Exercises in Epidemiology Applying Principles and Methods

EXERCISES IN EPIDEMIOLOGY
This page intentionally left blank

EXERCISES IN
EPIDEMIOLOGY
Applying Principles
and Methods
NOEL S. WEISS
1
1
Oxford University Press, Inc., publishes works that further
Oxford Universitys objective of excellence
in research, scholarship, and education.
Oxford New York

Auckland Cape Town Dar es Salaam Hong Kong Karachi
Kuala Lumpur Madrid Melbourne Mexico City Nairobi
New Delhi Shanghai Taipei Toronto
With offices in
Argentina Austria Brazil Chile Czech Republic France Greece
Guatemala Hungary Italy Japan Poland Portugal Singapore
South Korea Switzerland Thailand Turkey Ukraine Vietnam
Copyright 2012 by Oxford University Press, Inc.
Published by Oxford University Press, Inc.

198 Madison Avenue, New York, New York 10016
www.oup.com
Oxford is a registered trademark of Oxford University Press
All rights reserved. No part of this publication may be reproduced,

stored in a retrieval system, or transmitted, in any form or by any means,
electronic, mechanical, photocopying, recording, or otherwise,
without the prior permission of Oxford University Press.
____________________________________________
Library of Congress Cataloging-in-Publication Data
Weiss, Noel S., 1941-

Exercises in epidemiology : applying principles and
methods / Noel S. Weiss.
p. ; cm.
Includes bibliographical references and index.
ISBN 978-0-19-979678-6
1. EpidemiologyProblems, exercises, etc. I. Title.
[DNLM: 1. Epidemiologic MethodsProblems and Exercises. WA 18.2]
RA652.7.W45 2012
614.4dc23
2011017403
____________________________________________
9 8 7 6 5 4 3 2 1
Printed in the United States of America
on acid-free paper
CONTENTS
Introduction vii
1. Rates and Proportions 3

2. Causal Inference 43
3. Confounding 69
4. Cohort Studies 119
5. Case-Control Studies 147
6. Multiple Causal Pathways and
Effect Modification 175
7. Screening 207
References 237
Index 243
INTRODUCTION
There are a lot of texts that deal with the principles and meth-
ods of epidemiology. Ive been a coauthor of one of these myself.
All of the texts, to a greater or lesser extent, provide examples
of real or hypothetical epidemiologic studies to illustrate a given
principle or method. For many (probably most) readers of these
books, the examples help to solidify an understanding of the
topic at hand.
What the examples do not provide is the opportunity to
consider, on ones own, how a particular issue ought to be dealt
with, or how a particular question should be addressed. The
purpose of this book is to supplement the material contained in
the textbooks in such a way that the reader is forced to: (1) iden-
tify situations in which the validity or accuracy of a particular
design or analytic approach may be limited; and (2) determine
how that limitation might be overcome. Such actions are just
those that epidemiologists have to take when they are planning
research or are reviewing that of others.
The key word in the preceding paragraph is supplement. The
present book cannot stand alone as a means of learning about
epidemiology, or even as a means of being introduced to the
v i i i | INTRODUCTION
subject. My hope is that the exercises contained in it can extend

the knowledge of students of epidemiology, and equip them
more fully to deal with the real world problems and issues that
theyll encounter in their professional lives.
The book is organized into seven chapters, each of which
contains a set of questions and answers to those questions. Any
reader who believes a given answer is incomplete (or wrong!)
is welcome to communicate with me ([email protected]). With
the readers permission, his/her suggestions will be posted on
the discussion page of that section of the Oxford University
Press website devoted to this bookperhaps with an additional
comment from me. No doubt all readers of this portion of the
website will benefit from learning of different viewpoints.
To minimize the likelihood of an ambiguous question being
present in this book, or an incomplete or incorrect answer,
I enlisted the help of the following persons to review parts of
the draft manuscript: Peter Cummings, Paul Doria-Rose, Sarah
Lowry, Amanda Phipps, Gaia Pocobelli, and Ali Rowhani-
Rahbar. The contributions of each of them helped to make the
chapters of the book that you are reading better than the draft
chapters that they received from me.
EXERCISES IN EPIDEMIOLOGY
CHAPTER 1
Rates and Proportions
EPIDEMIOLOGISTS LOVE denominators. Sometimes we divide

the number of numerator events among exposed individuals by
the total number of exposed individuals, so that we can calcu-
late the proportion of (say) eaters of potato salad at a picnic who
were diagnosed with a Salmonella infection during the ensuing
48 hours. At other times, we use a person-time denominator
enabling us to calculate the rate of (say) lung cancer in persons
who have been employed in a given industry. Depending on the
question being addressed, we may seek to estimate a propor-
tion or a rate. The accuracy of that proportion or that rate will
depend on our ability to measure correctly both numerator and
denominator.
RATES AND PROPORTIONS | 5
Question 1.1 A recent study observed that 1 in 20 persons

with cancer later were diagnosed with a second cancer. In the
general population, the lifetime probability of being diagnosed
with cancer is considerably greater. Is this evidence of immu-
nity developed as a result of the first cancer?
6 | EXERCISES IN EPIDEMIOLOGY
Answer 1.1 This is not necessarily evidence of immunity.

Whats not being taken into account is the very different denom-
inator for each of the two groupsthe amount of person-time
at risk. Among cancer patients, person-time begins to accrue
as of the date of diagnosis, typically in mid- to late life. In the
general population, person-time begins to accrue at birth.
Question 1.2 A study on ovarian cancer observed the follow-

ing pattern of histologic type and race among its participants.
Type of tumor
Race Mucinous Other Total
Caucasian 33 (13%) 225 (87%) 258 (100%)

Asian 55 (27%) 151 (73%) 206 (100%)
The authors concluded that Asian women had a higher

incidence of mucinous tumors than did Caucasian women.
What reservations do you have regarding this conclusion?
8 | EXERCISES IN EPIDEMIOLOGY
Answer 1.2 The observed proportional distribution of histo-

logic type by race could be due to a relatively high incidence rate
of mucinous tumors in Asian women, or as well to Asian women
having a low rate of other ovarian tumors. For example, the
rates below would give rise to the numbers presented in this
question:
Type of tumor (rate per 100,000 woman-years)

Race Mucinous Other Total
Caucasian 1 6.7 7.7

Asian 1 2.7 3.7
Question 1.3 The following is excerpted from a news item in

the British Medical Journal:
The clinical features of more than 1000 patients with lung

cancer presenting to 46 UK hospitals have been analyzed. The
results showed that women under 65 are particularly at risk of
small cell lung cancer34% presented with this form of the
disease compared to 18% of men.
Assume that: (1) the distribution of histologic types of lung

cancer in the patients under 65 years in the 46 UK hospitals
accurately reflects that of all U.K. lung cancer patients; and
(2) the difference between the figure of 34% in women and
18% in men is not due to chance. Under what circumstance
could the observed difference not be indicative of a difference in
the incidence of small cell lung cancer between U.K. men and
women under 65 years?
1 0 | EXERCISES IN EPIDEMIOLOGY
Answer 1.3 The proportional incidence by gender will not be

an indication of the absolute incidence if the incidence of non
small cell lung cancer is different in men and women.
For example:
Men Women
Small cell cancers (%) 360 (18%) 340 (34%)

Other types 1640 660
Total 2000 1000
In the above example, assuming the numbers of men and

women in the population are similar to one another, the rate of
small cell lung cancer by gender is nearly identical. The disparity
in the proportional incidence comes from the disparity in the
rates of lung cancer that are not of the small cell type.
Question 1.4 The following statement appeared in a review

article:
In 1996 in the United States, a total of about 34,000 new cases

of endometrial cancer occurred, as well as approximately 6,000
deaths from this disease. The case-fatality is approximately 28%.
a. Assuming that the data described in the first sentence

are correct, why is it unlikely that the case-fatality from
endometrial cancer is truly 28%?
b. Describe a circumstance under which the data in the
first sentence and a case-fatality of 28% for endometrial
cancer in U.S. women could both be true.
Answer 1.4
a. If equilibrium existsin other words, no change in

number of cases or case fatality over timethe case
fatality among women with endometrial cancer should
be 6,000/34,000 = 18%.
b. The 6,000 deaths in 1996 occurred primarily in women
diagnosed with endometrial cancer prior to that year.
So, if the incidence of endometrial cancer had very
recently increased to a large degree, the appropriate
denominator for the calculation of case-fatality would
be a number much smaller than 34,000 (specifically,
21,429 cases in order to generate a case-fatality of 28%).
Question 1.5 Lets say youve conducted a cohort study to

determine some long-term consequences of surgical treatment
of patients with cataracts. For 174 patients who underwent
surgery and 103 other patients with cataracts who did not,
youve used records of the state department of motor vehicles
to determine who has been involved in a motor vehicle crash as
a driver. From periodic interviews with study subjects, you are
able to estimate the number of miles each one has driven during
a 2-year follow-up period.
The results of the study are as follows:
Study group No. of persons No. of crashes No. of miles driven
Surgery 174 27 5,677,867

No surgery 103 23 2,569,639
Assume the two groups of patients are exactly comparable

with respect to baseline characteristics that predict automobile
crash occurrence, including driving behavior, and that no mis-
classification is present in the study.
a. Estimate the influence of cataract surgery on crash rate

while driving.
b. Estimate the overall influence of cataract surgery on the
risk of an automobile crash, in other words, that which
would include a possible influence of the surgery on driv-
ing behavior.
Provide the rationale for your answers.

Answer 1.5
a. The answer is the relative rate based on the number of

miles driven
27 5,677,867
= 0.53 since it allows for the
23 2,569,639 number of driver-miles at risk.
b. Since the number of miles driven seems to have been

influenced by the receipt of surgery, the assessment of
the aggregate impact should not consider this, and the
27 174
relative risk of 0.69 should be used.
23 103
Question 1.6 The following is paraphrased from an article in

the British Medical Journal1:
Although the relative rate of myocardial infarction associated

with cigarette smoking is higher in women than in men, smok-
ing may well cause a higher rate of myocardial infraction in
men who smoke than in women who smoke.
Under what circumstance could this be true?

Answer 1.6 It could be true if, among nonsmokers, the inci-

dence of myocardial infarction (MI) in men were higher than
that in women. For example, assume that in a certain age group
the annual incidence of MI was 3 per 1,000 in men and 1 per
1,000 in women. Among men, a relative rate of 2 associated with
smoking would produce a rate difference of (2*3/1,000)
3/1,000 = 3/1,000 person-years. Among women, a higher rela-
tive rate3would produce a rate difference that is smaller than
this: (3*1/1,000)1/1,000 = 2/1,000 person-years.
Question 1.7 The following is excerpted from a letter to the

editor of a medical journal:
We have observed renal cell carcinomas in 6 out of 412 patients

with analgesic nephropathy (1.4%), treated over the past
12 years. The incidence of renal cell carcinoma in the general
population is 7.5 per 100,000 population per year, so the prev-
alence found in patients with analgesic nephropathy is highly
significant ( p <.005).
What additional information on these patients with analge-

sic nephropathy would be needed in order to better assess the
possibility that they are at increased risk of renal cell carcinoma?
Answer 1.7 At the very least, we would need the age-specific

person-time at risk for the diagnosis of cancer among partici-
pants with analgesic nephropathy. This would permit a compar-
ison of age-adjusted rates of renal cell carcinoma between these
patients and the general population.
Question 1.8 The Second National Health and Nutrition

Examination Survey was a cross-sectional survey conducted
from February 1976 to February 1980, with a probability sample
of 27,801 persons in the United States.
The following table presents some data obtained in the
survey:
Percentage of children 6 months through 4 years with a

history of eating unusual substances by selected
characteristics: United States, 19761980
Percentage with
history of eating
No. examined unusual substances
Blood lead level in
micrograms per
deciliter:
30 or more 117 16.2
2029 503 14.1
Less than 20 1,752 5.2
Earlier studies have shown that elevated blood lead levels

(30 g/dl or higher) are associated with slowed intellectual
development in children. At issue in the present analysis is
whether eating unusual substances (e.g., paint) contributes to
elevated blood lead levels.
In U.S. children 6 months through 4 years of age, can you
determine the likelihood of having a blood lead level of >30 g/dl
for those with a history of eating unusual substances relative to
the likelihood for those with no such history? If yes, what is it?
If no, why not?
Answer 1.8
Blood lead level

(microg/dl)
History of eating
unusual substances 30 <30 All
Yes 19 162 181

No 98 2,093 2,191
2,372
19 181
Relative risk = = 2.3
98 2,191
Question 1.9 The following data were obtained in a very

large cohort study conducted in Korea during 19932002 that
examined potential risk factors (including the prevalence of
hepatitis B surface antigen positivity (HbsAg+)) for mortality
from hepatocellular carcinoma (HCC).
Rate per 100,000

No. of HCC deaths person-years
Men
HbsAg+ 1522 405.2
HbsAg 734 21.8
Women
HbsAg+ 37 58.4
HbsAg 9 1.2
a. For men and women, separately, estimate the relative

mortality from HCC associated with being HbsAg+, and
also the mortality difference.
b. One of the above measures of excess mortality is greater
in men; the other greater in women. How can this be?
Answer 1.9
a. Relative mortality
Men: 405.2 21.8 = 18.6

Women: 58.4 1.2 = 48.7
Mortality difference (per 100,000 person-years)
Men: 405.2 21.8 = 383.4

Women: 58.4 1.2 = 57.2.
b. The annual mortality from HCC, in the absence of

active infection with hepatitis B, differs greatly by sex:
21.8 per 105 for men versus 1.2 per 105 for women.
Thus, an absolute increase in mortality of 57.2 per 105
experienced by Korean women is very large in relative
terms (relative mortality = 48.7). In men, the larger
absolute mortality difference (383.4 per 105) is not
nearly so large on a ratio scale, since it is superimposed
not on a baseline mortality rate of 1.2 per 105, but on
the higher male baseline rate of 21.8 per 105.
Question 1.10 Black men in the United States have a substan-

tially higher incidence of prostate cancer than U.S. white men.
Lets say theres a variant of the androgen receptor gene thats
more common in black than white men in the United States
50% versus 30%that is also associated with a doubling of
incidence of prostate cancer in American men of either race.
What would be the relative incidence of prostate cancer,
black versus white American men, if the genetic marker were
the sole risk factor for this disease that differed between the
two races?
Answer 1.10 If x = incidence of prostate cancer in men without

the variant genotype, the incidence of prostate cancer in white
men would be a weighted average of the incidence in the 70%
of men without the variant genotype and the 30% who have it:
7x + .3(2x) = 1.3x. The incidence in black men would be .5x +
.5(2x) = 1.5x, because half have the variant genotype and half do
not. If, in terms of prostate cancer risk, white and black men
were identical save for the prevalence of this genotype, black
men would have an incidence that was 1.5x/1.3x = 1.15 times
that of white men.
Question 1.11 You read a magazine article in which a medical

columnist has expressed concern that the mean age at which
colorectal cancer is diagnosed among Americans who smoke
cigarettes and consume alcohol is lower than among their fellow
citizens who neither smoke nor drink. Assume that the age dis-
tribution is the same between Americans who smoke and drink
and those who do not. Must it be true that, among relatively
young American adults, the incidence of colorectal cancer is
higher in cigarette smokers/alcohol drinkers than in other
persons? If yes, why? If not, why not?
Answer 1.11 No. For example, if among older persons the

incidence of colorectal cancer were relatively low in those who
smoked and consumed alcohol, with the incidence among
younger persons who smoked and consumed alcohol being the
same as that of young abstainers, the mean age of diagnosis of
smokers/drinkers also would be lower than that of abstainers.
Question 1.12 The following statement was made in a news-

paper article that sought to provide data bearing on the efficacy
of seat belts in preventing deaths that occur in automobile
crashes:
Of the 649 people who died in traffic accidents in Washington

last year, 55 percent were not wearing seat belts. In those same
fatal crashes, 73 percent of people who were belted in survived
without serious injury.
Does this statement support the hypothesis that seat belt

use saves lives? Explain.
Answer 1.12
Dead Alive
Belt 45% ?
No belt 55% ?
Total 100% 100%
The data do not bear on the hypothesis. What is needed

instead is information on the percentage of persons who sur-
vived these crashes who were unbelted. There would be evidence
of efficacy to the extent that this figure was smaller than 55%.
Alternatively, one could compare the percentage of unbelted
persons who survived without serious injury to the figure of
73% for belted individuals:
Dead Alive Total
Belt 27% 73% 100%

No belt ? ? 100%
Question 1.13 During a recent decade in the United States,

the annual proportion of all women ages 25 to 29 years who gave
birth to a first child rose from .031 to .039. In this same decade,
however, the annual proportion of childless women ages 25 to
29 years who gave birth to a child fell from .114 to .092. How is
it possible that these two trends can be in opposite directions?
Answer 1.13 The numerator for the two proportions is the

same, in other words, the annual number of 25- to 29-year-old
American women who gave birth to their first child. But the
denominator for the second proportionthe number of 25- to
29-year-old childless womenis but a part of the first denomi-
nator (all 25- to 29-year-old women). In order for the incidence
of first births to have risen overall but to have declined among
childless women, it must be true that the fraction of 25- to
29-year-old women who were childless must have risen during
the decade. This more than compensated for the declining first-
birth incidence in childless women and caused a rise in the first-
birth incidence in 25- to 29-year-olds as a whole.
Question 1.14 The rate of suicide among American physi-

cians, relative to the corresponding rate in the population as a
whole, varies by gender. Among men, the rate in physicians is
1.5 times higher, whereas among women the corresponding
relative rate is 3.0. It turns out that the rate of suicide in
American male and female physicians is identical. For American
men and women in general, what is the relative rate of suicide
in men compared to women?
Answer 1.14
Pm = rate of suicide in male physicians,

Pf = rate of suicide in female physicians.
M = rate of suicide in American men
W = rate of suicide in American women
Pm
RR, men = = 1.5
M
Pm
M=
1.5
So, for American men in general, their rate of suicide is that
of the male physicians divided by 1.5.
Pf
RR, women = = 3.0
W
Pf
W=
3.0
Similarly, American women have but one-third the rate of

suicide of female physicians.
Now, because Pm and Pf are the same (well label this
rate as P),
M P 1.5 3.0
= = = 2.
W P 3.0 1.5
American men, as a whole, have twice the rate of suicide as

American women.
Question 1.15 A study of suicide among men with cancer

was conducted in the United States.2 The goal of the study was
to enable health professionals to be aware of the potential for
suicide in cancer patients. Some of the site-specific data are
presented below.
Suicides
Type of No. of men per 100 men
cancer with cancer No. of suicides (95% CI)
Lung 102,940 215 0.21

(0.180.24)
Melanoma 19,377 46 0.24
(0.170.32)
Thyroid 5,339 14 0.26
(0.140.42)
It had been hypothesized that the risk of suicide during any

given period of time following diagnosis would be greatest for
types of cancer with a poor prognosis (e.g., lung) than types with
a good prognosis (e.g., melanoma, thyroid). Do the above data
argue against this hypothesis? (Assume that the distribution of
demographic characteristics bearing on suicide occurrence is
similar across the three types of cancer.)
Answer 1.15 The data do not argue against the hypothesis.

The analysis fails to consider person-time at risk. Because this
is, on average, considerably greater for a man with melanoma or
thyroid cancer than for a man with lung cancer, the rate of sui-
cide (i.e., number of suicides divided by person-time at risk) in
the latter group must be higher than the rate for the other two
groups.
Question 1.16 In a study of oral cancer, you observe that 17%

of the Hispanic cases are younger than 40 years, as compared
to 4.8% of non-Hispanic men with oral cancer (p <.05).
Assume that the ascertainment of cases of oral cancer was
equally complete in the Hispanic and non-Hispanic men, and
that the above difference was not due to chance. Does this find-
ing necessarily imply that in the population under study the
risk of developing oral cancer is elevated in Hispanic men under
40 years of age compared to non-Hispanic men of similar age?
Explain.
Answer 1.16 No. The age disparity could simply be a reflection

of the relatively younger age of Hispanic males in the population
under study.
For example:
Hispanic men Other men

Incidence Incidence
Age No. of Person- per No. of Person- per
(years) cases years 100,000 cases years 100,000
<40 17 100,000 17 17 100,000 17

40 83 100,000 83 337 406,000 83
% under 17% 4.8%
40
Or, beyond this, a high proportional incidence of oral cancer

in younger Hispanic men could be due to an atypically low abso-
lute incidence in older Hispanic men.
For example:
Hispanic men Other men

Incidence Incidence
Age No. of Person- per No. of Person- per
(years) cases years 100,000 cases years 100,000
<40 17 100,000 17 17 100,000 17
40 83 100,000 83 337 100,000 337
% under 17% 4.8%
40
Question 1.17 A population-based case-control study of

Guillain-Barr syndrome (a neurological disease) conducted in
19921994 in 4 states estimated the risk of this disease to be
1.7 times greater among adults who had received influenza vac-
cine in the prior 6 weeks than those who had not. The investiga-
tors also estimated that the added risk of Guillain-Barr
syndrome associated with the receipt of influenza vaccine was
about one per million persons during the first 6 weeks after
vaccination.
From these data, can you calculate the 6-week incidence
among adults in the 4-state population who did not receive the
vaccine? If yes, what is that incidence? If no, why not?
Answer 1.17 The difference of 1 per million in the 6-week

incidence between persons who did (Ie) and did not (Io) receive
the vaccine is Ie Io = 1.7Io Io. Therefore, Io = 1/0.7 = 1.43 per
million.
Question 1.18 The incidence of stomach cancer in country X

is 8.0 per 100,000 per year. The incidence rate in nearby coun-
try Y, with a similar age-sex-race composition as country X, is
10.0. You are concerned with explaining this difference. You
know that 5% of people in country Y drink tea containing sus-
pected carcinogen A, whereas nobody in country X drinks this
tea. In order for this to be the sole explanation of the difference
in the incidence rates of stomach cancer between the two coun-
tries, how strongly must carcinogen-A-tea drinking be associ-
ated with stomach cancer?
Answer 1.18 If all the difference were due to ingestion of car-

cinogen A in tea, the incidence of stomach cancer in country Y
could be described as follows:
10 = .95 ( 8 ) + .05t
where t = incidence in drinkers of A-containing tea

10 .95 ( 8 )
t= = 48
.05
48 8 = 6
Therefore, a relative risk of 6 (associated with drinking

