Cancer Epidemiology - Principles and Methods (PDFDrive)
Cancer Epidemiology - Principles and Methods (PDFDrive)
Cancer Epidemiology - Principles and Methods (PDFDrive)
æ1
~~
~~ 'I~
,~ Il~
WORLD HEAL TH ORGANIZATION
INTERNATIONAL AGENCY FOR RESEARCH ON CANCER
Cancer
Epidem iology:
Principles and
Methods
Sy
Isabel dos Santos Silva
\'~'ì."'..
International Agency for Research on Cancer
The International Agency for Research on Cancer (IARC) was established in 1965 by the
World Health Assembly, as an independently financed organization within the framework of
the World Health Organization. The headquarters of the Agency are at Lyon, France.
The Agency conducts a programme of research concentrating particularly on the
epidemiology of cancer and the study of potential carcinogens in the human environment.
Its field studies are supplemented by biological and chemical research carried out in the
Agency's laboratories in Lyon, and, through collaborative research agreements, in national
research institutions in many countries. The Agency also conducts a programme for the
education and training of personnel for cancer research.
The publications of the Agency are intended to contribute to the dissemination of
authoritative information on different aspects of cancer research. Information about IARC
publications and how to order them is also available via the Internet at: http://www.iarc-fr/
Cancer Epidemiology:
Principles and Methods
Published by the International Agency for Research on Cancer,
150 cours Albert Thomas, 69372 Lyon cédex 08, France
The designations employed and the presentation of the material in this publication do not imply
the expression of any opinion whatsoever on the part of the Secretariat of the World Health
Organization concerning the legal status of any country, territory, city, or area or of its authorities,
or concerning the delimitation of its frontiers or boundaries.
The mention of specific companies or of certain manufacturers' products does not imply
that they are endorsed or recommended by the World Health Organization in preference to others
of a similar nature that are not mentioned. Errors and omissions excepted,
the names of proprietary products are distinguished by initial capitalletters.
The authors alone are responsible for the views expressed in this publication.
The International Agency for Research on Cancer welcomes requests for permission to
reproduce or translate its publications, in part or in fulL. Applications and enquiries should be
addressed to the Editorial & Publications Service, International Agency for Research on Cancer,
which will be glad to provide the latest information on any changes made to the text, plans for new
editions, and reprints and translations already available.
v
Foreword
P. Kleihues
Director, IARC
Nubia Muñoz
D. Maxwell Parkin
Jacques Estève
Paolo Boffetta
vii
J \....
f
1
",
1
i
í
About this book
ix
them! 1 would also like to thank IARC for giving me the opportunity to
write this book and, in particular, the members of the editorial committee
(Ors Jacques Estève, Nubia Muñoz, Max Parkin and Paolo Boffetta), Dr.
Rengaswami Sankaranarayanan and Dr. Martyn Plummer, for their
suggestions and comments in the earlier stages of the project. 1 am also
grateful for aIl the support and encouragement given by my colleagues at
the London School of Hygiene and Tropical Medicine and, in particular,
to the thoughtful comments on earlier versions of various chapters
provided by Professor Peter Smith, Dr Noreen Maconochie and Dr Punam
Mangtani. 1 also wish to thank Ms Maria Quigley and Mr. Craig Higgins
for giving permission to include sorne of their teaching examples in
Chapter 14. Part of the material presented in Chapters 6 and 12 was
initially developed as teaching material for our intensive course in
collaboration with Dr Maconochie and Mr Jerry Wheeler-the stimulating
discussions we had at that time and the feedback received from our
students were very helpful in shaping those chapters. Lastly, 1 would like
to thank Dr John Cheney and Ms Helis Miido for aIl their editorial
assistance.
x
Chapter 1
Introduction to cancer
epidemiology
Interest in cancer has grown during the past century as infectious dis-
eases have increasingly been controlled as the result of improved sanita-
tion, vaccination and antibiotics. Although this interest is relatively
recent, cancer is not a new disease. Autopsies of an cie nt Egyptian mum-
mies have shown the presence of bone tumours and possibly other neo-
plasias (BrothweIl, 1967). Symptoms of what can be assumed to be malig-
nant diseases were also described in Chinese and Arabic medical writings.
By the time of Hippocrates in the 4th century BC, many types of tumour
were clinically recognized and described. Hippocrates introduced the term
carcinoma from the Greek ward karkinos, for crab: he saw cancer as crab-
like in its spread through the body and in its persistence (Long, 1928).
Sorne 600 years later, Galen distinguished three types of tumour: 'tumours
according to nature', which included aIl normal physiological swellngs,
such as enlargement of the breast with normal female maturation;
'tumours exceeding nature', which included the productive process fol-
lowing injury, such as the proliferation of bone that occurs during the
reuniting of a fracture; and 'tumours contrary to nature', which included
what we now define as neoplastic growths, as weIl as many inflammatory
lesions (Long, 1928).
However, it was not unti the end of the 18th century that cancer began
to be studied systematically and intensively. Bichat (1771-1802) described
the pathology of many neoplasms in humans and suggested that cancer
was an 'accidental formation' of tissue built up in the same manner as any
other part of the organism. Sorne decades later, Müller (1801-58) and
Virchow (1821-1902) extended Bichats findings, using the microscope to
show that cancerous tissue was made up of cells (Long, 1928).
Ever since, pathologists and clinicians have considered cancers in the var-
ious organs of the body as being in man y respects completely different dis-
eases with distinct morphologies, clinical manifestations and prognoses. But
only during the past few decades has it emerged that their causes also differ
enormously. As a discipline, epidemiology has been crucially important in
defining the causes of different cancers and in evaluating preventive mea-
sures.
1
Chapter 1
'\,
""\~.
Introduction to Cancer Epidemiology
3
Chapter 1
"The fate of these people seems singularly hard; in their early infancy
they are most frequently treated with great brutality, and almost
starved with cold and hunger; they are thrust up narrow, and some-
times hot chimneys, where they are buried, burned and almost suffo-
cated; and when they get to puberty, become liable to a most noisome,
painful and fatal disease." (Pott, 1775)
"A feature of the disease which exercised those interested in it, and
which was connected with its aetiology, was that chimney sweeps'
cancer seemed to be almost exclusively an English disease. Cases were
virtually unknown on the Continent, in America, or even in Scotland. Figure 1.1.
Protective c10thing worn by a German
This problem was fully investigated by Henry Butln. Butlin took him- sweep in 1785 (repraduced, by per-
self to the Continent during the course of his researches and, as the mission of BMJ Publishing Group, fram
result of meeting and talking to Continental sweeps, he considered Waldron, 1983).
that he had established the reasons for the virtual absence of scrotal
cancer among them. ln part it was due to the protective clothing they
wore. As early as 1785, the German sweep is depicted in a close fitting
suit complete with head covering (Figure 1. 1 J. There are no openings
for the soot to penetrate nor any loose clothing in which it can lodge.
The London sweep, as late as 1851 stands in sharp contrast, a waif-like
boy, dirty and in loose smock and trousers. (Figure 1.2J"
5
Chapter 1
Figure 1.3.
Mortality fram ail malignant neoplasms International Statistics of Cancer Mortality. 1908-1912
for certain countries and cities,
1908-1912 (repraduced fram Hoffman,
1915). CouFTrries CiÜes
.Jilzerläod /i'3
C $q /00
RaIe per 100. 000 Population
ISO
(jnM5'eI 1613
o 50 /00
Raie per /00.000 Popu/abon
~
7
Chapter 1
-8
~=o ",.~Introduction
~-..~_ ta Cancer Epidemialogy
Further reading
~ -" Box 1.1. Key issues - "-"_
* Buck et al. (1988) compiled and
-Cåocerapidemiologyisconcernédwiththestudy of the distribution of the dis-
discussed a collection of classic
aa~ecancerin populatipos.ltsultimate Qoal. is to identify. riskfactors that may epidemiological papers, includ-
laad tq.eady introductipnpfeffective preventivemeasures.
ing many of those quoted in
Section 1.2.
-Cancerepidemiology,andepidamiologyingeneral, is based on the comparison
otgroups of people.. FprethicaJreasons, however, epidemiological methods are
predominantly obseNational(i.e., non-experimental). The major challenge for
epidemiologistsis toidentifyand make use of 'natural experiments' that wil help
to answerthe questionunder investigation.
9
Chapter 2
Measurement of exposures
and outcomes
2.1 Introduction
Most epidemiological research involves the study of the relationship of
one type of event or characteristic to another. Consider the following
questions as examples:
11
Chapter 2
Suppose that a researcher observes that lung cancer occurs more often
in people who drink alcohol than in those who do not. It would not be
possible to conclude from this observation that exposure to alcohol
increases the probabilty of developing lung cancer, unless the researcher
can show that the observed relationship cannot be due to the fact that
those who drink alcohol smoke more heavily than non-drinkers. ln this
example, smoking is acting as a confounder. Confounding can be dealt with
when designing studies or when analysing the results provided that the
relevant data have been collected. These issues are discussed in detail in
Chapters 13 and 14.
Thus, most epidemiological studies must collect information on three
types of variable:
12
Measurement of exposures and outcomes
"
2.3.2 Dose
Exposure is seldom simply present or absent. Most exposures of interest
are quantitative variables. Smokers can be classified according to the num-
ber of cigarettes smoked daily; industrial exposures by the extent of expo-
sure (often achieved by classifying workers according to the duration of
employment and type of job); infections by dose of agent or age at expo-
sure; breast-feeding by duration; and psychological exposures by sorne
arbitrary scale of severity. Thus the simple situation of two groups, one
exposed and one unexposed, is rare, and the conclusions of a study are
greatly strengthened where there is a trend of increasing disease incidence
with increasing exposure-an exposure-response relationship.
13
Chapter 2
2.3.3 Time
As far as possible, each exposure should be characterized as to when it
began, when it ended (if at aIl), and how it was distributed during the
intervening period (was it periodic or continuous? did the dose vary over
time?). Similar details should also be obtained for any behaviour that may
protect against the exposure.
There is thought to be a restricted period, the critical time window, dur-
ing which the exposure could have caused cancer. Unfortunately, the
beginning and end of this critical time window are not known, and its
length is likely to vary between individuals. Collecting data on the timing
of exposure allows the possible extent of this window to be estimated.
Analyses may include examination of the effects of time since first expo-
sure and time since last exposure.
Pattern of exposure may also be important. Exposure that occurs peri-
odically in intense bursts may have a different effect from a similar total
amount of exposure that occurs continuously at low intensity (e.g., con-
stant versus intermittent exposure to sunlight; chronic exposure to low
levels of ionizing radiation versus acute exposure to high levels).
2.4.1 Questionnaires
Questionnaires are used to collect exposure data in epidemiological
studies by putting the same set of questions to each study participant in a
standardized form. Questionnaires can be self-administered or may be
administered by an interviewer.
The aim of a research questionnaire is to obtain, with minimal error,
measurements of the exposure variables of interest for the study. Thus, the
questions to be incIuded in a questionnaire should relate directly to the
objectives of the study. Sorne basic principles that should be taken into
account when designing a questionnaire are discussed in Appendix 2.1. To
14
Measurement of exposures and outcomes
ensure that the questions are properly understood and wil elicit appro-
priate answers, questionnaires should be pre-tested on a sample of subjects
from the population to be studied.
Se!f-administered questionnaires
Self-administered questionnaires are distributed to study subjects who
are asked to complete them. They can be delivered and returned either per-
sonally or by mail if this is feasible and more convenient. Such question-
naires are particularly appropriate when small amounts of reasonably sim-
ple data are required, or for documenting sensitive or socially undesirable
behaviour. They are one of the cheapest ways of collecting information,
but have the limitation that they can be used only in literate populations.
The investigator also has relatively little control on the quality of the data
collected.
2.4.2 Diaries
Diaries are detailed records of exposure kept by the subject. They are
generally open-ended and take the form of a booklet in which the subject
records each occurrence of a particular behaviour such as physical exercise,
alcohol consumption, dietary intake, sexual activity, use of medication,
etc. Diaries are assumed to be highly accurate in measuring current behav-
iour, because they do not rely on memory. They also allow more detailed
information about the exposure to be collected than with a questionnaire.
For example, foods can be weighed by the subject before being eaten.
The main limitation of diaries is that only current exposures can be
measured. ln addition, diaries generally demand more of subjects in terms
of time and skil than other methods, so compliance may be a problem.
Training of subjects in the skils needed to keep an accurate diary can be
time-consuming for both subjects and investigators. Thus, diaries are
rarely used in countries in which many people are iliterate.
2.4.3 Records
Data on the exposure of interest may be available from census, employ-
ment, medical (in- and out-patient), cancer registry, birth certification and
death certification records. Typically, as the data have already been col-
lected for purposes other than epidemiological research, the researcher has
no control over what items were recorded, how questions were phrased,
and so on. Records are also often produced by a large number of people
with litte uniform training. Moreover, the availabilty and quality of
records in many countries tends to be po or.
Despite these limitations, the use of records has several advantages over
other methods of data collection. Study costs are usually low, and the
duration of the study is shorter because sorne or aIl of the data have
already been collected. Records can also provide ne ar-complete data on a
weIl defined population, and information can be obtained without con-
tacting the subjects or their relatives. Certain data items (for example,
intake of medications or occupation al exposures) may be recorded more
accurately than information obtained in a personal interview; for instance,
errors caused by poor recall or lack of knowledge of the exposure are elim-
inated.
Characteristics and limitations of sorne such routine data-collection sys-
tems are discussed in more detail in Section 2.9 and Chapter 11.
16
-\'...~
Measurement of exposures and outcomes
17
Chapter 2
18
Measurement of exposures and outcomes
19
,¡
f
f
Chapter 2
Specificity = d/(b+d)
The predictive value of a positive study test result represents the probabil-
ty that someone with a positive study test result reaUy has the character-
istic of interest as determined by the gold standard:
The predictive value of a negative study test result represents the probabil-
ty that someone with a negative study test result does not have the char-
acte tic of interest as determined by the gold standard:
ris
Table 2.2.
Comparison of ViraPap(ß and Southern
hybridization methods in the diagnosis
of ceivicai HPV infection in a sarnple
of women who atlended a sexually
transmitled disease c1inic. a
20
Measurement of exposures and outcomes
Table 2.3.
Comparison of ViraPap(é and Southern
hybridization methods in the detection
of cervical HPV infection among
apparently healthy women: hypotheti-
cal data.
21
Chapter 2
Table 2.4.
Comparison of ViraPapQi and poly-
merase chain reaction (PCR) in the
detection of cervical HPV infection. a
ln Example 2.3 the validity of the ViraPap(I test (as measured by its sen-
sitivity and specificity) was much lower th an when Southern hybridiza-
tion was used as the gold standard method (Example 2.1). This is because
the PCR method is more sensitive and more specific than the Southern
hybridization technique.
Not aIl tests give a simple yes/no result. Sorne yield results that are
numerical values along a continuous scale of measurement. ln these situ-
ations, high sensitivity is obtained at the cost of low specificity and vice
versa. For example, the higher the blood pressure, the more probable is
hypertensive disease. If a diagnostic or screening test for hypertension is
22
- =._~-.~: =w.-.,..='I~~
Measurement of exposures and outcomes
23
Chapter 2
120a + + 173
+ - 25
- + 17
- - 145
2808 + + 95
+ - 0
- + 95
- - 170
8 Slood assay results equal to or greater th an the eut-off value were taken as positive: +, posi-
tive result; -, negative result
Figure 2.1.
Healthy subjects
The upper curve describes the distrib-
ution of results of the blood as say
among healthy individuals and the
lower curve the distribution among
cancer patients (as defined by the
anatomo-pathological examination).
Different eut-off values are used to
classify the results of the blood assay Cancer patients
as 'positive' or 'negative'. 1
1 1
1 1
1 1
1 1
1 1
1 1
o 280
24
L
i
í
Measurement of exposures and outcomes
~ ~ - --~il--~ ~_.._s;--~ ~r.
2.6.2 Reliability
Reliabilty, sometimes also called repeatabilty or repro- specificity
ducibilty, is a measure of the consistency of the performance 100 90 80 70 60 50 40 30 20 10 0
100
of a test when used under similar circumstances. To be valid, 40
90
a measurement must be reliable. However, reliability is not in
itself sufficient for validity: in other words, a test may yield the 80
same result consistently, but the result may not be the true 70
Consider the simple example of a test that can give only a positive or
negative (Le. binary) result. The agreement between pairs of measure-
ments carried out by two independent observers on the same subjects can
be presented as a 2x2 table (Table 2.7).
One measure of repeatabilty is the observed agreement (0) or mean
pair agreement index, which can be calculated as:
25
Chapter 2
This index has the disadvantage that sorne agreement would be expected
even if both observers simply guessed the result. The kappa statistic (K) is an
alternative measure that takes account of the agreement expected solely on
the basis of chance.
To calculate the kappa statistic, the number of pairs of observations that
would be expected on the basis of chance in ceIls (++) and (- -) must first be
calculated. The expected value in any cell is given by:
((a+b) x (a+c)JfN
((CHi) x (b+d)J/N
The expected agreement on the basis of chance (E) can now be calculat-
ed as:
(Expected value for cel! (++) + Expected value for cel! (- -)J / N
i -E
26
Measurement of exposures and outcomes
The kappa statistic indicates how much the actual agreement beyond
chance (O-E) represents relative to this potential (1 - E).
27
Chapter 2
28
Measurement of exposures and outcomes
..=r~",...,
reducing the confidence that can be placed in each particular test result.
Although this random misclassification has important implications in
clinical medicine, it is of less concern in epidemiology, where groups
rather than individuals are the main interest. Herein lies a great strength
of epidemiology. ln the above example, the association between smoking
and lung cancer was weakened because those classifying themselves as
'never smokers' were in fact a mixture of those who had never smoked and
those who had. Although this type of mis classification makes it more dif-
ficult to reveal an association between the exposure and the outcome of
interest, the problem can usually be overcome by increasing the sample
29
l....
Chapter 2
30
'\.
~'-i"'.
Measurement of exposures and outcomes
-
31
Chapter 2
32
Measurement of exposures and outcomes
33
,l
'...~
í
Chapter 2
Further reading
'"" '~~ ",,' ,¿' ,.'",:~ ;:.:',,' .,,: B,at~);"~~¥.is.s.a-~~~:.:::, . -- -- ~_. ,.'.,~-', ':'._.''" .-~ --;
* Comprehensive coverage of ':",:/:,'. ':,. ",-:";,:".,__:":,,,-,,',-.. '-:'., _ ,. -:';';"'_,"-:\;,:-:,',,-: :,._',:':-'.',':,'-",-,::~~',J),:",-C
* An often-referenced paper on
the validity and reliability of tests
that yield results on a continuous
scale (e.g., blood pressure mea-
surements) is that by Sland &
Altman (1986).
* For a further, more complex, ., _,'_ '_ _: _., ' ' . . _, ,",. h; '- _ :'~ ":_:,":" .- _ _:',:_'-- ':" _,:., .. ,_'.
discussion on the kappa statistic, · DifferentiaI measuremcmt errar occurs when the sensitivity andíorspecificityof
see Feinstein & Cicchetti (1990), the exposure measurement for the diseased subjects differsfrom thatfor the
Cicchett & Feinstein (1990) and
non-diseased subjects, or when the sensitivity and/orspecificity~f, theoutcome
measurement is different for exposed and unexposed sÙbjects: Differentiai lTE3a- '
Lantz & Nebenzahl (1996).
surement error can exaggerate, attenuate, or even rev~rse, th13relationphip
between the exposure and the stud ", ' 'b~
the outcome; 80. that the res(jlts of
34
Appendix 2.1
Designing a questionnaire
Types of question
There are two main types of question: 'open-ended' and 'closed-ended'.
Open-ended questions allow the respondents to answer on their own terms
and should be recorded in the respondents own words. Open-ended ques-
tions should be used for numerical data (for example, age, date of birth)
and for questions having many possible answers (e.g. country of birth).
35
Appendix 2.1
Not applicable 07
Wording of questions
Questions must be written in simple, non-threatening language, avoiding
the use of abbreviations and technical jargon. The wording should avoid any
suggestion that a particular answer is preferred by the researcher(s). Each
question should contain only one concept related to a clear time period.
Order of questions
Questions should follow a logical sequence resembling, as far as possi-
ble, the sequence that the respondents might expect to follow when
thinking about the topic. Questions about a particular subject should be
grouped together, and proceed from the general to the particular. When a
response to a general question makes further responses on that topic irrel-
evant (e.g., a woman who has never been pregnant need not answer ques-
tions about number and characteristics of pregnancies), a branching of the
question sequence may be introduced. This should be as simple as possi-
ble, with clear instructions given on the questionnaire (Example A2.1.3).
Questionnaire layout
Layout is important in both self- and interviewer-administered ques-
tionnaires. A pleasant appearance wil arouse interest and encourage cor-
rect completion. A separate page with a brief introduction, explanatory
notes and instructions should precede the first question. To help inter-
viewers and subjects, long questionnaires may be subdivided into sections,
each one corresponding to a specific topic. AIl questions should be
assigned a number.
If sorne questions are optional, this should be indicated on the ques-
tionnaire with clear instructions and appropriate branch and jump expla-
36
Designing a questionnaire
ûiirnping) .
Yes D 1
No D 2
If Yes,
2. How many pregnancies in total (including still births,
miscarriages and abortions) have you had? IT
nations. For questions that are repeated several times, su ch as questions
about each pregnancy, a tabular layout may be used (Example A2.1.4).
Space should be provided at the end of the questionnaire for any infor-
mation or comments that the subject may wish to add.
rn rn rn rn rn
Birth 01 Birth 01 Birth 01 Birth 01 Birth 01
Still birth 02 Still birth 02 Stil bi rth 02 Sti ii bi rth 02 Still birth 02
Miscarriage 03 Miscarriage 03 Miscarriage 03 Miscarriage 03 Miscarriage 03
Abortion 04 Abortion 04 Abortion 04 Abortion 04 Abortion 04
rn rn rn rn rn
Yes 01 Yes 01 Yes 01 Yes 01 Yes 01
No 02 No 02 No 02 No 02 No 02
Not applicable 0 7 Not applicable 07 Not applicable 0 7 Not applicableO 7 Not applicableO 7
Do not know 09 Do not know 09 Do not know 09 Do not know 0 9 Do not know 09
37
L
f
f
Í
Appendix 2.1
Method of administration
The questionnaire can be either self-administered or interviewer-adminis-
tered. ln general, self-administered questionnaires must be simpler and much
more carefully designed than those intended for use by interviewers.
variables.
Pre-testing
AIl questionnaires should be pre-tested. This involves testing the draft
questionnaire on samples of subjects similar to those who wil ultimately
be studied. Its purpose is to identify questions that are poorly understood,
ambiguous, or evoke hostile or other undesirable responses. Pre-tests
should be carried out using the same procedures that wil finally be used
in administering the questionnaire. Interviewers and study subjects
should be asked to provide feedback and the questions revised in the light
of their comments. Several rounds of pre-testing wil usually be necessary
before the final form of a questionnaire is developed.
Assessment of validity
The validity of the questionnaire as a measure of the variables of inter-
est should always be determined in a sam pIe of subjects before the main
study is undertaken. This requires comparison of the results obtained
using the questionnaire with those obtained using a go Id standard test (see
Section 2.6). For instance, questions on past hospitalizations and surgical
interventions may be validated against hospital records. Validation is usu-
ally difficult, often expensive, and may sometimes be impossible, when no
appropriate gold standard is available.
38
1
f
1
j
í
Designing a questionnaire
~,
39
.~~~~.n;--'-~'.'.,.
Appendix 2.2
Classification of diseases
Neoplasms can be classified in many ways, but the most important clas-
sifications for the epidemiologist are those based on:
(1) Topography-the site in the body where the tumour is located.
(2) Morphology (or histology)-the microscopic characteristics of the
tumour.
(3) Behaviour-the tendency to invade other tissues (malignant,
benign, in situ, and uncertain).
Uniform definitions and uniform systems of classification are funda-
mental to the quantitative study of diseases. Without a standard classifi-
cation tool that remains fixed for periods of time and is applied uniform-
ly, meaningful comparative analyses of morbidity and mortality data
would be impossible. The International Classification of Diseases (ICD), pub-
lished by the World Health Organization, is su ch a standard classification
tool. It is revised every ten years or so (Table A2.2.1); the lOth revis
ion
(ICD-lO) (WHO, 1992) is currently in use. An historical review of disease
classification from the first revision, the Bertilon Classification of Causes
of Death, unti 1947 can be found in the introduction to ICD-7 (WHO,
1957), and an account of classification in the years 1948-1985 is given by
Muir and Percy (1991).
41
Appendix 2.2
42
Classification of diseases
1956 Statistical code for Human Tumours World Health Topography codes fram ICD-7
(STAT GODE) Organization Morphology codes fram MOTNAC
Behaviour codes fram MOTNAC
1968 Manual of Tumor Nomenclature and American Cancer Topography codes fram ICD-8
Goding (MOTNAG) 2nd edition Society Morphology codes tram SNOP
1990 IGD-O, 2nd edition World Health Topography codes fram ICD-10
Organization Morphology codes fram ICD-O, 1 st edition
Behaviour codes fram ICD-O, 1 st edition
Table A2.2.3.
The major advantage of ICD is that it is truly international, being used Morphology and behaviour
byaIl WHO Member States for tabulating the causes of death and for most classifications of neoplasms.
health statistics. The main disadvantage is that, for the majority of sites,
no separation on the basis of morphology is possible. As a result, it is gen-
erally recommended that agencies interested in identifying both the site
and morphology of tumours, like cancer registries and pathology labora-
tories, use ICD-O, which is a dual-axis classification providing indepen-
dent coding systems for topography and morphology.
As new classifications and new revis ions of ICD and ICD-O have come
into use, data coded by previous classifications must be converted to the
new codes. The National Cancer Institute of the USA has produced a series
of conversion tables for neoplasms (e.g., Percy, 1980, 1981, 1983; Percy &
van Holten, 1979). Summary tables of equivalence between various revi-
sions of the ICD have also been published in certain volumes of Cancer
43
Appendix 2.2
Incidence in Five Continents (e.g., Waterhouse et aL., 1976; Muir et al., 1987).
Programs that perform conversions from ICD-O (lst edition) to ICD-O
(2nd edition) and vice versa, from ICD-O (lst and 2nd editions) to ICD-9,
and from ICD-O (2nd edition) to ICD-10 have been developed by the
International Agency for Research on Cancer (lARe) and are available on
diskette for use on microcomputers (Ferlay, 1994).
44
'\.
-"ì.,,,
Chapter 3
Describing and presenting
data
AIl epidemiological studies involve the collection of data on the expo-
sures and outcomes of interest. ln a weIl planned study, the raw observa-
tions that constitute the data contain the information that satisfies the
objectives of the study. The aim of data analysis is to extract the pertinent
information from these raw observations in a concise way.
The first stage in data analysis is the preparation of an appropriate form
in which the relevant data can be collected and coded in a format suitable
for entry into a computer; this stage is referred to as data processing. The
second stage is to review the recorded data, checking for accuracy, consis-
tency and completeness; this process is often referred to as data editing.
Next, the investigator summarizes the data in a concise form to allow sub-
sequent analysis-this is generally done by presenting the distribution of
the observations according to key characteristics in tables, graphs and
summary measures. This stage is known as data reduction. Gnly after data
processing, editing and reduction should more elaborate statistical manip-
ulation of the data be pursued.
45
Chapter 3
46
Describing and presenting data
47
..
\.~.....
Chapter 3
important features of a distribution are usuaIly the central value and the
spread about this value.
45.3,49.8,50.5,60.7,65.2,69.4,73.2,75.9,
(45.3 + 49.8 + 50.5 + 60.7 + 65.2 + 69.4 + 73.2 + 75.9)/8 = 490/8 = 61.5 = 61. kg
ln this example, the value of the median is the 4.Sth value, Le., the aver-
age of the fourth and fifth values, (60.7 + 65.2) /2= 63.0 kg.
The choice of measure used wil depend on the nature of the data and
the purpose of the analysis. The mean is often the preferred measure of cen-
tral value because it takes into account every observation and it is easy to
use in the most common types of statistical analysis. Its major disadvantage
is that it can be affected by outlers-single observations that are extreme in
comparison with most observations and whose inclusion or exclusion
changes the mean markedly.
The median is a useful descriptive measure when outliers make the mean
unrepresentative of the majority of the data. It is also particularly useful
when certain observations are not recorded precisely because they are
above or below a certain level; in these circumstances, the mean cannot be
calculated, but the median can determined so long as definite values are
known for more than one haU of aIl subjects. For instance, the mean sur-
vival time of a group of cancer patients can be calculated only when aIl the
patients have died; however, the median survival time can be calculated
while almost haU the patients are alive. The main disadvantage of this me
a-
sure is that it ranks the data, but does not make use of their individual val-
ues.
When the shape of a distribution is roughly symmetric about a central
axis, the mean and the median are approximately equal (as is the case for
the data on weight given in the example above).
Measures of variation
ln addition to a measure of the central value of a distribution, it is also
useful to have an ide a of the variation or spread of values around this cen-
tral value. Several measures of variation are used:
(1) The range is the interval between the largest and smallest values. The
major advantage of this measure is that it is very simple to calculate. The
main disadvantage is that it is based only on two extreme observations and
gives no indication of how other observations are distributed in between.
(2) The three values that divide a distribution into quarters are caIled the
qumties. Of the total number of observations, 25% lie below the lower
quartile, 50% below the middle quartie and 75% below the upper quartile.
The middle quartile is the median. The distance between the lower quartile
and the upper quartie is called the inter-quartile range and is sometimes
used to describe variabilty. The inter-quartile range contains 50% of the
observations.
Similarly, percentiles are values that divide the distribution into percent-
ages. The 50th percentile corresponds to the median, while the 2Sth and the
75th percentiles correspond to the lower and upper quarties, respectively.
A simple but useful semi-graphical way of summarizing data using per-
centiles is the box-and-whisker plot. Figure 3.2 shows a box-and-whisker
49
Chapter 3
plot for the data given in Table 3.1. The box indicates the lower and upper
180
quarties, while the central line is the median. The points at the ends of
175
the 'whiskers' are the 2.Sth and the 97.Sth percentile values.
170
Ê
2- 165
E'" (3) The standard deviation is the most commonly used measure of the
160
Qi
:i average spread of values about the mean. If the values of a variable do not
155
vary greatly within a population, observations wil lie closely around the
150
mean, whereas if there is substantial variation, the observations wil be
145
í
scattered widely about the mean.
This variabilty or variance can be measured in terms of how much, on
average, observations differ from the mean: in other words how far, on
Figure 3.2. average, each observation deviates from the mean. Figure 3.3 ilustrates
Box-and-whisker plot of the heights of this for the weights of the eight healthy women given in the example
a sample of 1250 women (data shown
in Table 3.1).
above. The deviations from the mean are shown by the lines dii d2, ... dg.
First, the sum of the these deviations is calculated; however, the sum of
the deviations from the arithmetic mean is, by definition, zero, since neg-
ative deviations cancel out positive deviations. ln calculating the disper-
sion of values around the arithmetic mean, it is irrelevant whether the
deviations are positive or negative: only their absolute numerical magni-
tude is of interest. Hence, to avoid getting zero when the deviations are
added together, the individu al deviations are first squared, eliminating
negative values. The average of these squared deviations is called the vari-
ance.
Mean = 61.3 kg
dB The variance is a very useful mathematical measure of the average
~ spread of values around the mean, but is difficult to interpret because it is
dG
expressed as the square of the units measured. ln our example, the vari-
ance of weight wil be expressed as kg2. As it is usually more convenient to
ds
express the variation in terms of the original, unsquared units, the square
d4
root of the variance is usually used. This is known as the standard deviation
d3
(SD). ln this example, the SD is equal to 10.8 kg.
d2 A small standard deviation indicates that most values are very close to
d,
the mean, whereas a large one indicates that many lie far from the mean:
Le., the more the values in a population vary, the bigger the standard devi-
40 45 50 55 60 65 70 75 80 ation.
Weight (kg)
As a general mIe, provided a distribution is roughly symmetrical and
has a bell-like shape, with a dome of central values and two tails at the
Figure.3.3. extremes (char tic of what statisticians calI a 'normal distribution'),
acte ris
Deviations fram the mean of the
weights of eight women: hypothetical the me an is the central value and the standard deviation is a measure of
data. spread su ch that one standard deviation either side of the mean includes
roughly 70% of the observations, and two standard deviations include
roughly 95% (Figure 3.4).
50
Describing and presenting data
Table 3.2.
Distribution of cervical cancer cases by
Education (schooling)
educational status and oral contracep-
Ever 317 72.7 tive use. a
Bar charts are often used to present the distribution of categorical vari-
ables (Figure 3.5). ln this type of graph, the value of the quantity of inter-
est is represented by the length of the bar.
51
f
¡
\
'.~
,,
,i
f
l
Chapter 3
Figure 3.5.
Distribution of cervical cancer cases by
Summarizing categorical data
oral contraceptive use (data fram For categorical data, percentages are the only sumIDary measures that
Bosch et al., 1992). can be calculated (as in Table 3.2). They are particularly useful when mak-
ing comparisons between different groups.
70
t 60
3.3.3 Two variables-quantitative or categorical
iß 50
'"
So far, we have considered the frequency distribution of a variable with-
'"
Õ
o 40 in a single group of subjects. Often, we need to compare frequency distri-
~ 30
i: butions between two or more groups defined by another variable; for
~ 20
eT
Cl
example, the distribution of oral contraceptive use among Colombian and
ù: 10
Spanish cases of cervical cancer. Thus, we wish to examine the association
o
Ever Never Unknown between two categorical variables: oral contraceptive use and country of
Oral contraceptive use
residence. One way to do this is to tabulate the data as in Table 3.3.
52
¡
f
,,
,
j
f
¡
Describing and presenting data
Quantitative variables (such as age) can also be tabulated by grouping Figure 3.6.
Distribution of cervical cancer cases by
the values of the variable. Table 3.5 shows the age distribution of cervical oral contraceptive use according to
cancer cases at the time of their enrollment into the study in each coun- country of residence (data from Bosch
try. The distribution of the response variable (age) is shown according to et al., 1992).
Colombia Spain
~30 15 (8.1) 7 (2.8)
30-39 48 (25.8) 41 (16.4)
40-44 27 (14.5) 30 (12.0)
45-54 50 (26.9) 61 (24.4)
;:55 46 (24.7) 111 (44.4)
53
Chapter 3
0 20 40 60 80 100 120
into a single table. Two or three sim-
ple tables may be better than a single
Vitamin C intake (mg/day)
large and complex one.
omitted without any loss of information. The same data are shown
in a mu ch simpler and clearer manner in graph (b).
(a) (b)
1000 41
i:
41
L_~.~___.____._,_~_________
Figure 3.8.
Two graphical presentations of data
fram 750 subjects on their country of
birth, smoking status and serum con-
centrations of beta-caratene:
(a) 'business' graph; (b) 'scientific'
graph (repraduced, with permission,
fram Jolley, 1993. (Q by The Lancet
Ud, 1993).
55
Chapter 3
Further reading
, .",'.' ," " ,.", ~ ,', a'p)cßA. 'ke'" issuês ': ,:,.., " . ,...:,;:,
'~ ,"'.' ~~ _', _-'_.'~," .~. ',- -,..'.' , ~ ~..,,!J./ '"_" ';.,~-w" "~~,-'';'::~ '.,," _,.)., -'~."
" ,,' . .,,_, _', _" '_',
, -,--_,.c-;:""-__,-,,,," .__ _,,,,..,,,,-
J~' , '-,,,-.-'-','
,"" __,_,,',', ,_,_ ,,' ,'.'",-",'"
',,-"-;,''',,'-''_'.u'',' ,',,, ,,''."__H'
_u __,'
-',,' ,.c_
-,' _',
"-',,""', "_C"',
,"C'_,"', ",_d'.',,,'','-''-,'-,';"'-,
-"-,,,-':, ' .-, ,_', ,. ;
* Further details of ways of pre-
· Afterthe datatiave been editep,they shouldbe ex~mined~sirigsimpletabula-
senting and summarizing data
are given by Altman (1991) and
tions,gràphs ~àndsulllTarymeastJres. '.' ...',
, ",:,' "" :, '_., -, ,:" '. _, _,' _, -:., , -: :_, "; - ;. ,:;;,~ '," ,_ ,,_' ,,_ '_'_""" -":_ '_, ;,' ", ' ,,:. -'._,,: '_, _' ,.',,, ';"_"" " ;_ '_ ,,_ . - _, '.",,' ';. '.':'.-1 ' ,
Sland (1987).
. Theèhoicgóitflt7cOJl'ectmêthocÎs: to present andsuinmarize thedataq8.pends. .
on the type of dàtäcollectßd:quaÌ1t1tativeor categoricaf. ' ,
* An excellent book on graphical
methods is that by Tufte (1983).
· The freqU~~¿~ distributiori ofaquantitative variablecan be presente'dintables, .'
56
Chapter 4
Measures of occurrence of
disease and other health-
related events
4.1 Introduction
Epidemiological research is based on the abilty to quantify the occur-
rence of disease (or any other health-related event) in populations. To do
this, the following must be clearly defined:
(1) What is meant by a case, Le., an individual in a population who has
the disease, or undergoes the event of interest (e.g., death).
(2) The population from which the cases originate.
(3) The period over which the data were collected.
57
Chapter 4
4.2.1 Prevalence
a Period prevalence is a variation that
represents the number of people who Point prevalence is the proportion of existing cases (old and new) in a
were counted as cases at any time dur- population at a single point in time.
ing a specified (short) period, divided
by the total number of people in that
population during that time. This mea-
sure is used when the condition is No. of existing cases in a defined population at one point in time
Point prevalence =
recurrent and non-fatal, and so is sel- No. of people in the defined population at the sa me point in time
dom used in cancer epidemiology. An
example of period prevalence would be
the proportion of women who have
used oral contraceptives at any time
du ring the 12-month period preceding
This measure is called point prevalencea because it refers to a single
the day of the survey. point in time. It is often referred to simply as prevalence.
58
¡
'\
L -..\~
!
1
l
Measures of occurrence of disease and other health-related events
o- - r- ------
--------------------- ------
- - - -0 . r- - - - -
--------------------- ------ o = disease onset
Figure 4.1.
Changes in the disease status and
migration of members of a population
over time, and how these changes
affect the prevalence of the disease in
-0 m r = recovery the population.
d = death
- - - - - - - - - - - - - - - - - -0 m = migration
- - - - - - - - - - - - - -m
--------------- 0 d- - - -
- - - - - - - -0- r- - - - - - --
t1
Time (t) ~ tz
59
Chapter 4
4.2.2 Incidence
The number of cases of a condition present in a population at a point
in time depends not only on the frequency with which new cases occur
and are identifed, but also on the average duration of the condition
(Le., time to either recovery or death). As a consequence, prevalence may
vary from one population to another solely because of variations in
duration of the condition.
Prevalence is therefore not the most useful measure when attempting
to establish and quantify the determinants of disease; for this purpose, a
measurement of the flow of new cases arising from the population is
more informative. Measurements of incidence quantify the number of
new cases of disease that develop in a population of individuals at risk
during a specified time interval. Three distinct measures of incidence
may be calculated: risk, odds of disease, and incidence rate.
Rísk
Risk is the proportion of people in a population that is initially free of
disease who develop the disease within a specified tIme interval.
60
Measures of occurrence of disease and other health-related events
Odds of disease
Another measure of incidence is odds of disease, which is the total num-
ber of cases divided by the total number of persons who remained disease-
free over the study period.
No. of new cases of disease ansing in a defined population
over a given period of time
Odds of disease =
No. of people in that population who remain disease-free during that period
Figure 4.2.
Number of Follow-up of the 100 disease-free indi-
new cases viduals described in Example 4.5.
of disease
Number of ¡ (10)
individuals Number of
initially at individuals
risk Number of
currently non-diseased
(disease-free) at risk individuals
(100)
(stil at risk)
(90)
l
Time (t)
Thus, it is possible to calculate the risk and the odds of
developing the disease during the study period as:
61
'-,
Chapler 4
Risk and odds of disease use the same nurnerator (nurnber of new cases) but
different denominators. ln the calculation of risk, the denominator is the total
number of disease-free individuals at the beginning of the study period,
whereas when calculating the odds of disease, it is the number of individuals
who remained disease-free at the end of the period (Exarnple 4.5).
Incidence rate
Calculations of risk and odds of disease assume that the entire population
at risk at the beginning of the study period has been followed up during the
specified time period. Often, however, sorne participants enter the study sorne
time after it begins, and sorne are lost during the foIlow-up; Le., the popula-
tion is dynamic. ln these instances, not aIl participants wiIl have been £01-
lowed up for the same length of time. Moreover, neither of these two mea-
sures of incidence takes account of the time of disease onset in affected indi-
viduals.
To account for varying lengths of foIlow-up, the denominator can be cal-
culated so as to represent the sum of the times for which each individual is at
risk, Le., the sum of the time that each person remained under observation
and was at risk of becoming a case. This is known as person-time at risk, with
time being expressed in appropriate units of rneasurement, su
ch as person-
years (often abbreviated as pyrs).
Figure 4.3.
Calculation of an individual's time at
risk and talaI persan-lime al risk for
lhe nine sludy subjects described in
Example 4.6.
62
Measures of occurrence of disease and other health-related events
group of nie persons. Subject (1) joined the studyat the begig of 1980 and was
followed up thoughout the study penod. Therefore, (1) was at nsk of becomig a
cae for the entie five year of the study. Subjec (4) also joined at the begig of
the study; but was last contacted at the end of 1983; thus, (4) was at nsk for only four
year. Subject (6) joined the studyat the begig of 1982, and developed the di-
ease by the end of that year; afer that, (6) was no longer at nsk (assumg there ca
be no recovery from the disease of interest). The total person-year at nsk is the SUf
of al the individuals' tie at nsk.
The incidence rate accounts for differences in person-time at risk and is
given by:
No. of new cases of disease arising in a defined population over a
given time period
Incidence rate =
Total person-time at risk during that period
When presenting an incidence rate, the time units must be specified; that
is, whether the rate measures the number of cases per person-day, pers on-
mon th, person-year, etc. Although the above definitions of risk, odds and
rate are now widely accepted, the terms risk and rate are used interchange-
ably in much of the literature, and especially in older publications.
63
Chapter 4
Risk depends both on the incidence rate and on the duration of the at-
risk period. It is also affected by mortality from diseases other than the
disease of interest; sorne of those who died from other diseases would
have been expected to develop the disease of interest had they survived.
If mortality from other diseases is disregarded, and if the incidence rate is
constant throughout the period at risk, the following relationship
applies:
..,,'.
'-' -,....',
'.-, --,'-'..-. -', .',-"
,-.-'.
-'.' "" '
, of developing this condition during a five- year period (assuming no. Qther .
',-,'.. ,'-',",-
,','-,-
causesofdeath) is given by: .
;,-
i-"'" -
:.,,:.:
" c_..
- :'.' ,-',.'
,: ",",,', -q',.._
- -,_ ..... ,"--
'_', ','-,.:-_ ..~......,,'
- , -. "_..
" ..' ,_:',"
:: -i :''.--::-:,.i,:.-..'~)_ - " - ~:-_:-..-; ,--,,~-..'_:_:-'_.' :-':::'. ':",','" __-),',;'.--:,;;/",: :,_:;;-,:'.-,::;' .'.'" ',\-,'_,;'- _:,:: ,: '-,"" ,'- ': - ;' :.-: _:': - :-:~'.".-,_--:,- :... ",," :'-"~ è,:: ,-_,,:-
ar'.,,,,,Tisk = 0.3per'
,.", ....,.",,,,',,,. '5 years=1.5 = 150% ".,.....
64
Measures of occurrence of disease and other health-related events
Thus the mean annual incidence rateofsto ...' ;' '.c.ancer .,'
Càli during the yeaTs 1982~86was 2 iper100p00.py,s. . .'
person-time at risk in each calendar year. Thus in the above example, the
sum of the annual population estimates for the years 1982-86 could
have been used to provide an estimate of the total person-years at risk
for the entire study period.
65
¡
,, ~"'l..
i
1
Chapter 4
Crude rates are widely used, in part because they are summary mea-
sures and so are easily interpreted, and in part because their calculation
requires relatively little information. Crude rates may obscure the fact
that subgroups of the population have marked differences in incidence;
for instance, people in different age groups have a different risk of
death. This should be borne in mind when comparing cru de rates from
various populations, as disparities might reflect differences in their pop-
ulation structure rather than in disease incidence (see Section 4.3.3).
To gain an understanding of certain epidemiological aspects of a dis-
ease, more detailed rates, specifie for sex and other demographic char-
acteristics such as age, are needed. For example, age-specific rates can be
calculated as follows:
66
i
L
i
f
Measures of occurrence of disease and other health-refated events
Figure 4.4.
Age-group (years) __ Lexis diagram showing the follow-up of
1950~0- 35- 40- 45- 50- 55 the study subject described in Example
Enln
~ 4.10 and the calculation of his person-
1955- months contribution to each calendar
Calendar "-
year 1960- "- period and age stratum.
"-
l 1965- ~
~Xil
1970
67
Chapter 4
Even when data on date of birth are not available, the Lexis diagram can
be used with routine data to describe the incidence of a disease in succes-
sive generations. Mortality rates from cancer of the lung in men in
England and Wales during 1941-78 are shown in Table 4.1: columns show
Table 4.1. changes in the incidence rates with age, and rows show changes in the
Mortality (per million person-years) age-specific rates over calendar time. ln any age x calendar-period two-way
from cancer of the lung, in males in table, diagonallines represent successive birth cohorts, although the earli-
England and Wales, 1941-78, byage
group and calendar period; the central est and the most recent birth cohorts (in the extremes of the table) wil
year of birth is indicated on the diago- have very few data points.
nais.
a
Age-group (years)
Year 01 death 0- 5- 10- 15- 20- 25- 30- 35- 40- 45- 50- 55- 60- 65- 70- 75- 80- 85 and
1946-50 1 '2 ..1 '4 'i2.. 18.. 41.. 97.. 242.. 555.. 972.. 1375 ~ 1749.. 1798. 1442.. 1130.. 791.. 497
1951-55 1 ::0:: 1 :: 2:: 8.. 14.. 37 ~OO .. 250.. 584.. 1232.. 2018 ~ 2575.. 2945. 2651.. 2087.. 1444.. 927"
1956-60 1 0 "" 0,2, 5.. 13.. 36.. 94.. 253.. 594.. 1254.. 2326.. 3330.. 3941. 3896.. 3345.. 2271 .. 1438" 1866
1961-65 o 1 ,2,5.. 11 .. 33.. 91 .. 225.. 566.. 1226.. 2290.. 3673.. 4861. 4994.. 4530.. 3423.. 2062" 1871
1966-70 o o 0 3, 4.. 11 .. 25.. 76.. 218.. 532.. 1165.. 2208 ~ 3703.. 5281. 6223.. 5931.. 4578.. 3490" 1876
1971-75
1976-78
o
o o 1 3 8"""7,4,1,2,2,,,1886
o 0 1 4,10.. 25.. 58 .. 178 .. 505 .. 1074.. 2082 ~ 3552.. 5185 . 6834.. 7284.. 6097.. 4384"'1881
1951 1946 1941 1936 1931 1926 1921 1916 1911 1906 1901 1896 1891
Birth cohort (diagonal)
a Data lrom OPCS, (1981),
The diagonals of Table 4.1 (from upper left to lower right), for instance,
define the lung cancer mortality experience for successive generations of
men who were born together and hence aged together. For example, a
man aged 40-44 years in 1941-45, was aged 45-49 in 1946-50, 50-54 in
1951-55, etc. To be 40-44 in 1941-45, he could have been born at any
time betweenJanuary 1896 (44 in 1941) and December 1905 (40 in 1945).
These so-called birth cohorts are typically identified by their central year
of birth; for example, the 1901 birth cohort, or more precisely the 1900/1
birth cohort, contains those men born during the lO-year period from
1896 to 1905. The diagonal just above this one shows the rates pertaining
to the 1896 cohort, Le., those men born between 1891 and 1900. As the
years of birth for each cohort are estimated from the age and calendar peri-
od data, adjacent cohorts inevitably overlap, Le., they have years of birth
in common. When data on exact year of birth are available, these esti-
mates need not be made, and so successive birth cohorts do not overlap.
Analyses by birth cohort thus use the same age-specific rates as in cal-
endar time period analyses, but these rates are arranged in a different way.
Comparison of rates in successive birth cohorts allows us to assess how
incidence may have changed from one generation to another.
68
Measures of occurrence of disease and other health-related events
(a) (b)
10000 10000
1000 1000
Qi Qi
ii0 ii0
Ul Ul Age-group
ci ci (yrs)
g g -- 25-9
~ ~ -x 35-9
;:
a. Year of death ;:
a.
i: 100 i: 100 -- 45-9
-+ 1941-45
,g ~ -- 55-9
'* 1956-60
.e .e -0 65-9
.. 1971-75
Qi Qi -- 75-9
a. a.
2al
II
õí
ex ex
10 10
1 ~ 1 1 1 l l 1 1 1 1 1 1 r 1 ..
25- 35- 45- 55- 65- 75- 85+ 1941- 1951- 1961- 1971- 1976-8
Age-group (years) Year of death
ln certain situations, cohort analysis gives the most accurate picture of Figure 4.5.
Mortality from lung cancer in men in
changes in the patterns of disease over time, for example, if exposure to a England and Wales, 1941-78. (a)
potential risk factor occurs very early in life and influences the lifetime risk Rates presented ta show differences in
of a particular disease, or if the habits of adults are adopted by successive age-specifie curves between three
selected calendar periods; (b) rates
generations (as with cigarette smoking and lung cancer, and exposure to presented ta show secular (calendar)
sunlight and malignant melanoma of skin). ln other situations, secular trends in age-specifie rates. For c1arity,
analysis might be more appropria te: for example, if exposure to the risk only rates for alternate age-groups are
shawn; the first five age-groups are
factor affects aIl age groups simultaneously (as with the introduction of a omitted because of the small number
new medical treatment). However, in most situations, it is not clear which of deaths (data from Table 4.1).
analysis is most appropriate to describe temporal trends, and the results of
both should be examined.
