American Journal of Epidemiology Volume169 Number8 April15 2009

American Journal ofEpidemiologyVolume 169 Number 8 April 15, 2009www.aje.oxfordjournals.org

issn 0002-9262 (print)ISSN 1476-6256 (online)printed in the u.s.a.

ox

fo

rd

AM

ER

ICA

N JO

UR

NA

L OF

EP

IDE

MIO

LOG

Y

ORIGINAL CONTRIBUTIONS

919 Parkinson’s Disease and Residential Exposure to Maneb and Paraquat FromAgricultural Applications in the Central Valley of California. Sadie Costello, Myles Cockburn,Jeff Bronstein, Xinbo Zhang, and Beate Ritz

927 Overweight and Obesity Over the Adult Life Course and Incident Mobility Limitation inOlder Adults: The Health, Aging and Body Composition Study. Denise K. Houston, JingzhongDing, Barbara J. Nicklas, Tamara B. Harris, Jung Sun Lee, Michael C. Nevitt, Susan M. Rubin, FrancesA. Tylavsky, and Stephen B. Kritchevsky for the Health ABC Study

937 Association of Diabetes With Prostate Cancer Risk in the Multiethnic Cohort. KevinM. Waters, Brian E. Henderson, Daniel O. Stram, Peggy Wan, Laurence N. Kolonel, and Christopher A. Haiman

946 Modification of the Effect of Vitamin E Supplementation on the Mortality of MaleSmokers by Age and Dietary Vitamin C. Harri Hemilä and Jaakko Kaprio

954 Dietary Acrylamide Intake and Risk of Premenopausal Breast Cancer. Kathryn M. Wilson,Lorelei A. Mucci, Eunyoung Cho, David J. Hunter, Wendy Y. Chen, and Walter C. Willett

962 Alcohol Intake and Cigarette Smoking and Risk of a Contralateral Breast Cancer: TheWomen’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 StudyCollaborative Group, and Jonine L. Bernstein.

969 Positive Associations Between Ionizing Radiation and Lymphoma Mortality AmongMen. David B. Richardson, Hiromi Sugiyama, Steve Wing, Ritsu Sakata, Eric Grant, Yukiko Shimizu,Nobuo Nishi, Susan Geyer, Midori Soda, Akihiko Suyama, Fumiyoshi Kasagi, and Kazunori Kodama

977 Biomarker-calibrated Energy and Protein Consumption and Increased Cancer RiskAmong Postmenopausal Women. Ross L. Prentice, Pamela A. Shaw, Sheila A. Bingham, ShirleyA. A. Beresford, Bette Caan, Marian L. Neuhouser, Ruth E. Patterson, Marcia L. Stefanick, SuzanneSatterfield, Cynthia A. Thomson, Linda Snetselaar, Asha Thomas, and Lesley F. Tinker

Published for the Johns Hopkins Bloomberg School of Public Health by OxfordUniversity Press Sponsored by the Society for Epidemiologic Research

Founded in 1920 by W. H. Welch and W. H. Howell as the American Journal ofHygiene at the Johns Hopkins School of Hygiene and Public Health

Continued on Inside Front Cover

Volum

e 169 N

umb

er 8 A

pril 15, 2009

Pag

es 919–1042

THE DHHS/NIH/NCI ARE EQUAL OPPORTUNITY EMPLOYERS

Associate Director for Epidemiology and Genetics Research ProgramDivision of Cancer Control and Population Sciences

The National Cancer Institute (NCI), a major research component of the National Institutes of Health (NIH) within the Department of Health and Human Services (DHHS), seeks a senior scientist to serve as an As-sociate Director in its Division of Cancer Control and Population Sciences (DCCPS). The Division provides national scientific leadership and oversight of NCI-funded research in the areas of cancer epidemiology, sur-veillance, health services, survivorship, and behavioral science. DCCPS also is committed to addressing health disparities through transdisciplinary research.

This challenging and highly visible position requires broad scientific expertise, a passion for public service, commitment to collaboration, and an ability to develop effective strategies to identify gaps in research and overcome barriers to scientific progress. Exploiting scientific opportunities requires visionary leadership and sound scientific judgment to ensure the greatest payoff from NCI’s investments. Therefore, a broad perspectiveon cancer epidemiology and other population sciences that informs development of creative and cost-effectivestrategies to advance science is essential.

The Associate Director will lead the Epidemiology and Genetics Research Program (EGRP), which includes itsOffice of the Associate Director and four branches: Clinical and Translational Epidemiology, Host SusceptibilityFactors, Methods and Technologies, and Modifiable Risk Factors. He/she provides scientific and administrativeleadership for the Program, supervises the staff, and represents NCI to a wide variety of professional, academic,and advocacy organizations. The Associate Director also develops and facilitates collaborations with funders ofother types of population science, including the NIH Institutes and Centers, Centers for Disease Control andPrevention (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 370research grants and numerous interagency agreements and consortia.

Qualifications: The successful applicant will be an experienced epidemiologist (M.D. or Ph.D.-level training re-quired) with leadership experience, excellent communication skills, and a strong record of peer-reviewed publica-tions. Strong leadership skills, an ability to work effectively across disciplinary boundaries, and a commitment to thehighest standards of scientific integrity and quality are required. Experience in managing complex research projects,scientific staff, training programs, interdisciplinary collaborations, and/or funded programs is highly valued.

Salary: This position is an excepted service position (Title 42) with a salary range of $160,000-$195,000.

How to Apply: Applications will be considered until the position is filled. Please submit a letter of interest and CV to the Search Committee Chair:

Rachel Ballard-Barbash, M.D., M.P.H., Associate DirectorApplied Research ProgramDivision of Cancer Control and Population SciencesNational Cancer Institute6130 Executive Blvd, Room 4005, MSC 7344 Bethesda, MD 20892-7344Express Mail: Rockville, MD 20852

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

BOARD OF EDITORS MOYSES SZKLO, Editor-in-ChiefDONNA K. ARNETT

TERRI BEATY

HARVEY CHECKOWAY

AARON R. FOLSOM

GARY D. FRIEDMAN

ROBERT J. GLYNN

ICHIRO KAWACHI

MUIN J. KHOURY

MARK KLEBANOFF

MARTHA S. LINET

POLLY MARCHBANKS

POLLY NEWCOMB

JONATHAN M. SAMET

DAVID VLAHOV

CLARICE WEINBERG

POLICY BOARD MICHAEL J. KLAG, ChairmanTODD HUMMEL

SHERMAN JAMES

MALCOLM MACLURE

NOEL R. ROSE

EDITOR-IN-CHIEF EMERITUS,IN MEMORIAM

GEORGE W. COMSTOCK

ASSOCIATE EDITOR INRESIDENCE

SHRUTI H. MEHTA

EDITORIAL PROJECTMANAGERS

DEBORAH J. ANDERSON

SANDRA L. BRYANT

AMY W. REDMON-NORWOOD

ADMINISTRATOR HARRIETT TELLJOHANN

ADMINISTRATIVEASSISTANT

CYNTHIA A. BARBRE

REVIEW COORDINATOR GAYLE FINI

ASSOCIATE EDITORS

ANTHONY ALBERG, Charleston, SCALBERTO ASCHERIO, Boston, MABRAD ASTOR, Baltimore, MDDONNA DAY BAIRD, Research Triangle

Park, NCOLGA BASSO, Research Triangle Park, NCEDWARD J. BOYKO, Seattle, WAROBERT F. BREIMAN, Atlanta, GAGERMAINE M. BUCK, Rockville, MDKENNETH P. CANTOR, Bethesda, MDWONG-HO CHOW, Bethesda, MDJOHN D. CLEMENS, Bethesda, MDSTEPHEN COLE, Chapel Hill, NCGLINDA S. COOPER, Research Triangle Park, NCJOSEF CORESH, Baltimore, MDSTEVEN S. COUGHLIN, Washington, DCROSA M. CRUM, Baltimore, MDRALPH B. D’AGOSTINO, JR., Winston-Salem, NCRALPH B. D’AGOSTINO, SR., Boston, MACONSTANTINE DASKALAKIS, Philadelphia, PAANA V. DIEZ-ROUX, Ann Arbor, MIBRENDA ESKENAZI, Berkeley, CAGUY D. ESLICK, Boston, MAEDUARDO FAERSTEIN, Rio de Janeiro, BrazilKATHERINE M. FLEGAL, Hyattsville, MDBETSY FOXMAN, Ann Arbor, MILIN FRITSCHI, Crawley, AustraliaSANDRO GALEA, Ann Arbor, MIMARILIE D. GAMMON, Chapel Hill, NCPETER J. GERGEN, Rockville, MDRICHARD F. GILLUM, Hyattsville, MDEDWARD GIOVANNUCCI, Boston, MA

MARLENE B. GOLDMAN, Lebanon, NHELISEO GUALLAR, Baltimore, MDM. ELIZABETH HALLORAN, Seattle, WABERNARD L. HARLOW, Minneapolis, MNPATRICIA HARTGE, Rockville, MDMAUREEN C. HATCH, Bethesda, MDRICHARD B. HAYES, New York, NYMIGUEL A. HERNÁN, Boston, MALISA HERRINTON, Oakland, CAIRVA HERTZ-PICCIOTTO, Davis, CADONALD HOOVER, Piscataway, NJJANE A. HOPPIN, Research Triangle Park, NCDAVID J. HUNTER, Boston, MAJOUNI JAAKKOLA, Birmingham, United KingdomSTEVEN J. JACOBSEN, Pasadena, CAKONRAD JAMROZIK, Nedlands, AustraliaSTANISLAV V. KASL, New Haven, CTFREYA KAMEL, Research Triangle Park, NCJAY S. KAUFMAN, Montreal, CanadaWILLIAM C. KNOWLER, Phoenix, AZALAN R. KRISTAL, Seattle, WAMICHAEL F. LEITZMANN, Regensburg, GermanyDE-KUN LI, Oakland, CARONGLING LI, Memphis, TNMARC LIPSITCH, Boston, MAJULIAN LITTLE, Aberdeen, United KingdomSTEPHANIE LONDON, Research Triangle

Park, NCMATTHEW P. LONGNECKER, Research Triangle

Park, NCGERALD MCGWIN, JR., Birmingham, ALJOSEPH K. MCLAUGHLIN, Rockville, MD

ROBERT C. MILLIKAN, Chapel Hill, NCLORENE M. NELSON, Stanford, CAROBERTA B. NESS, Houston, TXCRAIG J. NEWSCHAFFER, Philadelphia, PATHOMAS R. O’BRIEN, Bethesda, MDANDREW OLSHAN, Chapel Hill, NCNANCY PADIAN, San Francisco, CAJULIE R. PALMER, Brookline, MAJULIE PARSONNET, Stanford, CAWENDY S. POST, Baltimore, MDNANCY POTISCHMAN, Bethesda, MDERIC B. RIMM, Boston, MAHARVEY A. RISCH, New Haven, CTLESLIE ROBISON, Minneapolis, MNKARIN ROSENBLATT, Champaign, ILAUDREY F. SAFTLAS, Iowa City, IAGLEN A. SATTEN, Atlanta, GAENRIQUE F. SCHISTERMAN, Bethesda, MDTHOMAS A. SELLERS, Tampa, FLEYAL SHAHAR, Tucson, AZA. RICHEY SHARRETT, Baltimore, MDGARY SHAW, Oakland, CACOLIN L. SOSKOLNE, Edmonton, CanadaMEIR STAMPFER, Boston, MARACHAEL STOLZENBERG-SOLOMON,

Rockville, MDSHOLOM WACHOLDER, Bethesda, MDDOUGLAS L. WEED, Bethesda, MDSCOTT T. WEISS, Boston, MAPHYLLIS A. WINGO, Atlanta, GAJUN ZHANG, Bethesda, MDWEI ZHENG, Nashville, TN

Founded 1920 byW. H. Welch and

W. H. Howell as the American Journal of Hygiene

American Journal of EpidemiologyThe American Journal of Epidemiology (ISSN 0002-9262) is published

twice a month, 24 times a year in 2 volumes (12 issues per volume), byOxford University Press, 2001 Evans Road, Cary, NC 27513-2009, for theJohns Hopkins Bloomberg School of Public Health and is sponsored by theSociety for Epidemiologic Research. Information for authors appears on theJournal’s home page at www.aje.oxfordjournals.org. Authors should consultthese instructions before submitting manuscripts.

Oxford University Press is a department of the University of Oxford. Itfurthers the University’s objective of excellence in research, scholarship,and education by publishing worldwide.

Subscriptions: A subscription to the American Journal of Epidemiologycomprises 24 issues. Prices include postage; for subscribers outside theAmericas, issues are sent air freight.

Annual subscription rate (Volumes 169 and 170, 24 issues, 2009):Institutional: Print and online $762/£508/€762, Online only $724/£483/€724, and Print only $724/£483/€724; Personal: (Combination: the Journaland Epidemiologic Reviews) Online only $355/£237/€355, and Print only$373/£249/€373). Please note: UK£ rate applies to UK and Rest of World,except US ($) and Europe (€). There may be other subscription rates available.For a complete listing please see www.oxfordjournals.org/our_journals/aje/access_purchase/price_list.html.

Full pre-payment in the correct currency is required for all orders.Payment should be in US dollars for orders being delivered to the USA orCanada; Euros for orders being delivered within Europe (excluding theUK); GBP sterling for orders being delivered elsewhere (i.e. not beingdelivered to USA, Canada, or Europe). All orders should be accompaniedby full payment and sent to your nearest Oxford Journals office.Subscriptions are accepted for complete volumes only. Orders are regardedas firm, and payments are not refundable. Our prices include Standard Airas postage outside of the UK. Claims must be notified within four monthsof despatch/order date (whichever is later). Subscriptions in the EEC maybe subject to European VAT. If registered, please supply details to avoidunnecessary charges. For subscriptions that include online versions, a pro-portion of the subscription price may be subject to UK VAT. Subscribers inCanada, please add GST to the prices quoted.

Personal rate subscriptions are only available if payment is made by per-sonal cheque or credit card, delivery is to a private address, and is for personaluse only.

Methods of payment: (i) Check (payable to Oxford University Press,mailed to Oxford University Press, Cashiers Office, Great ClarendonStreet, Oxford OX2 6DP, UK) in GB£ Sterling (drawn on a UK bank), US$Dollars (drawn on a US bank), or EU€ Euros. (ii) Bank transfer to BarclaysBank Plc, Oxford Group Office, Oxford (bank sort code 20-65-18) (UK),overseas only Swift code BARC GB 22 (GB£ Sterling to account no.70299332, IBAN GB89BARC20651870299332; US$ Dollars to accountno. 66014600, IBAN GB27BARC20651866014600; EU€ Euros to accountno. 78923655, IBAN GB16BARC20651878923655). (iii) Credit card(Mastercard, Visa, Switch or American Express).

Back issues: The current year and two previous years’ issues are availablefrom Oxford University Press. Previous volumes can be obtained onlineat http://www.periodicals.com/oxford.html or from the Periodicals ServiceCompany, 11 Main Street, Germantown, NY 12526, USA. E-mail:[email protected]. Tel: (518) 537-4700. Fax: (518) 537-5899.

For further information, please contact: Journals Customer Service,Oxford University Press, Great Clarendon Street, Oxford OX2 6DP, UK.Telephone (& answer-phone outside normal working hours): +44 (0) 1865353907, Fax: +44 (0) 1865 353485, E-mail: [email protected] the US, please contact: Journals Customer Service, Oxford UniversityPress, 2001 Evans Road, Cary, NC 27513, USA, Telephone (& answer-phone outside normal working hours): 800-852-7323 (toll-free in USA/Canada), fax: 919-677-1714, e-mail: [email protected]. In Japan,

please contact: Journals Customer Services, Oxford University Press, 4-5-10-8F Shiba, Minato-ku, Tokyo, 108-8386, Japan. Telephone: +81 35444 5858, Fax: +81 3 3454 2929, e-mail: [email protected].

Advertising: Advertising, inserts, and artwork enquiries should beaddressed to Advertising and Special Sales, Oxford Journals, OxfordUniversity Press, Great Clarendon Street, Oxford OX2 6DP, UK.Telephone: +44 (0) 1865 353329. Fax: +44 (0) 1865 353774. E-mail: [email protected].

Digital Object Identifiers: For information on dois and how to resolvethem, please visit www.doi.org.

Oxford Journals Environmental and Ethical Policies: OxfordJournals is committed to working with the global community to bring thehighest quality research to the widest possible audience. Oxford Journalswill protect the environment by implementing environmentally friendlypolicies and practices wherever possible. Please see http://www.oxfordjournals.org/ethicalpolicies.html for further information on OxfordJournals’ environmental and ethical policies.

Requests for permissions, reprints, and photocopies: All rightsreserved; no part of this publication may be reproduced, stored in a retrievalsystem, or transmitted in any form or by any means (electronic, mechanical,photocopying, recording, or otherwise) without either prior writtenpermission of the publisher (Oxford University Press, Rights and BusinessDevelopment Journals Division, Great Clarendon St., Oxford OX2 6DP,UK. Tel. +44 1865 354490/353695; Fax: +44 1865 353485; E-mail: [email protected]; Web: www.oxfordjournals.org/jnls/permissions) or a license permitting restricted copying, issued in theUS by the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA01923 (Fax: 978-750-4470) or in the UK by the Copyright LicensingAgency Ltd., 90 Tottenham Court Road, London W1P 9HE, UK. Reprintsof individual articles are available only from the authors.

It is a condition of publication in the Journal that authors grant anexclusive license to the Johns Hopkins Bloomberg School of Public Health.However, requests for permission to reprint material found in the Journalshould be submitted to Oxford University Press. This ensures that requestsfrom third parties to reproduce articles are handled efficiently and consis-tently and will also allow the article to be disseminated as widely as possi-ble. As part of the license agreement, authors may use their own material inother publications provided that the Journal is acknowledged as the origi-nal place of publication and Oxford University Press is notified in writingin advance.

The author(s) of each article appearing in this Journal is/are solelyresponsible for the content thereof; the publication of an article shall notconstitute or be deemed to constitute any representation by the Editors, theJohns Hopkins Bloomberg School of Public Health, or the Society forEpidemiologic Research that the data presented therein are correct or suffi-cient to support the conclusions reached or that the experiment design ormethodology is adequate.

The American Journal of Epidemiology is printed on acid-free paperthat meets the minimum requirements of ANSI Standard Z39.48-1984(Permanence of Paper).

The Journal is indexed or abstracted in Index Medicus, the ScienceCitation Index, Current Contents: Life Sciences, and the BIOSIS database.

Copyright © 2009 by the Johns Hopkins Bloomberg School of PublicHealth.

The American Journal of Epidemiology (ISSN 0002-9262) is publishedsemimonthly in Jan., Feb., Mar., Apr., May, Jun., Jul., Aug., Sept., Oct., Nov.,and Dec. by Oxford University Press, 2001 Evans Road, Cary, NC 27513-2009. Periodical Postage paid at Cary, NC and at additional post offices.Postmaster: Send address changes to: American Journal of Epidemiology,Journals Customer Service Department, Oxford University Press, 2001Evans Road, Cary, NC 27513-2009.

American Journal ofEpidemiologyVolume 169 Number 8 April 15, 2009www.aje.oxfordjournals.org

issn 0002-9262 (print)ISSN 1476-6256 (online)printed in the u.s.a.

ox

fo

rd

AM

ER

ICA

N JO

UR

NA

L OF

EP

IDE

MIO

LOG

Y

ORIGINAL CONTRIBUTIONS

919 Parkinson’s Disease and Residential Exposure to Maneb and Paraquat FromAgricultural Applications in the Central Valley of California. Sadie Costello, Myles Cockburn,Jeff Bronstein, Xinbo Zhang, and Beate Ritz

927 Overweight and Obesity Over the Adult Life Course and Incident Mobility Limitation inOlder Adults: The Health, Aging and Body Composition Study. Denise K. Houston, JingzhongDing, Barbara J. Nicklas, Tamara B. Harris, Jung Sun Lee, Michael C. Nevitt, Susan M. Rubin, FrancesA. Tylavsky, and Stephen B. Kritchevsky for the Health ABC Study

937 Association of Diabetes With Prostate Cancer Risk in the Multiethnic Cohort. KevinM. Waters, Brian E. Henderson, Daniel O. Stram, Peggy Wan, Laurence N. Kolonel, and Christopher A. Haiman

946 Modification of the Effect of Vitamin E Supplementation on the Mortality of MaleSmokers by Age and Dietary Vitamin C. Harri Hemilä and Jaakko Kaprio

954 Dietary Acrylamide Intake and Risk of Premenopausal Breast Cancer. Kathryn M. Wilson,Lorelei A. Mucci, Eunyoung Cho, David J. Hunter, Wendy Y. Chen, and Walter C. Willett

962 Alcohol Intake and Cigarette Smoking and Risk of a Contralateral Breast Cancer: TheWomen’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 StudyCollaborative Group, and Jonine L. Bernstein.

969 Positive Associations Between Ionizing Radiation and Lymphoma Mortality AmongMen. David B. Richardson, Hiromi Sugiyama, Steve Wing, Ritsu Sakata, Eric Grant, Yukiko Shimizu,Nobuo Nishi, Susan Geyer, Midori Soda, Akihiko Suyama, Fumiyoshi Kasagi, and Kazunori Kodama

977 Biomarker-calibrated Energy and Protein Consumption and Increased Cancer RiskAmong Postmenopausal Women. Ross L. Prentice, Pamela A. Shaw, Sheila A. Bingham, ShirleyA. A. Beresford, Bette Caan, Marian L. Neuhouser, Ruth E. Patterson, Marcia L. Stefanick, SuzanneSatterfield, Cynthia A. Thomson, Linda Snetselaar, Asha Thomas, and Lesley F. Tinker

Published for the Johns Hopkins Bloomberg School of Public Health by OxfordUniversity Press Sponsored by the Society for Epidemiologic Research

Founded in 1920 by W. H. Welch and W. H. Howell as the American Journal ofHygiene at the Johns Hopkins School of Hygiene and Public Health

Continued on Inside Front Cover

Volum

e 169 N

umb

er 8 A

pril 15, 2009

Pag

es 919–1042

THE DHHS/NIH/NCI ARE EQUAL OPPORTUNITY EMPLOYERS

Associate Director for Epidemiology and Genetics Research ProgramDivision of Cancer Control and Population Sciences

The National Cancer Institute (NCI), a major research component of the National Institutes of Health (NIH) within the Department of Health and Human Services (DHHS), seeks a senior scientist to serve as an As-sociate Director in its Division of Cancer Control and Population Sciences (DCCPS). The Division provides national scientific leadership and oversight of NCI-funded research in the areas of cancer epidemiology, sur-veillance, health services, survivorship, and behavioral science. DCCPS also is committed to addressing health disparities through transdisciplinary research.

This challenging and highly visible position requires broad scientific expertise, a passion for public service, commitment to collaboration, and an ability to develop effective strategies to identify gaps in research and overcome barriers to scientific progress. Exploiting scientific opportunities requires visionary leadership and sound scientific judgment to ensure the greatest payoff from NCI’s investments. Therefore, a broad perspectiveon cancer epidemiology and other population sciences that informs development of creative and cost-effectivestrategies to advance science is essential.

The Associate Director will lead the Epidemiology and Genetics Research Program (EGRP), which includes itsOffice of the Associate Director and four branches: Clinical and Translational Epidemiology, Host SusceptibilityFactors, Methods and Technologies, and Modifiable Risk Factors. He/she provides scientific and administrativeleadership for the Program, supervises the staff, and represents NCI to a wide variety of professional, academic,and advocacy organizations. The Associate Director also develops and facilitates collaborations with funders ofother types of population science, including the NIH Institutes and Centers, Centers for Disease Control andPrevention (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 370research grants and numerous interagency agreements and consortia.

Qualifications: The successful applicant will be an experienced epidemiologist (M.D. or Ph.D.-level training re-quired) with leadership experience, excellent communication skills, and a strong record of peer-reviewed publica-tions. Strong leadership skills, an ability to work effectively across disciplinary boundaries, and a commitment to thehighest standards of scientific integrity and quality are required. Experience in managing complex research projects,scientific staff, training programs, interdisciplinary collaborations, and/or funded programs is highly valued.

Salary: This position is an excepted service position (Title 42) with a salary range of $160,000-$195,000.

How to Apply: Applications will be considered until the position is filled. Please submit a letter of interest and CV to the Search Committee Chair:

Rachel Ballard-Barbash, M.D., M.P.H., Associate DirectorApplied Research ProgramDivision of Cancer Control and Population SciencesNational Cancer Institute6130 Executive Blvd, Room 4005, MSC 7344 Bethesda, MD 20892-7344Express Mail: Rockville, MD 20852

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

Contents Continued from Front Cover

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

996 Serum Selenium and Peripheral Arterial Disease: Results From the National Health and NutritionExamination Survey, 2003–2004. Joachim Bleys, Ana Navas-Acien, Martin Laclaustra, Roberto Pastor-Barriuso, Andy Menke, Jose Ordovas, Saverio Stranges, and Eliseo Guallar

1004 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. DanaFlanders, James A. Mulholland, Csaba Siffel, Michele Marcus, and Paige E. Tolbert

1015 Maternal Urinary Metabolites of Di-(2-Ethylhexyl) Phthalate in Relation to the Timing of Labor in a US Multicenter Pregnancy Cohort Study. Jennifer J. Adibi, Russ Hauser, Paige L. Williams, Robin M. Whyatt, Antonia M. Calafat, Heather Nelson, Robert Herrick, and Shanna H. Swan

1025 Longitudinal Trends in Hazardous Alcohol Consumption Among Women With Human ImmunodeficiencyVirus Infection, 1995–2006. Robert L. Cook, Fang Zhu, Bea Herbeck Belnap, Kathleen Weber, Judith A. Cook, David Vlahov, Tracey E. Wilson, Nancy A. Hessol, Michael Plankey, Andrea A. Howard, Stephen R. Cole, Gerald B. Sharp,Jean L. Richardson, and Mardge H. Cohen

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

BOOK REVIEWS

1039 Hyping Health Risks: Environmental Hazards in Daily Life and the Science of Epidemiology By GeoffreyC. Kabat. David A. Savitz

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/.

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

1

To register for this free service, go towww.oupjournals.org/tocmail, selectthe journal(s) you are interested in,enter your email address, and press‘submit’.

We will automatically email you the forthcomingtables of contents for new issues of journalspublished by Oxford University Press.

Stay alert to forthcoming contents of OUP Journals... for free!

Central Arkansas Veterans Healthcare System (CAVHS)

Health Services Researcher Full-time position

The Geriatric Research Education and Clinical Center (GRECC) at the Central Arkansas Veterans Healthcare System (CAVHS) is seeking a full-time health services researcher. Faculty appointment in the Departments of Geriatrics and/or Health Policy and Management, University of Arkansas for Medical Sciences at the Assistant/Associate Professor level commensurate with experience. Position includes 90% protected time for research. The incumbent must be an American citizen, hold a doctoral degree (PhD or equivalent), and have: (1) completed an approved residency or postdoctoral training program in the area of health services research, or (2) have extensive training, research and dissemination experiences in the area of health services research as documented byprevious research investigations, senior mentor evaluations, presentations at national meetings, and peer-reviewed publications. The ideal candidate will have expertise and interest in the areas of nutrition, exercise, or functional outcomes of older adults.

Women and minorities are encouraged to apply. A nationally competitive salary isavailable and competitive recruitment packages are possible for superior candidates.

For information and application procedures please contact Brandy Fulmer, Human Resources Management Specialist, at [email protected].

American Journal of Epidemiology

ª The Author 2009. Published by the Johns Hopkins Bloomberg School of Public Health.

All rights reserved. For permissions, please e-mail: [email protected].

Vol. 169, No. 8

DOI: 10.1093/aje/kwp006

Advance Access publication March 6, 2009

Original Contribution

Parkinson’s Disease and Residential Exposure to Maneb and Paraquat FromAgricultural 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 toParkinson’s disease (PD). Human data are insufficient to support this claim for any specific pesticide, largelybecause of challenges in exposure assessment. The authors developed and validated an exposure assessmenttool based on geographic information systems that integrated information from California Pesticide Use Reportsand land-use maps to estimate historical exposure to agricultural pesticides in the residential environment. In1998–2007, the authors enrolled 368 incident PD cases and 341 population controls from the Central Valley ofCalifornia in a case-control study. They generated estimates for maneb and paraquat exposures incurred between1974 and 1999. Exposure to both pesticides within 500 m of the home increased PD risk by 75% (95% confidenceinterval (CI): 1.13, 2.73). Persons aged �60 years at the time of diagnosis were at much higher risk when exposedto 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 atyounger 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, toxicmetabolite 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 athigh rates among farmers and in rural populations, contrib-uting to the hypothesis that agricultural pesticides might becausal agents (1–4). Animal studies have linked certain pes-ticides to Parkinsonism and dopaminergic cell death. Thepesticide rotenone can produce the behavioral and neuro-pathologic features of PD in some rodent models throughchronic systemic inhibition of mitochondrial complex I (5,6). Exposure to a combination of the fungicide maneb andthe herbicide paraquat in mice leads to increased substantianigra neuronal pathology (7), age-dependent motor degen-eration, progressive reductions in dopamine metabolites andturnover (8), and reduced tyrosine hydroxylase and dopa-mine transporter immunoreactivity (9, 10).

Human evidence is insufficient to identify any particularpesticide compound, including those implicated by animalstudies, as being responsible for causing PD (11). Method-ological limitations have clouded the interpretation of mostepidemiologic studies exploring pesticide exposures and PDin humans. Past studies have generally relied on self-reportsand recall of chemical usage, making them vulnerable toinformation bias and differential recall bias (12).

Because pesticides applied from the air or ground maydrift from their intended treatment sites, with measurableconcentrations subsequently detected in the air, in plants,and in animals up to several hundred meters from applica-tion sites (13–15), accurate methods of estimating environ-mental exposures in rural communities are sorely needed.

Correspondence to Dr. Sadie Costello, Department of Environmental Health Sciences, School of Public Health, University of California, Berkeley,

50 University Hall, #7360, Berkeley, CA 94720-7360 (e-mail: [email protected]).

919 Am J Epidemiol 2009;169:919–926

Geographic information system (GIS)-based methods of as-sessing exposure to pesticides have become popular in re-cent years and may prove an effective solution whenpesticide data exist. We developed and employed a validatedGIS-based exposure assessment tool to estimate pesticideexposure from applications to agricultural crops, relyingon California Pesticide Use Reporting (PUR) data, land-use maps, and geocoded residential historical locations(16). We investigated whether exposure to the pesticidesmaneb and paraquat, alone and in combination, increasedthe risk of incident PD among residents of the Central Valleyof California, an area well-known for its intensive agricul-ture and potential for pesticide exposure.

MATERIALS AND METHODS

All procedures described have been approved by theUniversity of California, Los Angeles, institutional reviewboard for human subjects, and informed consent was ob-tained from all participants.

Subject recruitment

We used a population-based approach for recruiting casesand controls from a largely agricultural population in Cali-fornia. Details are provided elsewhere (17). Briefly, personswith PD newly diagnosed between January 1998 and January2007 who resided in 1 of 3 central California counties(Fresno, Tulare, or Kern county) and had lived in Californiafor at least 5 years prior to diagnosis were recruited into ourstudy within 3 years of diagnosis. Altogether, 28 (90%) ofthe 31 practicing local neurologists who provided care forPD patients assisted in recruiting cases for this study. Wesolicited collaboration from Kaiser Permanente MedicalCenter (Fresno, California), Kern Medical Center (Bakersfield,California), and Visalia Medical Clinic (Visalia, California)and from the Veterans Administration, PD support groups,local newspapers, and local radio stations that broadcastpublic service announcements.

Of the 1,167 PD cases who were initially invited, 604were not eligible: For 397, the case’s diagnosis date felloutside the 3-year range prior to contact, 51 denied havingreceived a PD diagnosis, 134 lived outside the tricountyarea, and 22 were too ill to participate. Of the 563 eligiblecases, 473 (84%) were examined by a University of Cali-fornia, Los Angeles, movement disorder specialist at leastonce and were confirmed to have clinically ‘‘probable’’ or‘‘possible’’ PD; the remaining 90 potential cases could notbe examined or interviewed (54% withdrew, 32% were tooill or died, and 14% moved out of the area prior to theexamination or did not honor a scheduled appointment).We examined but excluded another 96 patients because theyhad other causes of Parkinsonism. This left us with 377cases; of these, 368 provided all information needed foranalyses.

Controls aged 65 years or older were identified fromMedicare lists in 2001, but because of implementation ofthe Health Insurance Portability and Accountability Act,which prohibits the use of Medicare enrollees, 70% of ourcontrols were recruited from randomly selected tax assessor

residential units (parcels) in each of the 3 counties. Wemailed letters of invitation to a random selection of residen-tial living units and also attempted to identify head-of-household names and telephone numbers for these parcels,using the services of marketing companies and Internetsearches.

We contacted 1,212 potential population controls by mailand/or telephone for eligibility screening. Eligibility criteriawere: 1) not having PD, 2) being at least 35 years of age,3) currently residing primarily in 1 of the 3 designated coun-ties, and 4) having lived in California for at least 5 yearsprior to the screening. Only 1 person per household wasallowed to enroll. Of the potential controls contacted, 457were ineligible: 409 were too young, 44 were terminally ill,and 4 resided primarily outside of the study area. Of the 755eligible controls, 409 (54%) declined participation, were tooill to honor an appointment, or moved out of the area prior tointerview; 346 (46%) were enrolled, and 341 provided allinformation needed for analyses.

Assessment of environmental pesticide exposure

We conducted telephone interviews to obtain demo-graphic and exposure information. Detailed residential his-tory forms were mailed to subjects in advance of theirinterview and were reviewed in person or over the phone.We estimated pesticide exposures in the residential environ-ment from applications to agricultural crops employinga validated GIS-based system, which combined PUR dataand land-use maps (16, 18), to produce estimates of residen-tial ambient pesticide applications within a set distance ofsubjects’ homes. We recorded and geocoded lifetime resi-dential histories and estimated ambient exposures for allhistorical addresses at which participants had resided be-tween 1974 and 1999, the period covered by the PUR data.A technical discussion of our GIS-based approach is pro-vided elsewhere (16); here we briefly summarize the datasources and the exposure modeling process.

Residential addresses. Addresses were automaticallygeocoded to TigerLine files (NAVTEQ (Chicago, Illinois),unpublished data, 2006), and discrepancies were then man-ually resolved in a multistep process similar to that de-scribed by McElroy et al. (19). Resulting locations wererecorded, along with the relevant year range of residence,so they could be matched to the appropriate year-specificPUR and land-use data (below). For our GIS model, werelied on addresses in Fresno, Kern, and Tulare counties(the tricounty area) at which participants had resided be-tween 1974 and 1999. Out of 9,568 total residential yearscontributed by cases (26 years 3 368 cases), 7,593 years(79%) were spent at addresses within the tricounty area ascompared with 6,757 (76%) of 8,866 years contributed bycontrols (26 years 3 341 controls). We geocoded thesetricounty residential addresses for the period 1974–1999with similar precision for cases and controls; that is, bothhad spent 88% of their respective residential years at ad-dresses we considered to have been mapped with high pre-cision (i.e., at the level of a residential parcel, streetaddress, or street intersection rather than a zip code or citycentroid).

920 Costello et al.

Am J Epidemiol 2009;169:919–926

Pesticide use reporting. PUR data are recorded by theCalifornia Department of Pesticide Regulation for any com-mercial application of restricted-use pesticides (defined asagents with harmful environmental or toxicologic effects(20)) and, since 1990, for all commercial uses of pesticidesregardless of toxicologic profile. The location of each PURrecord is referenced to the Public Land Survey System(PLSS), a nationwide grid that parcels land into sectionsat varying resolutions. Each PUR record includes the nameof the pesticide’s active ingredient, the poundage applied,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-cultural pesticide application only to a whole PLSS gridsection, we added information from land-use maps to moreprecisely locate the pesticide application, as described indetail elsewhere (18). The California Department of WaterResources periodically (every 7–10 years) performs county-wide large-scale surveys of land use and crop cover, whichallowed us to identify the locations of specific crops withineach PLSS grid section. Digital maps from more recent(1996–1999) surveys are available (21), and paper mapswere manually digitized for earlier periods (1977–1995).The 1977 land-use survey was conducted closest in timeto 1974, when PUR data became available. We constructedhistorical electronic maps of land use and crop type, andusing the PLSS grid section and the crop type reported in thePUR record, we allocated pesticide applications to an agri-cultural site to which we assigned a GIS-based location.

Deriving estimates of residential pesticide exposure. Thetime-specific total exposure at each location, by pesticide,was derived through summation of exposures over a fixed500-m radius (suggested in previous literature (13, 15, 19))around the home for the relevant years of residence. Thenumbers of pounds of pesticide applied annually per acrewere summed for each residential buffer and weighted bythe proportion of treated acreage in each buffer, resulting inpesticide application rates that could be averaged over spe-cific calendar periods of each subject’s lifetime.

Statistical analysis

We estimated residential exposures to maneb and para-quat, alone and in combination, for the following time win-dows: 1) 1974–1999, 2) 1974–1989, and 3) 1990–1999, toassess the possibility of an extensive induction period priorto PD onset and the influence of age at exposure. We strat-ified models by sex and age (�60 years, >60 years) and, inadditional sensitivity analyses, controlled for exposure tosome groups of pesticides suspected to increase PD risk.

We controlled for occupational exposure to pesticidesamong subjects who had held jobs in the agricultural sector,assigning them to categories of ‘‘likely exposed to pesti-cides’’ when they reported pesticide handling and applica-tions or fieldwork and ‘‘possibly exposed to pesticides’’when they reported managerial, produce processing, andother nonfield farm work; all other subjects were considered‘‘not occupationally exposed to pesticides’’ (22). In somemodels, we also adjusted for residential exposures to groupsof other pesticides that some studies have found to be linked

to dopaminergic cell damage or possibly PD (organochlo-rines, organophosphates, and dithiocarbamates (23) and pro-teasome inhibitors (24)).

We considered the following demographic variables aspotential confounders in all analyses: age (age at diagnosisfor cases and age at interview for controls), sex, race (white,nonwhite), education (<12 years, 12 years, >12 years), andcigarette smoking (current, former, never). We used SAS 9.1(SAS Institute Inc., Cary, North Carolina) to perform un-conditional logistic regression analyses.

RESULTS

Study participants were predominantly Caucasian, overthe age of 60, and without a family history of PD (Table 1).Cases were slightly older than controls, were more oftenmale, and had completed fewer years of education. Theywere also more likely to have been occupationally exposedto pesticides and to be never or former smokers.

We did not find increased risks of PD among subjectsexposed to paraquat alone during the years 1974–1999(Table 2). While the rarity of sole maneb exposure (4 subjects)precluded any meaningful interpretation of the maneb-onlyresults, combined exposure to both maneb and paraquat in-creased the risk of PD by 75% (odds ratio (OR) ¼ 1.75,95% confidence interval (CI): 1.13, 2.73), an effect estimatewhich was essentially unchanged after adjustment foroccupational pesticide exposure (OR ¼ 1.74, 95% CI:1.11, 2.72).

When we examined 2 separate exposure time windows,the years 1974–1989 and 1990–1999, the risk increase ob-served for the whole period was found to be mainly attribut-able to exposures incurred during the earlier window(OR ¼ 2.14, 95% CI: 1.24, 3.68), while being exposedduring the later window did not seem to increase PD risk(Table 2). Furthermore, for younger (�60 years) subjects,exposure to both maneb and paraquat in both windows in-creased PD risk as much as 4- to 6-fold (Table 3). Exposureto either maneb or paraquat alone during 1974–1989 alsoincreased risk of PD in younger subjects (OR ¼ 2.27, 95%CI: 0.91, 5.70). When we examined exposure windowsamong our older subjects (>60 years), combined exposureto both pesticides in the earlier window only (1974–1989)was also associated with a 2-fold increase in PD risk(OR ¼ 2.15, 95% CI: 1.15, 4.02), but no increase was foundfor either the later window (1990–1999) or the combinedexposure periods (Table 3). Stratification by sex suggestedno differences in estimates between males and females.

DISCUSSION

In this population-based case-control study, agriculturalapplication of both maneb and paraquat within 500 m ofa residence during the period 1974–1999 greatly increasedthe risk of developing PD, especially when exposure oc-curred between 1974 and 1989 or when PD was diagnosedat a younger age (�60 years). Exposure to both pesticidesduring the earlier time window (1974–1989) also doubledthe risk for older cases. Associations were particularly

PD and Residential Maneb and Paraquat Exposure 921

Am J Epidemiol 2009;169:919–926

strong for younger-onset patients (�60 years), who wouldhave been children, teenagers, and young adults during theexposure period: Among those exposed in the earlier timewindow, risk was increased more than 4-fold with exposureto both pesticides and more than 2-fold with exposure to just1 of the pesticides. Consistent with some theories regardingthe progression of PD pathology (25), these data suggestthat the critical window of exposure to toxicants may be

years before the onset of motor symptoms which lead todiagnosis.

Pesticide and herbicide exposures have previously beenimplicated in idiopathic PD. Paraquat is structurally similarto the toxic metabolite (MPPþ) of the 1-methyl-4-phenyl-pyridinium ion (a metabolite of 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine), an agent known to induce Parkinsoniansymptoms in humans that has been widely used to study

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

Characteristics, Central Valley of California, 1998–2008

Variable

Cases(n 5 368)

Controls(n 5 341) Odds

Ratio

95%ConfidenceIntervalNo. or Mean % No. or Mean %

Mean age,years (range)

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

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 withParkinson’s disease

55 15 37 11 1.44 0.93, 2.25

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 occupationallyexposed to pesticides

232 63 240 70 1 Reference

Possibly occupationallyexposed to pesticides

26 7 26 8 1.03 0.58, 1.83

Likely occupationallyexposed to pesticides

110 30 75 22 1.52 1.08, 2.14

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 ofcigarette 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.

922 Costello et al.

Am J Epidemiol 2009;169:919–926

Parkinsonism in animal models (26). MPPþ is believed tocause cell death by interfering with mitochondrial respira-tion (27), because it concentrates in mitochondria and in-hibits complex I of the electron transport chain (28). Manylines of evidence point to possible mitochondrial dysfunc-tion in PD. Several genes have been identified in familialforms of PD that are linked to mitochondrial function(PINK1 and DJ1), and in sporadic cases of PD, pathologicfree radical reactions that damage mitochondria and de-crease electron transport activity have been described (29).Impaired electron transport hampers adenosine triphosphateproduction and leads to the diversion of electrons from theirnormal electron transport recipients and, thus, further for-mation of damaging free radicals (29).

Although paraquat is also used to induce Parkinsonism insome animal models, the mechanism by which it producessymptoms is not yet understood (30). Recent mammalianand yeast-cell experiments suggest that mitochondria takeup paraquat actively across their membranes, where com-plex I reduces it to the paraquat radical cation that subse-quently produces mitochondria-damaging superoxide (31).It has also been suggested that maneb may inhibit the ubiq-uitin proteasome system, thereby damaging the dopaminer-gic neuron (24, 32). Additionally, maneb has been linked toParkinsonism in mice also exposed to paraquat. In 3 recentstudies, investigators reported that only when mice wereexposed to a combination of the fungicide maneb and the

Table 2. Odds Ratio for Parkinson’s Disease According to

Residential Ambient Exposure to Maneb and/or Paraquat, Central

Valley of California, 1974–1999

Time Windowand Exposure

Cases(n 5 368)

Controls(n 5 341) Odds

Ratioa

95%ConfidenceIntervalNo. % No. %

1974–1999

Missing data 13 4 13 4

No exposure 115 31 126 37 1 Reference

Paraquat only 149 40 152 45 1.01 0.71, 1.43

Maneb only 3 1 1 0 3.04 0.30, 30.86

Both paraquatand maneb

88 24 49 14 1.75 1.13, 2.73

1974–1989



Paraquat ormaneb only

148 40 137 40 1.25 0.85, 1.85


74 20 39 11 2.14 1.24, 3.68

1990–1999




113 31 95 28 0.96 0.64, 1.43


25 7 18 5 0.93 0.45, 1.94

a Odds ratios were adjusted for age, sex, nonwhite race, education,

and smoking status. Results were mutually adjusted for exposure in

each time window.

Table 3. Odds Ratio for Parkinson’s Disease According to

Residential Ambient Exposure to Maneb and/or Paraquat, by Time

Window of Exposure and Age Group, Central Valley of California,

1974–1999

Age Groupand Exposure

Cases ControlsOddsRatioa

95%ConfidenceIntervalNo. % No. %

1974–1999 Time Window

�60 years




38 48 42 48 1.77 0.84, 3.75


21 27 7 8 5.07 1.75, 14.71

>60 years




114 39 111 44 0.90 0.60, 1.34


67 23 42 17 1.36 0.83, 2.23


�60 years




36 46 34 39 2.27 0.91, 5.70


14 18 6 7 4.17 1.15, 15.16

>60 years




112 39 103 41 1.18 0.75, 1.84


60 21 33 13 2.15 1.15, 4.02


�60 years




27 34 22 25 2.00 0.84, 4.74


7 9 2 2 5.74 0.55, 59.62

>60 years




86 30 73 29 0.78 0.49, 1.24


18 6 16 6 0.66 0.29, 1.50

a Age-stratified models with adjustment for sex, nonwhite race,

education, and smoking status. Results were mutually adjusted for

exposure in each time window.


Am J Epidemiol 2009;169:919–926

herbicide paraquat (paraquat þ maneb), not to either pesti-cide alone, did they exhibit increased neuronal pathology(7), age-dependent motor degeneration and progressive re-ductions in dopamine metabolites and dopamine turnover(8), and reduced tyrosine hydroxylase and dopamine trans-porter immunoreactivity (9).

The fungicide maneb and the herbicide paraquat are bothused in the Central Valley of California and are often usedon the same crops, including potatoes, dry beans, and toma-toes. The average amount of maneb applied near the homesof these study subjects was relatively stable throughout bothtime windows; however, annual paraquat exposure in-creased during the later (1990–1999) time window. Personsliving near fields sprayed with maneb and paraquat may alsobe exposed to a host of other agricultural chemicals. Whenwe controlled for the influence of other groups of pesticidessuspected a priori to be risk factors for PD in our study, theodds ratios for combined maneb and paraquat exposure andPD in the younger subjects were still in the 3- to 6-fold rangeand statistically significant; however, our precision de-creased, probably because of correlated exposures. Correla-tion between pesticides is an inherent problem whenassessing the effects of human exposure. However, sinceadjustment for other pesticides did not remove the associa-tion for maneb and paraquat, our data provide compellingevidence that these 2 pesticides may in fact affect PD risk inhumans, as has been suggested by animal experiments.

Paraquat and maneb are applied by ground, aerial, andbackpack methods; however, paraquat has a much longerfield half-life of 1,000 days (33), as compared with only12–36 days for maneb (34). Both chemicals bind stronglyto soil, though, and are not thought to be a threat to ground-water (35, 36). Such strong binding could result in contam-inated soil getting blown or tracked into homes by wind,pets, and shoes, thereby increasing exposure for personswho live closer to agricultural application sites (3, 37, 38).

In a previous validation study, our prediction model for aserum measure of dichlorodiphenyldichloroethylene (DDE)explained 47% of the biomarker’s variance (39). Addition-ally, our GIS-derived measure of organochlorine exposureidentified persons with high serum DDE levels reasonablywell (specificity of 87%) (39).

Although our GIS model allowed us to calculate the num-ber of pounds of each active ingredient applied per acrewithin a 500-m buffer, these quantities are not comparableacross pesticides. That is, a pound of active ingredient doesnot represent the same human neurotoxicity across pesti-cides, and no information currently exists that would allowus to standardize these measures. Thus, while we believethat our model provided us with an accurate indicator of anypesticide exposure from applications close to a residence,our exposure measure cannot be considered quantitativebeyond a crude rank ordering of low/medium likelihoodof 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 thestatistical power to perform extensive categorical analyses(note that only 3 cases and 1 control were exposed solely tomaneb). We conducted additional analyses after dichoto-mizing pounds per acre at their median and mean levels

and found that exposure to both pesticides at the highestlevel was associated with PD, especially in persons aged�60 years; however, wide confidence intervals surroundingour point estimates rendered these results generally uninfor-mative (results not shown).

In only 1 previous analysis, conducted within the Agri-cultural Health Study cohort (40), did researchers assess theeffects of maneb and paraquat exposures. Statistical powerwas limited by the small number (n ¼ 78) of incident casesidentified during follow-up and the very small number (n ¼4–10) of cases exposed to maneb/mancozeb (OR ¼ 2.1) andparaquat (OR ¼ 1.4). In a small Taiwanese study, the onlycase-control study to date with sufficient statistical power toexamine exposure to the herbicide paraquat, Liou et al. (41)reported a 4- to 6-fold increase in PD risk among long-termapplicators. In a case-control study from the Mayo Clinic(Rochester, Minnesota), Brighina et al. (42) presented asso-ciations between self-reported pesticide exposure and PD insubjects younger than 60 years only (for all pesticides,OR ¼ 1.80, 95% CI: 1.12, 2.87; for herbicides, OR ¼ 2.46,95% CI: 1.34, 4.52).

Our exposure estimates did not depend on the subject’srecall of pesticide exposure and are therefore unlikely tohave been biased by differential exposure misclassification.Since all of our PD diagnoses were clinically confirmed, weexpect disease misclassification to have been minimal. Non-differential exposure misclassification is a possibility in ourstudy and may have attenuated our effect estimates.

Our results may be biased if cases and controls selectedthemselves into our study according to their potential forpesticide exposure, but our subjects were not asked to self-report environmental exposures and probably were unawareof their true historical exposures. There is no reason tosuspect that cases and controls would have chosen to par-ticipate on the basis of their historical residence near certainagricultural plots. We saw no difference in estimated effectswhen we restricted analyses to only those subjects withmore (�12 years) or less (<12 years) education. Similarly,we saw no difference in our results when we restricted thesample to persons whose addresses had been mapped withhigh precision in the tricounty area during the period 1974–1999 (363 cases, 336 controls).

Our analysis has confirmed 2 previous observations fromanimal studies: 1) exposure to multiple chemicals may po-tentiate the effect of each chemical (of interest, since hu-mans are often exposed to more than 1 pesticide in theenvironment) and 2) the timing of exposure is important.To our knowledge, this is the first epidemiologic study toprovide strong evidence that 2 specific pesticides, suggestedby animal research as potentially acting synergistically tobecome neurotoxic, strongly increase the risk of PD inhumans, especially given combined exposure and whenencountered earlier in life.

ACKNOWLEDGMENTS

Author affiliations: Department of Environmental HealthSciences, School of Public Health, University of California,

924 Costello et al.

Am J Epidemiol 2009;169:919–926

Berkeley, Berkeley, California (Sadie Costello); Departmentof Preventive Medicine, Keck School of Medicine, Univer-sity of Southern California, Los Angeles, California (MylesCockburn, Xinbo Zhang); Department of Geography, Collegeof Letters, Arts and Sciences, University of SouthernCalifornia, Los Angeles, California (Myles Cockburn,Xinbo Zhang); Department of Neurology, School of Medicine,University of California, Los Angeles, Los Angeles, Cali-fornia (Jeff Bronstein); and Department of Epidemiology,School of Public Health, University of California, LosAngeles, Los Angeles, California (Beate Ritz).

This work was supported by the National Instituteof Environmental Health Sciences (grants ES10544,U54ES12078, and 5P30 ES07048), the National Institute ofNeurological Disorders and Stroke (grant NS 038367), andthe Department of Defense Prostate Cancer Research Pro-gram (grant 051037). In addition, initial pilot funding wasprovided by the American Parkinson’s Disease Association.

The authors thank the participating neurologists and med-ical centers in Fresno, Kern, and Tulare counties for theirsupport.

Conflict of interest: none declared.

REFERENCES

1. Ben-Shlomo Y, Finnan F, Allwright S, et al. The epidemiologyof Parkinson’s disease in the Republic of Ireland: observationsfrom routine data sources. Ir Med J. 1993;86(6):190–191, 194.

2. Burguera JA, Catala J, Taberner P, et al. Mortality fromParkinson’s disease in Spain (1980–1985). Distribution by age,sex and geographic areas. Neurologia. 1992;7(3):89–93.

3. Morano A, Jimenez-Jimenez FJ, Molina JA, et al. Risk-factorsfor Parkinson’s disease: case-control study in the province ofCaceres, Spain. Acta Neurol Scand. 1994;89(3):164–170.

4. Svenson LW, Platt GH, Woodhead SE. Geographic variationsin the prevalence rates of Parkinson’s disease in Alberta. Can JNeurol Sci. 1993;20(4):307–311.

5. Betarbet R, Sherer TB, MacKenzie G, et al. Chronic systemicpesticide exposure reproduces features of Parkinson’s disease.Nat Neurosci. 2000;3(12):1301–1306.

6. Sherer TB, Betarbet R, Greenamyre JT. Pesticides andParkinson’s disease. Scientific World Journal. 2001;1:207–208.

7. Norris EH, Uryu K, Leight S, et al. Pesticide exposure exac-erbates alpha-synucleinopathy in an A53T transgenic mousemodel. Am J Pathol. 2007;170(2):658–666.

8. Thiruchelvam M, McCormack A, Richfield EK, et al. Age-related irreversible progressive nigrostriatal dopaminergicneurotoxicity in the paraquat and maneb model of theParkinson’s disease phenotype. Eur J Neurosci. 2003;18(3):589–600.

9. Thiruchelvam M, Richfield EK, Baggs RB, et al. The nigro-striatal dopaminergic system as a preferential target of repeatedexposures to combined paraquat and maneb: implications forParkinson’s disease. J Neurosci. 2000;20(24):9207–9214.

10. Thiruchelvam M, Richfield EK, Goodman BM, et al. Devel-opmental exposure to the pesticides paraquat and maneb andthe Parkinson’s disease phenotype. Neurotoxicology. 2002;23(4-5):621–633.

11. Brown TP, Rumsby PC, Capleton AC, et al. Pesticides andParkinson’s disease—is there a link? Environ Health Perspect.2006;114(2):156–164.

12. Seidler A, Hellenbrand W, Robra BP, et al. Possible environ-mental, occupational, and other etiologic factors forParkinson’s disease: a case-control study in Germany.Neurology. 1996;46(5):1275–1284.

13. Chester G, Ward RJ. Occupational exposure and drift hazardduring aerial application of paraquat to cotton. Arch EnvironContam Toxicol. 1984;13(5):551–563.

14. Currier WW, MacCollom GB, Baumann GL. Drift residuesof air-applied carbaryl in an orchard environment. J EconEntomol. 1982;75(6):1062–1068.

15. MacCollom GB, Currier WW, Baumann GL. Drift compari-sons between aerial and ground orchard application. J EconEntomol. 1986;79(2):459–464.

16. Goldberg DW, Wilson JP, Knoblock CA, et al. An effectiveand efficient approach for manually improving geocoded data.Int J Health Geogr. 2008;7:60.

17. Kang GA, Bronstein JM, Masterman DL, et al. Clinicalcharacteristics in early Parkinson’s disease in a centralCalifornia population-based study. Mov Disord. 2005;20(9):1133–1142.

18. Rull RP, Ritz B. Historical pesticide exposure in Californiausing pesticide use reports and land-use surveys: an assess-ment of misclassification error and bias. Environ HealthPerspect. 2003;111(12):1582–1589.

19. McElroy JA, Remington PL, Trentham-Dietz A, et al. Geo-coding addresses from a large population-based study: lessonslearned. Epidemiology. 2003;14(4):399–407.

20. United States Senate Committee on Agriculture, Nutritionand Forestry. Federal Insecticide, Fungicide, andRotenticide Act [As Amended Through P.L. 110–246,Effective May 22, 2008]. Section 3(d)(1)(C).Washington, DC: US Senate, 2008:30. (http://agriculture.senate.gov/Legislation/Compilations/Fifra/FIFRA.pdf).(Accessed February 8, 2009).

21. California Department of Water Resources. California Landand Water Use. Sacramento, CA: California Department ofWater Resources; 2009. (http://www.landwateruse.water.ca.gov/). (Accessed February 8, 2009).

22. Young HA, Mills PK, Riordan D, et al. Use of a crop and jobspecific exposure matrix for estimating cumulative exposure totriazine herbicides among females in a case-control study inthe Central Valley of California. Occup Environ Med. 2004;61(11):945–951.

23. Elbaz A, Tranchant C. Epidemiologic studies of environmentalexposures in Parkinson’s disease. J Neurol Sci. 2007;262(1-2):37–44.

24. Wang XF, Li S, Chou AP, et al. Inhibitory effects of pesticideson proteasome activity: implication in Parkinson’s disease.Neurobiol Dis. 2006;23(1):198–205.

25. Braak H, Del Tredici K, Rub U, et al. Staging of brain pa-thology related to sporadic Parkinson’s disease. NeurobiolAging. 2003;24(2):197–211.

26. Langston JW, Ballard P, Tetrud JW, et al. Chronic Parkinson-ism in humans due to a product of meperidine-analog syn-thesis. Science. 1983;219(4587):979–980.

27. Sayre LM, Wang F, Hoppel CL. Tetraphenylborate potentiatesthe respiratory inhibition by the dopaminergic neurotoxinMPPþ in both electron transport particles and intact mito-chondria. Biochem Biophys Res Commun. 1989;161(2):809–818.

28. Singer TP, Ramsay RR. Mechanism of the neurotoxicity ofMPTP. An update. FEBS Lett. 1990;274(1-2):1–8.

29. Cassarino DS, Bennett JP Jr. An evaluation of the role ofmitochondria in neurodegenerative diseases: mitochondrialmutations and oxidative pathology, protective nuclear


Am J Epidemiol 2009;169:919–926

http://agriculture.senate.gov/Legislation/Compilations/Fifra/FIFRA.pdf

http://agriculture.senate.gov/Legislation/Compilations/Fifra/FIFRA.pdf

http://www.landwateruse.water.ca.gov/

http://www.landwateruse.water.ca.gov/

responses, and cell death in neurodegeneration. Brain ResBrain Res Rev. 1999;29(1):1–25.

30. Dinis-Oliveira RJ, Remiao F, Carmo H, et al. Paraquatexposure as an etiological factor of Parkinson’s disease.Neurotoxicology. 2006;27(6):1110–1122.

31. Cocheme HM, Murphy MP. Complex I is the major site ofmitochondrial superoxide production by paraquat. J BiolChem. 2008;283(4):1786–1798.

32. Zhou Y, Shie FS, Piccardo P, et al. Proteasomal inhibitioninduced by manganese ethylene-bis-dithiocarbamate: rele-vance to Parkinson’s disease. Neuroscience. 2004;128(2):281–291.

33. Extension Toxicology Network. EXTOXNET: Extension Tox-icology Network. Pesticide Information Profiles. Paraquat.Corvallis, OR: Oregon State University; 1996. (http://extoxnet.orst.edu/pips/paraquat.htm). (Accessed June 5, 2008).

34. Extension Toxicology Network. EXTOXNET: ExtensionToxicology Network. Pesticide Information Profiles. Maneb.Corvallis, OR: Oregon State University; 1996. (http://extoxnet.orst.edu/pips/maneb.htm). (Accessed June 5, 2008).

35. Environmental Protection Agency. R.E.D. [ReregistrationEligibility Decision] Facts. Paraquat Dichloride. Washington,DC: Environmental Protection Agency; 1997. (http://www.epa.gov/oppsrrd1/REDs/factsheets/0262fact.pdf). (AccessedJune 5, 2008).

36. Environmental Protection Agency. Reregistration Eligibility

Decision (RED) for Maneb. Washington, DC: Environmental

Protection Agency; 2005. (http://www.epa.gov/oppsrrd1/

REDs/maneb_red.pdf). (Accessed June 5, 2008).37. Jimenez-Jimenez FJ, Mateo D, Gimenez-Roldan S. Exposure

to well water and pesticides in Parkinson’s disease: a case-

control study in the Madrid area. Mov Disord. 1992;7(2):

149–152.38. Rybicki BA, Johnson CC, Uman J, et al. Parkinson’s disease

mortality and the industrial use of heavy metals in Michigan.

Mov Disord. 1993;8(1):87–92.39. Ritz B, Costello S. Geographic model and biomarker-derived

measures of pesticide exposure and Parkinson’s disease. Ann N

Y Acad Sci. 2006;1076:378–387.40. Kamel F, Tanner CM, Umbach DM, et al. Pesticide exposure

and self-reported Parkinson’s disease in the Agricultural

Health Study. Am J Epidemiol. 2007;165(4):364–374.41. Liou HH, Tsai MC, Chen CJ, et al. Environmental risk factors

and Parkinson’s disease: a case-control study in Taiwan.

Neurology 1997;48(6):1583–1588.42. Brighina L, Frigerio R, Schneider NK, et al. Alpha-synuclein,

pesticides, and Parkinson disease: a case-control study. Neu-

rology. 2008;70(16):1461–1469.