A-containing tea) is required for the whole of the difference
in rates between the two countries to be attributable to this
exposure.
Question 1.19 During 19932001, men at 10 U.S. study cen-

ters were invited at random to receive annual PSA screening for
6 years (plus annual digital rectal exams for 4 years; n = 38,343)
or no intervention (n = 38,350).3 Through 10 years from the
time of randomization, there were 3,452 cases of prostate
cancer diagnosed and 83 deaths from this disease in the group
invited for screening, versus 2,974 cases and 75 deaths from
prostate cancer among men in the control arm. The investiga-
tors noted that, among men diagnosed with prostate cancer
in the screening and control arms of the trial, there were 312
and 225 deaths from other causes, respectively, a difference
of 87 deaths. They went on to speculate that his latter differ-
ence was possibly associated with over-diagnosis of prostate
cancer.
What would be a better approach to quantifying the likeli-
hood of death from causes other than prostate cancer between
men who were invited and those who were not invited to be
screened?
Answer 1.19 Among men diagnosed with prostate cancer,

ideally the rate of death from causes other than prostate cancer,
not simply the number of such deaths, would be compared: that
is, the number of deaths divided by the number of person-years.
Even though the sizes of the groups invited and not invited to
be screened were nearly identical, the number of men diagnosed
with prostate cancer in the former group was larger (by 16%,
3452 vs. 2974), because screening identified a number of malig-
nancies that otherwise would not have been diagnosed during
the follow-up period. Failure to take into account the relatively
larger number of person-years in the invited men diagnosed
with prostate cancer would lead to a falsely high mortality rate
in that group.
To the extent that the age distribution of men with screen-
detected prostate cancer differed from that of men whose cancer
was diagnosed for other reasons, a valid comparison would
require age adjustment as well (given the strong association
between age and mortality rates).
CHAPTER 2
Causal Inference
EPIDEMIOLOGISTS PRIMARY responsibility to society is to

provide data relevant to the prevention of disease and injury.
We do this by examining the association between disease or
injury and potential etiologic factors. When across all relevant
epidemiologic studies there does appear to be an association, it
is necessary to infer whether that represents cause and effect
preventive measures would be appropriate only when a judg-
ment of cause and effect can be made. To a great extent, the
process of distinguishing causal and noncausal associations is
subjective: persons may disagree on the relative importance of
the various elements that go into such an inference, and on the
degree to which those elements are present. However, subjec-
tive or not, we cannot escape the need to try to draw causal
inferencesan efficient program of prevention depends on our
success in correctly distinguishing associations that are causal
from those that are not.
CAUSAL INFERENCE | 45
Question 2.1 The following is excerpted from an abstract of

an article on sudden infant death syndrome4:
Objective: To assess the role of parental bedsharing in sudden

infant death syndrome (SIDS)-like deaths, this study
examines the hypothesis that, compared with other SIDS
cases, the age distribution of deaths associated with bed-
sharing should be lower.
Methods: For 84 SIDS cases in Cleveland, Ohio, 1992 to
1996, age at death, maternal weight, and other risk
factors for SIDS were compared for cases grouped accord-
ing to bedsharing status.
Results: Mean ages at death were 9.1 weeks for 30 bedshar-
ing and 12.7 for 54 nonbedsharing cases.
Conclusion: By demonstrating that among an urban popula-
tion at high risk for SIDS, bedsharing is strongly associ-
ated with a younger age at death, independent of any
other factors, this study provides evidence of a relation-
ship between some SIDS-like deaths and parent-infant
bedsharing.
In this study, assume that information on bedsharing was

completely accurate. What is your primary reservation regard-
ing the authors assertion that their study provides evidence of
a relationship between some SIDS-like deaths and parental
bedsharing?
Answer 2.1 Parental bedsharing with an infant may simply be

an age-related phenomenon: the younger the infant, the more
likely the parents are to share a bed with him/her. If so, then
bedsharing may have no etiologic relevance to SIDS at any age.
Question 2.2 Quoted from a magazine of an insurance agency:
Does your automobile have body damage? Fix it, and youll sig-
nificantly reduce your probability of involvement in another
traffic mishap. There is a distinct psychological advantage to
having even minor auto collision damage repaired as soon as
possible. Studies have shown that drivers of newly repaired
automobiles tend to drive more defensively than those with
unrepaired damage.
Assume that the studies referred to are cohort (follow-up)

in type and that they have shown unequivocal differences in the
rate of second accidents between persons who did and did not
repair the damage resulting from an initial accident. What is
your main reservation concerning the conclusion that repairing
auto damage influences driving behavior?
Answer 2.2 The association may be due to a third (confound-

ing) factorpersons who choose to have damaged automobiles
repaired may drive more defensively than persons who choose
not to do so, no matter what the condition of their cars.
Question 2.3 For some time, women taking SSRIs (a class of

antidepressant drugs) during pregnancy have been encouraged
to stop this medication 14 days prior to the end of the preg-
nancy, due to concerns about a possible deleterious impact on
labor and delivery. A study conducted to examine this issue
observed an increased risk of prematurity (<37 weeks gesta-
tion) in infants born to mothers who had taken SSRIs and had
not stopped taking them 14 or more days prior to delivery
(women not taking SSRIs during their pregnancy served as the
basis for comparison). No corresponding association was seen
for discontinuation of SSRIs at least 14 days before delivery.
Even though you have no reason to be concerned with mis-
classification of exposure or outcome status in this study, or
with a difference in risk factors for prematurity at the start of
pregnancy between the three categories of women (nonusers of
SSRIs, users within the last two weeks of pregnancy, users only
prior to the last two weeks of pregnancy), you are concerned
that the study does not provide a valid result. What is the basis
for your concern?
Answer 2.3 These data may be indicative of a causal relation,

but one in the opposite direction. That is, premature delivery in
a woman using an SSRI during pregnancy is likely to prevent
her from stopping the drug, rather than the reverse.
Question 2.4 The following question pertains to a news item

that appeared in a medical journal:
People screened for lung cancer by spiral CT have accelerated

and prolonged quit rates of smoking, regardless of whether the
screening shows disease. Researchers found that 1 year after
scanning, 14% of smokers had stopped smoking; by contrast,
the rate among the general smoking population was 57%. The
findings suggest that screening is an ideal place to provide
cessation messages, say the researchers.
Assume that the difference between the figures of 14%

and 57% is not due to chance and is not due to differences in
demographic characteristics between persons who do and do
not receive spiral CT screening. What is an explanation for this
difference apart from a genuine impact of attending the screen-
ing program on the likelihood of smoking cessation?
Answer 2.4 Due to the likely presence of a great deal of con-

founding, the data obtained in this study do not have the ability
to assess the efficacy of cessation messages to persons who
smoke cigarettes. It could be that those smokers who are suffi-
ciently health conscious to receive screening for lung cancer are
overrepresented with persons who, apart from any additional
health education or advice, are likely to quit during the coming
year.
Question 2.5 The following quotation comes from a review of

research on diet in relation to levels of serum lipids:
Many (studies) were limited by having a small number of

participants. Diet studies quoted to this day as authoritative
had as few as five subjects. If any enduring truth has emerged
about human beings and diet, it is that everyone is remarkably
different, and studies that dont involve dozens if not hun-
dreds of participants are of extremely limited value.
How is it possible for an epidemiologic study (of any

question) that involves fewer than dozens or hundreds of
participants to be of more than limited value?
Answer 2.5 I had a professor of neuroanatomy who groused

one day in class about a manuscript of his that was not accepted
for publication, apparently because the results contained in it
were based on but two dogs. He said that if he had just one dog,
but could teach it to play the violin, there ought to be no
concern with the sample size! If a study has identified a strong
association (and otherwise has been designed in a way to pro-
vide a valid result), it need not be large to document this in
a convincing way. For example, it took just 8 cases of vaginal
adenocarcinoma and 32 controls to identify an unequivocal
association with maternal use of diethylstilbestrol (7 cases and
0 controls had been exposed in utero).
Of course, when associations are not so strongperhaps in
the case of diet in relation to serum lipidsstudies of dozens if
not hundreds (if not more) participants are needed to docu-
ment their presence and/or size.
Question 2.6 A study investigated the relationship between

a womans age at first birth and the presence of depression.
The study used data from the U.S. National Maternal and Infant
Health Survey, a National Center for Health Statistics-sponsored
nationally representative survey that selected a stratified sys-
tematic sample of 1988 live births from state vital statistics
records via a multistage cluster design. Respondents were inter-
viewed in person an average of 17 months after delivery, and
the data collected included (among other things) age, race, and
the respondents extent of depressive symptoms in the 4 weeks
prior to the interview. On the scale that was used, a score of 16
or higher (out of a possible 60) was the criterion for categoriz-
ing a woman as depressed. This analysis was limited to women
who had just delivered their first child.
The following table was presented:
Age, years % depressed Odds ratio (95% CI)*

African-American women
1517 48.1 2.7 (1.9, 3.9)

1819 36.8 1.7 (1.2, 2.5)
2534 25.3 1.0 (reference)
White women
1517 27.8 2.4 (1.4, 4.1)
1819 32.9 3.0 (2.0, 4.7)
2534 13.8 1.0 (reference)
*Adjusted for all relevant confounding variables.
Despite the strong association present in this study, you are

reluctant to conclude that giving birth to a child as a teenager is
a cause of depression. Why is this?
Answer 2.6
a. The prevalence of depression could be higher in

teenagers than in 25- to 34-year-old women even in
those without a recent first birth. Thus, the observed
association could be related to age per se and not to
having had a child at a given age.
and/or
b. Antecedent depression could be more common in

women who as teenagers attempt to conceive for the
first time (or who fail to prevent conception) than in
women who do so at ages 25 to 34.
Question 2.7 In a study conducted among members of a

prepaid health-care plan, an investigator observed that infants
of women who had obtained analgesic X from the plans phar-
macy sometime in the year prior to delivery had twice the risk
of certain major congenital anomalies compared to the infants
of women who did not obtain that drug. Additionally, pregnan-
cies in women who had obtained analgesic X ended in sponta-
neous abortion 1.8 times more commonly than did pregnancies
in women who had not.
The studys findings of increased risk were subsequently
criticized because of questions about the accuracy with which
infants were classified as exposed or unexposed. Because anal-
gesic X is relatively inexpensive and widely available without
prescription, one critic argued that it was likely that some mem-
bers purchased it outside the health plan. Another critic noted
that there was no evidence that the analgesics obtained were
actually used during pregnancy. In your judgment, could these
factors account for the studys findings, in the absence of a true
association? Explain.
Answer 2.7 No. The information bias resulting from these

factors would be expected to be nondifferential with regard
to the occurrence of a congenital anomaly or spontaneous
abortion. This would result in the observed association being
spuriously close to the null, rather than away from it.
Question 2.8 A study was conducted among women with

either of two types of breast cancer: (1) negative for both estro-
gen receptor (ER) and progesterone receptor (PR); and (2) all
other types combined.5 The two groups were compared with
regard to the use of a statin medication prior to diagnosis. The
results were as follows:
Type of breast cancer

Stat in use ER/PR All other
Yes 34 (11%) 269 (15%)

No 340 1498
374 1767
a. The difference in prior use of a statin between the two

groups of cases was unlikely to be the result of chance
( p = .02). The study investigators concluded that statin
use could well be related to a reduced incidence of ER/
PR breast cancer. Assuming that no confounding is
present, what is an alternative interpretation of these
results?
b. One analysis focused on use of lipophilic statins in rela-
tion to type of breast cancer, because of the investigators
concern that prior studies had not accounted for the
confounding effect of combining lipophilic and hydro-
philic statins when examining this question. Assuming
that the large majority of women use just one type of
statin (lipophilic or hydrophilic), it is likely that the inves-
tigators concern was not actually with confounding in
those earlier studies, but with another potential source of
bias. Which one? Explain?
Answer 2.8
a. It is possible that statin use increases the incidence of

breast cancer that is not ER/PR.
b. The investigators concern appears to be with exposure
heterogeneity, specifically that an association between
the proportional incidence of ER/PR breast cancer and
lipophilic statin use could be obscured by having included
users of hydrophilic statins in the exposed category.
Question 2.9 The following data were obtained in a study to

evaluate venous thromboembolism risk among women after
air travel.
Oral # of # of Odds
Air travel contraceptives cases controls ratio 95% CI
No No 54 94 1.0 Ref.
No Yes 95 48 3.5 2.15.8
Yes No 4 5 1.4 0.36.8*
Yes Yes 20 2 17.4 3.9157.0*
*Exact confidence limits.
The authors noted that they found a marked increase in

risk of venous thromboembolism among women who used oral
contraceptives and also had recent exposure to air travel, but
argued for a cautious interpretation of this finding because of
the small number of controls who reported recent air travel and
oral contraceptive use. Do you agree that the small number of
controls with both exposures argues for a cautious interpreta-
tion? Explain.
Answer 2.9 No. Because the lower limit of the confidence

interval greatly exceeds 1, chance is an unlikely explanation for
the association, notwithstanding the small number of controls
with both oral contraceptive use and recent air travel.
Question 2.10 In the United States in 1998, all enriched

grains and cereals were required to contain 140 micrograms of
folic acid per 100g of grain. Median levels of blood folate among
American women of childbearing age rose from 4.8 ng/ml in
1994 to 13.0 ng/ml in 2000.
One of the reasons for the folic acid fortification program
was the results of studies (both randomized and nonrandom-
ized) conducted prior to 1998 that documented a large increase
in risk of neural tube defects in the offspring of women with
low folic acid intake in the year before becoming pregnant
and/or blood levels of folic acid at the start of pregnancy.
However, a large, multisite, case-control study of neural tube
defects in U.S. children born during 19982003 failed to iden-
tify a difference between mothers of cases and controls with
regard to folic acid intake in the year preceding the pregnancy.6
Assume that the 19982003 study provided a valid result,
and that the confidence limits around the risk estimates were
so narrow as to exclude the possibility of a true case-control dif-
ference of any importance. Why do you believe this null result
does not detract from the hypothesis that maternal folic acid
deficiency is a cause of neural defects in her offspring?
Answer 2.10 It is likely that above a certain threshold of folic

acid intakeone effectively reached by all American women
after 1997there is no further reduction in risk of neural tube
defects in relation to maternal folic acid intake or blood levels.
The inverse association between maternal intake (or levels) and
neural tube defects already has been well demonstrated in
women below this threshold; the 19982003 study simply is
unable to address this question in the setting of population-
wide high folic acid intake. The cases in the 19982003 study
no doubt arose via a causal pathway not involving folic acid.
Question 2.11 In an effort to gauge the impact of surgery and

radiation therapy on mortality among men with organ-confined
prostate cancer that was well or moderately differentiated,
a large cohort study was conducted.7 Beginning one year after
diagnosis, all-cause mortality in 65- to 80-year-old American
men with this disease who received either surgery or radiation
therapy (n = 32,022) was 69% that of the 12,608 men who
received no definitive treatment, adjusted for tumor and demo-
graphic characteristics and for the presence of comorbidity
(as assessed in Medicare claims data). For mortality from pros-
tate cancer itself, the corresponding relative risk associated
with receipt of active treatment was 0.67.
a. Based on the adjusted relative risk for all-cause mortality

of 0.69 associated with receipt of active treatment for
prostate cancer, and assuming that this reduced risk was
indeed a result of treatment, how many of the 4,663 deaths
in the untreated men in this studys follow-up period
might have been prevented had they also been treated?
b. Perform the same calculation for the 314 deaths from
prostate cancer itself, assuming a relative risk associated
with receipt of treatment of 0.67.
c. Assume that classification of cause of death in men with
prostate cancer is 100% accurate, and that in truth treat-
ment of prostate cancer does not have an impact on mor-
tality from other causes. What do you believe to be the
most likely explanation for the difference in the numbers
obtained in (a) and (b) above? Why?
Answer 2.11
a. Deaths (all causes) in observation group = 4,663

% potentially averted = 100% 69% = 31%
4,663 x .31 = 1,446
b. Deaths from prostate cancer in observation group = 314
% potentially averted = 100% 67% = 33%
314 x .33 = 104
c. There were 1,446 104 = 1,342 fewer deaths from causes
other than prostate cancer in men in the observational
group than would have been predicted by the rates in the
treatment group. The difference between the observed
and expected number must be the result of confound-
ing. Despite the considerable efforts of the authors to
nullify confounding (e.g., by eliminating the first year
of follow-up postdiagnosis, and by adjusting for demo-
graphic characteristics, disease characteristics, and the
presence of comorbidity), an inherent mortality disad-
vantage must have been present in the untreated men.
Question 2.12 The following is an excerpt from an article in

the JAMA:
Age-at-menarche of 38 college female athletes was ascer-

tained and related to the age of initiating training. The
18 premenarche-trained athletes had a mean menarcheal age
of 15.1 0.5 years, whereas the 20 postmenarche-trained
athletes had a mean menarcheal age of 12.8 0.2 years. Thus,
it is apparent that strenuous exercise can delay the onset of
menses to a substantial degree.
The authors choice of comparison is likely to have produced

a biased result. Why, and in what direction?
Answer 2.12 The bias results from the fact that the effect
being measureddifference in age at menarcheis intertwined
with the definition of the groups being comparedpre-versus
postmenarche initiation of training. Thus, if a group of girls
with age-at-menarche distributed normally about a mean of
13 years all began to train at 13, the mean age at menarche
for the postmenarche trained would be <13 years while the
mean age at menarche for the premenarche trained would be
>13 years. For example:
12.6 13 yrs 13.4
12.6 = Mean age at menarche for "postmenarche trained"

13.0 = Age at which all started training
13.4 = Mean age at menarche for "premenarche trained"
This bias will exaggerate the influence of training in delaying

age-at-menarche.
CHAPTER 3
Confounding
WHEN THE measured relation between an exposure and the

occurrence of disease is distorted by the co-relation of each
of these to another exposure or characteristic, we say that con-
founding is present. Some people, when they learn of an asso-
ciation based on data from one or more nonrandomized studies,
do not consider the possibility of confounding, and automati-
cally take the results at face value. Others cannot imagine that
it is ever possible to disentangle the influence of one exposure
from that of others with which it might be correlated, and so
are unwilling to make any inferences from studies in humans
that do not entail randomization. Certainly there are nonran-
domized studies in which confounding is virtually absent, and
others in which it is substantial and uncontrollable. However,
most of the time we are in between these extremes, and it is
possible to evaluate where confounding might be coming from,
gauge its likely magnitude, and even explicitly take it into
account in the study design and/or analysis. The questions in
this chapter seek to illustrate how all of this can be done in spe-
cific circumstances, and also to illustrate more fully just what
are the properties of a confounding variable.
CONFOUNDING | 71
Question 3.1 The following questions concern the data pre-

sented in the figure below:
1,300 1,300
Rate per 100,000 population
1,200 1,200
1,100 1,100
Crude death rate
1,000 1,000
900 900
800 800
700 700
600 600
Age-adjusted death rate
500 500
0 0
1940 1950 1960 1970 1980 1990
Crude and age-adjusted death rates: United States, 1940-85
a. What is the reason for the increasing disparity between

crude and age-adjusted death rates?
b. For purposes of age adjustment, can you tell which popu-
lation distribution was used as a standard? Why, or why
not?
Answer 3.1
a) The sharp fall in the age-adjusted death rate must reflect

a fall in the age-specific rates on which it is based. The fact
that the crude death rate fell to a much lesser degree must
be a result of a shift in the age distribution over time,
with groups at higher risk of death (generally the older
segments of the population) being more heavily
represented in later years.
b) The age distribution of the 1940 population was chosen
as the standard, because only in that year were the crude
and age-adjusted rates identical.
CONFOUNDING | 73
Question 3.2 The following data are taken from an article on

the occurrence of out-of-hospital cardiac arrest in New York
City during April 2002-March 20038:
Incidence per 10,000

person-years
No. of cardiac Age-
Race/ethnicity Population* arrests Crude Adjusted
Black 1,393,859 1,257 9.0 10.1

Hispanic 1,489,208 636 4.2 6.5
White 2,345,564 1,908 8.1 5.8
Other 829,378 252 3.0 4.8
*18 years and older.
Can you draw any conclusion regarding the difference in

age distribution between white and nonwhite residents of
New York City ages 18 years and above during 20022003?
If yes, how would you characterize that difference? If no, why
not? (In all demographic subgroups of the population, the inci-
dence of cardiac arrest rises sharply with increasing age.)
Answer 3.2 The white population must be relatively older

because once the age difference between groups is accounted
for, their rate of out-of-hospital cardiac arrest falls. This can
happen only if the white populations age distribution put them
at a relatively higher risk than that of members of other seg-
ments of the population, in other words, if whites were older on
average.
CONFOUNDING | 75
Question 3.3 The following table presents the incidence of

hospitalized pneumonia in the United States among persons
65 years and above during two periods of time:
19881990 20002002
Incidence per Incidence
Age Population 1,000 person- Population per 1,000
(years) size years size person-years
6574 17,506,833 10 18,495,139 12

7584 9,939,111 21 12,409,949 26
85 2,939,646 49 4,413,680 51
An increased incidence was noted in each of the three

age groups during 20002002 relative to that in 19881990.
Could confounding by age have distorted the time trends within
each age group to some degree? If yes, why, and in what manner.
If no, why not?
Answer 3.3 Yes, residual confounding by age likely is present

to at least some degree in these data, leading to an exaggeration
of the true change over time:
a) Rates rise steeply with age.

b) Across the three age categories, there were proportion-
ately more persons in the older group(s) in 20002002
than in 19881990. Thus, within each age group it is likely
that the mean age was higher in 20002002 than it was
earlier. And so, even if the actual rate of pneumonia had
not risen at all, it would appear to have done so in each of
the three 10-year age groups.
CONFOUNDING | 77
Question 3.4 In a survey of Americans 18 years and older

conducted in 2004, 24.1% of white men and 23.9% of black
men were current smokers of tobacco. Irrespective of race, the
prevalence of tobacco smoking differed little by age, except for
a sharp drop in men 65 years and older.
The small difference in current smoking (0.2%) between
white and black men noted above does not take into account
the fact that in 2004 the average age of white men was some-
what higher. After age adjustment, would you expect the inter-
racial difference to be larger, smaller (or reversed), or unchanged?
Why?
Answer 3.4 The adjusted difference will be larger. A greater

proportion of white than black American men are in the
65-year age group, and these older men have the lowest
prevalence of smoking. Removing the age disparity by means
of adjustment will allow the higher prevalence of smoking
of white men to become more evident.
CONFOUNDING | 79
Question 3.5 The following data come from a Swedish study

of children born following in vitro fertilization during
19821995.9 Records of all 26 childhood disability centers in
Sweden were reviewed to identify children in this cohort who
received services from these centers. In addition, a comparison
cohort of children was selected from the entire Swedish Medical
Birth Registry so as to be similar to the IVF cohort with respect
to sex, year of birth, and hospital of birth.
Disability
Yes No
All Children
IVF 101 5,579
Comparison cohort 119 11,241
Singletons only
IVF 45 3,183
Comparison cohort 115 10,955
Nonsingletons
IVF 56 2,396
Comparison cohort 4 286
a. What is the cumulative incidence of disability in children

conceived by means of IVF relative to that in other
children, both adjusted and not adjusted for singleton/
nonsingleton birth?
b. Which of the above relative risks addresses the following
questions:
i. What is the impact of IVF on the incidence of disability?
ii. What is the impact of IVF on the incidence of dis-
ability, beyond the impact of IVF to lead to multiple
births?
Answer 3.5
101 5,680
a) Crude relative incidence = = 1.70
119 11,360
Adjusted relative incidence =

(45 11,070 14,298) + (56 290 2, 742)
= 1.38 *
(115 3,228 14,298) + (4 2,452 2,742)
b) i. The crude relative incidence. Since multiple preg-

nancy is a consequence of IVF, it is not appropriate to
treat multiple pregnancy as a confounding variable.
ii. The adjusted relative incidence. This tells us how
much of an association would remain if, hypotheti-
cally, the excess risk of multiple pregnancy associated
with IVF could be eliminated.
* Method of Mantel and Haenszel.10

CONFOUNDING | 81
Question 3.6 A randomized trial was conducted11 in which

2,763 women with a history of coronary heart disease were ran-
domly assigned in equal proportion to:
a. a regimen of 0.625 mg/day of conjugated estrogens and

0.25 mg/day of medroxyprogesterone acetate; or
b. a placebo.
During an average follow-up of 4.1 years the incidence of the

combined endpoint, myocardial infarction and death from
coronary heart disease was identical in the two groups (rela-
tive risk = 0.99, 95% CI = 0.801.22).
A critic of the study contended that the results may have

been confounded, since diagnostic tests to ascertain the extent
of coronary and other vascular disease at the outset of this trial
were not performed on the study participants. Do you share
this concern? If yes, why? If no, why not?
Answer 3.6 Given that this is a randomized trial of 2,763 per-

sons, the treatment and placebo groups would be expected to be
extremely similar with respect to their distribution of severity
of vascular disease. Failure to have documented that severity in
study participants, and thus the failure to adjust for it, should
have introduced no bias at all.
CONFOUNDING | 83
Question 3.7 Investigators at the CDC sought to determine

the efficacy of the drug zidovudine (AZT) in preventing HIV
infection in health-care workers who had sustained a percuta-
neous exposure (e.g., a needle stick) while caring for patients
with AIDS.12 Among health-care workers with a history of such
an exposure, 27 persons who developed HIV infection and a
sample of those who remained uninfected were compared with
regard to their receipt of AZT soon after the incident.
The analysis suggested that four characteristics of the expo-
sure, or of the patient being cared for, were particularly more
common among cases than controls: (a) Deep injury, (b)
Visible blood on the device that caused the wound in the health-
care worker, (c) Injury during a procedure involving a needle in
an artery or vein, and (d) Terminal illness in the source patient.
The results of the study pertaining to the receipt of AZT fol-
lowing percutaneous exposure are summarized in the following
table:
Cases Controls
Number of the other
four risk factors AZT No AZT AZT No AZT
0 0 0 40 88
1 0 3 51 73
2 2 9 33 22
3, 4 6 7 7 6
Among health-care workers with percutaneous exposure to

