Clinical Epidemiology

T H R D E D T o N
Clinical
Epidemiology
THE ESSENTIALS
Robert H. Fletcher, M.D., M.sc.

Professor
Department of Ambulatory Care and Prevention
Harvard Medical School and Harvard Community Health Plan
Department of Epidemiology, Harvard School of Public Health
Department of Medicine, Brigham and Women's Hospital
Boston, Massachusetts
Suzanne W. Fletcher, MD., M.sc.

Professor
Department of Ambulatory Care and Prevention
Harvard Medical School and Harvard Community Health Plan
Department of Epidemiology, Harvard School of Public Health
Department of Medicine, Brigham and Women's Hospital
Boston, Massachusetts
Edward H. Wagner, MD., M.P.H.

Director, Center for I Iealth Studies
Croup Health Cooperative of Puget Sound
Professor of Health Services
University of Washington
Seattle, Washington
Williams & Wilkins

A WAVEHY C()Mr,~"Y
'UJI,."lI'oI'II'' ' >:,"'ow:,,,

'''',"" ",,,; ° "'''''~''';
' o' ' ,"0""oL',<lIJ'· ""(;W'
IY'''''010<\(.'' "H'","
Edilo,- Tirnolily S, Sattcrueld
Managing Edilor: Crystal Taylm
Producti,,,, Coordin"lor; Haymond 1::. nctc-
Copy Editor, Candace B, Levy, Ph.D
Designer: Norman W. Oell
!/Justmli"" Pfilnn",: Mmio l-crnandcz
Typesetter; lapsco, Inc" Akron, Pennsvlvarua
Printer & Binder: VIctor Graphics, In<;., 8;JllirnorB, MHrylimd
Copyrightl1l 1996 Williams & Wilkins
351 West Camden Street

Baltimore, Maryland 2t 201-?436 USA
ROSH Tree Corporate CHnter

1~OO North Providence Road
Building II. Suite 502b
Media, Pennsylvania 190m 20~3 USA
All rights reserve<:j rhis book is protected by copyrighl No part ct this hook may be reprodlJced in
anytorm m by any means, including phctoccpyinq, or utilized by any intormation storagH and retrieval
system without written permission trom the copyright owner
Accurate indications, adverse reHdions and dos-agH schedules for dt1.Jgs are provided in this llOok,
but it is possible that they may changQ. The reader is urgod to review Ihe package informabon oata
of the rnHrlLJfacturers of the medications mentioned.
P"illad in Ihe Unil"d Slatos at America
First roaoo. 1982

Second Edition. 1988
Library of Congre~ Cataloging-in-Publicatlon Data
Fletcher, Robert H.
Clinical epidemloloqy: ue essentials/Bobert H. Fletcher.
Suzanne W. Fletcher, Edward II. Wagner, -- aro ed.
p, cm
Includes bibliographical references and index.
ISm, 06B3-03?69-0
1 Clinical epidemiology, I. Wa(1ner, Edward H. (lidwanl Harris).
t940- II litle.
[DNI M- t Epidemiologic Methods, WA %0 F6t~c 1996]
RAC52.2.C5c.Fb1 1996
6 t 4.4-<ic?O
DNI M/DLC
for Librarv of Cong",ss 8b-8382
CP
The publishers !>iNe mcao every elforl to tmce the cOpyrigll! holders for borrow~'CJ material Illhey
have inadvert"mliy overlooked any, limy WIJI be pleased 10 make the necesswy =angements al the
first opportunity,
9b 96 91 98 99
? 3 4 5 6 7 8 9 10
PREFACE
Since the Second Edition was written in 1988, the pace of change in
medicine has accelerated. Changes have brought greater recognition of the
perspectives and methods of clinical epidemiology.
Countries throughout the world have, in their efforts to provide high-
quality health care, experienced growing difficulties controlling the cost
of care. The tension between demands for care and resources to provide
it have increased the need for better information about clinical effectiveness
in setting priorities. It has become dearer that not all clinical care is effective
and that the outcomes of care are the best way of judging effectiveness.
Variations in care among clinicians and regions, not explained by patients'
needs and not accompanied by similar differences in outcomes, have wised
questions about which practices are best. All these forces in modern society
have increased the value of good clinical research and of those who can
perform and interpret this research properly.
Phenomenal advances in understanding the biology of disease, espe-
cially at the molecular level, have also occurred. Discoveries in the labora-
tory increase the need for good patient-based research. They must be tested
in patients before they can be accepted as clinically useful. Thus the two-
laboratory science and clinical epidemiology-complement each other and
are not alternatives or competitors.
Other aspects of medicine are timeless. Patients and physicians still face
the same kinds of questions about diagnosis, prognosis, and treatment and
still value the same outcomes: to relieve suffering, restore function, and
prevent untimely death. We rely on the same basic strategies (cohort and
case-control studies, randomized trials, and the like) to answer the ques-
tions. The inherent uncertainty of all clinical information, even that based
on the best studies, persists.
In preparing the third edition of this text, we have tried to take into
account the sweeping changes in medicine as well as what has not changed.
We have left the basic structure of the book intact. We updated examples
throughout in recognition that some diseases, such as AIDS, are new and
others, such as peptic ulcer disease, arc better understood.
We have tried to remember that the book's niche is as an introduction
v
vi PREFACE
to clinical epidemiology and to avoid pitching the presentation to our

colleagues who already have a firm grasp of the basics. The presentation
is meant to be as simple as the topic allows. However, the field is covered
in somewhat greater depth on the belief that readers expect more of the
book now than they did when the field was new.
This edition is still primarily for clinicians who wish to develop a sys-
tematic understanding of how the evidence base for patient care is devel-
oped and assessed. Researchers begin with many of the same basic needs.
The text should be useful at any level of clinical training: from medical
student to practicing clinician.
ACKNOWLEDGMENTS
We have been privileged to work with wonderful colleagues, many of

them clinical epidemiologists. This edition of the text distills what we
have learned from them and from our own efforts to use clinical epidemi-
ology. Our teachers were the pioneers of this field: Archie Cochrane, John
Cassell, Alvan Feinstein, David Sackett, Kerr White, and others. More
recently we have lear-ned from the many young physicians throughout
the world who have believed in this way of thinking about health and
medical care and have put their beliefs into practice. Many are members
of distinct programs-such as the International Clinical Epidemiology
Network (INC LEN) and the Robert Wood Johnson Clinical Scholars Pro-
gram-but the number of groups has become too numerous to list each
one here.
The Flcichers are especially grateful to the editorial staff of Annals of
Internal Medicine, our colleagues during the four ye<lfs we were Editors-
in-Chief. Articles in peer-reviewed clinical journals are the written record
of clinical epidemiology, and we learned a great deal from our efforts to
select and improve manuscripts. We were able to begin work on the third
edition of this text-c-a daunting task in the midst of a move from the
American College of Physicians in Philadelphia to Boston-thanks to the
hospitality of Drs. Steven Jones and Wendy Levinson of the Legacy Port-
land Hospitals, our hosts while we were on sabbatical in Oregon. Tom
Inui invited us to our new academic home-the Department of Ambula-
tory Care and Prevention at Harvard Medical School and Harvard Commu-
nity Health Plan-and provided all the fellowship and resources we
needed to finish the writing.
Ed Wagner is grateful for having been given the privilege to explore the
uses of clinical epidemiology in studying, planning, and improving care
for the 0.5 million enrollees of Group Health Cooperative of Puget Sound.
My contribution to this edition owes much to the stimulation and support
of my colleagues at the Center for Health Studies, the Group Health deliv-
ery system, at the University of Washington.
This edition reflects all of our experiences from the time the three of us
vii
viii ACKNOWLEDGMENTS
were together on the faculties of Medicine and Epidemiology at the Univer-

sity of North Carolina at Chapel Hill. Now, 13 years later and in different
places, we appreciate more than ever the importance of the many col-
leagues who have helped us improve our understanding of clinical epide-
miology and to express it more clearly.
CONTENTS
Preface , v
Acknowledgments , vii
Chapter 1
INTRODUCTION , . . .. 1
Chapter 2
ABNORMALITY. ...... 19
Chapter 3
DIAGNOSIS .... , 43
Chapter 4
FREQUENCY ..... 75
Chapter 5
H1SK. . 94
Chapter 6
PROGNOSIS , . . .. 111
Chapter 7
TREATMENT. .............. , ,.136
Chapter 8
PREVENTION .. . , , 165
Chapter 9
CHANCE .. . .... 186
Chapter 10
STU DYING CASES , , . . 208
Chapter 11
CAUSE . . ... 228
Chapter 12
SUMMING UP ... ......... , ... 249
Index 271
1
INTRODUCTION
A 51-year-old man asks to sec you because of chest pain. He was well
until 2 weeks ago, when he noticed tightness in the center of his chest
while walking uphill. The tightness stopped after 2 to 3 min of rest. A
similar discomfort occurred several times since then, sometimes during
exercise and sometimes at rest. He smokes one pack of cigarettes per day
and has been told that his blood pressure is "a little high." He is otherwise
well and takes no medications. However, he is worried about his health,
particularly about heart disease. A complete physical examination
and resting electrocardiogram are normal except for a blood pressure of
150/96.
This patient is likely to have many questions. Am 1 sick? How sure are
you? If Tam sick, what is causing my illness? How will it affect me? What
can be done about it? How much will it cost?
As the clinician caring for this patient, you must respond to these ques-
tions and use them to guide your course of action. Is the probability of
serious, treatable disease high enough to proceed immediately beyond
simple explanation and reassurance to diagnostic tests? How well do vari-
ous tests distinguish among the possible causes of chest pain: ilngina pecto-
ris, esophageal spilsm, muscle strain, anxiety, and the like. For example,
how helpful will an exercise electrocardiogram be in either confirming or
ruling out coronary artery disease? Tf coronary disease is found, how long
can the patient expect to have the pain? Will the condition shorten his
life? How likely is it that other complications-congestive heart failure,
myocardial infarction, or atherosclerotic disease of other organs-will oc-
cur? Will reduction of his risk factors for coronary diseasc-c-cignrctte smok-
ing and hypertension-reduce his risk? If medications control the pain,
should the patient have coronary artery bypass surgery anyway?
Clinicians use various sources of information to answer these questions:
their own experiences, the advice of their colleagues, and the medical
literature. In general, they depend on past observations on other similar
patients to predict what will happen to the patient at hand. The manner
1
2 CLINICAL EPIDEMIOLOGY
in which such observations are made and interpreted determines whether

the conclusions they reach are valid, and thus how helpful the conclusions
will be to patients.
Clinical Epidemiology
Clinical epidemiology is the science of making predictions about indi-
vidual patients by counting clinical events in similar patients, using strong
scientific methods for studies of groups of patients to ensure that the pre-
dictions are accurate. The purpose of clinical epidemiology is to develop
and apply methods of clinical observation that will lead to valid conclu-
sions by avoiding being misled by systematic error and chance. It is one
important approach to obtaining the kind of information clinicians need
to make good decisions in the care of patients.
CLINICAL MEDICINE AND EPIDEMIOLOGY
The term clinical epidemiology is derived from its two parent disciplines:
clinical medicine and epidemiology. It is "clinical" because it seeks to
answer clinical questions and to guide clinical decision making with the
best available evidence. It is "epidemiologic" because many of the methods
used to answer these questions have been developed by epidemiologists
and because the care of individual patients is seen in the context of the
larger population of which the patient is a member.
Clinical medicine and epidemiology began together (1). The founders
of epidemiology were, for the most part, clinicians. It is only during this
century that the two disciplines drifted apart, with separate schools, train-
ing, journals, and opportunities for employment. More recently, clinicians
and epidemiologists have become increasingly aware that their fields inter-
relate and that each is limited without the other (2).
TRADITIONAL CLINICAL PERSPECTIVE
Clinicians have a special set of experiences and needs that has condi-
tioned how they go about answering clinical questions. They are, by and
large, concerned with individual patients. They know all of their patients
personally; take their own histories; do their own physical examinations;
and they accept an intense, personal responsibility for each patient's wel-
fare. As a result, they tend to see what is distinctive about each one and
are reluctant to lump patients into crude categories of risk, diagnosis, or
treatment and to express patients' membership in these categories as a
probability.
Because their work involves the care of a succession of individual pa-
tients and is demanding in its own right, clinicians tend to be less interested
in patients who have not come to their attention because they are in some
other medical setting or are not under medical care at all-even though
these patients may be just as sick as the patients they see.
CHAPTER 1 I INTRODUCTION 3
Clinical training has been oriented toward the mechanisms of disease

through the study of biochemistry, anatomy, physiology, and other tradi-
tional basic sciences. These sciences powerfully influence medical students
during their formative years and are a predominant force in clinical re-
search and publications. This training fosters the belief that to understand
the detailed processes of disease in individual patients is to understand
medicine. The implication is that one can predict the course of disease and
select appropriate treatments through knowledge of the mechanisms of
disease.
THE NEED FOR AN ADDITIONAL "BASIC SCIENCE"
This traditional approach serves clinicians well under the right circum-
stances. It has identified many promising interventions-for example, vac-
cines, antimicrobial and vasoactive drugs, and synthetic hormones. It
works well for correcting acid-base abnormalities and diagnosing and
treating nerve compressions.
However, clinical predictions from knowledge of the biology of disease
should ordinarily be considered hypotheses, to be tested by clinical re-
search, about what might transpire in patients, because the mechanisms
are only partly understood and many other factors in the genetic, physical,
and social environments also affect outcome. For example, it has been
shown, despite predictions to the contrary, that feeding diabetics simple
sugars produces no worse metabolic effects than feeding them complex
sugars, that some antiarrhythmic drugs actually cause arrhythmias, and
that drugs that favorably affect the rheologic properties of sickle cells do
not necessarily reduce the frequency and severity of sickle cell crises.
Personal experience is also a guide to clinical decision making. However,
no one clinician can have enough direct experience to recognize all the
subtle, long-term, interacting relationships that characterize most chronic
diseases (see Chapter 6).
Therefore, for clinicians who intend to make up their own minds about
the soundness of clinical information, some understanding of clinical epi-
demiology is as necessary as an understanding of anatomy, pathology,
biochemistry, and pharmacology. Indeed, clinical epidemiology is one of
the basic sciences, a foundation on which modern medicine is practiced.
ELEMENTS OF CLINICAL EPIDEMIOLOGY
Personal experience and medicine's basis in the biology of disease are
both valuable, but they do not take into account some of the realities of
clinical science, which might be summarized as follows:
• In most clinical situations the diagnosis, prognosis, and results of treat-

ment are uncertain for individual patients and, therefore, must be ex-
pressed ,1S probabilities
• Probability for an individual patient is best estimated by referring to

past experience with groups of similar patients
• Because clinical observations are made on people who are free to do as
they please and by clinicians with variable skills and biases, the observa-
tions may be affected by systematic errors that can cause misleading
conclusions
• All observations, including clinical ones, are also influenced by the play
of chance
• To avoid being mislead, clinicians should rely on observations that are
based on sound scientific principles, including ways to reduce bias and
estimate the role of chance.
THE SOCIAL CONTEXT OF CLINICAL EPIDEMIOLOGY
Important forces in modern society have accelerated the recognition of
clinical epidemiologic methods and perspectives. The costs of medical care
are rising beyond the point where even the most affluent societies are able
to pay for all the care people want. Studies have shown wide variation in
clinical practices without corresponding variation in outcomes of care,
suggesting that not all common and expensive practices are useful. More
rigorous methods of evaluating clinical evidence are being developed and
are valued by decision makers. These observations have led to the consen-
sus that clinical care should be based on the strongest possible research
and should be judged by the outcomes it achieves at a cost society can
afford. Also, individual patients are increasingly seen in relation to the
larger group of which they are members, both to make accurate predictions
about them and to assist in deciding which uses of limited medical re-
sources do the most good for the most people.
Basic Principles
The basic purpose of clinical epidemiology is to foster methods of clini-
cal observation and interpretation that lead to valid conclusions. The most
credible answers to clinical questions are based on the following principles.
CLINICAL QUESTIONS
Types of questions addressed by clinical epidemiology are listed in
Table 1.1. These are the same questions confronting the doctor and patient
in the example presented at the beginning of this chapter. They are at issue
in most doctor-patient encounters.
HEALTH OUTCOMES
The clinical events of primary interest in clinical epidemiology are the
health outcomes of particular concern to patients and those caring for them
(Table 1.2). They are the events doctors try to understand, predict, interpret,
and change when caring for patients. An important distinction between
CHAPTER 1 ! INTRODUCTION 5
Table 1.1
Clinical Questions
Question
Abnormality Is the patient sick or well?

Diagnosis How accurate are tests used to diagnose disease?
Frequcrcv How otton docs a disease occur')
Risk What factors are associated with an increased risk ot disease?
Prognosis What are the consequences of having a disease?
Treatment How does treatment change the course of disease?
Prevention Docs an mtarconton on wcll people keep disease from arising? DOGS
early detection and treatment improve the course of disease?
Cause What conditions lead to disease? What me the pathogenetic
mocharusms ot disease')
Cost How much will care for an illness cost?
Table 1.2
Outcomes of Disease {the Five Ds)B
Dcath A bad outcome it untimely

Disease" A set of symptoms, physical signs, and laboratory abnormalities
Discomfort Symptoms such as pain, nausea, dyspnea, itching, and tinnitis
Disability Impaired ability to go about usual activities at home, work, or recreation
Dissatisfaction Emotionl reaction to disease and its care, such as sadness or anger
Perhaps a sixth 0, destitution, belongs on this list because the financial cost of illness (for individual patients
or society) is an important consequence of orseeso.
"Or illness, the patient's experience of disease
clinical epidemiology and other medical sciences is that the events of inter-
est in clinical epidemiology can be studied directly only in intact humans
and not in animals or parts of humans, such as humeral transmitters, tissue
cultures, cell membranes, and genetic sequences,
Biologic outcomes cannot properly be substituted for clinical ones with-
out direct evidence that the two are related. Table 1.3 summarizes some
biologic and clinical outcomes for the modern treatment of a patient with
human immunodeficiency virus (HIV) infection. It is plausible, from what
is known about the biology of HIV infection, that clinical outcomes such
as opportunistic infections, Kaposi's sarcoma, and death would be better
if an intervention reduced the decline in CD4+ cell counts and p34 antigen.
However, there is evidence that these are incomplete markers of disease
progression and response to treatment. It is too much to assume that patient
outcomes would improve as a result of the intervention just because bio-
logic markers do, because many other factors might determine the end
result. Clinical decisions should, therefore, be based on direct evidence
that clinical outcomes themselves arc improved.
Table 1.3
Biologic and Clinical Outcomes; Treatment of Human Immunodeficiency
Virus Infection
Outcomes
Interventions Biolcqlc Clinical
HIV infection Zidovudine CD4+ counts Opportunistic infections

001 p24 antigenemia Quality of life
DOC Viremia Death
Association known or assumed?
NUMBERS AND PROBABILITY

Clinical science is at its strongest when measurements are quantitative,
in part because numerical information allows better confirmation, more
precise communication among clinicians and between clinicians and pa-
tients, and estimation of error. Clinical outcomes such as death, symptoms,
or disability, can be counted and expressed as numbers. Although qualita-
tive observation is also important in clinical medicine, it is not part of
clinical epidemiology.
Individual patients will either experience a clinical outcome or not, but
predicting whether or not an individual will do so is seldom exact. Rather,
clinicians use the results of research to assign probabilities that the outcome
will occur. The clinical epidemiologic approach accepts that clinical predic-
tions are uncertain, but can be quantitated, by expressing predictions as
probabilities-c-for example, that symptomatic coronary disease occurs in
1 in 100 middle-aged men per year, that cigarette smoking doubles one's
risk of dying at all ages, and that exogenous estrogens reduce the risk of
fractures from osteoporosis by half.
POPULATIONS AND SAMPLES
In general, poputatione are large groups of people in a defined setting
(such as North Carolina) or with a certain characteristic (such as age >65
years). These include relatively unselected people in the community, the
usual population for epidemiologic studies of cause, as well as groups of
people selected because of their attendance in a clinic or hospital or because
of a characteristic such as the presence or severity of disease, as is more
often the case in clinical studies. Thus one speaks of the general population,
a hospitalized population, or a population of patients with a specific
disease.
A sample is a subset of a population and is selected from the population.
Clinical research is ordinarily carried out on samples. One is interested in
the characteristics of the defined population but must, for precttca1reasons,

estimate them by describing a sample.
BIAS
Bias is "a process at any stage of inference tending to produce results
that depart systematically from the true values" (3). Suppose, for example,
that treatment A is found La work better than treatment B. What kinds of
biases might have brought about this observation if it were not true? Per-
haps A is given to healthier patients than is B; then the results could be
due to the systematic difference in health between the groups of patients
whether or not they were treated rather than to differences in the effective-
ness of treatment. Or A might taste better than B so that patients take the
drug more regularly. Or A might be a new, very popular drug and B an
old one, so that researchers and patients are more inclined to think that
the new drug works better whether or not it really does. All of these are
examples of potential biases.
Observations on patients (whether for patient care or research) are par-
ticularly susceptible to bias. The process tends to be just plain untidy. As
participants in a study, human beings have the disconcerting habit of doing
as they please and not necessarily what would be required for producing
scientifically rigorous answers. When researchers attempt to conduct an
experiment with them, as one might in a laboratory, things tend to go
wrong. Some people refuse to participate, while others drop out or choose
another treatment. What is more, some of the most important things about
humans-feelings, comfort, performance-are generally more difficult to
measure than physical characteristics, such as blood pressure or serum
sodium. Then, too, clinicians arc inclined to believe that their therapies are
successful. (Most patients would not want a physician who felt otherwise.)
This attitude, so important in the practice of medicine, makes clinical obser-
vations particularly vulnerable to bias.
Although dozens of biases have been defined (4), most fall into one of
three broad categories (Table 1.4).
Sdection bias occurs when comparisons are made between groups of p"-
bents that differ in ways. other than the main factors under study, that affect
Table 1.4
Bias in Clinical Observation
• Selection bidS occurs whon comparisons are made between groups of patients that differ
in of outcome other than the one under study,
(jeterrnillant~;
• Measurement bl~gs
occurs when the methods of measurement me ossmsor among
groups of patients,
• Confounding bias occurs when two teeters arc ossooctoo {"'trilvet toqether"} and H18
effect of one is confused with or distorted by tile ottect of the ou-er
the outcome of the study. Croups of patients often differ in many ways-
age, sex, severity of disease, the presence of other diseases, and the care they
receive. If we compare the experience of two groups that differ on a specific
characteristic of interest (for example, a treatment or a suspected cause of
disease) but are dissimilar in these other ways and the differences are them-
selves related to outcome, the comparison is biased and little can be concluded
about the independent effects of the characteristic of interest. In the example
used earlier, selection bias would have occurred if patients given treatment
A were healthier than those given treatment B.
Measurement bias occurs when the methods of measurement are consis-
tently dissimilar in different groups of patients. An example of a potential
measurement bias would be the use of information taken from medical
records to determine if women on birth control pills were at greater risk
for thromboembolism than those not on the Pill. Suppose a study were
made comparing the frequency of oral contraceptive use among women
admitted to a hospital because of thrombophlebitis and a group of women
admitted for other reasons. It is entirely possible that women with throm-
bophlebitis, if aware of the reported association between estrogens and
thrombotic events, might report use of oral contraceptives more completely
than women without thrombophlebitis, because they had already heard
of the association. For the same reasons, clinicians might obtain and record
information about oral contraceptive use more completely for women with
phlebitis than for those without it. If so, an association between oral contra-
ceptives and thrombophlebitis might be observed because of the way in
which the history of exposure was reported and not because there really
is an association.
Confounding bias occurs when two factors are associated with each other,
or "travel together," and the effect of one is confused with or distorted by
the effect of the other. This could occur because of selection bias, by chance,
or because the two really are associated in nature.
example Is herpesvirus infection a cause of cervical cancer) It has been
consistently observed that the prevalence of herpesvirus infection is higher
in women with cervical cancer than in those without. However, both herpes-
virus and a number of other infectious agents, themselves possible causes of
cervica I cancer,arc transmitted by sexual contact.In particular, there i~ strong
evidence that human papillomavirus infection leads to cervical cancer. Per-
haps the higher prevalence of herpesvirus infection in women with cervical
cancer is only a consequence of greater sexual activity and so is indirectly
related to a true cause, which is also transmitted sexually (Fig. 1.1). To show
that herpesvirus infection is associated with cervical cancer independently
of other agents, it would be necessary to observe the effects of herpesvirus
Ircc of the other factors related to increased sexual activity (5).
Selection bias and confounding bias are not mutually exclusive. They arc
described separately, however, because they present problems at different
MAIN QUESTION
I
sexual activity
CONFOUNDING
FACTORS
!
HPV
I
Figure 1.1. Confounding bias: Is herpesvirus 2 (HSV-2) a possible cause of cervical
cancer? Only if its association with cervical cancer is independent of human papillo-
mavirus (HPV) infection, known to be a cause of cervical cancer. Both viruses are
related to increased sexual activity.
points in a clinical observation or study. Selection bias is at issue primarily

when patients arc chosen for observation, and so it is important in the
design of a study. Confounding bias must be dealt with during analysis
of the data, once the observations have been made.
Often in the same study more than one bias operates. i1S in the following
hypothetical example.
Example A study was done to determine whether reglJl;u exercise low-
ers the risk of coronary heart disease (CHD). An exercise program was offered
to employees of a plant, and the rate of subsequent coronary events was
compared between employees who volunteered for the program and those
who did not volunteer. Coronary events were determined by means of regu-
lar voluntary checkups, including a careful history, an electrocardiogram,
and a review of routine health records. The group that exercised had lower
rates of CHD. However, they also smoked cigarettes less.
In this example, selection bias could be present if volunteers for the

exercise program were at lower risk for coronary disease even before the
program began-for example, because they had lower serum lipids or less
family history of coronary disease. Measurement bias might have occurred
because the exercise group stood a better chance of having a coronary
event detected, because more of them were examined routinely. Finally,
the conclusion that exercise lowered the risk of coronary disease might be
the result of a confounding bias if the association between exercise and
coronary events in this particular study resulted from the fact that smoking
cigarettes is a risk factor for coronary disease and the exercise group
smoked less.
A potential for bias does not mean that bias is actually present in a
particular study. For a researcher or reader to deal effectively with bias,
it is first necessary to know where and how to look for it and what can
be done about it. But one should not stop there. It is also necessary to
determine whether bias is actually present and how large it is likely to be

to decide whether it is important enough to change the conclusions of the
study in a clinically meaningful way.
CHANCE
Observations about disease are ordinarily made on a sample of patients

rather than all those with the disease in question. Observations on a sample
of patients, even if unbiased, may misrepresent the situation in the popula-
tion as a whole because of chance. However, if the observations were
repeated on many such patient samples, results for the samples would
vary about the true value. The divergence of an observation on a sample
from the true population value, due to chance alone, is called random
rariaticn.
We are all familiar with chance as an explanation for why a coin does
not come up heads exactly 50'~<, of the time when it is flipped, say, 100
times. The same effect, random variation, applies when assessing the effects
of treatments A and 13, discussed earlier. Suppose all biases were removed
from a study of the relative effects of the two treatments. Suppose, further,
that the two treatments are, in reality, equally effective, each improving
about 50'10 of the patients treated. Even so, because of chance alone a single
study including small numbers of patients in each treatment group might
easily find A improving a larger proportion of patients than 13 or
vice versa.
Chance can affect all of the steps involved in clinical observations. In
the assessment of treatments A and B, random variation occurs in the
sampling of patients for the study, the selection of treatment groups, and
the measurements made on the groups.
Unlike bias, which deflects values in one direction or another, random
variation is as likely to result in observations above the true value as below
it. As a consequence, the mean of many unbiased observations on samples
tends to correspond to the true value in the population, even though the
results of individual small samples may not.
Statistics can be used to estimate the probability of chance (random
variation) accounting for clinical results. A knowledge of statistics can also
help reduce that probability by allowing one to formulate a better design
and analyses. However, random variation can never be totally eliminated,
so chance should always be considered when assessing the results of clini-
cal observations.
The relationship between bias and chance is illustrated in Figure 1.2.
The measurement of diastolic blood pressure on a single patient is taken
as an example. True blood pressure can be obtained by an intraarterial
cannula, which is 80 mm Hg for this patient. But this method is not possible
for routine measurements; blood pressure is ordinarily measured indi-
True Blood
blood pressure
.,r::: pressure
(intraarterial cannula)
measurement
(sphygmomanometer)
o
~
1lo l.. •••••• •• ••
•• .. t
'0
~
1:
I. Chance •
E
:>
z ----Bias----
80 90
Diastolic Blood Pressure (mm Hg)
Figure 1.2. Relationship between bias and chance: Blood pressure measurements
by intraarterial cannula and sphygmomanometer.
rectly, using a sphygmomanometer. The simpler instrument is prone to

error, or deviations from the true value. In the figure, the error is repre-
sented by all of the sphygmomanometer readings falling to the right of
the true value. The deviation of sphygmomanometer readings to the right
(bias) may have several explanations-for example, a poorly calibrated
sphygmomanometer, the wrong cuff size, or a deaf clinician. Bias could
also result if different sounds were chosen to represent diastolic blood
pressure. The usual end points-phase IV and phase V Korotkoff
sounds-tend to be above and below the true diastolic pressure, respec-
tively, and even that is unpredictable in obese people. Individual blood
pressure readings are also subject to error because of random variation
in measurement, as illustrated by the spread of the sphygmomanometer
readings around the mean value (90 mm Hg).
The two sources of error-bias and chance-are not mutually exclusive.
In most situations, both are present. The main reason for distinguishing
between the two is that they are handled differently.
Bias can in theory be prevented by conducting clinical investigations
properly or corrected through proper data analysis. Tf not eliminated, bias
often can be detected by the discerning reader. Most of this book is about
how to recognize, avoid, or minimize bias. Chance, on the other hand,
cannot be eliminated, but its influence can. be reduced by proper design

of research, and the remaining: error can be estimated by statistics. Statistics
can also help remove the effects of known biases. However, no amount of
statistical treatment can correct for unknown biases in data. Some statisti-
cians would go so far as to suggest that statistics not be applied to data
vulnerable to bias because of poor research design, for fear of giving false
respectability to misleading work.
INTERNAL AND EXTERNAL VALIDITY
When making: inferences about a population from observations on a

sample, two fundamental questions arise (Fig. 1.3): First, are the conclu-
sions of the research correct for the people in the sample? Second, if so,
does the sample represent fairly the population of interest?
Internal validity is the degree to which the results of a study are correct
for the sample of patients being studied. It is "internal" because it applies
to the conditions of the particular group of patients being observed and
not necessarily to others. The internal validity of clinical research is deter-
mined by how well the design, data collection, and analyses are carried
out and is threatened by all of the biases and random variation discussed
above. For a clinical observation to be useful, internal validity is a necessary
but not sufficient condition.
External validity (genl'ralizability) is the degree to which the results of an
All patients with the INTERNAL

condition of interest VALIDITY
selection
bias
~
measurement,
~
? - confounding bias
-- ... chance 1
EXTERNAL CONCLUSION
VALIDITY
(generalizability)
Figure 1.3. Internal and external validity,

observation hold true in other settings. for an individual physician, it is

an answer to the question, "Assuming that the results of a study are true,
do they apply to my patient as well?" Gcneralizability expresses the valid-
ity of assuming that patients in a study arc comparable with other patients.
An unimpeachable study, with high internal validity, may be totally
misleading if its results are genera lizcd to the wrong patients.
Example What is the risk that an abdominal aortic aneurysm will rup-
ture? Clinicians seeing patients with ant:'urysms must have this information
to make wise decisions about the need for elective surgical repair. The answer
depends on which kinds of patients arc described. Among patients with
aneurysms <5 cm in diameter, above which surgt:'ry is commonly advised,
those seen in referral centers have about a 10 times greater rate of rupture
during 5 yt:'ars of follow-up than those in the general population (Fig 1.4)
(6). This may be because patients in centers are referred for symptoms or
signs of impending rupture. If clinicians in office practice were to use the
results of research from rcfcrre 1centers to predict rupture, they would greatly
overestimate the risk and perhaps make the wrong decision about the need
for elective surgical repair.
The gcncraltzabtltty of clinical observations, even those with high inter-

nal validity, is a matter of opinion about which reasonable people might
disagree.
Example The Physician's Health Study showed that low-dose aspirin
(325 mg every other day) prevented myocardial infarction in male physicians
30
•"
••,..
-l
~
20
~
•
"' 10
Population Referral
centers
Figure 1.4. Sampling bias: Range of risk of rupture (shaded area) in the next 5
years of abdominal aortic aneurysm «5.0 cm in diameter) according to whether the
patient is from the general population or a referral center (6).
14 CLINICAL ~PIDEMIOLOGY
without known coronary heart disease (7). The 11,037 physicians randomly
assigned to take aspirin had a 44~;) lower rate of myocardial infarction than
the 11,034 assigned to take placebo. The study was carorullv conducted and
used a strong research design; its findings have stood up well to criticisms.
However, only healthy male physicians were in the study. When the results
of the study were first released, clinicians had to decide whether it was
justified to i;ive aspirin to women, people with many risk factors, and patients
who arc already known to have coronary disease. Subsequently, reviews of
evidence from all available studies have suggested that aspirin is also effec-
tiVIO in these other groups of people (8).
Ccncrafizabilitv can rarely be dealt with satisfactorily in anyone study.

Even a defined, geographically based population is a biased sample of
larger populations; for example. hospital patients are biased samples of
county residents; counties, of states: states, of regions, and so on. Doing <I
study in many centers m<lY improve gencralizability. but does not settle
the issue.
Usually, the best a researcher can do about generalizilbility is to ensure
internal validity, have the study population fit the research question, and
avoid studying groups so unusual that experience with them generalizes
to few other patients. It then remains for other studies, in other settings,
to extend generellzabiltty.
Sampling bias has occurred when the sample of patients in a study is
systematically different from those appropriate for the research question
or the clinical usc of the information. Because most clinical studies take
place in medical centers and because patients in such centers usually over-
represent the serious cod of the disease spectrum, sampling bias in clinical
research tends to result in an exaggerated view of the serious nature of
disease.
Uses of Clinical Epidemiology

Learning and applying clinical epidemiology adds time to an already
busy clinician's schedule. What can he or she expect in return?
Understanding the strengths and weaknesses of clinical evidence, such
as reports of research, gives intellectual satisfaction and confidence where
there might otherwise be bewilderment and frustration. It can increase
efficiency in acquiring sound information by allowing one to decide
quickly, from basic principles, which articles or sources of clinical informa-
tion are credible. During interaction with colleagues, it provides a sounder
alternative to other ways of deciding where to invest belief in an asser-
tion-c-fhe conviction, rhetoric, seniority, or specialty of the proponent. By
rdying on clinical epidemiology, clinicians of <111 backgrounds are on a
more equal footing, all depending mainly on the interpretation of the same
set of strong studies. Finally, clinical epidemiology gives clinicians a per-
spective on the extent to which their efforts, relative to other factors, such
as the biology of disease and the physical and social environment, deter-
mine health outcomes, so that they can know what they can and cannot
change.
For these reasons, we believe the time invested in learning clinical epide-
miology is more than repaid.
Information and Decisions

The primary concerns of this book are the quality of clinical information
and its correct interpretation. Making decisions is another matter. True,
good decisions depend on good information; but they involve a great deal
more as well, including value judgments and the weighing of competing
risks and benefits.
In recent years, medical decision making has become a valued discipline
in its own right. The field includes qualitative studies of how clinicians
make decisions and how the process might be biased and can be improved.
It also includes quantitative methods-decision analysis, cost-benefit anal-
ysis, and cost-effectiveness analysis-that present the decision-making
process in an explicit way so its components and the consequences of
assigning various probabilities and values to them can be examined.
Some aspects of decision analysis, such as evaluation of diagnostic tests,
are included in this book. However, we have elected not to go deeply into
medical decision making itself. Our justification is that decisions are only
as good as the information used to make them, and we have found enough
to say about the essentials of collecting and interpreting clinical information
to fill a book. Readers who wish to delve more deeply into medical decision
making can begin with some of the suggested readings listed at the end
of this chapter.
Organization of the Book

This book is written for clinicians who wish to understand for them-
selves the validity of clinical observations to be able to judge the credibility
of their own clinical observations, those of their colleagues, and research
findings in the medical literature. We have not written primarily for those
who do clinical research, but for all the rest who depend on it. However,
we believe that the basic needs of those who do and those who use clinical
research findings are similar.
Tn most textbooks of clinical medicine, information about disease is
presented as answers to traditional clinical questions: diagnosis, clinical
course, treatment, and the like. On the other hand, most books about
clinical investigation are organized around research strategies such as clini-
cal trials, surveys, and case-control studies. This way of organizing a book
may serve those who perform clinical research, but it is awkward for
clinicians.
Events Natural History Chapter Title
Population
at risk
Risk factors
Cause
Cigarettes
Risk
Asbestos
Prevention
Radon
Abnormality
Onset of disease Frequency
Diagnosis
Diagnosis Prevention
Symptoms and signs
Chest x-ray
Sputum cytology
Biopsy
Treatment Treatment
Outcomes
Death
Disease
Recovery •
Figure 1.5. Organization of the book.
Prognosis
We have organized the book primarily according to the clinical ques-

tions encountered when doctors care for patients. Figure 1.5 illustrates how
these questions correspond to the book chapters, taking lung cancer as an
example. The questions relate to the entire natural history of disease, from
the time people without lung cancer arc first exposed to risk, through
when some acquire the disease and emerge as patients, until the end results
of disease are manifest.
In each chapter, we describe research strategies used to answer that
chapter's clinical questions. Some strategies, such as cohort studies, are
useful for answering several different kinds of clinical questions. For the
purposes of presentation, we have discussed each strategy primarily in
one chapter and simply referred to the discussion when the method is
relevant to other questions in other chapters.
REFERENCES
I. White KL. Henling th", schism. Epidemiology, medicine ~nd public health. New York:
Spnnger-verlagfvst.
2. Pletcher RH. Clinical medidne m"""t~ modern epidemiology-and both profit. Ann Epide-
mini 1992;2:325-333.
3. Murphy EA. Th", logic of medicine. Baltimon" Jnhn~ Hopkins University Press, 1976.
4. Sackett DL. Bias in analytic research. J Chronic Dis 1Y7Y;3251-63.
5. [ha PKS, Beral V, Peto J, Hack S, Hermon C Deacon J, Mant D, Chilver~ C, Vessey MP,
Pik" Me, Muller M, Cissmann L. Antibodi"s to human papillomavirus and to other genital
infectious agents and inva~iv", cervical cancer risk. Lancet 1993;341:1116-1118.
6. Ballard D], Abdominal aortic aneurysm IIk'tterl· New Engl J Mcd 1993;329:1275.
7. Steering Committee of the Physicians' Ilealth Study Research Group. Final report on the
aspirin component of the ongoing Physicians' IleJlth Study. New Eng! J Med 1989;321·129
135.
8. Antiplatelet Trialists' Collaboration. Collaborative overview of randomized trials of anti-
platelet therapy-l. Prevention of d"ath, myocardial infarction, and stroke by prolonged
antiplatdd therapy in various categories of patients. Br Med J 1994;308:81-106.
SUGGESTED RRAI)[NCS
Andersen B. Methodological ",rrors in medical research. Boston: Blackwell Scientific Publica-
tions, 1990.
Eisenberg JM. Clinical economics. A guide to the ,,'nmomic analysis of clinical practices. JAMA
1Y89;262:2879- 28.%.
Facl~, {igllres, and fallacies series
Jolley T. The glitter of the I table. LancctI9Y3;342:27-29.
Victoria CG Whafs the denominator? Lancet 1')<)3;342:97-99.
Grisso JA. MJking compari~on~. I.ane",t 1993;342:157-160.
Datta M. You cannot exclude the expbnation you haven't considered. LancdI993;342:345-
347.
Mert<m~ TE. Estimating the effects of mtsclassification. I.imed 1993;342:418-42l.
Le~m D. Failed or misleading adjustment for confounding. Lancet 1993;342:479-48l.
Sitlhi-amorn C, 1"oshachinJa V. Bia~. Lancer 1993;34J:2116-2!l!l.
Glynn JR. A question of attribution. I.aned 1993;342:530-532.
Carpenter 1M. b the study worth doing? LancctIYY3;343:221-223.
Feinstein AK Clinical epidemiology. The architecture of clinical research. Philadelphia: WB

Saunders. 1985.
Feinst",in AR. Clinimdrics. N",w Hawn, CT: Yak University Pres.", 1987,
Friooman Gn. Primn of "'pid,,,miology. 4th cd. New York: McCraw-Hili, 1YY4.
Cchlbach Sf!. Interpreting the medical literature. 3rd cd. New York: McGraw-Hill, 1'193.
l Icnnckins Cll, Buring JE. Epidemiology in medicine. Boston: Little, Brown & Co, 1987,
Hulley S6, Cummings SR. Designing clinical research. An epidemiologic approach, Baltimore:
Williams & Wilkins, 1988-
Jenicek M, Clt'foux R. Fpidemiologie dinique, Clinimdrie. St-Hyacinthe, Qu",., Canada: Fdi-
sem,IYI\5.
Kramer MS, Clinical epidemiology and biostatistics, A primer for clinical investigators and
decision-makers. New York Springer-Verlag, 1989.
Riegelmnn RK, Hirsch [{P. Studying and study and testing a test. 2nd Ed. Boston: Litt!c,
Brown & Co, 1989,
Sackett DL, Haynes RB, Guyatt GH, Tugwell P. Clinical epidemiology. a basic science for
clinical medicine. 2nd ed. Boston: Little, Brown & and Co, 1991.
&lX He. Blatt MA, Higgins Me, Marton KI. Medical decision making. Stoneham, MA: Butter-
worth, 1988.
Users' guide 10 the mcdicai titeratnre seTies
Cuyatt CIl, Rennie D. U'l('rs' guides to the medical likraturp. JAMAIYY3;270:20Y7-20Y8.
Oxman AD, Sackett DL, Guyatt GH. I. How to get started, JAMA 1Y93;270:20Y3-20<J5,
Guyatt GH, Sackett DL, Cook DJ. II, How to use and article about therapy or prevention.
A. An' the results of the study valid) JAMA 1993;2702598-2601
Cuyatt CH, Sackett DJ., Cook DJ. 11. How to use and article about thl>rapy or pH'vention.
1:1. What were the results and will they help me in caring for my patients? JAMA
1994;271:59-63.
Laupakis A, Wells C, Richardson S, Tugwell P, V, How to use an artide about prognosis.
JAMA 1994;272:234-237.
[aeschke R, Cuvatt G, Sackett DL Ill. How to use and article about a diagnostic test.
A. Are th", results of th", study valid) JAMA 1994;271;389-391
Jr""schkc R, Cuyatt C, Sackl'lt DI.. lIl. How to USl' and article about a diagnostk t",st
ll. What are the results and will they help me in caring for my patients? JAMA
1994;271:703-707.
Levine M, Walter S, Lee H, Haines 1. Holbrook A, Moyer V. IV. How to use an article
about harm. JAMA 1994;271:1615-1619.
W",iss NS. Clinka] "'pid,,,miology. The study of the outcome of illness. New York: Oxford
University Press, 1986,
'Nhite KL. Healing the schism, Epidemiology, medicine and the public's health. New York:
Springcr-Verlag.1991
2
ABNORMALITY
Clinicians spend a great deal of time distinguishing "normal" from

"abnormal." When confronted with something grossly different from
the usual, there is Iittle difficulty telling the two apart. We all are fami liar
with pictures in textbooks of physical diagnosis showing massive hepa-
tosplenomegaly, huge goiters, or severe changes of rheumatoid arthritis
in the hand. We can take no particular pride in recognizing this degree
of abnormality. More often, however, clinicians must make subtler dis-
tinctions between normal and abnormal. Is fleeting chest pain angina
or inconsequential? Is a soft systolic heart sound a sign of valvular heart
disease or an innocent murmur? Is a slightly elevated serum alkaline
phosphatase evidence for liver disease, asymptomatic Paget's disease,
or nothing important?
Decisions about what is abnormal are most difficult among relatively
unsclected patients, usually found outside of hospitals. When patients
have already been selected for special attention, as is the case in most
referral centers, it is usually clear that something is wrong. The tasks
are then to refine the diagnosis and to treat the problem. Tn primary
carl.' settings, however, patients with subtle manifestations of disease
are mixed with those with the everyday complaints of healthy people.
It is not possible to pursue all of these complaints aggressively. Which
of many patients with abdominal pain have self-limited gastroenteritis
and which have early appendicitis? Which patients with sore throat and
hoarseness have a garden variety pharyngitis and which the rare but
potentially lethal Haemophilus epiglottitis? These are examples of how
difficult, and important, distinguishing various kinds of abnormalities
can be.
The point of distinguishing normal from abnormal is to separate out
those clinical observations that should be considered for action from
those that can be simply noted. Observations that are thought to be
normal arc usually described as "within norma] limits," "unremark-
able," or "noncontributory" and remain buried in the body of a medical
19
record. The abnormal findings are set out in a problem list or under the
heading "impressions" or "diagnoses" and are the basis for action.
Simply calling clinical findings normal or abnormal is undoubtedly
crude and results in some misclassification. The justification for taking this
approach is that it is often impractical or unnecessary to consider the raw
data in all their detail. As Bertrand Russell put it, "To be perfectly intelligi-
ble one must be inaccurate, and to be perfectly accurate, one must be
unintelligible." Physicians usually choose to err on the side of being intelli-
gible-to themselves and others-even at the expense of some accuracy.
Another reason for simplifying data is that each aspect of a clinician's
work ends in a decision-to pursue evaluation or to wait, to select a
treatment or reassure. Under these circumstances some sort of present/
absent classification is necessary.
Table 2.1 is an example of how relatively simple expressions of abnor-
mality are derived from more complex clinical data. On the left is a typical
problem list, a statement of the patient's important medical problems. On
the right are some of the data on which the decisions to call them problems
are based. Conclusions from the data, represented by the problem list, are
by no means noncontroversial. For example, the mean of the four diastolic
blood pressure measurements is 94 mm Hg. Some might argue that this
level of blood pressure does not justify the label "hypertension," because
it is not particularly high and there are some disadvantages to telling
patients they are sick and giving them pills. Others might consider the
label fair, considering that this level of blood pressure is associated with
an increased risk of cardiovascular disease and that the risk may be re-
duced by treatment. Although crude, the problem list serves as a basis
for decisions-about diagnosis, prognosis, and trcahnent-and clinical
Table 2.1
Summarization of Clinical Data: A Patient's Problem List and the Data
on Which It Is Based
Problurfl Li~l new nato
Hypertension Several blood pressure readings [mm Hg):

1(0/102, 150/86. 166/92, 172/96
2, Diabetes mellitus Glucose tolerance test:
Time (h) 0 0.5 1 ?
Plasma glucose 110 190 170 140
(rTlg/1 DD rill)
J Renal insufficiency Serum chemistries:
Crealinine 2.7 mg/100 mt
Urea nitrogen 40 mg/100 rnl
Bicarbonate 18 mEqll
CHAPTER 2 I ABNORMALITY 21
decisions must be made, whether actively (by additional diagnostic tests

and treatment) or passively (by no intervention).
This chapter describes some of the ways clinicians distinguish normal
from abnormal. To do so. first it will be necessary to consider how biologic
phenomena are measured, vary, and are summarized. Then it will be possi-
ble to consider how these data are used as a basis for value judgments
about what is worth calling abnormal.
Clinical Measurement
Measurements of clinical phenomena yield three kinds of data: nominal,
ordinal, and interval.
Nominal data occur in categories without any inherent order. Examples
of nominal data are characteristics that are determined by a small set of
genes (c.g.. tissue antigens, sex, inborn errors of metabolism) or are dra-
matic, discrete events (c.g., death, dialysis, or surgery). These data can be
placed in categories without much concern about misclasstfication. Nomi-
nal data that can be divided into two categories (e.g.. present/absent, yes/
no, alive/dead) are called dichotomous.
Ordinal data posse~s some inherent ordering or rank, such as small to
large or good to bad, but the size of the intervals between categories cannot
be specified. Some clinical examples include 1+ to 4+ leg edema, grades
I to VI murmurs (heard only with special effort to audible with the stetho-
scope off the chest), and grades 1 to 5 muscle strength (no movement to
normal strength)
For interral data, there is inherent order and the interval between successive
values is equal, no matter where one is on the scale. There are two types of
interval data. Continuous data can take on any value in a continuum. Exam-
ples include most serum chemistries, weight, blood pressure, and partial
pressure of oxygen in arterial blood. The measurement and descriptions of
continuous variables may in practice be confined to a limited number of points
on the continuum, often integers, because the precision of the measurement, or
its usc, does not warrant greater detail. For example, a particular blood glucose
reading may in fact be 1':)3.2846573 ... mg/lOO mL but simply reported as
193 mg/100 ml.. Discrete data, can take on only specific values and arc ex-
pressed as counts. Examples of discrete data arc the number of a woman's
pregnancies and live births and the number of seizures a patient has per
month.
It is for ordinal and numerical data that the following question arises:
Where docs normal leave off and abnormal begin? When, for example,
does a large normal prostate become too large to be considered normal?
Clinicians are free to choose any cutoff point. Some of the reasons for the
choices will be considered later in this chapter.
Performance of Measurements
Whatever the type of measurement, its performance can be described
in several ways, discussed below.
VALIDITY
As pointed out in Chapter 1, validity is the degree to which the data
measure what they were intended to measurec--that is, the results of a
measurement correspond to the true state of the phenomenon being mea-
sured. Another word for validity is IIcCliracy.
Por clinical observations that can be measured by physical means, it is
relatively easy to establish validity. The observed measurement is com-
pared with some accepted standard. For example, serum sodium can be
measured on an instrument recently calibrated against solutions made up
with known concentrations of sodium. Clinical laboratory measurements
are commonly subjected to extensive and repeated validity checks. For
example, it is a national standard in the United States that blood glucose
measurements be monitored for accuracy by comparing readings against
high and low standards at the beginning of each day, before each technician
begins a day, and after any changes in the techniques such as a new bottle
of reagents or a new battery for the instrument. Similarly, the validity
of a physical finding can be established by the results of surgery or an
autopsy.
Other clinical measurements such as pain, nausea, dyspnea, depression,
and fear cannot be verified physically. In clinical medicine, information
about these phenomena is obtained by "taking a history." More formal
and standardized approaches, used in clinical research, are structured in-
terviews and questionnaires. Groups of individual questions (items) are
designed to measure specific phenomena (such as symptoms, feelings,
attitudes, knowledge, beliefs) called "constructs." Responses to questions
concerning a construct are converted to numbers and grouped together to
form "scales."
There are three general strategies for establishing the validity of mea-
surements that cannot be directly verified by the physical senses.
Content validily is the extent to which a particular method of measure-
ment includes all of the dimensions of the construct one intends to measure
and nothing more. For example, a scale for measuring pain would have
content validity if it included questions about aching, throbbing, burning,
and stinging but not about pressure, itching, nausea, tingling, and the like.
Construct validity is present to the extent that the measurement is consis-
tent with other measurements of the same phenomenon. For example, the
researcher might show that responses to a srale measuring pain are related
to other manifestations of the severity of pain such as sweating, moaning,
writhing, and asking for pain medications.
CHAPTER 2 ! ABNORMALITY 23
Criterion villidity is present to the extent that the measurements predict

a directly observable phenomenon. for example, one might see if responses
on a scale measuring pain bear a predictable relationship to pain of known
severity: mild pain from minor abrasion, moderate pain from ordinary
headache and peptic ulcer, and severe pain from renal colic.
Validity is not, as is often asserted, either present or absent; Rather,
with these strategies one can build a case for or against the validity of a
scale, under the conditions in which it is used, so as to convince others
that the scale is more or less valid.
Because of their selection and training, physicians tend to prefer the
kind of precise measurements the physical and biologic sciences afford,
and they avoid or discount others, especially for research. Yet relief of
symptoms and promoting satisfaction and a feeling of well-being are
among the most important outcomes of care, central concerns of patients
and doctors alike. To guide clinical decisions, research must include them,
Jest the "picture" of medicine painted by the research be distorted.
As Feinstein (1) put it:
The term "hard" is usually applied to data that are reliable and preferably
dimensional k.g., laboratory data, demogrJphic data, and financial costs).
Hut clinical performance, convenience, anticipation, and familial data are
"soft." They depend on xubjer-tivc statements, usually expressed in words
rather than numbers, by the people who are the observers and the observed.
To avoid such soft data, the results of treatment arc commonly restricted
to laboratory information that can be objective, dimensional, and reliable-
but it is also dehumanized. If we are told lhat the serum cholesterol is 230
mg per 100 ml, that the chest X-rily shows cardiac enlargement, and that the
electrocardiogram has Q waves, we would not know whether the treated
object was a dog or a person. If we were told that capacity at work was
restored, that the medicine tasted good and W,lS easy to take, and that the
family was happy about the results, we would recognize a human set of
responses.
RELIABILITY
Reliability is the extent to which repeated measurements of a stable

phenomenon-by different people and instruments, at different times and
places-c-get similar results. Reproducibility and precision are other words
for this property.
The reliability of laboratory measurements is established by repeated
measures-for example, of the same serum or tissue specimen-some-
times by different people and with different instruments. TIle reliability of
symptoms can be established by showing that they are similarly described
to different observers under different conditions.
The relationships between reliability and validity are shown in Figure
2.1. An instrument (laboratory apparatus or a questionnaire] used to collect
VALIDITY
High low
, ,
A , B ,
, ,
, ,
High , ,
, ,
, ,
>-
c
,
RELIABILITY ~
c :1 ,
,
c- , ,
o C , D ,
"
u, , ,
, ,
, ,
low , ,
, ,
,
~A l.A ,
,
~~-
Measurement
Figure 2.1. Validity and reliability. A, High validity and high reliability. B, Low validity
and high reliability. C, High validity and low reliability, 0, Low validity and low reliability
Tile dotted lines represent the true values.
a large set of measurements can be valid (accurate) on the average but not
be reliable, because the measures obtained are widely scattered about the
true value. On the other hand, an instrument can be very reliable but be
systematically off the mark (inaccurate). A single measurement with poor
reliability has low validity because it is likely to be off the mark simply
because of chance alone.
RANGE
An instrument may not register very low or high values of the thing
being measured, limiting the information it conveys. Thus the "first-
generation" method of measuring serum thyroid-stimulating hormone
(1'Sl I) was not useful for diagnosing hyperthyroidism or for precise titra-
tion of thyroxine administration because the method could not detect low
levels of TSH. Similarly, the Activities of Daily Living scale (which mea-
sures people's ability at feeding, continence, transferring, going to the toi-
let, dressing, and bathing) docs not measure inability to read, write, or play
the piano-activities that might be very important to individual patients.
CHAPTFR 2 / ABNORMALITY 25
RESPONSIVENESS
An instrument is responsive to the extent that its results change as condi-
tions change. For example, the New York Heart Association scale-classes
1 to IV (no symptoms, symptoms with slight and moderate exertion, and
symptoms at rest)-is not sensitive to subtle changes in congestive heart
failure, ones patients would value, whereas laboratory measurements of
ejection fraction can detect changes too subtle for patients to notice.
INTERPRETABILITY
A disadvantage of scales based on questionnaires that is not generally
shared by physical measurements is that the results may not have meaning
to clinicians and patients. for example, just how bad is it to have a Zung
depression scale value of 72? To overcome this disadvantage, researchers
can "anchor" scale values to familiar phenomena-for example, by indi-
cating that people with scores below 50 are considered normal and those
with scores of 70 or over are severely or extremely depressed. requiring
immediate care.
Variation
Clinical measurements of the same phenomenon can take on a range of
values, depending on the circumstances in which they are made. To avoid
erroneous conclusions from data, clinicians should be aware of the reasons
for variation in a given situation and know which are likely to play a large
part, a small part, or no part at all in what has been observed.
Overall variation is the sum of variation related to the act of measure-
ment, biologic differences within individuals from time to time, and bio-
logic differences from person to person (Table 2.2).
MEASUREMENT VARIATION
All observations are subject to variation because of the performance of
the instruments and observers involved in making the measurement. The
conditions of measurement can lead to a biased result (lack of validity) or
Table 2.2
Sources of Variation
Source Dc'ipition
Measurement
Instrument The means of rnClking the rneaeuremera
Observer The person making the measurement
Biologic
Within individuals Changes in people with time and situation
Among individuals Biologic differences from person to person
simply random error (lack of reliability). It is possible to reduce this source

of variation by making measurements with great care and by following
standard protocols. However, when measurements involve human judg-
ment, rather than machines, variation can be particularly large and difficult
to control.
Example Fetal heart rate is often monitored by auscultation, which is
subject to observer error. Electronic monitoring gives the true rate.Petal heart
rates that are unusually high or low arc markers of fetal distress, suggt..sting
a need for early delivery.
Day et al. (2)compared fetal heart rates obtained by auscultation by hospi-
tal staff with rates obtained by electronic monitoring (Fig. 2.2). When the
true fetal heart rate was in the normal range, rates by auscultation wert;'
evenly distributed about the true value, i.e., there was only random error.
But when the true fetal heart rate was unusually high or low, rates by auscul-
tation were biased toward normal. Low rates tended to be reported as higher
than the true ratt;'s, and high rates as lower than the true rates.
This study illustrates both random and systematic errors in clinical ob-
servations. In this case, the bias toward normal rates might have arisen
because the hospital staff hoped the fetus was well and were reluctant to
undertake a major intervention based on their observation of an abnor-
mally high or low heart rate.
Variations in measurements also arise because measurements are made on
only a sample of the phenomenon being described, which mtly misrepresent
the whole. Often the samplingfraction (the fraction of the whole that is included
in the sample) is very small. For example, a liver biopsy represents only about
l/JOO,OOO of the liver. Because such a small part of the whole is examined,
there is room for considerable variation from one sample to another.
If measurements are made by several different methods (e.g., different
laboratories, technicians. or instruments) some of the determinations may
be unreliable and/or manifest systematic differences from the correct
value, contributing to the spread of values obtained.
BIOLOGIC VARIATION
Variation also arises because of biologic changes within individuals over
time. Most biologic phenomena change from moment to moment. A mea-
surement at a point in time is a sample of measurements during a period
of time and mtly not represent the usual value of these measurements.
Example Clinicians estimate the frequency of ventricular premature de-
polarization (V!'!)) to help determine the need for and effectiveness of treat-
mcnt. For practical reasons, they may do so by making relatively brief obser-
vations-perhaps feeling a pulse for 1 min or reviewing an electrocardiogram
(a record of about 10sec). llowever, the frequency of VI'Os in a given patient
varies over time. To obtain ,1 lilrger sample of VPO rate, a portable (Holter)
monitor is sometimes used. Hut monitoring even for extended periods of
time can be misleading. figure 23 shows observations on one patient with
15 Monitored fetal
heart rate
130-150
10
Ul
<:
o
~ 5
e
CI>
1l
o
'0
~
~ 11
E
::l Monitored fetal
Z 10
heart rate >150
oL,----,t"-,Iillibfl
-50 40 30 20 10 0 10 20 30 4050+
Underestimate Overestimate
ERROR (beats/min)
Figure 2.2. Observer variability. Error in reporting fetal heart rate according to
whether the true rate, determined by electronic monitor, is within the normal range,
low, or high. (Redrawn from Day E, Maddern L, Wood C. Auscultation of foetal heart
rate: an assessment of its error and significance. Br Med J 1968;4:422-424.)
VPDs, similar to other patients studied (3). VPDs per hour varied from less
than 20 to 380 during a 3-day period, according to day and time of day. The
authors concluded: "To distinguish a reduction in VPD frequency attribut-
able to therapeutic intervention rather than biologic or spontaneous variation
alone required a greater than 83% reduction in VPD frequency if only two
24-hour monitoring periods were compared."
Variation also arises because of differences among people. Biologic dif-

ferences among people predominate in many situations. For example, sev-
eral studies have shown that high blood pressure on single, casual mea-
400
~
~
C
300
0.
>
'0 200
~
1: Day 1
E
100
z"
, ,
Noon Midnight 6 A.M.
Figure 2.3. Biologic variability. The number of ventricular premature depolarizations

(VPDs) for one untreated patient on 3 consecutive days. (Redrawn from Morganroth
J, Michelson EL, Horowitz LN, Josephson ME, Pearlman AS, Dunkman WB. Limita-
tions of routine long-term electrocardiographic monitoring to assess ventricular ec-
topic frequency. Circulation 1978;58:408-414.)
surements, although subject to all other forms of variation, is related to

subsequent cardiovascular disease.
TOTAL VARIATION
The several sources or variation are cumulative. Figure 2.4 illustrates
this for the measurement of blood pressure. Variation from measurement
contributes relatively little, although it covers as much as a 12 mm Hg
range among various observers. On the other hand, each patient's blood
pressure varies a great deal from moment to moment throughout the day,
so that any single blood pressure reading might not represent the usual
for that patient. Much of this variation is not random: blood pressure is
generally higher when people are awake, excited, visiting physicians, or
taking over-the-counter cold medications, Of course, we are most inter-
ested in knowing how an individual's blood pressure compares with that
of his or her peers, especially if the blood pressure level is related to
complications of hypertension and the effectiveness of treatment.
EFFECTS OF VARIATION
Another way of thinking about variation is in terms of its net effect on the
validity and reliability of a measurement and what com be done about it.
Random variation-for example, by unstable instruments or many ob-
CHAPTER 2 I ABNORMAlI1Y 29
servers with various biases that tend to balance each other out-results
on average in no net misrepresentation of the true state of a phenomenon
if a set of measurements are made; individual measurements, however,
may be misleading. Inaccuracy resulting from random variation can be
reduced by taking a larger sample of what is being measured, for example,
by counting more cells on a blood smear, examining a larger area of a
urine sediment, or studying more patients. Also, the extent of random
variation can be estimated by statistical methods (see Chapter 9).
On the other hand, biased results are systematically different from the true
Conditions of Distribution of Source(s)

Measurement Measurements of Variation
One patient, one observer,

repeated observations
at one point in time
Measurement
One patient, many
observers, at one time
One patient, one observer,

J\
almanYI;meSOf~ ,
Biologic and
Measurement
Many patients
60 70 80 90 100 110 120 130
DIASTOLIC BLOOD PRESSURE (mm Hg)

-----------------
Figure 2.4. Sources of variation. The measurement of diastolic (phase V) blood
pressure, (Data from Fletcher RH and Fletcher SW; and Boe J, Humerfell S, Weder-
vang F, Oecon C, The blood pressure in a population [Special Issue]. Acta Med
Scand 1957;321:5·-313.)
value, no matter how mallY times they are repeated. For example, when investi-
gating a patient suspected of having an infiltrative liver disease (perhaps follow-
ing up an elevated serum alkaline phosphatase) a single liver biopsy may be
misleading, depending on how the lesions are distributed in the liver. If the
lesion is a metastasis in the left lobe of the liver, a biopsy in the usual place
(the right lobe) would miss it. On the other hand, a biopsy for miliary tubcrculo-
sis, which is represented by millions of small granulomata throughout the liver,
would be inaccurate only through random variation. Similarly, all of the high
values for VPDs shown in Figure 2.3 were recorded on the first day, and most
of the low values on the third. The days were bia.s. ed estimates of each other,
because of variation in VPD rate from day to day.
Distributions
Data that are measured on interval scales are often presented as a figure,
called a frequency distribution, showing the number (or proportion) of a
defined group of people possessing the different values of the measure-
ment (Fig. 2.5). Presenting interval data as a frequency distribution conveys
the information in relatively fine detail.
20 Mode
•
15
-"
c:
CJ 10
~
"
l1.
2 3 4 5 6 7 8 9 10
PSA (ng/mL)
Figure 2.5. Measures of central tendency and dispersion, The distribution of pros-
tate-specific antigen (PSA) levels in presumably normal men, (Data from Kane RA,
Littrup PJ, Babaian R, Drago JR, Lee F, Chesley A, Murphy GP, Mettlin C. Prostate-
specific antigen levels in 1695 men without evidence of prostate cancer. Cancer
1992;69:1201-1201,)
Table 2.3
Expressions of Central Tendency and Dispersion
Expression Definition Advantages Disadvantages
Central Tendency
Mean Sum of values Well suited tor Affected by extreme

for observations mathematical values
Number of manipulation
observations
Median The point where the number Not easily Not well suited for
ot observations above influenced by mathematical
equals the number below extreme values manipulation
Mode Most frequently occurring Simplicity of Sometimes there

value meaning are no, or many,
most frequent
values
Dispersion
Range From lowest to highest lncludes all values Greatly affected by

value in a distribution extreme values
Standard The absolute value of the Well suited tor For non-Gaussian
deviation" average difference 01 mathematical distributions,
individual values tram the manipulation does not describe
mean a known
proportion of the
observations
Percentile, decile, The proportion of all Describes the Not well suited for
quartile, etc, observations falling "unusualness" statistical
between specified values of a value manipulation
without
assumptions
about the shape
of a distribution
where X ~ each obse""Rtion; X = rT""'' of ali observations; and N ~ number of observations

DESCRIBING DISTRIBUTIONS
It is convenient to summarize distributions. Indeed, summarization is imper-
ative if a large number of distributions art:' to be presented and compared.
Two basic properties of distributions arc used to summarize them: cen-
tral tendency, the middle of the distribution, and dispersion, how spread out
the values are. Several ways of expressing central tendency and dispersion,
along with their advantages and disadvantages. arc summarized in Table
2.3 and illustrated in Figure 2.5.
ACTUAL DISTRIBUTIONS
The frequency distributions of four common blood tests (potassium,
alkaline phosphatase. glucose, and hemoglobin) are shown in Figure 2.6.
In general, most of the values appear near the middle, and except fOT the
30
20
Serum Alkaline
potassium 20 phosphatase
10
10
-
c
Gl
0
~
Gl
3.0 4.0
mEq/L
50 20 40 60 80 100120140
units
"- 30 40
Plasma 30
20 glucose Hemoglobin
20
10
10
100 150 200 B 9 10 11 12 13 14 15 16

mg/1 00 mL g/1 00 mL
Figure 2.6. Actual clinical distributions. (Data from Martin HF, Gudzinowicz BJ,
Fanger H. Normal values in clinical chemistry. New York: Marcel Dekker, 1975.)
central part of the curves, there are no "humps" or irregularities. The high
and low ends of the distributions stretch out into tails, with the tail at one
end often being more elongated than the tail at the other (i.e., the curves
are "skewed" toward the long end). Whereas some of the distributions
are skewed toward higher values, others are skewed toward lower values.
Tn other words, all these distributions are unimodal, are roughly bell-
shaped, and are not necessary symmetric; otherwise they do not resemble
each other.
The distribution of values for many laboratory tests changes with char-
acteristics of the patients such as agc, sex, race, and nutrition. figure 2.7
shows how the distribution of one such test, blood urea nitrogen (BUN),
changes with age. A BUN of 25 mg/lOO mL would be unusually high for
a young person, but not particularly remarkable for an older person.
THE NORMAL DISTRIBUTION
Mother kind of distribution, called the "normal" or Caussian distribution,

is sometimes assumed to approximate naturally occurring distributions,
though it is based in statistical theory and has no nC"Cl'ssary relationship to
natural distributions. The normal curve describes the frequency distribution
of repeated measurements of the same physical object by the same instrument.
Dispersion of values represents random variation alone. A normal curve is
shown in Figure 2.R. The curve is symmetrical and bell shaped. It has the
f¥f44---- 20-29 years old

'"
U
e
Ql ~80 years old
"
l:T
Ql
~
U.
10 20 30 40 50
BUN mg/l00 mL
Figure 2.7. The distribution of clinical variables changes with age: BUN for people
aged 20-29 versus those 80 or older, (Data from Martin HF, Gudzinowicz BJ, Fanger
H. Normal values in clinical chemistry. New York: Marcel Dekker, 1975.)
>-
.
e
e
.
:>
0'
~
LL
Standard -3 -2 -1 a +1 +2
deviations
2.14 13.59 34.13 34.13 13.59 2.14
Percent of
--68.26-:
area under
the curve - - - 95.44 --~
- - - - - 99.72 - - - - -•
Figure 2.8. The normal (Gaussian) distribution.
mathematical property that about two-thirds of the observations fall within