926 Costello et al.

Am J Epidemiol 2009;169:919–926

http://extoxnet.orst.edu/pips/paraquat.htm

http://extoxnet.orst.edu/pips/paraquat.htm

http://extoxnet.orst.edu/pips/maneb.htm

http://extoxnet.orst.edu/pips/maneb.htm

http://www.epa.gov/oppsrrd1/REDs/factsheets/0262fact.pdf

http://www.epa.gov/oppsrrd1/REDs/factsheets/0262fact.pdf

http://www.epa.gov/oppsrrd1/REDs/maneb_red.pdf

http://www.epa.gov/oppsrrd1/REDs/maneb_red.pdf




Vol. 169, No. 8

DOI: 10.1093/aje/kwp007



Overweight and Obesity Over the Adult Life Course and Incident MobilityLimitation in Older Adults

The Health, Aging and Body Composition Study

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

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/orobesity 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 measuredweight at ages 70–79 years. Mobility limitation (difficulty walking 1/4 mile (0.4 km) or climbing 10 steps) wasassessed 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 significantgraded 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 ofmobility 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 isexpected to grow to 20% of the population by the year2030 (1). In addition, adults are, on average, increasinglyheavier than earlier generations, with approximately one-third classified as obese (body mass index (BMI)�30 kg/m2)(2). Obesity has been consistently associated with increasedrisk of cardiovascular disease, diabetes, and several cancers,as well as other chronic conditions (3, 4). Obesity in middle-aged and older adults also increases the risk of physicaldisability (5–10), possibly as a result of these obesity-related chronic conditions or by other mechanisms. Thus,the growing prevalence of obesity, particularly among youn-

ger age groups, could reverse the recent declines in disabil-ity rates among future generations of older adults (11–13).

Although the association between obesity in middle andlate adulthood and physical disability is established (5–8,10), few studies have examined the effect of obesity amongyounger adults (14, 15) or the cumulative effect of obesityfrom young adulthood (9, 16) on disability in older adults.In the National Health and Nutrition Examination SurveyEpidemiologic Follow-up Study, being obese or becomingobese over 20 years of follow-up was associated with higherlevels of disability (9). Among Finnish adults aged 55 yearsor 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

on recalled weights at ages 20, 30, 40, and 50 years in-creased the likelihood of reporting walking limitations(16). However, neither of these studies included a measureof incident disability.

In the Health, Aging and Body Composition (HealthABC) Study, we previously showed that, compared withthose of normal weight, men and women who reported be-ing overweight or obese at ages 25, 50, and 70–79 years hadsignificantly worse physical performance at study baseline(ages 70–79 years) based on objective measures (17), pre-dictive of subsequent mobility disability (18). The primaryobjective of the present study was to examine the associa-tion of overweight and/or obesity in young, middle, and lateadulthood, as well as the cumulative effect of overweightand/or obesity across all 3 time points, with incident mobil-ity limitation in men and women in their 70s and 80s.


Study population

Data for this analysis were derived from the Health ABCStudy, a prospective cohort study investigating the associa-tions among body composition, weight-related health con-ditions, and incident functional limitations in older adults.The Health ABC Study enrolled 3,075 community-dwellingmale and female blacks and whites aged 70–79 years be-tween April 1997 and June 1998. Participants were recruitedfrom a random sample of white and all black Medicare-eligible residents in the Pittsburgh, Pennsylvania, andMemphis, Tennessee, metropolitan areas. Participants wereeligible if they 1) reported no difficulty walking one-fourthof a mile (0.4 km), climbing up 10 steps, or performingactivities of daily living; 2) were free of life-threateningillness; 3) planned to remain in the geographic area for atleast 3 years; and 4) were not enrolled in lifestyle interven-tion trials. All participants provided written informed con-sent, and all protocols were approved by the institutionalreview boards at both study sites.

Participants who were missing data on recalled weight atage 25 years (n ¼ 75) or age 50 years (n ¼ 68) or data onrecalled height at age 25 years (n ¼ 39) were excluded.Participants with an extremely low body weight (BMI<15 kg/m2) at ages 25, 50, or 70–79 years (n ¼ 15) or anabsolute weight change of more than 100 pounds (45.4 kg)from age 50 to ages 70–79 years (n¼ 8) were also excluded.Participants who were missing information on mobility lim-itation during follow-up (n ¼ 2) as well as other pertinentbaseline covariates (n ¼ 23) were also excluded. The finalanalysis sample was 2,845 participants.

Mobility limitation

Occurrence of mobility limitation during follow-up wasassessed during annual clinic visits alternating with tele-phone interviews every 6 months. Persistent mobility limi-tation was defined as 2 consecutive reports of having anydifficulty walking one-fourth of a mile (0.4 km) or climbingup 10 steps without resting because of health or a physical

problem. Follow-up included incident cases ascertainedthrough 6.9 years of follow-up, with a mean follow-up of4.2 (standard deviation, 2.4) years.

BMI history

At the baseline visit, participants were asked to recalltheir usual body weight (women were instructed to answerfor a time when not pregnant) and height (without shoes) atage 25 years and to recall their usual body weight at age 50years. Body weight at study baseline (ages 70–79 years) wasmeasured in kilograms by using a standard balance beamscale, and height was measured in millimeters by usinga Harpenden stadiometer (Holtain Ltd., Crosswell, UnitedKingdom). BMI at ages 25 and 50 years was calculated byusing recalled weight at age 25 or 50 years and recalledheight at age 25 years, while BMI at ages 70–79 yearswas calculated by using measured weight and height. BMIwas categorized as normal weight (<25.0 kg/m2), over-weight (25.0–29.9 kg/m2), and obese (�30.0 kg/m2). Addi-tional analyses that excluded underweight (BMI<18.5 kg/m2)participants at ages 25, 50, and 70–79 years were similarto those that included them with the normal-weight group(data not shown).

To assess the cumulative effect of overweight and/orobesity across all 3 time points, the following categorieswere created: normal weight/nonobese at ages 25, 50, and70–79 years; overweight and/or obese at ages 70–79 yearsbut not at ages 25 or 50 years; overweight and/or obese atages 70–79 and 50 years but not at age 25 years; over-weight and/or obese at all 3 time points; and overweightand/or obese at age 50 years but not at ages 70–79 years.Other BMI history patterns were categorized as other. Forthese analyses, recalled height at age 25 years was used tocalculate BMI at ages 70–79 years in addition to BMI atages 25 and 50 years to minimize possible misclassifica-tion bias due to systematic differences between recalledheight at age 25 years and measured height at study base-line. Results were similar when measured height at studybaseline was used to calculate BMI at ages 25, 50, and70–79 years (data not shown).

Potential confounders

Demographic characteristics (age, gender, race, andeducation), smoking status, alcohol consumption, andphysical activity were ascertained by an interviewer-administered questionnaire at study baseline. The timeand intensity of self-reported physical activities performedin the past 7 days, including walking for exercise, otherwalking, climbing stairs, aerobic dance, weight training,and other high- and medium-intensity activities, weresummed (kcal/week). Because BMI history may affect mo-bility as a consequence of weight-related health condi-tions, prevalent health conditions at study baseline wereexamined as potential mediators of the associations. Theprevalence of diabetes, coronary heart disease, congestiveheart failure, stroke, chronic obstructive pulmonary dis-ease, and knee pain were determined by using algorithmsbased on self-report and medication use at study baseline;

928 Houston et al.

Am J Epidemiol 2009;169:927–936

participants with definitive health conditions were coded as‘‘yes.’’ The 20-item Center for Epidemiologic StudiesDepression Scale was used as an indicator of depressedmood, and persons scoring 16 or higher were classifiedas depressed (19). The Modified Mini-Mental State Exam-ination was used as an indicator of general cognitive status,with a minimum score of zero and maximum score of100 (best) (20). Because mean scores on the ModifiedMini-Mental State Examination vary by education, a cut-point of <75 for individuals with less than a high schooleducation and a cutpoint of<80 for individuals with a highschool education were used to classify participants as cog-nitively impaired.

Statistical analyses

Cox proportional hazards regression models were used toexamine the associations between BMI status at ages 25, 50,and 70–79 years and risk of incident mobility limitationwith SAS version 9.1 software (SAS Institute, Inc., Cary,North Carolina). The cumulative effect of overweight and/orobesity across all 3 time points and incident mobility limi-tation was also examined. Participants who survived with noevidence of incident mobility limitation were censored attheir next-to-last 6-month contact. Participants who diedand had no evidence of incident mobility limitation werecensored at their time of death, and those who were lost tofollow-up were censored at their last visit. Two-way inter-actions between gender and BMI status at ages 25, 50, and70–79 years were tested, and interactions were found at thesignificance level of a¼ 0.10. Interactions between race andBMI status at ages 25, 50, and 70–79 years within gendergroup were also tested but were not significant. Thus, in thispaper, all analyses are presented for men and women sepa-rately, with race groups combined. Models were adjusted forage, race, field center, education, smoking, alcohol con-sumption, and physical activity at study baseline. Additionalmodels were also adjusted for prevalent health conditions atstudy baseline. Proportional hazards assumptions were as-sessed by examining log(�log S(t)) plots as well as testinginteractions of each variable with time in the model. Theproportional hazards assumptions were not violated. AllP values were 2 sided.

RESULTS

The mean age of the study population was 73.6 years,50.6% were women, and 39.7% were black. Participantsexcluded from the present analysis (n ¼ 230, 7.5%) weremore likely to be female, black, and older; have less thana high school education; have a higher BMI at ages 70–79years; and report incident mobility limitation during follow-up (P < 0.01). The descriptive characteristics of the studypopulation by gender and incident mobility limitation areshown in Table 1. Approximately 43.0% of men and 53.7%of women reported becoming limited in mobility over the7 years of follow-up. Participants who reported incidentmobility limitation were more likely to be black, have lessthan a high school education, report being a current smoker

and a former drinker, engage in less physical activity, andreport prevalent chronic conditions than those who did notreport incident mobility limitation. Overweight and obesityat ages 25, 50, and 70–79 years were more prevalent amongthose who reported incident mobility limitation. Participantswho reported mobility limitation were less likely to be nor-mal weight or nonobese at all 3 time points.

The hazard ratios and 95% confidence intervals of inci-dent mobility limitation by BMI status at ages 25, 50, and70–79 years for men and women are shown in Table 2. Menwho were obese at age 25 years and overweight or obese atages 50 and 70–79 years were at significantly increased riskof incident mobility limitation compared with men whowere normal weight at ages 25, 50, and 70–79 years, re-spectively. For women, those who were overweight at age25 years and overweight or obese at ages 50 and 70–79 yearswere at significantly increased risk of incident mobility lim-itation compared with women who were normal weight ateach of the 3 time points. Adjusting for prevalent healthconditions at study baseline attenuated the associations be-tween BMI status at ages 25, 50, and 70–79 years and in-cident mobility limitation slightly, but, in general, theassociations remained significant.

Figures 1 and 2 show the hazard ratios and 95% confi-dence intervals for incident mobility limitation associatedwith overweight and/or obesity from age 25 to ages 70–79years. Compared with that for those of normal weight at all3 time points, the risk of incident mobility limitation ap-peared to increase by duration of overweight or obesity(BMI �25 kg/m2) for those who were overweight or obeseat ages 70–79 years (Figure 1). Although fewer participantsreported being obese at age 50 years or earlier, similar asso-ciations were seen for duration of obesity (BMI�30 kg/m2),particularly among men (Figure 2). After excluding partic-ipants who were overweight and/or obese at age 50 years butnot at ages 70–79 years and those who had a BMI historypattern of ‘‘other,’’ we found a significant graded responsefor the cumulative effect of obesity in men (P for trend <0.0001) and overweight and/or obesity in women (P fortrend < 0.0001) on risk of mobility limitation. Womenwho reported being overweight and/or obese at age 50 yearsbut not at ages 70–79 years were at increased risk of incidentmobility limitation compared with those who were normalweight or nonobese at all 3 time points. Men who reportedbeing obese at age 50 years but not at ages 70–79 years werealso at increased risk of incident mobility limitation. Furtheradjustment for prevalent health conditions at study baselineattenuated the associations slightly, but, in general, the as-sociations remained significant.

The cumulative effect of overweight and/or obesity onincident mobility limitation may partly be explained by at-tained BMI at ages 70–79 years. Participants who reportedbeing overweight or obese at age 50 years or earlier hadhigher BMIs in late adulthood compared with those notclassified as being overweight or obese until ages 70–79years (30.8 kg/m2 vs. 28.0 kg/m2, P < 0.0001). Thus, weexamined the cumulative effect of overweight or obesity inanalyses limited to those who were overweight or obese atages 70–79 years, adjusting for age, race, education, fieldcenter, smoking, alcohol consumption, and physical activity

Weight History and Mobility Limitation 929

Am J Epidemiol 2009;169:927–936

at study baseline. Among men, the hazard ratios of incidentmobility limitation were 1.22 (95% confidence interval:0.96, 1.54) for those who were overweight or obese at ages25, 50, and 70–79 years and 1.15 (95% confidence interval:0.92, 1.44) for those who were overweight or obese at ages50 and 70–79 years compared with those who were over-weight or obese at ages 70–79 years but not at ages 50 or 25years (P for trend ¼ 0.15). Among women, the hazardratios of incident mobility limitation were 1.71 (95% con-fidence interval: 1.30, 2.23) for those who were overweightor obese at all 3 time points and 1.23 (95% confidenceinterval: 1.02, 1.47) for those who were overweight orobese at ages 50 and 70–79 years compared with thosewho were overweight or obese at ages 70–79 years butnot earlier (P for trend ¼ 0.0002). Further adjustment for

prevalent health conditions at study baseline attenuated theassociations slightly; however, the trend remained signifi-cant for women.

DISCUSSION

For both men and women, overweight and/or obesity inyoung, middle, and late adulthood were associated with anincreased risk of incident mobility limitation in late adult-hood compared with normal weight. The risk of incidentmobility limitation was approximately 2.8-fold higher forwomen and 1.6-fold higher for men who were overweight orobese at ages 25, 50, and 70–79 years compared with beingnormal weight at all 3 time points. For men, the risk ofincident mobility limitation was approximately 2.4-fold

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 MobilityLimitation(n 5 800)

MobilityLimitation(n 5 604)



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

930 Houston et al.

Am J Epidemiol 2009;169:927–936

higher among those who were obese compared with non-obese at all 3 time points. Although not significant, therewas a trend for women who were obese at all 3 time pointsto have a 1.7-fold higher risk of mobility limitation com-pared with those who were nonobese throughout. Further-more, the risk of incident mobility limitation appeared toincrease with duration of overweight and/or obesity. Amongthose who were normal weight at ages 70–79 years, havinga history of overweight and/or obesity in midlife was asso-ciated with an increased risk of incident mobility limitationcompared with those who were normal weight or nonobeseat all 3 time points. Thus, being overweight and/or obese atany time during adulthood increased the risk of mobilitylimitation later in life, and, the longer the duration of over-weight and/or obesity, the greater the risk. This finding ex-tends prior work showing that overweight and obesity in

young and middle-aged adults predicts physical disabilityin late life (14–16).

Overweight and obesity may lead to joint wear and tear,reduced exercise capacity, and a higher rate of chronic dis-ease such as cardiovascular disease, diabetes, and arthritis,thus resulting in physical disability. The onset of overweightor obesity in young and middle adulthood may result inlower physical activity levels, contributing to decreasedmuscle strength and cardiovascular fitness and greater de-clines in physical function because of longer duration ofexcess body weight and earlier onset of chronic disease.In the National Health and Nutrition Examination SurveyEpidemiologic Follow-up Study, being obese or becomingobese over 20 years of follow-up was associated with higherlevels of upper- and lower-body disability among adultsaged 25–74 years at baseline (9). Among Finnish adults

Table 1. Continued

Men Women





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 atages 70–79 years

24.2 23.0 29.8 31.3

Overweight or obese atages 50 and 70–79 years

23.5 27.5 21.6 32.7

Overweight or obeseat ages 25, 50, and70–79 years

17.2 22.7 3.0 10.0

Overweight or obese atage 50 years but normalweight at ages 70–79 years

6.8 5.5 1.2 3.1

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 and70–79 years

3.1 6.1 2.7 9.3

Obese at ages 25, 50,and 70–79 years

0.2 1.2 0.9 1.8

Obese at age 50 years butnonobese at ages70–79 years

2.2 5.0 0.9 2.8

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).


Am J Epidemiol 2009;169:927–936

aged 55 years or older, those who reported being obese atage 30, 40, or 50 years had a 4-fold or higher increasedlikelihood of walking limitations, and there was a significantlinear trend between duration of obesity and walking limi-tations (16). However, reverse causality cannot be ruled outbecause neither of these studies included a measure of in-cident disability. In the Health ABC cohort, participantswho had a history of being overweight and/or obese at ages25, 50, and/or 70–79 years had an increased risk of incidentmobility limitation compared with those who were normalweight at all 3 time points. Adjustment for prevalent di-abetes, cardiovascular disease, pulmonary disease, andknee pain attenuated the associations slightly; however,the associations between BMI history and incident mobil-ity remained.

Whether overweight and/or obesity in young and middleadulthood are independent of overweight and/or obesity inlate adulthood is difficult to untangle. In the Health ABCcohort, recalled BMI at ages 25 and 50 years and measuredBMI at ages 70–79 years were moderately correlated (r ¼0.33 and r¼ 0.64, respectively). This finding raises the issueof whether BMI in earlier life or attained BMI at ages 70–79years was the driving force behind the associations seenbetween recalled BMI at ages 25 and 50 years and mobilitylimitation in late adulthood. Thus, caution must be exercisedwhen interpreting the results. Participants who reported be-ing overweight or obese in midlife or earlier were heavier inlate adulthood than those who became overweight or obeselater in life. Nonetheless, there appeared to be a graded re-sponse between age at onset of overweight and/or obesityand increased risk of mobility limitation, particularly for

women. Furthermore, a history of overweight and/or obesityin midlife or earlier was associated with increased risk ofincident mobility limitation among those who were normalweight at ages 70–79 years.

A limitation of the work presented here is the use ofrecalled weight and height from the distant past. However,previous studies have shown high correlations (r �0.80)between recalled and measured weight in young adulthoodamong middle-aged and older men and women (21–23). Inthe Health ABC cohort, self-reported and measured weightat study baseline (ages 70–79 years) was highly correlated(r ¼ 0.98). The correlation between recalled height at age25 years and measured height at study baseline was alsohigh (r ¼ 0.93; mean difference, 3 cm). Other studies havealso shown high correlations between self-reported andmeasured height among older adults (r �0.77), with a meandifference of approximately 2–3 cm (24, 25). However, theoverreporting of height by older adults may be due, in part,to loss of height associated with aging (26) and may moreaccurately reflect height in younger adulthood. Results weresimilar when measured height at study baseline was used inthe analysis in place of recalled height at age 25 years tocalculate BMI.

Important characteristics of the Health ABC cohort limitgeneralization of these findings. First, participants were re-cruited to be well functioning and free of mobility limitationat study baseline. Thus, selection bias may have weakenedthe association between BMI history and incident mobilitylimitation because persons with a history of overweight and/or obesity may have developed mobility limitation prior tostudy entry and were excluded from the study. Another

Table 2. Incident Mobility Limitation AmongMen andWomen 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

No.Model 1b Model 2c

No.Model 1b Model 2c

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.

932 Houston et al.

Am J Epidemiol 2009;169:927–936

limitation is the use of self-reported mobility limitation asthe primary endpoint. However, previous studies have

shown that self-reported mobility limitation is valid andhas clinical significance (27). Furthermore, the use of

Figure 1. Hazard ratios and 95% confidence intervals for incident mobility limitation among men (A) and women (B) by history of overweight orobesity (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.


Am J Epidemiol 2009;169:927–936

2 consecutive reports of mobility limitation reduces the in-fluence of transient mobility limitation. Use of BMI as a sur-

rogate of body fatness is also a limitation of these analyses.Although BMI is correlated with body fatness in young and

Figure 2. Hazard ratios and 95% confidence intervals for incident mobility limitation among men (A) and women (B) by history of obesity (bodymass 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.

934 Houston et al.

Am J Epidemiol 2009;169:927–936

middle-aged adults, changes in body composition that ac-company aging alter the relation between BMI and bodyfatness in older adults such that older adults have more bodyfat for a given BMI compared with younger adults (28).Finally, the observational nature of our study did not enableus to evaluate a causal association between BMI history andmobility limitation. It is biologically plausible that a historyof overweight and/or obesity may result in mobility limita-tion. However, although the analyses were adjusted for be-havioral characteristics such as smoking and physicalactivity and for prevalent health conditions at study base-line, overweight and/or obesity may serve as a proxy mea-sure for other relevant participant characteristics that maylead to both overweight and obesity and mobility limitation.

In conclusion, overweight and obesity in young, middle,and late adulthood was associated with increased risk ofincident mobility limitation in this well-functioning cohortof older adults. Those with a history of overweight and/orobesity in midlife or earlier but not in late adulthood alsotended to have an increased risk of incident mobility limi-tation compared with those who were normal weightthroughout. Furthermore, there appeared to be a graded re-sponse between age at onset of overweight and/or obesityand risk of mobility limitation. These data suggest that in-terventions targeting prevention of overweight and obesityin young and middle-aged adults may be useful in prevent-ing or delaying the onset of mobility limitation later in life.

ACKNOWLEDGMENTS

Author affiliations: Sticht Center on Aging, Wake ForestUniversity School of Medicine, Winston-Salem, North Car-olina (Denise K. Houston, Jingzhong Ding, Barbara J.Nicklas, Stephen B. Kritchevsky); Laboratory of Epidemi-ology, Demography, and Biometry, National Institute onAging, Bethesda, Maryland (Tamara B. Harris); Departmentof Foods and Nutrition, University of Georgia, Athens,Georgia (Jung Sun Lee); Department of Epidemiology andBiostatistics, University of California, San Francisco, Cali-fornia (Michael C. Nevitt, Susan M. Rubin); and Departmentof Preventive Medicine, University of Tennessee Health Sci-ence Center, Memphis, Tennessee (Frances A. Tylavsky).

This work was supported in part by the IntramuralResearch Program of the National Institute on Aging,National Institutes of Health; National Institute on Aging,National Institutes of Health (contracts N01-AG-6-2101,N01-AG-6-2103, and N01-AG-6-2106); and the WakeForest University Claude D. Pepper Older AmericansIndependence Center (grant P30-AG21332 to D. K. H.).


REFERENCES

1. Federal Interagency Forum on Aging-Related Statistics. OlderAmericans 2004: Key Indicators of Well-Being. Washington,DC: US Government Printing Office; 2004.

2. Ogden CL, Carroll MD, Curtin LR, et al. Prevalence of over-weight and obesity in the United States, 1999–2004. JAMA.2006;295(13):1549–1555.

3. Willett WC, Dietz WH, Colditz GA. Guidelines for healthyweight. N Engl J Med. 1999;341(6):427–434.

4. Overweight, obesity, and health risk. National Task Force onthe Prevention and Treatment of Obesity. Arch Intern Med.2000;160(7):898–904.

5. LaCroix AZ, Guralnik JM, Berkman LF, et al. Maintainingmobility in late life. II, Smoking, alcohol consumption,physical activity, and body mass index. Am J Epidemiol.1993;137(8):858–869.

6. Launer LJ, Harris T, Rumpel C, et al. Body mass index, weightchange, and risk of mobility disability in middle-aged andolder women: the epidemiologic follow-up study of NHANESI. JAMA. 1994;271(14):1093–1098.

7. Jensen GL, Friedmann JM. Obesity is associated with func-tional decline in community-dwelling rural older persons.J Am Geriatr Soc. 2002;50(5):918–923.

8. Ostbye T, Taylor DH Jr, Krause KM, et al. The role of smokingand other modifiable lifestyle risk factors in maintaining andrestoring lower body mobility in middle-aged and olderAmericans: results from the HRS and AHEAD. Health andRetirement Study. Asset and Health Dynamics Among theOldest Old. J Am Geriatr Soc. 2002;50(4):691–699.

9. Ferraro KF, Su YP, Gretebeck RJ, et al. Body mass index anddisability in adulthood: a 20-year panel study. Am J PublicHealth. 2002;92(5):834–840.

10. Daviglus ML, Liu K, Yan LL, et al. Body mass index in middleage and health-related quality of life in older age. Arch InternMed. 2003;163(20):2448–2455.

11. Sturm R, Ringel JS, Andreyeva T. Increasing obesity ratesand disability trends. Health Aff (Millwood). 2004;23(2):199–205.

12. Lakdawalla DN, Bhattacharya J, Goldman DP. Are the youngbecoming more disabled? Health Aff (Millwood). 2004;23(1):168–176.

13. Alley DE, Chang VW. The changing relationship of obesityand disability, 1988–2004. JAMA. 2007;298(17):2020–2027.

14. Hubert HB, Bloch DA, Fries JF. Risk factors for physicaldisability in an aging cohort: the NHANES I EpidemiologicFollowup Study. J Rheumatol. 1993;20(3):480–488.

15. Houston DK, Stevens J, Cai J, et al. Role of weight history onfunctional limitations and disability in late adulthood: theARIC study. Obes Res. 2005;13(10):1793–1802.

16. Stenholm S, Rantanen T, Alanen E, et al. Obesity history asa predictor of walking limitation at old age. Obesity (SilverSpring). 2007;15(4):929–938.

17. Houston DK, Ding J, Nicklas BJ, et al. The association be-tween weight history and physical performance in the Health,Aging and Body Composition study. Int J Obes (Lond). 2007;31(11):1680–1687.

18. Guralnik JM, Ferrucci L, Pieper CF, et al. Lower extremityfunction and subsequent disability: consistency across studies,predictive models, and value of gait speed alone comparedwith the short physical performance battery. J Gerontol A BiolSci Med Sci. 2000;55(4):M221–M231.

19. Radloff L. The CES-D Scale: a self-report depression scale forresearch in the general population. Appl Psychol Meas. 1977;1:385–401.

20. Teng EL, Chui HC. The Modified Mini-Mental State (3MS)examination. J Clin Psychiatry. 1987;48(8):314–318.

21. Stevens J, Keil JE, Waid LR, et al. Accuracy of current, 4-year,and 28-year self-reported body weights in an elderly popula-tion. Am J Epidemiol. 1990;132(6):1156–1163.


Am J Epidemiol 2009;169:927–936

22. Rhoads GG, Kagan A. The relation of coronary disease,stroke, and mortality to weight in youth and in middle age.Lancet. 1983;1(8323):492–495.

23. Casey VA, Dwyer JT, Berkey CS, et al. Long-term memory ofbody weight and past weight satisfaction: a longitudinalfollow-up study. Am J Clin Nutr. 1991;53(6):1493–1498.

24. Kuczmarski MF, Kuczmarski RJ, Najjar M. Effects of age onvalidity of self-reported height, weight, and body mass index:findings from the Third National Health and Nutrition Exam-ination Survey, 1988–1994. J Am Diet Assoc. 2001;101(1):28–34.

25. Gunnell D, Berney L, Holland P, et al. How accurately areheight, weight and leg length reported by the elderly, and how

closely are they related to measurements recorded in child-hood? Int J Epidemiol. 2000;29(3):456–464.

26. Sorkin JD, Muller DC, Andres R. Longitudinal change in theheights of men and women: consequential effects on bodymass index. Epidemiol Rev. 1999;21(2):247–260.

27. Fried LP, Young Y, Rubin G, et al. Self-reported preclinicaldisability identifies older women with early declines in per-formance and early disease. J Clin Epidemiol. 2001;54(9):889–901.

28. Villareal DT, Apovian CM, Kushner RF, et al. Obesity in olderadults: technical review and position statement of the Ameri-can Society for Nutrition and NAASO, The Obesity Society.Am J Clin Nutr. 2005;82(5):923–934.

936 Houston et al.

Am J Epidemiol 2009;169:927–936




Vol. 169, No. 8

DOI: 10.1093/aje/kwp003

Advance Access publication February 24, 2009


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, andChristopher 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. Thebiologic basis of this association is unknown. The authors have examined whether the association is robust acrosspopulations in a population-based prospective study. The analysis included 5,941 prostate cancer cases identifiedover 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% confidenceinterval (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) amongEuropean Americans to 0.89 (95% CI: 0.77, 1.03; P ¼ 0.13) among African Americans. Mean PSA levels weresignificantly lower in diabetics than in nondiabetics (mean PSA levels, 1.07 and 1.28, respectively; P ¼ 0.003) aswere PSA screening frequencies (44.7% vs. 48.6%; P < 0.001); however, this difference could explain only a smallportion (~20%) of the inverse association between these diseases. Diabetes is a protective factor for prostatecancer 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 mostcommon chronic diseases that afflict the aging male popu-lation. Epidemiologic studies conducted primarily in popu-lations of European ancestry have provided evidence of aninverse relation between these diseases, with diabetics hav-ing ~20% lower risk of developing prostate cancer than donondiabetics (1–8). However, considerable effect heteroge-neity has been noted among studies, highlighting the needfor additional prospective analyses of these endpoints inlarge representative population-based studies. The biologicbasis of this suspected relation is currently unknown and,aside from age and perhaps obesity (4, 9, 10), these 2 dis-eases share no known nongenetic risk factors.

The inverse relation between these endpoints may be dueto direct effects on prostate cancer growth and development,as men with type 2 diabetes have been found to have lowerprostate-specific antigen (PSA) levels, on average, than do

men without type 2 diabetes (11, 12). The reported protec-tive effects of type 2 diabetes may also be attributed todifferences in prostate cancer screening in diabetic and non-diabetic patients. Differences in health maintenance, accessto medical care, and the presence of serious medical con-ditions may result in more (or less) medical attention andpreventive measures (13). Thus, examining the associationbetween type 2 diabetes status and prostate cancer screeningfrequencies is important to quantify the degree to whichdetection bias may explain the apparent relation.

The incidence rates of type 2 diabetes and prostate cancervary widely across populations. However, the ethnic dispar-ities for these common diseases are not correlated; that is,not all populations with high rates of diabetes are at low riskof prostate cancer (14–20). The extent to which these dis-eases are linked in non-European populations is not clear.To confirm the previously reported association in a large,

Correspondence to Dr. Christopher A. Haiman, Harlyne Norris Research Tower, 1450 Biggy Street, Room 1504A, Mail Code LG591 MC9601,

Los Angeles, CA 90033 (e-mail: [email protected]).

937 Am J Epidemiol 2009;169:937–945

representative population-based prospective study, as wellas to examine the consistency of the association acrossracial/ethnic populations with differing rates of prostatecancer, we evaluated prostate cancer incidence by type2 diabetes status in a multiethnic sample of 86,303 menfrom the Multiethnic Cohort (21). We also assessed the pre-sumed effect of diabetes status in the etiology of prostatecancer by examining PSA levels in a multiethnic sample ofmen with and without type 2 diabetes. We also evaluatedPSA screening frequencies in diabetics and nondiabetics todefine the role of screening bias in explaining the observedassociation between these common diseases.


Study population

The Multiethnic Cohort is a prospective cohort study thatincludes 215,251 men and women, the majority from 5racial/ethnic groups in Hawaii and Los Angeles, California(African Americans, European Americans, Native Hawaiians,Japanese Americans, and Latinos) (21). Between 1993 and1996, participants entered the cohort by completing a26-page, self-administered questionnaire that asked aboutdiet and demographic factors, personal behaviors (e.g.,physical activity), history of prior medical conditions(e.g., diabetes), and family history of common cancers.Potential cohort members were identified primarily throughDepartment of Motor Vehicles drivers’ license files and,additionally for African Americans, Health Care FinancingAdministration data files. Participants were between the agesof 45 and 75 years at the time of recruitment.

In the cohort, incident prostate cancer cases are identifiedannually through cohort linkage to the population-basedSurveillance, Epidemiology, and End Results (SEER) can-cer registries in Hawaii and Los Angeles County, as well asthe California Cancer Registry. Information on stage andgrade of disease is also obtained through these registries.Linkage with these registries is complete through December31, 2004, in Hawaii and December 31, 2005, in California.Over this period, 5,941 incident cases of invasive prostatecancer were identified. Deaths within the cohort are deter-mined from linkages to the death certificate files in Hawaiiand California, supplemented with linkages to the NationalDeath Index. In the Multiethnic Cohort, diabetes status isdefined on the basis of self-report on the baseline question-naire. This question did not differentiate between type 1diabetes and type 2 diabetes, and thus we expect a smallfraction (<10%) of the respondents to have type 1 diabetesand to be potentially misclassified (22).

In addition to self-reported race/ethnicity, the followingrisk factors were included in the analysis: body mass index(weight (kg)/height (m)2), educational level (�12 years,some college or vocational, and college graduate), first-degree family history of prostate cancer, and amount ofvigorous physical activity (0, >0–1.5, >1.5–5, and >5 hours/week). Vigorous activity includes both vigorous sports andvigorous work.

We limited our analysis to 91,018 men in the MultiethnicCohort from the 5 major racial/ethnic groups. We excluded

men with a prevalent report of prostate cancer (n ¼ 3,004)based on self-report or from the Surveillance, Epidemiol-ogy, and End Results registries. We excluded men withmissing information for body mass index (n ¼ 838), educa-tional level (n ¼ 872), and diabetes status (n ¼ 1). Theprospective analysis of the association between diabetesstatus and prostate cancer incidence in this study includes86,303 men.

PSA levels were previously measured on 4,623 men in theMultiethnic Cohort (23). These men were randomly selectedfrom the cohort to evaluate the distribution of PSA levelsacross ethnic groups. We excluded 194 men with prevalentprostate cancer at baseline. We also excluded 1,527 menwith incident prostate cancer during the follow-up period,to ensure that elevated PSA levels among undiagnosed casesdid not influence the results. Another 28 men with missingbody mass index data were excluded from the analysis,leaving 2,874 men who are included in the final analysisof the effect of type 2 diabetes on PSA levels.

In 2001, we sent a short follow-up questionnaire to cohortmembers. On this questionnaire, we also asked about PSAscreening prior to 1999. Of the 86,303 men included in theprimary analysis of type 2 diabetes and prostate cancer,23,768 (27.5%) did not complete the follow-up question-naire. We also excluded 4,649 men with incident prostatecancer. Finally, we excluded men under the age of 50 years(n ¼ 10,916) because annual PSA screening is recommen-ded to begin at age 50 (24). This leaves 46,970 men includedin the analysis of the association between type 2 diabetesand PSA screening.

The informed consent and study protocol were approvedby the institutional review boards at the University of South-ern California and the University of Hawaii.


Cox regression was used to estimate hazard ratios (re-ported as relative risks) for the effect of type 2 diabetes onprostate cancer incidence (STATA, version 8, software; Sta-taCorp LP, College Station, Texas). We adjusted for age,body mass index, educational level, and race/ethnicity (inpooled analyses). Neither body mass index nor educationallevel was associated with prostate cancer risk, but both re-mained in the model as the former was found to be associ-ated with PSA levels and the latter was found to be a highlysignificant predictor of PSA screening. Physical activity andfamily history of prostate cancer were left out of the finalmodel because neither had an effect on the association be-tween type 2 diabetes and prostate cancer. Stratified analy-ses were performed in older age groups to assess whethertype 2 diabetes duration and long-term exposure to declin-ing insulin levels may be important in prostate cancer de-velopment. Because men may be at increased risk ofprostate cancer within the first few years following a diabetesdiagnosis as a result of higher insulin levels, and since thedate of type 2 diabetes diagnosis is unknown for cohortmembers, we also performed a sensitivity analysis to exam-ine whether the association might be attenuated in recentlydiagnosed diabetics. In this analysis, we censored follow-upof incident prostate cancer cases incrementally by year from

938 Waters et al.

Am J Epidemiol 2009;169:937–945

1 to 5 years after cohort entry. We also examined the asso-ciation in analyses stratified by body mass index (�25 kg/m2 and <25 kg/m2). Analyses stratified by Gleason score todetermine the effect of type 2 diabetes status on prostatecancer severity were also conducted. This latter analysisexcludes 370 prostate cancer cases with missing informationon the Gleason score.

In the analysis of PSA levels, generalized linear modelswere used to estimate least-squared mean PSA levels bytype 2 diabetes status (SAS, version 9.1, software; SAS In-stitute, Inc., Cary, North Carolina). Models were adjustedfor the putative confounders of age, body mass index, andrace/ethnicity. We calculated PSA screening frequenciesadjusted for both age and educational level by type 2 di-abetes status, and we tested for a difference using logisticregression; body mass index was not found to influence theeffect of type 2 diabetes on PSA screening. The fraction ofthe association between type 2 diabetes and prostate cancerthat may be attributable to PSA screening was estimated.Assuming that prostate cancer incidence roughly doubled

since the initiation of PSA screening (25), with about 50%of men being screened, we estimate that incidence rateshave increased by 0.02 per 1% of the population screened.We then used this slope to estimate the relative impact ofscreening on prostate cancer incidence in diabetic and non-diabetic men as follows: relative risk (RR)PSA ¼ (1 þ 0.023screening frequency in nondiabetics)/(1 þ 0.02 3 screeningfrequency in diabetics), with (1 � RRPSA/1 � RRT2D) beingan estimate of the fraction of the association between type 2diabetes (T2D) and prostate cancer incidence attributable toPSA screening.

RESULTS

The mean age of the men (n ¼ 86,303) in this study was59.9 (standard deviation, 8.8) years and ranged from 56.6for Native Hawaiians to 61.3 for African Americans (Table 1).The age-standardized prostate cancer incidence rate of830.2 (per 100,000) was nearly 2 times greater in African

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

EuropeanAmericans

AfricanAmericans

NativeHawaiians

JapaneseAmericans

LatinosTotal

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 prostatecancer cases

60 1,193 215 1,295 35 213 160 1,302 198 1,270 5,941

Prostate cancerincidence ratesa

242.8 413.9 686.2 845.2 257.9 373.8 270.1 354.9 304.2 433.7

Family history ofprostate cancer, %b

7.0 7.9 7.9 8.8 5.6 5.7 6.1 6.2 5.1 5.8

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 orvocational

29.0 29.1 36.9 37.0 25.8 28.5 34.3 30.5 22.5 23.5

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.

Type 2 Diabetes and Prostate Cancer Risk 939

Am J Epidemiol 2009;169:937–945

Americans than in the other populations (Table 1). The age-adjusted prevalence of type 2 diabetes also varied widelyacross populations, from 6.9% in European Americans to18.0% in Native Hawaiians. The mean age of the diabeticmen in our study at baseline was slightly higher than that ofthe nondiabetic men for each racial/ethnic group, rangingfrom 59.2 years (vs. 56.1 in nondiabetics) in Native Hawai-ians to 63.6 years (vs. 60.6 in nondiabetics) in JapaneseAmericans. The age-standardized prostate cancer incidencerates were lower in diabetic men than in nondiabetic men foreach racial/ethnic group, ranging from 242.8 (per 100,000)in European-American diabetics (vs. 413.9 in nondiabetics)to 686.2 in African-American diabetics (vs. 845.2 in non-diabetics). First-degree family history of prostate cancerwas also less common in diabetic men than in nondiabeticmen for each racial/ethnic group, ranging from 5.1% inLatino diabetic men (vs. 5.8% in nondiabetics) to 7.9% inAfrican-American diabetic men (vs. 8.8% in nondiabetics).As expected, in each population, diabetic men were morelikely to be overweight and less physically active than menwithout type 2 diabetes (Table 1).

In multivariate analyses, men with type 2 diabetes hadsignificantly lower risk of prostate cancer than did menwithout type 2 diabetes (RR ¼ 0.81, 95% confidence in-terval (CI): 0.74, 0.87; P < 0.001) (Table 2). The inverseassociation was observed consistently in all 5 populationsand ranged from 0.65 (95% CI: 0.50, 0.84) in EuropeanAmericans to 0.89 (95% CI: 0.77, 1.03) in African Ameri-cans (Pheterogeneity ¼ 0.32). We also examined the effect oftype 2 diabetes status on prostate cancer incidence by age atentry into the cohort as a surrogate for duration of type 2diabetes, as the progressive decline of insulin levels with ageamong type 2 diabetics has been suggested to be protectivefor prostate cancer (1, 3, 6, 8). We found no evidence thatthe inverse association was strengthened among older men(Table 2). However, we did observe a slight, yet consistent,decrease in the relative risk when censoring the follow-up ofincident cases, by year, within the first 5 years of follow-up(Appendix Figure 1). We observed no significant differencein the association when stratified by body mass index (�25kg/m2: RR ¼ 0.78, 95% CI: 0.71, 0.86; <25 kg/m2: RR ¼0.86, 95% CI: 0.74, 1.00; Pinteraction ¼ 0.24). We also ob-served consistent effects by disease severity (Gleason score�7, n ¼ 3,853: RR ¼ 0.81, 95% CI: 0.73, 0.90; Gleasonscore >7, n ¼ 1,703: RR ¼ 0.76, 95% CI: 0.65, 0.89).

Association of PSA levels with type 2 diabetes statusand body mass index

In the subset of 2,874 men with PSA measurements, di-abetic men (n¼ 344) were found to have significantly lowerleast square geometric mean PSA levels than did nondia-betic men (n ¼ 2,530; 1.04 vs. 1.29 ng/mL; P < 0.001).Adjusting for body mass index had little effect on this as-sociation (1.07 vs. 1.28 ng/mL; P ¼ 0.003) (Table 3). Thisassociation was noted in all populations except NativeHawaiians and was statistically significant in EuropeanAmericans (0.62 vs. 1.21 ng/mL; P ¼ 0.003) and Latinos(0.99 vs. 1.27 ng/mL; P ¼ 0.02). Consistent with previousreports (11, 12), this report also shows an inverse relation

between body mass index and PSA levels (Table 3). Inethnicity-pooled analyses, compared with men with a bodymass index of <25 kg/m2, men with a body mass index of�30 had 13.8% lower mean PSA levels (P ¼ 0.009). Inmultivariate generalized linear models, adjusted for type 2diabetes status and age at blood draw, we estimated a 1-unitincrease in body mass index to be associated with a 1.6%decrease in mean PSA level (P < 0.001).

Association of type 2 diabetes status and educationwith PSA screening frequencies

In the sample of 46,970 men over 50 years of age withinformation on PSA screening, 48.2% reported a PSAscreening test prior to 1999. European-American men weremore likely to report having been screened (55.8%), whileNative Hawaiians (34.3%) were the least likely to have hadPSA testing (Pheterogeneity < 0.001). We observed a modest,yet highly statistically significant difference in age and ed-ucational level regarding standardized PSA screening fre-quencies between diabetics (44.7%) and nondiabetics(48.6%; P < 0.001) (Table 4). The lower PSA screeningfrequencies among diabetics were noted in all populationsexcept in African Americans and were statistically signifi-cant in Japanese Americans (P < 0.001) and Latinos (P ¼0.02). Men with higher educational levels were much morelikely to have had a PSA test than were men with �12 yearsof schooling (Table 4). We adjusted for educational level asa surrogate for PSA screening in the primary cohort analysesdiscussed above, yet it had little impact on the association.

Next, we examined the potential impact of detection biason the observed association between type 2 diabetes andprostate cancer. On the basis of the 3.9% difference inPSA screening frequencies observed between diabeticsand nondiabetics, we estimated that detection bias is likelyto account for only ~20% of the inverse association betweentype 2 diabetes and prostate cancer risk.

DISCUSSION

In this prospective analysis of 5 racial/ethnic populations,we found a highly significant association between type 2diabetes status and prostate cancer incidence, with diabeticshaving ~20% lower risk of developing prostate cancer. Thisinverse association was observed in all populations, with themagnitude of the effect being consistent with that of themajority of other studies conducted in men of Europeanancestry (1–8).

In this study, type 2 diabetes status was based on self-report, which may have led to misclassification. Previousstudies, however, have shown that self-reported responsesfor many common chronic diseases such as diabetes arereliable when compared with medical records (26–28).The analysis does not account for incident cases of type 2diabetes over the 8-year follow-up period. However, inci-dent cases of diabetes in the nondiabetes group would makethe 2 groups more similar and create an underestimation ofthe inverse association. Another limitation of our study isthat we cannot differentiate between cases of type 1 diabetes

940 Waters et al.

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

EuropeanAmericans

AfricanAmericans

NativeHawaiians

JapaneseAmericans

Latinos All

No. ofCases

RelativeRiska

95%ConfidenceInterval

No. ofCases

RelativeRiska


No. ofCases

RelativeRiska


No. ofCases

RelativeRiska


No. ofCases

RelativeRiska


No. ofCases

RelativeRiska


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

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.

Type2DiabetesandProstate

CancerRisk

941

Am

JEpidemiol2009;169:937–945

and type 2 diabetes. Although they have similar phenotypes,type 1 diabetes and type 2 diabetes have distinct mecha-nisms of pathogenesis and may have dissimilar associationswith prostate cancer incidence. However, we expect thisdifferential misclassification to be minimal as the preva-lence of type 1 diabetes is comparatively low in thesepopulations (22).

Diabetes and prostate cancer are both traditionally under-diagnosed diseases. In this study, undiagnosed type 2 dia-betes would result in prostate cancer incidence rates beingmore similar between the diabetic and nondiabetic groups.Undiagnosed cases of prostate cancer would result in lowerrates of prostate cancer among both diabetics and nondia-betics. As a result, we would expect these simultaneousevents of disease misclassification to counter the inverseassociation that we noted in this study toward the null. Itis also possible that men who do not receive frequent med-ical care would be underdiagnosed and misclassified forboth diseases. As a result, the underdiagnosis of prostatecancer and the lower risk of prostate cancer among themisclassified diabetics would also result in a bias towardthe null of the underlying association. Although we expectthat the underdiagnosis of these diseases is unable to explaintheir inverse association, future studies demanding regularblood glucose and PSA screening will be needed to quantifythe impact of this bias.

In this study, we also found that men with diabetes areless likely to report PSA screening than are men withoutdiabetes. These findings are contrary to those of a previousstudy that reported that men with diabetes are more likely toundergo screening for prostate cancer (3). PSA screeningfrequencies were lower among diabetics in all populationsexcept in African Americans. Education, which is a surro-gate for socioeconomic status and access to health care, wassignificantly associated with both PSA screening frequen-cies and diabetes status. However, further adjustment foreducation in the main cohort analyses did not change theresults. We estimated that the potential bias incurred bydifferential PSA screening (~4%) in diabetics and nondia-betics explained only ~20% of the protective effect of type2 diabetes on prostate cancer risk. In addition, if the asso-ciation between these conditions was influenced by detec-tion bias, then one would expect the inverse association todiminish among severe cases of prostate cancer, becausethey are likely to have been diagnosed without the use ofPSA screening. However, we observed only a minimalchange in the association between diabetes status and pros-tate cancer incidence when stratified by disease severity.Thus, detection bias associated with lower PSA levelsand/or lower PSA screening frequencies in diabetics is un-likely to explain the strong and highly significant inverseassociation between type 2 diabetes and prostate cancer inthis study.

Studies have suggested that the protective effect of di-abetes on prostate cancer incidence may be greater amongmen with longstanding type 2 diabetes (3, 6, 8). One theoryis that hyperinsulinemia, which is observed at onset, isassociated with increased levels of growth factors (e.g.,insulin-like growth factor-I) that may induce prostate cancerduring the first few years of type 2 diabetes. Subsequently,T

able

3.

GeometricMeanProstate-specificAntig

enLevelsbyEthnicity,DiabetesStatus,andBodyMassIndexin

theMultiethnicCohort(n

¼2,874),LosAngeles,California,andHawaii,

1993–2005

European

Americans

African

Americans

Native

Hawaiians

Japanese

Americans

Latinos

Alla

Mean,

ng/m

LNo.

PValue

Mean,

ng/m

LNo.

PValue

Mean,

ng/m

LNo.

PValue

Mean,

ng/m

LNo.

PValue

Mean,

ng/m

LNo.

PValue

Mean,

ng/m

LNo.

PValue

Pheterogeneity

Allmenb

1.19

446

1.50

916

1.14

313

1.22

485

1.26

714

1.26

2,874

<0.001

Diabetesstatusc

Yes

0.62

25

1.46

126

1.00

44

1.15

56

0.99

93

1.07

344

No

1.21

421

0.003

1.64

790

0.29

0.97

269

0.86

1.26

429

0.53

1.27

621

0.02

1.28

2,530

0.003

0.11

Bodymassindex(kg/m

2)d

<25

1.09

174

Referent

1.66

266

Referent

1.12

59

Referent

1.37

240

Referent

1.27

196

Referent

1.30

935

Referent

25–29.99

1.27

198

0.18

1.69

457

0.84

0.96

148

0.26

1.14

216

0.03

1.27

365

0.97

1.28

1,384

0.70

�30

1.12

74

0.84

1.38

193

0.07

0.92

106

0.18

1.07

29

0.18

1.09

153

0.16

1.12

555

0.009

0.49

Bodymassindex,%

changee

�0.6

0.65

�1.8

0.03

�1.9

0.06

�2.4

0.06

�1.6

0.06

�1.6

<0.001

0.93

aAdjustedforbodymassindex,diabetesstatus,ethnicity,andageatblooddraw.

bAdjustedfordiabetesstatus,bodymassindex,andageatblooddraw.AfricanAmericansversuseachethnic

group,P<

0.001.

cAdjustedforbodymassindexandageatblooddraw.

dAdjustedfordiabetesstatusandageatblooddraw.

ePercentchangein

geometric

prostate-specificantig

enlevels

with

increaseof1bodymassindexunitadjustedfordiabetesstatusandageatblooddraw.

942 Waters et al.

Am J Epidemiol 2009;169:937–945

prostate cancer rates would decrease in the later stages oftype 2 diabetes when insulin levels decrease and men be-come hypoinsulinemic. We do not have data on the date ofdiagnosis for type 2 diabetes, but we did analyze the asso-ciation between type 2 diabetes status and prostate cancerincidence by age at entry to the cohort as a surrogate forlongstanding type 2 diabetes. With both of these theories,one would expect the magnitude of the inverse associationbetween diabetes and prostate cancer to be greater amongolder men. However, we found no difference in the associ-ation in older men. We did, however, notice a modestincrease in the magnitude of the inverse association whenremoving incident cases within the first 5 years of follow-up,which supports the hypothesis that men with newly diag-nosed diabetes may have an increased risk of prostatecancer.

Most, but not all, studies have shown that, on average,men with diabetes have lower PSA levels than do thosewithout diabetes (11, 12, 29). In our multiethnic sample,PSA levels were lower in diabetic men. However, what thisindicates is not clear. Lower PSA levels in diabetics maysignal a lower prevalence of prostate cancer and an indica-tion of a biologic effect of type 2 diabetes status on prostategrowth and development. At the same time, the effectof diabetes status on PSA levels could result in decreasedfollow-up for prostate cancer diagnosis among diabetics,which may partially account for the inverse relation betweentype 2 diabetes and prostate cancer risk. Additional workwill be needed to understand whether type 2 diabetes statusinfluences the accuracy of PSA screening or directly con-tributes to prostate cancer risk. Consistent with previousstudies (11, 12, 30), our analysis also suggests an inverserelation between body mass index and PSA levels. Furtherstudies of this association are necessary to determine ifobese men should have lower PSA thresholds to indicatefurther work-up for prostate cancer.

Only a small number of studies have investigated the re-lation between type 2 diabetes and prostate cancer risk innon-European populations (2, 31–33). Most of these studieshave observed nonsignificant inverse associations; however,small sample sizes have limited interpretation of the find-ings. Our results, from a population-based prospective studyof over 5,900 prostate cancer cases from 5 racial/ethnicpopulations, provide strong support for the pan-ethnic na-ture of the association between these common diseases.

Recently, a common variant in the hepatocyte nuclearfactor-1 b gene (HNF1b) was found to be associated withan increased risk of prostate cancer. This same variant wasalso found to be associated with a decreased risk of type 2diabetes (34). Common genetic variation in another gene,JAZF1, has also been associated with risks of both prostatecancer and type 2 diabetes (35, 36). These findings, alongwith findings from other studies that have shown that di-abetes is inversely associated with a family history of pros-tate cancer (5, 37), which we also noted, point to bothshared genetic risk and common molecular and/or meta-bolic pathways that are important in the etiology of thesediseases.

In summary, in this large, multiethnic prospective study,we observed consistent inverse associations between type 2T

able

4.

Prostate-specificAntig

enScreeningFrequenciesbyLevelofE

ducatio

nandDiabetesStatusintheMultiethnicCohort(n

¼46,970),LosAngeles,C

alifornia,a

ndHawaii,1993–2005

European

Americans

African

Americans

Native

Hawaiians

Japanese

Americans

Latinos

All

No.a

%b

PValue

No.

%PValue

No.

%PValue

No.

%PValue

No.

%PValue

No.

%PValue

Pheterogeneity

Allmenc

12,035

55.8

4,946

52.1

2,995

34.3

16,021

46.3

10,973

44.7

46,970

48.2

<0.001

Educa

tionallevelc,d

�12years

2,747

44.5

Referent

1,954

46.3

Referent

1,549

30.1

Referent

6,247

37.1

Referent

6,856

41.1

Referent

19,353

40.0

Referent

Collegeorvoca

tional

3,347

53.6

<0.001

1,819

53.3

<0.001

874

35.6

0.001

4,846

46.0

<0.001

2,702

50.8

<0.001

13,588

49.1

<0.001

Collegegraduate

5,941

63.4

<0.001

1,173

59.8

<0.001

572

45.7

<0.001

4,928

56.4

<0.001

1,415

50.5

<0.001

14,029

58.6

<0.001

<0.001

Diabetesstatuse,f

Yes

796

51.4

742

52.5

487

35.1

1,947

42.1

1,793

44.3

5,765

44.7

No

11,239

53.0

0.46

4,204

52.1

0.84

2,508

36.4

0.37

14,074

46.0

<0.001

9,180

47.4

0.02

41,205

48.6

<0.001

0.45

aNo.ofmenwith

prostate-specificantig

enscreeningdata.

bPercentscreened.

cPercentagesare

agestandardized(5-yearagegroups)

tothetotalpopulatio

nincludedin

thestudy.

dPvaluesare

from

alogistic

regressionmodelandare

adjustedforage,educatio

nallevel,andrace/ethnicity

(pooledanalysis).

ePercentagesare

age(5-yearagegroups)

andeducatio

nstandardizedto

thetotalpopulatio

nin

thestudy.

fPvaluesare

from

alogistic

regressionmodelandare

adjustedforageandrace/ethnicity

(pooledanalysis).


Am J Epidemiol 2009;169:937–945

diabetes and prostate cancer risk across multiple racial/ethnic populations. These findings provide strong supportfor the hypothesis that type 2 diabetes is a protective factorfor prostate cancer. We also confirmed results from previousstudies, showing that PSA levels are decreased in diabeticmen. Our findings that diabetic men are less likely to bescreened for prostate cancer could not account for theseresults. Future studies aimed at determining the biologiclink between diabetes and prostate cancer are warrantedand should focus on common environmental and geneticfactors that are shared across populations.

ACKNOWLEDGMENTS

Author affiliations: Department of Preventive Medicine,Keck School of Medicine, University of Southern California/Norris Comprehensive Cancer Center, Los Angeles, Cali-fornia (Kevin M. Waters, Brian E. Henderson, Daniel O.Stram, Peggy Wan, Christopher A. Haiman); and Epidemi-ology Program, Cancer Research Center, University ofHawaii, Honolulu, Hawaii (Laurence N. Kolonel).

This work was supported by the National Cancer Institute(grant CA54281).


REFERENCES

1. Kasper JS, Giovannucci E. A meta-analysis of diabetes mel-litus and the risk of prostate cancer. Cancer Epidemiol Bio-markers Prev. 2006;15(11):2056–2062.

2. Calton BA, Chang SC, Wright ME, et al. History of diabetesmellitus and subsequent prostate cancer risk in the NIH-AARPDiet and Health Study. Cancer Causes Control. 2007;18(5):493–503.

3. Giovannucci E, Rimm EB, Stampfer MJ, et al. Diabetes mel-litus and risk of prostate cancer. Cancer Causes Control.1998;9(1):3–9.

4. Gong Z, Neuhouser ML, Goodman PJ, et al. Obesity, diabetes,and risk of prostate cancer: results from the prostate cancerprevention trial. Cancer Epidemiol Biomarkers Prev. 2006;15(10):1977–1983.

5. Velicer CM, Dublin S, White E. Diabetes and the risk ofprostate cancer: the role of diabetes treatment and complica-tions. Prostate Cancer Prostatic Dis. 2007;10(1):46–51.

6. Rodriguez C, Patel AV, Mondul AM, et al. Diabetes and risk ofprostate cancer in a prospective cohort of US men. Am JEpidemiol. 2005;161(2):147–152.

7. Pierce BL, Plymate S, Ostrander EA, et al. Diabetes mellitusand prostate cancer risk. Prostate. 2008;68(10):1126–1132.

8. Tavani A, Gallus S, Bosetti C, et al. Diabetes and the risk ofprostate cancer. Eur J Cancer Prev. 2002;11(2):125–128.

9. Giovannucci E, Rimm EB, Stampfer MJ, et al. Height, bodyweight, and risk of prostate cancer. Cancer Epidemiol Bio-markers Prev. 1997;6(8):557–563.

10. Huxley R, James WP, Barzi F, et al. Ethnic comparisons of thecross-sectional relationships between measures of body sizewith diabetes and hypertension. Obes Rev. 2008;9(suppl 1):53–61.

11. Fukui M, Tanaka M, Kadano M, et al. Serum prostate-specificantigen level in men with type 2 diabetes. Diabetes Care.2008;31(5):930–931.

12. Werny DM, Saraiya M, Gregg EW. Prostate-specific antigenvalues in diabetic and nondiabetic US men, 2001–2002. Am JEpidemiol. 2006;164(10):978–983.

13. Ross LE, Berkowitz Z, Ekwueme DU. Use of the prostate-specific antigen test among U.S. men: findings from the 2005National Health Interview Survey. Cancer Epidemiol Bio-markers Prev. 2008;17(3):636–644.

14. Abate N, Chandalia M. The impact of ethnicity on type 2diabetes. 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 InterviewSurvey: 2000–2003. Public Health Rep. 2007;122(5):616–625.

16. Mau MK, Glanz K, Severino R, et al. Mediators of lifestylebehavior change in Native Hawaiians: initial findings from theNative Hawaiian Diabetes Intervention Program. DiabetesCare. 2001;24(10):1770–1775.

17. Aluli N. Prevalence of obesity in a Native Hawaiian popula-tion. Am J Clin Nutr. 1991;53(6 suppl):1556S–1560S.

18. Brawley OW, Jani AB, Master V. Prostate cancer and race.Curr Probl Cancer. 2007;31(3):211–225.

19. Jemal A, Siegel R, Ward E, et al. Cancer statistics, 2008. CACancer J Clin. 2008;58(2):71–96.

20. Kolonel LN, Altshuler D, Henderson BE. The MultiethnicCohort Study: exploring genes, lifestyle, and cancer risk. NatRev Cancer. 2004;4(7):519–527.

21. Kolonel LN, Henderson BE, Hankin JH, et al. A multiethniccohort in Hawaii and Los Angeles: baseline characteristics.Am J Epidemiol. 2000;151(4):346–357.

22. National Center for Chronic Disease Prevention and HealthPromotion. National diabetes fact sheet. Atlanta, GA: Centersfor Disease Control and Prevention; 2005. (http://www.cdc.gov/diabetes/pubs/general.htm#what).

23. Cheng I, Yu MC, Koh WP, et al. Comparison of prostate-specific antigen and hormone levels among men in Singaporeand the United States. Cancer Epidemiol Biomarkers Prev.2005;14(7):1692–1695.

24. Smith RA, Cokkinides V, Eyre HJ, et al. American CancerSociety guidelines for the early detection of cancer, 2003. CACancer J Clin. 2003;53(1):27–43.

25. Collin SM, Martin RM, Metcalfe C, et al. Prostate-cancermortality in the USA and UK in 1975–2004: an ecologicalstudy. Lancet Oncol. 2008;9(5):445–452.

26. Midthjell K, Holmen J, Biøndal A, et al. Is questionnaire in-formation valid in the study of a chronic disease such as di-abetes? The Nord-Trøndelag Diabetes Study. J EpidemiolCommunity Health. 1992;46(5):537–542.

27. Okura Y, Urban LH, Mahoney DW, et al. Agreement betweenself-report questionnaires and medical record data was sub-stantial for diabetes, hypertension, myocardial infarction andstroke but not heart failure. J Clin Epidemiol. 2004;57(10):1096–1103.

28. Walitt BT, Constantinescu F, Katz JD, et al. Validation of self-report of rheumatoid arthritis and systemic lupus erythematosus:theWomen’sHealth Initiative. JRheumatol. 2008;35(5):811–818.

29. Chan JM, Latini DM, Cowan J, et al. History of diabetes,clinical features of prostate cancer, and prostate cancerrecurrence-data for CaPSURETM (United States). CancerCauses Control. 2005;16(7):789–797.

30. Fowke JH, Signorello LB, Chang SS, et al. Effects of obesityand height on prostate-specific antigen (PSA) and percentageof free PSA levels among African-American and Caucasianmen. Cancer. 2006;107(10):2361–2367.

944 Waters et al.

Am J Epidemiol 2009;169:937–945

http://www.cdc.gov/diabetes/pubs/general.htm&num;what

http://www.cdc.gov/diabetes/pubs/general.htm&num;what

31. Coker AL, Sanderson M, Zheng W, et al. Diabetes mellitusand prostate cancer risk among older men: population basedcase-control study. Br J Cancer. 2004;90(11):2171–2175.

32. Mishina T, Watanabe H, Araki H, et al. Epidemiological studyof prostatic cancer by matched-pair analysis. Prostate. 1985;6(4):423–436.

33. Rosenberg DJ, Neugut AI, Ahsan H, et al. Diabetes mellitusand the risk of prostate cancer. Cancer Invest. 2002;20(2):157–165.

34. Gudmundsson J, Sulem P, Steinthorsdottir V, et al. Two var-iants on chromosome 17 confer prostate cancer risk, and theone in TCF2 protects against type 2 diabetes. Nat Genet.2007;39(8):977–983.

35. Thomas G, Jacobs KB, Yeager M, et al. Multiple loci identifiedin a genome-wide association study of prostate cancer. NatGenet. 2008;40(3):310–315.