69
,l
f
!
Chapter 4
(a) (b)
10000 10000
íi
¡¡
0
UJ
el
1000
Year of birth
-+ 1871
íi
¡¡
0
UJ
el
1000 fr~ Age-group (yrs)
-+ 25-9
~
g, -- 1881 g,
UJ -- 1891 -- 35-9
:; ~
a. -- 1901 :-
a.
-- 45-9
c: 100 100 -- 55-9
-0 1911 c:
-0 65-9
~ -+ 1921 ,g
.e -+ 75-9
-- 1931 E
.. ..
~
CI -2 1941 CI
a. a.
CI CI
10 10
II II
10 10
1 / 1 .i 1 i 1 1 1 l 1 1 l ! 1 l 1 1 1 l 1 1 1
25- 35- 45- 55- 65- 75- 85+ 1á61 1881 1901 1921 1941 1961
Age-group (years) Year of birth
Figure 4.6. Descriptive analyses by age, calendar time and cohort are a popular epi-
Mortality fram lung cancer in men in
England and Wales, 1941-78. (a) demiological tool for examining temporal changes in the incidence of a dis-
Rates presented to show differences in ease. These analyses are based on the inspection of tables and graphs, in
age-specifie CUNes for successive birth
cohorts; (b) rates presented to show
much the same way as described here, although statistical models can also
cohort trends in age-specifie rates. For be used to assess whether there is a statistically significant trend in rates
ci a rity, only rates for alternate age- over calendar time, or between birth cohorts (Clay ton & Schiffers,
groups or cohorts are shown; the first
five age-graups are omitted because of
1987a,b).
the small number of deaths (data fram
Table 4.1). 4.3.3 Controlling for age
For comparison of incidence between populations, crude rates may be
misleading. As an example, let us compare stomach cancer incidence
among men living in Cali, Colombia and Birmingham, England. The data
are extracted from Cancer Incidence in Five Continents (Parkin et aL., 1992).
de incidence rate (the rate for aIl ages com-
Table 4.2 shows that the cru
bined) for Birmingham was much higher than that for Cali. However,
before concluding that the incidence of male stomach cancer in
Birmingham (1983-86) was higher than in Cali (1982-86), the age-specific
rates for the two must be compared. Surprisingly, age-specific rates were
higher for Cali in aIl age-groups. The discrepancy between crude and age-
specific rates is because these two populations had markedly different age-
structures (Table 4.3), with Birmingham having a much older population
than Cali.
70
Measures of occurrence of disease and other health-related events
~~~~- --~~~-
Cali Birmingham
Age No. of Male Mean Male
No. of Mean
(years)b cancers population annual rateC cancers population annual rateC
(1982-86) (1984) (1982-86)
0-44 39 ~~.~n~_.--~~,_~
(1983-86) (1985) (1983-86)
524 220 1.5 79 1 683 600 1.2
45-64 266 76 304 69.7 1037 581 500
65+ 315 22 398 281.3 2352 291 100
44.6
202.0
Ali ages 620 622 922 19.9d 3468 2 556 200 33.9
a Data fram Parkin et al. (1992)
b For simplicity, only three broad age-graups are used thraughout this example.
C Rate per 100 000 person-years.
d This crude rate is slightly lower than in Example 4.9 (21.03 per 100 000 person-years) because cases of unknown age (35 in total) were
excluded here. The exclusion of two cases of unknown age in Birmingham did not affect the value of the crude rate calculated here.
The lower crude rate for Cali is thus explained by its male population
Table 4.2.
being younger th an that of Birmingham, and the fact that younger people Incidence of stomach cancer in males
have a mu ch lower incidence of stomach cancer than older people (Figure by age group in Cali, 1982-86, and
4.7). ln this situation, age is a confounding variable, Le., age is related to Birmingham, 1983-86. a
exposure (locality) and it is itself a risk factor for the outcome of interest,
stomach cancer (see also Chapter 13).
0-44 84 66
45-64 12 23
65+ 4 11 Figure 4.7.
Ali ages 100 100 Age-incidence curve of stomach can-
a Data fram Parkin et al. (1992). cer in males in Cali, 1982-86, and
Birmingham, 1983-86 (data from
Parkin et al., 1992).
As incidence rates for stomach cancer change
considerably with age, differences in the age distri- 500
bution of populations need to be considered before -0 Cali
attempting to compare incidence. One approach is - + - Birmingham
to compare age-specific rates, as in the example 400
71
Chapter 4
for the effect of age, it can equally be used to control for any other con-
founding variable such as social class, area of residence, etc. There are two
methods of standardization: direct and indirect.
Figure 4.8. dardized incidence rates; they can be seen as the crude incidence rates
The direct method of standardization. that these populations would have if their age distributions were shifted
from their actual values in the
mid-1980s to the age distribution
of the standard population.
These standardized rates are a fic-
tion: they are not the stomach
cancer incidence rates that actu-
ally existed, but rather those that
these two populations would
have had if, while retaining their
own age-specific rates, they had a
hypothetical (standard) popula-
tion. The fiction is useful, how-
ever, because it enables the epi-
demiologist to make sumffary
comparisons between popula-
tions fram different areas, or dur-
ing different time periods, which
are free from the distortion that
arises fram age differences in the
actual populations.
72
,
f \,"-....,
J
i
í
~ -.,""
Measures of occurrence of disease and other health-related events
73
Chapter 4
Cali Birmingham
Age No. of Male Mean
annual rateb
No. of
cancers
Male
population
Mean
annual rateb
(years)b cancers population
(1983-86)
0-4 NA 524 220 (1982-86) (1984) (1982-86)
79
(1985)
1 683 600
(1983-86)
1.2
45-64 NA 76 304 1037 581 500 44.6
65+ NA 22 398
Ali ages 620 622 922 19.9
2352
3468
291 100
2 556 200
202.0
33.9
NA, data assumed to be not available (see Table 4.2).
8 Data fram Parkin et al., 1992.
b Rate per 100 000 person-years.
74
Measures of occurrence of disease and olher health-related evenls
each locality, the numbers of cases observed and expected can be com-
pared, because both refer to the same population. The ratio of the
observed number of cases to that expected is called the standardized inci-
dence ratio (SIR) or the standardized mortality ratio (SMR) if a case is
defined as death. These ratios are usually expressed as a percentage.
ln the example above, the SIR (%) for Birmingham is 100; by definition,
the number of observed cases of stomach cancer is equal to the number of
expected cases when using the age-specific stomach cancer incidence rates
for Birmingham as the standard rates. The SIR (%) for Cali is 145, mean-
ing that the number of cases observed was 45% higher than that expected
if Cali had the sa me incidence of stomach cancer as Birmingham. This
result is similar to that obtained using the direct method of standardiza-
tion.
This method is called the indirect method of standardization. As with
the direct method, the results de pend in part upon the standard chosen.
However, the indirect method of standardization is less sensitive to the
choice of standard than the direct one.
75
Chapter 4
(a)
mation. It must be emphasized that, under certain circumstances, it may
not be appropriate to summarize disease rates in a single summary mea-
160 sure. Consider this example. We wish to monitor trends in mortality from
~ 155
ovarian cancer in England and Wales. The age-adjusted death rates for
,.Q. this cancer in England and Wales increased slightly from 1970-74 to
c:
~ 150 1985-89, as shown in Figure 4.lOa. However, trends in the age-specific
.e
G; 145 rates for this period reveal that they did not increase across aIl age-groups.
Q.
.l
f! 140
Figure 4.10b shows the rate ratio ofthe 1985-89 to the 1970-74 age-spe-
.t
~
cific rates. It becomes apparent that while death rates were increasing at
¿g 135 older ages (rate ratios above 1), there was no increase in women below age
130 ' 55 years (rate ratios below 1). If age-standardized death rates for aIl ages
1970- 1975- 1980- 1985-9
Year of death
are calculated, this information is lost, because mortality rates from this
cancer at younger ages were so low that they were dominated by the
much higher mortality rates in the older age-groups. So in this case, the
age-adjusted summary measure is misleading.
(b)
Before age-adjusted summary measures are calculated, the age-specific
1.6 rate ratios should always be examined to determine whether this
1.4 approach is appropriate. If these ratios vary systematicaIly with age, this
1.2
o
~ 1 information would inevitably be lost in the summary age-adjusted me a-
;; 0.8
õí 0.6 sure.
a:
0.4
0.2
o 4.3.4 Cumulative rate
25- 35- 45- 55- 65- 75- 85+
Age-group (years) The cumulative rate is another measure of disease occurrence that is
increasingly used in cancer epidemiology. This measure has been includ-
Figure 4.10. ed in recent editions of Cancer Incidence in Five ContInents (see, for exam-
Ovarian cancer mortality in England pIe, Parkin et aL., 1992).
and Wales, 1970-89. (a) Rates are A cumulative rate is the sum of the age-specific incidence rates over a
age-standardized to the 1981 female
population of England and Wales; (b) certain age range. The age range over which the rate is accumulated must
age-specifie mortality rate ratios, be specified, and depends on the comparison being made. Thus for child-
1970-74 and 1985-89 (rates in
1970-74 taken as the baseline). hood tumours, this might be age 0-14 years. ln general, however, the
most appropriate measure is calculated over the whole life span, usually
taken as 0-74 years.
The cumulative rate can be calculated by the sum of the age-specific
incidence rates (provided they are expressed in the sa me person-time
units, e.g., 100000 pyrs), multplied by the width of the age-group. Thus
for five-year age-groups, the cumulative rate would be five times the sum
of the age-specific incidence rates over the relevant age range (Figure
4.11). If the age-groups are of different width, each age-specific rate
should be first multiplied by the width of the corresponding age-group;
the sum over aIl age-categories yields the cumulative rate. This measure
is usually expressed as a percentage.
The cumulative rate can be interpreted as a form of direct age-stan-
dardization with the same population size (Le., denominator) in each
age-group. Thus, it avoids the arbitrary choice of a standard population.
Another advantage of the cumulative rate is that it provides an esti-
mate of cumulative risk, Le., the risk an individual would have of devel-
76
Measures 01 occurrence 01 disease and other health-related events
oping a particular cancer over a defined life span in the absence of any
other cause of death. The cumulative risk can be calculated as follows:
77
Chapter 4
As with rates, these proportions can be standardize for age (or any other
potential confounding factor).
Note that a proportional measure is not equivalent to a rate, as the
denominator is derived from the total number of cases, and not from the
population at risk. While proportional measures reveal the proportion of
cases (or deaths) that can be attributed to a particular disease, a cause-spe-
cific rate reflects the risk of developing (or dying from) a particular disease
for members of a specific population.
Proportional measures can be misleading because their denominator is
the total number of cases (or deaths), a measure that depends on the num-
ber of cases (or deaths) from aIl causes, not just that being studied. For
example, although the proportion of deaths due to cancer is greater in
middle-aged women than in elderly women, death rates from cancer are
actually higher among the elderly (Figure 4.12). This is because the total
78
'\ \"'"'
Measures of occurrence of disease and other health-related events
Figure 4.12.
(a) Female deaths in England and Wales,
1993. (a) Proportion of deaths due ta
cancer and other causes by age; (b)
cancer mortality rates by age (data
from OPCS, 1995).
75
.rII
1i
CI
'0
ro
Õ 50
CI
01
a:
ëCI Circulatory disorders
0
G¡
a.
25
o
5- 15- 25- 35- 45- 55- 65- 75- 85
Age-group (years)
(b)
10000
ãì
ro 1000
0
II
01
g,
II
~
Q.
00 100
0
0..0
..
CI
Q.
CI
1i 10
cc
1-
5- 15- 25- 35- 45- 55- 65- 75- 85
Age-group (years)
79
"\
\'"l..,.
Chapter 4
Further reading
* Most of the measures of dis-
ease occurrence discussed here
are dealt with in more detail in
Breslow & Day (1987) and " ,,"__".'.'.'
"'. .' '-",' ," ,",_ .' ,',
", ",-' "-. '_,'_ ..,- -",,-"'--',"',-
'-'-, ',-,',.-'- '."',,,'" .H'_ ',,""_ _ .'
"''''''-''''-'--''''-''::''','-
,,_.'.' '-,: '., ',.' .', - ,-, .:\':. - ','- , .' -" .; - .: :. :,._'.: '~, -',,':-': ,,,'.- -,,', : ': ,: ""-::''-':: ,,-_: ::' .',,' /:-': ','/ ;.'-::.:' ~::.' :-::. ::,-:;'::'::~ .- '~'-~":~ ':',: '(':')"':"\'\:~';' ;.:';+:
,:"'.;,,--"'. "_'_"',':.' _ '_-' "', -,",_'_ '_ -'" .,..__ ",_, -_,-_ '-."",': .'i_(,_ ,"_,: '."_"-::":__'::_'.:,::::~:''':~:'':_:-,':,:,:\':.~:' '_:.'/:,,::.:'-:~\~,_,
Estève et al. (1994).
.There are two major measures ofdiseaseo~cyìien, .' .'.., ..' .... '.
· dence: Prèvalence refers, to the tàtal riumbèr òfexistin , . . " . se~ of, . '.
* A more elaborate discussion of
a condition in a population at a specificpointintinie, Incidence JefèJsto the
age, calendar time and cohort
occurrence of new cases in a population over a specific timeperiod:' .',' . .
effects can be found in Clay ton & , --.;_'__"" ,,', ','" ",'," - - "'H ,'. ..".. __
, - -',. -,.' ": ','
, , - ""'" \': .,' ,,' -,'..., :-,-. ,', ",; -:' ,-':, ,- '-~ " , ,'- -', ,,":-'-~
'.',,"' "".::',",'''.".'.:----,',-'',',.-
-:,," _ ',"" ".',',," - ,,': -,-:.:',-:::,-'-:-,-"''':'-:-'_:'''',:
' ;'-,':"" -:,- ,',,: ' - ,: :" ',--_.:" " ~ - ,--- ,. :. '" -- ',- ,,' ,,;
'--.- .:-,'.-
,-',.;---:.-',,-,:.,':-,":,-' -,',' ,,',",,'-;..,-::_,-,' --'-',-"-'''''-'-'
Schifflers (1987a,b).
· Incidence can be measured as eitherr!s.k, odds of diseasG,orrate. T 'calcu-. . .
lation of. risk and odds requireSco ., . tefollow-:upofallstu odhé
entire study period.lncontrast,th . culation'òfral "e 'tindi":
vidual differences in lengthoUollow~up~'" .'.
'-.:-",.,'" ---.' --" ,'-"'".,-,,,'''-,,''.-
"'-,
,__"",,_
"",-",-",,' -",," '',',
,_, _"',,',.
,,,"... .,.,," ','"
,".,-,- ,,- ....'_',,,
--- . .
,,', '.,-",'-,_.---"-,".'."-,:,:
'" ,'V,_.. --','-'
~_-",-,:
. .....- .-c_.:',"
-,: -, --::': .--' ',' '::, ._ -':",
_ .".__' '.,d',': ~:".:.~-':':'-;;::,
'...,,__", --::':".,.
- " ,.i.,~,'.!,:',:-:,-,;
"__,,,,.,.. ..'~_::,.::',:::,.;::,' :/: ~;
80
i
t
,.
i
1
1
Appendix 4.1
Conventional standard
populations
The choice of the standard population to be used in the direct method
of standardization is, to a certain extent, arbitrary. For example, if the aim
is to compare disease occurrence in several groups in England and Wales,
an appropriate standard might be the adult population of England and
Wales. On the other hand, this may not be an appropriate standard when
making comparisons between countries.
For international comparisons, various conventional standard popula-
tions have been used (Table A4. 1). These standard populations range from
an African population with a low proportion of old people, through an
intermediate world population, to a European population with a high pro-
portion of old people. ln the earliest volumes of Cancer Incidence in Five
Continents, rates were standardized ta these three populations; however,
the European and African standards were dropped in Volume iV
(Waterhouse et aL., 1982) and replaced by cumulative rates over the age
ranges 0-64 and 0-74 years.
The truncated population is derived from the world population but
comprises only the middle age-groups. This truncated population was
often used in the past, because data for older age-groups were likely to be
less reliable than those for the middle age-groups, and because for most
forms of cancer, virtually no cases arise in groups under 35 years.
81
Appendix 4.1
Table A4.1.
Age group African World European Truncated
Conventional standard populations
(years)
used for international comparisons.
0 2 000 2400 1600 -
1-4 8 000 9600 6400 -
5-9 1 a 000 1 a 000 7000 -
10-14 10 000 9 000 7 000 -
15-19 10 000 9 000 7 000 -
20-24 10 000 8 000 7 000 -
25-29 1 a 000 8 000 7 000 -
30-34 1 a 000 6 000 7 000 -
35-39 1 a 000 6 000 7 000 6 000
40-44 5000 6 000 7000 6 000
45-49 5 000 6 000 7 000 6 000
50-54 3 000 5 000 7000 5 000
55-59 2 000 4000 6 000 4 000
60-64 2 000 4 000 5 000 4 000
65-69 1000 3 000 4000 -
70-74 1 000 2 000 3 000 -
75-79 500 1 000 2 000 -
80-84 300 500 1 000 -
85+ 200 500 1000 -
Total 100 000 1 00 000 1 00 000 31 000
82
Chapter 5
Overview of study designs
83
,¡
\
f
í
Chapter 5
84
Overview of study designs
Figure 5.2.
Outline of an inteNention triaL.
/ Intervention group
( Outcome
No outcome
~ Control group
( Outcome
No outcome
Direction of inquiry
ln Example 5.1, women who had already developed breast cancer (and
had surgery) were randomized to standard treatment accompanied by a low-
fat dietary regimen ('intervention group') or standard treatment alone ('con-
trol group') to assess whether a low-fat diet reduces recurrence of the tumour
and increases survival.
85
Chapter 5
86
-~~~.- - _. ,'~ Overview of sludy designs
No out come
~ Unexposed group
( Outcome
No out come
Direction of inquiry
aged34to59
Cohort studies take individuals and classit them according to their expo-
sure status. Sometimes they can be simply classified as exposed/unexposed.
More usually, various degrees of exposure can be identified. ln Example 5.5, the
US nurses (the 'cohort') were classified into five groups (quintiles) according to
87
Chapter 5
their levels of fat intake. The incidence of breast cancer ('outcome') was then
measured and compared across these quintiles.
Figure 5.4.
Outline of a case-control study. Exposed 1
Cases
Unexposed ~
Exposed 1
Controls
Unexposed ~
Direction of inquiry
88
Overview of study designs
"._:.:......:............-.."..::..-.:.........'.::.;.".,'::.,:;;,.:,'.-.,.....,,:...,..-...:',.,:....:...
89
Chapter 5
which the cases were drawn. If not, the results wil be distorted by selection
bias (see Chapter 9). Secondly, one cannot be sure that the exposure did
precede the disease (except when the exposure is a fixed attribute such as
blood type that does not change over time). ln Example 5.7, it is conceiv-
able that the reported diet may be a consequence rather th an a cause of
breast cancer.
ln Example 5.9, a sample of 1815 study subjects was selected from a weIl
defined population (healthy blood donors in certain prefectures of Japan)
and a blood sample taken from each of them at a particular point in time to
measure both the levels of H. pylori antibodies ('exposure') and the levels of
pepsinogen 1 and II ('outcome').
ln this type of study, it is crucial to ensure that the sample of subjects who
participa te in the study is representative of the whole population to whom
the results wil be extrapolated. Otherwise, the results wil be distorted by
selection bias. The best way to safeguard against selection bias is to use ran-
dom sampling methods. These methods ensure that chance alone determines
who wil and who wil not be included in the sample (see Chapter 10).
Cross-sectional surveys are generaIly used to estimate the prevalence of
common conditions of reasonably long duration. Thus this design is not
appropriate for studying diseases such as cancer, since it would be necessary
to survey a very large population to identify enough cases ta be able to draw
any conclusions. Moreover, prevalent cancer cases are a biased sam pIe of aIl
cases, in which those with long survival tend to be over-represented.
The main use of cross-sectional surveys in cancer epidemiology has been
to examine the distribution and determinants of common conditions, such
as human papilomavirus infection or skin naevi, which are known (or sus-
pected) to be associated with cancer. This type of study has also been used
90
Overview of study designs
.-.
Individuallevel
Many routine data-collection systems collect data on personal attributes
such as age, sex, place of birth, place of residence, occupation, etc. Cancer
occurrence can then be examined in relation to the se variables. The objec-
tive is to search for patterns that might suggest or confirm specific etiologi-
cal hypotheses.
ln Example 5.10, data on place of birth, place of residence and age at
migration ('exposures') and on breast cancer ('outcome') were available
from population censuses, social security records and cancer registries for
each of the study individuals.
The key feature of this type of study is that data on both the exposure
and the outcome(s) of interest are obtained for each of the study subjects
from routine data-collection systems. ln terms of their analysis and inter-
pretation, they can be regarded as being similar to cohort or case-control
studies. For instance, Example 5.10 can be regarded as a cohort of women
who were classifed into four different exposure categories according to
their place of birth, residence and age at migration. These women were
th en foIlowed up in time and the occurrence of breast cancer in each
group was measured and compared.
The main advantage of routine-data-based studies in relation to other
observational studies is that they can be carried out relatively quickly and
cheaply because the data have already been collected and there is no need
to contact the individuals. A major limitation, however, is that few vari-
91
Chapter 5
Figure 5.5.
Age-adjusted incidence rates of
female breast cancer for USA residents
(1972-85) by birthplace and age at ~ 60 i: Homeland (Japan)
migration, and for Japan (1973-81) II
CI
(reproduced, by permission of Churchill Late immigrant
Livingston, from Shimizu et al., 1991).
o
;: 50
o
o 40 _ Early immigrant
oo .. US-barn
:: 30
CI
~ 20
u
i:
~ 10
"(3
i: o
1.0 1.7 2.4 2.6
Ratio
ables are usually available and they have not been collected with the spe-
cifie needs of the study in mind. ln Example 5.10, the variables 'country
of birth' and 'country of residence' are only proxy measures for more bio-
logically relevant exposures, such as reproductive factors, diet and other
lifestyle characteristics, for which no data were available in the routine
data-sets.
92
=--==-~-=~~ ". Overview of study designs
Figure 5.6.
Estimated daily consumption of fat per
cap ut in 1964-66 and age-adjusted
200 breast cancer incidence rates in
women aged 35-64 years in 1972-77
in 24 countries (reproduced, by per-
mission of Oxford Univesity Press,
(/
:; 150
Co SWI CAN
oo .. . NZ
o FRA
oo
....
al
Co
. UK
ISR FRG .. . . DEN
AUST.. SWE
al
o .NOR
i:
.g 100
'õ
.= . . FIN
al
CI
ca
.YUG
SPA .DDR
~
::
HK .CZE
ca ROM. ·
¡; .POL
:i . CUBA .HUN
ê 50 . SIN
-:
.JAP
o
o 50 100 150
Total fat intake per caput (g per day)
93
"'
Chapter 5
~ ','_: _' . - ~ _-:" ' :, _ _' _ ,_ ." '_ -' ': _ _,.,-, -.,. _ -'_,,_ d.. '" , ' ',- _-- ,'., _'" -"i'; ,': - \_": _,-,-":,::' _ ',":
Example 5.12. A study wàs carried out to examine theassociáJion between
prevalence ofHelicobacter pyloriínfection and mortality (roni gasúic can~
cer in 46 rural counties of the People's Republic of China. . The results are
plotted in Figure 5.7 (Forman et aL., 1990).
Figure 5.7.
Prevalence of H. pylori IgG antibodies
and gastric cancer mortality in 46 rural 70 .
Chinese counties; each county is rep-
resented by a dot in the graph (repro- .
duced by permission of Wiley-Liss, 60
Inc., a subsidiary of John Wiley &
Sons, Inc., from Forman et al., 1990).
0
0
0 50
. .
,.
;:;:
== ~
01 01 .
0;:
t: QI
E -: 40
.
~ ,
tD
QI 0
0i:
al .. 1
01 ..
o ~
o QI
E .~
fi ..
30 . .. . . ,.
01 11
C) ::
.. .
E
:: . .
.... .
~ 20
.. . .
10
. , . .
o
. .. . ... 1
.
.
o 10 20 30 40 50 60 70 80 90 100
Prevalence of H. pylori IgG antibodies
level are subject to the 1 ecological fallacy' (see Section 11.2). For instance,
in Example 5.11, we do not know whether the women who developed
breast cancer in each country were those who actually consumed a high-
fat diet.
Despite these limitations, ecological studies may be the best approach
to study exposures that are easier to measure at a group rather than at an
individuallevel, such as air pollution and water quality. They are also
useful for monitoring the effectiveness of population interventions such
as health education campaigns and mass screening programmes (see
Chapter 16).
95
Chapter 5
Table 5.1.
Hypotheticaf data from a cohort of
20 000 individuais followed up for a No
period of two years.
No. initially at risk 4000 16 000
Deaths 30 60
Person-years at riska 7970 31 940
30/(4000-30 ) 0.00756
Disease odds ratio 2.0076
60/(16 000-60) 0.00376
a Assuming that on average ail deaths occurred in the middfe of the follow-up period.
96
Overview of study designs
sible for, assuming that the relationship between exposure and disease is causaL.
It can be calculated either as the difference of risk (risk difference) or of rates
(rate difference).
or
97
\
\.i.~
Chapter 5
contrast to the relative risk, however, the magnitude of the excess risk can-
not be generalized to other populations because it depends on the baseline
incidence in the unexposed group, which tends to vary between popula-
tions.
It is useful to express the excess risk in relation to the risk (or rate) in the
exposed group. This measure is called the excess fraction (also known as the
ex cess risk percentage or attibutable risk percentage) and can be calculated as
Excess fraction (%) = 100 x (excess risk/risk (or rate) in the exposed)
ln the above example, the excess fraction for diseases A and B would
be
Excess fraction (%) for disease A = 100 x (15 per 100000 pyrs/20 per 100000
pyrs) = 75%
Excess fraction (%) for disease B = 100 x (40 per 100000 pyrs/80 per 100000
pyrs) = 50%
98
Overview of study designs
Example 5.14 shows that 44% of the deaths that occurred among
male British doctors who smoked could be attributed to smoking
(assuming causality). The proportion of deaths that could be attributed
to smoking varied by disease. For those diseases shown in Table 5.3, this
proportion was highest for lung cancer (93%) and lowest for vascular
diseases (37%). However, if smokers had never smoked, the total nu m-
ber of deaths prevented would have been much greater for vascular dis-
eases (606 per 100 000 pyrs) than for lung cancer (195 per 100 000
pyrs).
Similar absolute measures of effect can be calculated when those
exposed have a lower risk of developing the disease than those unex-
posed. ln these circumstances, we would have
Risk reduction = risk (or rate) in the unexposed - risk (or rate) in the
exposed
ln Example 5.15, 40% of ovarian cancer cases could have been pre-
vented among never-users if they had used oral contraceptives.
5.3 Conclusions
Before implementing a study, careful consideration must be given to
the appropriateness of the proposed study design, especially in terms
of practical feasibilty, information to be obtained, expected duration
of the study and total costs. The advantages and disadvantages of each
of these study designs are covered in more detail in subsequent chap-
ters (7-11).
99
Chapter 5
Further reading AlI the measures of effect discussed in this chapter can be directIy calcu-
lated from intervention and cohort studies, since the incidence of disease
* References to books and in those exposed and in those unexposed is known. This is not the case in
papers dealing with each study other study designs, where these measures of effect can be estimated only
design are given in the relevant indirectly. For instance, in case-control studies, the subjects are selected on
chapters. the basis of their disease status (sample of subjects with ('cases') and with-
out ('controls') a particular disease), not on the basis of their exposure sta-
tus. Therefore, it is not possible to calculate the incidence in exposed and
unexposed individuals. It is, however, possible to calculate the odds of
exposure among cases and the odds of exposure among controls and obtain
an odds ratio of exposure (see Chapter 9). It can be shown that, depending
on the sampling scheme used to select controls, this measure provides an
unbiased estimate of one of the three relative measures of effect considered
in Section 5.2.1. This is discussed in detail in later chapters (9-11, 16).
- ln inteivention studies, the allocation of the study subjects to the different study
groups is implemented by the investigator. Thus, if conducted properly, the Interven-
tion and the control groups would be similar in ail respects apart from the exposure
under study. There are two types of intervention study: "
Clinical trials, where the niain aÎm is toassessthevalÜe of new forms qftreatment.
Field trials, where the objective is to evaluate whether an intervention decreases the
risk of disease among disease-free people. Field trials can be conducted atan indi~
viduallevel, .if the unit of allocation is the individual, or atanaggregated level, if the
unit is a group of people. Community trials are an eXiimple of trials carried out at an.
aggregated level, where whole communities are",'.,
the unit of allocation.
, ,- '-,". ' -', -" , -', ,',;, ".-' -", "~,
:""'_ c'._,"',,'-_
;:,-",d""',-','-..".' ,"-': ":;",',,-",..',-'
._;"':.'_',_._ _.,- -'0,,',;,," ' -:,""'.,-. " ,,- -:,
""_,'_"','__',
': '-- " ",'- : ": --t ;. ',- ,,-: _ ':'~"-:'-- '-' -,.: "- - :::- :,' - : ,': ,,' " :", '. ,,' '- ' -,,, ,-:-'
-ln obseivational studies, the researchers limit themselves to observing theoccur':
rence' of disease in' people who are already segregated . into diferentexposure
groups. There are various types of observational study: ", . .
" -':""-',':,",":.'-,"'_.._' -' - '-.- ,-, .-..,:,', :
'_ - .". ,'_._n,
',' '.',,',,' - __.._
,,' ':' n'_ , '" ':_:.. '_,_'",.,_,",
,_",_" '
,':-',,-.,'"
, ,', ," ',' ,,:", ,-:"""':"",,::":
" ""'..' ',' ,,:,..', ''',"
"'"":,,,',,-,'.'-:',"':"'-----':":"",'-: ",::,", ".',:"""'-
100
Overview -of study designs
Routine-data-based studies, in which the data are derived from routine data-collec-
tion systems (e.g" cancer registration or death certification). They may be carried out
at an individual level if information on the exposure(s) and outcome(s) of interest is
available for each of the study subjects or at an aggregated level (ecological studies)
if the group rather than the individual is the unit of study.
· Once the data.,from a partîcular study have been collected, the association between
the exposure and the Qutcome of interest can be quì:mtified by calculating an appro-
priate meastJe of effect. This may be expressed as either the ratio of the measure
of disease occurrence in the exposed relative
.:........:..::'--,-'.:-:-':_-.......'_..........._-..:.:.:..-:.....:-:..-.....::........-:-.....:....._-.. to that in ..........-.:-............-.....
................................- the uriexposed (relative mea-
..... .................................-....
sure) or as the diference between the two (absolute measure). The first type of mea-
sure is particularly important when assessing etiology, whereas the second type is
more useful for evaluations of the public health impact of the association.
~Each .studyqesign ... has its. .own . limitations. and strengths. Theseare. considered .in
detail.in subsequentcha.pters of this book (Qhapters 7__13).
101
,l
i
1
Chapter 6
Evaluating the role of chance
of the true population proportion. This implies that our sample estimate 'l
Population Sample
is subject to sampling errar. The proportion of current smokers in the whole proportion proportion, p
town is unlikely to be exactly the 23% found in our sample, due to sam-
pling error. The question is, how far from 23% is it likely to bel Figure 6.1.
To try to answer this question, we first recognize that the sample we Sampling distribution of p for 100 000
picked was only one of a very large number of possible samples of 30 indi- repeated sam pies of size 30.
viduals. Suppose we were able to look at 100 000 samples. For each sam-
pIe, we interview 30 individuals and calculate the sam
pIe proportion of
current smokers p. The value of p wil vary from sample to sample. If we
plot aIl values of p, we would see a distribution like the one shown in
a To ensure representativeness, the
Figure 6.1.
sample of individuals should be
This distribution is called the sampling distribution of p. It shows that randomly selected from the popula-
although most sample estimates are likely to be concentrated around the tion of interest. That is, every individ-
true (population) proportion TI, sorne wil be a long way from this true ual in the population should have an
equal chance of being included in
value. The amount of spread tells us how precise our sample proportion p the sam pie. The different ways in
is likely to be, as an estimate of the true proportion TI. If the distribution which a random sample can be
drawn fram a specifie population are
is wide, there is a lot of sampling error and our p may be a long way from dealt with in Chapter 10.
103
Chapter 6
the true value TI. If it is narrow, there is little sampling error, and our pis
likely to be very close to TI.
We have already seen in Section 3.3.1 that the spread of a distribu-
tion can be measured by a quantity caUed the standard deviation. It
can be shown that the standard deviation of a sampling distribution of
p is given by
SE(P)= ~ 7l(1;7l)
where n represents the size of the sample. SE stands for standard error,
which is the term we generally use for the standard deviation of a sam-
pling distribution. The standard error is a measure of the precision
with which our sample value p estimates the true value TI. Notice that
pIe size, n, we decrease the standard error and the
if we increase the sam
SE(P)=~ p(1-p)
Il
SE(P)=~ p2(~_p)
(usually written p :1 1.96 x SE(P)) wil include the true proportion TI with
probabilty 95%. These limits, calculated from the sample data, are called
Iowa and upper confidence limits, respectively, and the interval between
them is called a 95% confidence interval of the unknown population pro-
portion TI.
ln our example for estimating the proportion of men currently smok-
ing,
A 95% confidence interval for the true proportion of men who current-
ly smoke in the whole town is therefore given by
105
Chapter 6
~ p2(1_p)
P l 2.58 x
a
p l 1.64 x ~ p2(~_p)
of 165 cm. Again, the mean height of the male adult population in the
whole town is unlikely to be exactly 165 cm. However, if a very large
number of samples of 30 individuals were drawn from this population
and for each one the mean height were calculated and plotted, a sam-
pling distribution of the mean would be obtained. This sampling distri-
bution wil have a shape similar to that of the sampling distribution of
a proportion (Figure 6.1), Le., the distribution would be bell-shaped,
with most sarnple estimates centred around the true population mean.
106
'\
".-i...
Evaluating the raie of chance
SE (mean) = SDI-I n
107
Chapter 6
Table 6.1.
Age-adjusted mortaliy rate ratios and
rate differences of cancer (ail sites
combined) by duration of employment:
hypothetical data.
The corresponding rate ratio was 9.3 with a 95% confidence interval rang-
ing from 6.0 to 14.6. Thus, our best guess is that men employed for 10 or
more years were nine times more likely to die from cancer than those
employed for less than one year, but the true rate ratio might lie somewhere
between 6.0 and 14.6 (and is unlikely to lie outside this range).
You might have noticed in this example that, whereas the confi-
140
dence limits of a rate difference are equidistant from the sample esti-
mate, this is not the case for the confidence limits of a rate ratio. This
Ü 120
~ can be seen clearly in Figures 6.3 and 6.4(a). ln contrast to the rate
:£ 100 difference, the sampling distribution of a rate ratio is not symmetric,
'C
i: since the minimum possible value it can take is zero, whereas the
~ 80
Ql maximum is infinity. To ob tain a more symmetric distribution, a log-
g 60
..
Ql arithmic transformation of the data was used. As a consequence of
Ql
:! 40 this transformation, the confidence limits are equidistant from the
'C
.S 20
CI
a:
o
~ ,
sample estimate of the rate ratio on the logarithmic scale (Figure
6.4(b)) but asymmetric when converted back to the original scale
1.0-1.9 2.0-4.9 5.0-9.9 10+ (Figure 6.4(a)) (see Appendix 6.1, at the end of this chapteri which
Duration of employment (years) provides formulae to calculate confidence intervals for difference
and ratio measures).
Figure 6.3.
Graphical display of rate differences
(indicated by the middle horizontal 6.1.3 Display of confidence intervals
lines) and their 95% confidence inter-
vals (verticallines) on an arithmetic If we have two or more groups, we can display the sample estimates and
scale (data from Table 6.1). their 95% confidence intervals in a graph. For instance, the rate ratios and
rate differences from Table 6.1 and their respective confidence intervals
are displayed in Figures 6.3 and 6.4.
The middle horizontallines show the observed rate differences and rate
ratios, while the verticallines indicate the 95% confidence intervals. Note
108
Evaluating the raie of chance
how the confidence interval is much narrower when the number (a)
of cases is large (e.g., category 1-1.9 years (based on 67 cases)). It
16
is the number of cases in the numerator of rates and proportions
which determines the size of the standard error and, therefore, the 8' 14
~ 12
width of the confidence interval. 10
al 10
'C
Æ- 8
6.1.4 Further comments ~ 6
CI
l
al
bias in the selection of the study subjects or in the measurement 'C
Table 6.2.
Number of incident cases of lip cancer
by year of registration, females,
England and Wales, 1979-89.8
109
Chapter 6
Even though the whole population of the country was studied to pro-
duce the data in Example 6.2, there was stil random variabilty in the
number of lip cancer cases from year to year, which cannot be explained
in terms of changes in underlying risk. ln these situations, we are sampling
'in time', that is the population in any particular year can be viewed as a
'sample'. The methods discussed above should stil be used to assess the
degree of precision of the observed rates. Of course, not aIl of the variation
from year to year may be random and there may, for example, be an
underlying trend, upwards or downwards, over a particular time period
(see Section 4.3.2).
When the rates are based on smaIl numbers of cases, so that the confi-
dence intervals are very wide, it may be useful to reduce the random vari-
ation by pooling data over five or even ten adjacent calendar years and cal-
culating the average annual incidence (or mortality) rates.
The data in Example 6.3 show that 63.0% (68 out of 108) patients
administered the new treatment were stil alive compared with 40.9% (45
out of 110) of those given the standard treatment. From these results the
new treatment appears superior but how strong is the evidence that this is
the case?
ln general, when comparing the effects of two levels of an exposure or
of two treatments, two groups are sampled, the 'exposed' and the 'unex-
posed', and their respective summary statistics are calculated. We might
wish to compare the two samples and ask: 'Could they both come from
the same population?' That is, does the fact that sorne subjects were
110
Evaluating the raie of chance
exposed and others were not influence the outcome we are observing? If
there is no strong evidence from the data that the exposure influences the
outcome, we might assume that both samples came from the same popu-
lation with respect to the particular out come under consideration.
Statistical significance testing is a method for quantification of the
chance of obtaining the observed results if there is no true diference
between the groups being examined in the whole population.
If the null hypothesis were true, what are the chances of getting a dif-
ference at least as large as that observed?
For example, in the cervical cancer trial, what is the probabilty of get-
ting a treatment difference at least as large as the observed 63.0% - 40.9%
= 22.1%? This probabilty, commonly denoted by P (capital P rather than
the small p we used earlier for the sample proportion), is determined by
applying an appropriate statistical test.
111
\
....'i.
Chapter 6
A simple example
To understand the basis for a statistical significance test, let us look first
at an example that uses numbers small enough to allow easy calculations.
Suppose that an investigator has a theory that predicts there wil be an
excess of male births among babies resulting from in-vitro fertilzation (IVF)
and he therefore wants to study the question 'Are there more boys than
girls among IVF babies?!
The investigator formulates a null hypothesis that there is an equal pro-
portion (0.5 or 50%) of males and females in the population of IVF babies.
Next, he samples five records from one IVF clinic and finds that they are aIl
males in this sample of births. We can now calculate the probabilty of
obtaining five males and no females if the null hypothesis of equal num-
bers of males and females were true.
new and the standard treatments were really equally effective, the
113
Chapter 6
test statistic exceeds in absolute value8 0.674 with probability 0.5 (50%)
1.282 0.2 (20%)
1.645 0.1 (10%)
1.960 0.05 (5%)
2.576 0.01 (1%)
3.291 0.001 (0.1 %)
8 Absolute value means that the plus or minus signs should be ignored; for example, -1 and +1
have the same absolute value of 1 .
If p ;: 0.1, the results are reasonably consistent with the nul! hypothesis.
j
,,
i
1
Evaluating the raie of chance
Table 6.4.
Distribution of benign ovarian tumour
cases and contraIs according to past
history of infertility: hypothetical data.
ln Example 6.4, the null hypothesis assumes that the two groups of
women have a similar risk of getting benign ovarian tumours, Le., the
true odds ratio in the population is equal to one. After using an appro-
priate statistical test, as described in Appendix 6.1, the researchers
obtained a P-value of 0.0003, Le., if the null hypothesis were true, the
probability of getting an odds ratio as large as, or larger than, 5.08
would be very small (less than 1 in 1000). Thus, the data from this
study provide strong evidence against the null hypothesis.
115
¡
f
!
í
Chapter 6
Table 6.5.
Cancer incidence among first-degree
relatives of mycosis fungoides patients
(unpublished data).
117
Chapter 6
Further reading
.' " '" .:-'.,. ::"_'~" ':"'~"'" :.~'. :~~~, ~.1,~:~eYJs~û.~s~~'~.~'-,~..:... _. .~:~ _; _ -; _ '-"_ ;,.' :.,. .'
* Gardner & Altman (1986) pro- ':",. - "" -,'..:"_...',,, ,--,',,'.,
'. - ""-""-,.-""""''','
., '.,'.':, ,",- :',:;,,':-- .". .... . .
.. ,'""'0" ,,',
. """ -.,~ -:. -_,'.' _,,,.'
,', ,', -''''- -"", - .' " ;"',
' ',',_"'_0_ _',
vide a simple overview of most of ',",'",
'-'. ",,0,
',-.,.,-',...".
0"', .... ''u,_:'...'.-.:'-
'-',.,'_"0'
.",," :-:':"-. ".o,'.-:',
',_'_ '" ';_, ''''.,_,'.' _'"."
.-","'.-.','."""""_ . ,_._. _ _"':""'" .'_
the concepts covered in this ..·
Epidemi()logical. studies are usually conducted' in subsets or samples of individ~
chapter and also give sugges-
. .' uals c1rawrifromthe population of interest. Asample estimate of a particularepi-
,....demiOlogicaIml;asure is,however, unlikelyto beequal to the true population
tions for presentation and graphi-
. value;duetó si;mpling .error.' .., ,
cal display of statistical results.
. '. Confidence intervals'and statistical significance testing deal only with sampling
, variation. It is assumed that non-sampling errors such as bias in the selectíon of
....:the subjeGtsin thesample and in the measurement of the variables ofinterest
.. . åreabsènt.. .' .
118
~~~~~-""=-~"'-.'"~~..
Appendix 6.1
Confidence intervals and
significance tests for
epidemiological measures
This appendix provides formulae for the calculation of confidence
intervals and statistical significance tests for the most commonly used epi-
demiological measures. The formulae presented here can only be applied
to 'crude' measures (with the exception of the standardized mortality (or
incidence) ratio). For measures that are adjusted for the effect of potential
confounding factors, see Chapter 14. For measures not considered here,
see Armitage and Berry (1994). Similar results may be easily obtained using
computer packages such as EPI INFO, STATA or EGRET.
SE(P)= ~ p2(~_p)
where a is the number of cases and p = a/n (n being the sample size).
A 95% confidence interval can be obtained as
119
Appendix 6.1
Single rate
If the number of cases that occur during the observation period is
denoted by a and the quantity of person-time at risk by y, the estimated
incidence rate (r) is
r = a/y
SE(r) = r/~a
The 95% confidence interval for the observed rate (r) can then be
obtained as
r :t 1.96 x SE(r)
120
\.
'~"l":
~
Confidence intervals and significance tests for epidemiological measures
These formulae are appropriate when the number of cases in the numer-
ator of the rate, a, is greater than 30. If the number of cases is small, 'exact'
confidence intervals, based on the Poisson distribution, can be obtained
from Table A6. 1.1. This table gives factors by which the observed rate is
multiplied to obtain the lower and the upper limit of a 95% confidence
interval:
SE(r) = 2.24 per 100 000 pyrs/~39 = 0.36 per 100 000 pyrs
95% CI(r) = 2.24 :! 1.96 x 0.36 = 1.3 to 2.95 per 100000 pyrs
For the 'exact' method, the lower limit factor (L) and the upper limit
factor (U) corresponding to 39 cases are obtained from the table by inter-
polation between the rows for 35 and 40 cases.
r. 39 - 35)
L = 0.697 + L(0.714 - 0.697) x 40 _ 35 = 0.711
U = 1.9 - fl
L(1.9 39
- 1.6) -
x 4035)
_ 35 = 1.7
Lower limit = 2.24 per 100000 pyrs x 0.711 = 1.9 per 100000 pyrs
Upper limit = 2.24 per 100000 pyrs x 1.7 = 3.07 per 100000 pyrs
ln this example, the 'exact' and the 'approximate' confidence limits are
relatively close to each other, because the rate was based on a suffciently
large number of cases. The larger the number of cases, the closer wil be
the confidence limits obtained by these two methods.
Let us now consider sorne data from Kuwait. The total number of deaths
from stomach cancer among men aged 45-54 years in this country in
121
l
i
¡
i
L
Appendix 6.1
1980 was only 3 in 74000 pyrs (WHO, 1983). The 'approximate' method
gives
SE(r) = 4.05 per 100 000 pyrs/-v3 = 2.34 per 100000 pyrs
95% CI(r) = 4.05 :t 1.96 x 2.34 = -0.54 to 8.64 per 100 000 pyrs
This method gives a negative value for the lower limit, which is mean-
ingless, as incidence and mortality rates cannot be negative. By the 'exact'
method, consulting again Table A6. 1. 1, the limits for the 95% confidence
interval are:
Lower limit = 4.05 per 100000 pyrs x 0.206 = 0.83 per 100000 pyrs
Upper limit = 4.05 per 100000 pyrs x 2.92 = 11.83 per 100000 pyrs
122
Confidence intervals and significance tests for epidemiological measures
.. .. p.'"~~......,~~~~~,,,
ln R = ln (Pdpo)
ri1 1 1
SE (ln R) = ~ ;; t - - - --
a b mi mo
ln R = ln 2.0 = 0.69
SE(lnR)= -t - - -- - =0.19
11 11000
60 45 1 1500
1
95% CI (ln R) = 0.69:t 1.96 x 0.19 = 0.318 to 1.062
The 'approximate' 95% confidence interval of the risk ratio (R) can then
be obtained by taking antilogarithms:
123
Appendix 6.1
a b
SE (Pl - po) = ~ p/ (1- Pl) + Po2 (1- Po)
where a and b are the numbers of cases in the two study groups.
ln the example shown in Table A6. 1 .2,
l a ~ 60 45
SE (P _ P ) = .1 0.062 (1 - 0.06) + 0.032 (1 - 0.03) = 0.0087
Rate ratio
Consider the results from another hypothetical cohort study, shown in
Table A6.1.3.
Table A6.1.3. Exposure Total
Results from a cohort study in which
rates were calculated as measures of Yes No
occurrence of the outcome of interest Cases 60 (a) 45 (b) 105 (n)
in each study group: hypothetical data. Person-years at risk (pyrs) 4150 (Yi) 6500 (yo) 10650 (y)
Rate per 1000 pyrs 14.5 ((,) 6.9 (ro) 9.9 (1)
As with a risk ratio, a rate ratio can only take values from zero to
infinity. Thus to construct a confidence interval for an estimated rate
ratio (RR), its naturallogarithm needs to be calculated first:
ln RR = ln (ri!ro)
where a and b are the numbers of cases in the exposed and unexposed
groups, respectively.
124
Confidence intervals and significance tests for epidemiological measures
ln RR =: ln 2.1 =: 0.74
We can then ob tain the 'approximate' 95% confidence interval for the
rate ratio by taking the antilogarithms of these values:
125
Appendix 6.1
Rate difference
The standard error of the difference between two estimated rates (ri and
ro) is given by
SE (ri - ro) = -l + _0
mz
a rZ
b
where a and b refer to numbers of cases in the two groups.
The 95% confidence interval is given by
ri - ro = 14.5 per 1000 pyrs - 6.9 per 1000 pyrs = 7.6 per 1000 pyrs
95% CI (ri - ro) = 7.6:t 1.96 x 2.14 = 3.41 to 11.79 per 1000 pyrs
Odds ratio
Data from a case-control study can be presented in a 2 x 2 table, as
shown below:
Table A6.1.4.
Results fram a case-control study:
Exposure Total
hypothetical data. Yes No
Cases 457 (a) 26 (b) 483 (nt)
Contrais 362 (e) 85 (ci 447 (no)
Total 819 (mi) 111 (mo) 930 (N)
ln OR = ln 4.13 = 1.42
126
a= T ¡¡Hi' ~~_ ~..
Confidence intervals and significance tests for epidemiological measures
Thus an 'approximate' 95% confidence interval for the odds ratio can
be obtained by taking antiogarithms:
O=a
E = m¡ni/N
n¡nOm¡mO
v=
NI (N - 1)
A special statistical test called the chi-squared (X2) test can then be
applied to measure the extent to which the observed data differ from
those expected if the two proportions were equal, that is, if the null
hypothesis were true.
x2 = (0 - E)2/V
0= 60
E = 1000 x 105/2500 = 42
v= =24.15
105 x 2395 x 1000 x 1500
25002 x (2500 - 1)
(60 - 42)2
x2 = = 13.42
24.15
127
Appendix 6.1
Large values of t suggest that the data are inconsistent with the null
hypothesis, and therefore that there is an association between exposure
and outcome. The P-value is obtained by comparing the calculated value
of X2 with tables of the chi-squared distribution.
ln referring the calculated t test statistics to these tables, we need to
know a quantity called the 'degrees of freedom' (dJ.), which takes into
consideration the number of sub-groups or 'ceIls' in the table which con-
tributed to the calculation. For 2 x 2 tables, the number of degrees of free-
dom (dJ.) is one.
