonio, Texas: Robert Brzyski; University of Wisconsin,
follows. Program Office—National Heart, Lung, and Blood Madison, Wisconsin: Gloria E. Sarto; Wake Forest University
Institute, Bethesda, Maryland: Elizabeth Nabel, Jacques School of Medicine, Winston-Salem, North Carolina: Mara
Rossouw, Shari Ludlam, Linda Pottern, Joan McGowan, Vitolins; and Wayne State University School of Medicine/
Leslie Ford, Nancy Geller. Clinical Coordinating Cen- Hutzel Hospital, Detroit, Michigan: Susan Hendrix.
ters—Fred Hutchinson Cancer Research Center, Seattle, Decisions concerning study design, data collection and
Washington: Ross Prentice, Garnet Anderson, Andrea analysis, interpretation of results, preparation of the manu-
LaCroix, Charles L. Kooperberg, Ruth E. Patterson, Anne script, or submission of the manuscript for publication resided
McTiernan; Wake Forest University School of Medicine, with committees comprising WHI investigators that included
Winston-Salem, North Carolina: Sally Shumaker; Medical National Heart, Lung, and Blood Institute representatives.
Research Labs, Highland Heights, Kentucky: Evan Stein; Conflict of interest: none declared.
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, REFERENCES
Texas: Aleksandar Rajkovic; Brigham and Women’s Hospi-
tal, Harvard Medical School, Boston, Massachusetts: JoAnn 1. Armstrong B, Doll R. Environmental factors and cancer in-
Manson; Brown University, Providence, Rhode Island: cidence and mortality in different countries, with special ref-
Annlouise R. Assaf; Emory University, Atlanta, Georgia: erence to dietary practices. Int J Cancer. 1975;15(4):617–631.
Lawrence Phillips; Fred Hutchinson Cancer Research Cen- 2. Kritchevsky D, Klurfeld DM. Influence of caloric intake on
experimental carcinogenesis: a review. Adv Exp Med Biol.
ter, Seattle, Washington: Shirley Beresford; George Wash- 1986;206:55–68.
ington University Medical Center, Washington, District of 3. Freedman LS, Clifford C, Messina M. Analysis of dietary fat,
Columbia: Judith Hsia; Los Angeles Biomedical Research In- calories, body weight, and the development of mammary tu-
stitute at Harbor-UCLA Medical Center, Torrance, California: mors in rats and mice: a review. Cancer Res. 1990;50(18):
Rowan Chlebowski; Kaiser Permanente Center for Health 5710–5719.
Research, Portland, Oregon: Evelyn Whitlock; Kaiser Per- 4. Willett WC, Sampson L, Stampfer MJ, et al. Reproducibility
manente Division of Research, Oakland, California: Bette and validity of a semiquantitative food frequency question-
Caan; Medical College of Wisconsin, Milwaukee, Wisconsin: naire. Am J Epidemiol. 1985;122(1):51–65.
Jane Morley Kotchen; MedStar Research Institute/Howard 5. Schoeller DA. Recent advances from application of doubly-
University, Washington, District of Columbia: Barbara V. labeled water to measurement of human energy expenditure.
J Nutr. 1999;129(10):1765–1768.
Howard; Northwestern University, Chicago/Evanston, 6. Heitmann BL, Lissner L. Dietary underreporting by obese
Illinois: Linda Van Horn; Rush Medical Center, Chicago, individuals—is it specific or non-specific? BMJ. 1995;
Illinois: Henry Black; Stanford Prevention Research Center, 311(7011):986–989.
Stanford, California: Marcia L. Stefanick; State University of 7. World Cancer Research Fund/American Institute for Cancer
New York at Stony Brook, Stony Brook, New York: Dorothy Research. Food, Nutrition and the Prevention of Cancer: A
Lane; The Ohio State University, Columbus, Ohio: Rebecca Global Perspective. Washington, DC: American Institute for
Jackson; University of Alabama at Birmingham, Birming- Cancer Research; 1997.
ham, Alabama: Cora E. Lewis; University of Arizona, 8. World Cancer Research Fund/American Institute for Cancer
Tucson/Phoenix, Arizona: Tamsen Bassford; University at Research. Food, Nutrition, Physical Activity, and the Preven-
Buffalo, Buffalo, New York: Jean Wactawski-Wende; Univer- tion of Cancer: A Global Perspective. Washington, DC:
American Institute for Cancer Research; 2007.
sity of California at Davis, Sacramento, California: John 9. Bingham SA. Urine nitrogen as a biomarker for the validation
Robbins; University of California at Irvine, Irvine, California: of dietary protein intake. J Nutr. 2003;133(suppl 3):921S–
F. Allan Hubbell; University of California at Los Angeles, 924S.
Los Angeles, California: Lauren Nathan; University of 10. Prentice RL, Sugar E, Wang CY, et al. Research strategies and
California at San Diego, LaJolla/Chula Vista, California: the use of nutrient biomarkers in studies of diet and chronic
Robert D. Langer; University of Cincinnati, Cincinnati, 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 dietary modification randomized controlled trial. J Natl Can-
exposure biomarkers and flexible measurement error. Bio- cer Inst. 2007;99(20):1534–1543.
metrics. 2007;63(1):143–151. 19. Hays J, Hunt JR, Hubbell FA, et al. The Women’s Health
12. Neuhouser ML, Tinker L, Shaw PA, et al. Use of recovery Initiative recruitment methods and results. Ann Epidemiol.
biomarkers to calibrate nutrient consumption self-reports in 2003;13(9 suppl):S18–S77.
the Women’s Health Initiative. Am J Epidemiol. 2008; 20. Langer RD, White E, Lewis CE, et al. The Women’s Health
167(10):1247–1259. Initiative Observational Study: baseline characteristics of
13. Women’s Health Initiative Study Group. Design of the participants and reliability of baseline measures. Ann Epide-
Women’s Health Initiative clinical trial and observational miol. 2003;13(9 suppl):S107–S121.
study. Control Clin Trials. 1998;19(1):61–109. 21. Kristal AK, Shattuck AL, Williams ME. Food frequency
14. Women’s Health Initiative Study Group. Dietary adherence in questionnaires for diet intervention research. In: 17th National
the Women’s Health Initiative dietary modification trial. J Am Nutrient Databank Conference Proceedings. Baltimore, MD:
Diet Assoc. 2004;104(4):654–658. International Life Sciences Institute; 1992:110–125.
15. Prentice RL, Caan B, Chlebowski RT, et al. Low-fat dietary 22. Schakel SF, Buzzard IM, Gebhardt SE. Procedures for esti-
pattern and risk of invasive breast cancer: the Women’s Health mating nutrient values for food composition databases. J Food
Initiative randomized controlled dietary modification trial. Compost Anal. 1997;10(2):102–114.
JAMA. 2006;295(6):629–642. 23. Curb JD, McTiernan A, Heckbert SR, et al. Outcomes ascer-
16. Beresford SA, Johnson KC, Ritenbaugh C, et al. Low-fat di- tainment and adjudication methods in the Women’s Health
etary pattern and risk of colorectal cancer: the Women’s Initiative. Ann Epidemiol. 2003;13(9 suppl):S122–S128.
Health Initiative randomized controlled dietary modification 24. Cox DR. Regression analysis and life tables (with discussion).
trial. JAMA. 2006;295(6):643–654. J R Stat Soc (B). 1972;34:187–220.
17. Howard BV, Van Horn L, Hsia J, et al. Low-fat dietary pattern 25. Subar AF, Kipnis V, Troiano RP, et al. Using intake biomarkers
and risk of cardiovascular disease: the Women’s Health Ini- to evaluate the extent of dietary misreporting in a large sample
tiative randomized controlled dietary modification trial. of adults: the OPEN study. Am J Epidemiol. 2003;158(1):1–13.
JAMA. 2006;295(6):655–666. 26. Kipnis V, Subar AF, Midthune D, et al. Structure of dietary
18. Prentice RL, Thomson CA, Caan B, et al. Low-fat dietary measurement error: results of the OPEN biomarker study. Am
pattern and cancer incidence in the Women’ s Health Initiative 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 whether HBV infection can explain the sex difference in the
of primary liver cancer (hereafter called liver cancer); it is occurrence of liver cancer. In this study, we examined effect
the fifth most common cancer and the third leading cause of modification by sex of the association between HBV infec-
cancer death worldwide (1, 2). China is one of the countries tion and liver cancer mortality, based on data from a 14-year
with the highest incidence of hepatocellular carcinoma in longitudinal study conducted in Haimen, China.
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 MATERIALS AND METHODS
ratio ranging from 2:1 to 4:1 (2). Hepatitis B virus (HBV) Study population
infection, which is endemic in many Asian countries, in-
cluding China, is a leading risk factor for hepatocellular Study methods have been described in detail previously
carcinoma and accounts for approximately 53% of hepato- (5). Briefly, the study was started in February 1992 in
cellular carcinoma cases worldwide (3). Other important Haimen, China, and enrollment continued until December
risk factors for hepatocellular carcinoma include hepatitis 1993. A total of 90,236 residents (60,306 men and 29,930
C virus infection, alcoholism, aflatoxin B1 ingestion, and women) voluntarily participated. The subjects were 22–80
family history of liver cancer (4). So far, it has been unclear 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% HBsAg carriers. Male sex and older age were also important
were HBsAg-positive (men: 15.3%; women: 16.9%). predictors of liver cancer mortality. Compared with women
A total of 13,529 deaths or losses to follow-up were iden- who were HBsAg-negative, mortality was significantly in-
tified, among which 1,803 participants died from liver cancer, creased in males and persons who were HBsAg-positive,
with 73.7% of them being HBsAg-positive. The median age and the combined effect of male sex and HBsAg positivity
at death for liver cancer patients was 51.7 years for men and was synergistic (Figure 1). After adjustment for potential
52.4 years for women. Mortality from liver cancer varied in confounders, the 3 interaction measures on the additive
the 4 groups defined by sex and HBsAg status (Table 2). Male scale all showed significantly higher values than the null
HBsAg carriers had the highest liver cancer mortality (RERI: 33.27 (95% confidence interval (CI): 22.54,
(1,035.2 deaths per 100,000 person-years), while female 43.99); attributable proportion of interaction: 0.59 (95%
non-HBsAg carriers had the lowest (16.1 deaths per CI: 0.55, 0.63); synergy index: 2.49 (95% CI: 2.13, 2.90)),
100,000 person-years). The mortality difference for liver can- suggesting a synergistic effect of male sex and HBsAg pos-
cer between HBsAg carriers and non-HBsAg carriers was itivity on liver cancer mortality. We further examined this
much greater in men than in women among all age groups. joint effect according to age (Table 3). The RERI decreased
HBsAg carriers had an increased risk of mortality from with increasing age, suggesting that the sex-HBsAg inter-
liver cancer, with a hazard ratio of 16 as compared with non- 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