] standard deviation of the mean, and about 95'\"0 within 2 standard deviations.
Although clinical distributions often resemble a normal distribution the
resemblance is superficial. As nne statistician (4) put it:
The experimental fact is that for most physiologic variables the distribution
is smooth, unimodal, and skewed, and that mean ::'::2 standard deviations
does not cut off the desired ':15%. We have no mathematical, statistical. or
other theorems that enable us to predict the shape of the distributions of
phySiologic measurements.
whereas the normal distribution is derived from mathematical theory

and reflects only random variation, many other sources of variation con-
tribute to distributions of clinical measurements, especially biologic differ-
ences among people. Therefore, if distributions of clinical measurements
resemble normal curves, it is largely by accident. Even so, it is often as-
sumed, as a matter of convenience (because means and standard deviations
are relatively easy to calculate and manipulate mathematically), that clini-
cal measurements are "normally" distributed.
Criteria for Abnormality

It would be convenient if the frequency distributions of clinical measure-
ments for normal and abnormal people were so different that these distri-
butions could be used to distinguish two or more distinct populations.
CHAPTER 2 / ABNORMALITY 35
This is the case for specific DNA and RNA sequences and antigens (Fig.
2.9A), which are either present or absent, although their clinical manifesta-
tions may not be so dear-cut.
However, most distributions of clinical variables are not easily di-
vided into "normal" and "abnormal," because they are not inherently
dichotomous and they do not display sharp breaks or two peaks that
Normal Mutant
A lIeles for Phenylalanine Hydroxylase
o 2 4 6 8 10
Blood Phenylalanine (mg IdL)
Figure 2.9. Screening for phenylketonuria (PKUj in infants: dichotomous and over-
lapping distributions of normal and abnormal. A, Alleles coding for phenylalanine
hydroxylase are either normal or mutant. B, The distributions of blood phenylalanine
levels in newborns with and without PKU overlap and are of greatly different magni-
tude. (The prevalence of PKU, actually about 1/10,000, is exaggerated so that its
distribution can be seen in the figure.)
36 CLINICAL EPIDeMIOLOGY
chnracterfze normal and abnormal results. There are several reasons

why this is so.
For m<lny laboratory tests there are not even theoretical reasons for
believing that distinct populations-e-well and dtseescd-c-cxtsr. Disease is
acquired by degrees, and so there is a smooth transition from low to high
values with increasing degrees of dysfunction. Laboratory tests reflecting
organ failure, such <IS serum creatinine for renal failure, behave in
this way
In other situations, well and diseased persons do in fact belong to sepa-
rate populations, but when the two populatiun-, are mixed together they
cannot be recognized as separate because values for the abnorrnals ve'Y.
they overlap those for normals, and there are few abnorrnals in the
population.
Example Phenylketonur-ia (I'KU) is <1 disease characterized by progres-
sive mental retardation in childhood. 1\ variety of mutant alleles coding for
phenylalanine hydroxylase results in dysfunction of the enzyme and, with a
normal diet, accumulation of phenylalanine. The diah'llosis, which becomes
apparent in the first year of life, is confirmed by persistently high phenylala-
nine levels (several times the usual range) and low tyrosine levels in
the blood.
11 is common practice to screen newborns for PKC with a blood test for
phenylalanine a few dilys after birth, in time to treat before there is irrevers-
ible damage. However, the test misclassifies some infants, because at that
age there is an overlap in the distributions of serum phenylalanine concenrra-
tions in infants with and without PKU and because infants with I'KU make
up only a small proportion of those screened, about IIlO,OOO (Fig. 2.9Ll).
Some newborns with PKU are in the normal range either because they have
not yet ingested enough protein or because they hac e a combination of alleles
associated with mild disease. Some children who are not destined to develop
PKU have relatively high levels .cfor eXMnple, because their mothers have
abnormal phenylalanine metabolism. The test is set to be positive at the lower
end of the overlap between normal and abnormal levels, to detect most
infants with the disease, even though only about lout of 5 infants with an
abnormal screening test turns out to have I'KL'.
In unsclccted populations, the diseased patients often do not stand out

because there are very few of them relative to normal people and because
laboratory values for the diseased population overlap those for normals. The
curve for diseased people is "swallowed up" by the larger curve for normal
people. If, on the other hand, normal and diseased populations are mixed in
more equal proportions-perhaps by selecting out for testing people with an
unusually high likelihood of disease-c-thcn the resulting distribution could
be truly bimodal. Eve11 so, it would not be possible to choose a test value
that dearly separates diseased and nnndiscased persons (see Chapter 3).
If there is no sharp dividing line between normal and abnormal, and
the clinician can choose where the line is placed, what ground rules should
be used to decide? Three criteria have proven useful: being unusual, being
sick, and being treatable. For a given measurement, the results of these
approaches bear no necessary relation to each other, so that what might
be considered abnormal using one criterion might be normal by another.
ABNORMAL ~ UNUSUAL
Normal often refers to the most frequently occurring or usual condition.
Whatever occurs often is considered normal, and whatever occurs infre-
quently is abnormal. This is a statistical definition, based on the frequency
of a characteristic in a defined population. Commonly, the reference popu-
lation is made up of people without disease, but this need not be the case.
For example, we may say that it is normal to have pain after surgery or
for eczema to itch.
It is tempting to be more specific by defining what is unusual in mathe-
matical terms. One commonly used way of establishing a cutoff point
between normal and abnormal is to agree, somewhat arbitrarily, that all
values beyond 2 standard deviations from the mean are abnormal. On
the assumption that the distribution in question approximates a normal
(Gaussian) distribution, 2.5°/" of observations would then appear in each
tail of the distribution and be considered abnormal.
Of course, as already pointed out most biologic measurements are not
normally distributed. So it is better to describe unusual values, whatever
the proportion chosen, as a fraction (or percentile) of the actual distribution.
In this way, it is possible to make a direct statement about how infrequent
a value is without making assumptions about the shape of the distribution
from which it came.
A statistical definition of normality is commonly used but there are
several ways in which it can be ambiguous or misleading.
First, if all values beyond an arbitrary statistical limit, say the 95th
percentile, were considered abnormal, then the prevalence of all diseases
would be the same, 5'%. This is inconsistent with our usual way of thinking
about disease frequency.
Second, there is no general relationship between the degree of statistical
unusualness and clinical disease. The relationship is specific to the disease
in question. For some measurements, deviations from usual are associated
with disease to an important degree only at quite extreme values, well
beyond the 95th or even the 99th percentile.
Example The World Health Organization (WHO) considers anemia to
be present when hemoglobin (lib) levels are below 12 g/lOO mL in adult
nonpregnant females. In a British survey of women aged 20-64, Hb was
below 12 g/lOO mL in 11% of 920 nonpregnant women, twice as many as
would be expected if the criterion for abnormality were exceeding 2 standard
deviations (5). But were the women with Hb levels below 12 g/lOO ml.
"diseased" in any way because of their relativalv low llb? Two possibilities
38 CLINICAL ePIDEMIOLOGY
come to mind: The low I Ib may be associated with symptoms or il may be

a manifestation of serious underlying disease. Symptoms such as fatigue,
dizziness, and irritabilitv were not correlated with Hb level, at least for
women whose Hb was above 8.0. Moreover, oral iron, given to women with
Hb between .s.o and 12.0, increased fib by an average of 2..iO g/100 mL but
did not lead to any greater improvement in symptoms than was experienced
by women given placebo. As for serious underlying disease, it is true that
occasionally low Hb may be a manifestation of cancer, chronic infection, or
rheumatic diseases. But only a very small proportion of women with low Hb
have these conditions.
Thus only at Hb levels below 8.0, which occurred in less than 1':;, of these
women, might anemia he an important health problem.
Third, many laboratory tests are related to risk of disease over their entire
range of values, from low to high. for serum cholesterol, there is an almost
threefold increase in risk from the "low normal" to the "high normal" range.
Fourth, some extreme values are distinctly unusual but preferable to
more usual ones. This is particularly true at the low end of some distribu-
tions. Who would not be pleased to have a serum creatinine of 0.4 mg/
100 mL or a systolic blood pressure of 105 mm Hg? Both are unusually
low but they represent better than average health or risk.
Finally, sometimes patients may have, for laboratory tests diagnostic of
their disease, values in the usual range for healthy people, yet clearly be
diseased. Examples include low pressure hydrocephalus, normal pressure
glaucoma, and normncnlremic hyperparathyroidism.
ABNORMAL --= ASSOCIATED WITH DISEASE
A sounder approach to distingulshmg normal from abnormal is to call
abnormal those observations that are regularly associated with disease,
disability, or death, i.e., clinically meaningful departures from good health.
Example What is a "normal" alcohol (ethanol) intake? Several studies
have shown a U-shaped relationship between alcohol intake and mortality:
high death rates in abstainers, lower rates in moderate drinkers, and high
rates in heavy drinkers (Fig. 2.10). It has been suggested that the lower death
rates with increasing alcohol consumption, at the lower end of the curve,
occur because alcohol raises high density lipoprotein levels, which protects
against cardiovascular disease. Alternatively, when people become ill they
reduce their alcohol consumption and this could explain the high rate of
mortality associated with low alcohol intake (6). High death rates at high
intake is less controversial: alcohol is a cause of several fatal diseases (heart
disease, cancer, and stroke). The interpretation of the causes for the Ll-shapcd
curve determines whether it is as abnormal to abstain as it is to drink heavily.
ABNORMAL - TREATABLE
Par some conditions, particularly those that are not troublesome in their
own right (i.e., are asymptomatic), it is better to consider a measurement
abnormal only if treatment of the condition represented by the measure-
CHAPTER 2 ! A8NORMALITY 39
16
14
••
~
12
~
•
~
0- 10
0
-
0
0
8
0;
0-
.••
s:
0
6
None Occasional Light Moderate Heavy
Figure 2.10. Abnormal as associated with disease. Ihe relationship between alco-
holconsumption and mortality. (From Shaper AG, weroametbee G, Walker M, Alco-
hol and mortality in British men: explaining the u-sbecec curve, Lancet
1888;2:12l';7-1 )73)
mcnt leads to a better outcome. This is because not everything that is

associated with an increased risk can be successfully treated: the removal
of the condition may not remove risk, either because the condition itself
is not a cause of disease bu t is only related to a ca use or because irreversible
damage has already occurred. Also, to label people abnormal can cause
adverse psychological effects that are not justified if treatment cannot im-
prove the outlook.
What we consider treatable changes with time. At their best, therapeutic
decisions are grounded on evidence from well-conducted clinical trials
(Chapter 8). As new knowledge is acquired from the results of clinical
trials, the level at which treatment is considered useful may change. For
example. accumulating evidence for treating hypertension has changed
the definition of what level is treatable. As more studies are conducted,
successively lower levels of diastolic blood pressure have been shown to
be worth treating.
Regression to the Mean
When clinicians encounter an unexpectedly abnormal test result, they
tend to repeat the test. Often the second test result is closer to normal.
Why doc-, this happen? Should it be reassuring?
Patients selected because they represent an extreme value in a distribu-

tion can be expected, on the average, to have less extreme values on subse-
quent measurements, This occurs for purely statistical reasons, not because
the patients have necessarily improved. The phenomenon is called regres-
sion to tile meall.
Regression to the mean arises in the following way (fig 2.11). People
are first selected for inclusion in a study or for further diagnosis or treat-
ment because their initial measurement for a trait fell beyond an arbitrarily
selected cutoff point in the tail of a distribution of values for all the patients
examined. Some of these people will remain above the cutoff point on
subsequent measurements, because their true values are usually higher
than average. But others who were found to have values above the cutoff
point during the initial screening usually have lower values. They were
selected only because they happened, through random variation, to have
a high value at the time they were first measured. When the measurement
is taken again, these people have lower values than they had during the
first screening. This phenomenon tends to drag down the mean value of
the subgroup originally found to have values above the cutoff point.
Thus patients who are singled out from others because of a laboratory
test result that is unusually high or low can be expected, on average, to
be closer to the center of the distribution if the test is repeated. Moreover,
subsequent values are likely to be more accurate estimates of the true
t First testing of
.,'"
(J the popUlation
c:
.,.,.
::J
~
LL
I
Patients with high values

~Mean
Patients retested
Mean
Figure 2.11. Regression to the mean.

CHAP! [-8 2 / ABNORMALITY 41
value, which could be obtained if measurements were repeated for a partic-

ular patient many times. So the time-honored practice of repeating labora-
tory tests that are found to be abnormal and of considering the second
one, which is often within normal limits, the correct one is not just wishful
thinking. It has a sound theoretical basis. It also has an empirical basis.
For example, it has been shown that half of serum T4 tests found to be
outside normal limits on screening were within normal limits when rc-
pcatcd (7). However, the more extreme the initial reading is, the less likely
it is to be normal if it is repeated.
Summary
Clinical phenomena are measured on nominal, ordinal, and interval
scales. Although many clinical observations fall on a continuum of values,
for practical reasons they are often simplified into dichotomous (normal!
abnormal) categories. Observations of clinical phenomena vary because
of measurement error, differences in individuals from time to time, and
differences among individuals. The performance of a method of measure-
ment is characterized by validity (Does it measure what it intends to mea-
sure"), reliability (Do repeated measures of the same thing give the same
result?), range, responsiveness, and interpretability.
Frequency distributions for clinical variables have different shapes,
which can be summarized by describing their central tendency and
dispersion.
Laboratory values from normal and abnormal people often overlap;
because of this and the relatively low prevalence of abnormals, it is usually
not possible to make a dean distinction between the two groups using the
test result alone. Choice of a point at which normal ends and abnormal
begins is arbitrary and is often related to one of three definitions of abnor-
mality: statistically unusual. associated with disease, or treatable. If patients
with extreme values of a test are selected and the test is repeated, the
second set of va lues is likely to fa11 closer to the central (statistically normal)
part of the frequency distribution, a phenomenon called regression to
the mean.
RHI+:RENCES
I. Feinstein AR. The need for humanized science in evaluating medication. Lancet 1972;2:421-
423.
2. [Jay E, Madd"rn L, Wood c:. Auscultation of fnetal h",ul ratc: an assessment of its error
and significance, IJr Mcd j 196H;4:422424,
3. Morgaru-oth J, Michelson EL, Horowitz LN, Josephson ME, Pearlman AS, Dunkrnan WB.
Limitations of routine long-term electrocardiographic monitoring to assess ventricular ec-
topic frequency. Circubtion 1978;58:408-414.
4. Flwhil(·k 1.1'. Cuillin CI .. Kl'ating FR. H""lth, normality, and th" ghc>st of <;auss. j;\Mi\
1970;211:69--75.
5. Elwood PC, W~ters WE, Greene WJW. sw",etnam P. Symptoms and circulating hemoglobin
level. J Chmn Dis 1%9;21:615-628.
6. Shaper AG, wannamerhee C, Walker M. Alcohol and mortality in British men: explaining
the If-shaped curve. I.aned 19HH;2: 1267-1273.
7. Epstein KA, Schneiderman LJ, Bush Iw, Zenner A. The "abnormal" screening serum thy-
roxine (T4 ) : una lysis of physician response. outcome, cost and health and dfediveness, J
Cham OisI9HI;34;175-190,
SUGGESTED READINGS
Department of Clinic~l Epidemiology and Biostatisti~s, McMaster University. Clinical db-
agre-ement L How often it occurs and why. Can M"d Assoc J 1980;123:499-504.
o..'partmenl of Clinical Epidemiology and Biostatistics, Mclvlaster Uni~Trsity, Clinical dis-
agreement II. How to avoid it and how to learn from on"·s mistakes, Can Med Assoc J
1980; 123:613-617.
Feinstein AR. Clinical judgment. &lltimore: Williams & Wilkins, 1967.
Feinstein AR. Problems in measurement, dinieal biostatistics. St. Louts: CV Mosby. 1977,
Feinstein AR. Clinimetrics. New Haven, CT: Yale University Press, 191>7.
Koran LM. TIle reliability of dinical methods, data and judgment. N Engl J Med 1975;293:642-
646,695-701.
Mainland D Remarks on clinical "norms." elin Chern 1971; 17:267-274,
Murphy EA. Th,' logic of medicine. Baltimore, Johns Hopkins University Press, 1976
Cuy~tt CJ J, t-eeny DH, Patrick DI.. M",,,suring health-related quality of lif". Ann Intern Med
1,),)3;111l:622-629.
3
DIAGNOSIS
Appearances to the mind are of four kinds. Things either arc what
they appear to be; or they neither arc, nor appear to be; or they arc,
and do not appear to be; or they are not, yet appear to be. Rightly
to aim in all these cases is the wise man's task
Epictetus, 2nd century A.D.
Clinicians devote a great deal of time to determining diagnoses for

complaints or abnormalities presented by their patients. They arrive at the
diagnoses after applying various diagnostic tests. Most competent clini-
cians use good judgment, a thorough knowledge of the literature, and
a kind of rough-and-ready approach to how the information should be
organized. However, there are also basic principles with which a clinician
should be familiar when interpreting diagnostic tests. This chapter deals
with those principles.
A diagnostic test is ordinarily taken to mean a test performed in a labora-
tory. But the principles discussed in this chapter apply equally well to
clinical information obtained from history, physical examination, and im-
aging procedures. They also apply where a constellation of findings serves
as a diagnostic test. Thus one might speak of the value of prodromal
neurologic symptoms, headache, nausea, and vomiting in diagnosing clas-
sic migraine or of hemoptysis and weight loss in a cigarette smoker as
indicators of lung cancer.
Simplifying Data
In Chapter 2, it was pointed out that clinical measurements, including
data from diagnostic tests, are expressed on nominal, ordinal, or interval
scales. Regardless of the kind of data produced by diagnostic tests, clini-
cians generally reduce the data to a simpler form to make them useful in
practice. Most ordinal scales are examples of this simplification process.
Obviously, heart murmurs can vary from very loud to inaudible. But trying
to express subtle gradations in the intensity of murmurs is unnecessary
for clinical decision making. A simple ordinal scale-grades I to VI-
43
serves just as well. More often, complex data are reduced to a simple
dichotomy, e.g" present/absent, abnormnlj'normal, or diseased/well. This
is particularly done when test results Me used to decide on treatment. At
any given point in time, therapeutic decisions are either/or decisions. Ei-
ther treatment is begun or it is withheld.
The use of blood pressure data to decide about therapy is an example
of how we simplify information for practical clinical purposes. Blood
pressure is ordina rily measured to the nearest 2 mm Hg, i.e. on an
interval scale. However, most hypertension treatment guidelines, such
as those of the Joint National Committee on the Detection, Evaluation,
and Treatment of Hypertension (1) and of most physicians, choose a
particular level (e.g.. % mm Hg diastolic pressure) at which to initiate
drug treatment. In doing so, clinicians have transformed interval data
into nominal (in this case, dichotomous) data. To take the example
further, the Joint National Committee recommends that a physician
choose a treatment plan according to whether the patient's diastolic
blood pressure is "mildly elevated" (90-94 mm Hg), "moderately ele-
vated" (95-114 mm Hg). or "severely elevated" (:2:115 mm Hg). an
ordinal scale,
The Accuracy of a Test Result

Establishing diagnoses is an imperfect procei:is, resulting in a probability
rather than a certainty of being right. In the past, the doctor's diagnostic
certainty or uncertainty was expressed by using terms such as rule Ollt or
possible before a clinical diagnosis. Increasingly, the modem clinician expresses
the likelihood that a patient has a disease by using a probability. That being
the case, it behooves the clinician to become familiar with the mathematical
relationships between the properties of diagnostic tests and the information
they yield in various clinical situations. In many instances, understanding
these issues will help the clinician resolve some uncertainty surrounding the
use of diagnostic tests. In other situations, it may only increase understanding
of the uncertainty. Occasionally, it may convince the clinician to increase his
or her level of uncertainty.
A simple way of looking at the relationships between a test's results
and the true diagnosis is shown in Itigure 3,1. The test is considered to be
either positive (abnormal) or negative (normal), and the disease is either
present or absent. There are then four possible interpretations of the test
results, two of which are correct and two wrong. The test has given the
correct answer when it is positive in the presence of disease or negative
in the absence of the disease. On the other hand, the test has been mis-
leading if it is positive when the disease is absent (false positive) or negative
when the disease is present (false negative).
CHAPTER 3 I DIAGNOSIS 45
DISEASE
Present
Positive
TEST f------+
True
Negative negative
Figure 3.1. The relationship between a diagnostic lest result and the occurrence
of disease, There are two possibilities for the test result to be correct (true positive
and true negative) and two possibilities for the result to be incorrect (false positive
and false negative),
THE GOLD STANDARD

Assessment of the test's accuracy rests on its relationship to some way
of knowing whether the disease is truly present or not-a sounder indica-
tion of the truth often referred to as the "gold standard." As it turns out,
the gold standard is often elusive. Sometimes the standard of accuracy is
itself a relatively simple and inexpensive test, such as a throat culture for
group A (:i-hemolytic streptococcus to validate the clinical impression of
strep throat or an antibody test for human immunodeficiency virus. How-
ever, this is usually not the case. More often, one must turn to relatively
elaborate, expensive, or risky tests to be certain whether the disease is
present or absent. Among these are biopsy, surgical exploration, and of
course, autopsy.
For diseases that are not self-limited and ordinarily become overt in a
matter of a few years after they are first suspected, the results of follow-
up can serve as a gold standard. Most cancers and chronic, degenerative
diseases fall into this category. For them, validation is possible even if on-
the-spot confirmation of a test's performance is not feasible because the
immediately available gold standard is too risky, involved, or expensive.
Some care must be taken in deciding the length of the follow-up period,
which must be long enough for the disease to manifest but not so long
that cases can arise after the original testing.
Because it is almost always more costly and more dangerous to use
these more accurate ways of establishing the truth, clinicians and patients
prefer simpler tests to the rigorous gold standard, at least initially. Chest
x-rays and sputum smears are used to determine the nature of pneumonia,
rather than lung biopsy with examination of :he diseased lung tissue.
Similarly, electrocardiograms and serum enzymes are often used to estab-
lish the diagnosis of acute myocardial infarction, rather than catheterization
or imaging procedures. The simpler tests arc used as proxies for more
elaborate but more accurate ways of establishing the presence of disease,
with the understanding that some risk of misclassiftcatlon results. This risk
is justified by the safety and convenience of the simpler tests. But Simpler
tests are only useful when the risks of misclass.ification are known and
found to be acceptably low. This requires sound data that compare their
accuracy to an appropriate standard.
LACK OF INFORMATION ON NEGATIVE TESTS
The goal of all clinical studies describing the value of diagnostic tests
should be to obtain data for all four of the cells shown in Figure 3.1.
Without all these data, it is not possible to assess the risks of misclassifica-
tion, the critical questions about the performance of the tests. Given that
the goal is to fill in all four cells, it must be stated that sometimes this is
difficult to do in the real world. It may be that an objective and valid
means of establishing the diagnosis exists, but it is not available for the
purposes of formally establishing the properties of a diagnostic test for
ethical or practical reasons. Consider the situation in which most informa-
tion about diagnostic tests is obtained. Published accounts come primarily
from clinical, and not research, settings. Under these circumstances, physi-
cians are using the test in the process of caring for patients. They feel
justified in proceeding with more exhaustive evaluation, in the patient's
best interest, only when preliminary diagnostic tests arc positive. They are
naturally reluctant to initiate an aggressive workup, with its associated
risk and expense, when the test is negative. As a result, information on
negative tests, whether true negative or false negative, tends to be much
less complete in the medical literature.
This problem is illustrated by an influential study of the utility of the
blood test that detects prostate specific antigen (PSA) in looking for prostate
cancer (2). Patients with PSAs above a cutoff level were subjected to biopsy
while patients with PSAs below the cutoff were not biopsied. The authors
understandably were reluctant to subject men to an uncomfortable proce-
dure without supporting evidence. As J result, the study leaves us unable
to determine the false-negative rate for PSA screening.
LACK OF INFORMATION ON TEST RESULTS IN THE NONDISEASED
As discussed above, clinicians are understandably loath to perform elab-
orate testing on patients who do not have problems. An evaluation of a
test's performance can be grossly misleading if the test is only applied to
patients with the condition.
Example Magnetic resonance imaging (MI~I) of the lumbar spine is fre-

quently used in the evaluation of patients with low back pain. Many patients
with back pain show herniated intervertebral disks on MRI, which often
serves to explain the pain and guide treatment.
MRls were performed on 98 asymptomatic volunteers (3). The studies
were read by radiologists who did not know the symptom status of the
patients. Bulging or protruding disks were found in nearly two-thirds of
asymptomatic subjects, only slightly lower than the frequency of similar
abnormality in patients with back pain. The authors concluded that such
findings "may frequently be coincidental."
LACK OF OBJECTIVE STANDARDS FOR DISEASE
For some conditions, there are simply no hard-and-fast criteria for diag-
nosis. Angina pectoris is one of these. The clinical manifestations were
described nearly a century ago, Yet there is still no better way to substanti-
ate the presence of angina pectoris than a carefully taken history. Certainly,
a great many objectively measurable phenomena arc related to this clinical
syndrome, for example, the presence of coronary artery stenoses seen on
angiography, delayed perfusion on a thallium stress test, and characteristic
abnormalities on electrocardiograms both at rest and with exercise. All are
more commonly found in patients believed to have angina pectoris. But
none is so closely tied to the clinical syndrome that it can serve as the
standard by which the condition is considered present or absent.
Sometimes, usually in an effort to be "rigorous," circular reasoning is
applied. The validity of a laboratory test is established by comparing its
results to a clinical diagnosis, based on a careful history of symptoms and
a physical examination. Once established, the test is then used to validate
the clinical diagnosis gained from history and physical examination! An
example would be the use of manometry to "confirm" irritable bowel
syndrome, because the contraction pattern demonstrated by manometry
and believed to be characteristic of irritablr- bowel was validated by clinical
impression in the first place.
CONSEQUENCES OF IMPERFECT STANDARDS
Because of such difficulties as these, it is sometimes not possible for

physicians in practice to find information on how well the tests they use
compare with a thoroughly trustworthy standard. They must choose as
their standard of validity another test that admittedly is imperfect but is
considered the best available. This may force them into comparing one
weak test against another, with one being taken as a standard of validity
because it has had longer use or is considered superior by a consensus of
experts. In doing so, a paradox may arise. If a new test is compared with
an old (but inaccurate) standard test, the new test may seem worse even
when it is actually bette-. For example, if the new test is more sensitive
than the standard test, the additional patients identified by the new test
would be considered false positives in relation to the old test. Just such a
situation occurred in a comparison of real-time ultrasonography and ora]
cholecystography for the detection of gallstones (4).]n five patients, ultra-
sound was positive for stones that were missed on an adequate cholccvsto-
gram. Two of the patients later underwent surgery and gallstones were
found, so that for at least those two patients, the standard oral cholecysto-
gram was actually less accurate than the newer real-time ultrasound. Simi-
larly, if the new test is more often negative in patients who really do not
have the disease, results for those patients will be considered false nega-
tives compared with the old test. Thus, if an inaccurate standard of validity
is used, a new test can perform no better than that standard and will seem
inferior when it approximates the truth more closely.
Sensitivity and Specificity
Figure 3.2 summarizes some relationships between a diagnostic test and
the actual presence of disease. It is an expansion of Figure 3.1, with the
addition of some useful definitions. Most of the rest of this chapter deals
DISEASE
Present
Positive a a+b +pv",~

a+b
TEST
Negative d c+d -PV = ---<L
c+d
a+c b+d a+b+c+d
Se",~ SP=~ P a+c

aec br d '" a e b e c e d
---'L -"---
LR+ = a + c LR_=a+c
~ ---<L
b+d b+d
Figure 3.2. Diagnostic test characteristics and definitions. Se - sensitivity; Sp -,--
specificty; P """ prevalence; PV - predictive value.
CHAPTER 3 ! DIAGNOSIS 49
with these relationships in detail. Figure 3.3 illustrates these relationships.

The diagnostic test is housestaff's clinical impression of whether patients
complaining of pharyngitis have a group A t3-hemolytic streptococcus in-
fection or not, and the gold standard is a throat culture.
DEFINITIONS
As can be seen in Figure 3.2 sensitivity is defined as the proportion of
people with the disease who have a positive test for the disease. A sensitive
test will rarely miss people with the disease. Specificity is the proportion
of people without the disease who have a negative test. A specific test will
rarely misdassify people without the disease as diseased.
Applying these definitions to the pharyngitis example (Fig. 3.3), we see
that 37 of the 149 patients with sore throats had positive cultures, and
Group A
B = Hemolytic
Streptococcus on
Throat Culture
Present
Clinical Yes 27 62 + PV == 27 == 44%

Diagnosis 62
of Strep
Pharyngitis No 77 87 -pV== 77 =88%
87
37 112 149
Se ==~~ == 73% so = 1~~== 69% p == ?I9 == 25%

27 10
LR+ == 27 + 10 == 2.3 LR- == 10 + 27
- - - - == 0.39
35 77
35 + 77 77 + 35
Figure 3.3. The accuracy of the clinical diagnosis of streptococcal pharyngitis

compared with the results of throat culture, (Data from Fletcher SW, Hamann C.
Emergency room management of patients with sore throats in a teaching hospital;
influence of non-physician factors. J Comm Health 1976; 1:196--204,)
housestaff correctly diagnosed 27 of these-for a sensitivity of 73%, On the

other hand, 112 patients had negative culture results; housestaff correctly
withheld antibiotics from 77, for a specificity of 69%.
USES OF SENSITIVE TESTS
Clinicians should take the sensitivity and specificity of a diagnostic test
into account when a test is selected. A sensitive test (Le.. one that is usually
positive in the presence of disease) should be chosen when there is an
important penalty for missing a disease. This would be so, for example,
when there is reason to suspect a dangerous but treatable condition, such as
tuberculosis, syphilis, or Hodgkin's disease. Sensitive tests are also helpful
during the early stages of a diagnostic workup, when a great many possi-
bilities are being considered, to reduce the number of possibilities. Diagnos-
tic tests are used in these situations to rule out diseases, i.e., to establish
that certain diseases are unlikely possibilities. For example, one might
choose an HIV antibody test early in the evaluation of lung infiltrates and
weight loss to rule out an AIDS~related infection. In sum, a sensitive test
is most helpful to the clinician when the test result is negative.
USES OF SPECIFIC TESTS
Specific tests are useful to confirm (or "rule in") a diagnosis that has
been suggested by other data. This is because a highly specific test is rarely
positive in the absence of disease, i.e., it gives few false-positive results.
Highly specific tests are particularly needed when false-positive results
can harm the patient physically. emotionally, or financially. Thus, before
patients are subjected to cancer chemotherapy, with all its attendant risks,
emotional trauma, and financial costs, tissue diagnosis is generally required
instead of relying on less specific tests. In sum, a specific test is most
helpful when the test result is positive.
TRADE-OFFS BETWEEN SENSITIVITY AND SPECIFICITY
It is obviously desirable to have a test that is both highly sensitive and
highly specific. Unfortunately, this is usually not possible, Instead, there
is a trade-off between the sensitivity and specificity of a diagnostic test.
This is true whenever clinical data take on a range of values. In those
situations, the location of a cut-off point, the point on the continuum be-
tween normal and abnormal, is an arbitrary decision. As a consequence,
for any given test result expressed on a continuous scale, one characteristic
(e.g.. sensitivity) can be increased only at the expense of the other (e.g.,
specificity). Table 3.1 demonstrates this interrelationship for the diagnosis
of diabetes. If we require that a blood sugar taken 2 ill after eating be
greater than 180 mg ''/0 to diagnose diabetes, all of the people diagnosed
as "diabetic" would certainly have the disease, but many other people
with diabetes would be missed using this extremely demanding definition
Table 3.1
Trade~Off between Sensitivity and Specificity when Diagnosing Diabetes"
mocc Sugar Level

? hr ~fter FOating Sfmsitivity Specificity
{mgilOO mL) (%) {%)
70 986 88
SO 97,1 25,5
90 94,3 41,6
100 88.6 69,8
110 85.7 84,1
120 71.4 92,5
130 64.3 96.9
140 57.1 99.4
150 500 99.6
160 47,1 99,(>
110 42,9 100.0
180 38,6 100,0
190 34,3 100,0
200 271 1000
Publ,c Health Service DiHhetHs program guide, Puolicnticn no. GOll. W~sh,nnton, DC: U,S, Gove",,,,ent
Printin(] Office, 1960,
of the disease. The test would be very specific at the expense of sensitivity.
At the other extreme, if anyone with a blood sugar of greater than 70 mg
'X, were diagnosed as diabetic, very few people with the disease would be
missed, but most normal people would be falsely labeled as having diabe-
tes. The test would then be sensitive but nonspecific There is no way,
using a single blood sugar determination under standard conditions,
that one can improve both the sensitivity and specificity of the test at the
same time.
Another way to express the relationship between sensitivity and speci-
ficity for a given test is to construct a curve, called a receiver operator charac-
teristic (1{00 curve. An ROC curve for the use of a single blood sugar
determination to diagnose diabetes mellitus is illustrated in Figure 3.4. It
is constructed by plotting the true-positive rate (sensitivity) against the
false-positive rate {l-spccificity) over a range of cut-off values. The values
on the axes run from a probability of 0 to 1.0 (or, alternatively, from 0 to
100'};,). Figure 3.4 illustrates the dilemma created by the trade-off between
sensitivity and specificity. A blood sugar cutoff point of 100 will miss only
11% of diabetics, but 30'~o of normals will be alarmed by a false-positive
report. Raising the cutoff to 120 reduces false-positives to less than 10% of
normals, but at the expense of missing nearly 30% of cases.
Tests that discriminate well crowd toward the upper left comer of the
ROC curve; for them, as the sensitivity is progressively increased (the
SPECIFICITY (%)
80 60 40 20
100
/
80 l"
'0
• , 00 /
20
--
_.-
80 110/ /
/
;1!
-"
;f''(U 'co Cutoff points
/
/
>
(mg/l00 mL)
> ">
>-
->".-
60 no
''0
I /
/
/ 40 >-
>
>-
'50 /
>-~ ,eo
60 '"
/ z
inc.
, 40 no / W
z" '80 /
'" >-.
W a
20
'"
'00
/
/
/
80
-
'",
/
/
/
0 -'
20 40 60 80 100
1-SPECIFICITY (%)
(False-positive rate)
-----
Figure 3.4. A ROC curve. The accuracy of 2-hr postprandial blood sugar as a
diagnostic test for diabetes mellitus. (Data Iron I Public Health Service, Diabetes
program guide. Publication no, 506 Washington, DC: U.S. Government Printing
Office. 1960.)
cutoff point is lowered) there is little or no loss in specificity until very

high levels of sensitivity <Ire achieved. Tests that perform less well have
curves that fall closer to the diagon<l] running from lower left to upper
right. The diagonal shows the relationship between true-positive and false-
positive rates that would occur for a test yielding no information, e.g., if
the clinician merely flipped a coin.
The ROC curve shows how severe the trade-off between sensitivity and
specificity is for a test and can be used to help decide where the best cutoff
point should be, Generally, the best cutoff point is at or near the "shoulder"
of the ROC curve, unless there are clinical reasons for minimizing either
false negatives or false positives.
ROC curves are particularly valuable ways of comparing alternative
tests for the same diagnosis. The overall accuracy of a test can be described
as the area under the ROC curve; the larger the area, the better the test.
CHAPTE'R 3 ! DIAGNOSIS 53
Figure 3.5 compares the ROC curves for two questionnaire tests used to
screen for alcoholism in elderly patients-the CAGE and the MAST (Michi-
gan Alcoholism Screening Test) (5). The CAGE is both more sensitive and
more specific than the MAST and includes a much larger area under its
curve.
Obviously, tests that are both sensitive and specific are highly sought
after and can be of enormous value. However, practicing clinicians rarely
work with tests that are both highly sensitive and specific. So for the
present, we must use other means for circumventing the trade-off between
sensitivity and specificity. The most common way is to use the results of
several tests together (as discussed below).
Establishing Sensitivity and Specificity

Not infrequently, a new diagnostic test is described in glowing terms
when first introduced, only to be found wanting later when more experi-
ence with it has accumulated. Enthusiasm for the clinical value of serum
100
90
CAGE
80
70
...
~ .
s 60
.... MAST
E
Ul 50
.
Z ,
W
Ul 40 ,
.
...
30
20 ,
10
o l_~~_~~~_ T-~~~-
o 10 20 30 40 50 60 70 80 90 100
1- SPECIFICITY
Figure 3.5. ROC curves for the CAGE and MAST questionnaires in elderly patients
with and without alcoholism. (Redrawn tram Jones TV, Lindsey BA, Yount P, Soltys
R, Farani-Enayat B, Alcoholism screening questionnaires: arc they valid in elderly
medical outpatients? J Gen Intern Med 1993;8:674-678,)
carcinocmbryonic antigen (CEA) waxed and then waned in this way. At

first, CEi\ was considered a very promising means of diagnosing colon
cancer. 13ut subsequently CEA was shown to be increased in a wide variety
of other conditions as well as in approximately 20% of smokers without
cancer. This kind of confusion-initial enthusiasm followed by disappoint-
ment-arises not from any dishonesty on the part of early investigators
or unfair skepticism by the medical community later. Rather, it is related
to limitations in the methods by 'which the properties of the test were
established in the first place. At the crudest level, the properties of a diag-
nostic test-sensitivity and specificity, for cxamplc-c.may be inaccurately
described because an improper standard of validity has been chosen, as
discussed previously. However, two other issues related to the selection
of diseased and nondiseased patients can profoundly affect the determina-
tion of sensitivity and specificity as well. They are the spectrum of patients
to which the test is applied and bias in judging the test's performance. A
third problem that can lead to inaccurate estimates of sensitivity and speci-
ficity is chance.
SPECTRUM OF PATIENTS
Difficulties may arise w hen patients used to describe the test's properties
arc different from those to whom the test will be applied in clinical practice,
Early reports often assess the test's value among people who are clearly
diseased compared with people who are clearly not diseased, e.g., medical
student volunteers. The test may be able to distinguish between these
extremes very well.
Even patients with the disease in question can differ in severity, stage,
or duration of the disease, and a test's sensitivity will tend to be higher in
more severely affected patients.
Example Pigure 3.6 illustrates how the performance of the test CEA
var-ies with the stage of colorertal cancer. CEA performs well for metastatic
disease and poorly for localized cancer. Thus the sensitivity (or "colorcctal
cancer" depends on the particular 111ix of stagc'S of patients with disease used
to describe the test, and its accuracy is 1110f(-' stable within stages (6).
Similarly, some kinds of people without disease, such as those in whom
disease is suspected, may have other conditions that cause a positive test,
thereby increasing the false-positive rate and decreasing specificity. For exam-
ple, CEA is also elevated in many patients with ulcerative colitis or cirrhosis.
If patients with these diseases were included in the nondiscascd group when
studying the performance or CEA for rolnrcctal cancer, false positives would
increase and the specificity of the test for cancer would fall.
In theory, the sensitivity and specificity of a test are said to be independent
of the prevalence of diseased individuals in the sample in which the test is
being evaluated. (Work with Pigurc 1.2 to confirm this for yourself.} In prac-
SPECIFICITY
0.8 0.6 0.4 0.2
1.00 ,1 "-:: ~ _1_ - ;-;;f
... StageD
... -- _-- .......
".
... Stagec ...

0.80 - 0.2
> , , , ,,, >

>- , , >-
0.60 Stage B, ... 0.4 >
> , . , >-
>- ,, ,,
'"zw 0.40 "Stage A '"z
0.6 w
'"
0.20
Cutoff Point
... 2.5 ng/mL
.. 0.8
-'"
,
• 5.0 ng/rnL
.10.0 ng/mL
0.20 0.40 0.60 0.80 1.00
1-SPECIFICITY
- - - - - ---
Figure 3.6. ROC curve for CEA as a diagnostic test for colorectal cancer, according
to stage of disease. The sensitivity and specmcuv of a test vary with the stace of
disease. (Redrawn from Fletcher RH. Carcinoembryonic antigen, Ann Intern Med
1986; 104:66-73,)
tice, however, several characteristics of patients, such as stage and severity

of disease, may be related both to the sensitivity and specificity of a test and
to the prevalence, because different kinds of patients are found in high- and
low-prevalence situations. Using a test to screen for disease illustrates this
point (see Chapter 8 for a fuller discussion of screening). Screening involves
the use of the test in an asymptomatic population where the prevalence of
the disease is generally low and the spectrum of disease favors earlier and
less severe cases. In such situations, sensitivity tends to be lower and specific-
ity higher than when the same test is applied to patients suspected of having
the disease, more of whom have advanced disease.
BIAS
Sometimes the sensitivity and specificity of a test are not established
independently of the means by which the true diagnosis is established,
leading to <J biased assessment of the test's properties. This may occur in
several ways, As already pointed out, if the test is evaluated using data
obtained during the course of a clinical evaluation of patients suspected
of having the disease in question, a positive test may prompt the clinician
to continue pursuing the diagnosis, increasing the likelihood that the dis-
ease will be found, On the other hand, a negative test may cause the
clinician to abandon further testing, making it more likely that the disease,
if present, will be missed.
In other situations, the test result may be part of the information used
to establish the diagnosis, or conversely, the results of the test may be
interpreted taking other clinical information or the final diagnosis into
account. Radiologists are frequently subject to this kind of bias when they
read x-rays. Because x-ray interpretation is somewhat subjective, it is easy
to be influenced by the clinical information provided. All clinicians experi-
ence the situation of having x-rays overread because of a clinical impres-
sion, or conversely, of going back over old x-rays in which a finding was
missed because a clinical event was not known at the time, and therefore,
attention was not directed to the particular area in the x-ray. Because of
these biases, some radiologists prefer to read x-rays twice, first without
and then with the clinical information. All of these biases tend to increase
the agreement between the test and the standard of validity. That is, they
tend to make the test seem more useful than it actually is, as, for example,
when an MRJ of the lumbar spine shows a bulging disc in a patient with
back pain (see earlier example in this chapter).
CHANCE
Values for sensitivity and specificity (or likelihood ratios, another char-
acteristic of diagnostic tests, discussed below) arc usually estimated from
observations on relatively small samples of people with and without the
disease of interest. Because of chance (random variation) in anyone sample,
particularly if it is small, the true sensitivity and specificity of the test can
be misrepresented, even if there is no bias in the study. The particular
values observed are compatible with a range of true values, typically char-
acterized by the "95% confidence intervals"] (sec Chapter lJ). The width
of this range of values defines the degree of precision of the estimates of
sensitivity and specificity. Therefore, reported values for sensitivity and
specificity should not be taken too literally if a small number of patients
is studied.
Tl", 9."'~·;, confidence interval of a prop"'lion is ~asily ~stim"led by the following formula, b""'-'d on the
binomiol th"orcm:
wh"", {' i, th" "h-",,,',,<I proportiOIl aml ,,,, i, the numbe, "f people ol"erved. Tn be more nearly cx"c!,
Illulliply by 1.Y6,
Figure 3.7 shows how the precision of estimates of sensitivity increases

as the number of people on which the estimate is based increases. In this
particular example, the observed sensitivity of the diagnostic test is 75'Yo.
Figure 1.7 shows that if this estimate is based on only 10 patients, by
chance alone the true sensitivity could be as low as 4SO;;, and as high as
nearly 100,];,. When more patients are studied, the 95°/., confidence interval
narrows, i.c., the precision of the estimate increases.
Predictive Value
As noted previously, sensitivity and specificity are properties of a test that
are taken into account when a decision is made whether or not to order the
test. But once the results of a diagnostic test are available, whether positive
or negative, the sensitivity and specificity of the test are no longer relevant,
because these values pertain to persons known to have or not to have the
disease. But if one knew the disease status of the patient, it would not be
necessary to order the test! For the clinician, the dilerruna is to determine
whether or not the patient has the disease, given the results of a test.
DEFINITIONS
The probability of disease, given the results of a test, is caned the pre-
dictive mlue of the test (see Fig. 3.2). Positivepredictive oaiue is the probability
of disease in a patient with a positive (abnormal) test result. Negative pre-
100
--
if-
80
>-
I- 60
>
l-
V>
z 40
w
V>
20
0
10 20 30 40 50
NUMBER OF PEOPLE OBSERVED
Figure 3.7. The precision of an estimate of sensitivity, The 95% confidence interval
for an observed sensitivity of 75%, according to the number of people observed,
dictive value is the probability of not having the disease when the test
result is negative (normal). Predictive value answers the question, "If my
patient's test result is positive (negative) what are the chances that my
patient does (does not) have the disease?" Predictive value is sometimes
called posterior (or posUest) probability, the probability of disease after the
test result is known. Figure 3.3 illustrates these concepts. Among the pa~
tients treated with antibiotics for streptococcal pharyngitis, less than half
(44%) had the condition by culture (positive predictive value). The negative
predictive value of the housestaff's diagnostic impressions was better; of
the 87 patients thought not to have streptococcal pharyngitis, the impres-
sion was correct for 77 (88~';,).
Terms summarizing the overall value of a test have been described.
One such term, accuracy, is the proportion of all test results, both positive
and negative, that are correct. (For the pharyngitis example in Figure
3.3, the accuracy of the houscstaff's diagnostic impressions was 70%.)
The area under the ROC curve is another useful summary measure of
the information provided by a test result. However, these summary
measures are too crude to be useful clinically because specific informa-
tion about the component parts-sensitivity, specificity, and predictive
value at specific cutoff points-is lost when they are aggregated into a
single index.
DETERMINANTS OF PREDICTIVE VALUE
The predictive value of a test is not a property of the test alone. It is
determined by the sensitivity and specificity of the test and the prevalence
of disease in the population being tested, where prevalence has its custom-
ary meaning-the proportion of persons in a defined population at a given
point in time with the condition in question. Prevalence is also called prior
(or pretest) probability, the probability of disease before the test result is
known. (For a full discussion of prevalence, see Chapter 4.)
The mathematical formula relating sensitivity, specificity, and preva-
lence to positive predictive value is derived from Bayes's theorem of condi-
tional probabilities:
Positive Sensitivity x Prevalence
predictive = (Sensitivity x Prevalence) + (I-Specificity) x (f-Prevalence)
value
The more sensitive a test is, the better will be its negative predictive
value (the more confident the clinician can be that a negative test result
rules out the disease being sought). Conversely, the more specific the test
is, the better will be its positive predictive value (the more confident the
clinician can be that a positive test confirms or rules in the diagnosis being
sought). Because predictive value is also influenced by prevalence, it is not
independent of the setting in which the test is used. Positive results even
for a very specific test, when applied to patients with a low likelihood of
having the disease, will be largely false positives. Similarly, negative re-
sults, even for a very sensitive test, when applied to patients with a high
chance of having the disease, are likely to be false negatives. In sum, the
Interpretation of a positive or negative diagnostic test result varies from
setting to setting, according to the estimated prevalence of disease in the
particular setting.
It is not intuitively obvious what prevalence has to do with an individual
patient. For those who arc skeptical it might help to consider how a test
would perform at the extremes of prevalence. Remember that no matter
how sensitive and specific a test might be (short of perfection), there will
still be a small proportion of patients who are misdassified by it. Imagine
a population in which no one has the disease. In such a group all positive
results, even for a very specific test, will be false positives. Therefore, as
the prevalence of disease in a population approaches zero, the positive
predictive value of a test also approaches zero. Conversely, if everyone in
a population tested has the disease, all negative results will be false nega-
tives, even for a very sensitive test. As prevalence approaches 100%, nega-
tive predictive value approaches zero. Another way for the skeptic to con-
vince himself or herself of these relationships is to work with Figure 3.2,
holding sensitivity and specificity constant, changing prevalence, and cal-
culating the resulting predictive values.
The effect of prevalence on positive predictive value, for a test at differ-
ent but generally high levels of sensitivity and specificity, is illustrated in
Figure 3.8. When the prevalence of disease in the population tested is
relatively high-more than several percent-the test performs well. But
at lower prevalences, the positive predictive value drops to nearly zero,
and the test is virtually useless for diagnosing disease. As sensitivity and
specificity fall, the influence of changes in prevalence on predictive value
becomes more acute.
Example The predictive value of PSA for diagnosing carcinoma of the

prostate has been studied in various clinical situations, corresponding to
different prevalences or prior probabilities. In older asymptomatic men,
where the prevalence of prostatic carcinoma is estimated to be 6~ 12%, only
about 15~~ of men with a PSA of 4 mfo;/dL or more actually had cancer. In
higher risk men (with symptoms or a suspicious rectal exam), where the
prevalence of prostatic carcinoma is 26'%, 4()'Yo of men with positive l'SAs
had cancer (7). If PSA were used as a screening test in asymptomatic men,
5 or 6 healthy men would have to undergo additional tests, often induding
biopsy, to find one man with cancer. However, when there is a strong clinical
suspicion of rnalignancv, nearly 50% of men with a positive test will have
prostatic cancer.
100~ __
--
ifl
w 80
w-'
:> \ \
80/80 90/90
,
sensitivity/specificity
99/99
>"
-> 60
I-
-w
"'>
0 - 40
0.1-
0
0 20
w
a:
0.
0 1/5 1/10 1/50 1/100 1/1000 1/10,000
PREVALENCE
Figure 3.8. Positive predictive value according to sensitivity, specificity, ano.oreva-

lence of disease.
Current efforts to prevent transmission of acquired immunodeficiency

syndrome (AIDS) through blood products is another example of the effect
of disease prevalence on positive predictive value.
Example A blood test for antibodies to human immunodeficiency virus
(HIV) is used to screen blood donors. At one cutoff point, the sensitivity is
97,8°/" and the specificity is 90.4%. In 1985, the positive predictive value of
the test was estimated from the prevalence of infectious units to be no more
than 1/10,000. Thus there would be 9,250 false-positive test results for every
true-positive result (8). Almost 10,000 units would have to be discarded or
investigated further to prevent one transfusion of contaminated blood. The
authors concluded that, for this emotionally charged subject, "careful adher-
ence to the principles of diagnostic test evaluation will avoid unrealistic
expectations."
But the situation changed. As the prevalence of HIV infection increased
in the general population, the positive predictive value of the screening test
improved. In a publication a year later, the prevalence of infected units
among 67,190 tested was 25/10,000, and at similar levels of sensitivity and
specificity, the positive predictive value would be 2.5°;':" much higher than a
few years before (9).
ESTIMATING PREVALENCE
How can clinicians estimate the prevalence or probability of disease in
a patient to determine the predictive value of a test result? There are several
sources of information: the medical literature, local databases, and clinical
judgment. Although the resulting estimate of prevalence is seldom very
precise, error is not likely to be so great as to change clinical judgments
that are based on the estimate. In any case, the process is bound to be
more accurate than implicit judgment alone.
Tn general, prevalence is more important than sensitivity and specificity
in determining predictive value (see fig. 3.8). One reason why this is so
is that prevalence commonly varies over a wider range. Prevalence of
disease can vary from a fraction of a percent to near certainty in clinical
settings, depending on the age, gender, risk factors, and clinical findings
of the patient. Contrast the prevalence of liver disease in a healthy, young
adult who uses no drugs, illicit or otherwise, and consumes only occasional
alcohol, with that of a jaundiced intravenous drug user. By current stan-
dards, clinicians arc not particularly interested in tests with sensitivities
and specificities much below 50'Yo, but if both sensitivity and specificity
are 99°/;" the test is considered a great one. In other words, in practical
terms sensitivity and specificity rarely vary more than twofold.
INCREASING THE PREVALENCE OF DISEASE
Considering the relationship between the predictive value of a test and
prevalence, it is obviously to the physician's advantage to apply diagnostic
tests to patients with an increased likelihood of having the disease being
sought. In fact as Figure 3.8 shows, diagnostic tests are most helpful when
the presence of disease is neither very likely nor very unlikely.
There Me a variety of ways in which the probability of a disease can
be increased before using a diagnostic test.
Referral Process
The referral process is one of the most common ways in which the
probability of disease is increased. Referral to teaching hospital wards,
clinics, and emergency departments increases the chance that significant
disease will underlie patients' complaints. Therefore, relatively more ag-
gressive use of diagnostic tests might be justified in these settings. In pri-
mary care practice, on the other hand, and particularly among patients
without complaints, the chance of finding disease is considerably smaller,
and tests should be used more sparingly.
Example While practicing in a military clinic, one of the authors saw
hundreds of people with headache, rarely ordered diagnostic tests, and never
encountered a patient with a severe underlying cause of headache. (It is
unlikely that important conditions were missed because the clinic was virtu-
.,lIy the only source of medical care for these patients and prolonged follow-
up was avaflahle.) However, during the first week back in a medical resi-
dency, a patient visiting the hospital's emergency department because of a
headache similar to the ones managed in the military was found to have a
cerebellar abscess!
Because clinicians may work at different extremes of the prevalence spec-
trum at various times in their clinical practices, they should bear in mind
that the intensity of diagnostic evaluation may need to be adjusted to suit

the specific situation.
Selected Demographic Groups
In a given setting, physicians can increase the yield of diagnostic tests
by applying them to demographic groups known to be at higher risk for
a disease. 1\ man of 65 is 15 times more likely to have coronary artery
disease as the cause of atypical chest pain than a woman of 30i thus the
electrocardiographic stress test, a particular diagnostic test for coronary
disease, is less useful in confirming the diagnosis in the younger woman
than in the older man (10). Similarly, a sickle-cell test would obviously
have a higher positive predictive value among blacks than among whites.
Specifics of the Clinical Situation
The specifics of the clinical situation are clearly the strongest influence
on the decision to order tests. Symptoms, signs, and disease risk factors
all raise or lower the probability of finding a disease. For example, a woman
with chest pain is more likely to have coronary disease if she has typical
angina and hypertension and she smokes. As a result, an abnormal electro-
cardiographic stress test is more likely to represent coronary disease in
such a woman than in persons with nonspecific chest pain and no coronary
risk factors.
The value of applying diagnostic tests to persons more likely to have a
particular illness is intuitively obvious to most doctors. Nevertheless, with
the increasing availability of diagnostic tests, it is easy to adopt a less
selective approach when ordering tests. However, the less selective the
approach, the lower the prevalence of the disease is likely to be and the
lower will be the positive predictive value of the test.
The magnitude of this effect can be larger than most of us might think.
Example Factors that influence the interpretation of an abnormal electro-
cardiographic stress test are illustrated in Figure 3.9. It shows that the positive
predictive value for coronary artery disease (CAD) associated with an abnor-
mal test can vary from 1.7 to 99.8%, depending on age, symptoms, and the
degree of abnormality of the test. Thus an exercise test in an asymptomatic
3.'i-year-old man showing 1 rnm ST seh'lllent depression will be a false-
positive test in more than 98% of cases. The same test result in a en-year-old
man with typical angina by history will be associated with coronary artery
disease in more than 90% of cases (10).
Because of this effect, physicians must interpret similar test results dif-
ferently in different clinical situations. A negative stress test in an asymp-
tomatic 35-year-old man merely confirms the already low probability of
coronary artery disease, but a positive test usually will be misleading if it
is used to search for unsuspected disease, as has been done among joggers,
airline pilots, and business executives. The opposite applies to the 65-
w
100
gO _>
L::'EI 0.5-1.0 mm
2.5 mm
::> 80
~
'w_
>" 70
:=~ 60
I-~
00
-'"
iilo
a: a:
50
40
"0
w"-
30
".'"
!::
0
20
10
0
Age 30-39 60-69 60-69 60-69
Symptom None None Atypical Typical
angina angina
Prevalence 1.g 94.3

of CAD (%)
Figure 3.9. Effect of disease prevalence on positive predictive value of a diagnostic

test. Probability of coronary artery disease in men according to age, symptoms, and
depression of $T segment on electrocardiogram. (Data from Diamond GA, Forrester
JS. Analysis of probability as an aid in the clinical diagnosis of coronary artery disease.
N Engl J Med 1979;300:1350-1358.)
year-old man with typical angina. In this case, the test may be helpful in
confirming disease but not in excluding disease. The test is most useful in
intermediate situations, in which prevalence is neither very high nor very
low. For example, a eu-ycar-old man with atypical chest pain has a 6n:,
chance of coronary artery disease before stress testing (see Fig. 3.9); but
afterward, with greater than 2.5 mm ST segment depression, he has a 99%
probability of coron<lry disease.
Because prevalence of disease is such a powerful determinant of how
useful a diagnostic test will be, clinicians must consider the probability of
disease before ordering a test. Until recently, clinicians relied on clinical
observations and their experience to estimate the pretest probability of a
disease. Research using large clinical computer data banks now provide
64 CLINICAL EPIDEMIQIOGY
quantitative estimates of the probability of disease, given various combina-

tions of clinical findings (11).
IMPLICATIONS FOR THE MEDICAL LITERATURE
Published descriptions of diagnostic tests often include, in addition to
sensitivity and specificity, some conclusions about the interpretation of a
positive or negative test, i.e., predictive value. This is done, quite rightly,
to provide information directly useful to clinicians. But the data for these
publications are often gathered in university teaching hospitals where the
prevalence of serious disease is relatively high. As a result, statements
about predictive value in the medical literature may be misleading when
the test is applied in less highly selected settings. What is worse, authors
often compare the performance of a test in a number of patients known
to have the disease and an equal number of patients without the disease.
This is an efficient way to describe sensitivity and specificity. However, any
reported positive predictive value from such studies means little because it
has been determined for a group of patients in which the prevalence of
disease was set by the investigators at 50()'0.
Likelihood Ratios
Likelihood ratios are an alternative way of describing the performance
of a diagnostic test. They summarize the same kind of information as
sensitivity and specificity and can be used to calculate the probability of
disease after a positive or negative test.
ODDS
Because use of likelihood ratios depends on odds, to understand them
it is first necessary to distinguish odds from probability. Probability-used
to express sensitivity, specificity, and predictive value-is the proportion
of people in whom a particular characteristic, such as a positive test, is
present. Odds, on the other hand, is the ratio of two probabilities. Odds
and probability contain the same information, but express it differently.
The two can be interconverted using simple formulas:
Odds = Probability of event -~ 1 - Probability of event
Probability - Odds -'-- I + Odds
These terms should be familiar to most readers because they are used in
everyday conversation. For example, we may say that the odds are 4:1 that
the Seattle Supersonic, will win tonight or that they have an 80% probabil-
ity of winning.
DEFINITIONS
The likelihood ratio for a particular value of a diagnostic test is defined
as the probability of that test result in people with the disease divided by
the probability of the result in people without disease. Likelihood ratios
express how many times more (or less) likely a test result is to be found
in diseased, compared with nondiscased, people. If a test is dichotomous
(positive/negative) two types of likelihood ratios describe its ability to
discriminate between diseased and nondiseased people: one is associated
with a positive test and the other with a negative test (sec Fig. 3.2).
In the pharyngitis example (see Fig. 3.3), the data can be used to calculate
likelihood ratios for streptococcal pharyngitis in the presence of a positive
or negative test (clinical diagnosis). A positive test is about 2.5 times more
likely to be made in the presence of streptococcal pharyngitis than in the
absence of it. If the clinicians believed streptococcal pharyngitis was not
present, the likelihood ratio for this negative test was 0.39; the odds were
about 1:2.6 that a negative clinical diagnosis would be made in the presence
of streptococcal pharyngitis compared with the absence of the disease.
USES OF LIKELIHOOD RATIOS
Pretest probability (prevalence) can be converted to pretest odds using
the formula presented earlier. Likelihood ratios can then be used to convert
pretest odds to posttcst odds, by means of the following formula:
Pretest odds X Likelihood ratio = Posttest odds
Posttest odds can, in turn, be converted back to a probability, using the

formula described earlier in this chapter. In these relationships, pretest
odds contains the same information as prior probability (prevalence), likeli-
hood ratios the same as sensitivity / specificity, and posttest odds the same
as positive predictive value (posttcst probability).
The main advantage of likelihood ratios is that they make it easier for
us to go beyond the simple and clumsy classification of a test result as
either abnormal or normal, as is usually done when describing the accuracy
of a diagnostic test only in terms of sensitivity and specificity at a single
cutoff point. Obviously, disease is more likely in the presence of an ex-
tremely abnormal test result than it is for a marginal one. With likelihood
ratios, it is possible to summarize the information contained in a test result
at different levels. One can define likelihood ratios for any number of test
results, over the entire range of possible values. In this way, information
represented by the degree of abnormality, rather than the crude presence
or absence of it, is not discarded. In computing likelihood ratios across a
range of test results, sensitivity refers to the ability of that particular test
result to identify people with the disease, not individuals with that result
or worse. The same is true for the calculation of specificity.
Thus likelihood ratios can accommodate the common and reasonable
clinical practice of putting more weight on extremely high (or low) test
results than on borderline ones when estimating the probability (or odds)
that a particular disease is present.
Example How accurate is serum thyroxine (T4 ) alone as a test for hypo-
thyroidism? This question was addressed in a study of 120 ambulatory gen-
eral medical patients suspected of having hypothyroidism ('2). Patients were
diagnosed as being hypothyroid if serum thyrotropin (TSH) was elevated
and if subsequent evaluations, including other thyroid tests and response to
treatment, were consistent with hypothyroidism. The authors studied the
initial Tvlevel in 27 patients with hypothyroidism and 93 patients whn were
found not to have it to determine how accurately the simple test alone might
have diah'l1osed hypothyroidism.
As expected, likelihood ratios for hypothyroidism were highest for low
levels of T1 and lowest for high levels (Table 3.2). The lowest values in the
distribution of Ls «4.0 pg/dL) were only seen in patients with hypothyroid-
ism, i.e., these levels filled in the diagnosis. The highest levels (>8.0 pg/dL)
were not seen in patients with hypothyroidism, i.e., the presence of these
levels ruled out the disease.
The authors concluded that "it may be possible to achieve cost savings
without loss of diagnostic accuracy by using a single total It measure-
ment for the initial evaluation of suspected hypothyroidism in selected
patients."
The likelihood ratio has several other advantages over sensitivity and
specificity as a description of test performance. The information contrib-
uted by the test is summarized in 011e number instead of two. The calcula-
tions necessary for obtaining posttest odds from pretest odds arc easy.
Table 3.2
Distribution of Values for Serum Thyroxine in Hypothyroid and Normal Patients,
with Calculation of Likelihood Ratios'
P;lt,eflts w',th Test Result

Total Serum
Illyroxine Ilypothyroid Nor",,,1 likelihood
l!-'gidLj (numbHr, pHrCHnt) fnurnbor. percent) RatiO
<11 2 (7.4) 1
1.1-2,0 3 (11.1) Ruled in
? 1·-3,0 1 (3.7)
::11-40 8 (29,6) J
4,1-5,0 4 (14,8) 1 (11) 13,8
5,1-6.0 4 (148) 6 (6,5) 23
61 7.0 3 (11 1) 11 (11.8) 9
7,1-8.0 2 {7.41 19 (20.4) A
8,1-9.0 1/ (1S,:J)
91-10
101-11
20
11
(21,5)
(11,8)
1
Ruled out
11.1-12
-,12
Tolal 27 (100)
4
4
93
(43)
!'1.3)
(WO)
j
, I,om Goldstein BJ, Musblin AI. Use 01 a 'linglH thyroxine test to evaluate ambulatory medical paliHnts for
suspected 11YlX}lIlyroidis",.,t Gellintern Med 19fJ7;2;20 2·1
Also, likelihood ratios are particu lady well suited far describing the overa 11
probability of disease when a series of diagnostic tests is used (see below).
Likelihood ratios (LR) also have disadvantages. One must use odds, not
probabilities, and most of us find thinking in terms of odds more difficult
than probabilities. Also, the conversion from probability to odds and back
requires math or the use of a nomogram, which partly offsets the simplicity
of calculating posttest odds using LRs. Finally, for tests with a range of
results, LRs use measures of sensitivity and specificity that are different
from those usually described.
Multiple Tests
Because clinicians commonly use imperfect diagnostic tests, with less
than IOO'Yo sensitivity and specificity and intermediate likelihood ratios, a
single test frequently results in a probability of disease that is neither very
high nor very low, e.g.. somewhere between 10')'0 and 90%. Usually it is
not acceptable to stop the diagnostic process at such a point. Would a
physician or patient be satisfied with the conclusion that the patient has
even a 20% chance of having carcinoma of the colon? Or that an asymptom-
atic 35-year~ald man with 2.5 mm 5T segment depression on a stress test
has a 42% chance of coronary artery disease (see Fig. 3.9)? Even for less
deadly diseases, such as hypothyroidism, tests with intermediate postrest
probabilities are of little help. The physician is ordinarily bound to raise
or lower the probability of disease substantially in such situations-unless,
of course. the diagnostic possibilities are all trivial, nothing could be done
about the result, or the risk of proceeding further is prohibitive. When
these exceptions do not apply, the doctor will want to proceed with
further tests.
When multiple tests are performed and all are positive or all are nega-
tive, the interpretation is straightforward. All too often, however, some
are positive and others are negative. Interpretation is then more compli-
cated. This section discusses the principles by which multiple tests are
applied and interpreted.
Multiple tests can be applied in two general ways (Fig. 3.10). They can
be used in parallel (i.c.. all at once), and a positive result of any test is
considered evidence for disease. Or they can be done serially (i.e., consecu-
tively), based on the results of the previous test. for serial testing, all
tests must give a positive result for the diagnosis to be made, because the
diagnostic process is stopped when a negative result is obtained.
PARALLEL TESTS
Physicians usually order tests in parallel when rapid assessment is necessary,
as in hospitalized or emergency patients, or for ambulatory patients who cannot
return easily because they have come from a long distance for evaluation.
STRATEGY SEQUENCE OF EVENTS CONSEQUENCES
Serial testing
Test A and test B and test C are positive

-l.Sensitlvity
"l'Speclnclty
Parallel testing
Test A or test 8 or test C is positive

A--,>
TSensitivity
B--,>
LSpeclttclty
c--'>
Figure 3.10. Serial and parallel testing.
Multiple tests in parallel generally increase the sensitivity and, therefore,

the negative predictive value for a given disease prevalence above those
of each individual test. On the other hand, specificity and positive pre-
dictive value are lowered. That is, disease is less likely to be missed (parallel
testing is probably one reason referral centers seem to diagnose disease
that local physicians miss), but false-positive diagnoses are also more likely
to be made (thus the propensity for ovcrdiagnoslng in such centers as
well). The degree to which sensitivity and negative predictive value in-
creases depends on the extent to which the tests identify patients with the
disease missed by the other tests used. Por example, if two tests are used
in parallel with 60 and 8(),~{, sensitivities, the sensitivity of the parallel
testing will be only 80% if the better test identifies all the cases found by
the less sensitive test. If the two tests each detect all the cases missed by
the other, the sensitivity of parallel testing is, of course, 10m!". Tf the two
tests are completely independent of each other, then the sensitivity of
parallel testing would be 92'\"0.
Parallel testing is particularly useful when the clinician is faced with
the need for a very sensitive test but has available only two or more
relatively insensitive ones that measure different clinical phenomena. By

using the tests in parallel, the net effect is a mere sensitive diagnostic
strategy. The price, however, is evaluation or treatment of some patients
without the disease.
Example PSi\. and digital rectal exam are both insensitive tests for the
diagnosis of prostate cancer (7). Table 3.3 shows their sensitivity, specificity,
and predictive values in the screening selling (men without symptoms).
When the two tests arc used in parallel, the sensitivity increases but the
specificity falls. The positive predictive value is lower than for PSA
testing alone.
SERIAL TESTING
Physicians most commonly use serial testing strategies in clinical sttua-
tions where rapid assessment of patients is not required, such as in office
practices and hospital clinics in which ambulatory patients are followed
over time. Serial testing is also used when some of the tests are expensive
or risky, these tests being employed only after simpler and safer tests
suggest the presence of disease. For example, maternal age and blood tests
(c-Ietoprotcin. chorionic gonadotropin and estriol) are used to identify
pregnancies at higher risk of delivering a baby with Down's syndrome.
Mothers found to be at high risk by those tests are then offered amniocente-
sis (D). Serial testing leads to less laboratory usc than parallel testing,
because additional evaluation is contingent on prior test results. However,
serial testing takes more time because additional tests are ordered only
after the results of previous ones become available.
Serial testing maximizes specificity and positive predictive value, but
lowers sensitivity and the negative predictive value (sec Table 3.3). One
ends up surer that positive test results represent disease, but runs an in-
creased risk that disease will be missed. Serial testing is particularly useful
when none of the individual tests available to a clinician is highly specific.
1£ a physician is going to use two tests in series, the process will be
Table 3.3
Tests Characteristics of PSA and Digital Rectal Examination (ORE)"
Positive Preddive
Sensitivily Specilicily Value
PSA 4,0 "g!ITlL 0.67 0,97 Cl.43