If the null hypothesis is true,
x' test statistic (with 1 d.f.) exceeds 0.45 with probabilty 0.5
1.2 0.25
2.71 0.1
3.84 0.05
5.02 0.025
6.63 0.01
7.88 0.005
10.83 0.001
Thus, if the X2 statistic with 1 d.f. exceeds 3.84, then P-(O.OS, indi-
cating sorne evidence of a real difference in the proportions. If it
exceeds 6.63, then P -( 0.01, and there is strong evidence of a differ-
ence.
ln the example shown in Table A6.1.2, the value of X2 was 13.42,
which corresponds to P -( 0.001. There is therefore strong evidence for
an association. Thus we can conclude that the observed risk ratio of 2.0
is statistically significantly different from 1, and that there is very
strong evidence that the risk is higher in those who were exposed than
in those who were not.
ln fact, we have already performed a test for difference in two pro-
portions in Section 6.2.2. We then used a different test statistic which
gave similar results to the more general Mantel-Haenszel type of test
statistic used here.
Note that the statistical test is the same regardless of the measure of
effect (risk (or prevalence) ratio or risk (or prevalence) difference) that we
are interested in. However, the confidence intervals are calculated in a
different way (see Section A6.1.2).
128
j
f
,,
i
1
Confidence intervals and significance tests for epidemiological measures
o = 457
819 x 483
E= = 425.35
930
(457 - 425.35)2
x2 = = 41.0
24.43
The X2 gives a measure of the extent to which the observed data difer
from those expected if the two odds of exposure were equal. This X2 value
(with one degree of freedom) corresponds to P-(O.OOl. Thus, there is strong
evidence against the null hypothesis.
Then
x2 = (0 - E)2/V
0=60
105 x 4150
E= = 40.92
10 650
(60 - 40.92)2
x2 = = 14.58
24.97
129
Appendix 6.1
v = 15
(20 - 1W
x2 = 15= 1.67
This X2 value, with one degree of freedom, corresponds to 0.1 .( P.( 0.25.
Thus, these results are consistent with the nuIl hypothesis of no difference
in the age-specific mortality rates from leukaemia between the town and the
whole country.
Note that the statistical test is the same regardless of the measure of effect
(rate ratio or rate difference) we are interested in. However, the confidence
intervals are calculated in a different way (see Section A6.1.2).
130
,
"
f
i
¡
L
,
Confidence intervals and significance tests for epidemiological measures
The first step is to assign a score x to each exposure category. For exam-
pIe, in Table A6.1.S '0' was assigned to those women with 1 partner, '1 to
those with 2-3 partners, and so on. The second step is to use the following
formulae to ob tain the values of Ti, Tz and T3. ln these formulae, the sym-
bol i means sum and the subscript i stands for the subscripts 0, 1, 2, 3, etc.
T3 = Im¡x? = (90 x 02) + (101 x 12) + (93 x 22) + (168 x 32) = 1985
The X2 test for trend has one degree of freedom and can be calculated as
N (NT¡ - n¡ Tz)Z
x2=
n¡nO (N T3 - TzZ)
131
.~"\..
Appendix 6.1
A X2 test for a linear trend can be used to test the hypothesis that there
is a decreasing risk of ovarian benign tumours with increasing parity. The
calculations are exactly the same as those used for the t test for a linear
trend in proportions.
This test result with 1 dJ. corresponds to 0.01 .( P.( 0.025. Thus there is
evidence that the risk of developing a benign ovarian tumour decreases
with increasing parity.
T3 = IYiX/ = (31 200 x 02) + (25 100 x F) + (11 600 x 22) = 71 500
The t test for a linear trend in rates, which has one degree of freedom,
can be calculated as follows:
132
Confidence intervals and significance tests for epidemiological measures
(Ti - (n/y)Tz))2
X2 =
(n/yZ) (yT3 - T2Z)
This test value with 1 d.f. corresponds to P? 0.5. Hence, the results of
the study provide no support for an upward or downward trend in breast
cancer rates with duration of oral contraceptive use.
Va/idity of x2 tests
If in a 2x2 table the total sample size (N) is less than 40 and the expect-
ed value in any of the cells is less than 5, the X2 test should not be used.
ln these circumstances, the Fisher's exact test wil be the appropriate sta-
tistical test (see Kirkwood (1988)). For larger tables, the X2 test is valid if no
more than 20% of the expected values are less than S, and none is less
than one.
Note that the expected value (E) for a particular cell is calculated as fol-
lows:
The X2 test is valid in this example since the total sample size (N) is
greater than 40 and aIl of the expected cell values are weIl above S.
133
Chapter 7
Intervention trials
Intervention trials are the epidemiological studies that most closely resem-
ble the experiments conducted by scientists in the laboratory. The essential
and distinguishing feature of such studies lies in the investigator's direct con-
trol over the aIlocation of subjects to study groups. ln contrast, in observa-
tional studies, the aIlocation is determined by the subjects themselves and the
researchers are just passive observers of what happens.
Intervention trials provide the strongest evidence with which to test
hypotheses. However, they are not the most usual study design in epidemiol-
ogy, mainly because of ethical constraints. It would be unacceptable to allo-
cate people to either be or not be exposed to a substance or to be subjected to
a procedure for which there is sorne suspicion that it may be harmfuL It is,
however, possible to conduct a trial to test whether removal of such an expo-
sure wil decrease subsequent incidence and mortality. Thus, intervention tri-
als in epidemiology are limited to interventions for which there are grounds
to believe that there wil be a potential benefit to individuals.
ln contrast, field trals deal with subjects who are disease-free. A field tral
involves evaluation of whether an agent or procedure reduces the risk of
developing disease among those free from that condition at enrolment.
Because these trials involve healthy rather than diseased people, they tend to
be logistically more difficult to carry out than clinical trials. They generally
135
L
i
f
Chapter 7
Example 7.2. A' randDmized trial' wascarried Dut amDng Whitehall . -.- , .
(English) civil servants tD measure in middle-aged men the health effects Df
stDpping smDking. AtDtalDf 1445 male cigarette smDkers aged40~59 years
who, wereat a high risk Df develDping cardiDrespiratDry diseases were rall~
dDmly allDcated tD interventiDn (714
men) Dr nDrmal care (731 men).. Thdse
in the' interventiDn grDUp' received individual advice, Dn the relatiDnDf smDk-
. :ing tD health. lvDSt then' expressed theirwish tD StDp smDking and received
, furthersl1ppDrt over the neit 12 mDnths~ The twD grDupS were then fDllDwed
. up fDr twenty years (RDse& CDlwell, 1992). . . . , .
Field trials can be carried out among individuals (as in Example 7.2) or grDupS
Df peDple (as in Examples 7.3 and 7.4). ln the first case, the unit of allocation
to the intervention is the individual, whereas in the second, it is the group.
The group may be a household, a block of housesi a school or a whole com-
munity. Field trials in which whole communities are the unit of allocation are
called cDmmunity trals (Example 7.3).
,~.' ,', " '.'-,'" ,', "- '-":": :,' , c, .'.:'.' :..:_, "f. ','. ::' ~,., .'. ; ,.-,:. _, : ; _ -.:' ':', ,;' ,', _ ,',,:" :: ,.'...,. ",.'C,','," - '",' :::'"
Example 7.3. The CDrnmunity Interventiòn Trial. fDr Smoking CessatiDn
, ,(COMMIT) was a rnulticentreproject designed tD evaluate. a cDmmunity~
.' wiqe smDking cessatiDnprogramme in the USA. This
trial began in 1989 ill
.,.. .1t niatche.d.pairsDfcDn;munitiès. OnecDmmunity Df eàch pair was fan;.
,... tdònilyasSignedtorec,eive .the.srnDking çes~ation prDgramme with the
Dthèr .
; .aCtirígàsa contrDI. The' '. rventinn wasde~iined tDpinmote srnDking ces~. .
..såtiQh bY1fslfJS awiClè '~.o. ,~Dfcommunity resDUtcestDpffeçt atttude.s' an.d .'
:P '. ; 'es t 'sirÚjktng-replvtT: Rèsea~ch GrDl1Pl1991);' ....'
,/ '.',' \\ ': :',::;r:"~,~ '~..:- ;.: -:;:;"..;~(::':_i,--:~:' .'.~' '::'_'.~:~::::_~:" :_-::- .:~' '\~:'~'~,::' ':,\/ ;,::0; ,0':',,-,,', - ': co,: ',', .~ -- ",-, ':',-.' - '.-' . -.'. 'i' - '- c, ',' "" ,,,," " " -, ' "'.
',--'\:',~:'" '-,-
136
--- --.-= - ---~~_. --
Intervention trials
There are various reasons for selecting groups rather than individuals as
the study unit. Many interventions are impossible to assign at an individual
leveL Environmental interventions such as water fluoridation or improve-
ment of air quality can be conducted only at a group level. Most health edu-
cation interventions also fall into this category. For instance, the interven-
tion (Le., the smoking cessation programme) in Example 7.3 was aimed pri-
marily at the community rather than the individual; thus, it was appropri-
ate to choose the community as the study unit. It may also be logistically
easier to conduct the trial among groups of people than among individuals.
ln Example 7.4, for instance, it was much easier to allocate flghts to either
the intervention group or the control group than it would have been to aIlo-
ca te individuals. By allocating flghts, it was also possible to minimize the
potential for 'contamination', that is, the possibilty that people in the con-
trol group would end up having access to the leaflets. Such contamination
would have made the two groups more alike and, consequently, wou Id have
decreased the abilty of the trial to reveal any true effect that the interven-
tion might have had (see Section 7.10).
137
Chapter 7
that the procedures used to measure the exposures and the outcomes of inter-
est do not involve unacceptable levels of discomfort, stress or risk for the par-
ticipants.
ln intervention trials, however, the situation is different. Researchers are
no longer simply observing what happens to the study subjects. Since the
investigator is deliberately intervening, ethical considerations are more
important than in any other type of epidemiological study. Intervention tri-
als are ethically justified only in a situation of uncertainty, when there is gen-
uine doubt concerning the value of a new intervention in terms of its bene-
fits and risks. The researcher must have sorne evidence that it may be of ben-
efit, for instance, from laboratory and animal studies, or from observational
epidemiological studies. Otherwise, there would be no justification for con-
ducting a triaL.
Unfortunately, many medical interventions have never been properly
evaluated in weIl conducted intervention trials. For instance, radical mastec-
tomy was used for more than a hundred years as the standard form of treat-
ment for early breast cancer. It was not until the late 1970s, when clinical tri-
als were finally conducted, that this form of treatment was replaced by more
conservative tyes of breast surgery. The clinical trials revealed that there
were no differences in recurrence or survival between patients who under-
went radical mastectomy and other (more conservative) types of surgery
(Veronesi et al., 1981; Fisher et al., 1985). Thus, women with early breast can-
cer were unnecessarily subjected to a very mutilating form of surgery for
decades because clinicians were convinced that it would have been unethical
to deprive women of the standard form of therapy. The lesson from this, and
many other examples, is that it is best to conduct a trial when any agent or
procedure is first introduced rather than after it has gained widespread accep-
tance and becomes considered standard practice. Failure to carry out a prop-
er trial, when it is needed and feasible, may also be unethicaL
Whether a study is considered to be ethical or unethical is a subjective
judgement based on cultural norms, which vary from society to society and
over time. A useful reference with proposed guidelines for research involving
human subjects is that published by the Council for International
Organizations of Medical Sciences (CI OMS) and the World Health
Organization (WHO) (1993).
138
Intervention trials
/~
Chapter 15). Third, it is important to
choose an experimental population that 1
Random allocation
wil experience a sufficient number of the
outcomes of interest to permit meaningful
Intervention Control
comparisons between various treatments or group group
procedures within a reasonable period of
time. Thus, most trials are carried out in
Follow-up
populations where the risk of developing
the outcome(s) of interest is high. For
instance, to assess the potential benefit of a
smoking cessation programme, it would
make sense to select as the experimental
population one with a high prevalence of Outcome Outcome Outcome Outcome Outcome Outcome
tobacco use and high incidence of lung known unknown known unknown known unknown
cancer.
The selection of the experimental popu-
lation also depends on logistic factors. The study should be carried out in an Figure 7.1.
area where it wil be possible to obtain support from the local authorities or General outlne of an intervention triaL.
leaders of the community and where it wil be possible to obtain complete
and accurate foIlow-up information for the duration of the triaL. For
instance, conducting a long-term trial among a highly mobile population
such as college students or nomads may result in low foIlow-up, which
would compromise the study.
139
Chapter 7
140
Intervention trials
141
Chapter 7
142
¡
l
f
!
,
f
l
Intervention trials
is, therefore, the only way of ensuring that any differences in the outcome
measures of the trial are due to the effects of the intervention rather than
to underlying differences between the groups. Randomization has two
major advantages in relation to other methods of aIlocation:
(2) Randomization tends to create groups that are comparable in terms of the
distribution of known andi more importantlyi unknown factors that may
influence the outcome.
143
Chapter 7
Simple randomization
Simple randomization is the most elementary method of randomization.
It is the equivalent of tossing a coin. However, randomization by tossing a
coin should not be used because it cannot be checked or reproduced. The
67 19 00 71 74 60 47 21 29 68
alternative is to use a table of random nurnbers (or a computer-
02 02 37 03 31
02 94 37 34 02 76 70 90 30 86 38 45 94 30 38 generated randomization list) (Figure 7.2).
79 78 45 04 91 16 92 53 56 16 02 75 50 95 98
87 75 66 81 41 40 01 74 91 62 48 51 84 08 32 The first step in determining random group assignments is to
34 86 82 53 91 00 52 43 48 85 27 55 26 89 62 set up a correspondence between the numbers in the table and
11 05 65 09 68 76 ~ 20 37 90 57 16 00 11 66 the study groups. Let us assume that odd numbers correspond to
52 27 41 14 86 22 98 12 22 08 07 52 74 95 80
07 60 62 93 55 59 33 82 43 90 49 37 38 44 59
the control group and even numbers to the new intervention.
04 02 33 31 08 39 54 16 49 36 47 95 93 13 30 The second step is to define a convenient way of reading the
01 90 10 75 06 40 78 78 89 62 02 67 74 17 33
table of random numbers, for instance, to read down the
92 03 51 59 77 59 56 78 06 83 52 91 05 70 74
61 71 62 99 15 06 51 29 16 93 58 05 77 09 51
columns or across the rows.
73 32 08 11 12 44 95 92 63 16 29 56 24 29 48 The third step is to select a starting point, for instance, by clos-
42 10 50 67 42 32 17 55 85 74 94 44 67 16 94
26 78 63 06 55 13 08 27 01 50 15 29 39 39 43ing your eyes and selecting a number with a pin. Once the start-
ing point is established, numbers are then read from the table
Figure 7.2. following the sequence defined in step two. Figure 7.2 is an extract of a table
Extract fram a table of random of random numbers (a full table is reproduced in Appendix 7.1). Suppose
numbers.
that the chosen starting point was the one circled in the table and that we
have decided that nurnbers should be read column by column down the
page. The first 10 numbers would have been 8, 9, 3, 5, 7, 5, 5, 9, 1, O. The
fourth step is to make the treatment assignments according to the system
defined above (Table 7.1).
Random number tables are generated in such a way that each of the dig-
its zero through ni ne is equally likely to occur. If equal numbers of partici-
pants are required in each intervention group, the same nurnber of one-digit
numbers should be assigned for each group, even if this me ans that sorne
digits do not correspond to any group. Thus, in the case of three groups,
144
Intervention trials
three of the ten one-digit numbers are assigned to each group (e.g., numbers
1, 2, 3 to group A; 4, 5, 6 to group B; and 7, 8, 9 to group C). The remaining
number (Le., zero in our example) in the random tables is ignored and selec-
tion moves to the next number.
One of the disadvantages of simple randomization is that it may result in
markedly unequal number of subjects being allocated to each group just by
chance. For instance, in the above example, only two persons out of ten
were assigned to the intervention group. Moreover, simple randomization
may also result in the compositions of the different intervention groups
being different with respect to factors that may affect the outcome measures
in the triaL. ln the above example, not only was the number of persons allo-
cated to the intervention smaIl but the sex distribution was also quite dif-
ferent in the two groups. This is particularly likely to happen when the total
number of subjects in a study is small. For trials involving several hundred
participants or more, any such imbalance is likely to be smaIl and can be
taken into account in the analysis of the study. ln a small trial, imbalance
may make the trial more difficult to interpret and, hence, it is advisable to
ensure balance by using the randomization methods described below.
No. Combination
1 AABB
2 ABAB
3 AB BA
4 BBAA
5 BABA
6 BAAB
145
\.......
Chapter 7
Stratified randomization
When the results of the trial are likely to vary between, say, the sexes or
different age-groups, stratified randomization should be used. ln this situ-
ation, strata or groups are formed and randomization occurs separately for
the subjects in each stratum. As subjects become eligible for inclusion in
the trial, their appropriate stratum is determined and they receive the next
random-number assignment within that stratum. For example, patients
may be classified according to sex and age (under 50, and 50 and over),
yielding a total of four strata. Within each stratum, each patient wil be
randomly assigned ta either the intervention or the control group. This
could be do ne by using either simple or restricted randomization.
Stratified randomization has the advantage of assuring balance between
the groups, at least on the characteristics that determine the strata. The
use of this method of randomization in the example described above
would ensure that the intervention and the control group would be bal-
anced with respect to sex and age. If stratification had not been employed,
the researcher wauld have run the risk that chance might produce imbal-
ance with regard to these important factors, especially if the number of
subjects in the trial was small. The disadvantage with stratified random-
ization is that it is administratively difficult and cumbersome to execute.
Matched-pair design
A matched-pair design is a special case of stratified randomization in
which the strata are each of size 2. Individuals (or communities) are
matched into pairs, chosen to be as similar as possible for potential con-
founding variables such that in the absence of any intervention they
would be expected to be at similar risk of the disease under study. The
intervention is assigned at random to one member of each pair, with the
other member acting as a control.
Matching is unnecessary in large trials, as it is likely that any imbalance
between the intervention groups, with respect to risk factors for the occur-
rence of the outcomes of interest, wil tend to even out. Furthermore, it is
possible to adjust for any residual imbalance during the data analysis with-
out substantialloss of statistical power.
For small trials, more serious imbalance can arise for which it may be
difficult to adjust fully in the analysis. This can be a special problem in tri-
als in which communities are randomized, as it is unusual to be able to
146
Intervention trials
Figure 7.3.
Outline of the facto rial design of the
Physicians' Health Study.
147
Chapter 7
Crossover trials
Most trials have a paraUd design, that is, a group of subjects receives the
intervention and another parallel group receives the standard treatment or
placebo. ln contrast, in crossover trials each subject acts as his/her own con-
trol by receiving at least two different interventions (e.g., a new drug
(treatment A) versus the standard drug (treatment B)) at different times
during the trial (Figure 7.4). The order in which each individual receives
them (A then B or B then. A) should be determined by random allocation
(Example 7.9)'. There should be a 'wash-out' period between each
'wash-out of the interventions to avoid 'carry-over effects' (also called
period
'spil-over effects'), that is, to ensure that there is no overlap of
A B effects between the first and the second interventions.
i --------1
Consequently, this design is suitable only when neither of the
B A interventions has long-term effects.
1 -------- i The main advantage of crossover trials is that each subject is
compared with himself/herself and, therefore, confounding is
. eliminated from the comparison of the effects of the two treat-
Time
ments (provided that there is no carry-over effect). This design
Figure 7.4. also increases statistical precision in the comparison, because it eliminates
Outline of a crossover trial inter-subject variabilty in the outcome response. Hence, fewer subjects are
needed th an in a corresponding parallel triaL.
Crossover trials are used mostly in the early phases of evaluation of new
drugs in which their pharmacokinetic properties are investigated in
healthy volunteers. They are not appropriate to assess the long-term
effects of a treatment, as the nature of the design implies that the treat-
ment period must be limited.
148
\.
..-i..
1 ntervention trials
7.9.4. Conclusions
It should be emphasized that allocation of subjects to the study groups
should be done only after having ascertained that individuals are both eli-
gible and wiling to participate. Otherwise subjects who refuse to partici-
pate or who withdraw from the study (because the treatment is inappro-
priate, etc.) wil have to be excluded after the randomization, so that the
groups may no longer be comparable. This is ilustrated in Example 7. la.
1
. ~~,', ~lU, W ,W i in ~
,
T:
_ ~ ii:ri:
-- .. ,
fi~/)'i '- ~'. ',,' ,.
i .
,
'1.al Y' y _. v¡p.
.,,'J.,
" "~ . Wi:1l
.. '¡; V¡;i!.
, ¡; .'..
. _;-il
l- c" U1 ~ :1,
.
Table 7.2.
Mortaliy fram causes other than
breast cancer by the end of the first
five years of follow-up. a
Table 7.2 shows that there was a large difference in mortality from caus-
es other than breast cancer between women who actually received the
intervention (i.e., who were screened) and those in the control group (42
versus 58 per 10 000 pyrs). Since the intervention under study (i.e., breast
screening) should not have affected mortality from causes other than
breast cancer, the observed difference seem to indicate that the two groups
were different in relation to important baseline characteristics. However, if
those who refused after randomization were included in the intervention
group, as they should be, there is no longer a mortality difference between
the two groups (57 versus 58 per la 000 pyrs).
Random allocation does not necessarily guarantee that the groups wil
be similar. Discrepancies between the groups may arise just by chance,
especially if the number of units being allocated (e.g., individuals, fami-
lies, communities) is relatively smalL Hence, it is essential to collect base-
line data on the subjects. These baseline data should include aIl the vari-
ables which are known or thought to affect the outcome(s) of interest and
can be used to check the degree of similarity of the groups. If the study
149
1
\...
f
/, -,"1"'..
j
!
Chapter 7
groups differ, statistical techniques can be used that yield results 'adjusted'
for any baseline differences (see Chapters 13 and 14).
150
Intervention trials
151
Chapter 7
152
Intervention trials
153
¡
, i
,,
,1
f
J'
¡
Chapter 7
7.12 Analysis
154
Intervention trials
155
Chapter 7
156
L
i
f
Intervention trials
beta-carotene
1 % were 10stto
the study was to
lung . cancer
(Test statistics and confidence intervals were calculated using the formulae given in
Appendix 6.1).
157
Chapter 7
(Test statistics and confidence intervals were calculated using the formulae given in
Appendix 6.1).
158
Intervention trials
Prevented fraction (%) = 100 X (rate (or risk) difference/rate (or risk) in the
unexposed)
If the aim of the trial is to assess the value of a vaccine, this measure is
caIled vaccine effcacy. For instance, if the risk of developing a particular
disease among those who were vaccinated was 40 per 100 000 and 70 per
100000 among those not vaccinated, the vaccine efficacy would be
Vaccine efficacy (%) = 100 X ((70 per 100000 - 40 per 100000)/70 per 100000)
=43%
Thus, 43% of cases could have been prevented among the unvaccinat-
ed if they had been vaccinated.
If baseline differences between the study groups need to be taken into
account in the analysis, one of the statistical techniques discussed in
Chapter 14 should be used.
If we are particularly interested in the distribution of time unti occur-
rence orthe event or interest (e.g., time from treatment to death or time from
treatment to recurrence), as is the case in many clinical trials, the most
appropriate approach is survival analysis. The techniques used in survival
analysis derive from the lie-table methods which are discussed in Chapter
12.
Subgroup analysis
It is usual for investigators to perform subgroup analyses to assess
whether the intervention has an effect on subgroups of individuals with
certain characteristics (e.g., males, elderly people, patients with particular
clinical features, etc.). These subgroup analyses raise important problems,
however. If the subgroups are defmed according to the baseline characteristics
of the patients, the main concern involves loss of abilty of the trial to
detect an effect (that is, loss of statistical power or precision (see Chapter
159
Chapter 7
15)), since the results wil be based on only a small proportion of the total
numbers of randomized subjects. On the other hand, if multiple analyses
are performed, sorne wil inevitably achieve 'statistical significance' just by
chance. Their interpretation wil de pend very much on the existence of a
priori hypotheses based on biological plausibilty, existence of supporting
evidence from laboratory experiments and from other epidemiological
observations.
Analyses performed on subgroups defìned on the basis of individual char-
acteristics which develop after randomization are of much greater concern,
because potential confounding variables wil no longer be distributed at
random among the subgroups. For instance, analyses restricted to persons
who reached a certain serum concentration of the active treatment or who
developed a weIl known secondary effect (e.g., skin yellowing after ingest-
ing beta-carotene) should be treated with extreme caution. Their findings
should never be reported as main results of the trial but just as interesting
observations that might be worth investigating in specifically designed tri-
als.
7.13 Interpretation
The interpretation of results from a weIl conducted intervention study
should be relatively straightforward, since the two major problems of con-
cern in observational studies, bias and confounding, are greatly reduced
by using an experimental design.
This is not to say that trials are exempt from problems. The lung cancer
beta-carotene story provides a good ilustration of this. Data from three
large-scale chemoprevention trials conducted in western countries to
assess this question have now been published. The ATBC Cancer
Prevention Study (Example 7.11) was set up in Finland to test the hypoth-
esis that a high intake of beta-carotene and alpha-tocopherol reduces the
risk of lung cancer. This was a reasonable hypothesis given the substantial
evidence available from observational epidemiological studies suggesting
that beta-carotene was associated with a lower risk of lung cancer. The
results of this trial failed to show any benefit of beta-carotene (or alpha-
tocopherol) in the prevention of this malignancy; instead, men who took
beta-carotene had an unexpected 'statistically significant increase in the
risk of lung cancer (as we saw in Example 7.17). The authors did consider
alternative explanations for this unexpected finding. Confounding cou Id
be discarded given the large sample size and the random allocation of sub-
jects to the various study groups. Moreover, the treatment groups were
weIl balanced in relation to relevant baseline characteristics. Since this
result was not supported by biological or previous epidemiological evi-
den ce, the au th ors were reluctant to reject the null hypothesis of no effect.
Results from two other trials were subsequently published. The active-
intervention phase of the CARET trial (Example 7.14) was terminated early
because its results confirmed the unexpected increase of lung cancer risk
among those who took beta-carotene reported by the ATBC triaL. There
was again no obvious explanation for this unexpected finding. The
Physicians' Health Study (Example 7.8) had a much longer foIlow-up (12
years) than the other two trials (average of 6 and 4 years for the ATBC and
CARET trials, respectively). Its results were consistent with the null
hypothesis of no effect of beta-carotene on the risk of lung cancer; in other
words, they did not provide evidence of either a beneficial or a harmful
effect of beta-carotene. Thus, the lung cancer beta-carotene story shows
that results from a single trial should not be considered in isolation.
The results of a trial cannot be translated directIy into public health
decisions. Gther factors that need to be taken into account include issues
such as generalizabilty of the results to different populations, acceptabil-
ity of the intervention, feasibilty, costs, available resources and compet-
ing public health priorities. Furthermore, the overall impact of an inter-
vention in a particular population depends not only on the magnitude of
the effect of the intervention on the risk of developing a particular condi-
tion, but also on the frequency (and severity) of the condition in the pop-
ulation. This issue is further discussed in Chapter 16.
161
Chapter 7
Further reading
, ". ....,... .o,. -'."., .":,;.8'0,, 7.~..~~~y'.ÎS~u~~(:.. ..:.'... . .......~. . .~:. p:."
',';.'_'.":'.. "0_",-':,,,' ,':.i"':': ;,'-,._,',,,--'_ -.'_ -,': __ ,",,:,:, :;-,.,,_'-:-,. -:"-,:-,:,,.'.:,-,_,- :;,-_ :.",--,,":. ,- -""':",:,,,' c',.'
* The book by Smith & Morrow
· .Intervention triàls arè charact~rized by the fact that investigatorsareresponsible
(1996) provides a comprehen-
for allocating subjects to the different study groups. '.' .
sive coverage of the design,
implementation and monitoring
of field trials, with particular
emphasis on practical aspects.
',:" _,',;'-"_:","'_,"__','. .,'_ :'-":,,,__, "',, .',' ,0- _',_._",,-.,':
,:,-:.,,- ",' ':",:'--, "'-';,'_:'::-' '_:::
-,",'-: :-,:,:',:,:--'--'-.'
"': '-",' :,"--"-":':_' ':,'.-,'--'-
.- _co, _:' '_..,__-:,-" "_ -, :,'-:"- ',' ',- :_ --,: ' d d _' ,-.,,__:;'_::-."-
,",,':- ,','
,'-- '---,
Although the focus is on develop- 1. Random allocatíon ofsubjects ensures that allocation of subjects to
ing countries, most of the issues the different study groupsis unaffected by selection bias.
discussed in this book are also
relevant to developed countries. 2. Random allocatíon ensures that the groups are weil balanced in rela-
tion to knownand, more importantly,unknown factors that may affect
the outcome(s) of the study (provided thestudyissuffciently large). .
* The book by Pocock (1983)
provides a good and accessible
reference for those interested in
the design, analysis and interpre-
tation of clinical trials.
162
¡
\~
J '-
!
1
Intervention trials
Appendix 7.1
Table of random numbers
16 22 77 94 39 49 54 43 54 82 17 37 93 23 78 87 35 20 96 43 84 26 34 91 64 Table A7.1.
84 42 17 53 31 57 24 55 06 88 77 04 74 47 67 21 76 33 50 25 83 92 12 06 76 Table of random numbers (fram Table
63 01 63 78 59 16 95 55 67 19 98 10 50 71 75 12 86 73 58 07 44 39 52 38 79
XXXiII of Fisher and Yates (1963)).
33 21 12 34 29 78 64 56 07 82 52 42 07 44 38 15 51 00 13 42 99 66 02 79 54
57 60 86 32 44 09 47 27 96 54 49 17 46 09 62 90 52 84 77 27 08 02 73 43 28
18 18 07 92 46 44 17 16 58 09 79 83 86 19 62 06 76 50 03 10 55 23 64 05 05
26 62 38 97 75 84 16 07 44 99 83 11 46 32 24 20 14 85 88 45 10 93 72 88 71
23 42 40 64 74 82 97 77 77 81 07 45 32 14 08 32 98 94 07 72 93 85 79 10 75
52 36 28 19 95 50 92 26 11 97 00 56 76 31 38 80 22 02 53 53 86 60 42 04 53
37 85 94 35 12 83 39 50 08 30 42 34 07 96 88 54 42 06 87 98 35 85 29 48 39
70 29 17 12 13 40 33 20 38 26 13 89 51 03 74 17 76 37 13 04 07 74 21 19 30
56 62 18 37 35 96 83 50 87 75 97 12 25 93 47 70 33 24 03 54 97 77 46 44 80
99 49 57 22 77 88 42 95 45 72 16 64 36 16 00 04 43 18 66 79 94 77 24 21 90
16 08 15 04 72 33 27 14 34 09 45 59 34 68 49 12 72 07 34 45 99 27 72 95 14
31 16 93 32 43 50 27 89 87 19 20 15 37 00 49 52 85 66 60 44 38 68 88 11 80
68 34 30 13 70 55 74 30 77 40 44 22 78 84 26 04 33 46 09 52 68 07 97 06 57
74 57 25 65 76 59 29 97 68 60 71 91 38 67 54 13 58 18 24 76 15 54 55 95 52
27 42 37 86 53 48 55 90 65 72 96 57 69 36 10 96 46 92 42 45 97 60 49 04 91
00 39 68 29 61 66 37 32 20 30 77 84 57 03 29 10 45 65 04 26 11 04 96 67 24
29 94 98 94 24 68 49 69 10 82 53 75 91 93 30 34 25 20 57 27 40 48 73 51 92
16 90 82 66 59 83 62 64 11 12 67 19 00 71 74 60 47 21 29 68 02 02 37 03 31
11 27 94 75 06 06 09 19 74 66 02 94 37 34 02 76 70 90 30 86 38 45 94 30 38
35 24 10 16 20 33 32 51 26 38 79 78 45 04 91 16 92 53 56 16 02 75 50 95 98
38 23 16 86 38 42 38 97 01 50 87 75 66 81 41 40 01 74 91 62 48 51 84 08 32
31 96 25 91 47 96 44 33 49 13 34 86 82 53 91 00 52 43 48 85 27 55 26 89 62
56 67 40 67 14 64 05 71 95 86 11 05 65 09 68 76 83 20 37 90 57 16 00 11 66
14 90 84 45 11 75 73 88 05 90 52 27 41 14 86 22 98 12 22 08 07 52 74 95 80
68 05 51 18 00 33 96 02 75 19 07 60 62 93 55 59 33 82 43 90 49 37 38 44 59
20 46 78 73 90 97 51 40 14 02 04 02 33 31 08 39 54 16 49 36 47 95 93 13 30
64 19 58 97 79 15 06 15 93 20 01 90 10 75 06 40 78 78 89 62 02 67 74 17 33
05 26 93 70 60 22 35 85 15 13 92 03 51 59 77 59 56 78 06 83 52 91 05 70 74
07 97 10 88 23 09 98 42 99 64 61 71 62 99 15 06 51 29 16 93 58 05 77 09 51
68 71 86 85 85 54 87 66 47 54 73 32 08 11 12 44 95 92 63 16 29 56 24 29 48
26 99 61 65 53 58 37 78 80 70 42 10 50 67 42 32 17 55 85 74 94 44 67 16 94
14 65 52 68 75 87 59 36 22 41 26 78 63 06 55 13 08 27 01 50 15 29 39 39 43
163
\"'-i.....
Chapter 8
Cohort studies
One of the great advantages of this type of cohort study is that it allows
a large number of common exposures to be considered in relation to a
large number of outcomes. The Framingham Study is a classical example
of this. Approximately 5000 residents of the town of Framingham, in
Massachusetts (USA), have been followed up since 1948 (Dawber et aL.,
1951). There were several reasons for selecting this location for the study,
mainly determined by logistic and other practical considerations to ensure
that it would be feasible to identify and follow participants for many
years. At the time the study was set up, Framingham was a relatively sta-
ble community including both indus trial and rural areas, with a number
of occupations and industries represented. The town was small enough to
allow residents to come to one central examining facility and there was
only one major hospitaL FoIlow-up of this cohort has permitted assess-
ment of the effects of a wide variety of factors (e.g., blood pressure, serum
166
Cohort studies
Similarly, when Richard Doll and Bradford Hil set up a cohort study in
England and Wales to assess the health effects of smoking, the choice of
the British physicians as the study population (Example 8.3) was deter-
mined mainly by logistic considerations. Physicians were registered with
the British Medical Association and were therefore easy to identify and fol-
low up. Besides, they were more likely to cooperate and the cause of death
to be properly investigated.
If the exposure is rare, a study of the general population wil have little
abilty to detect an effect (Le., the study would have insufficient statisti-
cal power (see Chapter 15)), since very few people would have been
exposed to the factor of interest. This problem can be overcome by delib-
erately selecting a highly exposed group of people as the study population.
For example, exposure to dyestuffs is rare in the general population.
However, by choosing a group of workers with high exposure, the full
range of effects of the exposure on their health can be studied, including
outcomes that are rare in the general population but not in those heavi-
ly exposed. The general public health impact of the exposure may be
smaIl, but such studies can give insight into common biological mecha-
nisms in disease.
167
1
,t
f
r
Chapter 8
168
-~--.
=~."'~--- -,~, ~~"'~~ Cohort studies
intake and the risk of breast cancer. For this purpose, nurses were asked to
complete a dietary questionnaire and the distribution of fat intake in the
whole cohort was examined and divided into five groups of equal size
('quinties'); women in the lowest quintile of fat intake were taken as the
'unexposed group' (Wilett et aL., 1987).
ln occupational cohorts, an internaI comparison group might consist of
workers within the same facilty with other types of job not involving
exposure to the factor under investigation.
169
Chapter 8
avoids the need to follow up a large number of unexposed individu aIs, but
ne only for outcomes for
it has several disadvantages. First, it can be do
170
L
,i
l
l
Cohort studies
b Only data for these two age-groups were available for ail the four populations.
sure, dates at which exposure started and stopped, PAST PRESENT FUTURE
dose and pattern of exposure (intermittent versus TIME
constant), and changes over time (see Section 2.3).
(b)
Information on the exposure(s) of interest may be Exposed ~ ;-r Outcome
Outcome
171
Chapter 8
There are two main types of cohort study, defined according to the
point in time when information on exposure was collected: present or
past. ln prospective cohort studies, data on exposure is collected now, once
the study population has been defined. ln this instance, it is possible to
use the most up-to-date methods of exposure measurement so that bias in
exposure classification can be minimized. The main disadvantage of this
type of cohort study, however, is that the time from exposure to onset of
disease (i.e., the induction period) may be too long (many decades for
most cancers). Examples 8.1 to 8.5 are examples of prospective cohort
studies which involved the follow-up of large numbers of people for very
long periods of time.
The alternative, particularly useful for conditions with long induction
periods, is to rely on exposure measurements made many years before the
study was set up, which may be available from medical, employment or
other personal records. By use of data from existing records, the time we
have to wait for the exposure to have any effect on the risk of disease may
be considerably reduced or even eliminated. This type of cohort study is
called a historical cohort study.
Example 8.8. ln the early 1950s1 Case and his co-workers set up a cohort
study to assess whether menengaged in the manufacture of certain dyestuff
intermediates had an excess risk of bladder cancer. . They began by con-
. structinga list ofàll men ..who had ever beenemployed in the chemiázl
industr in the United Kingdom for al least six months since 1920.' The age '
and the datesbetween which exposure to dyestuffoccurred were recorded. A
search was made retrospectively for all bladder cancer cases occurring among
men who had beenemployed in the chemical industri in or after 1921 unti.
1 February 1952. The number of observed bladder cancer cases among these
workers was then compared with the number that would have been expect-
ed if these workers had the samemortality experience as the general popu-
lation of the United Kingdom (Case etal.i 1954; Case & PearsonI1954).
The study described in Example 8.8 is a classic example of the use of this
historical approach. Examples 8.6 and 8.7 are also ilustrations of histori-
cal cohort studies, since both relied on preexisting employment records to
identify the cohort members and to classify them according to their expo-
sure status. Historical cohort studies are particularly useful in occupation-
al epidemiology because, if there is concern that a particular exposure may
be a hazard, it is not reasonable to wait decades for clarification in a
prospective cohort study. However, if at the time the historical cohort is
identified, a large proportion of members are stil alive, the follow-up peri-
od can be extended into the future (as in Example 8.6) to ensure that aIl
possible long-term health effects are properly assessed.
One of the main limitations of historical cohort studies is that the expo-
sure data available in past records are generally less accurate and detailed
172
Cohort studies
="~~""
173
Chapter 8
Table 8.2.
Smoking habits of male participants in
the British doctors study who replied to
the 1951 and 1990-91
questionnaires. a
disease within the cohort has to be set up. For instance, the ascertainment
of the outcomes of interest may be do ne through self-administered ques-
tionnaires sent regularly to aIl study subjects, through personal interviews,
or by regular physical examination of aIl members of the cohort.
Regardless of the method chosen to as certain the outcome(s) of interest,
it is vital to ensure that it is used identically for subjects exposed and those
not exposed. If possible, interviewers and any other persons involved in the
ascertainment of the outcomes should be kept blind to the exposure status
of the study subjects. Otherwise, there is potential to introduce measure-
ment bias (see Section 13.1.2).
Cohort studies focus on disease development. ln order for a disease to
develop, it must, of course, be absent at the time of entry into the study. An
initial examination of the potential study population may be required to
identify and exclude existing cases of disease. Even so, it may stil be impos-
175
Chapter 8
8.6 Follow-up
The criteria for entry into the cohort must be defined before the start of the
study in a clear and unambiguous way. Individuals should enter the
cohort, and contribute person-time at risk, only after aIl the entry crite-
ria have been satisfied. ln most cohort studies, participants wil join the
cohort at different points in time and therefore the exact date of entry of
each subject should be recorded.
Methods must be set up at the start of the study to ensure adequate fol-
low-up of the study subjects. ln general, these involve periodic contacts
with the individuals su Ils or mailed ques-
ch as home visits, telephone ca
8.7 Analysis
The first step in the analysis of a cohort study is to measure the incidence of
disease (or of any other outcome of interest) in those exposed and in those
unexposed and compare them.
176
¡ \.-.
L '~....
f
í
Cohorl studies
Table 8.3.
(a)
Exposure Analysis of a cohorl study (a) by risks;
(b) by rates.
Yes No
Outcome Yes a b
No c d
Risk in exposed group (P1) = a/(a+c)
Risk in unexposed group (Po) = bl(b+d)
Risk ratio = P11Po
Risk difference8 = P1 - Po
(b)
Exposure
Yes ~o
Cases a b
Person-lime at risk Y1 Yo
Rate in exposed group (r1) = a/Y1
Rate in unexposed group (ro) = blyo
Rate ratio = r11ro
Rate difference8 = r1 - ro
8 If the exposure is protective, the risk and rate differences should be calculated as Po - P1 or
ro - r1, respectively (see Section 5.2.2)
If aIl, or virtally aIl, cohort members were foIlowed up for the same peri-
od of time, we can calculate risk as the measure of disease occurrence in each
group (Table 8.3(a)). For example, if the period is uniformly five years, the five-
year risk can be computed separately for the exposed and unexposed groups.
Risk ratio and risk difference can then be calculated as measures of relative and
absolute effect, respectively.
If study subjects have unequal foIlow-up periods, this must be taken into
account in the analysis. FoIlow-up durations may differ markedly if subjects
were recruited into the study population over a relatively long period of time,
or if sorne were lost to foIlow-up during the course of the study (for example,
by moving out of the area). One way of handIing variable foIlow-up periods
is to calculate rates which use person-years at risk (or person-months or per-
son-days, etc.) as the denominator (Table 8.3
(b)). With this approach, each
subject contributes to the population at risk only as many years of observation
as he/she is actually observed; thus if the subject leaves after one year, he/she
contributes 1 person-year; if after 10, 10 person-years (see Section 4.2.2).
People may contribute person-years of observation to more than one sub-
group. Suppose, for example, that in a five-year study, disease incidence is
determined for each lO-year age subgroup. A person entering the cohort when
he or she reaches the age of 48 years wil contribute 2 person-years of obser-
vation to the 40-9 year-old subgroup and 3 person-years of observation to
the 50-59 year-old subgroup (see Section 4.3.2). This may also happen with
exposure categories if the study subjects change their exposure status over
time. For instance, a person may be a smoker for a few years and then give up.
177
\o~~..
Chapter 8
Rate ratio = 217 per 100 000 pyrs/187 per 100 000 pyrs = 1.16
95% confidence for the rate ratio = 0.96 to 1.40
Rate difference = 217 per 100 000 pyrs - 187 per 100 000 pyrs = 30 per 100 000 pyrs
95% confidence interval for the rate difference = - 8 to 68 per 100 000 pyrs
X2 = 2.48; 1 d.f.; P=0.12.
(Test statistics and confidence intervals were calculated using the formulae given in
Appendix 6.1).
179
Chapter 8
~~-=~...~""-===~.t
(fable 8.5). . .
Table 8.5. Timing Cases Person- Rate (per Rate ratio
Incidence of breast cancer among Duration of use years 100000 pyrs) (95% confidence interval)
nurses aged 45-49 years at the time
their entry into the cohort by timing
Non-usersb 240
Current users
128 528 187 1.00
and duration of oral contraceptive
use.a S; 48 months 4 328 1220 6.52 (2.43-17.53)
:: 48 months 4 2263 177 0.95 (0.35-2.55)
X2 test for trend = 0.46; P = 0.50
Past usersc
S; 48 months 106 54 080 196 1.05 (0.84-1.32)
:: 48 months 86 36 039 239 1.28 (1.00-1.64)
X2 test for trend = 3.33; P = 0.07
. a Data fram Romieu et al. (1989).
b Taken as the baseline category.
C Information on duration is missing for four past users.
(The 95% confidence intervals and the X2 test for a linear trend in rates were calculated
, using the formulae given in Appendix 6.1.)
short-term users, but this estimate was based on only four breast cancer
cases (hence, the wide confidence interval).
To assess whether there was a linear (positive or negative) trend in rates
with increasing duration of use, a special statistical test (X2 test for a linear
trend) was performed separately for current and past users (using the for-
mula given in Section A6.1.3). There was moderate evidence of a positive
trend among past users, but no evidence of a linear trend among current
users (Table 8.5).
It should be noted that the shape of an exposure-response relationship
does not have to be linear. This is clearly ilustrated in Example 8.14. For
instance, the relationship between alcohol consumption and aH-cause
mortality (Figure 8.2(a)) is U-shaped, with men who reported drinking
8-14 units of alcohol a week having the lowest mortality. By contrast, the
relationship between alcohol consumption and mortality from alcohol-
related disorders (Figure 8.2(b)) is basicaHy linear, with a progressive
increase in mortality with increasing alcohol consumption among regular
drinkers. Thus, lack of a linear trend (as assessed by the X2 test for trend)
does not imply absence of a relationship. The form of an exposure-out-
come relationship should primarily be identifed by plotting the data as in
Figure 8.2. If the shape is suggestive of a non-linear trend, special statisti-
cal procedures, which are outside the scope of this book, should be used
to assess its statistical significance.
Since the allocation of the study subjects to the different exposure cat-
egories is not random in cohort studies, the exposure groups are likely to
differ in many respects other than the exposure of interest. These differ-
ences rnust be taken into account in the analysis. For instance, age is an
180
Cohort studies
Figure 8.2.
40 (a) Ali causes (b) Alcohol-augmented causes Male mortality from various causes by
344t weekly alcohol consumption: (a) ail
1486 6
causes; (b) conditions known to be
,585 + +413 alcohol-related (e.g., cancers of the
, 383
liver, mouth, pharynx, oesophagus and
20
'675 442 4 ~¡ larynx, cirrhosis, alcoholism, and exter-
nal causes); (e) ischaemic heart dis-
2 t 19 t" ease; (di other known causes (cere-
brovascular diseases, respiratory dis-
+32 t26 eases, ail other cancers not included in
,24'35
0 0 (b), and others). Points and bars are
rates and 95% confidence intervals
15 f~.'''.m¡' "".rt d¡.oo~ 30 (d) Other causes adjusted for age, smoking habits and
history of previous disease; the values
are numbers of deaths (reproduced, by
c;
::
i:
i:
.:
10 +206 l
,235 f 127 87t
20 t 290
2171 permission of BMJ Publishing Group,
from DolI et al., 1994b).
,344 , +237+241
l.
125 111 '387 272
5 10
o
o 21 42 63 o 21 42 63
Weekly alcohol consumption
o
a 1 unit = half pint (237 mL) of ordinary beer or one glass (125 mL) of wine or one glass
(62 mL) of sherry or a glass (26 mL) of spirits.
181
L
\-'ï.."..
ì
!
Chapter 8
182
\"..
\"\..
"
Cohort studies
(Test statistics and confidence intervals were calculated using the formulae given in
Appendix 6.1.)
the cases (Le., those individuals in the cohort who contract the disease
under study) and of a subgroup of individuals who remained disease-free
(the controls), are analysed at the end of the foIlow-up.
,..--.:--"., "-;-',., , ,',' .--,
Exampie 8.16.Inl'
, 'west&ile bistr,.trio
"Epstetll':Båtrvirus" .
obtained (rom. each chUd. gt '
the fOllQw-up in1979,16né
a,mongthecohortme' S~.'
, pletakena(entrfro,
elsin the serâof fouf,'
in the neighbourhood'
who'd.id'nofdevel()p thèdiseàs.e.
ln Example 8.16, blood samples were obtained and stored from aIl
42 000 children who participated in the study but the rather complex and
expensive virus tests were carried out only on the serum samples from the
16 children who developed the lymphoma and from asample of about 80
selected disease-free members who acted ascontrols.
Similarly, in nutritional cohort studies, food dia
ries may be used to me a-
sure the subjects' usualdietary intake. As the coding and analysis of food
diaries is very labour-intensive, a nested case-control study may be con-
ducted in which only the diaries of the cases and of a sample of disease-
free members of the cohort (' controls') are examined.
183
Chapter 8
This type of study design and the analysis of its data are discussed fur-
ther in Sections 9.3 and 9.5.
8.9 Interpretation
The main advantage of cohort studies is that the time sequence between
exposure to a factor and occurrence of the disease is clearly observed.
Cohort studies may, however, suffer from important bias. Knowledge
of the presence or absence of exposure to a particular factor may affect
the subsequent assessment of disease and introduce measurement bias.
This can occur if the decision as to the presence or absence of disease is
made by persons who are aware of the subjects status with regard to the
study factor. Cohort studies are also potentially prone to selection bias
due to loss of study subjects. Such losses may occur initially if a portion
of the target study population does not participate, or later on as me m-
bers of the study population are lost to follow-up. These losses do not
necessarily invalidate the study. However, the investigators should con-
sider whether the reasons for loss might have affected the study results.
Sometimes it is possible to gather information concerning lost subjects,
particularly about whether they left because of ilness or death possibly
related to the exposures and outcomes under investigation.
As with any other observational study, confounding is a critical issue in
the interpretation of cohort studies. Special statistical techniques can be
used to take into account the effect of potential confounding variables,
but only if these variables were known at the time of the data collection
and if they were properly measured. If these data were not collecte di we
have to judge how much the observed findings are likely to have been
affected by confounding in the light of aIl available biological and epi-
demiological evidence.
ln Example 8.17, uranium miners had significantly elevated mortality
for cancers at three sites relative to the general population (Table 8.7).
Before concluding that these raised risks are due to exposures in the
mines, we need to consider alternative explanations for the observed
findings. A possible explanation for the high risk of lung cancer among
miners is that they were heavier smokers than the general population.
No information on smoking habits was available for the miners, but a
survey of 697 men in other Czech uranium mines in 1974 showed that
76% were smokers, slightly more than the average (66%) for
Czechoslovakian males at that time (see Tomášek et aL. (1994)). The
results from this survey indicate that differences in smoking habits are
unlikely to have fully accounted for the estimated five-fold higher lung
cancer risk in the miners th an in the general population. Moreover, Table
8.8 (Example 8.18) shows that there was a signifcant positive trend of
mortality from lung cancer with increasing cumulative exposure to
radon. This trend provides considerable support for a true association
between exposure to radon and lung cancer. Similar findings have been
found in other radon-exposed miners.
