Cohort studies

1
Cohort studies

• Design Concepts in Nutritional Epidemiology
• Barrie M. Margetts and Michael Nelson

2

• Contents
• General considerations
• Practical issues
• Infancy and childhood
• Adults
• The elderly
• Concluding remarks
• Appendix
• Checklist for planning a cohort study
• Analysis of cohort studies

3
General considerations

• A cohort
• A group of persons, identified at one
  point in time, who march off together into the
  future under the watchful eye of an investigator.
  
• A cohort study
• A group of persons is defined, certain
  characteristics about each individual are
  recorded, and they are then followed up in such a
  way that new events (such as disease and death)
  or other changes in their characteristics are
  detected.
• These new events and changes can then be
  related to the original observations in order to
  discover what aspects of the initial status of
  the subjects predict their subsequent experience.
  

4

• Two main types of cohort study
• Prospective (more usual type)
• --baseline information is collected
  when subjects enter the study and they are
  followed up over its duration
• Retrospective
• --- a 'historical cohort' is identified
  with reference to some point in the past
• --- This approach is possible only
  where adequate records exist it has the
  advantage that the effects of a long period of
  time can be observed during a relatively short
  period of study.

5
The use of the cohort study

• In the elucidation of aetiology.
• In studying the natural history of a disease, and
  how this is modified by treatment.

6
Reasons for preferring the cohort type of study
to the case-control approach

• It enables the investigator to obtain accurate
  information about the individuals before the
  onset of the disease being investigated.
• Information obtained prospectively is not only
  more accurate but is also less open to bias than
  that obtained retrospectively, when the outcome
  is known.
• It allows us to detect unexpected effects of the
  initial factors,whereas a case-control analysis
  is restricted to the selected condition.
• The findings show the strength of an association
  in public health terms, so that the 'population
  attributable risk' can be calculated.

7
The disadvantages of the cohort study

• Prospective aetiological studies are inevitably
  large, long, and expensive, even for the more
  common diseases, and not usually feasible for
  less common diseases.
• It may happen that the subjects alter their
  eating habits, either spontaneously or as a
  result of changes in dietary fashion, to such an
  extent that the original observations are
  irrelevant.
• The difficulties of tracing all the subjects,
  unforeseen changes in the personnel involved in
  the study, and the ultimate hazard that the
  hypothesis being tested becomes superseded, so
  that the investigator finds that the wrong
  baseline.information has been recorded.

8

• Occasionally, a 'nested' design is used,
  combining the cohort and case-control approaches.
  The sensitivity of the comparison is not much
  less than it would be if the specimens from all
  the subjects were used, and the expense of
  analysis is greatly reduced. The advantages of
  the cohort methods are thus retained, together
  with the greater efficiency of the case-control
  design.
• Ex. If a large cohort is being followed up,
  and a sample of every member's serum as been
  stored, a case-control analysis can be made
  within the cohort at a later date. Persons who
  acquire a given disease are identified, and
  controls (usually two or three per case) are
  selected from the cohort, each control being
  matched with a case in respect of age and gender.
  The sera of cases and controls are extracted and
  appropriate biochemical analyses are made.

9
Practical issues

• To obtain expert statistical advice at the
  planning stage so as to ensure that the study is
  likely to be large and long enough but not
  greatly in excess of the requirements.
• Calculations of size and duration will be based
  on the expected numbers of endpoints in the study
  (see appendix 14.1).
• Suitable forms are designed on which information
  can be collected and coded for easy
  computerization.
• To conduct a short pilot study to test the
  procedures that will be used in the survey.
• Information can be obtained from a sub-sample and
  will provide some estimate of the stability of
  the original measurements in the group being
  studied.

10
Practical issues

• To acquire information during the whole of the
  follow-up period, contact will then have to be
  made with the subject periodically, either by
  visiting or by telephone. If day-to-day
  information is required (e.g. on infant feeding
  methods) a diary may be used.
• Some of the subjects will have moved or died, and
  information may not be easily obtained,
  particularly in a free-living population. Plans
  should be made at the start of the study about
  how the subjects will be traced.
• It may be advisable to have periodic contact with
  all the subjects (say at annual intervals), by
  telephone or reply-paid cards, simply in order to
  detect those who have left the area.