HIV, estimate the risk of HIV infection in persons who received
postexposure AZT relative to the risk in those who did not: a)
Adjusted for the number of other risk factors for HIV infection
and b) Not adjusted for the other risk factors.
Why do these two estimates differ from one another?
Answer 3.7 From this case-control study, the relative risk

associated with AZT can be estimated by the odds ratio (OR).
The adjusted OR (calculated here by the method of Mantel
and Haenszel) considers only persons with at least one other
risk factor, because no case failed to have at least one.
0 73 2 22 6 6
+ +
Adjusted OR = 127 66 26 = 0.27
3 51 9 33 7 7
+ +
127 66 26
8 131
Crude OR = = 0.61
19 189
The adjusted OR is well below the null, suggesting a benefi-

cial influence of AZT in blocking transmission of HIV infection.
The crude OR is not as low as the adjusted OR, because of the
tendency (see the data in the table) for exposed health-care per-
sonnel with other risk factors for contracting HIV to have
received postexposure AZT.
CONFOUNDING | 85
Question 3.8 In white married U.S. males, the annual inci-

dence of prostate cancer is about 1 per 100,000 at ages 35 to 44
years, and about 100 per 100,000 at ages 55 to 64 years.
The following table presents the incidence of cancer of the
prostate in white U.S. men, ages 45 to 54 years, in relation to
marital status:
Marital status Rate per 100,000 per year
Married at present 12.5

Never married 11.2
Widowed 20.7
Disregarding the possible role of chance, what do you

believe to be the most likely noncausal explanation for the
observed high rate of prostate cancer among widowed men?
Answer 3.8 Because the incidence of prostate cancer rises dra-

matically with age (1 per 100,000 at 3544 years and 100 per
100,000 at 5564 years), the incidence rate is not likely to be
uniform within the 45- to 54-year age category. It may be
expected that the widowed men within this age category are
older on average than the married at present or never-married
men, thus accounting for their higher incidence rate. There is
residual confounding by age.
CONFOUNDING | 87
Question 3.9 A population-based case-control study of blad-

der cancer was conducted in a part of the western United States
in which some areas have arsenic levels in drinking water that
are relatively high (about 100 micrograms per liter).13 Persons
diagnosed with this condition during 19942000 were identi-
fied through the records of cancer registries serving a 6-county
area. Controls under age 65 years were recruited through
random digital dialing of phone numbers; controls 65 years and
over were obtained through the records of the Health Care
Financing Administration. Information concerning a history
of living or working in the areas in which arsenic levels were
elevated was obtained by means of interviews with the study
participants, as was information on potential confounding
factors.
On average, income levels in cases were lower than those of
controls.
In discussing their findings, the authors acknowledged
this imbalance, and suggested it was likely related to the
increased participation rates among [potential controls] in
higher socioeconomic brackets. Nonetheless, they concluded,
Other studies have shown little or no association between
socioeconomic status and bladder cancer, suggesting that this
variable is not likely to act as a substantial confounder.
In this study, do you believe it would be possible for annual
income to be a confounding variable when assessing the poten-
tial association between arsenic in drinking water and the inci-
dence of bladder cancer? If yes, under what circumstance? If no,
why not? (Assume that income level has been measured with-
out error.)
Answer 3.9 Income will be a confounder if it is related, posi-

tively or negatively, to ingestion of arsenic. Income is already
related to case/control status, probably by virtue of differences
in level of participation across socioeconomic strata (as the
authors suggest). The fact that income level is not truly a risk
factor for bladder cancer in the population at large is irrelevant,
given the case/control disparity in the distribution of income in
this study.
CONFOUNDING | 89
Question 3.10 The following table contains data from a New

Zealand study in which automobiles involved in crashes and a
random sample of other automobiles were compared for color.14
Association of car color with car crash injury in Auckland

No. (%) No. (%) of
of cases controls Univariate Multivariable
Car color (n = 567) (n = 588) odds ratio odds ratio*
White 145 (25.6) 146 (25.9) 1 1

Yellow 31 (5.5) 15 (2.8) 2.0 (1.0 to 4.0) 0.8 (0.3 to 2.3)
Grey 52 (9.2) 61 (10.0) 0.9 (0.6 to 1.5) 0.6 (0.3 to 1.3)
Black 36 (6.4) 34 (5.5) 1.2 (0.7 to 2.0) 2.0 (1.0 to 4.2)
Blue 91 (16.1) 96 (17.4) 0.9 (0.6 to 1.4) 0.9 (0.5 to 1.6)
Red 85 (15.0) 82 (13.3) 1.1 (0.7 to 1.8) 0.7 (0.4 to 1.4)
Green 42 (7.4) 44 (7.0) 1.1 (0.6 to 1.8) 1.8 (1.0 to 3.6)
Brown 55 (9.7) 49 (6.8) 1.4 (0.8 to 2.5) 2.1 (1.1 to 4.2)
Silver 30 (5.3) 61 (11.3) 0.5 (0.3 to 0.8) 0.4 (0.2 to 0.9)
*Adjusted for drivers age, ethnicity, alcohol consumption in past 6 hours, seat belt use, vehicle
speed, average driving time each week, driving license status, vehicle insurance status, and weather.
From the information contained in the table, which one of

the following statements do you believe to be true? Explain
your answer.
a) On the basis of their age, ethnicity, alcohol consumption,
and so forth (see footnote to the table), drivers of black or
brown cars in Auckland during 19981999 tended to be at
higher risk of a car crash injury than drivers of white cars.
b) On the basis of their age, ethnicity, alcohol consumption,
and so forth (see footnote to the table), drivers of black or
brown cars in Auckland during 19981999 tended to be at
lower risk of a car crash injury than drivers of white cars.
c) From the data presented, no inferences can be made from
these data regarding the underlying differences in risk
between drivers of black/brown versus white cars.
Answer 3.10 The correct answer is (b). Since adjustment for

the characteristics listed in the footnote led to an increase in
the OR associated with driving a black or brown car, the under-
lying risk of a car crash injury must have been low relative to
drivers of white cars.
CONFOUNDING | 91
Question 3.11 The following table is from an article that

describes the prevalence at birth of a particular congenital mal-
formation in Ontario, Canada, before and after a program of
folate fortification of cereal grain products was begun in
January, 199815:
Crude Age-adjusted
prevalence prevalence
Before After ratio ratio
fortification fortification (95% CI) (95% CI)
Length of 48 29
observation
(months)
Maternal age 30.1 30.9
(mean, SD) (0.16) (0.081)
(years)
Number of 218,977 117,986
women
Number of 248 69
women with
an affected child
Prevalence 1.13 0.58 0.52 (0.40 0.62
(per 1,000 0.67) (0.460.83)
infants)
From the data presented in the table, what can be concluded

regarding an association between maternal age and the occur-
rence of this malformation? Explain your answer.
a) On average, the risk rises with increasing maternal age.

b) On average, the risk falls with increasing maternal age.
c) The risk is unaffected by maternal age.
d) No conclusion can be drawn about a relation of maternal
age to the prevalence of the malformation from these data.
Answer 3.11 The crude and age-adjusted relative risks associ-

ated with being born after the introduction of fortification
differ, so maternal age must be related to the occurrence of the
malformation. And, since we know that:
a) women in the latter time period were, on average, older

than those in the earlier period;
and
b) adjustment for maternal age made the prevalence ratio

rise,
older women must be a lower risk age group. The correct

answer is (b).
CONFOUNDING | 93
Question 3.12 The following is an excerpt from an article,

Cancer beats a retreat, that appeared in a 1998 issue of US News
and World Report. It dealt with recent declines in cancer incidence
rates in the United States.
The researchers acknowledge that the cancer war might not be

doing as well as the new data suggest. They adjusted their
numbers on the basis of the age of the U.S. population in 1970,
a widely used standard in cancer studies. However, the popula-
tion has aged in the past 25 years. Since cancer disproportion-
ately strikes older people, using a current standard might yield
less encouraging results.
Is it possible that using the age distribution of the U.S. pop-

ulation in 1990 as a standard might yield less encouraging
results than using the 1970 population? If so, how could this
occur? If not, why not?
Answer 3.12 The use of the 1990 U.S. population as a stan-

dard (which gives more weight to the rates in older age groups)
would tend to give a relatively smaller estimate of the declining
rate only if the size of the decline were smaller in older than in
younger persons.
CONFOUNDING | 95

There is a sharp contrast in the consistency of success in stud-

ies that have sought genotype-phenotype associations in ani-
mals and in humans. For example, animal models of depression
and anxiety disorders have consistently demonstrated geno-
type-phenotype associations. By contrast, a recent genome-
wide association study (GWAS) of depression found no
significant associations. One central difference between these
2 research approaches lies in control over potentially relevant
environmental exposures. These exposures are effectively ran-
domized in animal models, but such control is absent from
observational human gene-hunting studies.
Assume that the genome-wide association studies were

conducted in racially homogeneous populations. Do you agree
that the lack of randomization in these studies is likely to be
responsible for the difference between their results and those
obtained in studies conducted in other species? If yes, why?
If no, why not?
Answer 3.13 No. Whatever environmental exposures are

related to the development of depression and anxiety disorders
in human beings, these almost certainly do not bear on a per-
sons genotype. Given the likely absence of confounding, the
lack of randomization of genotype in human studies of depres-
sion should not affect the validity of the results obtained.
CONFOUNDING | 97
Question 3.14 The data presented in the following table come

from in-person interviews of random samples of the U.S. popu-
lation ages 19 years and above. They indicate that the preva-
lence of cigarette smoking in 2008 was greater among American
Indian/Alaska Native persons than among non-Hispanic whites.
The data are not adjusted for age, however, and the proportion
of American Indians/Alaska Natives over 64 years of age in
those years was smaller than the corresponding proportion of
non-Hispanic whites. If age adjustment were to be done, would
you expect the difference in smoking prevalence between these
two demographic subgroups to increase, shrink, or remain the
same? Explain.
Percentage of persons aged 18 years who were current

cigarette smokers, by sex,race/ethnicity, and age National
Health Interview Survey, United States 2008
Men Women
(n = 9,387) (n = 12,138)
Characteristic % (95% CI) % (95% CI)
Race/Ethnicity
White, non-Hispanic 23.5 (22.224.9) 20.6 (19.321.9)
American Indian/ 42.3 (27.457.2) 22.4 (12.532.3)
Alaska Native
Age (yrs)
18-24 23.7 (20.327.1) 19.0 (16.221.8)
2544 26.4 (24.528.2) 21.1 (19.522.7)
4564 24.8 (22.826.7) 20.5 (18.922.1)
65 10.6 (8.812.3) 8.4 (7.19.6)
Answer 3.14 The difference would be expected to shrink.

Younger Americans tend to be current smokers more than older
ones (see table), and American Indians/Alaska Natives tend to
be younger than whites. Thus, a part of the difference in smok-
ing prevalence is attributable to age alone, and that part would
disappear with age adjustment.
CONFOUNDING | 99
Question 3.15 The following appeared in the November 13,

1999, issue of Lancet:
The general health of the East German population was below

the standard of their Western counterparts despite an empha-
sis on public health. In 1991, life expectancy was 3.2 years
shorter in men (69.9 years) and 2.3 years in women (77.2 years)
than in West Germans. Mortality in 1991 also showed impor-
tant differences: 1324 per 100,000 population in East German
women, 1215 in men compared with 1149 in West German
women and 1061 in men.
In trying to determine whether the risk of death differed

between East and West Germans in 1991, youd be loath to use
the mortality rates provided in this article. Why do you believe
they might be misleading?
1 0 0 | EXERCISES IN EPIDEMIOLOGY
Answer 3.15 The mortality rates for 1991, higher in women

than in men in both populations, must not be age-adjusted.
(We know this because the greater life expectancy in German
women than German men must be a result of the age-specific
mortality among women being low relative to that of men.)
Therefore, if the East German population had been older, on
average, than the West German population, some or all of the
difference in the crude mortality rates could be due to con-
founding by age.
CONFOUNDING | 101
Question 3.16 I came across the following statement (and

data) in a draft of a masters thesis a long time ago. Can you
think of a reason that the observed association between marital
status and suicide is almost certainly greatly overestimated?
In general, suicide is less frequent among the marriedexcept

for the young married population. As the table indicates, in
the under-20 age group, suicide rates are higher in married or
divorced persons than in single persons. Explanations for this
phenomenon may be a desire to escape unsatisfactory home
conditions or pregnancy resulting in an unplanned, unhappy
marriage.
Death rates from suicide by marital status

and sex, persons less than 20 years of age:
United States, 19491951
Total Single Married Divorced
White males 0.9 0.9 6.2 14.5

White females 0.4 0.3 3.1 13.8
Answer 3.16 There is an unusually great degree of confound-

ing by age here. Persons at the upper end of the 0- to 20-year
age groups are the ones relatively most likely to get married
(and divorced) and (irrespective of marital status) to commit
suicide.
CONFOUNDING | 103
Question 3.17 The following data describe the occurrence of

primary cesarean delivery (i.e., in women with no prior cesar-
ean delivery), expressed as a percentage of all live births, in the
United States in 1990:
Age of mother
Under 20 2024 2529 3034 3539 4049
Total years years years years years years
All births 16.0 14.7 15.0 16.0 16.5 19.0 23.5

1st 24.6 16.9 22.6 27.1 32.2 39.2 46.9
2nd 8.9 6.8 7.3 8.7 10.4 13.6 20.1
3rd 8.4 6.3 6.5 7.7 9.0 12.0 18.1
4th 8.8 7.6 6.8 7.4 8.8 11.2 15.1
and over
There is generally a 2- to 3-fold difference in the proportion

of cesarean deliveries between the youngest and the oldest
categories of maternal age within individual categories of live
birth order. However, the difference is considerably
smaller14.7 percent versus 23.5 percentwhen the compar-
ison is made for all birth orders combined.
What do you believe to be the explanation for the relative
difference in the frequency of cesarean delivery across maternal
age being so much smaller when examined overall than within
individual birth order categories? Why?
Answer 3.17 There is confounding by birth order: (a) at any

maternal age, first births are associated with relatively high
percentage of cesarean deliveries; (b) young mothers, relative to
older ones, would be expected to be in the lower birth orders,
and on that basis alone would be at increased risk of cesarean
section. The confounding attenuates to some degree the very
strong association seen in the birth-order specific results.
CONFOUNDING | 105
Question 3.18 In a case-control study that was based on

information obtained from spouses, a history of high-intensity
leisure-time physical activity (LTPA) during the prior year
was associated with a reduced risk of primary cardiac arrest.16
Relative to the risk in persons with light or no LTPA, that in
persons with high-intensity activity was 0.19; after adjusting
for age, smoking, education, diabetes, hypertension and self-
reported health status (fair, good, excellent, or very good),
the corresponding relative risk was 0.36.
There was a case-control difference in self-reported health
status, and it is reasonable to assume that adjustment for this
variable explained at least some of the difference between the
unadjusted and adjusted odds ratios associated with high-
intensity LTPA. However, health status is difficult to assess
from interview data alone, so some subjects were probably mis-
classified on this variable. Had no misclassification been pres-
ent, what would you predict the adjusted odds ratio associated
with high intensity LTPA to be?
a. 0.36
b. Greater than 0.36
c. Less than 0.36
d. No prediction is possible.
Explain your answer.

Answer 3.18 Inaccurate measurement of a confounding vari-

able will give rise to an adjusted risk estimate that is spuriously
close to the unadjusted one. So, more complete adjustment for
self-reported health status would be expected to lead to a rela-
tive risk associated with high LTPA that is greater than 0.36.
CONFOUNDING | 107
Question 3.19 Investigators in Australia surveyed 202 preg-

nant women who were undergoing abortion who were using
oral contraceptives (OCs) when they became pregnant. Among
other things, the investigators gathered information on type of
OC used by these 202 women, in comparison with the types of
OCs used by all OC users in Australia. They were specifically
interested in pregnancies among women using triphasic OCs.
They presented the following results:
Abortion-seeking All Australian

OC users OC users
Mean age 23 years 30 years
Type OC used
Triphasic 52% 42%
Monophasic
30 microgram ethinylestradiol 27.2% 29.2%
50 microgram ethinylestradiol 7.9% 13.2%
Norethisterone 9.4% 9.0%
Progestin-only 3.5% 6.2%
The investigators found that triphasic OC use was more

common (P <0.01) in OC users seeking abortion than would be
expected based on the national usage. They concluded that tri-
phasic OCs were more likely than other OCs to increase the risk
of inadvertent pregnancy (and they hypothesized this was
because they have a smaller margin of safety due to decreased
progestin content).
You are concerned that the age difference between the
two groups above could be distorting the results. Under what
circumstance would this be so?
Answer 3.19 These data indicate that OC users seeking abor-

tion are younger than OC users overall. If young age is associ-
ated with an increased risk of abortion (either because of a
greater risk of pregnancy, or among pregnant women, a greater
use of induced abortion) and if younger OC users tend to use
triphasic preparations relatively more than older women do, it
is possible that there is no true association between triphasic
OC use and inadvertent pregnancy. Information on the type of
OC use, by age, is necessary to determine if this is the case.
CONFOUNDING | 109
Question 3.20 A case-control study of the relationship

between asthma and a history of pertussis among children
(2 years of age) was carried out in one community. Although
overall participation was good, not all parents of cases and con-
trols could be interviewed. A difference in the percentage inter-
viewed based on case/control status and on day care enrollment
was noted. Only 63% of control parents with a child enrolled in
day care could be interviewed, versus 82% of control parents
with a child who was not enrolled in day care. In all, 93% of
case parents were interviewed, a percentage that was the same
for parents of children who were and were not enrolled in
day care.
In this study, under what circumstance would enrollment
in day care confound the association between asthma and
pertussis vaccination?
Answer 3.20 Because there was differential case-control

response by day care enrollment status, day care enrollment
would be a confounder if it were also associated with pertussis
vaccination status.
CONFOUNDING | 111
Question 3.21 A case-control study of breast cancer in rela-

tion to prior use of postmenopausal hormones was conducted
among American women enrolled in a network of health insur-
ance plans.17 Women who were newly diagnosed with breast
cancer and who had been enrolled for at least 2 years prior to
that date comprised the case group.
For each case, four women were selected as controls,
matched on year of birth and enrollment status as of the time
of the cases diagnosis (and for the 2 years prior to that time).
Data on hormone use were obtained from paid claims for
pharmaceuticals. While a similar proportion of cases and con-
trols had taken estrogen alone prior to the time of the cases
diagnoses (odds ratio = 0.96, referent category being hormone
nonusers), a higher portion of cases than controls had taken
combined estrogen-progestin therapy (odds ratio = 1.44, refer-
ent category being hormone nonusers).
The information available on study subjects did not include
whether they had a history of bilateral oophorectomy. Such
a history is associated with a reduced risk of breast cancer. If
a woman has not undergone bilateral oophorectomy and is pre-
scribed hormone therapy, generally that therapy will be in the
form of a combined estrogen-progestin regimen. Women whose
ovaries have been removed generally will be given estrogen
alone.
Is there reason to believe that the two odds ratios obtained
are biased due to the investigators inability to adjust for a his-
tory of bilateral oophorectomy? If yes, why, and in which direc-
tion? If no, why not?
Answer 3.21 The inability to adjust for a history of bilateral

oophorectomy will lead to bias (confounding), at least to some
degree. Based on their absence of ovarian tissue, the users of
estrogen alone are a low-risk group for breast cancer: therefore,
the odds ratio associated with use of estrogen alone that was
obtained in this study is spuriously low. In contrast, users of
combined therapy have an inherently elevated risk, since they
do have intact ovaries. As a result, this studys odds ratio associ-
ated with receipt of hormone therapy is likely to be falsely
high.
CONFOUNDING | 113
Question 3.22 Based on the figure below, which one of the

following statements is correct for American women giving
birth during 19791982:
80
69 72
70
63
60 Black
53
50 47
White
Percent
43
40
30
20
10
0
All women Never married Ever married
Marital status of mother at time of birth
Percentage of mothers receiving prenatal care in the first trimester,

by race and marital status: United States, 19791982
a) A greater proportion of white than black mothers had

never been married.
b) A greater proportion of black than white mothers had
never been married.
c) An equal proportion of black and white mothers had
never been married.
d) From the data provided, no conclusion as to the marital
status distribution of white and black mothers is possible.

Answer 3.22 The correct answer is (b). Overall, 69% of white

women received prenatal care, a weighted average of the figures
of 47% for never-married women and 72% for ever-married
women. The preponderance of these women must have been
ever-married,* to explain the fact that the overall percentage
and never-married percentage are so similar. In contrast, only
about half the black women had ever married: the percentage
receiving prenatal care overall53%lies halfway between
the percentages for never- and ever-married women.
* x = proportion of white women who were ever married

72 x + 47 (1 x ) = 69
x =0.88
CONFOUNDING | 115
Question 3.23 The following data come from a 1975 survey of

a representative sample of about 140,000 persons from the
United States population:
% cigarette smokers
Age (years) Male Female
2124 41.3 34.0

2534 43.9 35.4
3544 47.1 36.4
4554 41.1 32.8
5564 33.7 25.9
6574 24.2 7.1
% cigarette
smokers
Marital Status Male Female
Married 38.3 28.3

Single 37.5 30.6
Divorced or 60.1 50.0
separated
Widowed 35.7 19.3
From these data, can you conclude that there is an associa-

tion, beyond that which could be explained on the basis of age:
(a) between smoking and widowhood among females? (b) among
persons of both sexes, between smoking and being divorced or
separated? (No calculations are required.)
Answer 3.23
a) Widowed women are likely to be considerably older, on

average, than other women, and (at the time the data
were collected) older women were less often cigarette
smokers than younger women. Thus, it is possible that
after controlling for age, widows may not have had a
smaller proportion of cigarette smokers than women in
the other marital status categories.
b) The frequency of cigarette smoking among divorced/
separated persons was high50%60%relative to that
of any age groups in the population as a whole. No matter
what the age distribution of divorced/separated persons,
a prevalence of cigarette smoking of 50%60% cannot be
explained solely by of an overrepresentation of a particu-
lar age group. An association beyond that which could be
explained by age must have been present.
CONFOUNDING | 117
Question 3.24 This question is based on the following (slightly

paraphrased) abstract18:
Background. Among women pregnant for the second time, the
risk of pre-eclampsia is lower than in their first pregnancy, but
not if the mother has a new partner for the second pregnancy.
One explanation is that the risk is reduced with repeated
maternal exposure and adaptation to specific antigens from
the same partner. However, the difference in risk might instead
be explained by the interval between births.
Methods. We used data from the Medical Birth Registry of
Norway, a population-based registry that includes births that
occurred between 1967 and 1998. We studied 551,478 women
who had two or more singleton deliveries.
Results. The risk of pre-eclampsia in a second pregnancy
was directly related to the time that had elapsed since the pre-
ceding delivery, and when the interbirth interval was 10 years
or more, the risk approximated that among women pregnant
for the first time. In unadjusted analyses, a second pregnancy
involving a new partner was associated with higher risk of pre-
eclampsia than a second pregnancy with the same partner, but
after adjustment for the interbirth interval, the difference in
risk of pre-eclampsia was reduced.
From the foregoing, what can be concluded about the
relation between interbirth interval and the presence of a new
partner? Explain your answer.
a) Women with new partners tended to have a shorter
interbirth interval than women with the same partner.
b) Women with new partners tended to have a relatively
longer interbirth interval.
c) Women with or without new partners had, on average,
the same interbirth interval.
d) Nothing can be concluded about a possible association
between interbirth interval and the presence of a new
partner.
Answer 3.24 The correct answer is (b). The presence of a new

partner must be associated with a longer interbirth interval.
The crude elevated risk for pre-eclampsia present in women
with new partners was diminished once adjustment was made
for interbirth interval. Therefore, since long interbirth intervals
are associated with high risk, these must have been relatively
more common in women with new partners.
CHAPTER 4
Cohort Studies
COHORT STUDIES compare the occurrence of an illness or

injury between persons with and without an exposure or char-
acteristic. Threats to the validity of cohort studies can come
from inaccurate characterization of exposure status; incomplete
or inaccurate ascertainment of health outcomes; and differ-
ences between exposed and unexposed persons with respect to
factors that themselves bear on the occurrence of the outcome.
The forward-looking structure of cohort studiesstart with
exposure, follow subjects for the development of an outcome
eventresembles that of randomized trials, and that similarity
can put you off your guard when trying to interpret the results
of such studies. The questions that follow are intended to put
you back on guard.
COHORT STUDIES | 121
Question 4.1 In a study of more than 20,000 women who had

received a cosmetic breast implant (one that was not provided
in relation to breast cancer surgery), all-cause mortality begin-
ning 1 year later was only 74% that of socioeconomically and
demographically comparable women in the population as a
whole (95% CI = 0.680.81).19 However, all-cause mortality in
the implant recipients was nearly identical to that of some
16,000 women beginning 1 year after having undergone another
form of cosmetic surgery during the same period of time (rela-
tive risk = 1.02, 95% CI = 0.891.17, adjusted for demographic
characteristics).
Despite the observed 26% reduction in mortality compared
to women in general, the authors concluded that the receipt of
breast implants does not appear to influence death rates. Do
you agree? If yes, why? If no, why not?
Answer 4.1 Women who have a serious illness are unlikely

candidates for cosmetic implants; the proportion of women
with such an illness is undoubtedly lower in the implant cohort
than in the population as a whole (even after restricting the
period of observation by excluding the first year after the oper-
ation). Thus, even in the absence of any influence of breast
implants on the risk of death, the mortality experience of
women receiving them would be expected to be more favorable
than that of women in general. A better basis for comparison
would be the death rates for the women undergoing other forms
of cosmetic surgery, in which the same selection factors would
be expected to be present. The similarity of death rates between
implant recipients and other women undergoing cosmetic sur-
gery argues that receipt of breast implants has no bearing on
the risk of death.
Question 4.2 A study of the efficacy of pneumococcal vaccina-

tion in the elderly was described as follows20:
We conducted a 2-year retrospective cohort study among all

elderly members of a staff-model managed care organization
who had a baseline diagnosis of chronic lung disease. The study
outcomes were assessed over 2 years, from November 15,
1993, through November 14, 1995, and included hospitaliza-
tions for pneumonia and influenza.
Of 1898 subjects, 1280 (67%) had received pneumococcal
vaccination. This included 843 (44%) who were vaccinated
prior to November 15, 1993 and an additional 437 (23%)
vaccinated after that date. During the follow-up period there
were 174 hospitalizations for pneumonia and influenza. The
observed cumulative incidence was 138 per 1000 in the 618
unvaccinated persons and 70 per 1000 in the 1280 persons
following their receipt of vaccination. This represents a 49%
reduction in hospitalization for pneumonia and influenza.
Even if there were neither misclassification nor confounding

in this study, the estimate of benefit associated with pneumo-
coccal vaccination in elderly patients with chronic lung disease
must be biased.
a. Why? In which direction?