184
Cohort studies
~~'==.4'U"'~~~~~=-"".~~"'..;:;m.w:"~_,,~~.\r:",~\'~~=."~~"""'='""'''""~===;'-''-='''.=-.?;=-""=~=¡¡",:.;;;i:;,,i:i.''''W¡;":.,,,~;,.-.Ç:;¡--.~~(.\,;"'=';.:¡'¡:';';*;;"".;:~'",~~.i.t"'.i~W~~5"¡Qr.:e'-
o 7 3 4 5 3 0.57
OIEb 2.70 1.18 1.53 2.04 1.40
o a 2 1 3 6 0.003
gallbladder OIE b 0.00 1.92 0.93 3.03 6.73
and extrahepatic
bile ducts
a Data fram Tomášek et
al. (1993)
b Expected number of deaths (E) calculated using national male age-specifie
mortality rates for the former Czechoslovakia.
185
¡
J
,
ì
¡
f
¡
Chapter 8
There was also a positive trend in mortality from gaIlbladder and extra-
hepatic bile duct cancer with increasing levels of cumulative exposure to
radon (Table 8.8), but there is little supporting evidence in favour of this
finding from other epidemiological studies, and further investigation is
needed. By contras t, mortality froID liver cancer did not increase with
cumulative radon exposure, making it unlikely that the excess in the min-
ers was caused by radon. No information was available on the alcohol con-
sumption of the miners, but they were weIl paid compared with other
Czech workers and, therefore, it is likely that their alcohol consumption
was higher than in the general population. They also had a significant
excess of deaths from liver cirrhosis, probably caused by alcohol con-
sumption and, for six of the liver cancer deaths, cirrhosis was also men-
tioned on the death certifcate (Tomášek et aL., 1993).
The issues that need to be addressed in interpreting results froID cohort
studies are further discussed in Chapter 13.
186
Cohorl studies
Further reading
. . '. .,:' - .,- .' .' B.ox, 8.1.. Key issues' ..... . , _' .... ~
.. Cohort studies are studies in which subjects are selected on the basis of their * The book by Breslow & Day
exposure status and then follow~d up in time. ln contrast with intervention stud- (1987) provides a very compre-
ies, however, the allocation of exposure is not determined by the investigators. hensive coverage of the role,
design, analysis and interpreta-
· Themaiil advantages of thistypeofstudy are: tion of cohort studies in cancer
epidemiology. Some of the mate-
1. Exposure is measurèd betore disease onset and is thereforelikely to rial is presented at a relatively
be unbiased in terms ofdisease development. advanced levaI.
2. Rare exposures'.can' he examined ..by appropriate selection of study * Discussion of the healthy work-
cohorts. er effect can be found in papers
by Carpenter (1987) and
3,Multiple outcomes(diseases) c:anbestudied forany one exposure. McMichael (1976).
187
Chapter 9
Case-control studies
189
Chapter 9
":'.'
possible interactions between them). Example 9.2 clearly ilustrates this
feature.
_.... ," ,- ,,' ", ..... '-" .-.--'.-".,--:'.:'.'..,.-'-----.::.,-"" - ,.,. . -, . . ,
Example.9.2. A population-based case-control studywas carried out in'
Spain and. Colombia. to assess the relationship betweencervical cancer and
exposure to human papilomavirus (HP V), selected aspects of sexual and.
reproductive behaviour, use of oral contraceptives,screening praètices, smok-
ing' and possible interactions between them. The study íncltided.436 inci-
dent cases of histologically confirmed invasivesquamous~cell carcínomaof
the cervix and.387 controls of similarage randomlyselèdedJrom thegener~
al populationthat generated the. cases (Muñozèt aL/1992a). . " '
The results from these exploratory case-control studies may suggest spe-
cific hypotheses which can th en be tested in specifically designed studies.
190
¡ \''-
t
.. ~'.i...
f
l
f
1
Case-control studies
the disease or condition of interest, even if, as a result, sorne true cases
have to be eliminated. For instance, a histologically confirmed diagnosis
should be required for most cancers. By accepting less weIl documented
cases, the investigator runs the risk of diluting the case group with sorne
non-cases and lessening the chances of finding real exposure differences
between cases and controls.
It is sometimes impossible to eliminate aIl cases whose diagnosis is not
properly documented, particularly if the pool of available cases is rela-
tively small. ln these circumstances, it may be possible to classify the
cases according to diagnostic certainty. Such classification allows assess-
ment of the extent to which the results are likely to be affected by disease
rnisclassification (see Chapter 13). Suppose, for instance, that cases in a
particular case-control study are classified as 'definite', 'probable' or 'pos-
sible'. If there is disease misclassification, a graduaI decline in relative risk
froID the 'definite' to the 'possible' category should become apparent in
the analysis, since the probabilty that non-cases may have been mis di-
agnosed as cases increases from the 'definite' to the 'possible' category.
The case definition should be established in such a way that there is no
ambiguity about types of cases and stages of disease to be included in, or
excluded from, the study. The choice of cases should be guided more by
concern for validity than for generalizabilty. For example, in a study of
breast cancer, we may learn more by limiting the cases (and the controls)
to either pre- or post-menopausal women than by including women of aIl
ages (unless the number of cases in each group is large enough to allow
separate analyses), since the risk factors for pre- and post-menopausal
breast cancers may be different. By ensuring that the cases are a relative-
ly hornogeneous group, we maximize the chances of detecting important
etiological relationships. The abilty to generalize results to an entire pop-
ulation is usually less important than establishing an etiological relation-
ship, even if only for a small subgroup of the population.
Cases should also be restricted to those who have sorne reasonable pos-
sibilty of having had their disease induced by the exposure under inves-
tigation.
191
Chapter 9
ln Example 9.4, cases were restricted to women bom since the 1920s
because only women born since then could have been exposed to the fac-
tor of interest (oral contraceptives).
Although most case-control studies include only one case group, it is
possible to study simultaneously two or more cancers whose risk factors
are thought to share the same, or related, risk factors. Example 9.4 ilus-
trates this point. Such multiple-disease case-control studies may be regard-
ed as a series of case-control studies. This approach provides two main
advantages. First, it provides the possibilty of studying more th an one
cancer for relatively little extra cost. Second, the control groups may be
combined to give each case-control comparison increased statistical
power, that is, the abilty of the study to detect a true effect, if one reaIly
exists, is enhanced because of the larger number of controls per case (see
Chapter 15).
If the disease or condition of interest is very rare, the study may have to
be carried out in various participating centres, possibly located in various
countries. The study cited in Example 9.4 was conducted in 10 centres in
eight countries. Despite this, only 122 newly diagnosed liver cancers were
accrued during the seven-year study period. Sorne studies deliberately
include participating centres from low- and high-incidence areas to assess
whether the risk factors are similar. For instance, the cervical cancer study
mentioned in Example 9.2 was conducted in Colombia and Spain, coun-
tries with an eight-fold difference in cervical cancer incidence (Muñoz et
aL., 1992a).
The eligibility criteria should include not only a clear case definition but
also any other inclusion criteria (Example 9.5). Persons who are too il to
cooperate or for whom the study procedures may cause considerable phys-
ical or psychological distress should be excluded. It is also usual to exclu de
elderly people in cancer case-control studies because their diagnosis is like-
ly to be less valid and because of their diffculty in recallng past exposures.
UsuaIly, the inclusion of aIl patients who meet the eligibilty criteria is
not possible for a variety of reasons. Subjects may move out of the area,
die or simply refuse to cooperate. The investigator should report how
many cases met the initial criteria for inclusion, the reasons for any exclu-
sion, and the number omitted for each reason (as in Example 9.6). This
information allows us to assess the extent to which the results from the
study may have been affected by selection bias (see Chapter 13).
192
Case-control studies
193
Çhapter 9
a Data fram Buiatti et al. (1989a,b). (The numbers of cases on which these
percentages are based were not given in these papers).
definition, aIl those with a short duration leave the pool of prevalent cases
because of either recovery or death. Unless we can justify the assumption
that the exposure being studied is not associated with recovery or survival,
every effort should be made to limit recruitment to incident cases. By using
only newly diagnosed (incident) cases and selecting controls to be repre-
sentative of subjects from the population from which the cases arise, the
case-control study aims to identify factors responsible for disease develop-
ment, much like a cohort study. Moreover, prevalent cases may not be rep-
resentative of aIl cases if sorne affected patients are institutionalized else-
where or move to another city where there are special facilties for treat-
ment.
There are other advantages to the use of incident cases. Recall of past
events in personal histories tends to be more accurate in newly diagnosed
cases than in prevalent cases. Besides, incident cases are less likely to have
changed their habits (or 'exposures') as a result of the disease.
If constraints on time or resources make the use of prevalent cases
inevitable, we should choose those that were diagnosed as close as possible
to the time of initiation of the study. A check on the characteristics of the
prevalent cases may be possible by comparing the frequency (or level) of
exposure among subjects with different times of diagnosis. If, among the
cases, the frequency of exposure to a factor suspected of being associated
with the disease changes with time since diagnosis, we should suspect sur-
vival bias. For instance, if those cases who were exposed to the factor under
study have poorer survival than those unexposed, they wil become under-
represented in a prevalent case series as time since diagnosis increases. As a
result, the prevalence of exposure among the surviving cases wil decrease.
194
Case-control studies
Prevalent cases may have to be used for conditions for which it is difficult
to establish a specific date of onset. For instance, case-control studies to
examine risk factors for Helicobacter pylori infection have to be based on
prevalent cases, because it is difficult to establish the date of onset of this
condition.
195
\\-i..,
Chapter 9
women with breast cancer aged 45 years and over, the controls must be
selected from women in the same age group without the disease.
If the disease being studied is uncommon in the group serving as a
source of controls, litte, if any, diagnostic effort or documentation is
needed to mIe out the disease in the selected controls. A simple interview
question wil often suffice. However, if the disease is common, a greater
effort to minimize misclassification, such as a review of the individuals'
medical records, is desirable (as in Example 9.9).
196
"\
...'l..
Case-control studies
study, a study population can be defined from which aIl incident cases
are obtained; con troIs should be randomly selected from the disease-
free members of the same population. Consider, for example, aIl the
newly diagnosed cases of childhood cancer in the catchment area of a
regional cancer registry. ControIs for these cases would apprapriately
be drawn fram the population of the same area in the same sex- and
age-groups. Even when the cases are identified exclusively fram hospi-
taIs, it stil may be reasonable to assume that they represent aIl the
cases in the catchment area if the disease is serious enough that aIl
cases end up in hospital (which is likely to be true for most cancer
cases in countries with good health care).
It is generaIly expensive and time-consuming to draw contraIs from
a random sample of the catchment population. A list of aIl eligible sub-
jects or households must be available for sampling, or has to be creat-
ed (as in Example 9.11). (Methods to select a random sample from the
study population are discussed in Chapter 10.) Besidesi healthy people
may be disinclined to participa tei which may intraduce selection bias
due to non-response.
ControIs may also be selected fram close associates of the case, such
as friends and relatives who are from the same catchment population as
the cases. Although a relatively smaIl effort is required to identify
these contraIs and obtain their cooperation, there is a danger that they
wil be too similar (overmatched) to cases in terms of exposures and
other characteristics (see Section 9.3.4). Neighbourhood controls can also
be used, but people living in the same neighbourhood are likely to be
similar in many respects, so such contraIs may also be overmatched.
Moreover, if the interviewer has to visit each neighbourhood to con-
tact these contralsi the cost of the study may become extremely high.
When using hospital-based cases, it may not be possible to define
the population fram which the cases arase, either because the exact
catchment area of the hospital cannot be defined or because not aIl the
cases in the are a are referred to the hospital, and those referred may be
selected according to particular criteria (e.g., the more serious). ln
these circumstances, hospital-based controls may be used because the
study population can then be defined as potential 'hospital users/.
197
Chapter 9
198
Case-control studies
Table 9.3.
Recruitment levels among contrais
and reasons for non-participation by
recruitment centrea
199
!
f
f
í
Chapter 9
Example 9.13iSuppose that a cohort of 200 000 healthy women was fol-
lowed up, for ten years, to assess the relationship between 'lifestyle variables
and the risk of developing various ,types of cancer and that a total of 60
women were diagnosed withendometrial cancer during the follow-up period.
Suppose alsp thatthe. investigators decided to conduct a case.:ontrol study
nested within this cohort. to assess. whether oral contraceptive use protected
against endometrial cancer. The investigators couldsample the controls in
three differentwaysiasindicatedin Figure 9.1: (i)from those who were stil
disease-free by the end of the follow-up period (situation A); (2) from those
who were still at risk at the time each case was diagnosed (situation B); or
,(3) from those who were at risk at the starf of the study (situation C).
Figure 9. i .
Sampling schemes for contrais when New cases
the source of incident cases is a of disease
c10sed cohort with a fixed follow-up (60)
(x=incident case). Initially
at risk
1
(disease-free)
(C) Currently
at risk (B)
(200 000) Stil at risk
(disease-free)
(A)
j (199 940)
9.3.4. Matching
Individual matching refers to the procedure whereby one or more con-
troIs are selected for each case on the basis of similarity with respect to cer-
tain characteristics other than the exposure under investigation. Since
cases and controls are similar on the matching variables, their difference
with respect to disease status may be attributable to differences in sorne
other factors. It is, however, important that matching îs restricted to con-
200
Case-control studies
founding factors and is not performed for the exposure under investigation. The
characteristics generally chosen for matching are those that are known to
be strong confounders. Common matching variables are age, sex, and eth-
nicity but others might be place of residence, or socioeconomic status.
Let us suppose that we are interested in examining the relationship
between current use of oral contraceptives and ovarian cancer. ln this
example, it is appropriate to match on age, since age is associated with the
exposure of interest (current oral contraceptive use) and is an independent
risk factor for ovarian cancer. ln other words, age is a confounding factor.
Failure to match, or otherwise control, for age would result in a biased
assessment of the effect of oral contraceptive use.
use /
Oral contraceptive .. Ovarian cancer
~Age
When controls are chosen so as to be similar to the cases for a charac-
teristic and when this similarity tends to mask the disease's association
with the exposure of interest, cases and con troIs are said to be overmatched.
This can happen when controls are matched to cases for a characteristic
that is part of the pathway through which the possible cause of interest
leads to disease. Imagine a case-control study conducted in West Africa to
investigate the role of hepatitis B virus in the etiology of liver cancer in
which con troIs were matched to cases on the basis of previous history of
Hver disease.
201
Chapter 9
202
Case-control studies
9.5 Analysis
The analysis of data from case-control studies depends on their
design. Individual-matched studies require a different type of analysis
from ummatched (or frequency-matched) studies.
203
Chapter 9
Total mi ma N
204
,
i
,,
..
,j
f
¡
Case-control studies
low-up; odds of disease takes those who were stil disease-free by the end
of the follow-up; and the rate uses the total person-time at risk, which
takes into account the exact time when the cases occurred. Comparison
of these measures of incidence in those exposed relative to those unex-
posed yields three different measures of relative effect: the risk ratio, the
odds (of disease) ratio and the rate ratio, respectively (see Section 5.2.1).
ln case-control studies, it is not possible to directIy estimate disease
incidence in those exposed and those unexposed, since people are select-
ed on the basis of having or not having the condition of interest, not on
the basis of their exposure statusa. £t is, however, possible to calculate the
odds of exposure in the cases and in the controls:
(Confidence intervals and test statistics for the odds ratio were calculated as shown in
Appendix 6.1.
It can be shown algebraically that the odds (of exposure) ratio a Indirect calculations are possible in
obtained from a case-control study provides an unbiased estimate of population-based case-ontrol studies
one of the three relative measures of effect that can be obtained trom (see Appendix A 16.1)
205
Chapter 9
free by the end of the follow-up, the study wil estimate the odds (of
disease) ratio. If controls are selected from those stil at risk at the
time each case is ascertained, the study wil provided an unbiased
estimate of the rate ratio. ln this last instance, the analysis should
respect the fact that cases and con troIs are matched with respect to
time. (An unmatched analysis wil also yield an unbiased estimate of
the rate ratio if the rates of acquiring disease remain constant over
time among both the exposed and unexposed populations and the
total numbers at risk remain relatively constant in both populations.)
As we saw in Section 5.2.1, when the disease is rare, as with cancer,
cases constitute a negligible fraction of the population. The number
of people at risk in a cohort study remains practically constant over
time and therefore the three measures of effect yield similar results.
Consequently, the three sampling schemes used to select con troIs in
a case-control study wil also provide similar results. If the disease is
common, however, different control sampling schemes wil yield dif-
ferent results and the choice of the most appropriate one wil depend
on the specific problem being addressed (Smith et aL., 1984;
Rodrigues & Kirkwood, 1990).
ln strict terms, the odds ratio obtained from a case-control study
tells us how many more (or less, if the exposure is associated with a
reduced risk) times likely the cases are to have been exposed to the
factor under study compared with the controls.
ln Example 9.16, cervical cancer cases were 76% more likely to
have never attended school than con troIs. Since the odds ratio
obtained from a case-control study provides an estimate of one of the
three relative measures of effect that can be calculated from a cohort
study, we can also interpret it as an indication of the likelihood of
developing the disease in the exposed individuals relative to those
unexposed. ln our example, the odds ratio indicates that women who
never attended school were 76% more likely to develop cervical can-
cer than those who attended.
As in other types of study, inferences about the association
between a disease and a factor are considerably strengthened if there
is evidence of a gradient between the level (or intensity) of exposure
and risk of the disease in question. Odds ratios can be computed sep-
arately for each level of the exposure. The general approach is to treat
the data as a series of 2 x 2 tables, comparing con troIs and cases at
different levels of exposure, and then calculating the odds ratio at
each leveL.
ln Example 9.17, there is a trend of increasing risk of cervical can-
cer with increasing number of sexual partners.
206
Case-contrai studies
Table 9.7.
Number of lifetime sexual partners
among cases and contrais. a
125 265
74 305
207
Chapter 9
--"1l~ l~.c-?"'l":l'.'I!';:; - ......~~
cases and controls of any of these factors (in fact, the adjusted odds ratio
was higher than the crude odds ratio) (see Section 13.2 and Chapter 14 for
further discussion of these issues).
9.5.2/ndividua/-matched studies
Individual-matched studies require a special type of analysis, in which the
2 x 2 table takes a different form. Let us consider the simplest situation where
there is only one control per case. The status of the cases with regard to the
presence or absence of the exposure of interest is cross-tabulated against the
exposure status of their respective controls (rable 9.8).
Contrais Total
Table 9.8. Exposed Unexposed
layout of a 2 x 2 table with data from Cases Exposed s a
an individual-matched case-ontrol
Unexposed u b
study (control-to-case ratio = 1:1).
Total c d N/2
The marginal totals (a, b, c, d) of this table correspond to the entries in the
cells of the table for the unmatched studies. The total for the entire table is
NI2 pairs, where N represents the total number of paired individuals.
The matched odds ratio can be calculated as
This odds ratio calculation considers only the discordant pairs. It can be
explained intuitively: pairs where both case and control were exposed or
where both were unexposed give no information about the relationship of
the exposure to disease (Example 9.18).
The analysis is more complex than shown here if there is more than one
control per case (see Breslow & Day (1980), chapter 5).
208
Case-control studies
Table 9.9.
Frequency of exposure to artificial
Exposed Unexposed
sweeteners among 632 bladder can-
Exposed 468 (1 87 (s) 555 (a) cer cases and their individual-matched
Unexposed 73 (t) 4 (u) 77 (b) controls. a
whether the result could have arisen as a result of selection bias in the choice
of cases and controls, information bias in the gathering of exposure data, or
failure to take proper account of confounding factors.
The most serious potential problem in case-control studies is that the pro-
cedures used to select cases and controls may produce groups that are not
truly comparable.
ln Example 9.19, selection bias could have affected the results of this study
since only 62% (314/510) of aIl eligible patients were included in the final
analysis. Low participation levels can introduce bias if cases who used oral
contraceptives were more or less likely to participate in the study. If, for
instance, users of oral contraceptives were more likely to have a less aggres-
sive form of breast cancer than non-us ers and, hence, a better survival, this
would lead to over-estimation of the effect of oral contraceptives since a high
proportion of the deaths would have occurred among non-users.
Selection of an appropriate control group is one of the most difficult prob-
lems in case-control studies. Controls must come from the same defined
population as the cases. The use of hospital-based controls works only if
patients with different diseases came from the same general population (Le.,
209
\.~~......
Chapter 9
;. -_,,'.,
, _....'" " ,_ ,. - _, __._, '__n _ ,'.
Example9.19.The relation/Jetween use of oral contraceptives by young
women andtheir risk ofbreast cancerwas investigated in a population-'based
case-control. study conducted in Los Angeles County. The cases were patients
with histologically. confirmed breast' cancer, first diagnosed between fuly
1972 and May 1982, diagnostd before age 37 years, and without a prior his-
tory ofmalignancy.A total ot510eligiblecases were identified through the
local population~basedcancerregistr ofwhom 458 were still alive at the
time of the first contactthrough their doctors. Physicians gave permission to
contact 393 (86%) ofthese patients. Ofthese, 26 could not be locatedand
37 refused to be interviewed. Thus, completed questionnaires were obtained
(rom 330 patients. Sixteen of these patients were later excluded because no
suitable individuallymatched control was found (pike et aL., 1983).
if the referral patterns are the same for the disease under investigation and the
control diseases) and if the control diseases are themselves unrelated to the
exposure. ln many situations, it is diffcult to be sure that these conditions are
satisfied. The use of population-based controls avoids these problems, but
selection bias may stil be introduced if the levels of non-response are high
either because sorne eligible con troIs cannot be traced or because they refuse
to participate (as in Example 9.20). ln this instance, the control series may not
be representative of the population from which the cases arise.
Example 9.20. The possible association between oral contraceptive use and
the risk of breast cancer at young ages (under 45 years) was investigated in
a population-based case-control study conducted in Sweden and N01way. ln
Norway, where notification ofall cancer diagnoses is mandatory, cases were
identified (rom population-based cancer registries. A total of 114 eligible
women were identified ofwhom 105(92%) participated. For each case who
agreed to participate,two controls werechosen (rom an up-to-date national
population register. Potential controls were maile,d a request to participate.. If
an imswer was not received within foUr weeks or if the control refused to par ~
ticipate, a nèw controlwasselected.. Nine con troIs were never located; 34
, never answered the letter; 38 refused to participa te; 4 were either temporari-
'..' ly abroad.and couldnotbe reachedorhàd rrentaldisorders.Thuslto ob tain
'two' con
for, e(iCh~aseiit waseventuålly necessmy to select 295 controls.
troIs
210
-'\.
'~"ì..
Case-control studies
ln Example 9.21, the aim of the investigation was explained to the women
involved. This may have increased recall bias, particularly since the study was
carried out during a time of great public concern about oral contraceptives
and breast cancer. This problem could have been minimized to a certain
extent by not disclosing the study hypothesis to the study subjects.
The other potential source of bias in a case-control study is diagnostic bias.
For instance, if women using oral contraceptives are more likely than non-
users to examine their breasts, or to have them examined by a physician or
nurse, or to undergo mammography, diagnostic bias may be introduced.
Thus, if a positive association between oral contraceptives and breast cancer is
found in a study, it may just be due to the fact that oral contraceptive users
are more investigated and therefore more likely to be diagnosed with breast
cancer than non-users. One way of minimizing diagnostic bias is to obtain
information on the frequency of breast examinations for each of the study
subjects so that any effects of more frequent surveilance of oral contraceptive
users ean be controlled for in the analysis.
A weIl conducted case-control study that has taken into account aIl the
methodological concerns can yield valid and informative results. As dis-
cussed earlier in this chapter (see Sections 9.3.3 and 9.5), if cases and controls
are selected independently of exposure and controls are sampled randomly
from a defined study population from which the cases arose, the results from
a case-control study provide an unbiased estimate of the measure of effect
that would be obtained from an equivalent cohort study. Nevertheless, it is
important to remember that case-control studies always have the potential
for bias and that each study should be evaluated individually to determine
whether bias influenced the results. Usually, the diffculty lies in the fact that
although it is easy to identify potential sources of bias in any particular
case-control study, it is rarely possible to estimate the true impact that these
biases may have had on the results.
An important limitation of case-control studies is that they cannot pro-
vide direct estimates of the incidence of disease in those exposed and in those
unexposed (unless they are population-based; see Appendix 16.1). Thus, it is
usually not possible to calculate the absolute impact of the exposure on the
occurrence of the disease.
Case-control studies are not suitable for studying rare exposures because
very few cases wil have been exposed, unless a large proportion of the total
211
Chapter 9
Further reading cases of disease are attributable to that particular exposure (Le., the popula-
tion excess fraction is high (see Section 16.2.1)). For instance, the prevalence
* A brief history of the develop- of asbestos exposure is rare in the general population and accounts for a smaIl
ment and use of case-control proportion of lung cancers. Therefore a case-control study would not be
studies in epidemiology is given appropriate to investigate the relationship between this exposure and lung
in Lilienfeld & Lilienfeld (1979). cancer because very few cases would have been exposed to asbestos.
However, this study design would be appropriate to investigate the relation
* Breslow & Day (1980) provides between asbestos and pleural cancer because this exposure is responsible for
a comprehensive treatment of a large proportion of these cases.
the analysis of case-control
FinaIly, the temporal sequence between exposure and disease may be dif-
studies in cancer epidemiology.
ficult to establish. The possibilty that the exposure is the result (rather than
the cause) of the disease should always be considered (reverse causality). For
* Detailed discussions of sam-
pling schemes for selection of instance, even if an association between diet and stomach cancer is found in
controls are given by Miettinen
a case-control study, there is a possibilty that dietary differences are a con-
an a cause of the cancer.
(1976), Greenland & Thomas sequence rather th
212
Chapter 10
Cross..sectional surveys
L
f
1
Chapter 10
214
Cross-sectional surveys
215
Chapter 10
216
Cross-section al surveys
217
,
"
lr
J
¡
L
Chapter 10
1. The population is first divided into clusters, for example regions, vil-
lages or districts, and a list of the se first-stage units (or primary sam-
pIe units) drawn.
218
¡
f
,,
!
í
Cross-sectional surveys
"":'-""',-,;
, ..(
'. .:/:-::-,
: - '_:, ," '" ", . -,','.-,:, - ,,: ,..-",:;. - - :",,, ::', " .- '- -,'- ,,:,,:- - _. - .,-~
'-"-,-', -,',,-
,-".-,:,-,", ,'''-,-: ",":..'-',.",":
'_-""-'" . "-'; -:'"" " , '-',',..
,-:,:'..---"':--,
,"_.:. '",,,-----,,-,,:"- ",", ":.' .. -', ,,- "." ,
' 'tion teams jneach orthë13(jproduction brigades were
1 i.e.,aJotalof26ÔprQqut,tion teams.
219
Chapter 10
spread coverage of the town. The precise method used is not too impor-
tant, as long as it does not result in aIl the chosen households being very
close to one another, and as long as the mIe for selecting households after
the first is simple and unambiguous, to remove the possibilty of inter-
viewers introducing bias by avoiding certain areas.
If the objective of the study is to obtain an overall estimate of the preva-
lence or level of an attribute across the who le target population, it is sensi-
ble to select the various geographical areas in such a way that the probabil-
ity of selection is proportional to the size of their population. For instance,
a town with a population of 200 000 should stand ten times the chance of
being selected as a town with a population of 20 000. A similar number of
individuals or households should then be taken from both small and large
towns. This probabilty proportional to size sampling approach en sures that
individuals (or households) in both large and small towns stand an equal
probabilty of being selected at the start of the sampling procedure.
The advantages of multiple-stage sampling are obvious in terms of costs
and time. Thus, should all samples be obtained by selection of convenient
clusters? One drawback of this method is that in many situations the clus-
ters are likely to be formed by sets of individuals that are more homoge-
neous than the population as a whole. For instance, people living in the
same neighbourhood or vilage are likely to be similar in terms of their
lifestyle characteristics. If this is the case, individuals in a sample of neigh-
bourhoods provide less information than a sample of similar size obtained
from the whole study population.
ln Example 10.8, eight sex and age strata were formed (1, males born
1922; 2, males born 1932; 3, males born 1942; 4, males born 1952; 5,
females born 1922; 6, females born 1932; 7, females born 1942; 8, females
born 1952), and a random sample selected within each stratum.
220
Cross-sectional surveys
There are many situations where this type of sampling is the most
appropriate. Sometimes we may wish to have independent results for dif-
ferent stratai and to ensure an adequate sample size in each one. Also,
there may be indications that prevalence wil vary between strata (as in
Example 10.8). Since different sampling schemes can be used in different
strata, stratified random sampling is particularly convenient when, for
instance, sampling frames are available only for sorne subgroups of the
population (as in Example 10.9).
221
Chapter 10
222
Cross-section al surveys
Table 10.3.
Age distribution of the female resident
populations of Nykøbing Falster
(Denmark) and Nuuk (Greenland), of
the two random samples selected fram
them, and of the study participants. a
223
Chapter 10
224
\..,~,...
Cross-section al surveys
Table 10.4.
Breast cysts and Iietime use of oral
contraceptives: data from a hypotheti-
cal cross-sectional survey.
ln Example 10.13, the prevalence of breast cysts was 30% higher in ever-
users of oral contraceptives compared to never-users. It should be noted
that prevalence ratio is a good estimate of the incidence rate ratio only if
the prevalence of the outcome of interest among those unexposed is low
(less than 10%) and the duration of the disease is the same among those
who were exposed and those who were unexposed to the factor of inter-
est.
Most often the exposures we are interested in can be further classified
into various levels of intensity as in Example 10.14. We can then examine
trends in prevalence by level of exposure.
It is also usual to measure the strength of the association between a
putative risk factor and the outcome of interest in a cross-sectional study
by calculating the odds ratio. This is the ratio of the odds of exposure to a
putative risk factor in subjects with the outcome of interest to that in sub-
jects without the outcome. By calculating odds ratios, the cross-sectional
study is analysed as if it were a case-control study. However, cross-sec-
225
Chapter 10
(95% confidence intervals (CI) and X2 test for trend calculated using the formulae
given in Appendix 6.1.)
The prevalence ratio for each exposure level was calculated by forming 2 x 2 tables
as illustrated below for women with 10+ partners.
10+
70 1
Number of male sexual partners
19
Total
89
32 71
102 90
103
192
Prevalence among women with 10+ partners = 70/102 = 68.6%
Prevalence among women with one partner = 19/90 = 21.1 %
Prevalence ratio = 68.6% / 21.1 % = 3.3
tional studies differ from case-control studies in that the 'cases' and the
'controls' are defined a posteriori, Le., during the analysis and not at the
design stage. ln fact, if the outcome of interest is quantitative, it is even
possible to carry out several analyses using different definitions of 'cases'
and 'controls' by changing the cut-off point.
226
Cross-sectional surveys
Note, however, that the odds ratio wil yield a good estima te of the
prevalence ratio only if the baseline prevalence of the condition is low (as
in Example 10.15).
, '- _ '- -,', ':,:: ',.,' . ',' : , :,' - " '. - , - '- ',' , , -' ' ,'- -"~" . -, ".,
\,- ,", ,. " , .- ,'.__- .:"., ", _ ..:, "" ..~,;'; '.'_ n , ._: ,',' .", ,-' _.", '.: '_ _, : _ " ',',' '.
',-::,;:'-. :,:,.;, :-.-----,'_:-' :.~_:".-'-:..':',:::_-::-;-'--,),':,',:':;', :,-', '.-"'/, '-'.~,,"; ;':',: -',: :_:,-.: -, '-'.''- :-:-',"
; ::, ': ';' : " : _~:-,':,. '; "::': ,: :,': .-': "C-:':'.-.: _"', ~ ._:'~:. '", .,::"-::__: '_; - :: ',_ :' - :' _,-:- .,' "-':, ': ':. .', -; , :,:_ ' :: ,'.-" _ ,,'_: '_" - - ',,: n _ :-,~ ,_.-: :, '-. ,n:_ ' , '., .
10.4 Interpretation
Cross-sectional surveys are relatively easy and economical to conduct
and are particularly useful for investigating exposures that are fixed char-
acteristics of individuals, such as ethnicity and blood group.
Cross-sectional surveys are not, however, the appropriate study design
to investigate causal relationships because they are based on prevalent
rather than incident cases. Studies of this type can reveal the presence or
absence of a relationship between the study variables and prevalent (exist-
ing) cases. This implies a need for caution, since prevalent cases may not
be representative of all cases of the disease. Cases of short duration,
because of either rapid recovery or death, have a smaller chance of being
detected in a one-time prevalence survey than cases of longer duration. It
follows logically that cases of long duration are over-represented in a
cross-sectional study. The characteristics of these long-duration cases may,
on average, differ in a variety of ways from the characteristics of aIl cases
of the disease being studied. Prevalent cases can also become unrepresen-
227
Chapter 10
tative of aIl cases when certain types of case leave the community. Sorne
affected subjects may be institutionalized elsewhere or move to another
city where there are special facilties for treatment.
Cross-sectional surveys are also an inadequate approach to the study of
rare conditions, since it would be necessary to survey a very large popula-
tion to identify enough cases. Thus, their use in cancer epidemiology has
been limited to the investigation of factors associated with prevalence of
precursor lesions.
Another major limitation of cross-sectional studies is their difficulty in
establishing the time sequence of events. For instance, in our hypotheti-
cal example of breast cysts and oral contraceptive use, it cannot be
assumed that oral contraceptive use preceded the appearance of cysts. ln
fact, women with benign breast disorders are sometimes prescribed oral
contraceptives to improve their condition. ln contrast, there would be no
doubt about the time sequence of the cancer and such traits as blood type
or maternaI exposure to radiation.
228
Cross-section al surveys
Further reading
* A practical book on the role,
planning and conduct of surveys
in developing countries is that by
Casley & Lury (1987).
229
Chapter 11
Studies based on
routine data
Routine-data-based studies are characterized by the fact that data on
both the exposure(s) and the outcome(s) of interest are obtained from rou-
tine data-collection systems (e.g., cancer registries, hospital registries,
death notification, etc.). Thus, studies of this type can be carried out rela-
tively quickly and cheaply, without the need to contact the study subjects.
Their main limitation, however, is that the number of variables available
from routine surveilance systems is generally limited.
Studies based on routine data can be carried out at an individual or at an
aggregated leveL.
231
,l
f
1
Chapter 11
.,..:..:.':,':.,:.-,.....:,:-:'::,......-..:::..:_.,....-',:-.:.:',-':::':',''.
:"'......-..':_-....,.----.-':,.....d-:
Example .11.1., Data (rom population-based cancer registries .'.located d
throughout the world were used to examineinternationãl vafiations, in carL- '
cer incidence. For most cancer sites; there was a morethan tén-roId variation.
between the highest and the lowest recorded incidence rates (Table 11.1).
Table 11.1. Cancer site Males Females
International variations in recorded Highest Lowest H:L Highest Lowest H:L
incidence for selected cancer sites: the (ICD 9 code)
rate (H) rate (L)b ratio rate (H) rate (L)b rati 0
highest and lowest rate among ail the
population-based cancer registries Lip (140) 15.1 0.1 151 1.6 0.1 16
included in Cancer Incidence in Five Tongue (141) 9.4 0.4 24 3.4 0.2 17
Continents, Vol. V. Rates per 100 000
pyrs, age-standardized to the world
Mouth (143-145) 13.5 0.5 27 15.7 0.2 79
standard population. a Nasopharynx (147) 30.0 0.3 100 12.9 0.1 129
Pharynx (146. 148-149) 31.3 0.4 78 4.3 0.2 22
Oesophagus (150) 29.2 1.2 24 12.4 0.3 41
Stomach (151) 82.0 3.7 22 36.1 3.0 12
Colon (153) 34.1 1.8 19 29.0 1.8 16
Rectum (154) 21.5 3.0 7 13.4 1.3 10
Liver (155) 34.4 0.7 49 11.6 0.4 29
Larynx (161) 17.8 2.2 8 2.7 0.2 14
Lung (162) 110.0 5.8 19 68.1 1.2 57
Melanoma of skin (172) 30.9 0.2 155 28.5 0.2 143
Breast (174/175) 3.4 0.2 17 93.9 14.0 7
Cervix uteri (180) 83.2 3.0 28
Ovary (182) 25.8 3.3 8
Prostate (185) 91.2 1.3 70
Testis (186) 8.3 0.6 14
Bladder (188) 27.8 1.7 16 8.5 0.8 11
232
Studies based on routine data
thegeographicaldistril1utionof
triis .analysisshowed-thqtthr!
Uzbekistan and T'urkmenistan,
Incidence
per 100 000 pyrs.
'" 10
i-i 10-29
-i 30-79
_ ;,80
USSR
i;~:f.tY:: ~~~::~~:d high
USSR
China
host population can suggest or confirm the importance of environmen- Figure '11.1.
tal factors over genetic factors in the etiology of a particular cancer. Oesophageal cancer among men in
central Asia. Incidence rates age-stan-
This approach can be refined by adding a time component to it. The dardized ta the world standard popula-
degree of cultural integration into the host country can be measured indi- tion (reproduced, by permission of
rectly by information on age at migration and/or time since migration. Oxford University Press, from Muñaz &
Day, 1996).
ln Example 11.4, European immigrants were at lower risk of dying
fram melanoma than the Australian-born, particularly southern
Europeans. This finding may reflect the protective effect of darker com-
plexions. The data also suggested that migration in childhood was asso-
CÍated with a higher risk than migration later in life, but it is difficult to
separate this effect fram that of duration of stay (Le., migration at an
early age was inevitably assoCÍated with a longer stay th an migration at
an older age).
233
Chapter 11
_.~,- -- -,'
-'. -..' ."
-,'"- -".'
_.,
. ,--"-,--',-,
- ---.,"-,"''',,-''''
- -,'--. .'. -, -,-,,",
-," .-.:'
,~ ..
. Examplell.3. Routinely collected mortality datawere1fsedtô
geographical distribution of oesophageal cancer inBritt4. '
Normandydepartments of Calvados and Orne (Figure 11;2).A .
in most cantons were similar ta the average ratein France,.ins 'tons .
iti easternBrittany and north-western Orne, mortality. was fìveto tEn times ,
higher (Tuyns & Vernhes, 1981). '. ,.
Figure 11.2.
Age-standardized mortality rates for
cancer of the oesophagus among
males, by canton in (a) the Normandy
departments of Calvados and Orne
(repraduced with permission fram
Tuyns & Vernhes, 1981 t9 Masson
Editeur, 1981), and (b) Brittany,
1958-66 (modified fram Tuyns &
Massé, 1973).
234
Studies based on routine data
Figure 11.3.
Relative risks (RRs) of mortality from
melanoma in male European immi-
__ England
grants to Australia, by region of birth
-0 Ireland, Scotland and Wales and age at arrivai compared to
0.8 -7 Central Europe Australian-born (taken as the baseline
Eastern Europe group: RR=1); Auslralia, 1964-85
Southern Europe (reproduced with permission from Khlat
et aL., 1992).
.: 0.6
'"
.;:
Ql
;:
:;
ãi
a: 0.4
0.2
o
~15 15-24 25+
Age at arrivai (years)
ln Example 11.5, there was a consistent pattern: the risk for aIl four
cancer sites converged towards the risk in the host country. For stomach
and liver cancer, both of which were more common inJapan than among
whites in the United States, men who were born inJapan but migrated to
California had considerably lower risks of death than men of the same
age in Japan. But risks in the second generation were stillower. ln con-
trast, the risk of colon and prostatic cancers, which were comrnon in
California but rare in Japan, rose to approach that of white men in
California with migration. These findings clearly show the importance of
environmental factors over genetic factors linked to ethnicity. There was
stil a residual difference between rates in second-generation immigrants
and rates in California. These residual differences might reflect differ-
ences in genetic susceptibilty to these malignancies or, alternatively,
they might indicate that second-generation immigrants maintained sorne
of the 'traditional' Japanese lifestyle.
235
Chapter 11
Figure 11.4.
Age-adjusted mortality rate ratios
(RRs) for cancer of the stomach, liver,
colon and prostate among Japanese
men in Japan, and first- and second-
generation Japanese immigrants to
California, compared with white men in
California (taken as the baseline group:
RR=1); Japan, 1958-59 and California,
1956-62 (data from Bueil & Dunn,
1965).
\\
~")~.
~~~~Q~~ Studies based on routine data
AUSTRAlIA
S. Italy
Allitaly
Migrants
UK, Scotland
S.ltaly
Allitaly
Migrants
URUGUA y
S. Italy
Allitaly
Migrants
ARGENTINA S.ltaly
Allitaly
Migrants
FRANCE S.ltaly
Allitaly
Migrants
BRAZIL, São Paulo
S.ltaly
Allitaly 1.42
Migrants
237
Chapter 11
Colon (153) 338 7.1 1.0 1.4 2.1 2.2 13.3; P-:0.001
Cervix uteri (180) 645 7.2 1.0 0.7 0.4 0.2 50.2; P-:0.001
a Data from Bouchardy et al. (1993)
b Controls are ail cancers except the one under investigation (see Section 11.1.6 for
discussion of the method of analysis). Odds ratios are adjusted for age, civil status,
and country of birth.
cAfter exclusion of subjects with unknown level of education.
d Taken as the baseline category.
and absence from home for long periods, a situation known to be asso-
ciated with high risk of venereal diseases among men. This pattern sup-
ported the hypothesis that cervical cancer might be a sexually transmit-
ted infection long before any infectious agent was identified (Beral,
1974).
ln interpreting differences in incidence or mortality between socioe-
conomic groups, it must be remembered that health itself may deter-
mine entry into a specific group. For instance, people with poor health
are usually forced to work in physically less demanding jobs and more
demanding jobs selectively include only those in good health.
239
Chapter 11
Rx~nipleJ.l.9...Roütinelyco11ected....mortaUtJl.4ata.forEnglanii..andWales
were~sed.toexamine..riskpfmortalityfTom .cervicalcancer amongmarried
W. o.nien...by.hus.b.a.nd's. 0.. ccuP. a. tion .and....s.o.c...l...al.......c...la..s. s..0....ßera..l,. 1...9..74). (ln. this
countT...death... registrarsare requiredtoenterthe.hasbandls... lJccupation .on
the death certíficate ota married. womanor widow;socialclass for married
. women isdßtermined by the husbands occupation.) Theresults areshown
inTableJ 104.
b Age-specifie rates for ail married women in England and Wales taken as the
standard.
240
Studies based on routine data
Figure 11.6.
Age-specifie incidence rates for papil-
Year of incidence lary carcinoma of the thyraid in women
by calendar period, Connecticut (USA),
1965-75 1935-75 (repraduced with permission
4
/ \ "
1935-44
o
1 / \
ii --- ,"y--~,,--./ \
/' - / / / \
1/ // "-__~
\
decline in risk for generations born during the Second World War occurred
in Denmark, Norway and Sweden, which were apparently less affected by
the war th an Poland, the former East Germany and Finland.
Time trends can also be used to assess and monitor the effectiveness of
cancer control activities such as mass screening programmes. ln Example
Il.12, there was a close relationship between the decline in incidence of
cervical cancer in each country and the degree of coverage by organized
mass screening programmes. The decline in incidence was most marked
in Finland and Iceland, where national screening programmes were initi-
ated in the early 1960s. The fall was less marked in Sweden, where the
programme was introduced more graduaIly, and in Denmark where only
40% of the population lived in areas with organized mass screening.
There was no obvious decline in the incidence of cervical cancer in
Norway, the only country which did not have an organized programme
(except in one county).
241
\,..
'-"\..
Chapter 11
Figure 11.7.
Relative risk (RR) of developing testic-
ular cancer by country and year of
birth. (Men born in the study countries
between 1900 and 1909 were taken as
the baseline category: RR=1) (repra-
duced with permission fram Bergström
et aL., 1996).
11.1.6 Analysis
Studies based on routine data conducted at an individual level may be
regarded as cohort studies in which a group of people, or cohort, is followed
up in time. For instance, comparison of cancer incidence or mortality across
different occupational groups may be regarded as a study of groups of peo-
ple with different occupational 'exposures' who were followed up in time
and their cancer experience compared. Thus, the analysis of these studies is
similar to the analysis of any other cohort study and the methods are basi-
cally those described in Chapters 4 and 8. The analysis is based on calcula-
242
Studies based on routine data
Figure 11.8.
Trends in age-adjusted incidence of
cervical cancer in the five Nordic coun-
tries, 1943-78 (reproduced with per-
mission trom Hakama (1982)).
tion of rates as measures of disease occurrence, and of rate ratios and rate di(-
ferences as measures of effect. The analysis of time trends by calendar period
and birth cohort was considered in Section 4.3.2.
Calculation of rates requires information on person-time at risk (Le.,
denominators) for each of the groups of interest. Population estimates to
allow these calculations are not always available, however. For instance,
although national statistical offices in many countries are able to provide
population estimates for the whole country and for relatively large geo-
graphical areas, they may be unable to provide data for smaller geo-
graphical areas. Moreover, cancer registrations and death certificates may
contain information on many variables for which data are not collected
by the census (or any other population enumeration system). For
instance, cancer registries may collect information on ethnicity, but prop-
er denominators wil not be available if information on that variable is
not collected by the census. ln other situations, denominators are avail-
able from the census but case-finding is so incomplete (for instance, if on
a large proportion of death certificates, data on occupation of the
deceased is mIssing) that the numerator data are not comparable with the
available denominator data.
243
Chapter 11
~. ..~~~~..'2~&~:z.¡a:=-=~
Suppose also that the total number of incident cancers among maleprint-
ers during the same period (1970-74) was 1001 ofwhich 40 were Iungcan-
cers. The proportion of male Iung cancers amongprinters was
40/100= 0.40
number (E): .
the observed number of lung cancers among the printers (0) by the expec.ed
245
,
i
,,
!
í
Chapter 11
246
'''''''''''''='''''~''-~~ Studies based on routine data
nr:i=~ -----.....- """=-=;'.,.r~~9I-.,v.=~__""....=~=~__,-~~~~~.,,,~.V"';""!=~~~~-=='''~~~''-=''~~=i!'2'''
11.1.7 Interpretation
Routine surveilance systems usually coyer large catchment popula-
tions, sometimes of milions of people followed up over long periods of
time. Thus, one of the main advantages of routine-data-based studies is
that they allow the study of a very large number of people at a very low
cost.
A major limitation, however, is the restricted range of variables collect-
ed by these systems, that often tend to be just proxy measures of more bio-
logically relevant exposures. For instance, country of birth may act just as
marker of environmental factors (e.g., diet, reproductive variables) for
which data are not available in routine data systems.
Data qualiy
Potential data artefacts need to be considered when interpreting results
from these studies. The observation of differences in recorded cancer inci-
dence between populations does not necessarily reflect true underlying
variations in cancer risks. Differences in recorded incidence may arise
because of differences in health service access (including screening), diag-
nosis, and registration practices. ln addition, they may be due to varia-
tions in the accuracy of enumeration of the population. Sorne of these
issues were discussed briefly in Section 2.9 and Appendix A2.2.
An example is the difficulty in assessing rising trends in the recorded inci-
dence of prostatic cancer În many western countries due to uncertainty about
the effect of changes În diagnostic and registration practices. Strong parallels
have been observed in the USA between time trends in the recorded inci-
dence of localized prostate cancer and time trends in the use of transurethral
resection of the prostate (Potosky et al., 1990; Severson, 1990), a surgical pro-
cedure performed to relieve commonly occurring symptoms of benign pro-
247
Chapter 11
248
Studies based on routine data
Selection bias
People who migrate are not generally a representative group of their pop-
ulation of origin. For instance, migrants tend to be healthier than those who
stay in the home country. This 'healthy migrant effect wil affect compar-
isons with home population risks. A similar bias may occur in analyses by
occupation (the 'healthy worker effect) and this wil affect comparisons with
the general population (see Sections 8.2.2 and 13.1.1).
Another source of bias, that is more difficult to detect, results from changes
in 'exposure' which are related to the disease event itself. For example,
migrants may retum to their country of origin soon before death or people
move jobs as a consequence of being diagnosed with a particular condition.
ln these situations, risks in the host country or in the job held before diag-
nosis wil be under-estimated.
Confounding
One of the main limitations of routine-data-based studies is that informa-
tion on important confounding factors (with the exception of age and sex) is
generaIly not available. For instance, the high risks of lung cancer found in
sorne occupational groups are not due to occupational exposures but to high
prevalence of smoking. Unfortunately, data on smoking habits are rarely col-
lected by routine surveilance systems.
Final remarks
ln summary, the study of variations in cancer incidence and mortality by
place of residence and birth, ethnicity, socioeconomic status and over time is
a valid and useful exercise, provided that the investigator has a thorough
knowledge of the way data are collected and processed so that aIl possible
sources of data artefacts, bias and confounding are considered in the inter-
pretation of the findings.
249
,
i
f
t
¡
!
Chapter 11
11.2.1 Analysis
Ecological studies differ from individual-based epidemiological studies in
that the 'exposed' and 'unexposed' individuals in each of the populations
are not actually identified. Thus, it is not possible to measure the strength
of an association between exposure and outcome by using any of the
approaches described in Chapter 5 for individual-based studies.
Moreover, in contrast to other epidemiological studies, the outcome mea-
sure in an ecological study is usuaIly a quantitative variable (e.g., mortality
rate) rather than a binary one (such as 'diseased' versus 'non-diseased'). The
250
\..
...\....
Studies based on routine data
Figure 11.10.
Scattergram showing the relationship
between age-standardized mortality
ratios fram ovarian cancer and average
completed family size for women born
araund 1901 in 20 countries. The age-
specifie mortality rates in England &
Wales were used as the standard, i.e.
the mortality ratio for England & Wales
= 100. (Reproduced with permission
fram Beral et al., 1978. (Ç by The
Lancet Ud, 1978). BEL=Belgium;
BUL=Bulgaria; CAN=Canada;
CZE=former Czechoslovakia;
OEN=Oenmark; E&W=England and
Wales; FRA=France; FIN=Finland;
GER=former West Germany;
IRE=lreland; ITA=ltaly; JAP=Japan;
NET =Netherlands; NOR=Norway;
POL=Poland; SCO=Scotland;
SPA=Spain; SWE=Sweden;
USA=United States of America;
YUG=former Yugoslavia.
251
Chapter 11
Figure 11.11.
Scattergram showing the relationship
between age-standardized mortality
ratios fram ovarian cancer and average
completed family size for different gen-
erations of women born in England and
Wales and the USA. The average age-
structure of the combined population of
England & Wales and the USA fram
1931-73 was taken as the standard.