36. Zeggini E, Scott LJ, Saxena R, et al. Meta-analysis of genome-wide association data and large-scale replication identifiesadditional susceptibility loci for type 2 diabetes. Nat Genet.2008;40(5):638–645.

37. Meyer P, Zuern C, Hermanns N, et al. The association betweenpaternal prostate cancer and type 2 diabetes [electronic article].J Carcinog. 2007;6:14.

1

0 2 3 4 51

Cox Analysis, years after cohort entry

Re

lative

Ris

k

0.5

0.6

0.7

0.8

0.9

1.1

1.2

Appendix Figure 1. Ethnically pooled association between type 2diabetes and prostate cancer risk by analysis start point, Los Angeles,California, and Hawaii, 1993–2005. There were 5,941, 5,441, 4,933,4,456, 3,941, and 3,373 cases at 0, 1, 2, 3, 4, and 5 years after cohortentry, respectively. Bars, 95% confidence interval.


Am J Epidemiol 2009;169:937–945

American Journal of Epidemiologyª 2009 The AuthorsThis is an Open Access article distributed under the terms of the Creative Commons Attribution Non-CommercialLicense (http://creativecommons.org/licenses/by-nc/2.0/uk/) which permits unrestricted non-commercial use,distribution, and reproduction in any medium, provided the original work is properly cited.

Vol. 169, No. 8

DOI: 10.1093/aje/kwn413



Modification of the Effect of Vitamin E Supplementation on the Mortality of MaleSmokers by Age and Dietary Vitamin C

Harri Hemila 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 Finnishmale cigarette smokers, finding that vitamin E supplementation had no overall effect on mortality. The authors ofthis paper found that the effect of vitamin E on respiratory infections in ATBC Study participants was modified byage, 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 deathsoccurred. Age and dietary vitamin C intake had a second-order interaction with vitamin E supplementation of50 mg/day. Among participants with a dietary vitamin C intake above the median of 90 mg/day, vitamin E increasedmortality among those aged 50–62 years by 19% (95% confidence interval: 5, 35), whereas vitamin E decreasedmortality among those aged 66–69 years by 41% (95% CI: �56, �21). Vitamin E had no effect on participants whohad a dietary vitamin C intake below the median. Smoking quantity did not modify the effect of vitamin E. This studyprovides strong evidence that the effect of vitamin E supplementation on mortality varies between differentpopulation groups. Further study is needed to confirm this heterogeneity.

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

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

Taking vitamin E supplements is a common practice,particularly among older people. In the United States, abouta quarter of adults aged 60 years or older take supplementscontaining vitamin E (1). Such a common habit makes thehealth effects of this practice an important public healthissue: does vitamin E supplementation improve health ornot?

The rationale behind taking the lipid-soluble antioxidantvitamin E is to protect against oxidative stress, which con-tributes to the aging processes and may affect longevity(2, 3). However, 3 meta-analyses of randomized trials foundthat vitamin E supplementation did not reduce mortality,implying that vitamin E does not lead to universal systemicbenefits against the processes that lead to chronic disease(4–6).

Vitamin E is a major lipid-soluble antioxidant, whereasvitamin C is a major water-soluble antioxidant. They inter-act in vitro and in vivo (7–11). Smoking increases the

plasma a-tocopherol disappearance rate, which is normal-ized by vitamin C supplementation (10). Thus, since smok-ing seems to modify the interaction of these 2 antioxidants,intake of vitamin C is particularly important when examin-ing the effect of vitamin E on smokers.

The Alpha-Tocopherol, Beta-Carotene Cancer Prevention(ATBC) Study was a randomized, double-blind, placebo-controlled trial that examined the effects of vitamin E andb-carotene on lung cancer in male smokers (12, 13). Pre-viously, we found significant heterogeneity in the effects ofvitamin E supplementation in the ATBC Study; vitamin Eincreased the risk of tuberculosis for heavy smokers withhigh dietary vitamin C intake but had no effect on otherparticipants (14). The effect of vitamin E on common coldand pneumonia risk was modified by age and smoking (11,15–19).

In the ATBC Study, vitamin E supplementation had nooverall effect on mortality (12). In this study, we tested the

Correspondence to Dr. Harri Hemila, Department of Public Health, University of Helsinki, POB 41, Helsinki, FIN-00014, Finland (e-mail:

[email protected]).

946 Am J Epidemiol 2009;169:946–953

http://creativecommons.org/licenses/by-nc/2.0/uk/

hypothesis that the variables that modify the effect of vita-min E supplementation on respiratory infections would alsomodify the effect of vitamin E on mortality. We did notpresume that changes in the incidence of respiratory infec-tions would significantly affect overall mortality, but thatchanges in respiratory infections may reflect nonspecificsystemic effects of vitamin E that might also affectmortality.


Participants

The design and methods of the ATBC Study examiningthe effects of vitamin E (dl-a-tocopheryl acetate, 50 mg/day)and b-carotene (20 mg/day) on the incidence of lung cancerand other cancers have been described earlier (12, 13).The ATBC Study is registered at the website www.ClinicalTrials.gov under the identifier NCT00342992.

In brief, male participants aged 50–69 years had to smoke5 or more cigarettes per day at entry to be eligible, and thoseenrolled in the trial (N ¼ 29,133) were randomized to 1 of 4intervention arms and were administered placebo, vitaminE, b-carotene, or vitamin E þ b-carotene; a 2 3 2 factorialdesign was used. Compliance with supplementation washigh: some 90% of the participants took more than 90%of their prescribed capsules during their active participationin the trial; there were no differences in capsule consump-tion among the intervention groups (13). Supplementationincreased the serum level of a-tocopherol by 50% comparedwith baseline levels (12, 13). The intervention continueduntil April 30, 1993. The trial was approved by the institu-tional review boards, and all participants gave written in-formed consent.

Baseline characteristics

Before randomization at baseline, the men completedquestionnaires on their medical and smoking histories andgeneral background characteristics, and their weight wasmeasured. A detailed dietary history questionnaire provideddata regarding vitamin C, vitamin E, fruit, vegetable, andberry consumption (20, 21). The validity of the dietary his-tory questionnaire was assessed by comparing it with thefood consumption records of 190 participants for 12 sepa-rate 2-day periods distributed evenly over 6 months. Clas-sified by the dietary history questionnaire, 74%–76% ofparticipants categorized by food consumption records werein the same vitamin C intake quintile or in the within-one-quintile category (20). Dietary data were not available for2,022 of the 29,133 participants.

Outcome and follow-up time

Deaths were identified by using the National Death Reg-istry, as previously described (12). In the database we ana-lyzed was one death additional to those in the 1994 report.

Follow-up time for each participant began from the day ofrandomization and continued until death or the end of thetrial. The median follow-up time for the participants in the

present analysis was 6.1 years, and there was a total of169,731 person-years of observation.

Statistical models

We estimated the effect of vitamin E supplementation onmortality through proportional hazards regression models.We calculated the risk ratio and the 95% confidence intervalof the risk ratio by using the PROC PHREG procedure inSAS software (release 8.2; SAS Institute, Inc., Cary, NorthCarolina). The 2 3 2 factorial design of the trial permittedassessment of the effect of vitamin E independent ofb-carotene after confirming no statistical interaction be-tween the agents. Thus, we compared the trial participantsadministered vitamin E with those not receiving vitamin E(no-vitamin-E participants). We did not analyze the effectsof b-carotene in this study. Regarding supplementation, wecarried out the analyses following the intention-to-treatprinciple. Because deaths were identified in the NationalDeath Registry, which registers all deaths occurring inFinland, loss-to-follow-up is insignificant.

To test the statistical significance of interaction betweenvitamin E supplementation and potential modifying factors,we first added the supplementation andmodifying factor to theregressionmodel. The statistical significance of the interactionwas thereafter calculated from the change in�23 log(likeli-hood) when the interaction term of vitamin E supplementationand the modifying factor was added to the model.

In our subgroup analyses, we split dietary vitamin E and Clevels, and the residual of fruit, vegetable, and berry consump-tion, at rounded levels close to the medians. Dietary vitamin Cwas also used as a continuous variable because interactionwith a continuous variable refutes the possibility that dichot-omizing dietary vitamin C intake might cause a spurious in-teraction; to decrease the skewness of the distribution, thestatistical model included the logarithm of dietary vitamin Cintake.

Nelson-Aalen cumulative hazard functions were con-structed by using the STATA sts program (release 9.1; StataCorporation, College Station, Texas). Two-tailed P valueswere used.

Examination of the specificity of effect modification byvitamin C

The major sources of vitamin C in the diet of study par-ticipants were fruit, vegetables, and berries, on average 58%of dietary vitamin C originating from these foods. Total in-take of fruit, vegetables, and berries was strongly correlatedwith the calculated vitamin C intake (r ¼ 0.88). Thus, it ispossible that an association with dietary vitamin C is a sta-tistical artifact reflecting other substances in these foods orthe lifestyle related to eating these foods. To examine thepossible role of dietary compounds other than vitamin C inthese foods, we calculated the residual of fruit, vegetables,and berries intake by using linear regression to model fruit,vegetables, and berries as a function of dietary vitamin C, aspreviously (22). As designed, the residual of fruit, vegeta-bles, and berries intake has no correlation with dietary vita-min C. We assumed that any other putative compound that

Vitamin E Supplementation and Heterogeneity in Mortality 947

Am J Epidemiol 2009;169:946–953

www.ClinicalTrials.gov

www.ClinicalTrials.gov

might interact with vitamin E supplementation has no perfectcorrelation with vitamin C and, therefore, that variation in theother compound remains as variation in the residual of fruit,vegetables, and berries intake, which was split at 0 g/day,close to the median. High residual of fruit, vegetables, andberries intake (above zero) indicates that the participant witha given vitamin C level consumes more than the averagequantity of fruit, vegetables, and berries, whereas low resid-ual of fruit, vegetables, and berries intake (below zero) in-dicates less-than-average intake of these food classes.

RESULTS

During the 169,731 person-years of follow-up of theATBC Study participants, 3,571 deaths occurred, equivalentto 21.0 deaths per 1,000 person-years. The deaths wereequally divided between the vitamin E and no-vitamin-Egroups: 1,801 vs. 1,770, corresponding to a risk ratio of1.02 (95% confidence interval (CI): 0.95, 1.09).

Dietary vitamin C intake did not modify the effect ofvitamin E supplementation. For participants with a vitaminC intake of less than 90 mg/day, the effect of vitamin E onmortality was a risk ratio of 1.00 (95% CI: 0.92, 1.11); forthose with a higher vitamin C intake, the effect was a riskratio of 1.04 (95% CI: 0.94, 1.16). Age did not modify theeffect of vitamin E (P ¼ 0.06; interaction between vitamin Eand age as a continuous variable).

We found that vitamin E supplementation had a second-order interaction with dietary vitamin C and age (Table 1).Vitamin E did not affect mortality for participants with lowvitamin C intake. However, for participants with high dietaryvitamin C intake, the effect of vitamin E depended on age, sothat it increased mortality in the young (aged <63 years)participants by 19% but reduced mortality in the old (aged�66 years) participants by 41% (Table 1, Figures 1 and 2).

Division of participants into the age categories shown inTable 1 was based on the inspection of data, but the findingswere not sensitive to the cutpoints of age. For young partic-ipants with high vitamin C intake, the point estimate of thevitamin E effect was very similar in narrower age catego-ries: risk ratio ¼ 1.17 (95% CI: 0.93, 1.48; 297 deaths), riskratio ¼ 1.25 (95% CI: 1.02, 1.53; 371 deaths), and riskratio ¼ 1.13 (95% CI: 0.92, 1.40; 353 deaths) in the groupsaged 50–54, 55–58, and 59–62 years, respectively, whichjustified combining these age groups. In participants aged66–69 years with high dietary vitamin C intake, vitamin Ehad the same effect in 2-year categories: risk ratio ¼ 0.58(95% CI: 0.39, 0.86; 113 deaths) and risk ratio ¼ 0.58 (95%CI: 0.37, 0.91; 82 deaths) in the groups aged 66–67 and68–69 years, respectively.

There seemed to be an approximately 3-year lag periodbefore vitamin E started to increase the mortality of partic-ipants aged 50–62yearswith highvitaminC intake (Figure 1).During the first 3 years, vitamin E had no effect on mor-tality, but thereafter it increased mortality by 38%. Inclu-sion of the lag period in the regression model led tosignificant improvement in the model. The survival curvesfor participants aged 66–69 years suggested a 2-year lagperiod (Figure 2).

The main food sources of vitamin C are fruit, vegetables,and berries. Thus, modification of the vitamin E effect bydietary vitamin C might be explained by high levels of othersubstances in these foods. Residual fruit, vegetables, andberry intake (refer to the Materials and Methods section)did not modify the effect of vitamin E in participants aged50–62 years (Table 2), indicating that other substances inthese foods do not explain the effect modification by vitamin C.The vitamin E effect was modified by vitamin C as a con-tinuous variable, indicating that the cutpoint for dichotomi-zation was not crucial to the finding. Dietary vitamin Eintake and b-carotene supplementation did not modify theeffect of vitamin E supplementation. In the participants aged

Table 1. Effect of Vitamin E Supplementation on Mortality by Age

and Dietary Vitamin C Intake, Alpha-Tocopherol, Beta-Carotene

Cancer Prevention Study, 1985–1993

Age at Baseline

Vitamin C Test forVitamin CInteraction(P Value)

<90 mg/day(n 5 13,567)a

‡90 mg/day(n 5 13,544)a

50–62 years(n ¼ 22,413)

Risk ratiob 1.00 1.19 0.048

95% confidenceinterval

0.90, 1.13 1.05, 1.35

Deathsc 614/616 552/469

63–65 years(n ¼ 2,761)

Risk ratiob 0.95 0.89 0.7


0.75, 1.20 0.68, 1.17

Deathsc 142/139 106/110

66–69 years(n ¼ 1,937)

Risk ratiob 1.07 0.59 0.002


0.84, 1.36 0.44, 0.79

Deathsc 139/137 71/124

Test for interaction;age as a continuousvariable (P value)d

0.4 0.0003

a Information on dietary vitamin C intake was missing for 2,022

participants, with 177 deaths of vitamin-E and 175 deaths of no-

vitamin-E participants.b Proportional hazards regression model comparing participants

who received vitamin E with those who did not.c Number of deaths of vitamin-E participants/number of deaths of

no-vitamin-E participants.d The second-order interaction term between vitamin E supplemen-

tation, dietary vitamin C, and age improved the regression model by

v2(1 df) ¼ 10.1, P ¼ 0.0015. The uniformity of the vitamin E effect

was also tested by adding a dummy variable for vitamin E effect in

5 groups of the table, allowing each of the 6 groups its own vitamin E

supplementation effect. The regression model was improved by

v2(5 df) ¼ 22.2, P ¼ 0.0005 compared with the model with a uniform

vitamin E effect. Adding the vitamin E effect to only those groups aged

50–62 and 66–69 years with high vitamin C intake led to similar im-

provement in the regression model, by v2(2 df) ¼ 21.0 compared with

the model with a uniform vitamin E effect.

948 Hemila and Kaprio

Am J Epidemiol 2009;169:946–953

66–69 years, residual fruit, vegetables, and berry intake;dietary vitamin E intake; and b-carotene supplementationdid not modify the effect of vitamin E (Table 3).

Smokinghad amarginally significantmodification effect onthe participants aged 66–69 years with high vitamin C intake(P ¼ 0.051). The decrease in mortality with vitamin Esupplementation was more evident for those who smokedless than a pack of cigarettes per day (risk ratio (RR) ¼0.44, 95% CI: 0.28, 0.68; 89 deaths) compared witha pack or more (RR ¼ 0.78, 95% CI: 0.53, 1.15; 106deaths). For the participants aged 50–62 years with highvitamin C intake, the effect was more apparent for thosewho smoked a pack of cigarettes or more per day (RR ¼1.22, 95% CI: 1.05, 1.42; 689 deaths) than for those whosmoked less (RR ¼ 1.12, 95% CI: 0.90, 1.39; 332 deaths),but the confidence intervals overlapped broadly and thedifference was not statistically significant (P ¼ 0.5 inheterogeneity test).

Of the participants aged 66–69 years with high vitamin Cintake, 16.7% of the vitamin-E participants died duringfollow-up, whereas 27.7% of the no-vitamin-E participantsdied. Thus, based on the baseline cohort, the number neededto treat to prevent one death during follow-up was 9.1 in this

subgroup. On the other hand, the cumulative hazard esti-mates over the 8-year follow-up (Figure 2) indicate mor-tality of 28% in the vitamin-E group and 49% in theno-vitamin-E group, corresponding to the number neededto treat of 4.8 over an 8-year supplementation period.

DISCUSSION

Vitamin E supplementation had no overall effect on mor-tality among the ATBC Study participants, yet we foundstrong evidence of heterogeneity between subgroups.Vitamin E increased, decreased, or had no effect on mortal-ity depending on age and vitamin C intake. The numericalestimates calculated for the subgroups are less essential thanthe strong evidence of heterogeneity. When the effect ofvitamin E depends on the characteristics of people, it seemsobvious that the estimates of intervention effect obtained

0.00

0.05

0.10

0.15

Cum

ulat

ive

Haz

ard

of D

eath

0 2 4 6 8

Years of Follow-up

No Vitamin E

Vitamin E

Figure 1. Effect of vitamin E supplementation on mortality amongparticipants aged 50–62 years with a dietary vitamin C intake of>90 mg/day (n ¼ 11,448 with 1,021 deaths), Alpha-Tocopherol,Beta-Carotene Cancer Prevention Study, 1985–1993. Nelson-Aalencumulative hazard functions for the vitamin-E and no-vitamin-Egroups are shown. Each step indicates 1 death. For the differencebetween the 2 groups, log-rank-test P ¼ 0.006. The number of par-ticipants with follow-up time of�7 years was 2,316; the curves are cutat 7.8 years because the number of participants declined abruptlythereafter. The possibility of a lag period was examined by addinga different risk ratio term for vitamin E effect starting at variable timepoints. The best improvement in the regression model was achievedby adding the second vitamin E effect starting at 3.3 person-years,which improved the regression model by v2(1 df) ¼ 7.1, P ¼ 0.007.This model gives risk ratios of 0.99 (95% confidence interval: 0.82,1.19) during the first 3.3 years and 1.38 (95% confidence interval:1.17, 1.63) thereafter.

0.00

0.10

0.20

0.30

0.40

0.50

Cum

ulat

ive

Haz

ard

of D

eath

0 2 4 6 8

Years of Follow-up

No Vitamin E

Vitamin E

Figure 2. Effect of vitamin E supplementation on mortality amongparticipants aged 66–69 years with a dietary vitamin C intake of >90mg/day (n ¼ 872 with 195 deaths), Alpha-Tocopherol, Beta-CaroteneCancer Prevention Study, 1985–1993. Nelson-Aalen cumulative haz-ard functions for the vitamin-E and no-vitamin-E groups are shown.Each step indicates 1 death. For the difference between the 2 groups,log-rank-test P ¼ 0.0003. The number of participants with follow-uptime of�7 years was 128; the curves are cut at 7.8 years because thenumber of participants declined abruptly thereafter. The possibility ofa lag period was examined by adding 2 different risk ratio terms forvitamin E effect starting at variable time points because the 2 curvesdiverge at the initiation of supplementation and at about 2 years. Thebest improvement in the regression model was achieved by addingthe second vitamin E effect starting at 0.3 person-years and the thirdrisk ratio starting from 1.9 years, which improved the regressionmodelby v2(2 df) ¼ 5.2, P ¼ 0.073. This model gives risk ratios of 0.15(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 thefirst 0.3 years of follow-up, there were 5 deaths in the b-carotene arm,2 deaths in the placebo arm, 1 death in the vitamin E arm, and nodeaths in the vitamin E þ b-carotene arm.


Am J Epidemiol 2009;169:946–953

from one study cannot be confidently generalized to otherpopulation groups.

Although we divided the participants into several sub-groups, the multiple comparison problem did not seem to bea concern in our study. First, the subgroup heterogeneity in the6 groups defined by age and dietary vitamin C intake wassignificant even when we allowed each of the 6 subgroups tohave its own vitamin E effect (Table 1). Second, our currentsubgroup analyses tested our earlier subgroup findings for res-piratory infection outcomes in the ATBC Study (11, 14–19).

Our findings have several important implications. Sub-group analysis has been discouraged because it can lead tofalse-positive findings due to the multiple comparison prob-lem (23–25). It has even been argued that ‘‘believing thata treatment effect exists in one stratum of patients, eventhough no overall significant treatment effect exists, isa common error’’ (24, p. 15). Furthermore, large trials areset up after years of deliberation by experts, and it is as-sumed that all relevant knowledge is therefore incorporatedinto the study plans (25). However, biology is complex, andit seems unlikely that all important biologic knowledgecould ever be taken into account properly when setting upa pragmatic controlled trial. A single estimate of treatmenteffect calculated for a large population can be misleading if

it is thought to be valid for all participants, as shown in ourwork. The overall estimate for all ATBC Study participants,risk ratio ¼ 1.02, is inconsistent with our subgroup findingsfor nearly half of all participants, that is, the young and oldwith high dietary vitamin C intake.

Our current study and earlier subgroup analyses of theATBC Study suggest that subgroup analyses of large trialdatabases should be encouraged even if there is no overalleffect in the study population, keeping in mind the possibil-ity of spurious findings from multiple testing. Given thelong-term commitment of participants and the resourcesinvested, it might even be considered a moral imperativeto analyze the large trial databases as extensively as possiblerather than simply to calculate an overall effect. Evidently,subgroup findings should be considered cautiously, and theinterpretation of P values must be related to the number ofsubgroup analyses being carried out. Nevertheless, our sub-group findings in the ATBC Study point out the need forfurther research on vitamin E, whereas the overall estimateof effect suggests that no further studies would have beenworthwhile. We concur with Feinstein’s concern (26) thatanti-subgroup doctrines have become so entrenched thatthey often hamper investigation of socially and biologicallysound subgroups with public health relevance.

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

SubgroupNo. of

Participants

Vitamin E No Vitamin ERiskRatioa


Test ofInteraction(P Value)

No. ofDeaths

RatebNo. ofDeaths

Rateb

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.


Am J Epidemiol 2009;169:946–953

Three meta-analyses of controlled trials found no effect ofvitamin E supplementation on mortality (4–6). Combiningseveral large trials leads to estimates with narrow confidenceintervals; however, pooling is based on the assumption of thesame universal effect in all populations of the combinedstudies. Our strong evidence of heterogeneity in the ATBCStudy refutes the assumption of a uniform effect and castsdoubt on the validity of the pooled estimates of the meta-analyses. Although the pooled estimates reject the proposalthat vitamin E supplementation would be beneficial for wide-ranging population groups, our study suggests that importantsubgroups being harmed or benefiting from vitamin Emay behidden in the misleadingly narrow confidence intervals.

The recently published Physicians’ Health Study II foundno overall effect of vitamin E and C supplementation onmortality (27). However, the lack of average effect doesnot conflict with our findings of heterogeneity because vi-tamin E had no overall effect in the ATBC Study either. Thenumber of deaths in the ATBC Study (n ¼ 3,571) was twicethat in the Physicians’ Health Study II (n ¼ 1,661); conse-quently, there is more statistical power to analyze potentialsubgroup differences in the ATBC Study. Furthermore,ATBC Study participants were recruited from the generalpopulation, so there is greater potential for analyzing het-

erogeneity compared with the Physicians’ Health Study IIfocused on a single upper-class profession.

Our strongest findings concern vitamin E supplementationbecause the division of participants into vitamin-E and no-vitamin-E groups was random. The evidence of heterogeneityof the vitamin E effect is strong irrespective of the specificityof vitamin C as the modifier. Nevertheless, by using the re-sidual of fruit, vegetables, and berries, which has no corre-lation with dietary vitamin C, we were able to show thatmodification of the vitamin E effect is not explained by othersubstances in these foods, suggesting that vitamin C is thespecific substance explaining the modification. Furthermore,the subgroup divisions in our current study were based on theeffects of vitamin E on respiratory infections, and vitamin Chas also affected respiratory infections in certain controlledtrials, showing that its physiologic effects are not limited topreventing scurvy (28, 29). Finally, the interaction betweenvitamins E and C is well established (7–11). These argu-ments support the possibility that vitamin C specificallycauses the modification of the vitamin E effect, although ourvitamin C intake levels were based on observational data.

All ATBC Study participants were smokers, and no directconclusions can be drawn for nonsmokers. The benefit ofvitamin E supplementation on respiratory infections was

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

SubgroupNo. of

Participants

Vitamin E No Vitamin ERiskRatioa


Test ofInteraction(P Value)

No. ofDeaths

RatebNo. ofDeaths

Rateb

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.


Am J Epidemiol 2009;169:946–953

previously more evident in ATBC Study participants whosmoked less, and the harm was more apparent in those whosmoked more (14, 15, 18, 19). We saw similar trends in theeffect of vitamin E on mortality, but the modification of thevitamin E effect caused by smoking quantity was not statis-tically significant. Nevertheless, the greater reduction in mor-tality among old participants with high vitamin C intake whosmoked less suggests that vitamin E supplementation mightaffect mortality among older male nonsmokers as well.

Since vitamin E is a fat-soluble substance and it may takemonths before tissue levels are substantially elevated (30,31), short trials may be uninformative. In the current study,there seemed to be 2- and 3-year lag periods before vitaminE started to decrease or increase mortality (Figures 1 and 2).These lag periods are consistent with a delay in the effects ofthe fat-soluble vitamin. Nevertheless, the maximum of8 years of follow-up in the ATBC Study was long enoughto observe that vitamin E supplementation does have effects.On the other hand, the risk of tuberculosis increased signif-icantly within a year of supplementation being initiated (14)and the risk of pneumonia in participants who exercised intheir leisure time decreased without a lag (16), so not alleffects of vitamin E are delayed.

It has been suggested that the vitamin E doses in the ran-domized trials were too low to show any effect (32), and highdoses in some trials increased mortality (5). Given our ob-servations that 50 mg/day of vitamin E caused significantincrease and decrease in mortality in the ATBC Study pop-ulation, depending on the characteristics of participants, nojustification exists for claiming that the vitamin E doses inrandomized trials have been too low. Our study also suggeststhat it may be primarily subject characteristics and not doseof vitamin E that determines whether vitamin E causes harm.

Many people take vitamin E supplements because theybelieve that the vitamin protects them against oxidativestress, which has a role in the aging processes (2, 3). Ourfindings among the older ATBC Study participants are con-sistent with the concept that the antioxidant vitamins C andE can counteract oxidative stress processes in old peopleunder some conditions. Assuming that the benefit of vitaminE supplementation in old people is conditional on a highintake of vitamin C, it seems possible that old people withlow dietary vitamin C might benefit from a combination ofthe 2. Therefore, the most informative type of further studywould be a factorial design with vitamins E and C on oldpeople with low dietary vitamin C intake.

Finally, the US nutritional recommendations and a recentreview consider that vitamin E is safe in doses up to 1,000mg/day (33, 34). Our finding that 50 mg/day of vitamin Esignificantly increased mortality among half of the malesmokers in the ATBC Study aged 50–62 years contradictsthe universal safety of 1,000mg/day of vitamin E. Previously,we found that 50 mg/day of vitamin E significantly increasedthe risk of the common cold, pneumonia, and tuberculosis insubgroups of the ATBC Study, also indicating that such a lowdose can cause harmful effects in some people (14, 18, 19).

In conclusion, our subgroup analyses of the ATBC Studycohort support the conclusions of meta-analysis of con-trolled trials (4–6), in that vitamin E supplementation seemsineffective or harmful for middle-aged male smokers. Evi-

dently, in people younger than age 65 years, taking vitaminE supplements should be strongly discouraged until clearevidence emerges that some population groups of youngeror middle-aged people benefit. On the other hand, our studyindicates that vitamin E supplementation may lead to ben-eficial effects in some subgroups of old people, and thispossibility should be investigated by using a factorial designwith vitamin C supplementation. Finally, the substantial de-crease in mortality with vitamin E supplementation amongthe older participants with high dietary vitamin C intakeraises the question of whether the decrease in overall mor-tality is attributable to a single cause of death or a fewcauses, or whether it suggests a general decrease in frailtyreflected in lessened mortality from diverse causes.

ACKNOWLEDGMENTS

Authors affiliation: Department of Public Health, Univer-sity of Helsinki, Helsinki, Finland (Harri Hemila, JaakkoKaprio).

The authors thank the ATBC Study (the National PublicHealth Institute, Finland, and the National Cancer Institute,United States) for access to the data.

H. H. planned the study and wrote the draft of the manu-script. J. K. participated in planning the analyses and thecritical revision of the manuscript. H. H. had full access toall data in this study and takes responsibility for the integrityof the data and the accuracy of the data analysis.


REFERENCES

1. Radimer K, Bindewald B, Hughes J, et al. Dietary supplementuse by US adults: data from the National Health and NutritionExamination Survey, 1999–2000. Am J Epidemiol. 2004;160(4):339–349.

2. Beckman KB, Ames BN. The free radical theory of agingmatures. Physiol Rev. 1998;78(2):547–581.

3. Finkel T, Holbrook NJ. Oxidants, oxidative stress and the bi-ology of ageing. Nature. 2000;408(6809):239–247.

4. Vivekananthan DP, Penn MS, Sapp SK, et al. Use of antioxi-dant vitamins for the prevention of cardiovascular disease:meta-analysis of randomised trials. Lancet. 2003;361(9374):2017–2023.

5. Miller ER, Pastor-Barriuso R, Dalal D, et al. Meta-analysis:high-dosage vitamin E supplementation may increaseall-cause mortality. Ann Intern Med. 2005;142(1):37–46.(Comments: Ann Intern Med. 2005;143(2):150–158).

6. Bjelakovic G, Nikolova D, Gluud LL, et al. Mortality in ran-domized trials of antioxidant supplements for primary andsecondary prevention: systematic review and meta-analysis.JAMA. 2007;297(8):842–857. (Comments: JAMA. 2007;298(4):400–403).

7. Packer JE, Slater TF, Willson RL. Direct observation of a freeradical interaction between vitamin E and vitamin C. Nature.1979;278(5706):737–738.

8. Tanaka K, Hashimoto T, Tokumaru S, et al. Interactions be-tween vitamin C and vitamin E are observed in tissues of in-herently scorbutic rats. J Nutr. 1997;127(10):2060–2064.


Am J Epidemiol 2009;169:946–953

9. Hill KE, Montine TJ, Motley AK, et al. Combined deficiencyof vitamins E and C causes paralysis and death in guinea pigs.Am J Clin Nutr. 2003;77(6):1484–1488.

10. Bruno RS, Leonard SW, Atkinson J, et al. Faster plasmavitamin E disappearance in smokers is normalized by vitamin Csupplementation. Free Radic Biol Med. 2006;40(4):689–697.(Comments: Free Radic Biol Med. 2007;42(4):578–580).

11. Hemila H. Do Vitamins C and E Affect Respiratory Infections?[dissertation]. Helsinki, Finland: University of Helsinki;2006:10–11, 56–57. (Available at http://ethesis.helsinki.fi/julkaisut/laa/kansa/vk/hemila/). (Accessed December 15, 2008).

12. The effect of vitamin E and beta-carotene on the incidence oflung cancer and other cancers in male smokers. The Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study Group.N Engl J Med. 1994;330(15):1029–1035.

13. The alpha-tocopherol, beta-carotene lung cancer preventionstudy: design, methods, participant characteristics, and com-pliance. The ATBC Cancer Prevention Study Group. AnnEpidemiol. 1994;4(1):1–10.

14. Hemila H, Kaprio J. Vitamin E supplementation may tran-siently increase tuberculosis risk in males who smoke heavilyand have high dietary vitamin C intake. Br J Nutr. 2008;100(4):896–902. (Comments: Br J Nutr. 2009;101(1):145–147).

15. Hemila H, Virtamo J, Albanes D, et al. Vitamin E and beta-carotene supplementation and hospital-treated pneumonia in-cidence in male smokers. Chest. 2004;125(2):557–565.

16. Hemila H, Kaprio J, Albanes D, et al. Physical activity and therisk of pneumonia in male smokers administered vitamin Eand b-carotene. Int J Sports Med. 2006;27(4):336–341.

17. Hemila H, Kaprio J, Albanes D, et al. Vitamin C, vitamin E,and beta-carotene in relation to common cold incidence inmale smokers. Epidemiology. 2002;13(1):32–37.

18. HemilaH,Virtamo J,AlbanesD, et al. The effect of vitaminEoncommon cold incidence is modified by age, smoking and resi-dential neighborhood. J Am Coll Nutr. 2006;25(4):332–339.

19. Hemila H, Kaprio J. Vitamin E supplementation and pneu-monia risk in males who initiated smoking at an early age:effect modification by body weight and vitamin C [electronicarticle]. Nutr J. 2008;7(1):33.

20. Pietinen P, Hartman AM, Haapa E, et al. Reproducibilityand validity of dietary assessment instruments. I. A self-administered food use questionnaire with a portion size picturebooklet. Am J Epidemiol. 1988;128(3):655–666.

21. Ovaskainen ML, Valsta M, Lauronen J. The compilation offood analysis values as a database for dietary studies—theFinnish experience. Food Chem. 1996;57(1):133–136.

22. Hemila H, Kaprio J, Pietinen P, et al. Vitamin C and othercompounds in vitamin C rich food in relation to risk of tu-berculosis in male smokers. Am J Epidemiol. 1999;150(6):632–641.

23. Assmann SF, Pocock SJ, Enos LE, et al. Subgroup analysis andother (mis)uses of baseline data in clinical trials. Lancet.2000;355(9209):1064–1069.

24. Peto R, Pike MC, Armitage P, et al. Design and analysis ofrandomized clinical trials requiring prolonged observation ofeach patient. II. Analysis and examples. Br J Cancer. 1977;35(1):1–39.

25. Vandenbroucke JP. Observational research, randomised trials,and two views of medical science [electronic article]. PLoSMed. 2008;5(3):e67.

26. Feinstein AR. The problem of cogent subgroups: a clinico-statistical tragedy. J Clin Epidemiol. 1998;51(4):297–299.

27. Sesso HD, Buring JE, Christen WG, et al. Vitamins E and C inthe prevention of cardiovascular disease in men: the Physi-cians’ Health Study II randomized controlled trial. JAMA.2008;300(18):2123–2133.

28. Douglas RM, Hemila H. Vitamin C for preventing and treatingthe common cold [electronic article]. PLoS Med. 2005;2(6):e168.

29. Hemila H, Louhiala P. Vitamin C may affect lung infections.J R Soc Med. 2007;100(11):495–498.

30. Kitagawa M, Mino M. Effects of elevated d-alpha(RRR)-tocopherol dosage in man. J Nutr Sci Vitaminol (Tokyo). 1989;35(2):133–142.

31. Handelman GJ, Epstein WL, Peerson J, et al. Human adiposealpha-tocopherol and gamma-tocopherol kinetics during andafter 1 y of alpha-tocopherol supplementation. Am J Clin Nutr.1994;59(5):1025–1032.

32. Blumberg JB, Frei B. Why clinical trials of vitamin E andcardiovascular diseases may be fatally flawed. Commentary on‘‘The relationship between dose of vitamin E and suppressionof oxidative stress in humans’’. Free Radic Biol Med. 2007;43(10):1374–1376.

33. Food and Nutrition Board, Institute of Medicine. DietaryReference Intakes for Vitamin C, Vitamin E, Selenium, andCarotenoids. Washington, DC: The National Academies Press;2000: 249–260. (Available at http://www.nap.edu/books/0309069351/html/249.html). (Accessed December 15, 2008).

34. Hathcock JN, Azzi A, Blumberg J, et al. Vitamins E and C aresafe across a broad range of intakes. Am J Clin Nutr. 2005;81(4):736–745. (Comments: Am J Clin Nutr. 2005;82(5):1141–1143).


Am J Epidemiol 2009;169:946–953

http://ethesis.helsinki.fi/julkaisut/laa/kansa/vk/hemila/

http://ethesis.helsinki.fi/julkaisut/laa/kansa/vk/hemila/

http://www.nap.edu/books/0309069351/html/249.html

http://www.nap.edu/books/0309069351/html/249.html




Vol. 169, No. 8

DOI: 10.1093/aje/kwn421



Dietary Acrylamide Intake and Risk of Premenopausal Breast Cancer

Kathryn M. Wilson, Lorelei A. Mucci, Eunyoung Cho, David J. Hunter, Wendy Y. Chen, and WalterC. 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 commonlyconsumed foods. It is widespread; approximately 30% of calories consumed in the United States are from foodscontaining acrylamide. In animal studies, acrylamide causes mammary tumors, but it is unknown whether the levelof acrylamide in foods affects human breast cancer risk. The authors studied the association between acrylamideintake and breast cancer risk among 90,628 premenopausal women in the Nurses’ Health Study II. They calculatedacrylamide 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 assessthe 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 ofacrylamide intake (Ptrend ¼ 0.61). Results were similar regardless of smoking status or estrogen and progesteronereceptor 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 foundno evidence that acrylamide intake, within the range of US diets, is associated with increased risk of premeno-pausal breast cancer.

acrylamide; breast neoplasms; diet

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

Acrylamide is classified as a probable human carcino-gen (1), and it is formed during high-temperature pro-cessing of many commonly consumed foods (2). Thediscovery of acrylamide in foods in 2002 caused consider-able concern worldwide. Acrylamide is widespread in thefood supply, with approximately 38% of calories con-sumed in the United States coming from foods thatcontain acrylamide (3). Potatoes, cold breakfast cereal,coffee, and baked goods are major sources of acrylamideintake in the United States (4). Prior to the discovery ofacrylamide in foods, industrial use and tobacco use werethought to be the major sources of acrylamide exposure inhumans (1).

In animal tests, acrylamide administered in high levels indrinking water causes several types of hormone-sensitive

cancers, including mammary tumors in female rats (5, 6).Given the burden of breast cancer, it is of interest to studythe association between acrylamide intake in humans andthe risk of breast cancer. Epidemiologic studies have hadmixed results. Two prospective studies of dietary acrylam-ide exposure in humans found no association with pre- orpostmenopausal breast cancer risk (7, 8). Both of thesereports used food frequency questionnaires (FFQs) to as-sess acrylamide intake. A prospective study in the DanishDiet, Cancer, and Health Cohort used a biomarker of ac-rylamide exposure, acrylamide adducts to hemoglobin, andfound an increased risk of breast cancer among postmen-opausal women with higher adducts. The increased riskappeared limited to smokers and to estrogen receptor-positivecancers (9).

Correspondence to Dr. Kathryn M. Wilson, Channing Laboratory, Brigham and Women’s Hospital, 181 Longwood Avenue, Third Floor, Boston,

MA 02115 (e-mail: [email protected]).

954 Am J Epidemiol 2009;169:954–961

We used data from the Nurses’ Health Study II to assessthe association between acrylamide intake and premeno-pausal breast cancer risk. This cohort has repeated measuresof diet, which allows us to study acrylamide intake over anextended time. We previously reported on the creation of anacrylamide food composition database for this cohort (10);we found a moderate association between calculated acryl-amide intake and hemoglobin adducts of acrylamide ina subset of the cohort.


Study population

The Nurses’ Health Study II is a prospective cohortstudy of 116,671 female registered nurses aged 25–42years at the start of the study in 1989. Follow-up ques-tionnaires have been sent biennially to update informationon lifestyle and health. Beginning in 1991, and every4 years thereafter, a semiquantitative FFQ was sent toparticipants to assess their usual dietary intake over theprevious year. Women who completed the first FFQ in1991 (n ¼ 97,807) form the study population for thisanalysis.

We excluded women who had an implausible energy in-take (<800 or >4,200 kcal/day) or who left more than 70food items blank (n ¼ 2,361). We also excluded womenwho reported a diagnosis of cancer (excluding nonmela-noma skin cancer) before baseline in 1991 (n ¼ 1,308).

The analysis was limited to premenopausal women, sowomen who were postmenopausal at baseline were ex-cluded (n ¼ 3,462), and women were censored after theyreached natural or surgical menopause. Women who hada hysterectomy without a bilateral oophorectomy were ex-cluded (n ¼ 48) or censored at the time of surgery becausetheir menopausal status was unknown. This left a total of90,628 premenopausal women with baseline diet informa-tion for the analysis. The response rate was approximately90% among these women through the end of follow-up onJune 1, 2005. This study was approved by the human re-search committees at the Harvard School of Public Healthand Brigham and Women’s Hospital.

Assessment of acrylamide intake

FFQs with over 130 food items were completed in 1991,1995, 1999, and 2003. Participants were asked how fre-quently they had consumed a specified portion size of eachitem over the previous year with 9 possible responses,ranging from never or less than once a month to 6 or moretimes per day. The FFQ includes the major acrylamide-contributing foods according to US Food and Drug Admin-istration surveys: French fries, cold breakfast cereal, potatochips, cookies, coffee, breads, baked goods, and snackfoods (4).

We previously reported on the creation and validation ofan acrylamide food composition database for the FFQ (10).Briefly, 42 food items on the FFQ were assigned acrylam-ide contents based on published data from the US Food andDrug Administration and additional analyses of US foods

by the Swedish National Food Administration. We calcu-lated daily acrylamide intake for each participant by mul-tiplying the acrylamide content of 1 serving of food by thefrequency of consumption of that food and summing acrossall food items on the questionnaire. Acrylamide intakefrom cold breakfast cereal was based on participants’ re-porting of which brand they use most often. The correla-tion between 1999 acrylamide intake and a biomarker ofacrylamide exposure, the sum of hemoglobin adductsof acrylamide and its metabolite glycidamide, was 0.34(P < 0.0001) among 296 nonsmoking women from theNurses’ Health Study II cohort. The accuracy of reportingfor individual food items on a similar FFQ was measuredby comparing FFQ responses and 28 days of diet records ina subset of women in the Nurses’ Health Study (11). Thecorrelation between FFQ and diet records for the topacrylamide-contributing foods was 0.73 for French fries,0.78 for coffee, 0.60 for potato chips, and 0.79 for coldbreakfast cereal.

Because acrylamide may have an effect on carcinogenesisover an extended period of time, we used the cumulativeaverage intake of acrylamide to represent long-term dietaryintake. That is, 1991 intake was used for the 1991–1995follow-up period, the average of 1991 and 1995 intakeswas used for the 1995–1999 follow-up period, the averageof 1991, 1995, and 1999 intakes was used for the 1999–2003follow-up period, and the average of all 4 questionnaireswas used for the 2003–2005 follow-up period. Data fromthe previous FFQ were carried forward to the next timeperiod for participants with incomplete FFQ informationafter baseline. In secondary analyses, we examined the as-sociation between baseline acrylamide intake and breastcancer risk.

Ascertainment of breast cancer cases

Biennial follow-up questionnaires were used to identifynewly diagnosed cases of breast cancer. Deaths weredocumented by responses to questionnaires by familymembers, by the postal service, or through the NationalDeath Index. Cause of death was confirmed by medicalrecord review, information from relatives, or review ofdeath certificates.

When participants reported breast cancer, we asked theparticipant for confirmation of the diagnosis and permis-sion to obtain relevant medical records. Pathology reportsconfirmed 98% of the self-reported breast cancers. Infor-mation on estrogen and progesterone receptor status wasobtained from pathology reports and was available for78% of cases. A recent validation study in the Nurses’Health Study I cohort demonstrated that pathology reportsprovide accurate information on estrogen receptor status(12). Cases of carcinoma-in-situ were not included in theanalysis.


Each participant contributed person-time from the date ofreturn of the 1991 questionnaire until the time of breast can-cer diagnosis, menopause, death, or June 1, 2005, whichever

Acrylamide and Premenopausal Breast Cancer 955

Am J Epidemiol 2009;169:954–961

came first. Participants were divided into quintiles based ontheir acrylamide intake and their consumption of acrylamide-rich foods. Acrylamide intake was adjusted for total energyintake by using the residual method. Relative risks of breastcancer were calculated as the incidence rate for a givenquintile of consumption divided by the rate in the lowestquintile.

We used Cox proportional hazards regression to adjustfor potential confounding by other breast cancer risk fac-tors. To control as finely as possible for confounding byage, calendar time, and any possible 2-way interactionsbetween these 2 time scales, we stratified the analysisjointly by age in months at the start of each follow-upperiod and calendar year of the current questionnaire cycle.We used multivariable models to adjust for the followingfactors: 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), oralcontraceptive use (never, former use <4 years, former use�4 years, current use <8 years, and current use �8 years),

parity and age at first birth (nulliparous, 1–2 children andage at first birth <25 years, 1–2 children and age at firstbirth 25–<30 years, 1–2 children and age at first birth �30years, 3 or more children and age at first birth <25 years,3 or more children and age at first birth �25 years), age atmenarche (<12, 12, 13, or �14 years), family history ofbreast cancer (yes/no), history of benign breast disease(yes/no), smoking (never, former smoker of <25 ciga-rettes/day, former smoker of �25 cigarettes/day, currentsmoker of <25 cigarettes/day, and current smoker of �25cigarettes/day), physical activity (�18 and >18 metabolicequivalent (MET)-hours/week), animal fat (quintiles), glyc-emic load (quintiles), alcohol intake (continuous g/day), andtotal energy intake (continuous kcal/day). We adjusted foranimal fat and glycemic load as they have previously beenassociated with breast cancer risk in this cohort (13, 14).We also considered adjustment for quintile of vegetablefat intake, trans fat intake, and glycemic index, as thesedietary factors were most correlated with acrylamide in-take, but they were not included in final models because

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

Calorie-adjusted Acrylamide Intake

Quintile 1, Low(n 5 20,934)

Quintile 2(n 5 17,416)



Quintile 5, High(n 5 19,179)

Acrylamide intake, lg/day 10.8 16.6 20.2 24.6 37.8

Acrylamide by body weight,lg/kg/day

0.17 0.26 0.32 0.38 0.58

Age, years 36 36 36 36 36

Body mass index, kg/m2 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

Glycemic loadb,c 123 122 121 120 120

Intakes of high acrylamidefoods, 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.

956 Wilson et al.

Am J Epidemiol 2009;169:954–961

they had no substantial effect on the relative risk or stan-dard error estimates for acrylamide. All covariates exceptheight and age at menarche were updated in each question-naire cycle. The SAS Proc PHREG procedure (SAS Insti-tute, Inc., Cary, North Carolina) was used for all analyses,and the Anderson-Gill data structure was used to handletime-varying covariates efficiently. To test for a lineartrend across quintiles of intake, we modeled acrylamideintake as a continuous variable using the median value foreach quintile.

We examined whether the association between acryl-amide intake and breast cancer risk was modified by in-dividual characteristics, including age, smoking status,body mass index, alcohol intake, and glycemic load, bymodeling the association separately in each group. Wetested the significance of interactions by adding cross-product terms between acrylamide intake and the variableof interest to the model and comparing this model to themodel without the cross-product term using the likelihoodratio test.

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

Ptrendb

Quintile 1, Low(12 mg/day)

Quintile 2(17 mg/day)



Quintile 5, High(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



Former smokers

No. of cases 56 64 63 74 68



Current smokers

No. of cases 16 23 23 25 31



By ER and PR status

ERþ/PRþ breast cancer

No. of cases 105 129 111 138 114



ER�/PR� breast cancer

No. of cases 39 45 34 43 35



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).


Am J Epidemiol 2009;169:954–961

RESULTS

During 14 years (945,764 person-years) of follow-up, wedocumented 1,179 cases of invasive breast cancer among90,628 premenopausal women in the cohort. The age rangeof women in 1991 was 26–46 years. Ages at breast cancerdiagnosis ranged from 26 to 56 years. We had informationon estrogen receptor (ER)/progesterone receptor (PR) statusfor 916 (78%) cases. Of these, 597 were ER and PR positive(ERþ/PRþ), and 196 were ER and PR negative (ER�/PR�). Because of the small number of mixed ER/PR statustumors, we did not include these cases in our analysis byER/PR status.

Table 1 shows the characteristics of the cohort in 1991 byquintile of energy-adjusted acrylamide intake. The mean

acrylamide intake was 10.8 lg/day in the lowest quintileand 37.8 lg/day in the highest quintile. The major foodcontributors to acrylamide intake were French fries (23%),coffee (15%), cold breakfast cereal (12%), potato chips(9%), and other potatoes (baked, roasted, mashed; 5%).Those in the highest quintile of acrylamide consumptionwere more likely to be current smokers and were less likelyto exercise than those in the lowest quintile.

Intake of acrylamide was not associated with risk of pre-menopausal breast cancer (Table 2). The multivariable rel-ative risk of breast cancer was 0.92 (95% confidence interval(CI): 0.76, 1.11) in the highest quintile of intake comparedwith the lowest quintile. The P value for a linear trend acrossquintiles was 0.61. No association was found for ERþ/PRþor ER�/PR� cancers.

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 FoodsPtrend

a

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

Multivariable relative riskb 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


Breakfast cereal


No. of cases 207 254 226 272 220


Potato chips


No. of cases 219 313 204 216 227


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


Popcorn


No. of cases 256 220 247 273 183


Muffins


No. of cases 216 155 324 212 272


Crackers


No. of cases 244 195 204 285 251


Table continues

958 Wilson et al.

Am J Epidemiol 2009;169:954–961

Because tobacco use is a major source of acrylamideexposure, we examined the association among neversmokers, former smokers, and current smokers separately(Table 2). There was no indication of increased risk ofbreast cancer for higher acrylamide intakes in any of thesegroups.

We found no significant differences in the associationbetween dietary acrylamide intake and breast cancer riskwhen we stratified the population by age, body mass index,alcohol intake, glycemic index, or glycemic load (data notshown).

We repeated the analysis measuring acrylamide exposurerelative to body weight (i.e., lg/kg of body weight/day), as

this is the exposure measurement used in toxicology studies,and again found similar results. The relative risk for thehighest versus the lowest quintile of acrylamide by bodyweight was 1.00 (95% CI: 0.82, 1.22), with a P value forlinear trend of 0.95. Baseline acrylamide intake through dietwas also not associated with breast cancer risk. The relativerisks relative to the lowest quintile of baseline acrylamideintake were 0.96 (95% CI: 0.79, 1.15) for quintile 2,1.05 (95% CI: 0.88, 1.26) for quintile 3, 1.01 (95% CI:0.84, 1.21) for quintile 4, and 1.03 (95% CI: 0.86, 1.24)for quintile 5 (Ptrend ¼ 0.62).

Table 3 shows the association between consumptionof the major acrylamide-contributing foods and

Table 3. Continued

Intake of High-Acrylamide FoodsPtrend

a

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

Dark bread


No. of cases 219 267 280 199 214


English muffins, bagels, rolls


No. of cases 149 279 270 268 213


Pizza


No. of cases 185 268 324 271 131


All potatoesc


No. of cases 246 225 250 239 219


Breads/starchesc

Median intake, servings/day 0.7 1.2 1.6 2.2 3.4

No. of cases 253 223 228 258 217


Baked goodsc


No. of cases 257 223 220 237 242


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.


Am J Epidemiol 2009;169:954–961

premenopausal breast cancer risk. We examined all individ-ual foods that contributed at least 2% to the total estimatedacrylamide intake in our population, and none was posi-tively associated with breast cancer risk. We also examinedseveral food groups that are major sources of acrylamide: allpotatoes (French fries, potato chips, and baked/mashed/roasted potatoes), breads (white and dark bread, Englishmuffins/bagels/rolls, tortillas, pancakes, pizza, and crack-ers), and baked goods (cookies, brownies, donuts, cake,pie, and sweet rolls). None of these food groups was asso-ciated with breast cancer risk.

DISCUSSION

We found no association between acrylamide intake andpremenopausal breast cancer risk in this cohort. There wasno association for ER/PR-positive or -negative cancers or bysmoking status. In addition, there was no association be-tween intake of any major acrylamide-contributing foodsand breast cancer risk.

These findings are in line with both previous prospectivestudies of acrylamide intake and breast cancer risk usingFFQs. Mucci et al. (7) found no significant association be-tween acrylamide intake and breast cancer risk amongmostly premenopausal Swedish women. Hogervorst et al.(8) found no significant association among postmenopausalDutch women. The mean acrylamide intake and the contrastin intakes between high and low quintiles were quite similaracross the 3 studies. However, the main sources of acrylam-ide vary between populations. French fries, coffee, coldbreakfast cereal, and potato chips contribute the most tointake in our US population. Coffee is a larger contributorto acrylamide intake in both the Swedish and Dutch cohorts(7, 8). Fried potatoes and crisp bread are other major con-tributors in the Swedish women (7), and Dutch spice cakeand cookies are major contributors in the Dutch women (8).In addition, Pelucchi et al. found no association betweenacrylamide intake (15) or fried potato intake (16) and breastcancer risk in a hospital-based, case-control study in Italyand Switzerland.

In a nested case-control study, Olesen et al. (9) studied theassociation between acrylamide adducts to hemoglobin,a biomarker of acrylamide exposure, and postmenopausalbreast cancer risk. Current smokers with higher levels ofacrylamide adducts to hemoglobin at baseline had a signifi-cantly increased risk of breast cancer (for a 10-fold increasein adducts, relative risk (RR) ¼ 3.1, 95% CI: 1.0, 9.7). To-bacco use is an important source of acrylamide exposure,and smokers have acrylamide adduct levels 3–5 times higherthan do nonsmokers. Therefore, adduct levels among smok-ers reflect both tobacco use and dietary intake of acrylam-ide. Among nonsmoking women, whose adduct levels arethought primarily to represent dietary acrylamide exposure,Olesen et al. found no statistically significant associationbetween adduct levels and breast cancer risk (for a 10-foldincrease in adducts, RR ¼ 1.5, 95% CI: 0.6, 3.6). Themeaning of the positive association with adduct levelsamong smokers is not clear; however, their results amongnonsmokers seem in line with our finding that acrylamide

exposure in the range of dietary intakes is not clearly asso-ciated with breast cancer risk. Olesen et al. also found a sig-nificantly increased risk of ERþ cancers among those withhigher adduct levels (for a 10-fold increase in adducts, RR¼2.7, 95% CI: 1.1, 6.6). This result was among smokersand nonsmokers combined, with multivariable adjustmentfor smoking behavior; however, given the relative contribu-tions of smoking and diet to adduct levels, it is not clear thatsuch adjustment provides adequate control of confounding bysmoking. In our analysis of food frequency questionnaire-assessed acrylamide and ERþ cancers, we found no signifi-cant association.

Strengths of our study include its prospective design,large number of premenopausal cases, and high rates offollow-up. In addition, we are the first to study acrylamideintake and cancer risk using multiple FFQs administered tocollect updated dietary data throughout the follow-up pe-riod, rather than a single point in time. This improves ourassessment of long-term diet and reduces measurement error(17). Our study also uses an extensive acrylamide databasewith 42 acrylamide-contributing foods, approximately twiceas many as used in previous studies.

We have previously found that FFQ acrylamide intake issignificantly correlated with hemoglobin adducts of acryl-amide and its metabolite glycidamide in this population(10). However, misclassification of acrylamide intake re-mains a limitation of our study, as acrylamide poses uniquechallenges for FFQ assessment. Acrylamide formation isaffected by many parameters (for review, refer to Stadlerand Scholz (18)), such as cooking temperature and eventhe length and temperature of storage of ingredients suchas potatoes, which implies that acrylamide content varieswidely among different brands of prepared foods, and de-pends on the cooking methods used at home. Therefore, theassignment of a single acrylamide value for each food likelyresults in nondifferential misclassification of acrylamide in-take, which would bias our observed relative risks towardthe null. As a result, we may have missed modest associa-tions between acrylamide intake and cancer risk. We alsofound no association between breast cancer risk and intakeof any of the top 11 acrylamide-contributing foods, whichare well measured by the FFQ within the similar Nurses’Health Study cohort (11).

Residual confounding is also a concern in observationalstudies. Adjustment for known breast cancer risk factors hadvery little effect on the relative risk estimates, suggesting thatit is unlikely that confounding was a source of substantialbias. Finally, it is not clear that adult diet is the most rel-evant period of exposure. It is possible that high acrylam-ide intake in childhood or adolescence may increase breastcancer risk later in life, and our study cannot address thisquestion.

In conclusion, we found no association between acrylam-ide intake from the diet and risk of premenopausal breastcancer risk. Combined with the results of other prospectivecohort studies, this suggests that intake of foods high inacrylamide is not a major risk factor for breast cancer. How-ever, a modest association could have been missed, becausethe substantial variation in the acrylamide content of foodsmakes measurement of intake difficult.

960 Wilson et al.

Am J Epidemiol 2009;169:954–961

ACKNOWLEDGMENTS

Author affiliations: Department of Epidemiology,Harvard School of Public Health, Boston, Massachusetts(Kathryn M. Wilson, Lorelei A. Mucci, David J. Hunter,Walter C. Willett); Department of Nutrition, HarvardSchool of Public Health, Boston, Massachusetts (KathrynM. Wilson, Eunyoung Cho, David J. Hunter, Walter C.Willett); Channing Laboratory, Department of Medicine,Harvard Medical School and Brigham and Women’sHospital, Boston, Massachusetts (Lorelei A. Mucci,Eunyoung Cho, David J. Hunter, Wendy Y. Chen, WalterC. Willett); and Department of Medical Oncology, Dana-Farber Cancer Institute, Boston, Massachusetts (Wendy Y.Chen).

This work was supported by a grant from the NationalCancer Institute (CA050385) and by a National CancerInstitute/National Institutes of Health training grant (T32CA09001 to K. M. W.)


REFERENCES

1. International Agency for Research on Cancer.Monographs onthe Evaluation of Carcinogenic Risks to Humans. Vol 60.Lyon, France: International Agency for Research on Cancer;1994.

2. Tareke E, Rydberg P, Karlsson P, et al. Analysis of acrylamide,a carcinogen formed in heated foodstuffs. J Agric Food Chem.2002;50(17):4998–5006.

3. Petersen BJ, Tran N. Exposure to acrylamide: placingexposure in context. Adv Exp Med Biol. 2005;561:63–76.

4. DiNovi M. The 2006 Exposure Assessment for Acrylamide.Rockville, MD: Center for Food Safety and Applied Nutrition,Food and Drug Administration; 2006. (http://www.cfsan.fda.gov/~dms/acryexpo.html).

5. Johnson KA, Gorzinski SJ, Bodner KM, et al. Chronic toxicityand oncogenicity study on acrylamide incorporated in thedrinking water of Fischer 344 rats. Toxicol Appl Pharmacol.1986;85(2):154–168.

6. FriedmanMA, Dulak LH, StedhamMA.A lifetime oncogenicitystudy in rats with acrylamide. Fundam Appl Toxicol. 1995;27(1):95–105.

7. Mucci LA, Sandin S, Balter K, et al. Acrylamide intake andbreast cancer risk in Swedish women. JAMA. 2005;293(11):1326–1327.

8. Hogervorst JG, Schouten LJ, Konings EJ, et al. A prospectivestudy of dietary acrylamide intake and the risk of endometrial,ovarian, and breast cancer. Cancer Epidemiol BiomarkersPrev. 2007;16(11):2304–2313.

9. Olesen PT, Olsen A, Frandsen H, et al. Acrylamide exposureand incidence of breast cancer among postmenopausal womenin the Danish Diet, Cancer and Health Study. Int J Cancer.2008;122(9):2094–2100.

10. Wilson KM, Vesper HW, Tocco P, et al. Validation of a foodfrequency questionnaire measurement of dietary acrylamide in-take using hemoglobin adducts of acrylamide and glycidamide.CancerCausesControl.2008. (doi: 10.1007/s10552-008-9241-7).(http://www.springerlink.com/content/5470507675718735/fulltext.pdf).

11. Salvini S, Hunter DJ, Sampson L, et al. Food-based validationof 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 estrogenreceptor results from pathology reports with results fromcentral laboratory testing. J Natl Cancer Inst. 2008;100(3):218–221.

13. Cho E, Spiegelman D, Hunter DJ, et al. Premenopausal fatintake and risk of breast cancer. J Natl Cancer Inst. 2003;95(14):1079–1085.

14. Cho E, Spiegelman D, Hunter DJ, et al. Premenopausal dietarycarbohydrate, glycemic index, glycemic load, and fiber in re-lation to risk of breast cancer. Cancer Epidemiol BiomarkersPrev. 2003;12(11 pt 1):1153–1158.

15. Pelucchi C, Galeone C, Levi F, et al. Dietary acrylamide andhuman cancer. Int J Cancer. 2006;118(2):467–471.

16. Pelucchi C, Franceschi S, Levi F, et al. Fried potatoes andhuman cancer. Int J Cancer. 2003;105(4):558–560.

17. Willett WC. Nutritional Epidemiology. New York, NY: OxfordUniversity Press; 1998.

18. Stadler RH, Scholz G. Acrylamide: an update on currentknowledge in analysis, levels in food, mechanisms of forma-tion, and potential strategies of control. Nutr Rev. 2004;62(12):449–467.


Am J Epidemiol 2009;169:954–961

http://www.cfsan.fda.gov/&sim;dms/acryexpo.html



http://www.springerlink.com/content/5470507675718735/fulltext.pdf

http://www.springerlink.com/content/5470507675718735/fulltext.pdf




Vol. 169, No. 8

DOI: 10.1093/aje/kwn422



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, JohnD. 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. Fewstudies have evaluated risk factors for the development of asynchronous contralateral breast cancer in womenwith breast cancer. In the Women’s Environmental Cancer and Radiation Epidemiology Study (1985–2001), theroles 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 atfirst breast cancer diagnosis were identified from 5 population-based cancer registries in the United States andDenmark. 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 associatedwith 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 isa 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 breastcancer, a primary breast cancer occurring in the oppositebreast subsequent to a first breast cancer diagnosis in 1 breast,in female survivors of breast cancer is considerably higherthan the risk of developing a first primary breast cancer inunaffected women (1). The population of women with breastcancer have a higher prevalence of all breast cancer riskfactors, both genetic and nongenetic, than women withoutbreast cancer. It would be expected that women who go onto develop asynchronous contralateral breast cancer wouldhave an even higher prevalence of risk factors than thosewho develop only 1 primary, although there may be mitigat-ing factors such as the treatment received for the first primaryand changes in behavior. Few studies have examined the role

of alcohol intake or cigarette smoking, 2 potentially modifi-able risk factors, in the development of asynchronous contra-lateral breast cancer. Given the high level of risk in thesewomen, clarifying the role of modifiable risk factors in asyn-chronous contralateral breast cancer in women with breastcancer is important.