b. How could the analysis be conducted to remove this
source of bias?
Answer 4.2
a. Not all patients in this study were followed for the same
length of time. Specifically, the 437 patients vaccinated
during the 2-year study period were at risk for hospi-
talization only for the period after immunization. Thus,
even had the hospitalization rates been the same in
vaccinated and unvaccinated groups, the cumulative inci-
dence would be lower in the former group. This would lead
to an overestimate of the vaccines efficacy.
b. The analysis should employ a person-time denominator,
so that incidence rates can be calculated. The 437 persons
vaccinated after November 15, 1993, would contribute
person-time-at-risk to the experience of the unvacci-
nated persons until their date of vaccination, and to the
experience of vaccinated persons afterward. Because
the risk of hospitalization for pneumonia and influenza
may be relatively low in the several week period immedi-
ately after November 15 (i.e., before the seasonal peak),
adjustment for calendar period may be needed. Otherwise,
confounding could arise from the different calendar dis-
tribution of person-time between vaccinated and unvac-
cinated groups.
Question 4.3 The following is taken from the abstract of an

article on the mortality experience of employees at a polymer
manufacturing facility, the DuPont Washington Works plant in
West Virginia21:
Methods. The cohort comprised 6,027 men and women who had
worked at the facility between 1948 and 2002; these years delimit
the mortality follow-up period. Standardized mortality ratios
(SMRs) were estimated to compare the observed number of
deaths to expected numbers derived from mortality rates for
2 reference populations: the West Virginia state population and an
8-state regional employee population from the same company.
The results of the study for deaths from heart disease are
shown in the table below. (The SMRs presented are adjusted for
age, sex, and calendar time.)
SMR estimates with 95% confidence intervals for mortality for
heart disease for all Washington Works employees compared
to 2 external reference populations
DuPont 8-state
WW regional employee
cohort WV population population
95% 95%
Cause of death O E SMR CI E SMR CI
All heart disease 314 475.6 66.0 58.9, 284.5 110.4 98.5,
73.7 123.3
WW = Washington Works; WV = West Virginia; O = observed deaths; E = expected

deaths; SMR = standardized mortality ratio = O/E x 100%; CI = confidence interval.
The confidence intervals around the SMRs based on expected

deaths in the DuPont Region 1 workers are somewhat wider
than those based on expected deaths in West Virginia.
Nonetheless, when evaluating the possible impact of employ-
ment at the Washington Works plant on mortality from heart
disease, why might the SMRs based on death rates in DuPont
Region 1 workers provide a more valid estimate?
Answer 4.3 Persons with heart disease are less likely than
other persons to become employed and stay employed, and also
are more likely to die of heart disease. The proportion of WW
employees with heart disease is almost certainly smaller than in
the West Virginia population, leading to a spuriously low esti-
mate of relative mortality from heart disease when these popu-
lation death rates are used as a basis for comparison. A similar
distortion would not be expected to be present using mortality
rates of other workers as a means of determining the expected
number of deaths.
Question 4.4 In Paris in the 1830s, bloodletting was believed

to be efficacious in the treatment of pneumonia (and a number
of other diseases). Virtually all patients diagnosed with pneu-
monia underwent this treatment, some early in the course of
the disease and some later. Louis compared the case-fatality in
41 patients with pneumonia who were bled within the first
4 days of disease onset with that in 36 others who were bled
later. The number of deaths were 18 and 9 in the two groups
(case-fatality = 44% and 25%, respectively). While the early-
bled cases were somewhat older than the late-bled cases (41 vs.
38 years), Louis concluded that the effect of venesection
[i.e., blood letting] on the progress of pneumonitis is much less
than is commonly thought.
Presented with these data, todays epidemiologists would
exclude from the analysis the experience during the first 4 days
following disease onset, and would not calculate case-fatality
from disease onset but rather mortality rates starting on day 5
(for the early-bled groups) or from the time bloodletting began
(late-bled group), adjusting for number of days since disease
onset. What would be the rationale for this modified approach?
Do you expect that todays method would raise or lower the
estimate of relative risk of death associated with receiving early
blood letting?
Answer 4.4 A comparison of case-fatality from the time

of disease onset inflates then-relative case-fatality associated
with early bleeding, given that deaths that occur during the
4 days following disease onset can be only in the early-bled
group. The rate-based approach that takes this into account,
and also adjusts for time since disease onset (after four days),
will produce a relative risk of death associated with early blood
letting that is loweras long as some deaths do occur within
the first four days after illness onset.
Question 4.5 This question is based on the following abstract

(abridged):
Methods We conducted a retrospective cohort study of postwar

mortality according to cause among 695,516 Gulf War veterans
and 746,291 other veterans. The follow-up continued through
September 1993. A stratified, multivariate analysis (with Cox
proportional-hazards models) controlled for branch of service,
type of unit, age, sex, and race in comparing the two groups. We
used standardized mortality ratios to compare the groups of
veterans with the general population of the United States.
Results Among the Gulf War veterans, there was a small
but significant excess of deaths as compared with the veterans
who did not serve in the Persian Gulf (adjusted rate ratio, 1.09;
95 percent confidence interval, 1.01 to 1.16). In both groups of
veterans the mortality rates were lower overall than those in the
general population. The adjusted standardized mortality ratios
were 0.44 (95 percent confidence interval, 0.42 to 0.47) for Gulf
War veterans and 0.38 (0.36 to 0.40) for other veterans.
(N ENGL J MED 1996;335:1498-504.)
The veterans who served in the Persian Gulf War had

a postwar mortality that was 9% higher than that of other vet-
erans, but only 44% that of Americans in general. Which of
these figures do you believe is most likely to reflect the subse-
quent impact (if any) of having experienced the Persian Gulf
War? Why?
Answer 4.5 A comparison of mortality in Gulf War veterans

to that in similar-aged Americans in general would likely be
biased, since persons with a number of diseases would not be
eligible for military service and would also be at an increased
risk of death during 19911993. The more appropriate com-
parison group is that of veterans who did not serve in the
Persian Gulf (which produced the rate ratio of 1.09).
Question 4.6 A cohort study was conducted among 6,849

Swedish men with localized prostate cancer diagnosed during
19972002.22 Death rates through 2008 were compared
between those who did and did not undergo treatment with
curative intent (most commonly, a radical prostatectomy). The
cumulative 10-year mortality from prostate cancer was low
(2.7 per 100) in the men receiving an attempt at cure, and only
0.9 per 100 higher than this in those not so treated. About 10
per 100 of the actively treated men died of causes other than
prostate cancer; adjusting for age, the corresponding figure for
the men in whom a curative procedure was not attempted was
nearly twice as high.
The very large observed difference in mortality from causes
other than prostate cancer was almost certainly not the result
of chance, and it seems unlikely to be a result of an attempt to
cure localized prostate cancer. What do you believe to be the
explanation?
Answer 4.6 When interpreting the results of any cohort study,

it is always necessary to ask, Could some reason for the pres-
ence or level of exposure itself bear a relation to the outcome in
question? In this instance, it seems likely that one factor enter-
ing into the decision to attempt a curative procedure in a man
with localized prostate cancer is the perceived life expectancy of
that man: The presence of a life-shortening condition (e.g., heart
disease, some other form of cancer) would argue against an
intervention that (even though potentially curative) would entail
its own morbidity. As a result of this source of confounding, the
observed reduction in mortality from causes other than pros-
tate cancer associated with active treatment almost certainly
does not reflect a benefit of that treatment.
Question 4.7 This question is based on an excerpt of an

abstract of a published article23:
Background. Reports on the relation between anthropometric

variables (height, weight) and physical activity with ovar-
ian cancer have been inconclusive. The objective of the
current study was to extend investigation of potential
associations in the Iowa Womens Health Study cohort.
Methods. The relation between self-reported anthropomet-
ric variables and incident ovarian cancer was studied in
a prospective cohort of women ages 5569 years who
were followed for 15 years. Two hundred twenty-three
incident cases of epithelial ovarian cancer were identi-
fied by linkage to a cancer registry.
Results. No association was found overall between ovarian
cancer and height. Although current body mass index
(BMI) was not associated with ovarian cancer, a BMI 30
kg/m2 at age 18 years appeared to be associated positively
with ovarian cancer (multivariate-adjusted RR, 1.83 for
BMI 30 kg/m2 vs. BMI <25 kg/m2; 95% CI, 0.903.72),
and this association was stronger after exclusion of the
first 2 years of follow-up (RR, 2.15; 95% CI, 1.054.40).
Conclusions. Anthropometric variables were not major risk
factors for ovarian cancer in the cohort studied; how-
ever, high BMI in early adulthood may increase the risk
of ovarian cancer among postmenopausal women.
In some cohort studies, events and person-time in exposed

and unexposed persons begins only after a period of time has
elapsed following formation of the cohort. When evaluating the
possible influence of BMI at age 18 years as a predictor of risk of
ovarian cancer in the Iowa Womens Health Study, would you
recommend that approach, or instead one in which events and
person-time are tabulated right away? Why?
Answer 4.7 The exclusion of events for a period of time

following the formation of a cohort is appropriate if there is
concern that there are occult cases of disease among cohort
members in whom exposure status ascertained at the begin-
ning of follow-up does not reflect that at the time their disease
was being produced. For the variable BMI at age 18, this is of no
concern, given that the earliest a womans follow-up could begin
was at age 55. The appropriate analysis in this study should con-
sider events and person-time from the very start of follow-up.
Question 4.8 A British study enrolled a large number of post-

menopausal women attending a breast cancer screening pro-
gram and ascertained whether they were or were not taking
hormone replacement therapy (HRT) at that time.24 During the
follow-up period, 2,894 users of HRT at baseline developed
breast cancer, versus 3,202 never users, producing a relative
risk of 1.66. Deaths from breast cancer occurred in 238 users of
HRT at baseline and in 191 never users (relative risk = 1.22).
A letter to the editor (Lancet 2003;362:329) argued that
the data provided for breast-cancer mortality are somewhat

misleading. Compared with never-users of HRT the relative
risk of death from breast cancer was raised in current-users.
However, this finding should not be interpreted as evidence
that HRT increases the risk of mortality among women diag-
nosed with breast cancer; the overall breast-cancer mortality
rate will necessarily be higher in current-users as a result
of the higher frequency of breast cancer among such women.
To quote breast-cancer mortality figures for the subgroup of
women diagnosed with breast cancer would seem more appro-
priatei.e., never-users of HRT, mortality rate 8.2% (238 of
2894) versus current-users of HRT, mortality rate 6.0% (191
of 3202). These figures give a crude relative risk estimate of
0.725 for current-users versus never-users for breast-cancer
mortality, indicating a lower risk of death in women taking
HRT at the time of their diagnosis with breast cancer than in
never-users.
Which of the two approaches to gauging mortality from

breast cancer in women who use HRTthat of the authors of the
study, or that of the authors of the letter to the editorhas the
potential to provide data that are somewhat misleading? Why?
Answer 4.8 The data produced by the authors of the letter are
misleading. A comparison of the risk of death from breast
cancer in women who do and do not receive HRT needs to incor-
porate the possible influence of HRT on both incidence and on
case-fatality. The use of mortality rates in users and nonusers
by the authors of the article does this; the use of case-fatality
alone does not.
Question 4.9 A meta-analysis of 13 studies of mortality from

brain cancer among firefighters obtained a summary standard-
ized mortality ratio (SMR) of 1.09 (95% CI = 0.921.25).25 Using
data from six of these studies that provided results based on
duration of employment, a second meta-analysis conducted by
the same author showed the following:
Duration of
employment
(years) Observed Expected SMR
<10 8 6.50 1.23

1019 12 7.41 1.62
2029 11 6.30 1.75
30 11 5.24 2.10
Put aside the possible concern with healthy worker or

other form of bias in these studies. Also, assume that both dura-
tion of firefighting employment and death from brain cancer
have been ascertained without error. Explain why the SMRs
above likely overstate the association between any particular
duration of firefighting and mortality from brain cancer.
Answer 4.9 A weighted average of the SMRs in the table would

give an overall SMR considerably greater than the value of 1.09
that is based on the results of all studies. The results of the six
studies are not representative of the entire group of 13, most
likely as a result of a form of publication biasthe tendency of
authors reporting the results of individual studies to preferen-
tially present positive results. Plausibly, the authors of one or
more of the remaining seven studies examined SMRs by duration
of employment, obtained a null or inverse relation, and chose to
provide to readers only the data for all durations combined.
Question 4.10 In a randomized controlled trial of screen-

ing for two forms of cancer (breast and colon), more than
150,000 men and women were recruited to take part. Follow-up
of these individuals took place over an average of 5 years for
cancer incidence and cause-specific mortality.
In an analysis unrelated to those bearing on the efficacy of
screening, incidence and mortality rates in all trial participants
combined were compared to those of demographically compa-
rable individuals in the population as a whole. The incidence of
cancer (excluding cancer of the breast and colon) in the trial
participants was 89% that of the general population, whereas
the corresponding figure for cancer mortality was 56%.
Because of the large number of events, chance is a highly
unlikely explanation for the difference between the relative risk
for cancer incidence (0.89) and that for cancer mortality (0.56).
What do you believe to be the most likely explanation for the
difference?
Answer 4.10 Persons dying of cancer are unlikely to be par-

ticipants in a cancer screening program. Their exclusion from
the study participants, but not from the comparison popula-
tion, would be expected to have a large impact on relative cancer
mortality but little or none on relative cancer incidence. This
represents a form of healthy screenee bias.26
Question 4.11 The occurrence of transient neurologic symp-

toms (transient ischemic attacks, TIA) commonly heralds the
later occurrence of a stroke (with its associated nontransient
neurologic damage). In 1,707 patients with a TIA, a stroke took
place in the subsequent 90 days in 14% of the 235 who received
warfarin anticoagulation therapy at the time of the TIA and in
10% of the other TIA patients (p = .04). The latter patients gen-
erally received aspirin or no specific therapy.
The authors of this study warned that despite the excellent
quality of information on treatment received and on outcome
events, the study may not provide valid data on the (lack of) effi-
cacy of warfarin anticoagulation in persons with a TIA. What do
you believe is the primary basis for their cautious interpretation?
Answer 4.11 The authors likely are concerned with the

presence of confoundingamong patients with a TIA, warfarin
anticoagulation treatment may have been selectively adminis-
tered to those with an inherently greater stroke risk (perhaps
based on age or the severity of the transient episode).
Question 4.12 Investigators at a clinic specializing in post-

trauma care wished to determine whether persons who sustain
acute neck trauma are at an increased risk of developing a dif-
fuse pain syndrome (i.e., involving multiple parts of the body).
They identified 102 patients in their clinic an average of
12 months after neck trauma, in whom 22 (21.6%) had symp-
toms that met the criteria for diffuse pain syndrome (DPS).
In contrast, among 59 patients seen in the clinic an average
of 12 months after a lower extremity fracture, only 1 (1.7%)
reported symptoms consistent with DPS (p = 0.001). Based in
large part on the strong association observed in this study, a
review article concluded that the hypothesis that acute neck
trauma can trigger DPS meets established criteria for deter-
mining causality.
In this study, not all patients who sustained a neck injury
were included, but rather just those who sought care in the
clinic. Under what circumstance could this choice have led to a
spurious exaggeration of the association between neck injury
and DPS? (Assume that demographic and other risk factors for
DPA are comparable between the patients with neck injury and
those with lower extremity fracture.)
Answer 4.12 It may be that most or all patients with a lower

extremity fracture require continuing care through 12 months
postfracture, in other words, they will be seen even in the
absence of any new condition. If the same were not true for
neck trauma patients, then those seeking care from the clinic an
average of 12 months later will be overrepresented by patients
with new problems, including DPS. Only a cohort study that
monitors the later DPS status of all victims of neck and lower
extremity trauma can be trusted to provide a valid result.
Question 4.13 You are planning a cohort study of the possible

influence of elective induction of labor at 39 weeks gestation
on the risk of several adverse maternal and fetal outcomes,
including the need for cesarean section. As a basis for compari-
son to the women receiving elective induction, you are consid-
ering two possibilities:
(a) Women who delivered a child at 39 weeks who went

into labor spontaneously and (like the women who
were electively induced) did not have any indication for
induction (e.g., preeclampsia, gestational diabetes)
(b) Women who at 39 weeks had not yet delivered and
(as of that time) had no indication for induction.
Which of (a) or (b) would be the more likely to provide an

unbiased estimate?
Answer 4.13 The correct answer is (b), because in this clinical

scenario the alternative to a decision to electively induce labor
at 39 weeks is to allow the pregnancy to continue. While some
women whose delivery is not induced will indeed soon deliver
on their own, others will remain pregnant at 40 weeks and
beyond, during which time pregnancy complications often
develop. These complications may lead to the decision to induce
labor for medical reasons or to perform a cesarean section.
Women experiencing these outcomes would be inappropriately
excluded from comparison group (a). All of the women not
undergoing elective induction of labor must be retained in the
comparison population so as to obtain a valid assessment of the
experience associated with this alternative course of action.
This has been noted by Caughey et al.,27 who also provided
some illustrative data on the occurrence of cesarean section:
Induction of labor at % of women delivered

39 weeks?* by cesarean section
Yes 14.3%
No
Delivered at 39 weeks 9.1%
Delivered after 39 weeks 15.0%
*These generally were not elective inductions, but the presence of an indication
for induction was adjusted for in the analysis.
The use of noninduced women who delivered at 39 weeks as

a basis for the expected incidence of cesarean section would
incorrectly suggest that induction of labor at 39 weeks predis-
posed to this outcome.
The didactic message: Comparison groups in cohort studies
should not be subject to selection on the basis of events that
occur after the time that the exposure has been sustained.
CHAPTER 5
Case-Control Studies
IN CASE-CONTROL STUDIES, persons with a given illness

or injury (cases) are characterized as to whether or not they
previously had sustained a given exposure (and degree of that
exposure). Ideally, as a basis for comparison, we would like
to determine the probability and degree of exposure in a sample
of a population from which the cases were drawn. This ideal can
be hard to achieve in practice, and it is common for the results
of case-control studies to be ambiguous in their interpretation.
Nonetheless, because case-control studies potentially are able
to provide valid results, and because there are instances
in which the only data relevant to a particular etiologic relation
can be obtained from case-control studies, they are, warts and
all, an essential item in the epidemiologists toolbox.
CASE-CONTROL STUDIES | 149

The fact that several relatives sometimes stutter has led others
to assert that stuttering is inherited. Yet in forty years of expe-
rience I have met more stutterers who had no close relatives
or ancestors who stuttered than those who had. It is my obser-
vation that most stutterers are hypersensitive persons, and
I believe that hypersensitivity is acquired at an early age
through the childs environment.
Does the information in the letter address the possible

association between positive family history of stuttering and
stuttering itself? If yes, how? If no, why not?
Answer 5.1 Even though only a minority of stutterers have

a family history of this condition, that proportion may still
be considerably in excess of the proportion present among
nonstutterers. In the absence of data on the latter, no conclu-
sions can be drawn regarding the presence or absence of an
association.
Question 5.2 In a case-control study of pneumonia in infants

in southern Brazil, the mothers of 152 cases and 2,391 controls
were interviewed. A far higher proportion of cases than con-
trols were not being breast-fed during the week prior to the date
of onset of the cases illness (and the corresponding date for
controls): odds ratio = 17, 95% confidence interval = 7.736.0.
A commentator on this study indicated that a serious con-
cern is the relatively small sample size, with the result that very
few cases were exclusively breast fed. Do you agree that this
is a serious concern when interpreting the results of the
study? If yes, why? If no, why not?
Answer 5.2 No. The lower confidence limit of the odds ratio
associated with not breast-feeding is 7.7; thus, there is little
possibility that chance is solely responsible for the association.
There would be concern over the sample size only if it were
important to know more precisely where, within the range
of 7.736.0, the true odds ratio lies.
Question 5.3 You are planning to conduct a case-control study

that would examine a possible association between a genetic
characteristiccalled T1and the incidence of lung cancer.
About 10% of the population possesses T1. In this study, you
have two laboratory methods that potentially could be used
to measure a persons T1 status from a sample of his/her DNA.
Neither is perfect. Method A will correctly categorize everyone
who truly is positive for T1, but is expected to misclassify 5%
of truly negative persons as being T1-positive. Conversely,
method B will correctly classify everyone who is truly
T1-negative, but will misclassify 5% of T1-positive persons
as T1-negative. Which is the better choice of tests to minimize
bias in the odds ratio relating T1 status and lung cancer?
a. Method A
b. Method B
c. Neither is a better choice than the other.

Answer 5.3 Method B will provide a less biased result, since

its use will result in a relatively smaller proportion of misclassi-
fied individuals. As an example:
True T1 status Cases Controls Odds Ratio
+ 200 100
2.25
800 900
Using Method A, 5% 800 = 40 cases truly negative for T1
would be labeled as T1positive, as would 5% 900 = 45
controls.
Method A: Observed T1 status
+ 200 + 40 100 + 45
1.64
800 40 900 45
Using Method B, 5% 200 = 10 cases truly negative for T1,
would be labeled as T1-negative, as would 5% 100 =
5 controls.
Method B: Observed T1 status
+ 200 10 100 5
1.99
800 + 10 800 + 5
Question 5.4 The following is excerpted from an article

on potential risks associated with spray painting in the auto-
mobile industry28:
In the case-control analyses, cases are defined as all lung cancer

deaths among the automotive workers (n = 263). Controls are
defined as those deaths due to circulatory disease or to acci-
dents among those same workers, thus ensuring a valid repre-
sentation of the population under investigation.
In this study, a history of spray painting was ascertained

in cases and controls from company records.
a. What population is the one for which we would like to

know the proportion of employees who worked as spray
painters?
b. Apart from chance, how might the control group selected
not accurately characterize the proportion exposed
in that population?
Answer 5.4
a. The population is that of automotive workers at risk of

death in the period during which case accrual took place.
b. If work as a spray painter is associated with the rate
of death of the two causes selectedpositively or nega-
tivelythen the proportion of spray painters in this con-
trol group will not be representative of that in the
underlying population at risk.
Question 5.5 At the sexually transmitted infection (STI) clinic

where you work, you are planning a case-control study of acute
gonorrhea in men in relation to condom use. Because the cases
of gonorrhea seen at the clinic are not derived from any defined
population, you consider the use of male patients with an STI
other than gonorrhea as a sampling frame for controls. This
choice has the advantage of being feasible, and it is likely that
the accuracy of information obtained on recent condom use
would be similar between cases and controls. What do you
believe to be the primary threat to the validity of this study
if you were to select as controls men seen at the clinic for an STI
other than gonorrhea?
Answer 5.5 The primary concern here is selection bias.