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
Figure 1. Individual and joint effects of sex and hepatitis B surface
antigen (HBsAg) status on mortality from primary liver cancer, Haimen incidence. Sex hormones may play an important role in sex
Cohort Study, China, 1992–2006. HBV, hepatitis B virus. 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
DISCUSSION liver cancer. Naugler et al. (13) found that estrogen-
mediated inhibition of interleukin-6 production by Kupffer
It has been well documented that men are more likely cells reduced the risk of liver cancer in females. An intra-
than women to develop liver cancer (2), and HBV infection cellular signaling protein called MyD88 and the transcrip-
is a major risk factor (3). In this analysis, we focused on the tion factor nuclear factor jB were also found to be involved
joint effect of male sex and HBV infection on liver cancer in this signaling pathway (13, 14). Wang et al. (15) inte-
mortality. The proportion of HBsAg-positive men was com- grated the HBsAg gene into the mouse p21neomycin locus and
parable to that of women in our study. It has been suggested found that estrogen receptor b was extremely up-regulated,
that the higher rate of liver cancer in males than in females is which indicates that estrogen receptor b may play an im-
due to higher levels of exposure to risk factors (2). However, portant role in the development of liver cancer related to
after adjustment for covariates, including HBsAg status, HBsAg. In a case-control study, Yu et al. (16) and Sung et al.
age, history of hepatitis, occupation, regular alcohol drink- (17) found that there was a positive association between
ing, regular cigarette smoking, and family history of liver testosterone levels and liver cancer risk in HBV carriers
cancer, men still had higher mortality from liver cancer than and that genes involved in the regulation of testosterone,
women did. This may suggest that some susceptibility other such as SRD5A2 and V89L, may also play a role in the
than exposure level is more important. etiology of liver cancer.
We assessed additive interaction by calculating 3 mea- Liver cancer arises most frequently in the setting of
sures, including RERI, the attributable proportion of inter- chronic liver inflammation (18). After infection by the hep-
action, and the synergy index, all of which indicated that atitis virus, the host’s inflammatory immune response to the
there was a significant synergistic effect of HBsAg positivity viral antigens induces hepatocyte damage and is followed
and male sex on risk of liver cancer mortality. HBsAg pos- by the pathogenesis of liver cancer (19). Sex hormones may
itivity had an increased effect on liver cancer mortality in interact with HBV infection in the process and lead to a dom-
men compared with women. RERI, considered the best 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- (2006BAI02A03) and the Shanghai Leading Academic Dis-
crease with age. Previous study has demonstrated that liver cipline Project (B118). Na Wang received a scholarship
cancer tends to occur in men and postmenopausal women, from the China Scholarship Council for her training at the
which may result from lower production of estradiol and University of Ottawa.
a reduced response to the action of estradiol (20). That Conflict of interest: none declared.
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 REFERENCES
sex-related.
One limitation of our study is that HBsAg status was used 1. Parkin DM. Global cancer statistics in the year 2000. Lancet
as the sole indicator of HBV infection. Occult HBV infec- Oncol. 2001;2(9):533–543.
tion, a type of chronic HBV infection that is characterized 2. El-Serag HB, Rudolph KL. Hepatocellular carcinoma: epide-
by the absence of detectable HBsAg in the blood and very miology and molecular carcinogenesis. Gastroenterology.
2007;132(7):2557–2576.
low levels of HBV DNA in the blood and liver (21, 22), can
3. Lupberger J, Hildt E. Hepatitis B virus-induced oncogenesis.
also lead to liver cancer, with a mechanism similar to that of World J Gastroenterol. 2007;13(1):74–81.
overt cases (23). Therefore, part of the mortality among 4. Bosch FX, Ribes J, Dı́az M, et al. Primary liver cancer:
HBsAg noncarriers might be attributable to the effect of worldwide incidence and trends. Gastroenterology. 2004;
occult hepatitis B infection, resulting in underestimation 127(5 suppl 1):S5–S16.
of the effect of HBV infection as well as its interaction with 5. Evans AA, Chen G, Ross EA, et al. Eight-year follow-up of the
sex. Additional data on HBV infection, such as information 90,000-person Haimen City cohort: I. Hepatocellular carci-
on hepatitis B e antigen, HBV virus load, and HBV geno- noma mortality, risk factors, and gender differences. Cancer
type, would consolidate the finding (3, 23). Epidemiol Biomarkers Prev. 2002;11(4):369–376.
In summary, we found a synergistic effect of male sex and 6. Rothman KJ. Epidemiology: An Introduction. Oxford, NY:
Oxford University Press; 2002.
HBV infection on risk of liver cancer mortality, suggesting 7. Rothman KJ, Greenland S. Modern Epidemiology. Philadel-
that HBV infection may have a greater impact on liver can- phia, PA: Lippincott-Raven; 1998.
cer risk in men than in women, at least for persons living in 8. Kalilani L, Atashili J. Measuring additive interaction using
high-risk areas. Possible mechanisms for the interaction be- odds ratios [electronic article]. Epidemiol Perspect Innov.
tween sex and HBV infection in liver cancer mortality need 2006;3:5.
to be further explored. If the finding is confirmed by other 9. Andersson T, Alfredsson L, Källberg H, et al. Calculating
studies, a specific HBV vaccination program may be devel- measures of biological interaction. Eur J Epidemiol. 2005;
oped to reduce liver cancer risk. It has been well demon- 20(7):575–579.
strated that HBV infection frequently occurs as a result of 10. Assmann SF, Hosmer DW, Lemeshow S, et al. Confidence
vertical (mother-to-child) or horizontal (child-to-child) intervals for measures of interaction. Epidemiology. 1996;7(3):
286–290.
transmission in endemic areas (24). It has also been reported 11. Li R, Chambless L. Test for additive interaction in propor-
that HBV infection can be acquired in adulthood through tional hazards models. Ann Epidemiol. 2007;17(3):227–236.
invasive medical procedures or household contact with 12. Yu MW, Chang HC, Chang SC, et al. Role of reproductive
HBV carriers (25). In Haimen, where the current study factors in hepatocellular carcinoma: impact on hepatitis B- and
was conducted, an HBV vaccination program for all new- C-related risk. Hepatology. 2003;38(6):1393–1400.
borns was implemented in 1992. Adults, however, are not 13. Naugler WE, Sakurai T, Kim S, et al. Gender disparity in liver
included in the program. Because of the increased risk of cancer due to sex differences in MyD88-dependent IL-6
liver cancer in HBsAg-positive males, immunization strat- production. Science. 2007;317(5834):121–124.
egies might be expanded to also include male adults not yet 14. Lawrence T, Hageman T, Balkwill F. Cancer: sex, cytokines,
infected with HBV, to reduce the risk of developing liver and cancer. Science. 2007;317(5834):51–52.
15. Wang Y, Cui F, Lv Y, et al. HBsAg and HBx knocked into
cancer. 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:
ACKNOWLEDGMENTS a nested case-control study among men. J Natl Cancer Inst.
2001;93(21):1644–1651.
Author affiliations: Department of Epidemiology, School 17. Sung YM, Tang NL, Lai PB, et al. Re: Hormonal markers and
of Public Health, Fudan University, Shanghai, People’s Re- hepatitis B virus-related hepatocellular carcinoma risk:
public of China (Na Wang, Yingjie Zheng, Qingwu Jiang); a nested case-control study among men [letter]. J Natl Cancer
Haimen City Center for Disease Control and Prevention, Inst. 2003;95(7):559–560.
18. Prieto J. Inflammation, hepatocellular carcinoma and sex:
Haimen City, People’s Republic of China (Xinsen Yu, IL-6 in the centre of the triangle. J Hepatol. 2008;48(2):
Wenyao Lin); and Department of Epidemiology and Com- 380–381.
munity Medicine, Faculty of Medicine, University of Ot- 19. Wands J. Hepatocellular carcinoma and sex. N Engl J Med.
tawa, Ottawa, Ontario, Canada (Yue Chen). 2007;357(19):1974–1976.
This work was supported by a grant from the Ministry of 20. Shimizu I, Kohno N, Tamaki K, et al. Female hepatology:
Science and Technology of the People’s Republic of China favorable role of estrogen in chronic liver disease with