Thus, a ratio of 140 means that the
ovarian cancer mortality of the cohort is
40% higher than the average for
women in England & Wales and the
USA. (The mid-year of birth of each
generation is shown in brackets)
(repraduced with permission fram Serai
et al., 1978. (Q by The Lancet Ltd,
1978).
gram show a clear trend, upwards and to the right; there is said to be a
positive relationship between the two variables. High values of one vari-
able are associated with high values of the other, and low with low. To
summarize the relation between the two variables, so as to be able to pre-
dict the value of one variable when we only know the other variable, we
could just draw a straight line through the scatter of points. Any straight
line drawn on a graph can be represented by an equation. ln the ab ove
example, this relationship could be summarized as y = x, because each
252
Studies based on routine data
Figure 11.12.
Age- and sex-adjusted incidence rates
of squamous-cell carcinoma of the eye
in relation to measurements of ultravio-
. .. let B radiation expressed as minimum
erythemal dose (MED), a unit which
. reflects more c10sely the biologically
effective dose (reproduced with per-
mission trom Newton et al., 1996. iQ by
The Lancet, Ud, 1996).
4 6 8 10 12 14 16
UV-B (MED)
y = a + bx
where y refers to values of the outcome variable and x to values of the expo-
sure variable, a is the intercept of the line on the y axis, and b is the slope of
the line, the increase (or decrease) in y per unit increase in x (Figure 11.14).
ln the special case in which the line passes through the origin (0) of the two
axes, as in Figure 11.13(a), and each value of y is exactly equal to the corre-
sponding value of x, the equation reduces to y = x, since a = 0 and b = 1.
ln Figure 1 1. 13 (b), there is also a perfect association between x and y, but
253
Chapter 11
°000
o
o
o
oo 00 00
o
The relationships ilustrated in Figure
l1.13(a), (b) and (c) are perfect, in that aIl
o 0
o o0 00 the points on the scatter diagram lie on a
o o
o 0
line. ln most real situations, however, the
x x
points are scattered around it (as in Figures
l1.13(e) and (f). ln these circumstances, we
use a straight line that gives the 'best pre-
y=a+bx y=a-bx diction of y for any value of x. We could just
(e) O.:r-d O;.r;.-I
(1) draw a line 'by eye', but such a subjective
y
ó'
y method is unlikely to yield the best line. An
8
o
alternative is to use the statistical method
o 0 go 0 0 'l oP 00 called regression analysis, to find the best line
00 0000 0 a 0 °0°0 o o 0
o 0o o o
that 'tifs' the data. The equation of the
o 00 0 o
straight line obtained by this method is
o 00°
o o called the regression equation (see Armitage
x x
and Berry (1994) for ilustration of calcula-
tions).
Figure 11.13.
Scattergrams, regression equations
This statistical method was applied to the ovarian cancer mortality data
and correlation coefficients. shown in Figure 11.11 (using the data from both countries) and yielded
the following regression equation (Beral et al., 1978):
y= 182-30x
:5
;:
o 0o 00 0
o
This equation means that for any given value of x (Le., average family
()
o 0 a..b"
'f 0 000
t:
QI
::
c- o 00 0 o size), an associated value for y (Le., ovarian cancer mortality) can be cal-
QI
o culated. Thus, the age-standardized mortality ratio from ovarian cancer in
any particular birth cohort of women can be predicted from this equation.
H o
For example, the mortality ratio in a cohort with a mean family size of
three children can be predicted by substituting x = 3 into the regression
Level of exposure to factor (x)
equation
Figure 11.14.
Scattergram and regression line. y = 182 - 30 x 3 = 92
254
Studies based on routine data
which is the value we would have estimated by examining the graph. This
predicted value can be interpreted as the average value of y associated with
a given value of x.
If aIl the values in the scattergram layon the regression line, the pre-
dicted and observed values of y would be identical. The regression equa-
tion would describe exactly the relationship between average family size
and ovarian cancer mortality. ln other words, the variation in ovarian
cancer mortality would be completely explained by the variations in
family size. ln practice this is not the case. Ovarian cancer mortality can
vary independently of variations of family size, so that two populations
with the same average family size may have different ovarian cancer
mortality risks. Thus, the regression equation can only measure the aver-
age relationship between the two variables.
If the points on the scattergram lie close to the regression line, this
suggests that the observed values for y do not differ markedly from the
predicted values represented by the regression line. Thus, most of the
variation in y can be explained by the variation in x. If, on the other
hand, there is a wide scatter of points around the regression line, a con-
siderable amount of the variation in y is not explained by the variation
in x.
To quantify the degree of scatter around the regression line, we can
calculate a measure called a correlation coeffcient, r. The value of this
coeffcient always lies between - 1 and + 1:
0 00
0 A more appropriate method to quantify the effect of exposure in an eco-
:5 0 0
;: ¡¡..b¡(
0
0
0 logical study is to fit a regression line to the data, which predicts incidence
o 0 0
i: o 'F 0 or mortality as a function of the level (or prevalence) of exposure. ln con-
Ql
:i 0
0- o 00 0
Ql
o trast to the correlation coeffcient, a regression line remains unaffected by
Ya Yi changes in the range of the exposure variable. Moreover, a measure of rel-
ative effect can be estimated from the slope and the intercept of the regres-
H o Xo Xl sion line (Beral et aL., 1979).
Level of exposure to factor (x) Let us assume that in n different populations the level of disease (y) is
linearly related to the average level of exposure (x) to the factor being stud-
Figure 11.15. ied. This relationship can be expressed by a regression equation fitted to
Illustration of the association between the data from n populations:
disease frequency (y) and level of
exposure (x) to the factor being studied
in various populations. y = a + bx
Thus, if
Yi a + bxi
RR=
Yo a + bxo
For instance, in our ovarian cancer mortality example, the regression
equation was found to be y = 182 - 30x. Thus, the relative risk of ovarian
cancer in a population with an average family size of two children com-
pared to that with an average of four children can be estimated as:
182 - 30 x 2
RR= = 2.0
182 - 30 x 4
256
Studies based on routine data
Figure 11.16.
Age-standardized incidence rates of
60 breast and cervical cancers in groups
of municipalities defined according to
the levels of various socioeconomic
characteristics of each municipality,
0 o
12345 12345 12345 1234 12345 12345 12345 Finland, 1955-74. The average Finnish
CERVIX
population in 1955-74 was taken as
the standard population. (Reproduced
20
with permission from Hakama et al.,
10 1982).
0 o
ln come Social Secondary Tele- Crowded- Running Central
classes education vision ness water heating
1&2
257
Chapter 11
258
1
,f
f
f
J
Studies based on routine data
Confounding
A second major limitation of ecological studies
Czechoslovakia .
is the lack of abilty to control for the effects of
potential confounding factors. For example, in a . Denmark . New Zealand
study of average per caput daily intake of fat in 24 en
20.0
Belgium .
;;
a.
countries in relation to their breast cancer inci- o
oo . Austria
dence in women aged 35-64 years, there was a pos- oo .UK
. France . Switzerland . W. Germany
itive relationship between these two variables L
'"
a. 15.0
. Hungary
.USA
en
(Armstrong & Mann, 1985), suggesting a possible '"
ë . . Canada . Australia
association between fat intake and risk of develop- ¡;
::
Ireland
c: . Sweden
ing breast cancer. However, increased fat con- c:
'"
'" .Italy
sumption may merely be acting as a marker for '"
'"
. Israel
,.Q; 10.0
other factors that are related to elevated risk of et · Portugal . Netherlands
. Japan Norway
breast cancer, su ch as higher socioeconomic status,
lower fertilty and later age at first birth. Data on .Iceland \.. Finland
Pol and
population leveL
It should be noted that risk factors which are independent of exposure Figure 11.17.
at the individual level may become correlated with it, and therefore Scattergram showing relationship
become confounders, when aggregated at the population level. For exam- between estimated average annual
age-adjusted incidence rates for rectal
pIe, in an investigation of the relationship between the proportion of cancer among males in 24 countries
woodworkers and lung cancer across geographically defined areas, smok- and per caput beer consumption,
ing wil induce confounding if cigarette consumption changes with the 1960-62 (apprax.) (repraduced, by
permission of Oxford University Press,
proportion of woodworkers in each are a, even if the two factors in ques- tram Breslow & Enstram, 1974).
259
,i
,, ...\..
!
1
Chapter 11
Measurement errors
Exposure is most often estimated from data collected for other reasons,
which generally provide only an indirect measure of possible risk factors.
For example, data on smoking and alcohol are often based on sales, which
only partially reflect consumption, because los ses and unregistered
imports are not taken into account. Moreover, since the data are not col-
lected by the investigators themselves, it may be difficult to assess their
quality adequately.
Measurement of cancer incidence and mortality can also be affected by
errors, as discussed in Section 11. 1. 7.
Latent period
Most ecological studies compare exposure measured at one point in
time with disease measured at another (or the same) point in time. This is
ilustrated in Figure 11. 1 7, where data on both the disease rate and the
exposure of interest (per caput beer consumption) refer to approximately
the same period (1960-62).
IdeaIly, an appropriate time-lag period should be incorporated into the
analysis, so that exposure data refer to the relevant etiological period (e.g.,
10-20 years before the development of cancer). Data on past exposures are
not always available and, quite often, we are forced to rely on data from a
period far too recent. This constitutes a serious problem when exposures
are likely to have changed markedly over time (e.g., smoking and alcohol
consumption), particularly if the rate of change has been diferent in the
different groups. Even when relevant past exposure data are available, the
populations on which exposure and outcome are measured may not be
the same, as a result of dynamic changes introduced by births, deaths and
migrations.
A special situation arises when birth cohort changes in exposure are
related to birth cohort changes in disease risk or when both the exposure
and the disease data refer to a specific birth cohort of individuals.
Examples 11. 1 6 and 11. 1 7 ilustra te this point. The average family size
summarized the reproductive experience of each generation of women,
that is the total number of children achieved by the end of their repro-
ductive lives. The ovarian cancer cohort risks used in these examples also
summarized the mortality experience of each generation of women (see
Beral et aL. (1978) for details of the calculations). ln such situations, there
is no need to build any time lag into the analysis.
260
Studies based on routine data
261
Studies based on routine data
" " . r" : ~ '. " "'~'" BQj'11)1. i~:~nttl) ~'Kei Js.~pes. " ,': ""~' .. - ,,:"~' .~: " ' '. ,:
262
Chapter 12
Introduction to survival
analysis
12. i Introduction
ln this chapter we examine another method used in the analysis of inter-
vention trials, cohort studies and data routinely coIlected by hospital and
population-based cancer registries. Consider the foIlowing example:
ln Example 12.1, the patients entered and left the study at different points
in time (Table 12.1). We discussed in previous chapters (Chapters 4, 7 and 8)
one way of analysing data of this tye which takes into account the varying
individual lengths of foIlow-up. That approach involves the calculation of
rates based on person-time at risk. These calculations are based on the assump-
tion that the rate under consideration remains approximately constant over
time, so that 100 person-years of observation are treated identicaIly, whether
they involve 100 subjects foIlowed over one year or 50 subjects foIlowed over
two years.
ln many situations, however, the rate of occurrence of the event under
study does not remain constant over time. For instance, the probabilty of
dying may rise suddenly with the onset of disease and then decline gradual-
ly as time since diagnosis increases. The most appropriate approach in these
situations is to conduct survival analysis.
263
1
J
f
f
Chapter 12
Table 12.1. Patient Age Stage" Date of Date of Vital Cause of No. of No. of days
Follow-up of 40 women diagnosed with study (years) diagnosis last status at death C complete from
breast cancer in a certain hospital dur- number contact or last years of diagnosis
death contactb observation ta last
ing the years 1989-93: hypothetical from contact or
data. diagnosis ta death
last contact
or death
because this is the point when the treatment groups are comparable. ln
Example 12.1, the date of diagnosis was taken as the starting point.
The third requirement is a clear and well defined outcome. Often the out-
come of interest is death, but it need not be so. It can be recurrence of
264
"\..'1...
Introduction to survival analysis
Figure 12.1.
(a) Calendar year of diagnosis
Diagram illustrating how follow-up data
1989 1990 1991 1992 1993 1994 1995
from 8 of the 40 women with breast
Study
no.
cancer (see Table 12.1) can be pre-
sented (a) by calendar year of diagno-
1 --nA sis and (b) by time since entry into the
3 ---D
5 n---A study (A=alive; D=dead).
15 _nhA
22 ----0
25 -0
30 nA
37 0
0 2 3 4 5 6 7
Study
no.
1 ----A
3 --0
5 ---A
15 -A
22 ---D
25 ----D
30 A
37 ----D
265
Chapter 12
266
Introduction to survival analysis
D/(N - 0.5L)
Thus censoring reduces the effective size of the cohort by haU the size
of the group lost to foIlow-up (O.SL). This rather crude way of taking
account of censoring works adequately provided L is small compared with
N.
We can now re-calculate the probabilty of dying in the first two years
in Example 12.1. Thus, of the 40 breast cancer patients recruited into the 10 dead
study
40 6 censored
10 died during the two-year follaw-up period (D = 10)
years of diagnosis. Since the last attempt to contact patients was made in
1995, patients diagnosed after 31 December 1992 entered the study too
late to have been able to complete a three-year foIlow-up. Thus, the obser-
vations for four patients (Nos 34, 38-40) were censored (Table 12.1). Five
other women (Nos 13, 6, Il, 23, 26) did not complete the three-year obser-
vation period because they were lost to foIlow-up (Figure 12.2). ThusJ of
the 40 breast cancer patients recruited into the study:
267
Chapter 12
268
Introduction to survival analysis
ln the second year, six women were censored. The effective size of the
cohort in this year can then be estimated as 33 - (0.5 x 6) = 30. Thus
the probabilty of a subject dying during the second year, given that the subject
was alive at the start of the year, is estimated to be 3/30 = 0.10;
the probability of surviving the year is estimated to be 1 - 0.10 = 0.90.
ln the third year, three women were censored. The effective size of the
cohort is 24-(0.5 x 3) = 22.5. Thus
the probability of a subject dying during the third year, given that the subject was
alive at the start of the year, is estimated to be 4/22.5 = 0.178;
the probabilty of surviving the year is estimated to be 1 - 0.178 = 0.822.
The full tree with the branch (conditional) probabilties of dying in each
year given that the subject survived the previous years is shown in Figure 0.175
D
12.7. 0.100
D
There are now four possible outcomes of interest, corresponding to the 0.825 D
0.178
tips of the tree. The probabilty of each outcome can be calculated by mul- 0.900
tiplying down the branches of the tree. Therefore the probabilties for each 0.822
s
outcome are:
0- 1- 2- 3-
Years ollollow-up
These probabilties wil always add Up to 1, sin ce there are no other pos-
sible outcomes. The probability of dying at some point during the three-
year interval is equal to 0.175 + 0.083 + 0.132 = 0.390. This probabilty can
be found more conveniently by subtracting the probabilty of surviving
the whole three-year period from 1, giving 1 - 0.610 = 0.390.
The final probabilty of surviving (0.610) is an example of a cumulative
survival probability for the cohort, Le.) the probabilty of surviving three
consecutive years.
12.6 Actuariallife-table
The data from the previous calculations are often presented in the form
of an actuariallife table, which shows the numbers of deaths and cens Of-
ings occurring in each consecutive intervaL A life table for the 40 breast
cancer patients from Example 12.1 is shown in Table 12.2.
ln this table, the probabilty of dying during each year is calculated as
D/(N - O.SL). Thus, the probabilty of surviving the year is equal to 1 -
D/(N - 0.5L). The cumulative survival is found by multiplying the survival
probabilties for each of the consecutive years to obtain the cumulative
269
Chapter 12
Table 12.2. Year No. at start No. of No. of Effective Probabilty Probabilty Cumulative
Actuarial life table for the 40 breast of interval deaths losses denominator of dying of surviving survival
cancer patients of Example 12.1. (N) (0) (L) (N-.5L) during the the year
year
0- 40 7 0 40 0.175 0.825 0.825
1- 33 3 6 30.0 0.100 0.900 0.743
2- 24 4 3 22.5 0.178 0.822 0.610
3- 17 4 4 15.0 0.267 0.733 0.447
4- 9 2 3 7.5 0.267 0.733 0.328
5- 4 1 2 3.0 0.333 0.667 0.219
6- 1 0 1 0.5 0.0 1.00 0.219
Total 21 19
270
Introduction to survival analysis
0.25
the end of the foIlow-up period, the
Kaplan-Meier curve has a plateau commencing
at the time of the last death and continuing until 0.00
the censored survival time of this longest surviv- o 2 3 4 5 6 7
ing patient. ln Figure 12.9, the survival time of Years
each censored observation is marked in the
No. of
curve by a cross. Afer the last death (patient 16, patients 40 33 24 17 9 4 o
at day 1862 (5.1 years)), the curve remains flat
until the longest censored survival time (patient
5, at day 2390 (6.5 years)). Figure 12.9.
Survival curve produced by the
It is useful to give the number of patients at risk at selected time points (for Kaplan-Meier method for the 40
example, at the start of each year) under the graph and/or to present confi- breast cancer patients of Example
dence intervals around the survival probabilty estimates. This information is 12.1 (x indicates censoring times).
crucial for a sensible interpretation of any survival curve.
271
Chapter 12
Figure 12.10.
Kaplan-Meier sUNival CUNes for 1.00
patients with breast cancer by stage of
the tumour at the time of diagnosis
(group 1 = tumour without Iymph node
involvement or metastasis; group 2 = 0.75
tumour with Iymph node involvement ~
:a
and/or regional or distant metastasis). cu
J:
0
The numbers on the sUNival CUNes Ci 0.50
represent censored obseNations. (i
,.
.~
:i
en
0.25
Group 2
0.00
0 2 3 4 5 6 7
Years
No. of patients
Group 1 : 19 17 11 9 5 3 0 0
Group 2: 21 16 13 8 4 1 1 0
272
l
j
f
r
Introduction to survival analysis
Year No. at start No. of No. of Effective Probabilty Probabilty Cumulative Table 12.3.
of interval deaths losses denominator of dying of surviving survival Life-table probabilities of dying from
(N) (D) (L) (N-.5L) during the the year breast cancer for the 40 breast cancer
year patients of Example 12.1. ln this table,
0- 40 7 0 40 0.175 0.825 0.825 deaths from causes other than breast
1- cancer were considered as censored
33 3 6 30.0 0.100 0.900 0.743 observations.
2- 24 3 4 22.0 0.136 0.864 0.641
3- 17 3 5 14.5 0.207 0.793 0.509
4- 9 1 4 7.0 0.143 0.857 0.436
5- 4 0 3 2.5 0.0 1.0 0.436
6- 1 0 1 0.5 0.0 1.0 0.436
Total 17 23
ing for these 'competing' causes of death is to treat patients who died
from other causes as if they had been withdrawn alive (Le., censored at
the time of their death) and then carry out the lie-table calculations as
described above. The survival probabilties obtained by this method are
cause-specifie survival probabilties, since they take into account deaths due
to causes other th an the disease under study.
ln Example 12.1, four patients died from causes other than breast can-
cer (see Table 12.1). A new actuariallie-table can then be constructed by
considering these deaths as censored observations (Table 12.3). The total
number of deaths is decreased by 4 (17 instead of 21) and the number of
losses increased by 4 (23 instead of 19). Similarly, when the exact dates
at which deaths occur are known, it is possi-
ble to use the Kaplan-Meier method to esti-
mate these cause-specific survival probabil- 1.00
of people in the general population similar to the patient group with Kaplan-Meier breast cancer-specifie
survival curve for the 40 breast cancer
respect to race, sex, age and calendar period of observation. This expect- patients of Table 12.1.
273
\,~
~'.\,
Chapter 12
(c)
275
Chapter 12
Further reading
* The use of probability trees in
this chapter was based on the
approach suggested by Clay ton
& Hills (1993).
276
Chapter 13
Interpretation of
epidemiological studies
277
i \..
l
f
,!
l
t
Chapter 13
who were not. ln aIl instances where selection bias occurs, the result is a
difference in the relation between exposure and outcome between those
who entered the study and those who would have been eligible but did
not participa te. For instance, selection bias wil occur with volunteers (self-
selection bias). People who volunteer to participa te in a study tend to be
different from the rest of the population in a number of demographic and
lifestyle variables (usuaIly being more health-conscious, better educated,
etc.), sorne of which may also be risk factors for the outcome of interest.
Selection bias can be a major problem in case-control studies, although
it can also affect cross-sectional studies and, to a lesser extent, cohort stud-
ies and randomized trials.
The selection of an appropriate sample for a cross-sectIonal survey do
es
not necessarily guarantee that the participants are representative of the
target population, because sorne of the selected subjects may fail to par-
ticipate. This can introduce selection bias if non-participants differ from
participants in relation to the factors under study.
ln Example 13. l, the prevalence of alcohol-related problems rose with
increasing effort to recruit subjects, suggesting that those who completed
the interview only after a large number of contact attempts were different
from those who required less recruitment effort. Constraints of time and
money usually limit the recruitment efforts to relatively few contact
attempts. This may bias the prevalence estimates derived from a cross-sec-
tional study.
ln case-control studies, controls should represent the source population
from which the cases were drawn, Le. they should provide an estimate of
the exposure prevalence in the general population from which the cases
come. This is relatively straightforward to accomplish in a nested case-con-
trol study, in which the cases and the controls arise from a clearly defined
population-the cohort. ln a population-based case-control study, a source
population can also be defined from which aIl cases (or a random sample)
are obtained; controls wil be randomly selected from the disease-free
members of the same population.
The sampling method used to select the controls should ensure that they
are a representative sample of the population from which the cases origi-
nated. If they are not, selection bias wil be introduced. For instance, the
method used to select controls in Example 13.2 excluded women who were
part of the study population but did not have a telephone. Thus, control
selection bias might have been introduced if women with and without a
telephone differed with respect to the exposure(s) of interest. This bias
could be overcome by excluding cases who did not have a telephone, that
is, by redefining the study population as women living in households with
a telephone, aged 20-54 years, who resided in the eight selected areas.
Moreover, the ultimate objective of the random-digit diallng method was
not merely to provide a random sample of households with telephone but
a random sample of al! women aged 20-54 years living in these households
during the study period. This depended on the extent to which people
278
;
i
!
1
i
1
Interpretation 01 epidemiological studies
100
o ;~'i. ,~;~;,
123 4 5 6 7 8 91011121314151617181920212223242526_57
Contact attempts
279
Chapter 13
Hormone Study. The study population was women aged 20-54years who
resided in eight selected areas inthe USAduling the study period. Attempts
were made to.ideniífy aIl incident cases ofbreast, ovariari.ünd endometrial.
cancerthat occurred inthestudy population duling the stùdy periodthrough
local population-baseq cancer registries. Controls, were selected by. random~
digitdiallng ofhouseholds inthe dght locations. A random sample ofhouse-
hold telephone numbers were caIled; information on the age and sex of 'aIl
household members was requested and controls were selected among female
members aged 20-54 years according to strct rules (Stadel et aL, 1985).
Table 13.2 shows that among men of similar age, the percentage of non-
smokers was 7.0 in the hospital controls and 12.1 in the population sam-
280
Interpretation of epidemiological studies
pIe. The percentage smoking at least 25 cigarettes per day was 13.4 in the
hospital con troIs but only 8.5 in the population sample. The investigators
stated that this difference in smoking habits between the hospital controls
and the population random sample might be explained by previously
unknown associations between smoking and several diseases. Thus, the
strength of the association between lung cancer and cigarette smoking was
underestimated in the case-control study, because the prevalence of smok-
ing in the hospital con troIs was higher than in the general population
from which the cases of lung cancer were drawn.
A hospital control series may fail to reflect the population at risk
because it includes people admitted to the hospital for conditions caused
(or prevented) by the exposures of interest. Individuals hospitalized for
diseases related to the exposure under investigation should be excluded in
order to eliminate this type of selection bias. However, this exclusion
should not be extended to hospital patients with a history of exposure-
related diseases, since no such restriction is imposed on the cases. Thus,
patients admitted to hospital because of a smoking-related disorder (e.g.,
chronic bronchitis) should be excluded from the control series in a
case-control study 100 king at the relationship between smoking and lung
cancer, whereas those admitted for other conditions (e.g., accidents) but
with a history of chronic bronchitis should be included.
Selection bias is less of a problem in cohort studies, because the enrol-
ment of exposed and unexposed individuals is done before the develop-
ment of any outcome of interest. This is also true in historical cohort stud-
~ 150
~
~153 136 123
95~~
,g 125
E
ies, because the ascertainment of the exposure status was made sorne time ~100 _!OO~ /____ ~ 1~~_______1_0~_
in the past, before the outcome was known. However, bias may stil be t
ãi
83
E 75
introduced in the selection of the 'unexposed' group. For instance, in "0 ß Cancers
occupational cohort studies where the general population is used as the :! 50
:¡
o Ali causes
o Non-cancers
(¡
comparison group, it is usual to find that the overall morbidity and mor- "0
c:
CG
tality of the workers is lower than that of the general population. This is ii
because only relatively healthy people are able to remain in employment, o 5 10 15
Years after cohort identification
while the general population comprises a wider range of people including
those who are too il to be employed. This type of selection bias is called Figure 13.2.
Decline in the healthy worker effect
with passage of time after initial identifi-
cation of a cohort of active workers.
Graph based on mortality data among
a cohort of asbestos product workers
(Enterline, 1965) (reproduced with per-
mission from McMichael, 1976.
(Q American College of Occupation al
and Environmental Medicine, 1976).
281
Chapter 13
282
Interpretation of epidemiological studies
-
p* = px sensitivity + (1 - p) x (1 - specificity)
and thus depends on the true prevalence of the attribute (P) and on the sen-
sitivity and specificity of the measurement method. For example, if p =
0.1%, sensitivity = 90% and specificity = 90%, then p* = 10.1%. This means
that if the prevalence were estimated by the proportion that is classified as
having the attribute by the questionnaire, the estimate would be 10.1%,
compared with a true prevalence of only 0.1%. This misclassification corre-
sponds to a 100-fold overestimation.
It is possible to obtain a corrected estimate of the true prevalence in situ-
ations where the sensitivity and the specificity of the measurement proce-
dure are both known, or may be estimated, by re-arranging the above for-
mula as follows:
p* + specificity - 1
P = sensitivity + specificity - 1
283
Chapter 13
Table 13.3 shows the effects of different levels of sensitivity and speci-
ficity of a measurement procedure on the observed prevalence of a partic-
ular attribute. ln most cases, the observed prevalence is a gross overesti-
mation of the true prevalence. For instance, the observed prevalence at
90% sensitivity and 90% specificity assuming a true prevalence of 0.1 % is
only about one haU of that which would be obtainable if the true preva-
lence were 12.5% (i.e., 10% versus 20%). The bias in overestimation is
severely influenced by losses in specificity (particularly when the true
prevalence is less than 50%). ln contrast, los ses in sensitivity have, at
most, moderate effects on the observed prevalence.
284
Interpretation of epidemiological studies
-
Table 13.5.
Ever Bias due to nondifferential exposure
Never
misclassification: observed exposure
...; Pancreatic (200xO.8)+(1 00xO.1) = 170 (100xO.9)+(200xO.2) = 130 300 status (sensitivity = 80% and specificity
'0. cancer cases
= 90% for both cases and controls).
¡'Controls (150xO.8)+(150xO.1) = 135 (150xO.9)+(150xO.2) = 165
Total 305 295
d Observed odds ratio = (170/130) / (135/165) = 1.60
300
600
these records would be different for those who later developed the outcome
of interest and those who did not. As long as the misclassification is nondif-
ferential, it wil generally dilute any true association between the exposure
and the outcome.
The implications of nondifferential exposure misclassification depend
heavily on whether the study is perceived as 'positive' or 'negative'. This bias
is a greater concern in interpreting studies that seem to indicate the absence
of an effect. ln these circumstances, it is crucial that the researchers consider
the problem of nondifferential exposure misclassification in order to deter-
mine the probabilty that a real effect was missed. On the other hand, it is
incorrect to dismiss a study reporting an effect simply because there is sub-
stantial nondifferential mis classification, since an estimate of effect without
the misclassification would generally be even greater.
It should be noted that nondifferential exposure misclassification leads to
an underestimation of the relative risk (bias towards relative effect of 1) only
if the exposure is classified as a binary variable (e.g.l 'exposed' versus 'unex-
285
Chapter 13
,
L
!
Í
Interpretation of epidemiological studies
'....... Controls 150 200 350 = 100% and specificity =100% for both
: : :,'-,
..:-:_.
_ ,,"'. -"-"''.':-d'''
',' ,'d\ ".',' ,-: : ',:, ::',',,~:. "",'_"._._-:__':"".,-,-,.':,
_ _ _' _ _ _ " _, _ _ _ _: _ _ ,'_ _ _ _ _ _'_ _'-:-.,-",.:,-,-'-'
,':__
,,-.-,_ _ _ _ _ : _ _
.', '-. - .,\-, - '...' ,'--, p- -" -,,' -,,'.:.,'-' '-'-,-.--, -,,' -,--' - ','
;'-",' '- -,""":".'-': '-' '-/'-" '-;""'-'''. - -",.,_...',',.. - -,. '- : " -,". "',' ,- " - ,..-",' ,'_.,'," _, d'. " _ ' . _-'0" "
ficity=.100%). '
ln Example 13.7, patients with breast cancer were more likely to incorrect-
ly report having ever used oral contraceptives than healthy controls, resulting
in a spurious association between oral contraceptives and breast cancer. This
is a particular tye of differential misclassification caIled recall bias. One way
of minimizing recaU bias is to use hospital controls, because then the controls
would generaIly have the same incentive as the cases to remember events in
the past. Subjects should also be unaware of the specific study hypothesis.
ln Example 13.8, there is no biological evidence to suggest that the effect
of vasectomy would be different between Catholic and Protestant men.
Thus, the most likely explanation for the observed findings is that Catholic
con troIs were less likely to report that they had had a vasectomy than
Catholic men with testicular cancer. This differential exposure misclassifica-
tion biased the exposure effect upwards among Catholic men.
Sometimes, it is possible to obtain direct evidence of the presence and
magnitude of differential misclassification. ln Example 13.9, cases, but not
controls, considerably overreported their inabilty to tan after being diag-
nosed with skin melanoma: the proportion of cases reporting inabilty to
tan was 26% (=9/34) in 1982 before skin cancer was diagnosed but 44%
(=15/34) after the diagnosis.
287
.....,
Chapter 13
,
\"-:::-' ','--",. -'" -', -,:-,
.'''. ,_" '_':, '~'-': _' 0,'.''-':':','':,..-""".:'',:::.,::-:__:,,'.-'.:--.,':.':, '-"'_:;':'_:/\'-:':-:",-,...__":"':; , ,:",_-"_, ""-~",:":.'_-'._" n-", ,-"::-, " ';,. ',_ ".. . . .
',:, -""-:.:,,'
,,'. "',,,,",
"" "'.,,,
':,.',,-.:' .'H.':',"",',...",:;
,_' ",' 0,,', __. " , , ' .',
,Exaniplè 13.9. TheNur~e(HeaIthStuyis ..
.'121 700 ternale nurses who aged 3(L
a ssem bIedin1976.dThe. seWòmen h '
','" ',.":',' ,",'"",,':,',:,"":
",,,:',,-,',,,,,,,,,,,"':' ,'Y'".',
'allysincethen,. The 1982q ..,
. skin meIánorna. Á heste~ c
incIuded 34. skinmeIallotnd
tionnaire and 234contrnIsra
historyofcance~ l: .
ti()1Js. (rom Jhe( l~ß
. .. ~onnaire:Th~S€t¡
. ...1991rTh€/esJ1o.~sl?~giv
"'~:,,:::\":.":::':',::
.",,;:,,;:::-,;,:;-::.':
')::,::::~,:"~':':,:~:~:
OR=2.5 (95% CI=1.1-5.7) OR=2.3 (95% CI=1.0-5.0) ;;c-,'~' ,.:"
Tanning abilty
No tan or light tan 9 79 15 77
Medium, deep or darkb 25 155 19 157
OR=0.7 (95% CI=O.3-1 .5) OR=1.6 (95% CI=0.8-3.5)
a Adapted fram Weinstock et al. (1991).
b Baseline category.
288
Interpretation of epidemiological studies
Table 13.12.
New Placebo Bias due to nondifferential outcome
misclassification: observed outcome
Cases 25xO.80 = 20 50xO.80=40 60 status (sensitivity = 80% and specificity
Non-cases 9 975+(25xO.20) = 9 980 9 950+(50xO.20)=9 960 19 940 = 100% for both the new intervention
Total 10 000 10 000 20 000
and placebo groups).
Observed (1 = 20 per 10 000
Observed (0 = 40 per 10 000
Observed risk ratio = (/(0 = 0.50
Observed risk difference = (0 - (1 = 20 per 10 000
Observed prevented fraction = (ro - (1)/(0 = 20 per 10 000 / 40 per 10 000 =
0.50 = 50%
289
Chapter 13
Table 13.13.
Bias due to nondifferential outcome
New Placebo
misclassification: observed outcome
status (sensitivity = 100%; specificity = Cases 25 + (9 975xO.1 0) = 1 022 50 + (9 950xO.1 0) = 1 045 2067
90% for both the new intervention and Non-cases 9 975 x 0.90 = 8 978 9950 x 0.90 = 8 955 17 933
placebo groups).
Total 10000 10 000 20 000
Observed (1 = 1022 per 10 000
Observed (0 = 1045 per 10 000
Observed risk ratio = (1 / (0 = 0.98
Observed risk difference = (0 - (1 = 23 per 10 000
Observed prevented fraction = (ro - (1) / (0 = 23 per 10 000 / 1045 per 10 000 =
0.022 = 2.2%
290
Interpretation of epidemiological studies
ln Example 13.12, the risk of lung cancer among asbestos workers may
have been slightly overestimated because the cause of death is likely to
have been ascertained more carefully among asbestos workers than
among the general population.
291
¡
\....
f
f .~'\~...
i
Í
Chapter 13
ln case,:.ontrol studies:
- Did thecontrols represent the population frQm which the cases arose?
- Was the identification and selection of cases and controls influenced
,bytheir exposure status?
· Were strategies built into the study design to allow assessment of the Iikely
.' direction 'and magnitude of the bias? .
increased risk of lung cancer. This might le ad to the inference that the
occupation is a direct cause of lung cancer. However, without further
analysis, this inference would be invalid if those people employed in the
occupation smoked more heavily than members of the general popula-
tion.
For a variable to be a confounder, it must be associated with the expo-
sure under study and it must also be an independent risk factor for the
disease. ln Figure 13.3, confounding occurs only in example I-smoking
is associated with the particular occupation under study and it is on its
own a risk factor for lung cancer. ln example II, alcoholic cirrhosis of the
liver is an intermediate factor in the causal path between the exposure
(alcohol intake) and the disease (liver cancer). ln example II, alcohol
intake is associated with the exposure under study (smoking) but is not
a risk factor for the disease (lung cancer).
292
¡ \.~':
L .."...
f
í
l
Interpretation of epidemiological studies
A potential confounder is
any factor which is believed to EXPOSURE
~
1) . OUTCOME
have a real effect on the risk of (Occupation)
the disease under investiga-
tion. This includes both factors / (lun9 canee,)
CONFOUNDER
that have a direct causal link
(smoking)
with the disease (e.g., smoking
and lung cancer), and factors
that are good proxy measures
of more direct unknown causes Il) EXPOSURE . VARIABLE OUTCOME
(alcohol intake) (alcoholic cirrhosis (liver cancer)
(e.g., age and social class). of the liver)
Confounding can be dealt
with at the study design level
or, provided the relevant data II) EXPOSURE ) OUTCOME
have been collected, in the (smoking) (Iung cancer)
analysis. Three approaches may tt
be used to control for con- VARIABLE
(alcohol intake)
founding in the design of an
epidemiological study:
293
Chapter 13
fact that the true population exposure effect is large. Conversely, with a
large sam pIe, small effects, which may be clinically and epidemiologically
irrelevant, may easily achieve statistical significance. Confidence intervals
are more informative than P-values because they provide a range of values
for the exposure-outcome association, which is likely to include the true
population effect (usually with a probabilty of 0.95, Le. 95%). They also
indicate whether a non-significant result is or is not compatible with a
true effect that was not detected because the sample size was too small (see
Chapter 6). Sample size calculations should always be done at the design
stage to ensure that the study wil have enough power (Le., its sample size
wil be large enough) to detect the hypothetical effect or enough precision
to quantify it (see Chapter 15).
Secondly, investigators tend to collect data on many potential risk fac-
tors and to perform multiple statistical tests to determine whether any of
these variables is significantly associated with the disease or outcome of
interest. However, the probabilty of obtaining a statistically significant P_
value ¡ust by chance increases with the number of tests performed. It is
particularly important to keep this in mind when the data are interpreted,
particularly if an unexpected association is found that was not hypothe-
sized at the beginning of the study. Such an unexpected association wou
Id
have to be reproduced in studies specifically designed to test it before it
could be accepted as reaL
FinaIly, statistical methods assess only the effects of sampling variation
and cannot control for non-sampling errors such as confounding or bias
in the design, conduct or analysis of a study.
295
f
,,
i
!
Chapter 13
Temporal relationship
This is an essential aspect. For an exposure to be the cause of a dis-
ease, it has to precede its biological onset. This is generally easier to
establish from intervention or cohort studies than from cross-section-
al or case-control studies in which measurements of the exposure and
the outcome are made at the same time.
IIThe criterion of temporality requires that cigarette smoking ante-
date the onset of cancer. Support for this criterion is provided by aIl
the major prospective (cohort) studies in which an enormous num-
ber of initially disease-free subjects were followed over varying time
intervals." (US Surgeon General, 1982)
Biological plausibility
The association is more likely to be causal if consistent with other
biological knowledge (e.g., animal experiments, biological mecha-
nisms). However, this aspect should not be taken too seriously, because
lack of plausibility may simply reflect lack of scientific knowledge or
the fact that human beings are biologically different from animaIs.
IIBenign and malignant tumours have been induced in the larynx
of hamsters by long-term exposure to diluted cigarette smoke.
Attempts to induce significant numbers of bronchogenic carcino-
ma (lung cancer) in laboratory animaIs were negative in spite of
major efforts with several species and strains. Neither rats nor ham-
sters nor baboons inhale cigarette smoke as deeply and as intense-
ly as the cigarette smokers who have provided the data with the
consequences of their 'experiment in the form of clinical evidence
gathered by epidemiologists." (US Surgeon General, 1982)
Consistency
If similar results have been found in different populations using dif-
ferent study designs, the association is more likely to be causal, since
it is unlikely that aIl studies were subject to the same type of bias
and/or confounding. However, a lack of consistency does not exclude
a causal association, since different intensity levels and other condi-
tions may reduce the impact of the exposure in certain studies.
IIMore than 50 retrospective (case-control) studies have reported
smoking patterns (by type and quantity of tobacco smoked, dura-
tion of smoking, and inhalation al practice) in a variety of subjects
with lung cancer (e.g., males and females, different occupational
groups, hospitalized patients, autopsy cases, aIl individuals who
died from lung cancer in an area, nationwide sam pIe of individuals
who died from lung cancer, and different races and ethnic groups).
296
Interpretation of epidemiological studies
Strength
The strength of an association is measured by the magnitude of the
relative measure of effect. A strong association is not easily explained
by potential sources of bias or confounding and hence it is more like-
ly to be causal than is a weak association, which could more easily be
the result of confounding or bias. For example, a relative risk of mag-
nitude 10 is more likely to reflect a true causal association than one of
1.5.
"Prospective (cohort) studies have shown that the death rate from
lung cancer among cigarette smokers is approximately 10 times the
rate in non-smokers ( ... ). To account for such a high relative risk
in terms of an indirect association would require that an unknown
causal factor be present at least 10 times more frequently among
smokers ( ... ) than among non-smokers. Such a confounding fac-
tor should be easily detectable, and if it cannot be detected or rea-
sonably inferred, the finding of such a strong association makes a
conclusion concerning causality more probable." (US Surgeon
General, 1982)
Exposure-response relatíonshíp
Further evidence of a causal relationship is provided if increasing
levels of exposure are associated with increasing incidence of disease.
297
Chapter 13
Specificity
If a particular exposure increases the risk of a certain disease but not
the risk of other diseases, this provides evidence in favour of a
cause-effect relationship. However, one-to-one relationships between
exposure and disease are rare and lack of specificity should not be used
to refute a causal relationship.
Reversibility
When the removal of a possible cause results in a reduced incidence
of the disease, the likelihood of the association being causal is
strengthened. IdeaIly, this should be assessed by conducting a ran-
domized intervention trial, but for many exposures-disease associa-
tions, such randomized trials are impossible in practice.
298
\,, \-i"
l ~ ,-;- ~~ ~ ."
Interpretation of epidemiological studies
Coherence
The putative cause-effect relationship should not seriously conflct
with the natural history and biology of the disease.
Table 13.14 gives an idea of how much the results from different
types of study rnay be affected by bias and confounding.
Data from various studies that have addressed the same question may
be pooled together and re-analysed by using special techniques that take
299
Chapter 13
into account the fact that the original studies may differ in a large num-
ber of respects (e.g., characteristics of the study subjects, methods of
data collection). This approach is particularly useful wh
en the effect of
the exposure is likely to be smaIl, so that a very large sample size wil be
necessary to detect and quantify it precisely (as in Example 13.13).
Meta-analysis is another technique used to combine results from var-
ious studies. It differs from a re-analysis in that there is no access to the
raw data from each individual study. Only the published estima
tes of
exposure effect from each study are available and can be used to gener-
ate a pooled overall result.
It is essential to ensure that practically aIl studies (published and
unpublished) are included in any re-analysis or meta-analysis. This is
because published studies tend to be a biased sample with an overrep-
resentation of those which showed 'positive' findings.
, , -,
. ~'.' - " --..
/',.' -' - 'J--:.:.'
'. -', .' - " ,-.''- ,_-'
.'.' - .__;
_.' -, ,.'_- '-'"- ,.'."
",' ,'_,'
-:' .' -', .',_
-, -,. .-, -.-.'.',~
_..'.'
..'.. .' J3:A large ..... '. ei:ofcohortand ¿ase~ontrol;tudies werecQn-,
.. ' . 'ades toexalline therelatianship of oral contraceptiveuse
. " ,réas.c '. . eirresultswerelargelyinconsistent. To Clarify this issue,'
..datafrQrn mostofthtSe studies'were brought togethet andre~analysed. The
.. .54stúdiesincludeclin this. re-analysis were conducted in 25 ,coùnmes in..
Europe,.. North. Ameria1;,.llsiq, .Australasia", Africa and' Latin America.
,Ti .they in ',297 W011en~ithinvqsive ,breastcancer and
1 ' . ., .Ce~Figure J3A shgwsrelativ~risk~ ofbreast can~
er~1JsersPfcomlJined,'. .' ntrqcêptives (themost ....
.... ates. VåTie4 sÛ:~s .... lyfroifs/ldy tòstudy,
werewideând inc udedthenullhypothesis ,....
matewas slightly . bovéi(1.07Ji¡nd itsconfi-:
dethe nul! hyp , .,. 'svalue. Currellt users of com.: .
, , . stoppt4 lesfth.and 0 yeqrs previously.
,se.mtherislcoflJavin . b¡;.cancq diag-:
ut,this '.' . · l 'lO;ormore'
J "ÒnHô1monal
300
fI-
i
í
Interpretation of epidemiological studies
Figure 13.4.
Relative risk of breast cancer in ever-users compared with never-users of combined oral contraceptives (reproduced with permission from
Collaborative Group on Hormonal Factors in Breast Cancer, 1996, (Ç by The Lancet Ud, 1996; see this paper for full discussion of the meth-
ods and full reference to individual studies).
301
Chapter 13
(a)
Relative risk of breast cancer by total duration Relative risk of breast cancer by age at
of use of combined oral contraceptives first use of combined oral contraceptives
Total duration Age at
of use RRa~SD Cases/Contrais first use RRa~SD Cases/Contrais RRa & 99% Ci
a Relative risk (given with 99% CI) relative ta never-users, stratified by study, age at diagnosis, parity and, where appropriate, the age of a woman al the
birth of her first child and the age when her risk of conception ceased. Each analysis includes aggregated data from ail studies included in Figure 13.4.
The area of each square is proportional ta the amount of stalistical information and Cis are drawn as white Iines when they are so narrow that they lie
entirely within the width of the square.
Figure 13.5.
Relative risk of breast cancer for various indices of timing of use of combined oral contraceptives (reproduced with permission from
Collaborative Group on Hormonal Factors in Breast Cancer, 1996, (Q by The Lancet Ltd, 1996; see this paper for full discussion of the meth-
ods and full reference to individual studies).
302
-~~~~~ "'~--~~~~ "õ Interpretation of epidemíological studies
Further reading
,::."...:....:: ..:,:..: ~:....,,',.:.~, :~,:.:\P:Q~~:~.3..2:7.~ay)s;slié~..~ "..... ..:';.. ..:....... .:,: :".'" ç~;
· Theinterpretation of findings trom any epIdemiologicalstudy requires consider- * The aspects that need to be
ationof bias.,confounding and chanGe.
considered when assessing
whether a particular exposure-
· Bias tendstoleadto an incorrect estimate.otthe effect Of anexposure on the outcome relationship is likely to
outcome. ofinterest. There are two main types ofbias: be causal were presented by
Bradford Hill in a paper published
Selection bias occurs when there is a difference between.the character- in 1965.
isticsof the people who participated. in the study and the characteristics
of those who did not. Selection'. bias can be a major problem in
case-contral studies, but can also affect cross..sectional studies and, to
a lesser extent, cohort studies and randqmii:ed trials.
303
Chapter 14
Dealing with confounding
in the analysis
ln the previous chapter we discussed briefly how confounding could be
dealt with at both the design stage of a study and during the analysis of
the results. We then mentioned that there are two main statistical proce-
dures that we can use in the analysis: stratification and regression modellng.
ln this chapter we wil discuss the se two approaches in more detaiL
Obviously, these techniques can be applied only if data on potential con-
founding factors are available. Thus, potential confounding variables have
to be identified at the design stage of the study to ensure that valid infor-
mation on them is collected.
Table 14.1.
Results of a case-control study on
smoking and ovarian cancer: hypothet-
¡cal data.
ln Example 14.1, women with ovarian cancer had a much lower preva-
lence of smoking (24/60 = 40%) compared with the controls (58/98 =
59%). This suggests that smoking protects against ovarian cancer (odds
ratio (OR) = 0.46). As discussed in the previous chapter, there are several
possible explanations for this finding:
305
\\.._-
Chapter 14
(i) Bias: the observed odds ratio of 0.46 does not accurately repre-
sent the true odds ratio because of either selection or measurement
bias.
(ii) Chance: the observed association between smoking and ovarian
cancer arose by chance. The 95% confidence interval around the
observed odds ratio is equal to 0.23-0.93 and the X2 test yields P=0.02.
Thus, chance is an unlikely explanation for the finding.
(iii) Confounding: the observed odds ratio of 0.46 is due to the effect
of another variable. For example, it may be that women who smoked
were different in other respects from non-smokers and less likely to
develop ovarian cancer because of this, rather than because of smoking.
(iv) Causation: smoking reduces the risk of ovarian cancer and the
95% confidence interval indicates how precisely the sample estimate
corresponds to the true effect in the population.
ln Example 14.11 it is possible that the association between smoking
and ovarian cancer arase because of the confounding effect of other fac-
tors such as oral contraceptive use. The results shown in Table 14.1 are
for aIl women combined regardless of their history of oral contraceptive
use. To assess whether oral contraceptive use is a confounder, we need
to look at the association between smoking and ovarian cancer sepa-
rately for oral contraceptive users and never-users. This is shown in
Table 14.2.
1
,f \...".
f
Í
Dealing with confounding in the analysis
the population from which the cases were drawn and we need to
assess that association in the general population. ln a cohort or inter-
vention study, the association would be looked at by constructing a
similar table, replacing the number of contraIs with person-years at
risk if the apprapriate measure of effect was a rate ratio or numbers of
persons at risk at the start of the follow-up if the measure of effect was
a risk ratio.
2. Oral contraceptive use is considerably lower among ovarian can-
cer cases th an among controls. The data from Table 14.2 can be
rearranged so that smoking is ignored (Table 14.4).
Only 32% (=19/60) of the women with ovarian cancer were oral
contraceptive users, whereas 63% (=62/98) of the contraIs were users.
Since oral contraceptive use in these data was associated with both
the exposure (smoking) and the outcome of interest (ovarian cancer),
it acted as a confounding factor. As a result, when the data for both
users and never-users were combined, the result suggested an associa-
tion between smoking and ovarian cancer far stronger than really
existed (positve confounding). ln other situations (as in Example 14.2;
see next section), combining strata in the presence of a confounder
may mask an effect that really exists (negative confounding), or even
show an effect in the opposite direction to the true one.
307
Chapter 14
sis aIl the subjects in our study sample. This analysis provides a crude esti-
mate of the effect of the exposure on the outcome of interest. The next log-
ICal step is to divide our study sample into several subgroups or strata,
defined by a potential confounding variable, to see if the results are consis-
tent across the strata. This approach is very informative, as it describes how
the effect of the exposure on the outcome of interest varies across subgroups
of subjects with different characteristics. We can simply report the stratum-
specific effect estimates and their confidence intervals. Each of these stra-
tum-specific estimates is supposed to be homogeneous in relation to the
potential confounding variable and therefore they are unconfounded.
UsuaIly, however, we are not much interested in the stratum-specific
results per se and would rather have a single overall estimate. ln other
words, we wou Id like to be able to calculate a summary effect estimate
which, in contrast to the crude estimate, would take into account the con-
tes can be
founding effect of the stratifying variable. Such adjusted estima
calculated by pooling results across strata. But even if the true effect is the
same in aIl strata, we would expect our estimates to differ because of ran-
dom variation. Pooling takes this into account by giving greater weight to
effect estimates from larger strata. 1t involves calculating a weighted average
of the individual stratum-specific estimates by choosing a set of weights
that maximizes the statistical precision of the adjusted effect estimate.