Condom use (or the lack thereof) plausibly has a similar favor-
able impact on the incidence of both gonorrhea and other STIs.
Therefore, the prevalence of condom use in STI controls may
be considerably lower than in men in general. The use of such
a control group could lead to a spuriously low estimate of the
efficacy of condom use in the prevention of gonorrhea.
Question 5.6 This question is based on the following abstract.29
Study objective. To investigate the association of alcohol use

and night driving with traumatic snowmobile fatalities.
Design. Case-control study.
Participants. Traumatic deaths occurring while driving a
snowmobile during the years 1985 to 1990 were reviewed. A
sample of 1989 to 1990 fatal motor vehicle driver and motor-
cycle driver accidents were used as controls. Records were
obtained from the provincial coroner.
Results. One hundred eight snowmobile fatalities, 432
motor vehicle fatalities, and 108 motorcycle fatalities were
included. Young men (mean age, 30 years) made up the snow-
mobile fatalities population, with weekend fatalities predomi-
nating (67%). Snowmobile fatalities were associated with use
during times of suboptimal lighting (crude odds ratio, 1.9
[95% confidence interval, 1.13.3]; P <.01). Blood alcohol con-
centration exceeded provincial limits in 64% of cases. When
snowmobile fatalities were adjusted for occurrence during
suboptimal lighting conditions, only alcohol use was associ-
ated independently with fatal outcome (adjusted odds ratio,
4.3 [95% confidence interval, 2.57.0]; P <.0001).
Conclusion. Drivers in snowmobile fatalities are associ-
ated with an approximately fourfold greater use of alcohol
than are age- and sex-matched drivers in automobile and
motorcycle fatalities.
Do you believe the control group chosen for this study led
to bias in the estimate of the size of the association between
fatal snowmobile trauma and alcohol use? If yes, why, and would
an unbiased estimate be greater or smaller than that obtained
by the authors? If not, why not?
Answer 5.6 The control group for this study is comprised

of persons with alcohol-related causes of death, in other words,
those due to motor vehicle and motorcycle accidents. Therefore,
the observed odds ratio is almost certainly lower than the true
one.
Question 5.7 Please answer the questions below after reading

the following meeting abstract (modified).
Design. Population based case-control study of drivers with
known drug and alcohol concentrations who were
involved in fatal crashes from October 2001 to September
2003. The cases were the 6766 drivers considered at fault
in their crash; the controls were 3006 other drivers.
Results. 681 drivers were positive for cannabis (cases 8.8%,
controls 2.8%) (odds ratio 3.32, 95% confidence interval
2.63 to 4.18).
Conclusions. Driving under the influence of cannabis
increases the risk of involvement in a crash.
a. Blood specimens were readily available both for cases and

controls, and obtaining such specimens on a representative
sample of drivers would have posed formidable logistical
difficulties. Nonetheless, this choice could have led to a
result that was biased to at least some extent.
i. Under what circumstance(s) would bias arise?
ii. In what direction would the bias likely be operating, to
falsely increase or falsely decrease the odds ratio? Why?
b. In this study, the population attributable risk % (PAR%,
the percentage of this populations incidence of a fatal
automobile crash attributable to the exposure in ques-
tion) associated with positive detection of cannabis
in blood was 6.1%. The corresponding PAR% for positive
detection of alcohol in blood was 28.6%, despite the fact
that the proportion of controls in whom each substance
was detected was identical. What must be the explana-
tion for the disparity between the size of the PAR% for
cannabis and that for alcohol?
Answer 5.7
a. Bias would be present to the extent that the blood levels

of tetrahydrocannabinol in drivers killed in auto crashes,
who were judged not to be at fault in the crash, did not
reflect that of drivers in general. This could occur if
either one or both of the following were true:
The designation of at fault sometimes was in error,
so that true cases were intermixed into the control
group, giving that group a spuriously high proportion
of apparent cannabis users.
Even if not at fault, drivers who had consumed can-
nabis were impaired in their ability to avoid a crash.
In either circumstance above, the observed odds ratio asso-

ciated with evidence of cannabis consumption would be spuri-
ously low.
b. The odds ratio associated with alcohol consumption

must have been considerably higher than that for
cannabis consumption. (In fact, it was higher:
15.5 vs. 3.3.)
Question 5.8 This question pertains to the following

30
abstract :
Acute influenza infection may be transiently associated with

the risk of cardiovascular disease. We examined the associa-
tion between influenza vaccination and incident myocardial
infarction (MI) and stroke in a population-based case-control
study. Case subjects were members of Group Health Cooperative
(GHC) with incident MI or ischemic stroke during flu season
(November-March) of 19921998. Control subjects were
GHC members without history of MI or stroke who were fre-
quency matched to case subjects by age, sex, and calendar
year. The medical records of 584 case subjects with MI,
269 case subjects with ischemic stroke, and 1,415 controls
were reviewed. Receipt of each years influenza vaccine was
not associated with risk of incident MI (odds ratio [OR] = 0.95,
95% confidence interval [CI]: 0.77, 1.17) or ischemic stroke
(OR = 1.20, 95% CI: 0.91, 1.60) during the period of expected
influenza activity. This study suggests that . . . influenza vac-
cination is not associated with a reduction in risk of first MI
or ischemic stroke.
By restricting cases of MI and stroke to those that occurred

during November-March, the investigators obtained a sample
size that was considerably smaller than the one that would have
included persons diagnosed with these illnesses in other months
as well. What do you believe to have been the primary compen-
sating advantage of the choice they made?
Answer 5.8 It was hypothesized that vaccination might pre-

vent cases of MI and stroke that were precipitated by an influ-
enza infection. Such infections occur primarily in the winter
months. The ability of the study to find a true association would
have been lessened if MI and stroke cases that were not precipi-
tated by influenza, and thus did not have the potential to
be prevented by vaccination, were included. That is, if there
truly had been a beneficial impact of vaccination on the risk
of MI and/or stroke, the observed odds ratio, using cases in all
months, would have been closer to the null than that based
on cases diagnosed in November-March.
Question 5.9 The following is excerpted from an article

on mesothelioma in relation to employment 31:
In a case-control study, the occupational exposures of 259 meso-

thelioma patients were compared to those of an equal number of
controls. Several occupations known to entail substantial expo-
sure to asbestos were more common among cases than controls.
Longest held occupation Cases Controls
Insulator 47 13
Shipbuilder 31 21
Plumber 35 28
Furnace or boiler
installer or repairman 21 10
However, there were an identical number of cases and controls (15)

who had engaged in brake lining work or repair, indicating that no
increase in risk was associated with this type of employment.
Assume the following in this study:
a) Employment status was ascertained without error.

b) Cases and controls were completely comparable with regard
to nonoccupational determinants of mesothelioma.
c) Among persons whose longest-held job was not one
of the four listed in the above table, those who did brake
lining work or repair were no more likely than other
persons to also have engaged in one of those four occupa-
tions at some time in their lives.
d) Sampling variability is not an issue.
Do you agree with the authors conclusion above regarding

the relation of mesothelioma occurrence to brake lining work
or repair that was observed in their study? If yes, why? If no,
why not?
Answer 5.9 No. To evaluate employment in brake lining work

or repair as a possible risk factor for mesothelioma, it is neces-
sary to estimate the mesothelioma incidence in persons with
that exposure relative to the risk in a referent category com-
prised of persons not believed to be at increased risk of this
disease. Thus, the analysis should be restricted to persons who
had not been employed in furnace and boiler work, insulation,
and so forth. If persons who had been insulators, and so forth,
had been excluded from the analysis, the referent category
would be reduced by 47 + 31 + 35 + 21 = 134 cases and by
13 + 21 + 28 + 10 = 72 controls. The odds ratio associated with
employment in brake lining and repair would be as follows:
Brake work Cases Controls
+ 15 15
259 15 134 = 110 259 15 72 = 172
15 15
Odds ratio = = 1.56
110 172
The authors interpretation of the results of their study
is incorrect.
(It turns out that assumption (c) above is almost certainly
not a valid one, i.e., a relatively high proportion of men whose
longest held job involved brake lining work or repair indeed had
been employed at other times in a high risk occupation. When
in a reanalysis attention was restricted to men who had no such
history,34 no greater proportion of mesothelioma cases (1/33)
than controls (9/171) had been employed to do brake work.)
Question 5.10 Epidemiologists in California observed that

about 65% of infants who died of sudden infant death syn-
drome (SIDS) typically were put to sleep in the prone position,
in contrast to 60% of control infants. Most other studies of
SIDS and sleeping position have found a considerably greater
case-control difference. Commenting on this result, an editori-
alist wrote, One reason that prone sleeping may not have been
observed as a strong risk factor for SIDS [in this study] is that it
is difficult to measure risk for a characteristic present in 60
percent of the population. Do you agree with this assertion?
If yes, why? If not, why not? If not, why not?
Answer 5.10 No, you do not agree.

In terms of the frequency of the exposure, the only circum-
stance in which case-control studies find it is difficult to mea-
sure risk is when the exposure either is extremely common
(e.g., > 95%) or extremely uncommon (e.g., < 5%). (In those
instances, study power is reduced for a given sample size.)
An exposure frequency of 60% is nowhere near these
values.
Question 5.11 You have been sent a manuscript to review.

It describes the results of a case-control study of gastric ulcer
in relation to prior use of nonsteroidal anti-inflammatory drugs
(NSAIDS). Cases in this study were all 225 patients diagnosed
with a first gastric ulcer between 1982 and 1985 in a given pop-
ulation. All diagnoses were made by endoscopy that had been
performed at specialized centers for this purpose. It is believed
that gastric endoscopy was not done elsewhere in that part
of the world. Community controls were selected randomly from
electoral rolls for 19821985. After conducting a screening
interview with potential control subjects (matched to cases
on the basis of sex, age, and area of residence), those with a his-
tory of gastric ulcer were excluded. In addition, about 13%
of potential controls with dyspepsia were excluded as well: The
authors were concerned that: (a) some persons with this symp-
tom might have a gastric ulcer that had not yet been diagnosed;
and (b) because NSAID use is a likely cause of dyspepsia, the
occurrence of dyspepsia would be considerably more common
in NSAID users than in other persons.
Assume that only a very small percentage of those poten-
tial controls with dyspepsia actually had a gastric ulcer. Do you
agree with the authors choice to exclude persons with dyspep-
sia? If yes, why? If no, why not?
Answer 5.11 You disagree. The controls should be a sample

of the at-risk population. Cases were included whether or not
they had dyspepsia, so persons with dyspepsia were part of the
at-risk population. By restricting controls to persons free
of dyspepsia, the proportion of persons with a history of use of
NSAID would be spuriously low. This would lead to an overesti-
mate of the odds ratio relating a history of NSAID use to the
occurrence of gastric ulcer.
Question 5.12 The following is an excerpt from an abstract

of a case-control study 32:
Experimental and epidemiologic evidence has suggested that

phenacetin use increases the risk of transitional cell cancers
of the urinary tract. The drug is no longer marketed but a com-
monly used metabolite, acetaminophen, has been linked
recently to an increased risk of renal cancer. We assessed the
relation of acetaminophen use to the risk of transitional cell
cancer of the urinary tract with data from a hospital-based
study of cancers and medication use conducted from 197696
in the eastern United States. We compared interviews with
498 cases of transitional cell cancer with those of 8,149 non-
cancer controls, and controlled confounding factors with logis-
tic regression. For transitional cell cancer, the relative risk (RR)
estimate for regular acetaminophen use that had begun
at least a year before admission was 1.1 (95% confidence inter-
val (CI = 0.61.9). RR estimates for use that lasted at least five
years, and for non-regular use, were also close to 1.0. Our
results suggest that acetaminophen, as used in present study
population, does not influence the risk of transitional cell
cancer of the urinary tract.
The authors stated that, in choosing noncancer controls

they included only persons hospitalized for conditions that
were judged to be unrelated to acetaminophen use. For exam-
ple, we did not include patients admitted for gastric or duode-
nal ulcers, because such persons might have used acetaminophen
preferentially to aspirin for pain relief. Do you agree with this
strategy? Why? If not, why not?
Answer 5.12 Yes, the strategy is reasonable. By not including

patients who were hospitalized because of a condition that is
an indication for acetaminophen use, the authors assembled
a group of hospitalized patients whose exposure history possi-
bly could reflect that of the source population of the cases.
However, it is plausible that a relatively high proportion of
hospitalized persons have a history of analgesic use, whether
aspirin or acetaminophen. If this is true, the proportion of con-
trols selected for this study who previously had taken acetamin-
ophen would exceed that of the at-risk population, leading to
a falsely low odds ratio.
Question 5.13 The following question is based on an excerpt

of the abstract of an article Patent foramen ovale and crypto-
genic stroke in older patients33:
We prospectively examined 503 consecutive patients who had

had a stroke, and we compared the 227 patients with crypto-
genic stroke and the 276 control patients with stroke of known
cause. We examined the prevalence of patent foramen ovale
in all patients, using transesophageal echocardiography.
The prevalence of patent foramen ovale [a congenital
heart defect] was significantly greater among patients with
cryptogenic stroke than among those with stroke of known
cause, for both younger patients (43.9% vs. 14.3%; odds ratio,
4.70; 95% confidence interval [CI], 1.89 to 11.68; P<0.001)
and older patients (28.3% vs. 11.9%; odds ratio, 2.92; 95% CI,
1.70 to 5.01; P <0.001).
In theory, the control group against which cases of crypto-

genic stroke ought to be compared for the prevalence of patent
foramen ovale is a sample of persons who are demographically
similar to the cases but otherwise unselected. Do you believe
that the control group actually chosen in this study correctly
produced a positive association between patent foramen ovale
and the occurrence of cryptogenic stroke? If yes, why? If no,
why not?
Answer 5.13 The control group usedpatients with a stroke

of known causewill provide a valid result to the extent that
the prevalence of patent foramen ovale (PFO) in these persons
reflects that of the underlying population from which the cases
of cryptogenic stroke arose. If the attribution of a known cause
had been correct, then a valid result likely would have been
obtained, given that there is no reason to believe that the pres-
ence of PFO influenced the development of a stroke in these
persons. To the extent that some of the cases with a known
cause were in truth cryptogenic ones, then the observed asso-
ciation between PFO and cryptogenic stroke actually would be
spuriously small. (However, the prevalence of PFO in the con-
trols chosen11.9%14.3%was not elevated relative to that
expected based on autopsy studies.)
CHAPTER 6
Multiple Causal
Pathways and Effect
Modification
AN IMPORTANT THREAT to an epidemiologic studys sensitiv-

ity in identifying a genuine exposuredisease association is error
in the measurement of the exposure and/or the illness outcome.
Confounding can be another threat. A third, not as widely appre-
ciated as measurement error and confounding, is a studys fail-
ure to take into account the presence of causal pathways leading
to disease other than the one under consideration. The means
by which these other pathways can be accommodated vary
from study to study. Sometimes, illness outcomes can be subdi-
vided based on the presence or absence of a manifestation of the
condition in question, for example, estrogen receptor-positive
breast cancer from estrogen receptor-negative breast cancer in
studies of hormonal exposures. Alternatively, persons in whom
a known potent etiologic agent is present can be omitted from
consideration, so as to allow the influence of an exposure acting
to cause disease through a different means to be seen (e.g., in
studies of mental retardation, when children with microcephaly
are excluded when studying the potential influence of postnatal
exposure to low levels of metals in the environment). However,
interpreting the results of analyses based on subgroups of the
study population is not always straightforwardas well see in
some of the following examples.
MULTIPLE CAUSAL PATHWAYS | 177
Question 6.1 You come across an abstract of an article in a

medical journal, which in part reads as follows:
In our study of women diagnosed with endometrial cancer,

possible risk factors were identified through personal inter-
views. Among women with endometrial cancer who had and
had not taken unopposed estrogen therapy prior to the time
the tumor was diagnosed, there were no differences with
regard to parity [i.e., the number of children they had borne].
The results of this study do not support the hypothesis that
parity is differentially associated with endometrial cancer that
is and is not related to exogenous estrogens.
Based on the above summary, what do you believe to be the

main limitation of this study in evaluating whether parity has a
different relationship to the etiology of endometrial cancer that
is and is not related to unopposed estrogen use?
Answer 6.1 The differential relationship between a particular

characteristic of a womans reproductive history and endome-
trial cancer, according to a history of estrogen use, can be dis-
cerned only by measuring the size of the association between
that characteristic and endometrial cancer in estrogen users
and then again in nonusers. Since no comparison group is men-
tioned, the presence of an association cannot be assessed in
either group of women.
Question 6.2 Persons with a factor V Leiden mutation are

resistant to the anticoagulant effect of activated protein C.
The following table describes the incidence of first venous
thrombosis (VT) in women ages 15 to 49 years, according to
presence of the factor V Leiden mutation and the use of oral
contraceptives (OC):
Incidence of
VT per 10,000
Cases of VT Person-years person-years
Factor V Leiden negative

No OC use 36 437, 870 0.8
Current OC use 84 275, 858 3.0
Factor V Leiden positive
No OC use 10 17, 515 5.7
Current OC use 25 8757 28.5
Assume that the incidence of first and recurrent VT bear a

similar relation to oral contraceptive use and the factor V Leiden
mutation. In a 15- to 49-year-old woman who develops VT, do
the data presented above argue that her factor V Leiden status
should be considered in counseling about her future method of
contraception?
Answer 6.2
Incidence of Rate difference

Factor VT per 10,000 per 10,000
V Leiden person-years Rate ratio person-years
Negative
No OC 0.8
Current OC 3.0 3.7 2.2
Positive
No OC 5.4
Current 28.2 5.0 22.8
Yes. Although OC use is associated with an increased inci-

dence of VT regardless of factor V Leiden mutation status, the
recommendation is supported by the considerably larger abso-
lute increase in the rate of VT associated with OC use in factor
V Leiden-positive than in factor V Leiden-negative women: 22.8
per 10,000 person-years versus 2.2 per 10,000 person-years.
Question 6.3 A study observed that, for a given history of cig-

arette smoking, the relative risk for lung cancer among women
was 1.2- to 1.7-fold greater than it was in men. The disparity
between the sexes was well beyond that expected by chance, and
led the authors to conclude that women were more susceptible
than men to respiratory carcinogenesis from cigarette smoke.
A subsequent letter to the editor of the journal in which the
article was published suggested that in order to use the data
from the study to infer differential susceptibility between the
sexes, it would be necessary to know the incidence of lung
cancer in men who had not smoked and in women who had not
smoked. Why would this additional information be useful?
Answer 6.3 If the rate in female nonsmokers were lower than

in male nonsmokers, the same added risk associated with ciga-
rette smoking in the two sexes would produce a larger relative
risk in women.
For example:
Incidence in Incidence in Attributable Relative risk

nonsmokers* smokers* risk*
Men 10 30 20 3
Women 5 25 20 5
*Rate per 100,000 person-years.
Therefore, depending on how differential susceptibility is

definedgreater relative risk versus greater attributable risk
women and men may or may not differ in terms of the impact
of smoking on their incidence of lung cancer.
Question 6.4 This question pertains to the following excerpt

of an abstract35:
Background. Elderly people who have a fracture are at high

risk of another. Vitamin D and calcium supplements are
often recommended for fracture prevention. We aimed
to assess whether vitamin D3 and calcium, either alone
or in combination, were effective in prevention of sec-
ondary fractures.
Methods. In a factorial-design trial, 5,292 people aged
70 years or older (4481 [85%] of whom were women)
who were mobile before developing a low-trauma frac-
ture were randomly assigned 800 IU daily oral vitamin
D3, 1000 mg calcium, oral vitamin D3 (800 IU per day)
combined with calcium (1000 mg per day), or placebo.
Participants who were recruited in 21 UK hospitals were
followed up for between 24 months and 62 months.
Analysis was by intention-to-treat and the primary out-
come was a low-energy fracture.
In their trial, the investigators identified 698 participants

in whom a low-energy fracture occurred. Not included in the
analysis were an additional 34 fractures that resulted from sub-
stantial trauma, for example, a fracture sustained in an auto-
mobile crash.
The disadvantage of not including the 34 fractures involv-
ing more than a low level of trauma is a reduced sample size
(by about 5%). What do you believe to be the main advantage
of this choice?
Answer 6.4 The investigators believed that the relative influ-

ence of one or more of the intervention measures on the risk of
fracture could differ depending on the level of trauma to which
a participant was exposed. If very little trauma were present,
the relative benefit might be great; in the presence of substan-
tial trauma, there might be but little benefit from the interven-
tion. Therefore, to maximize the sensitivity of the study to
observe any relative change in fracture occurrence associated
with one or more treatments, the analysis excluded persons
with fracture in whom another factor, substantial trauma, likely
played a causal role.
Question 6.5 You observe that among persons 20 years of age

or older who developed a particular infectious disease, 32% had
been vaccinated against that disease, in contrast to 16% of per-
sons under 20 years with the disease. Is this necessarily evidence
of greater vaccine efficacy in younger persons?
Answer 6.5 No, such a conclusion would not be justified.

In order to assess vaccine efficacy in either age group, it would
be necessary to estimate the percentage of the population at
risk who had been vaccinated.
<20 20
Vaccinated Cases Population Cases Population
Yes 32% ? 16% ?

No 68% ? 84% ?
Question 6.6 In a study of the possible adverse effect of the

drug rosuvastatin on the incidence of venous thromboembo-
lism (VTE),36 the investigators separately analyzed provoked
and unprovoked cases. Provoked cases were defined as those
in which a strong risk factor had been present, such as recent
surgery, immobility or metastatic cancer. What do you believe
to have been the rationale for making this separation?
Answer 6.6 If rosuvastatin acted together with other VTE risk

factors in a causal pathway, the size of the association might
have been expected to be particularly great for provoked cases.
Alternatively, if the means by which rosuvastatin predisposed
to VTE were independent of other risk factors, the relative risk
associated with its use would be expected to be greatest for
unprovoked VTE. An analysis that examines each type of VTE
separately enables the evaluation of either possibility.
Question 6.7 The data presented in the table below appeared

in a publication of the results of a case-control study on the risk
of endometrial cancer in relation to physical activity and
obesity.37 Based on the data in the table, the authors stated that
the increase in risk associated with obesity [BMI 30] was
much lower in active women (OR = 1.57) than in women with
low physical activity (OR = 3.10). What is a more accurate way
of quantifying the difference in the relative risk of endometrial
cancer associated with BMI 30 within the two categories of
physical activity? Why might the approach used by the authors
be misleading?
Lifetime physical activity and

endometrial cancer risk, by BMI
Average lifetime physical activity
Low High
OR OR
BMI Cases Controls (95% CI) Cases Controls (95% CI)
<25 105 100 1.0 59 74 0.95

(0.571.58)
2529 81 72 1.0 52 84 0.58
(0.340.99)
30 113 49 1.0 62 64 0.50
(0.280.91)
<25 105 100 1.0 59 74 0.90
(0.561.43)
2529 81 72 1.39 52 84 0.85
(0.882.20) (0.521.38)
30 113 49 3.10 62 64 1.57
(1.915.01) (0.942.62)
*Below and above the median.
Odds ratio (OR) and 95% CI were adjusted for age, race/ethnicity, education, family
history of endometrial cancer, age at menarche, full-term pregnancies, duration of oral
contraceptive use, duration of hormone therapy use, menopausal status, and height.
Answer 6.7 We wish to know whether the size of the OR

associated with a high BMI on risk differs according to physical
activity. Among less active women, the OR associated with obe-
sity is 3.1. Among more active women, it is 1.57/0.90 = 1.74.
(The OR of 1.57 incorporates both the elevated risk associated
with obesity and the reduced risk associated with activity.)
Question 6.8 In a cohort study, postmenopausal women who

were taking combined estrogen-progestogen hormone therapy
at baseline had a 5-year cumulative incidence of breast cancer
of 5.6/1,000 if their mother had a history of breast cancer,
and 2.2/1,000 if not. Among hormone nonusers, the corre-
sponding cumulative incidences were 5.1/1,000 and 1.7/1,000,
respectively.
a. Among women with a maternal history of breast cancer,

what was the relative risk of breast cancer associated with
use of hormone therapy at baseline? The risk difference?
b. What were these same measures of excess risk in women
without a maternal history of breast cancer?
c. The two means of assessing the potentially modifying
influence of maternal breast cancer on the association
between hormone use and risk of breast cancer do not
produce the same qualitative result. Why is this?
You are trying to decide whether to differentially counsel

postmenopausal women with and without a maternal history
of breast cancer with regard to the impact of hormone therapy
(perhaps in terms of the frequency of breast screening exams).
Which contrast do you believe to be the most relevant for this
purpose, the one between the two relative risks or that between
the two risk differences? Explain.
Answer 6.8 a.
Risk difference
Maternal history Relative risk (per 1,000)
+ 5.6 per 1000/5.1 5.6 5.1 = 0.5

per 1000 = 1.1
2.2 per 1000/1.7 2.2 1.7 = 0.5
per 1000 = 1.3
b. In this instance, the same absolute increase in risk

associated with hormone therapy (0.5/1,000) is acting
upon a smaller background risk in women without a
positive maternal history, producing a greater relative
change in their risk (1.3) than the relative change (1.1) in
women with a positive maternal history.
c. The size of the absolute change in risk associated with
hormone therapy is what bears on personal decision
making. Because the additional risk associated with hor-
mone therapy is 0.5/1,000 irrespective of maternal his-
tory, counseling should not be differential between the
two groups of women with regard to actions that might
be taken.
Question 6.9 A study was conducted in sub-Saharan Africa

among heterosexual couples in whom just one member was
infected with HIV.38 During the course of a 24-month follow-up,
about 10% of participants who were infected at baseline initi-
ated antiretroviral therapy. The study observed that among per-
sons who were HIV-negative at baseline, the acquisition of an
HIV infection that was phylogenetically linked to that of their
HIV-infected partner was relatively less common among those
whose partners had received antiretroviral therapy (0.37 vs. 2.24
per 100 person-years).
Had the study endpoint been any new HIV infection, irre-
spective of its linkage to the HIV strain of the initially infected
partner (so as to include HIV infections that arose from sexual
contact with other persons), what would have been the expected
impact (if any) on the following measures of association:
a) Rate difference
b) Rate ratio
Explain.
Answer 6.9 If receipt of antiretroviral therapy by ones HIV-

infected partner were unrelated to the likelihood of acquiring
an HIV infection from a different partner, then the size of the
absolute increase in the incidence of any new HIV infection
above that of the linked infections would be the same whether
or not antiretroviral therapy had been administered. In this cir-
cumstance, the rate difference would be unchanged, but the
rate ratio associated with the partners receipt of antiretroviral
therapy would be closer to the null.
For example, assume an incidence of unlinked HIV infec-
tion in initially HIV-negative persons of 0.20 per 100 person-
years:
Antiretroviral Rate
treatment difference
Type of HIV (per 100 Rate
infection Yes No person-years) ratio
Linked 0.37 2.24 1.87 6.1