Abnormal LJRE 0.50 0.94 0,24
Abnormal PSA or ORE': OEl~ 0.92 0,28
Abnormal PSA one I)RE 034 0.995 0.49
"I 'SI\ and ORE, alone and in combi"alion (parallel a"d SG,illl lestinR) in tll" diaqnosis 01 pr08l11to cancer
(l\daptHd from Kramer RS et al prostate cancer sctacnlnq: what WH know and what we flood to know. Arm
Int MO{j 1993;1199H 923,)
more efficient if the test with the highest specificity is used first. Table 3.4
shows the effect of sequence on serial testing. Test A is more specific than
test B, whereas B is more sensitive than A. By using A first, fewer patients
are subjected to both tests, even though equal numbers of diseased patients
are diagnosed, regardless of the sequence of tests. However, if one test is
much cheaper or less risky, it may be more prudent to use it first.
ASSUMPTION OF INDEPENDENCE
When multiple tests are used, as discussed above, the accuracy of the
result depends on whether the additional information contributed by each
test is somewhat independent of that already available from the preceding
ones, i.e., the next test does not simply duplicate known information. In
fact, this premise underlies the entire approach to predictive value we have
discussed. However, it seems unlikely that the tests for most diseases
Table 3.4
Effect of Sequence ill Serial Testing: A Then B versus B Theil A"
Prevalence of Disease
Number of patients tested 1000

Number of patients with disease 200 (20% prevalence)
Sensilivity era Specificity of Tests
Test Sensitivity Specificity

A 80 90
8 gO 80
Sequence of I esting
Begin wittl Test A Begin with Test B
Disease Disease
+
A 160 80 240 B + 180 160 340
40 no l60 20 640 660
200 SUO 1000 200 600 1000
240 Patients Retested with B 340 Patients Retested with A
Disease Disease
+ +
B , 144 16 160 A + 144 16 160
16 6' 80 46 144 180
160 80 240 180 160 340
. , Note that in both sequenG8S the sarnc number of patients arc idenlili",d as diseased (160) arld the same
number of true positives (t44) arp identified. But wlmn tAst A (wrth tho higher specificity) is used first, fewpr
patents am retest"d. The lower sensitivity oJ test II does not ndwrs"ly affect uc final result
are fully independent of one another. If the assumption that the tests are
completely independent is wrong, calculation of the probability of disease
from several tests would tend to overestimate the tests' value.
SERIAL LIKELIHOOD RATIOS
When a series of tests is used, an overall probability can be calculated,
using the likelihood ratio for each test result, as shown in Figure 3.11. The
prevalence of disease before testing is first converted to pretest odds. As
each test is done, the posttest odds of one becomes the pretest odds for
the next. Tn the end, a new probability of disease is found that takes into
account the information contributed by all the tests in the series.
Responsiveness
The clinical status of patients changes continually either in response to
treatment or because of the effects of aging or illness. Clinicians regularly
face the question, "Has my patient improved or deteriorated?" The tests
used to monitor the clinical course (e.g.. symptom severity, functional
status) are often somewhat different from those used to diagnose disease,
but the assessment of their performance is very similar.
The ability of changes in the value of a test to identify correctly changes
in clinical status is called its reeponeiocnese. It is conceptually related to
the validity of a diagnostic test, except that the presence or absence of a
meaningful change in clinical status, not the presence or absence of disease,
is the gold standard. The magnitude of a test's responsiveness can be
expressed as sensitivity, specificity, and predictive value or as the area
under the ROC curve.
Example Several self-report measures of health and. functional status are
commonly used to monitor the health of populations and evaluate the effects
of treatment. Two such measures are restricted activity days-number of
~jj"ll!:i'iI!IIT "PRC>&A8ILIH
Test A
"
Pretest odds x LRA = Posttest odds
Test B "
Pretest odds x LRs = Posttest odds
"
PC>i"tTIiIlT PRC>IIA8ILIH
Figure 3.11. Use of likelihood ratios in serial testing.
72 CLINICAL FPIDEMIOLOGY
days on which usual activities were limited by illness or injury-s-and self-

reported health-a question asking respondents 10 rate their health from
excellent 10 poor compared with others their age. The responsiveness of these
measures administered 1 year apart was assessed by comparing changes in
each measure between older adults who experienced a major illness during
the year and those thai did not have a major illness (14). The l.:OC curves in
Figure 3.12 show that the changes in self-reported health performed slightly
better than chance in picking up changes in health associated with major
illness, while changes in restricted activity days performed much better, ac-
counting for 80''l;, of the area under the ReX: curve,
Summary
Diagnostic test performance is judged by comparing the results of the
test to the presence of disease in a two-by-two table. All four cells of the
table must be filled. When estimating the sensitivity and specificity of a
new diagnostic test from information in the med ical Iiteruturc, there must
be a gold standard to which the accuracy of the test is compared. The
diseased and nondtscascd subjects should both resemble the kinds of pa-
100 I-------~...,.
_ Restricted .:·If;:"
80 days ..
.... / '/
.. -:/
'
/
>- 60 -: /
--rr- - Health ... /
>- -: /
> /
>- /
40
'"w
z
/
/
/
'" 20 /
/
'/
,>
:/
100
1-SPECIFICITY ('Yo)
Figure 3.12. The responsiveness ot two questionnaire measures of health status,

Distinguishing between elderly patients with and without a major intervening illness.
(Adapted from Wagner EH, LaCroix IV, Grothaus Le, Hecht JA. Responsiveness
of health status measures to change among older adults. J Am Geriatr Soc
1993;41 :241-248.)
bents for whom the test might be useful in practice. In addition, knowledge
of the final diagnosis should not bias the interpretation of the test results
or vice versa. Changing the cutoff point between normal and abnormal
changes sensitivity and specificity. Likelihood ratios are another way of
describing the accuracy of a diagnostic test.
The predictive value of a test is the most relevant characteristic when
clinicians interpret test results. It is determined not only by sensitivity and
specificity of the test but also by the prevalence of the disease, which may
change from setting to setting. Usually it is necessary to use several tests,
either in parallel or in series, to achieve acceptable diagnostic certainty.
Responsiveness, a test's ability to detect change in clinical status, is also
judged by the same two-by-two table.
REFERENCES
Joint National Committee on Detection, Evaluation, and Treatment of High Blood Pres-
MIre. The fifth report of the Joint National Committee on I:kkdinn, Evaluation, and
Treatment of High J3l ood Pressure (JNC V), Arch lnkm Mcd 1993;1:i3:154-183.
2, Catalona WJ; d al. Measurement of pro,tilk-sp"cific antigen in serum as a screening kst
for prostate "ilncer. N Engl J Med 1991;324:11,~6-1161.
3 kmen Me Brant-Zawadzki Ml\, Obu"howski N, Modic MT, Millkasi~n D, Ross IS. Mag-
netic resonance imaging of the lumbar spine in people without back pain. N Engl J Mcd
1YY4;331:69-73.
4. Bartrum RJ Jr, Crow liC, Foote SR. Ultrasonic "nd radiographic cholecystography. N Fngl
J M"d 1977;2%:53!:1-54L
5. jones TV, Lindsey RA, Yount P, Soltys R, Paraui-Enayat 11 Alcoholism screening question-
naires: are th.,y vilild in elderly medical o\ltpiltimts? J Cen Intern .'vied 1993;8:674-fi78.
6. Ptctchcr Rl l. Carcinoembryonic antigen. Ann Intern Med J 986; 104:66-73.
7. Voss JD. Prostate cancer, scn",:ning, and prostate-specific ilntigen: promise or peril? J Cen
Intern Med 1994;9:46R·474,
H, Barry MJ, MIlII"y AC, Singer DE. Screening for HTI.V III antibodies: the relation between
prevalPT1ce and positive predictive valu" ~nd its social consequences. JAM A 19R5;2,)3:33%,
9 W,ml JW, Crindon AJ, feorino PM, Schable C, Parvin M, Allen JR. I ,ilboratory and epide-
miologic evaluation of an enzyme immunoassay for antibodies t" HTLV 111. JAMA
lYIl6;256:357-361
HI. Diamond GA, Forrester jS, Analysis of probability il" an aid in the clinical diagnosis of
coronary artery disease. N Engl J M..,d 1Y79;3(HI:1350-1358.
11 Tierney WM, :McDonald CJ. Prilctice databases and their uses ill dinieal research, Sial
Med 1991;10:541-557.
12. Goldstein "J, Mushlin AI. Use of a single thyroxine lest to evaluate ambulatory medical
pillients for suspected hypothyroidi~m.J Cen Intern Med 1987;2:20-24.
13, Haddow JE, Palomaki GE, Knight Cj, Williams], Pulkkin"n II, Canick jA, Saller D:'J Jr,
Bowers GB, Prenilt,,1 sUt,<,ning for Down's syndrome with uS{' of maternal serum marker~.
N Engl J Med 1Y92;327:5118-593.
14. Wagnn EH, LaCroix AZ, Crothaus I.c, H<.:cht JII. Responsiveness of heilah stalus mea-
sures to change among older "dulls, J Am Ccnatr Soc 1993;41:241-248.
SUGGESTED READINCS
Cebul l{U, Beck LH. Teaching clinical decision making. New York: Pm..'ger, 1985,
Department of Clinical Epidemiology and Biostatistics, McMilster University. Interpretation
of diagnostic clilta. V, How to do it with simple math. Can Med Assoc J 198..1;129:22-29.
(Reprinted in Sackett DC Haynes RB, Tugwdl P, eds. Clinical epidemiology: a bask scknce
for clinical medicine. Boston: Little, Brown, 1985.)
Fagan TJ. Nomogram for Bayes' theorem [Letter]. N Engl J Mud 1975;293:257.
Griner PI', Maycwski RJ, Mu~hlin AI, Greenland P. Selection and interpretdtion of diagnostic
tests and procedures. Principles and applications. Ann Intern Med 1981;94:5-',7---600.
Criner PF, Panzer R], Crccnland P. Clini<:al diagnosis and the laboratory: logi<:<ll strategies
for common medical problems. ChicJgo: Year Book, 1986.
McNeil BJ, Abr~ms HL, eds, Brigham and Women's Hospital handbook of diagnostic imaging.
Boston: Little, Brown, 1986,
Pauker SG, Kassircr JI'. C1inicJI ilpplication of decision analysis: a detailed illustration, Semin
Nucl Med 1978;8:324-335.
Sh"ps SI3, Schechter MT. The assessment of didgnostic tests. A survey of current medicdl
research. JAMAI984;252:2418-2422.
Sox HC Jr. Probability theory in the use of diagnostic tests. An introduction to critical study
of the literature, Arm Intern MedI986;104:60-66.
Sox He Jr. Blatt M, Higgins Me, Marton KI. Medical decision making. Kent, UK: Butterworth,
1987,
Wasson JIl, Sox He [r, Neff RK, Goldman L. Clinical prediction rules. applications and
methodological standnnh. N Fngl , Moo 1985;313:793-799.
Weinstein Me, Fineberg BV, Elstcin AS, Frazier HS, Neuhauser D, Nculra RR, McNeil BJ,
Clinical decision analysis. Philaddphid: WB Saunders, 1980,
4
FREQUENCY
In Chapter 1, we outlined the central questions facing clinicians as they

care for patients. In this chapter, we define and describe the quantitative
evidence that clinicians use to guide their diagnostic and therapeutic deci-
sions. Let us introduce the subject with a patient.
i\ 72-year-old man presents with slowly progressive urinary frequency,
hesitancy, and dribbling. Digital rectal examination reveals a symmetrically
enlarged prostate gland. Urinary flow measurements show significant re-
duction in flow rate and serum PSA is not elevated. A diagnosis of benign
prostatic hyperplasia (I3PH) is made. In deciding on treatment, the clinician
and patient must weigh the costs and benefits of various therapeutic op-
tions: for example, the risks of worsened symptoms or obstructive renal
disease with medical treatment versus operative mortality or sexual dys-
function with surgery.
The decisions have traditionally been made by "clinical judgment,"
which we learn at the bedside and in the clinics. In recent years, methods
for quantitative clinical decision making have been introduced into medi-
cine. The most commonly used clinical strategies are decision analysis,
cost-effectiveness analysis, and cost-benefit analysis. These methods usc
quantitative data about the frequency of key clinical events and the conse-
quences of those events to patients to derive the best course of action. The
methods, described in more detail at the end of the chapter, are only as
good as the estimates of the probability or frequency of clinical outcomes
on which they rely.
For the patient with BPH, sound clinical judgment requires accurate
information about the probability of symptom deterioration, acute reten-
tion or renal damage with medical treatment; and symptom relief, mortal-
ity, impotence, or retrograde ejaculation with surgery. These are, in general,
the kinds of evidence needed to answer most clinical questions. Decisions
are guided by the probability of outcomes under alternative circumstances:
in the presence of a positive test versus a negative test or after treatment
A versus treatment B. Because the probability of disease, improvement,
75
deterioration, cure, or death forms the basis for answering most clinical
questions, this chapter examines measures of clinical frequency.
Assigning Numbers to Probability Statements

Physicians often communicate probabilities as words-usually, some-
times, rarely, etc.-rather than as numbers. Substituting words for numbers
is convenient and avoids making a precise statement when one is uncertain
about a probability. However, it has been shown that there is little
agreement about the meanings of commonly used words for frequency.
Example Physicians were asked to estimate the likelihood of disease
for each of 30 expressions of probability found by reviewing radiology and
laboratory reports. There was great difference of opinion for each expression.
Probabilities for consistent with ranged from 0.18 to 0.98; for unlikely, the range
was 0.01-0.9-1. These data support the authors' assertion that "difference of
opinion among physicians regarding: the management of a problem may
reflect differences in the meaning ascribed to words used to define
probability" (1).
Patients also assign widely varying values for expressions of probability.

In another study, highly skilled and professional workers thought usually
referred to probabilities of 0.35-1.0 (::'::2 standard deviations from the
mean); rarely meant to them a probability of 0-0.15 (2).
Thus substituting words for numbers diminishes the information con-
veyed. We advocate using numbers whenever possible.
PERCEPTIONS OF FREQUENCY
Personal experience colors the clinician's perception of the probability
of conditions and outcomes. I laving a recent patient experience an outcome
will tend to make the clinician inflate the probability of that outcome.
Conversely, clinicians tend to underestimate the frequency of occurrences
that they have not yet experienced or that patients may be reluctant to
discuss. For example, systematic interviews of patients after transurethral
resection of the prostate gland (TUR..P) reveal that more than 50% of men
experience retrograde ejaculation (3). Most urologists would estimate the
frequency to be much lower, since many mall' patients are reluctant to
discuss sexual issues.
Prevalence and Incidence

Tn general, clinically relevant measures of the frequency or probability
of events are fractions in which the numerator is the number of patients
experiencing the outcome (cases) and the denominator is the number of
people in whom the outcome could have occurred. Such fractions are, of
course, proportions; but by common usage, they are referred to as "rates."
Clinicians encounter two measures of frequency-prevalence and
CHAPTER 4 / FREQUENCY 77
incidence. A preualencc is the fraction (proportion) of a group of people

possessing a clinical condition or outcome at a given point in time.
Prevalence is measured by surveying a defined population containing
people with and without the condition of interest, at a single slice in
time. There are two kinds of prevalence. Point preoatence is measured at
the time of the survey for each person, although not necessarily the
same point in time for all the people in the defined population. Period
preootence refers to cases that were present at any time during a specific
period of time.
An incidence is the fraction or proportion of a group initially free of the
condition that develops it over a given period of time. Incidence refers
then to new cases of disease occurring in a population initially free of the
disease or new outcomes, such as disability or death, occurring in patients
with a specific disease. As described later in this chapter and in greater
detail in Chapter 5, incidence is measured by identifying a susceptible
group of people (i.e., people frrx- of the disease or the outcome) and examin-
ing them periodically over an interval of time to discover and count new
cases that develop during the interval.
Example To illustrate the differences between prevalence and incidence,
Pigurc 4.1 shows the occurrence of disease in a group of 100 people over the
course of 3 years (1992-1994). As time passes, individuals in the group de-
velop the disease. They remain in this state until they either recover or die.
In the 3 years. In people suffer the onset of disease and 4 alrcadv had it. The
remaining 80 people do not develop disease and do not appear in the figure.
At the beginning of 1992, there arc four cases, so the prevalence at that
point in lime is 4/H](I. If all 100 individuals, including prior cases, are exam-
ined at the beginning of each year, one can compute the prevalence at those
points in time. At the beginning of 1993, the prevalence is 5/100 because two
of the pre-1992 cases Iinger.'d on into 1993 and Iwo of the new cases devel-
oping in 1992 terminated (hopefully in a cure) before the examination at the
start of 1993. Prevalences can be computed for each of the other two annual
examinations, and assuming that none of the original 100 people died, moved
away, or refused examination, these prevalences arc 7/100 at the beginning
of 1':1':14 and 5/100 at the beginning of 1995.
To calculate the incidence of new cases developing in the population, we
consider only the 9h individuals free of the disease at the beginning of 1'192 and
what happens to them over the next 3 years. Five new cases developed in 1992;
six new cases developed in 1993, and five additional new cases developed in
1'194. The 3-year incidence of the disease is all new cases developing in the 3
years (which is In) divided by the number of susceptible individuals at the
beginnin~ of the follow-up period (% people). or 16/% in 3 years. Whal are
the annual incidences for 1'192, 1'193, and 1994, respectively? Rrmrmbcring to
remove the previous cases from the denominator, we would calculate the annual
incidences as 5/% for 1992,6/91 for 1993, and 5/85 for 1994.
Every measure of disease frequency necessarily contains some indica-

tion of time. With measures of prevalence, time is assumed to be instants-
1992 1993 1994
0
0
0-
0-
0 .
0-----
.
0--
0 ---
-. 0 Onset
-
Duration
Figure 4.1. Occurrence of disease in 100 people at risk from 1992 to 1994.
neous. as in a single frame from a motion picture. Prevalence depicts the

situation at that point in time for each patient, even though it may, in
reality. have taken several weeks or months to collect observations on the
various people in the group studied. For incidence, time is the essence
because it defines the interval during which susceptible subjects were mon-
itored for the emergence of the event of interest.
Table 4.1 summarizes the characteristics of incidence and prevalence.
Although the distinctions between the two seem clear, the literature is
replete with misuses of the terms, particularly incidence (4).
Why is it important to know the difference between prevalence and
incidence? Because they answer two different questions: (n) What propor-
tion of a group of people have a condition? and (b) At what rate do new
cases arise in 8 group of people as time passes? The answer to one question
cannot be obtained directly from the answer to the other.
CHAPTER 4 ! FREOUENCY 79
Measuring Prevalence and Incidence

PREVALENCE STUDIES
The prevalence of disease is measured by surveying a group of people,
some of whom are diseased at that point in time while others are healthy
(Fig. 4.2). The fraction or proportion of the group that is diseased (i.e.,
cases) constitutes the prevalence of the disease.
Such one-shot examinations or surveys of a population of individuals,
including cases and noncascs, are called prevalence studies. Another term
is cross-sectional studies, because people a re studied at a point (cross-section)
in time. They are among the more common types of research designs
reported in the medical literature.
The following is an example of a typical prevalence study.
Table 4.1
Characteristics of Incidence and Prevalence
CI,8rflClerislic lncrdence Prevalence
Numerator New cases occurring during a All cases counted on a single

period of time among a surveyor examination of a
group initially free at group
disease
Denominator All susceptible people present All people examined. includirlg
at the beqinninq of the casas and nonC8S8S
period
Time Duration ot the period Single point
How measured Cohort study (sec Chapter 5) Prevalence (cross.sectoncn
study
Defined Representative Disease/Outcome

Population Sample Present?
Figure 4.2. The design of a prevalence survey.

Example What is the prevalence of dementia in the general population

of older adults? To answer this question, 1965 people 75 years of age and
older living in Cambridge, England, were surveyed. Each participant under-
went an examination that included the Mini-Mental State Examination
(MMSE), a test for cognitive impairment. The presence of dementia was
identified by a serial testing strategy: Those with MMSE scores of 25 or less
were examined using a standardized protocol by a psychiatrist who made
the final diagnosis. The prevalence of dementia was about 10% overall, and
rates doubled in each 5-year age band (5).
INCIDENCE STUDIES
Tn contrast to prevalence, incidence is measured by first identifying a
population free of the event of interest and then following them through
time with periodic examinations to determine occurrences of the event.
The population under examination in <In incidence study, referred to as <I
cohort, may be healthy individuals followed for the emergence of disease
or diseased individuals followed for outcomes of the disease. This process,
also called a cohort study, will be discussed in detail in Chapter 5.
To this point, the term incidence has been used to describe the rate of
new events in a group of people of fixed size, all of whom are observed
over a period of time. This is called cumulative incidence, because new cases
are accumulated over time,
Example To study the incidence of dementia, the Cambridge investiga-
tors identified a cohort by removing from the follow-up study population
those older individuals diagnosed with dementia in the prevalence study
described above (6). The remaining 1778 nondemented people were tracked.
Of these, 305 died, 190 refused further testing, and 88 could not be found or
were ton ill to be examined. The remaining-II\J5 were reexamined an average
of 2.5 year::; after the original examination. Overall, the annual incidence rate
of dcmcntion in this cohort was 4.3')(, and exceeded 8')\, per year for those
who were over age 85 at the time of the prevalence examination.
A second approach to estimating incidence is to measure the number

of new cases emerging in an ever-changing population, where people are
under study and susceptible for varying lengths of time. The incidence
measure derived from studies of this type is sometimes called incidence
density. Typical examples are clinical trials of chronic treatment in which
eligible patients are enrolled over several years so that early enrollees are
treated and followed longer than late enrollees. Tn an effort to keep the
contribution of individual subjects commensurate with their follow-up in-
terval, the denominator of an incidence density measure is not persons at
risk for a specific time period but person-time at risk of the event. An
individual followed for 10 years without becoming <I case contributes 10
person-years, whereas an individual followed for 1 year contributes only
one person-year to the denominator. Incidence density is expressed as the
number of IWW cases per total number of person-years at risk.
CHAPTER 4 I FREQUENCY 81
The person-years approach is also useful for estimating the incidence

of disease in large populations (If known size when an accurate count of
new cases and an estimate of the population at risk are available, e.g., a
population-based cancer registry.
A disadvantage of the incidence density approach is that it lumps to-
gether different lengths of follow-up. A small number of patients followed
for a long time can contribute as much to the denominator as a large
number of patients followed for a short time. Tf these long-term follow-up
patients are systematically different from short-term follow-up patients,
the resulting incidence measures may be biased.
Interpreting Measures of Clinical Frequency

To make sense of a prevalence or incidence rate, the first steps involve
careful definition of both the numerator and the denominator.
WHAT IS A CASE?-DEFINING THE NUMERATOR
Up to this point, the general term case has been used to indicate an
individual suffering from the disease or outcome of interest. In classical
epidemiology, cases tend to be individuals with a disease, and prevalence
and incidence refer to the frequency of cases among population groups
like the residents of a community. However, clinical decisions often depend
on information about the frequency or rate of disease manifestations, such
as symptoms, signs, or laboratory abnormalities, or the frequency of dis-
ease outcomes, such as death, disability, or symptomatic improvement. In
clinical practice, then, "cases" are often those patients with a disease who
manifest a particular clinical finding or experience a particular outcome.
To interpret rates, it is necessary to know the basis on which a case is
defined, because the criteria used to define a case can strongly affect rates.
Example One simple way to identify a case is to ask people whether they
have a certain condition. How does this method compare to more rigorous
methods? In the Commission on Chronic Illness study, the prevalences of
various conditions, as determined by personal interviews in the home, were
compared with the prevalences as determined by physician examination of
the same individuals. Figure 4.3 illustrates the interview prevalences and the
clinical examination prevalences for various conditions.
The data illustrate that these two methods of defining a case can generate
very different estimates of prevalence and in different directions, depending
on the condition (7).
For some conditions, broadly accepted, explicit diagnostic criteria are
available. The Centers for Disease Control and Prevention criteria for defi-
nite Lyme disease (Table 4.2) can be used as an example (8). These criteria
demonstrate the specificity required to define reliably a disease that is as
much in the public eye as is Lyme disease. They also illustrate a trade-off
between rigorous definitions and clinical reality. If only "definite" cases
Method of Defining Case
Clinical Examination Questionnaire
Hernia
Heart disease
Peptic ulcer
Diabetes
Hypertension
Arthritis
Asthma/hayfever
Chronic sinusitis
I I
10 8 6 4 2 o 2 4 6
PREVALENCE (%)
Figure 4.3. Prevalence depends on the definition of a case. The prevalence of

diseases in the general population based on people's opinions (survey) and clinical
evaluation. (Data from Sanders BS. Have morbidity surveys been oversold? Am J
Public Health 1962;52:1648-1659.)
Table 4.2
Criteria for Reporting L.yme Dlseaee"
Clinical Case Definition (Confirmerl)

Erythema Migrans or
At least one late manifestation and laboratory contirmation ot infection
Late Manilestation (VVhen Alternative Explanation Not Found)
Musculoskeletal
Recurrent bnot attacks of objective joint swelling
Nervous system: any of the following
Lymphocytic meningitis
Cranial neuritis (particularly facial palsy)
encephalomyelitis with antibody in CSF
Cardiovascular
Acute onset 2 or 3° atrioventricular conduction defecls that resolve.
Laboratory Confirmation (Any of the Following)
Isolation of Borrelia burgdorfei
Diagnostic levels of fgM and fgG antibodies to the spirochete in serum or CSF-
Significant change in antibody responses in paired acute- and convalescent-phase serum
sampros.
"Centers for Disease Control ami f'rovenliun crileria. (Adupled from U.S. OepHrtlTlHrlt of HeHlfh HnrllllJman
Services. C""e (!HrlrlitioflS for plJblic health SI.Jtv81I1ance. MMWR 198(1;c!Y; 19-~O.)
were included in a rate, most patients who ordinarily would be considered

to have the disease would not be included. On the other hand, including
"probable" cases could overestimate the true rate of disease.
Example The incidence rate of Lyme disease was estimated in Olmstead
County, Minnesota (9). Between 1980 and 1990, 68 cases had been clinically
diagnosed in residents of the COW1ty. Only 17 (25%) met CDC criteria. In
Minnesota. it is mandatory to report Lyme disease to a public health official,
yet only 7 cases were reported, of which four met CDC criteria. These data
illustrate how difficult it is to make accurate estimates of the frequency of
diseases whose diagnosis relies on multiple clinical criteria.
WHAT IS THE POPULATION?-DEFINING THE DENOMINATOR
A rate is useful only to the extent that the individual practitioner can
decide to which kinds of patients the rate applies. The size and characteris-
tics of the group of individuals in which the cases arose must be known.
Customarily, the group included in the denominator of a rate is referred
to as the population or, more particularly, the population at risk, where at
risk means susceptible to the disease or outcome counted in the numerator.
For example, the incidence or prevalence of cervical cancer will be underes-
timated if the population includes women who have had hysterectomies
or includes men.
The denominator of a rate should include the population relevant to
the question being asked, or a representative sample of them. But what is
relevant depends on one's perspective. For example, if we wanted to know
the true prevalence of rheumatoid arthritis in Americans, we would prefer
to include in the denominator a random sample of all people in the United
States. But if we wanted to know the prevalence of rheumatoid arthritis
in medical practice-perhaps to plan services-the relevant denominator
would be patients seen in office practice, not people in the population at
large. In one survey, only 25(~~ of adults found to have arthritic and rheu-
matic complaints (not necessarily rheumatoid arthritis) during a commu-
nity survey had received services for such complaints from any health
professional or institution (10).
It is customary for epidemiologists to think of a population as consisting
of all individuals residing in a geographic area. And so it should be for
studies of cause and effect in the general population. But in studies of
clinical questions, the relevant populations generally consist of patients
suffering from certain diseases or exhibiting certain clinical findings and
who are found in clinical settings that are similar to those in which the
information will be used. Commonly, such patients are assembled at a
limited number of clinical facilities where academic physicians see patients.
They may make up a small and peculiar subset of all patients with the
findings in some geographic area and may even be an unusual group for
office practice in general.
What difference might the choice of a population make? What is at issue

is the generallzabthty of observed rates.
SAMPLING
It is rarely possible to study all the people who have or might develop
the condition of interest. Usually, one takes a sample, so that the number
studied is of manageable size. This raises a question: Is the sample repre-
sentative of the population?
In general, there are two ways to obtain a representative sample. In a
random compte, every individual in the population has an equal probability of
being selected. The more general term probability sample is used if every person
has a known (not necessarily equal) probability of being selected. It is often
important that a study sample includes a sufficient number of members of
particular subgroups of interest such as ethnic minorities. If these subgroups
are small, a simple random sample of the entire population may not include
enough subgroup members. To remedy this, a larger percentage of each of
these subgroups is selected at random. TIle final sample will still be represen-
tative of the entire population if the different sampling fractions are accounted
for in the analysis. On the <lverage, the characteristics of people in probability
samples are similar to those of the population from which they were selected,
particularly if a large sample is chosen.
Other methods of selecting samples may well be biased and so do not
necessarily represent the parent population. Most groups of patients de-
scribed in the medical literature, and found in most clinicians' experience,
arc based on biased samples. Typically, patients arc included in studies
because they are under care in an academic institution, available. willing
to be studied, and perhaps also particularly interesting and/or severely
affected. There is nothing wrong with this practice-as long as it is under-
stood to whom the results do (or do not) apply.
Relationship among Incidence, Prevalence,

and Duration of Disease
Anything that increases the duration of the disease or clinical finding
in a patient will increase the chance that that patient will be identified in a
prevalence study. A glance at Figure 4.1 will confirm this. The relationship
among incidence and prevalence and duration of disease in <I steady
state-c-i.e., where none of the variables is changing much over time-is
approximated by the following expression:
Prevalence = Incidence x Average duration of the disease
Example Table 4.3 shows approximate annual incidence and prevalence

rates for asthma. Incidence falls with increasing age, illustrating the fact thai
the disease arises primarily in childhood. But prevalence stays fairly stable
over the entire age span, indicating that asthma tends to be chronic and is
especially chronic among older individuals. Also, because the pool of preva-
lent cases does not increase in size, about the same number of patients are
recovering from their asthma as new patients are acquiring it.
If we usc the following formula, we can determine that asthma has an
average duration of 10 years:
Average duration = Prevalence -;- Incidence
When the duration of asthma is determined for each age category by divid-
ing the prevalences by the incidences, it is apparent that the duration of
asthma increases with increasing age. This reflects the rfinical observation
that childhood asthma often clears with time, whereas adult asthma tends
to be more chronic.
Bias in Prevalence Studies

Prevalence studies can be used to investigate potentially causal relation-
ships between risk factors and a disease or prognostic factors and an out-
come. For this purpose, they arc quick but inferior alternatives to incidence
studies. Two characteristics of prevalence studies are particularly trouble-
some: uncertainty about the temporal sequence and biases associated with
the study of cases of longer durntionc--vnld" cases.
UNCERTAINTY ABOUT TEMPORAL SEQUENCES
In prevalence studies, disease and the possible factors responsible for
the disease are measured simultaneously, and so it is often unclear which
came first. The time sequence is obscured, and if it is important to the
interpretation it must be inferred. If the risk or prognostic factor is certain
to have preceded the onset of disease or outcome-c-e.g., family history or
a genetic marker-interpretation of the cause-and-effect sequence is Jess
worrisome. If the risk or prognostic factor can be a manifestation of the
Table 4.3
The Relationships among Incidence, Prevalence, and Duration of Disease: Asthma
in the United states-
Gurat,on _ -:;c" ""Cc"c>C~"CCc'=

Annual Incidence Prevalence ArHlIJdII""idence
",
b Ib
Ii/IOOO
3/1000
?9/1000
32/1000
4.8 veers
10./ YP;lrs
17 ·44 ?/1000 ?b/1(01) 13.0 ycms
·15 M 1/1000 :nnOOO aa.o veers
6c.+ C 30/1000 XU) YPA,lrs
:)/1000 :31)/11)1)1) 10.0 yews
"Approxinll!tcd nom severol Wur';es

disease or outcome-c-e.g.. an abnormal laboratory test or a psychological

state-determining the sequence of events is much more difficult. In con-
trast, studies of incidence have a built-in sequence of events because possi-
ble causes of disease are measured initially, before disease has occurred.
These relationships are illustrated in Figure 4.4.
BIASES STUDYING "OLD" CASES
The difference between cases found in the numerator of incidence rates
and of prevalences rates is illustrated in Figure 45. In an incidence study, all
G1SeS are new and most cases OCCUlTing in the population at risk can be
ascertained if followed carefully through time. Tn contrast, a prevalence study
includes a mixture of old and new cases that are available at the time of the
single examination-that is, they identify cases that happen to be both active
(i.c., diagnosable) and alive at the time of the survey. Obviously, prevalence
rates will be dominated by those patients who are able to survive their disease
without losing its manifestations. The differences between the kinds of cases
included in the numerator of an incidence and the kinds of cases included
in the numerator of a prevalence may influence how the rates are interpreted.
Possible Causes Disease or Outcome
Incidence Study
Measurement is
development of
new cases of
disease over time
causes
Prevalence Study
Measurement is
past or present
exposure to J
possible causes
Figure 4.4. Temporal relationship between possible causal factors and disease for
incidence and prevalence studies.
CHAPTEn 4 ! FREQUENCY 87
Enter population
\
\TI~\>
Early Cures Leave
deaths population:
Severe disease
Mild disease
Prefer other care, etc.
Figure 4.5. The difference in cases for incidence and prevalence studies.
Prevalence is affected by the average duration of disease. Rapidly fatal

episodes of a disease would be included in an incidence study, but most
would be missed by a prevalence sfudy. For example, 25-40~/;, of all deaths
from coronary heart disease occur within 24 hr of the onset of symptoms
in people with no prior evidence of disease. A prevalence study would,
therefore, underestimate cases of coronary heart disease. On the other
hand, diseases of long duration are well represented in prevalence surveys,
even if their incidence is low. For example, although the incidence of
Crohn's disease is only about 2 to 7 per 100,OOO/year, its prevalence is
more than 100 per 100,000, reflecting the chronic nature of the disease (11).
Prevalence rates also selectively include more severe cases of nonfatal
diseases. For example, patients with quiescent rheumatoid arthritis might
not be diagnosed in a study based on current symptoms and physical
findings. Similarly, patients with recurrent but controllable illnesses, such
as congestive heart failure or depression, may be well at a given point in
time and, therefore, might not be discovered on a single examination.
Unremitting disease, on the other hand, is less likely to be overlooked
and, therefore, would contribute disproportionately to the pool of cases
assembled by a prevalence study.
Uses of Incidence and Prevalence

What purposes do incidence and prevalence serve? Clinicians use them
in three different ways: predicting the future course of a patient, assigning
a probability to a patient, and making comparisons.
PREDICTING THE FUTURE

Incidence is a description of the rate at which a disease or a disease
outcome has arisen over time in a group of people known to be free of
the disease at the beginning of follow-up. It can be used to predict the
probability that similar people will develop the condition in the future.
Example The probabilities of outcomes following TURI', needed to de-
cide the most appropriate treatment for the man with BPH described at the
opening of this chapter, were estimated from a large cohort studv of older
men in New England (12). Interviews with more than 300 men undergoing
TURP revealed that symptom resolution varied with the severity of symp-
toms before surgery; 93'1., of men with severe symptoms improved with
surgery while only 79';;, of those with moderate baseline symptoms improved.
On the other hand, prevalence studies offer no sound basis for pre-
dicting the future. Tf a prevalence study finds that 30% of patients with
stroke are depressed, this does not mean that 30% of nondepressed stroke
patients will become depressed in the future. It may be that depression
predisposes to stroke, that stroke predisposes to depression, or that nonde-
pressed stroke patients recover quickly. To find out the percentage of stroke
patients who become depressed, new stroke patients must be followed
over time with repeat measures of depressive symptoms.
ASSIGNING A PROBABILITY THAT A PATIENT HAS THE CONDITION
Prevalence studies are particularly useful in guiding decisions about
diagnosis and treatment. As pointed out in Chapter 3, knowing that a
patient with a combination of demographic and clinical characteristics has
a given probability of having a disease influences the use and interpretation
of diagnostic tests. It may also may affect the selection among various
treatment options.
A patient with pharyngitis illustrate, how variations in prevalence or
prior probability can influence the approach to a clinical problem.
Example A study compared three approaches to the treatment of phar-
yngitis. The value of the approaches was judged by weighing the potential
benefits of preventing rheumatic fever against the costs of penicillin allergy.
The three options were to obtain a throat culture and treat only those patients
with throat cultures positive for group A (I-hemolytic streptococcus, treat all
patients without obtaining a culture. and neither culture nor treat any patient.
The analysis revealed that the optimal strategy depended on the likelihood
that a patient would have a positive culture, which can be estimated from
the prevalence of streptococcal infection in the community ill the time and
the presence or absence of fever. The authors concluded that if the probability
of a positive culture for an individual patient exceeds 20'};), the patient shou ld
be treated: if it is less than .S':;" the patient should not be cultured or treated;
and if the probability lies between 5 and 20%, the patient should be cultured
first and treated based on the result (13).
This study represents a rational approach to the use of prevalences as
CHAPTER 4 ! I'"REOUENCY 89
indicators of individual probabilities of disease in guiding clinical decision

making.
MAKING COMPARISONS
Although isolated incidences and prevalences are useful descriptions,
they become much more powerful tools in support of decision making
when used to make comparisons. It is the comparison between the frequen-
cies of disease among individuals with certain characteristics and individu-
als not sharing those characteristics that provides the strongest evidence.
For example, the risk (incidence) of lung cancer among males who smoke
heavily is of the order of 0.17% pcr year, hardly a common event. Only
when this incidence is contrasted with the incidence in nonsmokers (ap-
proximately 0.007':'0 per year) does the devastating effect of smoking
emerge. Clinicians use measures of frequency as the ingredients in compar~
ative measures of the association between a factor and the disease or dis-
ease outcome. Ways of comparing rates are described in more detail in
Chapter 5.
Clinical Decision Analysis

Quantitative approaches to assisting in decision making have been used
to define the most effective and efficient way to deal with specific problems
in individual patients (cliniCilI policy) or for allocating resources to larger
groups of people, such as communities or political jurisdictions (public
policy).
In decision analysis, one sets out alternative courses of action (e.g., surgery
versus medical treatment for BPII or culture then treat or treat everybody
for streptococcal pharyngitis) and then calculates which choice is likely to
result in the most valued outcome, based on estimates of frequencies for
each branch in the sequences of events and judgments about the relative
value of the possible outcomes. The basic steps are clearly presented else-
where (14) and are described only briefly below.
1. Create a decision tree. Clinical decision analysis begins with a patient who
poses a dilemma. Which of the possible courses of action should be
taken? The tree begins with these alternative decisions, then branches
out to include all of the important consequcnces of those decisions,
and ends with the clinically important outcomes. Branch points involve
either patient care decisions (r'choicc nodes," indicated by squares) or
spontaneous events (vchancc nodes," indicated by circles). Although
there is an infinite number of sequences of events and outcomes, usually
only a small number arc truly important and are reasonably likely to
occur. To make the analysis manageable, it is necessary to "prune" the
tree so that only the most important branches are included -typically
no more than several branch points.
2. Assign probabilities to chance nodes. These probabilities are assessments

of the frequency of clinical events, which are usually derived from the
medical literature.
3. Assign utilities to the oil/comes. Utilities are quantitative expressions of
the relative value of the various outcomes considered. They are best
obtained from patients who may confront the decision. The units <Ire
arbitrary, but must be on an interval scale, e.g., () to 100. It may seem
awkward to put <I number on the respective values of the various out-
comes (death, suffering, loss of function), especially when they are mea-
sured in different units, such as the length and quality of life. But pa-
tients attach values to outcomes in any case, and the numbers only
make the values explicit.
4. Calculate the expected utilities for the alternative courses of action. Starting
with utilities (at the end of the branches, to the right), multiply utilities
by probabilities for each branch and add branches at each node in suc-
cession until the expected utility at the main branch point, the decision
that hns to be made, is reached.
5. Select the choice with the highest expected utility.
6. Sensitivity analysis. Estimate, of probabilities and utilities are uncertain
in the first place. The final step in decision <Ina lysis is to see how the
results of the analysis change as these estimates are varied over a range
of plausible values. That is, one must find out how "sensitive" the
decision is to imprecision in the estimates. Sensitivity analysis indicates
which point in the sequence of events have the most effecl on the deci-
sion and how large the effect might be.
Example The therapeutic options facing the older man with urinary
symptoms from benign prostatic hyperplasia (described <It the opening of
this chapter) have been evaluated using decision analysis (15). Before drugs
and laser prostatectomy made the decision more complicated, the options
were surgery (transurethral resection of the prostate, TURP) or careful follow-
up, called "watchful waiting." Figure 4.6 shows the decision tree that the
authors used lo evaluate the options. The frequencies of the various outcomes
were derived in the incidence study of New England men described earlier
in the chapter (12) and other published sources (15). Note that the optimal
decision in this case is surgery (net utility 0.94). In this case, TURP is the
favored treatment because the risk of operative death is low and the utilities
i1ssigned 10 Incontinence or impotence are the same as that assigned to living
with stable moderate urinary symptoms. If stable moderate symptoms were
preferred over incontinence or impotence, the balance would shift.
Summary
Most clinical questions are answered by reference to the frequency of
events under varying circumstances. The frequency of clinical events is
indicated by probabilities or fractions, the numerators of which include
Incontinent or
impotert (0.06)
Symptoms
improved (0.80)
No complications (0.94) r;-;:;l
L!.:Qj
Survive 0.95
(0.99)
Symptoms not improved (O.2°L 10.801
Die
(0.01)
Symptoms improved (0.02)
Symptoms stable (0.93)
TURP
Figure 4.6. A decision tree. Management of a 70-year-old man with moderate

symptoms from benign prostatic hyperplasia. (Data adapted lrom Barry MJ, Mulley
AG, Fowler FJ, Wennberg JW. Watchful wailing vs. immediate transurethral resection
for symptomatic prostatism. JAMA 1988;259(20):3010-3017.)
the number of cases and the denominators of which include the number
of people from whom the cases arose.
There are two measures of frequency: prevalence and incidence. Preva-
lence is the proportion of a group with the disease at a single point in
time. Incidence is the proportion of a susceptible group that develops new
cases of the disease over an interval of time.
Prevalence is measured by a single survey of a group containing cases
and noncases. whereas measurement of incidence requires examinations
of a previously disease-free group over time. Thus prevalence studies iden-
tify only those cases who are alive and diagnosable at the time of the
survey, whereas cohort (incidence) studies ascertain all new cases. Preva-
lent cases, therefore, may be a biased subset of all cases because they do
not include those who have already succumbed or been cured. In addition,
prevalence studies frequently do not permit a clear understanding of the
temporal relationship between a causal factor and a disease.
To make sense of incidence and prevalence, the clinician must under-
stand the basis on which the disease is diagnosed and the characteristics
of the population represented in the denominator. The latter is of particular
importance in trying to decide if a given measure of incidence or prevalence
pertains to patients in one's own practice.
Incidence is the most appropriate measure of frequency with which to

predict the future. Prevalence serves to quantitate the likelihood that a
patient with certain characteristics has the disease at a single point in time
and is used for decisions about diagnosis and screening. The most powerful
use of incidence and prevalence, however, is to compare different clinical
alternatives.
Measures of disease or outcome incidence are essential ingredients in
methods for quantitative decision making. Approaches such as decision
analysis define alternative clinical strategies and then evaluate those strate-
gies quantitatively by comparing their expected utilities determined from
the frequencies and values assigned to the major outcomes associated with
each strategy.
Postscript
Counting clinical events as described in this chapter may seem to be
the most mundane of tasks. It seems so obvious that examining counts of
clinical events under various circumstances is the foundation of clinical
science. It may be worth reminding the reader that Pierre Louis introduced
the "numerical method" of evaluating therapy less than 200 years ago.
Louis had the audacity to count deaths and recoveries from febrile illness
in the presence and absence of blood-letting. He was vilified for allowing
lifeless numbers to cast doubt on the healing powers of the leech, powers
that had been amply confirmed by decades of astute qualitative clinical
observation.
REFERENCES
Bryant GO, .,\lormiln GR. Expressions of probability: words and numbers. N Engl J Med
1980;302:411,
2. Toogood JH. What do we mean by "usually"? Lancet 19&);1:1094.
3. ~cConn"]l ]0, Barry MJ, Bruskcwitz RC. Beni b'l1 prostatic hyperplasia: diilgnosis and
treatment. C1in Pract Guid(> Quick Ref Cuide Clin 1994;8:1-17.
4, Friedman GO. Medical usage and abusage, "prevalence" and "'incidPnce-" Ann Inkm
Med 1976;84:502-503.
5. O'Connor DW, PoEtt 1'1\, Hyde }B, Fellows JL, Miller NO, Brook CPB. Reiss BB, I~oth M.
The prevalence of dementia as measured by the Cambridge Mental Disorders of the
IJIderly Fxaminiation. Acta Psyehiatr Scand 1989;79:J90-191>.
6 Paykcl ES, Brayne C Huppert 1:1\, c.m C, Barkley C, GehlhaM E, Beardsall I., Girling
OM. PollittI', O'Connor D, Incidence of dementia in a popubtion older that 75 years in
the United Kingdom. Arch Cen Psychiatry 1994;51:325-332.
7. Sanders BS. Have morbidity surveys been oversold? Am J l'ublic Health 1962;52:1648-
1659.
8. Case definitions for public health surveillance. MlvmWI990:(RR-D);19-21
9, Matteson EL, Be<:kett VL, O'Fallon W1I.1, Melton 1,1 Ill, Duffy J. Epidemiology of Lyme
disease in Olmsted COllnty, MN,1975-19YO, J RheumatoI1992;19:1743-1745.
III Spitzer WO, Harth M, Coldsmith CII, Norman CR, Dickie CL, Bass M}, Newell Jr. TIle
arthritic complaint in primary care: prevalence, related disability. and costs. J Rheumatol
1976;3:88-',19,
11 Sedlack RE, Whisnant J, Elveback I.R, Kurland LT. Incidence of Crohn's dist'ase in Olmsted
County, Minnesota, 1935- 1975. Am J Epidemiol 19RO; 112:759-763.
12. Fowl(,r I'J, Wennlwrg JE, Timothy RP, Rarry \1J, Mulley AG Hanky D. Symptom status
,md quality of life following prosMectorny. JAMA 1988;259:3OTR-J022,
n. Tompkins RK, rlurnes DC Cable WE, An analysis of the cost-dfecliveness of pharyngitis
management and acute rheumatic f"ver prevention. Ann Intern Med 1977;R6:4Rl-492.
14. Sox He Blatt MA, Higgins Me Marton Kl. Medical decision milking. Stoneham, MA:
Butterworth,19!:1!:1,
15, Barry MJ, Mulley AG, Fowler FJ, weru-berg jw. Watchful waiting vs. immediate transur,,-
thral resection for symptomatic prostatism. JAMA 1988;259:3010-3017.
SUGGESTED KEADINGS
Fllmb<'rg JH, Nelson KB.Sample selection and the natural history of disease: studies of febrile
seizures, JAMA 19RO;24J(l):377- 1340.
Friedman GO. Medical usage und abusage, "prevalence" and "incidence." Ann Intern Med
1976;84:502-503,
IVlorgenstern H, Kleinbaum UG, Kupper 1.1 , Measures of disease incidence used in epidemio_
logic research. lnt J Epidcmiol 1980;9:97-104,
5
RISK
Risk generally refers to the probability of some untoward event. In this

chapter, the term risk is used in a more restricted sense to indicate the
likelihood that people who aft' exposed to certain factors ("risk factors")
will subsequently develop a particular disease.
People have a strong interest in their risk of disease. This concern has
spawned many popular books about risk reduction and is reflected in
newspaper headlines about the risk of breast cancer from exposure to toxic
chemicals, of AIDS from blood transfusions, or of prostatic cancer after
vasectomy.
This chapter describes how investigators obtain estimates of risk by
observing the relationship between exposure to possible risk factors and
the subsequent incidence of disease. We discuss several ways of comparing
risks, as they affect both individuals and populations.
Risk Factors
Characteristics that are associated with an increased risk of becoming
diseased are called riskfactors. Some risk factors arc inherited. For example,
having the haplotype HLA-B27 greatly increases one's risk of acquiring the
spondylarthropathies. Work on the Human Genome Project has identified
several other diseases for which specific genes are risk factors, including
colon cancer, osteoporosis, and amyotropic lateral sclerosis. Other risk
factors, such as infectious agents, drugs, and toxins, are found in the physi-
cal environment. Still others are part of the social environment. For exam-
ple, bereavement due to the loss of a spouse, change in daily routines,
and crowding all have been shown to increase rates of disease-not only
emotional illness but physical illness as well. Some of the most powerful
risk factors are behavioral; examples are smoking, drinking alcohol to ex-
cess, driving without seat belts, and engaging in unsafe sex.
Exposure to a risk factor means that a person has, before becoming ill,
come in contact with or has manifested the factor in question. Exposure
can take place at a single point in time, as when a community is exposed
94
CHAPTER 5 I RISK 95
to radiation during a nuclear accident. More often, however, contact with

risk factors for chronic disease takes place over a period of time. Cigarette
smoking, hypertension, sexual promiscuity, and SWl exposure are
examples.
There are several different ways of characterizing the amount of expo-
sure or contact with a putative risk factor: ever exposed, current dose,
largest dose taken, total cumulative dose, yeMs of exposure, years since
first contact, etc. (1). Although the various measures of dose tend to be
related to each other, some may show an exposure-disease relationship,
whereas others do not. For example, cumulative doses of sun exposure
constitute a risk factor for nonmelanorna skin cancer, whereas episodes of
severe sunburn <He a better predictor of melanoma. Choice of an appro-
priate measure of exposure to a risk factor is usually based on all that is
known about the biologic effects of the exposure and the pathophysiology
of the disease.
Recognizing Risk
Large risks associated with effects that occur rapidly after exposure are
easy for anyone to recognize. Thus it is not difficult to appreciate the
relationship between exposure and disease for conditions such as chick-
enpox, sunburn, and aspirin overdose, because these conditions follow
exposure relatively rapidly and with obvious effects. But most morbidity
and mortality is caused by chronic diseases. For these, relationships be-
tween exposure and disease are far less obvious. It becomes virtually im-
possible for individual clinicians, however astute, to develop estimates of
risk based on their own experiences with patients. This is true for several
reasons, which are discussed below.
LONG LATENCY
Many diseases have long latency periods between exposure to risk
factors and the first manifestations of disease. This is particularly true
for certain cancers, such as thyroid cancer in adults after radiation treat-
ment for childhood tonsillitis. When patients experience the conse-
quence of exposure to a risk factor years later, the original exposure
may be all but forgotten. The link between exposure and disease is
thereby obscured.
FREQUENT EXPOSURE TO RISK FACTORS
Many risk factors, such as cigarette smoking or eating a diet high in
cholesterol and saturated fats, are so common in our society that for many
years they scarcely seemed dangerous. Only by comparing patterns of
disease among people with and without these risk factors or by investigat-
ing special subgroups-e.g., Mormons (who do not smoke) and vcgetan-
ans (who eat diets low in cbolesterofj-c-dtd we recognize risks that are, in
fact, large.
LOW INCIDENCE OF DISEASE
Most diseases, even ones thought to be "common," are cctuallv quite
rare. Thus, although lung cancer is the most common cause of cancer
deaths in Americans, the yearly incidence of lung cancer even in heavy
smokers is less than 2 in 1000. In the average physician's practice, years
may pass between patients with new cases of lung cancer. It is difficult to
draw conclusions about such infrequent events.
SMALL RISK
If a factor confers only a small risk for a disease, a large number of
people are required to observe a difference in disease rates between
exposed and unexposed persons. This is so even if both the risk factor
and the disease occur relatively frequently. for example, it is still uncer-
tain whether birth control pills increase the risk of breast cancer, because
estimates of this risk are all small and, therefore, easily discounted as
resulting from bias or chance. In contrast, it is not controversial that
hepatitis B infection is a risk factor for hepatoma, because people with
hepatitis B infection are hundreds of times more likely to get liver cancer
than those without it.
COMMON DISEASE
If a disease is common-s-heart disease, cancer, or stroke-s-and some of
the risk factors for it are already known, it becomes difficult to distinguish
a new risk factor from the others. Also, there is les~ incentive to look for
new risk factors. For example, the syndrome of sudden, unexpected death
in adults is a common way to die. Many cases seem related to coronary
heart disease. However, it is entirely conceivable that there are other im-
portant causes, as yet unrecognized because an adequate explanation for
most cases is available.
On the other hand, rare diseases and unusual clinical presentations
invite efforts to find a cause. AIDS was surh an unusual syndrome that
the appearance of just a few cases raised suspicion that some new agent
(as it turned out, a retrovirus) might be responsible. Similarly, physicians
were quick to notice when several cases of carcinoma of the vagina, a very
rare condition, began appearing. A careful search for an explanation was
undertaken, and maternal exposure to diethylstilbestrol was found.
MULTIPLE CAUSES AND EFFECTS
There is usually not a dose, one-to-one relationship between a risk

factor and a particular disease. The relationship between hypertension and
congestive failure is an example (Fig. 5.1). Some people with hypertension
CHAPTER 5 I RISK 97
Ciqarettes Valvular Viral

disease infection Alcohol Thiamine
Cholesterol
deficiency
Diabetes
Family history Myopathies
1,:illffii:,:>,;:F;,:n;nHp<>""'"
'ill1K
''i:~:I :ii~ mUI ]!'i' ' ·ii;f :i!l'iI
,m ~"'-'
I::
.>
!.,.~·:,·" ~.;:."!.,''",.~':."!" 'iU!i~I·:·]
....i.,.'!. ..! '.f,!.'• . · .!,:.• :'.~! ·'.'].~.·'.'.·_
.. .. . ".·',!',1
Other
~ Coronary atherosclerosis
~Stroke
Renal failure
Myocardial infarction
Figure 5.1. Relationship between risk factors and disease: hypertension (1 BP)
and congestive heart failure (CHF). Hypertension causes many diseases, including
congestive heart failure, and congestive heart failure has many causes, including
hypertension.
develop congestive heart failure and many do not. Also, many people who
do not have hypertension develop congestive heart failure, because there
are several different causes. The relationship is also obscured because hy-
pertension causes several diseases other than congestive heart failure.
Thus, although people with hypertension are about 3 times more likely
than those without hypertension to develop congestive heart failure and
hypertension is the leading cause of the condition, physicians were not
particularly attuned to this relationship until the 1970s, when adequate
data became available after careful study of large numbers of people over
many years.
For all these reasons, individual clinicians are rarely in a position to
confirm associations between exposure and disease, though they may sus-
peet them. For accurate information, they must tum to the mcdtcal lltera-
ture, particularly to studies that are carefully constructed and involve 01
large number of patients.
Uses of Risk
PREDICTION
Risk factors are used, first and foremost, to predict the occurrence of
disease. In fact, risk factors, by definition, predict some future event. The
best available information for predicting disease in an individual person
is past experience with a large number of people with a similar risk factor.
The quality of such predictions depends on the similarity of the people
on whom the estimate is based and the person for whom the prediction
is made.
It is important to keep in mind that the presence of even a strong risk
factor does not mean that an individual is very likely to get the disease.
For example, studies have shown that a heavy smoker has a 20-fold greater
risk of lung cancer compared with nonsmokers, but he or she still has only
a 1 in a 100 chance of getting lung cancer in the next 10 years.
There is a basic incompatibility between the incidence of a disease in
groups of people and the chance that an individual will contract that
disease. Quite naturally, both patients and clinicians would like to answer
questions about the future occurrence of disease as precisely as possible.
They arc uncomfortable about assigning a probability, such as the chances
that a person will get lung cancer or stroke in the next 5 years. Moreover,
anyone person will, at the end of 5 years, either have the disease or not.
So in a sense, the average is always wrong because the two are expressed
in different terms, a probability versus the presence or absence of disease.
Nevertheless, probabilities can guide clinical decision making. Even if a
prediction does not come true in an individual patient, it will usually be
borne out in many such patients.
CAUSE
Just because risk factors predict disease, it docs not necessarily follow
that they cause disease. A risk factor may mark a disease outcome indi-
rectly, by virtue of an association with some other determinant(s) of dis-
ease, i.e., it may be confounded with a causal factor. For example, Jack of
maternal education is a risk factor for low birth weight infants. Yet, other
factors related to education, such as poor nutrition, less prenatal care,
cigarette smoking, etc., are more directly the causes of low birth weight.
A risk factor that is not a cause of disease is called a marker, because it
"marks" the increased probability of disease. Not being a cause does not
diminish the value of a risk factor as a way of predicting the probability
of disease, but it does imply that removing the risk factor might not remove
the excess risk associated with it. For example, as pointed out in Chapter
1, although there is growing evidence that the human papillomavirus
(HPV) is a risk factor for cervical cancer, the role of other sexually transmit-
CHAPTER 5 I RISK 99
ted diseases, such as herpes simplex virus and Chlalllydia, is not as dear.
Antibodies to these agents are more common among patients with cervical
cancer than in women without cancer, but the agents may be markers for
risk of cervical cancer rather than causes, If so, curing them would not
necessarily prevent cervical cancer. On the other hand, decreasing promis-
cuity might prevent the acquisition of both the causative agent for cervical
cancer and other sexually transmitted diseases (2).
There are several ways of deciding whether a risk factor is (I cause or
merely a marker for disease. These are covered in Chapter 1L
DIAGNOSIS
Knowledge of risk can be used in the diagnostic process, since the pres-
ence of a risk factor increases the prevalence (probability) of disease among
patients-one way of improving the positive predictive value of a diagnos-
tic test.
However, in individual patients, risk factors usually are not as strong
predictors of disease as art' clinical findings of early disease. As Rose (3)
put it:
Often the best predictor of future major diseases is the presence of existing
minor disease. A low ventilatory function today is the best predictor of its
future rate of decline. A high blood pressure today is the best predictor of
its future rate of rise. Early coronary heart disease is better than all of the
conventional risk factors as a predictor of future fatal disease.
Risk factors can provide the most help with diagnosis in situations
where the factor confers a substantial risk and the prevalence of the disease
is increased by clinical findings. For example, age and sex are relatively
strong risk factors for coronary artery disease, yet the prevalence of disease
in the most at risk age and sex group, old men, is only 12%. When specifics
of the clinical situation. such as presence and type of chest pain and re-
sults of an electrocardiographic stress test, arc considered as well, the
prevalence of coronary disease can be raised to 99'10 (4).
More often, it is helpful to use the absence of a risk factor to help rule
out disease, particularly when one factor is strong and predominant. Thus
it is reasonable to consider mesothelioma in the differential diagnosis of a
pleural mass in a patient who is an asbestos worker, but mesothelioma is
a much less likely diagnosis for the patient who has never worked with
asbestos.
Knowledge of risk factors is also used to improve the efficiency of
screening programs by selecting subgroups of patients at increased risk.
PREVENTION
If a risk factor is also a cause of disease, its removal can be used to
prevent disease whether or not the mechanism by which the disease takes
place is known. Some of the classic successes in the history of epidemiology
illustrate this point. For example, before bacteria were identified, Snow
found an increased rate of cholera among people drinking water supplied
by a particular company and controlJed an epidemic by cutting off that
supply. More recently, even before HIV had been identified, studies
showed that a lifestyle of multiple sexual partners among homosexual men
was a risk factor for acquiring AIDS (5). The concept of cause and its
relationship to prevention is discussed in Chapter 11.
Studies of Risk
The most powerful way of determining whether exposure to a potential
risk factor results in an increased risk of disease is to conduct an experi-
ment. People currently without disease would be divided into groups of
equal susceptibility to the disease in question. One group would be ex-
posed to the purported risk factor and the other would not, but the groups
would otherwise be treated the same. Later, any difference in observed
rates of disease in the groups could be attributed to the risk factor.
Unfortunately, the effects of most risk factors for humans cannot be
studied with experimental studies, in which the researcher determines who
is exposed. Consider some of the questions of risk that concern us today.
I low much are inactive people at increased risk for cardiovascular disease,
everything else being equal? Do cellular phones cause brain cancer? Does
alcohol increase the risk of breast cancer? For such questions as these, it
is usually not possible to conduct an experiment. first, the experiment
would have to go on for decades. Second, it would be unethical to impose
possible risk factors on a group of the people in the study. Finally, most
people would balk at having their diets and behaviors determined by
others for long periods of time. As a result, it is usually necessary to study
risk in less obtrusive ways.
Clinical studies in which the researcher gathers data by simply observ-
ing events as they happen, without playing an active part in what takes
place, are called observational studies. Most studies of risk are observational
studies, either cohort studies, described in the rest of this chapter, or case
control st1ldies, described in Chapter 10.
COHORTS
The term cohort is used to describe a group of people who have some-
thing in common when they are first assembled and who are then observed
for a period of time to sec what happens to them. Table 5.1 lists some of
the ways in which cohorts are used in clinical research. Whatever members
of a cohort have in common, observations of them should fulfill two criteria
if they are to provide sound information about risk.
First, cohorts should be observed over a meaningful period of time in
the natural history of the disease in question. This is so there will be
CHAPTER 5 I RISK 101
Table 5.1
Cohorts and Their purposes
fa Assess
Cttnrncterist«: in Common [!feci of bam pie
---'--- - - - - - - ~--
Agc Age Life expectancy tor people age
70 (regardless of when born)
Dale at birth Calendar time Tuberculosis rates for people
born in HJ10
Exposure Risk factor Lung cancer in people who
smoke
Disease Projrosu. Survival rate tor patients with
breast cancer
F'reV(ll!lve interventiOIl Prevention Reduction in incidence of
pneumonia atter
pneumococcal vaccination
Therapeutic intervention Treatment unproverront in survival for
patients with Hodgkin's
disease given combination
chemotherapy
sufficient time for the risk to be expressed. If we wish to learn whether

neck irradiation during childhood results in thyroid neoplasms, as-year
follow-up would not be a fair test of the hypothesis that thyroid cancer is
associated with irradiation, because the usual time period between irradia-
tion exposure and the onset of disease is considerably longer.
Second, all members of the cohort should be observed over the full
period of follow-up. To the extent that people drop out of the study and
their reasons for dropping out are related in some way to the outcome,
the information provided by an incomplete cohort can be a distortion of
the true state of affairs.
COHORT STUDIES
In a cohort study (Fig. 5.2), a group of people (a cohort) is assembled,
none of whom has experienced the outcome of interest, but all of whom
could experience it. (for example, in a study of risk factors for endometrial
cancer, each member of the cohort should have an intact uterus.] On entry
to the study, people in the cohort are classified according to those character-
istics (possible risk factors) that might be related to outcome. These people
are then observed over time to see which of them experience the outcome.
It is then possible to see how initial characteristics relate to subsequent
outcome events. Other names for cohort studies are longitudinal (emphasiz-
ing that patients are followed over time), pmc.-p1'ctive (implying tht.' forward
direction in which the patients are pursued), and incidence (calling attention
to the basic measure of new disease events over time).
No Disease Exposure to Disease

Risk Factor
I [Yes.!
Exposed I [j(Q]
I T I ME >
Not exposed
Figure 5.2. Design of a cohort study of risk
The following is a description of a classical cohort study, which has

made important contributions to our understanding of cardiovascular
disease.
Example The Framingham Study (6) was begun in 1949 to identify fac-
tors associated with an increased risk of coronary heart disease (CI-ID). A
representative sample of .'),209 men and women, aged 30-59, was selected
from approximately] 0,000 persons of that age living in Framingham, a small
town near Boston. Of these, 5,127 were free of CHlJ when first examined and.
therefore, were at risk of developing CHD. These people were reexamined
biennially for evidence of coronary disease. The study ran for 30 years and
demonstrated that risk of developing CHD is associated with elevated blood
pressure. high serum cholesterol, cigardte smoking, glucose intolerance, and
left ventricular hypertrophy. There was a large difference in risk between
those with none and those with all of these risk factors.
HISTORICAL COHORT STUDIES
Cohort studies can be conducted in two ways (Fig. 5.3). The cohort can
be assembled in the present and followed into the future (a concurrent
cohort study), or it can be identified from past records and followed forward
from that time up to the present (an historical cohort study).
Most of the advantages and disadvantages of cohort studies discussed
below apply whether the study is concurrent or historical. However, the
potential for difficulties with the quality of data is different for the two.
In concurrent studies, data can be collected specifically for the purposes
of the study and with full anticipation of what is needed. It is thereby
possible to avoid biases that might undermine the accuracy of the data.
On the other hand, data for historical cohorts arc often gathered for other
purposes-usually as part of medical records for patient care. These data
may not be of sufficient quality fur rigorous research.
.~ PRESENT FUTURE
Historical
Cohort Cohort
-o-Follow-up
assembled
Concurrent
Cohort ""~?~'~'~~"'~Follow-up
assembled
Figure 5.3. Historical and concurrent cohort studies.
ADVANTAGES AND DISADVANTAGES OF COHORT STUDIES

Some of the advantages and disadvantages of cohort studies, for the
purpose of describing risk factors, are summarized in Table 5.2. Cohort
studies of risk are the best available substitutes for a true experiment when
experimentation is not possible. They follow the same logic as a clinical
trial, and they allow determination of exposure to a possible risk factor
while avoiding any possibility of bias that might occur if exposure is deter-
mined after the outcome is already known.
The principal disadvantage is that if the outcome is infrequent (which
is usually the case) a large number of people must be entered in <J study
and remain under observation for a long time before results are available.
Pur example, the Framingham Study of coronary heart disease-one of
the most frequent of the chronic diseases in America-was the largest
Table 5.2
Advantages and tnsacvanteoes of Cohort Studies
Advantages
___-==oc rJbcHlvantdgcs
_
The only WaY of establishing incidence (i.e.. Inefficient because rnarw more subjects
ubsolute risk) directly must be enrolled than experience the
Follows tile same logic as tile clinical event of interest: Iherefore, cannot he
question: If persons exposed. then do u~;e{j (or rare diseases
they get the disease? Expensive because of resources necessary
Exposure can he elicited without the l)i,JS to study many people over time
that mighl occur if outcome were ,llreacJy Results not <Ivail<ltlle tor a long time
known Assessee; nl(o relationship between disease
Carl assess the relationship between and exposure to only relatively few teeters
exposure one many drsoases u.c., those recorded at the outset of the
~;tlJdy)
104 CLINICAL EPIDFMIOLOGY
study of its kind when it began. Nevertheless, more than 5000 people had
to be followed for several years before the first, preliminary conclusions
could be published. Only 5% of the people had experienced a coronary
event during the first oS years!
A related problem with cohort studies results from the fact that the
people being studied are usually "free living" and not under the control
of researchers. 1\ great deal of effort and money must be expended to keep
track of them. Cohort studies, therefore, are expensive, sometimes costing
many millions of dollars.
Because of the time and money required for cohort studies, this ap-
proach cannot be used for all clinical questions about risk. For practical
[('<I sons, the cohort approach has been reserved for only the most important
questions. This has led to efforts to find more efficient, yet dependable,
ways of assessing risk. (TIl(' most common of these, case control studies,
is discussed in Chapter 10.)
The most important scientific disadvantage of observational studies,
including cohort studies, is that they are subject to a great many more
potential biases than are experiments. People who are exposed to a certain
risk factor in the natural course of events arc likely to differ in a great
many ways from a comparison group of people not exposed to the factor.
1£ these other differences arc also related to the disease in question, they
could account for any association observed between the putative risk factor
and the disease.
This leads to the main challenge of observational studies: to deal with
extraneous differences between exposed and nonexposcd groups to mimic
as closely as possible an experiment. The differences arc considered "extra-
ncous" from the point of view of someone hying to determine cause-
and-effect relationships. The following example illustrates one approach
to handling such differences.
Example Although the presence of sickle-cell trait (HbAS) is generally

regarded as a benign condition, several studies have suggested that it is
associated with defects in physical growth and cognitive development. A
study was undertaken, therefore, to see if children born with HbAS experi-
enced problems in growth and development more frequently than children
with normal hemoglobin (HbAA), everything else being equal (7). It WdS
recognized that a great many other factors are related to growth and develop-
ment and also to having IIbAS, Among these are race, sex, birth date, hirth
wei;;ht, gestational age, 5-min Apgar score, and socioeconomic status. If these
other factors were not taken into account, one or more of them could bias
the results of the study, and it would not be possible to distinguish the effects
of HbAS, in and of itself, from the effects of the other factors. The authors
chose to deal with these other factors by matching. For each child with HbAS,
tht,y selected a child with IfbAA who was similar with respect to the seven
other factors. A total of 100 newborns--50 with HbAS and 50 with HbAA-
were followed from birth to 3-5 vcars old, No differences in growth and
development were found. •
Major biases in observational studies and ways of dealing them are

described in Chapter 6.
Comparing Risks
The basic expression of risk is incidence, defined in Chapter 4 as the
number of new cases of disease arising in a defined population during a
given period of time. But usually we want to compare the incidence of
disease in two or more groups in a cohort that differ in exposure to a
possible risk factor. To compare risks, several measures of the association
between exposure and disease, called measures of effect, are commonly used.
They represent different concepts of risk and are used for different pur-
poses. four measures of effect are discussed below (Tables 5.3 and 5.4).
ATTRIBUTABLE RISK
First, one might ask, "What is the additional risk (incidence) of disease
following exposure, over and above that experienced by people who are
not exposed?" The answer is expressed as attributable risk, the incidence
of disease in exposed persons minus the incidence in nonexposcd persons.
Attributable risk is the additional incidence of disease related to exposure,
taking into account the background incidence of disease, presumably from
other causes. Note that this way of comparing rates implies that the risk
factor is a cause and not just a marker. Because of the way it is calculated,
attributable risk is also called risk differcllce.
Table 5.3
Measures of Effect
Expression coeenon Definition"
Attributable risk What is the incidence 01 disease AR=I,,-li.-