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

hepatitis B virus infection. World J Gastroenterol. 2007;13(32): 23. Chemin I, Trépo C. Clinical impact of occult HBV infections.
4295–4305. J Clin Virol. 2005;34(suppl 1):S15–S21.
21. Vivekanandan P, Kannangai R, Ray SC, et al. Comprehensive 24. Marrero CR, Marrero JA. Viral hepatitis and hepatocellular
genetic and epigenetic analysis of occult hepatitis B from liver carcinoma. Arch Med Res. 2007;38(6):612–620.
tissue samples. Clin Infect Dis. 2008;46(8):1227–1236. 25. Zhang HW, Yin JH, Li YT, et al. Risk factors for acute hep-
22. Torbenson M, Thomas DL. Occult hepatitis B. Lancet Infect atitis B and its progression to chronic hepatitis in Shanghai,
Dis. 2002;2(8):479–486. China. Gut. 2008;57(12):1713–20.

Am J Epidemiol 2009;169:990–995
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- A recent meta-analysis of 14 prospective cohort studies
oxidant selenoenzymes such as glutathione peroxidases, found a modest, but statistically significant inverse associa-
has been hypothesized to prevent atherosclerotic disease tion between selenium levels and coronary heart disease (3),
(1–3). Glutathione peroxidase synthesis and activity, but the 2 US studies in this meta-analysis found no associ-
however, plateau at selenium levels above 70–90 ng/mL ation (3, 6, 7). Moreover, serum selenium levels were not
(4). In the United States compared with other countries, associated with cardiovascular disease mortality in a pro-
selenium intake is substantially higher (5), and most spective study in a representative US sample, although
adults have serum selenium levels above 95 ng/mL (4). a U-shaped relation was suggested (8). Given the current
At these high concentrations, selenium is incorporated interest in selenium supplements for chemoprevention of
nonspecifically as selenomethionine in the synthesis of cancer (5), it is important to understand the overall impact
other plasma proteins, with unknown health effects (4). of increased selenium intake on other health endpoints, in-
It is thus unclear whether increased selenium levels con- cluding vascular disease.
fer additional benefit for atherosclerosis prevention in the Peripheral arterial disease, which affects about 8 million
United States. 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


Selenium and Peripheral Arterial Disease 997

in the muscular arteries of the lower extremities and is an right arm (brachial artery) and both ankles (posterior tibial
important marker of generalized atherosclerosis (9–11). arteries) with a Doppler device, the Parks Mini-Lab IV, model
Data on the association of selenium levels with peripheral 3100 (Parks Medical Electronics, Inc., Aloha, Oregon). If the
arterial disease are very limited (12, 13). The objective of participant had a condition that would interfere with blood
the present study was to assess the association between pressure reading in the right arm, the left arm was used.
serum levels of selenium and reduced ankle-brachial Systolic blood pressure was measured twice at each site for
blood pressure index (ABI), a specific subclinical marker participants aged 40–59 years and once at each site for
for peripheral arterial disease, in the National Health and participants aged 60 years or older. The left and right
Nutrition Examination Survey (NHANES), 2003–2004. ABI ABI measurements were obtained by dividing the mean
values below 0.90 are considered diagnostic of peripheral systolic blood pressure in each ankle by the mean systolic
arterial disease. Furthermore, a low ABI is an independent blood pressure in the arm. Peripheral arterial disease was
predictor of cardiovascular risk after adjusting for traditional defined as an ABI value of less than 0.90 in at least one leg
cardiovascular risk factors (14, 15). (14, 15).

MATERIALS AND METHODS Other variables

NHANES is conducted by the National Center for Information about age, sex, race-ethnicity, education,
Health Statistics (Hyattsville, Maryland) by using a com- family income, menopausal status for women, cigarette
plex multistage sampling design to obtain a representative smoking, alcohol consumption, use of dietary supplements,
sample of the civilian, noninstitutionalized US popula- and use of cholesterol- and blood-pressure-lowering medi-
tion. We used data from NHANES 2003–2004 (16) be- cations was based on self-report. Body mass index was cal-
cause it was the first NHANES survey to measure culated by dividing measured weight in kilograms by
selenium and ABI levels simultaneously. In NHANES measured height in meters squared. Three to 4 systolic blood
2003–2004 interviews and physical examinations, the pressure measurements were taken and were averaged by
overall response rate was 76%. Serum selenium and using standardized protocols. Diabetes was defined as a fast-
ABI measurements were restricted to participants aged ing serum glucose concentration of 126 mg/dL or higher,
40 years or older (N ¼ 3,086). We excluded 2 pregnant a nonfasting serum glucose concentration of 200 mg/dL or
women, 183 participants without selenium measure- higher, a self-reported physician diagnosis, or current med-
ments, 514 participants without ABI measurements in ication use. Glomerular filtration rate was estimated by us-
both legs, and 314 participants with missing information ing the Modification of Diet in Renal Disease Study
on any adjustment covariate. We finally excluded 11 par- equation with serum creatinine values (19).
ticipants with left or right ABI measurements of more
than 1.5, usually due to vessel stiffness. The final sample Statistical methods
size was 2,062. The 2003–2004 NHANES study protocols
were approved by the National Center for Health Statis- Participants were grouped in quartiles of serum sele-
tics institutional review board. Oral and written informed nium levels based on the weighted population distribu-
consent was obtained from all participants. tion. Odds ratios and 95% confidence intervals for
peripheral arterial disease prevalence comparing the 3
Serum selenium highest quartiles of serum selenium with the lowest quar-
tile were estimated by using logistic regression. Tests for
Collection materials were screened for potential selenium linear risk trend across serum selenium quartiles were
contamination. After blood collection, serum aliquots were performed by including an ordinal variable with the
obtained, frozen at 20C, and shipped to the Trace Ele- median selenium level of each quartile in the logistic re-
ments Laboratory at the Wadsworth Center of the New York gression models. To further explore the shape of the dose-
State Department of Health for analysis. Serum selenium response relation between serum selenium levels and
levels were measured by using inductively coupled peripheral arterial disease prevalence, we used restricted
plasma-dynamic reaction cell-mass spectrometry. The lab- quadratic splines with knots at the 5th, 50th, and 95th
oratory procedures and quality control methods for serum percentiles of serum selenium distribution. These spline
selenium measurement have been described in detail else- models require the same number of parameters as the
where (17). The between-assay coefficients of variation for quartile analysis, but they can accommodate a wide vari-
quality-control pooled samples analyzed throughout the du- ety of smooth risk trends (20). Sensitivity analyses using
ration of the survey ranged from 2.5% to 2.9%. different numbers and locations of the knots, with cubic
instead of quadratic splines, and log-transforming serum
Peripheral arterial disease selenium levels gave similar results (not shown). Statisti-
cal analyses were performed with the survey package in R
A specific protocol was used to measure ABI in software (to account for the complex sampling design in
NHANES 2003–2004 (18). The measurements of blood NHANES 2003–2004) (21, 22). Strata, primary sampling
pressure used for ABI were additional to and different from units, and examination sample weights were used to ob-
other measurements of blood pressure used to evaluate hy- tain unbiased point estimates and robust linearized stan-
pertension. Systolic blood pressure was measured on the dard errors.