Unlinked 0.20 0.20 0 1.0
Total 0.57 2.44 1.87 4.3
The presence of a second causal factor leading to disease

(in this instance, sexual contact with another HIV-infected
person)one that does not join with the first factor in an etio-
logic pathway and is not accounted for in the study design and/
or analysiswill lead to an attenuation of the estimated rate
ratio of the disease associated with the first causal factor.
Question 6.10 The following letter, somewhat paraphrased,

appeared in a medical journal:
Bailar and Gornik report that the age-adjusted rate of mortal-

ity from all cancers in the United States declined by 1 percent
from 1991 through 1994. Our estimate for the same interval is
2.2 percent. The discrepancy in the two figures stems from the
use of different populations for age adjustment. Bailar and
Gornik used the relatively elderly 1990 U.S. population and by
doing so, minimized striking reductions in mortality that
occurred among young and middle-aged persons. We used the
U.S. (relatively younger) 1940 population, which reveals the
full downturn in cancer-related mortality.
Suppose that you also are interested in quantifying the

change in U.S. cancer mortality during 1991-1994, and have
access to age-specific U.S. mortality rates for those years. In
your analysis, would there be any virtue in presenting adjusted
rates within specific age categories (e.g., young and middle-
aged versus older)? If yes, why? If no, why not?
Answer 6.10 Yes, there would be value in presenting age-

specific rates. During the 4-year period under study there was,
apparently, a striking reduction in cancer mortality in the
United States in young and middle-aged persons, but not in
older persons. Only the presentation of age-specific differences
over time can illustrate this difference in trends. A measure
such as a difference in adjusted rates across all ages will provide
a summary that may not apply to any individual age group.
Furthermore, as is clear from the letter to the editor, the size of
the mortality trend will be influenced by the arbitrary choice
of the age distribution to be used as the standard.
Question 6.11 The following is an abstract of an article

Occupational asbestos exposure and the incidence of non-Hodg-
kin lymphoma of the gastrointestinal tract: An ecologic study.39
Purpose. A previous case-control study observed a strong asso-

ciation between occupational exposure to asbestos and the
incidence of non-Hodgkin lymphoma of the gastrointesti-
nal tract (GINHL). To test this hypothesis we sought to
determine whether the geographic pattern of the incidence
of GINHL in the US has paralleled that of mesothelioma.
Methods. Using data obtained from the nine US regions par-
ticipating in the National Cancer Institutes Surveillance,
Epidemiology, and End Results program, we examined
the incidence of malignancies among men ages 50 to 84
years between 1973 and 1984.
Results. The rates of mesothelioma, but not of GINHL, were
about two times higher in the areas of Seattle and San
Francisco than in the other regions. Overall, there was
no correlation between the rates of mesothelioma and of
GIHNL (Person correlation coefficient0.12, = 0.77).
Conclusions. This ecologic study finds no support for the
hypothesis that occupational asbestos exposure is related
to the subsequent incidence of GINHL.
In their analysis, the authors paid particular attention to

rates in men (in whom the likelihood of prior occupational
exposure was far greater than in women) and to rates in 50- to
84-year-olds (to allow for a potentially long induction period).
And, even though the Surveillance, Epidemiology, and End
Results program had data available through the 1990s, the
authors confined their analysis to cancer incidence through just
1984. What do you believe to have been their reason for this
latter choice? (Hint: The presence of HIV infection strongly
predisposes to the development of GINHL.)
Answer 6.11 GINHL has a number of causal pathways that

may lead to its occurrence. One of these involves HIV infection.
If occupational asbestos exposure adds to a persons risk of
GINHL to the same extent whether HIV infection is present or
not, then the relative increase in risk associated with asbestos
exposure will be greatest in HIV-uninfected persons. Therefore,
the most sensitive assessment of the potential role of asbestos
is to exclude cases related to HIV. One way of accomplishing
this is to restrict the time period being considered to that before
HIV infection was widespread, in other words, prior to 1985.
(Because the prevalence of HIV infection varies geographi-
cally across the United States, failure to control for HIV infec-
tion in this way also could lead to confounding. For example,
GINHL rates in San Francisco during the last decade of the
twentieth century might be high relative to other parts of the
United States because of the relatively high prevalence of HIV
infection there, and not because of a higher degree of occupa-
tional asbestos exposure.)
Question 6.12 A study sought to determine if pregnancy

intendedness is associated with intimate-partner physical vio-
lence, and to identify factors that modify this association. Three
to 6 months after delivery, the investigators mailed a question-
naire to a population-based sample of 12,612 mothers of infants
born in four states.
Some of the results of this study are shown in the following
table:
Percentage of women experiencing physical

violence during pregnancy, by pregnancy
intendedness and education
Unwanted Intended
Education (y) % (95% CI) % (95% CI)
All 12.1 (8.815.6) 3.2 (2.44.0)

<12 18.6 (9.627.6) 7.0 (4.19.9)
12 10.6 (5.915.3) 3.7 (2.54.9)
>12 9.2 (2.715.7) 1.3 (0.71.9)
The authors stated that the risk for experiencing violence in

women who had an unwanted pregnancy, relative to the risk in
women with an intended pregnancy, was particularly high
among women with more social advantage. For example, the
relative prevalence among women with fewer than 12 years of
education was 2.6, whereas the corresponding relative preva-
lence in women with >12 years of education 7.1.
In discussing the findings, the authors put forth some
possible explanations for the interaction between pregnancy
intendedness, social status, and physical violence. Could it be
argued that, apart from the issue of chance (i.e., sampling vari-
ability), there is no interaction to account for? If yes, why? If no,
why not?
Answer 6.12 The investigators used a relative measure of

association to assess the possibility of an interaction between
education and pregnancy intendedness as a predictor of physi-
cal violence. However, the data suggest that, in absolute terms,
the increased risk of experiencing violence associated with an
unwanted versus an intended pregnancy is about the same for
all levels of education. If anything, the observed difference in
prevalence was greater in women with <12 years of education
(18.6% 7.0% = 11.6%) than in women with >12 years of edu-
cation (9.2% 1.3% = 7.9%).
Question 6.13 The data presented in the following table are

from a hypothetical study of perinatal death among twins in
relation to order of delivery.*
No. of perinatal deaths in:

Gestational No. of twin first-born second-born
age (weeks) pregnancies twin (a) twin (b) b/a
2427 703 359 381 1.06

2831 1,371 98 141 1.44
3235 2,897 74 164 2.22
36 2,935 33 124 3.76
Would it be reasonable to infer from these results that the

increase in risk of perinatal death in the second-born twin is a
particular concern in a term or a near-term pregnancy, and less
so in a pregnancy that ends well prior to term? Explain.
* Infants born to women undergoing a planned caesarean section are excluded.

Restricted to pairs in which at least one twin survived.
Answer 6.13 At gestational ages 2427 weeks, the risk of per-

inatal death in first-born twins is 359/703 = 51/100, and that
in second-born twins is 381/703 = 54/100. The difference in
risk of death between second- and first-delivered twins at other
gestational ages also is about 3 per 100. Therefore, although the
relative difference in risk of death differs across gestational age
categories, the added risk among second-delivered twins is of
similar concern irrespective of gestational age.
Risk of perinatal
death (per 100)
Risk
Gestational First-born Second-born difference
age (weeks) twin twin Relative risk (per 100)
2427 51.1 54.2 1.06 3.1

2831 7.1 10.3 1.45 3.2
3235 2.6 5.7 2.19 3.1
36 1.1 4.2 3.82 3.1
Question 6.14 The following is adapted from a report of a

case-control study of primary liver cancer in relation to serum
levels of retinol, in which serum samples had been obtained
prior to the diagnosis of cancer.40
Men with low prediagnostic serum retinol levels had a rela-

tively high risk of liver cancer. A statistically significant inter-
action was observed between retinol levels and hepatitis B
surface antigen (HBsAg) seropositivity on cancer risk: HBsAg-
positive men in the lowest third of the distribution of serum
retinol had greater than a 70-fold higher risk than HBsAg-
negative men in the highest third of the distribution of serum
retinol (p for interaction = .018).
From the information provided, can you determine the

nature of the interaction between retinol levels and HBsAg
status with regard to the incidence of liver cancer? If yes, what
is it? If not, what additional information from the study would
you need?
Answer 6.14 No, the nature of the interaction cannot be

determined. In each of the following examples, assume (for sim-
plicity) that the odds ratio associated with the presence of both
low retinol levels and HBsAg+ relative to high retinol levels
and HBsAg is not 70, but instead is 100 (200/100 20/1000),
based on the hypothetical data shown below.
The risk of liver cancer associated with low retinol levels
might be particularly great in men who are HBsAg positive:
HBsAg+ HBsAg
Case Control Case Control
Low retinol 200 100 Low retinol 700 1000

High retinol 10 100 High retinol 20 1000
Odds ratio 20 3.5

Or, it might not:
HBsAg+ HBsAg
Case Control Case Control
Low retinol 200 100 Low retinol 700 1000

High retinol 10 100 High retinol 20 1000
Odds ratio 20 35
Only in the first example is the odds ratio associated with

low retinol levels greater in HBsAg+ men than HBsAg men.
Therefore, whats needed to understand the nature of the inter-
action is the size of the odds ratio (or the risk difference, if
examining a deviation from additivity) associated with retinol
levels within categories of HBsAg seropositivity.
Question 6.15 The following data come from a case-control

study of upper gastrointestinal bleeding (UGIB) in relation to
prior use of nonsteroidal anti-inflammatory drugs (NSAIDS).
History of
ulcer NSAID use Cases Controls
No No 607 15,242
No Yes 171 597
Yes No 405 1,430
Yes Yes 106 164
Do the data from this study suggest that, when deciding to

initiate treatment with an NSAID, the risk of UGIB should
weigh less heavily as a potential adverse effect in persons with a
history of an ulcer than in other persons? Explain your answer.
Answer 6.15 The decision to use an NSAID (or any other

drug) should be based in part on a comparison of the size of the
added benefits and added risks that such use would entail. From
the data obtained in this case-control study, it is possible to
estimate the relative risk of UGIB (by means of the odds ratio)
but not the added (attributable) risk. Nonetheless, the rela-
tive size of the added risk associated with NSAID use in persons
with, and in persons without, a history of ulcer can be calcu-
lated, as shown below:
Odds Odds Relative

Ulcer NSAIDS Cases Controls ratioa ratiob added risk
No No 607 15,242 1 1 7.19 1 = 6.19

No Yes 171 597 7.19 7.19
Yes No 405 1,430 1 7.11 16.23 7.11 =
Yes Yes 106 164 2.28 16.23 9.12
a
Separate referent category of nonusers of NSAIDs for persons with and without
a history of ulcer.
b
Common referent category = Persons with neither a history of ulcer nor of
NSAID use.
The relative added risks associated with use of an NSAID

obtained above have no units, but nonetheless can be meaning-
fully compared between persons with and without ulcers. Since
the added risk of UGIB incurred as a result of NSAID use is sim-
ilar in the two groups (if anything, it is greater in those with
ulcers [i.e., 9.12 vs. 6.19]), it should receive equal weight by
potential NSAID users irrespective of ulcer history. The fact
that the first odds ratio is so much smaller in patients with
ulcers than in patients without ulcers is attributable to the
higher underlying rate of UGIB in ulcer patients.
CHAPTER 7
Screening
AT A POPULATION LEVEL, the degree to which a screening

test can lead to improved health outcomes is related to:
a. The proportion of the population with the condition

(or predictor of that condition) that the test seeks to
detect;
b. The sensitivity of the test in identifying the condition or
predictor;
c. A low frequency of false positive tests, or a low frequency
of adverse health outcomes associated with the conse-
quences of a false positive test; and
d. The efficacy of treatment of persons who screen as
positive.
Often, elements a-d above are addressed in separate studies.

Occasionally, though, especially when we dont have available to
us the experience of untreated persons who have tested posi-
tive, a single study seeks to examine the aggregate impact of two
or more elements (e.g., a comparison of cancer mortality in
screened and unscreened persons). The exercises in this chapter
consider studies of the individual elements as well as those that
deal with several of these at once.
SCREENING | 209
Question 7.1 In 132 patients with cirrhosis of the liver, serum

levels of alpha-L-fucosidase were measured and the patients
were followed for 8 years for the occurrence of liver cancer.
In 12 of them, levels of this enzyme were high. Liver cancer
was diagnosed in 19 patients, and in 3 of the 19, serum alpha-L-
fucosidase levels were high.
a. From the data provided, calculate the

1. Sensitivity
2. Specificity
3. Predictive value of a positive test
4. Predictive value of a negative test
of a high level of serum alpha-L-fucosidase in patents
with cirrhosis for predicting liver cancer during the
ensuing eight years.
b. There was a significant increase 6 to 9 months before
evidence of liver cancer in 7 of the 16 patients who had
low serum levels of alpha-L-fucosidase at enrollment.
What other information is needed before concluding that
a change of this sort is useful in predicting the presence
of liver cancer?
c. The authors of this paper concluded, We recommend the
measurement of this enzyme activity in surveillance pro-
grams for cirrhotic patients. Even if the serum alpha-
L-fucosidase level in a person with cirrhosis perfectly
discriminated between those who did and did not develop
liver cancer, is it possible that this recommendation could
be misguided? How?
Answer 7.1
High activity Cancer No cancer Total
Yes 3 9 12
No 16 104 120
19 113 132
a. 1. Sensitivity = 3/19 = 0.16

2. Specificity = 104/113 = 0.92
3. PV+ = 3/12 = 0.25
4. PV = 104/120 = 0.87
b. Other information needed: In what proportion of patients
who did not develop liver cancer was there a significant
increase in levels?
c. The recommendation could be misguided if, on average,
early recognition of a liver tumor, by means of this test,
did not result in an improved outcome.
SCREENING | 211
Question 7.2 The following appeared in the News section of

the Journal of the National Cancer Institute, May 17, 2000:
Some Promising Biomarkers for Cancer

- LPA (lysophosphatidic acid). LPA is probably
the most accurate marker we have for detection of early
stage ovarian cancer, said Northwestern Universitys David
Fishman, M.D., who is heading a multi-center study of the
marker. A 1998 report from the Cleveland Clinic found 9 of
10 women with stage 1 disease, 24 of 24 with advanced dis-
ease, and 14 of 14 with recurrent ovarian cancer had elevated
blood LPA levels. In contrast, just 5 of 48 controls had elevated
LPA. A growth factor, LPA is not generally present in normal
ovary cells.
Based on the above information, you believe it unlikely that

blood LPA levels will be of practical use in the early detection of
ovarian cancer. What is your reasoning?
Answer 7.2 If the prevalence of ovarian cancer among

screened women is low, the number of false positive tests
many of which would lead to a surgical procedure to document
the absence of ovarian cancerwould greatly exceed the
number of true positives. If the prevalence of cancer is 1/2,000,
for example, these would be the expected results in 20,000
screened women:
Ovarian cancer
LPA Yes No Total
Positive 10 5/48 (19,990) = 2,082 2,092

Negative 0 43/48 (19,990) = 17,908 19,990
10 19,990 20,000
The PV+ would be 10/2,092 = 0.005, very likely too low

to warrant use of LPA for early detection. For conditions
whose prevalence is relatively low, this question illustrates the
strong influence of a tests specificity on the predictive value of
a positive test result.
SCREENING | 213
Question 7.3 A test was developed (based on levels of a cer-

tain peptide in blood) to identify persons with congestive heart
failure. After studies revealed that a negative test predicts the
absence of heart failure virtually 100% of the time, the manu-
facturer of the test concluded that a positive test is an unam-
biguous warning sign for the presence of congestive heart
failure. You disagree. Why?
Answer 7.3
Congestive heart failure

Test Results Yes No Total
+ A b a+b
c (0%) d (100%) c+d
In order for a positive test to represent an unambiguous

warning sign, its positive predictive value (a/(a + b)) would
have to be high. However, the presence of a negative predictive
value (d/(c + d)) of 100% does not speak to this issue.
SCREENING | 215
Question 7.4 The questions below pertain to the following

abstract (abridged):
Survival of Women Ages 40-49 Years with Breast Carcinoma

according to Method of Detection
Methods. Women ages 40-49 years diagnosed with invasive
breast carcinoma between 1986 and 1992 were identified.
Measures of tumor size, lymph node status, and overall sur-
vival were compared with breast carcinoma patients whose
tumors were detected by breast self-exam (BSE), clinical breast
exam (CBE), patient incidental finding (PI), or mammography.
Results. Mean tumor size among women in the mammog-
raphy group was smaller than that among women in the BSE,
CBE, and PI groups (P <0.002).Tumors detected by mammog-
raphy were significantly more likely to be localized than those
detected by other methods (P <0.0001). Patients whose tumors
were detected by mammography had significantly better sur-
vival than patients in the other detection method (P <0.0001).
Conclusions. Women ages 40-49 years whose invasive
breast carcinoma is detected by mammography have signifi-
cantly smaller tumors, more localized disease, and may have a
lower risk of mortality than women whose tumors are detected
by other methods.41
a. When interpreting their findings, the authors of the

study discussed the possibility of lead-time bias accounting
for the relatively better survival of women whose breast
cancer was detected by means of screening mammo-
graphy. Why might lead-time bias have been present?
b. To address this concern, the authors adjusted for tumor
size and extent of disease when comparing survival
across the groups of women with breast cancer defined by
method of detection. What is the primary limitation of
an analysis of this sort?
Answer 7.4
a. Mammography is a relatively more sensitive means of

breast cancer detection: It can identify cancers earlier in
their natural history than can self-exam or clinical exam.
Therefore, even if there were no effective treatment for
early breast cancer, the mean interval from diagnosis to
death would be greatest in women whose tumors were
identified via mammography, due to their added lead-
time prior to the age they otherwise would have been
diagnosed. In the absence of effective treatment, receipt
of screening mammography would not have extended
life, but rather it would have extended that portion of the
same life span during which a woman was known to have
breast cancer.
b. Screening has the potential to exert a favorable influ-
ence on the likelihood of cancer mortality only by identi-
fying tumors that are relatively early in their natural
history, in other words, small in size and limited in
their spread. Adjusting for these characteristics would
not allow whatever true benefit that screening has on
survival to be evident.
SCREENING | 217
Question 7.5 This question pertains to the following news

article that appeared in the April 20, 2005, issue of the Journal
of the National Cancer Institute:
Group Recommends Earlier Colorectal Cancer Screening for

African Americans
The American College of Gastroenterology has recommended
that physicians begin screening African Americans for colorec-
tal cancer at age 45 rather than at age 50, the general recom-
mendation made by several groups.
In the publication of the recommendation, which appears
in the March issue of the American Journal of Gastroenterology,
the authors point out that African Americans have the highest
incidence of colorectal cancer of any racial or ethnic group. In
addition, they note that the mean age of presentation among
African Americans is lower than whites.
The American College of Gastroenterology based its recom-

mendation (regarding a racial difference in the age at which
screening for colorectal cancer begins) on the high incidence
and low mean age at presentation of colorectal cancer in African
Americans. The latter is not a valid reason. Why?
Answer 7.5 It is possible that the basis for the relatively

younger age distribution of African-American cases is the
relatively younger age distribution of African Americans in
general. The incidence rate of colorectal cancer among 45- to
49-year-oldsthe relevant piece of data that would bear on
the recommendationcould be identical between the races
and the mean age at diagnosis would still be lower in African
Americans than in whites.
SCREENING | 219
Question 7.6 The following is excerpted from the abstract of

an article on computer-aided detection of early breast cancer.42
We determined the association between the use of computer-

aided detection at mammography facilities and the perfor-
mance of screening mammography from 1998 through 2002
at 43 facilities in three states. We had complete data for
222,135 women (a total of 429,345 mammograms), including
2351 women who received a diagnosis of breast cancer within
1 year after screening. We calculated the specificity, sensitivity,
and positive predictive value of screening mammography with
and without computer-aided detection.
Diagnostic specificity decreased from 90.2% before imple-
mentation to 87.2% after implementation (P < 0.001). The
increase in sensitivity from 80.4% before implementation of
computer-aided detection to 84.0% after implementation was
not significant (P = 0.32).
The use of computer-aided detection is associated with
reduced accuracy of interpretation of screening mammograms.
The increased rate of biopsy with the use of computer-aided
detection is not clearly associated with improved detection of
invasive breast cancer.
The size of the decrease in specificity associated with

computer-aided detection90.2% 87.2% = 3.0%was
slightly smaller than the size of the increase in sensitivity
84.0% 80.4% = 3.6%. However, the p value for the difference
in specificity (<0.001) was far smaller than that for the differ-
ence in sensitivity (0.32). What is the reason for the disparity
between the size of two p values?
Answer 7.6 The calculation of the specificity is based on

the women without breast cancer, whereas the calculation of
the sensitivity is based on the women with breast cancer. The
p value is heavily influenced by the number of subjects, and in
this study the women without cancer outnumbered the women
with cancer by nearly 100 to 1. Thus, the p value must be smaller
for the difference in specificity than that for the difference in
sensitivity.
SCREENING | 221
Question 7.7 A randomized trial of fecal occult blood screen-

ing for colorectal cancer was conducted in Minnesota.43
a. During a follow-up period of 18 years, the incidence of

colorectal cancer that was first diagnosed when metastatic
(stage D) in 15,570 persons assigned to annual screening
was only 53% that of 15,394 persons assigned to not be
screened; for 15,587 persons assigned to be screened
every 2 years, the corresponding figure was 68%. Are any
other data needed to show that screening was successful?
b. The table below summarizes the mortality experience of
the three study groups. In assessing the efficacy of fecal
occult blood screening, should attention be focused primar-
ily on total mortality or mortality from colorectal cancer?
Study group
Annual Biennial
screening screening Control
No. enrolled 15,570 15,587 15,394

Person-years of 240,325 240,163 237,420
observation
Deaths from all causes
No. of deaths 5,236 5,213 5,186
Cumulative mortality* 342 340 343
95% confidence 334350 333348 336531
interval (CI)
Deaths from colorectal cancer (CRC)
No. of deaths 121 148 177
Cumulative mortality* 9.46 11.19 14.09
95% CI 7.7511.17 9.3912.99 12.0116.17
* Per 1000.
Answer 7.7
a. The answer depends on the criterion for success. On

average, persons in the groups assigned to receive screen-
ing did have their tumors detected at a relatively earlier
stage than persons assigned not to be screened. How-
ever, unless treatment at this earlier stage is more
efficacious than that given later, no lives will have been
saved. Therefore, if one wishes to learn whether screen-
ing led to a reduction in mortality, it is necessary to
compare mortality rates in the screened and unscreened
participants.
b. If assessment of cause of death is believed to be largely
unaffected by screening history and the impact of screen-
ing is believed to be confined to averting deaths from col-
orectal cancer, then the focus of the analysis should be on
mortality from colorectal cancer. If there are potentially
fatal complications of the screening process (e.g., perfo-
rated colon during colonoscopy that would follow a posi-
tive stool exam) or treatment (e.g., death during cancer
surgery), deaths from these causes should be included
as well.
A comparison of all-cause mortality would be an
insensitive means of measuring the potential benefit
associated with screening: Even if annual screening pre-
vented every death from colorectal cancer and had no
bearing on any other cause of death, the all-cause mortal-
ity in the screened group would be decreased by only 3%
5,186 - 177
5,186 = 0.97