(risk difference) attributable to exposure?
Relative risk (risk ratio) How many times more likely are t,
RR
exposed persons to become I,
diseased. relative to nonexposeo
persons?
Population attributable risk whet is the incidence of disease in a AR~=ARxP
population, assocateo with the
occurrence of a risk factor?
Population attributable What traction of disease in CI
fraction population is attributable to
exposure to a risk factor?
'Where I, = incidence in exposed persons; r; - im;id,mce in noncxpcsed pl1rscms; f' - prevalence of

exposure to a nsk filet",; Clnd I,. - total inci<iHneH of disease ill a population,
Table 5.4
Calculating Measures of Effect: Cigarette Smoking and Death from Lung Cancer~
Simplp. risks
Death rille from lung cancer in cigarette smokers o 96/1000/yfW

Death rate from tung cancer in nonsmokers O,07/1000/year
Prevalence of cigarette smoking 56%
Total ooam rate trom lurlg cancer 0,56/ I OlJO/YC8r
Com[J8rp:!l risks
Attributable risk .'" O.96/1000/yeilr - Q,07/WOO/year

O.89/1000/year
Relative ri~;k ~ 0.96/WOO/Yf~ar ,. 0.07/1000/year
= 1:1.7
Population attributable risk ~ 0.89/WOO/yeilr x 0.56
= O.50/10oo/year
Population attributable traction = O.oO/Woo/year ~ O,5Ej/1000/year
= 0.89
'btirnated data nom Doll R, Hill A8. 8r Med J 196~; t 1399-1410,
RELATIVE RISK
On the other hand, one might ask, "How many times are exposed per-
sons more likely to get the disease relative to nonexposed persons?" To
answer this question, we speak of relative risk or risk ratio, the ratio of
incidence in exposed persons to incidence in noncxposed persons. Relative
risk tells us nothing about the magnitude of absolute risk (incidence). Even
for large relative risks, the absolute risk might be quite small if the disease
is uncommon. It does tell us the strength of the association between expo-
sure and disease and so is a useful measure of effect for studies of disease
etiology.
INTERPRETING ESTIMATES OF INDIVIDUAL RISK
The clinical meaning attached to relative and attributable risk is often
quite different, because the two expressions of risk stand for entirely differ-
ent concepts. The appropriate expression of risk depends on which ques-
tion is being asked.
Example Risk factors for cardiovascular disease are generally thought
to be weaker among the elderly than the middle-aged. This assertion was
examined by comparing the relative risks and attributable risks of common
risk factors for cardiovascular disease among different agc groups (8). An
example is the risk of stroke from smoking (Table 5..' ). The relative risk
decreases with age, from 4.0 in persons ages 45·49 to 1.4 in persons aged
65-69. However, the attributable risk increases slightly with age, mainly
because stroke is more common in the elderly regardless of smoking status,
Thus, although the causal link between smoking and stroke decreases with
age, an elderly individual who smokes increases his or her actual risk of
stroke to a similar, indeed slightly greater, degree than a younger person.
Table 5.5
Comparing Relative Risk and Attributable Risk in the Relationship of Smoking,
Stroke, and Age~
nxsocrcc (per 10001

Ilelative AttributablH
A!JH Nonsmoker,; Smokers Risk Risk
45--49 7.4 297 4.0 22,3

50-54 1/,2 37.D 2.2 19,8
55-59 27,9 64./ 2.3 367
60-64 47,~ 76.9 1.6 29.5
65- 69 1l0.2 110~ 14 30.2
'Adapted from r-ctv 8M el al. J Clin Epiderniol 1990; 43:961 -970
ln most clinical situations, because attributable risk represents the actua I

additional probability of disease in those exposed, it is a more meaningful
expression of risk for individuals than is relative risk. On the other hand,
relative risk is more useful for expressing the strength of a causal
relationship.
POPULATION RISK
Another way of looking at risk is to ask, "How much does a risk factor
contribute to the overall rates of disease in groups of people, rather than
individuals?" This information is useful for deciding which risk factors
are particularly important and which are trivial to the overall health of a
community, and so it can inform those in policy positions how to choose
priorities for the deployment of health care resources. A relatively weak
risk factor {i.e., one with a small relative risk) that is quite prevalent in a
community could account for more disease than a very strong, but rare,
risk factor.
To estimate population risk, it is necessary to take into account the
frequency with which members of a community are exposed to a risk
factor. Population attributable risk is the product of the attributable risk and
the prevalence of the risk factor in a population. It measures the excess
incidence of disease in a community that is associated with a risk factor.
One can also describe the fraction of disease occurrence in a population
that is associated with a particular risk factor, the papulation attributa/JI1'
fraction. It is obtained by dividing the population attributable risk by the
total incidence of disease in the population.
Figure 5.4 illustrates how the prevalence of a risk factor determines the
relationship between individual and population risk. Figure 5.4A shows
the attributable risk of death according to diastolic blood pressure. Risk
increases with increasing blood pressure. However, few people have ex-
tremely high blood pressure (Fig. 5.4B). When hypertension is defined as
A
o en 150
c..- ..co
c~
~>;-100 ~----~J=r...r='=tUill1
-co ..
.<:0
Q)
00.
~
~
Q) 50
Excess death
rate
attributable to
BP >90 mm Hg
01---,---,---,--1--,--,--,--
-
o 0~
-;
15 B Prevalence of
elevated BP at
lii..!
o :J
10 various levels
~ 0. 5
.. 0
0. 0. 0 ~C;C-r-_,_-jillllif
..ill 60 Percent excess
o ..
><.<:
c deaths attributable
to various levels of
w_
- ..co
~
40 hypertension
24.1
"0 20
f:!
~
o',--,----,----,---f=
50 60 70 ao 90 100 110 120 130
Diastolic Blood Pressure (mm Hg)
Figure 5.4. Relationships among attributable risk, prevalence of risk factor, and
population risk for hypertension, (Adapted from The Hypertension Detection and
Follow-up Cooperative Group, Mild hypertensives in the hypertension detection and
follow-up program. Ann NY Acad Sci 1978;304:254-266.)
having a diastolic blood pressure >',10 mm Hg, most hypertensive people

are just over 90 mm Hg, and very few are in the highest category (~115 mm
Hg). As a result, the greatest percentage of excess deaths in the population
(58.4°/<,) is attributable to relatively [ow-grade hypertension, 90-105 mm
Hg (Fig. S.4C). Paradoxically, then, physicians could save more lives by
effective treatment of mild hypertension than severe hypertension. This
fact, so counterintuitive to clinical thinking, has been termed "the preven-
tion paradox" (',I).
Measures of population risk are less frequently encountered in the clini-
cal literature than are measures of individual risk, e.g., attributable and
relative risks. But a particular clinical practice i;: as much a population
for the doctor as is a community for health policymakers. Also, how the
prevalence of exposu re affects community risk can be important in the care
of individual patients. for instance, when patients cannot give a history or
when exposure is difficult for them to recognize, we depend on the usual
prevalence of exposure to estimate the likelihood of various diseases. When
considering treatable causes of cirrhosis in a North American patient, for
example, it would be more profitable to consider alcohol than schtstosomes,
inasmuch as few North Americans are exposed to Schistosoma maneoni. Of
course, one might take a very different stance in the Nile delta, where
schistosomes are prevalent and the people, who are mostly Musl irns, rarcly
drink alcohol.
Summary
Risk factors are characteristics that are associated with an increased risk
of becoming diseased. Whether or not a particular risk factor is a cause
of disease, its presence allows one to predict the probability that disease
will occur.
Most suspected risk factors cannot be manipulated for the purposes of
an experiment, so it is usually necessary to study risk by simply observing
people's experience with risk factors and disease. One way of doing so is
to select a cohort of people, some members of which are and some of which
are not exposed to a risk factor, and observe the subsequent incidence of
disease. Although it is scientifically preferable to study risk by means of
cohort studies, this approach is not always feasible because of the time,
effort, and expense it entails.
When disease rates are compared among groups with different expo-
su res to a risk factor, the results can be expressed in several ways. Attribut-
able risk is the excess incidence of disease related to exposure. Relative
risk is the number of times more likely exposed people are to become
diseased relative to noncxposed people. The impact of a risk factor on
groups of people takes into account not only the risk related to exposure
but the prevalence of exposure as well.
REI;!':I.:ENCES
I. Weiss NS, Liff JM. An'ounting for th", multtcausal n~tur~ of disease in the design ~nJ
analysis of epidemiologic studies. Am J Epidemiol 19SJ;1l7:14-18.
2, Pr~bhat KSJ, Beral V, Pete J, Hack S, Hermon C. Deacon J. Antibodies to hum~n papillo-
rnnvirus and to other genitol infectious ~gl'nls and invasive cervical c~ncer risk. Lancet
1993;34l:1116- 1118.
3. Rose G. Sick individuals and sick populations. Int J Epldcmiol 1985;1432-38.
4. Diamond GA, Forrester jS, Analysis of probability as nn aid in the clinical diagnosis of
coronary-artery disease. N Engl J Med 1979;300:1350-1358.
110 CLINiCAL EPIDEMIOLOGY
5. Jaffe HW et al. National case-control study of Kaposi's S<lrcoma and Pncumocystis carinii
pneumonia in homosexual men. Part 1, Fpidemio!qgic results. Ann Intern Med
1983;99:145-151
6, Dawber TR. The Framingham Study. 'Ih., epidemiology of atherosclerotic disease. Cam-
bridge MA: Harvard University Press, 1980.
7. Kramer MS, Rooks Y, Pearson HA. Growth and development in children with ~ickJe-cell
trail. N Engl J Med 1978;299:686-689.
8. rsety BM ct nl. Risk ratio~ and risk differences in estimating the effect of risk factors for
cardiovascular disease in the elderly. 1 Clin EpidemioI1990;4J:961-970,
9. Hofmnn A, Vandenbroucke JP Geoffrey Rose's big id'''L Br MeJ J 1992;305:1519-1520,
SUGGESTFD READINGS
Detsky AS, O'Rourke K, Corey PN, Johnston N, fenton S, Jeejecbhoy KN, The hazards of
using adive clinic patients as " source of subjects for clinical stmlie~. J Ccn Intern Med
1988;3:260-266.
Feinstein AR. Scientific standards in epidemiologic studies of the menace of daily life. Science
19&\;242:1257-1263. Respo"-'<-: by Savitz DA, Greenland S, Stolley PD, Kelsey JL Scientific
standards of criticism: a reaction to "Scientific standards in epidemiologic studies of the
menace of daily life" by Feinstein. Epidemiology 1990;1:78-83.
Malenka DJ, Baron JA, Johansen 5, wahrenbergcr JW, Ross 1M. The framing effect of relative
and absolute risk. J Gen Intern Mcd 1993;8:543-·548-
Morganstern H, Kleinbaum LX~, Kupper LL. Measures of disease incidence used in epidemio-
logic re~eilrch, Int J EpidemioI1980;9:97-104.
Naylor CD, Chen E, Straus~ B. Measured enthusiasm: does the method of reporting trial
results alter perceptions of therapeutic effectiveness? Ann Intern Med 1992;117:916-921.
6
PROGNOSIS
When people become sick, they have a great many questions about how
their illness will affect them. Is it dangerous? Could I die of it? Will there
be pain? How long will I be able to continue my present activities? Will
it ever go away altogether? Most patients and their families want to know
what to expect, even if little can be done about their illness.
ProfillOsis is a prediction of the future course of disease following its
onset. In this chapter, we review the ways in which the course of disease
can be described. We then consider the biases that can affect these descrip-
tions and how these biases can he controlled. Our intention is to give
readers a better understanding of a difficult but indispensable task-pre-
dicting patients' futures as closely as possible. The object is to avoid ex-
pressing prognosis with vagueness when it is unnecessary, and with cer-
tainty when it is misleading.
Doctors and patients think about prognosis in several different ways.
First, they want to know the general course of the illness the patient has.
A young patient suffering from postherpetlc neuralgia associated with
herpes zoster can be assured that the pain usually resolves in less than a
month. Second, they usually want to know, as much as possible, the prog-
nosis in the particular case. Even though HIV infection is virtually univer-
sally fatal, individuals with the infection may live from a few months to
more than a decade; a patient wants to know where on this continuum
his or her particular case falls. Third, patients especially are interested to
know how <In illness is likely to affect their lives, not only whether it will
or will not kill them, but how it will change their ability to work, to walk,
to talk, how it will alter their relationships with family and friends, how
much pain and discomfort they will have to endure.
Prognosis Studies
Studies of prognosis tackle these clinical questions in ways similar to
cohort studies of risk. A gnup of patients having something in common
(a particular medical disease or condition, in the case of prognostic studies)
111
are assembled and followed forward in time, and clinical outcomes are
measured. Often, conditions that are associated with a given outcome of
the disease, i.c., prognostic factors, are sought.
CLINICAL COURSE/NATURAL HISTORY OF DISEASE
Disease prognosis can be described for either the clinical course or
the natural history of illness. The term clinical course has been used to
describe the evolution (prognosis) of disease that has come under medi-
cal care and is then treated in a variety of ways that might affect the
subsequent course of events. Patients usually come under medical care
at some time in the course of their illness when they have diseases that
cause symptoms such as pain, failure to thrive, disfigurement, or un-
usual behavior. Examples include type 1 diabetes mellitus, carcinoma
of the lung, and rabies. Once disease is recognized, it is also likely to
be treated.
The prognosis of disease without medical intervention is termed the
natural history of disease. Natural history describes how patients will
fare if nothing is done for their disease. A great many medical condi-
tions, even in countries with advanced medical care systems, often do
not come under medical care. They remain unrecognized, perhaps be-
cause they are asymptomatic or are considered among the ordinary
discomforts of daily living, Examples include mild depression, anemia,
and cancers that are occult and slow growing (e.g., some cancers of the
thyroid and prostate).
ZERO TIME
Cohorts in prognostic studies are observed starting from a point in time,
called zero time. This point should be specified clearly and be the same
well-defined location along the course of disease (e.g., the onset of symp-
toms, time of diagnosis, or beginning of treatment) for each patient. The
term inception cohurt is used to describe a group of people who are assem-
bled near the onset (inception) of disease.
Tf observation is begun at different points in the course of disease for
the various patients in the cohort, description of their subsequent course
will lack precision. The relative timing of such events as recovery, recur-
rence, and death would be difficult to interpret or misleading.
For example, suppose we wanted to describe the clinical course of pa-
tients with lung cancer. We would assemble a cohort of people with the
disease and follow them forward over time to such outcomes as complica-
tions and death. Hut what do we mean by "with disease"? If zero time
was detection by screening for some patients, onset of symptoms for others,
and hospitalization or the beginning of treatment for still others, then
observed prognosis would depend on the particular mix of zero times in
the study. Worse, if we did not explicitly describe when in the course of
CHAPTER 6 / PROGNOSIS 113
disease patients entered the cohort, we would not know how to interpret
or use the reported prognosis.
DESCRIBING OUTCOMES OF DISEASE
Descriptions of prognosis should include the full range of manifestations

that would be considered important to patients. This means not only death
and disease but also consequences of disease such as pain, anguish, and
inability to care for one's self or pursue usual activities. (The Five Os listed
in Table 1.2 are a simple way to summarize important clinical outcomes.)
In their efforts to be "scientific," physicians sometimes value certain
kinds of outcomes over others, at the expense of clinical relevance. Clinical
effects that cannot be directly perceived by patients (c.g.. reduction in
tumor size, normalization of blood chemistries, or change in serology)
are not ends in themselves. It is appropriate to substitute these biologic
phenomena for clinical outcomes only if the two are known to be related
to each other. Thus hypercalcemia is an important clinical outcome of
hyperparathyroidism only if it causes symptoms such as drowsiness or
thirst or if there is reason to believe that it will eventually lead to complica-
tions such as bone or kidney disease. If an outcome cannot be related to
something patients will recognize, the information should not be used to
guide patient care, although it may be of considerable value in understand-
ing the origins and mechanisms of disease.
HEALTH-RELATED QUALlTY·OF~L1FE MEASURES
There is growing recognition that "health" involves more than the

avoidance of negative aspects such as death and disease. Clinical activities
should have a positive Impact on how a person functions and lives. This
concept has been referred to as health-related quality of life, health status, or
functional status. Questionnaires have been developed to measure patients'
quality of life. Sometimes their use strengthens arguments for certain clini-
cal interventions. For example, a study showed that erythropoietin treat-
ment of patients with chronic renal failure not only increased patients'
hematocrits but improved their health-related quality of life (1). On the
other hand, sometimes quality-of-life measurements reveal complicated
trade-offs. A study of zidovudine (AZT) treatment in patients with mildly
symptomatic HIV infection showed that although the drug delayed pro-
gression to AIDS by an average of 0.9 months, the positive result was
offset by adverse effects of the drug. Thus patients receiving the drug had
an average of 14.5 months without disease progression or severe symptom-
atic adverse effects from AZT compared with an average of 14.7 months
for patients not receiving the drug (2). What looked like a small benefit in
delayed progression to AIDS was not so clear when quality-of-life mea-
sures were added to the study.
Prognostic Factors
Although most patients are interested in the course of their disease in
general, they Me even more interested in a prediction for their given case.
Prognostic factors help identify groups of patients with the same disease
who have different prognoses.
DIFFERENCES BETWEEN PROGNOSTIC FACTORS AND RISK FACTORS
Studies of risk factors usually deal with healthy people, whereas prog-
nostic factors-conditions that are associated with an outcome of disease-
are, by definition, studied in sick people. There arc other important differ-
ences as well, outlined below.
Different Factors
Factors associated with an increased risk are not necessarily the same
as those marking a worse prognosis and are often considerably different
for a given disease. for example, low blood pressure decreases one's
chances of having an acute myocardial infarction, but it is a bad prognostic
sign when present during the acute event (fig. 6.1). Similarly, intake of
exogenous estrogens during menopause increases women's risk of endo-
metrial cancer, but the associated cancers are found at an earlier stage and
seem to have a better-than-average prognosis.
Well Onset of Acute Outcomes