There are several alternative weighting procedures which achieve precise
summary effect estimates. ln this section, we concentrate on a procedure
derived by Mantel and Haenszel which is widely used and relatively sim-
ple to apply.
of the study in a way that removes the confounding effect of oral contra-
ceptive use. The Mantel-Haenszel odds ratio, denoted ORMH, gives a weight-
ed average of the odds ratios in the different strata, where those from larger
strata are given more weight.
To calculate the ORMH, we start by constructing 2 x 2 tables of exposure
by outcome for the separate strata of the confounder, as ilustrated in
Table 14.2. The ORMH can then be obtained by applying the following for-
mula:
aiddNi + azdz/Nz
o RMH =
biei/Ni + bzez/Nz
308
Dealing with confounding in the analysis
Thus, the odds ratio for smoking adjusted for oral contraceptive use is 0.95.
This adjusted odds ratio contrasts with the crude odds ratio of 0.46
obtained from Table 14.1. Adjusting for oral contraceptive use gives an
odds ratio much closer to unit y, which me ans that the protection afford-
ed by smoking, if any, is far less strong than the initial result led us to
think.
The above formula can easily be extended to more than two strata, by
summing both the numerator and the denominator over aIl strata:
Iaid/Ni
o RMH =
Ibic¡/N¡
ln this formula, L means sum and the subscript i stands for the sub-
scripts 1, 2, 3, ..., which represent each of the strata.
We can calculate a confidence interval around the OR¡.m and a
Mantel-Haenszel X2 test by using the formulae given in Appendix 14.1,
Section A14.1.1. ln our ovarian cancer example, the 95% confidence
interval is 0.42-2.16. The X2 is equal to 0.016 with one degree of free-
dom, which corresponds to P = 0.93. Thus, after adjusting for oral con-
traceptive use, the effect of smoking on ovarian cancer is no longer 'sta-
tistically significant.
309
,
l '~..
j
1
1
Chapter 14
Non-smokers
Exposure ta chemical substance Total
Yes No
Lung cancer Yes 400 (a2) 200 (b2) 600 (n12)
310
Dealing with contounding in the analysis
l.û¡mo¡/N¡
RMH =
'Lb¡m¡¡/N¡
Thus,
(80 x 16000)/20000 + (400 x 80 000)/160 000 264
RMH = = 2.0
(160 x 4 000)/20 000 + (200 x 80 000)/160 000 132
Table 14.8.
Results tram a cohorl study on Pap
smear testing and cervical cancer:
hypothetical data.
311
Chapter 14
ln Example 14.3, the crude rate ratio and its confidence interval are
consistent with a decrease in the incidence of cervical cancer among
women who reported in the initial questionnaire having ever had a Pap
smear test. Other studies have shown that there is a socioeconomic gradi-
ent in the incidence of cervical cancer, with women from poor socioeco-
nomic backgrounds being at higher risk. Thus, socioeconomic factors
might have confounded the association between Pap smear testing and
cervical cancer if, for instance, women from a high social background were
more likely to visit their doctors and had a Pap smear as part of their reg-
ular medical examination. To clarify this issue, we need to examine the
relationship between Pap smear testing and cervical cancer separately for
women from different socioeconomic backgrounds. This is shown in Table
14.9, where a woman's educationallevel is used as a marker of socioeco-
nomic status.
Table 14.9. High educationallevel
Hypothetical cohort study on Pap Pap smear Total
smear testing and cervical cancer
Ever Never
described in Example 14.3: results
stratified by women's educationallevel. Cases 13 (a,) 697 (b,) 710 (n,)
Person-years at risk 38 346 (Yii) 828 149 (Yo,) 866 495 (Y1)
Rate per 100 000 pyrs 33.9 (r11) 84.2 (r01) 81.9 (r1)
Rate ratio = OAO; 95% confidence interval = 0.23-0.69
Law educationallevel
Pap smear Total
Ever Never
Cases 4 (a2) 427 (b2) 431 (n2)
Person-years at risk 32 838 (Yd 690 552 (Y02) 723 390 (h)
Rate per 100 000 pyrs 12.2 (rd 61.8 (r02) 59.6 (r2)
Rate ratio = 0.20; 95% confidence interval = 0.08-0.54
YAyo¡/y¡
RRMH =
i'b¡ydYi
312
-\.'1...
Dealing with confounding in the analysis
313
l
¡
!
f
1
Chapter 14
Two other aspects of stratification should be noted. First, factors that are
on the causal pathway between an exposure and a disease should not be
regarded as confounding the association between the exposure and the
outcome. Controllng for a factor that is on the causal pathway leads to
underestimation of the strength of the effect of the exposure on the out-
come under study. OccasionaIly, a variable that lies on the causal pathway
may be adjusted for in the analysis if we want to assess whether the effect
of the exposure on the outcome under study is entirely mediated through
that intermediate variable or whether there may be other independent
biological mechanisms. For instance, if we believe that human papilo-
mavirus infection is on the causal pathway between number of sexual
partners and cervical cancer, the association with number of sexual part-
ners should disappear after adjusting for HPV infection. If the effect does
not disappear completely, it would suggest that the effect of number of
sexual partners on cervical cancer may be mediated through other biolog-
ical mechanisms not directIy related to HPV infection. ln practice, this rea-
soning may not be so straightforward because of errors in the measure-
ment of the intermediate variable.
Secondly, it is important to note that stratification assumes that each
stratum is homogeneous in relation to the confounding variable. This
assumption depends on both the validity of the data on the confounding
variable (see Section 13.2) and the way strata are defined. For instance, if
we control for age through stratification, this is better achieved if the stra-
ta are relatively narrow. Stratification into very broad age groups (e.g.,
0-45 and 46+ years) is unlikely to be fully effective since, within each stra-
tum, there are likely to be substantial differences in the age distribution of
cases and controls (in a case-control study) or exposed and unexposed (in
a cohort or intervention study).
314
Dealing with confounding in the analysis
315
Chapter 14
however, these tests are not very powerful. Epidemiological studies, unless
they are specifically designed to do this, rarely have enough statistical
power to detect interactions (see Section 15.4.4). Thus, more is usually
gained by visual inspection of the size and pattern of the effect estima tes
across strata than from tests for interaction (effect modification).
Table 14.11.
Results from a case-control study on
smoking and HPV infection: hypotheti-
cal data.
ln Example 14.5, the cru de odds ratio is 1.54. The 95% confidence
interval and the P-value suggest that chance is an unlikely explana-
tion of this finding. Thus, women who smoked were more likely to be
HPV-positive than those who did not.
Number of sexual partners is a weIl known risk factor for HPV infec-
tion and it may have confounded the association between smoking
and HPV infection. Table 14.12 shows the association found between
smoking and HPV stratifed by reported number of lifetime sexual
partners (categorized as .. 2 partners and ¿ 2 partners).
Examination of the stratum-specific odds ratios (and their confi-
dence intervals) suggests that the effect of smoking on HPV infection
in women who reported less than two sexual partners is similar to the
316
Dealing with confounding in the analysis
ORMH =
'La¡d¡! Ni
'Lbic¡! Ni
(31x437)/670 + (11x44)/89
(124x78)/670 + (22x12)/89
= -
25.66
17.40
= 1.47
317
Chapter 14
-
X2 test for heterogeneity = 7.37 on 5 d.f.;P = 0.20
Laid;! Ni 25.85
ORMH = = = 1.47
Lbic;! Ni 17.55
318
Dealing with confounding in the analysis
ORMH = - -
X2 test for heterogeneity = 12.44 on 10 d.f.;P = 0.26
'ia¡d¡! N¡
'ib¡c¡! N¡
=
24.60
17.46
= 1.41
There are many types of regression mode!. The choice of any particu-
lar model depends on the characteristics of the outcome variable (i.e.,
continuous or categorical) and on the way it is mathematically related to
the explanatory variables. The simplest mathematical model we could
use has already been introduced in Section 11.2.1 to describe the rela-
tionship between two quantitative variables.
A discussion of these models and the assumptions underlying them is
beyond the scope of this chapter. However, these modellng techniques
are commonly used in epidemiological studies and, therefore, in the rest
of this chapter we wil try to ilustrate how these techniques relate to the
Mantel-Haenszel method, to allow the reader to understand and inter-
pret results from published work where they have been used.
319
Chapter 14
Thus, after adjusting for number of sexual partners, the effect of smok-
ing on HPV infection became smaller (1.47 versus 1.54). This result is sim-
ilar to that obtained earlier wh en we used the Mantel-Haenszel technique
to control for the effect of number of sexual partners (Table 14.12). But in
contrast to this technique, this regression model also gives us the odds
ratio for number of sexual partners adjusted for smoking:
Odds ratio for ;:2 sexual partners versus -:2, adjusted for smoking
= 1.90 (95% CI = 1.9-3.03)
320
\..
"
Dealing with confounding in the analysis
Odds ratio for 2 2 sexual partners versus -: 2, adjusted for smoking and age
= 1.95 (95% CI = 1.9-3.17)
Odds ratio for age 20-24 versus age dO, adjusted for smoking and partners
= 4.14 (95% CI = 2.12-8.12)
Odds ratio for age 25-29 versus age dO, adjusted for smoking and partners
= 3.58 (95% CI = 1.77-7.24)
Odds ratio for age 30-34 versus age -:20, adjusted for smoking and partners
= 3.01 (95% CI = 1.47-6.15)
Odds ratio for age 35-44 versus age -:20, adjusted for smoking and partners
= 2.18 (95% CI = 1.07-4.46)
Odds ratio for age 45+ vers us age dO, adjusted for smoking and partners
= 1.49 (95% CI = 0.67-3.33)
321
Chapter 14
Table 14.16. Variables adjusted for Cervical cancer rate 95% confidence
Hypothetical cohort study on Pap ratio for Pap smear use interval
smear testing and cervical cancer
None 0.32 (crude) 0.20-0.52
described in Example 14.3. Results
obtained using the Mantel-Haenszel Educationalleveia 0.32 0.20-0.52
technique. Educationallevel 0.40 0.25-0.66
and marital statusb
Educationallevel, marital status 0.43 0.27-0.72
and age at tirst intercoursec
a Categorized as 'Iow educationallevel' and 'high educationallevel'.
b Categorized as marital status 1 =married; 2=single; 3=divorced/widowed.
C Categorized as age at tirst intercourse 1 =-:18 years; 2=18-22 years; 3=22+ years.
Cervical cancer rate ratio for high versus low educationallevel adjusted for Pap smear
use = 0.73 (95% CI = 0.65-0.82)
ln the last model (modeI4) shown in Table 14.17, we included Pap smear
use, educationallevel, marital status and age at first intercourse as explana-
tory variables. The Poisson estima te of the rate ratio for Pap smear use
adjusted for educationallevel, marital status and age at first intercourse is
0.46 (95% CI = 0.29-0.75), similar to that obtained with the
Mantel-Haenszel method (RRMH = 0.43; 95% Ci = 0.27-0.72) (Table 14.16).
But with the Poisson regression, we also obtained the following additional
rate ratios:
Cervical cancer rate ratio for high versus low educational level adjusted for Pap
smear use, marital status and age at first intercourse = 0.77 (95% CI = 0.68-0.87)
322
Dealing with contounding in the analysis
Model3
constant 0.005 0.004-0.005
Pap smear use8 0.41 0.25-0.66
Educational leveib 0.74 0.66-0.84
Marital status2C 2.68 2.28-3.15
Marital status3 1.89 1.61-2.21
Model4
constant 0.008 0.006-0.009
Pap smear use8 0.46 0.29-0.75
Educationalleveib 0.77 0.68-0.87
Marital status2C 2.68 2.27-3.15
Marital status3 1.60 1.36-1.87
Age at tirst intercourse2d 0.52 0.46-0.59
Age at tirst intercourse3 0.13 0.09-0.19
8 Categorized as 'never' (baseline) and 'ever'.
b Categorized as 'Iow education
al level' (baseline) and 'high educationallevel'.
C Categorized as marital status 1=married (baseline), 2=single, 3=divorced/widowed.
d Categorized as age at tirst intercourse 1=.c 18 years (baseline), 2=18-22 years, 3=22+ years.
Cervical cancer rate ratio for single versus married women adjusted for Pap smear use,
educationallevel and age at first intercourse = 2.68 (95% CI = 2.27-3.15)
Cervical cancer rate ratio for divorced/widowed versus married women adjusted for Pap
smear use, educationallevel and age at tirst intercourse = 1.60 (95% CI = 1.6-1.87)
Cervical cancer rate ratio for age at tirst intercourse 18-22 versus -( 18 years adjusted for
Pap smear use, educationallevel and marital status = 0.52 (95% CI = 0.46-0.59)
Cervical cancer rate ratio for age at first intercourse 22+ versus -( 18 years adjusted for
Pap smear use, educationallevel and marital status = 0.13 (95% CI = 0.09-0.19)
323
....
"
Chapter 14
14.7 Conclusions
ln summary, the Mantel-Haenszel method is a very useful technique to
adjust for confounders, and this approach is often adequate for data with
few confounders. However, in order to adjust simultaneously for several
confounders, regression modellng methods may be necessary.
It is important, however, to stress that any analysis should start by using
the Mantel-Haenszel method to obtain preliminary crude effect estima tes
and effect estima tes adjusted for each confounder separately. The cross-
tabulations used for stratification in this technique allow the investigator
to observe most of the important relationships and interactions that are
present and to detect errors and inconsistencies in the data that might not
otherwise be evident.
Regression models can then be used in a second stage of the analysis to
adjust simultaneously for several confounders. One of the main disadvan-
tages of regression modellng is that we lose sight of the data, so that it is
often regarded as a 'black box' approach. Statistical modellng should not
be used by people who are not familar with it and who do not understand
the assumptions upon which it is based.
324
Dealing with confounding in the analysis
Further reading
* ln this chapter, we have pre-
sented formulae to calculate
Mantel-Haenszel rate ratios, risk
ratios and odds ratios. Formulae
to calculate adjusted estimates of
risk and rate differences can be
found in Greenland & Robins
(1985).
325
Appendix 14.1.
Confidence intervals and
statistical tests for adjusted
relative measures of effect
Note that there are several methods to calculate confidence intervals
for adjusted relative measures of effect, which may yield slightly different
values from those obtained here. These calculations can easily be per-
formed by statistical computing packages such as EPI INFO, STATA or
EGRET.
I(b¡c¡IN¡)2 V¡
SE (ln ORMH) =
(Ib¡c¡IN¡)2
Thus, in the ovarian cancer study (Example 14.1; Table 14.2), we have
Thus,
95% CI (ln ORMH) = (ln 0.95) :t 1.96 x 0.42 = - 0.05 :t 1.96 x 0.42 = - 0.87 to 0.77
327
,f
f
f
Appendix 14.1.
O(ai) = ai
E(ai) = l1iimi¡/Ni
V(a¡)=niiI10¡mi¡mo¡/(N?(N¡ - 1))
We then sum each of these quantities over aIl the strata. ln Example
14.1 (Table 14.2), we ob tain
IO(a¡) = 9 + 15 = 24
We would expect the difference between our observed and expected val-
ues to be small if the null hypothesis were true. To test whether the dif-
ferences obtained are greater than would be expected by chance, we cal-
culate
and obtain a P-value by referring our result to the X2 distribution with one
degree of freedom (d.f.).
ln our example, X2 = (24 - 24.30)2/5.67 = 0.016 on 1 d.f. This gives
P = 0.93, from which we conclude that after adjusting for oral contra-
ceptive use, there is no evidence of any association between smoking
and ovarian cancer. So ORMH = 0.95 is not statistically significantly dif-
ferent from 1.
328
Confidence inteivais and slalislical lests for adjusled measures of effecl
.
I(nlimi¡mO¡ - a¡b¡Ni)/N¡2
SE (ln RMH) =
(Ia¡mo¡/ Ni) (Ibim 1 ¡/ Ni)
Ia¡mo¡/N¡ = 264
Ib¡m¡¡N¡ = 132
I(nl¡ml¡mOi-a¡b¡N¡)/N? = (240x4000x16 000 - 80x160x20 000)/20 0002 +
Thus,
95% confidence interval (ln RMH) = (ln 2.0) :: 1.96 x 0.073 = 0.547 to 0.833
329
~\
\."""
Appendix 14.1.
IV(ai)
SE (ln RRMH) =
(IaiYOi/ Yi) (IbiYi ¡/ Yi)
IaiYO;!Yi = 16.24
IbiYu/Yi = 50.23
330
Confidence intervals and statistical tests for adjusted measures of effect
where the summation is over aIl strata. The value calculated should be
looked up in tables of the X2 distribution with one degree of freedom.
Thus, in Example 14.3,
I,O:aJ = 13 + 4 = 17
I,V(a¡) = 48.71/ which was obtained above for the calculation of the confidence
interval.
331
Chapter 15
Size of a study
15.1 Introduction
It is important to ensure at the design stage that the proposed number
of subjects to be recruited into any study wil be appropriate to answer the
main objective(s) of the study. A smaIl study may fail to detect important
effects on the outcomes of interest, or may estimate them too imprecise-
ly, no matter how good its design may be in other respects. A study larger
than necessary, while less common in practice, may waste valuable
resources.
Table 15.1.
Number of patients with breast cancer
still alive one year after entry into the
trial by type of treatment administered:
hypothetical data.
ln Example 15. l, the X2 test of the difference between the se two pro-
portions gives a value of 2.65, which corresponds to P? 0.10. Thus the dif-
ference between the two proportions could easily have arisen by chance.
However, we cannot conclude from this that there is no true difference
between the treatments, since the 95% confidence interval for the difer-
ence between the proportions of patients stil alive one year after entry
into the trial is - 2% to +22%. Therefore the data from this trial are con-
333
Chapter 15
334
1
f
f
Í
l
Size of a study
cance' level (usually 0.05) if the true magnitude of the effect is as antici-
pated. However, as discussed in Section 6.3, there is a close link between
P-values and confidence intervals. Therefore, power can also be interpret-
ed as the probabilty of obtaining an estimate whose confidence interval does
not include the value stipulated by the null hypothesis. The null hypothesis
states that the exposure has no effect on the outcome of interest corre-
sponding to a value of zero, if the exposure effect is measured on an
absolute scale (e.g., risk or rate difference), or one, if measured on a ratio
scale (e.g., risk or rate ratio).
Figure IS.1(a) ilustrates the relationship between
the null hypothesis value, the anticipated effect esti- (a)
mate and its confidence interval when the exposure
is associated with an increase in the occurrence of the
: kSE :
J J ~J
ii ¡SE
J i
¡SE
l
J 1J( )1.
outcome of interest. For the study to have appropri- ) 1
confidence interval) and upon the significance (confidence) level chosen (j. (a) an increase or (b) a decrease in
the occurrence of the outcome of inter-
For a 95% confidence interval, ¡ would be equal to 1.96 (Table 15.2); that est (adapted from Clay ton & Hills,
1993).
is, the confidence interval wil extend 1.96SE to each side of the sam
pIe
estimate. For a 99% confidence interval, ¡ wil be 2.576 and, therefore, the
confidence interval wil be wider. The wider the confidence interval, the
lower the power of a study of a given size.
Suppose that the study were repeated several times. The effect estimates
obtained each time and their confidence intervals would differ because of
sampling variation. If the effect estimates obtained each time were plot-
ted, we would obtain a Normal sampling distribution with a standard error
of SE. Similarly, if the lower limits (or upper limits, if the exposure is pro-
tective) of each of the confidence intervals were plotted, we would obtain
a Normal distribution, with the same standard error SE, centred around
the anticipated value of the lower (or upper) limit. The power of a study is
335
Chapter 15
Significance level j the probabilty that the lower limit of the confidence interval would fall
0.10 1.645 above the null hypothesis value (or the upper limit would fall below it, if
0.05 1.960 the exposure is protective). This probability depends upon the number of
0.01 2.576 standard errors (k) between the null hypothesis and the anticipated posi-
Power k tion of the lower limit (or upper limit, if the exposure is protective) of the
confidence interval of the anticipated effect estima te (Figure 15.1). It can
0.95 1.645 be shown mathematically that if k is equal to 1.645, the study wil have
0.90 1.282 95% power (1àble 15.2). ln other words, if the study were to be conduct-
0.75 0.674 ed repeatedly, we would expect only 5 out of 100 resulting 95% confi-
0.50 0.0 dence intervals to include the null hypothesis value, if the true magnitude
.:0.50 .:0 of the effect is as anticipated. When the anticipated location of the lower
(or upper) confidence limit is exactly at the null hypothesis, so that k = 0,
Table 15.2.
the power is 0.50 and there is an even chance of obtaining a significant
Values of k and j for diHerent signifi- result. If k 0( 0, the power will be less than 50%. ln general, a power of less
cance levels and powers. than 80% is regarded as unacceptable.
1. The magnitude of the anticipated effect (Le., the distance between the
null hypothesis value and the anticipated effect). The greater the
effect, the higher the power to detect it as 'statistically signifcant for
a study of a given size.
2. The width of the confidence interval (jSE), which determines the position
of the lower limit (or the upper limit, if the exposure is protective). The
wider the confidence interval, the lower the power of a study of a given
size. This in turn depends on:
(a) The study size. The bigger the study, the smaller the standard
error (SE) and, therefore, the narrower the confidence interval.
It is useful to construct power curves to show how the power varies with
the study size for different signifcance levels and different magnitudes of
the anticipated effect. Figure 15.2 shows sorne examples of such curves for
the breast cancer trial described in Example 15.1.
ln most real situations, researchers have a very good idea of the number
of eligible subjects they wil be able to recruit into their study. This num-
ber is usuaUy determined by availabilty of eligible subjects, logis tics of
recruitment, costs, etc. ln these circumstances the relevant question is not
'How large should the study be?' but 'Is the avai/able number of subjects enough
to pravide a clear answer to the study objectives?'. To answer this last question,
we need to estimate the power of the study with the proposed number of
336
Size of a study
subjects. If these calculations reveal that the power wil be too low, it wil
be necessary to estimate by how much our sam pIe size needs to be 100
increased to ensure that the study wil achieve the desired power. This is
80
the approach suggested by Clay ton & Hils (1993), which we wil follow in
the rest of this section.
t:60 -a P=0.10
Qi
+ P=0.05
;: ~ P=0.01
15.2.1 Comparing two proportions (prevalences or risks) & 40
To calculate the power of our breast cancer trial (Example 15.1) to detect
a 10% difference in survival between the two treatments, we need to cal- 20
culate first the SE of the difference between the two proportions of women
o
stil alive by the end of the first year. As we saw in Section A6. 1 .2, the SE o 100 20 30 40 50 60 700 80 90
Study size
of the difference between two proportions can be estimated, approxi-
mately, as
Figure 15.2.
Power for detecting a one-year sur-
SE(Pi-PJ=~ Pi2~1-P) + Po2~1-Po) vival difference of 10% when baselíne
survival is 70% (equivalent ta 14%
increase in survival) for various sample
sizes and significance levels.
Thus, in our example,
~ 80 + 70 .
SE =,1 0.802(1-0.80) 0.702(1-0.70) = 0 061
Figure lS.l(a) shows that the distance between the anticipated effect
and the null hypothesis value (in our example, 0.10-0) is the sum of the
two components, one deriving from the width of the confidence interval
(jSE) and the other from the distance between the anticipated position of
the lower limit of the confidence interval and the null hypothesis (kSE).
Hence,
0.10 = í SE + k SE
k = - 0.32
337
\.
-,,\~.
Chapter 15
~
,'( ,,
: kSE:
,
,
¡SE
) I(
¡SE
)1
1.96 X SE + 1.645 X SE = 0.10
The initial sample size was 100 in each group. To ensure that the study
wil have 95% power to detect a 10% difference in survival at the 5% sig-
nificance level, we need to multiply this original sample size by 4.8. Thus,
we need to enrol 480 subjects in eaeh treatment group.
As we saw in Section A6. 1.3, the significance test is the same regardless
of the type of measure of effect (ratio or difference) used to compare the
two groups. Power calculations are based on the significance test that wil
be performed at the end of the study and, as a logical consequence, simi-
laI' sample size estimates would be obtained if the ealeulations were based on the
ratio of the two proportions (80%/70%) rather than on their difference
(80%-70%). The approach would be similar to the one followed above,
except that the calculations would be based on the SE of the ratio of the
two proportions rather than on the SE of their difference. Since the confi-
dence interval around a ratio of proportions is asymmetric, ta king only
values from zero to infinity (Section A6.1.2), we first convert the estimat-
ed survival risk ratio into its naturallogarithm (denoted by ln):
338
Size of a study
The distance between the logarithm of the anticipated effect and the
logarithm of the null hypothesis value of 1 is equal to
Thus,
0.134 = ; x SE + k x SE
The value of the required SE to ensure that the study wil have 95%
power (k = 1.645) to detect a risk ratio of 1. 1 4 at the 5% significance level
(j = 1.96) wil be
Therefore,
This value is exactly the one we obtained before when the calculation
was based on the difference between the two proportions rather than on
their ratio. Thus, the sample size required to detect a 10% increase in sur-
vival from 70% to 80% is equivalent to the sample size required to detect a
risk ratio of 80%/70% = 1.14.
339
Chapter 15
Since the anticipated rate ratio (RR) is equal to Z.O, the expected nu m-
bers of lung cancer cases among exposed and unexposed workers are
Table 15.3.
Results from a hypothetical cohorl
study.
We can now complete Table 15.3 with the results we expect to obtain
from this cohort study if our assumptions are correct (Table 15.4).
Table 15.4. Exposure Total
Anticipated results fram the proposed
cohorl study iIustrated in Example Yes No
15.2. No. of cases 80 (a) 60 (b) 140 (n)
Person-years at risk 80 000 (Y1) 120 OOO(Yo) 200000 (y)
Rate per 100 000 pyrs 100 (ri) 50 ((0) 70 (r)
The number of SEs between the logarithm of the anticipated rate ratio
(ln (Z.O) = 0.693) and the logarithm of the null hypothesis value (ln (1) =
0) is
340
'\.
~..'\..
Size of a study
í + k = 0.693/0.171 = 4.05
c = 200 x 0.335 = 67
d = 200 - 67 = 133
Table 15.5.
Yes Results trom a hypothetical case-con-
No
trol study.
Cases a b n1
Controls c d no
Total mi mû N
341
,l
j
1
Chapter 15
Since we anticipate an odds ratio equal to 0.5, we can calculate the odds
of exposure among the cases as
Thus,
a = 0.25 x b
b = 160
a = 200 - 160 = 40
0.693 = 1.96 x SE + k x SE
342
Size of a study
0.238 = k x 0.232
k = 1.03
Table 15.7.
Alive one year after Yes Anticipated results of a hypothetical
384 (a) 336 (b) 720 (n1) trial to assess the value of a new treat-
entry into the trial ment on the one-year survival from
No 96 (e) 144(d¡ breast cancer. Anticipated effect = 10%
240 (no)
Total 480 (m1)
(= 80%-70%) difference in survival;
480 (mo) 960 (N) significance level = 0.05; power = 0.95;
Pi = 384/480 = 80% sample size = 480 women in each
Po = 336/480 = 70% treatment group.
1 1 Figure 15.4.
1 1
1 kSE 1 Anticipated effect and its 95% confi-
1 1
1
1( .,1 dence interval for the hypothetical
1
1
1
1
breast cancer trial described in
1
1
1
1
¡SE ¡SE Example 15.4: anticipated effect =
1
1
1'(
) 1 10% (= 80%-70%) difference in sur-
1
1 .1 (
1
1
1
vival; significance level = 0.05; power
1
1
1
1 95% confidence interval = 0.95; sample size = 480 women in
1 1
1 1 each treatment group.
1 1
1 1
1 1
343
Chapter 15
p=a/n
As we saw in Section 6.1, this estimate is subject to sampling error, but the
95% confidence interval will give a range of values within which the true pop-
344
Size of a study
SE(P) = ~ pZ(~-p)
w = j x SE
For a 95% confidence interval, j = 1.96, that is the interval extends 1.96
standard errors either side of the estimate (w = 1.96 x SE).
Hence, we can estimate the prevalence of oral contraceptive use (p) with
a pre-defined degree of precision by choosing an appropriate sample siie
(n). We must first guess the value of p. Suppose that statistics on oral con-
traceptive sales indicate that the prevalence of use is about 50% and we
want to estimate it to within lS%. Thus
p = 0.50
w = j x SE = 0.05
1.96 x SE = 0.05
Since p = 0.5, we can estimate a from the formula for the SE:
a = 192
Thus, we need to enrol 384 womeii into the study. When planning a
study, it is a good idea to find out what sample siie wil be required for var-
ious levels of precision (Table 15.8 and Figure 15.5).
345
¡
i
¡
,,
ì
¡
f
!
Chapter 15
Table 15.8. Thus, to estimate a true prevalence of oral contraceptive use of 50%
Sample sizes required to estimate a
true prevalence of 0.50 with 95% confi- within :t1% (Le., from 49% to 51%), we would need ta recruit 9612
dence intervals of differenl widths (:tw). women.
It is also important to calculate the required sampIe size for different
Width (:!w) Sample size (n) values of the anticipated prevalence p. As we can see in Table 15.9, the
0.01 9612 sample size required does not vary mu ch for values of p between 0.3 and
0.02 2403 0.7, being greatest when p = 0.50. Thus, to be on the safe side, we can set
0.03 1068 p = 0.50 and obtain the maximum sam pIe size (n) required.
0.04 600
0.05 384
0.06 266
70
0.07 196
0.08 150
0.09 118
60
0.10 96
0.15 43 ~
0
CI
0
c:
50
CI
¡¡
,.CI l l
èi 1i
40
Figure 15.5.
Anticipated 95% confidence intervals
for a true prevalence of 0.50 for vari- 30 r 1 i i i i i
ous sample sizes. o 500 1000 1500 2000 2500 3000
Study size
Table 15.9. ln Example 15.6, we plan to take a random sample of individuals from the
Sample size required to ensure that study population. Thus, for a 95% confidence level (j = 1.96)
the 95% confidence interval for differ-
ent levels of prevalence (p) will extend
W = 0.05 to each side of the sample w = 1.96 x SE
estimate. 5 pei 10 000 pyiS = 1.96 x SE(r)
346
Size of a study
r
SE(r) =
,fa
where ris the estimated rate and a is the number of cases that occurred
during the observation period. Thus, in our example,
50 per 10 000 pyrs
a = 384
0.025/1.96 = 0.0128
and
mo = mi = a/0.80
b = 0.70 x (a/0.80) = 0.875 a
347
Chapter 15
The above formula for the standard error can then be re-written as
a 0.875a
SE = J- 0.802(1-0.80) + 0.702(1-0.70) = 0.0128
a = 1805
mi == ma = 1805/0.80 == 2256
based on the formula for the SE of a ratio of proportions (see Section A6.1.2).
ln Example 15.7,
348
1
,f
f
1
Size of a sludy
ln our example,
Since the same number of subjects are to be allocated to each arm of the
trial and we anticipate a rate ratio of 0.6, then
a = 0.6 x b
(0.134)2 = 1.6/0.6b
0.6b = 1.6/0.018 = 88.89
b = 88.89/0.6 = 148
a = 0.6 x 148 = 89
The stomach cancer incidence rate for those given placebo is believed to
be 65 per 100000 pyrs. Thus, we need 227 692 pyrs (= 148/0.000065 pyrs)
of foIlow-up in each arm of the trial in order to accumulate 148 stomach
cancer cases among those receiving the intervention and 89 among those
given placebo. This wil ensure that the result of the trial wil have the
desired level of precision. This can be achieved by foIlowing 227 692 men
for one year, 113 846 for two years and so on.
A similar approach can be used to calculate sample size estimates based
on precision in situations where the appropriate measure of effect is the rate
difference rather than the rate ratio, except that the calculations would be
based on the formula of the SE of a difference in rates (see Section A6.1.2).
349
Chapter 15
ln our example,
The odds of exposure among the cases is expected to be twice the odds
of exposure among the controls:
a/b = 2 x (eld)
e = 110 x 0.335
d = 110 x 0.665
Hence,
110 x 0.335
a/b = 2 x = 1.008
110 x 0.665
a = 1.008 x b
0.114= ~i11i
110 = III = a + b = 1.008b + b = 2.008b
- +- + +
1.008b b 2.008b x 0.335 2.008b x 0.665
Thus, we need to recruit 653 cases and 653 controls into our case-con-
trol study.
350
Size of a study
351
Chapter 15
\ .-...
Size of a study
/,
,lox ~ 5.~. Key issues - , , ;
studies are given respectively in
Breslow & Day (1987, 1980) and
Smith & Morrow (1996).
SaflPI~§i?ecalculationscan b~ made to .ensurèjhat:
353
Chapter 16
Cancer prevention
16.1 Introduction
The preceding chapters of this book have focused on principles and
methods needed to study the determinants of disease and their effects.
The ultimate goal of epidemiology, however, is to provide knowledge that
wil help in the formulation of public health policies aimed at preventing
the development of disease in healthy persons.
Nearly aIl important issues in cancer prevention are linked to the nat-
ural history of the disease 'cancer', which can be summarized as shown in
Figure i 6.1. Here, point A indicates the biological onset of the disease and
the start of the pre-clinical phase. This may be the point at which an irre-
versible set of events (e.g., gene mutation) takes place. As a result of pro-
gression of the disease, symptoms and/or signs appear that bring the
patient to medical attention and diagnosis at point C. This is the end of
the pre-clinical phase, which is the period from A to C, and the beginning
of the clinical phase of the natural history. The disease may then progress
to cure (Di), to permanent ilness and disabilty (Oz) or to death (03). The
time from initial symptoms and/or signs to
cure, permanent ilness or death may reflect Pre-clinical phase Clinical phase
the effects of treatments given, as weIl as the
Onset of
underlying characteristics of the untreated Onset of Early symptoms
disease. Exposure disease detection and/or signs
355
"\
~."\""
Chapter 16
imum size that can be seen by X-ray examination. Thus, the dis-
tance from point B to C represents the 'detectable pre-clinical
phase'. The location of point B varies markedly from one indi-
vidual to another, and also de pends on the screening technique
used.
Excess risk = risk (or rate) in the exposed - risk (or rate) in the unexposed
356
Cancer prevention
It is useful to express the excess risk in relation to the total risk (or rate)
of the disease among those exposed to the factor under study. This mea-
sure is called excess fraction (also known as excess risk percentage or attibut-
able risk percentage). It describes the proportion of disease in the exposed
group which is attribut able to the exposure.
Excess fraction (%) = 100 x (excess risk / risk (or rate) in the exposed)
Table 16.1.
Lung cancer incidence in smokers and
non-smokers: hypothetical data.
357
Chapter 16
incidence of a disease that would have been seen if the exposed had never
been exposed. Thus, in Example 16.1, a maximum of 91 % of lung cancer
cases in smokers could theoretically have been prevented if they had never
smoked.
If the exposure is protective, analogous measures can be calculated. They
are usually called risk reduction (also known as prevented risk) and prevented
fraction (also known as prevented risk percentage).
Risk reduction = risk (or rate) ¡n the unexposed - risk (or rate) in the exposed
Prevented fraction (%) = 100 x (risk reduction / risk (or rate) in the unexposed)
"" ,-'"
'-, ','-,''''. ,,-,',',-'',.'
. -,', ',_.,. -',:',',",'-" '--'.:;. ". ,- - -,' "', - - , ' , , ,
.. Exirmple16..lA lafgerai1Clomizedtral was çarred outtóassess the .value
o(asinokinK.cessatiòriprogramme (theinterierition)in reducing theoccùr~ .
TertCe o(lung¿anæramong smòkers.By the endofthf-tralltheincidence of
.. .lurig.cancer was 155. per .1 00000 pyrs mnong thpse who received the inter":.
. ventiol1 an.d.240 per 100.000 pyrs among thecontrols. Thus,' the maximum
.benefit achievedby the intervention was . ..
358
Cancer prevention
Population excess risk = risk (or rate) in the population - risk (or rate) in the unexposed
or, similarly, as
Population excess risk = excess risk x proportion of the population exposed to the
risk factor
359
!
J
,,
!
í
Chapter 16
360
Cancer prevention
stantiaIly higher for smoking because this exposure is far commoner than
exposure to aromatic amines. It has been estimated that the population
excess fraction for smoking in England is 46% in men (Morris
on et al.,
1984), whereas the population excess fraction for aIl occupational expo-
sures (including exposure to aromatic amines) is only between 4 and 19%
(Vineis & Simonato, 1986). Thus, a much larger number of bladder cancer
cases would be prevented by eliminating smoking th
an by eliminating
occupation al exposures.
Prevalence of exposure Relative risk Table 16.2.
(Pe) (%) 2 5 Population excGss fractions for differ-
10
ent levels of prevalence of the expo-
10 0.09 0.29 0.47 sure and various magnitudes of the rel-
25 0.20 0.50 0.69 ative risk.
50 0.33 0.67 0.82
75 0.43 0.75 0.87
95 0.49 0.79 0.90
361
Chapter 16
362
s
i
f
l!
f
1
Cancer prevention
;;,.,,::.l,,'\":_::'i::.'.':"_-,"':'-_::_:' _ ','..":".'."'~"-, ':":'.,"-', '. ". ,"'_-, ,'_ '.,:'. ._"','_.':.':';':,',_:.':'_\:'_','- '",".,
...,...,.:-..-.'._::...';..:..'...,......,:._:. '.....:.'.:.....: .......::.:.-:.. ," ...'.......: ....'.'......._.:..:...:'..-:..-.:.....:..:.-.,..
,':,,',.,,'-,',-
. · Adveitising restrictions
(e.g., banning of tobacco advertising or forcing the print-
ing of healthWarnings on cigarette packages). '
363
Chapter 16
(al
Increased marketing Fairness (except if the risk factor acts at la te stages of
of fiUer cigarettes Doctrine
5000
begins carcinogenesis), as ilustrated in Figure 16.2.
Third, if the measures undertaken are
4000 directed to only a small fraction of the pop-
"
,." ulation (e.g., a region), the evaluation
~ 3000
"" should be limited to the same sub-popula-
~ tion, otherwise the effect may be missed.
u.Q' 2000
Fourth, when implementation has been
WWIl
1000 confined to one area, comparisons of the
o T J 1 1 1 1 1 1 l 1 1 iiT
1925 1930 1935 1940 1945 1950 1955 1960 1965 1970 1975 1980 1985 1990 changes in the intervention area versus
Year
'control' areas may be possible.
Individual-based studies of subjects who
(b)
600 have adopted potentially healthier habits or
MALES
lifestyles are relatively few and were in the
Age-group (yrs)
500
-- 40-44
past confined mainly to the investigation of
'" -D 45-49 the risk of cancer in ex-smokers. Such stud-
~ 400
oo
--
--
50-54
55-59
ies show a marked decline in risk, which is
o
ê 300
-. 6lH4 related to the time since cessation of smok-
-t 65-69
Ol
c. -- 70-74
ing, and they constitute the most powerful
Ol
10 200
a:
--75-79 evidence for the effectiveness of stopping
-x- 80-84
-*- 85+
smoking in preventing cancer (Rose &
ColweIl, 1992; Gupta et aL., 1990). More
0
recently, the potential of other changes in
",'0'"
'"
",'" ,,'" ",ci lifestyle (e.g., changes in diet) in cancer pri-
"-","
!d'o'"
"-"," "-o,'ö
",'ö!d'à'"
"- ~c:
"- "-
'1"",
","1 ",'I
"- "-
!d'ô'"
",'Ö
mary prevention have also been (or are cur-
rently being) assessed in large intervention
180
FEMALES trials (e.g. Alpha-Tocopherol, Beta Carotene
160 Age-group (yrs) Cancer Prevention Study Group, 1994;
140 -- 40-44 Hennekens et aL., 1996; Omenn et aL., 1996;
-D 45-49
~ 120
,.c. -- 50-54
Chlebowski & Grosvenor, 1994).
00 -- 55-59 Individual-based studies have also helped
0 100
0 -. 60-64
s: -t 65-69 to assess the effectiveness of preventive
80
~
c. -- 70.74 measures in the workplace. For instance,
~'" 60 --75-79
a: -x- 80-84 women first employed before 1925 in the
40
-*- 85+ watch dial-painting industry in the USA
20
had greatly increased risks of mortality from
0 bone cancer and from leukaemia and other
'"
",'"
!d'o'"
",rJ
",'ö ",'ö!d'à'"
,,'"
c: '1"", fòC:
# !d'ô'" haematological diseases, but risks declined
"-"," "-"," "- "- "-0,"1 "-0,"1 "OJ "-0,'Ö
for those employed in subsequent years
Y ear of death
(Table 16.3). The reduction in risk coincided
with changes in work regulations in the
Figure 16.2. industry, which included the prohibition of
(a) Per caput consumption of cigarettes among persons aged 18 years and
over, United States of America, 1925-90 (repraduced fram US Department of tipping or pointing of brushes between the
Health and Human Services, 1991) and (b) age-specifie mortaliy trends fram lips in 1925-26 (Figure 16.3). These mea-
Jung cancer, United States of America, 1950-89 (repraduced fram Gilliland &
Samet, 1994).
sures greatly reduced the exposure of the
workers to radium.
364
Cancer prevention
Year of first Bone cancer Leukaemia and blood diseases Table 16.3.
employment SMR (O/E)b Observed no. SMR (O/E)b Observed no. Mortality of women employed in US
of cases (0) of cases (0) watch dial-painting industry and fol-
1915-19 233** 7 lowed to end of 1975, by year of first
7.4* 2
1920-24 154** employment. a
20 3.3* 4
1 925-29 10 1 1.0 1
a Data fram Polednak et al. (1978)
b Expected numbers derived fram cause-specifie mortality rates for US white females
* P.: 0.05; ** P.: 0.001.
16.3.1 Screening
Screening represents an important component of secondary preven-
tion. It involves application of a relatively simple and inexpensive test to
asymptomatic subjects In order to classify them as being likely or unlike-
ly to have the disease which is the object of the screen. The positive cases
can then be subjected to conventional diagnostic procedures and, if nec-
essary, given appropriate treatment. Screening activities are based on the
assumption that early detection and treatment wil retard or stop the pro-
gression of established cases of disease, while later treatment is likely to be
less effective. The ultimate objective of screening for a particular cancer is
to reduce mortality from that disease among the subjects screened. Figure 16.3.
The concept of screening is not as straightforward as it may at first New York newspaper cartoon alluding
to the radium poisoning of watch dial
appear, however. Early treatment does not always improve prognosis and, painters
even if it does, the true benefits of any type of screening have to be
assessed In relation to its risks and costs and in relation to the benefits that
may be derived from other public health activities. The final value of any
screening programme can be established only by rigorous evaluation.
Any cancer screening activity requires (1) a suitable disease; (2) a suit-
able test and (3) a suitable screening programme.
Suitable disease
To be suitable for control by a programme of early detection and treat-
ment, a disease must pass through a pre-clinical phase during which it is
detectable (see Figure 16.1), and early treatment must offer sorne advan-
tage over later treatment. ObviouslYJ there is no point in screening for a
365
Chapter 16
Early treatment
For screening to be of benefit, treatment given during the detectable
pre-clinical phase must result in a lower mortality th an therapy given after
symptoms develop. For example, cancer of the uterine cervix develops
slowly, ta king perhaps more than a decade for the cancer ceIls, which are
initially confined to the outer layer of the cervix, to progress to a phase of
invasiveness. During this pre-invasive stage, the cancer is usually asymp-
tomatic but can be detected by screening using the Papanicolaou (or Pap)
smear test. The prognosis of the disease is mu ch better if treatment begins
during this stage than if the cancer has become invasive.
On the other hand, if early treatment makes no difference because the
prognosis is equaIly good (or equally bad) whether treatment is begun
before or after symptoms develop, the application of a screening test wil
be neither necessary nor effective and it may even be harmful (see below).
Suitable test
For a screening programme to be successful, it must be directed at a suit-
able disease with a suitable test. ln order to assess the suitability of a
screening test, it is necessary to consider its validity and acceptabilty.
Validity
The preliminary assessment of a screening test should involve studies of
its reliability, which is evaluated as intra- and inter-observer variation (see
Section 2.6). Although even perfect reliabilty does not ensure validity, an
, '... The sensitivityof the screening test ìs the proportion of individuals c1assified as
'positivesbythegold standard whO are correctly identified by the screening test:
.._....,'.... c._.....,.. .....,.;.. ',:. ..:.,
367
Chapter 16
ã.
.,
Oisease-
test results in the disease-free and diseased
o.,
a.
free subjects overlap (Figure 16.4(b)). ln these
Õ situations, the location of the cut-off value
Q;
.a to classify screening test results as positive
E
::
z Oisease- or negative is arbitrary. ln Figure 16.4(b),
affected
sorne disease-affected subjects, having val-
ues below the chosen cut-off value, wil be
__ Negative Positive
results results missed by the screening test, whereas sorne
Screening test disease-free people wil be wrongly classified
as positive. Thus, there is generaIly a trade-
Figure 16.4. off between the sensitivity and specificity of a given screening test; its ab il-
Distribution of results from a screening ity to detect as many true cases as possible (high sensitivity) can only be
test in disease-free and disease-
affected individuals: (a) ideal distribu-
increased at the expense of an increase in the number of individuals with-
tion without overlap; (b) overlap of the out the disease who wil erroneously be classified as positive by the screen-
distributions with the inevitable trade- ing test (low specificity) and vice versa (see Section 2.6.1 for a numerical
off of sensitivity and specificity. ln this
example, where the screening test example of this).
tends to give higher values for people Sensitivity is an indicator of the yield of cases (i.e., the number of truly
with the disease than for those without diseased cases identified by the programme), whereas specificity is an indi-
it, moving the eut-off value to the left
(Le., lowering its value) increases sen-
cator of the number of false-positive test results. Although one would like
sitivity but decreases specificity; mov- to detect aIl the subjects with the disease in a screening programme by
ing it to the right (Le., increasing its
value) increases specificity but
using a test with a maximum sensitivity, such a policy might lead to an
decreases sensitivity. unacceptably low specificity, entailng high costs because of the need for
further investigation of large numbers of false positives and a risk of poor
motivation of subjects to participate in subsequent screening examina-
tions. Hence, the choice of the cut-off point depends on the relative costs
of false positives and false negatives.
ln practice, however, it is difficult to estimate the sensitivity of the
test, since it is not possible to apply a 'gold standard test' to the screened
368
Cancer prevention
would be acceptable because of the expense, the discomfort and the risk
of bowel perforation.
369
"\
"'-i~.
Chapter 16
· The. individuals. ta' be' screened are. identifiable (e;g.,list with names and'
addresses.ofalleligible individuals in the.tárgetpopulátion)~'
"-_;,;,_:.::,:',::-~,::-;::-!-,::,-,-;
.;._-:.'_:._''--.:':,'''''''''-: :_:"'_-"'//C'::'¿'/""_-:-":'" "_:._,
"':':''"''-;C'',:''',::-:o,':''':'''-'',,:,'__:_'_
'" " ,_---:'-"',,-,",,.,-".--,;:-,'":-;,;:
'''..,'-:'',' ,-,":;-- - -, ' "_1:',,- .-,,, ." ''-,:' ":~ ;", .. .:, -::
...,...'(mOdiri~dfr6~ Håkårn
370
Cancer prevention
Process measures
The feasibilty, acceptabilty and costs of a programme may be evaluated
by process measures, which are related to the administrative and organiza-
tional aspects of the programme such as identification of the target popu-
lation, number of persons examined, proportion of the target population
examined, facilities for diagnosis and treatment in the health services,
functioning of the referral system and its compliance, total costs, cost per
case detected, etc. The major advantage of these process measures is that
they are readily obtained and are helpful in monitoring the activity of the
programme. Their main limitation is that they do not provide any indica-
tion of whether those screened have lower mortality from the cancer being
targetted by the programme than those who were not screened.
A particularly useful process measure is the predictive value of a positive
test. The predictive positive value (PPV) represents the proportion of per-
sons found to have the disease in question after further diagnostic evalu-
ation out of aIl those who were positive for the screening test (a/(a+b) in
Box 16.2). A high PPV suggests that a reasonably high proportion of the
costs of a programme are in fact being spent for the detection of disease
during its pre-clinical phase. A low PPV suggests that a high proportion of
the costs are being wasted on the detection and diagnostic evaluation of
false positives (people whose screening result is positive but did not appear
to have the disease on subsequent diagnostic investigation). It is impor-
tant to emphasize, however, that the PPV is a proportional measure; a high
PPV might be obtained even if the frequency of case detection is unac-
ceptably low. For instance, the PPV may be 80% indicating that 80% of
those who screened positive were truly diseased. However, if only 10 sub-
jects screen positive, the number of cases detected by the programme wil
be only 8! The main advantage of this measure is that it is available soon
after the screening programme is initiated and, in contrast to sensitivity,
no foIlow-up is necessary for it to be estimated.
The PPV of a screening test depends upon both the number of true pos-
itives a and the number of false positives b (see Box 16.2). Thus, it can be
increased by either increasing the number of true positives or decreasing
the number of false positives. The number of true positives may be
increased by increasing the prevalence of detectable pre-clinical disease,
for instance, by screening less frequently so as to maintain the prevalence
of pre-clinical disease in the target population at a higher leveL The num-
ber of false positives may be reduced by increasing the specificity of the
371
Chapter 16
1-
A B
f- A B ciated with the use of these intermediate end-
points. Absence of a change in the parameters
may mean that the screening is not successful,
A B A B
but they do not provide an adequate measure of
A B
evaluation because they suffer from a number of
biases, namely, length bias, lead-time bias and
Time -+ overdiagnosis bias.
Figure 16.5. (a) Length bias. Length bias refers to the phenomenon occurring when
Diagram iIustrating length bias (see cases detected by a screening programme are not a random sample from
text). Each line represents a case tram
the point where it becomes detectable
the general distribution of aIl cases of pre-clinical disease in the screened
by the screening test (A) to the point population. This is likely to happen when screening tests are applied at
where c1inical symptoms or signs occur moderately long intervals (say once every 2-5 years), so that cases with a
long pre-clinical phase are more likely to be detected than those with
faster-growing tumours (Figure 16.5). Hence, the cases detected by screen-
ing may be those with lesions having a more favourable prognosis, while
cases with similar onset date but more rapid disease progression are detect-
ed by clinical symptoms. The resulting length bias could le ad to an erro-
372
Cancer prevention
neous conclusion that screening was beneficial when, in fact, observed dif-
ferences in survival (case-fatality) were a result merely of the detection of
less rapidly fatal cases through screening.