11

• Another way of minimizing the difficulties of
  finding subjects is to select them from certain
  occupations whose members are particularly easy
  to trace.
• Ex. Medical practitioners ,Civil Servants
• The main drawback of these groups is that
  they do not entirely represent the population at
  large, but are drawn from specific social
  classes. Furthermore, people who are employed or
  who volunteer always tend to be healthier than
  others of their age-group , so that the cohort
  may need to be larger. And the results may not
  apply exactly to the wider population.
• If the intention is to look for changes in
  biochemical variates (e.g. serum cholesterol),
  the study should be discussed with the biochemist
  at the start so as to avoid a change in
  laboratory methodology, which will invalidate
  comparisons.

12

• Changes in staff or collaborators may be more
  difficult to foresee and can have a disastrous
  effect if the newcomer does not have the same
  degree of interest or commitment as the person
  with whom the study was set up.
• If the final measurements (e.g. of blood
  pressure) are to be made by someone other than
  the person who made the initial measurements,
  attention should be paid to standardization and
  comparability.
• It may also be desirable for the person recording
  the final measurements or endpoints to be unaware
  of the initial data, so as to avoid bias.

13
Infancy and childhood

• Infancy provides perhaps the best possible
  opportunity for cohort studies.
• A cohort of babies can easily be defined in terms
  of place and time of birth. Some information
  (e.g. regarding birthweight and initial mode of
  feeding) is collected routinely and can be
  supplemented with other details as required. The
  duration of such studies can in principle be
  extended indefinitely even the longevity of the
  original investigator need not be a limitation!
• Large national birth cohorts
• The National Survey of Health and
  Development, the National Development Study, and
  the Child Health and Education Study
• It is clear that early influences have long-term
  predictive effects. It is difficult to rule out
  potential confounding variables (e.g. long-acting
  adverse environmental factors or genetic
  influences.

14
Adults

• Cohort studies in adults have been very useful in
  elucidating the role of diet in the causation of
  disease. Nutrition is a major determinant of
  health, so it is obviously important to
  investigate its influence on morbidity and
  mortality. Everybody eats food, but not
  necessarily the same food, so there are ample
  opportunities for comparing the effects of
  different dietary habits within the same
  population. Ischaemic heart disease (IHD) is a
  good example of a relationship between diet and
  health, and has been investigated in this way.
  Table 14.1 summarizes nine cohort studies that
  have reported associations between diet and heart
  disease.

15
(No Transcript)
16

• Cohort studies have yielded useful information
  about nutritional factors in relation to other
  common diseases
• --The Framingham study indicated a
  protective effect of fruit and vegetables against
  stroke.
• --Breast cancer has been positively
  associated with alcohol intake and negatively
  with vitamin A.
• --In the Prospective Basel Study, plasma
  anti-oxidants (carotene and vitamins A, C, and E)
  were measured in 2974 Swiss men in the following
  12 years, lung cancer was associated with
  initially low plasma carotene and stomach cancer
  with low vitamin C and lipid-adjusted vitamin A
  levels.

17
The elderly

• Reasons for expecting cohort studies to be
  particularly easy and
• fruitful in the elderly
• --- The diet of an old person is likely to be
  more uniform.
• --- Nutritional deficiency is more common in old
  age than at other
• periods of life, yet many elderly people eat
  very well. The dietary
• variation between individuals is therefore
  large in comparison
• with the variation within individuals .
• --- Morbidity and mortality are very high in the
  elderly, so more
• endpoints will occur per person-year of
  follow-up than at younger
• ages.

18

• The confounding effect of differential ageing is
  particularly important in old age,and is well
  illustrated by the paradoxical finding that
  obesity is a favourable prognostic index in the
  elderly. Over the age of 65, people tend on
  average to lose weight as they grow older. But
  not everybody ages at the same rate. Some tend to
  be leaner and to die earlier, while those who
  arebiologically young for their chronological age
  are on average fatter and less likely to die in
  the next few years. But it does not necessarily
  follow that thin old people would live longer if
  they managed to put on weight.

19
Conclusion remarks

• Cohort studies allow more rigorous testing of
  aetiological hypotheses than other observational
  studies.
• They also provide unique information about the
  natural history of disease.
• Their disadvantages firstly concern feasibility,
  because they tend to be large,long, and suitable
  only for the study of common diseases. Secondly,
  they share other weaknesses of the observational
  approach, in that subjects who choose to eat one
  type of diet probably differ from people who eat
  a different diet in other ways that could affect
  their risk of disease. This confounding is
  particularly important when the nutritional
  variables are likely to be associated with
  particular lifestyles or the initial state of
  health and senescence of the subjects.
• Insofar as it is not possible to conduct
  long-term, randomized controlled trails of
  dietary changes in free-living populations,cohort
  studies provide the best available evidence of
  aetiology.