Myocardia Infarction
~~~~~~~_. Death
1m!' UWj~~~ii~~~t.1 i;]L> ~:~~~arction
Risk Factors Prognostic (Poor) Factors
rAge rAge
Male Female
Cigarette smoking Cigarette smoking
Hypertension Hypotension
!LDLI lHDL Anterior infarction
Inactivity Congestive heart failure
Ventricular arrhythmia
Figure 6.1. Differences between risk and prognostic factors for acute myocardial
infarction
CHAPTER 6 I PROGNOSIS 115
Some factors do have a similar effect on both risk and prognosis. For
example, both the risk of experiencing an acute myocardial infarction and
the risk of dying of it increase with age.
Different Outcomes
Risk and prognosis describe different phenomena. For risk, the event
being counted is the onset of disease. For prognosis, a variety of consc-
quences of disease are counted, including death, complications, disability,
and suffering.
Different Rates
Risk factors generally predict low probability events. Yearly rates for
the onset of various diseases are on the order of 1/100 to 1/10,000. As a
result, relationships between exposure and risk usually elude even astute
clinicians unless they rely on carefully executed studies, often involving a
large number of people over extended periods of time. Prognosis, on the
other hand, describes relatively frequent events. Clinicians often can form
good estimates of prognosis on their own, from their personal experience.
For example, they know that few patients with lung or pancreatic cancer
survive as long as 5 years, whereas most with chronic lymphocytic leuke-
mia survive much longer.
MULTIPLE PROGNOSTIC FACTORS AND PREDICTION RULES
A combination of factors may give a more precise prognosis than each
of the same factors taken one at a time. Clinical prediction rules estimate the
probability of outcomes according to a set of patient characteristics.
Example Once patients with HIV infection develop A[DS, the prognosis
is poor and survival time is short. Even so, and before antiviral and prophy-
lactic therapy for opportunistic infections became standard treatment, it was
clear that some patients with AIDS survived much longer than others. A
study was done to determine which patient characteristics predicted survival
(3). Each of several physiologic characteristics was found to be related to
survival. Using these factors in combination, the investigators developed a
prognostic staging system, with [ point for the presence of each of 7 factors:
severe diarrhea or a serum albumin <2.0 gm/dL, any neurologic deficit, Pl.':2
less than or equal to 50 mm Hg, hematocrit <30(}:" lymphocyte count < 150/
mL, white count <2500/mL, and platelet count < 140,000/ mL. The total score
determined the prognostic stage (I, 0 points; II, 1 point; lll, greater than or
equal to 2 points). Figure 6.2 shows the survival of AIDS patients in each
prognostic stage. Using multiple prognostic factors together, the authors
noted that prediction for median length of survival varied from [[.5 months
for patients in stage I to 2.1 months for patients in stage Ill.
Describing Prognosis
PROGNOSIS AS A RATE
It is convenient to summarize the course of disease as a single number,
or rate: the proportion of people experiencing an event. Rates commonly
1.0
'; '
,
,
0.8 : ,
•-
'"
<::
'S;
:
,,
' -,
Stage II
,
':;;: 0.6 -,,
"<:: '-,,
,
'" ,
,- -
0 , '- - 1
:e0 0.' '- -,

0. ,
-- Stage III
,----- ,
~ ----.
,---
0.2 - '-. --,

'- ---- --,,
,- ---------' ... --- -',
O~---,--_---, _ _~_--,--_---,_ _
2
• 6 8
Survival in Months
10 12
Figure 6.2. Survival of AIDS patients according to prognostic stage. Median sur-
vival times (in months): stage I, 11.6; stage II, 5.1, stage III, 2.1. (Adapted from
JusliceAC, Feinstein AR, Weils CK. A new prognostic staging system for the acquired
immunodeficiency syndrome. N Eng J Med 1989: 320:1388-1393,)
used for this purpose are shown in Table 6.1. These rates have in common
the same basic components of incidence, events arising in a cohort of
patients over time.
All the components of the rate must be specified: zero time, the specific
clinical characteristics of the patients, definition of outcome events, and
length of follow-up. Follow-up must be long enough for all the events to
occur; otherwise, the observed rate will understate the true one.
A TRADE-OFF: SIMPLICITY VERSUS MORE INFORMATION
Expressing prognosis as a rate has the virtue of simplicity. Rates can be
committed to memory and communicated succinctly. Their drawback is
that relatively little information is conveyed, and large differences in prog-
nosis can be hidden within similar summary rates.
Figure 6.3 shows 5-year survival for patients with four conditions. For
each condition, about 10°;;, of the patients are alive at 5 years. But similar
Table 6.1
Rates Commonly Used to Describe Prognosis
Hntn Detiniti!)rr'
5-year survival Percent of patients surviving 5 years from some point in the
course of their disease
Case fatality Percent of patients with a disease who die of it
Disease-specific mortality Number of people per 10,000 {or 100,0(0) population dying
of a specific disease
Response Percent 01 patients showing some evidence of improvement
following an intervention
gomsson Percent of patients entering a phase in which disCilSlo is no
longer detectable
Recurrence Perceot ot patients who have return of disease after a
disease-free interval
"Timo under observation is cnncr stated or assumed to be \OufticiArltly 10"9 so tllil! "II ffi'Arlls lila! will OcCur
have teen observed.
summary rates of approximately 10% survival obscure differences of con-

siderable importance to patients. Early survival in patients with dissecting
(Ineurysms is very poor, but if they survive the first few months, their risk
of dying is not affected by having had the aneurysm (Fig. 6.3A). On the
other hand, HIV positive patients who develop AIDS die throughout the
S ymrs (Fig. 6.38). Chronic granulocytic leukemia is a condition that has
relatively little effect on survival during the first few ymrs after diagnosis
(Fig. 6.3C). Later, there is an acceleration in mortality rate until nearly all
patients are dead 5 years after diagnosis. Figure 6.30 is presented as a
benchmark. Only at age 100 do people in the general population have a
5-year survival rate comparable to that of patients with the three diseases.
SURVIVAL ANALYSIS
When interpreting prognosis, we would like to know the likelihood, on
the average, that patients with a given condition will experience an out-
come at any point in time. When prognosis is expressed as a summary
rate it does not contain this information. However. there are methods for
presenting information about average time to event for any point in the
course of disease.
SURVIVAL OF A COHORT
The most straightforward way to learn about survival is to assemble a
cohort of patients with the condition of interest at some point in the course
of their illness (c.g.. onset of symptoms, diagnosis, or beginning of treat-
ment) and keep them under observation until all could have experienced
the outcome of interest. For a small cohort, one might then represent the
experience with these patients' course of disease as shown in Figure fi.4A.
100 A. Dissecting 100 B. AIDS

aneurysm
80 80
60 60
40 40
tll
e 20 20
"S;
'S; . -,-----,
~
-.. o
::l 0 2 3 4 5 2 3 4 5
'".c: 100 c. Chronic 100
D.Age100
c
~ granulocytic years
80· leukemia 80
"-
60 60
40 40
20 20
0 2 3 4 5 o 1 2 3 4 5
Years
Figure 6.3. A limitation of 5-year survival rates: four conditions with the same ti-
year survival rate of 10%. (Data from Anagnostopoulos CD et al. Aortic dissections
and dissecting aneurysms. Am J Cardiology 1972:30:263-273; Sash JA, Hoover
DR et al. Factors influencing survival after AIDS: report from the Multicenter AIDS
Cohort Study (MACS). J Acquir Immune Defic Syndr 1994;7:287-?95; Ksrdinal
CG et al. Chronic qranulocvtic leukemia, Review of 536 cases. Arch Intern Med
1976; 136:30tJ-313: and American College of I ite Insurance. 1979 life insurance
fact book, Washington, DC: ACLI 1979,)
The plot of survival against time displays steps, corresponding: to the death
of each of the 10 patients in the cohort. If the number of patients were
increased (Fig. 6.48), the size of the steps would diminish. If a very large
number of patients were represented, the figure would approximate a
smooth curve. This information could then be used to predict the year-by-
year, or even week-by-week. prognosis of similar patients.
Unfortunately, obtaining the information in this way is impractical for
several reasons. Some of the patients would undoubtedly drop out of the
study before the end of the follow-up period. perhaps because of another
illness, a move to a place where follow-up was impractical, or dissatisfac-
tion with the study. These patients would have to be excluded from the
cohort, even though considerable effort may have been exerted to gather
data on them up to the point at which they dropped out. Also, it would
be necessary to wait until all of the cohort's members had reached each
point in time before the probability of surviving to that point could be
calculated. Because patients ordinarily become available for a study over
a period of time, at any point in calendar time there would be a relatively
long follow-up for patients who entered the study first, but only brief
experience with those who entered recently. The last patient who entered
the study would have to reach each year of follow-up before any informa-
tion on survival to that year would be available.
SURVIVAL CURVES
To make efficient use of all available data from each patient in the
cohort, a way of estimating the survival of a cohort over time, called
survival analysis, has been developed. (The usual method is called a Kaplan-
Meir analysis, after the originators.) The purpose of survival analysis is
not (as its name implies) only to describe whether patients live or die. Any
outcome that is dichotomous and occurs only once during follow-up-
e.g., time to coronary event or to recurrence of cancer-can be described
in this way. 'When an event other than survival is described, the term time-
to-c'vcnt analysis is sometimes used.
Figure 6.5 shows a typical survival curve. On the vertical axis is the
100 100
s
l:
A. 10 patients B. 100 patients
80
'"
~
-
n.
o
~
60
40
1: 20
E
::J
Z
o 1 2 3 4 5 0 2 3 4 5
Time (years)
Figure 6.4. Survival of two cohorts, small and large, when all members are ob-
served for the full period of tallow-up.
415=80% 112=50%
Probability of Surviving Interval: 100% III 100% 11 100%
"0_ 10 0
- "'-
>~
=CQ
...
.c>
-
.c~
0::>
~(J)
0.
2
Time (Years)
Figure 6.5. A typical survival curve, with detail for ono part of the curve,
probability of surviving, and on the horizontal axis is the period of time

following the beginning of observation. Often, the numbers of patients at
risk at various points in time are shown to give some idea of the contribu-
tion of chance to the observed rates.
The probability of surviving to any point in time is estimated from the
cumulative probability of surviving each of the time intervals that preceded
it. Time intervals can be made as small as necessary; in Kaplan-Meir analv-
scs, the intervals are between each new event (death) and the preceding
one. Most of the time, no one dies, and the probability of surviving is 1.
When one or more patients die, the probability of surviving is calculated
as the ratio of the number of patients surviving to the number at risk of
dying at that time. Patients who have already died, dropped out, or have
not yet been followed-up to that point are not at risk of dying and so are
not used to estimate survival for that time. When patients are lost from
the study at any point in time, for any reason, they are censored, i.c., they
arc no longer counted in the denominator. The probability of surviving
does not change during intervals in which no one dies; so in practice, the
probability of surviving is recalculated only for times when there is a

death. Although the probability assigned at any given interval is not very
accurate, because of the small number of events involved, the overall prob-
ability of surviving up to each point in time (which is the product of all
preceding probabilities) is remarkably accurate.
A part of the survival curve in Figure 65 (from 3 to 5 years after zero
time) is presented in detail to illustrate the data used to estimate survival:
patients at risk, patients no longer at risk (censored), and patients experi-
encing outcome events at each point in time.
INTERPRETING SURVIVAL CURVES
Several points must be kept in mind when interpreting survival curves.

First, the vertical axis represents the estimated probability of surviving for
members of a hypothetical cohort, not the percent surviving for an actual
cohort.
Second, points on a survival curve are the best estimate, for a given set
of data, of the probability of survival for members of a cohort. However,
the precision of these estimates depends, as do all observations on samples,
on the number of observations on which the estimate is based. One can
be more confident that the estimates on the left-hand side of the curve are
sound, because more patients are at risk during this time. But at the tail
of the curve, on the right, the number of patients on whom estimates of
survival are based often becomes relatively small because deaths, dropouts,
and late entrants to the study result in fewer and fewer patients being
followed for that length of time. As a result, estimates of survival toward
the end of the follow-up period are imprecise and can be strongly affected
by what happens to relatively few patients. For example, in Figure 6.5, the
probability of surviving is 8'1" at 5 years. If at that point the one remaining
patient happens to die, the probability of surviving would fall to zero.
Clearly, this would be a too literal reading of the data. Estimates of survival
at the tails of survival curves must, therefore, be interpreted with caution.
Finally, the shape of some survival curves, particularly those in which
most patients experience the event of interest, gives the impression that the
event occurs more frequently early on than later, when the slope reaches a
plateau and it appears that the risk of outcome events is considerably less.
But this impression is deceptive. As time passes, rates of survival are being
applied to a diminishing number of people, causing the slope of the curve
to flatten even when the rate of outcome events does not change.
Variations on the basic survival curve are found in the medical literature
(Fig. 6.(,). Often the proportion with, rather than without, the outcome
event is indicated on the vertical axis; the curve then sweeps upward and
to the right. Other variations increase the amount of information presented
with the curve. The number of patients at risk at various points in time
45
40
'"~~
.0_
35
,,-
--
"<II
o
" W~
~
30
25
Severe
(~75% stenosis)
1ii
c: .!:! 20
" E
"
1ii.r:: " 15
E "
:;::::;.!?
81 10
Normal
5 (:'030"1., stenosis)
6 12 18 24 30 36 42
Time from Detection of
Asymptomatic Neck Bruit (months)
N{severe 94 80 68 52 42 28 21 7
Normal 242 236 211 169 158 69 56 1
Figure 6.6. Survival curve showing comparison of two cohorts, number of people
at risk, and 95% confidence intervals for observed rates. These curves show the
cumulative probability of a cerebral ischemic event from time of diagnosis, according
to the initial degree of carotid stenosis. (Data from Chambers SR, Norris, JW. Out-
come in patients with asymptomatic neck bruits, N Engl J Med 1986; 315:860-865.)
can be included under the horizontal axis; the precision of estimates of

survival, which declines with time because fewer and fewer patients are
still under observation as time passes, can be identified by confidence
intervals (see Chapter 9); and survival curves for patients with different
characteristics (e.g., patients with different prognostic factors or treat-
ments) can be compared in the same figure. Sometimes tics (not shown in
Fig. 6.6) are added to the survival curves, to indicate each time a patient
is censored.
Survival curves can be constructed for combinations of prognostic fac-
tors. This can be done by stratifying patients according to the presence or

absence of a set of prognostic factors, as shown earlier in this chapter. A
statistical technique called the Cox proportional hazards regression model can
be used to identify a combination of factors that best predicts prognosis
in the group of patients under study or the effect of individual factors
independently (Chapter 9).
Bias in Cohort Studies

Potential for bias exists in any observation. Bias in cohort studies-
whether to study risk or prognosis-can create apparent differences when
they do not actually exist in nature or obscure differences when they really
do exist.
Bias can be recognized more easily when one knows where it is most
likely to occur in the course of a study. First, it is important to determine
if bias could be present under the conditions of the study. Second, deter-
mine if bias is actually present in the particular study being considered.
Third, decide if the consequences of bias are sufficiently large that they
distort the conclusions in a clinically important way. If damage to the
study's conclusions is not very great, then the presence of bias will not
lead to misleading results. Some of the characteristic locations of bias in
cohort research are illustrated in Figure 6.7 and described below.
SUSCEPTIBILITY BIAS
A form of selection bias, called susceptibility bias, occurs when groups
of patients assembled for study differ in ways other than the factors under
study. These extraneous factors, not the particular factors being studied,
may determine the outcome. A comparable term is assembly bias. Groups
Prognostic Outcomes
Factor
Present
I
~
Absent .. 1.1!Iii;!·1
LMeasurement~
Selection
Potential Sampling
Biases
Figure 6.7. Locations of potential bias in cohort studies.

100
80
60
o 3 6 9 12 15 18 21 24
Months
Figure 6.8. Disease-free survival according to CEA levels in colorectal cancer pa-
tients with similar pathological staging (Dukes B). (Redrawn from Wolmark N et
at The prognostic significance of preoperative carcinoembryonic antigen levels in
colorectal cancer. Results from NSABP clinical trials. Ann 8urg 1984; 199:375-382.)
being; compared are not equally susceptible to the outcome of interest,

other than the factor under study.
Susceptibility bias in prognosis studies may be due to one or more
differences among cohorts, including the extent of disease, the presence of
other diseases, the point of time in the course of disease, and prior treat-
ment. The following illustrates how susceptibility bias was assessed in a
study of the prognostic value of carcinoembryontc antigen results in pa-
tients with colorcctal cancer.
Example Increased levels of carcmoembryonic antigen (CEA), a tumor-
associated fetal antigen, arc found in several types of tumors, induding colo-
rectal cancer. /\. study was undertaken to determine if preoperative CEA
levels predict relapse of disease after surgical resection with the intent to cure
(4). CEA levels were found to correlate with the extent of disease (frequently
categorized according to "Dukes classification": A, tumors confined to the
bowel wall; H, tumors extending through the bowel wall but not to the lymph
nodes; C, tumors involving regional lymph nodes; and 0, tumors having
distant metastases). Mean CEA levels varied with extent of disease: 4 for
Dukes A, 9 for B, 32 for C, and 251 for D. Both Dukes classification and
CEA level strongly predicted disease relapse. But did the CEA level predict
independently of the Dukes classification or was susceptibility of relapse
explained by Dukes classification alone? To answer this question, the associa-
tion of preoperative CFA levels to disease relapse was examined for patients
in each Dukes classification. Figure 6.8 shows that for Dukes B classification,
CEA levels independently predicted relapse. Similar results were found for
patients with Dukes C tumors. Therefore, the association between CEA levels
and likelihood of relapse could not be explained by susceptibility bias for
patients with Dukes Hand C colorectal cancers, and CEA is an important
independent prognostic factor.
SURVIVAL COHORTS
True cohort studies should be distinguished from studies of survival
cohorts in which patients are included in a study because they both have
a disease and are currently available-perhaps because they are being
seen in a specialized clinic. Another term for such groups of patients is
available patient cohorts. Reports of survival cohorts arc misleading if they
are presented as true cohorts. Tn a survival cohort, people are assembled
at various times in the course of their disease, rather than at the beginning,
as in a true cohort study. Their clinical course is then described by going
back in time and seeing how they have fared up to the present (Fig. 6.9).
The experiences of survival cohorts are sometimes presented as if they
Observed True
True Cohort Improvement Improvement
Assemble Measure outcomes

cohort
75 50% 50%
Improved:
(N= 150)
Not improved: 75
Survival Cohort
Assemble
patients
1
Begin Measure outcomes
follow-up 80% 50%
Improved: 40
(N.501 Not improved: 10
-
1 Not 1
1
1
observed 1
1- - --I
1 (N=100) 1 ~
1 1 Dropouts
Improved: 35
Not improved: 65
Figure 6.9. Comparison of a true and a "survival" cohort: in the survival cohort,
some at the patients present at the beginning are not included in tile follow-up.
were descriptions of the course of disease from its inception. However,

they may represent a biased view, because they include only those patients
who are available for study some time after their disease began. For lethal
conditions, the patients in a survival cohort are the ones who are fortunate
enough to have survived and so are available for observation years later.
For diseases that remit, the patients arc the ones who are unfortunate
enough to have persistent disease. In effect, survival cohorts describe the
past history of prevalent cases and not what one would expect over the
time following the onset of disease. Thus a survival cohort is a special case
of assembly bias.
Reports of survival cohorts are relatively common in the medical litera-
ture, particularly in the form of "case series" (discussed in Chapter 10).
Such reports can make an important contribution, primarily by describing
early experiences with newly defined syndromes, but they represent tenta-
tive, not conclusive, observations.
Example Concern has been raised about the possibility that silicone
breast implants may cause autoimmune symptoms of rheumatic disease. A
study was, therefore, done of 156 women with silicone breast implants and
rheumatic disease complaints (5). The patients were consecutive referrals to
three rheurnatologists who were known for their interest in silicone implants
and rheumatic disease. Serologic tests in the women were compared to those
of women without implants but with fibromyalgla and to tests in women
with implants but no rheumatic symptoms. The clinical findings in the
women with implants and complaints were described; most did not fulfill
criteria for rheumatoid arthritis and most had normal immunologic tests.
However, l4 patients had scleroderma-like illness and abnormal serology
that was not found in the comparison groups. Because of the possible biases
that can occur in the assembly of patients for this case series, the authors
were cautious about their findings, concluding that "the hypotheses raised
in this study and others should be tested in large, population-based studies."
Publication of the first such study does not support the hypothesis (6).
MIGRATION BIAS
Migration bias, another form of selection bias, can occur when patients
in one group leave their original group, dropping out of the study alto-
gether or moving to one of the other groups under study. If these changes
take place on a sufficiently large scale, they can affect the validity of
conclusions.
In nearly all studies, some members of an original group drop out over
time. If these dropouts occur randomly, such that the characteristics of lost
subjects in one group are on the average similar to those lost from the
other, then no bias would be introduced. This is so whether or not the
number of dropouts is large or the number is similar in the groups. But
ordinarily the characteristics of lost subjects arc not the same in various
groups. The reasons for dropping out-death, recovery, side effects of
treatment, etc.v-ere often related to prognosis and may also affect one
group more than another. As a result, groups in a cohort that were compa-
rable at the outset may become less so as time passes.
As the proportion of people in the cohort who are nnt followed up
increases, the potential for bias increases. It is not difficult to estimate how
large this bias could be. All one needs is the number of people in thc
cohort, the number not accounted for, and the observed outcome rate.
Example Thompson et al. described the long-term outcomes of gastro-
gastrostomy {7)./\ cohort of 123 morbidly obese patients was studied 19-47
months after surgery. Success was defined as having lost more than 30% of
excess weight.
Only 103 patients (84%) could be located. In these, the success rate of
surgery was 60/lO3 (58°1<,). To determine the range within which the true
success rate must lie, the authors did a best case/worst case analysis. Success
rates were calculated, assuming that all of the patients lost to follow-up were,
on the one hand, successes (best case) and, on the other hand, failures (worst
case). Of the total cohort of 123 patients, 103 were followed up and 20 were
lost to follow-up. The observed success rate was 60/103, or 58%. In the best
case, all 20 patients lost to follow-up would be counted as successes, and the
success rate would be (60 + 20)/123, or 65':0. In the worst case, all 20 patients
would be counted as failures, and the success rate would be 60/123, or 49"/0.
Thus the true rate must have been between 49 and 65%; probably, it was
closer to 58%, the observed rate, because patients not followed up are unlikely
to be all successes or all failures.
Patients may also cross over from nne group to another in the cohort
during their follow-up. Whenever this occurs, the original reasons for pa-
tients being in one group or the other no longer apply. Tf exchange of
patients between groups takes place on a large scale, it can diminish the
observed difference in risk compared to what might have been observed
if the original groups had remained intact. Migration bias due to crossover
is more often a problem in risk than in prognosis studies, because risk
studies often go on for many years. On the other hand, migration from
one group to another can be used in the analysis of a study.
Example The relationship between lifestyle and mortality was studied
by classifying 10,269 Harvard College alumni by physical activity, smoking
status, weight, and blood pressure in 1966 and again in 1977 (8). Mortality
rates were then observed over a a-veer period from 1977 to 1985. It was
recognized that original classifications might change, obscuring any relation-
ship that might exist between lifestyle and mortality. To deal with this, the
investigators defined four categories: men who maintained high-risk life-
styles. those who changed from low- to high-risk lifestyles, those who
changed from high- to low-risk life~tyles, and those who maintained 101'.'-
risk lifestyles. After adjusting for other risk factors, men who increased their
physical activity from low to moderate amounts, quit smoking, lost weight
to normallevels, and/or became normotensive all had lower mortality than
men who maintained or adopted high-risk characteristics, but not as low as
the rates for alumni who never had any risk factors.
MEASUREMENT BIAS
Measurement bias is possible if patients in one ,jroup stand a better chance
of having their outcome detected than those in another group. Obviously,
some outcomes, such as death, cardiovascular catastrophes, and major
cancers, arc so obtrusive that they are unlikely to be missed. But for less
dear-cut outcomes-the specific cause of death, subclinical disease, side
effects, or disability-measurement bias can occur because of differences
in the methods with which the outcome is sought or classified.
Measurement bias can be minimized in three general ways. One can
ensure that those who make the observations are unaware of the group to
which each patient belongs, can set careful rules for deciding whether or
not an outcome event has occurred (and follow the rules), and can apply
efforts to discover events equally in all groups in the study.
Example Chambers and Norris studied the outcome of patients with
asymptomatic neck bruits (9). A total of 500 asymptomatic patients with
cervical bruits were observed for up to 4 years. Patients were classified ac-
cording to the degree of initial carotid artery stenosis by Doppler ultrasonog-
raphy. Outcomes were change in degree of carotid stenosi s and incidence of
cerebral ischemic events.
To avoid biased measurements, the authors estimated carotid stenosis
using established, explicit criteria fur mterprottng Doppler scans and made
the readings without knowledge of the auscultatory or previous Doppler
findings. Clinical and Doppler assessments were repeated every 6 months,
and all noncomplying patients were telephoned to determine whether out-
comes had occurred.
This study showed, among other things, that patients with >75'1" carotid
stenosis had a >2IY:;, incidence of cerebral ischemic events in 3 years, more
than 4 times the rate of patients with <30% stenosis (see Fig. fl.fl).
Dealing with Selection Bias

To determine how a factor is related to prognosis, ideally we would
like to compare cohorts with and without the factor, everything else being
equal. But in Teal life "everything else" is usually not equal in cohort
studies.
What can be done about this problem? There are several possible ways
of controlling for differences during either designing or analyzing research
(Table 6.2).1 For any observational study, if one or more of these strategies
have not been applied, the reader should be skeptical. The basic question
is, "Are the differences in prognosis in the groups related to the particular
factor under study or to some other factorrs)?"
I Co"I,,,1 h,,, ,,'v,'r~1 m,'~nings in r""'arch: I,,) gcner"I term ror any Im,,-,'," -_.n..stridion. !notching, ~l," lifi-
Glth",. "djll'lrHt'!lt-~im"d ~l ,,'mo,'in,. the' efi,'cts of exl,"neo", vari,'ble' while ,,,,,milling the' ind,'pen
d,'nl di,'cts 01 one variabk. (bJ lhe non'''I"""d p~"rl" in" cohort ,tudy (0 conlusing use of H", It"",),
(<.'! th .. npnt"',lt,'{i p"tic,~ts i~" clinical tl'i<11. "nd (.1.1 ,,,,ndi,,',,,,·d I'~"pl,' (non,·"",,) ;n ,1 (~",' conll'ol study
(sec Chapter "111).
Table 6.2
Methods for Controlling Selection Bias
Phase of Study
Method Description lJesign Analys,s
Randomization Assign patients to groups in a way that

gives each patient an equal chance
of falling into one or the other group
Restriction Unlit the renqe at characteristics of
patients in the study
Matching For each patient in one group, select +
one or more patients with the same
charactensncs {except for the one
under study) for a comparison group
Strctucaton Compare rates within subgroups
(strata) with otherwise similar
probability of the outcome
Adjustment
Simple Mathematically adjust crude rates for
one or a few characteristics so that
equal weight is given to strata at
similar nsk
Multiple Adjust for differences in a large number +
of factors related to outcome. using
mathematical modeling techniques
Best case/ Describe how different the results +
worst case could be under the most extreme (or
simpty very unlikely) conditions at
selection bias
RANDOMIZATION
The only way to equalize all extraneous factors, or "everything else," is to
assign patients to groups randomly so that each patient has an equal chance
of falling into the exposed or unexposed group. A special feature of random-
ization is that it not only equalizes factors we think might affect prognosis,
it also equalizes factors we do not know about. Thus randomization goes a
long way in protecting: us from incorrect conclusions about prognostic factors.
However, it is usually not possible to study prognosis in this way. The special
situations in which it is possible to allocate exposure randomly, usually to
study the effects of treatment on prognosis, will be discussed in Chapter 7.
RESTRICTION
Patients who are enrolled in a study can be restricted to only those
possessing a narrow rilnge of characteristics, to equalize important extrane-
ous factors. For example, the effect of age on prognosis after acute myocar-
dial infarction could be studied in white males with uncomplicated anterior
myocardial infarctions. However, one should keep in mind that although

restriction on entry to a study can certainly produce homogeneous groups
of patients, it does so at the expense of generalizability. In the course of
excluding potential subjects, cohorts may be selected that are unusual and
not representative of most patients with the condition.
MATCHING
Patients can be matched as they enter a study so that for each patient
in one group there are one or more patients in the comparison group with
the same characteristics except for the factor of interest. Often patients are
matched for age and sex, because these factors are strongly related to the
prognosis of many diseases. But matching for other factors may be called
for as well, such as stage or severity of disease, rate of progression,
and prior treatments. An example of matching in a cohort study of sickle-
cell trait was presented in the discussion of observational studies in
Chapter 5.
Although matching is commonly used and can be very useful, it controls
for bias only for those factors involved in the match. Also, it is usually not
possible to match for more than a few factors, because of practical difficul-
ties in finding patients who meet all the matching criteria. Moreover, if
categories for matching are relatively crude, there may be room for sub-
stantial differences between matched groups. For example, if a study of
risk for Down's syndrome were conducted in which there was matching
for maternal age within 10 years, there could be a nearly lO-fold difference
in frequency related to age if most of the women in one group were 30
and most in the other 39. Also, once one restricts or matches on a variable,
its effects on outcomes can no longer be evaluated in the study.
STRATIFICATION
After data are collected, they can be analyzed and results presented
according to subgroups of patients, or strata, of similar characteristics.
Example Let us suppose we want to compare the operative mortality
rates for coronary bypass surgery at hospitals A and B. Overall, hospital A
noted 48 deaths in 1200 bypass operations (4'\,), and hospital B experienced
64 deaths in 2400 operations (2.6%,).
The crude rates suggest that hospital 13 is superior. Or do they? Perhaps
patients in the two hospitals were not otherwise of comparable prognosis.
On the basis of age, myocardial function, extent of occlusive disease, and
other characteristics, the patients can be divided into subgroups based on
preoperative risk (Table 6.3); then the operative mortality rates within each
category Of stratum of risk can be compared.
Table 6.3 shows that when patients arc divided by preoperative risk, the
operative mortality rates in each risk stratum are identical in two hospitals:
6';;', in high-risk patients, 4uj" in medium-risk patients, and 0.67~j, in low-
risk patients. The obvious source of the misleading impression created by
Table 6,3
Example of Stratification: Hypothetical Death Rates after Coronary BYP8SS Surgery
in Two Hospitals, Stratified by Preoperative Risk
Hospital A Hospital B
Preoperative ._--_.
Ri"k Patients Deaths R~te (%) Patients Dpatlls Rate {%)
High SOO so 6 400 24 6

Medujrrt 400 16 4 800 32 4
L.ow 300 2 0.67 1200 !J 0,67
ToWI 1200 48 4 2400 64 2.6
examining only the crude rates is the important differences in the risk charac-
teristics of the patients treated at the two hospitals: 42% of hospital A's
patients and only 17% of hospital B's patients were high risk.
Stratification is one of the most common and most revealing ways of

examining for bias,
STANDARDIZATION
Two rates can be compared without bias if they are adjusted so as to
equalize the weight given to another factor that could be related to out-
come, This process, called standardization (or adjustment), shows what the
overall rate would be if strata-specific rates were applied to a population
made up of similar proportions of people in each stratum. Tn the previous
example, the mortality rate of 6% for high-risk patients receives a weight
of 500/1200 in hospital A and a much lower weight of 400/2400 in hospital
B, and so on, such that the crude rate for hospital A = (500/1200 x 0.06)
+ (400/1200 x 0.04) + (300/1200 - 0.0067) = 0.04 and the crude rate for
hospital 13 equals (400/2400 x 0.06) + (800/2400 x 0.04) + (1200/2400 x
O. 0067) = 0.026.
If equal weights are used, let us say 1/3 (but they could be based on
one or the other hospital or any reference population), then the standard-
ized rate for hospital A = (1/3 x 0,(6) + (1/3 x 0.(4) + (1 /3 x 0.0067)
= 0.036, which is exactly the same as the standardized rate for hospital B.
The consequence of giving equal weight to strata in each group is to remove
totally the apparent excess risk of hospital A.
The difference between the crude operative mortality rates in the two
hospitals results from the bias introduced by the differences in patients'
preoperative risk. We are only interested in differences attributable to the
hospitals and their surgeons, not to the patients per se. The difference in
the crude mortality rates is confounded by the differences in patients,
whereas standardized mortality rates equalize the weight of patients' pre-
operative risk in the two hospitals. Standardization is found much more
commonly in studies of risk (in which rates <Ire frequently standardized

for <lge, sex, and/or race) than in studies of prognosis. In contrast to strati-
fication (which is often used in prognosis studies), standardization removes
the effect of the extraneous factor. With stratification, the effect can still be
examined, even if controlled for. Thus, with standardization, we found
that patients had similar prognoses in hospitals A and 13. With stratification,
we also learned the mortality rates among patients in different risk strata.
MULTIVARIABLE ADJUSTMENT
In most clinical situations, many factors act together to produce effects.
The associations among these variables are complex. They may be related
to each other as well as to the outcome of interest, the effect of one might
be modified by the presence of others, and the joint effects of two or more
might be greater than the sum of their individual effects.
Multivariable analysis is a method for simultaneously considering the
effects of many variables (Chapter (1). It is used to adjust (control) simulta-
neously for the effects of many variables to determine the independent
effects of one. Also, the method can select from a large set of variables a
smaller subset that independently and significantly contributes to the over-
<111 variation in outcome and can arrange variables in order of the strength
of their contribution. Cox's proportional hazard analysis is a type of multivari-
able analysts used when the outcome is the time to an event (as in survival
analyses).
Multtvanablc analysis is the only feasible way to deal with many vari-
ables at one time during the analysis phase of a study. (Randomization
also controls for multiple variables, but during the design and conduct
phases of a study.) Simpler methods, such as stratification or matching,
can only consider <I few variables at a time and then only by sacrificing
statistical power.
SENSITIVITY ANALYSIS
When data on important prognostic factors are not available, it is possi-
ble to estimate the potential effects on the study by assuming various
degrees of maldistribution of the factors between the groups being com-
pared and seeing how that would affect the results. The general term
for this process is sensitivity analysis. The best case/worst case analysis.
described earlier in this chapter, is a special type of sensitivity analysis in
which one compares results assuming the best and worst possible maldis-
tribution of a prognostic variable.'
Assuming the worst is a particularly stringent test of how a factor might
affect the conclusions of <I study. A less conservative approach is to assume
th<lt the factor is distributed between the groups in an unlikely way.
) Sen~itivit}' ;ll1~Jysis can ,,1.,0 lx, used to assess the )Jol,·nti,,1 l'ifects of in"eeur"de, in the dat" o,~d in
J~("i-,ion analysi, '1-' discussed in Ch,'pkr 4.
Example A studv of treatment for mild diabetes found that patients

given the sulfonylurea tolbutamide experienced '1 greater risk of dying from
cardiovascular disease than those given insulin or diet alone. The results
were criticized because data on smoking-known to be associated with car-
diovascular death-were not collected and not taken into account in the
analysis. It was suggested that if cigarette smokers were unequally distrib-
uted among the groups, such that there were more smokers among those
receiving tolbutamide than in the other groups, then the difference in death
rates might be related to smoking, not tolbutamide. However, Cornfield (IO)
pointed out that even if cigarette smokers in the tolbutamide group exceeded
those in the control group by 20%, a situation that would have been extremely
unlikely by chance (1/50,000), an increased risk in the tolbutamide group
would have persisted. Thus bias in the distribution of smokers was unlikely
to have accounted for the observed differences.
OVERALL STRATEGY
Except for randomization, all ways of dealing with extraneous differ-
ences between groups have a limitation: They are effective against only
those factors that are singled out for consideration. They do not deal with
prognostic factors that are not known at the time of the study or are known
but not taken into account.
Ordinarily, one does not rely on only one or another method of control-
ling for bias; one uses several methods together, layered one on another.
Thus in a study of whether the presence of ventricular premature contrac-
tions decreases survival in the years following acute myocardial infarction,
one might (a) restrict the study to patients who arc not very old or young
and do not have unusual causes (c.g.. mycotic aneurysm) for their in-
farction; (b) match for age, a factor strongly related to prognosis but extra-
neous to the main question; (c) examine the results separately for strata of
differing clinical severity (e.g., the presence or absence of congestive heart
failure or other diseases, such as chronic obstructive pulmonary disease);
and (d) using multivariable analysis, adjust the crude results for the effects
of all the variables other than the arrhythmia, taken together, that might
be related to prognosis.
Generalizability and Sampling Bias

Published accounts of disease prognosis that are based on experience in
special centers can paint a misleading picture of prognosis in less selected
patients. This is so even if a study is well done, biases arc carefully con-
trolled for, and the reported prognosis for a medical condition is correct
for the particular sample of patients. Because of the sample of patients
used, it may be that the study findings are not generalizable to most other
patients with the condition, or to your patient.
Sometimes, patients in randomized controlled trials who are assigned
to the control group are studied to better determine the usual clinical
course of a disease. But such patients may not be representative of most
134 CLINICAL EPIOEMIOLOGY
patients because volunteers for studies tend to do better than patients who
do not volunteer. For example, in a large Canadian study of breast cancer
screening among women in their 40s, 90'1" of women who were in the
control group and had invasive breast cancer were alive 7 years later, and
the number of deaths from breast cancer were lower than for Canadian
women generally (11).
Summary
Prognosis is a description of the course of disease from its onset. Com-
pared to risk, prognostic events arc relatively frequent and often can be
estimated by personal clinical experience. However, cases of disease ordi-
narily seen in medical centers and reported in the medical literature are
often biased samples of all cases and tend to overestimate severity.
Prognosis is best described by the probability of having experienced an
outcome event at any time in the course of disease. Tn principle, this can
be done by observing a cohort of patients until all who will experience
the outcome of interest have done so. However, because this approach is
inefficient, another method-called survival, or time-to-event analysis-
is often used. TIle onset of events over time is estimated by accumulating
the rates for all patients at risk during the preceding time intervals.
As for any observations on cohorts, studies comparing prognosis in
different groups of patients can be biased if differences arise because of
the way cohorts arc assembled, if patients do not remain in their initial
groups, and if outcome events arc not assessed equally. A variety of strate-
gies arc available to deal with such differences as might arise, so as to
allow fair (unbiased) comparisons. These include restriction, matching,
stratification, standardization, multivariable analysis, and sensitivity analy-
sis. One or more of these strategies should be found whenever comparisons
are made.
RFI'ERENCES
I. Laupacis A. Changes in quality of lik ~nd functional capacity in hemodialysis patient.;
treated with recombinant human erythropoietin. The C:,,"~dian Erythropoietin Study
Group. Semin '\!~phroI199();2{SuppJ 1)1:1119.
2. Celber RD ct al. Quality-of-lin- evaluation in " clinical trial of zidovidine therapy in
patients with mildly symptomatic I IIV infection. Ann Inkm 'vied 1992;116:961-'0166.
3. Justice AC, Feinstein AR, Wells CK. A ,,,'W prognostic staging system for thl' acquired
immunodeficiency syndrome. N Fng ] Med 19S9;320:lJHH-1393
4. Wolmark N et <11. Thl' prognostic significance of preoperanvc carcinoembryonic antigen
IeveIs in coloredal cancer. Results i",m the f\SABP cllnicaltnals. Ann Sutg 19M;199:375-
382.
5. Bridg,,'s AI, Conley C. Wang G, BlIrn~ DF. Vasey FIJ. A clinical and immunologic evalua-
tion of women with silicone breast implants and s}'mpt()m~ of rheumatic disease. Ann
Intern Med 1993;]]8:929 <HIi.
Ii. Cabri~l SC, O'Fullon Wlvl, Kurland IT, Beard CM, Woods [E, Ml'iton LJ Ill. Risk of
CHAPTER 6 ! PROGNOSIS 135
connective-tissue ui~eascs and other di~(lrd..r~ after breast implantation. N Engl J Med
1994;330:1fi97 1702.
7. Thompson KS, Fletcher SW, ()'I\bli~y MS, Buckwalter JA. Long-term outcomes of mor-
bidly obese patients tr""ted with gastrogastrostomy. J Cen Intern Med 1986;1:85-99.
H. l'affcnbarger RS, Hyde RT, Wing AL, lee 1M, Jung lJL, Kempen JB The a~~ociati()n of
changes in physical-activity level and other lifestyle characteristics with mortality among
men. N FngJ Med 1993;328:538-545.
9 Chambers 1m, Norris JW Outcome in patients with a~ymrtom"tic neck bruits, N Engl J
'vkd 1<J1l6;315:860-865.
10. Cornfield J. The University Croup Diabetes Pmgram, A further statistical analy~i~ of the
mortality findings, .lAMA 1971;217:1676-1687.
11 Miller AB, Baines Ct. Teress Tc, W,,11 Co Canadian Nalional llr"ast Screening Study. 1.
Breast cancer detection "nd death rates among wom"n aged 40 to 49 year~. Cim Med
Assoc J 1992:147:14_~<J 1470.
SUCCESTED READINGS
C"lton T. Statistics in medi,in". 13oston; Liulc, Brown, 1975, PI" TJ7-25(),
Concato J, Feinstein AI<, Hulford TR. '111e risk of determining risk with multivariable models.
Ann Intern M,-dI993;1 HUOI-21O.
Feinstein AI<, Sosin DM, Wells CK. The INili I<og"rs phenomenon-stage migration and new
diJgn()~tic techniques as a source of misleading statistics for ~urvivJI in Cclnn'L N Lngl J
M"d 19H5;312;1604-1608.
Cuyatt Cll, Feeny DH, P~trkk D!.. 'v1easuring health-related gu,dity of life. Ann Intern M~d
19')3;118:622-629.
Horwitz RL Th" "xperimental paradigm and observational studies of cause-died relationships
in dini~al medicine, J Chron Dis 19R7;40:91--99,
l.aupacis A, Wells C, Rirhardson S, Tugwell P, Users' guides to the medical literature. V
How to 118e ~n "rticle about prognosis. lAMA 1994;272;234-237
Peto R d al.l.ksign and analysis of r,mdumiz,xi clinicaltrials requiring prolonged observation
of each patient. 11. Analysis and .., xmnples, l::lr J Cancer 1977;35:1-39.
W"ssonJH, Sox HC, Ndl I~K, Goldman L. Clinical pn,dktion rules; applications and method·
ological standards. N Engl J Med 19S5;313:793-799
weiss NS: Cliniccll epidemiology: the study "I' the outcomes of illness. :\"w York: Oxford
University I'n"s, ')<JS6,
7
TREATMENT
Once the nature of a patient's illness has been established and its ex-
pected course predicted, the next question is, 'What can be done about it?
Is there a treatment that improves the outcome of disease? This chapter
describes the evidence used to decide whether a well-intentioned treatment
is effective.
Ideas and Evidence

The discovery of new treatments requires both rich sources of promising
possibilities and ways of establishing that the treatments arc in fact useful.
IDEAS
Ideas (hypotheses) about what might be useful treatment arise from
virtually any activity within medicine. Some therapeutic hypotheses are
suggested by the mechanisms of disease at the cellular or molecular level.
Drugs against antibiotic resistant bacteria are developed through knowl-
edge of the mechanism of resistance. Hormone analogues are based on the
structure of native hormones. The effectiveness of afterload reduction in
congestive heart failure was suggested by studies of the importance of
afterload in the pathophysiology of heart failure.
Other hypotheses about treatments have come from astute observations
by clinicians. Two examples are the discovery that patients with Parkin-
son's disease who arc given amantadine to prevent influenza show im-
provement in their neurologic status and the reports that colchicine, given
for gout, reduces the frequency of attacks of familial Mediterranean fever.
The value of these treatments was not predicted by an understanding of
the mechanism of these diseases, and the ways in which these drugs work
arc not yet understood. Similarly, folk remedies from throughout the
world, bolstered by centuries of experience but few scientific studies, are
potentially useful treatments.
Other ideas come from trial and error. Some anticancer drugs have been
found by methodically screening huge numbers of substances for activity.
136
CHAPTER 7 I TREATMENT 137
Ideas about treatment, but more often prevention, also come from epide-
miologic studies of populations. Burkitt observed that colonic diseases are
less frequent in African countries, where diet is high in fiber, than in
developed countries, where intake of dietary fiber is low. This observation
has led to efforts to prevent bowel diseases-c-imtable bowel syndrome,
diverticulitis, appendicitis, and colorectal cancer-with high-fiber diets.
Comparisons across countries have also suggested the value of red wme
to prevent heart disease and fluoride to prevent dental caries.
TESTING IDEAS
Some treatment effects are so prompt and powerful that their value is
self-evident even without formal testing. Clinicians do not have reserva-
tions about the value of penicillin for pneumonia, surgery for appendicitis,
or colchicine for gout. Clinical experience has been sufficient.
Usually, however, the effects of treatment are considerably less dra-
matic. Tt is then necessary to put ideas about treatments to a formal test,
through clinical research, because a variety of conditions-coincidence,
faulty comparisons, spontaneous changes in the course of disease, wishful
thinking-can obscure the true relationship between treatment and effect.
Sometimes knowledge of mechanisms of disease, based on work with
laboratory models or phvsiolugic studies in humans, has become so exten-
sive that it is tempting to predict effects in humans without formal testing.
However, relying solely on our current understanding of mechanisms,
without testing ideas on intact humans, can lead to unpleasant surprises
because the mechanisms are only partly understood.
Example Many strokes are caused by cerebra! infarction in the area distal
to an obstructed segment of the internal carotid artery. It should be possible
to prevent the rnarnfestatinns of disease in people with these lesions by
bypassing the diseased segment so that blood can flow to the threatened area
normally. It is technirallv feasible to enastamosc the superficial temporal
artery to the internal carotid distal to an obstruction. Because its value seemed
self-evident on physiologic grounds and because of the documented success
of an analogous procedure, coronary arterv bypass, the surgery became
widely used.
The EC/IC Bypass Study Croup {l] conducted a randomized controlled
trial of temporal artery bypass surgery. Patients with cerebral ischemia and
an obstructed internal carotid artery were randomly allocated to surgical
versus medical treatment. The operation was a technical success; 96°;', of
anastomoses were patent just after surgery_ Yet, the surgery did not help the
patients. Mortality and stroke rates after 5 years were nearly identical in
the surgically and medically treated patients, but deaths occurred earlier in
the surgically treated patients.
This study illustrates how treatments that make good sense, based on
what we know about the mechanisms of disease, may be found ineffective
in human terms when put to a rigorous test. Of course, it is not always the
case that ideas arc debunked; the value of carotid endarterertomv, suggested
on similar grounds, has been confirmed (2).
Therefore, it is almost always necessary to test therapeutic hypotheses

by means of clinical research, in which datiJ are collected on the clinical
course of treated and untreated patients. As one author (3) put it, treat-
ments should be given "not because they ought to work, but because they
do work."
Studies of Treatment Effects
Treatment is usually considered to be what physicians prescribe for
patients with established disease: surgery, drugs, diet and exercise. But
there are a great many other ways of intervening to improve health. Among
these are efforts to prevent disease in individual patients (counseling and
early detection with treatment, discussed in Chapter 8), intervention on
communities and changes in the organization and financing of health care.
Regardless of the nature of a well-intentioned intervention, the principles
by which it is judged superior to its alternatives are the same.
There are two general ways to establish the effects of treatment: observa-
tional and experimental studies. They differ in their scientific strength and
feasibility.
Observational studies of treatment are a special case of studies of prog-
nosis in general, where the prognostic factor of interest is a therapeutic
intervention. What has been said about cohort studies (Chapters 5 and 6)
applies to observational studies of treatment as well. The main advantage
of these studies is that they are feasible. The main drawback is the likeli-
hood that there are systematic differences in treatment groups other than
the treatment itself, which lead to misleading conclusions about the effects
of treatment.
Clinical trials arc a special kind of cohort study in which the conditions of
study-selection of treatment groups, nature of interventions, management
during follow-up, and measurement of outcomes-are specified by the inves-
tigator for the purpose of making unbiased comparisons. Clinical trials arc
more highly controlled and managed than arc cohort studies. The investiga-
tors are conducting an experiment, analogous to those done in the laboratory.
They have taken it upon themselves (with their patients' permission) to isolate
for study the unique contribution of one factor by holding constant, as much
as possible, all other determinants of the outcome. Hence, other names for
clinical trials are experimental and intervention studies.
Randomized controlled trials are the standard of excellence for scientific
studies of the effects of treatment. We will consider them in detail first,
then consider alternative ways of answering the same question.
Randomized Controlled Trials
The structure of a clinical trial is shown in Figure 7.1. The patients to
be studied arc first selected from a larger number of patients with the
condition of interest. They are then divided, using randomization, into two
CHAPTER 7 ! TREATMENT 139
Outcomes
Experimental
intervention ,liiipr{>,," I
l ~.lITIpr<lIlOO1
I""" """~,
f'liflpriMill:.TI
-Compariso';-
Intervention
-c;~vOOl
. ,
Figure 7.1. The structure of a clinical trial.
groups of comparable prognosis. One group, called the experimental or

treated group, is exposed to an intervention that is believed to be helpful.
The other group, called a control or comparison group, is treated the same
in all ways except that its members are not exposed to the intervention.
The clinical course of both groups is then observed and any differences in
outcome are attributed to the intervention.
The main reason for structuring clinical trials in this way is to avoid
bias (systematic error) when comparing the respective value of the two or
more kinds of treatments. The validity of clinical trials depends on how
well they result in an equal distribution of all determinants of prognosis,
other than the one being tested, in treated and control patients.
In the following discussion, we will describe the design and interpreta-
tion of clinical trials in detail, with reference to Figure 7.2.
SAMPLING
The kinds of patients that are included in a trial determine the extent
to which conclusions can be generalized to other patients. Of the many
reasons why patients with the condition of interest may not be part of a
trial, three account for most of the losses: They do not meet specific entry
criteria, they refuse to participate, or they do not cooperate with the con-
duct of the trial.
The first, entry criteria, is intended to restrict the heterogeneity of pa-
tients in the trial. Common exclusion criteria are atypical disease, the pres-
ence of other diseases, an unusually poor prognosis (which may cause
patients to drop out of the assigned treatment group), and evidence of
unreliability. Patients with contra indications to one of the treatments arc
also excluded, for obvious reasons. As heterogeneity is restricted in this
way, the internal validity of the study is improved; there is less opportuni ty
for differences in outcome that are not related to treatment itself. Also,
generalizing the results is more precise because one knows exactly to whom
_-~lr_:T",~-~.~~ L
]---"---> OUTCOMES
Allocation
))
..
... - - - - - - - - - - - - - - - -
Generalizing ......
-~
Statistical
analysis
Figure 7.2. Bias in clinical trials.
the results apply. But exclusions come at the price of diminished scope of
generaltzability, because characteristics that exclude patients occur com-
monly among those ordinarily seen in clinical practice, limiting generaliz-
ability to these patients, the very ones for whom the information is needed.
Second, patients can refuse to participate in a trial. They may not want
a particular type of treatment or to have their medical care decided by a
flip of a coin or by someone other than their own physician. Patients who
refuse to participate are usually systematically different-in socioeconomic
class, severity of disease, other health-related problems, and other ways-
from those who agree to enter the trial.
Third, patients who are found to be unreliable during the early stages
of the trial are excluded. This avoids wasted effort and the reduction in
internal validity that would occur if patients moved in and out of treatment
groups or out of the trial altogether.
for these reasons, patients in clinical trials are usually a highly selected,
biased sample of all patients with the condition of interest (Fig. 7.3). Be-
cause of the high degree of selection in trials, it often requires considerable
faith to generalize the results of clinical trials to ordinary practice settings.
INTERVENTION
The intervention itself can be described in relation to three general char-
acteristics: gcneralizebllity, complexity, and strength.
First, Is the intervention in question one that is likely to be implemented
Percent of Patients
.IiIIIiIiIiIiIi Population Sampled

100 '1JI1J11J11J11J11J11J11PlIJln' Patients with noninsulin-
dependent diabetes mellitus in one
hospital
Inclusion Criteria
Age >40 years
Diabetes diagnosed after 30 years old
Require medication for hyperglycemia
Plan to remain in practice >2 years
Other illness, disability, etc.
24 Eligible
Uncooperative
Refused to participate
Did not keep appointments
13 Randomized
Dropped Out
Death
Change of residence
Illness, etc.
12 Completed Study
Figure 7.3. Sampling for a clinical trial. A study of the effectiveness of a program
to reduce lower extremity problems in patients with diabetes. (Data from Litzelman
OK, Slemenda CW, Langfeld CD, Hays LM, Welch MA, Bild DE, Ford ES, Vinicor F.
Reduction in lower extremity clinical abnormalities in patients with non-insulin depen-
dent diabetes mellitus. A randomized controlled trial. Ann Intern Med 1993;119:
36-41.)
in usual clinical practice? In an effort to standardize therapy so it can be

easily described and reproduced in other setting:" some investigators end
up studying treatments that are so unlike those in usual practice that the
results of the trial are not usefu L
Second, single, highly specific interventions make for tidy science, be-
cause they can be described precisely and applied in a reproducible way.
However, clinicians regularly make choices among alternative treatments
that involve many elements. Multifaceted interventions arc amenable to
careful evaluation as long as their essence can be communicated and repro-
duced in other settings.
Example Falls are a major problem in the elderly, have a variety of
causes, and tend to recur. Rubenstein et al. (4) studied the effects of a program
to prevent falls in the elderly. Elderly people in a long-term residential care
facility were randomized after a fall to a special progrilm or to usual care.
The program included a detailed examination, laboratory tests, and environ-
mental assessment; therapeutic recommendations were given to the patient's
primary physician. Over the next 2 years, the intervention group had fewer
falls, 26% fewer hospitalizations, and a 52% reduction in hospital days corn-
pared with controls.
Third, Is the intervention in question sufficiently different from alterna-
tive managements that it is reasonable to expect that outcome will be
affected? Some diseases can be reversed by treating a single, dominant
cause, e.g., treating hyperthyroidism with radioisotope ablation or surgery.
But most diseases arise from a combination of factors acting in concert.
Interventions that change only one of them, and only a small amount,
cannot be expected to show strong treatment effects. If the conclusion of
a trial evaluating such interventions is that a new treatment is not effective,
it should come as no surprise.
COMPARISON GROUPS
The value of a treatment can only be judged by comparing the results
of the treatment to those of some alternative course of action. The question
is not whether a point of comparison is used, but how appropriate it is.
Results among patients receiving an experimental treatment can be mea-
sured against one or more of several kinds of comparison groups.
No Intervention
Do patients receiving the experimental treatment end up better than
those receiving nothing at all? Comparing treatment with no treatment
measures the total effects of health care, both specific and nonspecific.
Observation
Do treated patients do better than other patients who are simply ob-
served? A great deal of special attention is directed toward patients 111
clinical trials, and they are well aware of it. People have a tendency to
change their behavior because they are the target of special interest and
attention in a study, regardless of the specific nature of the intervention
they might be receiving, a phenomenon called the Hawthorne effect. The
reasons for this changed behavior are not clear. Patients are anxious to
please their doctors and make them feel successful. Also, patients who
volunteer for trials want to do their part to see that "good" results are
obtained. Thus comparison of treatment with simple observation measures
treatment effect over and above the Hawthorne effect.
Placebo Treatment
Do treated patients do better than similar patients given a placebo, an
intervention that is intended to be indistinguishable from the active treat-
ment-e-in physical appearance, color, taste, and smell-s-but does not have
a specific, known mechanism of action? Sugar pills and saline injections
are examples of placebos. It has been shown that placebos, given with
conviction, relieve severe, unpleasant symptoms, such as postoperative
pain, nausea, or itching, of about one-third of patients, a phenomenon
called the placebo effect.
Example Patients with chronic severe itching were entered in a trial of

antipruritic drugs. During each of ::; weeks, 46 patients received in random
order either cyproheptadine Hel, trimeprazine tartrate, or placebo. There
was a I-week rest period, randomly introduced into the sequence, in which
no pills were given. Results were assessed without knowledge of medication
and expressed as "itching scores"; the higher the score, the worse the itching.
Itching scores for the various treatments were cyproheptadine Hel, 28; trime-
praztne tartrate, 35; placebo, 30; and no treatment, 50. The two active drugs
and placebo were all similarly effective and all gave much better results than
no treatment (5).
Placebo effects have different meaning for researchers and clinicians.

Researchers are more likely to be interested in establishing specific ef-
fects-ones that are consistent with current theories about the causes of
disease. They consider the placebo effect the baseline against which to
measure specific effects. Clinicians, on the other hand, should welcome
the placebo effect and attempt to maximize it or any other way of helping
patients.
Many clinical interventions have both specific and nonspecific effects
(Fig. 7.4). What is important to clinicians and their patients is the total
effect of the intervention beyond what would have otherwise occurred in
the course of disease without treatment. However, it is also useful to know
what part of the total effect is specific and what is nonspecific so as to
avoid dangerous, uncomfortable. or costly interventions when relatively
little of their effect can be attributed to their specific actions.
Action Examples
Mostly Antimicrobials
_i
specific Antimetabolites
Narcotics
Mixed Placebo
Antidepressants
Mostly Antipruritics
non- •
specific L- _ Injection of
trigger points
Figure 7.4. The effects of most drugs are partly attributable to the placebo effect.
(Redrawn from Fletcher RH. The clinical importance of placebo effects. Fam Med
Rev 1983; 1:40-48.)
Effects
i- Specific
..
I:
E
treatment
Placebo
>
0
~
a. Hawthorne
E
Natural
I history
Figure 7.5. Total effects of treatment are the sum of spontaneous improvement,
nonspecific responses, and the errocts of specific treatments.
Usual Treatment
Do patients given the experimental treatment do better than those re-
ceiving usual treatment? This is the only meaningful (and ethical) question
if the usual treatment is already known to be efficacious.
The cumulative effects of these various reasons for improvement in
treated patients are diagrammed in Figure 7.5.
CHAPTER I ! TREATMENT 145
ALLOCATING TREATMENT
To study the effects of a clinical intervention free of other effects, the best
way to allocate patients to treatment groups is by means of randomization.
Patients are given either the experimental or the control treatment by one
of a variety of disciplined procedures-analogous to flipping a coin-
whereby each patient has an equal (or at least known) chance of being
assigned to anyone of the treatment groups.
Random allocation of patients is preferable to other methods of alloca-
tion, because randomization assigns patients to groups without bias. That
is, patients in one group are, on the average, as likely to possess a given
characteristic as patients in another. Only with randomization is this so
for all factors related to prognosis, whether or not they are known before
the study takes place.
In the long run, with a large number of patients in the trial, randomization
usually works as described. above. However, random allocation does not
guarantee that the groups will be similar. Although the process of random
allocation is unbiased. the results may not be. Dissimilarities between groups
can arise by chance alone, particularly when the number of patients random-
ized is small. To assess whether this kind of "bad luck" has occurred, authors
of randomized controlled trials often present a table comparing the frequency
of a variety of characteristics in the treated and control groups, especially
those known to be related to outcome. Tt is reassuring to see that important
characteristics have, in fact, fallen out nearly equally in the groups being
compared. If they have not, it is possible to see what the differences are and
attempt to control them in the analysis (see Chapter 6).
Some investigators believe it is best to make sure, before randomization,
that at least some of the characteristics known to be strongly associated
with outcome appear equally in treated and control groups, to reduce the
risk of bad luck. They suggest that patients first be gathered into groups
(strata) of similar prognosis and then randomized separately within each
stratum-a process called stratified ronaomiuuion. The groups arc then
bound to be comparable, at least for the characteristics that were used to
create the strata. Others argue that whatever differences arise by bad luck
arc unlikely to be large and can be dealt with mathematically after the
data are collected.
DIFFERENCES ARISING AFTER RANDOMIZATION
Not all patients in a clinical trial participate as originally planned. Some
are found not to have the disease they were thought to have when they
entered the trial. Others drop out, do not take their medications, are taken
out of the study because of side effects or other illnesses, or somehow
obtain the other study treatment or treatments that are not part of the
study at all. The result is comparison of treatment groups that might have
been comparable just after randomization but have become less so by the
time outcomes are counted.
Patients Do Not Have the Disease
Tt may be necessary to decide which treatment to give (in a clinical trial
or in practice) before it is certain the patient actually has the disease for
which the treatment is designed.
Example To study whether a monoclonal antibody against endotoxin im-
proves survival from sepsis, 543 patients with sepsis and suspected Gram-
negative infection were randomized to receive antiendotoxin or placebo (6). In
the subgroup of patients who actually had Gram-negative bacteremia, death
rate was reduced from 49 to 3ln:" a large difference that was well beyond what
could be accounted for by chance. However. only 200 patients (37'1<,) had Crem-
negative bacteremia, confirmed by blood culture. There is no known reason
why the other 63% would be helped by the drug. For all patients with sepsis
(some of whom had bacteremia and others did not) mortality rate was 43)}~ in
the placebo group and 39'% in the group receiving antiendotoxin, a small differ-
ence that was not beyond that expected by chance alone.
Thus, from this trial, there was evidence that the drug was effective against
Cram-negative bacteremia, but not for sepsis. Both are important questions:
the former for researchers, who are interested in the biologic effect of antien-
dotoxin in bacteremia, and the latter for clinicians, who needed to know the
clinical effects of their decision to give the drug to patients with scpsis->-a
decision that musl be made before it is known whether or not bacteremia is
actually present.
When patients suspected of having the specific disease in question later

turn out not to have it, there is a price to pay. Studying additional patients
who could not benefit from the specific action of the treatment decreases
the efficiency of the trial; more patients must be studied to see the effect.
Looked at another way, because patients experiencing the specific effect
are mixed with others who cannot, the effect size is reduced relative to a
trial including only patients with the disease. This decreases the chances,
for a given number of patients in the trial, that an effect will be found (see
Chapter 9). However, this kind of trial has the important advantage of
providing information on the consequences of a decision as the clinician
encounters it (see "Management and Explanatory Trials," later in this
chapter).
Compliance
Compliance is the extent to which patients follow medical advice. Some
prefer the term adherence, because it connotes a less subservient relationship
between patient and doctor. Compliance is another characteristic of pa-
tients that can arise after randomization. Although noncompliance sug-
gests a kind of willful neglect of good advice, in medicine other factors
also contribute. Patients may misunderstand which drugs and doses are
intended, run out of prescription medications, confuse various prepara-

tions of the same drug, or have no money or insurance to pay for drugs.
Taken together, these may limit the usefulness of treatments that have
been shown to work under specially favorable conditions.
Compliance is particularly important in medical care outside the hospi-
ta 1. In the hospital, many factors act to constrain patients' personal behavior
and render them compliant. Hospitalized patients are generally sicker and
more frightened. They arc in strange surroundings, dependent on the skill
and attention of the staff for everything-even their life. What is more,
doctors, nurses, and pharmacists have developed a well-organized system
for ensuring that patients receive what is ordered for them. As a result,
clinical experience and a medical literature developed on the wards may
underestimate the importance of compliance outside the hospital, where
most patients and doctors are and where doing what clinicians advise is
more difficult.
Comparing responses among compliant and noncompliant patients in
a randomized trial can be misleading.
Example During a large study of the effects oi several lipid-lowering
drugs on coronary heart disease, 1103 men were given clofibrate and 2789
men were g-iven placebo. The fi-year mortality rate was 20.()')';, for the clofi-
brate group and 20.9'>;) for the placebo group, indicating that the drug was
not effective.
It was recugnized that not all patients took their medications. Was clofi-
brate effective among patients who actually took the drug? The answer ap-
peared to be yes. Among patients given clofibrate, Scycar mortality for pa-
tients taking most oi their prescribed dnl';; was 15.0%,compared with 24.m;,
for the less cooperative patients (p < 10 '}. However, taking the prescribed
drug was also related to lower mortality rates among patients prescribed
placebo. For them, 5-year mortality was 15.1% ior patients taking most of
their placebo medication and 28.3 ior patients who did not (I' < 10- 15) . Thus
there was an association between dnJg taking and prognosis that was not
related to the active drug itself.
The authors (7) cautioned against evaluating treatment effects in sub-
groups determined by patient responses to the treatment protocol after
randomization.
Cointerventions
After randomization, patients may receive a variety of interventions
other than the ones being studied. If these occur unequally in the two
groups and affect outcomes, they can introduce systematic differences
(bias) between the groups compared.
Example The care of AIDS is emotional, in part because it affects young
adults and is universally fatal within a few years of the onset of symptoms.
Efforts to study the effectiveness of treatment have been hindered by disrup-
tion of the usual procedures of randomized trials, as patients try to maximize
their chances of survival. Patients in randomized trials sometimes exchange
the drugs being studied in the trial (researchers call the exchange of treatment
regimens among study participants "contamination") or obtain drugs that
are not part of the trial through"drug clubs." Information about this behavior
is usualiv not shared with the researchers and so cannot be accounted for in
the study. The result is to bias the study toward observing no effect, since
the contrast between the treatment of the "treated" group and the comparison
group is diminished.
Comparing Responders with Nonresponders
In some clinical trials, particularly those about cancer, the outcomes of
patients who initially improve after treatment (responders) are compared
with outcomes in those who do not (nonrcsponders). The implication is
that one Gill learn something about the efficacy of treatment in this way.
This approach is scientifically unsound and often misleading, because
response and non response might be associated with milny characteristics
related to the ultimate outcome: stage of disease, rate of progression, corn-
pliance. dose and side effects of drugs, and the presence of other diseases.
If no patient actually improved because of the treatment, and patients were
destined to follow various clinical courses for other reasons, then some
(the ones who happened to be doing well) would be called "responders"
and others (the ones having a bad course) would be considered "nonrc-
spenders." Responders would, of course, have a better outcome whether
or not they received the experimental treatment.
BLINDING
Participants in a trial may change their behavior in a systematic way
(i.e., be biased) if they are aware of which patients receive which treatment.
One way to minimize this effect is by blillding, an attempt to make the
various participants in a study unaware of which treatment patients have
been offered, so that the knowledge does not cause them to act differently,
thereby damaging: the internal validity of the study. "Masking" is a more
appropriate metaphor, but blinding is the time-honored term.
Blinding can take place at four levels in a clinical trial. First, those
responsible for allocating patients to treatment groups should not know
which treatment will be assigned next so that the knowledge does not
affect their willingness to enter patients in the trial or take them in the
order they arrived. Second, patients should be unaware of which treatment
they are taking: they arc thereby less likely to change their compliance or
their reporting of symptoms because of this information. Third, physicians
who take care of patients in the study should not know which treatment
each patient has been given; then they will not, perhaps unconsciously,
manage them differently. Finally, if the researchers who assess outcomes
cannot distinguish treatment groups, that knowledge cannot affect their
measurements.
The terms sil1x'dllind (patients] and double-blind (patients and research-
ers) are sometimes used, but their meaning is ambiguous. Tt is better simply
to describe what was done. A trial in which there is no attempt at blinding
is called open or open label.
When blinding is possible, mainly for studies of drug effects, it is usually
accomplished by means of a placebo. However, for many important clinical
questions-the effects of surgery, radiotherapy, diet, or the organization
of medical care-blinding of patients and managing physicians is not
possible.
Even when blinding appears to be possible, it is more often claimed
than successful. Physiologic effects, such as lowered pulse rate with beta-
blocking drugs, or bone marrow depression with cancer chemotherapy,
are regular features of some medications. Symptoms may also be a clue.
Example In the Lipids Research Clinics (B) trial of the prim<lry preven-
tion of cardiovascular disease, a nearly perfect placebo was used. Some Pl.'O-
pie received cholestyramine and others a powder of the same appearance,
odor, and taste. However, side effects were substantially more common in
the cholcstyramine group. At the end of the Tst year of the trial, there were
much higher rates in the experimental (chulestyrarnine] group than the con-
trol group for constipation (39 versus 10%), heartburn (27 versus 1m:.),
bdching and bloating (27 versus 16%), and nausea (16 versus 8'1.,). Patients
might have been prompted by new symptoms to guess which treatment they
were getting.
There is also objective evidence that patients and physicians in some

blinded trials can guess who received what treatment.
Example A double-blind, randomized trial was conducted to ~ee if pro-
pranolol could prevent another myocardial infarction in patients who had
already had one (9). At the conclusion of the trial, but before unblinding,
patients and clinic personnel were asked to guess the treatment group assign-
ment of each patient. For patients, 79.9% guessed propranolol correctly and
';7.2% placebo correctly. Physicians and clinic personnel were similarly accu-
rate. Clinical personnel seemed to he aided in their guessing by observation
of heart rate; it was unclear how patients knew.
ASSESSMENT OF OUTCOMES
When the outcome of a trial is measured in unequivocal terms, such as
being alive or dead, it is unlikely that patients will be misclassified. On
the other hand, when outcomes are decided by the opinion of one of
the participants, there is much greater opportunity for bias. For example,
although the fact of death is usually dear, the cause of death is often not.
Most people die for a combination of reasons or for obscure reasons,
allowing some room for judgment in assigning cause of death. This judg-
ment can be influenced by knowledge of what went before, including the
treatments that were given. Opportunities for bias are even greater when
assessing symptoms such as pain, nausea, or depression. Bias in assessing
150 CLINICAL EPIDFMIOLOGY
outcomes is avoided by searching for outcome events equally in all pa-

tients, using explicit criteria for when an outcome has occurred, and by
blinding.
Short-term, easily measurable "outcomes" may be substituted for clini-
cal ones so as to speed the rate at which trials can be completed and
reported. For example, it has been common in clinical trials of treatment
of HIV infection to take as the main outcome measures biologic tests that
reflect the extent of infection (such as ClJ4+ counts and p32 antigen) rather
than clinical progression of disease (opportunistic infections and death).
However, it has been shown that CD4+ counts are an imperfect marker
of clinical treatment effect. As discussed in Chapter 1, the practice of substi-
tuting biologic outcomes for clinical ones in studies that Me to guide patient
care is defensible only if the proxy is known to be itself strongly related
to the clinical outcome.
There are several options for summarizing the relative effects of two
treatments (Table 7.1). It has been suggested that the most clinically rele-
vant expression is number needed to treat, the number of patients that must
be treated to prevent one outcome event (10). Number needed to treat is
the reciprocal of absolute risk reduction.
Perception of the size of a treatment effect, both by patients and clini-
cians, is influenced by how the effect is reported. In general, effects reported
as relative risks seem larger than the same effects described as attributable
risks, which in turn seem larger than reports of the number needed to treat
(11,12). Also, patients told their probability of survival believe they have
a better chance than those told the complement, their probability of dying
(13). Thus, to understand and communicate treatment effects, it is neces-
sary to examine the main results of a trial in several ways. It is moot which
is the "correct" statistic.
Table 7.1
Summarizing Treatment Effects"
Summary Measure' DHfi"itiOrl
Control evcnt rllte- Treated event rate

Relative risk reduction
Control event rate
Absolute risk reduction Control event rate - Treated event rato
1
Number needed to treat
Control event rate I rcctoc event rate
-t.aooos A Sackelt DL llolJHrfs RS. An H~,*,~SrTlf1flt of clinically useful measures of the consequences of
treatment. New Engl J Med t988;3 t8 t 728 t 733
I, ,or continuous data, when there are measurements at bascmc and alter treatment, analogous rlIeilsurH~
are basHd Oil thH "'Ha" values for treated and control oroupr either aftof treatment or for the rjftcrcnco
between baseline and posttreatment "llluo,;
CHAPltR 7 ! TREATMENT 151
MANAGEMENT AND EXPLANATORY TRIALS

The results of a randomized controlled trial can oe analyzed and pre-
sented in two ways: according to the treatment to which the patients were
randomized or to the one they actually received. The correct presentation
of results depends on the question being asked.
If the question is which treatment policy is best at the time the decision
must be made, then analysis according to the assigned (randomized) group
is approprintc-c-whether or not some patients did not actually receive the
treatment they were supposed to receive. Trials analyzed in this way are
called intention to treat or management trials (14). The advantages of this
approach are that the question corresponds to the one actually faced by
clinicians and the patients compared are really randomized. The disadvan-
tage is that if many patients do not receive the treatment to which they
were randomized differences between experimental and control groups
will tend to be obscured, increasing the chances of a negative study. Then
if the study shows no difference, it is uncertain whether the experimental
treatment is truly ineffective or was just not received.
Another question is whether the experimental treatment itself is better?
For tills question, the proper analysis is according to the treatment each
patient actually received, regardless of the treatment to which they were
randomized. Trials analyzed in this way are called explanatory trials because
they emphasize the mechanism by which effects are exerted. The problem
with this approach is that unless most patients receive the treatment to
which they are assigned the study no longer represents a randomized
trial; it is simply a cohort study. Therefore, one must be concerned about
dissimilarities between groups, other than the experimental treatment, and
must use one or more methods (restriction, matching, stratification, or
adjustment) to achieve compnrcbilitv. just as one would for any nonexpert-
mental study. These two approaches are illustrated in Figure 7.6.
EFFICACY AND EFFECTIVENESS
A trial's results are judged in relation to two broad questions. Can
the treatment work under ideal circumstances? Does it work in ordinary
settings? The words efficacy and efj('cfivcl1e~s have been applied to these
concepts (Fig. 7.7).
The question of whether a treatment can work is one of Ifjicacy. An
efficacious treatment is one that has the desired effects among those who
receive it. Efficacy is established by restricting patients in a study to those
who will cooperate fully with medical advice.
In contrast, a treatment is Ilfective if it does more good than harm in those
to whom it is offered. Effectiveness is established by offering a treatment or
program to patients and allowing them to accept or reject it as they might
ordinarily do. Only a small proportion of clinical trials set out to answer
Intention to Treat Analysis
Drop out
,, ,
, ,, Analysis
according to
Cross over ,,
,,
treatment
,,
assigned
. . . . . . . .4,
--- - - - - ~
Drop out
Explanatory Analysis
Drop out
~/ Cross over
Analysis
according to
treatment
.--.0."'\ .,
received
\'rop out
Figure 7.6. Intention to treat and explanatory trials.
questions of effectiveness. This is in part because of the risk that the result
will be inconclusive. If a treatment is found to be ineffective, it could be
because of a lack of efficacy, lack of patient acceptance, or both.
Tailoring the Results of Trials to Individual Patients
Clinical trials involve pooling the experience of many patients who arc
admittedly dissimilar and describing what happens to them on the average.
How can we obtain more precise estimates for individual patients? Two
ways are to examine subgroups and to study individual patients using
rigorous methods similar to those in randomized trials.
SUBGROUPS
The principal result of a clinical trial is a description of the most im-
portant outcome in each of the major treatment groups. But it is tempting
to examine the results in more detail than the overall conclusions afford.
CHAPTFR 7 I TREATMENT 153
Does receiving treatment

+ work under ideal conditions?
Efficacy
Trial
1
Internal
Validity Noncompliance
Less selected patients
Less selected clinicians
Generalizability + Costs
Impracticality
Other
Does offering treatment help

under ordinary
+ Effective-
circumstances?
nessTrial
Figure 7.7. Efficacy and effectiveness.
We look at subgroups of patients with special characteristics or with partic-

ular outcomes. In doing so, however, there are some risks of being misled
that are not present when examining the principal conclusions alone, and
these should be taken into account when interpreting information from
subgroups. (Some of the concepts on which this section is based are dis-
cussed in Chapter 9.)
One danger in examining subgroups is the increased chance of finding
effects in a particular subgroup that are not present, ill the long run, in
nature. This arises because multiple comparisons lead to a greater chance
of a false-positive finding than is estimated by the individual p value for
that comparison alone (see Chapter 9). Table 7.2 lists some guidelines for
deciding whether a finding in a subgroup is real.
A second danger is of a false-negative conclusion. Examining subgroups
in a clinical trial-either certain kinds of patients or specific kinds of out-
comes-involves a great reduction in the data available, so it is frequently
impossible to come to firm conclusions. Nevertheless, the temptation to
look is there, and some tentative information can be gleaned.
Example The Physicians' Health Study (15) is a randomized controlled
trial designed to assess whether daily aspirin prevents mortality from cardlo-
vascular disease in healthy male physicians. Another aspect of the trial is to
studv the effect of l3-carotene on the incidence of cancer. The aspirin part of
the study was stopped kmg before there were enough deaths to determine
if aspirin affected mortality, because the physicians had a much lower than
expected death rate. The trial was also stopped because there were fewer
myocardial infarctions in the treated than the control group. The authors
thought that the effect on myocardial infarction, although not the answer to
a main study question at the outset, was real because it was biologically
plausible, because it was found in other studies, and because the chance of
a false-positive conclusion was estimated to be very small (1/10,000). On the
other hand, although the authors observed a small increase in risk of stroke
in the treated group, they could not be certain whether this effect was real
or not, as there were too few physicians with this end point. Thus, in a study
that could not address the main research question, the authors interpreted
the validity of findings in subgroups (both positive and negative) in relation
to the totality of information that might bear on the validity of these findings.
EFFECTIVENESS IN INDIVIDUAL PATIENTS
A treatment that is effective on the average may not work on an individ-
ual patient. The results of valid clinical research provide a good reason to
begin treating a patient, but experience with that patient is a better reason
to continue therapy. Therefore, when conducting a treatment program it
is useful to ask the following series of questions:
• Is the treatment known to be efficacious for any patients?
• Is the treatment known to be effective, on the average, in patients like
mine?
• Are the benefits worth the discomforts and risks?
• Is the treatment working in my patient?
By asking these questions, and not simply following the results of trials
alone, one can guard against ill-founded choice of treatment or stubborn
persistence in the face of poor results.
Table 7.2
Guidelines for Deciding Whether Apparent Differences in Effects
within Subgroups Are Real"
From the stuuv itself

• b the magnitude of the observed difference clinically important?
• How likely is the effect to have arisen by chance, taking into account
The number of subgroups examined?
lhe rTldgnitucJe ot the p value?
• Was a hvpothesis that the eltect would be observed
Made before its discovery (or was justification for the effect argued for after it was
found)?
One of a smCll1 number of hypotheses')
From other intormaticn
• Was the difference suooesrec by comparisons within rather than between studies?
• Has the effect been observed in other studies?
• Is there indirect evidence that supports the existence of the effect?
. , Adapted from Oxrndo Arl. Guyatt GIL A corwurners glJidp. to subgrolJp arlHlysis A"" I"t"", M"d
1992: 116:78 8~_
TRIALS OF N = 1
Rigorous clinical trials, with proper attention to bias and chance, can
be done with individual patients, one at a time (16). The method-called
trials of N = 1-is an improvement in the more informal process of trial
and error that is so common in clinical practice. A patient is given one or
another treatment (e.g.. active treatment or placebo) in random order, each
for a brief period of time, such as a week or two. Patients and physicians
are blind to which treatment is given. Outcomes (c.g.. a simple preference
for a treatment or a symptom score) are assessed after each period and
subjected to statistical analysis.
This method is useful when activity of disease is unpredictable, response
to treatment is prompt, and there is no carryover effect from period to
period. Examples of diseases for which the method can be used include