373
i
,f \.."..,
f
iÍ
Chapter 16
pIe (rom the time o( screening. ln the above example, the mortality rate in
the screened group is two deaths in 496 person-years (98 persons x 5 years,
plus 1 person x 4 years, plus 1 person x 2 years). The rate is the same in
the unscreened group, since the number of person-years, counted from
the time the screening would have taken place had it been done, is iden-
tical to that for the screened group.
There are two ways in which the effect of lead time on the evaluation
of the efficacy of a screening programme can be taken into account. The
first is to compare not the length of survival from diagnosis to death, but
rather the mortality rates in the screened and unscreened groups (as done
in the above example). Alternatively, if the lead time for a given disease
can be estimated, it can be taken into account, allowing comparison of the
survival experience of screen- and symptom-detected cases. For example,
the average lead time for breast cancer has been estimated as approxi-
mately one year (Shapiro et aL., 1974). Thus, to evaluate the effcacy of a
breast cancer screening programme, the two-, three-, four-, five- and six-
year survival risks of the screened cases should be compared, respectively,
with the one-, two-, tluee-, four- and five-year survival risks of the
unscreened cases. However, determinations of lead time are diffcult to
carry out and cannot be generalized, since they depend on the ability of
the screening procedure to detect pre-clinical conditions.
374
l¡
,J
j
¡
f
Cancer prevention
Observational studies
While randomized trials can provide the best and most valid evidence
conceming the effcacy of a screening programme, as with the evaluation
of etiological hypotheses, most evidence on the effects of screening pro-
375
Chapter 16
Figure 16.7.
Change in cervical cancer mortality
between 1950-54 and 1965-69 in rela-
tion to estimated average annual per-
centage of women aged 19 years and
over screened during 1953-68, by
state of the USA. The stars indicate the
less populous states (Le., accumulated
female population aged 19 years or
over of Jess than five millon during the
study period) (reproduced, by permis-
sion of Wiley-Liss Inc., a subsidiary of
John Wiley & Sons, Inc., trom Cramer,
1974).
376
¡
f
,,
i
í
Cancer prevention
377
¡
f
,,
ì
i
f
1
Chapter 16
compared the history of Pap smear screening in 212 hospital cases of inva-
sive cervical cancer with age-matched neighbourhood controls (Clarke &
Anderson, 1979). Fewer cases than controls had received a Pap smear dur-
ing the five years before the year of diagnosis. The risk of invasive cervical
cancer among women who had not had a Pap smear was about three times
that of women who were screened, after controllng for socioeconomic
status. The authors attempted to examine the impact of potential recall
bias by comparing the data obtained through a sample of personal inter-
views with data from physicians' records. There was no evidence that the
information obtained during the interviews was affected by recall bias.
- -", -, '- - . ,-- --, - '-.,,',-
: :_'., , ',- -, ,,:' -
- - '" '-,':,-,
-- ' -,' .:: .',-'.', -
,-', -~',
378
Cancer prevention
379
Chapter 16
Further reading
',-~,_:~'- - '- ,':," ': 'Box:16.4/(G'ónt.)"-:.' -',', ';~,
_ ~ _-: _ v '", ~ - ~ ,r -_ " ~ ~ - _ _~- _~ ~,:, ~, J ~. .
* A comprehensive discussion of
cancer screening programmes is
· Screenínt¡ involves the use of a simple and inexpensive test tc) detect early
stages of cancer at which treatment' is more effectiv~ than at,the .time of usual diag-
given in Cole & Morrison (1978).
nosis. Mass screening programmes should only be ,directed towards the control of
cancers for which there is àn effective treatment that wil reduce mortality,if applied
at early stages. There should also be valid, inexpensive and acceptable tests for the
detection of the cancer at early stages.
· The performance of a screening test In terms of its acceptabilty, feasibility and costs
can be monitored by process measures related to the administrative and organiza-
tional aspects of the programme' (e.g., proportion of the target population examined,
functioning Qf the referralsysterr, cost percase detected, etc.): Predictive value of a
positve te~t. is a particularly' useful measure, because ' it . provides an indication of'
whether most of the effort of the programme is being used to identity cases at an early
stage or whether they are mainlywasted on the evaluation of false positives. Although
process measuresare useful for monitoring the àctivity of the programme they do not'
indicate whether those screened wil have.lower mortality than those'not screened.
380
Appendix 16.1
Calculation of absolute
measures of exposure
effect and measures of
population impact in
case-control studies
ln population-based case-control studies in which the incidence rate
in the total population of interest is known and the distribution of
exposure among the controls is assumed to be representative of the
whole population, these parameters can be used to estimate inci-
dence rates in the exposed and unexposed groups.
A population contains a mix of exposed and unexposed people.
Thus, the overall incidence rate (r) of the disease in a population is
equal to the weighted average of the incidence rates in its exposed (ri)
and unexposed (ro) groups, the weights being the proportions of indi-
viduals in each group. Suppose that a proportion Pe of the population
is exposed to the factor under study. Thus, the proportion of unex-
posed people in that population is equal to (1 - Pe). Hence, the rate
in the population wil be
r = riPe + ro (1 - Pe)
r
ro =
(OR x Pe) + (1 - Pe)
381
Appendix 16.1
.~. ..' .,", . . "'. '; ..., .......,. '.... '., .'"..' ,... .-
tthe exceSi fraètion.:cO Uldaa.:.,:~.J ,~,' een ",~latea by using the
en mSection 16..:2__:.:-1: .'..' " . ' ,... : "
",
".."."...'.,'
." ...' ' '...
'..' '.....
",...
": " .'. .,.'
'.',' .
,'.. ,..... "...'~,...'...(Q ".
essfractlOn (%)=100x ,,',
,,' .,.. ,." '.. ,...".'.'.., . '
'.' .,' ": '.
ess. Fraction
fI' . " (%)::JOOx
,...:., . ," ,,=89%
" .., .',
:',
, J9" ,.' H .'.'.','"
, , ".' "
,.:'.','.
, . '.,... .. .' " " '. ' .' ,,'.-
.' ,,' '. '
,,' ".'".. . .',' .
..', "'"',.",,. , ,', ..','..,.... "..' '...' .:'.. ."
The population ex cess risk and the population excess fraction can then be
determined as
Population excess risk = excess risk x proportion of the population exposed to the
factor (Pe)
The population excess fraction can also be calculated by using the formu-
la given in Section 16.2.1:
Pe (OR - 1)
Population excess fraction (%) = 100 x
Pe (OR - 1) + 1
382
L
¡
\ ...i......
li
(
f
l
Calculation of absolute measures of exposure effect and measures of population impact in case-ontrol studies
(OR - 1)
Excess fraction (%) = 100 x
OR
383
Chapter 17
The role of cancer registries
(1) They describe the extent and nature of the cancer burden in the
community and assist in the establishment of public health prior-
ities.
Sorne of these functions can be fulfiUed using mortality data derived from
vital statistics systems. Cancer registration data, however, provide more com-
prehensive, more valid and more detailed information on patient characteris-
385
Chapter 17
tics than can be obtained from death certificates. Moreover, reliable cause-spe-
cific mortality data are available in most developed countries but in only a few
developing countries. Thus, cancer registries may be the only way of obtain-
ing information on the burden and patterns of cancer in developing coun-
tries, as weIl as providing a focus for research into etiology and prevention.
The discussion in the rest of this chapter wil focus on population-based
cancer registries unless otherwise specified.
386
The raie of cancer registries
(c) GeneraUy available medical care and ready access to medical facil-
ties, so that the great majority of cancer cases wil come into contact
with the health care system at sorne point in their ilness and, the re-
fore, wil be correctly diagnosed.
387
\ ~"i.~,
Chapter 17
388
The raie of cancer registries
Demographie
Address Usual residence
Ethnie group If relevant
The tumour
Incidence date
Most valid basis of diagnosis Non-microscopie or microscopie
Topography (site) Coded using ICD-Ob
Morphology (histology) Coded using ICD-O
Behaviour Coded using ICD-O
Source of information Type of source: physician, laboratory, hospital, death
certificate or other
Actual source: name of physician, laboratory,
hospital, etc.
Dates (e.g. dates of relevant appointments,
hospital admissions, diagnostic procedures)
a Modified from MacLennan (1991).
b International Classifieauon of Diseases for Oneology (Percy et al., 1990).
389
Chapter 17
Treatment
Initial treatment
Follow-up
Date 01 last contact
Status at last contact (alive, dead, emigrated, unknown)
Date of death
Cause of death
Place of death
a Modified from Maclennan (1991).
390
L
!
1
The raie of cancer registries
391
J
,f
f
1
Chapter 17
0/0
Site (ICD-9) Number of cases by age group Total Incidence rate
Unknown 0- 15- 25- 35- 45- 55- 65- 75+ Crude AS Rb
Ali sites 10 69 89 255 241 266 362 264 74 1630 100.0 101.1 238.5
Ali sites but skin 8 69 88 253 236 264 359 259 72 1608 99.8 234.6
Oral cavity (140-145) 1 - 1 1 2 5 2 3 1 16 1.0 1.0 2.5
Nasopharynx (147) - - 1 5 1 1 1 4 - 13 0.8 0.8 2.0
Other pharynx (148-149) - - 1 - - - 2 1 - 4 0.2 0.2 0.6
Oesophagus (150) - - - 1 16 25 63 35 13 153 9.4 9.5 30.4
Stomach (151) - - - 2 10 15 14 20 6 67 4.1 4.2 13.5
Colon (153) - - 5 2 6 6 9 9 2 39 2.4 2.4 6.6
Rectum (154) - - 2 3 6 8 4 4 1 28 1.7 1.7 3.8
Liver (155) - 2 10 22 37 41 46 46 9 213 13.1 13.2 34.6
Pancreas (157) - - - 1 2 7 13 7 1 31 1.9 1.9 5.9
Larynx (161) 1 - - - - 4 12 4 2 23 1.4 1.4 4.5
Bronchus, Jung (162) - - - 1 6 30 50 32 6 125 7.7 7.8 24.6
- - - - - - 1 - - 1 0.1 0.1 0.1
Pleura (163)
Connective tissue (171) - 4 4 4 2 2 1 - - 17 1.0 1.1 1.1
Melanoma of skin (172) - - 2 2 2 2 4 1 1 14 0.9 0.9 1.8
Other skin (173) 2 - 1 2 5 2 3 5 2 22 1.3 1.4 4.0
Breast (175) - - - - 1 3 1 - - 5 0.3 0.3 0.6
Prostate (185) 3 - - - 2 11 37 41 18 112 6.9 6.9 29.2
Penis (187) - - - - 1 4 3 2 3 13 0.8 0.8 2.8
Bladder (188) - 1 - 3 5 19 18 16 6 68 4.2 4.2 13.2
Kidney (189) - 10 1 - 1 2 1 2 - 17 1.0 1.1 1.7
Eye (190) - 5 1 1 - 1 1 1 - 10 0.6 0.6 0.9
Brain, nervous system (191-192) - 7 6 4 5 2 4 1 - 29 1.8 1.8 2.4
Thyroid (193) - - 1 1 1 1 4 1 - 9 0.6 0.6 1.2
- 2 1 4 2 2 2 - - 13 0.8 0.8 1.0
Hodgkin's disease (201)
Non-Hodgkin Iymphoma (200, 202) - 12 4 13 11 10 6 2 - 58 3.6 3.6 4.7
Multiple myeloma (203) - - - 5 4 8 1 1 20 1.2 1.2 2.7
Lymphoid leukaemia (204) - 8 3
1
392
The raie of cancer registries
Identity of individuals
The abilty to distinguish individuals from events (e.g., hospital admis-
sions) is a key feature of a cancer registry. Thus, the registry should have
suffcient information on each individual to avoid multiple registrations
of the same subject. The most universal and generally used identifier is
the name. The utilty of using names will vary depending on local cus-
tom. For instance, surname (or family name) may not be used-persons
may be known only by their first name. Individuals may change theÍr
name when they get married or for other social reasons. Variations in
spellng of names is a frequent problem, particularly if a large percentage
of the population is iliterate. This is aggravated if there is a need to
transliterate names to the Roman alphabet, in order to use computerized
database systems.
393
Chapter 17
Lack of follow-up
Active follow-up usually means that the registry attempts to contact physi-
cians or patients on a regular basis to see if the patient is stil alive. Because
this is expensive, many registries rely on passive follow-up, matching with
death certificates and assuming patients are alive otherwise. Mixed systems
use death certificates plus updating the 'date last known alive' from hospital
admissions, consultations, and other sources of data.
Active foIlow-up of the patients is usuaIly very difficult in developing coun-
tries. Few registries have the necessary facilties for regular foIlow-up of
patients. There are also problems with unreliable postal services, unstable
addresses and mobilty of the population. Passive follow-up is possible only in
the few countries where a reliable death registration system exists.
(DoIl et aL., 1966; Waterhouse et al., 1970, 1976, 1982; Muir et aL., 1987; Parkin
et aL., 1992, 1997). These data are of great value for international comparisons.
ln addition to incidence figures, population-based cancer registries that
conduct adequate foIlow-up of their patients are able to estimate the preva-
lence of cancer. Prevalence figures give an indication of the burden of the dis-
ease in the community. Cancer registries generally assume that once diag-
nosed with cancer, an individual remains a prevalent case unti death. Thus,
prevalence may be estimated from data on incidence and survivaL When a
registry has been in operation for many years, so that aIl patients diagnosed
with cancer before the establishment of the registry have died, the prevalent
cases may simply be enumerated from the registry file! provided, of course,
that the registry receives information on deaths and emigrations for aU regis-
tered cases.
The cancer registry provides an economical and effcient method of ascer-
taining cancer occurrence in intervention trals (Example 17.1) and cohort stud-
ies, as long as the cancer patients are properly identified in their files so that
case matching can be performed.
Population-based registries can also provide a source of cases for case-con-
trol stuies. However! in general, cancer registries are not regarded as weIl suit-
ed for the conduct of these studies because of delays in registration. The main
value of the registry is rather to evaluate the completeness and representa-
tiveness of the case series.
394
The raie of cancer registries
The registry may, however, carry out its own case-control studies using its
database, comparing one tye of cancer with a selection of the other cancers
('controls') (see Section 11. 1.6). The variables usually available for these analy-
sis are limited to those routinely collected by the registry. Registries may sup-
plement these variables with additional information (e.g., smoking, diet,
Table 17.5.
Risk factors for oesophageal cancer in
men, south-western Zimbabwe,
1963-77.8
395
Chapter 17
'. .,~ .
,
, .~
. ,
.. J,:";¡
.' .!__
" ) j, ~f
otal
,
;' Dn.." - , . .
..
... '.
1
396
i \"0,.\...
,l
f
í
The raie of cancer registries
Site ICD-10 No. of incident No. of prevalent cases at Five-year relative Table 17.6.
cases, 1992b 31 December 1992b survival ("I)e Number (and ranking) of male incident
and prevalent cancer cases, and five-
Lung C33-34 6434 (1) 8201 (4) 9
year relative survival ratios for selected
Prostate C61 4096 (2) 13564 (3) 49 sites. South-east England, 1992. a
Colorectal C18-21 3492 (3) 14470 (2) 43
Bladder C67 2183 (4) 16538(1) 70
Stomach C16 1516 (5) 2407 (7) 13
Non-Hodgkin C82-85 1009 (6) 4582 (5) 51
Iymphoma
Oesophagus C15 858 (7) 805 (10) 8
Pancreas C25 833 (8) 586 (11) 6
Kidney C64 644 (9) 2296 (8) 40
Brain C71 523 (10) 1558 (9) 22
Melanoma of skin C43 367 (11) 2855 (6) 71
Ali malignant neoplasms 28732 109 637 36
(excluding non-melanoma
skin cancer)
a Data from Thames Cancer Registry (1995).
b Ranking of sites by decreasing frequency is given in parentheses.
C Patients aged 15 years and over, diagnosed during the years 1986-89.
poor survival have lower prevalence even if their incidence is higher. Figure 17.2.
Up-to-date cancer statistics provide information on the present burden Annual age-adjusted incidence rates of
of cancer to the health care system in a population. To develop long-term cancers at selected primary sites in
Finland: actual rates from 1953 ta 1979
programmes for cancer control, it is necessary to predict what the needs and predictions up to the year 2000
wil be in the future. ln other words, it is necessary to have reliable esti- based on a statistical model which
mates of the numbers of incident and prevalent cases that wil occur in includes age, period and cohort effects
(reproduced with permission from
coming years. Cancer registries are an important source of data upon Läärä, 1982).
which to base such
predictions. The
simplest predictions
of cancer incidence
rates are based on
continuing the pre- Q) 100
200
~ MALES
Ali sites
200
FEMALES
-- ~ Ali sites
~~LUng
-¡ Q) 100
sent age-specific o
'"
o
-¡
o
'"
~ Breast
time trends into the .e 50
o
.e 50
.s
.¡: '..__.... __ Prostate .s
future. The forecast .¡:
CI
CI CI
CI
can be imp-roved if ~ Colon
r?~::::"m ,
g,
~ 20
g, ~ and rectum
birth cohort effects c.
o
o Corpus uteri
can also be taken into o Lung
account by using
g 10
~ , _ ,_'''" oftheskin ~ 10 Gallbladder
age-period-cohort ~
c.
'"
.'CI 5
\ "\'-,
A~//-"-'-- $"i 205
;;;;~a Stomach
Figure 17.2.
An even more 1 .( 1 .(
sophisticated 1950 1960 1970
Year
1980 1990 2000 1950 1960 1970 1980 1990 2000
Year
approach to predic-
397
Chapter 17
398
The raie of cancer registries
Tertiary prevention
Survival statistics can be produced by population-based cancer reg-
istries that follow up theÍr cases, either actively or passively. Although
survival analysis of data from population-based registries cannot evaluate
specific treatments (this can be done only in clinical randomized trials),
it provides a useful evaluation of cancer care in the are
a covered by the
registry, since aIl cancer cases wil be included regardless of the type of
treatment they may have received.
The methods used in survival analyses are those discussed in Chapter
12. The first requirement for the application of the se methods is a clear
and well defìned case defìnition. This should clearly specify the site of the
cancer and/or histology, age and sex of the patient and, if available, the
extent of disease (stage) at the time of diagnosis. The nature of the cases
to be included should also be defined. For example, a decision must be
taken on whether to include cases for which the most valid basis of diag-
nosis is solely clinicaL. A decision should also be taken regarding cases
registered on the basis of a death certificate only (DCO), for whom no
information is available on the date of diagnosis of the cancer. The most
usual practice is to omit these cases from the analysis, but if they repre-
sent a large proportion of registrations, it may be better to present two
survival analyses, one including DCO cases and another excluding them.
ln both cases, the proportion of DCa registrations should be stated in
survival reports.
399
Chapter 17
The second requirement is a clear and weIl defined starting point. For popu-
lation-based cancer registries, the starting date (from which the survival is
calculated) is the incidence date (see Section 17.3.1).
The third requirement is a clear and weIl defined outcome. Death is gener-
ally the outcome of interest, but sorne registries collect enough data to allow
them to conduct analyses using recurrence of tumour, or first recurrence of
a particular complication, as the outcome of interest. It is also necessary to
formulate clear criteria for deciding who should be considered 'lost to fol-
low-up'. For instance, certain registries would assume that cases for which it
was not possible to obtain foIlow-up data for more th an 15 months should
be taken as 'lost to follow-up'.
There are several problems in the interpretation of time trends in survival.
Firstly, improvements in survival may be due, at least in part, to better ascer-
tainment and recording of incident cases. Secondly, if there has been a trend
towards earlier diagnosis (e.g., through introduction of a screening pro-
gramme), survival may improve but the gain may be due entirely to
increased lead time, with no change in mortality rate (see Section 16.3.1).
Despite these caveats, time trends in survival are useful to assess the
extent to which advances in treatment have had an effect in the population.
For instance, the dramatic improvements in survival observed in clinical tri-
als in the treatment of childhood cancers conducted in the 1960s do seem
to have been transposed into the community in many developed countries,
as the population-based survival from many of these cancers shows signifi-
cant increases over time (Table i 7. 7).
Comparisons of cancer survival estimates derived from population-based
cancer registries are increasingly used to compare the effectiveness of cancer
treatment across populations. However, survival reflects not only treatment
but also prognostic factors such as stage at diagnosis, histological type and
other characteristics of the disease. When data on su ch factors are not avail-
Lymphoid leukaemia 14
39
20 39
Five-year survival risk (%)
1962-64b 1971-74
79
25
Hodgkin's disease
Non-Hodgkin Iymphomas
Wilms' tumour
Malignant bone tumours
26
22 57
29
a Data fram Draper et al. (1982).
b 1968-70 data for Iymphoid leukaemia.
ln Example 17.4, survival in Khon Kaen was equal toi or better than, that in
the USA for stomach, liver and lung cancers. Thus, improvements in treatment
may be of reduced benefit in the control of these cancers in Thailand compared
with the potential benefits of primary and secondary prevention (e.g., control
400
The raie of cancer registries
of hepatitis B infection and Hver fluke infestation for Hver cancer; anti-smok-
ing campaigns for lung cancer). ln contrast, survival was lower for Khon Kaen
residents than USA white residents for those cancers whose prognosis is asso-
ciated with early diagnosis (breast, cervi and large bowel), indicating that
interventions to promote early detection may provide potential benefits.
Survival from leukaemia and lymphoma was also lower for residents in Khon
Kaen, probably because of poor access to complex therapeutic regimens.
401
"\
\.\.,
Chapter 17
(1) They are institution-based and not population-based. This means that
no attempt is made to register aIl cancer cases occurring in any
defined population; thus incidence rates cannot be determined.
Patients who are hospitalized in more than one hospital are counted
more than once in an area's hospital tumour registries. Information
may not be shared among hospitals caring for the patient at different
times. Changes over time in numbers of any type of cancer or patient
characteristics may only reflect shifts by patients (or doctors) from
one institution to another. The cancer cases in any one hospital (or
group of hospitals) may not be representative of aIl cancer cases that
are occurring in the area. For instance, certain institutions are referral
centres for specific types of cancer or for particularly diffcult or
extensive tumours.
402
'\..,,-,
The role of cancer registries
Further reading
* Jensen et al. (1991) describe in
great detail the planning of can-
cer registries in both developed
and developing countries and the
uses of registration data in epi-
demiology and public health
planning.
"',:.,,'.,"",:";',,'" .
,",.:, - -,0 '-_"
',- '.",_.-,
-'\",," '.,' _:-'--',' ',-'.'; -' '-, - .,'-,: .. :", .-','-' ,:_,
403
Chapter 18
Designing, planning and
conducting epidemiological
research
18.1 Introduction
ln previous chapters of this book, we covered the basic methodological
principles of study design, analysis and interpretation. ln this chapter, we
concentrate on practical aspects of how to design, plan and conduct an epi-
demiological study. Only general issues are covered. Each research project
and each study setting presents unique problems which cannot be dealt
with here.
The first step in the design of any epidemiological study is to have an
'idea' for a study. This is a creative process for which no guidelines or advice
can be given. The ide a for a study usually cornes from one's own work and
experience and from the realization that there is a need to obtain a clear
answer to a particular research question. Once the idea for the study has
been identified, the next step is a critical review of the existing literature to
find out what exactly is known about the subject and to make sure that the
question has not already been answered. Where appropria te, registers of
clinical trials and/or directories of on-going research (such as that of the
International Agency for Research ()n Cancer) should be consulted. It may
also be helpful to contact experts in the subject. ln addition to revealing
what is already known about the question, this wil help the investigator to
become familar with problems that other researchers have faced, when
using various study methods.
Once aIl the relevant information on the topic has been gathered, the
next step is to state the study hypotheses in a clear and practical way. There
is a tendency to formulate hypotheses that are broad and vague. Instead,
they should be stated in terms of clear, simple, answerable questions. AIl
main and secondary hypotheses should be formulated at the beginning of
a study to ensure that aIl the necessary data are collected. They should
include a concise description of the exposure(s) or intervention(s) to be
studied, the outcome(s) of interest, and the magnitude of the anticipated
effect(s) in the population in which the study wil be conducted.
405
,
¡ -........
f
!
¡
Chapter 18
Writing a protocol
The protocol of the study should be written according to the specifi-
cations of the funding body. Although the layout of the application
forms varies from one funding body to another, they are generally divid-
ed into the following sections:
Study hypotheses
The main and secondary study hypotheses should be written in a clear
and concise way, indicating the exposure(s) and outcome(s) of interest
and the magnitude of the anticipated effect(s).
406
Designing, planning and conducting epidemiological research
Study design
It must be made explicit whether this wil be an intervention, cohort,
case-control, cross-sectional or routine-data-based study. The choice of
design needs to be justified (see Chapters 5 and 7-11).
Sample size
It is necessary to show that the proposed number of subjects in the
study wil provide adequate power or precision to detect or estimate a
particular effect. Sample size estimates should be presented under differ-
ent assumptions. A single estimate is rarely convincing (see Chapter 15).
Statistical analysis
A concise description of the statistical methods planned for use in the
analysis of data should be given (see Chapters 6 and 14).
Ethics
The protocol of any study involving human subjects should provide
answers to the following questions: What wil the subjects be told?
What wil their collaboration entail? Wil invasive procedures be
used? Are there any risks for participants? How wil consent be
obtained (e.g., at an individual or community level; written or ver-
bal)? What steps wil be taken to ensure confidentiality of the data? It
wil be worth consulting at this stage the ethical guidelines proposed
by the Council for International Organizations of Medical Sciences
(CIOMS) and the World Health Organization (WHO) (1993). Most
funding bodies wil fund only projects that have been approved by
the relevant ethical committees and sorne wil wish to see samples of
the information sheet and consent form to be given to the study sub-
j ects.
Timetable
A realistic timetable for carrying out the various activities of the
study should be provided. This should incorporate milestones to be
achieved at regular time intervals (e.g., every six months), which wil
help in monitoring the progress of the study.
References
References to key publications should be included in the grant pro-
posaI.
407
Chapter 18
Budget
The budget is generaIly divided into staff costs, equipment (e.g., com-
puters, freezers, centrifuges) and running expenses (e.g., office expenses,
computer and laboratory consumables, travel costs). Each item must be jus-
tified and of reasonable cost. This wil often involve discussions with the
personnel office about staff costs and consulting price lists to get the best
prices for supplies and equipment. ln certain countries, aIlowances should
be made for inflation and currency fluctuations when calculating the final
budget.
It is also important to check whether the institution where the study is
going to be based wil charge 'overheads'. These correspond to the costs to
the institution of administering the grant, organizing the payment of
salaries, ordering supplies, providing office space, heat, electricity, air-con-
ditioning etc. Sorne funding bodies refuse payment of overheads (e.g., the
World Health Organization), whereas others wil pay only up to a certain
proportion of the total cost (e.g., the European Commission pays only 20%
of the total cost). If the funding body does not pay overheads, it may be
possible to include as running costs in the budget sorne expenses that
might otherwise be covered by the overheads, such as telephone, fax and
mailng costs.
Dissemination of results
Sorne funding bodies require applicants to state how they intend to
inform study participants about the findings of the study and how they
wil be disseminated to relevant health authorities and the scientific com-
munity. This topic is further discussed in Section 18.4.
408
Designing, planning and conducting epidemiological research
'eJ()rs~bniiSs,ionOfapplbi- ,
'--.- . ..... .,' .' ,,' -" -'" ' -'''" ,. '.
ri .charg~sQv~rhl3ads.
.-- ;C" ,'- :;~",:-: _-'.'/_" ;..;,:-;'_,;: : ::~~'.__:. :' ';'''',:';'. '.\' _,_:"__:.:_:''';, :::, -.', -. ;" _" .'_- _,:-, \:, _,: _.', '._ ,"_' _, :"__ :'" .'_ '" "_ , '
',: .,-~;.--,:. -",:';,.-\:,:: :::,':
,:":::': :" '- ::' :/:,-::::i, ::- '"y, -, - _;,":,;'::,-" ~:': ,,,"-', :'~~ -';,:_::',:,"..,'::\'~: ,;'. :'":.",,,::')"':.:';.:_, .": '-:, - - . , /- '-:::':_ .
409
,
l-
f
!
Chapter 18
..
ii )i
Conduct of interviews
and collection of 4( .-
blood sam pies
Laboratory measurements ~
Data entry and checking il .
Data analysis 0( ..
Reports to Ministry of
Health and funding bodies X X X
Preparation of scientific or )t
papers
410
¡
i \\,~,
J
f
¡
Designing, planning and conducting epidemiological research
the study. Copies of the relevant parts of the manual should be given to
aIl research team members. ln addition to the study manual, it is also
important to keep a study diary in which aIl problems encountered are
noted and the solutions adopted are recorded. This wil be useful in main-
taining consistency of decisions throughout the study.
A detailed record of expenditure should be kept to monitor expenses
throughout the study and to faciltate submission of expenditure state-
ments to the funding body. It is important to keep copies of salary pay-
ment sheets, invoices, receipts, order forms, etc. Grant accounts can be
easily monitored by using a computer spreadsheet.
18.3.3 Equipment
The equipment required wil depend on the design of the study and the
setting where it wil be carried out. Routine-data-based studies may require
only access to good computing facilties. ln contras t, a large intervention
trial conducted in a remote area may require everything from computer
equipment to staff accommodation and transport facilties.
Computing equipment
The choice of computer equipment ('hardware') and computer pro-
grams ('software') (see Box 18.3) wil be determined by the design and size
of the study as weIl as by the availabilty of technical expertise for data-
processing and statistical analysis and of local servicing facilties. However,
the rate of development of computer equipment and software is such that
any advice or guidelines soon become out-of-date. Thus, it is essential that
professional advice, from a computing specialist, should be sought at the
planning stage of the study so that advantage is taken of the most recent
developments.
Hardware
Computers differ in their type of microprocessor, the size of their ran-
dom access memory (RAM) and their storage capacity. The most com-
monly used microcomputers are based on the 486 and Pentium micro-
411
Chapter 18
processors. The RAM governs the size of program and the amount of data
that the machine can actively work with at any time. The bigger the RAM
the better. Most machines have at least 4 Mb of RAM, but if Windows is to
be used, at least 16 Mb is desirable. The data storage capacity is determined
by the capacity of the ha rd disk. ln general, the hard disk capacity should
be at least 500 megabytes (Mb), especiaUy if Windows-based software is to
be used, as these programmes tend to use a considerable amount of space
on the ha rd disk.
Once the appropriate type of machine for the study has been selected,
it is necessary to obtain price lists from different manufacturers.
Machines with similar specifications are produced by many different
manufacturers and the prices tend to differ widely. However, value for
money is not the same as cheapness. Equipment reliabilty, warrant y and
free maintenance from the dealer are aU important factors which need to
be considered. The availabilty of good local servicing and repair facilties
is another important consideration, particularly for studies in developing
countries.
Microcomputers with similar features are available in two basic mod-
els: desktop/tower and laptop (portable) computers. Laptop computers
have the advantage of being portable, but they are more expensive than
desktop machines of equivalent computing capacity. However, since they
work from rechargeable batteries, they avoid the need for an uninter-
ruptible power supply (see below).
It is absolutely vital to take adequate precautions against the possibil-
ty of losing data that have been entered onto the computer. AU the infor-
mation on the computer should be regularly copied onto sorne other stor-
age device to guard against the possibilty of hard disk or computer fail-
ure. Floppy disks are adequate to store relatively smaU data-sets, but
removable hard-disk units may be needed for very large data-sets.
Alternatively, a 'tape-streamer' may be used to back-up the data onto a
sm aU magnetic tape. It is essential to make back-up copies of the data reg-
ularly, on both a daily and weekly basis, and ta create multiple copies as
a back-up to the back-up (see Section 18.3.6).
Unreliable power supplies can lead to loss of data and serious damage
to the computing equipment. It is advisable to avoid sharing the mains
circuit with equipment such as electric motors and air-conditioning sys-
tems that makes heavy but intermittent demands on the power supply.
Power stabilzers (or mains filters) are smaU devices, sometimes incorpo-
rated into the plugs, which are designed to de al with rapid voltage fluc-
tuations. However, they offer no protection against loss of data if there is
a power failure. Uninterruptible power supplies (UPS) are devices that, in
the event of a power failure, provide power (from batteries) for short peri-
ods, until the batteries run down. This gives enough time to save the
work and switch off the computer without loss of information. Sorne
UPSs shut down the computer automatically should the loss of the main
power supply be prolonged.
412
Designing, planning and conducting epidemiological research
Printers vary in price enormously. ln general, the better the print qual-
ity, the higher the price. Dot-matrix printers are relatively cheap and the
quality of their output is usuaIly good enough for most requirements. An
ink-jet or a laser printer may be required if there is a need to produce high-
er-quality documents such as grant proposaIs, questionnaires and study
reports.
\::::: ': :' ,;:,_" ;,:, ::-_',;"'i. ;'-:; ',':. _: :,:: ':::'" ,,:,: .-,,- '" -' - '_ ',:': ,;:, , ,- ",:', .::: ,: : :_:, '::,:,::. ',,_:':', '., ,-' ,_: ',::. ; """-,~
Software
ln order for a computer to work, it requires a set of instructions which
control its operations. This set of instructions is called an operating system.
ln the past, most IBM-tye microcomputers used the DOS operating system,
but today the Windows operating system is almost universaL Windows
allows several programs to be run simultaneously (which is not possible
with DOS unless a special software package has been installed) and, there-
fore, it is easier to move data, text and figures between diferent programs.
Apart from the operating system, most epidemiological studies require
a database package to enter and check the data, a statistical package to per-
413
Chapter 18
Other equipment
It may be necessary to obtain offce equipment such as desks, filng cab-
inets, a photocopying machine, a fax machine, etc. Studies involving col-
lection of biological specimens may require laboratory equipment such as
freezers, centrifuges, etc. (see Section 18.3.5). Field studies in remote areas
may require staff accommodation and transport facilties.
414
Designing, planning and conducting epidemiological research
415
Chapter 18
sive labels with each identification number duplicated several times are avail-
able commercially. Alternatively, they can easily be produced by a microcom-
puter. Waterproof marker pens should be used, except if samples are going to
be frozen in salt-alcohol mixtures, in which case plain pencils should be used
instead. The labellng must be done on the body of the container and never
on the cap only. Numbers and letters must be written in a clear and stan-
dardized form (see Box 18.5). It is important to pre-test the labels to check how
they stick and how pen/pencil writing is preserved during transport and stor-
age.
Samples may be stored temporarily in a laboratory located where the field-
work is being carried out before being sent to the laboratory where the assays
will be conducted and/or to their permanent storage place. If a large number
of sam pIes is collected and stored, it is important to store them in a way that
aIlows rapid retrieval of any particular sample. For instance, samples may be
stored in batches according to the date they were collected or frozen. This
information may be computerized so that any particular specimen can be eas-
ily and rapidly located.
\.......
Designing, planning and conducting epidemiological research
417
Chapter 18
418
Designing, planning and conducting epidemiological research
the original response but never writing over the top of the incorrect
response. Field workers and supervisors may use pens of different colours
so that it is clear where corrections were made.
The data-collection activities should be supervised and the quality of
the data monitored on a regular basis. Every data-collection activity
should be designed in su ch a way that it can be checked. Checks should
be designed with the objective of detecting errors rather than proving the
high quality of the data.
419
\'.'''
Chapter 18
. .... ", '.',' Box UÎ.6. Ert$'u,\i'igj;ióDd Ctåì~q:¡ùal'Îty ',.' ; ::. ,',
~ / ' ~ ~ l _ ,,~ J ~ r '~ ~~ '" ~ ~- ,
- Ensure adequate training and supervision of field workers.
,-.-.- .,......-...
Coding
Coding is the term used to describe the conversion of data from a data-
collection form into a format that is suitable for analysis on a computer,
by assigning a numerical code to every possible answer on the data-col-
lection form. For instance, sex may be coded 1 for males and 2 for females;
only the number 1 or 2 wil be entered onto the computer fie. Numerical
data (e.g., number of children) do not require coding since the exact num-
ber can be entered.
Many data collection forms are 'pre-coded', that is, they are designed so
that every possible answer is assigned a code in the form (see Appendix
2.1). The field worker selects one (or more) of these 'pre-coded' options.
Such pre-coded forms have the advantages of being almost ready for com-
puter entry by the data clerk and of minimizing transcription errors.
However, later coding is stil required for 'open' questions and for answers
which do not fit into the pre-coded categories. These answers may be
given additional codes if necessary.
An alternative approach is to code the data only after the form has been
completed. The data-collection forms are designed so that there is a spe-
420
Designing, planning and conducting epidemiological research
cifc column for coding on the right- hand side of the page. The field
workers (or the study subjects if the questionnaire is self-administered) are
asked to ignore this column and fill in the answers in the spaces provided
on the left-hand side. The answers are coded later by the field workers or
by specially trained coding clerks and the codes are inserted in the right-
hand column. The time between data recording and coding should be
kept as short as possible so that inconsistencies in the data are revealed
and the field worker can be asked to re-contact the subject to correct them.
Sorne aspects deserve special attention when developing coding instruc-
tions. Firstly, numerical variables such as number of children, weight or
height should not be coded into pre-determined categories. TheÍf actual
values should be recorded so as to preserve the full detail of the data in
analyses. Data may then be grouped in different ways, if appropriate.
Secondly, specific codes should be given if a particular question is not
applicable to a respondent (e.g., number of pregnancies for a man) or the
answer is not known or the response was mistakenly left blank by the field
worker (for example, 1 = yes; 2 = no; 7 = not recorded; 8 = not relevant; 9
= not known). Similarly, specific result codes should be developed in lab-
oratory forms to allow for the coding of technical problems such as bro-
ken samples, insuffcient material, error in laboratory procedures, etc. To
minimize coding errors, it is advisable, if possible, that the meaning of any
particular code remains constant throughout the data-coIlection form
(e.g., a code of 9 should be given to the 'not known' answer of every ques-
tion). Thirdly, a coding manual with aIl the coding values and rules should
be compiled and given to field workers, data-coding and data-entry clerks.
This manual should be updated if any changes in coding are made during
the study.
421
Chapter 18
Data c1eaning
Range and consistency checks should be mn for each variable, either while
they are being entered into the computer or after a whole batch has been
entered. Range checks identify inadmissible values. For instance, if the vari-
able sex was coded '1 for males and '2' for females, values other th an 1 and
2 would be flagged as errors. Consistency checks cross-check the consistency
of data for related variables. For example, ilales should not have a diagnosis
of ovarian cancer.
ln addition to checking for incorrect or unusual values, the distribution of
each variable should be examined. Interim tabulations of the data and scat-
ter plots of quantitative variables, such as height against weight, are effective
in revealing unlikely outlying observations. These should be checked against
the data collection forms, since they are unlikely but not necessarily impos-
sible. ln sorne cases, it may be possible to correct the data. ln other cases, it
wil be necessary to insert a 'missing value' code, if it is certain that the data
were in error (e.g., an impossible height).
422
Designing, planning and conductîng epîdemiological research
complete and updated list should be made of what data are on aIl stored disks
and tapes, and copies should be kept at more than one location.
The data collection forms should also be stored in a secure place to ensure
confidentiality of the data. The forms should be filed in batches or in seriaI
order (either by date of collection or by study identification number) to facil-
itate their retrieval for checking apparent errors detected in records on com-
puter. IdeaIly, the forms should be kept weIl after the end of the study, but
id-
space restrictions may preclude this. ln large studiesi it may be worth cons
ering the possibilty of copying the data collection forms onto microfiches.
423
i
,f
r
t
Chapter 18
Box 18.7 lists sorne of the most important international journals of rele-
vance to cancer epidemiologists. A useful way of choosing an appropriate
journal is to see where most previous work on the same topic was published.
The journal's requirements-style, headings, referencing system, etc.
must be followed closely. 'Instructions for Authors' are generally published
in most issues of journals. The draft of the paper should be prepared exact-
ly according to instructions. It is important to think carefully about how
to ilustrate the results (tables, figures, etc.) so that the main findings of
the study are clearly presented to potential readers.
Most journals ask the authors to structure the paper into the following
sections:
Tit/e page
This page should include an informative and concise title for the paper;
me of the institution in which the
the name(s) of the author(s); the na
work was performed; the institutional affilation of each author; and the
na me and address of the author to whom correspondence should be
addressed.
Abstract
The abstract should summarize in no more than a few hundred
words the reason for the study, the major findings, and the principal
conclusions. Sorne papers (e.g., British Medical Journal) require a struc-
tured abstract which is divided into special sections. It is helpful to
ob tain sorne examples of abstracts published in the journal before
attempting to write one. Sorne journals ask the authors to provide key
words from the Medical Subject Headings of Index Medicus for indexing
purposes.
Introduction
This section should include a description of the background of the
investigation (citing only essential references) and the reason(s) for per-
forming the study.
Results
The findings should be presented in a clear and concise way. Repetition
should be avoided by presenting data in only one format, that is, results
displayed in a table should not be presented in the text.
424
j
.~\
,f \':
f
¡
J
1
Designing, planning and conducting epidemiological research
References
The references should be written according to the specifications of the
journal where the paper is going to be submitted.
18.4.2 Review
If the editor thinks the journal might accept the paper, he/she wil send
it out to referees. The reports from these referees wil then determine
whether the paper is accepted or rejected. Papers are rarely accepted with-
out alterations. Most commonly, it wil be either rejected or accepted sub-
ject to modifications based on the referees' comments. The authors should
submit a revised version of the paper which takes into account these com-
ments, together with a covering letter to the editor explaining exactly how
this was done.
425
References
Acheson, E.D. (1967) Medical Record Linkage, Oxford, Oxford Beasley, R.P., Hwang, L.-Y., Lin, C-C & Chien, CoS. (1981)
University Press Hepatocellular carcinoma and hepatitis B virus. A prospec-
Alpha-Tocopherol, Beta Carotene Cancer Prevention Study tive study of 22 707 men in Taiwan. Lancet, ii, 1129-1133
Group (1994) The effect of vitamin E and beta carotene on Belcher, D.w', Wurapa, EK., Neumann, A.K. & Lourie, LM.
the incidence of lung cancer and other cancers in male (1976) A household morbidity survey in rural Africa. Int. J.
smokers. New Engl. 1. Med., 330, 1029-1035 Epidemio!., 5, 113-120
Altman, D.G. (1991) Practical Statistics for Medical Researcli, Beral, V. (1974) Cancer of the cervix: a sexually transmitted
London, Chapman & Hall infection? Lancet, 1, 1037-1040
American Cancer Society (1951) Manual ofTUlnor Nomenclature BeraL, V., Chilvers, C & Fraser, P. (1978) Does pregnancy pro-
te et against ovarian cancer? Lancet, 1, 1083-1087
and Coding, first edition. Washington, DC
Andersen, L.P., Rosenstock, S.J., Bonnevie, O. & Jørgensen, T. Beral, v., Chilvers, C & Fraser, P. (1979) On the estimation of
(1996) Seroprevalence of immunoglobulin G, M, and A an
ti- relative risk from vital statistical data. J. Epidemio!.
bodies to Helicobacter pylori in an unselected Danish popula- Commimity Healtli, 33, 159-162
tion. Am. f. Epidemiol., 143, 1157-1164 Bergström, R., Adami, H.-O., Möhner, M., Zatonski, w', Storm,
Anon. (1901) Verhandlungen des Comités für Krebsforschung. H., Ekbom, A., Tretli, S., Teppo, L., Akre, O. & Hakulinen, T.
Dtscli. Med. Wocliensclir., suppl., 305-312 (1996) Increase in testicular cancer incidence in six European
Aristizabal, N., Cuello, C, Correa, P., Collazos, T. & Haenszel, countries: a birth cohort phenomenon. f. Nat!. Cancer. Inst.,
W. (1984) The impact of vaginal cytology on cervical cancer 88, 727-733
risks in Cali, Colombia. Int. f. Cance, 34, 5-9 Bland, M. (1987) An Introduction to Medical Statistics, Oxford,
Armitage, P. & Berry, G. (1994) Statistical Metliods in Medical Oxford University Press
Researcli, third edition, Oxford, Blackwell Scientific Bland, J.M. & Altman, D.G. (1986) Statistical methods for
Publications assessing agreement between two methods of clinical mea-
Armstrong, B. & Mann, J,L (1985) Diet. ln: Vessey, M.P. & Gray, surement. Lancet, 1,307-310
M. (eds), Cancer Risks and Prevention, Oxford, Oxford Bleehen, N.M. & Stenning, S.P. (The Medical Research Council
University Press, p. 71 Brain Tumour Working Party) (1991) A Medical Research
Armstrong, B.K., White, E. & Saracci, R. (1992) Principles of Council trial of two radiotherapy doses in the treatment of
Exposure Measurement in Epidemiology (Monographs in grades 3 and 4 astrocytoma. Br. f. Cancer, 64, 769-774
Epidemiology and Biostatistics, VoL. 21), Oxford, Oxford Bone, M.R., Bebbington, A.C, Jagger, C, Morgan, K. &
University Press Nicolaas, G. (1995) Healtli Expectancy and its Uses, London,
Ashley, J. & Devis T. (1992) Death certification from the point H.M.S.O.
of view of the epidemiologist. Population Trends, 67, 22-28 Bosch, EX., Muñoz, N., de Sanjosé, S., Izarzugaza, L, Gili, M.,
Baldwin, J.A., Acheson, E.D. & Graham, W.J (eds.) (1987) Viladiu, P., Tormo, M.J Moreo, P., Ascunce, N., Gonzalez,
Textbook of Medical Record Linkage, Oxford, Oxford University L.C, Ta fur, L., Kaldor, J,M., Guerrero, E., Aristizabal, N.,
Press Santamaria, M., Alonso de Ruiz, P. & Shah, K. (1992) Risk fac-
Balzi, D., Buiatti, E., Geddes, M., Khlat, M., Masuyer, E. & tors for cervical cancer in Colombia and Spain. Int. J. Cancer,
Parkin, D.M. (1993) Summary of the results by site. ln: 52, 750-758
Geddes, M., Parkin, D.M., Khlat, M., Balzi, D. & Buiatti, E. Bouchardy, C, Parkin, D.M., Khlat, M., Mirra, A.P., Kogevinas,
(eds.), Cance in Italian Migrant Populations (IARC Scientific
M., de Lima, ED. & de Cravalho Ferreira, CE. (1993)
Publications No. 123), Lyon, International Agency for Education and mortality from cancer in São Paulo, Brazil.
Research on Cancer, pp. 193-284 Ann. Epidemiol., 3, 64-70
Bassett, M.T., Chokunonga, E., Mauchaza, B., Levy, L., Ferlay, J, Bouwes Bavinck, J,N., Vermeer, B.J van der Woude, EL,
& Parkin, D.M. (1995) Cancer in the African population of Vandenbroucke, J.P., Schreuder, G.M.Th., Thorogood, J.,
Harare, Zimbabwe, 1990-1992. Int. J. Cancer, 63, 29-36 Persijn, G.G. & Claas, EH.J, (1991) Relation between skin
Beaglehole, R., Bonita, R. & Kjellström, T. (1993) Basic cancer and HLA antigens in renal-transplant recipients. New
Epidemiology, Geneva, World Health Organization Engl. J. Med., 325, 843-848
427
References
Breslow, N.E. & Day, N.E. (1980) Statistical Methods in Cancer Casley, D.j & Lury, D.A. (1987) Data Collection in Developing
Research. Volume I. The Analysis of Case-Control Studies (IARC Countries, second edition, Oxford, Clarendon Press
Scientific Publications No. 32), Lyon, International Agency for Chen, J., Campbell, T.C, Junkvao, L. & Peto, R. (1990) Diet, Life-
Research on Cancer Style and Mortality in China. A Study of the Characteristics of 65
Breslow, N.E. & Day, N.E. (1987) Statistical Methods in Cancer Chinese Coimties, Oxford, Oxford University Press
Research. Volume II. The Design and Analysis of Cohon Studies Chlebowski, R.T. & Grosvenor, M. (1994) The scope of nutrition
(IARC Scientific Publications No. 82), Lyon, International intervention trials with cancer-related endpoints. Cancer, 74,
Agency for Research on Cancer 2734-2738
Breslow, N.E. & Enstrom, J.E. (1974) Geographie correlations Chuang, w.-L., Chang, w.-y., Lu, S.-N., Su, w.-P., Lin, Z.-Y,
between cancer mortality rates and alcohol-tobacco consump- Chen, S.-C, Hsieh, M.-Y., Wang, L.-Y, You, S.-L. & Chen, C-
tion in the United States. l. Natl. Cancer Inst., 53, 631-639 J (1992) The role of hepatitis Band C viruses in hepatocellu-
Broders, A.C (1920) Squamous-cell epithelioma of the lip. A lar carcinoma in a hepatitis B endemic area. A case-control
study of five hundred and thirty-seven cases. l. Am. Med. Assoc., study. Cancer, 69, 2052-2054
74, 656-664 Cicchetti, D.V. & Feinstein, A.R. (1990) High agreement but low
Brothwell, D. (1967) The evidence for neoplasms. ln: Brothwell, Kappa: II. Resolving the paradoxes. l. Clin. Epidemiol. 43,
D. & Sandison, A.T. (eds.) Diseases in Antiquity: A SU/vey of 551-558
Diseases, Injuries and Surgery in Early Populations, Springfield, IL, ClOMS (Council for International Organizations of Medical
Charles C Thomas, pp. 320-345 Sciences) and WHO (World HeaIth Organisation) (1993)
Buck, C, Llopis, A., Nájera, E. & Terris, M. (1988) The Challenge of Intemational Ethical Guidelines for Biomedical Research Involving
Epidemiology. Issues and Selected Readings. Pan American Health Human Subjects, Geneva, CIOMS
Organization (PAHO) (Scientific Publication No. 505), Clarke, E.A. & Anderson, T.w. (1979) Does screening by "Pap"
Washington, DC, PAHO smears help prevent cervical cancer? A case-control study.