Am J Epidemiol 2009;169:996–1003
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


associated with body mass index and serum cotinine levels.
The weighted prevalence of peripheral arterial disease in The age-, sex-, and race-adjusted prevalence of peripheral
the study population was 4.9%. Compared with participants arterial disease decreased with increasing serum selenium
without peripheral arterial disease, those with disease were levels (P for linear trend ¼ 0.02) (Table 3). However, there
more likely to be older, black, ever smokers, nondrinkers, and was an indication of an upturn in risk trend in the highest
diabetic; to have a lower educational level and family in- quartile of serum selenium, particularly after adjusting for
come; and to use cholesterol- and blood-pressure-lowering cardiovascular risk factors. The fully adjusted odds ratios
medications (Table 1). Participants with peripheral arterial for peripheral arterial disease comparing selenium quartiles
disease, compared with those without disease, also had higher 2, 3, and 4 with the lowest quartile were 0.75 (95% confi-
average levels of body mass index, systolic blood pressure, dence interval: 0.37, 1.52), 0.58 (95% confidence interval:
and C-reactive protein and lower high density lipoprotein 0.28, 1.19), and 0.67 (95% confidence interval: 0.34, 1.31),
cholesterol levels and glomerular filtration rate. respectively. In spline regression models, peripheral arterial
Participants in the highest quartile of serum selenium disease prevalence decreased with increasing serum sele-
levels, compared with those in the lowest quartile, were nium levels up to 150–160 ng/mL (80th–91st percentiles
more likely to be older, men, white, and nonsmokers and of the serum selenium distribution in the study population),
to use dietary supplements and cholesterol-lowering med- followed by a gradual increase at higher selenium levels
ications (Table 2). Serum selenium levels were also positive- (Figure 1). Consistently, there was a marginally significant
ly associated with total cholesterol, high density lipoprotein 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 selenium in the soil (1, 23). In the United States, estimated
spline terms ¼ 0.06 and 0.12 after adjustment for age, sex, selenium intake ranges from 60 lg/day to 220 lg/day (23),
and race and after full adjustment, respectively). 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–
DISCUSSION 2004, the median selenium level was 134 ng/mL, and 99%
of study participants had serum selenium levels above 95
In this cross-sectional study, conducted in a representative ng/mL. These concentrations are considerably higher than
sample of the US population, the association between serum in other countries. In Europe, for instance, average serum
selenium levels and the prevalence of peripheral arterial selenium levels range from 50 ng/mL to 90 ng/mL (23, 24).
disease was not statistically significant, although a U-shaped Very limited data are available on the association of
relation was suggested: the prevalence of peripheral arterial selenium with peripheral arterial disease. Two small studies
disease decreased with increasing serum selenium levels up found similar selenium levels in patients with peripheral
to 150 ng/mL but increased with increasing selenium levels arterial disease compared with controls, but the dose-
above 160 ng/mL. Selenium intake varies around the world response relation was not evaluated (12, 13). For other car-
primarily because of geographic variation in the amount of 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-


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-
0.5 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
0.25 0 (8). Finally, in the Health Professionals Follow-up Study, the
100 120 140 160 180 200 odds ratios for incident coronary heart disease comparing
Serum Selenium, ng/ml 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
Figure 1. Odds ratios for peripheral arterial disease by serum sele- dose-response relation (7). Both serum and toenail selenium
nium levels, National Health and Nutrition Examination Survey, 2003– levels reflect selenium status, although toenails reflect longer-
2004. Curves represent adjusted odds ratios (solid line) and their 95% term exposure. It is unclear, however, whether both biomarkers
confidence intervals (dashed lines) based on restricted quadratic are comparable in their ability to capture the different types of
splines for serum selenium levels with knots at the 5th, 50th, and selenium compounds.
95th percentiles. The reference value (odds ratio ¼ 1) was set at the
10th percentile of serum selenium distribution (116 ng/mL). Odd ratios Few randomized trials have evaluated the effect of selenium
were adjusted for age, sex, race-ethnicity, education, family income, supplementation on cardiovascular outcomes or atherosclero-
postmenopausal status, smoking, serum cotinine, alcohol consump- sis progression, and most of these studies combined selenium
tion, body mass index, dietary supplement use, C-reactive protein, total with other vitamins and minerals (3, 37). Only 2 of these trials
cholesterol, high density lipoprotein cholesterol, cholesterol-lowering
medication use, systolic blood pressure, blood-pressure-lowering were conducted in the United States, both reporting null results
medication use, diabetes, and glomerular filtration rate. Bars represent (38, 39). In the Nutritional Prevention of Cancer trial, the
the weighted histogram of serum selenium distribution. relative risk for cardiovascular disease incidence comparing

Am J Epidemiol 2009;169:996–1003
Selenium and Peripheral Arterial Disease 1001

200 lg/day of selenium supplementation with placebo was metabolic abnormalities, and to determine whether the
1.03 (95% confidence interval: 0.78, 1.37) (38). In the HDL- change point in risk associated with elevated selenium levels
Atherosclerosis Treatment Study, the progression of athero- depends on genetic polymorphisms in candidate genes for
sclerosis measured by coronary angiography in patients with selenium metabolism (48).
coronary artery disease was similar among participants ran- Several limitations of our study need to be considered.
domized to an antioxidant supplement containing 100 lg/day The use of a cross-sectional design and of prevalent cases of
of selenium, 800 IU/day of vitamin E, 1 g/day of vitamin C, peripheral arterial disease limited our ability to determine
and 25 mg/day of b-carotene and participants randomized to the direction and the causality of the observed association. It
placebo (39). Overall, limited evidence from randomized trials is possible that the pathophysiologic changes of atheroscle-
has not shown a protective effect of selenium supplementation rosis could modify serum selenium levels or that partici-
in US studies. With respect to observational studies, those in pants with peripheral arterial disease change their health
the United States have not been able to detect a significant behaviors, including selenium intake through diet and di-
linear association between serum selenium and cardiovascular etary supplements. As a consequence, our findings must be
outcomes, but the dose-response associations in these studies confirmed in prospective studies with incident cases of pe-
were U-shaped. ripheral arterial disease. Another limitation of our study is
The biologic mechanisms underlying a potential effect of the use of a single measurement of serum selenium, which
selenium on cardiovascular disease are likely complex, but reflects short-term selenium intake and may be subject to
they may be related to the dual role of selenium as an es- high within-person variability (49). Furthermore, our study
sential and toxic element. Selenium is an essential micro- measured only total serum selenium, and we did not have
nutrient that is incorporated into glutathione peroxidases information on selenoprotein levels or activity or about non-
and other selenoproteins (4). Increasing serum selenium specific incorporation of selenium as selenomethionine in
levels increase the concentration and activity of glutathione other plasma proteins. More detailed analysis of different
peroxidases, but this dose-response relation reaches a pla- compartments of serum selenium will be needed to better
teau at serum selenium levels of 70–90 ng/mL (4). As a con- understand the association of selenium with peripheral
sequence, higher selenium levels could potentially prevent arterial disease. The strengths of our study come from
atherosclerosis development and progression in populations the rigorous sampling design and the quality of the study
whose selenium exposure is below the levels needed to measurements used in NHANES; the representativeness of
maximize glutathione peroxidases (1–3). In selenium- the NHANES sample; and the use of ABI, a noninvasive
replete populations such as in the United States, in which measure of subclinical atherosclerosis.
virtually all participants have serum selenium levels above In summary, the association between serum selenium lev-
70–90 ng/mL, the mechanisms underlying a potential ben- els and the prevalence of peripheral arterial disease in
eficial effect of increased selenium levels are unclear. Since NHANES 2003–2004 was not statistically significant,
selenium supplementation is actively promoted in the although a U-shaped relation was suggested. Other sources
United States, and large randomized controlled trials testing of evidence (6, 8) also suggest a U-shaped relation between
the efficacy of selenium supplementation in prostate cancer serum selenium levels and cardiovascular outcomes in the
prevention are under way (40, 41), mechanistic studies are United States, a selenium-replete population. In many pop-
urgently needed to establish the biologic basis for a protec- ulations worldwide, selenium intake is lower than in the
tive effect of selenium in populations whose selenium status United States (1, 23). At these lower levels of selenium
is already high. intake, the association of selenium with peripheral arterial
Selenium, however, has a narrow therapeutic range (4), disease remains unknown. Prospective studies of selenium
and it may even be harmful at intake levels below the current status across populations with different levels of selenium
tolerable Upper Intake Level of 400 lg/day (2). In fact, intake and randomized trials stratified by baseline sele-
some selenium compounds have been documented to gen- nium status are needed to establish the optimal selenium
erate reactive oxygen species (42–44), and the upturn in levels to minimize the risk of cardiovascular and other
peripheral arterial disease prevalence that we observed at chronic diseases.
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 ACKNOWLEDGMENTS
high selenium levels in US populations. For instance, the
Nutritional Prevention of Cancer trial showed an increased Author affiliations: Department of Epidemiology, Johns
risk of diabetes for participants receiving 200 lg/day of Hopkins Bloomberg School of Public Health, Baltimore,
selenium compared with placebo (hazard ratio ¼ 1.50, Maryland (Joachim Bleys, Ana Navas-Acien, Martin
95% confidence interval: 1.03, 2.33) (46). Interestingly, Laclaustra, Andy Menke, Eliseo Guallar); Department of
the excess risk was limited to participants in the upper tertile Medicine, Johns Hopkins School of Medicine, Baltimore,
of the serum selenium distribution (>121.6 ng/mL), who Maryland (Eliseo Guallar); Welch Center for Prevention,
had a hazard ratio for diabetes of 2.70 (95% confidence Epidemiology and Clinical Research, Johns Hopkins
interval: 1.30, 5.61). Further research is needed to establish Bloomberg School of Public Health and Johns Hopkins
the mechanisms underlying the association of high-normal School of Medicine, Baltimore, Maryland (Joachim Bleys,
selenium levels with peripheral arterial disease and with Ana Navas-Acien, Martin Laclaustra, Andy Menke, Eliseo