migraine, bronchospasm, fibrositis, and functional bowel disease.
N of 1 trials can be useful for guiding clinical decision making, although
for a relatively small proportion of patients. It can also be used to screen
interesting clinical hypotheses to select some that are promising enough
to be evaluated using a full randomized controlled trial involving many
patients.
Alternatives to Randomized Trials

Randomized, controlled, blinded trials are the standard of excellence
for comparisons of treatment effects over time. They should be given prece-
dence over other information about treatment effects whenever they are
available. However, it is not always possible to rely on clinical trials.
LIMITATIONS OF RANDOMIZED TRIALS
Clinical trials are limited for several reasons. There may not be enough
patients with the disease of interest, at one time and place, to carry out a
scientifically sound trial. Clinical trials are costly, more than $50-100 mil-
lion for some large trials. Years pass before results are available, which
may be politicalIy unacceptable for severe, emotion-laden diseases such as
AIDS.
Sometimes a practice may have become so well established, in the ab-
sence of conclusive evidence of its benefit, that it is difficult to convince
physicians and potential participants that a trial is needed. It could be
argued that if the treatment effect is not really known then the only ethical
thing is to do the study (and it is unethical to continue to use treatments
of uncertain benefit), but this argument demands a level of analytic reason-
ing that is uncommon among patients and their physicians. Because of this
problem, some physicians have advocated "randomization from the first
patient," beginning trials just after a new treatment is introduced. Others
argue that it is better to conduct rigorous clinical trials somewhat later,
after the best way to deliver the treatment has been worked out, so that a
good example of the intervention is tested. In any case, it is generally
agreed that if a controlled trial is postponed too long, the opportunity to
do it all may be lost.
For these reasons, guidance from clinical trials is not available for many
treatment decisions. But the decisions must be made nonetheless. What
are the alternatives and how credible are they?
ADVANTAGES AND DISADVANTAGES
Alternatives to randomized trials usually make use of large databases

such as those collected for patient care, billing, or ad ministration. Sometime
data collected to answer another research question are used. A research
question, and a study to answer it, can be devised after the data have been
collected so that most of the resources needed for the study go into analyses
of the data. This process is called secondary data anI/lysis, because answering
the research question was not the primary reason for collecting the data.
Using secondary data for research has several advantages, all of them
practical. First, if the database includes experience from a large number of
patients, as is often the case, then the research question can be answered
with a high degree of confidence that the results are not just by chance. It
may even be possible to examine subgroups (e.g., elderly women taking
estrogens or young men with a first anterior myocardial infarction) with
statistical confidence. Most clinical trials are not designed with such an
abundance of patients because of the cost; the best trials are sufficient to
answer the main research question for aU patients in the study but are
rarely sufficient to answer questions about subgroups of patients.
Second, these databases are collected in more natural settings than clini-
cal trials. They reflect experience in health care organizations or perhaps
entire regions or nations, rather than a highly selected group of experimen-
tal subjects. Therefore, the results are more generalizable.
Third, it costs less to use existing data than to collect new data in a clinical
trial. Randomized trials cost thousands of dollars per patient to recruit, evalu-
ate, enroll, and follow up each patient, whereas analyses of existing data can
be relatively inexpensive.
Finally, by using existing data, it is possible to have an answer to an
important question in a relatively short time. Clinical trials often take years
from enrollment of the first patient to the end of follow-up. Sometimes
clinicians need an answer, however imperfect, sooner because they are
making high-stakes decisions on alternative treatments every day.
Balanced against all of these practical advantages are disadvantages.
The data are usually not collected and classified with as much care as they
would be for a well-run clinical trial. For example, a claims data diagnosis
of "hypertension" stands for whatever the responsible physicians believes
hypertension is, whereas a research definition of hypertension would spec-

ify a level of blood pressure, method, and frequency, perhaps adjusted for
age. Some important variables may be missing from the database because
they were not important for the database's original purposes, though they
arc important for the research. Of course, there is also the problem of
making unbiased comparisons.
The trade-off between speed and ease, on the one hand, and validity,
on the other, will be discussed for each alternative design in the next
sections.
COMPARISONS ACROSS TIME AND PLACE
Control patients can be chosen from a time and place different from the
experimental patients. For example, we may compare the prognosis of
recent patients treated with current medications to experience with past
patients who were treated when current medications were not available.
Similarly, we may compare the results of surgery in one hospital to results
in another, where a different procedure is used. This approach is conve-
nient. The problem is that time and place are almost always strongly related
to prognosis. Clinical trials that attempt to make fair comparisons between
groups of patients arising in different eras, or in different settings, have a
particularly difficult task.
The results of current treatment are sometimes compared with experi-
ence with similar patients in the past, called historical or nonconcurrent
controte. Although it may be done well, this design has many pitfalls.
Methods of diagnosis change with time, and with them the average prog-
nosis. It has been shown that new diagnostic technologies have created
the impression that the prognosis of treated lung cancer have improved
over time when it has not (17). With better ability to detect occult metasta-
ses, patients are classified in a worse stage than they would have been
earlier, and this "stage migration" has resulted in a better prognosis in
each stage than was reported in the past. Supporting treatments (e.g.,
antibiotics, nutritional supplementation and peptic ulcer prevention) also
improve with time, creating a general improvement in prognosis that might
not be attributable to the specific treatment given in a later time period.
Example Sacks et al. (18) reviewed d;nical trials of six therapies 10 sec
if trials with concurrent controls produced different results than studies of
the same treatments with historical controls. They studied 50 randomized
trials and 56 studies with historical controls. A total of 7';l'}:, of trials with
historical controls but onlv 20'%, of trials with a concurrent, randomized con-
trol group found the experimental treatment to be better. Differences between
the two kinds of trials occurred mainly because the control patients in the
historical trials did worse. Adjustment for prognostic factors, when possible,
did not change the results, i.e., the differences were probably because of
genera! improvements in therapy or to selection of less ill patients.
Therefore, if concurrent, randomized controlled trials <Ire taken as a

standard of validity it seems that published historical trials are biased in
favor of the experimental treatment and that the bias cannot be overcome
by adjusting for known prognostic variables.
If historical controls arc used, the shorter the period of time between
selection of treated and control groups and the less other aspects of medical
care have changed during the interval, the safer the comparison. Thus
some oncology centers study a succession of chemotherapeutic regimens
by comparing results of the newest regimen with those of the immediately
preceding one, often given as recently as the previous year. In general,
however, choosing concurrent controls (i.e., patients treated during the same
period of time) is a better way of avoiding bias.
Experience in other settings, using different treatments, can serve as a
standard of comparison. However, it is preferable to choose both treated
and control patients from the same setting, because a variety of factors-
referral patterns, organization and skill of staff, etc.-often result in very
different prognoses in different settings, independently of the treatment
under study.
Example The mortality rate for hospitals where coronary bypass surgery
was done varied almost threefold across hospitals in central Pennsylvania
(Fig. 7.8) (19). The severity of illness, and therefore prognosis, of patients in
these hospitals varied too. After taking into account the number of deaths
expected, by considering patients' prognostic factors, one hospital had fewer
than expected deaths, another the expected number, and a third more than
expected. Any fair comparison of treatment effects across these hospitals
would have to take into account not only the differences in severity of the
patients in these hospitals but also the skills of the surgeons.
UNCONTROLLED TRIALS
Uncontrolled trials describe the course of disease in a single group of

patients who have been exposed to the intervention of interest. Another
name for this design is a "before-after study." The assumption of this
approach is that whatever improvement is observed after treatment is
because of treatment. This assumption may be unwarranted for several
reasons.
Unpredictable Outcome
When the clinical course of a disease is quite predictable, a separate
control group is less important. We know that subacute bacterial endocar-
ditis without antibiotics and rabies without vaccine invariably lead to
death, that most patients with hypothyroidism will only get worse without
exogenous thyroid hormone, and that bawd infarction will rarely improve
without surgery.
However, most therapeutic decisions do not involve conditions with
CHAPTeR 7 I TREATMENT 159
6.93
6.50
6.0 Expected
range
"c• 5.2 5.25
5.0
0
s:
;::
-••
<
4.0
..-• 3.0
2.48 Actual 1.~ffL.J
-•
0
e
2.0
number-
.•.
0
"
1.0
L---_~ _
1 2 3
Hospital
Figure 7.8. Severity of illness and skill of surgeons vary by location. Observed and
expected (taking into account case mix) death rate from coronary bypass surgery
in 1I1reB hospitals. (Data from Topol EJ, Califf RM. Scorecard cardiovascular medicine.
Its impact and future direction. Ann Intern Med 1994; 120:65- 70,)
such predictable outcomes. In situations where the clinical course is ex-

tremely variable for a given patient and from one patient to another, as-
sessing treatment effects by observing changes in the course of disease
after treatment is unreliable.
Many severe diseases that are not self-limited may nevertheless undergo
spontaneous remissions in activity that call be misinterpreted as treatment
effects. figure 7.Y shows the clinical course of a patient with systemic lupus
erythematosus over a 'l O-vcar period, 1955-1%4 (20). Although powerful
treatments were not given (because none was available during most of
these years), the disease passed through dramatic periods of exacerbation,
followed by prolonged remissions. Of course, exacerbations, such as those
illustrated, are alarming to both patients and doctors, so there is often a
feeling that something must be dune at these times. If treatment were
160 CLINICAL ePIDEMIOLOGY
4+
3+
-",
-> 2+
()
<I:
1+
0
1955 1960 1964
Time
Figure 7.9. The unpredictable course of disease, The natural history of systemic
lupus erythematosus in a patient observed before the advent of immunosuppressive
drugs, (Redrawn from Ropes M. Systemic lupus erythematosus. Cambridge, MA:
Harvard University Press, 1976.)
begun at the peak of activity, improvement would have followed. Without

any better comparison than the previous activity of the disease, the treat-
ment would have received credit for the improvement.
Nonspecific Effects
In uncontrolled trials, there is no way of separating Hawthorne and
placebo effects from treatment effects. But if there are control patients who
receive the same attention as the treated ones and a placebo, then these
effects cancel out in the comparison.
Regression to the Mean
Treatments are often tried because a manifestation of disease, e.g., a
particularly high blood pressure or fever, is extreme or unusual. In this
situation, subsequent measurements are likely to show improvement for
purely statistical reasons. As discussed in Chapter 2, patients selected be-
cause they represent an extreme high value in a distribution are likely, on
the average, to have lower values for later measurements. If those patients
are treated after first being found abnormal and the effects of treatment
are assessed by subsequent measurements, improvement could be ex-
pected even if treatment were ineffective.
Predictable Improvement
The usual course of some diseases is to improve; if so, therapeutic efforts
may coincide with improvement but not cause it. For example, patients
tend to seek care for many acute, self-limited diseases, such as upper
respiratory infections or gastroenteritis, when symptoms are at their worst.

They often begin to recover after seeing the doctor because of the natural
course of events regardless of what was done.
NONRANDOM ALLOCATION OF TREATMENT
One way to allocate patients to treated and control groups is to have
the physicians caring for the patients decide. When this is done, the study
has all the advantages and disadvantages of cohort studies.
Studies of treated cohorts take advantage of the fact that therapeutic
decisions must be made for sick patients regardless of the quality of ex-
isting evidence on the subject. In the absence of a clear-cut consensus
favoring one mode of treatment over others, various treatments are often
given. As a result, in the course of ordinary patient care large numbers of
patients receive various treatments and go on to manifest their effects. If
experience with these patients can be captured and properly analyzed, it
can be used to guide therapeutic decisions.
Unfortunately, it is often difficult to be sure that observational studies
of treatment involve unbiased comparisons. Decisions about treatment are
determined by a great many Iacrors-c-scvcnty of illness, concurrent dis-
eases, local preferences, patient cooperation, etc. Patients receiving the
different treatments are likely to differ not only in their treatment but in
other ways as well. Efforts to determine the results of treatment alone, free
from other factors, are thereby compromised.
Phases of Studies of Treatment

For studies of drugs, it is customary to define three phases of trials, in
the order they are undertaken (21). Phase [ trials arc intended to identify
a dose range that is well tolerated and safe (at least for high-frequency,
severe side effects) and include very small numbers of patients (perhaps
a dozen), without a control group. Phase II trials provide preliminary infor-
mation on whether the drug is efficacious and the relationship between
dose and efficacy; these trials may be controlled but include too few pa-
tients in treatment groups to detect any but the largest treatment effects.
They milY not be blinded. Phase III trials provide definitive evidence of
efficacy and the presence of common side effects. They include enough
pntients-c-dozcns to thousands-to detect clinically important treatment
effects and are commonly published in clinical journals and used by regula-
tory agencies to decide whether to license drugs.
Phase III trials are not large enough to detect differences in the rate-or
even the existence-of uncommon side effects. (See discussion of statistical
power, Chapter Y.) For this, it is necessary to follow up very large numbers
of patients after a drug is in general use, a process called "postmarkcting
surveillance" or, sometimes, phase IV of drug development.
Summary
Promising ideas about what might be good treatment should be put to
a rigorous test before being used as a basis for clinical decisions. The best
test is a randomized controlled trial, a special case of a cohort study in
which the intervention is allocated randomly and, therefore, without bias.
Patients in clinical trials are usually highly selected, reducing gcneralizabil-
ity. They are randomly allocated to receive either an experimental interven-
tion or some comparison management: usual treatment, a placebo, or sim-
ple observation. On the average, the compared groups have a similar
prognosis just after randomization (and before the interventions). How-
ever, differences not attributable to treatment can arise later, including not
taking the assigned treatment, dropping out of the study, receiving the
other treatment, being managed differently in other ways, or getting treat-
ments that are not part of the study. Blinding all participants in the trial
can help minimize bias in how patients are randomized, managed, and
assessed for outcomes but is not always possible or successful.
The results of randomized trials can be summarized according to the
treatment assigned; an intention-to-treat analysis, which is a test of the
clinical decision and maintains a randomized trial design; or according to
the treatment actually received, which bears on the biology of disease, but
not directly on the clinical decision, and has the disadvantage that patients
may not remain with the treatment they were originally assigned. To obtain
information more closely tailored to individual patients than the main
results of randomized trials afford, clinicians can use results in subgroups
of patients, which carry the additional risk of being misleading, or do trials
on their own patients, one at a time.
For many clinical questions it is not possible, or not practical, to rely
on a randomized controlled trial. Compromises with the ideal include
making comparisons to experience with past patients, to past experience
with the same patients, or to a concurrent group of patients who are not
randomly allocated. When these compromises arc done, the internal valid-
ity of the study is weakened.
REFIiRENCES
l. ECjlC Ilyp~~~ Study Group, Failun, "f ",xtr~cranial-intracranial arterial bypass to reduce
the risk of ischemic ~troke. N Engl J Mcd 19S5;313:1191-1200,
2. Mayberg MR, Wilson E, Y~t~u F, Weiss DG, Messina I., Hershey LA, Colling C bkri(lgt'
L Deykin D, Winn UK Carotid endarterectomy and prevention of cerebral ischemic' in
symptomatic carotid stenosis. JAMAI991 ;266:3289-3294.
:I, Opie On the heartllidnorial]. LancctI9S0;1:692.
4, Rubenstein LZ, R(lhhin~ AS, Josephson KI':, Schulm~n ilL, Dsterweil 0. l'hc valu", of
assessing falls in an elderly population. A randomized controlled trial. Ann Intern Med
1990;113:30t!-316.
S. Fisch"r RW Comparison of antipruritic drugs administered orally. JAMA 1968;2U3:4Hl-
419.
6. 7iegler Ej, et ~l. Treatment of gram neg"tive bacteremia and septic shock with HA-IA
human monoclonal antibody against endotoxin. New Engl J Med 1991,.124:429-43(,.
7, Coronary Drug Project Research Croup. Influence of adherence to treatment ~nd response
of cholesterol on mortality in th" coronary drug project. r-..: Engl I Med 19S0;303:1038-
1041-
R l.ipid Research Clinics Program. The Lipid Research Clinics coronary primary prevention
trial results. 1 Reduction in incidence of coronary heart disease, jAMAI9M;251:351-1M,
9. Bvlngton RP, ct al. A~s;,~~ment of doublc-blindne,s at the conclusion of the bda·blocker
heart "tbck trial. JAMA 1985;253:1733-1736.
10. l,aupacis A, S"ckelt DL, Roberb R.S. An ~~s;,~sment of clinically useful measures of the
«ooscqvences of treatment New Engl.l Med 1988;318:I72H-1733,
11. Naylor CD, Chen E, Str"u~s 1:1. Mea~ured enthusiasm: does the method of reporting trial
results alter perceptions of tht'f~peutic effectiveness? Ann Inkrn Med 1992;117:916-921
12, Malenka OJ, Hamn [A, joh"ns;,n S, Wahr"nberger jW, Ross .1M. The framing effect of
relative "nd absolute rbk ] Gcn Intern Med 1993;8:541-54H.
13. McNeil B], Paukcr SG, Sox I IC [r. Tvt'fsky A. On th" elicitation of preference~ for altcrna-
tive th"rapies.. New Engl.l Med 19H2;306:12.S9.
14. Sackett DL, Gent M. Controversy in counting ~nd attributing events in clinical trials. N
Fngl j Med 1979;301:1410-1412.
15 Steering Commilte<:e of tile Physicians' Ht'alth Study R,,~eMCh Group !'inal report of the
aspirin component of the ongoing Physicians' Hl'alth Study. New Engl .I Mcd
1989;321 :129-135.
16. Guyatt C, Sackett D, Taylor OW, Chong [. Roberts R, Pugsley S. Determining optimal
thsrapv-c-randomizcd trials in individual patients. N Engl.l Mcd 1986;314:RS9-S92.
17 Feinstein AR, Sosin DM, W"lls CK, The Will Rogers pht'nornenon, Stage migration ~l1d
new diagnostic techniques as a source of misleading statistics for survival in cancer. New
Engl.l Med 1985;312:1(,04-2608.
18. S<:t\K.s II, Chalmers TC, Smith II Jr. Randomized versus historical controls for clinical
trials, Am I Mcd 1982;72:211-240.
19. Topol Fj, C~liff RM. Scorecard cardiovascular medicine. lis impart and future direction.
Ann Intern Med 1994;120:65-70.
21l, Ropes M. Systemic lupus erythematosus. Cambridge, MA: H"vnrd University Press, 1976.
21 Spilker B. Guide to clinical interpretation of data, New York: Raven Pn'ss, 19!16.
SUGGESTl-iD READINGS
Chalmers TC, Smith II [r, Blackbum 1:1, Silverman 1:1, Schroeder 1:1, Reitman U, Arnbroz A.
A method for assessing tht, quality of a randomized control trial. Control Clin Tri~lsI9Hl;
2:31-49
lJ€partmeTlt of Clinical Epidemiology and Biostatistics, Mclvlaster University, H~milton, Ont.
How to read clinical journals. V: To distinguish useful from useless Dreven harmful thnllpy,
Can /l.1ed Assoc .I 1981,124:1l:i6-1162.
DerSimonian R, Charette Lj, McP"",k 1:1, Mosteller F. Reporting on methods ill clinical trial~.
N Engl J Med 19H2;30fi:1332-B37.
Feinstein AR An additional basic science for clinical medicine, J1: The limitations of r"ndom-
izcd trials. Ann Intern Mcd 19f\3;99:544-550.
Friedman LM, l'urberg CD, De Mets DL. Fundamentals of dinical trials, 2nd ed, Littleton,
MA: john Wright PSC, 1985.
Guyatt CII, Sackett DL, Cook D.l. How to read clinical journals. II: How tu use and article
about therapy or prevention. A: Are the results of the ~tudy valid? .lAMA 1993;270:259H-
2601.
Cuyatt CH, Sackett DI" Cook DJ. How to read clinical journals. II: I low to use ilnd article
about therapy or prevention, B: What were the n'~ult5 and will they help me in caring for
my patient~? JAMA 1994;271:59-63,
Cuyan G, Sackett 0, Taylor OW, Chong J, Roberts I{, PUI-,,,,Iey S. Determining optimal ther-
apy-randomized trials in indiviuual patients. N Engl J Mcd I'JH6;::l14889-892.
Hdlman 5, Hellman DS, Of mice anu men Problems of the randomized clinkal trial. New
Eng! J MedI991;324:1585-15H'J,
Laupakis A, Welb G, Richardson S, Tugwdl P Users' guide to the medicalliteratun:, V: How
to use an article about prognosis. 1994;272:234-2::17,
Lavon PW, Louis I'A, lJaiiar Je 1J1, Polansky M. Design for experim"nt~-parallel comparisons
of tn,atmenL N Engl J Med 19H::l;109:1291-1298
Meinert (:1.. Clinical trials: design, conduct and analysis. New York Oxford Univn~ity Press,
1986.
Mostelller F, Cllbertlt'. Mcf'cck B. I''''porting standards and research ~trategie~ for controlled
triab, Control CEn Trials I'JHO;L37-'i8.
Oxman AO, Cuyutt CH. A consumer's guide to ~ubgrOlJp analysis. AJUl Intern M",d
1992;116:7H-Il4.
Peto R, Pike MC, Armitage P, Bre~low NE, Cox DR, Howard SV, Mantel N, Mcl'herson K,
Pete J, Smith PC, Design and analysi8 of randomized clinical triab requiring prolonged
observation of each patient, part l. Br J C~n,,,,r, 1976;34:5lJ5-612,
Peto R, Pike Me, Armitage P, Breslow NE, Cox DR, Howard SV, Mantel N, MrPher~on K,
I'eto J, Smith Pc. Design and analysis of randomized clinical trials requiring prolong",d
observation of each patitnt, port 2. Br J Cancer, 1'J77;::l5:1-,~9.
Pocock 5J. Clinical trials: a practical approach, New York; John Wiky & Sons. 1983.
Yusuf S, Collin~ R, Peto R Why do we n,'"d ~ome large, simple randomized tria1s7 Stat Med
19H4;3:409-420.
8
PREVENTION
lmc sensibly-among Ii ttioueana people, only Dill' dies a natural
death, tile rest succumb to irmuouat modes of living.
Maimonides 1135-1204 A.I).
Most doctors are attracted to medicine because they look forward to

curing disease. But all things considered, most patients would prefer never
to contract a disease in the first place-Of, if they cannot avoid an illness,
they prefer that it be caught early and stamped out before it causes them
any harm. To accomplish this, procedures are performed on patients with-
out specific complaints, to identify and modify risk factors to avoid the
onset of disease or to find disease early in its course so that by intervening
patients can remain well. Such activity is referred to as health maintenance
or till' periodic health examination.
Health maintenance constitutes a large portion of clinical practice (1).
Often, health maintenance activities can be incorporated into the ongoing
care of patients, as when a doctor checks the blood pressure in a patient
complaining of a sore throat; sometimes, a special visit just for health
maintenance is scheduled.
Physicians should understand the conceptual basis and content of the
periodic health examination. They should be prepared to answer questions
from patients such <IS "Why do J have to get a Pap smear again this year,
Doctor?" or "My neighbor gets <I chest x-ray every year; why aren't you
ordering one for me?"
This chapter concentrates on prevention activities clinicians undertake
with individual patients. However, prevention at the community level is
also effective. Immunization requirements for students, no-smoking regu-
lations in public buildings, and legislation restricting the sale of firearms
are examples of community wide prevention. For some problems, such as
injury prevention from firearms, community prevention works best. For
others, such as colorectal cancer, screening in clinical settings works best.
For still others, clinical efforts can complement communltywidc activities,
<IS in smoking prevention efforts by which clinicians help individual pa-
165
tients stop smoking and public education, regulations, and taxes prevent
teenagers front starting to smoke.
Much of the scientific approach to prevention in clinical medicine, par-
ticularly the principles underlying the use of a diagnostic tests, disease
prognosis, and euectlvcncss of interventions, has already been covered in
this book. This chapter expands on those principles and strategies as they
specifically relate to prevention.
Levels of Prevention
Webster's (2) dictionary defines prevention as "the act of keeping from
happening." With this definition in mind, almost all activities in medicine
could be defined as prevention. After all, clinicians' efforts are aimed at
preventing the untimely occurrences of death, disease, disability, discom-
fort, dissatisfaction, and destitution (Chapter 1). However, in clinical medi-
cine, the definition of prevention is usually restricted, as outlined below.
Although more prevention is practiced than ever before, clinicians still
spend most of their time in diagnosing and trcating rather than in pre-
venting disease.
Depending on when in the course of disease interventions are made,
three types of prevention are possible (pig. 8.1).
PRIMARY PREVENTION
Primary preoention keeps disease from occurring at all, by removing its

causes. Folic acid administration to prevent ncura I tube defects, immuniza-
tions for many communicable diseases, and counseling patients to adopt
healthy lifestyles (e.g., helping patients to stop smoking, to eat foods low
in saturated fats and cholesterol and high in fiber, to exercise appropriately,
and to engage in safe sexual practices) are examples of primary prevention.
Clinical
Onset Diagnosis
~f
~
ASYMPTOMATIC
DISEASE
CLINICAL COURSE
PRIMARY SECONDARY TERTIARY

Remove risk Early detection Reduce
factors and treatment complications
Figure 8.1. Levels of prevention.

CHAPTER 8 / PREVENTION 167
Primary prevention is often accomplished outside the health care system

at the community level, as noted above. Chlorination and fluoridation of
the water supply and laws mandating seat belt use in automobiles and
helmets for motorcycle use <In' examples. Certain primary prevention activ-
ities occur in specific occupational settings (use of ear plugs or dust masks),
in schools (immunizations), or in specialized health care settings (usc of
tests to detect the hepatitis B or HIV in blood donations in blood banks).
SECONDARY PREVENTION
Secondary prevention detects disease early when it is asymptomatic and
when early treatment can stop it from progressing; Pap smears, mammo-
grams, and fecal occult blood tests are examples. Most secondary preven-
tion is done in clinical settings, and all physicians, especially those caring
for adults, undertake secondary prevention. There <Ire a few community-
wide programs (shopping mall fairs for glaucoma screening are an
example).
SCREENING
Screening is the identification of <In unrecognized disease or risk factor
by history taking (c.g.. asking if the patient smokes), physical examination
(e.g., a prostate examination), laboratory test (c.g.. a serum phenylalanine
determination}, or other procedure (e.g.. a sigmoidoscopy) that can be
applied rapidly. Screening tests sort out apparently well persons who have
a disease or a risk factor for a disease from those who do not. It is part of
many primary and all secondary prevention activities. A screening test is
not intended to be diagnostic. If the clinician is not cunurrittcd to further
investigation of abnormal results and treatment, if necessary, the screening
test should not be performed at all.
TERTIARY PREVENTION
Tertiary prevention refers to those clinical activities that prevent further
deterioration or reduce complications after a disease has declared itself.
An example is the use of beta-blocking drugs to decrease the risk of death
in patients who have recovered from myocardial infarction. The bound-
aries of tertiary prevention blend into curative medicine, but well-
performed tertiary prevention goes beyond treating the problems patients
present with. For example, in diabetic patients, tertiary prevention requires
more than good control of blood glucose; patients need regular ophthalmo-
logic examinations for early diabetic retinopathy, education for routine
foot care, searches for and treatment of other cardiovascular risk factors,
and monitoring for urinary protein so that angiotensin-converting enzyme
inhibitors can be used to prevent renal failure.
Tertiary prevention is particularly important in the management of pa-
tients with fatal disease. The goal here is not to prevent death but to
maximize the amount of high-quality time a patient has left. For example,
presently there is no specific therapy for patients with amyotrophic lateral
sclerosis, a neurologic condition ending in paralysis of respiratory and
swallowing muscles. But careful medical management C,In lead to early
intervention with a gastrostomy for administering food and liquids to
prevent dehydration and weakness from starvation, a tracheostomy for
better suctlontng to prevent pneumonia for as long as possible, and if the
patient wishes, a portable respirator to rest respiratory muscles. Without
such a proactive approach, the patient may present with acute respiratory
failure due to the combined effects of the underlying disease, dehydration,
and pneumonia. Patient, family, and physician are then faced with endotra-
cheal intubation and admission to the intensive care unit, with the hope
of reversing enough of the processes to reestablish decent quality of life
for a little longer. Tertiary prevention can help avoid this scenario.
There are few, if any, tertiary prevention programs outside the health
care system, but many health care professionals in addition to physicians
are active in these programs.
Approach to the Periodic Health Examination

When considering what to do routinely for patients without specific
symptoms for a given disease, the clinician must first decide which medical
problems or diseases he or she should try to prevent. This statement is so
straightforward that it would seem unnecessary. But the fact is that many
preventive procedures, especially screening tests, are performed without
a clear understanding of what is being sought. For instance, a urinalysis
is frequently ordered by physicians performing routine checkups on their
patients. However, a urinalysis might be used to search for ilny number
of medical problems, including diabetes, asymptomatic urinary tract infec-
tions, and renal calculi. It is necessary to decide which, if any, of these
conditions is worth screening for before undertaking the test.
Three criteria are important when deciding what condition to include
in a periodic health examination (Table 8,1): (a) the burden of suffering:
caused by the condition, (b) the quality of screening test if one is to be
performed, and (c) the effectiveness of the intervention for primary preven-
tion (e.g., counseling patients to practice safe sex) or the effectiveness of
treatment for secondary prevention after the condition is found on screen-
ing (e.g., prostate cancer treatment).
Burden of Suffering
Is screening justified by the severity of the medical condition in terms
of mortality, morbidity, and suffering caused by the condition? Only condi-
tions posing threats to life or health (the six Ds) should be sought. The
severity of the medical condition is determined primarily by the risk it
CHAPTER 8 I PREVENTION 169
Table 8.1 •
Criteria for Deciding Whether a Medical Condition Should Be Included in Periodic
Health Examinations
1. How qreat is the burden of $utterirlg caused by lhe condition in terms of:
Death Discomfort
Disease Dissatisfaction
Disability Destitution
2. How good is the screening test, it cnc is to tJC performed, in terms 01:
Sensitivity end Labeling effects
Spccncrtv Safety
Simplicity Acceptability
3. a. For primary prevention, how effcctivl} i~; Ihe intervention?
"
b For secondary prevention, if the condition is found. how cttccnvo i~; the en~;uirlg
treatment in terms ot
Efficacy
Patient compliance
Farly treatment being more effective than later treatment
poses or its prognosis (discussed in Chapters 5 and 6). For example, except
during pregn<lncy and before urologic surgery, the health consequences
of asymptomatic bacteriuria are not dear. We do not know if it causes
renal failure and Zor hypertension. Even so, bacteriuria is frequently sought
in periodic health examinations.
Burden of suffering takes into account the frequency of a condition.
Often a particular condition causes great suffering for individuals unfortu-
nate enough to get it, but occurs too rarcly-c-pcrhaps in the individual's
particular age group-for screening to be considered. Breast cancer and
colorectal cancer are two such examples. Although both can occur in much
younger people, they primarily occur in persons older than 50 years. For
women in their early 20s, breast cancer incidence is 1 in 100,000 (one-fifth
the rate for men in their early 70s) (3). Although breast cancer should be
sought in periodic health examinations in women over 50, it is too uncom-
mon in 20-ye<IT-old women (or 70-year-old men) for screening. Screening
for very rare diseases means not only that at most very few people will
benefit but, because of false-positive tests, that many people may suffer
harm from labeling and further workup (sec below).
A particularly difficult dilemma faced by clinicians and patients is the
situation in which a person is known to be at high risk for a condition,
but there is no evidence that early treatment is effective. What should the
physician and patient do? For example, there is evidence that people with
Barrett's esophagus (a condition in which the squamous mucosa in the
distal esophagus is replaced by columnar epithelium) run a 30- to 40-fold
greater risk of developing esophageal cancer than persons without Barrett's
esophagus (4). However, the effectiveness of screening such people with

periodic endoscopic examinations followed by early treatment if cancer
occurs is unknown.
There is no easy answer to this dilemma. But if physicians remember
that screening will not work unless early therapy is effective, they can
weigh carefully the evidence about therapy with the patient. If the evidence
is against effectiveness, they may hurt rather than help the patient by
screening.
Which Tests?
The following criteria for a good screening test apply to all types of
screening tests, whether they are history, physical examination, laboratory
tests or procedures.
SENSITIVITY AND SPECIFICITY
The very nature of searching for a disease in people without symptoms
for the disease means prevalence is usually very low, even among high-
risk groups selected because of age, sex, and other characteristics. A good
screening test must, therefore, have a high sensitivity, so it does not miss
the few cases of disease that are present, and a high specificity, to reduce
the number of people with false-positive results who require further
workup.
Sensitivity and specificity arc determined for screening tests much as
they are for diagnostic tests, except that the gold standard for the presence
of disease usually is not another test but rather a period of follow-up. For
example, in a study of fecal occult blood tests for colorectal cancer, the
sensitivity of the test was determined by the ratio of the number of c010-
rectal cancers found during screening to that number plus the number of
intervalcancers, colorectal cancers subsequently discovered over the follow-
ing year in the people with negative test results (the assumption being that
interval cancers were present at screening but were missed, i.e., the test
results were false negative) (5). Determination of sensitivity and specificity
for screening tests in this way is sometimes referred to as the detection
method.
The detection method for calculating sensitivity works well for many
screening tests, but there are two difficulties with the method for some
cancer screening tests. First, it requires that the appropriate amount of
follow-up time is known; often it is not known and must be guessed. The
method also requires that the abnormalities detected by the screening test
would go on to cause trouble if left alone. This second issue is a problem
in screening for prostate cancer. Because histologic prostate cancer is so
common in men (it is estimated that 25')'0 of 50-year-old men have histologic
foci of prostate cancer, and by the age of 90, virtually all men do), screening
CHAPTER 8 / PREVENTION 171
tests can find such cancers in milny men, but for most, the cancer will
never become malignant. Thus, when the sensitivity of prostate cancer
tests such as prostate-specific antigen (PSA) is determined by the detection
method, the test may look quite good, since the numerator includes all
cancers found, not just those with malignant potential.
To get around these problems, the incidence method, a new method,
calculates sensitivity by using the incidence in persons not undergoing
screening and the interval cancer rate in persons who are screened. The
rationale for this approach is that the sensitivity of a test should affect
interval cancer rates but not disease incidence. For prostate cancer, the
incidence method defines sensitivity of the test as 1 minus the ratio of the
interval prostate cancer rate in a group of men undergoing periodic screen-
ing to the incidence of prostate cancer in a group of men not undergoing
screening (control group). The incidence method of calculating sensitivity
gets around the problem of counting "benign" prostate cancers, but it may
underestimate sensitivity because it excludes cancers with long lead times.
True sensitivity of a test is, therefore, probably between the estimates of
the two methods.
Because of the low prevalence of most diseases, the positive predictive
value of most screening tests is low, even for tests with high specificity.
Clinicians who practice preventive health care by performing screening
tests on their patients must accept the fact that they will have to work up
many patients who will not have disease. However, they can minimize
the problem by concentrating their screening efforts on people with a
higher prevalence for disease.
Example The incidenceof breast cancer increases with age, from approx-
imately 1 in 100,000/yl:'ar at age 20 tal in 200/year over age 71). Therefore,
a lump found during screening in a young woman's breast is more likely to
be nonmalignant than a lump in an older woman. In a large demonstration
project 011 breast cancer screening, biopsy results of breast masses varied
markedly according to the age of women (6); in women under age 40, more
than 16 benign lesions were found for every malignancy, but in women over
age 70 fewer than .'1 benign lesions were found for every malignancy (Fig.
8.2). Sensitivity and specificity of the clinical breast examination and mam-
mography arc better in older women as well, because of changl:'s in breast
tissue as women grow older.
The yield of screening decreases as screening is repeated over time in
a group of people. Figure 8.3 demonstrates why this is true. The first time
that screening is carried out-the prevalence screen-cases of the medical
condition will have been present for varying lengths of time. During the
second round of screening, most cases found will have had their onset
between the first and second screening. (A few will have been missed by
the first screen.) Therefore, second and subsequent screenings are called
incidence screens, Figure 8.3 illustrates how, when a group of people are
172 CLINICAL EPIDEMIOWGY
18
14
-=
~ '"
rJl:=: 1a
Ill';':
a:c:
>'"
U)c: 6
'c..~
0-
.- '"
lIlE
o
c: 2 o
o
Z
40 50 60 70 80
Age (years)
Figure 8.2. Yiold of a screoning test according to patients' age. Ratio of nonmalig-
nantmalignant biopsy results among women screened for breast cancer. (Data from
Baker LH. Breast Cancer Detection Demonstration Project: five-year summary report.
CA 1982;32:195-231.)
periodically rescreened, the number of cases of disease present in the group

drops after the prevalence screen. This means that the positive predictive
value for test results will decrease after the first round of screening.
SIMPLICITY AND LOW COST
An ideal screening test should take only a few minutes to perform,

require minimum preparation by the patient, depend on no special ap-
pointments, and be inexpensive.
Simple, quick examinations such as blood pressure determinations are
ideal screening tests. Conversely, complicated diagnostic tests such as co-
lonoscopv. which are expensive and require an appointment and bowel
preparation, arc reasonable in patients with symptoms and clinical indica-
tions but may be unacceptable as screening tests, especially if they must
be repeated frequently. Other tests, such as visual field testing for the
detection of glaucoma and audtograms for the detection of hearing loss,
fall between these two extremes. Even if done carefully, such tests, <11-
Round
2 3
D,
(}-- D,
D,
D,
~D,
D,
-D,
0,
0 f-D
P- ~--D,
0,
533
Number of Cases
Newly Detected
though not as difficult as colonoscopy, are probably too complex to be

used as screening tests.
The financial "cost" of the test depends not only on the cost of (or
charge for) the procedure itself but also on the cost of subsequent evalua-
tions performed on patients with positive test results. Thus sensitivity,
specificity, and predictive value affect cost. Cost is also affected by whether
the test requires a special visit to the physician. Screening tests performed
while the patient is seeing his or her physician for other reasons (as is
frequently the case with blood pressure measurements) are much cheaper
for patients than tests requiring special visits, extra time off from work,
and additional transportation.
SAFETY
Tt is reasonable and ethical to accept a certain risk for diagnostic tests
applied to sick patients seeking help for specific complaints. The patient comes
asking for help, sometimes with a problem about which little is known.
The physician cannot postpone action and does his or her best. It is quite
another matter to subject presumably well people to risks when there is
no known problem. In such circumstances, the procedure should be espe-
cially safe. This is partly because the chances of finding disease in healthy
people arc so low. Thus, although colonoscopy is hardly thought of as a
"dangerous" procedure when used on patients with gastrointestinal com-
plaints, it may be too dangerous to use as a screening procedure because
of the possibility of bowel perforation. In fact, if colonoscopy, with a perfo-
ration rate of 0.2°,{" were used to screen for colorcctal cancer in women in
their 50s, almost two perforations would occur for every cancer found. For
women in their 70s, the ratio would reverse, because colorectal cancer is
so much more common (7).
ACCEPTABLE TO BOTH PATIENTS AND CLINICIANS
The importance of acceptability is illustrated by experience with tests
for early cervical cancer and early colon cancer. Women at greatest risk
for cervical cancer are least likely to get routine Pap smears. The same
problem holds true for colorectal cancer. Studies indicate there is a strong
reluctance among asymptomatic North Americans to submit to periodic
examinations of their lower gastrointestinal tracts-a finding that should
be no surprise to any of us!
Table 8.2 shows acceptance of screening for colorectal cancer by various
kinds of people. People who voluntarily attended a colon-eta I cancer
screening clinic were very cooperative; they were willing to collect stool
samples, smear the samples on guaiac-impregnated paper slides, and mail
the slides to their doctors for clinical testing. Patients who did not volunteer
were less willing to participate. Older persons, who arc at greatest risk for
culun-ctal cancer because of their age, were least willing to be screened.
Table 8.2
Patients' Acceptance of Screening Tests: Reported Response Rates for Returning
Guaiac-impregnated Slides in Different Settings·
I·'mticirilnts
Returning SlidHS
SHitin'1 fPArGent\
- - --'----
Colorectal cancer screening proqram 85
Breast cancer ~;cr(~cninlJ proqram 70
HMO members agee) 50-74 years 27
HMO members aged [iO-74 years cent kit. rorrnnoer letter, and sell- 4e'
help booklet ,md who were VllIGd with instrucnons and renmcors
"Ofltd trom Myc,,'; 11[. Ross EA, Wult TA. necrcm A, ,JAIH1I1 C, Millner I Behavioral intmvAfltions to
ocrcrc,e adhArAllce in CGlorHc1,,1 scrccns-q. M"d Colee 1YY1;2'1: I fXN-1ilbil: c-o HHld\er SW, llaupllillAe
Wi!. Should culorAclal carcinonli; bH oouqht in periodic IlAillth 0Xlllllindtiofls'! 1\11 apptoach 10the evktcnce
eli" IrlVHst Med t')SI:'1Z]-Jl
Substantial extra effort can result in getting more people (but still fewer
than half) to participate.
TIlt' acceptability of the test to clinicians is a criterion usually overlooked
by all but the ones performing it. After one large, well-conducted study
on the usefulness of screening, sigmoidoscopy was abandoned because the
physicians performing the procedure-gastroenterologists, at that-found
it too cumbersome and time-consuming to be justified by the yield (8).
(Patient acceptancc, 1RClj,), was not good either.]
LABELING
The labl'lillg effect describes the psychological effect of test results or

diagnoses on patients. Studies of labeling suggest that test results can
sometimes have important psychological effects on patients.
Labeling can either help or hurt patients. A positive labeling effect may
occur when a patient is told that all the screening test results were normal.
Most clinicians have heard such responses as, "Great, that means I can keep
working for another year." If being given a dean bill of health promotes a
positive attitude toward one's daily activities, a positive labeling effect has
occurred.
On the other hand, being told that something is abnormal may have an
adverse psychological effect. A study of women who had false-positive
mammograms (women with suspicious mammograms who on subsequent
evaluation were found not to have cancer) found that several months later
almost half reported mammography-related anxiety (47%) and worries
about breast cancer (4l'~;,); 17'};, said the worries affected their daily func-
tion (9).
Labeling effects of screening tests may become a major concern with
progress made in genetic screening. A gene has been identified for Hun-
tington's chorea, and relatives of affected individuals can be tested to sec
if they carry the dominant, universally fatal gene. Such a test may help
people who wonder if they should marry and have children. More compli-
cated are the much more common situations in which genes are associated
with a risk, not a certainty, of future disease. For example, several genes
are known to be associated with colorectal and breast cancer. In these
situations, many people with the genes will not get cancer, and many
without the particular genes will get the cancer. Because the events are in
the future, persons who have been told they have one of these genes will
have to live with the possibility of a dire event for a long time.
Negative labeling effects are particularly worrisome ethically when they
occur among patients with false-positive tests. In such situations, screening
efforts might promote a scn ic of vulnerability instead of health and might
do more harm than good.
RISK OF A FALSE-POSITIVE RESULT

The previous discussion applies to each of the individual screening
tests that a clinician might consider performing during a periodic health
examination. However, most clinicians do not perform only one or two
tests on patients presenting for routine checkups. In one study, practicing
internists believed that 57 different tests should be performed during peri-
odic health examinations (10). Modern technology, and perhaps the threat
of lawsuit, has fueled this propensity to "cover an the bases." Automated
blood tests allow physicians to order up to several dozen tests with a few
checks in the appropriate boxes.
When the measurements of screening tests are expressed on interval
scales (as most are) and when normal is defined by the range covered by
95% of the results (as is usual), the more tests the clinician orders, the
greater the risk of a false-positive result. In fact, as Table 8.3 shows, if
the physician orders enough tests, "abnormalities" will be discovered in
virtually all healthy patients.
Effectiveness of Treatment
"Treatments" in primary prevention are immunizations, such as tetanus
toxoid to prevent tetanus; drugs, such as aspirin to prevent myocardial
infarction; and behavioral counseling, such as helping patients stop smok-
ing or adopt low-cholesterol diets. Whatever the intervention, it should be
efficacious (produce a beneficial result in ideal situations) and effective
(produce a beneficial result under usual conditions). Efficacy and effective-
ness of pharmaceuticals are usually better documented than they are for
behavioral wunseling. Federal laws require rigorous evidence of efficacy
before pharmaceuticals are approved for use. The same is not true for
behavioral counseling methods, but clinicians should require scientific evi-
dence before incorporating routine counseling into health maintenance.
Health behaviors are among the most important determinants of health in
modern society; effective counseling methods could promote health more
Table 8.3
Relation between Number of Tests Ordered and Percentage of Normal People with
at Least One Abnormal Test Resutt"
PeoplH with at I east Ore

Number of Tests Abnornmlily {I-'ercenl)
5
:<3
20 (;4
100 99.4
"rom Sackett OL, Clinical Jiagnosis Hfl(j the c1imcallaboratofY. C1in Invesl MH<t 19713: l:M -4:3,
than most anything else a clinician can do, but counseling that does not
work wastes time, costs money, and may harm patients.
Example Two different smoking cessation counseling strategies-
weekly hour-long group counseling sessions for 8 weeks and weekly 10- to
20-min individual counseling sessions for 8 weeb-were combined with
nicotine patch therapy and evaluated for their effectiveness in promoting
smoking cessation (11,12). Compared with patients randomized to control
groups, the patients receiving the interventions did somewhat better, with a
third of patients in the group counseling sessions having stopped smoking
at 6 months follow-up. However. fewer than 20% of patients receiving indi-
vidual counseling had stopped smoking. furthermore, the authors found
that most failures at fi months could be predicted by patients smoking at
some time during the first 2 weeks after trying to stop. These findings suggest
that counseling should be "front loaded" By carefully evaluating behavioral
counseling, studies such as this are determining what approaches work.
Treatments for secondary prevention are generally the same as treat-
ments for curative medicine. Like interventions for primary prevention,
they should be both efficacious and effective. If early treatment is not
effective, it is not worth screening for a medical problem regardless of how
easily it can be found, because early detection alone merely extends the
length of time the disease is known to exist, without helping the patient.
Another criterion important for treatments in secondary prevention is
that patient outcome must be better if the disease is found by screening,
when it is asymptomatic, than when it is discovered later, after the condi-
tion becomes symptomatic and the person seeks medical care. If outcome
in the two situations is the same, screening is not necessary.
Example In a study of the use of chest x-rays and sputum cytology to
screen for lung cancer, mille cigarette smokers who were screened every 4
months and treated promptly if cancer was found did no better than those
not offered screening (13); at the end of the study, death rates from lung
cancer were the same in the two groups-3.2 per lOOO person-years in the
screened men versus 3.0 per 1000persons·years in men not offered screening.
Farly detection and treatment did nol help patients with lung cancer more
than treatment of people at the time they presented with symptoms.
BIASES
As discussed in Chapter 7, the best way to establish the efficacy of
treatment is with a randomized controlled trial. This is true for all interven-
tions but especially for early treatment after screening. To establish that a
preventive intervention is effective typically takes years and requires large
numbers of people to be studied. for example, early treatment after colo-
rectal cancer screening can decrease colorectal cancer deaths by approxi-
mately one-third. But to show this effect, a study with 13 years of follow-
up was required (5). A "clinical impression" of the effect of screening
simply does not suffice in this situation.
DIAGNOSIS
Onset Early Usual Death
UNSCREENED
SCREENED
Early treatment 0 - - - - Dx ",*i!'i;i#ii*i!'i;i#ii*#ii*#ii*#ii*#ii*t
not effective
Improved
Survival
SCREENED
Early treatment 0 - - - - Dx 7,,~~~~~~~;;;, t
is effective
Figure 8.4. How lead time affects survival lime after screening; shaded areas
indicate length of survvat after diagnosis (Ox).
Careful studies an' also necessary because of biases that are specific to
studies of the effectiveness of screening programs. Three such biases are
described below.
lead Time Bias
read time is the period of time between the detection of a medical condi-
tion by screening and when it ordinarily would be diagnosed because a
patient experiences symptoms and seeks medical care (Fig. 8.4). The
amount of lead time for a given disease depends on both the biologic rate
of progression of the disease and on the ability of the screening test to
detect early disease. When lead time is very short, as is presently the case
with lung cancer, treatment of medical conditions picked up on screening
is likely to be no more effective than treatment after symptoms appear.
On the other hand, when lead time is long, as is true for cervical cancer
(on average, it takes approximately 30 years to progress from carcinoma
in situ to clinically invasive disease), treatment of the medical condition
found on screening can be very effective.
How can lead time cause biased results in a study of the efficacy of
early treatment? As Figure 8.4 shows, because of screening, a disease is
found earlier than it would have been after the patient developed symp-
toms. As a result, people who are diagnosed by screening for a deadly
disease will, on average, survive longer from the time of diagnosis than
people who are diagnosed after they get symptoms, even if eilrly treatment
is no more effective than. treatment at clinical presentation. In such a situa-
tion screening would appear to help people live longer, when in a reality
they would be given not more "survival time" but more "disease time."
Screening
0 0,
0--0,
0 0,
0,
0 0,
0,
0,
0,
0-0,
0,
0,
0,
Figure 8.5. Length time bias. Cases that progress rapidly from onset (0) to symp-
toms and diagnosis (Ox) are less likely 10 be detected during a screening examination.
An appropriate method of analysis to avoid lead time bias is to study both

a screened group of people and a control group of people and compare age-
specific mortality rates rather than survival rates from the time of diagnoses.
We can be confident that early diagnoses and treatment of colorectal cancer
are effective, because studies have shown that mortality rates of screened per-
sons are lower than those of a comparable group of unscrecned people (5).
Length Time Bias
Lengthlime bias (see Figs. 8.5 and 8.6), another bias that can affect studies
of screening, occurs because the proportion of slow-growing lesions diag-
nosed during screening programs is greater than the proportion of those
diagnosed during usual medical care. The effect of including a greater
number of slow-growing cancers makes it seem that screening and early
treatment are more effective than usual care.
Length time bias occurs in the following way. Screening works best
when a medical condition develops slowly. Most types of cancers, how-
ever, demonstrate a wide range of growth rates. Some cancers grow slowly,
some very fast. Screening tests arc likely to find mostly slow-growing
tumors because they are present for a longer period of time before they
cause symptoms; fast-growing tumors are more likely to cause symptoms
that lead to diagnosis in the interval between screening examinations.
Screening, therefore, tends to find tumors with inherently better prognoses.
As a result, the mortality rates of cancers found on screening may be better
than those not found on screening, but it is not because of the screening
itself.
180 CLINICAl, EPIDEMIOLOGY
Compliance Bias
Compliance bias, the third major type of bias that can occur in effective-
ness studies of presymptomatic treatment is the result of the extent to
which patients follow medical advice. Compliant patients tend to have
better prognoses regardless of screening. If a study compares disease out-
comes among volunteers for a screening program with outcomes in a group
of people who did not volunteer, better results for the volunteers might not
be due to treatment but be the result of other factors related to compliance.
Example In a study of the effect of a health maintenance program, one
group of patients was invited fur an annual periodic health examination and
a comparable g:roup was not invited (14). Over the years, however, some of
the control group asked for periodic health examinations. As seen in Figure
8.7, those patients in the control group who ilctivdy sought out the examina-
tions had better mortality rates than the patients who were invited for screen-
ing. The latter group contained not only compliant patients but also ones
who had to be persuaded to participate.
Biases due to length time and patient compliance can be avoided by

relying on randomized controlled trials that count all the outcomes in the
groups, regardless of the method of diagnosis or degree of participation.
Groups of patients that are randomly allocated will have comparable num-
1 Diagnosis after
symptoms
Onset "
Screened
-----Time-----~
Figure 8.6. Length time bias. Rapidly growing tumors come to medical attention
before screening is oerrorrneo, whereas more slowly growing tumors allow time for
detection. D, diagnosis after symptoms; S, diagnosis after screening.
2.0
•
0 0 0
1.0 • 0 0 0
• • 0
0
0.5 • • •
•
o 2 3 4 5
Number of MHCs
Figure 8.7. Effect of patient compliance on a screening program. The control group
of patients (e) had 8 lower standardized mortality ratio (observed:expected deaths,
standardized for age) than the study group offered {OJ screening; but the control
group included only patients who requested screening, whereas the study group
included all patients offered screening. MI-JCs, multiphasic health checkups, (Re-
drawn trom Friedman GO, Collen MF, Fireman BH, Multiphasic health checkup evalu-
ation: a 16-year follow-up. J Chron Dis 1986; 39:453-463.)
bers of slow- and fast-growing tumors and, on average, comparable levels

of compliance. These groups then can be followed over time with mortality
rates, rather than survival rates to avoid lead time bias.
Because randomized controlled trials are difficult to conduct, take so
long. and are expensive, investigators sometimes try to use other kinds of
studies, such as cohort studies (Chapter 5) or case control studies (Chapter
10), to investigate preventive maneuvers and effectiveness of treatment
after screening.
Example To test whether periodic screening with sigmoido:;copy rl;'-
duces mortality from cclorectal cancer within the reach of the sigmoidoscope,
Selby et al. (15) investigated the frequency of screening sigmoidoscopy over
the previous 10 years among; patients dying of colorccta! cancer and alllong
well patients, matched for age and Sl;'X. To deal with lead time and length
time biases, they investigated screening only in people who were known to
have died (case group) or not to have died (control group) from colorectal
cancer. To deal with compliance bias. they adjusted their results for the
number of general periodic health examinations each person had. They also
adjusted the results for the prescnCl' of medic-ill conditions that could have
led to both increased screening and increased likelihood of culorectal cancer.
Patients dying of colorer-tal cancer in the rectum Of distant sigmoid were less
likely \0 have undergone a screening sigmoidoscopy in the previous 10 years
(8.8'\.) than those in the control group (24.2%), and sigmoidoscopy followed
by early therapy prevented almost 60(:'~, of deaths [rom distal colorcctat cancer.
Also, by showing that there was no protection for colorectal cancers above
the level reached by sigmoidoscopy, the authors suggested that "it is difficult
to conceive of how such anatomical specificity of effect could be explained
by confounding."
Case series, in which a group of people participating in a screening

program arc followed over time, Me a common but inappropriate method
of evaluating the effectiveness of screening progrilms; they Me subject to
all the biases discussed (see Chapter Hj for more on case series).
How Much Harm for How Much Good?

Health promotion and disease prevention are becoming increasingly
popular. The goal of keeping people as healthy as possible is laudable, but
as this chapter points out, the concepts behind the goal are complex. Most
important, health promotion activities can cause harm. In filet, it is probably
fair to say that they usually do cause harm, even though totally unintended.
At the least, they cost money, patients' time and often discomfort. At the
worst, they can cause serious physical harm in the rare patient. either
because of complications of the screening test itself or because of adverse
consequcnces of subsequent tests or treatment, particularly in patients with
false-positive test results. False-positive tests can cause psychological dam-
age as well. Thus it is important that the clinician have solid evidence
about how much good and how much harm health promotion activities
accomplish. Good intentions are not enough.
Before undertaking a health promotion procedure on a patient, espe-
cially if the procedure is controversial ilm(mg expert groups, the clinician
should discuss both the pros (probability of known end hoped-for health
benefits) and cons (probability of unintended effects) of the procedure with
the patient.
Example Although clinical breast examinations and mammography

screening for breast cancer arc universally recommended for older women,
there is controversy about screening for women ages 40 to 49; randomized
controlled trials show that screening does not work \\'10'11 in this age group,
but il protective effect of about 15':;, may still be possible after many years.
Expert groups are divided in their recornrncndations. When discussing this
dilemma with" patient, it is useful to demonstrate both benefits and harms
resulting from screening (Fig 8.8). Such an approach not only is more honest
with the patient but helps clarify the situation for her so that her consent for
whatever is chosen is trulv informed. Cost effectiveness analvsis formalizes
this approach for policvmakcrs (16). .
CHArTER 8 I PREVENTION 183
Screened Not Screened
1000
False-positive
mammograms
300
-
1 50
Procedures because
of mammograms
.1
15
New breast cancers
.115
Deaths from
breast cancer
5-B 7-8
Figure 8.8. Weighing tenent and harm from screening. What happens during a
decade of annual mammography in 1OOll women starting at ago 40,
Current Recommendations
With progress in the science of prevention, current recommendations
on health maintenance are quite different from those of the past. Several
groups have recommended abandoning routine annual checkups in favor
of a selective approach in which the tests to be done depend on a person's
agc, sex, and clinical characteristics (thereby increasing prevalence and
positive predictive value). They have also tended to recommend fewer

tests than previously (thereby decreasing the percentage of patients with
false-positive results). Several groups have turned their attention to the
selection process for deciding what medical conditions should be sought.
There is increasing concern for dear delineation of the criteria that tests
should meet before they are incorporated into periodic health examina-
tions. Croups with explicit criteria for selecting medical conditions are
more conservative in their recommendations than groups without such
criteria.
Summary
Disease can be prevented by keeping it from occurring in the first place
(primary prevention), with interventions such as immunization and behav-
ioral counseling. Such interventions should be evaluated for effectiveness
as rigorously as other kinds of clinical interventions.
Il] effects from disease can also be prevented by conducting screening
tests at a time when presymptomatic treatment is more effective than treat-
ment when symptoms occur (secondary prevention). A disease is sought
if the disease causes a substantial burden of suffering, if a good screening
test is available, and if prcsymptomatic treatment is more effective than
treatment at the usual time. Screening tests should be sensitive enough to
pick up most cases of the condition sought, specific enough that there are
not too many false-positive results, inexpensive, safe, and well accepted
by both patients and clinicians.
In secondary prevention, three potential biases threaten studies of the
effectiveness of presymptomatic treatment: failure to account for the lead
time gained by early detection, the tendency to detect a disproportionate
number of slowly advancing cases when screening prevalent cases, and
confounding the good prognosis associated with compliance with the ef-
fects of the preventive intervention itself.
Based on these criteria, i\ limited number of primary prevention inter-
ventions and screening tests for secondary prevention are recommended
for health maintenance, according to the age, sex, and clinical status of the
patient.
REFERENCES
I. Schappert SM, National arnbulJlory medical care survey: 1991 summary, Advance data
from vital and health statistics, No. 210. Hyattsville, MD: National Center for II<,alth
Statistics, 1993.
2. Webskr's ninth new collegiate dictionary, Springfield, MA: M,ni"m-Webster, 1991.
J. Rics LAC, Miller 13A, Hanh:y BF. Kosary Cl, Harras A, Edwards 13K, eds. SEER ,ancer
statistics review, 1973 1991, tanks ond grophs. NIH Publication No. 94-27H9. Ikthesda,
MD: National Cancer Institute, 1994.
4. Cameron A), Ott B). Payne WS, The incidence of ad"n",'arcinoma in columnar-lined
(l:Iarrdt's) esophagus. N Fngl J Med 1985;J13:H57-H5<J.
CIIAPTER 8 I PREVENTION 185
:;, Mandel JS, ct al. R"ducing mortality from colorectal cancer by screening tor t",.,,1 occult
blood:--'- Engl j Med 1';l93;J2Ii:D6S·-1371
h. Ihk",r l.H. RW,1bl Cancer Detection Demonstration Project: fiv(~ye~r bUmnlMY report. CA
19H2;J2:1% - 231.
7. Eddy OM, S.:r"ening f"r mlor"d~1 ('~ncer Ann Intern IvIed 1<J<J1l;1Ll:171 3H4,
8. Dales LG, Friedman CO, Collen MF, F.valu~ting periodic multiphasic health ched("ups:
a controlled trial. ] Chron Oisl<J79;,12:3H.5-404.
'I. lerman C Trod: B, Rimer BK Boyce A,Jepson C Engstrom PF. Psychological and behav-
ioral Impli"atiom "f ~hllormal mammograms, Ann Intern Med '199'1;'1'14:6-57-661.
lO, komm 1'.1, l'ktcht'rSW, 1'lulL! llS, The periodic health examination: comparison "fr",com-
mendaticns and internists' performance, South Med J 1981;74:263-271.
11 Fiore MC, Kenford SL, lorenby Dr, Weller OW, Smith S5, llaker TE. The nicotine patch;
clinical effedivcn,,~s with dlffewnt counseling treatments. Chestl<J<J4;1O,S:,S24-533.
12, Kenford, SL, Fiore MC, Jorcnby DE, Smitl1 S5, "Vetter D, Baker TEl. Prediding sm"klng
(ess~ti"n: who will quit with and without the nicotine patch. JAMA. 1994;2715H'J-5<J4.
13, Fontana RS, Sanderblll1 DI~, Woolner LB, Taylor WF, Miller We, Muhm JR. Lung cancer
screening: the 'Vby" progrilm. J Occup Med 1986;2H:746-750.
14, friedman GD, Collen MF, Fin'JIlan BH. Multiphasic health checkup evaluation: il 16-yeilr
follow-up. J Chron Ois 19H6;39:453 4h3.
15. Sdhy JV, Frledm,'n GD, Quesenberry C!', ~Veis-, NS. A cilse-control study of screening
slgm"ido~c()py and mortality from colorcctal cancer. N Fng J Med 1992;326:65J-657,
16. l'ddy OM, Hilsselhbd V, McGivney W, Hendee W. The value "f m~mmography screening
in women under agc Sil yt'art']AMA 1988;239:1512-1.519,
SUCCESTED READrI\;GS
Eddy DM fed) Common soecntng tests, I'hiladelphia, An",riciln College of Physicians. 1991
Goldbloom lUI, Lawn'n"" RS, <,ds, Pr<,v",nting disease: beyond the rhetoric, Nt'w York:
Springer-Verlag, I'JtJIl,
Cuy"tt CH, Sarkdt D1., Cook DJ. Users' guides to Ih~ m"di"al Iiter~hJre. 11: How to use an
article about therapy or prevention. A: Are the results of tl1<' sh"ly valid? jAylA
1'I'IJ;270:259H-2601
Guyall GH, Sackett DL, Cook OJ. Users' guide" to th", medical literature. II: I low to use an
arlic1e about therapy or prcvention, 5: What w,,,e the results and will they help me in
eming for my patlenh 7 JAMA 1994;27159-6.1.
H"yw,lrd "';, Slelnherg FP, Ford DE, Roizcn MI', Roach KW, I'reventiv'.' C,1r", guidelines: 1491.
Ann Intern Mcd 1991,1l47.SR-7!l3
Miller AB, cd. Screening tor Gu",er. Orbndo, FI Academic Press, 1%5.
Miller All, Chamberlain J, Oay NE, llakdn,a M, Prorok PC eds. Cancer screening. Cambridg",
England: Inkrniltional Union Against Cancer, 1991
Russell LI:\. Educ,lted guesses: making policy about medical screening tesls. R"rkel('Y' Univ('r-
blty of California t'ross, [9<J4.
u.s. Preventive Services Task rorcc. Cuide to pn"',,ntive st'rvin'b: an assessment of the effec-
tiveness of '169 Intervention,;. Raltimore: Williams & VVilkins, '1989,
9
CHANCE
When clinicians attempt to learn from clinical experience, whether dur-

ing forma] research or in the course of patient care, their efforts are im-
peded by two processes: bias and chance. As we discussed (Chapter 1),
bias is systematic error, the result of any process that causes observations
to differ systematically from the true values. In clinical research, a great
deal of the effort is aimed at avoiding bias where possible and controlling
for and estimating its effects when bias is unavoidable.
Random error, on the other hand, is inherent in all observations. It can
be minimized but never avoided altogether. Random variation can arise
from the process of measurement itself or the biologic phenomenon being
measured (Chapter 2). This source of error is called "random," because
on average it is as likely to result in observed values being on one side of
the true value as on the other.
Most of us tend to overestimate the importance of chance relative to
bias when interpreting data. We might say, in essence, "If pis <0.001, a
little bit of bias can't do much harm!" However, if data are assembled
with unrecognized bias, no amount of statistical elegance can save the day.
As one scholar put it, perhaps taking an extreme position, "A well de-
signed, carefully executed study usually gives results that are obvious
without a formal analysis and if there arc substantial flaws in design or
execution a formal analysis will not help" (1).
In this chapter, chance is discussed in the context of a controlled clinical
trial, because that is a simple way of presenting the concepts. However,
application of the concepts is not limited to comparisons of treatments in
clinical trials. Statistics are used whenever one makes inferences about
populations based on information obtained from samples.
Random Error
The observed differences between treated and control patients in a clini-
cal trial cannot be expected to represent the true differences exactly because
of random variation in both of the groups being compared. Statistical tests
186
CHAPTER 9 ! CHANCE 187
help estimate how well the observed difference approximates the true
one. Why not measure the phenomenon directly and do away with this
uncertainty? Because research must ordinarily be conducted on a sample
of patients and not all patients with the condition under study. As a result,
there is always a possibility that the particular sample of patients in a
study, even though selected in an unbiased way, might not be similar to
the population of patients as a whole.
Two general approaches are used to assess the role of chance in clinical
observations. The first, called hypothesis testillg, asks whether an effect (dif-
ference) is present or not by using statistical tests to examine the hypothesis
that there is no difference (the "null hypothesis"). This is the traditional
way of assessing the role of chance, popular since statistical testing W<lS
introduced at the beginning of this century and associated with the familiar
"p values." The other approach, called estimation, uses statistical methods
to estimate the range of values that is likely to include the true value.
This approach has gained popularity recently and is now favored by most
journals for reasons that we describe below.
We begin with a description of the traditional approach.
Hypothesis Testing
In the usual situation, where the principal conclusions of a trial are
expressed in dichotomous terms (e.g., the treatment is considered to be
either successful or not) and the results of the statistical test is also dichoto-
mous (the result is either "statistically slgnificantv-c-i.e., unlikely to oe
purely by chance-c-or not), there are four ways in which the conclusions
of the test might relate to reality (Fig. 9.1).
Two of the four possibilities lead to correct conclusions: (a) when the
treatments really do have different effects and that is the conclusion of the
study and (b) when the treatments really have similar effects and the studv
makes that conclusion.
There are also two ways of being wrong. The treatments under study
may actually have similar effects, but it is concluded that the study treat-
ment is better. Error of this kind, resulting in the "false-positive" conclu-
sion that the treatment is effective, is referred to as an a or Type 1 error.
Alpha is the probability of saying that there is a difference in treatment
effects when there is not. On the other hand, treatment might be effective,
but the study concludes that it is not. This "false-negative" conclusion is
called a f3 or Type II error. Beta is the probability of saying that there is
no difference in treatment effects when there is one. "No difference" is a
simplified way of saying that the true difference is unlikely to be larger
than a certain size. It is not possible to establish that there is no differenco
at all between two treatments.
Figure 4.1 is similar to the two-by-two table comparing the results of a
188 CLINICAL cPIDl:MIOLOGY
TRUE
DIFFERENCE
Present Absent
-
§ ••••.• ,,,
CONCLUSION Significant Correct l~~!&f!X1
OF
STATISTICAL
TEST Not
Significant .• (1~~~r~r· Correct
Figure 9.1. The relationship between the results of a statistical test and the true
difference between two treatment groups. 'Absent is a simplification. It really means
that the true difference is not greater than a specified arnount.)
diagnostic test to the true diagnosis (Chapter 3). Here the "test" is the
conclusion of a clinical trial, based on a statistical test of results from a
sample of patients. Reality is the true relative merits of the treatments
being compared-if it could be established, for example, by making obser-
vations on all patients with the illness under study or a large number of
samples of these patients. Alpha error is analogous to a false-positive and
(i error to a false-negative test result. ln the absence of bias, random varia-
tion is responsible for the uncertainty of the statistical conclusion.
Because random variation plays a part in all observations, it is an over-
simplification to ask whether or not chance is responsible for the results.
Rather, it is a question of how likely random variation is to account for
the findings under the particular conditions of the study. The probability
of error due to random variation is estimated by means of inferel1tial statis-
tics, a quantitative science that, based on assumptions about the mathemati-
cal properties of the data, allows calculations of the probability that the
results could have occurred by chance alone.
Statistics is a specialized field with its own jargon-null hypothesis,
variance, regression, power, and modeling-Lhat is unfamiliar to many
clinicians. However, leaving aside the genuine complexity of statistical
methods, inferential statistics should be regarded by the nonexpert as a
useful means to an end. Statistical tests are the means by which the effects
of random variation arc estimated.
The next two sections discuss a and f3 error, respectively. We will at-
tempt to place hypothesis testing, as it is used to estimate the probabilities
of these errors, in context. However, we will make no attempt to deal with
CHAPTER 9 I CHANCE 189
these subjects in a rigorous, quantitative fashion. for that, we i;uggest that

readers consult any of a number of excellent textbooks of biostatistics (see
"Suggested Readings," later in this chapter).
CONCLUDING THAT A TREATMENT WORKS
Most of the statistics encountered in the current medical literature con-
cern the likelihood of an a error and are expressed by the familiar p value.
The p value is a quantitative estimate of the probability that observed
differences in treatment effects in the particular study at hand could have
happened by chance alone, assuming that there is in fact no difference
between the groupi;. Another way of expressing this is that p is an answer
to the question, If there were no difference between treatments and the
trial was repeated many times, what proportion of the trials would
lead to the conclusion that a treatment is as or more effective than found
in the study?
We will call the p value "p,," to distinguish it from estimates of the
other kind of error due to random variation, f3 error, which we will refer
to as "Pi/' When a simple P is found in the scientific literature it ordinarily
refers to what we call p".
The kind of error estimated by p" applies whenever it is concluded that
one treatment is more effective than another. If it is concluded that the p"
exceeds some limit and so there is no difference between treatments, then
the particular value of p., is not as relevant; in that situation, PiJ (probability
of f3 error) applies.
DICHOTOMOUS AND EXACT P VALUES
It has become customary to attach special significance to p values falling
below 0.05 because it is generally agreed that less than 1 chance in 20 is a
small risk of being wrong. A rate of 1 in 20 is so small, in fact, that it is
reasonable to conclude that such an occurrence is unlikely to have arisen
by chance alone. It could have arisen by chance, and 1 in 20 times it will.
But it is unlikely.
Differences associated with p" less than 0.05 are called "statistically
significant." Tt is important to remember, however, that setting a cutoff
point at 0.05 is entirely arbitrary. Reasonable people might accept higher
values or insist on lower ones, depending on the consequences of a false-
positive conclusion in a given situation.
To accommodate various opinions about what is and is not unlikely
enough. some researchers report the exact probabilities of p"s (e.g., 0.03,
0.07, 0.11, etc.). rather than lumping them into two categories. <U.05 or
>0.05. The interpretation of what is statistically significant is then left to
the reader. However, p values greater than 1 in 5 are usually reported as
simply p > U.20, because nearly everyone can agree that a probability of
an a error that is greater than one in five is unacceptably high. Similarly.
below very low Ii values (such as p (l.OOl) chance if; a vcry unlikely
explanation for the observed difference. and little further information is
imparted by describing this chJnce more precisely.
STATISTICAL SIGNIFICANCE AND CLINICAL IMPORTANCE
A statistically significant difference, no matter how smJII the p, does
not mean that the difference is climcallv important. A p < 0.0001, if it
emerges from a well-designed study, conveys a high degree of confidence
that a difference really exists. But this Ii value tells us nothing about the
magnitude of that difference or its cliniral importance. In fact, entirely
trivial differences may be highly statistically significant if a large enough
number of patients '''/(IS studied.
Example In the carlv 1990s there was a heated debate about which
thrombolytic ilgt'nl, streptokinase or tissue plasminogen activator (tPA), is
most effective during acute n-yorardial infarction, Large trials had shown a
difference in reperfuxion rates but nol mortality. The two were rnmpared
(along with subcutaneous or intravenous heparin) in a large randomized
controlled trial. called CUSTO, involvini-\ 4'1,021 patients in 15 countries (2),
tl'A was givcn by a more aggressive regimen than in ear-lier studies. The
death rate at 30 days was lower among patients receiving: tl'A (63\,) than
among those receiving: stre-ptokinase (7.2 or 7.4%, depending on how hepa-
rin was given) and this difference was highly unlikely to be by chance
(p ().OOI), However, the difference is not large; one would have to treat
about 100 patients with tl'A instead of with streptokinase to prevent one
short-term death. Because tf'A is muc-h more expensive than streptokinase-
it would cost nearly $250 thousand to prevent that death (3)-and because
tl'A is more likely to cause hemorrhagic strokes, some have questioned
whether the marginal benefit of tl' A is worl hwhi lc, i.c.. whether the difference
in mortality between tl'A and xh-eptokinnsc treatment. all things considered,
is "clinically slgntflcanr."
On the other hand, very unimpressive p values can result from studies
showing strong treatment effects if there are few patients in the study (sec
the following section),
STATISTICAL TESTS
Commonly used statistical tests, familiar to many readers, are used to
estimate the probability of an a error. The tests arc applied to the data to
give a test statistic, which in turn can be used to corm- up with a probability
of error (Pig. 9.2). The tests Me of the l1ul/ h.llpofhe.c.i.s, the proposition that
there is no true difference in outcome between the 1\'/0 treatment groups.
This device is for mathematiral reasons, not because "no difference" is the
working scientific hypothesis of the study. One ends up rejecting the null
hypothesis (concluding there is a difference) or failing to reject it (conclud-
ing there is no difference).
Some commonly used st,ltisticallcsts are listed in Table 9.1. The validity
Data • Test
statistic Compare to
• Estimate of
probability that
Statistical
standard observed value
test
distribution could be by
(using tables, chance alone
etc.)
Figure 9.2. Statistical testing.
Table 9.1
Some Statistical Techniques Commonly Used in Clinical Research
To test the statistical signifJ'cance of a difference
Chi square (x?) Between two or more proportions (wilen there is a large number
of observations]
Fisher's exact Between two proportions (when there is a small number of
observationsj
Mann-Whitney U Between two medians
Student I Between two means
Ftest Between two or more means
To describe the extent of association
Regression coefficient Between an independent (predictor) variable and a dependent

(outcome) variable
Pearson's r Between two variables
To model the effects of multiple variables
Logistic regression On a dichotomous outcome

Cox proportional hazards On a lime-to-event outcome
of each test depends on certain assumptions about the data. If the data at
hand do not satisfy these assumptions, the resulting p" may be misleading.
A discussion of how these statistical tests are derived and calculated and
of the assumptions on which they rest can be found in any biostatistics
textbook.
Example The chi square (X') test, for nominal data (counts) is more easily
understood than most and so can be used to illustrate how statistical testing
works. Consider the following data from a randomized trial of two ways of
initiating anticoagulation with heparin: a weight-based dosing nomogram
and standard care (4). The outcome was a partial thromboplastin time (PTT)
exceeding the therapeutic threshold within 24 hr of beginning anticoagula-
tion. In the nomogrilm group hO of h2 (97(:~) did so; in the standard care
grO\Jp, 37 of 4,s (77%).
Observed Rates
I 'TT [xceedrng -rhw~h(Jld
Yes No rulHI
Nomogram 60 , 62
Stannard care 3i 11 48
Total 97 13 110
How likely would it be for a study of this size to observe a difference

in rates as great as this or greater if there were in fact no differences in
etfecttvcncss? That depends on how far the observed results depart from
what might have been expected if the treatments were of similar value
and only random variation caused them to differ in the samples studied.
If treatment had no effect on outcome, applying the success rate for the
patients as a whole (8W'!<,) to the number of patients in each treatment
group gives the expected number of successes in each group:
Expected Rates (Rounded to Nearest Integer)
PTI I::xceedinq Threshold
Yes No Total
Nomogram 55 7 62
Standard care 42 6 48
total 97 13 110
The X 2 statistic, which quantitates the difference between the observed

and expected numbers, is the sum for all four ce]]s of:
(Obse~ved number - Expected~umber)2

Expected number
The magnitude of the X 2 statistic is determined by how different all of

the observed numbers are from what would be expected if there were no
treatment effect. Because they are squared, it does not matter whether the
observed rates exceed or fall short of the expected. By dividing the squared
difference in each cell by the expected number, the difference for that cell
is adjusted for the number of patients in that cell.
The X2 statistic for these data is
~(6~Ucc-;c'5~5,-f + (2 - 7)2 + (37 - 42)2 + (11 - 6f = R.79

55 7 42 6
This X2 is then compared to a table relating X2 values to probabilities

(available in books and computer programs) for that number of cells, to
obtain the probability ofaX2 that large or larger. It is intuitively obvious
that the larger the X2 , the less likely chance is to account for the observed
differences.The result in this case is p = 0.004, which is the probability of
a false-positive conclusion that the treatments had different effects.
CONCLUDING THAT A TREATMENT DOES NOT WORK
Some trials come to the conclusion that neither treatment is better than the
other. There are some very influential examples, including studies showing
that coronary artery bypass surgery does not prolong life in patients with
chronic stable angina (except for those with left main coronary artery obstruc-
tion), that antioxidents do not prevent cancer, and that antibodies against
endotoxin do not improve the prognosis of most patients with septic shock.
The question arises, could results like these have occurred by chance
alone? Could the findings of such trials have misrepresented the truth
because these particular studies had the bad luck to turn out in relatively
unlikely ways? Specifically, what is the probability of a false-negative result
(a f3 or Type II error)? The risk of a false-negative result is particularly
large in studies with relatively few patients.
Beta error has received less attention than a error for several reasons.
It is more difficult to calculate. Also, most of us simply prefer things that
work. Negative results arc unwelcome: authors are less likely to submit
negative studies to journals and if negative studies are reported at all, the
authors may prefer to emphasize subgroups of patients in which treatment
differences are found, even if the differences are not statistically significant.
Authors may also emphasize reasons other than chance for why true differ-
ences might have been missed. Whatever the reason for not considering
the probability of f3 error, it is the main question that should be asked
when the results of a study indicate no difference.
The probability that a trial will find a statistically significant difference
when a difference really exists is called the statistical power of the trial.
Statistical power = 1 -- pf3
Power and prJ arecomplell1l'!ltary ways ofe:xpressing the saine concept. Power
is analogous to the sCllsitivity of a diasnostic test. ln [ace. one speaks of u study
being powe/illl if it has a high probability of ddectins as di;ffercnt treatments that
rcally arc differenl.
HOW MANY PATIENTS ARE ENOUGH?
Suppose you are reading about a clinical trial comparing a promising
new therapy to the current form of treatment. You are aware that random
variation can be the source of whatever differences are observed, and you
wonder if the number of patients (sample _~ize) in this study is large enough
to make chance an unlikely explanation for what was found. How many
patients would be necessary to make an adequate comparison of the effects
of the two treatments? The answer depends on four characteristics of the
study: the magnitude of the difference in outcome between treatment
groups, p", pfj, and the nature of the study's data. These are taken into
account when the researcher plans the study and when the reader decides
whether the study has a reasonable chance of giving a useful answer.
Effect Size
Sample size depends on the magnitude of the difference to be detected.
We are free to look for differences of any magnitude, and of course, we
hope to be able to detect even very small differences. But more patients
are needed to detect small differences, everything else being equal. So it
is best to ask only that there is a sufficient number of patients to detect
the smallest degree of improvement that would be clinically meaningful.
On the other hand, if we Me interested in detecting only very large differ-
ences between treated and control groups (t.c.. strong treatment effects),
then fewer patients need be studied.
Alpha Error
Sample size is also related to the risk of an a error (concluding that
treatment is effective when it is not). The acceptable size for a risk of this
kind is a value judgment; the risk could be as large as 1 or as small as O.
If one is prepared to accept the consequences of a large chance of falsely
concluding that the therapy is valuable, one can reach conclusions with
relatively few patients. On the other hand, if one wants to take only a
small risk of being wrong in this way, a larger number of patients will be
required. As we discussed earlier, it is customary to set p, at 0.05 (1 in 20)
or sometimes 0.01 (I in 100).
Beta Error
The chosen risk of a /j error is another determinant of sample size. An
acceptable probability of this error is also a judgment that can be freely
made and changed, to suit individual tastes. Probability of 13 is often set
at 0.20, a 20% chance of missing true differences in a particular study.
Conventional (J errors are much larger than a errors, reflecting the higher
value usually placed on being sure an effect is really present when we say
it is.
Characteristics of the Data
The statistical power of a study is also determined by the nature of the
data. When the outcome is expressed on a nominal scale and so is described
by counts or proportions of events, its statistical power depends on the
rate of events: the larger the number of events, the greater the statistical
power for a given number of people at risk. As Pete et al. (5) put it,
In clinical trials of time to death (or of the time to some other particular
"event" -rdapst', metastasis, first thrombosis, stroke, recurrence, or time to
death from a particular cause), the ability of the trial to distinguish between
the merits of two treatments depends on how many patients die (or suffer a
relevant event), rather than on the number of patients entered. A study of
100 patients, 50 of whom die, is about as sensitive as a study with 1000
patients, 50 of whom die.
If the outcome is a continuous variable, such as blood pressure or serum

cholesterol, power is affected by the degree to which patients vary among