20
Appendix I Checklist for planning a cohort study

• 1. Purpose of the study
• 2. Value of the study
• 3. Definition of the cohort
• 4. Numbers
• 5. Recruitment of the subjects
• 6. Baseline data
• 7. Tracing of subjects
• 8. Collection of follow-up data
• 9. Analysis of data
• 10. General considerations

21

• 1. Purpose of the study
• (a) What hypotheses will the study examine?
• (b) What other specific questions will it
  address?
• 2. Value of the study
• (a) If the hypotheses are confirmed, will we be
  any better off ?
• (e.g. in our ability to understand disease
  or treat patients)
• (b) If the hypotheses are not confirmed, will
  other scientists be interested?

22

• 3. Definition of the cohort
• (a) Is the cohort to be identified
  retrospectively or prospectively?
• (b) What are the inclusion criteria (age, gender,
  area of residence, etc.)?
• (c) What exclusion criteria apply ?
• (e.g. presence of certain diseases,
  residence in institutions)
• (d) If the cohort is recruited over a period of
  time, at what point do the subjects
• have to meet the age and other criteria
  (e.g. at the start of the study, or when
• the subjects are seen)?
• (e) Are there any ambiguities in the way the
  criteria are defined or recorded?
• (f) Will the cohort comprise a total population
  defined as above, or will it be a sample
• of the population, and, if so, how will the
  sample be selected ?
• (e.g. randomly or by volunteering)

23

• 4.Numbers
• (a) How large will the cohort be and how has its
  size been calculated
• (expected differences, statistical
  power,etc.)?
• (b) What allowances have been made for
  non-response and migration of subjects?
• 5. Recruitment of the subjects
• (a) How will the subjects be identified?
• (b) How accurate and up-to-date is the sampling?
• (c) Over what period will recruitment continue?
• (d) Can we foresee any biases arising during
  recruitment ?
• (e.g. from selective identification or
  response of subjects)

24

• 6. Baseline data
• (a) What data are to be collected at baseline
  (including potential aetiological
• factors and possible confounders)?
• (b) What checks could be conducted on
  reproducibility, validity, comprehensibility
• of questionnaires, etc., so that the
  findings will be accepted as true?
• (c) How soon can the data be checked, coded, and
  computerized so as to allow early
• detection and correction of errors and
  omissions?
• (d) If blood is taken, should specimens of
  serum/plasma be kept deep-frozen for
• future analysis in case further hypotheses
  are suggested?

25

• 7. Tracing of subjects
• (a) How will the subjects be traced and when?
• (b) What secondary methods of tracing are
  available for subjects who cannot be
• traced by the primary methods?
• (c) What biases are likely to arise from
  incomplete tracing?
• 8. Collection of follow-up data (next page)
• 9. Analysis of data
• (a) Is a statistician (preferably the person who
  will undertake the analysis)
• involved in the design of the study?
• (b) What analyses of the data will be performed?
  10. General considerations

26

• 8. Collection of follow-up data
• (a) After what interval(s) will follow-up data be
  collected?
• (b) What information will be required
• (repeat baseline data, outcome events, new
  tests)?
• (c) What checks should be made on the quality of
  the data to be collected
• (e.g. reproducibility, validity,
  comparability with the baseline data)?
• (d) Can we ensure that the outcome events are
  recorded 'blind' with regard to the
• initial observations?
• (e) If some subjects are not available (e.g.
  through migration or refusal), is there
• any useful information that we can obtain
  about them?
• 9. Analysis of data

27

• 10. General considerations
• (a) What ethical issues arise (e.g. concerning
  explanation and information
• given to subjects signed consent forms for
  tests and follow-up procedures)?
• (b) What issues of professional etiquette must be
  considered (e.g. whose
• permission needs to be obtained who
  should be informed as a courtesy)?
• (c) If the data collection could disclose
  abnormalities in the subjects (e.g. a high
• serum cholesterol), what is our criterion of
  abnormality and what do we do
• when we find it?
• (d) What are the costs of the study and what
  personnel will be required ?
• (next page)