A number of meta-analyses confirm that alcohol is a riskfactor for both pre- and postmenopausal first primary breastcancer, although the magnitude of increased risk associatedwith consuming 1 or more drinks per day is moderate (2–4).Recent studies confirm this finding (5, 6). Thus far, a signif-icant association between alcohol intake and second pri-mary breast cancer has not been observed (7–9), although2 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: [email protected]).

962 Am J Epidemiol 2009;169:962–968

1.11 associated with recent or ever drinking prior to the firstdiagnosis, respectively (7, 9). A variety of mechanisms,which may all contribute to the relation, have been proposedto explain the association between alcohol and breast cancerincluding estrogen metabolism, acetaldehyde mutagenesis,oxidation and free radicals, and 1-carbon metabolism (10).

Cigarette smoking has been more controversial as a possi-ble risk factor for breast cancer with inconsistent results in theliterature (11, 12). Conflicting results may be due to the com-peting effects of smoking at different ages. In some studies,increased risk has been associated specifically with smokingat an early age during breast development, although observa-tions regarding the effect of early smoking vary (13–15). Therelation between smoking and second primary breast cancerhas also been inconsistent, with some studies observingevidence of an association (16, 17) and others not (7–9).

In this study, we evaluate the evidence for an associationbetween alcohol intake and cigarette smoking and the devel-opment of asynchronous contralateral breast cancer amongwomen with a first diagnosis of breast cancer from theWomen’s Environmental Cancer and Radiation Epidemiol-ogy (WECARE) Study.


The WECARE Study is a multicenter, population-based,nested case-control study where women with asynchronouscontralateral breast cancer serve as cases and women withunilateral breast cancer serve as matched controls (18). Ad-ditional detail on data collection has been reported previously(19). All participants were identified through 5 population-based tumor registries, 4 in the United States (Los AngelesCounty Cancer Surveillance Program, Cancer SurveillanceSystem of the Fred Hutchinson Cancer Research Center,State Health Registry of Iowa, Cancer Surveillance Programof Orange County/San Diego-Imperial Organization forCancer Control) and 1 in Denmark (the Danish Breast CancerCooperative Group Registry supplemented by data from theDanish Cancer Registry).

Study population

Women were eligible as cases if they were diagnosed be-tween January 1, 1985, and December 31, 2000, and wereaged less than 55 years with a first primary invasive breastcancer that did not spread beyond the regional lymph nodes atdiagnosis and a second primary in situ or invasive breastcancer diagnosed in the contralateral breast at least 1 yearafter the first breast cancer diagnosis. The asynchronous con-tralateral breast cancer had to have been diagnosed no laterthan December 31, 2001. Case patients were required to haveresided in the same reporting area at the time of diagnosis ofboth cancers, to have had no prior or intervening cancer di-agnosis between their first and second primary breast cancers,and to be alive at the time of contact. Two control subjectswere individually matched to each case on year of birth, yearof diagnosis, registry region, and race and were 1:2 counter-matched on registry-reported radiation exposure, so that eachtriplet consisted of 1 radiation-unexposed and 2 radiation-exposed subjects. In addition, controls had to meet the

following criteria: 1) diagnosed since January 1, 1985, witha first primary invasive breast cancer while residing in 1 ofthe study reporting areas; 2) residing on the reference date(the date of first diagnosis plus ‘‘at-risk interval,’’ the timebetween the first and second diagnoses, of the matched case)in the same registry reporting area where they were diagnosedwith their breast cancer; 3) never diagnosed (by referencedate) with a second primary breast cancer or any other cancer;4) alive at the time of contact; and 5) without prophylacticmastectomy of the contralateral breast following diagnosis oftheir first primary. The control sampling is accounted for inthe analysis by the inclusion of sampling weights. The designhas been discussed in detail previously (18).

A total of 998 women with asynchronous contralateralbreast cancer were eligible and approached for inclusionin the study as cases, and 2,112 women with unilateralbreast cancer were eligible as controls. Of these potentialparticipants, 708 cases (71%) and 1,399 controls (66%)completed the interview and had a blood sample drawn.

Data collection

All participants in the WECARE Study were interviewedover the telephone to obtain information on known andsuspected risk factors for breast cancer including the follow-ing: personal demographics; age at menarche, first birth, andmenopause; parity and lactation; body size; and family his-tory of cancer. All questions were asked in reference to theperiod before the reference date (defined above). Thefollowing descriptions reflect the actual wording used inthe questionnaire or by the interviewer. The specific infor-mation collected on alcohol and smoking consisted ofwhether the women had ever smoked cigarettes or drunkany alcoholic beverages regularly (at least 1 cigarettea day for 6 months or longer or at least 1 drink per month,respectively) before the reference date, how old they werewhen they first started smoking cigarettes or drinking alco-holic beverages regularly, whether they had stopped smok-ing or drinking regularly, at what age they last stopped, howmany total years they smoked or drank regularly, and, dur-ing periods when they smoked regularly, on average, howmany cigarettes they usually smoked per day, week, ormonth. With respect to alcohol consumption, they were toldthat 1 drink is equal to 1 bottle or can of beer, 1 glass of wineor bottle of wine cooler, or 1 cocktail, shot, or mixed drinkof liquor and then asked to describe their average alcoholconsumption before the reference date in categories (neveror less than 1 drink each month, 1–3 drinks each month, 1drink each week, 2–4 drinks each week, 5 or 6 drinks eachweek, 1 drink each day, 2 or 3 drinks each day, or 4 or moredrinks each day). Medical records, pathology reports, andhospital charts were used to collect detailed information ontreatment and tumor characteristics. The study protocol wasapproved by the institutional review boards at each studysite and by the ethical committee system in Denmark.


We used conditional logistic regression analysis with theinclusion of a log weight covariate in the model where the

Alcohol and Smoking and Contralateral Breast Cancer 963

Am J Epidemiol 2009;169:962–968

coefficient of this log weight is fixed at 1 (i.e., an offset inthe model). The weights used the numbers of registry-reported, radiation-exposed and -unexposed women in therisk set to account for the countermatched sampling design(18, 20). Further, because controls were independently sam-pled from the failure time risk sets, the estimated parametersare rate ratios in the proportional hazards model for cohortdata (21), and standard likelihood methods apply (22).Smoking and alcohol consumption were each examined asever regular use (yes/no), regular use during the at-risk pe-riod (yes/no), lifetime duration in tertiles defined in controlsversus never use, and age at starting drinking or smoking.

Consumption during the at-risk period was defined as start-ing prior to or during the period between first diagnosis andthe reference date and stopping during or after the periodbetween the first diagnosis and the reference date. The av-erage smoking amount was defined as half a pack per day orless and greater than half a pack per day versus neversmoked and also as pack-years in tertiles versus neversmoked. Average alcohol consumption, which was collectedas categories described above, was defined as less than 1drink per day and 1 drink per day or more compared withnever drinking. After adjusting for age in models examiningever versus never drinking and smoking, we tested the fol-lowing potential confounders: education, first-degree familyhistory of breast cancer, body mass index (both at first di-agnosis and at the reference date), age at menarche, age atmenopause, stage of first primary, histology of first primary,exposure to chemotherapy, exposure to radiation treatment,use of tamoxifen, number of full-term pregnancies, age atfirst full-term pregnancy, and ever breastfeeding. In addi-tion, to test for mutual confounding, ever regular smokingwas added to models of alcohol drinking, and ever regulardrinking was added to models of smoking. As there was noindication of confounding, defined as a change in the rateratio estimate of 10% or more over the rate ratio from the

Table 1. Characteristics of Women With Unilateral and

Asynchronous Contralateral Breast Cancer, the Women’s

Environmental Cancer and Radiation Epidemiology Study,

1985–2001

Breast Cancer

Unilateral(n 5 1,399)

AsynchronousContralateral(n 5 708)

No. %a No. %

Matched characteristics

Registry

Iowa 222 15.9 113 16.0

Orange County/SanDiego, California

231 16.5 118 16.7

Los Angeles County,California

390 27.9 199 28.1

Seattle, Washington 198 14.2 99 14.0

Denmark 358 25.6 179 25.3

Race

Non-Hispanic white 1,288 92.1 649 91.7

Hispanic white 48 3.4 24 3.4

Black 39 2.8 21 3.0

Other 24 1.7 14 2.0

Mean age at first diagnosis,years (range)

45 (23–55) 46 (24–55)

Mean age at reference date,yearsb (range)

51 (27–69) 51 (27–71)

Mean at-risk period, years(range)

5 (1–16) 5 (1–16)

Countermatched characteristic

Radiation treatment

No 266 50.2 362 51.1

Yes 1,133 49.8 346 48.9

Other characteristics

Year of seconddiagnosis

1986–1988 21 3.0

1989–1991 75 10.6

1992–1994 137 19.4

1995–1997 202 28.5

1998–2001 273 38.6

Table continues

Table 1. Continued

Breast Cancer

Unilateral(n 5 1,399)

AsynchronousContralateral(n 5 708)

No. %a No. %

Family historyc of breastcancer

No 1,088 77.9 472 66.7

Yes 285 20.4 225 31.8

Unknown 26 1.8 11 1.6

Histology of first breastcancer

Lobular 131 8.7 90 12.7

Medullar 51 3.4 33 4.7

Ductal and other 1,213 87.9 584 82.6

Stage of first breast cancer

Localized 916 64.3 506 71.5

Regional 483 35.7 202 28.5

Chemotherapy

No 629 42.5 386 54.5

Yes 770 57.5 322 45.5

Hormone therapy

No 909 66.3 511 72.2

Yes 488 33.7 197 27.8

a Proportions are weighted for the countermatching with the excep-

tion of factors contributing to the study matching schema.b ‘‘Reference date’’ is the date of diagnosis for asynchronous con-

tralateral breast cancer and the corresponding date for unilateral

breast cancer.c First-degree family history.

964 Knight et al.

Am J Epidemiol 2009;169:962–968

age-adjusted model, only the age-adjusted models are pre-sented. We performed tests for trend across categories ofduration and consumption. All analyses were conducted usingSAS, version 9.1, software (SAS Institute, Inc., Cary, NorthCarolina), and a 2-sided P < 0.05 was considered significant.

RESULTS

Table 1 shows selected characteristics of the WECAREStudy population. The 2 groups were similar on all matchedcharacteristics. Women with asynchronous contralateralbreast cancer were more likely to have a family history ofbreast cancer. The majority (67%) of second diagnoses oc-curred in 1995 or later. In addition, 56% of cases and 49% ofcontrols were interviewed within 5 years of the referencedate, and 88% of cases and 86% of controls were inter-viewed within 10 years. Table 2 shows that ever regulardrinking was associated with an elevated risk of developingasynchronous contralateral breast cancer (rate ratio (RR) ¼1.3, 95% confidence interval (CI): 1.0, 1.6), although the

increased risk associated with drinking specifically afterthe first diagnosis (the at-risk period) did not achieve statis-tical significance (RR ¼ 1.2, 95% CI: 0.9, 1.5). The riskincreased with increasing lifetime duration of drinking(Ptrend ¼ 0.03). We observed no apparent trend associatedwith the average amount of alcohol consumed. The risk ofasynchronous contralateral breast cancer increased withlater initiation of drinking (at age 20 or more years) butnot with early initiation (before age 20 years). We did notobserve any differences in the risks associated with the du-ration of drinking prior to the first pregnancy, adjusted forthe age at first full-term pregnancy: began drinking after firstpregnancy (RR ¼ 1.5, 95% CI: 1.1, 2.1); drank <7 yearsbefore first pregnancy (RR ¼ 1.4, 95% CI: 1.0, 1.9); drank�7 years before first pregnancy (RR ¼ 1.2, 95% CI: 0.8,1.7), versus parous and never drank. There was no evidencefor any association with smoking (Table 3) nor any consis-tent pattern with smoking before the first pregnancy, ad-justed for age at first full-term pregnancy: began smokingafter first pregnancy (RR ¼ 1.6, 95% CI: 1.0, 2.5); smoked

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

Rate Ratioa 95% ConfidenceInterval

UnilateralAsynchronousContralateral

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

Ever drank regularly during at-risk periodc

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

Ptrendd 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.


Am J Epidemiol 2009;169:962–968

<6 years before first pregnancy (RR ¼ 1.0, 95% CI: 0.7,1.4); smoked �6 years before first pregnancy (RR ¼ 1.3,95% CI: 1.0, 1.8), versus parous and never smoked. We didnot observe any difference in the effect of smoking or alco-hol by categories of radiation exposure or any difference inthe effect of alcohol by tamoxifen use (data not shown). Wealso did not observe consistent differences in results acrosscategories when we conducted the analysis by the time be-tween reference date and interview in 3 categories, 0–5years, 6–10 years, and >10 years (data not shown).

DISCUSSION

In the WECARE Study, we found that consuming alco-hol, particularly over longer periods of time, was associated

with an increased risk of asynchronous contralateral breastcancer. We did not find evidence that smoking cigarettesincreased the risk of this disease. Our results differ fromprevious studies that did not observe an increased risk ofasynchronous contralateral breast cancer associated withalcohol (7–9). However, as with a first primary breast can-cer, the effect of alcohol is modest and could be missed if thesample size was insufficient. Previous studies included 488,77, or 136 cases (7–9) compared with the 708 cases in ourstudy. Varying definitions and prevalences of drinking mayalso have contributed to the inconsistency among studies. Inthe largest previous study, only information on recent drink-ing prior to the first diagnosis was available, and the authorswere unable to evaluate lifetime duration of drinking (7).We did not find a relation with reported average amount

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

Rate Ratioa95% Confidence

IntervalUnilateral

AsynchronousContralateral

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

Ever smoked regularly during at-risk periodc

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.

966 Knight et al.

Am J Epidemiol 2009;169:962–968

consumed, but the women were asked to average theirlifetime consumption, and changes in intake were not cap-tured, including changes occurring around the time of thefirst diagnosis. Relatively few women in this population(12%) reported consuming 1 drink per day or more onaverage.

Previous studies of smoking and asynchronous contralat-eral breast cancer have yielded inconsistent results (7–9, 16,17), similar to studies of first primary breast cancer (11, 12).As with alcohol, this inconsistency may be due to issues ofstudy design, such as variation in sample size, smokingdefinitions, and smoking prevalences. Our results from theWECARE Study do not support the hypothesis that smokingis a risk factor for asynchronous contralateral breast cancer,although a small increased risk associated with smokingcannot be ruled out. Previous studies have not evaluatedchanges in drinking and smoking behavior. In the WECAREStudy, we found that only a small proportion of womenchanged their drinking status after their first breast cancerdiagnosis (10% of cases and 9% of controls), but the pro-portion who stopped smoking was greater (28% of cases and29% of controls).

The WECARE Study included women who were agedless than 55 years at first diagnosis (mean age, 45 yearsamong controls and 46 years among cases), although theywere somewhat older at the reference date (mean age in bothgroups, 51 years). These women were younger than those insome other studies (7) but not others (8). However, there isno evidence that the effect of smoking or alcohol on the riskof first primary breast cancer differs by menopausal status(2–4, 10).

An important strength of the WECARE Study is that it isthe largest case-control study conducted of asynchronous con-tralateral breast cancer to date that includes direct patientinterview. We have also been able to detect other associa-tions in their expected directions for fewer full-term preg-nancies and early menarche (19), radiation treatment (23),and treatment of the first primary breast cancer with chemo-therapy or tamoxifen (24). However, although we did cap-ture some information on lifetime duration of smoking andalcohol consumption, the level of detail of the informationcollected was limited. We did not capture the changes in thepatterns of consumption at various times of life. Althoughrecall bias is a problem in case-control studies in general,whether or not it is an issue in our study is unclear, as boththe cases and controls have been affected by breast cancer.In this study, cases and controls were matched on time sincethe first diagnosis, and the majority (56%) of cases wereinterviewed within 5 years and most (88%) within 10 yearsof the second diagnosis. It is likely that women can recallaccurately whether they drank or smoked, and the broadcategories used in the analysis of duration of drinking andsmoking minimize the potential for misclassification errorsin recall of when drinking and smoking began and ended. Aswith amount, we were unable to examine duration in greaterdetail. We were also unable to consider results by estrogenand progesterone status, as a considerable proportion of thecases were missing information on receptor status. Resultsfrom studies relating alcohol to first breast cancer accordingto hormone receptor status are inconsistent (10, 25). Further,

if drinking and/or smoking adversely affects survival, and iffewer women who drink and/or smoke survive to be diag-nosed with a second primary or to be included in the studyafter the second diagnosis, this could affect the relative risksreported here.

Women with a first primary unilateral breast cancer havean elevated risk of developing cancer in the contralateralbreast. Although we did not observe an increased risk ofasynchronous contralateral breast cancer associated withcigarette smoking, there are many other reasons to quitsmoking including reducing the risk of smoking-relatedcancer or heart disease. Alcohol appears to be associatedwith an increased risk of asynchronous contralateral breastcancer.

ACKNOWLEDGMENTS

Author affiliations: Samuel Lunenfeld Research Institute,Mount Sinai Hospital, Toronto, Canada (Julia A. Knight);City of Hope, Duarte, California (Leslie Bernstein); Univer-sity of California at Irvine, Irvine, California (Joan Largent);Memorial Sloan Kettering Cancer Center, New York, NewYork (Marinela Capanu, Colin Begg, Anne S. Reiner,Xiaolin Liang, Jonine L. Bernstein); Danish Cancer Society,Copenhagen, Denmark (Lene Mellemkjær); University ofIowa, Iowa City, Iowa (Charles F. Lynch); Fred HutchinsonCancer Research Center, Seattle, Washington (Kathleen E.Malone); University of Southern California, Los Angeles,California (Robert W. Haile); and International Epidemiol-ogy Institute, Rockville, Maryland, and Vanderbilt Univer-sity, Nashville, Tennessee (John D. Boice, Jr.).

This work was supported by the National Institutes ofHealth (grants U01 CA83178 and R01 CA97397).

WECARE Study Collaborative Group: Memorial SloanKettering Cancer Center (New York, New York): Jonine L.Bernstein (WECARE Study Principal Investigator), ColinBegg, Marinela Capanu, Xiaolin Liang, Anne S. Reiner,Irene Orlow, Tracy Layne; City of Hope (Duarte, Califor-nia): Leslie Bernstein (subcontract Principal Investigator),Laura Donnelly-Allen (some work performed at Universityof Southern California, Los Angeles, California); DanishCancer Society (Copenhagen, Denmark): Jørgen H. Olsen(subcontract Principal Investigator), Michael Andersson,Lisbeth Bertelsen, Per Guldberg, Lene Mellemkjær; FredHutchinson Cancer Research Center (Seattle, Washington):Kathleen E. Malone (subcontract Principal Investigator),Noemi Epstein; International Epidemiology Institute(Rockville, Maryland) and Vanderbilt University (Nashville,Tennessee): John D. Boice, Jr. (subcontract Principal Investi-gator), David P. Atencio; National Cancer Institute (Bethesda,Maryland): Daniela Seminara; New York University (NewYork, New York): Roy E. Shore (subcontract Principal Inves-tigator);University of California at Irvine (Irvine, California):Hoda Anton-Culver (subcontract Principal Investigator),Joan Largent (Coinvestigator); University of California atLos Angeles (Los Angeles, California): Richard A. Gatti(Coinvestigator); University of Iowa (Iowa City, Iowa):Charles F. Lynch (subcontract Principal Investigator), Jeanne


Am J Epidemiol 2009;169:962–968

DeWall;University of Southern California (Los Angeles, Cal-ifornia): RobertW. Haile (subcontract Principal Investigator),Bryan M. Langholz (Coinvestigator), Duncan C. Thomas(Coinvestigator), Anh T. Diep (Coinvestigator), ShanyanXue, Nianmin Zhou, Yong Liu, Evgenia Ter-Karapetova;University of Southern Maine (Portland, Maine): W. DouglasThompson (subcontract Principal Investigator); University ofTexas, M. D. Anderson Cancer Center (Houston, Texas):Marilyn Stovall (subcontract Principal Investigator), SusanSmith (Coinvestigator); and University of Virginia (Charlot-tesville, Virginia): Patrick Concannon (subcontract PrincipalInvestigator), Sharon Teraoka (Coinvestigator), Eric R. Olson(some work performed at Benaroya Research Institute at Vir-ginia Mason, Seattle, Washington), V. Anne Morrison, KarenM. Cerosaletti, Lemuel Navarro, Jocyndra Wright.


REFERENCES

1. Chen Y, Thompson W, Semenciw R, et al. Epidemiology ofcontralateral breast cancer. Cancer Epidemiol BiomarkersPrev. 1999;8(10):855–861.

2. Ellison RC, Zhang Y, McLennan CE, et al. Exploring the re-lation of alcohol consumption to risk of breast cancer. Am JEpidemiol. 2001;154(8):740–747.

3. Smith-Warner SA, Spiegelman D, Yaun SS, et al. Alcohol andbreast cancer in women: a pooled analysis of cohort studies.JAMA. 1998;297(7):535–540.

4. Collaborative Group on Hormonal Factors in Breast Cancer.Alcohol, tobacco and breast cancer—collaborative reanalysisof individual data from 53 epidemiological studies, including58515 women with breast cancer and 95067 women withoutthe disease. Br J Cancer. 2002;87(11):1234–1245.

5. Berstad P, Ma H, Bernstein L, et al. Alcohol intake and breastcancer among young women. Breast Cancer Res Treat. 2007;108(1):113–120.

6. Tjønneland A, Christensen J, Olsen A, et al. Alcohol intakeand breast cancer risk: the European Prospective Investigationinto Cancer and Nutrition (EPIC). Cancer Causes Control.2007;18(4):361–373.

7. Trentham-Dietz A, Newcomb PA, Nichols H, et al. Breastcancer risk factors and second primary malignancies amongwomen with breast cancer. Breast Cancer Res Treat. 2007;105(2):195–207.

8. Li CI, Malone KE, Porter PL, et al. Epidemiologic andmolecular risk factors for contralateral breast canceramong young women. Br J Cancer. 2003;89(3):513–518.

9. Bernstein JL, Thompson WD, Risch N, et al. Risk factorspredicting the incidence of second primary breast cancer

among women diagnosed with a first primary breast cancer.Am J Epidemiol. 1992;136(8):925–936.

10. Dumitrescu RG, Shields PG. The etiology of alcohol-inducedbreast cancer. Alcohol. 2005;35(3):213–225.

11. Terry PD, Rohan TE. Cigarette smoking and the risk of breastcancer in women: a review of the literature. Cancer EpidemiolBiomarkers Prev. 2002;11(10 pt 1):953–971.

12. Morabia A. Smoking (active and passive) and breast cancer:epidemiologic evidence up to June 2001. EnvironMol Mutagen.2002;39(2-3):89–95.

13. Reynolds P, Hurley S, Goldberg DE, et al. Active smoking,household passive smoking, and breast cancer: evidence fromthe California Teachers Study. J Natl Cancer Inst. 2004;96(1):29–37.

14. Ha M, Mabuchi K, Sigurdson AJ, et al. Smoking cigarettesbefore first childbirth and risk of breast cancer. Am J Epidemiol.2007;166(1):55–61.

15. Prescott J, Ma H, Bernstein L, et al. Cigarette smoking is notassociated with breast cancer risk in young women. CancerEpidemiol Biomarkers Prev. 2007;16(3):620–622.

16. Fowble B, Hanlon A, Freedman G, et al. Second cancersafter conservative surgery and radiation for stages I–IIbreast cancer: identifying a subset of women at increasedrisk. Int J Radiat Oncol Biol Phys. 2001;51(3):679–690.

17. Horn PL, Thompson WD. Risk of contralateral breast cancer:associations with factors related to initial breast cancer. Am JEpidemiol. 1988;128(2):309–323.

18. Bernstein JL, Langholz B, Haile RW, et al. Study design:evaluating gene–environment interactions in the etiology ofbreast cancer—theWECARE Study. Breast Cancer Res. 2004;6(3):R199–R214.

19. Largent J, Capanu M, Bernstein L, et al. Reproductive historyand risk of second primary breast cancer: the WECARE Study.Cancer Epidemiol Biomarkers Prev. 2007;16(5):906–911.

20. Langholz B, Borgan Ø. Counter-matching: a stratified nestedcase-control sampling method. Biometrika. 1995;82(1):69–79.

21. Cox DR. Regression models and life tables. J R Stat Soc (B).1972;34:187–202.

22. Borgan Ø, Goldstein L, Langholz B. Methods for the analysisof sampled cohort data in the Cox proportional hazards model.Ann Statist. 1995;23(5):1749–1778.

23. Stovall M, Smith SA, Langholz BM, et al. Dose to the con-tralateral breast from radiotherapy and risk of second primarybreast cancer in the WECARE Study. Int J Radiat Oncol BiolPhys. 2008;72(4):1021–1030.

24. Bertelsen L, Bernstein L, Olsen JH, et al. Effect of systemicadjuvant treatment on risk for contralateral breast cancer in theWomen’s Environment, Cancer and Radiation EpidemiologyStudy. J Natl Cancer Inst. 2008;100(1):32–40.

25. Althuis MD, Fergenbaum JH, Garcia-Closas M, et al. Etiologyof hormone receptor-defined breast cancer: a systematic re-view of the literature. Cancer Epidemiol Biomarkers Prev.2004;13(10):1558–1568.

968 Knight et al.

Am J Epidemiol 2009;169:962–968




Vol. 169, No. 8

DOI: 10.1093/aje/kwp018



Positive Associations Between Ionizing Radiation and Lymphoma MortalityAmong Men

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


The authors investigated the relation between ionizing radiation and lymphoma mortality in 2 cohorts: 1) 20,940men in the Life Span Study, a study of Japanese atomic bomb survivors who were aged 15–64 years at the time ofthe bombings of Hiroshima and Nagasaki, and 2) 15,264 male nuclear weapons workers who were hired at theSavannah River Site in South Carolina between 1950 and 1986. Radiation dose-mortality trends were evaluatedfor all malignant lymphomas and for non-Hodgkin’s lymphoma. Positive associations between lymphoma mortalityand radiation dose under a 5-year lag assumption were observed in both cohorts (excess relative rates per sievertwere 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. Thesefindings 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, likelihoodratio test; LSS, Life Span Study; ND, not determined; NHL, non-Hodgkin’s lymphoma; SRS, Savannah River Site.

Ionizing radiation has been considered as a cause of lym-phoma by a number of investigators. In a review of thisliterature, Boice (1) concluded that the evidence of associ-ation between ionizing radiation and non-Hodgkin’s lym-phoma (NHL) is extremely weak and that there is noevidence of association between radiation and Hodgkin’sdisease. The United Nations Scientific Committee on theEffects of Atomic Radiation noted that studies of NHL fol-lowing external exposure to ionizing radiation have yieldedmixed results and concluded that overall there is little evi-dence of an association between NHL and external exposureto ionizing radiation (2). Ron (3) reached a similar conclu-sion, noting that evidence of association between radiationand NHL has been inconsistent and Hodgkin’s disease hasrarely been related to radiation exposure; and Melbye andTrichopoulos (4) stated that there is no evidence that ioniz-ing radiation causes NHL. However, this conclusion is not

universally shared. Hartge et al. argued that the evidencesuggests that ionizing radiation probably causes lymphoma(5) and observed that high doses of ionizing radiation appearto be associated with lymphoma risk in some studies ofradiotherapy (6).

Lack of a consistent association between ionizing radia-tion and lymphoma could mean that there is no causal re-lation or that a causal relation is obscured by bias ordeficiencies in exposure measurement, case classification,duration of follow-up, or some combination of these factors.Given that lymphoma is often an indolent disease, long-termstudies of radiation-exposed populations may be needed toobserve an effect. The development of nuclear weapons inthe early 1940s led to 2 types of epidemiologic studies thatcan now provide evidence regarding the radiation-lymphoma association: studies based on follow-up of work-ers exposed to ionizing radiation during nuclear weapons

Correspondence to Dr. David B. Richardson, Department of Epidemiology, School of Public Health, University of North Carolina at Chapel Hill,

Chapel Hill, NC 27599 (e-mail: [email protected]).

969 Am J Epidemiol 2009;169:969–976

production and studies based on follow-up of people ex-posed to ionizing radiation from the use of nuclear weapons.Most prominent among the latter is the Life Span Study(LSS), a study of Japanese survivors of the atomic bombingsof Hiroshima and Nagasaki. Radiation risk estimates fromstudies of nuclear workers are often compared with esti-mates from the LSS in order to evaluate the consistency ofrisk estimates in a population that includes people exposedto acute high doses with estimates from populations that arechronically exposed to low doses (7–9).

We examined the association between ionizing radiationand lymphoma mortality in a US occupational cohort and ina sample of LSS atomic bomb survivors and compared find-ings from the 2 populations. Follow-up of each cohort com-menced in 1950 and spanned approximately 5 decades. Tothe extent possible, we conducted these analyses as parallelanalyses employing comparable methods. We focused, inparticular, on variation in the associations between radiationdose and lymphoma mortality by time since exposure.


The LSS cohort includes 86,611 people who were alive atthe time of the 1950 Japanese census, reported being inHiroshima or Nagasaki at the time of the bombings (August1945), and had dose estimates based on the DS02 dosimetrysystem (10). Follow-up for ascertainment of vital status andcause-of-death information started on October 1, 1950, andcontinued until December 31, 2000.

The Savannah River Site (SRS) was constructed nearAiken, South Carolina, in 1950 as a facility to producematerials for the US nuclear weapons program. A cohortof 18,883 workers who were hired at the SRS prior to1987, who worked there for at least 90 days, who werenot known to have been employed at another US Depart-ment of Energy facility, and who had complete informationon name, Social Security number, sex, date of birth, and dateof hire was enumerated (11). Vital status and cause-of-deathinformation were ascertained through December 31, 2002.

Cohort restrictions for comparability

Since over 95% of the collective dose at SRS was in-curred by males, there was little ability to estimate riskdue to radiation exposure among female SRS workers. Wetherefore restricted the analyses to males in both cohorts.Since the youngest age at hire at SRS was 15 years and mostSRS workers terminated their employment by age 65 years,LSS analyses were restricted to people who were aged 15–64 years at the time of the bombings. This resulted in a co-hort of 15,264 male SRS workers and a cohort of 20,940male LSS subjects who were aged 15–64 years at the time ofthe bombings.

Dosimetry data

For the LSS, we used DS02 revised colon dose estimatesadjusted for dosimetry errors, with shielded kerma estimates

above 4 Gy truncated to 4 Gy (12). For consistency withanalyses of the SRS cohort, dose estimates calculated as thesum of the c-radiation dose plus 10 times the neutron doseare expressed in sieverts; some recent reports on LSS ana-lyses refer to this quantity as the weighted dose in grays (13,14). Interactions between radiation and lymphocytes mayoccur in the lymphatic or circulatory system at a varietyof anatomic sites; the choice of target organ for dose esti-mation may depend on the characteristics of the lymphoma,including anatomic location (15, 16). The colon dose hasbeen taken as a representative dose to the organs involved ata variety of anatomic locations, similar to the approachemployed in prior analyses of solid cancers (17). The colondose estimate has been used by previous investigators as anestimate comparable to the quantity estimated by the radi-ation dosimeters worn by nuclear industry workers (i.e., the‘‘deep dose’’).

For SRS workers, the exposure of interest was defined ascumulative whole-body radiation dose equivalent from ex-ternal sources and tritium received during employment atthe site, expressed in sieverts; neutron doses were multipliedby a factor of 10. Personal radiation monitoring data wereavailable for the period 1950–1999. Whole-body doses wereestimated for work-years with missing dose data using doseestimates from adjacent time periods and average values forsimilar workers; estimated annual doses constituted 4% ofemployment years for male workers (18).

Outcome definitions

In the LSS, underlying cause of death was coded accord-ing to the International Classification of Diseases, NinthRevision (ICD-9), which was issued in 1977. In the SRSstudy, underlying cause of death was coded according to theEighth Revision of the ICD (ICD-8) for deaths occurringprior to 1979 and according to the ICD revision in effect atthe time of death for deaths occurring in 1979 or later. (TheTenth Revision of the ICD (ICD-10) was issued in 1992.)

As in prior analyses (17, 19), we examined the broadcategory of malignant lymphoma (ICD-8 and ICD-9 codes200–202; ICD-10 codes C81–C85). In addition, we exam-ined the subcategory of NHL (ICD-8 and ICD-9 codes 200and 202; ICD-10 codes C82–C85). There were too fewdeaths due to Hodgkin’s disease to support separate analysesof that outcome in these cohorts.

Statistical methods

Poisson regression methods were used. The analyticaldata file for the LSS cohort consisted of a tabulation ofperson-time and numbers of deaths by city, age at exposure(in 5-year intervals), attained age (in 5-year intervals), cal-endar time (1950–1952, 1953–1955, and then 5-year inter-vals up to 1995, 1996–1997, and 1998–2000), and dose(<0.005, 0.005–<0.02, 0.02–<0.04, 0.04–<0.06, 0.06–<0.08, 0.08–<0.1, 0.1–<0.125, 0.125–<0.150, 0.150–<0.175, 0.175–<0.2, 0.2–<0.25, 0.25–<0.3, 0.3–<0.5,0.5–<0.75, 0.75–<1, 1–<1.25, 1.25–<1.5, 1.5–<1.75,1.75–<2, 2–<2.5, 2.5–<3, and �3 Sv). The analytical data

970 Richardson et al.

Am J Epidemiol 2009;169:969–976

file for the SRS cohort consisted of a tabulation of person-time and events by attained age (in 5-year intervals), race(black vs. other), year of birth (before 1915, 1915–1924,1925–1929, 1930–1934, 1935–1949, or 1950 or later), paycode (paid monthly, weekly, or hourly), employment status(employed, terminated within the last 2 years, or terminatedmore than 2 years prior, classified separately for riskages <62 years and �62 years) (20–22), and dose(0,>0–<0.005, 0.005–<0.02, 0.02–<0.04, 0.04–<0.06, 0.06–<0.08, 0.08–<0.1, 0.1–<0.125, 0.125–<0.150, 0.150–<0.175, 0.175–<0.2, 0.2–<0.25, 0.25–<0.3, and �0.3 Sv).

Covariate control was achieved through background strat-ification of regression models. In analyses of the LSS co-hort, the stratifying factors were attained age, age atexposure, and city; in analyses of the SRS cohort, the strat-ifying factors were attained age, birth cohort, race, pay code,and employment status. Radiation dose-mortality associa-tions were estimated via a regression model of the form

rate ¼ eaið1 þ b xÞ;

where ai indexes the stratum-specific mortality rate in theabsence of radiation exposure and b provides an estimate ofthe excess relative rate (ERR) per sievert (23, 24).

In analyses of the LSS cohort, x represents the estimatedradiation dose delivered at the time of the bombings inAugust 1945. Since follow-up of the LSS cohort began inOctober 1950, this implies a minimal lag of approximately5 years between exposure and its effect. We also present re-sults from analyses in which we assumed that there was noexcess risk during the period 1950–1955; that is, a minimumlatency period of approximately 10 years was assumed. A10-year lag assumption has been used in previous nuclearworker studies that examined lymphoma mortality (25, 26).We refer to analyses of LSS data that examine excess mor-tality risk since 1950 and since 1956 as analyses carried outunder 5- and 10-year lag assumptions, respectively. In anal-yses of the SRS cohort, x represents the cumulative radiationdose under a 5- or 10-year lag assumption. Lagging doseassignment by L years means that an increment of dose wasincluded in the calculation of cumulative dose at time t if ithad been received at or before time t � L years; person-timeand events at time t were then classified according to thatcategory of lagged cumulative dose.

The dose range in the LSS, 0–4 Sv, was wider than the doserange in the SRS study (0–<0.5 Sv). In order to evaluatedose-response associations over a comparable range of doses,we also conducted analyses based upon LSS data limited tothe 19,183 survivors with doses in the range of 0–<0.5 Sv.

In analyses of the LSS cohort, we assessed variation inradiation risk with time since exposure via a regressionmodel of the form

rate ¼ eaið1 þ b1xPeriod1 þ b2xPeriod2

þ b3xPeriod3 þ b4xPeriod4Þ;

where Period1–Period4 are indicator variables for the cal-endar time periods 1950–1970, 1971–1980, 1981–1990, and1991–2000, respectively. The values b1; b2; b3; and b4 pro-

vide estimates of the ERR per 1-Sv dose during the periods5–25, 26–35, 36–45, and 46–55 years after the bombings. Inanalyses of the SRS cohort, we fitted a model of the form

rate ¼ edið1 þ /1d1 þ /2d2 þ /3d3Þ;

where d1–d3 represent the cumulative radiation doses ac-crued in the exposure time windows 5–25, 26–35, and�36 years prior to observation of a person-year orevent and /1; /2; and /3 provide associated estimates ofthe ERR per 1-Sv dose.

We estimated parameters using the EPICURE statisticalpackage (Hirosoft International Corporation, Seattle,Washington); for consistency with recent reports (2, 26),we generated 90% confidence intervals for estimated param-eters via the likelihood method (27). In some analyses, con-fidence bounds could not be determined (designated ‘‘notdetermined’’ (ND)). In order to aid interpretation of modelfittings, we report the 1-sided P value derived via a likeli-hood ratio test (LRT) for each reported point estimate.Tabulations of observed versus expected numbers of deathsby category of cumulative dose are reported; we calculatedexpected counts for each cell of the person-time table usinga regression model that included all variables except thedose term.

RESULTS

With follow-up through 2000, 90 malignant lymphomadeaths were observed among the male atomic bomb sur-vivors exposed at ages 15–64 years, including 6 deaths fromHodgkin’s disease (Table 1). Sixty-three malignant lym-phoma deaths occurred among residents of Hiroshima (58due to NHL) and 27 malignant lymphoma deaths occurredamong residents of Nagasaki (26 due to NHL). No deathsdue to malignant lymphoma occurred among survivors atattained ages less than 30 years. In the SRS cohort, 56lymphoma deaths were observed; 5 of these deaths weredue to Hodgkin’s disease. One death due to malignant lym-phoma was observed among black males (it was a case ofNHL), and 18, 14, and 24 deaths due to malignant lym-phoma were observed among workers paid monthly,weekly, and hourly, respectively. Three deaths due to ma-lignant lymphoma occurred among actively employed SRSworkers (all were cases of NHL) and 6 deaths occurredwithin 2 years of termination of employment (all were casesof NHL), while the remaining 47 deaths due to malignantlymphoma occurred 2 or more years after termination ofemployment at SRS (42 due to NHL).

In the LSS, the estimated ERR of malignant lymphomaper sievert, under a 5-year lag assumption, was 0.79 (90%confidence interval (CI): 0.10, 1.88). The goodness of modelfit was slightly improved, and the magnitude of associationwas slightly increased, upon exclusion of deaths due toHodgkin’s disease (Table 2). Under a 10-year lag assump-tion, these estimated associations were slightly larger inmagnitude. In the SRS study, the estimated ERRs of malig-nant lymphoma per sievert under 5- and 10-year lag assump-tions were 6.99 (90% CI: 0.96, 18.39) and 8.18 (90% CI:1.44, 21.16), respectively. Upon exclusion of deaths due to

Ionizing Radiation and Lymphoma Mortality 971

Am J Epidemiol 2009;169:969–976

Hodgkin’s disease, these estimated associations wereslightly smaller in magnitude. The SRS cohort includeda single death due to malignant lymphoma among blackworkers; upon restriction to nonblack workers, the esti-mated ERRs of malignant lymphoma per sievert under5- 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 dosesin the range of 0–<0.5 Sv, estimates of radiation-lymphomamortality associations were of greater magnitude than esti-mates obtained from model fittings over the entire doserange. Under a 5-year lag assumption, the estimated ERRsof malignant lymphoma and NHL per sievert were 3.02(90% CI: 0.33, 7.22) and 2.86 (90% CI: 0.10, 7.24), respec-tively. While this suggests nonlinearity in the dose-responseassociation, comparison of a linear-quadratic dose-responsefunction with a purely linear dose-response function indi-cated that inclusion of a quadratic term resulted in very littleimprovement in model fit (LRT ¼ 0.07, 1 df; P ¼ 0.79).Under a 10-year lag assumption, the estimated ERRs ofmalignant lymphoma and NHL per sievert were 4.54(90% CI: 1.16, 9.93) and 4.24 (90% CI: 0.83, 9.76),respectively.

In the LSS, there was no evidence of an association be-tween radiation dose and lymphoma mortality during theperiods 5–25 years or 26–35 years after irradiation (Table 3).Positive associations between lymphoma mortality anddose were observed during the periods 36–45 years and 46–55 years after irradiation. Analyses of associations betweenradiation dose and NHL led to risk estimates similar to thoseobtained via analyses of all malignant lymphoma (Table 3).In a nested model, defined post hoc, we evaluated the asso-

ciation between dose and malignant lymphoma mortalityduring the periods 5–35 years postexposure and 36–55 yearspostexposure. There was no evidence of association 5–35years after exposure (ERR/Sv ¼ 0.03, 90% CI: ND, 1.15;LRT ¼ 0.00, P ¼ 0.96); however, there was a positive

Table 2. Estimated Association Between Lymphoma Mortality and

Ionizing Radiation Dose Under 5- and 10-Year Exposure Lags

Among Male Atomic Bomb Survivors (1950–2000) and Male

Workers at the Savannah River Site (1950–2002), Japan and South

Carolina

ExposureLag andERR

Atomic BombSurvivorsa

Savannah RiverSite Workers

MalignantLymphoma

Non-Hodgkin’sLymphoma

MalignantLymphoma


5 years

ERR perSv

0.79 0.86 6.99 6.45

90% CI 0.10, 1.88 0.13, 2.03 0.96, 18.39 0.48, 17.95

P valueb 0.05 0.04 0.04 0.07

10 years

ERR perSv

1.06 1.12 8.18 7.62

90% CI 0.24, 2.38 0.26, 2.51 1.44, 21.16 0.93, 20.77

P value 0.02 0.02 0.03 0.05

Abbreviations: CI, confidence interval; ERR, excess relative rate.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.

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

AttainedAge, years

Atomic BombSurvivorsb

Savannah RiverSite Workers

Person-Yearsof Follow-Up

No. of DeathsPerson-Yearsof Follow-Up

No. of Deaths

MalignantLymphoma


MalignantLymphoma


<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.


Am J Epidemiol 2009;169:969–976

association between dose and lymphoma mortality �36years after exposure (ERR/Sv ¼ 1.93, 90% CI: 0.48, 4.66;LRT ¼ 6.83, P < 0.01).

In analyses of the SRS cohort, there was a highly impre-cise positive association between lymphoma mortality anddoses accrued during the periods 5–25 and 26–35 yearsprior. The association with doses accrued �36 years priorwas of the largest magnitude and contributed most to thegoodness of model fit. The estimated dose-response associ-ation within each exposure time window was based on thetotal number of lymphoma deaths. Similar estimates wereobtained in analyses restricted to NHL (Table 4).

When the LSS data were limited to those survivors withdoses in the range of 0–<0.5 Sv, there were positive, albeitimprecise, estimates of association between radiation doseand malignant lymphoma mortality during the periods 5–25years 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(ERR/Sv ¼ 2.52, 90% CI: �1.48, 11.71; LRT ¼ 0.7, P ¼0.40), 36–45 years (ERR/Sv ¼ 7.08, 90% CI: �0.08, 22.86;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 ¼0.12). Results for analyses of NHL were similar to those forall lymphoma mortality. There was a negative associationbetween radiation dose and NHL mortality during the period5–25 years after irradiation (ERR/Sv ¼�0.41, 90% CI: ND,5.00; LRT ¼ 0.03, P ¼ 0.85) and positive associations be-tween radiation dose and mortality during the periods 26–35 years after irradiation (ERR/Sv ¼ 2.46, 90% CI: �1.50,11.55; LRT ¼ 0.68, P ¼ 0.41), 36–45 years after irradia-tion (ERR/Sv ¼ 7.07, 90% CI: �0.08, 22.83; LRT ¼ 2.61,P ¼ 0.11), and 46–55 years after irradiation (ERR/Sv ¼6.42, 90% CI: �0.23, 23.11; LRT ¼ 2.41, P ¼ 0.12).

Table 5 shows observed and expected numbers of malig-nant lymphoma deaths by dose category under 5- and10-year lag assumptions. The distribution of events amongSRS workers with respect to dose was relatively narrow incomparison with the LSS data. Over the dose range at whichthe ratio of observed to expected numbers of malignantlymphoma deaths could be compared in these 2 cohorts(i.e., 0–<0.5 Sv), these ratios were similar in magnitudefor analyses of the 2 cohorts, although values tended to beslightly greater for the SRS cohort than for the LSS cohort.Ratios of observed to expected numbers of deaths were

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

Lymphoma Typeand ERR

Time Since Exposure, years(Calendar Period)

5–25(1950–1970)

26–35(1971–1980)

36–45(1981–1990)

46–55(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.

Table 4. Estimated Association Between Radiation Dose and

Lymphoma Mortality Among Male Workers at the Savannah River

Site, by Time Since Exposure, South Carolina, 1950–2002

Lymphoma Typeand ERR

Time Since Exposure, years

5–25 26–35 36–52

Malignant lymphoma

ERR per Sv 1.18 4.06 33.28

90% CI ND, ND ND, 25.34 4.83, 107.9

P valuea 0.85 0.64 0.03

Non-Hodgkin’s lymphoma

ERR per Sv 1.51 0.58 38.35

90% CI ND, 16.02 ND, 22.83 7.02, 121.57

P value 0.80 0.95 0.02

Abbreviations: CI, confidence interval; ERR, excess relative rate;

ND, not determined.a P value from a likelihood ratio test that the reported parameter for

the estimated ERR was equal to 0.


Am J Epidemiol 2009;169:969–976

minimally affected by exclusion of deaths due to Hodgkin’sdisease (results not shown).

DISCUSSION

In a previous analysis of lymphoma mortality among sur-vivors in the LSS, Pierce et al. (17) reported evidence ofa nonsignificant positive association with radiation doseamong males (ERR/Sv ¼ 0.27, 90% CI: ND, 1.49) anda nonsignificant negative association among females(ERR/Sv ¼ �0.17, 90% CI: ND, 0.30). In those analyses,a time-constant ERR model was fitted to mortality follow-up through 1990. In the present paper, time-window ana-lyses helped to explain the observation of a significant positiveassociation between radiation dose and lymphoma mortalityamong male atomic bomb survivors with more recent fol-low-up, showing that positive associations have been ob-served only since 1980. Such findings suggest a protractedinduction and latency period. If considered within theframework of a multistage model of carcinogenesis, therelatively long empirical induction period for lymphoma

following radiation exposure may be consistent with actionat an early stage of a multistage process.

The point estimates for the radiation dose-lymphomamortality association under 5- and 10-year lag assumptionsderived from analysis of the SRS cohort are larger than theestimates derived from analysis of the LSS cohort (Table 2).Differences in the magnitude and rate of exposure may in-fluence the comparability of dose-response estimates. Thesecohorts also differ with regard to potential biases from con-founding, selection, and exposure measurement error. Whileit is not an established cause of NHL, benzene is suspectedto be related to NHL (28). However, benzene was not usedin the production process at SRS, nor was it routinely usedas a degreaser. Plutonium-239 is a radiologic hazard at SRS.While a recent study suggested that the contribution of plu-tonium doses to total dose estimates for these workers wasrelatively small (29), we did not directly assess confoundingby plutonium exposure. Selection bias could have influ-enced these estimates of association—for example, via the‘‘healthy worker’’ survivor effect (20). Although we ad-justed for employment status, such an approach is sub-optimal if employment status is an intermediate variable

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

Assumed Lag and CohortRadiation Dose, Sv

<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

Obs/Exp ratioc 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


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


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


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.


Am J Epidemiol 2009;169:969–976

as well as a confounder of the association of interest. How-ever, in studies of chronic diseases with long latency peri-ods, cumulative exposure will typically not appreciablyinfluence employment termination rates; under such condi-tions, employment status will play a minor role as anintermediate variable but could have a strong role as a con-founder of the association (22). Frequent reading of dosim-eters could have led to dose underestimation if dosimeterswere not sufficiently exposed to reach a minimum detect-able dose. However, prior work suggests that the impact ofthis source of measurement error on estimates of radiationdose-response trends is modest (30–32).

Problems of bias could also influence estimates ofradiation-mortality associations among atomic bomb survi-vors. DS02 estimates account for the initial radiationreleased from the detonation of the weapons but not radia-tion from fallout or neutron activation of the ground andstructures (33). The available data suggest that most peoplein Hiroshima and Nagasaki had low cumulative externaldoses from fallout, with maximum estimates in the rangeof 0.2–0.4 Sv for several hundred people who were in anarea of Nagasaki approximately 3 km from the hypocenter(33, 34). Selective survival in the LSS cohort is anotherconcern and is a generic consideration when trying to un-derstand the temporal evolution of exposure-related risk(35). A relation between short-term survival after the bomb-ings and later risk of lymphoma could lead to bias indose-response estimates. Evidence of selection has beensuggested by some empirical analyses (36, 37); however,values for the magnitude of dose-related selective survivalassumed in a recent study suggested a modest potential forbias in dose-response estimates (38).

These analyses provide evidence of a positive associationbetween ionizing radiation dose and malignant lymphomamortality among male Japanese atomic bomb survivors andSRS workers. We did not address risk estimates for females,for whom there was no evidence of a positive associationbetween radiation dose and lymphoma mortality in follow-up through 1990 (17). The radiation-NHL mortality associ-ations among these male atomic bomb survivors and SRSworkers are of larger magnitude than the estimate reportedin a 15-country study of nuclear workers (under a 10-yearlag assumption, ERR/Sv ¼ 0.44, 90% CI: <0, 4.78) (7);however, in the current analyses, positive dose-response as-sociations were primarily observed more than 35 years afterirradiation. These findings underscore the importance ofcontinued follow-up of the LSS cohort and nuclear workercohorts.

ACKNOWLEDGMENTS

Author affiliations: Department of Epidemiology, Schoolof Public Health, University of North Carolina at ChapelHill, Chapel Hill, North Carolina (David Richardson, SteveWing); Department of Epidemiology, Radiation Effects Re-search Foundation, Hiroshima, Japan (Hiromi Sugiyama,Ritsu Sakata, Eric J. Grant, Yukiko Shimizu, Nobuo Nishi,Fumiyoshi Kasagi); Department of Statistics, Radiation

Effects Research Foundation, Hiroshima, Japan (Susan Geyer);Department of Epidemiology, Radiation Effects ResearchFoundation, Nagasaki, Japan (Midori Soda, Akihiko Suyama);and Radiation Effects Research Foundation, Hiroshima, Japan(Kazunori Kodama).

This project was supported in part by grant R01OH007871 from the US National Institute for OccupationalSafety and Health. Support was also provided by the Radi-ation Effects Research Foundation (Hiroshima and Naga-saki, Japan), which is funded by the Japan Ministry ofHealth, Labour and Welfare and the US Department of En-ergy, the latter partly through the National Academy ofSciences.


REFERENCES

1. Boice JD Jr. Radiation and non-Hodgkin’s lymphoma. CancerRes. 1992;52(19 suppl):5489s–5491s.

2. United Nations Scientific Committee on the Effects of AtomicRadiation. Sources and Effects of Ionizing Radiation. NewYork, NY: United Nations; 2000.

3. Ron E. Ionizing radiation and cancer risk: evidence fromepidemiology. Radiat Res. 1998;150(5 suppl):S30–S41.

4. Melbye M, Trichopoulos D. Non-Hodgkin’s lymphomas. In:Adami H-O, Hunter D, Trichopoulos D, eds. Textbook ofCancer Epidemiology. New York, NY: Oxford UniversityPress; 2002:535–555.

5. Hartge P, Smith MT. Environmental and behavioral factors andthe risk of non-Hodgkin lymphoma. Cancer Epidemiol Bio-markers Prev. 2007;16(3):367–368.

6. Hartge P, Wang SS, Bracci PM, et al. Non-Hodgkin lym-phoma. In: Schottenfeld D, Fraumeni JF, eds. Cancer Epide-miology and Prevention. New York, NY: Oxford UniversityPress; 2006:898–918.

7. Cardis E, Vrijheid M, Blettner M, et al. Risk of cancer afterlow doses of ionising radiation: retrospective cohort study in15 countries. BMJ. 2005;331(7508):77.

8. Muirhead CR, Goodill AA, Haylock RG, et al. Occupationalradiation exposure and mortality: second analysis of the Na-tional Registry for Radiation Workers. J Radiol Prot. 1999;19(1):3–26.

9. Schubauer-Berigan M, Daniels RD, Fleming DA, et al. Risk ofchronic myeloid and acute leukemia mortality after exposureto ionizing radiation among workers at four U.S. nuclearweapons facilities and a nuclear naval shipyard. Radiat Res.2007;167(2):222–232.

10. Preston DL, Pierce DA, Shimizu Y, et al. Effect of recentchanges in atomic bomb survivor dosimetry on cancermortality risk estimates. Radiat Res. 2004;162(4):377–389.

11. Richardson DB, Wing S. Leukemia mortality among workersat the Savannah River Site. Am J Epidemiol. 2007;166(9):1015–1022.

12. Pierce DA, Stram DO, Vaeth M. Allowing for random errors inradiation dose estimates for the atomic bomb survivor data.Radiat Res. 1990;123(3):275–284.

13. Preston DL, Ron E, Tokuoka S, et al. Solid cancer incidence inatomic bomb survivors: 1958–1998. Radiat Res. 2007;168(1):1–64.

14. International Commission on Radiological Protection. 1990Recommendations of the International Commission on


Am J Epidemiol 2009;169:969–976

Radiological Protection. Oxford, United Kingdom: PergamonPress; 1991.

15. National Institute for Occupational Safety and Health.Changes to the dose reconstruction target organ selection forlymphoma under the Energy Employees Occupational IllnessCompensation Program Act of 2000. Fed Regist. 2006;71(31):7969–7970.

16. Office of Compensation Analysis and Support, National In-stitute for Occupational Safety and Health. Selection for In-ternal and External Dosimetry Target Organs for Lymphatic/Hematopoietic Cancers. (Technical Information Bulletin).Cincinnati, OH: National Institute for Occupational Safety andHealth; 2006. (Document no. OCAS-TIB-0012).

17. Pierce DA, Shimizu Y, Preston DL, et al. Studies of the mor-tality of atomic bomb survivors. Report 12, part I. Cancer:1950–1990. Radiat Res. 1996;146(1):1–27.

18. Richardson DB, Wing S, Daniels RD. Evaluation of externalradiation dosimetry records at the Savannah River Site,1951–1989. J Expo Sci Environ Epidemiol. 2007;17(1):13–24.

19. Shimizu Y, Kato H, Schull WJ. Studies of the mortality ofA-bomb survivors. 9. Mortality, 1950–1985: part 2. Cancermortality based on the recently revised doses (DS86). RadiatRes. 1990;121(2):120–141.

20. Arrighi HM, Hertz-Picciotto I. The evolving concept of thehealthy worker survivor effect. Epidemiology. 1994;5(2):189–196.

21. Steenland K, Stayner L. The importance of employment statusin occupational cohort mortality studies. Epidemiology. 1991;2(6):418–423.

22. Steenland K, Deddens J, Salvan A, et al. Negative bias inexposure-response trends in occupational studies: modelingthe healthy workers survivor effect. Am J Epidemiol. 1996;143(2):202–210.

23. Thomas D. General relative-risk models for survival time andmatched case-control analysis. Biometrics. 1981;37(4):673–686.

24. Greenland S. Introduction to regression models. In:Rothman K, Greenland S, eds. Modern Epidemiology. 2nd ed.Philadelphia, PA: Lippincott, Williams & Wilkins; 1998:359–399.

25. Gilbert ES, Cragle DL, Wiggs LD. Updated analyses ofcombined mortality data for workers at the Hanford Site, OakRidge National Laboratory, and Rocky Flats Weapons Plant.Radiat Res. 1993;136(3):408–421.

26. Cardis E, Vrijheid M, Blettner M, et al. The 15-CountryCollaborative Study of Cancer Risk among Radiation Workersin the Nuclear Industry: estimates of radiation related cancerrisks. Radiat Res. 2007;167(4):396–416.

27. Preston DL, Lubin JH, Pierce DA, et al. Epicure: User’sGuide. Seattle, WA: Hirosoft International Corporation; 1993.

28. Smith MT, Jones RM, Smith AH. Benzene exposure and riskof non-Hodgkin lymphoma. Cancer Epidemiol BiomarkersPrev. 2007;16(3):385–391.

29. Daniels RD, Lodwick CJ, Schubauer-Berigan MK, et al.Assessment of plutonium exposures for an epidemiologicalstudy of US nuclear workers. Radiat Prot Dosimetry. 2006;118(1):43–55.

30. Daniels RD, Yiin JH. A comparison of statistical methods forestimation of less than detectable ionising radiation exposures.Radiat Prot Dosimetry. 2006;121(3):240–51.

31. Daniels RD, Schubauer-Berigan MK. Bias and uncertainty ofpenetrating photon dose measured by film dosimeters in anepidemiological study of US nuclear workers. Radiat ProtDosimetry. 2005;113(3):275–289.

32. Shin H, Ramsay T, Krewski D, et al. The effect of censoring oncancer risk estimates based on the Canadian National DoseRegistry of occupational radiation exposure. J Expo AnalEnviron Epidemiol. 2005;15(5):398–406.

33. Committee for the Compilation of Materials on DamageCaused by the Atomic Bombs in Hiroshima and Nagasaki.Hiroshima and Nagasaki: The Physical, Medical, and SocialEffects of the Atomic Bombings. Tokyo, Japan: IwanamiShoten, Publishers; 1981.

34. Radiation Effects Research Foundation. Frequently AskedQuestions. Hiroshima, Japan: Radiation Effects ResearchFoundation; 2007. (http://www.rerf.or.jp/general/qa_e/qa12.html). (Accessed June 30, 2008).

35. Flanders WD, Klein M. Properties of 2 counterfactual effectdefinitions of a point exposure. Epidemiology. 2007;18(4):453–460.

36. Stewart AM, Kneale GW. A-bomb survivors: factors that maylead to a re-assessment of the radiation hazard. Int J Epide-miol. 2000;29(4):708–714.

37. Shimizu Y, Pierce DA, Preston DL, et al. Studies of the mor-tality of atomic bomb survivors. Report 12, part II. Noncancermortality: 1950–1990. Radiat Res. 1999;152(4):374–389.

38. Pierce DA, Vaeth M, Shimizu Y. Selection bias in cancer riskestimation from A-bomb survivors. Radiat Res. 2007;167(6):735–741.


Am J Epidemiol 2009;169:969–976

http://www.rerf.or.jp/general/qa_e/qa12.html

http://www.rerf.or.jp/general/qa_e/qa12.html




Vol. 169, No. 8

DOI: 10.1093/aje/kwp008



Biomarker-calibrated Energy and Protein Consumption and Increased CancerRisk Among Postmenopausal Women

Ross L. Prentice, Pamela A. Shaw, Sheila A. Bingham, Shirley A. A. Beresford, Bette Caan, MarianL. 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 theWomen’s HealthInitiative, that produce calibrated estimates of energy, protein, and percentage of energy from protein consumptionfrom 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 womenenrolled in the Women’s Health Initiative dietary modification trial comparison group and 59,105 women enrolled inthe observational study. These estimates were related prospectively to total and site-specific invasive cancerincidence (1993–2005). In combined cohort analyses that do not control for body mass, uncalibrated energy was notassociatedwith total cancer incidence or site-specific cancer incidence formost sites, whereas biomarker-calibratedenergywaspositively associatedwith 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 of1.24, 1.35, 1.83, and 1.47). Calibrated protein was weakly associated, and calibrated percentage of energy fromproteinwas inversely associated, with total cancer.Calibrated energy and bodymass index associationswere highlyinterdependent. 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 positiveassociation between energy consumption and the incidenceand mortality from cancer. Among women, associationswere reported for breast, colon, rectal, endometrial, ovarian,and kidney cancer (1). Rodent feeding experiments indicatethat underfeeding typically inhibits the development of site-specific and overall cancer (2, 3).

Analytical epidemiologic studies of diet, nutrition, andcancer date to the 1970s. Initial case-control studies useda range of dietary assessment procedures, including foodrecords, recalls, and frequencies. Concern about dietary re-call bias subsequently led to cohort studies as the predom-inant design for dietary association studies. Because these

studies typically involve tens of thousands of enrollees,a self-administered, machine-readable food frequency ques-tionnaire (FFQ) has been the principal dietary assessmenttool in cohort studies.

However, like other dietary assessment methods, the mea-surement properties of FFQs remain substantially unknown.Comparison of FFQ assessments with food records revealsnoteworthy differences (4) that imply an important errorcomponent to self-reported nutrient intake. Small-scalestudies using a doubly labeled water biomarker (5) of en-ergy consumption suggest important systematic biases also,as obese persons may systematically underreport energyconsumption (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: [email protected]).

977 Am J Epidemiol 2009;169:977–989

especially systematic biases, may substantially distort dietand cancer associations. It is important to examine nutrientand disease associations in a manner that appropriately ac-commodates FFQ measurement errors.