themselves: The greater the variation from patient to patient with respect
to the characteristic being measured, the more difficult it is to be confident
that the observed differences (or lack of difference) between groups is not
because of this variation, rather than a true difference in treatment effects.
In other words, the larger the variation among patients, the lower the
statistical power.
In designing a study, the investigator chooses the size of treatment effect
that is clinically important and the Type I and Type II errors he or she
will accept. It is possible to design studies that maximize power for a given
sample size-e.g., by choosing patients with a high event rate or similar
characteristics-as long as they match the research question. But for a
given data set and question the investigator cannot control the way that
the characteristics of the data determine statistical power.
INTERRELATIONSHIPS
The interrelationships among the four variables discussed above are

summarized in Table 9.2. The variables can be traded off against each
other. Tn general, for any given number of patients in the study there is a
trade-off between a and f3 error. Everything else being equal. the more
Table 9,2
Determinants of Sample Size
uctermoeo by
jnvustqator The Data
V I
N varies according to and - or-
I P
WllElre n
= number or patients studied, .0. = size of difference in outcome between groups;
P" probability of an" [Type I) error, i.e., talse-posltlve results: p" = probability of a {3
=
(Type II) error, te.. talse-negative result; V variability ot observations (tor interval data):
and P proportion ot patients experiencing outcome of interest (for nominal data)
one is willing to accept one kind of error, the less it will be nece~sary to
risk the other, Neither kind of error is Inherently Norse than the other. The
consequences of accepting erroneous information depend on the clinical
situation. When a better treatment is badly needed (c.g., when the disease
is very dangerous and no satisfactory alternative treatment is available)
and the proposed treatment is not dangerous, it would be reasonable to
accept a relatively high risk of concluding a new treatment is effective
when it really is not (large u enol') to minimize the pos~ibility of missing
a valuable treatment (small /3 error). On the other hand, if the disease is
less serious, alternative treatments are available, or the new treatment is
expensive or dangerous, one might want to minimize the risk of accepting
the new treatment when it is not really effective (low IX error), even at the
expense of a relatively large chance of missing an effective treatment (large
/3 error). Tt is of course possible to reduce both a and (i errors if the number
of patients is increased, outcome events are more frequent, variability is
decreased. OJ a larger treatment effect is sought.
For conventional levels of p.. and Pi)' the relationship between the size
of the treatment effect and the number of patients needed for a trial if>
illustrated by the following examples, one representing a situation in which
a relatively small number of patients was sufficient and the other in which
a very large number of patients WdS required.
Example Small sample size: Case series suggest that the nonsteroidal
antiinflammatory drug sulindac is active "gainSl colonic polyps. This possi-
bility was tested in a randomized trial (6). A total of 22 patients with familial
adcnomatous polyposis were randomized; 11 received sulindac and II pla-
cebo. After Y months, patients receiving sulindac had an averaae of 44'1'0
fewer polyps than those receiving placebo. This difference was statistically
significant (p =- 0.(14)- Because of the larK" effect size and tile large number
of polyps per patient (some had more than IOU), few patients were needed
to establish that the effect was beyond cham:". (In lhis arialyxis it was neccs-
sary to assume that treatment affected polyps independently of which patient
tht,y occurred in-c-an unlikely, but probahly not damilging, assumption.)
Example Large sample size: The GUSTO Irial. described above, was de-
signed to include 41,000 patients to have a YO':;, chance of detecting a 15'%.
reduction in mortality or a 1'%, dec-rease in mortahtv rate, whichever was
larger, between the experimental and control treat!ll~nts with a pa of ll.05,
assuming the mortality rate in the control patients was at least R% (2). The
sample stzc had to he so large because a relatively small proportion of pa-
tients experienced the outcome event (death), the effect size was small (L,)'~(,),
and the investigators wanted a relatively high chance of detecting the effect
if it were present (90%).
For most of the therapeutic questions encountered today, a surprisingly
large number of patients is required. The value of dramatic, powerful
treatments, such as insulin for diabetic ketoacidosis or surgery for append i-
citis, could be established by studying a small number of patients. But

such treatments come along rarely and many of them CHI.' already well
established. We arc left with diseases, many of them chronic and with
multiple, interacting causes, for which the effects of new treatments are
generally small. This places special importance on whether the size of
clinical trials is adequate to distinguish real from chance effects.
Clinicians should be able to estimate the power of published studies.
Toward that end, Figure 9.3 shows the relationship between sample size
and treatment difference for several baseline rates for outcome events. It
is apparent that studies involving fewer than 100 patients have a rather
c. 1500 -
~
o
~
e
J:
c
"
W Outcome event
~
.. 1000 - 0.50 0.20 0.05 rate in the

untreated group
..
c.
o
-.
l1.
o
.... 500
.0
E
~
z
20 40 60 80 100
Proportional Reduction in Event Rate
(% of P)
Figure 9.3. The number of people required in each of two treatment groups (of
equal size) to have an 80% chance of detecting a difference (p = 0.05) in a given
outcome rate (P) between treated and untreated patients, for various rates of out-
come events in the untreated group, (Calculated from formula in Weiss NS, Clinical
epidemiology. The study of the outcome of illness, New York: Oxford University
Press, 1986.)
198 CLINICAl.. [PIDEMIOlOGY
poor chance of detecting statistically significant differences of even large

treatment effects. Also, it is difficult to detect dfect sizes of less than 25'\,.
Tn practice, statistical power can be estimated by means of readily available
formulas, tables, nomograms, or computer programs.
Point Estimates and Confidence intervals

The effect size (e.g., treatment dfect in a clinical trial or relative risk in
a cohort study) observed in ,1 particular study is called the poiJlII'slimlllc
of the effect. It is the best estimate from the study of the true effect size
and is the summary statistic usually given the most emphasis in reports
of research.
l lowcvcr, the true effect size is unlikely to be exactly that observed in
the study. Because of random variation, Jny one study is likely to find il
result higher or lower than the true value. Therefore, a summary measure
of the extent of variation that might be expected by chance is needed.
The statistical precision (stability of the estimate) of an observed effe-ct
size is expressed as a (ou(J'dencc intcmal, usually the 9.5'};, confidence interval,
around the point estimate. Confidence intervals around an effect size are
interpreted in the following mariner: if the study is unbiased, there is ,1
95'\,;, chance that the interval includes the true effect size. The narrower
the confidence interval, the more certain one can be about the size of the
true effect. The true value is most likely to be dose to the point estimate,
less likely to be near the outer limits of the interval, and could (5 times
out of 100) fall outside these limits altogether. Statistical precision increases
with the statistical power of the study.
Confidence intervals contain information similar to statistical signifi-
cance. If the value corresponding to no effect (such as a relative risk of 1
or a treatment difference of 0) falls outside the 95'~;) confidence intervals
for the observed effect, it is likely that the results are statistically significant
at the 0.05 level. If the confidence intervals include this point, the results
are not statistically significant.
But confidence intervals have other advilntages. They put the emphasis
where it belongs, on the size of the effect. Confidence intervals allow the
reader to sec the range of plausible values and so to decide whether an
effect size they regard as clinically meaningful is consistent with or ruled
out by the data (7). They also provide information about statistical power;
if the confidence interval barely includes the value corresponding to no
effect and is relatively wide, a signifiGll1t difference might have been found
if the study had had more power.
Example Figure 9.4 illustrates point csurnate, and confidence intervals
for the estimated relative risk of exogenous estrogens for three diseases:
endometrial cancer, breast cancer, and hip frarlurc. (!'-Joticc that the risk is
on a log scale, giving the sUjlc!·licial impression that confidenc-e intervals for
CHAPTtH 9 I CHANCE 199
10 I Endometrial
l' cancer
"
iii
c
(/l
Cl
5 (> 8 years)
o Breast cancer
-
..J
o
~
Risk
Protection
0:
U> 0.5
"C
"C
o
Hip fracture
0.1
Figure 9.4. Point estimates (0) and confidence intervals (1): the risks ami benefits
of exogenous fJstrogens for postmenopausal women. (Data from Grady D, Rubin
8M, Petitti DB, Fox GS, Black lJ, tttinger R, Ernster VL, Cummings SR. Ilormonal
therapy to prevent disease and prolong life in postmenopausal women, Ann Intern
Mod 1992; 117:1016-1037; Colot- GA, Stampter MJ, Willett we, Hennekens CH,
Rosner [3, Speizer FE. Prospective study of estrogen replacement therapy and risk
of breast cancer in postmenopausal women. JAMA 1990;264:2648-2653; and
Paganini-Hill A, nos; RK, Oerkins VR, Henderson t)c, Arthur M, Mack TM Meno-
pausal estrogen therapy and hip fractures. Ann Intern Med 1981 ;95:?8-31.)
the higher risks arc narrower than they really are.] The estimate of risk lor
endometrial cancer (ntlcr 8 or rnure year~ of estrogens) is 8.22, but the true
value i~ not precisely estimated and could easily be as high <JS 10.61 or as
low as 6.2,:;, In any case, it is unlikclv to be as low as 1.0 (no risk). In contrast,
this one study suggests that e~lrog<'ns are unlikely to be a risk factor for
breast cancer; the best estimate of relative risk is nearly 1.0, although the
data Me consistent with either a small harmful or a small protective effect.
Finally, estrogens Me likely to protect against hip fracture. That the upper
boundary of the confidence interval falls below ].0 is another way of indicat-
ing that the protective cffccl is ~l,lli~tically significant at the 0.05 level.
Point estimates and confidence intervals are used to characterize the

statistical precision of any rate (incidence and prevalence], comparisons of
rates (relative and attributable risks), and other summary statistics. Per
example, individual studies have shown that :14% of Ij.S. adults have used
unconventional therapy (95'X, confidence interval 31-37%) (8), that inten-
sive treatment of insulin-dependent diabetes lowers the risk of develop-
ment of retinopathy by 76(~;' (95°/', confidence interval 62-85%) relative to
conventional therapy (9), and that the sensitivity of clinical examination
for splenomegaly is 2n;, (95% confidence interval 19-36')';,) (10).
Confidence intervals have become the usual way of reporting the main
results of clinical research because of their many advantages over the hy-
pothesis testing (p value) approach. The P values are still used because of
tradition and as a convenience when many results are reported and it
would not be feasible to include confidence intervals for all.
Statistical Power before and after a Study Is Done
Ca lculation of statistical power based on the hypothesis testing approach
is done by the researchers before a study is undertaken to ensure that
enough patients will be entered to have a good chance of detecting a
clinically meaningful effect if it is present. However, after the study is
completed this approach is no longer as relevant (11). There is no need to
estimate effect size, outcome event rates, and variability among patients;
they are now known.
Therefore, for researchers who report the results of clinical research
and readers who try to understand their meaning, the confidence interval
approach is more relevant. One's attention should shift from statistical
power for a somewhat arbitrarily chosen effect size, which may be relevant
in the planning stage, to the actual effect size observed in the study and
the statistical precision of that estimate of the true value.
Detecting Rare Events
It is sometimes important to detect a relatively uncommon event (e.g.,
1/1000), particularly if that event is severe, such as aplastic anemia or life-
threatening arrhythmia following a drug. In such circumstances, a great
many people must be observed in order to have a good chance of detecting
even one such event, much less to develop a relatively stable estimate of
its frequency.
Figure 95 shows the probability of detecting an event as a function of
the number of people under observation. A rule of thumb is as follows:
To have a good chance of detecting a l/x event one must observe 3x people
(12). For example, to detect a 1/1 OUO event, one would need to observe
3000 people.
Multiple Comparisons
The statistical conclusions of research have an aura of authority that
defies challenge, particularly by noncxpcrts. But as many skeptics have
suspected, it is possible to "lie with statistics," even if unintentionally.
CHAPTtJ-1 9 CHANCE 201
1 1 1 1
Risk --
100
--
1,000 10,000 100,000
1 .0
~
-
.g
-
u
~
0.8
-
~
C
0.6
0
>-
'"co
s:l
0.4
s:l
0 0.2
~
D.
- I I
1,000 10,000 100,000 1,000,000
Size of Treatment Group
Figure 9.5. The probability of detecting one event according to the rate of the
event and the number of people observed, (From Guess HA, Rudnick SA, Use at
cost effectiveness analysis in planning cancer chemoprophylaxis trials, Control Clin
Trials 1983:4:89-100.)
What is more, this is possible even if the research is well designed, the
mathematics flawless, and the investigators' intentions beyond reproach.
Statistical conclusions can be misleading because the strength of statisti-
cal tests depends on the number of research questions considered in the
study and when those questions were asked. If many comparisons arc
made among the variables in a large set of data, the p value associated
with each individual comparison is an underestimate of how often the
result of that comparison, ilmong the others, is likely to arise by chance.
As implausible as it might seem, the interpretation of the p value from a
single statistical test depends on the context in which it is done.
To understand how this might happen, consider the following example.
Suppose a large study has been done in which there arc multiple subgroups
of patients and many different outcomes. Por instance, it might be a clinical
trial of the value of a treatment for coronary artery disease for which
patients arc in several clinically meaningful groups (c.g., 1-, 2-, and 3-
vessel disease; good and bad ventricular function; the presence or absence
of arrhythmias; and various combinations of these) and several outcomes
are considered (c.g.. death, myocardial infarction, and angina). Suppose
also that there arc no true associations between treatment and outcome in
any of the subgroups and for any of the outcomes. Finally, suppose that
the effects of treatment are assessed separately for each subgroup and for
each outcome-a process that involves a great many comparisons. As
pointed out earlier in this chapter, I in 20 of these comparisons is likely
to be statistically significant at the 0.05 level. In the general case, if 20
comparisons are made, on the average, 1 would be found to be statistically
significant; if 100 comparisons are made, about 5 would be likely to emerge
as significant, and so on. Thus, when a great many comparisons have been
made, a few will be found that are unusual enough, because of random
variation, to exceed the level of statistical significance even i r no true associ-
ations between variables exist in nature. The more comparisons that are
made, the more likely that one of them will be found statistically
significant.
This phenomenon is referred to as the multiple comparisons problem.
Because of this problem, the strength of evidence from clinical research
depends on how focused its questions were at the outset.
Unfortunately, when the results of research are presented, it is not al-
way~ possible to know how many comparisons really were made. Often,
interesting findings arc selected from a larger number of uninteresting
ones. This process of deciding what is and is not important about a mass
of data can introduce considerable distortion of reality.
How can the statistical effects of multiple comparisons be taken into
account when interpreting research? Although ways of adjusting p., have
been proposed, probably the best advice is to be aware of the problem
and to be cautious about accepting positive conclusions of studies where
multiple comparisons were made. As one statistician (13) put it:
If you dredge the data sufficiently deeply and sufficiently often, you will
find something odd. Many of these bizarre findings will be due to chance. I
do not imply thai data dredging is not an occupation for honorable persons,
but rather that discoveries that were not initially postulated as among the
major objectives of the [rial should be treated with extreme caution. Statistical
theory may in due course show us how to allow for such incidental findings.
At present, I think the best attitude to adopt is caution, coupled with an
attempt to confirm or refute the findings by further studies.
An approach to assessing the validity of statistically significant findings
in subgroups was presented in Chapter 7.
Describing Associations
Statistics are also used to describe the degree of association between
variables, e.g., the relationship between body mass and blood pressure.
Familiar expressions of association are Pearson's product moment correla-
tion (r) for interval data and Spearman's rank correlation for ordinal data.
Each of these statistics expre~ses in quantitative terms the extent to which
the value of one variable is associated with the value of another. Each has
a corresponding statistical test to assess whether the observed association

is greater than might have arisen by chance alone.
Multivariable Methods
Most clinical outcomes are the result of many variables acting together
in complex way,;. For example, coronary heart disease is the joint result
of lipid abnormalities, hypertension, cigarette smoking, family history, dia-
betes, exercise, and perhaps personality. It is appropriate first to try to
understand these relationships by examining relatively simple arrange-
ments of the data, such as 2-by-2 tables (for one variable at a time) or
contingency tables (stratified analyses, examining whether the effect of
one variable is changed by the presence or absence of one or more other
variables), because it is easy to understand the data when they arc dis-
played in this way. However, it is usually not possible to account for more
than a few variables using this method, because there arc not enough
patients with each combination of characteristics to allow stable estimates
of rates. For example, if 120 patients were studied, 60 in each treatment
group, and just one additional dichotomous variables were taken into
account, there would only be at most about 15 patients in each subgroup;
if patients were unevenly divided, there would be fewer in some.
What is needed then, in addition to contingency tables, is a way of
examining the effects of several variables at a time. This is accomplished
by multivariable modeling, developing a mathematical expression of the ef-
fects of many variables taken together. It is "multi variable" because it
examines the effects of multiple variables simultaneously. It is "modeling"
because it is a mathematical construct, calculated from the data but also
based on simplifying assumptions about characteristics of the data (e.g.,
that the variables arc all normally distributed and have the same variance).
Mathematical models can be used in studies of cause, when one wants
to define the independent effect of one variable by adjusting for the effects
of several other, extraneous variables. They can also be used to give more
precise predictions than individual variables allow by including several
variables together in a predictive model.
The basic structure of a multivariable model is
Outcome variable = constant + (/31 X variable.] + (/32 X vanable.) + .
where /31, /32< . arc coefficients that are determined by the data; and
variable I, varlable-, . . are the predictor variables that might be related
to outcome. The best estimates of the coefficients are determined mathe-
matically, depending on the powerful calculating ability of modern
computers.
Modeling involves several steps.

• Identify and measure all the variables that might be related to the out-
come of interest.
• Reduce the number of variables to be considered in the model to a
manageable number, usually no more than several. Often this is done
by selecting variables that are, when taken one at a time, most strongly
related to outcome. If a statistical criterion is used at this stage, it is usual
to err on the side of including variables, e.g., by choosing all variables
showing an association with the outcome of interest at a cutoff level of
p < 0.10. Evidence for the biologic importance of the variable is also
considered in making the selection.
• Some variables may be strongly related to each other. If so, only one is
included since both contain about the same information.
• The remaining variables are entered in the model, with the strategy for
the order in which they aIC tried determined by the research question.
For example, if some are to be controlled for in a causal analysis, they
are entered in the model first, followed by the variable of primary inter-
est. The model will then identify the independent effect of the main
variable. On the other hand, if the investigator wants to make a predic-
tion based on several variables, the variables can be entered in order of
the strength of their association to the outcome variable, as determined
by the model.
Modeling is now a regular feature of the medical literature, appearing
in about 18% of articles in major journals (14) and in nearly all large studies
of cause. Some commonly used kinds of the models are logistic regression
(for dichotomous outcome variables such as occur in case-control studies)
and Cox proportioned hazards models (for time-to-event studies).
Multivariable modeling is an essential part of many clinical studies;
there is no other way to adjust for or to include many variables at the
same time. However. this advantage comes at a price. Models tend to be
black boxes, and it is difficult to "get inside" them and understand how
they work. Their validity is based on assumptions about the data that may
not be met. They are clumsy at recognizing effect modification (different
effects in different subgroups of patients). An exposure variable may be
strongly related to outcome yet not appear in the model because it occurs
rarely-and there is little direct information on the statistical power of the
model for that variable. Final1y, model results are easily affected by quirks
In the datil, the results of random variation in the characteristics of patients
from sample to sample. It has been shown, for example, that a model
frequently identified a different set of predictor variables and produced a
different ordering of variables on different random samples of the same
data set (15). To protect against this possibility, a rule of thumb is that
there should be at least 10 outcome events for each predictor variable in

the model.
for these reasons, the models themselves cannot be taken as a standard
of validity; they must be independently validated. Commonly, this is done
by seeing if the model predicts what is found in another, independent
sample of patients (see Chapter 12). The results of the first model are
considered a hypothesis, to be tested by new data. If random variation is
mainly responsible for the results of the first model, it is unlikely that the
same random effects will occur in the validating data set too. Other evi-
dence for the validity of a model is its biologic plausibility and its consis-
tency with simpler, mOTC transparent analyses of the data such as stratified
analyses.
Summary
Clinical information is based on observations made on samples of pa-
tients. Even samples that are selected without bias may misrepresent events
in a larger population of such patients because of random variation in its
members.
Two general approaches to assessing the role of chance in clinical obser-
vations arc hypothesis testing and estimation. With the hypothesis testing
approach, statistical tests are used to estimate the probability that the ob-
served result was by chance. When two treatments arc compared, there
are two ·ways in which the conclusions of the trial can be WTOng: The
treatments may be no different, and it is concluded one is better; or one
treatment may be better, and it is concluded there is no difference. The
probabilities that these errors will occur in a given situation arc called pI)
and p~, respectively.
The power of a statistical test (1 - fill) is the probability of finding a
statistically significant difference when a difference of a given size really
exists. Statistical power is related to the number of patients in the trial,
size of the treatment effect, J!", and the rate of outcome events or variability
of responses among patients. Everything else being equal, power can be
increased by increasing the number of patients in a trial, but that is not
always feasible.
Estimation involves using the data to define the range of values that is
likely to include the true effect size. Point estimates (the observed effects)
and confidence intervals are used. This approach has many advantages
over hypothesis testing: It emphasizes effect size, not II value; indicates the
range of plausible values for the effect, which the user can relate to clini-
cally meaningful effects; and provides information about power.
Individual studies run an Increased risk of reporting a false-positive
result if many subsets of the data are compared; they are at increased risk
206 CLINICAL ErlDFMIOLOGY
of a false-negative result if they lack statistical power, usually because they

include too few patients or outcome events are uncommon.
Most clinical studies concern the effects of multiple interacting variables.
With multivariablc modeling, it is possible to take all into account simulta-
neously, either to control for extraneous variables in a causal study or to
provide a stronger prediction than would be possible by including nne
variable at a time. However, these models must be interpreted with caution
because their inner workings arc relatively inaccessible, they are sensitive
to random variation, and they are based on assumptions that may not
be met.
R[·:FEREl\'CES
1, Johnson AF. Beneath the technological fix. Outliers and probability stclkments. J Chmn
Dis 1Y85;38:957-96L
2. The GUSTO Investigators. An international randomized tri,,1 comparing fpur thrombolytic
strategies for acute myocardial infJrdion. New Engl J Mcd 1<f93;329:673-liIi2.
3. Parkouh ME, LangJD, Sackett lJl Thrombolytic agents: till' science of the art of choosing
the better treatment. Ann Intern Med 1994;120:886-88R
4, Raschke RA, Reilly 8M, GUidry JR, Fontana JR, Srinivus ,;, The weight-based heparin
dosing nomogram compared with" "standard care" nomof\'ram. A randomized controlled
trial. Ann Intern Med 1993;1t9:874-·Rfll
5. Peto R, Pike MC, Armitage P, Breslow ;\;E, Cox DR, Howard SV, Mantel N, Mcl'hcrson
K, Peto J, Smith Pc. Design and annlysis of randol11i7ed clinical trials requiring: prolonged
observation of each patient l. Introduction and dpsign, Br J emu"r 1976;34:51i5-612,
6. Giardiel10 Fl'vl, Hamilton SR, Krush AJ, Piantadosi S, Hylind 1.\1, Celano P, Booker SV,
Robinson CR, OfferhJus GJA. Tr<'cltm~nl of colonic and rectal adenomas \vith sulindac in
familial adenomatous polyposis. New Eng! J MdJ 1993;328:1313-1316.
7, Braitman LE. Confidence intervals assess both clinical significance and statistical signifi-
cance. Ann Intern Med 1991;114:515-517.
8. Eisenberg OM, Kessler RC, foster C, No-lock FE, Cillkins OK Delbaucc TL. Unconven-
tional medicine in the United StJte~. Prevalence, costs and piltterns of liS". New Engl J
Med 1993;328:246·-252.
9, LJiabetes Control and Complications Trial Research Croup. The effect of intensive treot-
ment of diabetes on the d"vdopl1lent anu progression of long-term complications in
insulin-dependent diabetes mellitis New lingl J Med 199:1;329:977-'186.
10. Grover SA, Barkun A"-;, Sackett LJI.. Does the pcltient have spl<'nomegaly? JAMA
1993;270:2218- 2221.
11 Goodm"n Si\, Berlin JA. The usc of predicted confidence intervals when planning experi-
ments and the misuse of power when interpreting results. Ann l"lPm IvIed 19',14; l2l :200-
206.
12. Sackett DL, llaynes RI3, Gent \1, T"ylor DVV, COl11pli"JKC. I,,: Inm,ln \-VIIW, "d. Monitoring
lor drug safety, I.Jncasler, L'K: MTP Press,IY8il.
13 Armitage 1', Importance of prognostic f"ctors in the ;m"ly~is of d"t" from clinical trials,
Control Clin Triab 1'J81;I:J47-J5J.
14. ("oneato J, reinstl'in AR, Ilolford TR. The risk of determilling risk with mult'variabk-
JIlodds. Ann Intern Med l'N:1;llS:2111-2W,
I.~ Diamond C.-\. Future iJllpl'rf<'ct: the limit"tipns or clinical prediction [llo<!"b and thl' limits
of clinical pr"dktion,1 Am Coil Cardiel 19S9;14:12/\ 221\
CHAPTER 9 ! CHANCF 207
SUGGESTElJ l':EAlJINCS
Altman DC, Gore 511.1, GJrdner MJ, Pocock SJ. Statisti,,,l guidelines ~or contributors 10 medical
jOllrnJIs, n- Med J 1983;286;1489·-1493,
Ilailer Je Ill, Mostdler r, cds..Medical uses of statistics. Waltham, MA: New England ]ournJI
of Medicine Books, 1986.
Cupples LA, Heeren T, Schatzkin A, Colton T. Multiple kbting uf hypotheses in comparing
two groups. Ann Intern Med 1984; 100:122-129.
Detsky AS, Sackett 01 .. When WclS" "negJtive" clinical trial big enough? How many patients
you net'd dep"nd~ on whclt you found, Arch Intern Mcd 1985;145:70'J-712,
facts, figurc>, and fallacies ,cries
Clrpenkr LM, Is Ihe study worth doing? Lancet 1993;343:221-223
Dalla M, You cannot exclude the explanJlion you haven't considered. Lmcet 1993;342:345-
347
Clyrl1l JI\. A question of attribution, Lancet 1993;342:530-532.
Crisso JA. Making comparisons. Lancet 1993;342:157-160.
Jolley T.TIle glitter of the I table IJncel 1993;342:27-29.
Leon D. Failed or mi~leading 'ldJustment for confounding. Lancet 1993;342:479-481.
Mertens TF. Estimating the effccts of nusclassification. Lancet 1'J'J3;342:418-421.
Sitthi-amom C l'oshachinda V. Bias. Lancet 1993;343:286-288.
Victoria CG. what's the denominator? Lancet 1993;342:97-99.
Friedman GD. Primer of epidemiology. 4th ed. New York: McCmw·Hill, 1994.
Gardner Mj. Altman DG. Stilti~tks with ("(lIlfidence confidcnce intervJls and statistical
guidelines. LonJon: 13r>'1J l3ooKs, 1989,
Coodman SN, lJerlin JA, The usc of predicted confidence intervals when pbnning expnim'mts
and the misuse of power when inte'lweting r~SIlIts. Ann Intern Med '1994;'12'1:2(KI-206,
Hilnky [A, Lil'mcln.IIJnd A. If nothing goes wrong is everything right? Interpreting zero
nUlllerators. JA1I.1'\ 1'J~3:24'J:1743.
Hennekens CH, Bllring IF. Fpid~miology in medicine. Boston: Little, Brown & Co., 1987,
lngelflnger JA. r>-l%tdkr F. Thihodeau I,A, VI/Me Jll, biostatistics in clinical medicine, New
York: Ma"milklll, 1983,
Moses 1£ Statistical concepts fundamental to investigations N Engl J Med 1985;312:890··H97.
Ricgclman I{K, Hirsch RP Studying and stLldy ~nd testing a tebt. 2nd "d. Boston; Little,
Brown & C». 1989,
Rothman KI A bhow of confidence. N Fllgi J MedI97i'l;299:B62 1361.
Young l\ilJ, Bresnitz EA, Strom ilL. Sample size nomograms for interpreting negative clinic,,1
studie~, /\nn Intern Mcd 1'JH3;9'J;24B-251.
10
STUDYING
CASES
Eacii case lias its lesson-a lesson which may be but is not always
learned.
-Sir William Osler
Most medical knowledge has emanated from the intensive study of sick
patients. The exhausted but engrossed physician at the bedside of the
febrile child, chin in hand, is a favorite medical image. The presentation
and discussion of a "case" is the foundation of modern medical education.
Most clinicopathologic conferences and grand rounds begin with the pre-
sentation of an interesting case and then use the case to illustrate general
principles and relationships. So, too, much of the medical literature is
devoted to studying cases, whether narrative descriptions of a handful of
cases (case reports), quantitative analyses of larger groups of patients (case
series), or comparisons of groups of cases with noncascs (case control
studies).
Case Reports
Case reporls Me detailed presentations of a single case or a handful of
cases. They represent an important way in which new or unfamiliar dis-
eases, or manifestations or associations of disease Me brought to the atten-
tion of the medical community. Approximately 20-30~;, of the original
articles published in major general medical journals are studies of 10 or
fewer patients.
USES OF CASE REPORTS
Case reports serve several different purposes. First, they arc virtually
our only means of describing rare clinical events. Therefore, they Me a rich
source of ideas (hypotheses) about disease presentation, risk, prognosis,
and treatment. Case reports rarely can be used to test these hypotheses,
but they do place issues before the medical community and often trigger
208
CHAPTeR 10 ! STUDYING CASES 209
more decisive studies. Some conditions that were first recognized through
case reports include birth defects from thalidomide, fetal alcohol syndrome,
toxic shock syndrome, Lyme disease, and HANTA virus infection.
Case reports also serve to elucidate the mechanisms of disease and
treatment by reporting highly detailed and methodologically sophisticated
clinical and laboratory studies of a patient or small group of patients. In
this instance, the complexity, cost, and often experimental nature of the
investigations limit their application to small numbers of patients. Such
studies have contributed a great deal to our understanding of the genetic,
metabolic, and physiologic basis of human diseases. These studies repre-
sent the bridge between laboratory research and clinical research and have
a well-established place in the annals of medical progress.
The following is an example of how a report of a single case can reveal
a great deal about the mechanism of a disease.
Example The anesthetic halothane was suspected of causing hepatitis.

However, because the frequency of hepatitis after exposure to halothane was
low and there were many other causes of hepatitis after surgery, "halothane
hepatitis" was controversial.
Experience with a single individual helped clarify the problem (I). An
anesthetist was found to have recurrent hepatitis, leading to cirrhosis. Attacks
of hepatitis regularly occurred within hours of his rotum to work. When he
was exposed to small doses of halothane under experimental conditions,
his hepatitis recurred and was well documented by clinical observations.
biochemical tests, and liver histology.
Because of this unusual case, it was clear that halothane can cause hepa-
titis. But the case report provided no information as to w hcthcr this reaction
was rare or common. Subsequent studies showed that it was not a rare
reaction, which contributed to the replacement of halothane with less hepa-
totoxic agents.
Another use of the case report is to describe unusual manifestations of
disease. Sometimes this can become the medical version of Ripley'S Beueoe
l! or Not, an informal compendium of medical oddities, with the interest
lying in the sheer unbclicvability of the case. The larger the lesion and the
more outrageous the foreign body, the more likely a case report is to find
its way into the literature. Oddities that are simply bizarre aberrations
from the usual course of events may titillate, but usually are less clinically
important than other types of studies.
Some so-caned oddities are, however, are the result of a fresher, more
insightful look at a problem and prove to be the first evidence of a subse-
quently useful finding. The problem for the reader is how to distinguish
between the freak and the fresh insight. There are no rules. When all else
fails, one can only rely on common sense and a well-developed sense of
skepticism.
210 CLINICAL EPIDEMIOI_OGY
BIASED REPORTING
Because case reports involve a small and highly selected group of pa-
tients, they are particularly susceptible to bias. For example, case reports
of successful therapy may be misleading because journals are unlikely to
receive or publish case reports of unsuccessful therapy. Perhaps the wisest
stance to take when reviewing a case report is to use it as ,1 signal to look
for further evidence of the described phenomenon in the literature or
among your patients.
Example A case report (2) described a 23-year-old woman who devol-
oped severe abdominal pain while on treatment with enalapril for essential
hypertension. An elevated serum lipase led to a diagnosis of pancreatitis.
Symptoms resolved, and the lipase returned to normal shortly after discontin-
uing the drug. The investigators found only one published case and began
an exhaustive search of the published and unpublished literature. The search
revealed an additional flO cases, the majority of which were unpublished
cases reported to the drug manufacturer. The additional cases lent strength
to the possibility of a causal association between enalapril treatment and
pancreatitis.
With very few exceptions, case reports on their own should not serve
as the basis for altering clinical practice because of their inability to estimate
the frequency of the described occurrence or the role of bias or chance.
THE JOINT OCCURRENCE OF RARE EVENTS
Case reports often describe the joint occurrence of uncommon events,

particularly if the observed association lends itself to an interesting biologic
explanation. But even rare events occur together by chance alone; simply
observing this occurrence does not mean they are biologically related. As
one author (3) put it, "In a large population the issue is not whether rare
events occur, but whether they occur more frequently than expected by
chance."
Table 10.1 illustrates how often two relatively uncommon conditions-
end-stage renal failure and use of a specific nonsteroidal antiinflammatory
drug-might occur together by chance alone. If there were no biologic
association between the two (and, as discussed later in this chapter, there
may well be such an association), then the probability that they would
occur together is the product of their separate frequencies. In the United
States alone, 100 cases would occur annually, more than enough to spawn
several case reports.
There an.' also reasons why such cases might be seen in medical centers
and be reported in the literature out of proportion to their frequency in
the population at large. Patients with two severe diseases might be more
likely to come to hospitals than those with either disease alone, simply
because they are sicker. It has also been shown that two diseases not
CHAPTER 10 / STUDYING CASES 211
Table 10.1
The Joint Occurrence of Two Rare Conditions: An Estimate of the Frequency and
Number of Cases of Exposure to a Nonsteroidal Antiinflammatory Drug and End-
Stage Renal Failure Occurring Together if the Two Were Not Biologically Related B
Frequency separately
Prevalence of use of the drug (hypothetical) 1/100 persons
Incidence of eno-stoqo ronal osccsc 40/1.000,000/year
ocoocc of joint occurrence t/lOO x 40/1,000,000/year
= 4/1 O,OOO,OOO/year
Population of the United States 250.000,000

Cases in the United States '1/10,OOO,OOO/year x 250,rXJO,OOO
100/year
., Data from Hiatt RA, Friedman GD, Characteristics 01patients referred for treatment 01end-stage renal disease
'n a defined population, Am J Public I iealth 1911?: 7:::':8?9-8,'3,').
associated in the general population can be associated in hospitals if pa-

tients with two diseases are admitted at different rates (4). Moreover, pa-
tients with two diseases are more interesting and so are more likely to be
written up in articles, submitted to journals, and accepted for publication.
Therefore, one should be skeptical about reports of association that arc
based on case reports only. They are simply hypotheses to be tested by
stronger methods before being believed.
Case Series
i\ case series is a study of a larger group of patients (c.g.. 10 or more) with
a particular disease. The larger number of cases allows the investigator to
assess the play of chance, and p values and other statistics often appear in
case series, unlike in case reports. A case series is a particularly common
way of delineating the clinical picture of a disease and serves this purpose
well-but with important limitations.
Case series suffer from the absence of a comparison group. Occasionally,
this is not a major problem.
Example Between june J981 and February 1983, a few years after A]])S
was first recognized and while its manifestations were being defined, re-
searchers from the Centers for Disease Control gathered information on 1000
patients living in the United States who met a surveillance definition for
the disease. They described demographic and behavioral characteristics of
patients and complications of the disease.
Pnellnlocystis carinii pneumonia (PCP) was found in smo, Kaposi's sarcoma
in 28%, and both in 8% of patients: 14% had opportunistic infections other
than PCP. All bur e'x, of the patients could be classified into one or more of the
following groups: homosexual or bisexual men, intravenous drug abusers,
Haitian natives, and patients with hemophilia (.'i).
212 CLINICAL EPID[MIOlOGY
This report includes no comparison group of people without AIDS.

Also, the definition of cases excluded some patients who have AIDS by
later standards. Nevertheless, because the complications arc so uncommon
in otherwise well people and the pattern of at risk groups so striking, the
report clarified our view of ATDS and set the stage for more detailed studies
of its manifestations and risk factors.
On the other hand, often a relatively frequent association and the ab-
sence of a comparison group have led to erroneous conclusions.
Example Manv physici,lIls attribute low back pain to protrusion of one

or more intervertebral disks. Several case series used magnetic resonance
imaging (MI\I) to define the anatomy of the lumbosacral spine in patients
with low back pain. Th~'se studies found that the majority of patients had
disk abnormalities, providing- apparent support for the importance of disk
abnormalities in low back pain. However, as described in Chapter 3, MI<.I
studies of asymptomatic individuals revealed similar prevalences of disk
abnormalities, undermining the argument that protruding disks seen on MRT
arc the cause of back pain (6).
Another limitation of case series is that they generally describe the clini-
cal manifestations of disease and its treatments in a group of patients
assembled at one point in time, a survival cohort (see Chapter 6). They
must be distinguished, therefore, from cohort studies or trials of treatment
for which <In inception cohort of patients with a disease is followed over
time with the purpose of looking for the outcomes of the disease. Case
series often look backward in time and that restricts their value as a means
of studying prognosis or causr-and-cffect relationships.
Case-Control Studies
To find out whether a finding or possible cause really is more common
in patients with a given disease, one needs a study with several features.
First and foremost, in addition to a series of cases there must be <I compari-
son group that does not have the disease. Second, there must be enough
people in the study so that cbance is less likely to playa large part in the
observed results. Third, the groups must be similar enough, even though
one is nondiscascd, to produce a credible comparison. Finally, if one wants
to show that a risk factor is independent of others-and, therefore, a
possible c<luse- it is necessary to control for a II other important differences
in the analysis of the findings.
Case reports and case series cannot take us this far. Neither can cohort
studies in many situations, because it is not feasible to accrue enough cases
to rule out the pl<lY of chance. Case-control studies, studies that compare
the frequency of a purported risk factor (generally called the "exposure")
in a group of cases and it group of controls, have these features.
CHAPTER 10 I STUDYING CASES 213
Exposure to
Risk Factor Uisease
l\'jj 1-
INo 1 1- - - - -
iYes 1
j\lJl~ .I
Figure 10.1. The design of case-control studies.
DESIGN
The basic design of a case-control study is diagrammed in Figure 10.1.
Patients who have the disease and a group of otherwise similar people who
do not have the disease are selected. The researchers then look backward in
time to determine the frequency of exposure in the two groups. These data
can be used to estimate the relative risk of disease related to the characteris-
tic of interest.
Example Does the use of nonsteriodal antiinflammatory drugs (NSAJDs)
Increase the risk of renal disease? Researchers have addressed this question
using a case-control study (7). How did they go about it?
First, they had to define renal disease and find a sizable group of cases
available to he interviewed. For obvious reasons, they looked in tertiary care
hospitals, where many such cases are gathered. The cases, of course, included
only patients in whom the diagnosis had been made in the course of usual
medical care. For example, asymptomatic patients with mild renal failure
were much less likelv to be included among the cases.
Once the cases were assembled and the diagnosis confirmed, a compari-
son, or control,' group was selected. Before deciding which people to choose
as controls, the investigators considered the purpose of the study. They
wanted to ascertain whether patients with renal failure were more likely to
have received NSAJD therapy in the past than a similar group of people with
no evidence of renal disease.
The investigators found that thc csfirnatcd relative risk of NSAID exposure
for renal failure was 2.1, using data on the rates of exposure in cases and
controls, and that the excess risk was largely confined to older men.
What is meant by similar? There is some controversy about this. In a

cohort study of the risk of NSI\IDs for renal disease, similarity would
mean membership in the cohort from which the cases arose, all of whom
were initially free of renal disease at the inception of the study, e.g., people
residing in the same community or enrolled in the same HMO. Is there a
natural cohort from which a group of cases receiving care at a given tertiary
care hospital can emerge?
Because of referral practices. cases assembled at hospitals and other
treatment centers usually reside in many communities, receive their care
from many physicians, and belong to no common group before becoming
ill. Therefore, there was no obviously similar group of people without
renal disease, and one had to be created.
This was done by randomly sampling people who resided in the vicinity
of each hospital. Tn this way, controls were assembled who, it was hoped,
would provide an accurate estimate of the likely prevalence of NSAllJ usc
ilmong the cases if there were no association between renal disease and
the use of the drugs (8).
Once the cases and controls were selected and their consent obtained,
the next step was to measure exposure to the risk factor of interest. The
drug-taking history of each case and each control had to be reconstructed.
As opposed to a cohort study where drug taking can be tracked over time,
assessment of drug exposure in this case-control study relied on memory_
It is often the past that is important in case-contra] studies, and therein
lies a potential for bias. it is difficult not to interpret the past in the light
of nne's present condition. For cases, this is particularly so when the present
includes a disease as serious as renal failure. Investigators can attempt to
avoid bias by using objective data such as computerized pharmacy records,
blinding subjects to the purpose of the study, blinding observers to case
status if possible, and by using carefully defined criteria to decide which
of the cases and controls received prior NSAIO therapy.
COHORT VERSUS CASE-CONTROL RESEARCH
Cohort and case-control studies arc both observational studies of risk

factors. Sometimes the two are confused. A distinguishing feature of the
case-control design is that cases have the outcome of interest at the time
that information on risk factors is sought. In cohort research, on the other
hand, people are free of disease at the beginning of observation when
the measurement of the risk factors is made. Figure 10.2 summarizes the
differences between case-control ,1Od cohort designs. Since the temporal
relationship between putative cause and effect is ,10 important criterion
for causality (sec Chapter 11), cohort studies provide ,1 stronger basis for
a causal interpretation.
Table 10.2 summarizes the essential characteristics of cohort, case-
CIIAPTER 10 / STUDYING CASES 215
CASE-CONTROL STUDV
Exposure to
NSAIDs Renal Failure
•Yes" I
,-NO]
['les 1
<R ES E AR C HI
I No I
COHORT STUDV
Exposure to Renal Failure

NSAIDs
YES
Population
'---- ---j[ Ves I

NO ,r'jgl
Figure 10.2 A comparison of case-control and cohort studies. Studies of NSAIDs
as a risk factor for retial failure.
control, and prevalence research designs and illustrates their differences.

As will be discussed later, it is these differences that make the case-control
study particularly susceptible to bias
THE ODDS RATIO
How do we decide whether there is an increased risk? Figure 10.3 shows
the calculation of risk for cohort and case-control studies. In a cohort study,
the susceptible population is divided into two groups-exposed to
NS/\lDs (/\ + B) and unexposed (C t D)-at the outset. Cases of renal
disease emerge naturally over time in the exposed group (A) and the
"l
m
o
c
z
~
q:
Table 10.2
Summary of Characteristics of Cohort, Case-Control, and Prevalence Designs
em
ColKlrt Case-Control Prevalence
"o
Begins with a defined population at risk
Cases not selected but ascertained by

Population at risk often undefined
Cases selected by investigator from 00

Begins with a defined population
Cases not selected buf ascertained by a single

B
continuous surveillance {presumably all available pool of patients examination of the population
cases)
Controls, the comparison group (i.e.. Controls selected by investigator to resemble rconcases include those free at disease at the
nonc.."ISes), not selected-evolve naturally cases single examination
Exposure measured before the development Exposure measured, reconstructed, or Exposure measured, reconstructed. or
of disease recollected after development of disease recollected after development of disease
Risk or incidence of disease and relative risk Risk or incidence 01 disease cannot be Risk or incidence of disease cannot be
measured directly measured directly: relatrve nsk of exposure measured directly: relative risk of exposure
can be estimated by the odds ratio can be estimated by the odds ratio
CHAPTER 10 ! STUDYING CASF'S 217
Cases Noncases
..............
Exposed A i
Not )
exposed D
.....
A+C B+D
Cohort Study Case-control Stu~r

Relative risk == Odds ratio ==
AI (A +B) A/(A+C)
C I (C + D) C/(A+C) A/C AD
B I (B + D) = BID = BC
D/(B+ D)
Figure 10.3 Calculation at relative risk for a cohort study and odds ratio (estimated
relative risk) for a case-control study.
unexposed group (e). This provides us with appropriate numerators and

denominators to calculate the incidences of renal disease in the exposed
[A/(Al B)] and unexposed [C/(C + D)] cohorts. It is also possible to
calculate the relative risk.
Incidence of disease in the exposed A/(A + B)

Relative risk =
Incidence of disease in the unexposed C/(C + lJ)
Case-control studies, on the other hand, begin with the selection of a
group of cases of renal disease (A-I C) and another group of controls
(B + 0). There is no way of knowing disease rates because these groups
are determined not by nature but by the investigators' selection criteria.
Therefore, an incidence rate of disease among those exposed to NSAIOs
and those not exposed cannot be computed. Consequently, it is not possible
to obtain a re-lative risk by dividing incidence among users by incidence
among nonusers. What does have meaning, however, are the relative fre-
quencies of people exposed to NSAIDs among the cases and controls.
It has been demonstrated that one approach for comparing the fre-
quency of exposure among cases and controls provides a measure of risk
that is conceptually and mathematically similar to the relative risk. This is

the odds ratio, defined as the odds- that a case is exposed divided by the
odds that a control is exposed
[AliA + C) 7 C/IA + C)[
[B/IB + OJ O/(B + 0)1
The odds ratio simplifies to
A/C AD
BID BC
As is seen in Figure 10.3, the odds ratio can be obtained by multiplying
diagonally across the table and then dividing these cross-products.
Note that if the frequency of exposure is higher among cases, the odds
ratio win exceed 1, indicating increased risk. Thus the stronger the associa-
tion between the exposure and disease, the higher the odds ratio. Con-
versely, if the frequency of exposure is lower among cases, the odds ratio
will be less than 1, indicating protection. The meaning of the odds ratio,
therefore, is analogous to the relative risk obtained from cohort studies.
The similarity of the information conveyed by the odds ratio and the rela-
tive risk has led some investigators to report odds ratios as "estimated
relative risks" or simply "relative risks."
The odds ratio is approximately equal to the relative risk only when
the incidence of disease is low, because of assumptions that must be made
in the calculations. How low must the rates be? The answer depends in
part on the size of the relative risk (9). In general, however, distortion of
the relative risk becomes large enough to matter at disease rates in unex-
posed people of greater than about 1/100. Fortunately, most diseases, par-
ticularly those examined by means of case-control studies, are considerably
less common than that rate.
ADVANTAGES OF CASE-CONTROL STUDIES
The case-control design has become a common and important method
used to study etiology and clinical questions. What are its advantages?
first, the investigators can identify cases unconstrained by the natural
frequency of disease and yet can still make a comparison. Cohort studies
<Ire quite inefficient for this purpose. For example, to gather information
about the risk of NSAID use in 100 individuals with end-stage renal dis-
ease, one would have to follow a cohort of 1,000,000 for about 2 1/ 2 years
(see Table 10.1). Obviously, because of the expense and logistic difficulties
of such a study, it would usually not be feasible. In contrast, it has been
relatively inexpensive and easy to assemble 100 or more cases from hasp i-
'For a l'cmindcl' of whal odd, mca"5, scc pagc O4.

tals and other treatment facilities, find similar groups without the disease,
and compare frequencies of past NSAID use. In this way, several hundred
study subjects can be interviewed in a matter of weeks or months, and an
answer can be obtained at a fraction of the cost of a cohort study.
A real advantage of the case-control study in exploring the effect of
some causal or prognostic factors is that one need not wait for a long time
for the answer. Many diseases have a long latency-the period of time
between exposure to a risk factor and the expression of its pathologic
effects. For example, it has been estimated that at least 10-20 years must
pass before the carcinogenicity of various chemicals becomes manifest. It
would require an extremely patient investigator and scientific community
to wait so many years to see if a suspected risk to health can be confirmed.
Because of their ability to address important questions rapidly and effi-
ciently, case-control studies play an increasingly prominent role in the
medical literature. If one wants to study cause and effect using a relatively
strong method, the case-control approach is the only practical way to study
some diseases. Case-control studies comprise a growing percentage of all
original articles and the majority of epidemiologic articles. Their quickness
and cheapness justify their popularity as long as their results are valid;
and here is the problem, because case-control studies are particularly prone
to biased results. These biases are discussed in the next section.
Avoiding Bias in Case-Control Studies

In many case-control studies, the investigators create the comparison
groups rather than allow nature to determine who in a population becomes
a case and who remains a noncase or control as in cohort or prevalence
studies. This element of manipulation is a necessary evil because the valid-
ity of a case-control study depends on the comparability of cases and
controls.
Cases and controls are comparable if the controls would have been
captured as cases if they developed the condition under study. In other
words, to be comparable, cases and controls must be members of the same
base population. A second, more controversial issue is whether to be com-
parable, cases and controls must have an equal opportunity to receive the
exposure (10). For example, the opportunity to have received NSAIDs
(discussed earlier) would presumably be greater among those who have
received regular medical care and perhaps still greater among those with
joint symptoms. Should both cases and controls have similar symptoms
and medical care experiences. Opinions differ, but it is clear that if one
insists that cases and controls have the same degree of arthritic symptoms
and the same doctor, the opportunity to evaluate risk may be impossible
if the doctors involved tend to either prescribe or not prescribe NSAIDs
to most of their patients with common causes of musculoskeletal pain.
Therefore, ensuring comparability between cases and controls requires

careful consideration of the circumstances under which an individual be-
comes exposed.
SELECTING CASES
In the past, most case groups in case-control studies were assembled
from among patients receiving care in hospitals or other medical treatment
facilities. The proliferation of disease registries, such as the National Cancer
Institute's Cancer Surveillance System, and computerized diagnostic data
from health plans has made it much more feasible to select all or a represen-
tative sample of all cases occurring in a defined population. Population-
based cases should be more typical and include a wider spectrum of disease
severity.
The cases in case-control research should, if possible, be new (incident)
cases, not existing (prevalent) ones. The reasons are based on the concepts
discussed in Chapter 4. The prevalence of a disease at a point in time is a
function of both the incidence and duration of that disease. Duration is in
turn determined by the rate at which patients leave the disease state (be-
cause of recovery or death) or persist in it because of a slow course or
successful palliation. Tt follows from these relationships that risk factors
for prevalent disease may be risk factors for either incidence and duration
or both; the relative contributions of the two cannot be distinguished. An
exposure that causes a particularly lethal form of the disease, thereby
lowering the proportion of prevalent cases that are exposed, would result
in a lowered relative risk if prevalent cases were studied. The reader can
be somewhat reassured that the results of a case-control study are not
biased by the selection of prevalent cases if the odds ratios obtained are
similar in short- and long-duration cases.
SELECTING CONTROLS
A major potential for bias exists in many case-control studies because
the controls are not a naturally occurring group, but one constructed for
the study by the investigators. Which controls are appropriate in relation
to the cases?
There are several strategies for choosing the right controls. First, the
best way to minimize selection bias is by selecting both cases and controls
from the same defined population. If cases comprise all cases or an un-
biased sample of all cases arising in the population, whether accrued in a
cohort study or identified in a prevalence survey, then controls can be a
random sample of all the other people in the same population. This strategy
is called a population-based or nested (in a cohort) case-control study. Controls
should meet the same general inclusion/exclusion criteria as the cases and
be sampled from the population or cohort at about the same times as the
cases arose.
Example Does habitual, vigorous physical activity protect against pri-

mMy cardiac arrest in people without apparent heart disease? An emergency
medical information system facilitated the conduct of a population-based
case-control study to answer this question (11). Cases were selected from
1250 people living in Seattle and suburban King County, Washington, who
had suffered out-of-hospital primary cardiac arrest (PCA) during a defined
period of time. Cases were chosen from paramedic reports; paramedics at-
tended nearly all instances of peA in the area at the time.
Controls were selected by dialing randomly selected telephone numbers
in the same area; most people in the area had telephones in their homes.
Both cases and controls had to meet criteria for entry: ag-e 25-7.'i years; no
clinically recognizable heart disease; no prior disease that limited activity;
and a spouse who could provide information about habitual exercise, the
exposure of interest. Controls were matched to cases on age, sex, marital
status, and urban or suburban residence. Spouses of both cases and controls
were asked about leisure-time activity. The entry criteria sought to ensure
that cases and controls were members of the same base population and had
opportunities to engage in physical activity.
The results, based on lb.')eligible cases and controls, confirmed previous
studies. The risk of PCA was 6_'i-75~;) lower in persons with high-intensity
leisure-time activity than in more sedentary people.
Although selecting cases and controls from a defined population or

cohort is preferable, selecting both from hospitals or other health organiza-
tions is often more feasible. But studying people in health care settings is
also more fallible because patients are usually a biased sample of all people
in the community, the people to whom the results should apply.
A second set of strategies for having controls who are comparable to
cases include the ones illustrated by the examples in this chapter and
presented in Chapter 6: restriction, matching, stratification, and adjust-
ment. Matching poses the greatest challenges and will be discussed here,
Cases can be matched with controls so that for each case one or more
controls are selected that possess characteristics in common with the case.
Researchers commonly match for age, sex, and residence, because these
are frequently related to disease. But matching often extends beyond these
demographic characteristics when other factors are known to be important.
Matching increases the useful information obtainable from a set of cases
and controls because it reduces differences between groups in determi-
nants of disease other than the one being considered and thereby allows
for a more powerful (sensitive) test of association. But matching: carries a
risk. If the investigator happens to match on a factor that is itself related
to the exposure under study, there is an increased chance that the matched
case and control will have the same history of exposure. For example, if
cases and controls were matched for the presence of arthritic symptoms,
which are commonly treated with NSA1Us, more matched pairs would
likely have the same history of NSAJD use. This process, called ooennatch-
ing, will bias the odds ratio toward 1 and diminish the ability of a study
to detect a significantly increased or decreased cdds ratio.
A third strategy is to choose more than one control group. Because of
the difficulties attending the selection of truly comparable control groups,
a systematic error in the odds ratio may arise for <lny one of them. One
way to guard against this possibility is to choose more than one control
group from different sources.' One approach used when cases are drawn
from a hospital is to choose one control group from other patients in the
same hospital and a second control group from the neighborhoods in which
the cases live. If similar odds ratios are obtained using different control
groups, this is evidence against bias, because it is unlikely that bias would
affect otherwise dissimilar groups in the same direction and to the same
extent. If the estimates of relative risks arc different, that is a signal that
one or both are biased, and an opportunity exists to investigate where the
bias lies.
Example In a case-control study of l:'strogl:'n and endometrial cancer,
cases were identified from a single teclC"hing hospital. Two control groups
wen' selected: one from among gynecologic admissions to the same hospital
and the second from a random sample of women living in the area served
by the hospital.
The presence of other diseases, such as hypertension, diabetes, and gall-
bladder disease, was much more common among the cast's and the hospital
control group, presumably reflecting the various forces that lead to hospital-
ization. Despite these differences, the two control groups reported much less
long-term t'strogen usc than did the cases and yielded very similar odds
ratios (4.1 and 16).
The authors (l2l conc-luded that "this consistency of results with two very
different comparison groups suggests that neither is significantly biased and
that the results .. are reasonably accurate."
Options for selecting cases and controls are summarized in Figure 10.4.
If cases are all occurring in a defined population (or a representative sample
of all cases), then controls should be too. This is the optimal situation. If
cases are a biased sample of all cases, as they arc in most hospital samples,
then controls should be selected with similar biases.
MEASURING EXPOSURE
Even if selection bias can be avoided in choosing cases and controls,
the investigator faces problems associated with validly measuring expo-
sure after the disease or outcome has occurred, i.c., avoiding measurement
bias. Case-control studies are prone to three forms of measurement bias
1 Ch",."ing tw" or more conlr,,1 gmup' ,'cr cose t;r0up, i, dilf~"·,,t f",m ch"'''ing Iw(' or m(m.' c"ntr,,1s
pcr case, Til., btkr i, d"",·, to illn",,,,, ,r,lh,ti,al pow"r (or prc",i,ion "I the c,t,male of rd"ti"" 'i,kl. Tn
W'n"ral, using mOl'e thon ,'n,' contr"T ,uhi~ct T)("""", ,,-,ults in ,m"IT I:>ulu,dul gain, in pnwer. bullherc
i, Tittle u",fuT ,d"ant"I:<' to "dding mor,' controls ,'er C05C beyond thr~~ <>r lour.
CHAPTER 10 ! STUDYING CASES 223
Samples
Random Investigator Selected

Population (unbiased) (matching biases)
Altc!iSes
ora
CASES>
ra:f'ldOin
sarrple
Ratloom
•
•
ICONTROlS> samplePt
ncncases
• •
Figure 10.4 Two strategies for selecting cases and controls from the general popu-
lation: unbiased samples and samples with rnatchinq biases.
because the exposure is measured after the onset of the disease or outcome
under study.
l. The presence of the outcome directly affects the exposure;
2. The presence of the outcome affects the subject's recollection of the
exposure; and
3. The presence of the outcome affects the measurement or recording of
the exposure.
The first bias is particularly problematic if the exposure under study is
a medical treatment, since the early manifestations of the illness may lead
to treatment. This is sometimes referred to as confounding by indication.
for example, a case-control design was used to determine whether beta-
blocker drugs prevented first myocardial Infarctions in patients being
treated for hypertension (13). Because angina is a major indication for use
of beta-blockers, the investigators carefully excluded any subjects with a
history suggesting angina or other rnani festation of coronary heart disease.
They found that hypertensive patients treated with beta-blockers still had
a significantly reduced risk of nonfatal myocardial infarctions, even after
those with angina or other evidence of coronary disease were carefully
excluded.
Second, people with a disease may recall exposure differently from those
without the disease. With all the publicity surrounding the possible risks
of various environmental exposures or drugs, it is entirely possible that
victims of disease would remember their previous exposures more acutely
than nonvfctirns or even overestimate their exposure. The influence of
disease on memory status, called recall bias, must be considered when

measurement of exposure relics on memory.
There arc two protections against biased remembering. First, there
should be alternative sources of the same information, whether written
documents, such as medical or other records, or interviews with n-ln-
tives or other knowledgeable individuals. Second, the specific purpose
of the study should be concealed from the study subjects. It would be
unethical not to inform subjects of the general nature of the study ques-
tion. But to provide detailed information to subjects about the specific
hypotheses could so bias the resulting information obtained as to com-
mit another breach of ethics-involving subjects in a worthless research
project.
The third problem, whether the presence of the outcome influences
the way in which the exposure is measured or recorded, should be
understandable to all students of physical diagnosis. If a resident admit-
ting a patient with renal disease to the hospital is aware of a possible
link between NSAID usc and renal failure one could expect the resident
to question the patient more intensely about previous analgesic use and
to record the information more carefully. Interviewers who are aware
of a possible relationship between exposure and disease and also the
outcome status of the interviewee would be remarkable indeed if they
conducted identical interviews for cases and controls. The protections
against these sources of bias are the same as those mentioned above:
multiple sources of information and blinding the data gatherers, 1.('.,
keeping them in the dark as to the hypothesis under study.
SCIENTIFIC STANDARDS FOR CASE-CONTROL RESEARCH
It has been suggested that one should judge the validity of a case-
control study by first considering how a randomized controlled trial of
the same question would have been conducted (14). Of course, one could
not actually do the study that \ViJY. But a randomized controlled trial
would be the scientific standard against which to consider the effects
of the various compromises that are inherent in a case-control study.
One would enter into a trial only those patients who could take the
experimental intervention if it were offered, so in a case-control study
one would select cases and controls who could have been exposed. For
example, a study of whether NSl\lDs are a cause of renal failure would
include men and women who had no contratndtcations to taking
NSI\IDs, such as peptic ulcer. Similarly, both cases and controls should
have been subjected to equal efforts to discover renal disease if it were
present. These and other parallels between clinical trials and case-con-
trol studies can be exploited when trying to think through just what
CHAPTER 10 ! STUDYING CASES 225
could go wrong, how serious a problem il; it, and what can be done
about it.
There have also been efforts to set out criteria for sound case-control
studies (15). To apply these guidelines requires an in-depth understand-
ing of the lllany possible determinants of exposure and disease, as well
as the detection of both, in actual clinical situations.
USING CASE·CONTROL STUDIES TO EXAMINE HEALTH CARE
The major use of case-control studies has been to test hypotheses about
the etiology of disease. More recently, investigators have exploited the
advantages of the case-control design to study questions related to the
provision and quality of health care.
Example Is cerebral palsy and fetal death preventable? British investiga-
tors (16) used a case-control design to compure the antepartum care received
by 141 babies developing cerebral palsy and 62 dying intrapartum or neona-
tally. Each case was matched with two healthy babies born at the same time
and place. A failure to respond to signs of severe fetal distress was more
common among cases than controls but only accounted for a v,'ry small
percentage of babies with cerebral palsy.
Because most serious adverse effects of poor-quality medical care are

relatively rare, the case-control design provides an efficient strategy for
examining the relationship between deviations from guidelines or other
protocols and poor outcomes.
Summary
Much of medical progress is derived from the careful study of sick
individuals. Case reports are studies of just a few patients, e.g.. < l O. They
are a useful means of describing unusual presentations of disease, examin-
ing the mechanisms of disease, and raising hypotheses about causes and
treatments. However, case reports are particularly prone to bias and
chance. Case series-studies of larger collections of patients-still suffer
from the absence of a reference group with which to compare the experi-
ence of the cases and from sampling cases at various times in the course
of their disease.
In case-control studies. a group of cases is compared with a similar
group of noncascs (controls). A major advantage resides in the ability to
assemble easel; from treatment centers or disease registries as opposed to
finding them or waiting for them to develop in a defined population at
risk. Thus case-control studies are much less expensive and much quicker
to perform than cohort studies and the only feasible strategy for studying
risk factors for rare diseases. Relative risk can be estimated by the odds
ratio, although it is not possible to compute incidences or relative risk
directly. The disadvantages of the case-control design all relate to its con-
sidcrablc susceptrbfhty to bias. This problem is most related to two chm-ar-

tcristics of case-control research. First, the grouFs to be compared are com-
monly constructed by the researcher and are not constituted naturally;
second, the exposure is measured after the disease has already occurred.
Given the vulnerability of case-control studies to bias, what place do
they have in clinical epidemiologic research? To some, case-control studies
are unscientific, illogical, and a curse. To others, they are viewed as the
essential first step in studying many medically important questions. There
is nearly universal agreement that cohort studies provide stronger, more
valid evidence and, if feasible, are the design of choice. But with appro-
priate attention to possible sources of bias, case-control studies can provide
a valid and efficient method to answer many clinical and health services
questions.
REFERENCFS
1. Klatskin C, Kimherg nv. Recurrent hepatitis atlrihut~hle to halothane sensitization in <In
anesthetist. N Engl J Mpd 1969;280:515-522,
2. Roush \1K, McNutt RA, CrilY TF. TIle adverse effect dilpmma: quest far accessible infor-
mation. Aim Intern Mcd 1991;114:298-299.
3. Mulvihill JJ. Clinical ecogpndics. Cancer in families,?\i Engl J MpdI985;312:1569-157(),
4. ll"rk-,on 1. 1.imitations of the application of fourfold table analysis to ho,pital d>lta. Ripmed
Bull 1946;2:47-_~3.
5. Jaffe HW, llregman DJ, Selik RM. Acquired immune deficiency syndrome in the United
St>lks: the first 1,000 cases, J Inf<:d Dis 1983;148:339-345.
6, Jemen ~1C.l3ri1nt-ZawadzkiMN, OBuchowski N. Modic MT, Malkasian D, RossJS. \lag-
netic resonance im>lging of the lumbar spine in p<,opl" without back pain, :-.,: Lngl J l\led
1994;33\:69-73,
7 Sandler Dr, Burr lR Weinb"rg CR. Non-aspirin nonsteroidal antiinfiammMory drugs
and risk of chronic renal disease. Ann Inkrn Med 1991;115:\65-172.
H. Schl"ss"lm<ln jJ. Case control studies; d"sign, nmduct, analysis, New York: Oxford Univer-
sity Press, 19H2,
<J. Feinstein AI'. Thl' biils caused by high values of incid",,'ce for P, in the odds rati" as,ump-
tion that 1 I', "" 1 J Chron Dis 1986;39:485-41l7.
10. Wacholder S, McLlUghlin JK, Silverman DT, Mandel JS, Seb·tion of controls in case-
,"ontrol studies. Am J Epidemioll<J92;BS:l019-1028.
11, SisUlvick I)S. weiss NS. Hallstrom AI', lnui T~, I''''en;on DR. Physical activity and primary
cardiac arrest. JAMA 1982;248:3113-3117.
12 Hulka BS, Fowler we Jr, Kaufman DG, Greenberg Be, Hogue CjR, Berger GS, Pulliam
CC E~lrogen and endometrial Cann,r: C>l~es and two control groups from North Carolina
Am J Obstd Cyn<:col 1980;137:92-101.
13. I'saty BM, Koepsdl TD. Lorjerto JP, Wagner U I, Inui 'fS, Beta-blockers and primary
prevention of coronary h"art dbease in patients with high blood pressure. lAMA
'1989;261:2087-2094,
14. F<:inskin AR, Horwitz RJ. Double standards. sc-ientific methods and epidemiologic ,",,'-
search, N Engl J \1ed 1982;307:1611-1617,
\5. Horwitz I~, F<:in~teil1 AR Methodologic standards and contradictory results in cnsc control
research, Am J Med 1979;6fi:556-564.
16. Gaffney G. Sellers S, Flav<:ll V, Squier M, Johnson A. Case-control study of intrapartum
car". c"r,,'hrill palsy. and perinatal d"ath. Br Med J 1994;301:1:743-750,
SUGGESTED READINGS
Ieinstetn AR. Clinical biostatistics XX: th(, epidemiologic trohoc, the ablative risk ratio, and
"retrospective TtN,arch." Clin Pharmacol Ther 1973;14:291-307.
Feinstein AR, Horwitz Rl, Spitzer WO, Battista RN. Coffee and pancreatic cancer: the problems
of etiologic science and epidemiological case-(:ontrol research. JAMA 19111;246:957-961.
Hayden GF, Kramer MS, Horwitz RI. The case-control study. A practical review for the
clinician. JAMA 1982;247:326-331.
Ibrahim MA, Spitzer WOo 1he case-control study: consensus and controversy. New York:
Pergamon Press, 1979.
Rothman KJ. Modern epidemiology. Boston: Little, Brown & Co., 1986.
11
CAUSE
- - _... --
Example Some vvars ago, medical students were presented a study of

the relationship between the cigarette smoking habits of obstetricians and
the vigor of babies they delivered. Infant vigor is measured by an Apgar
score; a high score (9-10) indicates that the babv is healthy, whereas a low
score indicates the baby might be in trouble and require close monitoring.
The study suggested that smoking bv obstetricians (not in the delivery suitcl]
had an ,{dvcrse effect on Apgar scores in newborns. The medical students
were then asked to comment on what was wrong with this study. After
many suggestions, someone finally said that the conclusion simply did not
make sense,
It W,lS then acknowledged that. although the study was real. the "expo-
sure" and "disease" had been altered for the presentation. Instead of compar-
ing smoking habits of obstetricians with Apgar scores of newborns, the study,
published in 1R4.1 by Oliver Wendell Holmes (then professor of anatomy
and physiology and later dean of Harvard Medical School), concerned hand
washing habits by obstetricians and subsequent puerperal sepsis in mothers.
The observations led Holmes (1) to conclude: "The disease known as puer-
peral fever is so far contagious, as to be frequently carried from patient to
patient by physicians and nurses."
One mid-19th century response to Holmes's assertion that unwashed
hands caused puerperal fever was remarkably similar to that of the medical
students: The findings made no sense. "1 prefer to attribute them [puerperal
sepsis cases] to accident, or Providence, of which I em form a conception,
rather than to contagion of which I cannot form any dear idea, at least as to
this particular malady," wrote Dr. Charles D. Meigs, professor of midwifery
and the diseases of women and children at Jefferson Medical College (1).
Holmes and Meigs were confronted with a question about cause.

Holmes was convinced by his data that the spread of puerperal sepsis was
caused by obstetricians not washing their hands between deliveries. He
could not, however, supply the pathogenetic mechanism by which hand
washing was related to the disease, as bacteria had not yet been discovered.
Meigs, therefore, remained unconvtnccd that the cause of puerperal sepsis
had been established (and presumably did not bother to wash his hands).
Clinicians frequently arc confronted with information about possible
228
CHAPTER 11 ! CAUSE 229
causal relationships. Tn fact, most of this book has been about methods
used to establish cause, although we have not called special attention to
the term.
Tn this chapter, we review concepts of cause in clinical medicine. We
then outline the kinds of evidence that, when present, strengthen the likeli-
hood that an association represents a causal relationship. We also deal
briefly with a kind of research design not yet considered in this book:
studies in which exposure to a possible cause is known only for groups
and not specifically for individuals in the groups.
Concepts of Cause
Webster's (2) defines canse as "something that brings about an effect or
a result." In medical textbooks, cause is usually discussed under such
headings as "etiology," "pathogenesis," "mechanisms." or "risk factors."
Ciluse is important to procticing physicians primarily in guiding their
approach to three clinical tasks: prevention, diagnosis, and treatment. The
clinical example at the beginning of this chapter illustrates how knowledge
of cause-and-effect relationships can lead to successful preventive strate-
gies. Likewise, when we periodically check patients' blood pressures, we
are reacting to evidence that hypertension causes morbidity and mortality
and that treatment of hypertension prevents strokes and myocardial in-
farction. The diagnostic process, especially in infectious disease, frequently
involves a search for the causative agent. Less directly, the diagnostic
process often depends on information about cause when the presence of
risk factors is used to identify groups of patients in whom disease preva-
lence is high (see Chapter 3). Finally, belief in a causal relation.s. hip under-
lies every therapeutic maneuver in clinical medicine. Why give penicillin
for pneumococcal pneumonia unless we think it will cause a cure? Or
advise a patient with metastatic cancer to undergo chemotherapy unless
we believe the antimetabolite will cause a regression of metastases and a
prolongation of survival, comfort, and/or ability to carryon daily
activities?
By and large, clinicians are more interested in treatable or reversible
than immutable causes. Researchers, on the other hand, might also be
interested in studying causal factors for which no efficacious treatment
or prevention exists, in hopes of developing preventive or therapeutic
interventions in the future.
SINGLE AND MULTIPLE CAUSES
In 1882, 40 years after the Holmes-Meigs confrontation, Koch set forth
postulates for determining that an infectious agent is the cause of a disease.
Basic to his approach was the assumption that a particular disease has one
cause and a particular cause results in one disease. He stipulated that:
230 CUNICAL EPIDf:'MIOLOGY
1. The organism must be present in every case of the disease;

2. The organism must be isolated and grovvn in pure culture;
3. The organism must cause ,1 specific disease when inoculated into an
animal: and
4. The organism must then be recovered from the animal and identified.
Interestingly, he did not consider the effect of treatment in establishing
cause, something he might have added a century later when effective
treatments had become more common in medicine.
Koch's postulates contributed greatly to the concept of cause in medi-
cine. Before Koch, it was believed that many different bacteria caused any
given disease. The application of his postulates helped bring order out of
chaos. They are still useful today. That a given organism causes a given
disease was the basis for the discovery in 1977 that Legionnaire's disease
is caused by a Gram-negative bacterium, and the determination in the
1980s that the newly discovered HTV causes ATns.
For most diseases, however, cause cannot be established simply by
Koch's rules. Sometimes, too much reliance on Koch's approach has gotten
the medical community into trouble by narrowing our perspectives. Would
that disease was so simple that we always had a single cause-single disease
relationship. Smoking causes lung cancer, chronic obstructive pulmonary
disease, peptic ulcers. bladder cancer, and coronary artery disease. Coro-
nary artery disease has multiple causes, including cigarette smoking, hy-
pertension, hypercholesterolemia, and heredity. It is also possible to have
coronary artery disease without any of these known risk factors.
Usually, many factors act together to cause disease in what has been
called the "web of causation" (3). A causal web is well understood in
conditions such as coronary artery disease, but is also true for infectious
diseases, where presence of the organism is necessary for disease to occur
but not necessarily sufficicnt. ,\Ins cannot occur without exposure to TTlV,
but exposure to the virus does not necessarily result in disease. For exam-
ple, exposure to HIV rarely results in AIDS after needlcxticks (3 or 4/10(0),
because the virus is not nearly as infectious as, say, the hepatitis U virus.
PROXIMITY OF CAUSE TO EFFECT
When biomedical scientists study cause, they usually search for the
underlying pathogenetic mechanism or final common pathway of disease.
Sickle-cell anemia is an example, with the genetic change associated with
hemoglobin S (HbS) leading to polymerization and erythrocytic sickling
when fTbS gives up its oxygen. Elucidating pathogenesis of disease has
played a crucial part in the advancement of medical science in this century.
However, disease is abo determined by less specific, more remote
causes, or risk factors, such as people's behavior or characteristics of their
environments. These factors may be even more important causes of disease
CIIAPTE::.f-:: 11 ! CAUSE 231
than are pathogenetic mechanisms. For example. a large proportion of

cardiovascular and cancer deaths in the United States can be traced to
behavioral and environmental factors; the spread of AIDS is due primarily
to sexual behaviors and drug use; and deaths from violence and acci-
dents are rooted in social conditions, access to guns, and alcohol and
»catbelt use.
To view cause in medicine exclusively as cellular and subcellular pro-
cesses restricts the possibilities for useful cl i nical interventions. Tf the patho-
genetic mechanism is not clear, knowledge of risk factors milY still lead to
very effective treatments and preventions. Thus Holmes was right in his
assertion that obstetricians should wash their hands, even though he had
little notion of bacteria.
For many diseases, both pathogenetic mechanisms and nonspecific risk
factors have been important in the spread and control of the diseases.
Sometimes the many different causes interact in complicated ways.
Example Koch's postulates were orig-inilily used to establish that tuber-
culosis is caused by inoculation of the acid-fast bacillus Mycobacterium tuber-
Cillosis into susceptible hosts. TIll' final common pathway of tuberculosis is
the invasion of host tissue by the bacteria. horn a pathogenetic perspective,
conquering the disease- required antibiotics or vaccines that were effective
against the organism. Through biomedical research efforts, both have been
achieved.
However, the development of the disease tuberculosis is far more com-
plex. Other important causes arc the susceptibility of the host and the degree
of exposuH' (Fif;. 11.1). In fad, th,'se causes determine whether invasion of
host tissue r-an occur.
~xP9slJreto
Crowding MYCOBACTtRiUM
Malnutrition
Vaccination
Genetic Tissue Invasion and Reaction
~',r, INFECTION TUBERCULOSIS
~4iil'
Risk Factors for Mechanisms of

Tuberculosis Pathogenesis Tuberculosis
'II Distant from Outcome ----. + - Proximal to Outcome -----+
Figure 11.1. Causes of tuberculosis.

232 CLINICAL. EPIDEMIOLOGY
Some clinicians would be hesitant to label host susceptibility and level

of exposure as causes of tuberculosis, but they <Ire very important compo-
nents of cause. Tn fact, social and economic improvements influencing- host
susceptibility, such as less crowded living space and better nutrition, may
have played a more prominent role in the decline in tuberculosis rates in
developed countries than treatments developed through the biomedical-
pathogenetic research model. Figure 11.2A shows that the death rate from
tuberculosis had dropped dramatically long before antibiotics were intro-
duced. (The vaccine came even latcr.]
Since 1985 the number of TB cases in the United States has increased
(4) (Fig. 11.2B). 'Why is this so? A total of 60°/,) of the increase occurred in
foreign-born persons. HIV infections arc also a factor, with an increasing
number of susceptible people, as AIDS spreads and weakens the immune
systems of its victims. These susceptible hosts are more likely than the
general population to be exposed to tuberculosis, because both AIDS and
tuberculosis are more common in economically depressed populations.
Finally. changes have occurred in the bacillus itself, with the evolution of
multidrug resistant strains. To complicate the picture further, multidrug
resistance also is caused by a web of circumstances. Genetic changes in
the Mycobacterium are more likely to occur with medication noncompliance
(5), which is more likely among- intravenous drug users, an important risk
group for ATDS. Changes in the bacterium's genetic makeup may also be
related to high replication rates in immunodeficient hosts. Thus the inter-
play of environment, behavior and subcellular biology may be incredibly
complex when thinking about cause.
Another example of the importance of both pathogenetic and epidemio-
logic approaches to cause is the recent decline in deaths from coronary
artery disease in the United States.
Example During the polsl two decades, the death rate from coronary
artery disease has dropped more than a third. This decline accompanied
decreased exposure, in the population as a whole, to several risk factors for
cardiovascular disease: A larger proportion of people with hypertension are
being treated effectively, middle-aged men are smoking less, and fat and
cholesterol consumption has declined. These developments were, at least in
part, the result of both epidemiologic and biomedicaIstudies and have spared
tens of thousands of lives per year. It is doubtful that they would have
occurred without understanding"of both the proximal mechanisms and the
more remote origins of cardiovascular disease (6).
INTERPLAY OF MULTIPLE CAUSES
When more than one cause act together, the resulting risk may be greater
or kss than would be expected by simply combining the effects of the
separate causes. Clinicians call this phenomenon synergism if the joint
400 A
300
Tubercle bacillus identified
200 I
100 Chemotherapy introduced
I
o'---,-_,-~,---~,--~,--~~-"'.,..~,
1838 1860 1880 1900 1920 1940 1960 1980
Year
B
50000 -
40000
....
Q)
30000
25000
o'"
lD
20000
f0-
e; 52,100 excess cases
... 15000
~
E
::J
Z
1980198119821983198419851986198719881989199019911992
Year
-------~-----
Figure 11.2. A, Declining death rate from respirafory tuberculosis in England and
Wales over the past 150 years. (From McKeown T The role of medicine: dream,
mirage or nemesis. London: Nuffield Provincial Hospital Trust, 1976.) B, Excess
tuberculosis cases in the United States, 1985-1992. Difference between expected
and observed number of cases. Dotted line, observed cases; solid line, expected
cases. (From Cantwell MF, et al. Epidemiology of tuberculosis in the United States,
1985 through 1992. JAMA 1994;272:535-539.)
234 CLINICAL EPIDEMIOI_OGY
400
.,
III
~
~ 317
ee 300
r::
r::
0
"
::;; 200
0
0
~
100
US 61
".,
U
III
12 20
0
Serum cholesterol 185 185 185 ':l:;l!i q
(mg %)
Systolic blood 105 105 jill;

, '"''
105 llli!;
pressure (mm Hg)
Cigarette smoking 0 0 0
Figure 11.3. Interaction of multiple causes of disease, The risk of developing car-
diovascular diseaso in men according to the level at several risk factors alone and
in combination. Abnormal values are in shaded boxes. (Redrawn from Kannel WB,
Preventive cardiology. Postgrad Med 1977;61:74-85.}
effect is greater than the sum of the effects of the individual causes, and
antagonism if it is less.'
Example Figure 11.3 shows the probabilitv of developing cardiovascular
disease over an 8-pcilf period among men aged 40. Men who did not smoke
cigarettes, had low serum cholesterol values, and had low systolic blood
pressure readings were at low risk of developing disease (l2/IOOO). Risk
increased, in the range of 20 to 61/1000, when the various factors were
present individually. But when all three factors were present, the absolute
risk of cardiovascular disease (317/1000) was almost three times greater than
the sum of the individual risks (7).
Elucidation of cause is more difficult when many factors playa part

than when a single one predominates. However, when multiple causative
i Slnli,licu! ;"lemc/ioll is pr~,~nt when combination' of "ariah"', in a malhemoti""l modd add to the
model's expldnah>ry power aller lhe nd ~fh~·t, of the individual pr",lict,,, variables have been t"k~n into
.m"",lllt. It is eoneeplu"lly rt'btl'd to biologic SY"~Tgy ond Jnla8oni,m but is " m<1themillical con,trud,
not <1n "b"'rv"bh' phenomenon in nalur<'
CHAPTER 11 I CAUSE 235
factors are present and interact, it may be possible to make a substantial

impact on a patient's health by changing only one, or a small number, of
the causes. Thus, in the previous example, getting patients to give up
smoking and treating hypertension might substantially lower the risk of
developing cardiovascular disease in men, even in the continuing presence