SCREENING | 223
Question 7.8 The following is excerpted from a manuscript

describing the results of a cohort study of screening for stom-
ach cancer (by means of photofluorography, a highly sensitive
test) in Japan in relation to mortality from stomach cancer:
The present study focused on 100,562 subjects who were ages

4079 years at the time of a baseline survey, which asked for
their screening experience during the past twelve months.
Of these, 219 subjects with a history of stomach cancer prior
to the time of the baseline survey were excluded, as were
8,386 subjects who did not provide information on their
participation in stomach-cancer screening.
Death rates from stomach cancer during a period follow-
ing the baseline survey, adjusted for differences in age and
other demographic characteristics, were compared between
participants with and without a history of screening for this
disease.
Do you believe this study is likely to provide a valid esti-

mate of the ability of screening photofluorography to lead to a
reduction in mortality from stomach cancer? If yes, why? If not,
why not, and in which direction do you believe the bias will
occur?
Answer 7.8 The design of this study does not permit an unbi-
ased estimate of the potential impact of photofluorographic
screening on mortality from stomach cancer. The problem stems
from what has been referred to as healthy screenee bias.26 The
exposed cohort in the present study has received a screening
exam in the prior year that was negative: Had it been positive
for cancer, such persons would have been removed from the
cohort. The unexposed cohort has not had such persons iden-
tified or excluded. Thus, even if screening led to completely
ineffective treatment, or to no treatment at all, the observed
relative mortality from stomach cancer would be lower in the
screened group.
SCREENING | 225
Question 7.9 You find the following in an article in a medical

journal:
Pap smear screening may be less effective among black women

than among white women. Laboratory-based evidence of Pap
smear screening (i.e., a Pap smear performed in the absence of
symptoms of cervical cancer) at least once during the past
5 years was found for 48% of the black population with inva-
sive cervical cancer versus only 32% of the white population
with invasive cervical cancer (p <0.05).
Assuming that ascertainment of invasive cervical cancer

was equally complete for the black and white populations
and that management of the Pap smear abnormalities was sim-
ilar for the two groups, what reason might there be for the
observed result other than a differential effectiveness of Pap
smear screening between the two races?
Answer 7.9 It is possible that in the underlying population

from which the women with cervical cancer were drawn, a
higher percentage of black women than white women had
undergone Pap screening at least once in the past 5 years. What
is needed in order to gauge the efficacy of this screening modal-
ity in the prevention of invasive cervical cancer is the propor-
tion of women in the population at risk who had been screened.
Suppose the following results would have been observed in a
case-control study of invasive cervical cancer:
Black women White women

Invasive Invasive
cancer Controls cancer Controls
% screened 48 60 32 44
% not screened 52 40 68 56
Odds ratio 0.6 0.6
In this situation, the relative impact of screening in reduc-

ing the incidence of invasive cervical cancer would have been
identical in black and in white women.
SCREENING | 227
Question 7.10 A randomized trial was performed to deter-

mine whether screening women with ovarian cancer in clinical
remission for serum levels of a tumor marker every 3 months,
followed by treatment of recurrences identified through this
means, could reduce mortality. Women were recruited into
the trial at the onset of the remission induced by their initial
treatment. Those who later were found to have elevated levels
of the marker were randomized to one of two approaches to
management:
(a) Immediate chemotherapy; or

(b) No additional treatment until the recurrence became
apparent for other reasons (typically, the development
of symptoms), on average 5 months later. (Values for
serum levels of the tumor marker in patients in group
(b) were not provided to these womens physicians.)
In comparing mortality between women in groups (a) and

(b), should rates be calculated from the time of randomization
or from the time of recognition of the recurrence? Explain the
reason for your answer.
Answer 7.10 The mortality rates in both groups should be

based on follow-up beginning at the time of randomization, in
other words, when the elevation in serum levels of the tumor
marker was first apparent. Follow-up beginning at the time
of recognition of the recurrence would give rise to lead-time
bias: Only in group (a) would person-time be accrued during
(on average) the first 5 months after tumor marker elevation
was noted, and in this period of time mortality from ovarian
cancer would be expected to be considerably smaller than in the
period following symptom development.
(A randomized trial of the efficacy of screening for the
tumor marker CA 125, designed in the manner described in this
question, did indeed tabulate mortality rates in both groups
from the time of randomization.47)
SCREENING | 229
Question 7.11 You are designing a case-control study to esti-

mate the efficacy of PSA screening for prostate cancer. From the
records of a large prepaid health-care plan, you are going to
ascertain screening histories of men who died as a result of
prostate cancer during 2007-2009. As a basis for comparison,
you are considering two possible control groups (who would be
individually matched to the fatal cases):
1. Members of the health-care plan who are demographically

comparable to the fatal cases and also were diagnosed
with prostate cancer at about the same time, but who
were still alive at the time of the matched cases death
2. Demographically comparable members of the health plan
who had not been diagnosed with prostate cancer as of
the date of diagnosis of their matched fatal case
Which of the two groups above would provide the more

valid result? Explain.
Answer 7.11 Option 2 conforms to the goal of control defini-

tion, in other words, a sample of the population from which the
cases were derived. The problem with option 1 is that, even if no
effective treatment were available for screen-detected prostate
cancer (in which case the correct odds ratio should be one),
there would be a higher proportion of screened men in controls
than cases (assuming that PSA screening has some sensitivity
for detecting prostate cancer).48
SCREENING | 231
Question 7.12 A case-control study was conducted to assess

the impact of cervical screening in reducing the incidence of
invasive cervical cancer.44 Because such screening can identify
cervical precancerous lesions, it was hypothesized that a smaller
proportion of cases than controls would have been screened at
some point during the prior 4 years.
The women with cervical cancer had been diagnosed during
20002003; they were identified using records of a cancer reg-
istry serving the area in which the study was done. The receipt
of screening during 19962003 was ascertained by means of
a registry that enumerated screening exams in that area.
Age-matched controls were chosen from women identified
in the screening registry during 19962004. The investigators
explained that each case would still have the possibility of
being matched to control(s) with no screening history prior to
the cases diagnosis, since some controls would have received
their first screen during 2004 and so would have been
unscreened during 19962003.
What do you perceive to be the greatest threat to the
validity of this study?
Answer 7.12 The purpose of a control group in a case-control

study is to estimate the proportion of exposed persons in the
population (who were at risk of the disease in question) from
which the cases arose. In this instance, the investigators needed
screening histories on a sample of women with a cervix who
resided in the area from which the cases had been derived. It
would be fortuitous if the sampling from the controls who were
chosen by the investigators succeeded in accomplishing this.
Almost certainly, the level of screening in the controls actually
selected (from a screening registry) overstated that in the
population at risk, leading to an observed odds ratio (approxi-
mately 0.1) that was falsely low.
SCREENING | 233
Question 7.13 Hackam et al. ascertained prior use of antihy-

pertensive and other medications among patients over 65 years
of age hospitalized with a ruptured abdominal aortic aneurysm
(cases, n = 3,379) and those hospitalized with an unruptured
aneurysm (controls, n = 11,947).45 In the course of reviewing
electronic files of medical services that preceded the aneu-
rysm diagnosis, the investigators incidentally noted receipt of
abdominal imaging (ultrasound, CT, or MRI) in nearly all of the
patients with an unruptured abdominal aortic aneurysm, but
in only a small fraction of those with a ruptured aneurysm.
Screening by means of abdominal imaging has the poten-
tial to lead to identification of an aortic aneurysm, and in turn
to surgical intervention that could avert a rupture. However,
despite the large observed case-control difference in the prior
receipt of screening by means of abdominal imaging, it would
not be appropriate to use the data of Hackam et al. to support
the hypothesis that screening could lead to a reduced risk of
a ruptured abdominal aortic aneurysm. Why? What would be
a better basis for comparison to the patients with a ruptured
abdominal aortic aneurysm?
Answer 7.13 In order for a case-control study of screening

efficacy to provide a valid result, the level of screening in the
controls must reflect that of the population at risk from which
the cases were derived. All of the controls in this study were
known to have an abdominal aortic aneurysm, many, no doubt,
as a result of screening, and so the level of screening in them
would be expected to be atypically high. Thus, even if there truly
were no beneficial impact of early recognition and treatment
of abdominal aortic aneurysms on the occurrence of rupture,
a (far) smaller proportion of persons with a ruptured aneu-
rysm than a known unruptured one would have a history of
screening.
If one were to design a case-control study to estimate the
ability of aneurysm screening to reduce the occurrence of rup-
tured abdominal aortic aneurysm (this was not the goal of
Hackam et al.), controls would need to be selected from an
entirely different sampling frame. If (as in this study) the cases
of ruptured aneurysm were all those over age 65 diagnosed in a
specific geographic population, the controls ideally would be
demographically similar members of that same population.
(Fortunately, there are data available from randomized
trials from which we can gauge the efficacy of ultrasound screen-
ing for abdominal aortic aneurysm.49)
SCREENING | 235
Question 7.14 In order to gauge the efficacy of screening by

means of a digital rectal exam (DRE) or measurement of levels
of prostate-specific antigen (PSA) in serum against mortality
from prostate cancer, a case-control study was conducted.46
A comparison was made between receipt of either of these
screening modalities in 74 men who died as a result of prostate
cancer (cases) and in a sample of year-of-birth matched men
who resided in the same county as did the cases. Because of the
concern that tests performed close to the time of the diagnosis
of prostate cancer might have been prompted by symptoms, the
primary analysis focused on screening tests performed during
the 1 to 5 years prior to diagnosis of prostate cancer, and during
the corresponding period of time in control men. Screening
tests had been performed on 81.3% of controls during this
interval, in contrast to just 60.8% of the fatal cases (odds
ratio = 0.35, 95% confidence interval = 0.170.71).
Because of the means by which exposure was defined in
the primary analysis, it is likely that a smaller proportion of
cases than controls would have been screened, even had early
identification of prostate cancer failed to lead to treatment that
would favorably influence the risk of death from prostate cancer.
Why is this?
Answer 7.14 For a screening modality that is sensitive

(such as a serum PSA measurement for the presence of pros-
tate cancer), excluding the 12 months prior to diagnosis from
consideration will effectively exclude most positive tests. And,
because positive tests are proportionally far more numerous in
men with prostate cancer than in men in general, exclusion of
the 12 months prior to diagnosis means that proportionally
more cases than controls who had been screened will be classi-
fied as not having been screened. Therefore, even if screening
failed to lead to any mortality reduction, this analytic approach
would give rise to an odds ratio less than 1, suggesting efficacy
against cancer mortality when in truth none had been present.
A valid result in case-control study of screening requires
accurate assessment of receipt of screening in cases and con-
trols during the whole of the period prior to diagnosis or symp-
toms (whichever comes first) that correspond to the period of
potential detectability.50
REFERENCES
1. Prescott E, Hippe M, Schnohr P, et al. Smoking and the risk of

myocardial infarction in women and men: longitudinal popula-
tion study. BMJ 1998;316:10431046.
2. Kendal WS. Suicide and cancer: a gender-comparative study.
Ann Oncol 2007;18:381387.
3. Andriole GL , Crawford ED, Grubb RL , et al. Mortality results
from a randomized prostate-cancer screening trial. N Engl J Med
2009; 360:13101309.
4. Carroll-Pankhurst C, Mortimer EA . Sudden infant death syn-
drome, bedsharing, parental weight, and age at death. Pediatrics
2001; 107: 530536.
5. Kumar AS, Benz CC, Shim V, et al. Estrogen receptor-negative
breast cancer is less likely to arise among lipophilic statin users.
Cancer Epidemiol Biomarkers Prev 2008; 17:10281033.
6. Mosley BS, Cleves MA , Siega-Riz AM, et al. Neural tube defects
and maternal folate intake among pregnancies conceived af-
ter folic acid fortification in the United States. Am J Epidemiol
2009;169:917.
7. Wong YN, Mitra N, Hudes G, et al. Survival associated with treat-
ment vs observation of localized prostate cancer in elderly men.
JAMA 2006;296:26832693.
2 3 8 | REFERENCES
8. Galea S, Blaney S, Nandi A , et al. Explaining racial disparities

in incidence of and survival from out-of-hospital cardiac arrest.
Am J Epidemiol 2007;166:534543.
9. Stromberg , B, Dahlquist G, Ericson A , et al. Neurological seque-
lae in children born after in-vitro fertilisation: a population-
based study. Lancet 2002;359:461465.
10. Koepsell TD, Weiss NS. Epidemiologic Methods. Oxford,
New York, 2003.
11. Hulley S, Grady D, Bush T, et al. Randomized trial of
estrogen plus progestin for secondary prevention of coronary
heart disease in postmenopausal women. JAMA 1998;280:
605613.
12. Cardo DM, Culver DH, Ciesielski CA , et al. A case-control
study of seroconversion in health care workers. Centers for
Disease Control and Prevention Needlestick Surveillance Group.
N Engl J Med 1997;337:14851490.
13. Steinmaus C, Yuan Y, Bates MN, et al. Case-control study
of bladder cancer and drinking water arsenic in the western
United States. Am J Epidemiol 2003;158:11931201.
14. Furness S, Connor J, Robinson E, et al. Car colour and risk
of car crash injury: population based case control study. BMJ
2003;327:14551456.
15. Ray JG, Meier C, Vermeulen MJ, et al. Association of neural
tube defects and folic acid food fortification in Canada. Lancet
2002;360:20472048.
16. Lemaitre RN, Siscovick DS, Raghunathan TE, et al. Leisure-
time physical activity and the risk of primary cardiac arrest.
Arch Int Med 1999;159:686690.
17. Jick SS, Hagberg KW, Kaye JA , et al. Postmenopausal
estrogen-containing hormone therapy and the risk of breast
cancer. Obstet Gynecol 2009;113:7480.
18. Skjaerven R , Wilcox AJ, Lie RT. The interval between pregnancies
and the risk of preeclampsia. N Engl J Med 2002;346:3338.
19. Villeneuve PJ, Holowaty EJ, Brisson J, et al. Mortality among
Canadian women with cosmetic breast implants. Am J Epidemiol
2006;164:334341.
20. Nichol KL , Baken L , Wuorenma J, et al. The health and
economic benefits associated with pneumococcal vaccination
REFERENCES | 239
of elderly persons with chronic lung disease. Arch Intern Med

1999; 159: 24372442.
21. Leonard RC, Kreckmann KH, Sakr CJ, et al. Retrospective
cohort mortality study of workers in a polymer production
plant including a reference population of regional workers.
Ann Epidemiol 2008;18:1522.
22. Stattin P, Holmberg , E, Johansson, JE, et al. Outcomes in
localized prostate cancer: National Prostate Cancer Register of
Sweden followup study. J Nat Cancer Inst 2010;102:950958.
23. Anderson JP, Ross JA , Folsom AR . Anthropometric variables,
physical activity, and incidence of ovarian cancer. Cancer 2004;
100:15151521.
24. Beral V, Million Women Study Collaborators. Breast cancer and
hormone-replacement therapy in the Million Women Study.
Lancet 2003;362:419427.
25. Youakim S. Risk of cancer among firefighters: a quantitative
review of selected malignancies. Arch Environ Occup Health
2006;61:223231.
26. Weiss NS, Rossing MA . Healthy screenee bias in epidemiologic
studies of cancer incidence. Epidemiology 1996;7:319322.
27. Caughey AB, Nicholson JM, Cheng YW, et al. Induction of labor
and cesarean delivery by gestational age. Am J Obstet Gynecol
2006;195:700705.
28. Chiazze L , Ference LD, Wolf PH: Mortality among automo-
bile assembly workers: I. Spray painters. J Occup Med 1984;22:
520526.
29. Rowe B, Milner R , Johnson C, et al. The association of alcohol
and night driving with fatal snowmobile injury: a case-control
study. Ann. Emerg Med 1994;24:842848.
30. Heffelfinger JD, Heckbert SR , Psaty BM, et al. Influenza vacci-
nation and risk of incident myocardial infarction. Hum Vaccin
2006;2:161166.
31. Spirtas R , Heineman EF, Bernstein L , et al. Malignant mesothe-
lioma: attributable risk of asbestos exposure. Occup Environ Med
1994; 51:804811.
32. Rosenberg L , Rao RS, Palmer JR , et al. Transitional cell cancer
of the urinary tract and renal cell cancer in relation to acetamin-
ophen use. Cancer Causes Control 1998;9:8388.
2 4 0 | REFERENCES
33. Handke M, Harloff A , Olschewski M, et al. Patent foramen

ovale and cryptogenic stroke in older patients. N Engl J Med.
2007;351:22622268.
34. Hessel PA , Teta MJ, Goodman M, et al. Mesothelioma among
brake mechanics: an expanded analysis of a case-control study.
Risk Anal 2004;24:547552.
35. Grant AM, Avenell A , Campbell MK , et al. Oral vitamin D3 and
calcium for secondary prevention of low-trauma fractures in
elderly people: a randomised placebo-controlled trial. Lancet
2005; 365:16211628.
36. Glynn RJ, Danielson E, Fonseca FA , et al. A randomized trial
of rosuvastatin in the prevention of venous thromboembolism.
N Engl J Med 2009;360:18511861.
37. John EM, Koo J, Horn-Ross PL . Lifetime physical activity and
risk of endometrial cancer. Cancer Epidemiol Biomarkers Prev
2010;19:12761283.
38. Donnell D, Baeten JM, Kiarie J, et al. Heterosexual HIV-1 trans-
mission after initiation of antiretroviral therapy: a prospective
cohort analysis. Lancet 2010;375:20922098.
39. Treggiari MM, Weiss NS. Occupational asbestos exposure and
the incidence of non-Hodgkin lymphoma of the gastrointestinal
tract: an ecologic study. Ann Epidemiol 2004;14:168171.
40. Yuan JM, Gao YT, Ong CN, et al. Prediagnostic level of serum
retinol in relation to reduced risk of hepatocellular carcinoma.
J Natl Cancer Inst 2006;98:482490.
41. McPherson CP, Swenson KK , Jolitz G, et al. Survival of women
ages 40-49 years with breast carcinoma according to method
of detection. Cancer 1997;79:19231932.
42. Fenton JJ, Taplin SH, Carney PA , et al. Influence of computer-
aided detection on performance of screening mammography.
N Engl J Med 2007; 356:13991409.
43. Mandel JS, Church TR , Ederer F, et al. Colorectal cancer mor-
tality: effectiveness of biennial screening for fecal occult blood.
J Natl Cancer Inst 1999;91:434437.
44. Yang B, Morrell S, Zuo Y, et al. A case-control study of the protec-
tive benefit of cervical screening against invasive cervical cancer
in NSW women. Cancer Causes Control 2008;19:569576.
REFERENCES | 241
45. Hackam DG, Thiruchelvam D, Redelmeier DA . Angiotensin-

converting enzyme inhibitors and aortic rupture: a population-
based case-control study. Lancet 2006; 368:659665.
46. Bergstralh EJ, Roberts RO, Farmer SA , et al. Population-based
case-control study of PSA and DRE screening on prostate cancer
mortality. Urology 2007;70:936941.
47. Rustin GJ, van der Burg ME, Griffin CL , et al. Early versus
delayed treatment of relapsed ovarian cancer. Lancet 2010;376:
11551163.
48. Weiss NS. Control definition in case-control studies of the
efficacy of screening and diagnostic testing. Am J Epidemiol
1983;118:457460.
49. Thompson SG, Ashton HA , Gao L , Scott RA . Screening men for
abdominal aortic aneurysm: 10 year mortality and cost effec-
tiveness results from the randomised Multicentre Aneurysm
Screening Study. BMJ 2009;338:b2307.
50. Weiss NS, Dhillon PK , Etzioni RS. Case-control studies of the
efficacy of cancer screening. Overcoming bias from nonrandom
patterns of screening. Epidemiology 2004;15:409413.
INDEX
Abdominal imaging, aortic crude and age-adjusted death

aneurysm, 233234 rates, 7172
Abortion depression and womans, at first
analgesic and pregnancy, 5758 birth, 5556
oral contraceptives and, 107108 health status and physical
Accident, automobile damage, activity, 105106
4748 infectious disease and
Accuracy, proportion or rate, 3 vaccination, 185186
Acetaminophen, transitional marital status and suicide,
cell cancer of urinary tract, 101102
171172 maternal, and cesarean delivery,
Africa, HIV infection and 103104
antiretroviral therapy, maternal age and congenital
193194 malformation, 9192
African Americans, colorectal menarche, 6768
cancer screening, 217218 men with oral cancer,
African-American women, 3536
depression and age at first oral contraceptives and abortion,
birth, 55 107108
Age pneumonia, 7576
cigarette smokers, 9798 prostate cancer and marital
cigarette smoking and gender, status, 8586
115116 renal cell carcinoma, 18
colorectal cancer, 2526 stomach cancer, 3940
2 4 4 | INDEX
Age (Continued) Asbestos, occupational, and

sudden infant death syndrome non-Hodgkin lymphoma of
(SIDS), 4546 gastrointestinal tract (GINHL),
U.S. cancer mortality, 195196 197198
AIDS exposure, health-care Asian women, ovarian cancer, 78
workers, 8384 Association
Air travel, oral contraceptives and, analgesic and abortion, 5758
6162 anthropometric variables and
Alaska Natives, cigarette smokers, ovarian cancer, 133134
9798 asthma and pertussis
Alcohol vaccination, 109110
colorectal cancer and drinking, automobile damage, 4748
2526 car color and automobile crash,
snowmobile deaths and, use of, 8990
159160 causal inference, 43
Alpha-L-fucosidase, serum levels confounding , 69
and liver cancer, 209210 diet and serum lipids, 5354
American College of family history of stuttering,
Gastroenterology, colorectal 149150
cancer, 217218 folic acid and pregnancy, 6364
American Indians, cigarette HIV infection and antiretroviral
smokers, 9798 therapy, 193194
Analgesic nephropathy, renal cell marital status and suicide,
carcinomas and, 1718 101102
Analgesics, pregnancy, 5758 neck trauma and diffuse pain
Anthropometric variables, ovarian syndrome, 143144
cancer, 133134 oral contraceptives and air travel,
Anticoagulant therapy, transient 6162
ischemic attacks (TIA) and snowmobile death and alcohol
stroke, 141142 use, 159160
Antidepressant drugs, pregnancy, SSRIs and pregnancy, 4950
4950 Asthma, children with history of
Antiretroviral therapy, HIV pertussis, 109110
infection, 193194 Athletic training, age-at-menarche,
Anxiety disorders, genome-wide 6768
association study (GWAS), Australia, oral contraceptives and
9596 abortion, 107108
Aortic aneurysm, abdominal Automobile crashes
imaging, 233234 car color, 8990
Arsenic, bladder cancer and, in cataract surgery, 1314
water, 8788 seat belts, 2728
INDEX | 245
Automobile industry, spray Brazil, breast-fed infants and

painting and lung cancer pneumonia, 151152
deaths, 155156 Breast cancer
Automobiles, body damage, 4748 computer-aided detection,
AZT (zidovudine), exposed health- 219220
care workers, 8384 detection and survival, 215216
estrogen-progestogen hormone
Bedsharing, sudden infant death therapy, 191192
syndrome (SIDS), 4546 estrogen receptor (ER) and
Bilateral oophorectomy, breast progesterone receptor (PR),
cancer and hormone use, 5960
111112 estrogen receptor-positive vs.
Births receptor-negative, 175
congenital malformation and hormone replacement therapy
folate fortification, 9192 (HRT), 135136
maternal age and cesarean hormone use and bilateral
delivery, 103104 oophorectomy, 111112
order of delivery of twins and mammography, 215216,
death, 201202 219220
Black men screening and mortality,
prostate cancer, 2324 139140
smoking, 7778 Breast feeding, pneumonia in
Black women infants, 151152
depression and age at first Breast self-exam (BSE), cancer
birth, 55 detection, 215216
Pap smear screening for cervical British Medical Journal
cancer, 225226 lung cancer by gender, 9
prenatal care and marital status, myocardial infarction, 1516
113114
Bladder cancer, arsenic in drinking Calcium supplements, fractures in
water, 8788 elderly people, 183184
Blood lead levels, children eating Cancer
unusual substances, 1920 alpha-L-fucosidase levels and
Bloodletting, pneumonia, 127128 liver, 209210
Blood levels, folic acid and anthropometric variables and
pregnancy, 6364 ovarian, 133134
Body mass index (BMI) biomarker lysophosphatidic acid
endometrial cancer risk, (LPA) for ovarian, 211212
189190 bladder, and arsenic in drinking
ovarian cancer, 133134 water, 8788
Brain cancer, firefighters, 137138 brain, in firefighters, 137138
2 4 6 | INDEX
Cancer (Continued) screening , 207