The accumulated data on diet and cancer were reviewedby an international panel of experts in 1997 (7). Rather few‘‘definite’’ or ‘‘probable’’ dietary associations emerged. Theauthors wrote, ‘‘The significance of the data on energy in-take and cancer risk in humans remains unclear’’ (7, p. 371),and ‘‘In the view of the panel, the effect of energy intake oncancer is best assessed by examining the data on relatedfactors: rate of growth, body mass, and physical activity’’(7, p. 371). This state of affairs has evidently not changed inthe intervening decade (8) and reflects considerable uncer-tainty about energy consumption estimates and related as-sociation study findings. The 1997 panel also assessed that

protein consumption was not ‘‘probably or convincingly’’related to the risk of any cancer (7, p. 394).

Good-quality biomarkers of both total energy consump-tion (5) and protein consumption (9) have been developedbut, for cost and logistics reasons, have received little use inepidemiologic research. These biomarkers involve urinaryrecovery of metabolites produced when these nutrients areexpended. In weight-stable persons, they provide objectiveestimates of short-term energy and protein consumption.The associated measurement error plausibly adheres toa simple classical measurement model,

W ¼ Zþ e; ð1Þ

where Z is the targeted (log-transformed) nutrient consump-tion, W is the (log-transformed) biomarker-measured

Table 1. Subject Characteristics for Women in the Women’s Health Initiative Dietary

Modification Trial Comparison Group and Observational Study, 1993–1998

Characteristic

Dietary ModificationTrial Comparison

Group (n 5 21,711)a

ObservationalStudy

(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

Body mass index, kg/m2

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

Otherb 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

978 Prentice et al.

Am J Epidemiol 2009;169:977–989

consumption, and e is measurement error that is assumed tobe independent of Z and of all other study subject character-istics. The cost to ascertain these biomarkers for each par-ticipant in a cohort study would be excessive. Instead,a substudy that includes both the biomarker and FFQ canbe used to produce calibrated consumption estimates for allcohort members.

The measurement model for the self-reported data typi-cally needs to be more complex than the classical measure-ment model (equation 1). Other factors, such as body mass,ethnicity, and age, may affect the assessment, and measure-ment errors may be correlated if the assessment is repeatedfor specific study subjects. Hence, we consider the measure-ment model (10, 11),

Q¼ S0 þ S1Zþ S2V þ S3VZþ rþ u; ð2Þ

for the (log-transformed) self-reported nutrient assessment

Q, where V is a set of characteristics that may relate tosystematic bias in the assessment, r is a person-specific errorvariable that will be present in each self-reported assessmentfor a study subject, and u is an independent measurementerror term. In addition, S0, S1, S2, and S3 are constants to beestimated, and all variables on the right sides of equations 1and 2 are assumed to be independent, given V.

We have recently reported (12) FFQ measurement errorfindings from the Nutritional Biomarker Study among 544women enrolled in the Women’s Health Initiative (WHI)dietary modification (DM) trial. FFQ estimates of energy,protein, and percentage of energy from protein were eachfound to incorporate important systematic bias, and corre-sponding calibration equations were developed. Here, weuse these equations to produce calibrated estimates of en-ergy, protein, and percentage of energy from protein forwomen in the DM trial comparison (control) group andfor women in the WHI observational study. The 2 cohorts

Table 1. Continued

Characteristic


Group (n 5 21,711)a

ObservationalStudy

(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 progesteroneuse ever, yes

28 6,054 31 18,395

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).

Biomarker-calibrated Energy and Cancer 979

Am J Epidemiol 2009;169:977–989

will be used, separately and combined, to assess associa-tions between calibrated nutrient consumption and cancerincidence as observed during WHI follow-up. Cancer riskamong DM intervention group women may depend ina complex manner on baseline and follow-up dietary pat-terns, so that intervention group women were excluded fromthe present analyses.


Study cohorts

Detailed accounts of the design of the WHI Clinical Trialand Observational Study and of the DM trial findings havebeen presented (13–18). This paper uses a subset of womenassigned to the DM trial comparison group (n¼ 29,294) anda subset of the observational study cohort (n ¼ 93,676).Both cohorts included only women who were 50–79 yearsof age at recruitment (1993–1998), were postmenopausal,and had no medical condition associated with less than 3years’ predicted survival. Both provided common core ques-tionnaires at baseline on medical history, reproductive history,family history, personal habits, psychosocial attributes, andfood frequency (19, 20).

DM trial women, who could be assigned to overlappingtrials of postmenopausal hormone therapy and of calciumand vitamin D supplementation, also satisfied additionalexclusionary criteria. To maximize commonality with theDM cohort, the 76,987 observational study women consid-ered here were those remaining after imposing additionalDM trial baseline exclusionary criteria as follows: priorhistory of breast cancer, colorectal cancer, or other cancer(except nonmelanoma skin cancer) within the preceding10 years; a stroke or myocardial infarction in the preceding6 months; severe hypertension (systolic blood pressure>200 mm or diastolic blood pressure>105 mm); already fol-lowing a low-fat diet; underweight (body mass index<18); orFFQ-reported daily energy of<600 kcal or >5,000 kcal.

WHI food frequency questionnaire

All DM trial and observational study women completedFFQs at baseline. DM trial women repeated the FFQ at 1year following enrollment and approximately every 3 yearsthereafter, while observational study women repeated theFFQ at 3 years following enrollment. FFQs were providedin connection with visits to the 40 participating clinicalcenters, where completeness and quality control checkswere applied. The self-administered FFQ included 122 lineitems for individual foods/food groups and 19 adjustmentitems regarding fat intake, as well as summary questions.Nutrition Data System for Research, version 2005, software(University of Minnesota, Minneapolis, Minnesota) wasused to compute daily average nutrient consumption esti-mates (21, 22).

Nutrient Biomarker Study

The WHI Nutrient Biomarker Study was conducted in2004–2005 to assess measurement properties of this FFQ

and to produce calibrated consumption estimates for energyand protein. The eligibility and recruitment methods for theNutrient Biomarker Study have been described (12); 544representative women from the DM trial cohort were en-rolled (276 comparison group, 268 intervention group).These weight-stable women participated in a doubly labeledwater protocol to estimate daily total energy expenditureover a 2-week period, as well as a urinary nitrogen protocolto estimate daily protein consumption over a 24-hour period,and also provided a concurrent FFQ and other questionnairedata. Twenty percent (n ¼ 111) repeated the entire NutrientBiomarker Study protocol an average of 6 months later toprovide reliability data for measurement error componentestimation (12). FFQ total energy and protein were found tobe underestimated, while the percentage of energy fromprotein was overestimated. Women having high body massindex (weight (kg)/height (m)2) and younger women under-estimated energy consumption to a comparatively greaterextent. Calibration equations were developed for each ofenergy, protein, and percentage of energy from protein bylinear regression of log-biomarker estimates on correspond-ing log-FFQ estimates, body mass index, age, ethnicity,and other factors (12). For example, the calibrated log-energy consumption is given by 7.61 þ 0.062 (log-FFQenergy � 7.27) þ 0.013 (body mass index � 28.2) �0.005 (age � 70.9 years), plus some less influential termsinvolving ethnicity, family income, and physical activity.DM intervention group assignment did not meet inclusioncriteria for any of the 3 calibration equations.

Nutrient Biomarker Study application to WHI cohorts

Here, we apply these calibration equations to FFQ datathat were collected earlier in the WHI and relate the cali-brated consumption estimates to subsequent cancer inci-dence. Doing so is complicated by the use of the FFQ inparticipant screening for the DM trial. The exclusion ofabout 50% of the women who had baseline FFQ percent-age of energy from fat of less than 32, in conjunction withFFQ measurement error, implies that the baseline FFQpercentage of energy from fat is overestimated in theDM trial (by about 3% on average), with correspondingestimates of energy likewise distorted. Observationalstudy baseline estimates are distorted in the opposite di-rection because many women screened out from the DMtrial enrolled in the observational study. In terms of equa-tion 2, these distortions arise because women tend to meetthe FFQ inclusion criteria when the independent randomerror term (u) that attends a particular FFQ application ispositive. Later FFQs for a woman, following a sufficientperiod of time (e.g., 6 months) to avoid carry-over effectson this measurement component, can be expected to befree of this measurement effect. Hence, our analyses relyon FFQs obtained at year 1 in the DM trial comparisongroup and at year 3 in the observational study, and onlycancer diagnoses that follow these FFQ collections areincluded in analyses. These FFQs were collected an aver-age 6.5 years (DM trial comparison group) and 4 years(observational study) prior to the Nutrient BiomarkerStudy data collection.

980 Prentice et al.

Am J Epidemiol 2009;169:977–989

Dietary consumption and disease risk associations wereestimated for total invasive cancer, as well as for invasivecancers of the breast, colon, rectum, ovary, endometrium,bladder, kidney, pancreas, and lung, and for lymphoma andleukemia. The ovarian cancer analyses were restricted towomen without bilateral oophorectomy at baseline, andthe endometrial cancer analysis was restricted to womenwith a uterus at baseline.

DM comparison group women were queried twice peryear, and observational study women annually, concerningdiagnosis of any cancer other than nonmelanoma skin can-cer. Cancer reports were verified by medical record andpathology report review by centrally trained physician ad-judicators at participating clinical centers (23).


Log-consumption estimates were calibrated directly fromthe biomarker assessments (equation 1) for the few womenincluded in the Nutrient Biomarker Study and for otherwomen by using the calibration equations previously devel-oped (12).

Hazard ratio estimates were based on Cox regression(24). Follow-up times extended from year 1 (DM trial com-parison group) or year 3 (observational study) to the earliestof cancer occurrence, death, lost to follow-up, or March 31,2005, when the intervention phase of WHI ended. To min-imize mammographic screening influences on results, thebreast cancer analyses censored the follow-up time fora woman the first time she exceeded 2 years without a mam-mogram. The Cox model baseline hazard rates for each

cancer outcome were stratified on baseline age in 5-yearcategories and, for the DM trial comparison group, alsoon hormone therapy trial participation (active estrogen; es-trogen placebo; active estrogen plus progestin; estrogen plusprogestin placebo; not randomized). Analyses that combinethe 2 cohorts stratify also on cohort. Analysis for specificcancer outcomes included standard risk factors in theCox regression model to control confounding, as shown inAppendix Table 1. Women having missing confoundingfactors were excluded from analysis.

Principal analyses modeled the log-hazard ratio linearlyon log-nutrient consumption, so that the hazard ratio fora fractional increase in the nutrient is independent of theconsumption. For display purposes, we present hazard ratiosfor a 20% increase in consumption. For a woman with me-dian consumption, a 20% increment corresponds to about413 kcal of energy, 15 g of protein, or 2.9 units in percentageof energy from protein.

Usual Cox model standard error estimates were calcu-lated for uncalibrated consumption regression coefficients.A more complex standard error estimation procedure isneeded for the calibrated consumption coefficients to ac-knowledge uncertainty in the calibration parameter esti-mates, as well as in the ‘‘regression calibration’’ hazardratio estimation procedure (11), which has been shown tobe free of practically important biases in extensive simula-tion studies. A bootstrap procedure (500 bootstrap samples),with bootstrap sampling stratified on cohort and member-ship in the Nutrient Biomarker Study and in the NutrientBiomarker Study reliability subset, was applied for cali-brated standard error estimation. A bootstrap procedure

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

Cancer


Group (n 5 21,711)a

Observational Study(n 5 59,105)a

Total(N 5 80,816)a

Incidence/1,000Person-Years

No. ofCases


No. ofCases


No. ofCases

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.


Am J Epidemiol 2009;169:977–989

(500 samples) was also used to test the equality of hazardratios in the DM trial comparison group and observationalstudy cohorts.

Calibrated energy turns out to be strongly positively cor-related with body mass index. The data analyzed here do notallow one to determine whether a high body mass should beregarded as a consequence of a high-energy diet, in whichcase body mass index should be excluded from the set ofpotential confounding factors to avoid overcorrection, orwhether a high body mass may arise for other reasons(e.g., sedentary lifestyle), in which case energy consumptionmay be high as a result of related energy requirements, andbody mass index control would be needed in regressionanalyses. Hence, we present hazard ratio estimates forenergy and for body mass index separately and jointly.Two-sided P values are used throughout.

RESULTS

A total of 26,531 (91%) DM trial comparison groupwomen and 66,788 (87%) observational study women pro-vided FFQs (year 1 DM, year 3 observational study) andwere without a prior cancer diagnosis during WHI follow-up. Of these, 21,711 (82%) DM trial comparison group and59,105 (88%) observational study women had all the dataneeded for energy calibration and for confounding controlfor total cancer. Table 1 shows some demographic and life-style characteristics for these women. Analyses of othercancer outcomes or other nutrients involve a slightly differ-ent set of women, because of different confounding factorsand, hence, missing data exclusions.

Table 2 shows incidence rates and the number of invasivecancers through March 31, 2005, for energy analyses foreach cancer site. Incidence rates are similar between the 2cohorts. A total of 5,041 invasive cancers contribute to thetotal cancer analyses, but the number of incident cancers is<300 for specific cancers other than breast, colon, endome-trial, and lung.

Table 3 shows the geometric mean consumption and 95%confidence interval for the consumption of energy, protein,and percentage of energy from protein for both cohorts, withand without calibration. The distribution of calibrated con-sumption estimates is similar in the 2 cohorts. The narrowerconfidence intervals for the calibrated versus uncalibratedestimates 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% increasein total energy consumption under a linear log-hazard ratiomodel 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 1.3standard deviations for calibrated protein. For comparison,extreme quartile medians differ by about 2.3 standard devi-ations, and extreme tertile medians differ by about 1.9 stan-dard deviations, for normally distributed exposures.

Separate hazard ratio estimates are given for the DM trialcomparison group and observational study cohorts, withoutand with biomarker calibration of consumption estimates.Biomarker calibration clearly has a major impact on hazard T

able

3.

GeometricMeanConsumptio

nand95%

ConfidenceIntervalsforUncalibratedDietary

Consumptio

n,a

sEstim

atedbytheWomen’sHealth

InitiativeFoodFrequencyQuestio

nnaire,

andforCalibratedConsumptio

nUsingNutritionalBiomarkerData,in

theWomen’s

Health

InitiativeDietary

Modifica

tionTrialComparisonGroupandObservatio

nalStudy,1993–2005

Energy(kcal/day)

Protein

(g/day)

PercentageofEnergy

From

Protein

Uncalibrated

Calibrateda

Uncalibrated

Calibrated

Uncalibrated

Calibrated

Mean

95%

Confidence

Interval

Mean

95%

Confidence

Interval

Mean

95%

Confidence

Interval

Mean

95%

Confidence

Interval

Mean

95%

Confidence

Interval

Mean

95%

Confidence

Interval

Dietary

modificatio

ntrialcomparisongroup

(n¼

21,711)

1,477.2

676.6,3,224.9

2,140.6

1,786.9,2,564.2

61.2

26.3,142.1

78.1

58.4,104.4

16.6

11.5,24.0

14.4

11.9,17.3

Observatio

nalstudy

(n¼

59,105)

1,384.3

641.0,2,989.4

2,055.8

1,722.3,2,453.9

58.6

24.8,138.1

74.2

54.8,100.5

16.9

11.5,25.0

14.4

11.8,17.6

aCalibratedbyusingthemeasurementmodel(equatio

n1)forwomenin

theNutritionBiomarkerStudyandequatio

n2otherw

ise.

982 Prentice et al.

Am J Epidemiol 2009;169:977–989

ratio estimates, with evidence for positive associations be-tween calibrated energy and total cancer, as well as certainsite-specific cancers, in both the DM trial comparison groupand observational study cohorts, but with little evidence ofassociation for uncalibrated energy. There is also little evi-dence of difference in hazard ratios between the 2 cohorts,with or without calibration, with the possible exception ofleukemia.

Figure 1 shows corresponding hazard ratio estimates and95% confidence intervals from the analysis of the 2 cohortscombined. 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, withthe exception of an inverse association with colon cancer.The wider confidence intervals for calibrated versus uncali-brated energy hazard ratios reflect both uncertainty in thecoefficients of the calibration equations and deattenuationthat arises from acknowledging dietary assessment mea-surement error in the hazard ratio estimation procedure.

Analyses of calibrated protein and percentage of energyfrom protein similarly yielded little evidence of hazard ratiodifferences between the 2 cohorts (each P > 0.05). Figures2 and 3 show corresponding combined cohort hazard ratiosand 95% confidence intervals for a 20% increase in thesenutritional factors. The hazard ratios for a 20% increase incalibrated protein are above 1 for total cancer (HR ¼ 1.06,95% CI: 1.01, 1.12), breast cancer (HR ¼ 1.09, 95% CI:

1.01, 1.19), endometrial cancer (HR ¼ 1.37, 95% CI: 1.16,1.61), and leukemia (HR¼ 1.39, 95% CI: 1.05, 1.83). Thesepositive associations may be substantially attributable tocorrelation between protein and energy consumption, sincethe hazard ratio estimates for percentage of energy fromprotein are less than 1 for total and most specific cancers,and the inverse association is significant for total cancer

Table 4. Hazard Ratio Estimates for a 20% Increase in Energy (kcal/day) Consumption in theWomen’s Health Initiative Dietary Modification Trial

Comparison Group and Observational Study, Without and With Biomarker Calibration, 1993–2005

Cancer

Dietary Modification TrialComparison Group

ObservationalStudy

Test of Equality ofHazard Ratios

Uncalibrated Calibrated Uncalibrated CalibratedUncalibratedP Valuea

CalibratedP ValueaHazard

Ratiob95% Confidence

IntervalHazardRatiob

95% ConfidenceIntervalc

HazardRatiob

95% ConfidenceInterval

HazardRatiob

95% ConfidenceInervalc

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.

Figure 1. Estimated hazard ratios and 95% confidence intervals fora 20% increase in energy consumption (kcal/day), from combinedanalysis of data from the Women’s Health Initiative dietary modifica-tion trial comparison group and observational study, without and withbiomarker calibration of consumption, 1993–2005. Unfilled square,uncalibrated; filled circle, calibrated.


Am J Epidemiol 2009;169:977–989

(HR ¼ 0.92, 95% CI: 0.85, 0.99 for a 20% increase inpercentage of energy from protein). Results correspondingto Figures 1–3 by quartile of calibrated consumption aregiven in Appendix Tables 2–4.

The correlation coefficients for body mass index with log-transformed energy, protein, and percentage of energy fromprotein in the combined cohorts were, respectively, 0.07,0.10, and 0.07 without calibration and 0.81, 0.46, and�0.12 following calibration. Hence, it may be difficult todistinguish between total energy and body mass index asso-ciations, with total or site-specific cancer. Table 5 examines

the effect of including body mass index in the log-hazardratio model on the calibrated energy hazard ratios shown inFigure 1 and also shows the effect of including calibratedenergy on the hazard ratio for body mass index. Hazardratios 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 correctfor 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, FFQestimates of energy, protein, or percentage of energy fromprotein were not significantly associated with total invasivecancer incidence. In contrast, following biomarker calibra-tion, the associations with total cancer incidence werestrong for energy (P < 0.0001), moderate for protein (P ¼0.01), 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, calibratedenergy consumption was found to be positively associatedwith the risk of breast, colon, endometrial, and kidney cancer,whereas uncalibrated energy was not.

These comparisons suggest that systematic bias in dietaryassessment could have a profound effect on nutritional ep-idemiology findings. Total energy assessment is a recog-nized weak aspect of FFQs. Uncalibrated FFQs aregenerally believed to be more reliable for nutrient densitythan for absolute consumption estimates. However, bio-marker calibration also qualitatively affected the findingsfor protein density in relation to total cancer (Figure 3).

Measurement error has typically been acknowledged inepidemiology reporting through a simple deattenuation fac-tor, as befits measurement equation 2 in the absence ofsystematic bias (i.e., S2 ¼ S3 ¼ 0). Such deattenuation typ-ically has little effect on significance levels. The presence ofsystematic bias changes this feature, however, because re-gression coefficients are corrected for distortions beyondsimple attenuation, possibly leading to substantially alteredP values.

To help interpret the calibrated energy variable definedhere, we note that calibrated energy can be viewed as esti-mated actual short-term energy consumption, as determinedby FFQ energy, body mass index, age, and other factors. Thecorrelations 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 withage and especially with body mass index imply that log-FFQenergy does not adhere to a simple classical measurementmodel. A linear regression of body mass index on log-calibrated energy gives a projected body mass index increaseof 9.2 units corresponding to a 20% increase in calibratedenergy, suggesting, in conjunction with Table 5, that much ofthe observed dependence of cancer incidence rates on totalenergy can be explained by body mass associations with

Figure 2. Estimated hazard ratios and 95% confidence intervals fora 20% increase in protein consumption (g/day), from combined anal-ysis of data from the Women’s Health Initiative dietary modificationtrial comparison group and observational study, without and with bio-marker calibration of consumption, 1993–2005. Unfilled square, un-calibrated; filled circle, calibrated.

Figure 3. Estimated hazard ratios and 95% confidence intervals fora 20% increase in percentage of energy from protein, from combinedanalysis of data from the Women’s Health Initiative dietary modifica-tion trial comparison group and observational study, without and withbiomarker calibration of consumption, 1993–2005. Unfilled square,uncalibrated; filled circle, calibrated.

984 Prentice et al.

Am J Epidemiol 2009;169:977–989

these diseases. Table 5 likewise suggests that much of thedependence of cancer incidence rates on body mass indexcan be explained by energy consumption associations withthese diseases.

Our analyses yielded similar results when calibrationequations were applied in the DM cohort where they werederived and when exported to the observational study. How-ever, this extrapolation is under near-optimal conditions asthe 2 cohorts were drawn from essentially the same popu-lations, with much commonality in eligibility and exclusion-ary criteria. Comparison with calibration equations fromnutritional biomarker studies in other populations (25, 26)could be informative.

As noted above, the Nutrient Biomarker Study was con-ducted in 2004–2005, an average of about 6.5 years after the1-year FFQ data collection for the DM trial comparisongroup women and about 4 years on average after the 3-yearFFQ data collection for observational study women. Ourapplication assumes that the calibration equations devel-oped from Nutrient Biomarker Study data apply to FFQsat these earlier time points. Moreover, the biomarker dataprovide consumption estimates over a rather short period oftime (e.g., 6 months between initial and repeat applicationsin the 20% subsample). However, dietary patterns are ex-pected to track over longer time periods for most women inthese cohorts.

On the substantive side, we observe strong positive asso-ciations between calibrated energy consumption and the riskof total and certain site-specific cancers. There are also sug-gestions of a positive association between protein consump-tion and leukemia and an inverse association betweenpercentage of energy from protein and bladder cancer(Figures 2 and 3) that would be worth examining in othersettings. More comprehensive temporal data on the interplay

between a high-energy diet and body fat accumulation willbe needed to understand the mechanisms leading to elevatedcancer risk with a high level of energy consumption. How-ever, regardless of whether body fat accumulation resultsfrom a history of high-energy consumption, or whethera high body mass leads to increased energy requirements,or both, it is evident that a high body mass index is animportant aspect of total and site-specific cancer risk, andefforts to prevent obesity deserve a continued high priorityin national cancer control efforts.

ACKNOWLEDGMENTS

Author affiliations: Division of Public Health Sciences, FredHutchinson Cancer Research Center, Seattle, Washington(Ross L. Prentice, Marian L. Neuhouser, Ruth E. Patterson,Lesley F. Tinker); Biostatistics Research Branch, NationalInstitute of Allergy and Infectious Diseases, Bethesda,Maryland (Pamela A. Shaw); Medical Research CouncilDunn Human Nutrition Unit, University of Cambridge,Cambridge, United Kingdom (Sheila A. Bingham); Depart-ment of Epidemiology, University of Washington, Seattle,Washington (Shirley A. A. Beresford); Kaiser PermanenteDivision of Research, Oakland, California (Bette Caan);Stanford Prevention Research Center, Palo Alto, California(Marcia L. Stefanick); University of Tennessee Health SciencesCenter, Memphis, Tennessee (Suzanne Satterfield); Depart-ment of Nutritional Sciences, University of Arizona, Tucson,Arizona (Cynthia A. Thomson); Department of Communityand Behavioral Health, University of Iowa, Iowa City, Iowa(Linda Snetselaar); Medstar Research Institute, Washington,District of Columbia (Asha Thomas); and Johns Hopkins

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

Cancer

Calibrated Energy Body Mass Index

Body Mass Index Unadjusted Body Mass Index Adjusted Energy Unadjusted Energy Adjusted

HazardRatio

95% ConfidenceIntervala

HazardRatio


HazardRatio


HazardRatio


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.


Am J Epidemiol 2009;169:977–989

University/Sinai Hospital, Baltimore, Maryland (AshaThomas).

This work was supported by the National Heart, Lung, andBlood Institute, National Institutes of Health, US Departmentof Health and Human Services (contracts N01WH22110,24152, 32100-2, 32105-6, 32108-9, 32111-13, 32115, 32118-19, 32122, 42107-26, 42129-32, and 44221). Clinical TrialsRegistration: ClinicalTrials.gov identifier: NCT00000611. Dr.Prentice’sworkwaspartially supportedbygrantCA53996 fromthe National Cancer Institute.

The authors thank the WHI investigators and staff fortheir outstanding dedication and commitment. A full listingof WHI investigators can be found at the following website:http://www.whi.org.

A list of key WHI investigators involved in this researchfollows. Program Office—National Heart, Lung, and BloodInstitute, Bethesda, Maryland: Elizabeth Nabel, JacquesRossouw, Shari Ludlam, Linda Pottern, Joan McGowan,Leslie Ford, Nancy Geller. Clinical Coordinating Cen-ters—Fred Hutchinson Cancer Research Center, Seattle,Washington: Ross Prentice, Garnet Anderson, AndreaLaCroix, Charles L. Kooperberg, Ruth E. Patterson, AnneMcTiernan; Wake Forest University School of Medicine,Winston-Salem, North Carolina: Sally Shumaker; MedicalResearch Labs, Highland Heights, Kentucky: Evan Stein;University of California at San Francisco, San Francisco,California: Steven Cummings. Clinical Centers—AlbertEinstein College of Medicine, Bronx, New York: SylviaWassertheil-Smoller; Baylor College of Medicine, Houston,Texas: Aleksandar Rajkovic; Brigham and Women’s Hospi-tal, Harvard Medical School, Boston, Massachusetts: JoAnnManson; Brown University, Providence, Rhode Island:Annlouise R. Assaf; Emory University, Atlanta, Georgia:Lawrence Phillips; Fred Hutchinson Cancer Research Cen-ter, Seattle, Washington: Shirley Beresford; George Wash-ington University Medical Center, Washington, District ofColumbia: Judith Hsia; Los Angeles Biomedical Research In-stitute at Harbor-UCLA Medical Center, Torrance, California:Rowan Chlebowski; Kaiser Permanente Center for HealthResearch, Portland, Oregon: Evelyn Whitlock; Kaiser Per-manente Division of Research, Oakland, California: BetteCaan;Medical College of Wisconsin, Milwaukee, Wisconsin:Jane Morley Kotchen; MedStar Research Institute/HowardUniversity, Washington, District of Columbia: Barbara V.Howard; Northwestern University, Chicago/Evanston,Illinois: Linda Van Horn; Rush Medical Center, Chicago,Illinois: Henry Black; Stanford Prevention Research Center,Stanford, California:Marcia L. Stefanick; StateUniversity ofNew York at Stony Brook, Stony Brook, New York: DorothyLane; The Ohio State University, Columbus, Ohio: RebeccaJackson; University of Alabama at Birmingham, Birming-ham, Alabama: Cora E. Lewis; University of Arizona,Tucson/Phoenix, Arizona: Tamsen Bassford; University atBuffalo, Buffalo, New York: JeanWactawski-Wende;Univer-sity of California at Davis, Sacramento, California: JohnRobbins;University of California at Irvine, Irvine, California:F. Allan Hubbell; University of California at Los Angeles,Los Angeles, California: Lauren Nathan; University ofCalifornia at San Diego, LaJolla/Chula Vista, California:Robert D. Langer; University of Cincinnati, Cincinnati,

Ohio: Margery Gass; University of Florida, Gainesville/Jacksonville, Florida: Marian Limacher; University ofHawaii, Honolulu, Hawaii: David Curb; University of Iowa,Iowa City/Davenport, Iowa: Robert Wallace; University ofMassachusetts/Fallon Clinic, Worcester, Massachusetts:Judith Ockene; University of Medicine and Dentistry of NewJersey, Newark, New Jersey: Norman Lasser; University ofMiami, Miami, Florida: Mary Jo O’Sullivan; University ofMinnesota,Minneapolis,Minnesota:KarenMargolis;Univer-sity of Nevada, Reno, Nevada: Robert Brunner; University ofNorth Carolina, Chapel Hill, North Carolina: Gerardo Heiss;University of Pittsburgh, Pittsburgh, Pennsylvania: LewisKuller; University of Tennessee, Memphis, Tennessee: KarenC. Johnson; University of Texas Health Science Center, SanAntonio, Texas: Robert Brzyski; University of Wisconsin,Madison, Wisconsin: Gloria E. Sarto;Wake Forest UniversitySchool of Medicine, Winston-Salem, North Carolina: MaraVitolins; and Wayne State University School of Medicine/Hutzel Hospital, Detroit, Michigan: Susan Hendrix.

Decisions concerning study design, data collection andanalysis, interpretation of results, preparation of the manu-script, or submission of the manuscript for publication residedwith committees comprising WHI investigators that includedNational Heart, Lung, and Blood Institute representatives.


REFERENCES

1. Armstrong B, Doll R. Environmental factors and cancer in-cidence and mortality in different countries, with special ref-erence to dietary practices. Int J Cancer. 1975;15(4):617–631.

2. Kritchevsky D, Klurfeld DM. Influence of caloric intake onexperimental carcinogenesis: a review. Adv Exp Med Biol.1986;206:55–68.

3. Freedman LS, Clifford C, Messina M. Analysis of dietary fat,calories, body weight, and the development of mammary tu-mors in rats and mice: a review. Cancer Res. 1990;50(18):5710–5719.

4. Willett WC, Sampson L, Stampfer MJ, et al. Reproducibilityand validity of a semiquantitative food frequency question-naire. Am J Epidemiol. 1985;122(1):51–65.

5. Schoeller DA. Recent advances from application of doubly-labeled water to measurement of human energy expenditure.J Nutr. 1999;129(10):1765–1768.

6. Heitmann BL, Lissner L. Dietary underreporting by obeseindividuals—is it specific or non-specific? BMJ. 1995;311(7011):986–989.

7. World Cancer Research Fund/American Institute for CancerResearch. Food, Nutrition and the Prevention of Cancer: AGlobal Perspective. Washington, DC: American Institute forCancer Research; 1997.

8. World Cancer Research Fund/American Institute for CancerResearch. Food, Nutrition, Physical Activity, and the Preven-tion of Cancer: A Global Perspective. Washington, DC:American Institute for Cancer Research; 2007.

9. Bingham SA. Urine nitrogen as a biomarker for the validationof dietary protein intake. J Nutr. 2003;133(suppl 3):921S–924S.

10. Prentice RL, Sugar E, Wang CY, et al. Research strategies andthe use of nutrient biomarkers in studies of diet and chronicdisease. Public Health Nutr. 2002;5(6A):977–984.

986 Prentice et al.

Am J Epidemiol 2009;169:977–989

http://www.whi.org

11. Sugar EA, Wang CY, Prentice RL. Logistic regression withexposure biomarkers and flexible measurement error. Bio-metrics. 2007;63(1):143–151.

12. Neuhouser ML, Tinker L, Shaw PA, et al. Use of recoverybiomarkers to calibrate nutrient consumption self-reports inthe Women’s Health Initiative. Am J Epidemiol. 2008;167(10):1247–1259.

13. Women’s Health Initiative Study Group. Design of theWomen’s Health Initiative clinical trial and observationalstudy. Control Clin Trials. 1998;19(1):61–109.

14. Women’s Health Initiative Study Group. Dietary adherence inthe Women’s Health Initiative dietary modification trial. J AmDiet Assoc. 2004;104(4):654–658.

15. Prentice RL, Caan B, Chlebowski RT, et al. Low-fat dietarypattern and risk of invasive breast cancer: the Women’s HealthInitiative randomized controlled dietary modification trial.JAMA. 2006;295(6):629–642.

16. Beresford SA, Johnson KC, Ritenbaugh C, et al. Low-fat di-etary pattern and risk of colorectal cancer: the Women’sHealth Initiative randomized controlled dietary modificationtrial. JAMA. 2006;295(6):643–654.

17. Howard BV, Van Horn L, Hsia J, et al. Low-fat dietary patternand risk of cardiovascular disease: the Women’s Health Ini-tiative randomized controlled dietary modification trial.JAMA. 2006;295(6):655–666.

18. Prentice RL, Thomson CA, Caan B, et al. Low-fat dietarypattern and cancer incidence in the Women’ s Health Initiative

dietary modification randomized controlled trial. J Natl Can-cer Inst. 2007;99(20):1534–1543.

19. Hays J, Hunt JR, Hubbell FA, et al. The Women’s HealthInitiative recruitment methods and results. Ann Epidemiol.2003;13(9 suppl):S18–S77.

20. Langer RD, White E, Lewis CE, et al. The Women’s HealthInitiative Observational Study: baseline characteristics ofparticipants and reliability of baseline measures. Ann Epide-miol. 2003;13(9 suppl):S107–S121.

21. Kristal AK, Shattuck AL, Williams ME. Food frequencyquestionnaires for diet intervention research. In: 17th NationalNutrient Databank Conference Proceedings. Baltimore, MD:International Life Sciences Institute; 1992:110–125.

22. Schakel SF, Buzzard IM, Gebhardt SE. Procedures for esti-mating nutrient values for food composition databases. J FoodCompost Anal. 1997;10(2):102–114.

23. Curb JD, McTiernan A, Heckbert SR, et al. Outcomes ascer-tainment and adjudication methods in the Women’s HealthInitiative. Ann Epidemiol. 2003;13(9 suppl):S122–S128.

24. Cox DR. Regression analysis and life tables (with discussion).J R Stat Soc (B). 1972;34:187–220.

25. Subar AF, Kipnis V, Troiano RP, et al. Using intake biomarkersto evaluate the extent of dietary misreporting in a large sampleof adults: the OPEN study. Am J Epidemiol. 2003;158(1):1–13.

26. Kipnis V, Subar AF, Midthune D, et al. Structure of dietarymeasurement error: results of the OPEN biomarker study. AmJ 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

TotalCancer

Cancer SiteLymphoma,LeukemiaBreast

Colon,Rectum

Ovary EndometriumBladder, Kidney,Pancreas, Lung

Raceb (white/other, black, Hispanic) x x xc x xd

Education (high school or less, beyondhigh school, college degree)

x x

Exercise (METs/week) x x x

Smokingb (never, past, current) x x x x xe

Alcoholb (never, past, <1/week, 1–7/week,>7/week)

x x x x

Breast cancer family history(no, yes)

x x

Gail 5-year risk (5-year absolute risk %) x

Unopposed estrogen use ever(no, yes)

x x xc x x

Estrogen plus progesterone use ever(no, yes)

x x xc x x

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

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, 1993�2005a

Cancer

Energy (kcal/day)

Quartile 2 Quartile 3 Quartile 4

HazardRatio


HazardRatio


HazardRatio


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, 1993�2005a

Cancer

Protein (g/day)


HazardRatio


HazardRatio


HazardRatio


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


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.

988 Prentice et al.

Am J Epidemiol 2009;169:977–989

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, 1993�2005a

Cancer

Percentage of Energy From Protein


HazardRatio


HazardRatio


HazardRatio


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


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




Vol. 169, No. 8

DOI: 10.1093/aje/kwn418



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

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, theirpossible biologic interaction has not been well studied. The authors examined data from 89,789 subjects aged25–69 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 combinedeffect of sex and HBV infection on liver cancer mortality by calculating 3 interaction measures: the relative risk dueto interaction, the attributable proportion of interaction, and the synergy index. There was a greater risk differencebetween hepatitis B surface antigen carriers and noncarriers among men than among women. After adjustment forpotential confounders, the relative risk due to interaction, the attributable proportion of interaction, and the synergyindex 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 onliver cancer mortality. HBV infection had a larger impact on liver cancer mortality in men than in women, which mayexplain 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 tointeraction.

Hepatocellular carcinoma represents the majority of casesof primary liver cancer (hereafter called liver cancer); it isthe fifth most common cancer and the third leading cause ofcancer death worldwide (1, 2). China is one of the countrieswith the highest incidence of hepatocellular carcinoma inthe world—approximately 35 cases per 100,000 populationfor males and 13 per 100,000 for females (2). Hepatocellularcarcinoma is amale-predominant cancer, with amale:femaleratio ranging from 2:1 to 4:1 (2). Hepatitis B virus (HBV)infection, which is endemic in many Asian countries, in-cluding China, is a leading risk factor for hepatocellularcarcinoma and accounts for approximately 53% of hepato-cellular carcinoma cases worldwide (3). Other importantrisk factors for hepatocellular carcinoma include hepatitisC virus infection, alcoholism, aflatoxin B1 ingestion, andfamily history of liver cancer (4). So far, it has been unclear

whether HBV infection can explain the sex difference in theoccurrence of liver cancer. In this study, we examined effectmodification by sex of the association between HBV infec-tion and liver cancer mortality, based on data from a 14-yearlongitudinal study conducted in Haimen, China.


Study population

Study methods have been described in detail previously(5). Briefly, the study was started in February 1992 inHaimen, China, and enrollment continued until December1993. A total of 90,236 residents (60,306 men and 29,930women) voluntarily participated. The subjects were 22–80years 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: [email protected]) 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: [email protected]).

990 Am J Epidemiol 2009;169:990–995

and provided a blood specimen at entry. The questionnairecollected information about name, residence, date of birth,occupation (peasant, nonpeasant), regular cigarette smoking(daily), regular alcohol drinking (�4 drinks per week), regu-lar tea drinking (�4 times per week), past pesticide exposure,drinking water source (ditch, pond, shallow well, deep well)in every decade from the 1960s to the 1990s, staple food type(corn, rice, wheat) in every decade from the 1960s to the1990s, history of acute hepatitis, history of jaundice, historyof cirrhosis, and family history of liver cancer. The subjectswere tested for hepatitis B surface antigen (HBsAg) byradioimmunoassay.

The study protocol was approved by the InstitutionalReview Board of Fox Chase Cancer Center (Philadelphia,Pennsylvania), the Medical Ethics Review Group of HaimenCity (Haimen, China), and the Ethics Review Committee ofShanghai Medical University (Shanghai, China). Question-naire data were collected by trained interviewers employedat the Haimen City Anti-Epidemic Station. In this analysis,we included all 89,789 participants (60,076 men and 29,713women) who were aged 25–69 years at entry.

All of the subjects were followed up every year for vitalstatus and liver cancer mortality until December 31, 2006.Death certification information was reported to the deathregister system at Haimen City Anti-Epidemic Stationmonthly by village or township doctors. Information oncause of death for persons who had temporarily moved outof Haimen was also collected. Persons who had a permanentchange of residency were identified and no longer followed.


Person-years of follow-up were calculated for each par-ticipant from the date of the baseline survey to the date ofdeath, the date of loss to follow-up, or, if neither of thoseevents occurred, the end of the study period (December 31,2006). Subjects who were lost to follow-up were censoredon the last confirmed date of their presence in Haimen.A Cox proportional hazards model was used to estimatethe sex-specific hazard ratio and 95% confidence intervalfor liver cancer mortality associated with HBsAg status.Time-dependent covariates were used to assess proportion-ality. The baseline hazard in the proportional hazards re-gression models was stratified by 1-year age category, andthen the proportional hazards assumption for the other var-iables was satisfied. Estimated hazard ratios were adjustedfor other potential confounders, including occupation, cig-arette smoking, alcohol consumption, history of hepatitis,and family history of liver cancer.

The modifying effect of sex on the relation betweenHBsAg and liver cancer mortality was examined on a mul-tiplicative scale; it was not statistically significant in a pre-vious analysis (5) and also was not significant in this study.It has been argued that interaction on an additive scale ismore meaningful for assessing public health and biologicsignificance (6, 7) and also more indicative of the underly-ing causal mechanism (8). In this analysis, we assessed theeffect of the additive interaction of HBsAg status and sex onrisk of liver cancer mortality. We calculated 3 measures(9, 10)—the relative excess risk of interaction (RERI) (calcu-

lated asRRAþBþ�RRAþB��RRA�Bþ þ 1), the attributableproportion of interaction (calculated as RERI/RRAþBþ),and the synergy index (calculated as (RRAþBþ �1)/[(RRAþB�� 1)þ (RRA�Bþ� 1)])—and their 95%confidenceintervals to assess the combined effect of male sex and HBVinfection asmeasured by HBsAg status. Subjects were groupedinto 4 categories: femaleswhowereHBsAg-negative (A�B�),females who were HBsAg-positive (AþB�), males who wereHBsAg-negative (A�Bþ), and males who were HBsAg-positive (AþBþ). If there is no interaction, RERI and the at-tributable proportion of interaction are both 0 and the synergyindex is 1 (9).

Statistical analysis was performed using the StatisticalAnalysis System, version 9.1 (SAS Institute, Inc., Cary,North Carolina). The SAS program provided by Li andChambless (11) was used to calculate the 3 measures ofadditive interaction and to test for significance.

RESULTS

The 89,789 subjects had a total of 1,147,713 person-yearsof follow-up. Table 1 shows the demographic characteristics

Table 1. Characteristics of Participants at Study Entry, Haimen

Cohort Study, China, 1992–2006

CharacteristicMen Women Total

No. % No. % No. %

Age, years

<40 23,081 38.4 14,179 47.7 37,260 41.5

40–54 27,511 45.8 12,362 41.6 39,873 44.4

�55 9,484 15.8 3,172 10.7 12,656 14.1

Hepatitis B surfaceantigen status

Positive 9,171 15.3 5,015 16.9 14,186 15.8

Negative 50,905 84.7 24,698 83.1 75,603 84.2

Occupation

Peasant 43,307 72.1 24,792 83.4 68,099 75.8

Nonpeasant 16,769 27.9 4,921 16.6 21,690 24.2

Regular cigarettesmoking (daily)

Yes 38,864 64.7 1,014 3.4 39,878 44.4

No 21,212 35.3 28,699 96.6 49,911 55.6

Regular alcoholdrinking (�4drinks/week)

Yes 35,985 59.9 3,812 12.8 39,797 44.3

No 24,091 40.1 25,901 87.2 49,992 55.7

History of acutehepatitis

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

Family history ofliver cancer

Yes 1,992 3.3 1,394 4.7 3,386 3.8

No 58,084 96.7 28,319 95.3 86,403 96.2

Sex-Modified Effect of HBV on Liver Cancer Mortality 991

Am J Epidemiol 2009;169:990–995

of the participants at baseline. Among the subjects, 15.8%were HBsAg-positive (men: 15.3%; women: 16.9%).

A total of 13,529 deaths or losses to follow-up were iden-tified, amongwhich 1,803 participants died from liver cancer,with 73.7% of them being HBsAg-positive. The median ageat death for liver cancer patients was 51.7 years for men and52.4 years for women. Mortality from liver cancer varied inthe 4 groups defined by sex andHBsAg status (Table 2).MaleHBsAg carriers had the highest liver cancer mortality(1,035.2 deaths per 100,000 person-years), while femalenon-HBsAg carriers had the lowest (16.1 deaths per100,000 person-years). Themortality difference for liver can-cer between HBsAg carriers and non-HBsAg carriers wasmuch greater in men than in women among all age groups.

HBsAg carriers had an increased risk of mortality fromliver cancer, with a hazard ratio of 16 as compared with non-

HBsAg carriers. Male sex and older age were also importantpredictors of liver cancer mortality. Compared with womenwho were HBsAg-negative, mortality was significantly in-creased in males and persons who were HBsAg-positive,and the combined effect of male sex and HBsAg positivitywas synergistic (Figure 1). After adjustment for potentialconfounders, the 3 interaction measures on the additivescale all showed significantly higher values than the null(RERI: 33.27 (95% confidence interval (CI): 22.54,43.99); attributable proportion of interaction: 0.59 (95%CI: 0.55, 0.63); synergy index: 2.49 (95% CI: 2.13, 2.90)),suggesting a synergistic effect of male sex and HBsAg pos-itivity on liver cancer mortality. We further examined thisjoint effect according to age (Table 3). The RERI decreasedwith increasing age, suggesting that the sex-HBsAg inter-action was stronger in younger people.

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

Sex or Age Group (years)and HBsAg Status

% HBsAg-Positive

No. ofDeaths

Person-Yearsof Follow-Up

Mortality per100,000

Person-Years

RiskDifference

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.

992 Wang et al.

Am J Epidemiol 2009;169:990–995

DISCUSSION

It has been well documented that men are more likelythan women to develop liver cancer (2), and HBV infectionis a major risk factor (3). In this analysis, we focused on thejoint effect of male sex and HBV infection on liver cancermortality. The proportion of HBsAg-positive men was com-parable to that of women in our study. It has been suggestedthat the higher rate of liver cancer in males than in females isdue to higher levels of exposure to risk factors (2). However,after adjustment for covariates, including HBsAg status,age, history of hepatitis, occupation, regular alcohol drink-ing, regular cigarette smoking, and family history of livercancer, men still had higher mortality from liver cancer thanwomen did. This may suggest that some susceptibility otherthan exposure level is more important.

We assessed additive interaction by calculating 3 mea-sures, including RERI, the attributable proportion of inter-action, and the synergy index, all of which indicated thatthere was a significant synergistic effect of HBsAg positivityand male sex on risk of liver cancer mortality. HBsAg pos-itivity had an increased effect on liver cancer mortality inmen compared with women. RERI, considered the best

measure of additivity in the hazards model (11), indicatedthat 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 mortalityamong male HBsAg carriers was attributable to the interac-tion. Because of the additive interaction, the excess risk ofliver cancer mortality resulting from combined exposure tomale sex and HBsAg positivity was 2.49 times higher thanthe sum of their independent effects. These results suggestthat men are more sensitive to the effect of HBV infection onliver cancer mortality than women. An increased impact ofHBV infection may explain, at least in part, the elevated riskof liver cancer in men.

Many studies have explored potential mechanisms ofliver cancer development, as well as the sex difference inincidence. Sex hormones may play an important role in sexdisparity in the process of oncogenesis. In a multicentercase-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 againstliver cancer. Naugler et al. (13) found that estrogen-mediated inhibition of interleukin-6 production by Kupffercells reduced the risk of liver cancer in females. An intra-cellular signaling protein called MyD88 and the transcrip-tion factor nuclear factor jB were also found to be involvedin this signaling pathway (13, 14). Wang et al. (15) inte-grated the HBsAg gene into the mouse p21neomycin locus andfound that estrogen receptor b was extremely up-regulated,which indicates that estrogen receptor b may play an im-portant role in the development of liver cancer related toHBsAg. In a case-control study, Yu et al. (16) and Sung et al.(17) found that there was a positive association betweentestosterone levels and liver cancer risk in HBV carriersand that genes involved in the regulation of testosterone,such as SRD5A2 and V89L, may also play a role in theetiology of liver cancer.

Liver cancer arises most frequently in the setting ofchronic liver inflammation (18). After infection by the hep-atitis virus, the host’s inflammatory immune response to theviral antigens induces hepatocyte damage and is followedby the pathogenesis of liver cancer (19). Sex hormones mayinteract with HBV infection in the process and lead to a dom-inant sex disparity in liver cancer risk.

Figure 1. Individual and joint effects of sex and hepatitis B surfaceantigen (HBsAg) status onmortality from primary liver cancer, HaimenCohort Study, China, 1992–2006. HBV, hepatitis B virus.

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,years

Unadjusted Adjusteda

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

In our analysis, the sex-HBsAg interaction tended to de-crease with age. Previous study has demonstrated that livercancer tends to occur in men and postmenopausal women,which may result from lower production of estradiol anda reduced response to the action of estradiol (20). Thatmay contribute to a lower sex disparity in the HBV-relatedrisk of liver cancer in older people. Another possibility ispotentially competing causes of death, which may be moreprominent in older subjects than in younger ones and is alsosex-related.

One limitation of our study is that HBsAg status was usedas the sole indicator of HBV infection. Occult HBV infec-tion, a type of chronic HBV infection that is characterizedby the absence of detectable HBsAg in the blood and verylow levels of HBV DNA in the blood and liver (21, 22), canalso lead to liver cancer, with a mechanism similar to that ofovert cases (23). Therefore, part of the mortality amongHBsAg noncarriers might be attributable to the effect ofoccult hepatitis B infection, resulting in underestimationof the effect of HBV infection as well as its interaction withsex. Additional data on HBV infection, such as informationon hepatitis B e antigen, HBV virus load, and HBV geno-type, would consolidate the finding (3, 23).

In summary, we found a synergistic effect of male sex andHBV infection on risk of liver cancer mortality, suggestingthat HBV infection may have a greater impact on liver can-cer risk in men than in women, at least for persons living inhigh-risk areas. Possible mechanisms for the interaction be-tween sex and HBV infection in liver cancer mortality needto be further explored. If the finding is confirmed by otherstudies, a specific HBV vaccination program may be devel-oped to reduce liver cancer risk. It has been well demon-strated that HBV infection frequently occurs as a result ofvertical (mother-to-child) or horizontal (child-to-child)transmission in endemic areas (24). It has also been reportedthat HBV infection can be acquired in adulthood throughinvasive medical procedures or household contact withHBV carriers (25). In Haimen, where the current studywas conducted, an HBV vaccination program for all new-borns was implemented in 1992. Adults, however, are notincluded in the program. Because of the increased risk ofliver cancer in HBsAg-positive males, immunization strat-egies might be expanded to also include male adults not yetinfected with HBV, to reduce the risk of developing livercancer.

ACKNOWLEDGMENTS

Author affiliations: Department of Epidemiology, Schoolof Public Health, Fudan University, Shanghai, People’s Re-public of China (Na Wang, Yingjie Zheng, Qingwu Jiang);Haimen City Center for Disease Control and Prevention,Haimen City, People’s Republic of China (Xinsen Yu,Wenyao Lin); and Department of Epidemiology and Com-munity Medicine, Faculty of Medicine, University of Ot-tawa, Ottawa, Ontario, Canada (Yue Chen).

This work was supported by a grant from the Ministry ofScience and Technology of the People’s Republic of China

(2006BAI02A03) and the Shanghai Leading Academic Dis-cipline Project (B118). Na Wang received a scholarshipfrom the China Scholarship Council for her training at theUniversity of Ottawa.


REFERENCES

1. Parkin DM. Global cancer statistics in the year 2000. LancetOncol. 2001;2(9):533–543.

2. El-Serag HB, Rudolph KL. Hepatocellular carcinoma: epide-miology and molecular carcinogenesis. Gastroenterology.2007;132(7):2557–2576.

3. Lupberger J, Hildt E. Hepatitis B virus-induced oncogenesis.World J Gastroenterol. 2007;13(1):74–81.

4. Bosch FX, Ribes J, Dıaz M, et al. Primary liver cancer:worldwide incidence and trends. Gastroenterology. 2004;127(5 suppl 1):S5–S16.

5. Evans AA, Chen G, Ross EA, et al. Eight-year follow-up of the90,000-person Haimen City cohort: I. Hepatocellular carci-noma mortality, risk factors, and gender differences. CancerEpidemiol Biomarkers Prev. 2002;11(4):369–376.

6. Rothman KJ. Epidemiology: An Introduction. Oxford, NY:Oxford University Press; 2002.

7. Rothman KJ, Greenland S. Modern Epidemiology. Philadel-phia, PA: Lippincott-Raven; 1998.

8. Kalilani L, Atashili J. Measuring additive interaction usingodds ratios [electronic article]. Epidemiol Perspect Innov.2006;3:5.

9. Andersson T, Alfredsson L, Kallberg H, et al. Calculatingmeasures of biological interaction. Eur J Epidemiol. 2005;20(7):575–579.

10. Assmann SF, Hosmer DW, Lemeshow S, et al. Confidenceintervals for measures of interaction. Epidemiology. 1996;7(3):286–290.

11. Li R, Chambless L. Test for additive interaction in propor-tional hazards models. Ann Epidemiol. 2007;17(3):227–236.

12. Yu MW, Chang HC, Chang SC, et al. Role of reproductivefactors in hepatocellular carcinoma: impact on hepatitis B- andC-related risk. Hepatology. 2003;38(6):1393–1400.

13. Naugler WE, Sakurai T, Kim S, et al. Gender disparity in livercancer due to sex differences in MyD88-dependent IL-6production. Science. 2007;317(5834):121–124.

14. Lawrence T, Hageman T, Balkwill F. Cancer: sex, cytokines,and cancer. Science. 2007;317(5834):51–52.

15. Wang Y, Cui F, Lv Y, et al. HBsAg and HBx knocked intothe p21 locus causes hepatocellular carcinoma in mice.Hepatology. 2004;39(2):318–324.

16. Yu MW, Yang YC, Yang SY, et al. Hormonal markers andhepatitis B virus-related hepatocellular carcinoma risk:a nested case-control study among men. J Natl Cancer Inst.2001;93(21):1644–1651.

17. Sung YM, Tang NL, Lai PB, et al. Re: Hormonal markers andhepatitis B virus-related hepatocellular carcinoma risk:a nested case-control study among men [letter]. J Natl CancerInst. 2003;95(7):559–560.

18. Prieto J. Inflammation, hepatocellular carcinoma and sex:IL-6 in the centre of the triangle. J Hepatol. 2008;48(2):380–381.

19. Wands J. Hepatocellular carcinoma and sex. N Engl J Med.2007;357(19):1974–1976.

20. Shimizu I, Kohno N, Tamaki K, et al. Female hepatology:favorable role of estrogen in chronic liver disease with

994 Wang et al.

Am J Epidemiol 2009;169:990–995

hepatitis B virus infection.World J Gastroenterol. 2007;13(32):4295–4305.

21. Vivekanandan P, Kannangai R, Ray SC, et al. Comprehensivegenetic and epigenetic analysis of occult hepatitis B from livertissue samples. Clin Infect Dis. 2008;46(8):1227–1236.

22. Torbenson M, Thomas DL. Occult hepatitis B. Lancet InfectDis. 2002;2(8):479–486.

23. Chemin I, Trepo C. Clinical impact of occult HBV infections.J Clin Virol. 2005;34(suppl 1):S15–S21.

24. Marrero CR, Marrero JA. Viral hepatitis and hepatocellularcarcinoma. Arch Med Res. 2007;38(6):612–620.

25. Zhang HW, Yin JH, Li YT, et al. Risk factors for acute hep-atitis B and its progression to chronic hepatitis in Shanghai,China. Gut. 2008;57(12):1713–20.


Am J Epidemiol 2009;169:990–995




Vol. 169, No. 8

DOI: 10.1093/aje/kwn414



Serum Selenium and Peripheral Arterial Disease: Results From the NationalHealth 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 ofperipheral arterial disease among 2,062 US men and women 40 years of age or older participating in the NationalHealth 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 diseasedecreased with increasing serum selenium (P for linear trend ¼ 0.02), but there was an indication of an upturn inrisk 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 splineregression models, peripheral arterial disease prevalence decreased with increasing serum selenium levels upto 150–160 ng/mL, followed by a gradual increase at higher selenium levels. The association between serumselenium levels and the prevalence of peripheral arterial disease was not statistically significant, althougha U-shaped relation was suggested.

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

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

Selenium, an essential micronutrient involved in anti-oxidant selenoenzymes such as glutathione peroxidases,has been hypothesized to prevent atherosclerotic disease(1–3). Glutathione peroxidase synthesis and activity,however, plateau at selenium levels above 70–90 ng/mL(4). In the United States compared with other countries,selenium intake is substantially higher (5), and mostadults have serum selenium levels above 95 ng/mL (4).At these high concentrations, selenium is incorporatednonspecifically as selenomethionine in the synthesis ofother plasma proteins, with unknown health effects (4).It is thus unclear whether increased selenium levels con-fer additional benefit for atherosclerosis prevention in theUnited States.

A recent meta-analysis of 14 prospective cohort studiesfound a modest, but statistically significant inverse associa-tion between selenium levels and coronary heart disease (3),but the 2 US studies in this meta-analysis found no associ-ation (3, 6, 7). Moreover, serum selenium levels were notassociated with cardiovascular disease mortality in a pro-spective study in a representative US sample, althougha U-shaped relation was suggested (8). Given the currentinterest in selenium supplements for chemoprevention ofcancer (5), it is important to understand the overall impactof increased selenium intake on other health endpoints, in-cluding vascular disease.

Peripheral arterial disease, which affects about 8 millionAmericans, 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: [email protected]).

996 Am J Epidemiol 2009;169:996–1003

in the muscular arteries of the lower extremities and is animportant marker of generalized atherosclerosis (9–11).Data on the association of selenium levels with peripheralarterial disease are very limited (12, 13). The objective ofthe present study was to assess the association betweenserum levels of selenium and reduced ankle-brachialblood pressure index (ABI), a specific subclinical markerfor peripheral arterial disease, in the National Health andNutrition Examination Survey (NHANES), 2003–2004. ABIvalues below 0.90 are considered diagnostic of peripheralarterial disease. Furthermore, a low ABI is an independentpredictor of cardiovascular risk after adjusting for traditionalcardiovascular risk factors (14, 15).


NHANES is conducted by the National Center forHealth Statistics (Hyattsville, Maryland) by using a com-plex multistage sampling design to obtain a representativesample of the civilian, noninstitutionalized US popula-tion. We used data from NHANES 2003–2004 (16) be-cause it was the first NHANES survey to measureselenium and ABI levels simultaneously. In NHANES2003–2004 interviews and physical examinations, theoverall response rate was 76%. Serum selenium andABI measurements were restricted to participants aged40 years or older (N ¼ 3,086). We excluded 2 pregnantwomen, 183 participants without selenium measure-ments, 514 participants without ABI measurements inboth legs, and 314 participants with missing informationon any adjustment covariate. We finally excluded 11 par-ticipants with left or right ABI measurements of morethan 1.5, usually due to vessel stiffness. The final samplesize was 2,062. The 2003–2004 NHANES study protocolswere approved by the National Center for Health Statis-tics institutional review board. Oral and written informedconsent was obtained from all participants.

Serum selenium

Collection materials were screened for potential seleniumcontamination. After blood collection, serum aliquots wereobtained, frozen at �20�C, and shipped to the Trace Ele-ments Laboratory at the Wadsworth Center of the New YorkState Department of Health for analysis. Serum seleniumlevels were measured by using inductively coupledplasma-dynamic reaction cell-mass spectrometry. The lab-oratory procedures and quality control methods for serumselenium measurement have been described in detail else-where (17). The between-assay coefficients of variation forquality-control pooled samples analyzed throughout the du-ration of the survey ranged from 2.5% to 2.9%.

Peripheral arterial disease

A specific protocol was used to measure ABI inNHANES 2003–2004 (18). The measurements of bloodpressure used for ABI were additional to and different fromother measurements of blood pressure used to evaluate hy-pertension. Systolic blood pressure was measured on the

right arm (brachial artery) and both ankles (posterior tibialarteries) with a Doppler device, the Parks Mini-Lab IV, model3100 (Parks Medical Electronics, Inc., Aloha, Oregon). If theparticipant had a condition that would interfere with bloodpressure reading in the right arm, the left arm was used.Systolic blood pressure was measured twice at each site forparticipants aged 40–59 years and once at each site forparticipants aged 60 years or older. The left and rightABI measurements were obtained by dividing the meansystolic blood pressure in each ankle by the mean systolicblood pressure in the arm. Peripheral arterial disease wasdefined as an ABI value of less than 0.90 in at least one leg(14, 15).

Other variables

Information about age, sex, race-ethnicity, education,family income, menopausal status for women, cigarettesmoking, alcohol consumption, use of dietary supplements,and use of cholesterol- and blood-pressure-lowering medi-cations was based on self-report. Body mass index was cal-culated by dividing measured weight in kilograms bymeasured height in meters squared. Three to 4 systolic bloodpressure measurements were taken and were averaged byusing standardized protocols. Diabetes was defined as a fast-ing serum glucose concentration of 126 mg/dL or higher,a nonfasting serum glucose concentration of 200 mg/dL orhigher, a self-reported physician diagnosis, or current med-ication use. Glomerular filtration rate was estimated by us-ing the Modification of Diet in Renal Disease Studyequation with serum creatinine values (19).

Statistical methods

Participants were grouped in quartiles of serum sele-nium levels based on the weighted population distribu-tion. Odds ratios and 95% confidence intervals forperipheral arterial disease prevalence comparing the 3highest quartiles of serum selenium with the lowest quar-tile were estimated by using logistic regression. Tests forlinear risk trend across serum selenium quartiles wereperformed by including an ordinal variable with themedian selenium level of each quartile in the logistic re-gression models. To further explore the shape of the dose-response relation between serum selenium levels andperipheral arterial disease prevalence, we used restrictedquadratic splines with knots at the 5th, 50th, and 95thpercentiles of serum selenium distribution. These splinemodels require the same number of parameters as thequartile analysis, but they can accommodate a wide vari-ety of smooth risk trends (20). Sensitivity analyses usingdifferent numbers and locations of the knots, with cubicinstead of quadratic splines, and log-transforming serumselenium levels gave similar results (not shown). Statisti-cal analyses were performed with the survey package in Rsoftware (to account for the complex sampling design inNHANES 2003–2004) (21, 22). Strata, primary samplingunits, and examination sample weights were used to ob-tain unbiased point estimates and robust linearized stan-dard errors.