Am J Epidemiol 2009;169:996–1003
1002 Bleys et al.

Guallar); Department of Environmental Health Sciences, 12. Ondrus P, Alberty R, Vassanyiova Z. Importance of lipid per-
Johns Hopkins Bloomberg School of Public Health, oxidation, protective enzymes and trace elements in chronic
Baltimore, Maryland (Ana Navas-Acien); Department of leg ischaemia. Eur J Clin Chem Clin Biochem. 1996;34(6):
Cardiovascular Epidemiology and Population Genetics, 471–475.
National Center for Cardiovascular Research (CNIC), 13. Mansoor MA, Bergmark C, Haswell SJ, et al. Correlation
between plasma total homocysteine and copper in patients
Madrid, Spain (Martin Laclaustra, Eliseo Guallar); National
with peripheral vascular disease. Clin Chem. 2000;46(3):
Center for Epidemiology, Instituto de Salud Carlos III, and 385–391.
the CIBER in Epidemiology and Public Health (CIBERESP), 14. Hirsch AT, Haskal ZJ, Hertzer NR, et al. ACC/AHA 2005
Madrid, Spain (Roberto Pastor-Barriuso); Friedman School Practice Guidelines for the management of patients with
of Nutrition Science and Policy at Tufts University, Boston, peripheral arterial disease (lower extremity, renal, mesen-
Massachusetts (Jose M. Ordovas); and Clinical Sciences teric, and abdominal aortic): a collaborative report from the
Research Institute at Warwick Medical School, Coventry, American Association for Vascular Surgery/Society for
United Kingdom (Saverio Stranges). Vascular Surgery, Society for Cardiovascular Angiography
Supported by grants 1 R01 ES012673 from the National and Interventions, Society for Vascular Medicine and
Institute of Environmental Health Sciences and 0230232N Biology, Society of Interventional Radiology, and the ACC/
from the American Heart Association. AHA Task Force on Practice Guidelines (Writing Committee
Conflict of interest: none declared. 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-
REFERENCES cular Nursing; TransAtlantic Inter-Society Consensus; and
Vascular Disease Foundation. Circulation. 2006;113(11):
1. Rayman MP. Food-chain selenium and human health: e463–e654.
emphasis on intake. Br J Nutr. 2008;100(2):254–268. 15. Fowkes FG, Murray GD, Butcher I, et al. Ankle brachial index
2. Navas-Acien A, Bleys J, Guallar E. Selenium intake and combined with Framingham Risk Score to predict cardiovas-
cardiovascular risk: what is new? Curr Opin Lipidol. 2008; cular events and mortality: a meta-analysis. JAMA. 2008;
19(1):43–49. 300(2):197–208.
3. Flores-Mateo G, Navas-Acien A, Pastor-Barriuso R, et al. 16. National Center for Health Statistics, Centers for Disease
Selenium and coronary heart disease: a meta-analysis. Am J Control and Prevention. National Health and Nutrition
Clin Nutr. 2006;84(4):762–773. Examination Survey, 2003–2004. 2007. (http://www.cdc.gov/
4. Food and Nutrition Board, Institute of Medicine. Dietary nchs/about/major/nhanes/nhanes2003-2004/nhanes03_04.htm).
Reference Intakes for Vitamin C, Vitamin E, Selenium, and (Accessed August 24, 2008).
Carotenoids. A report of the Panel on Dietary Antioxidants 17. National Center for Health Statistics, Centers for Disease
and Related Compounds, Subcommittees on Upper Reference Control and Prevention. NHANES 2003–2004 Data Docu-
Levels of Nutrients and Interpretation and Uses of Dietary mentation. Laboratory Assessment: Lab 39—Erythrocyte
Reference Intakes, and the Standing Committee on the Sci- Protoporphyrin and Selenium. 2007. (http://www.cdc.gov/
entific Evaluation of Dietary Reference Intakes. Washington, nchs/data/nhanes/nhanes_03_04/l39epp_c.pdf). (Accessed
DC: The National Academies Press; 2000. August 24, 2008).
5. Rayman MP. Selenium in cancer prevention: a review of the 18. National Center for Health Statistics, Centers for Disease
evidence and mechanism of action. Proc Nutr Soc. 2005; Control and Prevention. NHANES 2003–2004 Data Docu-
64(4):527–542. mentation. Exam Component: Ankle-Brachial Blood Pressure
6. Salvini S, Hennekens CH, Morris JS, et al. Plasma levels of Index (ABPI) Section of the Lower Extremity Disease Ex-
the antioxidant selenium and risk of myocardial infarction amination (LEXAB_C). 2005. (http://www.cdc.gov/nchs/
among U.S. physicians. Am J Cardiol. 1995;76(17): data/nhanes/nhanes_03_04/lexab_c.pdf). (Accessed August
1218–1221. 24, 2008).
7. Yoshizawa K, Ascherio A, Morris JS, et al. Prospective study 19. Levey AS, Coresh J, Greene T, et al. Using standardized serum
of selenium levels in toenails and risk of coronary heart dis- creatinine values in the Modification of Diet in Renal Disease
ease in men. Am J Epidemiol. 2003;158(9):852–860. Study equation for estimating glomerular filtration rate. Ann
8. Bleys J, Navas-Acien A, Guallar E. Serum selenium levels and Intern Med. 2006;145(4):247–254.
all-cause, cancer, and cardiovascular mortality among US 20. Greenland S. Dose-response and trend analysis in epidemiol-
adults. Arch Intern Med. 2008;168(4):404–410. ogy: alternatives to categorical analysis. Epidemiology. 1995;
9. McDermott MM, Liu K, Criqui MH, et al. Ankle-brachial 6(4):356–365.
index and subclinical cardiac and carotid disease: the 21. R Development Core Team. R: A Language and Environment
Multi-Ethnic Study of Atherosclerosis. Am J Epidemiol. 2005; for Statistical Computing. Vienna, Austria: R Foundation for
162(1):33–41. Statistical Computing; 2008. (ISBN 3-900051-07-0). (http://
10. Murabito JM, Evans JC, Larson MG, et al. The ankle-brachial www.R-project.org).
index in the elderly and risk of stroke, coronary disease, and 22. Lumley T. The Survey Package. (http://cran.fhcrc.org/web/
death: the Framingham Study. Arch Intern Med. 2003;163(16): packages/survey/survey.pdf). (Accessed August 24, 2008).
1939–1942. 23. Combs GF Jr. Selenium in global food systems. Br J Nutr.
11. Newman AB, Shemanski L, Manolio TA, et al. Ankle-arm 2001;85(5):517–547.
index as a predictor of cardiovascular disease and mortality in 24. Rayman MP. The importance of selenium to human health.
the Cardiovascular Health Study. The Cardiovascular Health Lancet. 2000;356(9225):233–241.
Study Group. Arterioscler Thromb Vasc Biol. 1999;19(3): 25. Salonen JT, Alfthan G, Huttunen JK, et al. Association
538–545. between cardiovascular death and myocardial infarction and