breast, and estrogen-progestogen stomach, 3940, 223224
hormone therapy, 191192 suicide and type of, 3334
breast, diagnosis, 5960 survival and detection of breast,
breast and colon, 139140 215216
cigarette smoking and lung, by thyroid, 3334
gender, 181182 transition cell, of urinary tract
colorectal, 2526, 217218 and acetaminophen, 171172
comorbidity of prostate, 6566 tumor marker for ovarian,
computer-aided detection of 227228
breast, 219220 Cannabis, driving under influence,
endometrial, 1112 161162
endometrial, and estrogen Carcinogen A in tea, stomach
therapy, 177178 cancer, 3940
endometrial, and physical Carcinomas, renal cell, and
activity, 189190 analgesic nephropathy, 1718
estrogen receptor-negative and Car color, automobile crashes,
receptor-positive breast, 175 8990
genetics and incidence of lung, Cardiac arrest, occurrence by race,
153154 7374
incidence and mortality, 139140 Cardiovascular disease, influenza
incidence rates in U.S., 9394 vaccination, 163164
lifetime probability, 56 Case-control studies, 147
lung, by gender, 910 abdominal imaging for aortic
lung, deaths in automobile aneurysm, 233234
workers, 155156 acetaminophen and transitional
melanoma, and suicide, 3334 cell cancer of urinary tract,
mortality in U.S., 195196 171172
oral, 3536 cervical cancer, 225226
ovarian, and race, 78 cervical cancer screening,
Pap smear screening for cervical, 231232
225226 crashes involving drugs and
prostate and marital status, alcohol, 161162
8586 incidence of lung cancer, 153154
prostate and race, 2324 liver cancer and serum levels of
prostate case-control study, retinol, 203204
235236 nonsteroidal anti-inflammatory
prostate diagnosis, 4142 drugs and gastric ulcers,
prostate mortality, 131132 169170
PSA screening for prostate, occupation and mesothelioma,
229230 165166
INDEX | 247
physical activity and endometrial Cigarette smoking

cancer risk, 189190 age and gender, 115116
pneumonia and breastfeeding , colorectal cancer, 2526
151152 lung cancer, 5152
prostate cancer, 235236 lung cancer and gender,
sexually transmitted infection 181182
and condom use, 157158 marital status and gender,
snowmobile deaths and alcohol 115116
use, 159160 myocardial infarction, 1516
spray painting and lung cancer sex, race/ethnicity and age,
deaths, 155156 9798
sudden infant death syndrome Cirrhosis of liver, alpha-L-
(SIDS) and sleeping position, fucosidase levels, 209210
167168 Clinical breast exam (CBE), cancer
Case-fatality, endometrial cancer, detection, 215216
1112 Cohort studies, 119
Cataracts, surgical treatment, hepatocellular carcinoma (HCC),
1314 2122
Caucasian women, ovarian stomach cancer, 223224
cancer, 78 College female athletes, age-at-
Causal inference, 43 menarche, 6768
Cervical cancer Colon cancer, screening and
Pap smear screening, 225226 mortality, 139140
screening registry, 231232 Colorectal cancer
Cesarean delivery age and behavior, 2526
elective induction of labor, mortality and screening,
145146 221222
maternal age and, 103104 Comorbidity, prostate cancer,
Chemotherapy, ovarian cancer, 6566
227228 Comparison, case-control
Childless women, first births, 2930 studies, 147
Children Computer-aided detection, breast
age of women for first, 2930 cancer, 219220
asthma and pertussis Condom use, sexually transmitted
vaccination, 109110 infection, 157158
blood lead levels and eating Confounding , 69, 175
unusual substances, 1920 Congestive heart failure, peptide in
disability and in vitro blood, 213214
fertilization, 7980 Co-relations, confounding, 69
Chronic lung disease, pneumococcal Coronary heart disease, women and
vaccination, 123124 hormones, 8182
2 4 8 | INDEX
Cosmetic breast implants, Diagnosis

mortality, 121122 breast cancer type, 5960
Cryptogenic stroke, patent foramen probability of cancer, 56
ovale, 173174 prostate cancer, 4142
Diet, serum lipids and, 5354
Damage, automobiles, 4748 Diethylstilbestrol, maternal use
Day care enrollment, asthma and of, 54
pertussis, 109110 Diffuse pain syndrome (DPS), neck
Death trauma and, 143144
breast cancer and hormone Digital rectal exam, prostate cancer,
replacement therapy (HRT), 235236
135136 Disability, in vitro fertilization in
colorectal cancer screening , Sweden, 7980
221222 Distortions, confounding, 69
coronary heart disease, 81 Drinking water, arsenic in, and
DuPont Washington Works bladder cancer, 8788
plant, 125126 Driving
East and West Germany, cataract surgery, 1314
99100 crashes involving drugs and
lung cancer in automobile alcohol, 161162
workers, 155156 snowmobile deaths and alcohol
order delivery of twins, use, 159160
201202 DuPont Washington Works plant,
prostate cancer, 4142, 6566 standardized mortality ratios
seat belts and automobile (SMRs), 125126
crashes, 2728 Dyspepsia, nonsteroidal anti-
snowmobile, and alcohol use, inflammatory drugs and
159160 gastric ulcers, 169170
sudden infant death syndrome
(SIDS), 4546 Eating unusual substances, blood
Death rates lead levels in children, 1920
crude and age-adjusted, 7172 Education
women and breast implants, health status and physical
121122 activity, 105106
Depression pregnancy intendedness and
genome-wide association study physical violence, 199200
(GWAS), 9596 Effect modification, 175
womans age at first birth, Elderly patients, fractures in, and
5556 vitamin D and calcium, 183184
Diabetes, health status and physical Elective labor induction, pregnancy,
activity, 105106 145146
INDEX | 249
Employment Fractures, vitamin D and calcium

brain cancer in firefighters, supplements, 183184
137138
exposure to mesothelioma, Gastric ulcers, nonsteroidal anti-
165166 inflammatory drugs (NSAIDs),
Endometrial cancer 169170, 205206
case-fatality, 1112 Gastrointestinal tract, non-
estrogen therapy, 177178 Hodgkin lymphoma of,
physical activity, 189190 and occupational asbestos,
Epidemiologists, rates and 197198
proportions, 3 Gender
Estrogen-progestogen hormone cigarette smokers, 9798
therapy, breast cancer, cigarette smoking and age,
191192 115116
Estrogen receptor (ER), breast cigarette smoking and lung
cancer type, 5960 cancer, 181182
Estrogen receptor-negative, breast cigarette smoking and marital
cancer, 175 status, 115116
Estrogen receptor-positive, breast lung cancer, 910
cancer, 175 myocardial infarction and
Estrogens smoking , 1516
breast cancer and bilateral suicide by American physicians,
oophorectomy, 111112 3132
endometrial cancer and, therapy, Genetic characteristic, incidence of
177178 lung cancer, 153154
women and coronary heart Genome-wide association
disease, 8182 study (GWAS), depression,
Ethnicity 9596
cardiac arrest, 7374 Germany, life expectancy in East
cigarette smokers, 9798 and West, 99100
Gestational age, perinatal death
Factor V Leiden mutation, venous among twins, 201202
thrombosis (VT), 179180 GINHL (non-Hodgkin lymphoma
Family history, stuttering , 149150 of gastrointestinal tract),
Female athletes, age-at-menarche, occupational asbestos,
6768 197198
Firefighters, brain cancer, Gonorrhea, condom use, 157158
137138 Group Health Cooperative (GHC),
Folate fortification, congenital influenza vaccine and
malformation and, 9192 myocardial infarction (MI),
Folic acid, pregnancy, 6364 163164
2 5 0 | INDEX
GuillainBarr syndrome, influenza breast cancer and estrogen-

vaccine, 3738 progestogen, 191192
Gulf War veterans, postwar Hospitalization, pneumonia and
mortality, 129130 influenza, 123124
Hydrophilic statin, breast cancer,
Health Care Financing 5960
Administration, arsenic Hypersensitivity, stuttering,
levels, 87 149150
Health-care plan Hypertension, health status and
analgesic and pregnancy, 5758 physical activity, 105106
PSA screening for prostate
cancer, 229230 Immunity, cancer diagnosis, 56
Health-care workers, AZT for Incidence
exposure, 8384 breast and colon cancer, 139140
Heart disease, mortality at DuPont myocardial infarction and
Washington Works plant, smoking , 1516
125126 ovarian cancer by race, 78
Heart failure, peptide in blood and rosuvastatin and, of venous
congestive, 213214 thromboembolism (VTE),
Hepatitis B surface antigen 187188
(HBsAg), liver cancer risk, Income levels, bladder cancer in
203204 arsenic in water, 8788
Hepatitis B surface antigen Inducing labor, pregnancy,
positivity (HbsAg+), mortality, 145146
2122 Infants
Hepatocellular carcinoma (HCC), breast-feeding and pneumonia,
mortality, 2122 151152
Heterosexual couples, HIV SIDS (sudden infant death
infection and antiretroviral syndrome), 4546
therapy, 193194 SIDS and sleeping position,
Hispanic men, oral cancer, 3536 167168
HIV infection Infectious disease, vaccination
antiretroviral therapy, 193194 against, 185186
health-care workers, 8384 Inference, causal, 43
non-Hodgkin lymphoma Influenza, pneumococcal
of gastrointestinal tract, vaccination, 123124
197198 Influenza vaccine
Hormone therapy GuillainBarr syndrome, 3738
breast cancer, 135136 myocardial infarction (MI) and
breast cancer and bilateral ischemic stroke, 163164
oophorectomy, 111112 Inheritance, stuttering, 149150
INDEX | 251
Insurance agency, automobile retinol levels and hepatitis B

damage, 4748 surface antigen (HBsAg),
Interbirth interval, pre-eclampsia, 203204
117118 Lung cancer
In vitro fertilization, disability in deaths among automobile
Sweden, 7980 workers, 155156
Iowa Womens Health Study, gender, 910
ovarian cancer, 133134 gender and cigarette smoking,
181182
Japan, stomach cancer screening, genetics and incidence, 153154
223224 smoking , 5152
Journal of the National Cancer suicide among men, 3334
Institute Lung disease, chronic, and
colorectal cancer screening , pneumococcal vaccination,
217218 123124
lysophosphatidic acid (LPA), Lysophosphatidic acid (LPA),
211212 biomarker for ovarian cancer,
211212
Korea, mortality from
hepatocellular carcinoma Mammography, breast cancer,
(HCC), 2122 215216, 219220
Marital status
Labor and delivery, SSRIs and cigarette smoking and gender,
pregnancy, 4950 115116
Labor induction, pregnancy, prostate cancer and age, 8586
145146 suicide, 101102
Lancet Maternal age
breast cancer and hormone cesarean delivery and,
135136 congenital malformation and
life expectancy in East and West folate fortification, 9192
Germany, 99100 Medical Birth Registry of Norway,
Leisure-time physical activity interbirth interval and pre-
(LTPA), health status, 105106 eclampsia, 117118
Life expectancy, East and West Medication, transition cell cancer
Germany, 99100 of urinary tract, 171172
Lifetime probability, cancer, 56 Medroxyprogesterone acetate,
Lipophilic statin, breast cancer, women and coronary heart
5960 disease, 8182
Liver cancer Melanoma cancer, suicide among
alpha-L-fucosidase, 209210 men, 3334
2 5 2 | INDEX
Men. See also Black men; Myocardial infarction

White men cigarette smoking , 1516
oral cancer, 3536 influenza vaccination, 163164
prostate cancer and marital women, 81
status, 8586
smoking and race, 7778 National Cancer Institute,
Menarche, age-at-, and female Surveillance, Epidemiology,
athletes, 6768 and End Results, 197198
Mesothelioma, occupational National Health Interview Survey,
exposure, 165166 smoking , 9798
Minnesota, colorectal cancer Neck trauma, diffuse pain
screening, 221222 syndrome (DPS), 143144
Monophasic oral contraceptive, Neurological disease, Guillain
abortion, 107108 Barr syndrome, 3738
Mortality New York City, cardiac arrest by
bloodletting and pneumonia, race, 7374
127128 New Zealand, automobile crashes
brain cancer in firefighters, and car color, 8990
137138 Night driving, snowmobile,
breast and colon cancer, 139140 159160
breast cancer and hormone Non-Hispanic men, oral cancer,
135136 Non-Hodgkin lymphoma of
colorectal cancer, 221222 gastrointestinal tract (GINHL),
cosmetic breast implants, occupational asbestos,
121122 197198
DuPont Washington Works Nonsteroidal anti-inflammatory
plant, 125126 drugs (NSAIDs)
East and West Germany, 99100 gastric ulcers, 169170
Gulf War veterans, 129130 upper gastrointestinal bleeding
hepatocellular carcinoma (HCC), (UGIB), 205206
2122 Norway, interbirth interval and
prostate cancer, 131132 pre-eclampsia, 117118
screening ovarian cancer,
227228 Occupation, exposure to
U.S. cancer, 195196 mesothelioma, 165166
Motorcycle accidents, snowmobile Occupational asbestos, non-
deaths and alcohol use, Hodgkin lymphoma of
159160 gastrointestinal tract (GINHL),
Mucinous tumors, race, 78 197198
Multiple causal pathways, 175 Odds ratio (OR), risk of AZT, 84
INDEX | 253
Oral cancer, Hispanic and non- Pneumococcal vaccination, elderly

Hispanic men, 3536 patients, 123124
Oral contraceptives Pneumonia
abortion, 107108 bloodletting , 127128
air travel, 6162 breast-fed infants, 151152
venous thrombosis and factor V incidence by age, 7576
Leiden mutation, 179180 vaccination, 123124
Ovarian cancer Postmenopausal women, breast
anthropometric variables, cancer and hormone therapy,
133134 191192
biomarker lysophosphatidic acid Postwar mortality, Gulf War
(LPA), 211212 veterans, 129130
race, 78 Pre-eclampsia, interbirth interval
screening tumor marker, and, 117118
227228 Pregnancy
analgesic and abortion, 5758
Pap smear screening, cervical antidepressant drugs, 4950
cancer, 225226 elective induction of labor,
Parental bedsharing, sudden infant 145146
death syndrome (SIDS), folic acid, 6364
4546 interbirth interval and pre-
Patent foramen ovale, cryptogenic eclampsia, 117118
stroke and, 173174 oral contraceptives and abortion,
Patient incidental finding (PI), 107108
breast cancer, 215 perinatal death among twins,
Peptide in blood, congestive heart 201202
failure, 213214 physical violence, education and,
Persian Gulf War, postwar intendedness, 199200
mortality of veterans, Prematurity, SSRIs and pregnancy,
129130 4950
Pertussis, children with asthma Prevention. See also Vaccination
and, 109110 causal inference, 43
Photofluorography, stomach cancer, Probability, cancer, in lifetime, 56
223224 Progesterone receptor (PR), breast
Physical activity, endometrial cancer type, 5960
cancer risk, 189190 Proportion, term, 3
Physical violence, education and Proportional incidence, lung cancer,
pregnancy intendedness, 910
199200 Prostate cancer
Physicians, suicide by gender, age and marital status, 8586
3132 American men by race, 2324
2 5 4 | INDEX
Prostate cancer (Continued) colorectal cancer, 217218,

case-control study, 235236 221222
diagnosis and death rate, 4142 lysophosphatidic acid (LPA) for
mortality, 131132 ovarian cancer, 211212
PSA screening , 229230 ovarian cancer by tumor marker,
surgery and radiation therapy, 227228
6566 prostate cancer, 4142
PSA (prostate-specific antigen) PSA, for prostate cancer,
screening, prostate cancer, 229230, 235236
4142, 229230, 235236 smoking and lung cancer, 5152
stomach cancer, 223224
Race Seat belts, automobile crashes,
cardiac arrest, 7374 2728
cigarette smokers, 9798 Second National Health and
ovarian cancer, 78 Nutrition Examination
Pap smear screening for cervical Survey, 19
cancer, 225226 Sensitivity, screening test, 207
prenatal care and marital status, Serum lipids, diet and, 5354
113114 Sexually transmitted infection
prostate cancer in U.S., 2324 (STI), condom use,
smoking by men, 7778 157158
Radiation therapy, prostate cancer Sleeping position, sudden infant
and mortality, 6566 death syndrome (SIDS),
Rate, term, 3 167168
Renal cell carcinoma, patients with Smoking. See also Cigarette
analgesic nephropathy, 1718 smoking
Retinol levels, liver cancer risk, colorectal cancer, 2526
203204 health status and physical
Rosuvastatin, venous activity, 105106
thromboembolism (VTE), lung cancer, 5152
187188 myocardial infarction, 1516
sex, race/ethnicity and age,
Screening, 207 9798
abdominal imaging for aortic white and black men, 7778
aneurysm, 233234 Snowmobile deaths, alcohol use
alpha-L-fucosidase and liver and, 159160
cancer, 209210 Spray painting, lung cancer deaths
breast and colon cancer, among automobile workers,
139140 155156
breast cancer, 215216 SSRIs, pregnancy, 4950
INDEX | 255
Standardized mortality Thyroid cancer, suicide among men,

ratio (SMR) 3334
brain cancer in firefighters, Tobacco smoking. See also Cigarette
137138 smoking; Smoking
DuPont Washington Works white and black men, 7778
plant, 125126 Traffic accidents, seat belts,
Statin, breast cancer, 5960 2728
Stomach cancer Transient ischemic attacks (TIA),
carcinogen A in tea, 3940 warfarin anticoagulation and
screening, 223224 stroke risk, 141142
Stroke Transitional cell cancer of urinary
cryptogenic, and patent foramen tract, acetaminophen and,
ovale, 173174 171172
ischemic, and influenza vaccine, Travel, oral contraceptive and air,
163164 6162
transient ischemic attacks (TIA) Triphasic oral contraceptive,
and warfarin, 141142 abortion, 107108
Stuttering, family history of, Tumor marker, ovarian cancer,
149150 227228
Sudden infant death syndrome Twins, perinatal death and birth
(SIDS) order, 201202
bedsharing, 4546
sleeping position, 167168 Ulcers, gastric, nonsteroidal anti-
Suicide inflammatory drugs (NSAIDs),
American physicians by gender, 169170, 205206
3132 United Kingdom, lung cancer by
marital status, 101102 gender, 910
men with cancer, 3334 United States
Surgery age of women for first birth,
cataract patients, 1314 2930
prostate cancer and mortality, blood lead levels and eating
6566 unusual substances, 1920
Swedish Medical Birth Registry, cancer incidence, 9394
disability and in vitro cesarean delivery and maternal
fertilization, 7980 age, 103104
Swedish men, prostate cancer cigarette smokers, 115116
mortality, 131132 death rates, 7172
endometrial cancer, 1112
Tea with carcinogen A, stomach folic acid and pregnancy,
cancer, 3940 6364
2 5 6 | INDEX
United States (Continued) Warfarin, transient ischemic

pneumonia by age, 7576 attacks (TIA) and stroke,
prostate cancer by race, 2324 141142
suicide of physicians by gender, Water, arsenic in, and bladder
3132 cancer, 8788
Upper gastrointestinal bleeding West Virginia, mortality at DuPont
(UGIB), nonsteroidal anti- Washington Works plant,
inflammatory drugs (NSAIDs), 125126
205206 White men
Urinary tract, transition cell cancer marital status and suicide,
of, and acetaminophen, 101102
171172 prostate cancer, 2324
U.S. National Maternal and Infant prostate cancer and marital
Health Survey, 55 status, 8586
US News and World Report, cancer smoking , 7778
incidence, 9394 White women
depression and age at first
Vaccination birth, 55
asthma and pertussis, 109110 marital status and suicide,
cardiovascular disease and 101102
influenza, 163164 Pap smear screening for cervical
GuillainBarr syndrome and cancer, 225226
influenza, 3738 prenatal care and marital status,
infectious disease and, 185186 113114
pneumonia, 123124 Women. See also Black women;
Vascular disease, women, 8182 White women
Venesection, pneumonia, 127128 age at birth of first child,
Venous thromboembolism 2930
oral contraceptives and air travel, breast cancer and hormone
6162 replacement therapy (HRT),
rosuvastatin, 187188 135136
Venous thrombosis (VT), factor breast cancer and hormone use,
V Leiden mutation and oral 111112
contraceptives, 179180 breast cancer type, 5960
Veterans, mortality of Gulf War, cervical cancer screening,
129130 231232
Violence, education and pregnancy coronary heart disease,
intendedness, 199200 8182
Vitamin D, fractures in elderly cosmetic breast implants and
people, 183184 mortality, 121122
INDEX | 257
depression and age at first birth, survival and detection of breast

5556 cancer, 215216
endometrial cancer, 1112 venous thrombosis with
endometrial cancer and estrogen factor V Leiden mutation
therapy, 177178 and oral contraceptives,
ovarian cancer, 78 179180
prenatal care and marital status,
113114 Zidovudine (AZT), exposed health-
SSRIs and pregnancy, 4950 care workers, 8384

Exercises in Epidemiology Applying Principles and Methods

Uploaded by

Copyright:

Available Formats

Exercises in Epidemiology Applying Principles and Methods

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Exercises in Epidemiology Applying Principles and Methods

Uploaded by

Copyright:

Available Formats

What is the purpose of the book?

What is the purpose of the book?

What does the book aim to supplement?

What does the book aim to supplement?

EXERCISES IN EPIDEMIOLOGY

This page intentionally left blank

Oxford New York

Copyright 2012 by Oxford University Press, Inc.

Published by Oxford University Press, Inc.

Oxford is a registered trademark of Oxford University Press

All rights reserved. No part of this publication may be reproduced,

Weiss, Noel S., 1941-

1. Rates and Proportions 3

subject. My hope is that the exercises contained in it can extend

Rates and Proportions

EPIDEMIOLOGISTS LOVE denominators. Sometimes we divide

Question 1.1 A recent study observed that 1 in 20 persons

Answer 1.1 This is not necessarily evidence of immunity.

Question 1.2 A study on ovarian cancer observed the follow-

Caucasian 33 (13%) 225 (87%) 258 (100%)

The authors concluded that Asian women had a higher

Answer 1.2 The observed proportional distribution of histo-

Type of tumor (rate per 100,000 woman-years)

Caucasian 1 6.7 7.7

Question 1.3 The following is excerpted from a news item in

The clinical features of more than 1000 patients with lung

Assume that: (1) the distribution of histologic types of lung

Answer 1.3 The proportional incidence by gender will not be

Small cell cancers (%) 360 (18%) 340 (34%)

In the above example, assuming the numbers of men and

Question 1.4 The following statement appeared in a review

In 1996 in the United States, a total of about 34,000 new cases

a. Assuming that the data described in the first sentence

a. If equilibrium existsin other words, no change in

Question 1.5 Lets say youve conducted a cohort study to

Study group No. of persons No. of crashes No. of miles driven

Surgery 174 27 5,677,867

Assume the two groups of patients are exactly comparable

a. Estimate the influence of cataract surgery on crash rate

Provide the rationale for your answers.

a. The answer is the relative rate based on the number of

b. Since the number of miles driven seems to have been

Question 1.6 The following is paraphrased from an article in

Although the relative rate of myocardial infarction associated

Under what circumstance could this be true?

Answer 1.6 It could be true if, among nonsmokers, the inci-

Question 1.7 The following is excerpted from a letter to the

We have observed renal cell carcinomas in 6 out of 412 patients

What additional information on these patients with analge-

Answer 1.7 At the very least, we would need the age-specific

Question 1.8 The Second National Health and Nutrition

Percentage of children 6 months through 4 years with a

Earlier studies have shown that elevated blood lead levels

Blood lead level

Yes 19 162 181

Question 1.9 The following data were obtained in a very

Rate per 100,000

a. For men and women, separately, estimate the relative

Men: 405.2 21.8 = 18.6

Mortality difference (per 100,000 person-years)