Selenium and Peripheral Arterial Disease 997

Am J Epidemiol 2009;169:996–1003

RESULTS

The weighted prevalence of peripheral arterial disease inthe study population was 4.9%. Compared with participantswithout peripheral arterial disease, those with disease weremore likely to be older, black, ever smokers, nondrinkers, anddiabetic; to have a lower educational level and family in-come; and to use cholesterol- and blood-pressure-loweringmedications (Table 1). Participants with peripheral arterialdisease, compared with those without disease, also had higheraverage levels of body mass index, systolic blood pressure,and C-reactive protein and lower high density lipoproteincholesterol levels and glomerular filtration rate.

Participants in the highest quartile of serum seleniumlevels, compared with those in the lowest quartile, weremore likely to be older, men, white, and nonsmokers andto use dietary supplements and cholesterol-lowering med-ications (Table 2). Serum selenium levels were also positive-ly associated with total cholesterol, high density lipoprotein

cholesterol, and systolic blood pressure and were inverselyassociated with body mass index and serum cotinine levels.

The age-, sex-, and race-adjusted prevalence of peripheralarterial disease decreased with increasing serum seleniumlevels (P for linear trend ¼ 0.02) (Table 3). However, therewas an indication of an upturn in risk trend in the highestquartile of serum selenium, particularly after adjusting forcardiovascular risk factors. The fully adjusted odds ratiosfor peripheral arterial disease comparing selenium quartiles2, 3, and 4 with the lowest quartile were 0.75 (95% confi-dence interval: 0.37, 1.52), 0.58 (95% confidence interval:0.28, 1.19), and 0.67 (95% confidence interval: 0.34, 1.31),respectively. In spline regression models, peripheral arterialdisease prevalence decreased with increasing serum sele-nium levels up to 150–160 ng/mL (80th–91st percentilesof the serum selenium distribution in the study population),followed by a gradual increase at higher selenium levels(Figure 1). Consistently, there was a marginally significantU-shaped dose-response relation between serum selenium

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

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

PeripheralArterial Disease

(n 5 169)

No PeripheralArterial Disease(n 5 1,893)

P value

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 lipoproteincholesterol, mg/dL

51.8 (1.1) 54.8 (0.5) 0.03

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<60 mL/minute per 1.73 m2

26.1 (4.5) 10.3 (0.8) <0.001

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.

998 Bleys et al.

Am J Epidemiol 2009;169:996–1003

and peripheral arterial disease prevalence (P for quadraticspline terms ¼ 0.06 and 0.12 after adjustment for age, sex,and race and after full adjustment, respectively).

DISCUSSION

In this cross-sectional study, conducted in a representativesample of the US population, the association between serumselenium levels and the prevalence of peripheral arterialdisease was not statistically significant, although a U-shapedrelation was suggested: the prevalence of peripheral arterialdisease decreased with increasing serum selenium levels upto 150 ng/mL but increased with increasing selenium levelsabove 160 ng/mL. Selenium intake varies around the worldprimarily because of geographic variation in the amount of

selenium in the soil (1, 23). In the United States, estimatedselenium intake ranges from 60 lg/day to 220 lg/day (23),higher than the recommended dietary allowance for healthyadults (55 lg/day) (4). As a consequence, serum seleniumlevels in the United States are high: in NHANES 2003–2004, the median selenium level was 134 ng/mL, and 99%of study participants had serum selenium levels above 95ng/mL. These concentrations are considerably higher thanin other countries. In Europe, for instance, average serumselenium levels range from 50 ng/mL to 90 ng/mL (23, 24).

Very limited data are available on the association ofselenium with peripheral arterial disease. Two small studiesfound similar selenium levels in patients with peripheralarterial disease compared with controls, but the dose-response relation was not evaluated (12, 13). For other car-diovascular outcomes, most prospective studies have been

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

Linear TrendbQuartile 1(<125)

Quartile 2(125–134)

Quartile 3(135–147)

Quartile 4(‡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<60 mL/minute per 1.73 m2

5.6 6.0 6.3 7.3 0.24

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.


Am J Epidemiol 2009;169:996–1003

conducted in populations with suboptimal selenium levels inEurope or China (3, 25–36). These studies tended to reportinverse associations between serum selenium levels and cor-onary heart disease incidence, but their sample sizes weretoo small for detailed dose-response analyses.

Findings from the only 2 prospective studies of serumselenium levels and coronary heart disease conducted inthe United States, however, are consistent with a U-shapedrelation (6, 8). In the Physicians’ Health Study, the relativerisks for incident myocardial infarction comparing quintiles2–5 of plasma selenium with the lowest quintile were 0.87,0.82, 0.60, and 1.53, respectively (6). The cutoff levels forquintiles 1 and 5 of serum selenium in this study were 92ng/mL and 134 ng/mL, respectively. In the NHANES IIIMortality Study, the relative risks for cardiovascular dis-ease mortality comparing tertiles 2 and 3 of serum sele-nium with the lowest tertile were 0.90 and 0.98,respectively; for stroke mortality, the corresponding rel-ative risks were 0.73 and 1.23 (8). The cutoff levels forserum selenium tertiles in NHANES III were 117.3 ng/mL and130.4 ng/mL, respectively. In this study, a dose-response anal-ysis showed that cardiovascular and coronary heart diseasemortality decreased with increasing serum selenium levelsup to 120 ng/mL followed by an increase at higher levels,although the U-shaped relation was not statistically significant(8). Finally, in the Health Professionals Follow-up Study, theodds ratios for incident coronary heart disease comparingquintiles 2–5 of toenail selenium levels with the first quintilewere 1.03, 0.99, 1.32, and 0.86, respectively, with no cleardose-response relation (7). Both serum and toenail seleniumlevels reflect selenium status, although toenails reflect longer-term exposure. It is unclear, however, whether both biomarkersare comparable in their ability to capture the different types ofselenium compounds.

Few randomized trials have evaluated the effect of seleniumsupplementation on cardiovascular outcomes or atherosclero-sis progression, and most of these studies combined seleniumwith other vitamins and minerals (3, 37). Only 2 of these trialswere conducted in the United States, both reporting null results(38, 39). In the Nutritional Prevention of Cancer trial, therelative risk for cardiovascular disease incidence comparing

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 forLinear TrendaQuartile 1 (<125) Quartile 2 (125–134) Quartile 3 (135–147) Quartile 4 (‡148)

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

Model 2c 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).

100 120 140 160 180 200

0.25

0.5

1

2

0

5

10

15

20

25

Serum Selenium, ng/ml

Odd

s R

atio

for P

erip

hera

l Arte

rial D

isea

se

Wei

ghte

d Pe

rcen

tage

Figure 1. Odds ratios for peripheral arterial disease by serum sele-nium levels, National Health and Nutrition Examination Survey, 2003–2004. Curves represent adjusted odds ratios (solid line) and their 95%confidence intervals (dashed lines) based on restricted quadraticsplines for serum selenium levels with knots at the 5th, 50th, and95th percentiles. The reference value (odds ratio ¼ 1) was set at the10th percentile of serum selenium distribution (116 ng/mL). Odd ratioswere adjusted for age, sex, race-ethnicity, education, family income,postmenopausal status, smoking, serum cotinine, alcohol consump-tion, bodymass index, dietary supplement use,C-reactive protein, totalcholesterol, high density lipoprotein cholesterol, cholesterol-loweringmedication use, systolic blood pressure, blood-pressure-loweringmedication use, diabetes, and glomerular filtration rate. Bars representthe weighted histogram of serum selenium distribution.

1000 Bleys et al.

Am J Epidemiol 2009;169:996–1003

200 lg/day of selenium supplementation with placebo was1.03 (95% confidence interval: 0.78, 1.37) (38). In the HDL-Atherosclerosis Treatment Study, the progression of athero-sclerosis measured by coronary angiography in patients withcoronary artery disease was similar among participants ran-domized to an antioxidant supplement containing 100 lg/dayof selenium, 800 IU/day of vitamin E, 1 g/day of vitamin C,and 25 mg/day of b-carotene and participants randomized toplacebo (39). Overall, limited evidence from randomized trialshas not shown a protective effect of selenium supplementationin US studies. With respect to observational studies, those inthe United States have not been able to detect a significantlinear association between serum selenium and cardiovascularoutcomes, but the dose-response associations in these studieswere U-shaped.

The biologic mechanisms underlying a potential effect ofselenium on cardiovascular disease are likely complex, butthey may be related to the dual role of selenium as an es-sential and toxic element. Selenium is an essential micro-nutrient that is incorporated into glutathione peroxidasesand other selenoproteins (4). Increasing serum seleniumlevels increase the concentration and activity of glutathioneperoxidases, but this dose-response relation reaches a pla-teau at serum selenium levels of 70–90 ng/mL (4). As a con-sequence, higher selenium levels could potentially preventatherosclerosis development and progression in populationswhose selenium exposure is below the levels needed tomaximize glutathione peroxidases (1–3). In selenium-replete populations such as in the United States, in whichvirtually all participants have serum selenium levels above70–90 ng/mL, the mechanisms underlying a potential ben-eficial effect of increased selenium levels are unclear. Sinceselenium supplementation is actively promoted in theUnited States, and large randomized controlled trials testingthe efficacy of selenium supplementation in prostate cancerprevention are under way (40, 41), mechanistic studies areurgently needed to establish the biologic basis for a protec-tive effect of selenium in populations whose selenium statusis already high.

Selenium, however, has a narrow therapeutic range (4),and it may even be harmful at intake levels below the currenttolerable Upper Intake Level of 400 lg/day (2). In fact,some selenium compounds have been documented to gen-erate reactive oxygen species (42–44), and the upturn inperipheral arterial disease prevalence that we observed atselenium levels above 160 ng/mL could be associated withselenium-induced increased oxidative stress. This upturn inrisk is also consistent with recent reports showing increasedrisk of diabetes (45, 46) and elevated lipid levels (47) withhigh selenium levels in US populations. For instance, theNutritional Prevention of Cancer trial showed an increasedrisk of diabetes for participants receiving 200 lg/day ofselenium compared with placebo (hazard ratio ¼ 1.50,95% confidence interval: 1.03, 2.33) (46). Interestingly,the excess risk was limited to participants in the upper tertileof the serum selenium distribution (>121.6 ng/mL), whohad a hazard ratio for diabetes of 2.70 (95% confidenceinterval: 1.30, 5.61). Further research is needed to establishthe mechanisms underlying the association of high-normalselenium levels with peripheral arterial disease and with

metabolic abnormalities, and to determine whether thechange point in risk associated with elevated selenium levelsdepends on genetic polymorphisms in candidate genes forselenium metabolism (48).

Several limitations of our study need to be considered.The use of a cross-sectional design and of prevalent cases ofperipheral arterial disease limited our ability to determinethe direction and the causality of the observed association. Itis possible that the pathophysiologic changes of atheroscle-rosis could modify serum selenium levels or that partici-pants with peripheral arterial disease change their healthbehaviors, including selenium intake through diet and di-etary supplements. As a consequence, our findings must beconfirmed in prospective studies with incident cases of pe-ripheral arterial disease. Another limitation of our study isthe use of a single measurement of serum selenium, whichreflects short-term selenium intake and may be subject tohigh within-person variability (49). Furthermore, our studymeasured only total serum selenium, and we did not haveinformation on selenoprotein levels or activity or about non-specific incorporation of selenium as selenomethionine inother plasma proteins. More detailed analysis of differentcompartments of serum selenium will be needed to betterunderstand the association of selenium with peripheralarterial disease. The strengths of our study come fromthe rigorous sampling design and the quality of the studymeasurements used in NHANES; the representativeness ofthe NHANES sample; and the use of ABI, a noninvasivemeasure of subclinical atherosclerosis.

In summary, the association between serum selenium lev-els and the prevalence of peripheral arterial disease inNHANES 2003–2004 was not statistically significant,although a U-shaped relation was suggested. Other sourcesof evidence (6, 8) also suggest a U-shaped relation betweenserum selenium levels and cardiovascular outcomes in theUnited States, a selenium-replete population. In many pop-ulations worldwide, selenium intake is lower than in theUnited States (1, 23). At these lower levels of seleniumintake, the association of selenium with peripheral arterialdisease remains unknown. Prospective studies of seleniumstatus across populations with different levels of seleniumintake and randomized trials stratified by baseline sele-nium status are needed to establish the optimal seleniumlevels to minimize the risk of cardiovascular and otherchronic diseases.

ACKNOWLEDGMENTS

Author affiliations: Department of Epidemiology, JohnsHopkins Bloomberg School of Public Health, Baltimore,Maryland (Joachim Bleys, Ana Navas-Acien, MartinLaclaustra, Andy Menke, Eliseo Guallar); Department ofMedicine, Johns Hopkins School of Medicine, Baltimore,Maryland (Eliseo Guallar); Welch Center for Prevention,Epidemiology and Clinical Research, Johns HopkinsBloomberg School of Public Health and Johns HopkinsSchool of Medicine, Baltimore, Maryland (Joachim Bleys,Ana Navas-Acien, Martin Laclaustra, Andy Menke, Eliseo


Am J Epidemiol 2009;169:996–1003

Guallar); Department of Environmental Health Sciences,Johns Hopkins Bloomberg School of Public Health,Baltimore, Maryland (Ana Navas-Acien); Department ofCardiovascular Epidemiology and Population Genetics,National Center for Cardiovascular Research (CNIC),Madrid, Spain (Martin Laclaustra, Eliseo Guallar); NationalCenter for Epidemiology, Instituto de Salud Carlos III, andthe CIBER in Epidemiology and Public Health (CIBERESP),Madrid, Spain (Roberto Pastor-Barriuso); Friedman Schoolof Nutrition Science and Policy at Tufts University, Boston,Massachusetts (Jose M. Ordovas); and Clinical SciencesResearch Institute at Warwick Medical School, Coventry,United Kingdom (Saverio Stranges).

Supported by grants 1 R01 ES012673 from the NationalInstitute of Environmental Health Sciences and 0230232Nfrom the American Heart Association.


REFERENCES

1. Rayman MP. Food-chain selenium and human health:emphasis on intake. Br J Nutr. 2008;100(2):254–268.

2. Navas-Acien A, Bleys J, Guallar E. Selenium intake andcardiovascular risk: what is new? Curr Opin Lipidol. 2008;19(1):43–49.

3. Flores-Mateo G, Navas-Acien A, Pastor-Barriuso R, et al.Selenium and coronary heart disease: a meta-analysis. Am JClin Nutr. 2006;84(4):762–773.

4. Food and Nutrition Board, Institute of Medicine. DietaryReference Intakes for Vitamin C, Vitamin E, Selenium, andCarotenoids. A report of the Panel on Dietary Antioxidantsand Related Compounds, Subcommittees on Upper ReferenceLevels of Nutrients and Interpretation and Uses of DietaryReference Intakes, and the Standing Committee on the Sci-entific Evaluation of Dietary Reference Intakes. Washington,DC: The National Academies Press; 2000.

5. Rayman MP. Selenium in cancer prevention: a review of theevidence and mechanism of action. Proc Nutr Soc. 2005;64(4):527–542.

6. Salvini S, Hennekens CH, Morris JS, et al. Plasma levels ofthe antioxidant selenium and risk of myocardial infarctionamong U.S. physicians. Am J Cardiol. 1995;76(17):1218–1221.

7. Yoshizawa K, Ascherio A, Morris JS, et al. Prospective studyof selenium levels in toenails and risk of coronary heart dis-ease in men. Am J Epidemiol. 2003;158(9):852–860.

8. Bleys J, Navas-Acien A, Guallar E. Serum selenium levels andall-cause, cancer, and cardiovascular mortality among USadults. Arch Intern Med. 2008;168(4):404–410.

9. McDermott MM, Liu K, Criqui MH, et al. Ankle-brachialindex and subclinical cardiac and carotid disease: theMulti-Ethnic Study of Atherosclerosis. Am J Epidemiol. 2005;162(1):33–41.

10. Murabito JM, Evans JC, Larson MG, et al. The ankle-brachialindex in the elderly and risk of stroke, coronary disease, anddeath: the Framingham Study. Arch Intern Med. 2003;163(16):1939–1942.

11. Newman AB, Shemanski L, Manolio TA, et al. Ankle-armindex as a predictor of cardiovascular disease and mortality inthe Cardiovascular Health Study. The Cardiovascular HealthStudy Group. Arterioscler Thromb Vasc Biol. 1999;19(3):538–545.

12. Ondrus P, Alberty R, Vassanyiova Z. Importance of lipid per-oxidation, protective enzymes and trace elements in chronicleg ischaemia. Eur J Clin Chem Clin Biochem. 1996;34(6):471–475.

13. Mansoor MA, Bergmark C, Haswell SJ, et al. Correlationbetween plasma total homocysteine and copper in patientswith peripheral vascular disease. Clin Chem. 2000;46(3):385–391.

14. Hirsch AT, Haskal ZJ, Hertzer NR, et al. ACC/AHA 2005Practice Guidelines for the management of patients withperipheral arterial disease (lower extremity, renal, mesen-teric, and abdominal aortic): a collaborative report from theAmerican Association for Vascular Surgery/Society forVascular Surgery, Society for Cardiovascular Angiographyand Interventions, Society for Vascular Medicine andBiology, Society of Interventional Radiology, and the ACC/AHA Task Force on Practice Guidelines (Writing Committeeto Develop Guidelines for the Management of Patients WithPeripheral Arterial Disease): endorsed by the AmericanAssociation of Cardiovascular and Pulmonary Rehabilitation;National Heart, Lung, and Blood Institute; Society for Vas-cular Nursing; TransAtlantic Inter-Society Consensus; andVascular Disease Foundation. Circulation. 2006;113(11):e463–e654.

15. Fowkes FG, Murray GD, Butcher I, et al. Ankle brachial indexcombined with Framingham Risk Score to predict cardiovas-cular events and mortality: a meta-analysis. JAMA. 2008;300(2):197–208.

16. National Center for Health Statistics, Centers for DiseaseControl and Prevention. National Health and NutritionExamination Survey, 2003–2004. 2007. (http://www.cdc.gov/nchs/about/major/nhanes/nhanes2003-2004/nhanes03_04.htm).(Accessed August 24, 2008).

17. National Center for Health Statistics, Centers for DiseaseControl and Prevention. NHANES 2003–2004 Data Docu-mentation. Laboratory Assessment: Lab 39—ErythrocyteProtoporphyrin and Selenium. 2007. (http://www.cdc.gov/nchs/data/nhanes/nhanes_03_04/l39epp_c.pdf). (AccessedAugust 24, 2008).

18. National Center for Health Statistics, Centers for DiseaseControl and Prevention. NHANES 2003–2004 Data Docu-mentation. Exam Component: Ankle-Brachial Blood PressureIndex (ABPI) Section of the Lower Extremity Disease Ex-amination (LEXAB_C). 2005. (http://www.cdc.gov/nchs/data/nhanes/nhanes_03_04/lexab_c.pdf). (Accessed August24, 2008).

19. Levey AS, Coresh J, Greene T, et al. Using standardized serumcreatinine values in the Modification of Diet in Renal DiseaseStudy equation for estimating glomerular filtration rate. AnnIntern Med. 2006;145(4):247–254.

20. Greenland S. Dose-response and trend analysis in epidemiol-ogy: alternatives to categorical analysis. Epidemiology. 1995;6(4):356–365.

21. R Development Core Team. R: A Language and Environmentfor Statistical Computing. Vienna, Austria: R Foundation forStatistical Computing; 2008. (ISBN 3-900051-07-0). (http://www.R-project.org).

22. Lumley T. The Survey Package. (http://cran.fhcrc.org/web/packages/survey/survey.pdf). (Accessed August 24, 2008).

23. Combs GF Jr. Selenium in global food systems. Br J Nutr.2001;85(5):517–547.

24. Rayman MP. The importance of selenium to human health.Lancet. 2000;356(9225):233–241.

25. Salonen JT, Alfthan G, Huttunen JK, et al. Associationbetween cardiovascular death and myocardial infarction and

1002 Bleys et al.

Am J Epidemiol 2009;169:996–1003

http://www.cdc.gov/nchs/about/major/nhanes/nhanes2003-2004/nhanes03_04.htm

http://www.cdc.gov/nchs/about/major/nhanes/nhanes2003-2004/nhanes03_04.htm

http://www.cdc.gov/nchs/data/nhanes/nhanes_03_04/l39epp_c.pdf

http://www.cdc.gov/nchs/data/nhanes/nhanes_03_04/l39epp_c.pdf

http://www.cdc.gov/nchs/data/nhanes/nhanes_03_04/lexab_c.pdf

http://www.cdc.gov/nchs/data/nhanes/nhanes_03_04/lexab_c.pdf

http://www.R-project.org

http://www.R-project.org

http://cran.fhcrc.org/web/packages/survey/survey.pdf

http://cran.fhcrc.org/web/packages/survey/survey.pdf

serum selenium in a matched-pair longitudinal study. Lancet.1982;2(8291):175–179.

26. Miettinen TA, Alfthan G, Huttunen JK, et al. Serum seleniumconcentration related to myocardial infarction and fatty acidcontent of serum lipids. Br Med J (Clin Res Ed). 1983;287(6391):517–519.

27. Salonen JT, Salonen R, Penttila I, et al. Serum fatty acids,apolipoproteins, selenium and vitamin antioxidants and therisk of death from coronary artery disease. Am J Cardiol.1985;56(4):226–231.

28. Virtamo J, Valkeila E, Alfthan G, et al. Serum selenium andthe risk of coronary heart disease and stroke. Am J Epidemiol.1985;122(2):276–282.

29. Ringstad J, Thelle D. Risk of myocardial infarction in relationto serum concentrations of selenium. Acta Pharmacol Toxicol(Copenh). 1986;59(suppl 7):336–339.

30. Kok FJ, de Bruijn AM, Hofman A, et al. Selenium status andchronic disease mortality: Dutch epidemiological findings. IntJ Epidemiol. 1987;16(2):329–332.

31. Ringstad J, Jacobsen BK, Thomassen Y, et al. The TromsøHeart Study: serum selenium and risk of myocardial infarctiona nested case-control study. J Epidemiol Community Health.1987;41(4):329–332.

32. Suadicani P, Hein HO, Gyntelberg F. Serum selenium con-centration and risk of ischaemic heart disease in a prospectivecohort study of 3000 males. Atherosclerosis. 1992;96(1):33–42.

33. Marniemi J, Jarvisalo J, Toikka T, et al. Blood vitamins,mineral elements and inflammation markers as risk factors ofvascular and non-vascular disease mortality in an elderlypopulation. Int J Epidemiol. 1998;27(5):799–807.

34. Kilander L, Berglund L, Boberg M, et al. Education, lifestylefactors and mortality from cardiovascular disease and cancer.A 25-year follow-up of Swedish 50-year-old men. Int J Epi-demiol. 2001;30(5):1119–1126.

35. Wei WQ, Abnet CC, Qiao YL, et al. Prospective study ofserum selenium concentrations and esophageal and gastriccardia cancer, heart disease, stroke, and total death. Am J ClinNutr. 2004;79(1):80–85.

36. Akbaraly NT, Arnaud J, Hininger-Favier I, et al. Selenium andmortality in the elderly: results from the EVA study. ClinChem. 2005;51(11):2117–2123.

37. Bleys J, Miller ER III, Pastor-Barriuso R, et al. Vitamin-mineral supplementation and the progression of atherosclerosis:

a meta-analysis of randomized controlled trials. Am J Clin Nutr.2006;84(4):880–887.

38. Stranges S, Marshall JR, Trevisan M, et al. Effects of seleniumsupplementation on cardiovascular disease incidence andmortality: secondary analyses in a randomized clinical trial.Am J Epidemiol. 2006;163(8):694–699.

39. Brown BG, Zhao XQ, Chait A, et al. Simvastatin and niacin,antioxidant vitamins, or the combination for the prevention ofcoronary disease. N Engl J Med. 2001;345(22):1583–1592.

40. Lippman SM, Goodman PJ, Klein EA, et al. Designing theSelenium and Vitamin E Cancer Prevention Trial (SELECT).J Natl Cancer Inst. 2005;97(2):94–102.

41. Marshall JR, Sakr W, Wood D, et al. Design and progress ofa trial of selenium to prevent prostate cancer among men withhigh-grade prostatic intraepithelial neoplasia. CancerEpidemiol Biomarkers Prev. 2006;15(8):1479–1484.

42. Spallholz JE. On the nature of selenium toxicity and carcino-static activity. Free Radic Biol Med. 1994;17(1):45–64.

43. Spallholz JE, Palace VP, Reid TW. Methioninase andselenomethionine but not Se-methylselenocysteine generatemethylselenol and superoxide in an in vitro chemiluminescentassay: implications for the nutritional carcinostatic activityof selenoamino acids. Biochem Pharmacol. 2004;67(3):547–554.

44. Drake EN. Cancer chemoprevention: selenium as a prooxi-dant, not an antioxidant. Med Hypotheses. 2006;67(2):318–322.

45. Bleys J, Navas-Acien A, Guallar E. Serum selenium anddiabetes in U.S. adults. Diabetes Care. 2007;30(4):829–834.

46. Stranges S, Marshall JR, Natarajan R, et al. Effects of long-term selenium supplementation on the incidence of type 2diabetes: a randomized trial. Ann Intern Med. 2007;147(4):217–223.

47. Bleys J, Navas-Acien A, Stranges S, et al. Serum selenium andserum lipids in US adults. Am J Clin Nutr. 2008;88(2):416–423.

48. Meplan C, Crosley LK, Nicol F, et al. Genetic polymorphismsin the human selenoprotein P gene determine the response ofselenoprotein markers to selenium supplementation in a gender-specific manner (the SELGEN study). FASEB J. 2007;21(12):3063–3074.

49. Longnecker MP, Stampfer MJ, Morris JS, et al. A 1-y trial ofthe effect of high-selenium bread on selenium concentrationsin blood and toenails. Am J Clin Nutr. 1993;57(3):408–413.


Am J Epidemiol 2009;169:996–1003


Published by the Johns Hopkins Bloomberg School of Public Health 2009.

Vol. 169, No. 8

DOI: 10.1093/aje/kwp011



Ambient Air Pollution and Cardiovascular Malformations in Atlanta, Georgia,1986–2003

Matthew J. Strickland, Mitchel Klein, Adolfo Correa, Mark D. Reller, William T. Mahle, TiffanyJ. 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 cardiovascularmalformations were investigated among the cohort of pregnancies reaching at least 20 weeks’ gestation that wereconceived during January 1, 1986–March 12, 2003, in Atlanta, Georgia. Surveillance records obtained from theMetropolitan Atlanta Congenital Defects Program, which conducts active, population-based surveillance on thiscohort, were reviewed to classify cardiovascular malformations. Ambient 8-hour maximum ozone and 24-houraverage 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. Temporalassociations 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 ofthe great arteries, muscular ventricular septal defect, perimembranous ventricular septal defect, conotruncaldefects, left ventricular outflow tract defect, and right ventricular outflow defect were modeled by using Poissongeneralized linear models. A statistically significant association was observed between PM10 and patent ductusarteriosus (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 withan average aerodynamic diameter of <10 lm; RR, risk ratio.

A growing body of epidemiologic evidence suggests as-sociations between ambient air pollution and adverse preg-nancy outcomes (1–6). Associations between air pollutionlevels during pregnancy and risks of cardiovascular malfor-mations among the offspring were investigated in 2 previouspopulation-based case-control studies (7, 8). In these stud-ies, cardiovascular malformations were classified by usingpreexisting surveillance database codes, and analyses werebased on contrasts in ambient pollution levels over spaceand time.

In the first study, conducted in southern California, inves-tigators reported an association between ambient carbonmonoxide levels and risk of ventricular septal defects

(fourth quartile vs. first quartile, odds ratio ¼ 2.95, 95%confidence interval (CI): 1.44, 6.05). Elevated risks of aorticartery and valve defects, pulmonary artery and valve de-fects, and conotruncal defects with increasing ambientozone levels were also reported (7). The second investiga-tion, conducted in Texas, did not corroborate the southernCalifornia findings, although a suggestive association be-tween ozone and pulmonary artery and valve defects wasobserved. The Texas investigators reported positive associ-ations for carbon monoxide and tetralogy of Fallot, particu-late matter with an average aerodynamic diameter of <10 lm(PM10) and atrial septal defects, and sulfur dioxide and ven-tricular septal defects (8).

Correspondence to Dr. Matthew Strickland, Department of Environmental and Occupational Health, Emory University, 1518 Clifton Road NE,

Atlanta, GA 30322 (e-mail: [email protected]).

1004 Am J Epidemiol 2009;169:1004–1014

We conducted a retrospective cohort study in Atlanta,Georgia, to explore temporal associations between ambientair pollution levels during pregnancy and risks of cardiovas-cular malformations. We did not closely replicate the meth-odologies of the previous studies. In addition to theretrospective cohort design and the temporal analytical ap-proach, we reviewed and reclassified each surveillance rec-ord using a modified version of the International Pediatricand Congenital Cardiac Code implemented in the Society ofThoracic Surgeons Congenital Heart Surgery Database (9–13).This activity permitted classification of cardiovascularmalformations according to embryologic (rather than onlyanatomic) considerations.


Study population

Vital records for the cohort of liveborn and stillborn in-fants of at least 20 weeks’ gestation whose mothers residedin 1 of 5 central Atlanta counties at delivery were obtainedfrom the Office of Health Information and Policy, GeorgiaDivision of Public Health. The records of infants with an

indication of a cardiovascular malformation were obtainedfrom the Metropolitan Atlanta Congenital Defects Program(MACDP), which conducts active, population-based birthdefects surveillance on the cohort of pregnancies reachingat least 20 weeks’ gestation to mothers residing in 1 of 5central Atlanta counties at delivery/termination (14).MACDP ascertains livebirths, stillbirths, and elective termi-nations with major structural defects, chromosomal abnor-malities, and clinical syndromes diagnosed by the age of6 years. When available, details from echocardiography,catheterization, and surgical reports are abstracted; generalpregnancy information, such as gestational age and birthweight, is also collected.

Pregnancies with an estimated date of conception duringJanuary 1, 1986–March 12, 2003, were included in the anal-ysis. For each conception date, we estimated the number ofpregnancies with a cardiovascular malformation (numera-tor) and total pregnancies (denominator). Approximately0.3% of pregnancies with cardiovascular malformationsand 1.9% of total pregnancies were excluded because ofmissing or implausible gestational age information (definedas <20 weeks or >44 weeks). For each pregnancy, we es-timated the date of conception (assuming that conception

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 aorticarch 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 arteriosuspersisting for �6 weeks following delivery. Infants were excluded if thepatent ductus arteriosus was an obligatory shunt lesion or if patencywas maintained by prostaglandin infusion.

Pulmonary stenosis, valvar 312 Includes valvar and unspecified pulmonary stenosis, as well as dysplasticpulmonary 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 majoraortopulmonary collateral arteries, and tetralogy of Fallot-type double-outletright ventricle.

Transposition of the great arteries 165 Includes all types of transposition with concordant atrioventricular connections anddiscordant ventricular arterial connections, with or without ventricular septaldefect or left ventricular outflow tract obstruction. Also includes double-outletright 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 includesaortopulmonary window defect, all other double-outlet right-ventricle variants,interrupted aortic arch type B, unspecified interrupted aortic arch, vascularrings, 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 andatresia 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 ventricularseptum, tricuspid valve atresia, double-chambered right ventricle, and isolatedsupravalvar 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.

Ambient Air Pollution and Cardiovascular Malformations 1005

Am J Epidemiol 2009;169:1004–1014

occurred 14 days after the last menstrual period date) usingvital records data. Unfortunately, these estimates were un-reliable; for 30% of pregnancies, the last menstrual perioddate was on the 15th of the month. When we subtracted theclinical gestational age estimate from the birth date, toomany last menstrual period dates fell between the 12thand 18th of the month.

To compensate for this data quality limitation, we usedgestational age estimates from MACDP surveillance rec-ords (which are obtained from medical chart review andare frequently based on ultrasound measurements) to es-timate conception dates for pregnancies with cardiovascu-lar malformations, because day-of-month patterns werenot evident among these estimates. For the denominators,we modeled the daily count of conceptions. We calculatedthe average daily number of conceptions for each monthusing vital records information. We then created a dailytime-series data set, in which each day was assignedits monthly average, and fit a cubic spline with 6 knotsper year to it. The predicted values from this model wereused as the daily estimates of conceptions (n ¼ 715,500pregnancies).

Cardiovascular malformation outcome groups

Each MACDP surveillance record with a diagnosis ofa cardiovascular malformation was reviewed by a pediatriccardiologist and classified by using the Society of ThoracicSurgeons Congenital Heart Surgery Database, version 2.30,nomenclature (11–13, 15). This nomenclature is specific tocardiovascular malformations and is more detailed thanwhat is typically used in birth defects surveillance (16–18). We classified infants with isolated transient newborncardiac conditions (e.g., patent foramen ovale) as physio-

logically normal, and we placed infants with more than 1cardiovascular malformation in multiple outcome groupsonly when these malformations were thought to be embry-ologically independent; otherwise, we coded only the majorcardiovascular malformation. Further details about this ac-tivity are available (12).

For our analyses, we excluded infants who had normalcardiac physiology and those with identified trisomies, evi-dence of heterotaxy syndrome, and abnormal cardiac loop-ing. Results are presented for 12 outcome groups; 3 of theseare aggregate groupings of cardiovascular malformations(Table 1).

Ambient air quality data

Ambient air quality measurements of daily 8-hour max-imum ozone and 24-hour average carbon monoxide, nitro-gen dioxide, PM10, and sulfur dioxide were obtained fromthe US Environmental Protection Agency Air Quality Sys-tem, Georgia Department of Natural Resources, and theMetro Atlanta Index. For each pollutant, we selected 1 cen-tral monitoring station for use in analyses. When centralstation measurements of carbon monoxide, nitrogen diox-ide, and sulfur dioxide were missing, pollution levels at thecentral station were modeled by using measurements fromother stations. The central station for ozone did not operateduring November–February. During 1993–2003, wintertimeozone levels were modeled by using ozone measurementsfrom a nearby monitor, with the model developed by use ofdata during the later time period when wintertime ozonelevels were available. During 1986–1992, wintertime ozonelevels were modeled by using maximum temperature and1-hour maximum nitrogen dioxide measurements froma nearby monitor. Measurements of PM10 were available

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

No. ofCases

Prevalence,a by Season of ConceptionPrevalence,a by Year of

Conception OverallPrevalenceaMarch–

MayJune–August

September–November

December–February

1986–1991

1992–1997

1998–2003

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,perimembranous

546 6.6 7.8 8.4 7.7 4.8 8.1 9.5 7.6

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.

1006 Strickland et al.

Am J Epidemiol 2009;169:1004–1014

every sixth day during 1986–1992, Sunday–Thursday during1993–1995, and daily during 1996–2003; linear interpolationbetween measurements was used to estimate missing PM10

levels. The location of the PM10 central monitoring stationchanged on January 1, 1993, and on January 1, 1998; onJanuary 1, 1998, the measurement method changed fromthe federal reference method to the tapered element oscillat-ing microbalance method.


For each conception date (January 1, 1986–March 12,2003), we estimated the number of pregnancies with a par-ticular cardiovascular malformation (numerator) and totalpregnancies (denominator). All pregnancies on a particularconception date were assigned the same pollutant metric,which was a weighted average of the 35 daily ambient airpollution measurements during weeks 3–7 of pregnancy(a period when the 4 chambers, inflow tract, and outflow tractof the heart develop (19)). Relative weights were 0.7 formeasurements during weeks 3 and 7, 0.9 for measurementsduring weeks 4 and 6, and 1.0 for measurements during

week 5. We chose this a priori weighting scheme, whichemphasizes pollution levels during the center of the window,because of uncertainty in the date of conception estimates.

We then created 52 strata representing the week-of-yearas follows: across all calendar years we grouped January 1–January 7 as the first week of the year, January 8–January 14as the second week of the year, and so on. We includedFebruary 29, when present, in the ninth week of the year(February 26–March 4). The 52nd week of the year was8 days long (December 24–December 31).

We modeled temporal associations between ambient airpollution and daily risks of cardiovascular malformationsusing Poisson generalized linear models with a log linkand scaled variance estimates. We modeled the pollutionmetric as a continuous variable and used the natural loga-rithm of total conceptions as the offset. We included indi-cator variables for the 52 strata representing week-of-year tocontrol for potential confounding by factors with seasonalvariation, and we included a cubic spline for day of follow-up with 1 knot per year to control for long-term trends. Allrisk ratios and confidence intervals corresponded to an in-terquartile range increase in the ambient pollutant metric.

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

8-HourOzone,ppbb,c

24-HourPM10,

mg/m3 c,d

24-HourNitrogenDioxide,ppbc

24-HourCarbon

Monoxide,ppmc

24-HourSulfurDioxide,ppbc

Interquartile range 29.9 14.2 5.7 0.3 4.0

Mean value, by season ofconception

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 ofconception

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.


Am J Epidemiol 2009;169:1004–1014

Models were created by using R statistical software, version2.5.0 (20).

We performed several sensitivity analyses. In one sensi-tivity analysis, we relaxed the seasonal and long-term tem-

poral controls by replacing the week-of-year indicatorvariables with a cubic spline for day-of-year that had 3knots; instead of including yearly knots in the cubic splinefor day of follow-up, we placed knots once every 3 years.

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

8-HourOzone,ppb

24-HourPM10,mg/m3

24-HourNitrogenDioxide,

ppb

24-HourCarbon

Monoxide,ppm

24-HourSulfurDioxide,

ppb

8-Hour ozone, ppb 1.00

24-Hour PM10, lg/m3 0.49*** 1.00

24-Hour nitrogendioxide, ppb

0.45*** 0.46*** 1.00

24-Hour carbonmonoxide, ppm

0.07*** 0.32*** 0.41*** 1.00

24-Hour sulfurdioxide, ppb

0.30*** 0.41*** 0.39*** 0.23*** 1.00

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.

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

No. ofCases

8-HourOzone, ppb

24-HourPM10, mg/m

324-Hour Nitrogen

Dioxide, ppb24-Hour CarbonMonoxide, ppm

24-Hour SulfurDioxide, ppb

RiskRatio


RiskRatio


RiskRatio


RiskRatio


RiskRatio


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 heartsyndrome

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

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 greatarteries

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

Ventricular septal defect,muscular

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


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

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 outflowtract defect

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

Right ventricular outflowtract defect

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

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

We also investigated the effect of limiting analyses to singlegestation pregnancies, limiting analyses to infants with only1 cardiovascular malformation, using unweighted 5-weekpollution metrics, and using unweighted 9-week pollutionmetrics (the average of measurements during the first 63days of pregnancy).

RESULTS

Surveillance records of pregnancies reaching 20 weeks’gestation with an estimated date of conception duringJanuary 1, 1986–March 12, 2003, indication of a cardiovas-cular malformation, and no evidence of trisomy were iden-tified (n ¼ 7,102). Of these, only 4,639 (65%) were deemedto have a cardiovascular malformation following review andclassification. Infants with complex malformations involvingabnormal cardiac looping (n ¼ 86) or heterotaxy syndrome(n ¼ 96) were excluded from analysis, irrespective of theother specific cardiovascular lesions that might have beenpresent. From the remaining 4,457 infants, 3,338 (75%) wereincluded in 1 or more of the outcome groups shown in Table 1.Overall, there were 715,500 liveborn and stillborn infants withan estimated date of conception during this period.

The prevalence of cardiovascular malformations, shownin Table 2 by season and year of conception, suggested someseasonal variation across outcome groups. Whereas the ob-served prevalence of the relatively severe lesions (hypoplas-

tic left heart syndrome, transposition of the great arteries,and tetralogy of Fallot) remained stable over time, the ob-served prevalence of the less severe lesions (secundum atrialseptal defect, valvar pulmonary stenosis, muscular ventric-ular septal defect, and perimembranous ventricular septaldefect) increased markedly.

Descriptive statistics for the air pollution metrics, whichare weighted averages of daily ambient pollution measure-ments during weeks 3–7 of pregnancy, indicated that allpollutants had seasonal variation (Table 3). Levels ofPM10, nitrogen dioxide, and sulfur dioxide declined overtime. Pairwise Spearman partial correlation coefficientsfor the pollution metrics are presented in Table 4.

As shown in Table 5, we observed a statistically signifi-cant positive association between PM10 and patent ductusarteriosus (for an interquartile range increase in PM10 levels,risk ratio (RR) ¼ 1.60, 95% CI: 1.11, 2.31). The 95%confidence intervals for all other associations, presented inTable 5, included the null value.

Results from 5 sensitivity analyses are presented in Tables6–10. We consistently observed positive associations be-tween PM10 and patent ductus arteriosus (less stringent sea-sonal and long-term temporal control, RR ¼ 1.40, 95% CI:1.01, 1.95; limited to single gestation pregnancies, RR ¼1.57, 95% CI: 1.07, 2.28; limited to infants with only 1cardiovascular malformation, RR ¼ 1.70, 95% CI: 1.12,2.56; using an unweighted 5-week metric, RR ¼ 1.53, 95%

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 theWeighted 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

No. ofCases

8-HourOzone, ppb

24-HourPM10, mg/m

324-Hour Nitrogen



RiskRatio


RiskRatio


RiskRatio


RiskRatio


RiskRatio





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





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


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


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



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


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





Am J Epidemiol 2009;169:1004–1014

CI: 1.02, 2.29; using an unweighted 9-week metric, RR ¼1.71, 95% CI: 0.98, 3.00). We observed a positive, statisti-cally significant association between nitrogen dioxide andpatent ductus arteriosus in the analysis with less stringentseasonal and long-term temporal control (RR ¼ 1.40, 95%CI: 1.07, 1.83). Using unweighted 9-week pollution metrics,we observed positive, statistically significant associationsbetween nitrogen dioxide and both secundum atrial septaldefect (RR ¼ 1.58, 95% CI: 1.02, 2.29) and muscular ven-tricular septal defect (RR ¼ 1.20, 95% CI: 1.02, 1.41). Wealso observed negative, statistically significant associationsbetween ozone and right ventricular outflow tract defects inthe analysis limited to infants with only 1 cardiovascularmalformation (RR ¼ 0.52, 95% CI: 0.29, 0.93) and betweensulfur dioxide and right ventricular outflow tract defects inthe analysis based on an unweighted 5-week metric (RR ¼0.73, 95% CI: 0.54, 0.98).

DISCUSSION

We investigated temporal associations between ambient airpollution levels during weeks 3–7 of pregnancy and risks ofcardiovascular malformations. Except for the association be-tween PM10 and patent ductus arteriosus, all 95% confidenceintervals from the primary analysis were consistent with littleor no association. Many confidence intervals, particularly

those for ozone, were wide and were therefore compatiblewith both no effect and a harmful effect of air pollution.

Patent ductus arteriosus was not investigated in the 2 pre-vious studies (7, 8), likely because identification of infantswith congenital patent ductus arteriosus is difficult. In ourstudy population, we used restrictive criteria to exclude patentductus arteriosus in premature or newborn infants and when itoccurred as an obligate shunt lesion in the presence of othercardiovascular malformations. Of the 2,273 surveillance rec-ords we reviewed that contained a code for patent ductusarteriosus, the records for only 219 infants met these criteria.

We are unaware of experimental animal evidence support-ing an association between PM10 and patent ductus arterio-sus. Analogous epidemiologic evidence comes fromSwedish registry data, wherein a positive association wasobserved between first trimester maternal smoking and riskof patent ductus arteriosus among term infants (21). Oneplausible biologic mechanism for our result relates to fetalgrowth and development. Relative to term infants, prematureinfants have a much higher incidence of patent ductus (22,23). Associations between reduced fetal growth and mater-nal smoking (24), environmental tobacco smoke (25), andambient particulate matter (26, 27) have been observed inepidemiologic studies. If high PM10 levels in utero restrictfetal development, this could explain an association betweenPM10 and patent ductus arteriosus among term infants.

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

No. ofCases

8-HourOzone, ppb

24-HourPM10, mg/m

324-Hour Nitrogen



RiskRatio


RiskRatio


RiskRatio


RiskRatio


RiskRatio





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





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


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


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



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


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





Am J Epidemiol 2009;169:1004–1014

Similar to our study, most results reported in the 2 previousstudies were consistent with no association (7, 8). In southernCalifornia, investigators reported 4 significant positive associ-ations from 144 models. Twenty-four of these models aver-aged pollution levels over weeks 5–8 of pregnancy; the 4positive associations were observed during this window (7).The Texas study investigators observed 3 positive associationsfrom 75 models; air pollution levels were averaged duringweeks 3–8 of pregnancy (8). No significant association fromthe southern California study was replicated in the Texas study.

Because of our review and classification of surveillancerecords, comparison of results from our study with thosefrom previous studies was difficult. Through our review,we excluded infants with structurally normal hearts and tran-sient newborn conditions; 35% of the records we reviewedwere reclassified as ‘‘structurally normal.’’ Outcome groupswere based on embryologic considerations, and infants withmultiple congenital heart defect codes were included in mul-tiple outcome groups only when the malformations werethought to be embryologically independent. Consequently,our outcome groups differed from those used in previousstudies (7, 8). For example, in the southern California study(7), an association was observed between carbon monoxideand isolated ventricular septal defects (fourth quartile vs.first quartile, odds ratio ¼ 2.95, 95% CI: 1.44, 6.05). Thisoutcome group incorporated the 4 major types of ventricularseptal defects, as well as pulmonary atresia with ventricular

septal defect. In our study, we distinguished among the 4types of ventricular septal defects, because each is thought todevelop through unique embryologic mechanisms (28, 29).We analyzed perimembranous and muscular ventricular sep-tal defects as distinct outcome groups, and we included sub-arterial ventricular septal defects in the conotruncal defectsoutcome group (Table 1). There were too few inlet ventricularseptal defects to permit analysis. We grouped pulmonaryatresia with ventricular septal defect, which is the extremeend of the anatomic spectrum of tetralogy of Fallot (30), in thetetralogy of Fallot outcome group. To provide a more directcomparison with the southern California result, we performeda secondary analysis in which we ignored our classificationsand defined isolated ventricular septal defect according to thepreexisting codes in the MACDP database. Using our primaryanalytical approach, we observed no evidence for an associ-ation between carbon monoxide and isolated ventricular sep-tal defects (RR ¼ 0.98, 95% CI: 0.82, 1.16).

Our analytical approach and exposure assignment meth-ods were also different. Whereas the previous study inves-tigators conducted spatial-temporal analyses using pollutionmeasurements assigned to women on the basis of residentiallocation, we opted for a temporal approach using measure-ments from centrally located monitors. We implementeda temporal analysis because of our desire to preclude con-cerns about our study results based on arguments of spatialconfounding. We characterized pollution levels using

Table 8. Sensitivity Analysis Limited to Infants and Fetuses With Only 1 Cardiovascular Malformation, With Risk Ratios and 95% Confidence

Intervals for Associations Between theWeighted 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

No. ofCases

8-HourOzone, ppb

24-HourPM10, mg/m

324-Hour Nitrogen



RiskRatio


RiskRatio


RiskRatio


RiskRatio


RiskRatio





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





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


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


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



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


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





Am J Epidemiol 2009;169:1004–1014

a centrally located monitor because of practical limitations.Namely, geocoded data were available dating back only to1994, whereas the central monitor approach allowed us toutilize data back to 1986. Throughout much of follow-up,there were only 2 or 3 monitors for each primary pollutant,and we determined that some monitors were unduly im-pacted by local sources. We were hesitant to use these mea-surements, which likely better reflect local conditions thanpopulation exposure. The use of 1 central monitor for eachpollutant also provided some consistency throughoutfollow-up; otherwise, modeling assumptions would havebeen needed to account for changes in the number and lo-cation of monitoring stations. For the PM10 analyses, werelied on such assumptions, and cubic spline knots wereplaced on the dates when the monitor location and measure-ment method changed. To examine the effect of these as-sumptions, we stratified the data according to measurementmethod (federal reference method: 1986–1997; tapered el-ement oscillating microbalance: 1998–2003). For PM10 andpatent ductus arteriosus, we observed a strong associationduring 1986–1997 (RR ¼ 1.89, 95% CI: 1.26, 2.84) and noassociation during 1998–2003 (RR ¼ 0.78, 95% CI: 0.28,2.04). Perhaps the strong association observed during 1986–1997 was attributable to average PM10 levels being 42%higher during this period of follow-up (Table 3).

A limitation of our study, also a limitation in previousstudies (7, 8), was that results were based on the cohort of

pregnancies reaching at least 20 weeks’ gestation. Given thatour gestational window of interest spanned weeks 3–7 ofpregnancy, we would have preferred our cohort to consist ofall pregnancies at week 3. The consequence of this limitationmight depend on the causal effect of air pollution on cardio-vascular malformations. For example, atrioventricular septaldefect, Ebstein’s anomaly, and tricuspid valve dysplasia canall cause intrauterine congestive heart failure, increasing therisk of intrauterine fetal death (31). If air pollution were toincrease the risk of these malformations, in turn increasing therisk of fetal loss before week 20, then our study would havebeen unable to detect this harmful effect of pollution.

Measurement error or its discrete counterpart, misclassifi-cation, which was present in the air quality, vital records, andsurveillance data, was another limitation in our study as inprevious studies (7, 8). Our use of ambient air pollution mea-surements from stationary monitors as proxies for personalexposure was likely the largest component of measurementerror. If this measurement error was nondifferential, a bias to-ward the null could explain some null results (32). If the mea-surement error was differential, for example, if the magnitudeof measurement error varied according to meteorologic con-ditions, then risk ratio estimates may have been biased toward,away from, or across the null. We did not observe broad pat-terns of consistently positive or negative risk ratios suggestingthe possibility of differential measurement error, although dif-ferential measurement error may have biased some results.

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,


No. ofCases

8-HourOzone, ppb

24-HourPM10, mg/m

324-Hour Nitrogen



RiskRatio


RiskRatio


RiskRatio


RiskRatio


RiskRatio





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





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


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


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



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


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

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




Am J Epidemiol 2009;169:1004–1014

Misclassification in the vital records data was evident bythe strong day-of-month pattern observed in the date-of-conception estimates. Although a statistical model was cre-ated to remove this day-of-month pattern, uncertainty in theestimates remained. The date-of-conception estimates fromMACDP surveillance records, although lacking an obviousday-of-month pattern, likewise had uncertainty. Althoughwe selected our exposure window to coincide with the pe-riod of cardiac morphogenesis, it is possible that exposuresearlier or later in pregnancy could have affected the devel-opment of certain malformations (33). Some results weresensitive to the window definition; for pollution levels dur-ing the first 9 weeks of pregnancy, we observed positive,statistically significant associations between nitrogen diox-ide and both secundum atrial septal defect and muscularventricular septal defect. Although the point estimates forthese 2 associations were elevated in our primary analysis,neither was statistically significant (Table 5).

Ours is the third epidemiologic study of air pollutionand cardiovascular malformations reported to date (7, 8).Results from these 3 studies do not seem consistent. Thisinconsistency could be due to the absence of true associa-tions between ambient air pollution and risks of cardiovas-cular malformations; it also could be due to differences inpopulations, pollution levels, outcome definitions, or analyt-ical approaches. If ambient air pollution levels do causecardiovascular malformations, then the lack of consistency

of results might be due to issues relating to statistical powerand measurement error. Although our study population waslarge, we nevertheless would have struggled to detect verysmall increases in risk.

ACKNOWLEDGMENTS

Author affiliations: Department of Environmental andOccupational Health, Rollins School of Public Health,Emory University, Atlanta, Georgia (Matthew J. Strickland,Mitchel Klein, Paige E. Tolbert); Department of Epidemi-ology, Rollins School of Public Health, Emory University,Atlanta, Georgia (Matthew J. Strickland, W. Dana Flanders,Michele M. Marcus); National Center on Birth Defects andDevelopmental Disabilities, Centers for Disease Controland Prevention, Atlanta, Georgia (Matthew J. Strickland,Adolfo Correa, Tiffany Riehle-Colarusso, Csaba Siffel);Oregon Health and Science University, Portland, Oregon(Mark D. Reller); Children’s Healthcare of Atlanta, EmoryUniversity, Atlanta, Georgia (William T. Mahle); Depart-ment of Pediatrics, University of Utah, Salt Lake City, Utah(Lorenzo D. Botto); School of Civil and Environmental En-gineering, Georgia Institute of Technology, Atlanta, Georgia(James A. Mulholland); and Computer Sciences Corpora-tion, Atlanta, Georgia (Csaba Siffel).

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,


No. ofCases

8-HourOzone, ppb

24-HourPM10, mg/m

324-Hour Nitrogen



RiskRatio


RiskRatio


RiskRatio


RiskRatio


RiskRatio





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





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


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


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



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


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

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




Am J Epidemiol 2009;169:1004–1014

Supported by the National Institute of EnvironmentalHealth Sciences (R01-ES012967-01A1), the Health Resour-ces and Services Administration (T03MC07651), and Envi-ronmental Public Health Tracking at the Centers for DiseaseControl and Prevention.

The findings and conclusions in this report are those ofthe author(s) and do not necessarily represent the officialposition of the Centers for Disease Control and Prevention.


REFERENCES

1. Glinianaia SV, Rankin J, Bell R, et al. Particulate air pollutionand fetal health: a systematic review of the epidemiologicevidence. Epidemiology. 2004;15(1):36–45.

2. Maisonet M, Correa A, Misra D, et al. A review of the liter-ature on the effects of ambient air pollution on fetal growth.Environ Res. 2004;95(1):106–115.

3. Sram RJ, Binkova B, Dejmek J, et al. Ambient air pollutionand pregnancy outcomes: a review of the literature. EnvironHealth Perspect. 2005;113(4):375–382.

4. Lacasana M, Esplugues A, Ballester F. Exposure to ambientair pollution and prenatal and early childhood health effects.Eur J Epidemiol. 2005;20(2):183–199.

5. Perera FP, Rauh V, Whyatt RM, et al. A summary of recentfindings on birth outcomes and developmental effects of pre-natal ETS, PAH, and pesticide exposures. Neurotoxicology.2005;26(4):573–587.

6. Choi H, Jedrychowski W, Spengler J, et al. International studies ofprenatal exposure to polycyclic aromatic hydrocarbons and fetalgrowth. Environ Health Perspect. 2006;114(11):1744–1750.

7. Ritz B, Yu F, Fruin S, et al. Ambient air pollution and risk ofbirth defects in southern California. Am J Epidemiol. 2002;155(1):17–25.

8. Gilboa SM, Mendola P, Olshan AF, et al. Relation betweenambient air quality and selected birth defects, seven county study,Texas, 1997–2000. Am J Epidemiol. 2005;162(3):238–252.

9. Mavroudis C, Jacobs JP. Congenital Heart Surgery Nomen-clature and Database Project: overview and minimum dataset.Ann Thorac Surg. 2000;69(4 suppl):S2–S17.

10. Beland MJ, Franklin RC, Jacobs JP, et al. Update from theInternational Working Group for Mapping and Coding ofNomenclatures for Paediatric and Congenital Heart Disease.Cardiol Young. 2004;14(2):225–229.

11. Jacobs JP, Mavroudis C, Jacobs ML, et al. Lessons learnedfrom the data analysis of the second harvest 1998–2001 of theSociety of Thoracic Surgeons (STS) Congenital Heart SurgeryDatabase. Eur J Cardiothorac Surg. 2004;26(1):18–37.

12. Riehle-Colarusso T, Strickland MJ, Reller MD, et al. Improv-ing the quality of surveillance data on congenital heart defectsin the Metropolitan Atlanta Congenital Defects Program. BirthDefects Res A Clin Mol Teratol. 2007;79(11):743–753.

13. The International Working Group for Mapping and Coding ofNomenclatures for Pediatric and Congenital Heart Disease.International pediatric and congenital cardiac code. Montreal,Quebec, Canada: International Society for Nomenclature ofCongenital and Paediatric Heart Disease; 2004. (http://www.ipccc.net).

14. Correa A, Cragan JD, Kucik JE, et al. Reporting birth defectssurveillance data 1968–2003. Birth Defects Res A Clin MolTeratol. 2007;79(2):65–186.

15. Society of Thoracic Surgeons. STS Congenital Heart SurgeryDatabase, version 2.30. Chicago, IL: Society of Thoracic Sur-geons; 2004. (http://www.sts.org/sections/stsnationaldatabase/datamanagers/congenitalheartsurgerydb/datacollection/index.html).

16. Centers for Disease Control and Prevention. MetropolitanAtlanta Congenital Defects Program defect code list. Atlanta,GA: Centers for Disease Control and Prevention; 2004. (http://www.cdc.gov/ncbddd/bd/macdp_resources.htm).

17. Jacobs JP, Mavroudis C, Jacobs ML, et al. Nomenclature anddatabases—the past, the present, and the future: a primer forthe congenital heart surgeon. Pediatr Cardiol. 2007;28(2):105–115.

18. National Center on Birth Defects and Developmental Dis-abilities, Centers for Disease Control and Prevention. Statebirth defects surveillance program directory. Birth Defects ResA Clin Mol Teratol. 2006;76(12):837–893.

19. Sadler TW. Langman’s Medical Embryology. Philadelphia,PA: Lippincott Williams & Wilkins; 2006.

20. R Development Core Team. R: A Language and Environmentfor Statistical Computing v 2.5.0. Vienna, Austria: R Foun-dation for Statistical Computing; 2007.

21. Kallen K. Maternal smoking and congenital heart defects. EurJ Epidemiol. 1999;15(8):731–737.

22. Schneider DJ, Moore JW. Patent ductus arteriosus. Circula-tion. 2006;114(17):1873–1882.

23. Kitterman JA, Edmunds LH Jr, Gregory GA, et al. Patentducts arteriosus in premature infants. Incidence, relation topulmonary disease and management. N Engl J Med. 1972;287(10):473–477.

24. US Department of Health and Human Services. Womenand Smoking: A Report of the Surgeon General. Atlanta,GA: Centers for Disease Control and Prevention; 2001.

25. US Department of Health and Human Services. TheHealth Consequences of Involuntary Exposure to TobaccoSmoke: A Report of the Surgeon General. Atlanta, GA:Centers for Disease Control and Prevention; 2006.

26. Wilhelm M, Ritz B. Local variations in CO and particulate airpollution and adverse birth outcomes in Los Angeles County,California, USA. Environ Health Perspect. 2005;113(9):1212–1221.

27. Parker JD, Woodruff TJ, Basu R, et al. Air pollution and birthweight among term infants in California. Pediatrics. 2005;115(1):121–128.

28. Clark EB. Pathogenetic mechanisms of congenital cardiovas-cular malformations revisited. Semin Perinatol. 1996;20(6):465–472.

29. Jacobs JP, Burke RP, Quintessenza JA, et al. CongenitalHeart Surgery Nomenclature and Database Project: ventricularseptal defect. Ann Thorac Surg. 2000;69(4 suppl):S25–S35.

30. Jacobs ML. Congenital Heart Surgery Nomenclature andDatabase Project: tetralogy of Fallot. Ann Thorac Surg. 2000;69(4 suppl):S77–S82.

31. Hoffman JIE. Epidemiology of congenital heart disease: eti-ology, pathogenesis and incidence. In: Yagel S, Silverman NH,Gembruch U, eds. Fetal Cardiology: Embryology, Genetics,Physiology, Echocardiographic Evaluation, Diagnosis andPerinatal Management of Cardiac Diseases. London, UnitedKingdom: Martin Dunitz; 2003:79–88.

32. Zeger SL, Thomas D, Dominici F, et al. Exposure measurementerror in time-series studies of air pollution: concepts and con-sequences. Environ Health Perspect. 2000;108(5):419–426.

33. Trines J, Hornberger LK. Evolution of heart disease in utero.Pediatr Cardiol. 2004;25(3):287–298.


Am J Epidemiol 2009;169:1004–1014

http://www.ipccc.net

http://www.ipccc.net

http://www.sts.org/sections/stsnationaldatabase/datamanagers/congenitalheartsurgerydb/datacollection/index.html



http://www.cdc.gov/ncbddd/bd/macdp_resources.htm

http://www.cdc.gov/ncbddd/bd/macdp_resources.htm




Vol. 169, No. 8

DOI: 10.1093/aje/kwp001



Maternal Urinary Metabolites of Di-(2-Ethylhexyl) Phthalate in Relation to theTiming of Labor in a US Multicenter Pregnancy Cohort Study

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

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

Di-(2-ethylhexyl) phthalate (DEHP) is a plasticizer used in consumer and medical products that can cross theplacenta, disrupt steroid hormone synthesis, and activate peroxisome proliferator-activated receptor c. The authorsexamined DEHP exposure in relation to the timing of labor in a pregnancy cohort study of 283 women recruited in 4US states (California, Iowa, Minnesota, and Missouri) between 2000 and 2004. The authors estimated associationsbetween concentrations of DEHP metabolites and gestational age at delivery using linear regression models andassociations between DEHP metabolites and clinical outcomes using logistic regression models. After covariateadjustment, women at the 75th percentile of DEHP metabolite concentrations had a 2-day-longer mean length ofgestation than women at the 25th percentile (95% confidence interval: 1.4, 3.3). Log-unit increases in mono-2-ethylhexyl phthalate and mono-2-ethyl-5-oxohexyl phthalate concentrations were associated with increasedodds of cesarean section delivery (30% and 50% increased odds, respectively), increased odds of delivering at41 weeks or later (100% and 120% increased odds), and reduced odds of preterm delivery (50% and 60%decreased odds). These data suggest that DEHP may interfere with signaling related to the timing of parturition.

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

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

Di-(2-ethylhexyl) phthalate (DEHP) is a plasticizer that iswidely used in consumer products, such as polyvinyl chlo-ride flooring, carpeting, roofing, vinyl, upholstery, clothing,and packaging (1). DEHP exposure occurs through inges-tion of food and water (2, 3) and inhalation of householddust (4), as well as parenterally from medical devices (1).DEHP metabolites have been detected in 95% of the USpopulation aged �6 years (5).