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

serum selenium in a matched-pair longitudinal study. Lancet. a meta-analysis of randomized controlled trials. Am J Clin Nutr.
1982;2(8291):175–179. 2006;84(4):880–887.
26. Miettinen TA, Alfthan G, Huttunen JK, et al. Serum selenium 38. Stranges S, Marshall JR, Trevisan M, et al. Effects of selenium
concentration related to myocardial infarction and fatty acid supplementation on cardiovascular disease incidence and
content of serum lipids. Br Med J (Clin Res Ed). 1983; mortality: secondary analyses in a randomized clinical trial.
287(6391):517–519. Am J Epidemiol. 2006;163(8):694–699.
27. Salonen JT, Salonen R, Penttila I, et al. Serum fatty acids, 39. Brown BG, Zhao XQ, Chait A, et al. Simvastatin and niacin,
apolipoproteins, selenium and vitamin antioxidants and the antioxidant vitamins, or the combination for the prevention of
risk of death from coronary artery disease. Am J Cardiol. coronary disease. N Engl J Med. 2001;345(22):1583–1592.
1985;56(4):226–231. 40. Lippman SM, Goodman PJ, Klein EA, et al. Designing the
28. Virtamo J, Valkeila E, Alfthan G, et al. Serum selenium and Selenium and Vitamin E Cancer Prevention Trial (SELECT).
the risk of coronary heart disease and stroke. Am J Epidemiol. J Natl Cancer Inst. 2005;97(2):94–102.
1985;122(2):276–282. 41. Marshall JR, Sakr W, Wood D, et al. Design and progress of
29. Ringstad J, Thelle D. Risk of myocardial infarction in relation a trial of selenium to prevent prostate cancer among men with
to serum concentrations of selenium. Acta Pharmacol Toxicol high-grade prostatic intraepithelial neoplasia. Cancer
(Copenh). 1986;59(suppl 7):336–339. Epidemiol Biomarkers Prev. 2006;15(8):1479–1484.
30. Kok FJ, de Bruijn AM, Hofman A, et al. Selenium status and 42. Spallholz JE. On the nature of selenium toxicity and carcino-
chronic disease mortality: Dutch epidemiological findings. Int static activity. Free Radic Biol Med. 1994;17(1):45–64.
J Epidemiol. 1987;16(2):329–332. 43. Spallholz JE, Palace VP, Reid TW. Methioninase and
31. Ringstad J, Jacobsen BK, Thomassen Y, et al. The Tromsø selenomethionine but not Se-methylselenocysteine generate
Heart Study: serum selenium and risk of myocardial infarction methylselenol and superoxide in an in vitro chemiluminescent
a nested case-control study. J Epidemiol Community Health. assay: implications for the nutritional carcinostatic activity
1987;41(4):329–332. of selenoamino acids. Biochem Pharmacol. 2004;67(3):
32. Suadicani P, Hein HO, Gyntelberg F. Serum selenium con- 547–554.
centration and risk of ischaemic heart disease in a prospective 44. Drake EN. Cancer chemoprevention: selenium as a prooxi-
cohort study of 3000 males. Atherosclerosis. 1992;96(1): dant, not an antioxidant. Med Hypotheses. 2006;67(2):
33–42. 318–322.
33. Marniemi J, Jarvisalo J, Toikka T, et al. Blood vitamins, 45. Bleys J, Navas-Acien A, Guallar E. Serum selenium and
mineral elements and inflammation markers as risk factors of diabetes in U.S. adults. Diabetes Care. 2007;30(4):829–834.
vascular and non-vascular disease mortality in an elderly 46. Stranges S, Marshall JR, Natarajan R, et al. Effects of long-
population. Int J Epidemiol. 1998;27(5):799–807. term selenium supplementation on the incidence of type 2
34. Kilander L, Berglund L, Boberg M, et al. Education, lifestyle diabetes: a randomized trial. Ann Intern Med. 2007;147(4):
factors and mortality from cardiovascular disease and cancer. 217–223.
A 25-year follow-up of Swedish 50-year-old men. Int J Epi- 47. Bleys J, Navas-Acien A, Stranges S, et al. Serum selenium and
demiol. 2001;30(5):1119–1126. serum lipids in US adults. Am J Clin Nutr. 2008;88(2):
35. Wei WQ, Abnet CC, Qiao YL, et al. Prospective study of 416–423.
serum selenium concentrations and esophageal and gastric 48. Méplan C, Crosley LK, Nicol F, et al. Genetic polymorphisms
cardia cancer, heart disease, stroke, and total death. Am J Clin in the human selenoprotein P gene determine the response of
Nutr. 2004;79(1):80–85. selenoprotein markers to selenium supplementation in a gender-
36. Akbaraly NT, Arnaud J, Hininger-Favier I, et al. Selenium and specific manner (the SELGEN study). FASEB J. 2007;21(12):
mortality in the elderly: results from the EVA study. Clin 3063–3074.
Chem. 2005;51(11):2117–2123. 49. Longnecker MP, Stampfer MJ, Morris JS, et al. A 1-y trial of
37. Bleys J, Miller ER III, Pastor-Barriuso R, et al. Vitamin- the effect of high-selenium bread on selenium concentrations
mineral supplementation and the progression of atherosclerosis: in blood and toenails. Am J Clin Nutr. 1993;57(3):408–413.

Am J Epidemiol 2009;169:996–1003
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- (fourth quartile vs. first quartile, odds ratio ¼ 2.95, 95%
sociations between ambient air pollution and adverse preg- confidence interval (CI): 1.44, 6.05). Elevated risks of aortic
nancy outcomes (1–6). Associations between air pollution artery and valve defects, pulmonary artery and valve de-
levels during pregnancy and risks of cardiovascular malfor- fects, and conotruncal defects with increasing ambient
mations among the offspring were investigated in 2 previous ozone levels were also reported (7). The second investiga-
population-based case-control studies (7, 8). In these stud- tion, conducted in Texas, did not corroborate the southern
ies, cardiovascular malformations were classified by using California findings, although a suggestive association be-
preexisting surveillance database codes, and analyses were tween ozone and pulmonary artery and valve defects was
based on contrasts in ambient pollution levels over space observed. The Texas investigators reported positive associ-
and time. ations for carbon monoxide and tetralogy of Fallot, particu-
In the first study, conducted in southern California, inves- late matter with an average aerodynamic diameter of <10 lm
tigators reported an association between ambient carbon (PM10) and atrial septal defects, and sulfur dioxide and ven-
monoxide levels and risk of ventricular septal defects 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


Ambient Air Pollution and Cardiovascular Malformations 1005

Table 1. Cardiovascular Malformation Outcome Groups, Number of Cases,a and Outcome Group Definitions