Buell, P. & Dunn, JE. (1965) Cancer mortality among Japanese Lancet, 2, 1-4
Issei and Nisei of California. Cancer, 18, 656-664 Clay ton, D. & Hils, M. (1993) Statistical Models in Epidemiology,
Buiatti, E., Pall, D., Amadori, O., Marubini, E., Puntoni, R., Oxford, Oxford University Press
Avellni, C, Bianchi, S., Cipriani, E, Cocco, P., Decarli, A., Clay ton, D. & Schiffers, E. (1987a) Models for temporal varia-
Vindigni, C & Blot, W. (1989a) Methodological issues in a mul- tions in cancer rates. 1. Age-period and age-cohort models.
ticentric study of gastric cancer and diet in Italy: Study design, Stat. Med., 6, 449-467
data sources and quality controls. Tumori, 75,410-19 Clay ton, D. & Schiffers, E. (1987b) Models for temporal varia-
Buiatti, E., Pall, D., Decarli, A., Amadori, D., Avellni, C, Bianchi, tions in cancer rates. II. Age-period-cohort models. Stat. Med.,
S., Biserni, R., Cipriani, E, Cocco, P., Giacosa, A., Marubini, E., 6, 469-481
Puntoni, R., Vindigni, C, Fraumeni Jr, J. & Blot, W. (1989b) A Cole, P & Morrison, AS (1978) Basic issues in cancer screening.
case-control study of gastric cancer and diet in Italy. Int. l. ln: Miler, A.B. (ed.) Screening in Cancer (A report of a UICC
Cancer, 44, 611-616 International Workshop, Toronto, Canada, April 24-27,
Buring, J,., Bain, CJ & Ehrmann, R.L. (1986) Conjugated estro- 1978), Geneva, International Union Against Cancer, pp. 7-39
gen use and risk of endometrial cancer. Am. l. Epidemiol., 124, Collaborative Group on Hormonal Factors in Breast Cancer
434-441 (1996) Breast cancer and hormonal contraceptives: collabora-
Cameron, H.M. & McGoogan, E. (1981) A prospective study of tive reanalysis of individual data on 53 297 women with
1152 hospital autopsies: I. Inaccuracies in death certificates. l. breast cancer and 100 239 women without breast cancer fram
Pathol., 113, 273-283 54 epidemiological studies. Lancet, 347, 1713-1727
Carpenter, L.M. (1987) Some observations on the healthy worker College of American Pathologists (1965) Systematized
effect. Br. 1. Ind. Med., 44, 289-291 Nomenclature of Pathology (SNOP), Ist edition, Skolkie, IL
Case, R.A.M. & Pearson, JT. (1954) Tumours of the urinary blad- College of American Pathologists (1977) Systematized
der in workmen engaged in the manufacture and use of certain Nomenclature ofMediCIne (SNOMED), Ist edition, Skolkie, IL
dyestuff intermediates in the British chemical industry. Part II. COMMIT Research Group (1991) Community Intervention
Further consideration of the role of anilne and of the manu- Trial for Smoking Cessation (COMMIT): summary of design
facture of auramine and magenta (fuchsine) as possible and intervention. l. Natl. Cancer Inst., 83, 1620-1628
causative agents. Br. l. Ind. Med., 11,213-216 Cornfield, J. (1951) A method of estimating comparative rates
Case, R.A.M., Hosker, M.E., McDonald, D.B. & Pearson,JT. (1954) from clinical data. Applications to cancer of the lung, breast,
Tumours of the urinary bladder in workmen engaged in the and cervix. l. Natl. Cancer Inst., 11, 1269-1275
manufacture and use of certain dyestuff intermediates in the CottIer, L.B., Zipp, JE, Robins, L.N. & Spitznagel, E.L. (1987)
British chemical industry. Part 1. The role of anilne, benzidine, Difficult-to-recruit respondents and their effect on prevalence
alpha-naphthylamine, and beta-naphthylamine. Br. l. Ind. estimates in an epidemiologic survey. Am. 1. Epidemiol., 125,
Med., 11, 75-104 329-339
428
References
Cox, D.R. & Snell, E.j. (1981) Applied Statistžcs: Principles and Enterline, P.E. (1965) Mortality among asbestos products work-
Examples, London, Chapman & Hall ers in the United States. Ann. N. Y. Acad. Sci., 132, 156-165
Cramer, D.W. (1974) The role of cervical cytology in the declin- Estève, 1. Benhamou, E. & Raymond, L. (1994) Statistical
ing morbidity and mortality of cervical cancer. Cancer, 34, Methods in Cancer Research. Volume iv. Descriptive Epidemiology,
2018-2027 (lARC Scientific Publications No. 128). Lyon, International
Dawber, T.R., Meadors, G.E & Moore, EE., Jr. (1951) Agency for Research on Cancer
Epidemiological approaches to heart disease: the Farr, W. (1975) Vital Statistics: A Memorial Volume of Selections
Framingham Study. Am. f. Public Health, 41, 279-286 (rom the Reports and Writings of Wiliam Farr (The History of
Day, N.E. & Boice, J.D., Jr (eds.) (1983) Second Cancer in Relation Medicine Series No, 46), Metuchen, NJ, Scarecrow Press.
to Radiation Treatment for Cervical Cancer (IARC Scientific Feinleib, M. (1984) Data bases, data banks and data dredging:
Publications No. 52), Lyon, International Agency for Research the agony and the ecstasy. f. Chrono Dis., 37, 783-790
on Cancer Feinsteini A.R & Cicchetti, D.V. (1990) High agreement but low
De Stavola, B. (transL) (1987) Statistical facts about cancers on Kappa: i. The problems of two paradoxes. f. Clin. Epidemiol.
which doctor Rigoni-Stern based his contribution to the 43, 543-549
Surgeon's subgroup of the IV Congress of the Italian scientists Ferlay, J. (1994) ICD Conversion Programs for Cancer, (IARC
on 23 September 1842. Stat. Med., 6, 881-884 Technical Report No. 21), Lyon, International Agency for
Dey, P., Collns, S., Wil, S. & Woodman, e.B.j. (1995) Research on Cancer
Randomised controlled trial assessing effectiveness of health Fisher, RA & Yates, E (1963) Statistical Tables for Biological,
education leaflets in reducing incidence of sunburn. Br. Med. Agricultural and Medical Research, London, Longman
T., 31l, 1062-1063 Fisher, B., Redmond, c., Fisher, E.R., Bauer, M., Wolmark, N.,
Doll, R. (1955) Mortality from lung cancer in asbestos workers. Wickerham, D.L., Deutsch, M., Montague, E., Margolese, R. &
Br. f. Ind. Med., 12, 81-86 Foster, R. (1985) Ten-year results of a randomised clinical trial
DoU, R. & Hil, A.B. (1950) Smoking and carcinoma of the lung. comparing radical mastectomy and total mastectomy with or
Preliminary report. Br. Med. f., ÍI, 739-748 without radiation. New Engl. f. Med., 312, 674-681
DoU, R. & Hil, A.B. (1952) A study of the aetiology of carcino- Flegal, KM., Brownie, e. & Haas, J.D. (1986) The effects of expo-
ma of the lung. Br. Med. f., ÍI, 1271-1286 sure misclassification on estimates of relative risk. Am. f.
Doll, R. & Hil, A.B. (1954) The mortality of doctors in relation Epidemiol. 123, 736-751
to their smoking habits. A preliminary report. Br. Med. f. i, Forman, D., Sitas, E, NeweU, D.G., Stacey, A.R., Boreham, 1.
1451-1455 Peto, R., Campbell, T.c., Li, J. & Chen, J. (1990) Geographie
DoU, R. & Peto, R. (1976) Mortalìty in relation to smoking: 20 association of Helicobacter pylori antibody prevalence and gas-
years' observations on male British doctors. Br. Med. f., 2, trie cancer mortality in rural China. Int. f. Cancer, 46, 608-611
1525-1536 Fukao, A., Komatsu, S., Tsubono, Y., Hisamichi, S., Ohori, H.,
DoU, R., Payne, P. & Waterhouse, JAH. (1966) Cancer Incidence Kizawa, T., Ohsato, N., Fujino, N., Endo, N. & Iha, M. (1993)
in Five Continents. VoL. 1. Geneva, UICe. Berlin, Springer Helicobacter pylori infection and chronic atrophie gastritis
DoU, R., Gray, R., Hafner, B. & Peto, R. (1980) Mortality in rela- among Japanese blood donors: a cross-sectional study. Cancer
tion to smoking: 22 years' observations on female British doc- Causes Control, 4, 307-312
tors. Br. Med. f., 280, 967-971 Gambia Hepatitis Study Group (1987) The Gambia Hepatitis
DoU, R., Peto, R., Wheatley, K, Gray, R & Sutherland, 1. (1994a) Intervention Study. Cancer Res., 47, 5782-5787
Mortality in relation to smoking: 40 years' observations on Gardner, M.J & Altman, D.G. (1986) Confidence intervals
male British doctors. Br. Med. f., 309, 901-911 rather than P values: estimation rather than hypothesis test-
DoU, R.i Peto, R., HaU, E., Wheatley, K & Gray, R (1994b) ing. Br. Med. f., 292, 746-750
Mortality in relation to consumption of alcohol: 13 years' Geser, A., de Thé, G., Lenoir, G., Dayi N.E. & Wiliams, E.H.
observations on male British doctors. Br. Med. f., 309, 911-918 (1982) Final case reporting from the Ugandan prospective
Dosemeci, M., Wacholder, S. & Lubin, J.H. (1990) Does nondif- study of the relationship between EBV and Burkitts lym-
ferential misclassification of exposure always bias a true effect phoma. Int. f. Cancer, 29, 397-400
toward the null value? Am. f. Epidemiol., 132, 746-748 Giliand, ED. & Samet, J.M. (1994) Lung cancer. ln: Doa Ri
Douglas, A.1. Omar, R.Z. & Smith, P.G. (1994) Cancer mortalìty Fraumeni, J.E, Jr & Muir, e.S. (eds.) Trends in Cancer Incidence
and morbidity among workers at the Se Ua field plant of British and Mortality, Cancer Surveys, VoL. 19/20, New York, Cold
Nuclear Fuels. Br. J. Cancer, 70, 1232-1243 Spring Harbor Laboratory Press, pp. 175-195
Draper, G.J, Birch, J.M., BitheU, J.E, Kinnier Wilson, L.M., Leck, Gore, S.M. (1981) Assessing methods - survivaL Br. Med. f., 283,
L, Marsdeni H.B., Morris Jones, P.H., Stiler, e.A. & Swindell, 840-843
R, (1982) Childhood Cancer in Britain: Incidence, Survival and Greenland, S. & Morgenstern, H. (1989) Ecological bias, con-
Mortality (Studies in Medical and Population Subjects No. 37), founding, and effect modification. Int. f. EpidemioI., 181
London, H.M.S.O. 269-274
429
References
Greenland, S. & Robins, 1.M. (1985) Estimation of a common Service to Obtain Information on the Extent of Agreemen
effect parame ter from sparse follow-up data. Biometries, 41, Between Clinical and Post-mortem Diagnoses (Studies or
55-68 Medical and Population Subjects No. 20), London
Greenland, S. & Robins, 1. (1994) Invited commentary: H.M.S.O.
Ecologie studies - biases, misconceptions, and counterex- Henderson, M.M. (1995) Nutritional aspects of breast can
amples. Am. f. Epidemiol., 139, 747-760 ceL Cancer, 76 (suppl. 10), 2053-2058
Greenland, S. & Thomas, D.C. (1982) On the need for the Hennekens, C.H. & Eberlein, K. (Physicians' Health Stud)
rare disease assumption in case-control studies. Am. f. Research Group) (1985) A randomized trial of aspirin ane
Epidemiol., 116, 547-553 ß-carotene among U.S. physicians. Prev. Med., 14, 165-16i
Grulich, A.E., Swerdlow, A.j. dos Santos Silva, I. & Beral, V. Hennekens, C.H., Buring, 1.E., Manson, J.E., Stampfer, M.
(1995) Is the apparent rise in cancer mortality in the elder- Rosner, B., Cook, N.R., Belanger, c., LaMotte, E, Gaziano
ly real? Analysis of changes in certification and coding of J.M., Ridker, P.M., Willett, W & Peto, R. (1996) Lack 0
cause of death in England and Wales, 1970-1990. Int. f. effect of long-term supplementation with beta caroten(
Cancer, 63, 164-168 on the incidence of malignant neoplasms and cardiovas
Gupta, P.c., Bhonsle, R.B., Mehta, ES. & Pindborg, J.1. (1984) cular disease. New f. Engl. Med., 334, 1145-1149
Mortality experience in relation to tobacco chewing and Herbst, A.L., Ulfelder, H. & Poskanzer, D.C. (1971
smoking habits from a 10-year follow-up study in Adenocarcinoma of the vagina. Association of materna
Ernakulam District, Kerala. ¡nt. f. Epidemiol., 13, 184-187 stilbestrol therapy with tumor appearance in youn¡
Gupta, P.c., Mehta, ES., Pindborg, J.., Aghi, M.B., Bhonsle, women. New Engl. f. Med., 284, 878-881
R.B., Daftary, D.K., Murti, P.R., Shah, H.T & Sinor, P.N. Hill, A.B. (1965) The environment and disease: associatior
(1986) Intervention study for primary prevention of oral or causation? Proc. R. Soc. Med., 58, 295-300
cancer among 36000 Indian tobacco users. Lancet, i, Hoffman, EL. (1915) The Mo rtali ty From Cancer Throughou
1235-1239 the World, Newark, NJ, Prudential Press
Gupta, P.c., Mehta, ES., Pindborg, 1.j. Daftary, D.K., Aghi, Howe, G.R., Burch, J.D., Miler, A.B., Morrison, B., Gordon
M.B., Bhonsle, R.B. & Murti, P.R. (1990) A primary preven- P., Weldon, L., Chambers, L.W, Fodor, G. & Winsor, G.M
tion study of oral cancer among Indian villagers. Eight- (1977) Artificial sweeteners and hum an bladder cancer
year follow-up results. ln: Hakama, M., Beral, v., Cullen, Lancet, 2, 578-581
1.W, Parkin, D.M. (1990) Evaluating Effectiveness ofPrimary IARC (International Agency for Research on Cancer) (1986
Prevention of Cancer (IARC Scientific Publications No. 103), IARC Monographs on the Evaluation of Carcinogenic Risks t(
Lyon, International Agency for Research on Cancer, pp. Humans. VoL. 38. Tobacco Smoking, Lyon, Internationa
149-156 Agency for Research on Cancer
Haenszel, W, Loveland, D. & Sirken, M.G. (1962) Lung can- IARC (International Agency for Research on Cancer) (1994a
cer mortality as related to residence and smoking histories. ¡ARC Monographs on the Evaluation of Carcinogenic Risks t(
f. Nat! Cancer Inst., 28, 947-1001 Humans. VoL. 61. Schistosomes, Liver Flukes am
Hakama, M. (1982) Trends in the incidence of cervical cancer Helicobacter pylori. Lyon, International Agency fOi
in the Nordic Countries. ln: Magnus, K. (ed.), Trends in Research on Cancer
Cancer Incidence. Causes and Practical Implications, New IARC (International Agency for Research on Cancer) (1994b
York, Hemisphere, pp. 279-292 IARC Monographs on the Evaluation of Carcinogenic Risks t(
Hakama, M., Hakulinen, T., Pukkala, E., Saxén, E. & Teppo, L. Humans. VoL. 59. Hepatitis Viruses, Lyon, Internationa
(1982) Risk indicators of breast and cervical cancer on eco- Agency for Research on Cancer
logic and individual levels. Am. f. Epidemiol., 116, Jensen, O.M., Parkin, D.M., MacLennan, R., Muir, C.S. E;
990-1000 Skeet, R.G. (eds.) (1991) Cancer Registration. Princip les am
Hakama, M., Miller, A.B. & Day, N.E. (eds.) (1986) Screening Methods (IARC Scientific Publications No. 95), Lyon
for Cancer of the Uterine Cervix (IARC Scientific Publications International Agency for Research on Cancer
No. 76), Lyon, International Agency for Research on Joint Iran-International Agency for Research on Canee)
Cancer, p. 289 Study Group (1977) Esophageal cancer studies in tht
Hakulinen, T. & Pukkala, E. (1981) Future incidence of lung Caspian littoral of Iran: results of population studies - ,
cancer: forecasts based on hypothetical changes in the prodrome. f. Natl. Cancer. Inst., 59, 1127-1138
smoking habits of males. ¡nt. f. Epidemiol., 10, 233-240 Jolley, D (1993) The glitter of the t table. Lancet, 342, 27-2Ç
Heasman, M.A. & Clarke, 1.A. (1979) Medical record linkage Jones, WH.S. (transI.) (1923) Hippocrates, VoL. l, London
in Scotland. Hea/th Bull., 37, 97-103 Wiliam Heinemann, pp. 71-13 7
Heasman, M.A. & Lipworth, L. (1966) Accuracy of Kaldor, J.M., Day, N.E., Bell, J., Clarke, E.A., Langmark, E
Certifìcation of Cause of Death. A Report on a Survey Karjalainen, S., Band, P., Pedersen, D., Choi, W., Blair, V.
Conducted in 1959 in 75 Hospitals of the National Hea/th Henry-Amar, M., Prior, P., Assouline, D., Pompe-Kim, V.
430
~\-......
References
Cartwright, R.A., Koch, M., Arslan, A., Fraser, P., Sutcliffe, Ley, c., Bauer, H.M., Reingold, A., Schiffman, M.H., Chambers,
S.B., Host, H., Hakama, M. & Stovall, M. (1992) Lung can-
J.C, Tashiro, CJ, & Manos, M.M. (1991) Determinants of gen-
cer following Hodgkin's disease: a case-control study. ¡nt. ital human papilomavirus infection in young women. J. Natl.
J. Cancer, 52, 677-681 Cancer ¡nst., 83, 997-1003
Khlat, M., Vail, A., Parkin, M. & Green, A. (1992) Mortality Lilienfeld, A.M. & Lilienfeld, D.E. (1979) A century of case-con-
from melanoma in migrants to Australia: variation by age at trol studies: progress? J. Chron. Dis., 32, 5-13
arrivaI and duration of stay. Am. J. Epidemiol., 135, Lippman, S.M., Lee, J,S., Lotan, R., Hittelman, w', Wargovich,
1103-1113 M,j & Hong, w'K. (1990) Biomarkers as intermediate end
Kirkwood, B.R. (1988) Essentials of Medical Statistics, London, points in chemoprevention trials. J. Nat!. Cancer ¡nst., 82,
Blackwell Science 555-560
Kiviat, N.B., Koutsky, L.A., Critchlow, C w" Galloway, DA, Long, E.R. (1928) A History of Pathology, London, Bai1ière, Tindall
Vernon, DA, Peterson, M.L., McElhose, P.E., Pendras, S,J, & Cox
Stevens, C.E. & Holmes, K.K. (1990) Comparison of Southern MacLennan, R. (1991) Items of patient information which may
transfer hybridization and dot filter hybridization for detec- be collected by registries. ln: Jensen, O.M., Parkin, D.M.,
tion of cervical human papilomavirus infection with types MacLennan, R., Muir, CS. & Skeet, R.G. (eds.) Cancer
6,11,16,18,31,33, and 35. Am. J. Clin. Patho!., 94, 561-565
les and Methods (IARC Scientific
Registration: Princip
Kjaer, S.K., de Viliers, E.-M., Haugaard, B,J, Christensen, R.B., Publications, No. 95), Lyon, Intemational Agency for Research
Teisen, C, Møller, KA, Poli, P., Jensen, H., Vestergaard, B.F., on Cancer, pp. 43-63
Lynge, E. &Jensen, O.M. (1988) Human papilomavirus, her- MacMahon, B., Cole, E, Lin, T.M., Lowe, CR., Mina, A.P.,
pes simplex virus and cervical cancer incidence in Greenland Ravnihar, R, Salber, E,j, Valaoras, v.G. & Yuasa, S. (1970) Age
and Denmark. A population-based cross-sectional study. ¡nt. at first birth and breast cancer risk. BulL. World Health Organ.,
J. Cancer, 41, 518-524 43,209-221
Kubik, A., Parkin, D.M., Khlat, M., Erban, j. Polak, J, & Malaty, H.M., Kim, J,G., Kim, S.D. & Graham, D.Y. (1996)
Adamec, M. (1990) Lack of benefit from semi-annual screen- Prevalence of Helicobacter pylori infection in Korean children:
ing for cancer of the lung: follow-up report of a randomised inverse relation to socioeconomic status despite a uniformly
controlled trial on a population of high-risk males in high prevalence in adults. Am. J. Epidemio!., 143,257-262
Czechoslovakia. ¡nt. J. Cancer, 45, 26-33 Mantel, N. & Haenszel, W, (1959) Statistical aspects of the analy-
Kurihara, M., Aoki, K. & Hisamichi, S. (eds.) (1989) Cance sis of data from retrospective studies of disease. J. Nat!. Cancer
Mortality Sta tis tics in the World 1950-1985, Nagoya, ¡nst., 22, 719-748
University of Nagoya Press Mayans, M.V., Hall, A,j, Inskip, H.M., Lindsay, S.w', Chotard,1.
Läärä, E. (1982) Development of Cancer Morbidity in Finland up to Mendy, M. & Whittle, H.C (1994) Do bedbugs transmit hepati-
the Year 2002 (in Finnish) (Health Education Series, Original tis B? Lancet, 343, 761-763
Reports 3/1982), Helsinki, National Board of Health McMichael, A,j (1976) Standardized mortality ratios and the
Landis, J.R. & Koch, G.G. (1977) The measurement of observer "healthy worker effect": scratching beneath the surface. J.
agreement for categorical data. Biometries, 33, 159-174 Occup. Med., 18, 165-168
Lane-Claypon, J,E. (1926) A Further Report on Cancer of the McMichael, A,j, Spirtas, R. & Kupper, L.L. (1974) An epidemio-
Breast, with Special Reference to its Associated Antecedent
logic study of mortality within a cohort of rubber workers,
Conditions (Reports on Public Health and Medical Subjects 1964-72 J. Occup. Med., 16,458-464
No. 32), London, H.M.S.O. Meirik, O., Lund, E., Adami, H.-O., Bergström, R., Christoffersen,
Lantz, CA. & Nebenzahl, E. (1996) Behavior and interpretation T. & Bergsjö, P. (1986) Oral contraceptive use and breast cancer
of the kappa statistic: resolution of the two paradoxes. J. Clin. in young women. A joint national case-control study in
Epidemio!., 49, 431-434 Sweden and Norway. Lancet, 2, 650-54
Last, J.M. (1995) A Dictionary of Epidemiology, third edition,
Miettnen, O. (1976) Estimabilty and estimation in case-referent
Oxford, Oxford University Press studies. Am. J. Epidemiol., 103, 226-235
Lee, H.P., Day, N.E. & Shanmugaratnam, K. (1988) Trends in Miettinen, O.S. & Wang, J.-D. (1981) An alternative to the pro-
Cancer Incidence in Singapore 1968-1982 (IARC Scientific portionate mortality ratio. Am. J. Epidemiol., 114, 144-148
Publications No. 91), Lyon, International Agency for Møller, R, Knudsen, L.B. & Lynge, E. (1994) Risk of testicular can-
Research on Cancer cer after vasectomy: cohort study of over 73 000 men. Br. Med.
Lee, H.P., Gourley, L., Duffy, S.w" Estève, j. Lee., J, & Day, N.E.
J., 309, 295-299
(1991) Dietary effects on breast-cancer risk in Singapore. Monison, A.S., Buring, J.E., Verhoek, w'G., Aoki, K., Leck, L,
Lancet, 337, 1197-1200 Ohno, Y. & Obata, K. (1984) An international study of smoking
Levi, M.L, Goldstein, H. & Gerhardt, P.R. (1950) Cancer and
and bladder cancer. J. Urology, 131,650-54
tobacco smoking. A preliminary report. J. Am. Med. Assoc., 143, Mould, R.E (1983) Cancer Statistics (Medical Sciences Series),
336-338 Bristol, Adam Hilger
431
References
Mueller, N., Evans, A., Harris, N.L., Comstock, G.W., Jellum, OPCS (Offce of Population Censuses and Surveys) (1983a,
E., Magnus, K., Orentreich, N., Polk, B.F. & Vogelman, J. 1983b, 1985a, 1985b, 1986, 1988, 1990, 1991, 1993, 1994a,
(1989) Hodgkin's disease and Epstein-Barr virus. Altered 1994b) Cancer Statistics. Registrations (Series MB1 Nos. 11-16,
antibody pattern before diagnosis. New Engl. f. Med., 320, 18-22), London, H.M.S.O.
689-695 OPCS (Office of Population Censuses and Surveys) (1995)
Muir, C.S. & Percy, C. (1991) Classification and coding of neo- Mortality Statistics. Review of the Registrar General on Deaths by
plasms. ln: Jensen, O.M., Parkin, D.M., MacLennan, R., Cause, Sex and Age, in England and Wales, 1993 (Series DH2 No.
Muir, C.S. & Skeet, R.G. (eds.) Cancer Registration. Princip les 20), London, H.M.S.O.
and Methods (IARC Scientific Publications, No. 95), Lyon, Omenn, G.S., Goodman, G.E., Thornquist, M.D., Balmes, J.
International Agency for Research on Cancer, pp. 64-81 Cullen, M.R., Glass, A., Keogh, J.P., Meyskens, EL., Valanis, B.,
Muir, c., Waterhouse, J. Mack, T, Powell, J. & Whelan, S. Wiliams, J.H., Barnhart, S. & Hammar, S. (1996) Effects of a
(eds.) (1987) Cancer Incidence in Five Continents. VoL. V (lARC combination of beta carotene and vitamin A on lung cancer
Scientific Publications, No. 88), Lyon, International Agency and cardiovascular disease. New. Engl. f. Med., 334, 1150-1155
for Research on Cancer Orr, LM. (1933) Oral cancer in betel nut chewers in Travancore.
Muñoz, N. & Day, N.E. (1996) Oesophageal cancer. ln: Its aetiology, pathology and treatment. Lancet, ii, 575-580
Schottenfeld, D., Fraumeni Jr., J.F. (eds.) Cance Epidemiology Parkin, D.M. & Hakulinen, T (1991) Analysis of survival. ln:
and Prevention, second edition. Oxford, Oxford University Jensen, O.M., Parkin, D.M., MacLennan, R., Muir, C.S. & Skeet,
Press, pp. 681-706 R.G. (eds.) Cancer Registration: Princip les and Methods (IARC
Muñoz, N., Wahrendorf, J., Bang, L.J. Crespi, M., Thurnham, Scientific Publications No. 95), Lyon, International Agency for
D.L, Day, N.E., Ji, l.H., Grassi, A., Yan, L.w., Lin, L.G., Research on Cancer, pp. 159-176
Quan, L.Y, Yun, l.C., Fang, L.S., Yao, L.J., Correa, P., Parkin, D.M. & Khlat, M. (1996) Studies of cancer in migrants:
O'Conor, G.T & Bosch, X. (1985) No effect of riblofavine, rationale and methodology. Eur. f. Cancer, 32A, 761-771
retinol, and zinc on prevalence of precancerous lesions of Parkin, D.M., Muir, C.S., Whelan, S.L., Gao, YT, Ferlay, J. &
oesophagus. Randomised double-blind intervention study Powell, J. (1992) Cancer Incidence in Five Continents. VoL. VI
in high-risk population of China. Lancet, 2, 111-114 (lARC Scientific Publications No. 120), Lyon, International
Muñoz, N., Bosch, EX., de Sanjosé, S., Ta fur, L., Izarzugaza, L, Agency for Research on Cancer
Gili, M., Viladiu, P., Navarro, c., Martos, c., Ascunce, N., Parkin, D.M., Pisani, P. & Ferlay, J. (1993) Estimates of the world-
Gonzalez, L.c., Kaldor, J.M., Guerrero, E., Lörincz, A., wide incidence of eighteen major cancers in 1985. Int. f.
Santamaria, M., Alonso de Ruiz, P., Aristizabal, N. & Shah, K. Cancer, 54, 594-606
(1992a) The causallink between human papilomavirus and Parkin, D.M., Chen, V.w., Ferlay, J. Galceran, J. Storm, H.H. &
invasive cervical cancer: a population-based case-control Whelan, S.L. (1994) Comparability and Quality ControL in Cancer
study in Colombia and Spain. Int. f. Cancer, 52, 743-749 Registration (IARC Technical Reports No. 19), Lyon,
Muñoz, N., Bosch, EX., Shah, K.V. & Meheus, A. (eds.) (1992b) International Agency for Research on Cancer
The Epidemiology of Cervical Cancer and Human Parkin, D.M., Whelan, S.L., Ferlay, J., Raymond, L. & Young, J.
Papilomavirus (IARC Scientific Publications, No. 119), Lyon, (1997) Cancer Incidence in Five Continents. VoL. VII (IARC
International Agency for Research on Cancer Scientific Publications No. 143), Lyon, International Agency
Newton, R., Ferlay, J. Reeves, G., Beral, V. & Parkin, D.M. for Research on Cancer
(1996) Effect of ambient solar ultraviolet radiation on inci- Pearce, N., Matos, E., Vainio, H., Boffetta, P. & Kogevinas, M.
dence of squamous-cell carcinoma of the eye. Lancet, 347, (eds) (1994) Occupational Cancer in Developing Countries (IARC
1450- 1 451 Scientifc Publications No. 129), Lyon, International Agency
Nomura, A., Stemmermann, G.N., Chyou, P.H., Kato, L, Perez- for Research on Cancer
Perez, G.L & Blaser, M,J (1991) Helicobacter pylori infection Percy, C. L. (ed.) (1980) Conversion of Malignant NeopLasms by
and gastric carcinoma among Japanese Americans in Topography and MorphoLogy (rom the International
Hawaii. New Engl. f. Med., 325, 1132-1136 Classification of Diseases for Oncology (lCD-O) to Chapter II,
OPCS (Office of Population Censuses and Surveys) (1970) MaLignant Neoplasms, International Classification of Diseases
Classifìcation of Occupations, 1970. London, H.M.S.O. (ICD-8), 1965 (NIH Publication No. 80-2136), Bethesda, MD,
OPCS (Office of Population Cens uses and Surveys) (1978) US Department of Health, Education and Welfare
Occupational Morta ii Percy, C. L. (ed.) (1981) Conversion of Neoplasm Section, 8th
ty. The Registrar Generals Decennial
Supplement for England and Wales, 1970-72 (Series DS, No. Revision ofInternational Classification of Diseases (1965): and
1), London, H.M.S.O. 8th revis ion International Classification of Diseases Adapted for
OPCS (Office of Population Cens uses and Surveys) (1981) Use in the United States, to Neoplasm Section, 9th Revision of
Cancer Statistics. Incidence, Survival and Mortality in England International Classification of Diseases (1975) (NIH
and Wales (Studies on Medical and Population Subjects, No. Publication No. 83-2448), Bethesda, MD, US Department of
43), London, H.M.S.O. Health and Human Services
432
References
Percy, c.L. (ed.) (1983) Conversion of Neoplasm Section, 9th Prince, A.M., Szmuness, W, Michon, J. Demaile, J. Diebolt,
Revision ofInternational Classification of Diseases (1975): to G., Linhard, J. Quenum, C. & Sankale, M. (1975) A case/con-
Neoplasm Section, 8th Revision of International Classification trol study of the association between primary liver cancer
of Diseases (1965) and 8th Revision of International and hepatitis B infection in Senegal. Int. J. Cancer, 16,
Classification of Diseases, Adapted for Use in the United States 376-383
(NIH Publication No. 83-2638), Bethesda, MD, US Puffer, R.R. & Griffth, G. W (1967) Patterns of Urban Mortality.
Department of Health and Human Services Report of the Inter-American Investigation ofMortality (Scientìfc
Percy, c.L., Berg, j.W & Thomas, L.B. (eds.) (1968) Manual of Publication No. 151), Washington, DC, Pan American Health
Tumour Nomenclature and Coding, second edition. Organization
Washington, DC, American Cancer Society Qian G.-S., Ross, R.K, Yu, M.C., Yuan, 1.-M., Gao, Y-T.,
Percy, c.L. & van Holten, V. (eds.) (1979) Conversion of Henderson, B.E., Wogan, G.N. & Groopman, 1.D. (1994) A
Neoplasms by Topography and Morphology from the fol1ow-up study of urinary markers of aflatoxin exposure and
International Classifiction of Diseases for Oncology (ICD-O): liver cancer risk in Shanghai, People's Republic of China.
to Chapter II, Neoplasms, 9th Revision of the International Cancer EpidemioL. Biomarkers Prev., 3, 3-10
Classification of Diseases (NIH Publication No. 80-2007), Rodrigues, L. & Kirkwood, B.R. (1990) Case-control designs in
Bethesda, MD, US Department of Health, Education and the study of common diseases: Updates on the demise of the
Welfare, Public Health Service rare disease assumption and the choice of sampling scheme
Percy, c., van Holten, V. & Muir, C. (eds.) (1990) International for controls. Int. J. Epidemiol., 19, 205-213
Classification of Diseases for Oncology (ICD-O), second edition. Romieu, L, Wìlett, Wc., Colditz, G.A., Stampfer, M.f, Rosner,
Geneva, World Health Organization B., Hennekens, C.H. & Speizer, EE. (1989) Prospective study
Peto, R., Pike, M.C., Armitage, P., Breslow, N.E., Cox, D.R., of oral contraceptive use and risk of breast cancer in women.
Howard, S.V., Mantel, N., McPherson, K., Peto,
1. & Smith, J. Natl. Cancer Inst., 81, 1313-1321
P.G. (1976) Design and analysis of randomised clinical trials Rose, G. (1985) Sìck individuals and sìck populations. ¡nt. J.
requiring prolonged observation of each patient. I. Epidemiol., 14,32-38
Introduction and design. Br. J. Cancer, 34, 585-612 Rose, G. & Colwel1, L. (1992) Randomised control1ed trial of
Peto, R., Pike, M.C., Armitage, P., Breslow, N.E., Cox, D.R., anti-smoking advice: final (20 year) results. J. EpidemioL.
Howard, S.V., Mante!, N., McPherson, K., Peto,
1. & Smith, Comm. Health, 46, 75-77
P.G. (1977) Design and analysis of randomised clinìcal trials Ross, DA & Vaughan, J.P. (1986) Health interview surveys in
requiring prolonged observation of each patient. II. Analysis developing countries: a methodologìcal review. Stud. Fam.
and examples. Br. J. Cancer, 35, 1-39 Plann., 17, 78-94
Pike, M.C., Henderson, B.E., Krailo, M.D., Duke, A. & Roy, S. Ross, R.K., Yuan, 1.-M."Yu, M.C., Wogan, G.N., Qian, G.-S., Tu,
(1983) Breast cancer in young women and use of oral con- j.-T., Groopman, J.D., Gao, Y-T. & Henderson, B.E. (1992)
traceptives: possible modifying effect of formulation and age Urinary aflatoxin biomarkers and risk of hepatocel1ular car-
at use. Lancet, 2, 926-930 cinoma. Lancet, 339, 943-946
Pocock, S.J. (1983) Clinical Trials: A Practical Approach, Sackett, D.L. (1979) Bias in analytic research. J. Chrono Dis., 32,
Chichester, John Wiley 51-63
Polednak, A.P., Stehney, A.E & Rowland, R.E. (1978) Mortality Schatzkin, A., Freedman, L.S., Schiffman, M.H. & Dawsey, S.M.
among women first employed before 1930 in the U.S. radi- (1990) Validation of intermadiate end points in cancer
um dial-painting industry. A group ascertained from employ- research. J. Natl. Cancer.¡nst., 82, 1746-1752
ment lists. Am. J. Epidemiol., 107,179-195 Schrek, R. & Lenowitz, H. (1947) Etiologie factors in carcinoma
Potosky, A.L., Kessler, L., Gridley, G., Brown, c.c. & Horm, 1. W of the penis. Cancer Res., 7, 180-187
(1990) Rise in prostatic cancer incidence associated with Schrek, R., Baker L.A., Ballard, G.P. & Dolgoff, S. (1950) Tobacco
increased use of transmethral resection. J. NatL. Cancer Inst., smoking as an etiologic factor in disease. I. Cancer. Cancer
82, 1624-1628 Res., 10, 49-58
Pott, P. (1775) Chirurgical Observations. VoL 3, London, Hawes, Schwartz, D. & Lel1ouch, 1. (1967) Explanatory and pragmatic
L., Clark, W, Collns, R., pp. 177-183. attitudes in therapeutical trials. J. Chrono Dis., 20, 637-648
Pottern, L.M., Stone, B.J. Day, N.E., Pickle, L.W. & Fraumeni, Scott, C. & Harpham, T. (1987) Sample design. ln: Cleland,1. &
J.E, Jr (1980) Thyroid cancer in Connecticut, 1935-1975: an Scott, C. (eds.), The World Fertility Survey. An Assessment,
analysis by cell tye. Am. J. Epidemiol., 112, 764-774 Oxford, Oxford University Press, pp. 309-324
Powles, T.J., Jones, A.L., Ashley, S.E., O'Brien, M.E.R., Tidy, V.A., Severson, R.K (1990) Have transurethral resections contributed
Treleavan, J., Cosgrove, D., Nash, A,G., Sacks, N., Baum, M., to the increasing incidence of prostatic cancer? J. NatL. Cancer
McKinna, j.A. & Davey, 1.B. (1994) The Royal Mardsen Inst., 82, 1597-1598
Hospital pilot tamoxifen chemoprevention triaL. Breast Shapiro, S. (1977) Evidence on screening for breast cancer from
Cancer Res. Treat., 31, 73-82 a randomised triaL. Cance, 39 (suppL 6), 2772-2782
433
'\ """
References
Shapiro, S., Strax, P. & Venet, L. (1971) Periodic breast cancer Swerdlow, Al (1989) Interpretation of England and Wales cancer
screening in reducing mortality from breast cancer. 1. Am. Med. mortality data: the effect of enquiries to certifiers for further
Assoc., 215, 1777-1785 information. Br. 1. Cancer, 59, 787-791
Shapiro, S., Goldberg, J.O. & Hutchison, G.B. (1974) Lead time in Swerdlow, Al, Marmot, M.G., Grulich, A.E. & Head, J. (1995)
breast cancer detection and implications for periodicity of Cancer mortality in Indian and British ethnic immigrants from
screening. Am. 1. Epidemiol., 100,357-366 the Indian subcontinent to England and Wales. Br. 1. Cancer, 72,
Shimizu, Y., Schull, Wl & Kato, H. (1990) Cancer risk among 1312-1319
atomic bomb survivors. The RERF Life Span Study. 1. Am. Med. Tabár, L., Fagerberg, e.J.G., Gad, A., Baldetorp, L. Holmberg, L.H.,
Assoc., 264, 601-604 Gröntoft, O., Ljungquist, U., Lundström, B., Månson, lC,
Shimizu H., Ross, R.K., Bernstein, L., Yatani, R, Henderson, B.E. & Eklund, G., Day, N.E. & Pettersson, F. (1985) Reduction in mor-
Mack, T.M. (1991) Cancers of the prostate and breast among tality from breast cancer after mass screening with mammogra-
japanese and white immigrants in Los Angeles County. Br. 1. phy. Randomised trial from the Breast Cancer Screening
Cancer, 63, 963-966 Working Group of the Swedish National Board of Health and
Shu, X.O., Nesbit, M.E., Buckley, j.D., Krailo, M.D. & Robison, L.L. Welfare. Lancet, 1,829-832
(1995) An exploratory analyis of risk factors for childhood Thames Cancer Registry (1995) Cancer in South East England 1992.
rnalignant germ-cell tumors: report from the Childrens Cancer Cancer Incidence, Prevalence, Swvival and Treatment for Residents of
Group (Canada, United States) Cancer Causes Control, 6, the DistJict Health Authorities in South East England, Sutton,
187-198 Thames Cancer Registry
Smith, P.G. & Day, N.E. (1984) The design of case-control studies: Tomášek, L., Darby, S.e., Swerdlow, Al, Placek, V. & Kunz, E.
the influence of confounding and interaction effects. ¡nt. 1. (1993) Radon exposure and cancers other than lung cancer
Epidemiol., 13, 356-365 among uranium miners in West Bohemia. Lancet, 341, 919-923
Smith, P.G. & Douglas, Al (1986) Mortality of workers at the Tomášek, L., Swerdlow, Al, Darby, S.e., Placek, V. & Kunz, E.
Sellafield plant of British Nuclear Fuels. Br. Med. 1., 293, (1994) Mortality in uranium miners in West Bohemia: a long
845-854 term cohort study. Occup. Environ. Med., 51, 308-315
Smith, P.G. & Morrow, RH. (eds.) (1996) Field Trials of Health Toniolo, P., Boffetta, P., Shuker, D.E.G., Rothman, N., Hulka, B. &
¡nteiventions in Developing Countres: A Toolbox, second edition. Pearce, N. (1997) Application ofBiomarkers in Cancer Epidemiology
London, Macmillan Education (IARC Scientific Publications No. 142), Lyon, Intemational
Smith, D.e., Prentice, R., Thompson, Dl & Herrmann, w.L. Agency for Research on Cancer
(1975) Association of exogenous estrogen and endometrial car- Trichopoulos, D., Tabor, E., Gerety, R.j., Xirouchaki, E., Sparros, L.,
cinoma. New Engl. 1. Med., 293, 1164-1167 Muñoz, N. & Linsell, CA (1978) Hepatitis Band primary hepa-
Smith, P.G., Rodrigues, L.e. & Fine, P.E.M. (1984) Assessment of tocellular carcinoma in a European population. Lancet, ii,
the protective efficacy of vaccines against common diseases 1217-1219
using case-control and cohort studies. ¡nt. 1. Epidemio!., 13, Tufte, E.R (1983) The Visual Display of Quantitative Information,
87-93 Cheshire, CT, Graphic Press
Smith, D.B., Newlands, E.S., Rustin, GlS., Begent, R.Hl, Howells, Tuyns, Al & Massé, L.M.F., (1973) Mortality from cancer of the
N., McQuade, B. & Bagshawe, K.O. (1991) Comparison of oesophagus in Brittany. Int. 1. Epidemio!., 2, 241-245
ondansetron and ondansetron plus dexamethasone as Tuyns, Al & Vernhes, j.e. (1981) La mortalité par cancer de l'oe-
antiemetic prophylaxis during cisplatin-containing chemother- sophage dans les départements du Calvados et de l'Orne.
apy. Lancet, 338, 487-490 Gastroenterol. Clin. Bio!., 5, 257-265
Sparén, P. (1996) Early Detection and Screening for Cancer of the Tuyns, AL, Pequignot, G. &jensen, O.M. (1977) Le cancer de l'oe-
Cervix in Sweden during the 20th Century (Doctoral thesis. sophage en Ile-et-Vilaine en fonction des niveaux de consom-
University of Uppsala, Sweden), Stockholm, Almqvist & Wiksell mation d'alcool et de tabac. Des risques qui se multiplient. Bull.
Sriamporn, S., Black, Rl, Sankaranarayanan, R, Kamsa-AD, S., Cancer, 64, 45--0
Parkin, D.M. & Vatanasapt, V. (1995) Cancer survival in Khon Tuyns, Al, Hu, M.X. & Pequignot, G. (1983) Alcohol consump-
Kaen Province, Thailand. ¡nt. 1. Cancer, 61, 296-300 tion patterns in the department of Calvados (France) Rev.
Stadel, ß.v., Rubin, G.L., Webster, L.A, Schlesselman, j. & Epidémiol. Santé Publique, 31, 179-197
Wingo, PA (1985) Oral contraceptives and breast cancer in United Nations (1987) Fertility Behaviour in the Context of
young women. Lancet, 2, 970-973 Development (Population Studies No. 100), New York, United
Stampfer, Ml, Buring, J.E., Wilet, w., Rosner, B., Eberlein, K. & Nations, Department of International Economic and Social
Hennekens, e.H. (1985) The 2x2 factorial design: its applica- Mfairs
tion to a randomzied trial of aspirin and carotene in U.S. physi- Uppsala-Örebro Breast Cancer Study Group (1990) Sector resec-
cians. Stat. Med., 4,111-116 tion with or without postoperative radiotherapy for Stage l
Strader, CH., Weiss, N.S. & Daling, J.R. (1988) Vasectomy and the breast cancer: A randomized triaL. J. Nat!. Cancer ¡nst., 82,
incidence of testicular cancer. Am. J. Epidemiol., 128, 56-63 277-282
434
References
US Department of Health, Education, and Welfare (1969) Use of Weinstock, MA, Colditz, G.A., Willett, WC, Stampfer, M.J.,
Tobacco: Practices, Atttudes, Knowledge, and Beliefs. United States Rosner, B. & Speizer, EE. (1991) Recall (report) bias and reli-
- Fall1964 and Spring 1966, Washington, DC, US Department ability in the retrospective assessment of melanoma risk.
of Health, Education, and Welfare Am. f. Epidemio!., 133, 240-245
US Department of Health and Human Services (1991) Smoking Whelan, S.L., Parkin, D.M. & Masuyer, E. (eds.) (1990) Patterns
and Tobacco Control Monographs 1. Strategies to Control Tobacco of Cancer in Five Continents (lARC Scientific Publications
Use in the United States: A Blueprint for Public Health Action in the No. 102), Lyon, International Agency for Research on
1990's (NIH Publication No. 92-3316), Bethesda, MD, US Cancer
Department of Health and Human Services, US Government WHO (World Health Organization) (1956) Statistical Code for
Printing Office Human Tumours (WHO/HS/CANC, 24.1 and 24.2, August
US Surgeon General (1982) The Health Consequences of Smoking: 15, 1956), Geneva, WHO
Cancer, Bethesda, MD, US Department of Health and Human WHO (World Health Organization) (1957) Manual of the
Services International Statistical Classification of Diseases, Injuries, and
Veronesi, U., Saccozzi, R., Del Vecchio, M., Banfi, A., Clemente, Causes of Death, seventh revision, Geneva, WHO
C, De Lena, M., Gallus, G., Greco, M., Luini, A., Marubini, E., WHO (World Health Organization) (1967) Manual of the
Muscolino, G., Rilke, E, Salvadori, B., Zecchini, A. & Zucali, R. International Statistical Classification of Diseases, Injuries, and
(1981) Comparing radical mastectomy with quadrantectomy, Causes of Death, eighth revis ion, Geneva, WHO
axilary dissection, and radiotherapy in patients with small WHO (World Health Organization) (1976) International
cancers of the breast. New Eng!. f. Med., 305, 6-11 Classification of Diseases for Oncology (ICD-OY, first edition,
Vineis, P. & Simonato, L. (1986) Estimates of the proportion of Geneva, WHO
bladder cancers attributable to occupation. Scand. f. Work WHO (World Health Organization) (1977) Manual of the
Environ. Health, 12, 55-60 International Statistical Classification of Diseases, Injuries, and
Vizcaino, A.P., Parkin, D.M. & Skinner, M.E.G. (1995) Risk factors Causes of Death, ninth revision, Geneva, WHO
assocìated with oesophageal cancer in Bulawayo, Zimbabwe. WHO (World Health Organization) (1983) World Health
Br. f. Cancer, 72, 769-773 Statistics Annual, 1983, Geneva, WHO
Wagner, G. (1991) History of cancer registration. ln: Jensen, O.M., WHO (World Health Organization) (1986) World Health
Parkin, D.M., MacLennan, R., Muir, CS. & Skeet, R.G. (eds.) Statistics Annual, 1986, Geneva, WHO
les and Methods (IARC Scìentific
Cancer Registration: Princip WHO (World Health Organization) (1992) International
Publications, No. 95), Lyon, International Agency for Research Statistical Classification of Diseases and Related Health
on Cancer, pp. 3-6 Problems, tenth revision (ICD-lO), Geneva, WHO
Waldron, HA (1983) A brief history of scrotal cancer. Br. f. Ind. WHO Collaborative Study. of Neoplasia and Steroid
Med., 40, 390-401 .., Contraceptives (1989) Combined oral contraceptives and
Walter, S.D. (1991a) The ecologic method in the study of envi- liver cancer. Int. f. Cancer, 43, 254-259
ronmental health. 1. Overview of the method. Environ. Health. Wilett, WC, Stampfer, M,J, Colditz, G.A., Rosner, B.A.,
Perspect., 94, 61-65 Hennekens, CH. & Speizer, EE. (1987) Dietary fat and the
Walter, S.D. (1991b) The ecologic method in the study of envi- risk of breast cancer. New Eng!. f. Med., 316, 22-28
ronmental health. II. Methodologie issues and feasibilty. Wright, Wc. (transI.) (1964) Diseases of Workers, by
Environ. Health. Perspect., 94, 67-73 Ramazzini, B. (The History of Medicine Series No. 23), New
Waterhouse, JAH., Doll, R. & Muir, CS. (eds.) (1970) Cancer York, Hafner
Incidence in Five Continents. VoL. II, Berlin, Springer Wynder, E.L. & Graham, E.A. (1950) Tobacco smoking as a
Waterhouse, JAH., Muir, CS. & Corra, P. (eds.) (1976) Cancer possible etiologic factor in bronchiogenic carcinoma. A
Incidence in Five Continents. VoL. II (IARC Scìentific Publications study of six hundred and eighty-four proved cases. f. Am.
No. 15), Lyon, International Agency for Research on Cancer Med. Assoc., 143, 329-336
Waterhouse, j.A.H., Muir, CS., Shanmugaratnam, K. & Powell, j. Ziel, H.K. & Finkle, WD. (1975) Increased risk of endometrial
(eds.) (1982) Cancer llUidence in Five Continents. VoL. iv (IARC carcinoma among users of conjugated estrogens. New Eng!.
Scientific Publications No. 42), Lyon, International Agency for f. Med., 293, 1167-1170
Research on Cancer
435
Subject index
437
1 \._",.
j
1
Subject index
438
¡
f
,,
i
í
Subject index
439
Subject index
440
Subject index
441
Subject index
442