The hydrolytic metabolite of DEHP, mono-(2-ethylhexyl)phthalate (MEHP), can cross the placenta and enter fetalcirculation (6, 7). Metabolites of DEHP have been measuredin umbilical cord blood (8), amniotic fluid (9), maternalurine (10), placental tissue (7, 11), and neonatal urine andmeconium (12, 13). MEHP is further metabolized into oxi-dative metabolites, including mono-2-ethyl-5-hydroxyhexyl

phthalate (MEHHP) and mono-2-ethyl-5-oxohexyl phthal-ate (MEOHP). MEHP and its oxidative metabolites canbe glucuronidated and excreted in urine and feces. These3metabolites account for approximately 50% of the receiveddose of DEHP (14).

Among rodents, DEHP suppresses fetal testosterone syn-thesis in males (15, 16) and inhibits ovarian aromatase tran-scription in adult females (17, 18). Aromatase, which isresponsible for the conversion of androgens to estradiol, isexpressed in several tissues, including the human placenta(19). Inhibition of steroid hormone synthesis in the ovary ofthe adult rodent, fetal testis, and fetal brain has been linkedto binding and activation of the transcription factor peroxi-some proliferator-activated receptor c (PPARc) (20–22).MEHP can activate PPARc in primate fibroblasts, in rodent

Correspondence to Dr. Jennifer J. Adibi, Center for Reproductive Sciences, Department of Obstetrics/Gynecology and Reproductive Sciences,

University of California, San Francisco, 513 Parnassus Avenue, Box 0556, San Francisco, CA 94143 (e-mail: [email protected]).

1015 Am J Epidemiol 2009;169:1015–1024

granulosa and trophoblast cells, and in human cervical can-cer cells (21–24).

PPARc activation increases early in pregnancy to pro-mote zygote implantation and placental development(25–27). During pregnancy, high levels of placental PPARccontribute to uterine quiescence by down-regulating cyclo-oxygenase 2 and inhibiting prostaglandin synthesis. By con-trast, at parturition, PPARc activation decreases in order toallow for increased prostaglandin production, which is es-sential in the stimulation of uterine contractions (26, 28). Ithas been hypothesized that exposure to PPARc ligands dur-ing pregnancy reduces the risk of preterm labor by suppress-ing the inflammatory response in fetal tissues (27).

We tested the hypothesis that exposure to DEHP alters thetiming of labor by measuring the associations between ma-ternal urinary concentrations of DEHP metabolites duringpregnancy and gestational age at delivery, as well as be-tween these metabolites and risk of cesarean delivery,preterm birth, and delivering at greater than 41 weeks’gestation.


Women included in this analysis were recruited into theStudy for Future Families (SFF) between March 2000 andAugust 2004 at 4 US sites. Details on the SFF are availableelsewhere (29). Briefly, women were recruited at prenatalclinics associated with university hospitals in Los Angeles,California (Harbor-UCLA and Cedars-Sinai); Minneapolis,Minnesota (University of Minnesota Health Center); Co-lumbia, Missouri (University of Missouri School of Medi-cine); and Iowa City, Iowa (University of Iowa School ofMedicine). Women were eligible if they were at least 18years of age, conception had been natural, and the preg-nancy was not medically threatened.

Of the 783 pregnant women enrolled in the SFF between2000 and 2004, only those who were fully enrolled ina follow-up study for postnatal evaluation of their babieswere eligible for inclusion in the phthalate study (n¼ 441).For a woman to be eligible for follow-up, the pregnancyhad to end in a livebirth, the baby had to be 2–36 months ofage, the mother and baby had to live within 50 miles (80 km)of the clinic, and the mother had to attend at least 1 studyvisit. Reasons for exclusion were unlikely to be relatedto either phthalate exposure or gestational age at delivery(except possibly a nonviable birth outcome, which was ex-tremely rare). The current analysis (n ¼ 283) was restrictedto mothers for whom we had urinary phthalate concentra-tions (n ¼ 304) and complete medical record data (n ¼ 298)after excluding twin births (12 babies), 2 babies with miss-ing data, and 1 baby born at 30 weeks with eclampsia.Women provided 1 urine sample at the time of recruitment,which was on average 12.2 weeks (standard deviation,7.6 weeks) before delivery or the beginning of the thirdtrimester.

Human subject committees at each of the participatingcenters approved study procedures, and all participants gavesigned informed consent. Data were stripped of identifyinginformation before analysis.

Gestational age at delivery was calculated in 2 ways. First,the date of the last menstrual period as reported by thewomanat study entry was subtracted from the date of delivery toobtain an estimate in days. Second, the clinical estimate re-corded by the obstetrician was abstracted directly from thebirth record. The clinical estimate was based on ultrasounddata, examination of the newborn, and dates reported bythe mother and was rounded up to the next-highestweek. There was a high correlation between the 2 estimates(r ¼ 0.92, n ¼ 283) after we substituted the clinical estimatefor cases for which therewas a discrepancy of 14 ormore days(30). We used the clinical estimate in the current analysis.

Urinary phthalate metabolite concentrations were mea-sured using analytical methods described in detail else-where (31). Briefly, the phthalate metabolites were firstenzymatically deconjugated and then extracted from theurine using automated on-line solid phase extraction, sep-arated by high-performance liquid chromatography, anddetected by isotope-dilution tandem mass spectrometry.Each analytical run included calibration standards, reagentblanks, and quality control materials of high and low con-centration to monitor for accuracy and precision. The lim-its of detection were 0.98 ng/mL (MEHP), 0.95 ng/mL(MEHHP), and 1.07 ng/mL (MEOHP).

Because of the high correlation and pharmacokinetic sim-ilarities between MEOHP and MEHHP (r ¼ 0.99), onlyestimates for MEOHP are described below.


Geometric mean values for urinary phthalates and their95% confidence intervals, unadjusted and adjusted for cre-atinine concentration, were calculated and compared withUS population estimates for women of reproductive age(i.e., 18–40 years) derived from 1999–2000 and 2001–2002 National Health and Nutrition Examination Survey(NHANES) data (32, 33). We also calculated geometricmean values and 95% confidence intervals for US pregnantwomen using the NHANES data (n ¼ 209 for MEHP andn ¼ 104 for MEOHP, MEHHP, and %MEHP (defined be-low)). Because the NHANES sample was nonrandom, weused the recommended weights to correctly estimate varian-ces (34). Phthalate metabolite concentrations, which wereall right-skewed, were transformed using the natural loga-rithm. For concentrations below the limit of detection, weassigned a value equal to the limit of detection divided bythe square root of 2 (35). A potential phenotypic marker ofDEHP metabolism, %MEHP, was calculated as the ratio ofMEHP concentration to the sum of the 3 DEHP metaboliteconcentrations (in nanomoles) and transformed using thenatural logarithm (36).

We adjusted for urinary dilution in 2 ways. For compar-isons with NHANES data, we used metabolite concentra-tions divided by creatinine concentrations (lg/g creatinine).However, because creatinine is associated with demo-graphic and physiologic parameters that may be on thecausal pathway from exposure to outcome (37, 38), we in-cluded a square root transformation of creatinine (mg/dL) asa covariate in all regression models rather than dividing theconcentration by creatinine.

1016 Adibi et al.

Am J Epidemiol 2009;169:1015–1024

Spearman correlation coefficients were used to estimatepairwise associations among phthalate concentrations. Mul-tivariate linear regression models were used to evaluate as-sociations between urinary phthalate concentrations andgestational age at delivery. Although the distribution of ges-tational ages was slightly skewed, a sensitivity analysis us-ing generalized estimating equations, which is valid evenunder departure from normality, produced similar results.Logistic regression was used to calculate associations ofphthalate concentrations with binary outcomes, such as ce-sarean section delivery.

Covariates, including demographic characteristics (race,geographic site, mother’s education, mother’s age, mother’semployment status during pregnancy), sample characteris-tics (timing of urine sample, creatinine), previous pregnancyhistory (parity, history of miscarriage), sex of the baby, ma-ternal and paternal smoking, job-related stress, and mother’sprepregnancy health (nongestational diabetes, thyroid dis-orders, fibroids, high blood pressure, and respiratory condi-tions (asthma and chronic obstructive pulmonary disease))were considered as potential confounders (Table 1) andwere retained in the model if they were significant at P �0.15. Potential effect modifiers evaluated included sex of thebaby, %MEHP status (low/high), geographic location, tim-ing of the urine sample, and parity. We had no informationon mother’s weight prior to pregnancy, body mass index, orweight gain during pregnancy.

Statistical significance was defined by a (2-sided) P valueof 0.05 or lower and was expressed as a 95% confidenceinterval. SAS 9.1 software (SAS Institute Inc., Cary, NorthCarolina) was used to conduct all analyses.

RESULTS

Demographic and other sample characteristics varied bygeographic site (Table 1), with significant differences inparity, race, education, history of fibroids, and gestationalage. After creatinine adjustment, DEHP metabolite concen-trations varied across geographic locations, with signifi-cantly higher concentrations of MEHP among womenfrom Missouri and Minnesota than among women fromCalifornia and Iowa. White non-Hispanic mothers, whomade up 84% of the study population, had marginally higherconcentrations of MEHP and MEOHP than Hispanics.Women with some college education had somewhat lowerconcentrations of MEOHP than women with a high schooleducation or less. Higher levels of job-related stress wereslightly associated with higher MEHP concentrations.Women with a previous miscarriage had significantly lowerurinary concentrations of MEHP. Metabolite and creatinineconcentrations were not related to the timing of the urinesample.

The unadjusted mean concentrations of all 3 DEHP me-tabolites in the SFF women were somewhat lower than theNHANES US population estimates for pregnant womenand women of reproductive age (Table 2), but only themean MEHHP concentration in SFF women was signifi-cantly lower than that for nonpregnant women (11.9 ng/mLvs. 19.8 ng/mL). Neither the unadjusted nor the creatinine-

adjusted mean values for pregnant women differed signif-icantly between the SFF and the NHANES.

Associations of the DEHP metabolites with gestationalage, unadjusted and adjusted for covariates, are presentedin Table 3 and Figures 1 and 2. Overall, a log-unit increasein the urinary concentrations of metabolites was associatedwith a 1.3- to 1.5-day increase in gestational age. The effectsize was slightly greater for MEOHP than for MEHP, andwe saw no association between %MEHP and gestationalage. Effect estimates were somewhat attenuated after adjust-ment for potential confounders. Each tertile increase inMEHP concentration was associated with a 1.8-day (95%confidence interval (CI): 0.2, 3.4) increase in gestationalage, and each tertile increase in MEOHP concentrationwas associated with a 1.9-day increase (95% CI: 0.1, 3.8).We compared estimates after removing 1 outlier for urinaryMEHP concentration (1,600 ng/mL) and 1 outlier for ges-tational age (32 weeks) (Figure 1). The effect estimates didnot change appreciably. Since we did not see any reasons,medical or otherwise, for excluding these subjects from theanalysis, they were included.

The odds of giving birth at greater than 41 weeks in-creased by 100% and 120% with each log-unit increase inMEHP and MEOHP concentrations, respectively, after ad-justment for covariates (Table 4). Conversely, we observed50% and 60% reductions in the odds of preterm deliverywith a log-unit increase in MEHP and MEOHP concentra-tions, respectively. The frequency of premature delivery(6%) was slightly lower than the NHANES US populationestimate of 11.5% for whites, because we excluded twinsand an eclampsia case, which would bring the overall prev-alence to 9.5%. The higher socioeconomic status of thiscohort also contributed to the slightly lower prevalence.The odds of a cesarean section delivery increased by 50%with each log-unit increase in MEOHP concentration and by30% with each log-unit increase in MEHP concentration.We explored potential explanations for the association ofcesarean section with DEHP metabolite concentrations.Women who had ‘‘failure of labor to progress’’ listed asa reason for cesarean section delivery had a mean MEOHPconcentration that was 0.33 log units higher (95% CI: 0.08,0.57) than that of all other women, after controlling for otherreasons for cesarean section. We found no associations ofDEHP metabolite concentrations with breech presentation(n ¼ 8), repeat cesarean section (n ¼ 15), or medically in-duced (n¼ 85) or augmented (n¼ 55) labor or with durationof labor measured in minutes.

We conducted stratified analyses by geographic site and%MEHP status (low/high). In Missouri, Iowa, and Califor-nia, gestational age at delivery tended to increase withhigher MEHP and MEOHP concentrations, while in Min-nesota gestational age tended to decrease (though notsignificantly). After adjustment for creatinine, mother’seducation, parity, and job-related stress, the difference inslopes for the 2 groups (California, Iowa, and Missouri vs.Minnesota) was significant for MEOHP (P ¼ 0.02) andmarginally significant for MEHP (P ¼ 0.06).

The association of MEOHP concentrations with gesta-tional age among women in the low-%MEHP categorywas approximately 3 times stronger in magnitude than in

Maternal DEHP Metabolites and Timing of Labor 1017

Am J Epidemiol 2009;169:1015–1024

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

CharacteristicOverall

Study Center

California Iowa Minnesota Missouri

No. or Mean % No. or Mean % No. or Mean % No. or Mean % No. or Mean %

No. of subjects 283 100 48 17 80 28 87 31 68 24

Time between urine collection and delivery, weeksa

Mean 12 (8)b 14 (8) 11 (8) 12 (8) 12 (6)

First trimester 5 2 1 2 0 0 4 5 0 0

Second trimester 119 42 24 51 34 43 34 39 27 40

Third trimester 157 56 22 47 45 57 49 56 41 60

Sex of baby

Male 143 51 24 50 38 47 45 52 36 53

Female 140 49 24 50 42 53 42 48 32 47

Parity (no. of prior livebirths)***

0 145 51 30 63 32 40 52 60 31 46

1 87 31 11 23 25 31 27 31 24 35

�2 51 18 7 15 23 29 8 9 13 19

Racec,***

White 238 84 22 46 72 91 84 97 60 88

Hispanic 26 9 20 42 2 3 0 0 4 6

Other 18 6 6 13 5 6 3 3 4 6

Education*

High school or less 25 9 9 19 6 8 3 3 7 10

Some college 63 22 19 40 14 18 20 23 10 15

College graduation 106 38 10 21 34 43 33 38 29 43

Graduate school 89 31 10 21 26 33 31 36 22 32

History of miscarriage (yes/no) 60 21 9 19 20 25 14 16 17 25

Employed during pregnancy (yes/no) 231 82 33 69 63 79 79 91 56 82

Job-related stress

No work-related stress 52 18 15 31 17 21 8 9 12 18

Not at all stressful 17 6 2 4 4 5 6 7 5 7

Not too stressful 80 28 14 29 21 26 29 33 16 24

Somewhat stressful 105 37 12 25 32 40 35 40 26 38

Very stressful 29 10 5 10 6 8 9 10 9 13

Maternal health

Diabetes 14 5 4 8 6 8 1 1 3 4

Thyroid disorders 17 6 5 10 5 6 2 2 5 7

Fibroids* 16 6 1 2 8 10 2 2 5 7

High blood pressure 19 7 4 8 5 6 5 6 5 7

Respiratory conditions 26 9 4 8 10 13 5 6 7 10

Mean creatinine level, mg/dL 92 (61) 99 (63) 93 (64) 82 (51) 97 (66)

Mean maternal age, years* 30.2 (6.0) 28.3 (6.0) 30.6 (4.6) 30.8 (5.3) 30.3 (4.6)

Gestational age at delivery (clinical estimate), weeksd

Mean** 39.2 (1.5) 38.9 (1.3) 38.9 (1.3) 39.7 (1.3) 39.1 (1.8)

<37 14 5 3 6 4 5 2 2 5 7

37–41 262 93 45 94 76 95 81 93 60 88

>41 7 2 0 0 0 0 4 5 3 4

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

Information was missing for 1 participant.

1018 Adibi et al.

Am J Epidemiol 2009;169:1015–1024

the high-%MEHP group after adjustment for creatinine,geographic site, mother’s education, and job-related stress.When modeled as a multiplicative interaction term, theP value was 0.10 for differences in slope between the2 groups (Figure 3). There was no difference in MEHP slopebetween the low- and high-%MEHP categories.

DISCUSSION

In approximately 300 pregnant women from 4 US loca-tions, we found that gestation was 1.1 days and 1.3 dayslonger with each log-unit increase in urinary concentrationsof the DEHP metabolites MEHP and MEOHP, respectively.Women at the 75th percentile of urinary MEHP concen-tration had a duration of gestation that was 2.3 days(95% CI: 1.4, 3.3) longer than that of women at the 25thpercentile of exposure, after we controlled for urinary

creatinine concentration, demographic factors, maternalhealth, stress, and parity. DEHP exposure in this cohortwas similar to the NHANES US population estimate forpregnant females.

The clinical or population significance of an exposure-related shift of 2–3 days in gestational length is difficult toevaluate. For this reason, we also estimated associationswith clinical outcomes and saw increased odds of deliveringafter 41 weeks, decreased odds of preterm delivery, and in-creased odds of delivering by cesarean section. Deliveryafter 41 weeks is associated with an increase in perinatalmortality due to meconium aspiration, fetal distress, as-phyxia, pneumonia, malformations, shoulder dystocia, andtraumatic injuries (39, 40). Women who undergo cesareansection are at increased risk in subsequent pregnancies ofmalpresentation, abnormal placentation, antepartum hemor-rhage, placenta accreta, prolonged labor, uterine rupture,preterm birth, low birth weight, and stillbirth (41). The

Table 2. Urinary Concentrations (ng/mL) of Di-(2-Ethylhexyl) Phthalate Metabolites Among Participants (n ¼ 283) in the Study for Future

Families (2000–2004) as Compared With Participants in the National Health and Nutrition Examination Survey (1999–2002)

Di-(2-Ethylhexyl)PhthalateMetabolite

% of Samples WithValues Greater ThanLimit of Detection

PercentileStudy for

Future FamiliesNHANES Pregnant

Womena

NHANESReproductive-Age Womenb

5th 25th 50th 75th 95th GM 95% CI GM 95% CI GM 95% CI

MEHP 76 0.6 1.1 3.5 8.2 40.2 3.6 3.1, 4.3 4.8 3.8, 6.0 4.5 4.0, 5.0

MEHHP 97 1.1 5.6 11.2 25.5 99.4 11.9 10.1, 13.9 19.0 13.5, 26.7 19.8 14.9, 26.2

MEOHP 96 1.2 5.1 9.9 24.6 68.4 10.9 9.3, 12.6 15.4 11.3, 21.1 13.8 10.4, 18.3

%MEHPc NA 4 8 14 23 48 14 12, 15 17 13, 20 13 11, 14

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

ate; MEOHP, mono-2-ethyl-5-oxohexyl phthalate; NA, not applicable; NHANES, National Health and Nutrition Examination Survey.a NHANES data on pregnant women, 1999–2000 (MEHP) and 2001–2002 (MEHP, MEOHP, and MEHHP) (n ¼ 209 for MEHP and n ¼ 104 for

MEHHP, MEOHP, and %MEHP).b NHANES data on women aged 18–40 years, 1999–2000 (MEHP) and 2001–2002 (MEHP, MEOHP, and MEHHP) (n ¼ 853 for MEHP and

n ¼ 437 for MEHHP, MEOHP, and %MEHP).c MEHP/R(MEHP, MEOHP, MEHHP) 3 100.

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

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

ModelMEHP MEOHP MEHHP

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

Adjusted for creatinine 0.18 0.04, 0.32 0.21 0.05, 0.38 0.19 0.05, 0.34

Adjusted for creatinine þdemographic factorsb

0.17 0.03, 0.32 0.20 0.03, 0.36 0.18 0.02, 0.33

Adjusted for above factors þmaternal healthc

0.17 0.03, 0.31 0.19 0.03, 0.35 0.17 0.02, 0.32

Adjusted for above factors þparity

0.16 0.02, 0.30 0.19 0.03, 0.35 0.16 0.01, 0.31

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

phthalate; MEOHP, mono-2-ethyl-5-oxohexyl phthalate.

* P < 0.05 for all results shown in table.a Change in gestational age at delivery (weeks) per log-unit increase in urinary di-(2-ethylhexyl) phthalate metab-

olite concentration.b Geographic center, mother’s educational level, and job-related stress.c Nongestational diabetes, thyroid disorders, and fibroids.


Am J Epidemiol 2009;169:1015–1024

observed decrease in the risk of preterm delivery may beprotective or may be indicative of abnormal function of theplacenta (26); in this study, we could not distinguish be-tween these possibilities.

A prior study using MEHP concentration in umbilicalcord blood from 84 Italian mother/newborn pairs found anassociation that pointed in the opposite direction than wasobserved here (8). Latini et al. (8) reported that gestational

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

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

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

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

Families, 2000–2004a

OutcomeNo. of

Subjects%

MEHP MEOHP MEHHP

OR 95% CI OR 95% CI OR 95% CI

Gestational age >41 weeks 7 2

Adjusted for creatinine 1.8* 1.1, 3.0 2.1** 1.2, 3.4 1.9* 1.1, 3.2

Adjusted for creatinine þ covariatesb 2.0* 1.1, 3.5 2.2** 1.3, 4.0 2.1* 1.3, 3.7

Preterm delivery (<37 weeks) 14 5

Adjusted for creatinine 0.5* 0.3, 0.9 0.5* 0.2, 0.9 0.5* 0.3, 0.9

Adjusted for creatinine þ covariatesc 0.5* 0.3, 0.9 0.4* 0.2, 0.9 0.5* 0.3, 0.9

Cesarean section delivery 62 22

Adjusted for creatinine 1.3 1.0, 1.6 1.4* 1.1, 1.9 1.4* 1.1, 1.8

Adjusted for creatinine þ covariatesd 1.3* 1.0, 1.6 1.5** 1.1, 1.9 1.4* 1.1, 1.8

Failure to progress as reasonfor cesarean section delivery

18 6

Adjusted for creatinine 1.3 0.9, 1.8 1.6* 1.1, 2.4 1.5* 1.1, 2.2

Adjusted for creatinine þ covariatese 1.2 0.9, 1.7 1.6* 1.1, 2.3 1.5* 1.0, 2.1

Abbreviations: CI, confidence interval; MEHHP, mono-2-ethylhexyl phthalate; MEHP, mono-2-ethylhexyl phthal-

ate; MEOHP, mono-2-ethyl-5-oxohexyl phthalate; OR, odds ratio.

* P < 0.05; **P < 0.01.a Change in log odds per log-unit increase in urinary di-(2-ethylhexyl) phthalate metabolite concentration.b Adjusted for geographic site (Minnesota vs. California, Iowa, and Missouri) and respiratory conditions.c Adjusted for high blood pressure and nongestational diabetes.d Adjusted for mother’s age (�35 years vs. <35 years), geographic site (Minnesota vs. California, Iowa, and

Missouri), nongestational diabetes, and fibroids.e Adjusted for parity, thyroid conditions, and high blood pressure. Information on the reason for cesarean section

was missing for 2 subjects.

1020 Adibi et al.

Am J Epidemiol 2009;169:1015–1024

age was shorter by 1.2 weeks in the MEHP-positive pairsthan in the MEHP-negative subjects. In that study, theymeasured MEHP and DEHP in umbilical cord blood, whichmay have been subject to contamination by DEHP in thesampling and analytic equipment (42). Measuring MEHP inblood is not recommended because of its short half-life (14).In that analysis, exposure status was dichotomized as ex-posed and nonexposed. Given that more than 95% of thegeneral US population has detectable urinary metabolites ofDEHP (5), this approach could have resulted in misclassifi-cation among the nonexposed subjects. The discrepant re-sults between these 2 studies could be due to differences instudy design, exposure assessment, and/or the underlyingcharacteristics of the populations.

When we stratified SFF subjects by geographic loca-tion, gestational age tended to increase with increasingphthalate metabolite concentrations for all sites exceptMinnesota, where it tended to remain flat or decreaseslightly. The women from Minnesota tended to havehigher MEOHP concentrations and %MEHP, higher ges-tational age, higher maternal age, and more education thanwomen from other study centers, most markedly relativeto California, and they were predominantly white (97%).We can speculate on 2 possible explanations. It is possiblethat the dose-response curve was nonlinear and essentiallyreversed at the higher doses among the Minnesota sub-jects. It is also possible that the site-specific populationsdiffered in ways that modified the relation between DEHPexposure and placental function. The significant differen-ces in race, education, maternal age, and parity betweenstudy centers could be proxies for other unmeasured effect

modifiers, such as nutritional factors, coexposures, and/orlifestyle factors.

Other known causes of prolonged gestation include fishoil consumption during pregnancy (43, 44), deficiency inplacental sulfatase, which is another cause of decreasedestrogen synthesis (45), and living in a highly polluted area(46). Fish oil contains n-3 long-chain polyunsaturated fattyacids, which are also ligands of PPARc and may contributeto a suppressed inflammatory response late in pregnancy(47, 48). If this were the case, it is possible that a competitiveinteraction between DEHP and fatty acids in the diet exists.We could not test this hypothesis, since fish consumptionamong our subjects was generally low (88% of those whoconsumed fish had 2 or fewer servings per week). We did nothave information on the type of fish consumed or on how itwas prepared.

We hypothesized that the association between DEHP ex-posure and timing of labor could also differ depending ona woman’s ability to metabolize and excrete DEHP. To testthis, we dichotomized %MEHP values at the median andcompared metabolite associations within the low and highstrata. The association of MEOHP concentrations withlonger gestation was 3-fold stronger in the low-%MEHPgroup than in the high-%MEHP group. The association ofMEOHP concentration with the timing of labor might bedue to disruption in parturition and signaling by DEHP me-tabolites, but it might also be due to differences in a woman’sability to metabolize and excrete DEHP. In a previousreport, we found %MEHP to be approximately twice as re-producible within a woman over the last 6 weeks of preg-nancy as DEHP metabolites (10), suggesting that %MEHPmay reflect stable interindividual differences that couldbe relevant to the metabolism and excretion of these com-pounds in pregnancy.

Concern exists about the potential for systematic error inestimating gestational age using the last-menstrual-periodapproach (49, 50). We addressed this by also using theclinical estimate, which takes into account ultrasound dataand examination of the newborn in cases where there areinconsistencies in last-menstrual-period dating and clinicalpresentation, but the clinical estimate may have been lessprecise, given that it was rounded up. We found results tobe consistent using both measures when including all ges-tational ages and less consistent when modeling associa-tions with preterm and postterm delivery. Misclassificationof gestational age by last menstrual period is most prob-lematic among preterm and postterm births, with the de-gree of misclassification being associated with maternalrace, age, education, parity, month that prenatal care began(51), and regularity of the menstrual period (52). Giventhat some of these factors are also related to phthalateexposure and pregnancy outcomes, we relied on the clini-cal estimate to model associations with DEHP metaboliteconcentrations.

Some misclassification of DEHP exposure was present inour data, given that we had a single spot urine sample forcharacterizing exposure. In a previous analysis, we showedthat DEHP metabolite concentrations are not highly repro-ducible in pregnant women over the last 6 weeks of preg-nancy, probably because of physiologic changes occurring

Figure 3. Clinically estimated gestational age as a function of logurinary mono-2-ethyl-5-oxohexyl phthalate (MEOHP) concentration(ng/mL), by percent mono-2-ethylhexyl phthalate (%MEHP) status(low/high) (n ¼ 283), after adjustment for creatinine and covariates,Study for Future Families, 2000–2004. %MEHPwas calculated as theratio of MEHP concentration to the sum of the concentrations of3 di-(2-ethylhexyl) phthalate metabolites (MEHP, MEOHP, and mono-2-ethylhexyl phthalate) (in nanomoles) and transformed using thenatural logarithm.


Am J Epidemiol 2009;169:1015–1024

in the third trimester (10). Of the SFF subjects, 55% weresampled in the third trimester. In addition to misclassifica-tion, there may have been other exposures and risk factorsassociated with urinary DEHP metabolite concentrationsand birth outcomes that we were not able to adequatelycontrol for, resulting in residual confounding.

In conclusion, we observed an association between in-creased concentrations of DEHP metabolites in maternalurine measured during pregnancy and gestational age in aUS multicenter pregnancy cohort study. The direction andsize of the effect appeared to differ depending in part ongeographic location and a woman’s ability to metabolizeand eliminate DEHP. Our results support the hypothesis thatDEHP exposure may alter the dialogue between thematernaland fetal compartments that is essential for normal labor.Consistent with this hypothesis, urinary DEHP metaboliteconcentrations were associated with an increased risk of ce-sarean section delivery and of delivering at more than 41weeks. The binding affinity of the metabolite MEHP forPPARc and the central role PPARc plays in regulating pla-cental function may provide an explanation for this associa-tion, but this was not directly explored in the present study.

These results need to be replicated in other populations.There is likewise a need for more in vitro and in vivo re-search to better understand molecular mechanisms by whichDEHP may alter placental development and/or function.

ACKNOWLEDGMENTS

Author affiliations: Department of Environmental Health,Harvard School of Public Health, Boston, Massachusetts(Jennifer J. Adibi, Russ Hauser, Heather Nelson, RobertHerrick); Department of Biostatistics, Harvard School ofPublic Health, Boston, Massachusetts (Paige L. Williams);Department of Environmental Health Sciences, MailmanSchool of Public Health, Columbia University, New York,New York (Robin M. Whyatt); National Center for Environ-mental Health, Centers for Disease Control and Prevention,Atlanta, Georgia (Antonia M. Calafat); and Department ofObstetrics and Gynecology, University of Rochester Medi-cal Center, Rochester, New York (Shanna H. Swan).

This research was funded by the Environmental Protec-tion Agency (Star Grant R-82943601-0) and the NationalInstitute of Environmental Health Sciences (grant R01ES013543). Jennifer Adibi’s doctoral training was fundedby the Harvard Education and Research Center for Occupa-tional Safety and Health (grant T42 OH008416).

The authors thank Maureen Nealon and Fan Liu for theprocessing and transfer of data, Dr. Sonia Hernandez-Diazfor her mentoring with regard to epidemiologic methods,Dr. Shruthi Mahalingaiah for her help in interpreting clinicaldata, and Dr. Rick Stalhut for his assistance in accessingand analyzing data from the National Health and NutritionExamination Survey. The authors also thank Dr. ManoriSilva, Dr. Jack Reidy, Jim Preau, and Ella Samandar ofthe Centers for Disease Control and Prevention for thephthalate metabolite measurements.

The findings and conclusions presented in this report arethose of the authors and do not necessarily represent theviews of the Centers for Disease Control and Prevention.


REFERENCES

1. Schettler T. Human exposure to phthalates via consumerproducts. Int J Androl. 2006;29(1):134–139.

2. Kavlock R, Barr D, Boekelheide K, et al. NTP-CERHR expertpanel update on the reproductive and developmental toxicityof di(2-ethylhexyl) phthalate. Reprod Toxicol. 2006;22(3):291–399.

3. Agency for Toxic Substances and Disease Registry, Centersfor Disease Control and Prevention. Toxicological Profile forDi(2-ethylhexyl)phthalate (DEHP). Atlanta, GA: Centers forDisease Control and Prevention; 2002. (http://www.atsdr.cdc.gov/toxprofiles/tp9.html#9). (Accessed December 15, 2007).

4. Rudel RA, Camann DE, Spengler JD, et al. Phthalates,alkylphenols, pesticides, polybrominated diphenylethers, and other endocrine-disrupting compounds inindoor air and dust. Environ Sci Technol. 2003;37(20):4543–4553.

5. Kato K, Silva MJ, Reidy JA, et al. Mono(2-ethyl-5-hydroxyhexyl) phthalate and mono-(2-ethyl-5-oxohexyl)phthalate as biomarkers for human exposure assessmentto di-(2-ethylhexyl) phthalate. Environ Health Perspect.2004;112(3):327–330.

6. Singh AR, Lawrence WH, Autian J. Maternal-fetal transfer of14C-di-2-ethylhexyl phthalate and 14C-diethyl phthalate inrats. J Pharm Sci. 1975;64(8):1347–1350.

7. Mose T, Mortensen GK, Hedegaard M, et al. Phthalatemonoesters in perfusate from a dual placenta perfusion system,the placenta tissue and umbilical cord blood. Reprod Toxicol.2007;23(1):83–91.

8. Latini G, De Felice C, Presta G, et al. In utero exposureto di-(2-ethylhexyl)phthalate and duration of humanpregnancy. Environ Health Perspect. 2003;111(14):1783–1785.

9. Silva MJ, Reidy JA, Herbert AR, et al. Detection of phthalatemetabolites in human amniotic fluid. Bull Environ ContamToxicol. 2004;72(6):1226–1231.

10. Adibi JJ, Whyatt RM, Williams PL, et al. Characterization ofphthalate exposure among pregnant women assessed by repeatair and urine samples. Environ Health Perspect. 2008;116(4):467–473.

11. Poole CF, Wibberley DG. Determination of di-(2-ethylhexyl)phthalate in human placenta. J Chromatogr. 1977;132(3):511–518.

12. Kato K, Silva MJ, Needham LL, et al. Quantifying phthalatemetabolites in human meconium and semen using automatedoff-line solid-phase extraction coupled with on-line SPE andisotope-dilution high-performance liquid chromatography–tandem mass spectrometry. Anal Chem. 2006;78(18):6651–6655.

13. Weuve J, Sanchez BN, Calafat AM, et al. Exposure to phthal-ates in neonatal intensive care unit infants: urinary concen-trations of monoesters and oxidative metabolites. EnvironHealth Perspect. 2006;114(9):1424–1431.

14. Koch HM, Bolt HM, Preuss R, et al. New metabolites of di(2-ethylhexyl)phthalate (DEHP) in human urine and serum aftersingle oral doses of deuterium-labelled DEHP. Arch Toxicol.2005;79(7):367–376.

1022 Adibi et al.

Am J Epidemiol 2009;169:1015–1024

http://www.atsdr.cdc.gov/toxprofiles/tp9.html&num;9

http://www.atsdr.cdc.gov/toxprofiles/tp9.html&num;9

15. Parks LG, Ostby JS, Lambright CR, et al. The plasticizerdiethylhexyl phthalate induces malformations by decreasingfetal testosterone synthesis during sexual differentiation in themale rat. Toxicol Sci. 2000;58(2):339–349.

16. Akingbemi BT, Youker RT, Sottas CM, et al. Modulation of ratLeydig cell steroidogenic function by di(2-ethylhexyl)phthal-ate. Biol Reprod. 2001;65(4):1252–1259.

17. Davis BJ, Maronpot RR, Heindel JJ. Di-(2-ethylhexyl)phthalate suppresses estradiol and ovulation in cycling rats.Toxicol Appl Pharmacol. 1994;128(2):216–223.

18. Lovekamp TN, Davis BJ. Mono-(2-ethylhexyl) phthalatesuppresses aromatase transcript levels and estradiol productionin cultured rat granulosa cells. Toxicol Appl Pharmacol.2001;172(3):217–224.

19. Fournet-Dulguerov N, MacLusky NJ, Leranth CZ, et al. Im-munohistochemical localization of aromatase cytochromeP-450 and estradiol dehydrogenase in the syncytiotrophoblastof the human placenta. J Clin Endocrinol Metab. 1987;65(4):757–764.

20. Borch J, Metzdorff SB, Vinggaard AM, et al. Mechanismsunderlying the anti-androgenic effects of diethylhexylphthalate in fetal rat testis. Toxicology. 2006;223(1-2):144–155.

21. Lovekamp-Swan T, Jetten AM, Davis BJ. Dual activation ofPPARa and PPARc by mono-(2-ethylhexyl) phthalate in ratovarian granulosa cells. Mol Cell Endocrinol. 2003;201(1-2):133–141.

22. Xu Y, Cook TJ, Knipp GT. Effects of di-(2-ethylhexyl)-phthalate (DEHP) and its metabolites on fatty acid homeo-stasis regulating proteins in rat placental HRP-1 trophoblastcells. Toxicol Sci. 2005;84(2):287–300.

23. Feige JN, Gelman L, Rossi D, et al. The endocrine disruptormonoethyl-hexyl-phthalate is a selective peroxisomeproliferator-activated receptor c modulator that promotesadipogenesis. J Biol Chem. 2007;282(26):19152–19166.

24. Hurst CH, Waxman DJ. Activation of PPARa and PPARc byenvironmental phthalate monoesters. Toxicol Sci. 2003;74(2):297–308.

25. Fournier T, Tsatsaris V, Handschuh K, et al. PPARs and theplacenta. Placenta. 2007;28(2-3):65–76.

26. Froment P, Gizard F, Defever D, et al. Peroxisome proliferator-activated receptors in reproductive tissues: from game-togenesis to parturition. J Endocrinol. 2006;189(2):199–209.

27. Schaiff WT, Barak Y, Sadovsky Y. The pleiotropic function ofPPARc in the placenta. Mol Cell Endocrinol. 2006;249(1-2):10–15.

28. Dunn-Albanese LR, AckermanWE IV, Xie Y, et al. Reciprocalexpression of peroxisome proliferator-activated receptor-gamma and cyclooxygenase-2 in human term parturition. Am JObstet Gynecol. 2004;190(3):809–816.

29. Swan SH, Main KM, Liu F, et al. Decrease in anogenitaldistance among male infants with prenatal phthalate exposure.Environ Health Perspect. 2005;113(8):1056–1061.

30. Qin C, Dietz PM, England LJ, et al. Effects of different data-editing methods on trends in race-specific preterm deliveryrates, United States, 1990–2002. Paediatr Perinat Epidemiol.2007;21(suppl 2):41–49.

31. Silva MJ, Slakman AR, Reidy JA, et al. Analysis of humanurine for fifteen phthalate metabolites using automated solid-phase extraction. J Chromatogr B Analyt Technol Biomed LifeSci. 2004;805(1):161–167.

32. National Center for Health Statistics. National Health andNutrition Examination Survey. NHANES 1999–2000. Hyatts-ville, MD: National Center for Health Statistics; 2006. (http://

www.cdc.gov/nchs/about/major/nhanes/nhanes99_00.htm).(Accessed July 11, 2006).

33. National Center for Health Statistics. National Health and Nu-trition Examination Survey. NHANES 2001–2002. Hyattsville,MD: National Center for Health Statistics; 2006. (http://www.cdc.gov/nchs/about/major/nhanes/nhanes01-02.htm).(Accessed July 11, 2006).

34. National Center for Health Statistics. Continuous NHANESWeb Tutorial. Hyattsville, MD: National Center for HealthStatistics; 2007. (http://www.cdc.gov/nchs/tutorials/NHANES/index_current.htm). (Accessed June 18, 2007).

35. Hornung RW, Reed LD. Estimation of average concentrationin the presence of nondetectable values. Appl Occup EnvironHyg. 1990;5(1):46–51.

36. Hauser R, Meeker JD, Singh NP, et al. DNA damage in humansperm is related to urinary levels of phthalate monoester andoxidative metabolites. Hum Reprod. 2007;22(3):688–695.

37. Barr DB, Wilder LC, Caudill SP, et al. Urinary creatinineconcentrations in the US population: implications for urinarybiologic monitoring measurements. Environ Health Perspect.2005;113(2):192–200.

38. Schisterman EF, Whitcomb BW, Louis GM, et al. Lipid ad-justment in the analysis of environmental contaminants andhuman health risks. Environ Health Perspect. 2005;113(7):853–857.

39. Olesen AW, Westergaard JG, Olsen J. Perinatal and maternalcomplications related to postterm delivery: a national register-based study, 1978–1993. Am J Obstet Gynecol. 2003;189(1):222–227.

40. Hilder L, Costeloe K, Thilaganathan B. Prolonged pregnancy:evaluating gestation-specific risks of fetal and infant mortality.Br J Obstet Gynaecol. 1998;105(2):169–173.

41. Kennare R, Tucker G, Heard A, et al. Risks of adverse out-comes in the next birth after a first cesarean delivery. ObstetGynecol. 2007;109(2):270–276.

42. Calafat AM, Needham LL. Factors affecting the evaluation ofbiomonitoring data for human exposure assessment. Int J An-drol. 2008;31(2):139–143.

43. Makrides M, Duley L, Olsen SF. Marine oil, and other pros-taglandin precursor, supplementation for pregnancy uncom-plicated by pre-eclampsia or intrauterine growth restriction.Cochrane Database Syst Rev. 2006;3:CD003402.

44. Szajewska H, Horvath A, Koletzko B. Effect of n-3 long-chainpolyunsaturated fatty acid supplementation of women withlow-risk pregnancies on pregnancy outcomes and growthmeasures at birth: a meta-analysis of randomized controlledtrials. Am J Clin Nutr. 2006;83(6):1337–1344.

45. Cunningham FG, Williams JW. Williams Obstetrics. NewYork, NY: McGraw-Hill Professional; 2005.

46. Shea KM,Wilcox AJ, Little RE. Postterm delivery: a challengefor epidemiologic research. Epidemiology. 1998;9(2):199–204.

47. Olsen SF, Østerdal ML, Salvig JD, et al. Duration of preg-nancy in relation to fish oil supplementation and habitual fishintake: a randomised clinical trial with fish oil. Eur J Clin Nutr.2007;61(8):976–985.

48. Schaiff WT, Knapp FF Jr, Barak Y, et al. Ligand-activatedperoxisome proliferator activated receptor gamma alters pla-cental morphology and placental fatty acid uptake in mice.Endocrinology. 2007;148(8):3625–3634.

49. Hediger M, Kiely J. Foreword. Paediatr Perinat Epidemiol.2007;21(suppl 2):1–3.

50. Savitz DA, Terry JW Jr, Dole N, et al. Comparison of preg-nancy dating by last menstrual period, ultrasound scanning,and their combination. Am J Obstet Gynecol. 2002;187(6):1660–1666.


Am J Epidemiol 2009;169:1015–1024

http://www.cdc.gov/nchs/about/major/nhanes/nhanes01-02.htm




http://www.cdc.gov/nchs/tutorials/NHANES/index_current.htm

http://www.cdc.gov/nchs/tutorials/NHANES/index_current.htm

51. Dietz PM, England LJ, Callaghan WM, et al. A comparison ofLMP-based and ultrasound-based estimates of gestational ageusing linked California livebirth and prenatal screening rec-ords. Paediatr Perinat Epidemiol. 2007;21(suppl 2):62–71.

52. Ananth CV. Menstrual versus clinical estimate of gestationalage dating in the United States: temporal trends and variabilityin indices of perinatal outcomes. Paediatr Perinat Epidemiol.2007;21(suppl 2):22–30.

1024 Adibi et al.

Am J Epidemiol 2009;169:1015–1024




Vol. 169, No. 8

DOI: 10.1093/aje/kwp004



Longitudinal Trends in Hazardous Alcohol Consumption Among Women WithHuman Immunodeficiency Virus Infection, 1995–2006

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


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

alcohol drinking; HIV; longitudinal studies; women

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

Between 1990 and 2000, the US human immunodefi-ciency virus (HIV) epidemic expanded to increasinglyinclude women. Today, women account for more thanone-quarter of all new diagnoses of HIV infection and ac-quired immunodeficiency syndrome in the United States (1).As a growing number of women become infected with HIV,the need to identify and address modifiable factors that af-fect disease progression or survival becomes more urgent.

Hazardous alcohol use is one such modifiable behavior.Hazardous alcohol use is defined as a pattern of alcoholconsumption that is associated with an increase in alcohol-related or other health problems but may not necessarily beclassified as alcoholism (2). The National Institute on Alco-hol Abuse and Alcoholism defines hazardous alcohol use for

women as an average consumption of>7 drinks per week or�4 drinks at 1 sitting at least weekly. It is important todistinguish hazardous alcohol use from nonhazardous use,since low-to-moderate levels of alcohol consumption areoften associated with improved health outcomes and sur-vival, whereas hazardous consumption is associated withpoorer outcomes (3, 4).

Although studies carried out before the era of highly ac-tive antiretroviral therapy (HAART) did not find alcoholconsumption to be associated with disease progression (5),more recent studies have demonstrated a significant associ-ation between hazardous alcohol consumption and severaladverse health outcomes and behaviors in HIV-positivewomen. These include increased HIV viral load, lower

Correspondence to Dr. Robert L. Cook, Department of Epidemiology and Biostatistics, College of Public Health and Health Professions,

University of Florida, P.O. Box 100231, Gainesville, FL 32607 (e-mail: [email protected]).

1025 Am J Epidemiol 2009;169:1025–1032

medication adherence, increased risky sexual behavior, andmore rapid disease progression (6–12). Women who arecoinfected with HIVand hepatitis C virus (HCV) are at evengreater risk of alcohol-associated health problems.

Previous studies have suggested that 6%–54% of HIV-positive women meet criteria for hazardous drinking in dif-ferent settings, depending on the specific measure of haz-ardous drinking used (6, 13–16). These studies were limitedby assessment at only 1 time point or recruitment froma single setting. Longitudinal assessment of alcohol con-sumption is important, because drinking behavior maychange with increasing age or with HIV disease progression.Drinking patterns may also be influenced by period effects(factors common to all age groups at a particular time point)or cohort effects (factors common to all women enrolled ina study in a given year) (17). Furthermore, many drinkers donot show consistent behavior, varying their drinking patternand volume over time (17).

Here we describe longitudinal patterns of alcohol con-sumption in HIV-positive women. Identification of demo-graphic, clinical, or behavioral characteristics of women atincreased risk for hazardous drinking might help cliniciansrecognize those women, so that increased screening, pre-vention, and therapeutic services could be provided. Ourobjectives in this study were to determine the proportionof HIV-positive women with hazardous alcohol consump-tion over 11 years of follow-up and to identify factors asso-ciated with hazardous drinking in these women.


Study population

Data were obtained from women participating in theWomen’s Interagency HIV Study (WIHS), the largest ongo-ing longitudinal cohort study of HIV-positive women in theUnited States. The WIHS is a multicenter, prospective studythat was established in 1994 to carry out comprehensiveinvestigations of the impact of HIV infection among womenaged �18 years. In a second recruitment phase, new mem-bers were enrolled in the study during 2000–2001. Womenare seen at semiannual study visits. Our analyses were lim-ited to 2,770 women who were confirmed by serologicanalysis to be HIV-positive at study entry.

Participating study sites are centered in the Bronx andBrooklyn, New York; Washington, DC; Los Angeles andSan Francisco, California; and Chicago, Illinois. WIHS par-ticipants were recruited from a variety of sources, includingHIV primary care clinics, hospital-based programs, researchprograms, community outreach sites, women’s supportgroups, drug rehabilitation programs, HIV testing sites,and referrals from enrolled participants. The study designhas been described previously (18, 19), and additional infor-mation is available at the study’s Web site (https://statepiaps.jhsph.edu/wihs/).

The WIHS participants undergo semiannual physical ex-aminations, complete study questionnaires, and provideblood for biomarker measurements. The semiannual ques-tionnaire data are obtained face to face by interviewers whoreceive extensive standardized training regarding how to ask

questions on sensitive issues and how to minimize socialdesirability bias. Participants who report heavy alcohol con-sumption or drug use during the interviews are referred tosubstance abuse treatment programs in the community withestablished linkages to the WIHS sites. The number ofwomen who pursue such referrals is not known.

Institutional review boards at each of the study centersand their community affiliates approved the WIHS proto-cols, and informed written consent was obtained from allparticipants.

Measures

Alcohol consumption. The WIHS has collected data onquantity and frequency of alcohol consumption since itsinception. At each visit, WIHS participants are asked toreport the average number of days per week on which theyhave consumed an alcoholic drink since their previous as-sessment 6 months earlier. A drink is defined as ‘‘1 can,bottle, or glass of beer, a glass of wine, a shot of liquor,a mixed drink with that amount of liquor, or any other kindof alcoholic beverage.’’ Response options are every day, 5–6days per week, 3–4 days per week, 1–2 days per week, lessthan once per week, and none. Next, participants are askedabout the usual number of drinks they have consumed perday since their previous assessment. Open-ended responseswere elicited between 1996 and 2002, and choices werecategorized for visits taking place after October 2002(0, 1–2, 3–4, 5–6, or�7 drinks per day).Open-ended responsessuch as ‘‘a pint of vodka’’ (0.5 L) are converted to numbersof standard drinks. At baseline, women also provided in-formation on whether they had ever received treatment forproblematic alcohol consumption, and if so, which type oftreatment (e.g., inpatient alcohol detoxification, outpatient al-cohol detoxification, attendance at Alcoholics Anonymous).

Using the WIHS data from visits 1–23 (1995–2006), wecreated 2 types of alcohol consumption measures based onthe 2 questions relating to hazardous drinking and based thecutpoint for these analyses on guidelines issued by theNational Institute onAlcoholAbuse andAlcoholism (2). First,we calculated the average quantity of drinks per week bymultiplying quantity by frequency. For response items thatinvolved a range, we used the midpoint of the range. Weeklyconsumption was also categorized as >7 drinks per week(excessive weekly consumption) or �7 drinks per week.Second, we used the average number of drinks consumedper day to categorize women into those who consumed �4drinks per day (excessive daily consumption) and those whoconsumed <4 drinks per day. Then, using these cutpoints ateach visit, we categorized a woman’s alcohol consumptioninto 1 of 3 groups: hazardous drinking (>7 drinks per weekor �4 drinks per day), moderate drinking (any drinking thatdid not qualify as hazardous drinking), or nondrinking. Wealso created a dichotomous variable for any hazardousdrinking in the past year (i.e., current visit or previous visit).

Additional measures. Women provided informationabout their age, race, educational attainment, marital status,and employment status at the time of study enrollment. Wecategorized the women according to their geographic site ofenrollment and study enrollment time (original cohort,

1026 Cook et al.

Am J Epidemiol 2009;169:1025–1032

https://statepiaps.jhsph.edu/wihs/

https://statepiaps.jhsph.edu/wihs/

recruited in 1994–1995, or second cohort, recruited in2001–2002). Illicit drug use was assessed at each visit byasking participants to indicate the quantity and frequency ofuse of the following substances since their previous visit:tobacco, marijuana, cocaine, ‘‘crack’’ or freebase cocaine,heroin, and methadone. Current drug use was defined astaking the drug at least once per month. Participants alsoreported whether they had ever injected any drugs andwhether they were currently injecting drugs.

At each visit, a CD4 cell count was determined, with theresults categorized as <200 cells/mL, 200–500 cells/mL,and >500 cells/mL. Hepatitis C infection was determinedby antibody testing from blood collected at study enrollment(actual testing was done several years after enrollment).Symptoms of depression were measured at each visit withthe 20-item Center for Epidemiologic Studies DepressionScale (20); women were classified as having symptoms ofdepression if they had a score of �16.


We compared characteristics of HIV-positive women withand without hazardous drinking behavior at baseline or thefirst follow-up visit. We used the t test to compare meanvalues and the chi-squared test to compare categorical var-iables (SAS, version 9.1; SAS Institute Inc., Cary, NorthCarolina). Next, we used multivariable logistic regressionto assess the effect of sociodemographic and clinical vari-ables on the risk of hazardous drinking at each visit. Weused a generalized estimating equations approach, allowingparticipants to contribute data from each time point, withadjustment for repeated measures within subjects. All socio-demographic factors and clinical variables were selected forthe initial model, with the exception of drug use, which wasleft out of the model because of significant collinearity withthe alcohol measure. Using stepwise elimination, the leastsignificant variable was dropped from the current modeluntil all of the variables left in the model were significant(P < 0.05). The corresponding estimates were comparedwith the initial estimates from the full model, and thechanges were negligible. These analyses were repeated ina stratified manner according to WIHS enrollment cohortdate (1995–1996 or 2001–2002); since the results were sim-ilar, we present the combined data as representative of theentire sample.

We plotted the proportion of women with hazardousdrinking or moderate drinking in the past year, overall andstratified by HCV status at baseline. To assess whether theproportion of women who were hazardous drinkers might beaffected by study loss to follow-up or differential survivalrates, we repeated these analyses using only data fromwomen who completed 1 of the last 2 visits. The observedtrends were similar; therefore, we present results from theentire cohort.

RESULTS

Baseline characteristics of the 2,770 HIV-infectedwomen, overall and stratified by baseline hazardous drink-ing behavior pattern, are shown in Table 1. Women report-

ing hazardous drinking at baseline were significantly morelikely to be older than 40 years, to be African-American, tobe unemployed, to have not completed high school, or to bepart of the original (1994–1995) cohort. Women with haz-ardous drinking patterns were also significantly more likelyto be HCV-seropositive and to have high levels of depressivesymptoms. They were also more likely to smoke cigarettes,to use other drugs, and to report current or past injectiondrug use.

Among women who did not engage in hazardous drinkingat baseline, 981 (45%) were drinking at moderate levels, 380(17%) were current nondrinkers with a history of hazardousdrinking, and 825 (37%) were current nondrinkers with nohistory of hazardous drinking. Among women who did en-gage in hazardous drinking at baseline, 138 (25%) exceededweekly consumption levels only, 142 (25%) exceeded dailyconsumption levels only, and 282 (50%) exceeded bothweekly and daily consumption levels. Overall, 17% of thewomen had received 1 or more types of treatment for alco-hol problems in the past, with the most common beinginpatient alcohol detoxification (13%), Alcoholics Anony-mous attendance (12%), and outpatient alcohol detoxifica-tion (8%).

Figure 1 shows the proportion of women who had con-sumed alcohol in the previous year during the period 1995–2006. Over this interval, 14%–24% of women met thecriteria for hazardous drinking, and 32%–48% of womenmet the criteria for moderate drinking. There was a gradualdecline in the proportion of women who met the criteria forhazardous drinking over time, with a corresponding increasein the proportion of women who met the criteria for mod-erate drinking during follow-up.

Figure 2 shows the proportion of HIV-positive womenwho reported hazardous drinking in the previous year ac-cording to their HCV serostatus at study enrollment. In themid-to-late 1990s, HCV-seropositive women were muchmore likely to be hazardous drinkers than HCV-seronegativewomen. However, this difference decreased over time, andthere was no significant difference in hazardous drinkingbehavior according to HCV status at the most recent timepoints.

In multivariable analysis, women were slightly but statis-tically significantly less likely to engage in hazardous drink-ing with each follow-up visit, a finding that was independentof age (Table 2). Women were also significantly less likelyto be hazardous drinkers if they were employed at baselineor if they had completed high school. In contrast, womenwere significantly more likely to be hazardous drinkers ifthey had CD4 cell counts of 200–500 cells/mL, were HCV-seropositive, or had high levels of depressive symptoms.There was no significant association of hazardous drinkingwith age, race, or enrollment cohort.

DISCUSSION

In this paper, we report patterns of alcohol consumptionbetween 1995 and 2006 over time in a geographically di-verse sample of 2,770 HIV-positive women from the WIHScohort. We found that at any time point, 14%–24% of

Alcohol Consumption in HIV-Positive Women 1027

Am J Epidemiol 2009;169:1025–1032

Table 1. Baseline Characteristics of HIV-Positive Women According to Self-reported Alcohol Consumption Status During the Past Year,

Women’s Interagency HIV Study, 1995–2006

Total (n 5 2,770)

Alcohol Consumption Statusa

P ValueHazardous

Drinking (n 5 562)

NonhazardousDrinking

(Moderate/None)(n 5 2,208)

No. % No. % No. %

Sociodemographic factors

Mean age,b years 36 (7.8)c 37 (7.3) 36 (7.9) 0.06

Age group,b years

<30 642 23.2 102 18.2 540 24.5

30–40 1,334 48.2 257 45.7 1,077 48.8

>40 791 28.6 203 36.1 588 26.7 <0.0001

Race

White (non-Hispanic) 422 15.2 79 14.1 343 15.5

African-American (non-Hispanic) 1,606 58.0 366 65.1 1,240 56.2

Hispanic 654 23.6 104 18.5 550 24.9

Other 88 3.2 13 2.3 75 3.4 0.001

Marital statusb

Single (never married) 910 33.4 200 36.0 710 32.7

Married 1,010 37.0 192 34.6 818 37.7

Separated/divorced/widowed 807 29.6 163 29.4 644 29.7 0.27

Employedb 671 24.3 84 15.0 587 26.6 <0.0001

More than a high school educationb 878 31.8 141 25.1 737 33.5 0.0002

Study site

Bronx, New York 545 19.7 103 18.3 442 20.0

Brooklyn, New York 454 16.4 77 13.7 377 17.1

Washington, DC 411 14.8 79 14.1 332 15.0

Los Angeles, California 569 20.5 99 17.6 470 21.3

San Francisco, California 419 15.1 109 19.4 310 14.0

Chicago, Illinois 372 13.4 95 16.9 277 12.5 0.0006

Enrollment cohort

1994–1995 2,041 73.7 493 87.7 1,548 70.1

2001–2002 729 26.3 69 12.3 660 29.9 <0.0001

Clinical characteristics

CD4 cell count,b cells/mL

<200 660 24.6 119 22.0 541 25.2

200–500 1,164 43.4 254 47.0 910 42.4

>500 860 32.0 167 30.9 693 32.3 0.13

Hepatitis C-seropositiveb 948 35.3 290 53.0 658 30.7 <0.0001

Depressive symptoms (CES-D score �16)b 1,467 54.1 366 66.7 1,101 50.9 <0.0001

Drug useb

Current use

Cocaine 321 11.6 170 30.2 151 6.8 <0.0001

Heroin 282 10.2 121 21.5 161 7.3 <0.0001

Crack/freebase cocaine 437 15.8 222 39.5 215 9.7 <0.0001

(Illicit) methadone 38 1.4 19 3.4 19 0.9 <0.0001

Marijuana/hashish 591 21.4 236 42.1 355 16.1 <0.0001

Tobacco 1,415 51.1 437 77.8 978 44.4 <0.0001

Ever use of injected drugs 914 33.0 292 52.0 622 28.2 <0.0001

Abbreviations: CES-D, Center for Epidemiologic Studies Depression [Scale]; HIV, human immunodeficiency virus.a Hazardous drinking was defined as excessive weekly consumption (>7 drinks per week) or excessive daily consumption (�4 drinks per

occasion). Information on drinking status for the first 2 visits was missing for 45 women.b Information on this measure was missing for some women.c Numbers in parentheses, standard deviation.

1028 Cook et al.

Am J Epidemiol 2009;169:1025–1032

women reported consumption of alcohol at hazardous levelsin the past year, while over 50% of women had consumedany alcohol.

The proportion of women with hazardous drinking at anytime point in our study was similar to or greater than theproportion in other cross-sectional studies that evaluatedhazardous alcohol consumption among HIV-positivewomen (6%–54%) (6–12). However, it is difficult to directlycompare proportions of women with hazardous drinkingacross these studies because of differences in the measure-ment of hazardous drinking. The proportions in our samplewere also similar to a recent (2001–2002) general popula-

tion estimate for US women aged 18 years or older, whichfound that 22% of US adult women engage in hazardousdrinking (21).

Our analysis differs from previous studies of hazardousdrinking in HIV-positive women by documenting trendsover time. This allowed us to see, for example, that theproportion of women with hazardous drinking behavior de-clined gradually during the time period of 1995–2006,whereas the overall proportion of women with any alcoholconsumption remained stable. Several other studies havedemonstrated a gradual decrease in hazardous drinking withincreasing age (17, 22). However, there is also evidence of

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

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


Am J Epidemiol 2009;169:1025–1032

mixed longitudinal patterns, with a significant proportion ofolder women initiating hazardous drinking in middle to latelife (22). Our finding that the decrease in hazardous drinkingover time was independent of age or study cohort suggeststhat the trend may represent a period effect, which refers togeneral changes in the population related to year of datacollection (23). We also considered whether the gradual de-crease in hazardous drinking could be related to differentialloss to follow-up or shorter survival among women withhazardous drinking, but additional analyses did not supportthis explanation. Although researchers in prior analyses ofretention rates among women enrolled in the original cohortdid not find a difference between heavier drinkers and thosewho drank less (24), further investigation of the relationbetween hazardous drinking and survival is warranted.

To date, conclusions about the relative impact of hazard-ous drinking on major health outcomes such as survival ordisease progression in HIV-positive persons have beenmixed. In the pre-HAART era, most studies showed no ev-idence of an effect of alcohol consumption on survival (5,25). However, in more recent studies, investigators haveconcluded that hazardous alcohol consumption is correlated

with more rapid disease progression (9, 10, 12). One possi-ble hypothesis regarding this shift is that in the HAART era,people with HIV infection are feeling better and thereforecontinuing or increasing their use of recreational substancessuch as alcohol. Alternatively, hazardous drinking could di-rectly shorten survival through its effects on the immunesystem, adverse health behaviors such as nonadherence tomedication protocols, or HIV-related comorbid conditionssuch as hepatitis.

The fact that women with HCV infection drank as muchas or more than women without HCV is concerning, al-though the steeper decline in rates of hazardous drinkingamong women infected with both HIV and HCV is reassur-ing. Liver disease, mostly related to HCV, is now the leadingcause of non-HIV-related death in women with HIV/acquired immunodeficiency syndrome (26). Hazardous al-cohol consumption is also a significant barrier to receivingtreatment for HCV infection (27). The more rapid declineamong women dually infected with both HIV and HCVcould be due to increased awareness and clinical interven-tion, to decreased survival, or to some combination of thesefactors. In this study, baseline HCVantibody testing was notconducted for several years. Therefore, many women werenot initially aware of their HCV status, and they may havereduced their drinking only after learning they were HCV-seropositive (28).

The overlap of hazardous alcohol consumption with druguse is also important to note. At baseline, over one-third ofthe women with hazardous drinking also had a history ofinjection or noninjection drug use. The relative contributionof hazardous alcohol consumption versus other drugs to ma-jor health outcomes in HIV infection can be difficult to dis-cern. Noninjection drug use has been associated with HIVprogression and all-cause mortality in this cohort (29). Inaddition, nearly all of the women with HCV infection in thissample had a history of current or past injection drug use,which could party explain the increased prevalence of haz-ardous drinking among HCV-seropositive women.