Outcome Group No. of Cases Definition

Atrial septal defect, secundum 379 Includes secundum-type atrial septal defect.
Coarctation of the aorta 275 Includes coarctation of the aorta, aortic arch hypoplasia, and interrupted aortic
arch type A.
Hypoplastic left heart syndrome 175 Includes hypoplastic left heart syndrome with or without ventricular septal defect.
Patent ductus arteriosus 219 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.
Pulmonary stenosis, valvar 312 Includes valvar and unspecified pulmonary stenosis, as well as dysplastic
pulmonary valve.
Tetralogy of Fallot 299 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.
Transposition of the great arteries 165 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.
Ventricular septal defect, muscular 1,108 Includes muscular-type ventricular septal defect.
Ventricular septal defect, perimembranous 546 Includes perimembranous-type ventricular septal defect.
Conotruncal defectb 661 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.
Left ventricular outflow tract defectb 558 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.
Right ventricular outflow tract defectb 421 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.
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, indication of a cardiovascular malformation were obtained
Georgia, to explore temporal associations between ambient from the Metropolitan Atlanta Congenital Defects Program
air pollution levels during pregnancy and risks of cardiovas- (MACDP), which conducts active, population-based birth
cular malformations. We did not closely replicate the meth- defects surveillance on the cohort of pregnancies reaching
odologies of the previous studies. In addition to the at least 20 weeks’ gestation to mothers residing in 1 of 5
retrospective cohort design and the temporal analytical ap- central Atlanta counties at delivery/termination (14).
proach, we reviewed and reclassified each surveillance rec- MACDP ascertains livebirths, stillbirths, and elective termi-
ord using a modified version of the International Pediatric nations with major structural defects, chromosomal abnor-
and Congenital Cardiac Code implemented in the Society of malities, and clinical syndromes diagnosed by the age of
Thoracic Surgeons Congenital Heart Surgery Database (9–13). 6 years. When available, details from echocardiography,
This activity permitted classification of cardiovascular catheterization, and surgical reports are abstracted; general
malformations according to embryologic (rather than only pregnancy information, such as gestational age and birth
anatomic) considerations. weight, is also collected.
Pregnancies with an estimated date of conception during
January 1, 1986–March 12, 2003, were included in the anal-
MATERIALS AND METHODS ysis. For each conception date, we estimated the number of
Study population pregnancies with a cardiovascular malformation (numera-
tor) and total pregnancies (denominator). Approximately
Vital records for the cohort of liveborn and stillborn in- 0.3% of pregnancies with cardiovascular malformations
fants of at least 20 weeks’ gestation whose mothers resided and 1.9% of total pregnancies were excluded because of
in 1 of 5 central Atlanta counties at delivery were obtained missing or implausible gestational age information (defined
from the Office of Health Information and Policy, Georgia as <20 weeks or >44 weeks). For each pregnancy, we es-
Division of Public Health. The records of infants with an timated the date of conception (assuming that conception