of other causative factors.
EFFECT MODIFICATION
Effect modifiration is a special type of interaction. It is present when the
strength of the relationship between two variables is different according
to the level of some third variable, called an effect modifier.
Example Because of conflicting results of studies evaluating the cffcc-
tiveness of thiazide diuretics in preventing coronary heart disease, a study
was done to examine whether there was a relationship between the dose of
thiazide and the risk of sudden death, and whether adding potassium-
sparing therapy modified the effect. Figure 11.4 summarizes the results. TIle
dose of thiazide determines its effect, with a low dose, 25 mg, protecting
against sudden death and a high dose, IOU rng, increasing the chances of
sudden death. Adding potassium-sparing therapy modifies the effectat sev-
eral doses, adding a protective effect (8).
Establishing Cause
ln clinical medicine, it is not possible to prove causal relationships be-
yond any doubt. It is only possible to increase one's conviction of a cause-
and-effect relationship, by means of empiric evidence, to the point at which,
for all intents and purposes, cause is established. Conversely, evidence
against a cause can be mounted until a cause-and-effect relationship be-
comes implausible. The possibility of a postulated cause-and-effect rela-
tionship should be examined in as many different ways as possible. This
usually means that several studies must be done to build evidence for or
against cause.
ASSOCIATION AND CAUSE
Two factors-the suspected cause and the effect-obviously must
appear ttl be associated if they are to be considered as cause and effect.
However, not all associations are causal. Figure 11.5 outlines other
kinds of associations that must be excluded. Pirst. a decision must be
made as to whether an apparent association between a purported cause
and an effect is real or merely all artifact because of bias or random
variation. Selection and measurement biases and chance are most likely
to give rise to apparent associations that do not exist in nature. If these
problems can be considered unlikely, a true association exists. But
before deciding that the association is causal, it is necessary to know
if the association occurs indirectly, through another (confounding) fac-
236 CLINICAL [PIDEMIOlOGY
Odds Ratio
0.1 0.2 0.5 1,0 2.0 50 10.0
Therapy Odds Ratio

(95%CI)
1OO-mg thiazide 2.4 (0.7-13.0)
SO-mg thiazide 1.1 (0.5-2.5) :- --.~.
2S-mg thiazide 0.7 (0.2-2.5)
50-mg thiazide and 0.5 (0.1-2.2)

potassium-sparing diuretic
2S-mg thiazide and

potassium-sparing diuretic
0.3 (0.1-1.0) ..
•
Favors Favors
diuretic B-adrenergic antagonist
therapy therapy
Figure 11.4. Example of effect rnoduication: how the risk of cardiac arrest in oa
tients using thiazide diuretics Icornpared with the risk of cardiac arrest in patients
using beta-blockers) changes according to use of potassium-sparing diuretics, Odds
ratios-with 95% confidence intervals (GI) ---increase with increasing dose of diuretic,
suggesting that it is safer to use beta-blockers than thiazide diuretics. However, with
the addition of potassium sparing diuretics, thiazide diuretics cause a lower risk of
cardiac arrest than beta-blocker therapy, (Redrawn from Siscovick OS, et a. Diuretic
therapy for hypertension and the risk of primary cardiac arrest. N Engl J Med
1994; 330:1852-1857,)
tor, or directly. If confounding is not found, a causal relationship is

likely.
At some future time another factor may be found that is more di-
rectly causal. Por example, several studies found that women fared
more poorly than men after coronary bypass surgery and it was
thought that sex was related to postoperative prognosis. On further
study, small body surface area-which correlated with small-diameter
coronary arteries-was found to be an important variable leading to
heart failure and death, not being female per se (9). Thus factors that
are considered causes at one time are sometimes found to be indirectly
related to disease later, when more evidence is available.
CHAPTER 11 ! CAUS~ 237
EXPLANATION FINDING _
Association
Bias in selection
or measurement
LY-';J ~o
.. "
Chance [Dkel1' tUn'.
~
Confounding
Cause Cause
Figure 11.5. Association and cause.
RESEARCH DESIGN
When considering a possible causal relationship, the strength of the
research design used to establish the relationship is an important piece of
evidence.
Of the research designs so far discussed in this book, well-conducted
randomized controlled trials, with adequate numbers of patients; blinding
of therapists, patients, and researchers; and carefully standardized meth-
ods of measurement and analysis are the best evidence for a cause-and-
effect relationship. Randomized controlled trials guard against differences
in the groups being compared, both for factors already known to be im-
portant, which can be overcome by other methods, and for unknown con-
founding factors.
We ordinarily usc randomized controlled trials to provide evidence
about causal relationships for treatments and prevention, However, as
pointed out in Chapter 6, randomized controlled trials are rarely feasible
when studying causes of disease. Observational studies must be used
instead.
In general, the further one must depart from randomized trials, the less
the research design protects against possible biases and the weaker the
evidence is for a cause-and-effect relationship. Well-conducted cohort stud-
ies are the next best design, because they can be performed in a way that
minimizes known confounding, selection and measurement biases. Cross-
sectional studies are vulnerable because they provide no direct evidence
of the sequence of events. True prevalence surveys- cross-sectional stud-
ies of a defined population-guard against selection bias but are subject
238 CUNICAL EPIDeMIOLOGY
to measurement and confounding biases. Case-control studies Me vulnera-

ble to selection bias ZlS well. Weakest of all arc cases series, because they
have no defined population and no comparison group.
This hierarchy of research designs is only ZI rough guide, based on
extent of susceptibility to bias. The manner in which an individual study
is performed can do a great deal to increase or decrease its validity, regard-
less of the type of design used.
POPULATION STUDIES
Up until now, we have discussed evidence for cause when exposure

and disease status are known for each individual in the study. In a different
kind of research, most often used for epidemiologic studies of large popula-
tions, exposure is known only for the groups, not for the individuals in
the groups.
Studies in which exposure to a risk factor is characterized by the average
exposure of the group to which individuals belong are called aggregate risk
studi!'s. Another term is ecological studies, because people are classified by
the general level of exposure in their environment.
Example What factors are axsociafcd with rarriiac mortality in developed
countries? 51. I.eger et al. (10) gathered data on rates of ischemic heart disease
mortality in 18 developed countries to explore the contribution of various
economic, health services, and dic(Jry variables. One finding that was not
anticipated was a strong negative association between ischemic h,'ar! disease
death and wine consumption (Fig. 11.6).
This study raist's the hypothesis that alcohol protects against ischemic
heart disease. Since then, studies on individuals have shown that levels of
serum high-density lipoprotein. a protec-tive factor for cardiovascular disease,
are increased by alcohol consumption.
Aggregate risk studies arc rarclv definitive in and of themselves. TIle

main problem is a potential bias called the ecological jI1/1I1(y: Affected indi-
viduals in a generally exposed group may not themselves have been ex-
posed to the risk. Also, exposure may not be the only characteristic that
distinguishes people in the exposed group from those in the nonexposcd
group, i.e.. there may be confounding factors. Thus aggregate risk studies
arc most useful in raising hypotheses, which must then be tested with
more rigorous research.
TIME SERIES STUDIES
Evidence from aggregate risk studies that a factor is actually responsible

for an effect can be strengthened if observations are made at more than
two points in time (before and after) and in more than one place. In a time
sail'S study, the effect is measured at various points in time before and
after the purported cause has been introduced. 1t is then possible to see if
the effect varies as expected. If changes in the pu rportcd ca use arc followed
11
• Finland
10
U,S.A.
9 • ••
Scotland
N.Z.
• Australia
8
• Canada
E.B.W.
7 •
• Ireland
e Norway
6
• Netherlands
• Denmark
5 • eBelgium
Sweden .
• e Austna
4 West Germany
3 •
Switzerland
•
Italy
2 •
France
0.40 0.80 1.20 1.60 2.00

Wine Consumption (Log l o Scale)
Figure 11.6. Example of an aggregate risk study: relationship between wine con-
sumption and cardiac mortality in developed countries. (Drawn from 81. Leger AS,
Cochrane AL, Moore F. Factors assoceteo with cardiac mortality in developed coun-
tries with particular reference to the consumption of wine. Lancet 19/9; 1.1017-
1020,)
by changes in the purported effect, the association is less likely to be

spurious.
Example The risk of Clostridium difficilc-associated diarrhea and pseudo-
membranous col itls has been shown to be related to the use of antibiotics,
particularly clindamycin, ampicillin, and ccphelosponns. An epidemic of C.
dijJicilc diarrhea broke out in a hospital in 1990 after usc of dindamycin
increased sharply (Hg. 11.7) (Il). Education, infection control, and environ-
mental hygiene efforts were irnrnodiatelv instituted, but the epidemic contin-
ued unabated. Clindamycin was then removed from the hospital formulary,
and its usc plummeted, along with the number of cases of C. d@cile-diarrhea.
To investigate the association further, the authors conducted a case-control
study, which corroborated the findings of the time series analysis.
In a multiple time series study, the suspected cause is introduced into

several different groups at various times. Measurements of effect are then
made among the groups to determine if the effect occurs in the same
sequential manner in which the suspected cause was introduced. If the
effect regularly follows introduction of the suspected cause at various times
and places, there is stronger evidence for cause than if this phenomenon
is observed only once, because it is even more improbable that the same
extraneous factor(s) occurred at the same time in relation to the cause in
many different places and eras.
Example Because there were no randomized controlled trials of cervical
cancer screening programs before they became widely accepted, their effec-
tiveness must be evaluated by means of observational studies. A multiple
time series study has provided some of the most convincing evidence of their
effectiveness (12). Data were gathered on screening programs begun in the
various Canadian provinces at various times during a 10-year period in the
1960s and 1970s. Reductions in mortality regularly followed the introduction
of screening programs regardless of time and location. With these data, it
12 800
C
UTI
~.
Eo
'" -c
TI_
c •
~.
9 O~
600
I Cl
~
::>
6
tv ~ 400 ·0
e
J(
>-
E
..
-e
3 200 <3"
I~
1Iqj LJ V
o 'I o
1988 1989 1990 1991 1992
Years of Surveillance
Figure 11.7. A time-series study of the relationship of c1indamycin use and Clostrid·
ium diffie/Ie-associated diarrhea (Redrawn from Pear SM, Williamson TH, Bettin KM,
Gerding ON, Galgiani IN. Decrease in nosocomial Clos/ridium difficile-associated
diarrhea by restricting c1indamycin use. Ann Intern Medl1 94;120:272--277_)
was concluded that "screening had a significant effect on reduction in mortal-

ity from carcinoma of the uterus."
ELEMENTS FOR OR AGAINST CAUSE
In 1965, the British statistician Sir Austin Bradford Hill (13) proposed a
set of features that should be sought when deciding whether a relationship
between a sickness and some environmental factor is causal or just an
association. His proposals have been widely used, sometimes with modifi-
cations (Table 11.1). We will comment briefly on the individual elements.
They are not all of equal weight in deciding about cause.
Temporal Relationships between Cause and Effect
Causes should obviously precede effects. This fundamental principle
seems self-evident, but it can be overlooked when interpreting most cross-
sectional studies and case-control studies, in which both the purported
cause and the effect are measured at the same point in time. It is some-
times assumed that one variable precedes another without actually estab-
lishing that this is so. In other cases, it may be difficult to establish which
came first.
Example It has long been noted that overweight persons are at higher
risk of death, especially cardiovascular death, than people with normal
weight. Thus it is reasonable to assume that weight loss would be protective
among overweight people. However, several cohort studies have found ex-
cess mortality among people who lose weight, even among people without
any apparent preexisting disease. These distressing findings may be ex-
plained if a subtle, prcrllnlcal effect of fatal illness is weight loss (14). Thus
fatal conditions may precede and cause weight loss, not vice versa. (This
possibility could be excluded if it was known whether the weight loss was
voluntary in those losing weight.)
Table 11.1
Evidence That an Association Is Cause and Effect"
Criteria CormTlHrlts
'tomporouv Cause precedes effect

Strength Large relative risk
Dose-response Largerexposures 10 cause associated with higher rates of disease
Reversibility Reduclion in exposure associated with lower rates of disease
Consistency Repeatedly observed by different persons, in different places,
circumstances. and times
Biologic plausibility Makes sense. according 10 biologic knowledge of the time
Specificity One cause leads to one errect
AnCllogy Cause-and-effect relationship already estatssnoo lor a similar
exposure or disease
"Modlttcd from Bradford-Hill A8 The erlVirorHT1Hni and disease ussoccton ere ccuseuoo. Proc R Soc Med
190S; :,B:?9:"i-300
Although it is absolutely necessary for a cause to precede an effect in

clinical medicine-and, therefore, the lack of such a sequence is powerful
evidence against cause-an appropriate temporal sequence alone is weak
evidence for cause.
Strength of the Association
A strong association between a purported cause and an effect, as ex-
pressed by a large relative or absolute risk, is better evidence for a causal
relationship than a weak association. Thus the 4- to 16-fold higher incidence
of lung cancer among smokers than nonsmokers in many different prospec-
tive studies is much stronger evidence that smoking causes lung cancer
than the findings in these same studies that smoking may be related to
renal cancer, for which the relative risks are much smaller (1.1-1.6) (15).
Similarly, that the relative risk of hepatitis B infection for hepatocellular
cancer is nearly 300 leaves little doubt that the virus is D cause of liver
cancer (16). Bias can sometimes result in large relative risks. However,
unrecognized bias is less likely to produce large relative risks than to
prod uce small ones.
Dose-Response Relationships
A dose-response relationship is present when varying amounts of the
purported cause are related to varying amounts of the effect. If a dose-
response relationship can be demonstrated, it strengthens the argument
for cause and effect. Figure 11.8 shows a clear dose-response curve when
lung cancer death rates (responses) are plotted against number of cigarettes
smoked (doses).
Although a dose-response curve is good evidence for a causal relation-
ship, especially when coupled with a large relative or absolute risk, its
existence does not exclude confounding factors.
Example Both the strong association between smoking and lung cancer
and the dose-response relationship have been dismissed by the tobacco in-
dustry as examples of confounding. According to this argument, there is
some unknown variable that both causes people to smoke and increases their
risk of developing lung cancer. The more the factor is present, the more both
smoking and lung cancer are found-hence, the dose-response relationship.
Such an argument is it theoretically possible explanation for the association
between smoking and lung cancer, although just what the confounding factor
might he has never been clarified. Short of a randomized controlled trial
(which would, on the ilverilge, allocate the people with the confounding
factor equally to smoking and nonsmoking groups) the confounding argu-
ment is difficult to refute completely.
Reversible Associations
A factor is more likely to be a cause of disease if its removal results in
a decreased risk of disease, i.e , the association between suspected cause
and effect is reversible. Figure 11.9 shows that when people give up smok-
ing they decrease their likelihood of getting lung cancer.
300
~
~"
251
-
~ Gl"
Gl:; 200
"
''0
,,0
O~ 127
",0 -
,,0
, , 0 100
...J~ ~
-"
.t:
e
C 0 ~
0 1-14 15-24 25+
Cigarettes Smoked, Number/Day
Figure 11.8. Example of a dose-response relationship: lung cancer deaths ac

cording to dose of cigarettes in male physicians, (Drawn from Doll R, Peto R, Mortality
in relation to smoking: 20 years' observations on male British doctors, Br Med J
1976; 2: 1525-1536,)
Reversible associations are strong, but not infallible, evidence of a causal

relationship. Confounding could conceivably explain a reversible associa-
tion. For example, in Figure 11.9 it is possible (but unlikely) that people
willing to give up smoking have smaller amounts of the unidentified factor
than those who continue to smoke.
Consistency
When several studies conducted <It different times in different settings
and with different kinds of patients <Ill come to the same conclusion, evi-
dence for a causal relationship is strengthened. That screening for colorectal
cancer is effective becomes more plausible when <I randomized controlled
trial of fecal occult blood testing (17) and a case-control study of sigmoidos-
copy (18) both find a protective effect. Causation is particularly supported
when studies using several different research designs all lead to the same
result, because studies using the same design can all make the same
mistake.
It is often the case that different studies produce different results. Lack
of consistency does not necessarily mean that the results of a particular
study are invalid. One good study should outweigh several poor ones.
Biologic Plausibility
Whether the assertion of cause and effect is consistent with our knowl-
edge of the mechanisms of disease as they are currently understood is
often given considerable weight when assessing causation. When we have
absolutely no idea how an association might have arisen, we tend to be
skeptical that the association is real. Such skepticism often serves us well.
For example, the substance Laetrile was touted as a cure for cancer in the
early 1980s. However, the scientific community was not convinced, because
they could think of no biologic reason why an extract of apricot pits not
chemically related to compounds with known anticancer activity should
be effective against cancer cells. To nail down the issue, Laetrile was finally
submitted to a randomized controlled trial in which it was shown that the
substance was, in fact, without activity against the cancers studied (19).
It is important to remember, however, that what is considered biologi-
cally plausible depends on the state of medical knowledge at the time. In
Mcig's day, contagious diseases were biologically implausible. Today, a
biologically plausible mechanism for puerperal sepsis, the effects of strep-
tococcal infection, has made it easier for us to accept Holmes's observations.
-"
0""
" 0
- E
~cn
"0 20 -
-
>~
- ""
.- >
~ Z 10 ~
-"
0"
:;; l!!
0"" ~ 4.7
o ~ ~
~~)(
a:
2.0
w 0 r-l
o <5 5-9 10-14 15+
Years Since Stopped Smoking
Figure 11.9. Reversible association: declining mortality from lung cancer in ex-
cigarette smokers. The data exclude people who stopped smoking after gettin9
cancer. (Drawn from 0011 R, Petro R. Mortality in relation to smoking: 20 years'
observations on male British doctors Br Med J 1976;2:1525-1536.)
CHAPTER 11 ! CAUSE: 245
On the other hand, the mechanism by which acupuncture causes anesthesia

is far Jess clear. To many scientists in the Western world, the suggestion
that anesthesia is caused by sticking needles into the body and twirling
them seems biologically implausible, and so they do not believe in the
effectiveness of acupuncture.
In sum, biologic plausibility, when present, strengthens the case for
causation. When it is absent, other evidence for causation should be sought.
If the other evidence is strong, the lack of biologic plausibility miJY ind reate
the limitations of medical knowledge, rather than the Jack of a causal
association.
Specificity
Specificity-one cause, one effect-is more often found for acute infec-
tious diseases (such as poliomyelitis and tetanus) and for inborn errors of
metabolism (such as gout, ochronosis, and familial hypercholesterolemia).
I\s we pointed out, for chronic, degenerative diseases there are often many
causes for the same effect or many effects from the same cause. Lung
cancer is caused by cigarette smoking, asbestos, and radiation. Cigarettes
cause not only lung cancer but also bronchitis, peptic ulcer disease, peri-
odontal disease, and wrinkled skin. The presence of specificity is strong
evidence for cause, but the absence of specificity is weak evidence iJgainst
a cause-and-effect relationship.
Analogy
The argument for a cause-and-effect relationship is strengthened if there
arc examples of well-established causes that arc analogous to the one in
question. Thus if we know that a slow virus can cause iJ chronic, degenera-
tive central nervous system disease (subacute sclerosing panencephalitis).
it is easier to accept that another virus might cause degeneration of the
immunologic system (acquired immunodeficiency syndrome). In general,
however, analogy is weak evidence for cause.
Weighing the Evidence

Most of this chapter has been a discussion of what to look for in individ-
ual studies when considering the possibility of a causal relationship. Hut,
when deciding about cause, one must consider all the available evidence,
from all available studies. After examining the pattern of evidence, the
case for causality can be strengthened or eroded. This calls for a good deal
of judgment, especially when the evidence from different studies is con-
flicting. In such cases, clinicians must decide where the weight of the
evidence lies.
Figure 11.10 summarizes the different types of evidence for and against
cause, depending on the research design, and features that strengthen or
Randomized controlled trial Randomized controlled trial

Multiple time series Multiple time series
Nonranoorruzec trial DESIGN Nonrandomized trial
Cohort Cohort
Case-control Case-control
Cross-sectional Cross-sectional
Case series Case series
Case report
Strong AGAINST Weak FOR Strong
Not specific Temporal sequence

Not reversible Small effect
No dose-response Specificity
No analogy Analogy
Not biologically plausible FINDING Biologic plausibility
No effect Consistency
Incorrect temporal sequence Large effect
Dose response
Reversibility
Figure 11.10. Relative strength of evidence for and against a causal effect Note
that with study designs, except for case reports and time series, the strenqth of
evidence for a causal relationship is a mirror image of that iJ9iJinsl. Wilh findings,
evidence for a causal effect docs not mirror evidence against an effect.
weaken the evidence for cause. The figure roughly indicates relative
strengths in helping to establish or discard a causal hypothesis. Thus a
carefully done cohort study showing a strong association and a dose-
response relationship is strong evidence for cause, while a cross-sectional
study finding no effect is weak evidence against cause.
Summary
Cause-and-effect relationships underlie diagnostic, preventive, and ther-
apeutic activities in clinical medicine.
Diseases usually have many causes, although occasionally one might
predominate. Often, several causes interact with one another in such a
way that the risk of disease is more than would be expected by simply
combining the effects of the individual causes taken separately. In other
cases, the presence of a third variable, an effect modifier, modifies the
strength of a cause-and-effect relationship between two variables.
Causes of disease can be proximal pathogenetic mechanisms or more
remote genetic, environmental, or behavioral factors. Medical interventions
to prevent or reverse disease can occur at any place in the development
of disease, from remote origins to proximal merh.misrns.
CHAPTER II/CAUSE 247
The case for causation is usually built over time with several different
studies. It rests primarily on the strength of the research designs used to
establish it. Because 'we rarely have the opportunity to establish cause
using randomized controlled trials, observational studies are necessary.
Some studies of populations (time series and multiple time series studies)
may suggest causal relationships when a given exposure of groups of
people is followed by a given effect.
features that strengthen the argument for a cause-and-effect relation-
ship include an appropriate temporal relationship, a strong association
between purported cause and effect, the existence of a dose-response rela-
tionship, a fall in risk when the purported cause is removed, and consistent
results among several studies. Biologic plausibility and coherence with
known facts are other features that argue for a causal relationship.
RI':FHI':bJ\CES
I. Holmes OW. Oil the contagious.ness of puerpelill fever. Med Classics 1936;1:207-2(-,8
IOriginally puhlishedI84.1.1
2. W"b~kr'~ ninth new collegi"te didionary. SrringH,·ld. r'ilA: Mcrriam-Web~ter, 199'1
.'. IliLlCM"hon B, Pugh TE, Epid"miol"gy' principles and methods. Bost"n: Littl", Ilmwn &
CO.,1'!70,
4. Cantwell MI;, Snider Dc. Cauthen Grv1. Onorato IIv\. Epidemiology of tuberculosis in the
Un;led States, 1985 through 1992. lAMA 199·1,272:535-539.
5. Wds SE, d al. Th" ,.ff"d of directly "bs"rved therapy (\11 thl' rat(,s of dmg rt'si~tan("(' ilnd
relapse in tuberculosis. N eng J Mcd 1,)()4;.l.lO:1l79-11li4.
h. Goldman L An,'lyzing the decline in the CAD death rail'. H05p I'ract 1988;23(2):109-
117.
7. Kimnl'l WB. I'rl'vl'ntivl' clrdiol"gy l'llstgr"d r-.kd '1977;6'1:74 as
Ii. Si'>Covick LlS, ct ul. Lliuretic therapy for hyperknsion and the risk of primary cardi"e
arrest. N Eng J I,led 1994;330:1852-1857.
'). O'Conner GT, Morton JR, Diehl \-11, Olmslead f'1\1, Coffin LH, Levy DC- Differences
bdw,'cn ml'n and wumen in Iw~pital mortality "ot'o("iil!t'd ,,--ilh "on>l1"ry artery bypass
graft ~urg('ry CirnllJtion 199,);88(1):2104-2110.
10. St. I."ger J\S, Cochra"" AI., Moor" E, I'adors as~"ciilted with cardi,,,, mortality ill devel-
oped countries with pJrticular rd"r"nee tn th" consumpti"n of wine, Lancet '1979; I,1017-
HJ20.
11 PeilT 511.1, William~(>l1 TH, Ilettin KM, Cl'rding DN, Gllgiill1lN. Decrease in no~ocomiJI
Clo,/ridilUlI difH('il"-asso"i,,tcd diilrrlw;, by r"stridillg din<\;Hnycin US", I\lln Inkrn M"d
1994;12ll:272-277.
12. Cervical Cancer Screening Programs. I. Epidemiology and natural history of carcinoma
"f th" """,'ix, {:;lI\ Med Ass'" J 197(';114:IOD.'··llFl3,
13. IJradford-J lill All. The "l1vironn",nt and dis"ase, a.%llciati"n Or causation? Proc R Soc
Med 1%5;5i'l:2'!5-301l.
14. Kuller L, Wing R. Weigh I loss and mortality. Ann Intern IVIed 1993;119:630-632,
15. Centers for Disease Control. Higlllight~ of the ~mgeon g<'lwral's report on smoking and
health, MMIVRI97<J;28:1-11
11,. 1J""sley 1,1', I.i" Cc, Hwal1g I.Y, Chit'n CS. H"patocellular c,m,inmna and hepatitis lJ
virus. Lancet 1<Ji'l1;2:1129-11J3.
17. Mandel JS, et al. Reducing mortality from cclorectal cancer by screening for fecal occult
blood. N Engl] Med 1993;.328:131'6-1371
18. Selby JV, Friedman GO, Quesenberry CP, Weiss NS. A case-control study of screening
slgrnotdoscopy and mortality from coiorectai cancer. N Engl J Med 1992;326:653-657.
19. Moertei CG, et al. A clinical triai of amygdalin (laetrile] in the treatment of human cancer.
N Eng] J Med 1982;306:201-206.
SUGGESTED READINGS
Buck C. Popper's philosophy for epidemiologists, lnt J Epidemiol 1975;4:159-168
Chalmers AF. What Is this Thing Calied Science? 2nd cd. New York University of Queensland
Press, 1982.
Department of Clinical Epidemiology and Bio~tatistics, McMaster University Health Sciences
Centre. How to read clinical journals. Iv: To determine etiology or causation. Can Med
Assoc J 1981;124:985-990.
Rothman KJ (ed). Causal inference. Chestnut Hill, MA: Epidemiology Resources, 1988,
12
SUMMING UP
Where is the knowledge we hmx: lost in informlltiol/?
T. S. Eliot, "The l{ock"
The usefulness of clinical research depends on its scientific credibility-

its believability to thoughtful, unbiased scientists-and its relevance to
the problems faced by clinicians and their patients. Both clinicians, who
base their decisions on the medical Iitcraturc, and researchers, who create
it, need to understand what adds to and subtrnrts from the strength of
scientific research.
To judge scientific credibility, readers must take an active role. They
must decide what they wanL to discover from the medicalliterature and
then sec if the information is present and meets their standards of scientific
credibility. By just reading: passively, without considering the basic scien-
tific principles systematically and ill advance, they will be less likely to
notice shortcomings and more likely to be misled.
This chapter describes how the mctbodologic principles discussed in
previous chapters can be applied by busy clinicians to the lifelong task of
trying to practice evidence-based medicine. First, we discuss how research
articles pertaining to a given clinical question are identified and how their
numbers can be reduced to manageable proportions without sacrificing
needed information. Next, we summarize basic rules for judging the
strength of individual articles; that section deals with concepts that have
been discussed throughout the book. Third, we consider how the many
articles on a given research question, as a group, are summarized to dis-
cover where the best available estimate of the truth lies. It is on this estimate
that clinicians must base their clinical decisions until better information
becomes available. Throughout the chapter we consider how these steps-
article identification, study evaluation, and evidence synthesis-relate to
strategies for keeping abreast of the literature throughout one's life as a
clinician.
Whatever the strength of the best available evidence, clinicians must
use it <IS a basis for action-sometimes rather bold action-yet regard it as
249
fallible and subject to revision. One scholar (1) has distinguished between
"decisions" and "conclusions." We decide something is true if we will act
as if it is so, for the present, until better information comes along. Conclu-
sions, on the other hand, are settled issues and are expected to be more
durable. Clinicians are mainly concerned with decisions. The integrity of
the scientific enterprise rests on the willingness of its participants to engage
in open-minded, well-Informed arguments for and against a current view
of the truth, to accept new evidence, and to change their minds.
Approaches to Reading the Literature

Clinicians examine the medical literature from different perspectives,
depending on their purpose. They browse to sec what is interesting, they
read articles of clinical interest to keep up, they look up the answers to
specific clinical questions, and they systematically review the literature
about a clinical issue to develop or change a clinical policy. We mainly
deal with the full review of the literature. We understand that clinicians
rarely have the time to do a full-blown review of existing information.
However, if they understand the basic principles by which literature
searches are done, they are in a better position to identify credible articles
efficiently and judge the results for themselves when they browse, keep
up, or look up information.
WHICH ARTICLES ARE IMPORTANT FOR CLINICAL
DECISION MAKING?
All articles are not equally important for clinical decision making.
Thoughtful clinicians must find and value the soundest articles in the face
of an almost overwhelming body of available information.
figure 12.1 summarizes <In approach to distinguishing articles of funda-
mental importance to clinical decision making from those that are not.
Many articles-reviews, teaching articles, editorials-arc written to de-
scribe what is generally believed to be true but are not themselves reports
of original research aimed at establishing that truth. These articles are a
convenient source of summary information, but they are interpretations
of the true knowledge base and arc not independent contributions to it.
Moreover, they are usually written by people with an established point of
view, so that there is the potential for bias.
Example How well do review articles and textbook chapters summarize
the available body of scientific evidence about a clinical question? Investiga-
tors produced estimates of the effectivenessof various interventions to reduce
morbidity and mortality from myocardial infarction (MI) by performing
meta-analyses (described later in this chapter) of randomized clinical trials
(r~CTs) (2). The estimates were compared with expert recommendations pub-
lished at the same point in time in review articles and textbook chapters.
They found that "expert opinion" generally lagged behind the cumulative
CHAPTER 12 / SUMMING UP 251
Contribution to Answering
the Clinical Question
T
Secondary
r t
Laboratory ("bench," T
Analogy
"basic") research
T
Weak
Primary
Direct
t
Strong
Figure 12.1. The literature on a research question: the relative value of various
kinds of articles for answering a clinical question.
evidence by several vears and not infrequently disagreed with it. For exam-
pie- by 1980 there were 12 RCTs in the literature that had examined the
efficacy of prophylactic lidocaine in the treatment of acute MI. Essentially,
all showed that treatment with lidocaine was no better and often worse than
placebo, yet the majority of review articles and chapters published during
the 19805 continued to recommend routine or selective use of lidocaine.
Other articles describe original research done in laboratories for the
purpose of understanding the biology of disease. These studies provide
the richest source of hypotheses about health and disease. Yet, "bench"
research cannot, in itself, establish with certainty what will happen in
humans, because phenomena in actual patients, who are complex organ-
isms in a similarly complex physical and social environment, involve vari-
ables that have been deliberately excluded from laboratory experiments.
Research involving intact humans and intended to guide clinical deci-
sion making ("clinical research") is, of course, conducted with varying
degrees of scientific rigor. Even by crude standards, most studies are rela-
tively weak. For example, a recent review of the methods of clinical studies
in three surgical journals revealed that more than 80% had no comparison
group, much less a randomized control group (3).
Throughout this book we have argued that the validity of clinical re-
search depends on the strength of its methods (internal validity) and the
extent to which it applies to a particular clinical setting (generalizability).
252 CLINICAL EPIDEMIOL.oGY
If this is so, a few good articles are more valuable than many weak or
inappropriate ones. Thus the overall conclusion frcrn the medical literature
often depends on how a relatively few articles are interpreted. A review
of the literature should involve selecting these articles carefully, identifying
their scientific strengths and 'weaknesses, and synthesizing the evidence
when their conclusions differ.
FINDING USEFUL ARTICLES
When systematically reviewing the literature, the first task is to locate

articles that may be useful. This is most challenging when reviewing the
literature, where it is first necessary to sort through a large number of
titles, often thousands, to find the small number of articles that are useful.
The objective is to reduce the literature to manageable proportions with-
out missing important articles. The task can be intimidating and time-
consuming. We describe a plan of attack, starting first with the full review
of the literature.
The first step is to develop a set of criteria for screening titles to select
articles that may be relevant while excluding a much larger number that
dearly are not. The criteria should provide a sensitive test for the articles
that one hopes to find in the same sense as a screening test should be
sensitive for a disease, i.e., few useful articles should be missed. Initially,
specificity can be sacrificed to achieve sensitivity, with the understanding
that it will be necessary to evaluate many "false-positive" articles in more
detail for each one that meets the final criteria. Often a useful screening
algorithm is defined by the joint occurrence of a few key words in the
title, e.g., sarcoidosis, pUlmonary, and corticosteroid or cancer, pancreatic. and
dillglwsis.
Second, the screening criteria are applied to a list of journal titles, gener-
ally the list maintained by the National Library of Medicine, MEDUNE.
Although available in bound volumes, many clinicians are currently ac-
cessing the medical literature electronically via modem, CD-ROM, or other
computer-based systems. Because computer searching usually misses some
important articles, one should also identify articles from other sources of
titles such as recent review articles, other articles on the same topic, text-
books, and suggestions from experts in the field. The result of this search
is a large number of titles, some of which represent relevant articles and
many of which do not.
Third, one must apply specific criteria to identify the articles that arc
actually appropriate for the question <It hand. Three kinds of criteria are
often used:
• Does the article address the specific clinical question that was the reason
for the search in the first place?
• Does the article represent original research, not secondary information

or opinion?
• Is the research based on relatively strong methods?
Many inappropriate articles can be excluded by examination of the full

title. Often, however, more information is required, and the abstract serves
this purpose well. Perusal of abstracts should reveal to the reader whether
a study of treatment used a comparison group and random allocation, a
study of prognosis was on an inception cohort, whether a study has suffi-
cient statistical power or precision, and so on.
Structured abstracts (4), which have been adopted by many of the leading
medical journals, provide a better opportunity to judge potentially useful
articles. The structured abstract summarizes in outline format those ele-
ments of a study-the research question, study design, setting, patients,
interventions, measurements, results, and conclusions-required to distin-
guish valid and informative studies from the larger number that are not
original or are inadequately rigorous.
Fina]]y, one must actually look at the articles that remain to see which
meet the final criteria. By this time, the number of articles should have
been reduced enough that the task is feasible. Figure 12.2 summarizes
these steps and illustrates the search process for a specific question: the
outcome of total knee replacement (5).
If there is not sufficient time for a full, broadly based search for articles
or the reader is browsing or trying to stay abreast of important develop-
ments, the early steps of this process must be abbreviated. One can examine
only those journals that publish original research with high methodologic
standards. However, this is an insensitive strategy: one would have to
examine at least 11 of the world's best journals just to find 80% of the best
articles on a question (6).
Another approach, is to have the screening of articles done by others.
One would want them to be experts in both clinical medicine and clinical
research methods, to examine all of the world's articles, and to make their
criteria for inclusion explicit. The journal ACr Journal Club presents struc-
tured abstracts for the scientifically strong, clinically relevant original re-
search in internal medicine, selected by explicit criteria published in each
issue. The results of selection in this way are powerful: in 1993, there were
more than 6 million articles published in clinical journals, of which about
350-a manageable number-met the criteria. Another option, the Cocb-
rane Collaboration, is being developed. Expert groups from throughout
the world are working together to select the best studies of clinical interven-
tions, summarize them in a standard form, disseminate the information
electronically, correct the database when errors arc found, and keep it up
Articles Reasons for

Remaining Exclusion
OOOOD 1631
OODDD 369 no original dates
~
DOODD 362 non-English language
D 70 with 10 or fewer patients
27 no outcomes reported
DDOOD 803
DDD
~
115 not a cohort undergoing the replica
96 no pertinent outcomes reported
37 different surgical procedures
DOODD 553
~
336 inadequate outcome assessment
DD 217
DO
~
130
87 other exclusions
Figure 12.2. Literature search: identifying the few most important articles from the
medical literature as a whole, (Callahan CM, Drake BG, Heck DA, Dittus RS. Patient
outcome following Iricompartmental total knee replacement. JAMA 1994; 271 1349-
1357,)
to date. The results, <It least for some clinical questions, should be available
in the rrud-Issus.
Judging Individual Articles

Once individual articles of interest have been identified, the next task
is to evaluate the quality of the evidence they contain. The approach uses
the scientific principles described in this book.
CHArTeR 12 i SUMMING UP 255
We have outlined basic criteria for the scientific credibility of specific

research designs in the appendix to this chapter. We summarized a set of
questions that should be asked of most studies-by investigators and read-
ers alike. They concern the nature of the research question, the generaliz-
ability and clinical usefulness of the results, and two processes that can
affect internal validity: bias and chance. We considered ground rules for
studies of diagnosis, frequency, prognosis, treatment, and cause. The ap-
pendix describes basic issues that should be considered when deciding
whether an article might be sufficiently strong to be useful in a literature
review and in setting criteria to assign weights to articles when preparing
a synthesis of the results of several articles.
DOES THE DESIGN FIT THE QUESTION?
One cannot speak of "good" or "bad" research designs in general with-

out reference to the question they are intended to answer. Many clinically
oriented methodologlc assessment schemes give lower grades to observa-
tional designs such as the prevalence survey. This may be justified if the
clinical question concerns preventive or therapeutic interventions, but in-
appropriate if considering studies of diagnostic tests. Table 1201 matches
clinical questions to the best research designs used to answer them. The
table is meant to offer a guideline; it should not preclude creative but
scientifically sound approaches other than those listed. For example, the
best available evidence that periodic screening sigmoidoscopy may reduce
deaths from culorectal cancer came from a rigorous case-control study (7).
Because of the large numbers of patients and lengthy follow-up period
required to test the efficacy of sigmOidoscopy in a randomized trial, this
may be the only form of evidence available for some time.
Table 12.1
Matching the Strongest Research Designs to Clinical Questions
Oue-uon
Dtaqnose Prevalence
Prevalence Prevalence
Incidence Cr)~IOr1
Hisk Cohort
Case control
PIO~_JrI{)sis Cohort
T reatment CliniCdl trial
I'reven1ion Clnucal tnal
Cause Cohort
Case control
Table 12.2
Characteristics of a Study That Detennine Whethe, It Can Test
or Only Raise Hypotheses
Characteristic
Dcsiqn Weak Strong

Hypotheses None (or otter data collected Stated before study begun
ano clrlalysed)
Comparisons Many Few
p value Laroe Small
Results confirmed on separate No Yes
data set
RAISING OR TESTING A HYPOTHESIS

The conclusions of an individual piece of research fall on a spectrum
of believability according to the decisiveness of the scientific strategy used
(see Chapter 11). At one end of the spectrum are reports that only suggest
relationships. albeit potentially useful ones, without putting these ideas to
the test. Most case reports serve this function. The conclusions of these
studies are tentative; many are later refuted. At the other end of the spec-
trum are studies-e.g., large randomized controlled trials-that have put
ideas to a rigorous test. Conclusions from these studies are more definitive.
Most studies fall between these extremes.
A priori hypotheses are important. Without them, false-positive findings
can make their way into the literature in the following way. Suppose one
examines a large number of variables in a data set none of which is associ-
ated with any of the others in nature. As discussed in Chapter 9, if a large
number of associations between variables are examined, some of them will
be extreme enough to appear "real." even though the associations are only
by chance. At a conventional level of statistical significance, p < 0.05, about
I in 20 such comparisons will be statistically significant, by definition. Of
course, the observed associations are "real" for the particular data set <It
hand-but not necessarily in the population-because the current sample
may misrepresent all such samples from the population of interest.
Now suppose that one of these comparisons is selected out of the larger
set of all possible comparisons and given special emphasis, perhaps be-
cause it fits well with existing biomedical theories. Suppose the other com-
parisons are minimized in the final report. Then the association, taken out
of context. can <lppe<Jr very important. This process-random (chance)
occurrence of associations followed by biased selection of interesting
ones-is not unusual in published research.
There are several dues that signal the degree to which a given study is
hypothesis testing rather than hypothesis raising (Table 12.2).
The first, a strong research design, is not a strictly separate factor from
the others. Making hypotheses in advance and limiting the number of
comparisons examined reduce the number of apparently "significant"
comparisons that emerge from a study. The exploration for effects in vari-
ous subgroups of a larger study population is a common analytic strategy
that may result in chance or spurious associations. When hypotheses made
in advance, a priori hypotheses, are confirmed, one can place more confi-
dence in the findings. Alternatively, investigators can simply limit the
number of comparisons made after the fact, so that there is less chance of
false-positive findings for the study as a whole. Or they can insist on a
particularly small p value before ruling out the role of chance in explaining
particular findings.
Another strategy to protect against the acceptance of spurious or chance
associations is to raise hypotheses on one set of data and test them on a
separate one (fig. 12.3). The availability of large data sets and statistical
computer software makes it relatively easy for the analysis to include
multiple variables, considered either separately or together in models. The
analysis of multiple variables should be viewed as raising hypotheses, as
the investigators rarely specify in advance what the model will find, much
less the weight given to each finding. 1£ the data set is large enough. it can
be divided randomly in hillf, with one half being used to develop the
model and the second half used to confirm it. Or it can be tested in a
different setting. This latter process is illustrated in the Following example.
Example Investigators developed an index, including seven physical

signs, for predicting the carty recurrence of acute asthma after discharge
from 'In emergency department (tl). Among 20.S patients at the investigators'
medical center, patients from whom the index was developed, the index had
a sensitivity of 9.S':,;, and a specificity of 97'1;,. The results were so striking that
the index began to be put into clinical practice elsewhere.
Liller, two other groups of inVl'stigiltors independently tested the index
in other settings (9,1(]). The results were disappointing. The sensitivity and
sped ficity were 40% and 71 '~;" respectively, in one study, and 1K.l)~;) ami
82.4'::, in the other.
Process Hypothesis Hypothesis

Raising Testing
Data Set Derivatidn Validation
Figure 12.3. Developing a hypothesis on one data set and testing it on another.
These studies illustrate the dangers of placing too much confidence in

a relationship that has been suggested in one data set but not tested in
another, independent one. There are several possible reasons for the differ-
ence in performance. Patients in the original sample might have been sys-
tematically different, the index might have been applied differently, or
chance might have resulted in unusual findings in the initial study.
Whatever the strategies used to increase the hypothesis-testing character
of a study. it is the author's responsibility to make it clear where a particular
study stands on the hypotheses-raising, hypothesis-testing: spectrum and
why. The readers' task is to seek out this information or reconstruct it, if
it is not apparent. However, one should not eschew studies that mainly
raise hypotheses; they are important, just not definitive.
Summarizing the Results of Many Studies

The current state of knowledge on a question is usually decided by the
pattern of results from all studies addressing the question, rather than by
one definitive study. Until recently, the commonest way of establishing
this pattern was by implicit judgment, i.e., opinion, without having stated
in advance the ground rules by which the contributions of individual
studies would be weighted. Judgments of this sort often take the form of
a traditional ("narrative") review article by an expert in the relevant field
or a consensus of scholars representing the many points of view that bear
on a question, e.g., the National Institutes of Health's Consensus Develop-
ment Conferences.
A variety of more structured methods of summarizing published re-
search is now used. These methods have the advantage of making explicit
the assumptions behind the relative weights given to the various studies.
They also follow the scientific method more directly; setting criteria in
advance, gathering data (in this case, the results of individual studies),
analyzing the data, and allowing the conclusions to follow from the criteria
and data.
The process of summing up the research on a question, using structured
methods, is referred to as meta-analysis-literally, analysis of analysis-or
information synthesis. This approach is particularly useful when there is one
specific question and at least a few relatively strong studies with apparently
different conclusions. The use of these methods has exploded in the last
few years. MEDLTNE listed nearly 2000 articles under the subject heading
"meta-analysis" between 1990 and mid-1994.
There are three general steps in performing a meta-analysis. first is to
identify the best articles from all possible articles, as described earlier in
this chapter. Second is to evaluate each study according to how well it
meets methodological criteria, which are decided on in advance. In some
meta-analyses. this evaluation results in the assignment of an overall qual-
CHAPTER 12 I SUMMING UP 259
Relative Risk
Type of Study or Odds Ratio
Treatment Treatment
- Better Worse --.---.
Randomized trials
RCT 1 0.41 x
RCT2 0.20 x
RCT3 0.26 x
RCT 4 0.24 x
RCT5 0.20 x
RCT6 1.01 x
RCT7 0.63
Nonrandomized trials
Study 1 0.80 X
Study 2 0.46 X
Study 3 0.25 X
Study 4 1.56 X
Study 5 0.71 ·x
Study 6 0.98 X
Overall relative risk

95%CI
0.49
0.34--0.70
•
10 Case-control studies
Overall odds ratio
95%CI
0.50
0.39-0.64
•
0.5 1.0 1.5
Figure 12.4. Results of a meta-analysis of the effectiveness of BeG vaccination
to prevent tuberculosis. C/, confidence interval. (Based on Colditz GA Brewer TF,
Berkey cs, Wilson ME, Burdick E. Fineberg HV, Mosteller F, Efficacy of BeG vaccine
in the prevention of tuberculosis: meta-analysis of the published literature. JAMA
1994;271 :698-702.)
ity score; in others, quality-related study characteristics of design, number

and source of patients, and data collection methods are considered sepa-
rately. The third step is to summarize, with numbers, the results of many
studies to form, in effect, one large study with more statistical power than
any of the individual studies alone. Each individual study is weighted by
its sample size, i.c., large studies get more weight than ones with smaller
260 CLINICAL EPIDeMIOLOGY
numbers of patients. In addition, many meta-analyses also include the

quality score as a weighting factor. Statistical methods. usually some form
of regression analysis, arc then used to estimate an overall effect measure,
such as a relative risk or percentage reduction in mortality. Reports of
meta-analyses include graphical displays of the results of the individual
studies as well as the overall measure of effect.
Although meta-analysis has become the standard against which other
approaches to literature synthesis are judged, there continues to be contro-
versy about many of its dements, particularly the evaluation of quality
and its inclusion in the overall assessment. But quality measures, while
lumping disparate metbodologtc features in a single number, may help
explain differences among studies.
Example Although Bacillc Calmettc-Cuerin (HeG) vaccinehas been used

to prevent tuberculosis for more than 50 years and is required in many
countries, iIs efficacy is controversial. In part this is because the several Iarge-
sealcclinical trials to evaluatc Bee; have reported conflictingresults. An early
meta-analysis compared the methods used in these trials to their r<-'sults (11 j.
The investigators found that the unbiased detection of tubcrcutos», in Bee
and control grnups was available only for the three trials reporting 75'1,. or
greater efficacy.
A more recent meta-analysis of the same question summarized the re-

sults of studies examining vaccine efficacy (12). figure 12.4 shows the
results of the seven randomized trials, six nonrandom trials, and the overall
findings of the 10 case-control studies published as of 1YY4. Overall, in
studies using each of the three designs, the risk of tuberculosis was found
to be reduced by about half for those receiving the vaccination compared
with those who did not. To help explain. differences in observed magnitude
of effect, the investigators developed overall scores fur the quality of each
study's methods. Using regression analysis, they found that better quality
scores predicted findings of greater vaccine effect. The meta-analyses
clearly establish the efficacy of BCe vaccine.
Often, however, there is no dear relationship between global quality
ratings and their results. In this case, the meta-analysts must look at the
specific mcthodologtc features of studies to see why they are reaching
disparate conclusions (13).
POOLING
Not uncommonly, the results of various individual studies are indeci-

sive because each study describes too few patients or too few outcome
events to have sufficient statistical power or precision. Consequently, esti-
mates of rates from these studies are unstable, and each study's comparison
of rates runs an unacceptably high risk of missing true effects (Type II
error).
Pooling refers to the process of aggregating the data from several rela-
tively small studies of the same question to form, in effect, one large one.
It is permissible when it can be shown t11M the studies are sufficiently
similar to each other (in patients, intervention and outcome measures) to
treat them as if they are part of a Single study. Pooling attempts to assemble
enough observations to generate a precise overall estimate of effect, not to
account for differences in conclusions among studies. The advantage of
pooling is that it can result in adequate xtatisfica 1power to detect meaning-
ful differences, if they exist. Pooling is particularly useful when the disease
and/or the outcome events of interest occur infrequently. Under these
circumstances there are no other feasible ways to achieve statistical power.
Example There arc many reports of peptic ulcer disease during cortico-
steroid therapy. Yet, it has been difficult to establish by means of observa-
tional studies whether corticosteroids cause ulcers, because manv or the situa-
tions in which they are given-e.g., during stress and in conjunction with
gastric-irritating drugs-may themselves predispose to peptic ulcer disease.
Abo, ulcers may be sought more diligently in patients receiving corticoste-
roids and go undetected in other patients.
Randomized controlled trials are the best wav to determine cause and
effect. There have been manv randomized trials in ;....hich corticosteroids were
used to treat various lunditiolls and peptic ulcer disease was a side effect.
None of these studies was large enough in itself to test the corticosteroid!
ulcer hypothesis. But together they provide an opportunity to examine the
rate of rare event.
In one review of 71 controlled trials of corticosteroids in which patients
were randomized (or its equivalent} and peptic ulcer disease was considered,
there were about 86 patients and 1 case of peptic ulcer disease per study;
only 31 of the trials reported any patients with ulcers (14). The investigators
pooled the results of these 71 trials to Increase statistical power. In the pooled
study, there were 6111 patients and about 80 ulcers. The rate of peptic ulcer
disease was 1.8 in the corticosteroid group and 0.8 in the control group
(relative risk, 2.3; 95°;;. confidence interval, 1.4-3.7). The results were similar
when examined separately according to the presence and absence of other
risk factors; various doses, routes of administration, and duration of therapy;
and whether the disease was suspected, defined as bleeding, or specifically
diagnosed.
Thus the combined results of many studies, each with relatively sound
design but too small to answer the question, gave sufficient statistical power
to detect risk.
Advocates of pooling point out that examination of the pattern of evi-

dence, effectively summarized, can give new insights into the strengths
and weaknesses of the evidence. For example, a single figure can show
the number of strong studies. the point estimate and statistical precision
of each study's observed effect size, the relationship between effect size
and precision, and the point estimate and precision of their pooled effect
(see Fig. 12.4).
262 CLINICAL EPIDtMIOLOGY
Opponents of pooling argue that the ways in which patients, interven-

tions, and outcomes were selected in the various studies are so dissimilar
that it is not reasonable to combine them. "Splitters" are not as satisfied
with pooling as "Jumpers." Also, pooling deals only with statistical power.
It does not correct for whatever biases existed in the designs of the various
individual studies, nor can it be assumed that these biases cancel each
other out when the studies are aggregated. In any cast', meta-analyses only
supplement and do not replace the insights gained by examining each of
the best studies of a clinical question carefully.
Publication and Bias

Clinicians prefer good news, as does everyone else. Thus such words
as t:ffiCllcy, predicting, and correlation are the order of the day in journal
titles. It is considerably less appealing to contemplate things that do not
work. In fact, such observations arc often considered failures. Researchers
with the bad fortune to make such observations arc likely to be advised
by their friends, with gentle malice, to seck publication in the /ol/I"IUlI (~f
Negative Results.
It may be that our penchant for positive results leads to bias in the kinds
of articles selected for publication in medical journals.
Example The final disposition of 285 studies that had been reviewed by
an English Research t.uucs Committee and broughr to conclusion hy the
investigators was studied (15). Steflsticallv significant results were found in
54')'" of studies. a nonsignificant trend in 16'(;" and null results in .'I()':;,. Of the
studies with significant results, 85% were either published or presented as
opposed to only 56':;, of studies with neg;ilivc results (odds ratio. 434; 95':;,
confidence interval, 2.4-R6). Studies with nul! results were not of poorer
quality, nor were they 1l101·e likely 10 be unpublished because of editorial
rejection.
Articles actually reaching publication arc a biased sample of all research

findings, tending to represent efforts to find causes, diagnostic tests, and
treatments as being more effective than they actually are. For example, a
meta-analysis of the relative effectiveness of single versus multiple drugs
in ovarian cancer found a large survival advantage with multiple drugs
in published data. When the investigator (16) added unpublished results,
the difference disappeared. There is no reason to assert that biased judg-
ments arc made deliberately. Everyone does his or her part to put the
"best" work forward, but publication is not a random process. There are
forces favoring positive over negative results that are quite independent
of the relative proportions of these results among all research projects
undertaken. Readers should be aware of this bias lest they become unrealis-
tically impressed with the many new and promising findings that appear
in medical journals.
CHAPTER 12 ! SUMMING UP 263
One way to avoid this bias is to give more credibility to large studies
than to small ones. Most large studies, having required great effort and
expense in their execution, will be published regardless of whether they
have a positive or negative finding. Smaller studies, on the other hand,
require less investment and so are more easily discarded in the selection
process.
DifferCllt Answers: The Same Questions?

Until now, we have emphasized how studies can come to different
conclusions because they have different methods, better for some than for
others. Rut there is an alternative explanation: The research questions,
although superficially the same, may actually be fundamentally different.
Rather than one or the other study being misleading, both might be right.
It may be that human biology, not research methods, accounts for the
difference.
Example Several authors (17,lH) have performed meta-analyses i1S-
sessing the effectiveness of drug treatment for hypertension. Early trials dem-
onstrated substantial and st,ltistically significant reductions in stroh'S but
smaller and often not statisticallv significant reductions in coronary heart
disease (CHD) (Fig. 12.5). The confidence interval for the pooled relative risk
for (HD across JII trials includes I.
More recent trials have focused on or at least included older adults. These
newer trials, also summarized in Figure 12.5,showed the same degree of risk
reduction for stroke but larger and consistently significant reductions in the
risk of coronary heart disease. While the larger effectiveness of treatment
observed in more recent trials suggests that drug therapy for hypertension
is more effective in older patients than was previously believed, these newer
trials tended to use diuretics and beta-blockers, the principal drugs in the
trials, in lower doses, This may also have contributed to the greater effect of
treatment on (HD.
Studies of cause and effect that seem to be asking similar questions

can in fact present different questions in at least four ways: The patients,
interventions, follow-up, and end results may not be the same. Differences
among studies in anyone of these may be enough to give different results.
Other Sources of Information

Until now the main source of information we have considered is journal
articles reporting original research and meta-analyses based on them. What
about other sources of information?
Textbooks are convenient and trustworthy for reporting well-estab-
lished facts. But they have the disadvantage of being out of date (as much
as 1 year old at time of publication) and reflecting the opinions of single
authors, with little external review. Colleagues, espeda11y those specialized
in the area of the clinical questions, an- also practical sources of informa-
Relative Risk (and Intervals)

Outcome 95% Confidence
Treatment Treatment
- better worse - -
Stroke (fatal and nonfatal)
Middle-aged subjects
HOFP 0.64 x
MRC·younger 0.55 :X
10 other trials 0.67 X
Pooled RR 0.62
(0.53--{).73) +
Older subjects
STOP 0.55 X
MRC-older 0.76 X
SHEP 0.65 X
3 other trials 0.43 X:
Pooled RR 0.61
(0.53-0.70)
:+
Coronary heart disease

(fatal and nonfatal)
Middle-aged subjects
HOFP 0.90 X
MRC-younger 0.94 X
10 other trials 0_92 X
Pooled RR 0.92
(0.82-1.04)
+
Older subjects
STOP 0.90
X
MRC-older 0.74 X
SHEP 0.82 X
3 other trials 0.76
Pooled RR 0.78 X
(0.68-0.90) +
0.5 1.0 1.5
Figure 12.5. Results of a meta-analysis of the efficacy of hypertension control on

the risk 01 stroke and coronary heart disease. (Adapted from Cutler JA, Pstay 8M,
MacMahon S, Furberg CD. Public health issues in hypertension control. What has
been learned from clinical trials? In: Laragh JH, Brenner 8M, eos. Hypertension:
physiology, diagnosis, and management. 2nd ed. New York: Raven Press, 1995.)
CIIAPH:H 1:.' / SUMMING UP 265
tion, but their opinions <Ire only as good <IS the consultant, who m<lY be
biased by the beliefs and financial interests of his or her field. For example,
it is natural for gastroenterologists to believe in endoscopy more than
radiologic contrast studies and surgeons to believe in surgery over medical
therapy.
A growing number of databases are complete, up to date, and widely
available by telephone, fax, floppy disks, CD-ROM, and e-mail, the In-
ternet, and comptcr bulletin board. Examples include a 24-hr telephone
connection to the Centers for Disease Control and Prevention for informa-
tion about disease prevention offered to those traveling to any part of the
world; To-cline for information on poisonings; PDQ for current recommen-
dation for cancer chemotherapy; and an array of databases on drugs, their
toxicities, and adjustment of dose in renal failure. These databases contain
information that is essential to the practice of medicine but are too infre-
quently needed and too extensive for clinicians to carry around in their
heads. Clinicians should find ways to access them in their location. 'Ihey
should also usc these databases with the lessons of this book in mind: The
data are only as good as the methods used to select them. Many of the
databases, such as guidelines of the Agency for Health Care Policy and
Research, the U.S. Preventive Services Task Force, and the American Col-
lege of Physicians, are created by excellent methods and make the process
clear. Some are the results of individuals or industries with conflict of
interest, and they should be used with skepticism.
Clinical Guidelines
Throughout the book we have argued that clinical research provides
the soundest grounds for establishing one's approach to clinical practice
and making decisions about patients. The shift away from anecdote and
personal experience has been called "evidence-based medicine" (19). An
important element in evidence-based med icinc is the translation of research
findings into clear. unambiguous recommendations for clinicians. Practice
guidelines are systematically developed statements to assist clinicians in
deciding about appropriate health care for specific clinical problems (19).
Their development and use are now commonplace in many organized
medical settings. At their best, the validity of guidelines is established by
including in the panel that prepZlres them people who represent all relevant
aspects of the question (ranging from highly specialized researchers to
clinicians, economists .and patients) to cover all important aspects of the
question and to balance, if not eliminate, the vested interests of anyone
or another participant. The best guidelines are based on research evidence,
not just expert opinion, and so often usc formal processes of literature
review and synthesis, as described in this chapter (20).
Guidelines are meant to guide, not prescribe clinical judgment There

are good reasons not to follow guidelines in the care of some individual
patients.
Do guidelines change physicians' behavior? 1\ meta-analysis identified
and summarized 59 published articles that evaluated the impact of explicit
guidelines using more rigorous research designs-randomized trials, non-
randomized comparative trials, and interrupted time series designs (21).
More than 90% of the studies demonstrated significant changes in care in
accordance with the guidelines, and 9 out of the 11 studies examining
patient outcomes showed improvements.
REFERRNCES
I. Tukey IW. Conclusions ~l1d decisions. T"dmomelrics IY60;2:42.1-.J33.
2. Antrn~n EM, Lau J, Kupelnick B, Mosteller 1', Chalmers 'l'C. A comparison of n'~ults of
mct~-,maly~es of randomized trials and recommendalions of clinical experts: treatment
of myocardial infarction. JAMAI992;26lJ:240-241i.
3 Soloman Mj, McLeod RS. Clinic~1 studies in surgic~1 joumals-c-havc 'A'" improved: Dis
C"lon Rectum 1993;36:43-48.
4. H~ynes RB, Mulmw CD, Huth EJ, Altm~n OC, Cardncr M]. More informative abstracts
revisited. Ann Intern Mcd 1990;113:69-75.
5. CallahJn eM, Drake Be, H('('k DA, Dittus RS. Patient outcome following tnrompnrtmcntal
total knee replacement. A mctd-an~lysis, JAMA 1994;271:134'1-1357,
6. Haynes RB. Where'~ the meat in dinicnl journals? [Editorial]. ACI' J Club 1993;i\iov-
DecA22-A23. [Originally published in Ann Intern Mcd J993;114(Suppl J).]
7. Selby JV, Friedman GD, Qu<,s"nherry cr Ir, weiss NS, A case control ~t\ldy of screl'ning
sigmoidoscopy and mortality from colorectal cancer. N Engl I Mcd 1992;32fi:653-657,
Ii. Fisrhl MA, Pitchenik A, Gardner LB, An ind",x predicting relap~e and need for h()spit~li7a
tion in patients with acute bronchial asthma. N Eng! J Mcd 19RI;.105:783-7!l9.
9, Rose Cc, Murphy JG, SchwMtz jS. Performance of an index predicting the response of
patients with acute bronchial asthma to intensive emergency department treatment. N
Engl J Med 19H4;31O:573-577.
10. Centor RM, Yarbrough B, Wood [I', Inability to predict feldPSe in acute asthma. N Engl
J l\1<'d 1984;310:577--'>1>0.
II. Clemens [D, Chuong jH, Feinstein AR The l:lCCcontroversy: a methodological and statis-
tical reapprnis<lL JAMA 191>3;249:2%2.
12 Coldit>: GA, Brewer TF, Berkey CS, Wilson ME, Burdick F, Fineberg IIV, rVlos!eller F.
Ufica(y of BCG vaccine in the prevention of tuberculosis: meta-,mil[v~i5 of the published
lilerature. lAMA 1994;271:69!l 702.
IT Greenland S. A nitical look at some popular meta-analytic mel hods [Invited Commen-
tdryl. Am I Epidcnuol 1994;140:290-2%,
H. Messer I. R"itmnn D, Sacks I IS, Smith H jr, Chalmers TC Association of udrenocorticostc-
roid therapy and peptic-ulcer disease. N Eng! J MedI983;309:21-24.
15, E~st('rhrook I'J, l:lerlin JA, Copalan K Matthews DR. Publication hias in clinicalres"arch.
L,mcd 1991;337:867-1>72,
16. Simes RJ. Confronting public~tion bias: a cohlll't design for meta·~naly~i~. Slat .'VIed
1987;6:]]-29.
17. Cutler JA, I'sclly flM, Macfvlahon S, Furberg CD, I'ublic health issues in hypertension
controL whal has b<'en learned from clinical trials? In: Laragh JH, Brenner l:lM, (,d~.
Hyp<,rten~ion: physiology, diagnosis. and lllJnag"ment. 2nd cd. New York: Raven Press,
1995.
18. Hebert PI{, Fiebach NH, Eberlein KA, T<lylor jO, Hennekens CH. The community ba~"d
randomized trial~ of pharmacological tr,'atment of mild-tlHnoderate hyperten~ion. Am J
EpidemioI19HR;127:581-589.
19. Evidence based medicine working group. Evid"nce based medicine: a new approach to
teJching the practice of m,,,licine. JAMA 1992;26H:242U-2425.
20, Hayward [{SA, Wilbon MC, Tunis SR, BJ~~ EI), Rubin HR, Haynes RI), More informiltive
ab~tracts of articles d",~nibing clinical pra,·ti"" guidelines. Ann Intern Med 1993;118:731 ~
737.
21. Crimshaw JM, Russell IT. Effect of clinical guidelines on medical practice. A systematic
review "f rigorous evaluJtion~. Lancet 1993;342:1117-1321.
SUGGESTED lUiADING
Epidemiology Work Croup of the lnkragency Regulatory Liaison Group. Cuiddines [or
documentation of epidemiologic studies, Am J Fpid."ninl lY81;114:6O'J-713,
Goldschmidt Pc. Information ~ynthesis: a practical guide. Health Services I,,,, lY!l6;21:215-237.
Haynes Rll McKibbon KA, Fit~.gerJld D, GllyJll CI-l, Walker C/,SJck"tt OL. How to keep up
with the medical literature. I: Why try to 1<",,1" up and how to get started. Ann Inkrn Mcd
IYH6;105: 149-153.
Haynes RB, McKibbon KA, Fitzgemld 0, Cuyatt GH, Walk"r Cj. Sackett m. How to keep
up with the m.,dical literature. 11: I:kciding which jOlJrnJls 10 read regulJrly. Ann Intern
Med 19R6;10S:JIIY-312.
Haynes RB, Mckibben KA, Fitzgerald D, Cuyatt CH, Walker C/, Sackett OL. How to keep
up with the medical lit('rature. III: Expanding the number of journals you read regubrly.
Ann Intern )vl",JIYH6;105:474-47R.
Ilaynes RB, McKibbon KA, Fitzgerald D, Cuvatt GH, Walker Cj. Sackett 01 .. How to keep
up with the medical literaturc. IV: Using th., literature to solve clinical problems. Ann
Intern Mcd 1986;105:636· MD.
Haynes RB,\·kKibbon KA, Walker Cj. Mousseau J, Baker I.M, Pitzgcrald D, Gl-lyatt C, Norman
CR Computer searching of th~ medical literature: an evaluation of MFDLlNE searching
systems. Ann Intern M",d 19H5;lIlJ:!l12-8J6.
Irwig I" 'tostcson AN, Gastonis C Lau J, Colditz C, Chalmers TC. Mostcllar f. Cuiddines
for meta-analyses eVilluating diagnostic te~ts. Ann Intern Med lYY4;120:667-676.
I 'Abbe KA, Detsky AS, O'Rourke K. Meta-analysis in clinical research, Ann Intcrn Med
1987;W7:224-2J1
Light RJ, Pillmer Oil. Summing up, The science of reviewing research. Cambridge. MA:
Harvard University Press, 19H4.
Mulrow CU, Stale of the ~;cience [The Medical Review Article]. Ann Intern Med 1987;106:485-
488.
Thompson SC, Po<ock 5J. CJn meta-analyses be lru~ted? Lancet 19YI;338:1127-IUD,
APPENDIX 12.1. BASIC CUlIJEUNES POR lJETERMININC THE VALIDITI

OF CUNICAL STUDIES
ALL STU1)1 ES
1. What kind of clinical question is the research intended to answer?
The research design should match the clinical question (see Table 12.1)
2. What iatiente. variables, and outcomes were studied?
These determine the generallzability of the results.
3. How likely is it that the findings arc the result of bias?

Systematic differences between compared groups (e.g.. in patients' char-
acteristics, interventions, or risk factors: outcomes; or measurement
methods) diminish internal validity.
4. How big is the effect?
Clinical decisions depend on the magnitude (not just on the existance)
of effect.
5. How likely is it that the findings occurred by chance?
Clinicians need to know the range of values within which the true
effect is likely to fan (confidence interval) or (less useful) how likely the
observed effect is by chance alone (p value for "positive" results and
power for "negative" results).
STUDIES OF DIAGNOSTIC TESTS

1. Is the test clearly described (including the point at which it is considered
abnormal)?
ff the test result can take on a range of values, the performance varies
according to the choice of cutoff point.
2. Is the true presence or absence of disease (gota standard) established for
all patients?
It is possible to know all important aspects of test performance only if
there are data for all four cells of the 2 by 2 table.
3. Does the spectrum (~f patients with and without disease match the charac-
teristics of patients for whom the test 'will be used?
Sensitivity is often affected by the severity of disease and specificity by
the characteristics of those in the study without the disease.
4. Is there an unbiased assessment of test and disease status?
Bias can occur if the test result is determined with knowledge of disease
status and vice versa.
5. Ts test perjonnance summarized by sensitivity and specificity or likelihood
ratio?
This information is needed to decide whether to use the test.
6. For tests with a range of values. how does mOI,ing the cutolfpoint affect
test performance?
The information conveyed by the test depends on the degree of
abnormality.
7. 1£ predictive value is reported, is it in relation to a clinically sensible
prevalence?
Predictive value depends on prevalence (as well as the sensitivity and
specificity of the test). If people with and without the disease are chosen
separately, without relation to the clinicaly occurring prevalence, the
resulting predictive value has no clinical meaning.
PREVALENCE STUDIES
1. What arc the criteria for being a case?
Prevalence depends on what one calls a case.
2. In what population arc the cases found?
Prevalence depends on the group of people in which it is described.
3. Is prevalence described for an unbiased sample of the population?
Prevalence for the sample estimates prevalence for the population to
the extent that the sample is unbiased.
COHORT STUDIES
1. Arc all members of the cohort:
a. Entered at the beginning of follow-up (inception cohor!)?
Otherwise people who do unusually well or badly will not be
counted in the result.
b. At risk for developing the outcome?
It makes no sense to describe how outcomes develop over time in
people who already have the disease or cannot develop it.
e. At a similar point (zero time) in the course of disease?
Prognosis varies according to the point in the course of disease at
which one begins counting outcome events.
2. Is there complete follow-up on all members?
Drop-outs can bias the results if they on average have a better or worse
course than those who remain in the study.
3. Are <Ill members of the cohort assessed for outcomes with the sameillll'nsity?
Otherwise differences in outcome rates might be from measurement
bias, not true differences.
4. Are comparisons unbiased? (would members of the cohorts have the same
outcome rate except for the variable of Interest")
To attribute outcome to the factor of interest other determinants of
outcome must occur equally in the groups compared.
KANDOMIZED TRIALS
1. Are the basic guidelines for cohort studies satisfied?
Clinical trials are cohort studies
2. Were paiierus randomly al/ocaled to treated and control groups?
This is the only effective way to make a completely unbiased comparison
of treatments.
3. Were patients, caregivers, and researchers unaware of the treatment
group (masked) to which each patient belonged?
Masking participants in a trial helps assure that they are unbiased.
4. Were cointenxnsions the same in both groups?
Treating patients differently can destroy the comparability that was
achieved by randomization.
5. Were rcsu Its described according to the tremment allocated or the treatment
actually received?
If not all patients receive the treatment assigned to them there are two
kinds of analyses with different objectives and scientific strengths. "In-
tention to treat" analyses arc for management decisions and are of the
randomly constituted groups. "Efficacy" analyses <He to explain the
effect of the intervention itself, are of the treatment actually received,
and are a cohort study.
CASE-CONTl~OL STU 1)1 ES
1. Were cases entered at the onsl'! of disease?

Risk factors for prevalent Gl[;CS may be related to onset or duration of
disease.
2. Were controls similar to cases except for exposure?
A valid estimate of relative risk depends on an unbiased comparison.
3. Were there similar and unbiased efforts to dl'kcf exposure in Glses and
controls?
Biased measurement of exposure can increase or decrease the estimate
of relative risk.
MET A-ANALYSFS
1. Is all relevant research (both published and unpublished studies) found?
The objective is to summarize the results of OJ]] completed research, not
a biased sample of it.
2. Does the meta-analysis include only sciflltijically strong studies (those
with a low probability of bias)?
The objective is to summarize the most credible evidence.
3. Tf a summary estimate of effect is calculated
a. Are the studies homoeeneau« (are patients, interventions, and out-
comes similar)?
It is inappropriate to seck a single, overall measure of effect from
inherently dissimilar studies.
b. Are the studies weigh/I'd by their size?
Larger (more precise) studies deserve more weight than smaller (less
precise) ones.
4. Are study quality and result relil/ed?
Better studies arc more believable.
INDEX
Page numbers In italics denote figures; those followed by a "1" denote tables.
Abnormality, 19-42 Biologic plausibility, 244-245

criteria [or, 34-39, 35 Blinding, 148-149
overview, 41 Burden of suffering, 168-170
Accuracy, of test result, 44-48, 45
Adherence, randomized controlled
Case
trials, 146
defining, 81-83, 82t
Adjustment, multivariable, 132
selection of, 220
Aggregate risk studies, 238
Case reports, 208-211
Allocation of treatment, 145
biased reporting, 210
nonrandom, 16"1
uses of, 208-209
Alpha error, 187
Case series, 211-212
Association
Case-control studies, 212-225
cause and, 235-236, 237
advantages of, 211 t. 218-219
describing, 202-203
bias in, 219-n'i
reversible, cause and, 242-243, 244
cases, selection of, 220
Bela error, 187 vs. cohort research, 214-215, 215,
Bias, 7, 177-182 2161
assembly, 123 controls, selection of, 220··222, 223
in case-control studies, 219-225 design, TI3, 213-214
chance, 11 exposure, measuring, 222-224
in cohort studies, 123, 123-128 scientific standards, 224-225
confounding, 8-10, 9 Cause, 228-248, 246-247
compliance, 180--182, 181 analogy, 245
diagnosis and,~5-56 association, 235-236, 237, 242
lead time, JiS, 17K-1i9 reversible, 242-243, 244
length time, 179, 179 180 biologic plausibility, 244-245
measurement, 8, 128 concepts of, 229-235
migration, 126-127 consistency, 243
overview, 7, 7t, 7-10 dose-response relationships, 242
in prevalence studies, HS-Hl elements, for or against, 241t,
publication and, 262-263 241-245
relationship with chance, 11 establishing,2:'t'i-245
in reporting, case reports, 210 evidence, weighing, 245-246, 246
sampling, 1J, 133~134 multiple, 96-97, 97, 229-230,
selection, 7-8, "J2S-133,129t 232-235
susceptibility, 123-12.'), 124 interplay of, 1'12-235, 234
271
272 INDEX
Cause-continued Data
single, 229-230 interval,21
specificity, 245 nominal, 21
temporal relationships, 241-242 ordinal, 21
Chance, 10-12, 11, 186-207, 189 simplifying, 43-44
diagnosis and, 56-57, 57 Decision, clinical, analysis, 89-90
overview, 205-206 Dt'lllographic groups, diagnosis and,
Clinical course, 112 62
Clinical data, summarization of, 20 Diagnosis, 43-74
Clinical epidemiology bias, 55-56
basic principles of, 4-14 chance, 56-57, 57
elements of, 3-4 cut-off point, 50
health outcomes, 4-5, 5-6t independence, assumption of,
need for, 3 70-71
overview, 2-4 multiple tests, 63, 67-71, 68
social context, 4 overview, 72-73
uses of, 14-15 standards
Clinical medicine, epidemiology for disease, lack of, 47
and,2 imperfect, consequence~ of,
Clinical perspective, traditional, 2-3 47-48
Cohort, 100-101 Distribution
vs. case-control research, 214-215, abnormality and, 30, 30-34
n5,216t actual, 32, 32-33, 32-33
inceptions, 112 describing, 30, 31 t. 32
survival, 125-126, 125 frequency, 30
Cohort studies, 100, 101-105, 102 Gaussian, 34
normal, 33-34, 34
advantages, l03t, 103-105
Dose-response relationships, 242
bias in, 123-128, 123
Double-blind, randomized controlled
disadvantages, 103t, 103-105
tria Is, 148
historical, 102, 103
Duration of disease, incidence,
Cninterventions, 147-148
prevalence, relationship, 78,
Comparison 84-85, Wit
groups, randomized controlled
trials,142-144
multiple, 200-202 Ecologic studies, 2.18
Compliance, 146-147 Effect, modification, 235, 236
Confidence interval, 198-200, 199 Effectiveness, 151-152, 153
Confounding bias, 8-10, 9 Efficacy, 151-152, 15,'!
Consistency, cause and, 243 Error, random, 18fi-187
Controls Explanatory trials, 151, 152
concurrent, 158
selection of, 220-222, 223 False-positive results, 44-45, 176,
Cost, 172-173 17fit
Cox proportional hazards regression Frequency, 7.')-93
model, 123 distribution, 30
Cut-off point, diagnosis and, 50 incidence, 76-89, 78
INDEX 273
overview, 90-92 Multiple tests, diagnosis and, 63,

perceptions of, 76 67-71,68
Multivariablc adjustment, 132
Gaussian distribution, 34 Multivariablc methods, 203-20S
Gold Standard, 45-41'>
Guidelines Natural history, of disease, 112
c1inical,2h'J-21i1i Negative tests, lack of information
detennining, validity of clinical on,46
studies, 267- 270 Nonresponders, in randomized
controlled trials, 148
Hawthorne effect, 143 Normal distribution, 33-34, 34
Hypothesis testing, 187 -19K Null hypothesis, 190
raising, testing, 256-2S8, 257 Number of patients, hypothesis
testing and, 193-198
Ideas, treatment and, 131i-137
Observation, 7, 142-143
testing, 137-138
Odds, h4
Inception cohort, 112
Odds ratio, 215-218, 217
Incidence, 76-78, 78, 79t, 80
Open label, 149
cumulative, SO
Outcome, 4-6
density, 80
assessment of, randomized
measuring, 79-81
controlled trials, 149-150, L'iOt
prevalence, duration of disease,
biologic, 6
relationship, 78, H4-KS, 85t
clinical, 4, 5, 5t
studies of, 80-81
of disease. describing, 113
uses of, 87-89
randomized trials, 158-100, 160
Independence, assumption of, 7ll-71
Information, sources of, 21i3-2hS
I' \',1Iucs,18,)-191l
Intervention, 140-142
Parallel tests. 67-69, 69t
Periodic health examination. 168,
Kaplan-Muir analysis, 119 169t
Phases of trials, 161
labeling, 175 Place, comparison across, "157-158
Latency,9S Placebo treatment, 143, /44
Likelihood ratios Placebo effec-t, 143
defined,64-65 Plausibility, biologic, cause and.
serial, 71, 71 244-245
uses of, 6S-67, 66t Point estimates, 198-200, /99
Literature Population, 6
approaches to, 250-254 defining, 83-84
importance of, 25()-2S2, 251 risk and, 107-109, 108
usefulness of, 252-234, 254 samples and, 6
studies, 238, 239
Management. trials, 151, 1.'i2 Prediction
Matching, no prognosis and, 115, 116
Measurement, interpreting, 81-84 of risk, 98
Multiple effects, risk, 96-97, 97 rules, 11.'), 116,
274 INDEX
Predictive value, 57-64 Proportional hazards regression

determinants of, 58-60, 60 model, Cox's, 123, 132
negative, 57-58 Publication bias, 262-263
positive, 57-58
Prevalence, 76-78, 78, 79t
Quality-of-life measures, health-
estimating, 60-61
related, 113
incidence, duration of disease,
Questions, clinical, 4, 5, 51
relationship, 78, 84-85, 851
increasing, 60, 61-64
measuring, 79-81 Random error, 186-187
period, 77 Randomized trials
point, 77 advantages, 156-157
studies, 79, 79-80 alternatives to, 155-161
uses of, 87-89 controlled, 138-152, 139-140
Prevention, 165-185 differences arising after, 145-148
bias, 177-182 disadvantages, 156-157
compliance, 180-182, 181 explanatory, 151, /52
lead time, 178, 178-179 limitations of, 15.'j-156
length time, 179, 179-180 management,151,152
levels of, 166, 166-168 randomization, 145
overview, 184 Range, 24
primary, 166-167 Rare events
recommendations, 183-184 detecting, 200, 201
screening, 167 joint occurrence of, case reports,
21O-211,211t
secondary, 167
Receiver operator characteristic
tertiary, 167-168
curve, 51
tests and, 170-176
Referral process, diagnosis and,
acceptability of, 1741, 174-175
61-62
cost, 172-173
Regression to mean, 39-41, 40, 40,
false-positive results, 176, 176t
160
incidence method, 171
Relative risk, 106
interval cancers, 170
Reliability,23-24
safety, 173-174
Research design, cause and, 237-238
sensitivity, 170-172, 172-173 Responders, non responders,
simplicity,I72-173 comparison, randomized
specificity, 170-172, 172-/73 controlled trials, 148
Probability, 6 Responsiveness
sample, 84 abnormality and, 25
statements, as~igning numbers to, diagnosis and, 71-72, 72
76 Restriction, 139-140
Prognosis, 111-135 Results
overview, 134 of many studies, summarizing,
as rate, 115-116, 117t 258-262,259
studies, 111-113 of trial, tailoring, to individual
Prognostic factors, 114, 114-115 patients, 152-155
Proportional factors, 114-115 Risk, 94-100, 100-101, lOIt
INDEX 275
attributable, 105 establishing, 53-57

estimates of, interpreting, 106-107, of tests, 170-172, 172-173
107t Spectrum of patients, diagnosis and,
overview, 109 43,54-55,55
population, 107-109, 108 Standardization, 13] -132
prediction, 98 Standards
recognition, 95-97 for disease, lack of, 47
relative, 106 imperfect, consequences of, 47-48
small,96 scientific, 224-225
studies, 100-105 Statistical power, 193
uses of, 98-100 before/after study done, 200
Risk factors, 94~95 Statistical significance, 189-190
Statistical tests, 190-]93, 191, 191t
Stratification, 130-131, 131t
Sample size, 193-198 Subgroups, tailoring results of
Samples, overview, 6 treatment, to individual
Sampling- patients, 132-154, 154t
frequency and, 84 Survival analysis, 1"17
randomized controlled trials, Survival cohorts, prognosis and, 125,
139-140,141 125-126
Sampling fraction, 26 Survival curves, 119-12-1, 120, 122
Scientific standards, case-control
research and, 224-225 Tailoring of results, to individual
Screening, 167 patients, 152-155
effectiveness, 176-182 Temporal relationships, cause and,
Secondary prevention, 167 241-242
Selection bias Tertiary prevention, 167-·168
overview, 7-8 Tests
prognosis and, 128" 133, 129t acceptability of, 174t, 174-175
Sensitive tests, uses of, 50 accuracy of, diagnosis and, 44-4R,
Sensitivity, 45, 48-49, 48-53 45
analysis, prognosis arid, 132-1-13 cost of, 172-173
defined,48 false-positive results, prevention
establishing, 53-57 and, 176, 17M
spcculcny. trade-offs. 50-53, 5It, lack of information on accuracy,
5253 46-47
Sequences, temporal safety, 173-174
cause and, 241-242 simplicity, 172-173
uncertainty and, 85-86, 86 Time, compat-ison across, 157-158,
Serial testing. diagnosis and, 69-70, 159
69-70t Time series study, 238 241,240
Simplicity, of tests, 172-173 multiple, 239
Specific tests, uses of, 50 Time-to event analysis, "1"19
Specificity Treatment, 136-164
CaUSl.' and, 245 effectiveness, 151- 152, 153
defined,48 efficacy, 151-152, 153
diagnosis and, 45, 48 49,48-53 overview, 162
276 INDEX
Treatment -C(JIj Iinlied construct, 22

prevention and, 176-182 content, 22
randomized controlled trial, criterion, 23
138-152,139-140 internal, external, 12, 12-14
Trial Variation, 25t, 25-30
explanatory, 151, 152 biologic, 26-28, 28
management. 151, 152 effects of, 28, 28-30
phases, 161 measurement, 25- 26, 27
randomized controlled, 138-152,
observer, 27
139-140
sources of, 25, 29
Uncontrolled trials, 15H- -161 total, 28, 29
Validity, 22-23
clinical studies, determining,
guidelines for, 267-270 Zero time, 112-113

Clinical Epidemiology

Uploaded by

Copyright:

Available Formats

Clinical Epidemiology

Uploaded by

Document Information

Original Title

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

Clinical Epidemiology

Uploaded by

Copyright:

Available Formats

What is the book about?

What is the book about?

What topics are covered in the book?

What topics are covered in the book?

T H R D E D T o N

Robert H. Fletcher, M.D., M.sc.

Suzanne W. Fletcher, MD., M.sc.

Edward H. Wagner, MD., M.P.H.

Williams & Wilkins

'UJI,."lI'oI'II'' ' >:,"'ow:,,,

Copyrightl1l 1996 Williams & Wilkins

351 West Camden Street

ROSH Tree Corporate CHnter

P"illad in Ihe Unil"d Slatos at America

First roaoo. 1982

Library of Congre~ Cataloging-in-Publicatlon Data

to clinical epidemiology and to avoid pitching the presentation to our

We have been privileged to work with wonderful colleagues, many of

were together on the faculties of Medicine and Epidemiology at the Univer-

in which such observations are made and interpreted determines whether

Clinical training has been oriented toward the mechanisms of disease

• In most clinical situations the diagnosis, prognosis, and results of treat-

• Probability for an individual patient is best estimated by referring to

Abnormality Is the patient sick or well?

Dcath A bad outcome it untimely

Interventions Biolcqlc Clinical

HIV infection Zidovudine CD4+ counts Opportunistic infections

Association known or assumed?

NUMBERS AND PROBABILITY

the characteristics of the defined population but must, for precttca1reasons,

points in a clinical observation or study. Selection bias is at issue primarily

In this example, selection bias could be present if volunteers for the

determine whether bias is actually present and how large it is likely to be

Observations about disease are ordinarily made on a sample of patients

rectly, using a sphygmomanometer. The simpler instrument is prone to

cannot be eliminated, but its influence can. be reduced by proper design

When making: inferences about a population from observations on a

All patients with the INTERNAL

Figure 1.3. Internal and external validity,

observation hold true in other settings. for an individual physician, it is

The gcncraltzabtltty of clinical observations, even those with high inter-

Ccncrafizabilitv can rarely be dealt with satisfactorily in anyone study.

Uses of Clinical Epidemiology

Information and Decisions

Organization of the Book

Events Natural History Chapter Title

We have organized the book primarily according to the clinical ques-

Feinstein AK Clinical epidemiology. The architecture of clinical research. Philadelphia: WB

Clinicians spend a great deal of time distinguishing "normal" from

Problurfl Li~l new nato

Hypertension Several blood pressure readings [mm Hg):

decisions must be made, whether actively (by additional diagnostic tests

Criterion villidity is present to the extent that the measurements predict

Reliability is the extent to which repeated measurements of a stable

simply random error (lack of reliability). It is possible to reduce this source

Variation also arises because of differences among people. Biologic dif-

Figure 2.3. Biologic variability. The number of ventricular premature depolarizations

surements, although subject to all other forms of variation, is related to

Conditions of Distribution of Source(s)

Se == == 73% so = 1== 69% p == ?I9 == 25%