Several issues related to self-reported alcohol consump-tion measures warrant mention. Only a few alcohol con-sumption measurements were obtained at each semiannualstudy visit (i.e., quantity and frequency), limiting our abilityto examine participants for additional drinking patterns suchas periodic binge drinking. Measures of quantity and fre-quency also occasionally lead to underestimates in compar-ison with other drinking assessment methods (30). Inaddition, the definition of a standard drink of alcohol mayhave been inconsistent across women. Early in the study,women could indicate specific types of beverages andamounts that did not easily translate into numbers of stan-dard drinks (e.g., a pint (0.5 L) of vodka). However, later inthe study, we did not obtain this open-ended information,which could account for part of the observed shift fromhazardous drinking to moderate drinking during the lastfew years of follow-up. Hazardous drinking was estimatedfor some women because of response items that includedranges, and because drinking patterns during the past monthcould have varied. Social desirability can also result in un-derestimation of heavy drinking in some women. Althoughthese measurement biases could point in either direction,

Table 2. Significant Predictors of Hazardous Alcohol Consumptiona

at Each Time Point in HIV-Positive Women (Multivariate Analyses),

Women’s Interagency HIV Study, 1995–2006b

ParameterOddsRatio


P Value

Current visit (vs. previous visit) 0.96 0.96, 0.97 <0.0001

Employment (yes vs. no) 0.80 0.72, 0.90 0.0002

More than a high schooleducation (vs. high schoolor less)

0.72 0.59, 0.87 0.0009

Study site

Brooklyn, New York 1 Referent

Bronx, New York 1.15 0.87, 1.52 0.33

Washington, DC 1.65 1.23, 2.21 0.0009

Los Angeles, California 1.18 0.87, 1.59 0.28

San Francisco, California 1.49 1.12, 1.99 0.007

Chicago, Illinois 1.31 0.96, 1.77 0.09

CD4 cell count, cells/mL

>500 1 Referent

200–500 1.10 1.00, 1.21 0.05

<200 1.06 0.95, 1.24 0.22

Hepatitis C-seropositive(yes vs. no)

1.62 1.35, 1.95 <0.0001

Depressive symptoms(CES-D score �16)(yes vs. no)

1.31 1.21, 1.43 <0.0001

Abbreviations: CES-D, Center for Epidemiologic Studies Depres-

sion [Scale]; HIV, human immunodeficiency virus.a Hazardous drinking was defined as excessive weekly consump-

tion (>7 drinks per week) or excessive daily consumption (�4 drinks

per occasion).b Other variables in the model that were not statistically significant

included age, race, and enrollment cohort (1994–1995 or 2001–

2002).

1030 Cook et al.

Am J Epidemiol 2009;169:1025–1032

they are probably more likely to underestimate drinking,and thus the true rate of hazardous drinking could be higher.

In summary, 14%–24% of HIV-positive women in theWIHS met criteria for past-year hazardous drinking overa decade of follow-up. Although more than half of theHIV-positive women in this study consumed at least somealcohol, interventions should attempt to target those womenwho are drinking at hazardous or unhealthy levels. In thecurrent era of HAART, hazardous drinking could havea greater impact on health outcomes because of its associ-ation with nonadherence to medication protocols. Data col-lection for the WIHS started at the beginning of the HAARTera, and it is possible that some women cut down on theirdrinking in order to improve their chances of survival.Women with additional risk factors such as depression orHCV coinfection may need even more intensive screeningand targeted intervention. Advice on how to screen and in-tervene is available (2, 31, 32); yet despite the consistentevidence of harm associated with hazardous drinking, inmany clinical settings hazardous alcohol consumption isoften neither detected nor addressed (33). Although thestudy design does not allow for formal cause-effect conclu-sions, these findings suggest that women with identified riskfactors such as depression or HCV infection may need ad-ditional or more aggressive screening for hazardous drink-ing. Further research and guidance are needed to translateeffective alcohol-related interventions into clinical practiceand to determine the impact of such interventions on thehealth outcomes of women with HIV infection.

ACKNOWLEDGMENTS

Author affiliations: Department of Epidemiology andBiostatistics, College of Public Health and Health Profes-sions, University of Florida, Gainesville, Florida (Robert L.Cook); Department of Medicine, School of Medicine, Uni-versity of Pittsburgh, Pittsburgh, Pennsylvania (Fang Zuy,Bea Herbeck Belnap); The CORE Center, Cook CountyBureau of Health Services, Chicago, Illinois (KathleenWeber, Mardge H. Cohen); Department of Internal Medicine,Rush School of Medicine, Rush University, Chicago,Illinois (Kathleen Weber, Mardge H. Cohen); Departmentof Psychiatry, College of Medicine at Chicago, University ofIllinois, Chicago, Illinois (Judith A. Cook); New York Acad-emy of Medicine, New York, New York (David Vlahov);Department of Preventive Medicine and CommunityHealth, SUNY Downstate Medical Center, Brooklyn, NewYork (Tracey E. Wilson); Departments of Clinical Pharmacyand Medicine, Schools of Pharmacy and Medicine, Univer-sity of California, San Francisco, San Francisco, California(Nancy A. Hessol); Department of Medicine, GeorgetownUniversity Medical Center, Washington, DC (MichaelPlankey); Department of Epidemiology and PopulationHealth, Montefiore Medical Center–Albert Einstein College ofMedicine, Bronx, New York (Andrea A. Howard); Depart-ment of Epidemiology, School of Public Health, Universityof North Carolina, Chapel Hill, North Carolina (Stephen R.Cole); National Institute of Allergy and Infectious Diseases,

Bethesda, Maryland (Gerald B. Sharp); and Department ofPreventive Medicine, Keck School of Medicine, Universityof Southern California, Los Angeles, California (Jean L.Richardson).

The Women’s Interagency HIV Study (WIHS) is fundedby the National Institute of Allergy and Infectious Diseases(grants UO1-AI-35004, UO1-AI-31834, UO1-AI-34994,UO1-AI-34989, UO1-AI-34993, and UO1-AI-42590) andthe National Institute of Child Health and Human Develop-ment (grant UO1-HD-32632). The study is cofunded by theNational Cancer Institute, the National Institute on DrugAbuse, and the National Institute on Deafness and OtherCommunication Disorders. Funding is also provided bythe National Center for Research Resources (University ofCalifornia, San Francisco–Clinical and TranslationalScience Institute grant UL1 RR024131).

Data for this analysis were collected by the WIHSCollaborative Study Group at the following centers: NewYorkCity/Bronx Consortium (Kathryn Anastos, Principal Inves-tigator (PI)); Brooklyn, New York (Howard Minkoff, PI);Washington, DC, Metropolitan Consortium (Mary Young,PI); The Connie Wofsy Study Consortium of NorthernCalifornia (Ruth Greenblatt, PI); Los Angeles County/Southern California Consortium (Alexandra Levine, PI);Chicago Consortium (Mardge Cohen, PI); and Data Coor-dinating Center (Baltimore, Maryland) (Stephen Gange, PI).

The contents of this publication are solely the responsi-bility of the authors and do not necessarily represent theofficial views of the National Institutes of Health.


REFERENCES

1. Centers for Disease Control and Prevention. HIV/AIDS AmongWomen. (CDC HIV/AIDS Fact Sheet). Atlanta, GA: Centersfor Disease Control and Prevention; 2008. (http://www.cdc.gov/hiv/topics/women/resources/factsheets/pdf/women.pdf).(Accessed January 4, 2009).

2. National Institute on Alcohol Abuse and Alcoholism. HelpingPatients Who Drink Too Much: A Clinician’s Guide. Bethesda,MD: National Institute on Alcohol Abuse and Alcoholism;2005. (NIH publication no. 07-3769). (http://pubs.niaaa.nih.gov/publications/Practitioner/CliniciansGuide2005/guide.pdf).(Accessed January 4, 2009).

3. Caetano R, Ramisetty-Mikler S, Floyd LR, et al. The epide-miology of drinking among women of child-bearing age. Al-cohol Clin Exp Res. 2006;30(6):1023–1030.

4. Mukamal KJ, Conigrave KM, Mittleman MA, et al. Roles ofdrinking pattern and type of alcohol consumed in coronaryheart disease in men. N Engl J Med. 2003;348(2):109–118.

5. Kaslow RA, Blackwelder WC, Ostrow DG, et al. No evidencefor a role of alcohol or other psychoactive drugs in acceler-ating immunodeficiency in HIV-1-positive individuals. A re-port from the Multicenter AIDS Cohort Study. JAMA. 1989;261(23):3424–3429.

6. Cook RL, Sereika SM, Hunt SC, et al. Problem drinking andmedication adherence among persons with HIV infection.J Gen Intern Med. 2001;16(2):83–88.

7. National Institute on Alcohol Abuse and Alcoholism. Alcoholand HIV/AIDS. (Alcohol Alert no. 57). Bethesda, MD:


Am J Epidemiol 2009;169:1025–1032

http://www.cdc.gov/hiv/topics/women/resources/factsheets/pdf/women.pdf

http://www.cdc.gov/hiv/topics/women/resources/factsheets/pdf/women.pdf

http://pubs.niaaa.nih.gov/publications/Practitioner/CliniciansGuide2005/guide.pdf

http://pubs.niaaa.nih.gov/publications/Practitioner/CliniciansGuide2005/guide.pdf

National Institute on Alcohol Abuse and Alcoholism; 2002.(http://pubs.niaaa.nih.gov/publications/aa57.htm). (AccessedJanuary 4, 2009).

8. Wilson TE, Massad LS, Riester KA, et al. Sexual contracep-tive, and drug use behaviors of women with HIV and those athigh risk for infection: results from the Women’s InteragencyHIV Study. AIDS. 1999;13(5):591–598.

9. Samet JH, Horton NJ, Traphagen ET, et al. Alcohol con-sumption and HIV disease progression: are they related?Alcohol Clin Exp Res. 2003;27(5):862–867.

10. Theall KP, Clark RA, Powell A, et al. Alcohol consumption,ART usage and high-risk sex among women infected withHIV. AIDS Behav. 2007;11(2):205–215.

11. Braithwaite RS, McGinnis KA, Conigliaro J, et al. A temporaland dose-response association between alcohol consumptionand medication adherence among veterans in care. AlcoholClin Exp Res. 2005;29(7):1190–1197.

12. Chander G, Lau B, Moore RD. Hazardous alcohol use: a riskfactor for non-adherence and lack of suppression in HIV in-fection. J Acquir Immune Defic Syndr. 2006;43(4):411–417.

13. Galvan FH, Bing EG, Fleishman JA, et al. The prevalence ofalcohol consumption and heavy drinking among people withHIV in the United States: results from the HIV Cost and Serv-ices Utilization Study. J Stud Alcohol. 2002;63(2):179–186.

14. Seth P, Wingood GM, Diclemente RJ. Exposure to alcoholproblems and its association with sexual behavior andbiologically-confirmed Trichomonas vaginalis among womenliving with HIV. Sex Transm Infect. 2008;84(5):390–392.

15. Theall KP, Amedee A, Clark RA, et al. Alcohol consumptionand HIV-1 vaginal RNA shedding among women. J StudAlcohol Drugs. 2008;69(3):454–458.

16. Metsch LR, Pereyra M, Colfax G, et al. HIV-positive patients’discussion of alcohol use with their HIV primary care pro-viders. Drug Alcohol Depend. 2008;95(1-2):37–44.

17. Moore AA, Gould R, Reuben DB, et al. Longitudinal patternsand predictors of alcohol consumption in the United States.Am J Public Health. 2005;95(3):458–465.

18. Barkan SE, Melnick SL, Preston-Martin S, et al. The Women’sInteragency HIV Study. WIHS Collaborative Study Group.Epidemiology. 1998;9(2):117–124.

19. Bacon MC, von Wyl V, Alden C, et al. The Women’s Inter-agency HIV Study: an observational cohort brings clinicalsciences to the bench. Clin Diagn Lab Immunol. 2005;12(9):1013–1019.

20. Radloff LS. The CES-D Scale: a self-report depression scalefor research in the general population. Appl Psychol Meas.1977;1(3):385–401.

21. National Institute on Alcohol Abuse and Alcoholism. AlcoholUse and Alcohol Use Disorders in the United States: MainFindings from the 2001–2002 National Epidemiologic Surveyon Alcohol and Related Conditions (NESARC). (U.S. Alcohol

Epidemiologic Data Reference Manual, vol 8, no. 1).Bethesda, MD: National Institute on Alcohol Abuse andAlcoholism; 2006. (NIH publication no. 05-5737). (http://defeataddictions.org/files/2001_2002_National_Epidemiologic_Survey.pdf). (Accessed January 4, 2009).

22. Liberto JG, Oslin DW, Ruskin PE. Alcoholism in older per-sons: a review of the literature. Hosp Community Psychiatry.1992;43(10):975–984.

23. Wilsnack RW, Kristjanson AF, Wilsnack SC, et al. Are U.S.women drinking less (or more)? Historical and aging trends,1981–2001. J Stud Alcohol. 2006;67(3):341–348.

24. Hessol NA, Schneider M, Greenblatt RM, et al. Retention ofwomen enrolled in a prospective study of human immunode-ficiency virus infection: impact of race, unstable housing, anduse of human immunodeficiency virus therapy. Am J Epide-miol. 2001;154(6):563–573.

25. Chandiwana SK, Sebit MB, Latif AS, et al. Alcohol con-sumption in HIV-I infected persons: a study of immunologicalmarkers, Harare, Zimbabwe. Cent Afr J Med. 1999;45(11):303–308.

26. Cohen MH, French AL, Benning L, et al. Causes of deathamong women with human immunodeficiency virus infectionin the era of combination antiretroviral therapy. Am J Med.2002;113(2):91–98.

27. Nunes D, Saitz R, Libman H, et al. Barriers to treatment ofhepatitis C in HIV/HCV-coinfected adults with alcohol prob-lems. Alcohol Clin Exp Res. 2006;30(9):1520–1526.

28. Tsui JI, Saitz R, Cheng DM, et al. Awareness of hepatitis Cdiagnosis is associated with less alcohol use among personsco-infected with HIV. J Gen Intern Med. 2007;22(6):822–825.

29. Kapadia F, Cook JA, Cohen MH, et al. The relationship be-tween non-injection drug use behaviors on progression toAIDS and death in a cohort of HIV seropositive women in theera of highly active antiretroviral therapy use. Addiction.2005;100(7):990–1002.

30. Allen JP, Wilson VB, eds. Assessing Alcohol Problems: AGuide for Clinicians and Researchers. 2nd ed. Bethesda, MD:National Institute on Alcohol Abuse and Alcoholism; 2003.(NIH publication no. 03-3745). Available at URL: (http://pubs.niaaa.nih.gov/publications/Assesing%20Alcohol/index.htm).(Accessed January 4, 2009).

31. Saitz R. Unhealthy alcohol use. N Engl J Med. 2005;352(6):596–607.

32. Aharonovich E, Hatzenbuehler ML, Johnston B, et al.A low-cost, sustainable intervention for drinking reductionin the HIV primary care setting. AIDS Care. 2006;18(6):561–568.

33. Conigliaro J, Gordon AJ, McGinnis KA, et al. How harmful ishazardous alcohol use and abuse in HIV infection: do healthcare providers know who is at risk? J Acquir Immune DeficSyndr. 2003;33(4):521–525.

1032 Cook et al.

Am J Epidemiol 2009;169:1025–1032

http://pubs.niaaa.nih.gov/publications/aa57.htm

http://defeataddictions.org/files/2001_2002_National_Epidemiologic_Survey.pdf



http://pubs.niaaa.nih.gov/publications/Assesing%20Alcohol/index.htm

http://pubs.niaaa.nih.gov/publications/Assesing%20Alcohol/index.htm


Published by the Johns Hopkins Bloomberg School of Public Health 2009.

Vol. 169, No. 8

DOI: 10.1093/aje/kwp010



Suicide Mortality Among Patients Receiving Care in the Veterans HealthAdministration Health System

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

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

Understanding and reducing mortality from suicide among veterans is a national priority, particularly for individ-uals receiving care from the US Veterans Health Administration (VHA). This report examines suicide rates amongVHA patients and compares themwith rates in the general population. Suicide mortality was assessed in fiscal year2001 for patients alive at the start of that fiscal year and with VHA use in fiscal years 2000–2001 (n ¼ 4,692,034).Deaths from suicide were identified by using National Death Index data. General population rates were identified byuse of the Web-based Injury Statistics Query and Reporting System. VHA rates were 43.13/100,000 person-yearsfor men and 10.41/100,000 person-years for women. For male patients, the age-adjusted standardized mortalityratio was 1.66; for females, it was 1.87. Male patients aged 30–79 years had increased risks relative to men in thegeneral population; standardized mortality ratios ranged from 2.56 (ages 30–39 years) to 1.33 (ages 70–79 years).Female patients aged 40–59 years had greater risks than did women in the general population, with standardizedmortality ratios of 2.15 (ages 40–49 years) and 2.36 (ages 50–59 years). Findings offer heretofore unavailablecomparison points for health systems. Prior to the conflicts in Afghanistan and Iraq and before recent VHAinitiatives, rates were higher among VHA patients than in the general population. Female patients had particularlyhigh relative risks.

military medicine; mortality; suicide; veterans

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

Understanding and reducing suicide risks among veteransis a national concern and a major priority for the Departmentof Veterans Affairs Veterans Health Administration (VHA)(1–4). Each year, the VHA provides health services to approx-imately 5 million veterans—roughly one-fifth of all veteransliving in the United States. The characteristics of the VHApatient population include important suicide risk factors. TheVHA serves a predominantly male older population, withsubstantial medical morbidities, high levels of substanceabuse and mental illness, and increased knowledge of andaccess to firearms (5–7). However, there have been few com-prehensive assessments of suicide mortality among veteransgenerally (8) and none to our knowledge among veterans re-ceiving services in the VHA health system.

In 2007, Congress required the Department of VeteransAffairs to implement a comprehensive suicide prevention

program (9). This legislation reflects concerns regardingthe high mental health needs observed among veterans ofOperations Enduring Freedom and Iraqi Freedom (10–12),the influx of these veterans now eligible for VHA healthservices (13, 14), and indications that suicide risks may beelevated among veterans compared with nonveterans. Nota-bly, a highly publicized study observed that, in the period1986–1997, male respondents to the US National HealthInterview Survey (NHIS) who indicated veteran status hadgreater suicide risks than those who did not (adjusted hazardratio ¼ 2.04, 95% confidence interval (CI): 1.10, 3.80) (8).

Even before passage of this legislation, however, theVHA had implemented key provisions and system-levelmental health enhancements. Based on its Mental HealthStrategic Plan (3), the VHA has worked to ensure accessto quality mental health services and to develop programs

Correspondence to Dr. John F. McCarthy, US Department of Veterans Affairs Serious Mental Illness Treatment Research and Evaluation Center,

P.O. Box 13017, Ann Arbor, MI 48113-0170 (e-mail: [email protected]).

1033 Am J Epidemiol 2009;169:1033–1038

specifically designed to prevent suicides. For example, since2005, the VHA has added more than 3,600 mental healthstaff, and VHA mental health service expenditures in-creased from $2 billion in fiscal year 2001 to nearly$4 billion in fiscal year 2008 (15). Evaluation of the impactof ongoing changes in the VHA patient population and ofrecent VHA mental health and suicide prevention initiativeswill require assessment of baseline suicide rates amongVHA patients.

Much of the existing research regarding suicide mortalityamong veterans has focused on combat theater veteran co-horts or veteran subgroups with service-related injuries.These have not found increased rates among combat theatercohorts (16–18), but risks have been shown to be higher inassociation with combat trauma and post-traumatic stressdisorder (19, 20).

The few studies that have examined suicide mortalityamong VHA patients have been limited to high-risk patientsubgroups. Increased risks have been demonstrated in pa-tients receiving depression treatment (21), those receivingoutpatient mental health services (22), and those who havereceived psychiatric discharge (23). However, no study hasexamined rates for the entire VHA patient population. In-deed, to date there have been no comprehensive assessmentsof suicide mortality for all patients of any health system ormanaged care organization.

Patterns of suicide risk across age groups may differamong VHA patients compared with risks in the generalpopulation. Older persons are conventionally understoodto have higher risks for suicide than younger persons (24).However, among individuals receiving VHA treatment fordepression, Zivin et al. (21) found that suicide risks werehighest among younger patients (age, <45 years).

It is also important to consider gender differentials insuicide mortality among patients receiving VHA care.Women veterans are one of the fastest growing segmentsof the population of veterans eligible for VHA care (25).In the general population, men are 4 times as likely to diefrom suicide, although women are 3 times as likely to at-tempt suicide (15, 26). However, little is known about gen-der differences in suicide risks among individuals receivingVHA services.

This paper presents new information regarding suicidemortality among all patients receiving VHA services. Weexamined suicide mortality during fiscal year 2001 (October 1,2000–September 30, 2001) among all individuals whowere alive at the start of fiscal year 2001 and who receivedVHA services during either fiscal year 2000 or fiscal year2001. We assessed suicide mortality overall and by age andgender categories. VHA suicide rates were compared withthose of the general US population.


Data sources

This study used data from the VHA’s National Patient CareDatabase to identify all veterans with inpatient or outpatientservices utilization in any VHA facility during fiscal year2000 or fiscal year 2001. Measures of vital status and cause

of death were based on information from the National DeathIndex, which is considered the ‘‘gold standard’’ for mortalityassessment (27). It includes national data regarding dates andcauses of death for all US residents, derived from death cer-tificates filed in state vital statistics offices. National DeathIndex searches were performed for cohorts of VHA patientswho received any VHA services use during fiscal year 2000 orfiscal year 2001 and had no subsequent use through June2006. This cost-efficient method for conducting NationalDeath Index searches enables comprehensive assessment ofvital status and cause of death among all individuals whoreceived VHA services in fiscal year 2000 or fiscal year2001. National Death Index search results often include mul-tiple records that are potential matches. ‘‘True matches’’ wereidentified on the basis of established procedures (28).

Measures

Patients’ age and gender were identified from administra-tive files included in the National Patient Care Database.The age at the start of fiscal year 2001 was categorized asbeing 18–29, 30–39, 40–49, 50–59, 60–69, 70–79, or �80years. Information regarding race/ethnicity was not consis-tently available in the National Patient Care Database for allVHA patients; therefore, it was not possible to calculaterace-adjusted rates.

Using National Death Index data, we identified dates andcauses of death. Suicide deaths were identified by usingInternational Classification of Diseases, Tenth Revision,codes X60–X84, Y87.0, and U03 (29). We calculated thetotal suicide deaths observed in fiscal year 2001 divided bythe person-years at risk for an observed suicide death. Thisquotient was then multiplied by 100,000 to generate suiciderates per 100,000 person-years.

Person-years at risk for an observed suicide death weremade operational as follows. For those veterans who receivedVHA services in fiscal year 2000 (and survived that year), therisk period began on the first day of fiscal year 2001. Forveterans with no VHA services in fiscal year 2000, the riskperiod began on the day of their first VHA use in fiscal year2001. Risk periods ended at either the date of death (from anycause) or the end of the year, whichever came first.

Cohort

The VHA patient population at risk for observed suicidedeaths during fiscal year 2001 included all 4,692,034 indi-viduals with VHA inpatient or outpatient services use infiscal year 2001 or fiscal year 2000 and who were alive atthe start of fiscal year 2001. Of these patients, 21,066 (0.4%)were excluded because of missing or out-of-range values forage or gender. Among the excluded patients, 2 suicides wereobserved. The final cohort consisted of 4,670,968 patients.This project was reviewed and approved by a VHA humansubjects committee.

Rates in the US general population

Suicide rates in the general population were identifiedfrom the Web-based Injury Statistics Query and Reporting

1034 McCarthy et al.

Am J Epidemiol 2009;169:1033–1038

System (WISQARS) (26). These data are derived from theNational Vital Statistics System. Suicide rates in the generalpopulation were assessed overall and by age and gendercategories. These are available by calendar year. Therefore,rates in the general population were calculated for calendaryear 2001.

Analyses

Standardized mortality ratios for suicide risks amongVHA patients were calculated for age and gender subgroupscompared with those in the general US population (30). Wecalculated 95% confidence intervals for the standardizedmortality ratios using the exact method based on Poissondistribution (31).

RESULTS

Table 1 presents characteristics of the VHA patients in-cluded in the risk cohort. The mean age was 59.8 years(standard deviation, 15.7) at the start of fiscal year 2001.The mean number of risk days in fiscal year 2001 was 316.7(standard deviation, 97.0). There were 1,613 deaths bysuicide.

Table 2 presents suicide rates among VHA patients com-pared with the general US population. Overall, for men andwomen combined, suicide risks among VHA patients were66% higher than those observed in the general population(age- and gender-adjusted standardized mortality ratio ¼1.66, 95% CI: 1.58, 1.75). The crude suicide rate amongmale VHA patients was 43.1/100,000 person-years, com-pared with 23.2/100,000 person-years among males in thegeneral population. The standardized mortality ratio formale VHA patients relative to males in the general popula-tion, with adjustment for differences in age distributions,was 1.66 (95% CI: 1.58, 1.74). That is, suicide risks amongmale patients were 66% greater than for males in the gen-eral population. Analyses by age subgroups indicated thatmale VHA patients between 30 and 79 years of age hadgreater suicide risks than did similarly aged men in thegeneral population. Standardized mortality ratios rangedfrom 1.33 among VHA patients aged 70–79 (95% CI:1.20, 1.47) years to 2.56 among patients aged 30–39(95% CI: 2.13, 3.03) years.

Among female patients, the crude suicide rate was 10.41/100,000 person-years, compared with 5.22/100,000 person-years in the general population, with a standardized mortal-ity ratio of 1.87 (95% CI: 1.35, 2.47). Compared withsimilarly aged women in the general population, femalepatients had significantly higher risks: For those aged40–49 years, the standardized mortality ratio was 2.15(95% CI: 1.25, 3.29), and for those aged 50–59 years, itwas 2.36 (95% CI: 1.22, 3.88).

DISCUSSION

Government agencies and the public acknowledge a re-sponsibility to monitor adverse outcomes among individualsdischarged from military service. This responsibility is

heightened when the government is also responsible fortheir health care. This study presents new information re-garding suicide mortality among the entire patient popula-tion of the VHA, representing approximately 5 millionveterans. The VHA is the only large health-care organiza-tion to assess suicide mortality among its entire patient pop-ulation. Thus, comparable data regarding suicide mortalityin other health system populations are not available—eventhough health systems are likely to be important venues forsuicide prevention efforts. This report from the nation’slargest integrated health system provides a heretofore un-available comparison point for other health systems thatmay be initiating surveillance of suicide mortality in theirpopulations. Study findings provide a baseline for evalua-tions of the most comprehensive suicide prevention programever put into place in a health-care system.

Suicide rates among VHA patients were significantlyhigher than those in the general US population. However,the differentials between suicide rates for VHA patients and

Table 1. Characteristics of US VHA Patients With Risk Time in

Fiscal Year 2001a

All Patients Female Male

No. 4,670,968 470,057 4,200,911

% 100.0 10.1 89.9

Age, years

Mean (SD) 59.8 (15.7) 48.4 (16.4) 61.1 (15.1)

Median 61.1 46.6 63.1

Age distribution, %b

18–29 4.3 13.9 3.3

30–39 7.5 18.7 6.2

40–49 15.0 26.3 13.8

50–59 21.6 18.0 22.1

60–69 19.8 10.5 20.9

70–79 24.1 9.1 25.8

�80 7.6 3.5 8.1

Total 99.9 100.0 100.2

With VHA use

In fiscal year 2001but not in fiscalyear 2000, %

28.0 35.6 27.3

In fiscal year 2000but not in fiscalyear 2001, %

17.6 33.1 15.9

Risk days, no.

Mean (SD) 316.7 (97.0) 320.6 (92.0) 316.3 (97.5)

Median 365.0 365.0 365.0

25% quartile 342.0 346.0 342.0

Total suicidedeaths, no.

1,613 43 1,570

Abbreviation: SD, standard deviation; VHA, Veterans Health

Administration.a Patients who had some VHA services utilization in fiscal year

2000 or fiscal year 2001 and who were alive at the start of fiscal year

2001.b Does not sum to 100% because of rounding.

Suicide Mortality Among VHA Health System Patients 1035

Am J Epidemiol 2009;169:1033–1038

for the general US population were not as large as the differ-entials previously estimated for veterans compared withnonveterans in the community (8). Whereas Kaplan et al.(8) reported that risks for male veterans were twice thoseof male nonveterans in the community (adjusted hazardratio ¼ 2.04, 95% CI: 1.10, 3.80), we observed an overallstandardized mortality ratio of 1.66 (95% CI: 1.58, 1.74) formale VHA patients compared with men in the general pop-ulation. However, it is difficult to make direct comparisonsacross studies because of differences in timeframes, popu-lations, and methodologies. Given that VHA patients hadsought health services and likely had greater medical andpsychiatric morbidity than did veterans in the community,one could expect higher relative risks in this treatment-seeking patient population. Alternatively, treatment popula-tions may have lower suicide risks as a function of increasedmental health and substance use care access and treatment.

This study provides new information on age-specificdifferentials in suicide risks among VHA patients relativeto age- and gender-matched comparisons in the generalpopulation. Among male patients, rates were highestamong those aged 30–49 years and lowest among patientsaged 18–29 and 60–69 years. The differential in suiciderates for male patients, relative to men in the general pop-

ulation, was highest for patients aged 30–39 years, and itfell among older patient groups. Further research is neededto evaluate whether these patterns reflect cohort effects orage-patterned selection of VHA services among veterans.Among female patients, rates were highest among thoseaged 50–59 years.

Consistent with the literature, crude rates among femaleVHA patients were substantially lower than those amongmale patients. However, compared with women in the gen-eral population, female VHA patients had high suicide mor-tality, with a standardized mortality ratio of 1.87. This isa larger differential than the standardized mortality ratio of1.66 for male patients.

Increased relative risks among female VHA patients mayresult from not only the characteristics of treatment-seekingpopulations generally (e.g., greater morbidity) but also thefactors unique to women who have served in the military.Female VHA patients may have greater experience with andaccess to firearms than do women in the general population,and they may be more likely to have experienced traumaticepisodes, increasing risks for both post-traumatic stress dis-order and suicide (32–37). In designing and implementingsuicide prevention initiatives, VHA providers must addresssuicide risks among both male and female patients.

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

Group

General US Populationa VA PatientsStandardized

MortalityRatio


Sex- andAge-adjusted

RateSuicides,

no.

GeneralUS Population,

no.

Crude Rate/100,000

Person-Years

Suicides,no.

Person-Years

at Riskb

Crude Rate/100,000

Person-Years

Males, years 23,850 102,872,290 23.18 1,570 3,639,873 43.13 1.66c 1.58, 1.74 38.46

18–29 4,830 23,983,737 20.14 31 111,234 27.87 1.38 0.94, 1.91

30–39 4,702 21,583,911 21.78 126 225,768 55.81 2.56c 2.13, 3.03

40–49 5,146 21,538,269 23.89 287 514,375 55.80 2.34c 2.07, 2.61

50–59 3,554 15,875,192 22.39 381 820,818 46.46 2.08c 1.87, 2.29

60–69 2,030 9,692,852 20.94 247 756,268 32.66 1.56c 1.37, 1.76

70–79 2,072 6,961,055 29.77 366 924,752 39.58 1.33c 1.20, 1.47

�80 1,516 3,237,274 46.83 132 287,458 45.92 0.98 0.82, 1.15

Females, years 5,730 109,719,004 5.22 43 412,873 10.41 1.87c 1.35, 2.47 9.76

18–29 782 22,972,302 3.40 1 52,972 1.89 0.55 0.01, 2.05

30–39 1,155 21,449,920 5.38 7 76,959 9.10 1.69 0.68, 3.15

40–49 1,573 22,024,321 7.14 17 110,730 15.35 2.15c 1.25, 3.29

50–59 1,113 16,725,057 6.65 12 76,254 15.74 2.36c 1.22, 3.88

60–69 514 10,945,634 4.70 3 44,241 6.78 1.44 0.30, 3.48

70–79 367 9,273,850 3.96 3 37,412 8.02 2.03 0.42, 4.88

�80 226 6,327,920 3.57 0 14,306 0.00 0.00 NA

Total 29,580 212,591,294 13.91 1,613 4,052,746 39.80 1.66c 1.58, 1.75 23.15

Abbreviations: ICD-10, International Statistical Classification of Diseases and Related Health Problems, 10th Revision, 2nd ed; NA, not avail-

able; VA, Department of Veterans Affairs; VHA, Veterans Health Administration; WISQARS, Web-based Injury Statistics Query and Reporting

System.a Source: WISQARS, Centers for Disease Control and Prevention (http://www.cdc.gov/ncipc/wisqars/). Suicide identified by ICD-10 diagnosis

codes X60–X84, Y87.0, and U03 (29).b Person-years at risk for being observed with a suicide death during fiscal year 2001. Person-years of risk time are presented here as rounded

numbers.c Standardized mortality ratios where the confidence interval did not contain 1.0.


Am J Epidemiol 2009;169:1033–1038

http://www.cdc.gov/ncipc/wisqars/

Several issues should be considered when interpretingstudy findings. First, there are longstanding concerns re-garding potential undercounting of suicide deaths. Researchis needed to assess potential differential recording of suicidedeaths by veteran status. Second, although we comparedrates among VHA patients with those among the generalUS population, we note that VHA patients are included inthe general population. Third, we note that the VHA patientpopulation includes some nonveterans (e.g., widows or de-pendents of veterans), whose utilization accounts for ap-proximately 1% of VHA inpatient admissions and 2% ofoutpatient visits (38). This study examined suicide mortalityamong all VHA patients, of whom the vast majority areveterans. Finally, it was not possible to adjust rates for race.In the general population, suicide rates are lower amongminority populations, and minorities may be overrepre-sented among VHA patients.

Study findings indicate that VHA patients were at increasedrisks for suicide, yet the differential in risks was less than whatmight be expected given previous comparisons (8). Future re-search is necessary to assess patients’ sociodemographic andclinical factors that may be associated with suicide mortality,to assess the impact of specific health-care services on suiciderisks among veterans, and to evaluate trends in suicide mor-tality among VHA patients and among veterans who do notreceive VHA care. For health-care providers, and particularlythe Veterans Health Administration, research on suicide ratesand risk factors among veterans may directly inform treat-ment and suicide prevention interventions.

ACKNOWLEDGMENTS

Author affiliations: US Department of Veterans AffairsSerious Mental Illness Treatment Research and EvaluationCenter, Ann Arbor, Michigan (John F. McCarthy, MarciaValenstein, Mark Ilgen, Kara Zivin, Frederic C. Blow);US Department of Veterans Affairs Center for ClinicalManagement Research, Ann Arbor, Michigan (John F.McCarthy, Marcia Valenstein, H. Myra Kim, Mark Ilgen,Kara Zivin, Frederic C. Blow); and University of MichiganDepartment of Psychiatry, Ann Arbor, Michigan (John F.McCarthy, Marcia Valenstein, H. Myra Kim, Mark Ilgen,Kara Zivin, Frederic C. Blow).

The authors thank Robert Bilgrad and Michelle Goodier(National Center for Health Statistics) for assistance withsearching the National Death Index and Dr. Ira R. Katz(Veterans Health Administration) and Dr. Katherine J. Hoggatt(University of Michigan) for comments on an earlier versionof this manuscript.


REFERENCES

1. Institute of Medicine. Reducing Suicide: A National Impera-tive. Washington, DC: National Academies Press; 2004.

2. US Department of Health and Human Services. The SurgeonGeneral’s Call to Action to Prevent Suicide. Washington, DC:National Institutes of Health; 1999.

3. Department of Veterans Affairs. The Comprehensive VHAMental Health Strategic Plan. Washington, DC: Departmentof Veterans Affairs; 2005.

4. Office of Inspector General, Department of Veterans Affairs.Healthcare inspection: implementing VHA’s Mental HealthStrategic Plan initiatives for suicide prevention. Report 06-03706-126. Washington, DC: VA Office of Inspector General;2007. (http://www.va.gov/oig/54/reports/VAOIG-06-03706-126.pdf). (Accessed December 4, 2007).

5. Agha Z, Lofgren RP, VanRuiswyk JV, et al. Are patients atVeterans Affairs medical centers sicker? A comparative analy-sis of health status and medical resource use. Arch Intern Med.2000;160(21):3252–3257.

6. Lambert MT, Fowler DR. Suicide risk factors among veterans:risk management in the changing culture of the Department ofVeterans Affairs. J Ment Health Adm. 1997;24(3):350–358.

7. Hankin CS, Spiro A, Miller DR, et al. Mental disorders andmental health treatment among US Department of VeteransAffairs outpatients: the Veterans Health Study. Am J Psychi-atry. 1999;156(12):1924–1930.

8. Kaplan MS, Huguet N, McFarland BH, et al. Suicide amongmale veterans: a prospective population-based study. J Epi-demiol Community Health. 2007;61(7):619–624.

9. The Joshua Omvig Veterans Suicide Prevention Act, Pub LNo. 110-110, 121 Stat 1031-1034. (http://npl.ly.gov.tw/pdf/6173.pdf).

10. Kang HK, Hyams KC. Mental health care needs among recentwar veterans. N Engl J Med. 2005;352(13):1289.

11. Milliken CS, Auchterloni JL, Hoge CW. Longitudinal assess-ment of mental health problems among active and reservecomponent soldiers returning from the Iraq war. JAMA.2007;298(18):2141–2148.

12. Seal KH, Bertentha D, Miner CR, et al. Bringing the war backhome: mental health disorders among 103,788 US veteransreturning from Iraq and Afghanistan seen at Department ofVeterans Affairs facilities. Arch Intern Med. 2007;167(5):476–482.

13. Kang HK. Committee on Evaluation of the VA’s PresumptiveDisability Decision-Making Process. VA health care utiliza-tion among OIF/OEF veterans. Washington, DC: Institute ofMedicine; 2006. (http://www.iom.edu/Object.File/Master/36/181/Kang1.pdf). (Accessed November 17, 2007).

14. Department of Veterans Affairs. Statement of Ira Katz, MD,Deputy Chief Patient Care Services Officer for Mental Health,Department of Veterans Affairs, before House Committee onVeterans Affairs July 25, 2007. Washington, DC: Departmentof Veterans Affairs; 2007. (http://www.va.gov/OCA/testimony/hvac/070725IK.asp). (Accessed December 4, 2007).

15. Office of the Inspector General, Department of Veterans Af-fairs. Statement of Michael Shepherd, MD, Physician, Office ofHealthcare Inspections, Office of Inspector General, Depart-ment of Veterans Affairs, before Special Committee on Aging,United States Senate, on the Department of Veterans Affairsimplementation of suicide prevention initiatives from theMental Health Strategic Plan, October 3, 2007. Washington,DC: Department of Veterans Affairs; 2007. (http://www.va.gov/oig/pubs/VAOIG-statement-20071003-shepherd.pdf).(Accessed December 6, 2007).

16. Boehmer TK, Flanders WD, McGeehin MA, et al. Postservicemortality in Vietnam veterans: 30-year follow-up. Arch InternMed. 2004;164(17):1908–1916.

17. Thomas TL, Kang HK, Dalager NA. Mortality among womenVietnam veterans, 1973–1987. Am J Epidemiol. 1991;134(9):973–980.

Suicide Mortality Among VHA Health System Patients 1037

Am J Epidemiol 2009;169:1033–1038

(http://www.va.gov/oig/54/reports/VAOIG-06-03706-126.pdf).

(http://www.va.gov/oig/54/reports/VAOIG-06-03706-126.pdf).

http://npl.ly.gov.tw/pdf/6173.pdf

http://npl.ly.gov.tw/pdf/6173.pdf

http://www.iom.edu/Object.File/Master/36/181/Kang1.pdf

http://www.iom.edu/Object.File/Master/36/181/Kang1.pdf

http://www.va.gov/OCA/testimony/hvac/070725IK.asp

http://www.va.gov/OCA/testimony/hvac/070725IK.asp

(http://www.va.gov/oig/pubs/VAOIG-statement-20071003-shepherd.pdf).

(http://www.va.gov/oig/pubs/VAOIG-statement-20071003-shepherd.pdf).

18. Kang HK, Bullman TA. Mortality among US veterans of thePersian Gulf War: 7-year follow-up. Am J Epidemiol. 2001;154(5):399–405.

19. Bullman TA, Kang HK. The risk of suicide among woundedVietnam veterans. Am J Public Health. 1996;86(5):662–667.

20. Bullman TA, Kang HK. Posttraumatic stress disorder and therisk of traumatic deaths among Vietnam veterans. J Nerv MentDis. 1994;182(11):604–610.

21. Zivin K, Kim HM, McCarthy JF, et al. Suicide mortalityamong individuals receiving treatment for depression in theVeterans Affairs Health System: associations with patient andtreatment setting characteristics. Am J Public Health. 2007;97(12):2193–2198.

22. Desai MM, Rosenheck RA, Desai RA. Time trends and pre-dictors of suicide among mental health outpatients in theDepartment of Veterans Affairs. J Behav Health Serv Res.2008;35(1):115–124.

23. Desai RA, Dausey DJ, Rosenheck RA. Mental health servicedelivery and suicide risk: the role of individual patient andfacility factors. Am J Psychiatry. 2005;162(2):311–318.

24. National Institute of Mental Health. Older Adults Depressionand Suicide Facts. Washington, DC: National Institute ofMental Health; 2003.

25. Meehan S. Improving health care for women veterans. J GenIntern Med. 2006;21(suppl 3):S1–S2.

26. National Center for Injury Prevention and Control. WISQARSfatal injuries: mortality reports. Atlanta, GA: Centers forDisease Control and Prevention; 2007. (http://webapp.cdc.gov/sasweb/ncipc/mortrate.html). (Accessed November 26,2007).

27. Cowper DC, Kubal JD, Maynard C, et al. A primer and com-parative review of major US mortality databases. Ann Epide-miol. 2002;12(7):462–468.

28. Sohn MW, Arnold N, Maynard C, et al. Accuracy andcompleteness of mortality data in the Department of VeteransAffairs [electronic article]. Popul Health Metr. 2006;4:2.

29. World Health Organization. International Statistical

Classification of Diseases and Related Health Problems. 10th

Revision. 2nd ed. (ICD-10). Geneva, Switzerland: World

Health Organization; 2004.30. Hennekens CH, Buring JE. Epidemiology in Medicine.

Boston, MA: Little, Brown and Company; 1987:82–85.31. Ulm K. A simple method to calculate the confidence interval

of a standardized mortality ratio. Am J Epidemiol. 1990;

131(2):373–375.32. Fontana A, Schwartz LS, Rosenheck R. Posttraumatic stress

disorder among female Vietnam veterans: a causal model of

etiology. Am J Public Health. 1997;87(2):169–175.33. Frayne SM, Seaver MR, Loveland S, et al. Burden of

medical illness in women with depression and post-

traumatic stress disorder. Arch Intern Med. 2004;164(12):

1306–1312.34. Himmelfarb N, Yaeger D, Mintz J. Posttraumatic stress dis-

order in female veterans with military and civilian sexual

trauma. J Trauma Stress. 2006;19(6):837–846.35. Kang H, Dalager N, Mahan C, et al. The role of sexual assault

on the risk of PTSD among Gulf War veterans. Ann Epidemiol.

2005;15(3):191–195.36. Lehmann LL, McCormick RA, McCracken L. Suicidal be-

havior among patients in the VA Health Care System. Psy-

chiatr Serv. 1995;46(10):1069–1071.37. Qin P, Agerbo E, Mortensen PB. Suicide risk in relation to

socioeconomic, demographic, psychiatric, and familial

factors: a national register-based study of all suicides in

Denmark, 1981–1997. Am J Psychiatry. 2003;160(4):

765–772.38. King S, Shane A, Smith MW. A Guide to Identifying Non-

Veteran Records in the Inpatient and Outpatient Databases.

(Technical report #20). Menlo Park, CA: VHA Health

Economics Resource Center (HERC); 2006.


Am J Epidemiol 2009;169:1033–1038

http://webapp.cdc.gov/sasweb/ncipc/mortrate.html

http://webapp.cdc.gov/sasweb/ncipc/mortrate.html




Vol. 169, No. 8

Book Reviews

Hyping Health Risks: Environmental Hazards in Daily Life and theScience of Epidemiology

By Geoffrey C. Kabat

ISBN: 978-0-231-14148-2, Columbia University Press, Irvington, New York (Telephone: 914-591-9111, Fax: 1-800-944-1844, E-mail: [email protected], World Wide Web: http://cup.columbia.edu), 2008, 272 pp., $27.95 Hardcover

For a scholarly book in epidemiology, Hyping HealthRisks by Geoffrey C. Kabat is more reflective and opinion-ated than most and even has somewhat of a plot. Few booksabout epidemiology generate a visceral in addition toa solely cerebral response. On reading the book, I foundpassages causing offense and others eliciting support forKabat’s thoughtful and articulate laments. For the most part,the story of truth and misrepresentation of evidence onhealth risks was engaging, including personal portraits ofgood and evil intent. However, I did not always agree withthe author on who belonged to which category and oftenfound myself thinking, ‘‘that’s not quite the whole story;it’s more complicated than that,’’ or ‘‘I can see why you’dinterpret it that way, but I don’t.’’ Perhaps, being provocativein that way is a real credit to the author.

Kabat has written a fairly succinct treatise (272 pages, butsmall print) making the case that environmental health con-cerns have been given undue attention and too much fundingand have been subjected to insufficient critical evaluationfor a variety of reasons that run counter to good science. Thebulk of the book consists of 4 lengthy, self-containedvignettes on the pursuit of environmental contaminants asa cause of breast cancer and the evaluation and interpreta-tion of health effects of exposures to electromagnetic fields,indoor radon, and environmental tobacco smoke. The exag-gerations by advocates to generate support for these issuesand claimed censoring of critics are portrayed as the productof self-serving, funding-hungry researchers;well-intentioned,but misguided political activists; exploiters from the mediaor real politics; and, more broadly, societal tenacity in pur-suing the wrong things in the wrong way rather than theright things in the right way. While the author does alludein passing to the countervailing arguments—that remotepossibilities of harm from ubiquitous environmental agentsneed careful scrutiny, that the public pays the bills and hassome right to influence funding priorities, that researchersbelieve they are serving public health—these modest at-tempts at balance fall short of a comprehensive apprecia-tion of the complexity and ambiguity of the issues underdiscussion.

The stories of the rise and fall of the controversies arevery nicely written, capturing a time in our recent historythat the author (and many of the potential readers of thebook) lived through. The dynamics and personalities comeacross vividly, with clarity and accuracy regarding the tech-

nical issues behind the controversies and a thoughtful, ifhighly subjective, answer to the question, What happened?The distance of at most a couple of decades and the benefitsof hindsight regarding the science and politics allow theauthor to tell a story with a beginning, middle, and end.Though I suspect that Dr. Kabat will be flattered andoffended by the comparison, it reminds me very much ofPaul Brodeur, who wrote extensively with great influ-ence for the New Yorker, and, to some extent, Gary Taubes,who also weaves science and personality into a compellingstory.

The book provokes reflection and reexamination of ourroles and motives as epidemiologists (my direct experienceconcerned the electromagnetic field saga, for which I wasinterviewed by Dr. Kabat) and, more broadly, of the inter-section of epidemiology with social, political, and ideolog-ical issues. The urge to reflect on our place in society hastended to be one-sided, coming largely from the politicalleft, focusing on the moral obligation to use our science topromote social justice and encouraging us to be activists inorder to change and not just study the world. This bookserves as a powerful counterbalance in reminding us thatsome of the causes for which epidemiologists and their ad-vocates join forces may not be grounded in good science andtherefore not likely to provide public health benefit. To beuseful in advancing public health, epidemiologic findingshave to be valid; when that is not the case, then the enthu-siastic support of the public and regulatory agencies can beharmful to public health by diverting resources and attentionfrom more important issues. This book forcefully examinesthat question—What goes wrong when the good intentionsof scientists and activists are based on weak epidemiologicfindings?

The specific ‘‘problem’’ varies across the 4 topics consid-ered. Regarding the first one, breast cancer activists gener-ated interest from politicians and epidemiologists about thepotential role of environmental pollutants in the causation ofelevated breast cancer rates, particularly on Long Island,New York. In fact, there was no elevation beyond that ex-pected based on the demographics of the affected commu-nities. In the case of health effects of electromagnetic fields,epidemiologists persisted despite the forceful argumentsfrom physicists that biologic or health effects were implau-sible. Indoor radon and environmental tobacco smoke areboth characterized by Kabat as real, but small hazards that

1039 Am J Epidemiol 2009;169:1039–1042

http://cup.columbia.edu

were exaggerated by regulatory agencies (particularly theUS Environmental Protection Agency) and the antitobaccolobby, respectively. In fact, the National Academy of Sci-ences estimated that more than 10% of lung cancer deathsin the United States are attributable to indoor radon, but theauthor notes (correctly) that most of these deaths occuramong smokers, such that if smoking were eliminated, theeffect of radon would be markedly reduced. In the case ofboth indoor radon and environmental tobacco smoke, theauthor expresses deep concern about the distraction fromaddressing active smoking, the real culprit in his view. Alsocommon to the latter 2 issues is the contention that certaintyof evidence and the magnitude of possible effects were ex-aggerated for political reasons. In the case of environmentaltobacco smoke, there is a perceived conspiracy to suppresscriticism of the research or those who generate contradictorydata regarding the impact of environmental tobacco smokeon risk of lung cancer, unfairly lumping such researcherswith the long lineage of tobacco industry shills.

To the best of my knowledge, the author was not incor-rect about any of the important technical points, yet theway he bundled the facts and drew broader inferences wasincomplete and oversimplified to make his point clearer. Tosuggest that topics such as environmental contributors tobreast cancer or lung cancer risks associated with indoorradon should not have been addressed at all or as persis-tently as they were, or that incomplete data should not havebeen used to promote regulatory action on radon and en-vironmental tobacco smoke, fails to capture the complexityof policy decisions. Even within the purely scientific arena,how does one combine limited direct epidemiologic evi-dence on low-level indoor radon with compelling evidencefrom high exposures among uranium miners and a strongtheoretical and empirical basis for extrapolation to the lev-els encountered in residences? Is the combination suffi-cient for drawing firm conclusions about indoor radoneven with more uncertain evidence directly from the stud-ies of indoor radon?

Determining which topics deserve initial and continuingattention and funding is ultimately a policy decision, a blendof science—considering a rangeof disciplinary perspectives—and political and societal concerns. Early in the evolution ofnew topics with environmental relevance, there is oftena set-aside of research funding (generated by advocates),after which the scientific merit and promise either movesthe topic forward or allows it to atrophy. Evidence on envi-ronmental pollutants as a major contributor to breast cancerhas not taken hold, and any proposed new studies on thattopic need to stand or fall based on their merits, not becausethere continues to be advocacy for pursuing this possibility.Similarly, electromagnetic fields continue to be of limitedinterest to epidemiologists after a period of extensive fund-ing, but those advocating new proposed research must ac-knowledge the findings and shortcomings of the studies thathave come before and make the case that the new studieswill be better. The question of when the funding and atten-tion to topics exceed their objective, public health merit isimportant, but we need to be wary of a stopping rule thatprecludes low-plausibility topics from being investigated atall, particularly when there are forceful advocates for having

a look. Topics not considered in the book have taken othercourses, such as the evolution of information on particulateair pollution at levels many once thought too low to be ofbiologic interest. In hindsight, of course we would do well toinvest exclusively in topics that will improve public health,but those cannot be predicted in advance. The question ofhow much to invest in novel, unexplored areas versus thoseof proven public health importance is complex. Some effortneeds to be devoted to topics other than those representingcompelling global epidemics—tobacco, HIV, and obesity,for example.

I completely concur with Dr. Kabat in his effort to stim-ulate more explicit examination of the issues involved in theevolution of these topics, assessing how the perspectives ofepidemiologists blend and conflict with those of other sci-entists and how the views of scientists are integrated byadvocates who call for funding and use research findingsfor their particular purposes. We should be more aware ofthe epidemiologist’s important but circumscribed role insetting research and regulatory policy. Through such under-standing, epidemiologists could become more effective con-tributors to the broader debates and more accepting whenour personal perspective on the epidemiology, or even theconsensus of epidemiologic thinking, does not carry theday. If the argument of activists overrides that of the epi-demiologists, or the empirical epidemiologic evidence ismore persuasive than the theoretical arguments of physi-cists or toxicologists (or vice versa), or public health acti-vists override the views of epidemiologists, it is notnecessarily an injustice. In the policy arena, where fundingand regulatory decisions are made, it is presumptuous tosuggest that in the face of ambiguity (common to each ofthe issues in the book), the current evidence from epide-miology (or physics, or toxicology, or molecular biology)should always prevail.

Those of us who have chosen to operate in the realm ofresearch value the relative clarity of the terms of battle,drawing on the tools of evidence, new studies, and refine-ment of hypotheses. Science necessarily isolates and ab-stracts from the broader world, and, within the realm ofscience, epidemiology represents a further narrowing ofthe frame of reference. Narrower still is the direct evidenceon a specific question, for example, indoor radon and lungcancer, versus the broader epidemiologic evidence on ra-don and lung cancer generally. By those standards, intru-sions and distractions from scientific principles areharmful. On the other hand, if suggestive evidence regard-ing environmental tobacco smoke and lung cancer is a valu-able tool for antismoking activists and those who generatescientific challenges weaken their efforts, the critics of thatevidence are naturally seen as acting counter to publichealth even if the criticisms are scientifically valid. Theethical issues in slanting or selectively attending to thescience to make wise policy are applicable more broadlythan to epidemiology.

Reading and reflecting on the thesis of this book can onlyhelp epidemiologists be more aware of our place in societyand thus be more effective contributors as we venture be-yond the technical aspects of epidemiology into the broader,messier world. Many of us recognize that regulatory policy

1040 Book Reviews

Am J Epidemiol 2009;169:1039–1042

requires a framework encompassing considerations otherthan epidemiologic evidence, and this book helps to extendthat perspective to include research funding policy. Kabatidentifies real issues that need closer examination and moreopen debate than has take place up to now. In the bipartisanspirit of the era, there is a need to examine the points ofcontention (between advocates and scientists, epidemiolo-gists and physicists, regulators and researchers) to find com-mon ground and enable good science and sound policy tomove forward, if not always hand in hand.

ACKNOWLEDGMENTS


David A. Savitz (e-mail: [email protected])Mount Sinai School of Medicine, Center of Excellence inEpidemiology, Biostatistics, and Disease Prevention,New York, NY 10029-6574

DOI: 10.1093/aje/kwp013; Advance Access publication February 10, 2009

Book Reviews 1041

Am J Epidemiol 2009;169:1039–1042

http://www.oup.com/us/

Concepts of Epidemiology: Integrating the Ideas, Theories, Principles andMethods of Epidemiology, 2nd Edition

By Raj Bhopal

ISBN-10: 0-1995-4314-3, ISBN-13: 978-0-1995-4314-4, Oxford University Press, Inc., New York, New York (Telephone:212-726-6000, Website: http://www.oup.com/us/), 2008, 472 pp., $55.95 Softcover

Several years ago, as the Department of Epidemiology ofthe Johns Hopkins Bloomberg School of Public Health re-vised its core methods courses, the faculty scanned thenavailable texts, seeking to find books that could be used overthe 1-year sequence. We found an unexpectedly large num-ber of introductory texts, including the first edition of RajBhopal’s Concepts of Epidemiology (1); of these, most wereintended for a broad introductory course and for the diversestudent groups that might take such courses. Bhopal hadthoughtfully reviewed 25 such introductory texts in 1997(2), and 2 years later he commented on the role that suchbooks play in defining the conceptual domain of the field(3). His review of these books, along with extensive teach-ing experience, contributed to his writing an introductorytext that differed from many long-used standards.

The first edition of Concepts of Epidemiology, befitting itstitle, emphasized the conceptual basis of epidemiology and itslinks to population health. It was intended primarily for thebroad audience of postgraduate students throughout the worldwho take an introductory course. The second edition (4) buildson the foundation set by the first, with some expansion and theaddition of questions and answers at the end of each chapter.The didactic approach of this edition is a strength. The bookhas numerous embedded exercises that challenge readers toconsider the implications and applications of the concepts thatare discussed. Case studies are abundant, and the questionsand answers that end the chapters should be particularly usefulfor those using the book outside of a course setting. Figureshave been used creatively to explain key concepts.

In its organization, the book follows a structure distinctfrom that of most texts, which generally begin with broadconcepts and the indicators of population health, followedby chapters on various study designs and measures of asso-ciation, and ending with discussions of bias, causal infer-ence, and application. Concepts of Epidemiology, by

contrast, begins after a framing introductory chapter withconsideration of the ‘‘epidemiological concept of popula-tion’’ (4, p. 17) followed by a chapter on disease variation bytime, place, and person that does not introduce either in-cidence and mortality rates or prevalence. These measuresare covered in the seventh chapter. The book’s fourth chap-ter covers bias and effect modification, while the fifth ad-dresses causation; this sequence appears out of order, ascausal association is the target of much epidemiologicresearch and bias is an unwanted source of association.Additionally, these 2 chapters come before study designand measures of association that are reviewed in chapters8 and 9, respectively. The concluding 10th chapter offersa perspective on the future of epidemiology, considers ethicsand practice, and provides historical vignettes.

We also faced the challenge of how best to order the pre-sentation of concepts and methods when we revised the coresequence of epidemiology courses at Johns Hopkins a fewyears ago. There are different ways to structure an introduc-tory course: The course can be anchored on association andcausation and on ‘‘descriptive’’ and ‘‘analytical’’ methods,for example. However, the rationale for the sequencing inConcepts of Epidemiology is not transparent and leads topotential difficulties in using the book for a course. For ex-ample, variation in disease by person, place, and time iscovered without illustration by incidence or mortality rates.Causation and bias are considered prior to the introduction ofmeasures of association. Bhopal’s intent is likely to ensurethat readers give emphasis to concept rather than calculations,but in my view a more integrated approach would be prefer-able. In the Johns Hopkins’ courses, we settled on conceptsand models of causation as the starting point, followed bymeasures of occurrence, risk, and rates per person-time.Measures of association and study designs followed. Onlythen were bias and effect modification covered. The sequence

Book Reviews 1041

Am J Epidemiol 2009;169:1039–1042

http://www.oup.com/us/

for the Hopkins courses is still unsettled as opportunities forimprovement have been identified with each iteration.

Beyond approach and organization, the utility of an intro-ductory text is determined by its content. Bhopal’s extensiveexperience in population health is an asset to the book, and herepeatedly emphasizes the linkages of epidemiology to pop-ulation health and draws in examples from his work. Therange of topics is broad, updated, and appropriate for an in-troductory text. On the other hand, I found laxity in the def-inition of some key concepts andmeasures, perhaps reflectinga tendency to largely offer descriptions, rather than sharpdefinitions or formulae. For example, person-time is not ex-plicitly defined nor sufficiently covered, and a lengthy discus-sion still does not clarify the relation between incidence rateand cumulative incidence. I have found the same deficiency inother introductory texts. Selection bias is vaguely defined asfollows: ‘‘Bias can result from the choice of populations to bestudied. This is known as selection bias’’ (4, p. 90). The sub-sequent discussion emphasizes external validity rather thandistortion of measures of association as a consequence ofselection processes. The Dictionary of Epidemiology (5)would have been a useful resource for ensuring that definitionswere better aligned with the most up-to-date formulations.Chapter 9, covering study designs, introduces the concept ofthe base population but then deviates from offering a unifyingformulation of study design; the various major designs arecovered as though not related, even though the chapter’s titlespoke to ‘‘an integrated suite of methods.’’ These are onlysome of the examples of the book’s content that caught myattention; of course, opinions vary among epidemiologists onthe field’s methods, and others may findBhopal’s presentationto be better aligned with their views.

The bottom line for any review of an introductory text iswhether the book can be recommended for use in an intro-

ductory course. Concepts of Epidemiology is sufficientlydistinct in its approach that anyone considering its useshould read it closely to determine if it will fit the needsof his/her particular course and students. I am certain that,when used by Dr. Bhopal, it works well.

ACKNOWLEDGMENTS


REFERENCES

1. Bhopal R. Concepts of Epidemiology: An IntegratedIntroduction to the Ideas, Theories, Principles and Methods ofEpidemiology. New York, NY: Oxford University Press, Inc,2002.

2. Bhopal R. Which book? A comparative review of 25introductory epidemiology textbooks. J Epidemiol CommunityHealth. 1997;51(6):612–622.

3. Bhopal R. Paradigms in epidemiology textbooks: in thefootsteps of Thomas Kuhn. Am J Public Health. 1999;89(8):1162–1165.

4. Bhopal R. Concepts of Epidemiology: Integrating the Ideas,Theories, Principles and Methods of Epidemiology. SecondEdition. New York, NY: Oxford University Press, Inc, 2008.

5. Last JM, ed. A Dictionary of Epidemiology. 4th ed. New York,NY: Oxford University Press, Inc, 2001.

Jonathan M. Samet (e-mail: [email protected])Department of Preventive Medicine and Institute for GlobalHealth, Keck School of Medicine, University of SouthernCalifornia, Los Angeles, CA 90089

DOI: 10.1093/aje/kwp009; Advance Access publication February 10, 2009

1042 Book Reviews

Am J Epidemiol 2009;169:1039–1042

American Journal of Epidemiology Volume169 Number8 April15 2009

Documents