Am J Epidemiol 2009;169:1004–1014
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 logically normal, and we placed infants with more than 1
vital records data. Unfortunately, these estimates were un- cardiovascular malformation in multiple outcome groups
reliable; for 30% of pregnancies, the last menstrual period only when these malformations were thought to be embry-
date was on the 15th of the month. When we subtracted the ologically independent; otherwise, we coded only the major
clinical gestational age estimate from the birth date, too cardiovascular malformation. Further details about this ac-
many last menstrual period dates fell between the 12th tivity are available (12).
and 18th of the month. For our analyses, we excluded infants who had normal
To compensate for this data quality limitation, we used cardiac physiology and those with identified trisomies, evi-
gestational age estimates from MACDP surveillance rec- dence of heterotaxy syndrome, and abnormal cardiac loop-
ords (which are obtained from medical chart review and ing. Results are presented for 12 outcome groups; 3 of these
are frequently based on ultrasound measurements) to es- are aggregate groupings of cardiovascular malformations
timate conception dates for pregnancies with cardiovascu- (Table 1).
lar malformations, because day-of-month patterns were
not evident among these estimates. For the denominators, Ambient air quality data
we modeled the daily count of conceptions. We calculated
the average daily number of conceptions for each month Ambient air quality measurements of daily 8-hour max-
using vital records information. We then created a daily imum ozone and 24-hour average carbon monoxide, nitro-
time-series data set, in which each day was assigned gen dioxide, PM10, and sulfur dioxide were obtained from
its monthly average, and fit a cubic spline with 6 knots the US Environmental Protection Agency Air Quality Sys-
per year to it. The predicted values from this model were tem, Georgia Department of Natural Resources, and the
used as the daily estimates of conceptions (n ¼ 715,500 Metro Atlanta Index. For each pollutant, we selected 1 cen-
pregnancies). tral monitoring station for use in analyses. When central
station measurements of carbon monoxide, nitrogen diox-
Cardiovascular malformation outcome groups ide, and sulfur dioxide were missing, pollution levels at the
central station were modeled by using measurements from
Each MACDP surveillance record with a diagnosis of other stations. The central station for ozone did not operate
a cardiovascular malformation was reviewed by a pediatric during November–February. During 1993–2003, wintertime
cardiologist and classified by using the Society of Thoracic ozone levels were modeled by using ozone measurements
Surgeons Congenital Heart Surgery Database, version 2.30, from a nearby monitor, with the model developed by use of
nomenclature (11–13, 15). This nomenclature is specific to data during the later time period when wintertime ozone
cardiovascular malformations and is more detailed than levels were available. During 1986–1992, wintertime ozone
what is typically used in birth defects surveillance (16– levels were modeled by using maximum temperature and
18). We classified infants with isolated transient newborn 1-hour maximum nitrogen dioxide measurements from
cardiac conditions (e.g., patent foramen ovale) as physio- 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 week 5. We chose this a priori weighting scheme, which
1993–1995, and daily during 1996–2003; linear interpolation emphasizes pollution levels during the center of the window,
between measurements was used to estimate missing PM10 because of uncertainty in the date of conception estimates.
levels. The location of the PM10 central monitoring station We then created 52 strata representing the week-of-year
changed on January 1, 1993, and on January 1, 1998; on as follows: across all calendar years we grouped January 1–
January 1, 1998, the measurement method changed from January 7 as the first week of the year, January 8–January 14
the federal reference method to the tapered element oscillat- as the second week of the year, and so on. We included
ing microbalance method. February 29, when present, in the ninth week of the year
(February 26–March 4). The 52nd week of the year was
Statistical analyses 8 days long (December 24–December 31).
We modeled temporal associations between ambient air
For each conception date (January 1, 1986–March 12, pollution and daily risks of cardiovascular malformations
2003), we estimated the number of pregnancies with a par- using Poisson generalized linear models with a log link
ticular cardiovascular malformation (numerator) and total and scaled variance estimates. We modeled the pollution
pregnancies (denominator). All pregnancies on a particular metric as a continuous variable and used the natural loga-
conception date were assigned the same pollutant metric, rithm of total conceptions as the offset. We included indi-
which was a weighted average of the 35 daily ambient air cator variables for the 52 strata representing week-of-year to
pollution measurements during weeks 3–7 of pregnancy control for potential confounding by factors with seasonal
(a period when the 4 chambers, inflow tract, and outflow tract variation, and we included a cubic spline for day of follow-
of the heart develop (19)). Relative weights were 0.7 for up with 1 knot per year to control for long-term trends. All
measurements during weeks 3 and 7, 0.9 for measurements risk ratios and confidence intervals corresponded to an in-
during weeks 4 and 6, and 1.0 for measurements during 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 tic left heart syndrome, transposition of the great arteries,
gestation pregnancies, limiting analyses to infants with only and tetralogy of Fallot) remained stable over time, the ob-
1 cardiovascular malformation, using unweighted 5-week served prevalence of the less severe lesions (secundum atrial
pollution metrics, and using unweighted 9-week pollution septal defect, valvar pulmonary stenosis, muscular ventric-
metrics (the average of measurements during the first 63 ular septal defect, and perimembranous ventricular septal
days of pregnancy). defect) increased markedly.
Descriptive statistics for the air pollution metrics, which
RESULTS are weighted averages of daily ambient pollution measure-
ments during weeks 3–7 of pregnancy, indicated that all
Surveillance records of pregnancies reaching 20 weeks’ pollutants had seasonal variation (Table 3). Levels of
gestation with an estimated date of conception during PM10, nitrogen dioxide, and sulfur dioxide declined over
January 1, 1986–March 12, 2003, indication of a cardiovas- time. Pairwise Spearman partial correlation coefficients
cular malformation, and no evidence of trisomy were iden- for the pollution metrics are presented in Table 4.
tified (n ¼ 7,102). Of these, only 4,639 (65%) were deemed As shown in Table 5, we observed a statistically signifi-
to have a cardiovascular malformation following review and cant positive association between PM10 and patent ductus
classification. Infants with complex malformations involving arteriosus (for an interquartile range increase in PM10 levels,
abnormal cardiac looping (n ¼ 86) or heterotaxy syndrome risk ratio (RR) ¼ 1.60, 95% CI: 1.11, 2.31). The 95%
(n ¼ 96) were excluded from analysis, irrespective of the confidence intervals for all other associations, presented in
other specific cardiovascular lesions that might have been Table 5, included the null value.
present. From the remaining 4,457 infants, 3,338 (75%) were Results from 5 sensitivity analyses are presented in Tables
included in 1 or more of the outcome groups shown in Table 1. 6–10. We consistently observed positive associations be-
Overall, there were 715,500 liveborn and stillborn infants with tween PM10 and patent ductus arteriosus (less stringent sea-
an estimated date of conception during this period. sonal and long-term temporal control, RR ¼ 1.40, 95% CI:
The prevalence of cardiovascular malformations, shown 1.01, 1.95; limited to single gestation pregnancies, RR ¼
in Table 2 by season and year of conception, suggested some 1.57, 95% CI: 1.07, 2.28; limited to infants with only 1
seasonal variation across outcome groups. Whereas the ob- cardiovascular malformation, RR ¼ 1.70, 95% CI: 1.12,
served prevalence of the relatively severe lesions (hypoplas- 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 those for ozone, were wide and were therefore compatible
1.71, 95% CI: 0.98, 3.00). We observed a positive, statisti- with both no effect and a harmful effect of air pollution.
cally significant association between nitrogen dioxide and Patent ductus arteriosus was not investigated in the 2 pre-
patent ductus arteriosus in the analysis with less stringent vious studies (7, 8), likely because identification of infants
seasonal and long-term temporal control (RR ¼ 1.40, 95% with congenital patent ductus arteriosus is difficult. In our
CI: 1.07, 1.83). Using unweighted 9-week pollution metrics, study population, we used restrictive criteria to exclude patent
we observed positive, statistically significant associations ductus arteriosus in premature or newborn infants and when it
between nitrogen dioxide and both secundum atrial septal occurred as an obligate shunt lesion in the presence of other
defect (RR ¼ 1.58, 95% CI: 1.02, 2.29) and muscular ven- cardiovascular malformations. Of the 2,273 surveillance rec-
tricular septal defect (RR ¼ 1.20, 95% CI: 1.02, 1.41). We ords we reviewed that contained a code for patent ductus
also observed negative, statistically significant associations arteriosus, the records for only 219 infants met these criteria.
between ozone and right ventricular outflow tract defects in We are unaware of experimental animal evidence support-
the analysis limited to infants with only 1 cardiovascular ing an association between PM10 and patent ductus arterio-
malformation (RR ¼ 0.52, 95% CI: 0.29, 0.93) and between sus. Analogous epidemiologic evidence comes from
sulfur dioxide and right ventricular outflow tract defects in Swedish registry data, wherein a positive association was
the analysis based on an unweighted 5-week metric (RR ¼ observed between first trimester maternal smoking and risk
0.73, 95% CI: 0.54, 0.98). of patent ductus arteriosus among term infants (21). One
plausible biologic mechanism for our result relates to fetal
DISCUSSION growth and development. Relative to term infants, premature
infants have a much higher incidence of patent ductus (22,
We investigated temporal associations between ambient air 23). Associations between reduced fetal growth and mater-
pollution levels during weeks 3–7 of pregnancy and risks of nal smoking (24), environmental tobacco smoke (25), and
cardiovascular malformations. Except for the association be- ambient particulate matter (26, 27) have been observed in
tween PM10 and patent ductus arteriosus, all 95% confidence epidemiologic studies. If high PM10 levels in utero restrict
intervals from the primary analysis were consistent with little fetal development, this could explain an association between
or no association. Many confidence intervals, particularly 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 septal defect. In our study, we distinguished among the 4
studies were consistent with no association (7, 8). In southern types of ventricular septal defects, because each is thought to
California, investigators reported 4 significant positive associ- develop through unique embryologic mechanisms (28, 29).
ations from 144 models. Twenty-four of these models aver- We analyzed perimembranous and muscular ventricular sep-
aged pollution levels over weeks 5–8 of pregnancy; the 4 tal defects as distinct outcome groups, and we included sub-
positive associations were observed during this window (7). arterial ventricular septal defects in the conotruncal defects
The Texas study investigators observed 3 positive associations outcome group (Table 1). There were too few inlet ventricular
from 75 models; air pollution levels were averaged during septal defects to permit analysis. We grouped pulmonary
weeks 3–8 of pregnancy (8). No significant association from atresia with ventricular septal defect, which is the extreme
the southern California study was replicated in the Texas study. end of the anatomic spectrum of tetralogy of Fallot (30), in the
Because of our review and classification of surveillance tetralogy of Fallot outcome group. To provide a more direct
records, comparison of results from our study with those comparison with the southern California result, we performed
from previous studies was difficult. Through our review, a secondary analysis in which we ignored our classifications
we excluded infants with structurally normal hearts and tran- and defined isolated ventricular septal defect according to the
sient newborn conditions; 35% of the records we reviewed preexisting codes in the MACDP database. Using our primary
were reclassified as ‘‘structurally normal.’’ Outcome groups analytical approach, we observed no evidence for an associ-
were based on embryologic considerations, and infants with ation between carbon monoxide and isolated ventricular sep-
multiple congenital heart defect codes were included in mul- tal defects (RR ¼ 0.98, 95% CI: 0.82, 1.16).
tiple outcome groups only when the malformations were Our analytical approach and exposure assignment meth-
thought to be embryologically independent. Consequently, ods were also different. Whereas the previous study inves-
our outcome groups differed from those used in previous tigators conducted spatial-temporal analyses using pollution
studies (7, 8). For example, in the southern California study measurements assigned to women on the basis of residential
(7), an association was observed between carbon monoxide location, we opted for a temporal approach using measure-
and isolated ventricular septal defects (fourth quartile vs. ments from centrally located monitors. We implemented
first quartile, odds ratio ¼ 2.95, 95% CI: 1.44, 6.05). This a temporal analysis because of our desire to preclude con-
outcome group incorporated the 4 major types of ventricular cerns about our study results based on arguments of spatial
septal defects, as well as pulmonary atresia with ventricular 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. pregnancies reaching at least 20 weeks’ gestation. Given that
Namely, geocoded data were available dating back only to our gestational window of interest spanned weeks 3–7 of
1994, whereas the central monitor approach allowed us to pregnancy, we would have preferred our cohort to consist of
utilize data back to 1986. Throughout much of follow-up, all pregnancies at week 3. The consequence of this limitation
there were only 2 or 3 monitors for each primary pollutant, might depend on the causal effect of air pollution on cardio-
and we determined that some monitors were unduly im- vascular malformations. For example, atrioventricular septal
pacted by local sources. We were hesitant to use these mea- defect, Ebstein’s anomaly, and tricuspid valve dysplasia can
surements, which likely better reflect local conditions than all cause intrauterine congestive heart failure, increasing the
population exposure. The use of 1 central monitor for each risk of intrauterine fetal death (31). If air pollution were to
pollutant also provided some consistency throughout increase the risk of these malformations, in turn increasing the
follow-up; otherwise, modeling assumptions would have risk of fetal loss before week 20, then our study would have
been needed to account for changes in the number and lo- been unable to detect this harmful effect of pollution.
cation of monitoring stations. For the PM10 analyses, we Measurement error or its discrete counterpart, misclassifi-
relied on such assumptions, and cubic spline knots were cation, which was present in the air quality, vital records, and
placed on the dates when the monitor location and measure- surveillance data, was another limitation in our study as in
ment method changed. To examine the effect of these as- previous studies (7, 8). Our use of ambient air pollution mea-
sumptions, we stratified the data according to measurement surements from stationary monitors as proxies for personal
method (federal reference method: 1986–1997; tapered el- exposure was likely the largest component of measurement
ement oscillating microbalance: 1998–2003). For PM10 and error. If this measurement error was nondifferential, a bias to-
patent ductus arteriosus, we observed a strong association ward the null could explain some null results (32). If the mea-
during 1986–1997 (RR ¼ 1.89, 95% CI: 1.26, 2.84) and no surement error was differential, for example, if the magnitude
association during 1998–2003 (RR ¼ 0.78, 95% CI: 0.28, of measurement error varied according to meteorologic con-
2.04). Perhaps the strong association observed during 1986– ditions, then risk ratio estimates may have been biased toward,
1997 was attributable to average PM10 levels being 42% away from, or across the null. We did not observe broad pat-
higher during this period of follow-up (Table 3). terns of consistently positive or negative risk ratios suggesting
A limitation of our study, also a limitation in previous the possibility of differential measurement error, although dif-
studies (7, 8), was that results were based on the cohort of 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 inter