28

• (e) Is this the best time to start the study, or
  would it be better to wait
• (e.g. until the relevant technology has
  improved)?
• (f) Should the methodology of the study be made
  comparable to that of any
• other study (e.g. by using similar
  questionnaires)?
• (g) What experts should be consulted to increase
  the likelihood that the findings
• will be accepted as conclusive?
• (h) Is somebody keeping a list of all the people
  we promise to inform about the
• conclusions of the study?
•  

29
Appendix II Analysis of cohort studies (Clive
Osmond)

• Cohort studies may be classified
  according to both the
• type of data that are collected at baseline and
  the nature of the
• eventual outcome measure. The combination
  determines the
• appropriate strategy for analysis. Below we
  consider four
• common combinations, mention the usual method of
  analysis,
• and give an example of each. (Table 14.2)

30

• Design I
• Measurements are made on individuals.
  The outcome is the time to an event. The time may
  be censored (that is, the event is known not to
  have occurred up to the time specified). The
  usual method of analysis is by the Cox
  proportional hazards model. The effect size is
  measured by the hazard ratio, rather similar to a
  relative risk.
• Example
• Gale et al studied 730 elderly men
  and women who had completed a 7-day dietary
  record in 1973. They followed up the cohort for
  20 years, noting when subjects died from stroke
  or other causes. Allowing for age, gender, and
  known cardiovascular risk factors, those who were
  in the highest third of vitamin C intake had a
  relative risk of 0.5 for stroke (95 confidence
  interval 0.3-0.8) compared with those in the
  lowest third. A similar gradient in risk was
  present for plasma ascorbic acid concentrations.

31

• Design 2
• Measurements are made on individuals.
  The outcome is a notionally continuous
  measurement. The usual method of analysis is by
  multiple linear regression and analysis of
  covariance. The effect size is measured by the
  regression coefficient. This assesses the change
  in the outcome variable for a unit change in a
  predictor variable. Thus when a predictor
  variable is binary the regression coefficient
  describes the contrast between two groups.
• Example
• Lucas et al. studied 502 pre-term
  babies who were randomized to receive one of two
  different diets during their early weeks - mature
  donor breast milk or pre-term formula.At age 18
  months the survivors were given mental
  development assessments. A development score with
  mean close to 100 (standard deviation about 20)
  was obtained from each child. Regression
  adjustment was made for gender, gestational age,
  and social class. No clear difference in mental
  development was found between the two feeding
  groups.
•  
•  
•  
•  

32

• Design 3
• Measurements are made on individuals.
  The outcome is a binary 'yes/no' variable. The
  usual method of analysis is by multiple logistic
  regression. The effect size is measured by the
  regression coefficient, and this can be
  transformed into an odds ratio.
• Example
• Richardson and Baird studied the milk
  intake and calcium supplement use of a cohort of
  9291 pregnant women in California. 268 women
  experienced pre-eclampsia, the 'yes/no' outcome
  variable in the study. Allowing for possible
  confounders such as number of previous
  pregnancies and body mass index, the odds ratio
  for pre-eclampsia was 1.9(95 confidence interval
  1.2-2.9) in those who drank less than one glass
  of milk a day relative to those who drank two
  glasses of milk a day. The odds ratio was also
  higher(1.8 95 Cl 1.1-3.0) in those who drank
  four or more glasses a day, again using as the
  comparison group the two glass drinkers. The
  authors interpreted the association of low levels
  of milk consumption with pre-eclampsia as
  consistent with data on calcium and hypertension.
  The association of high milk consumption with
  pre-eclampsia was unexpected. The authors
  suggested that it needed to be replicated.

33

• Design 4
• Comparisons are made at a group
  level. The outcome is a survival time (possibly
  censored). This is known at the individual level.
  Thus rates can be calculated for the groups.
• Age and gender standardization of
  rates is necessary. The analysis can be with an
  internal comparison group, when Poisson rates
  models are useful. An external comparison group,
  often implied by the use of national
  cause-specific mortality rates, leads to
  standardized mortality ratios as the measures of
  effect size. These are often scaled so that 100
  corresponds to the rates in the external standard
  population.
• Example
• Thorogood et al studied 6115 members
  of the United Kingdom Vegetarian Society and 5015
  of their meat-eating friends and relatives. The
  subjects were followed up for 12 years.
  Standardized mortality ratios (England and Wales
  100) for ischaemic heart disease were 51 (95
  confidence interval 38-66) for the meat-eaters
  and 28 (20-38) for the vegetarians. Figures for
  all cancer were 80 (64-98) and 50 (39-62) for
  meat-eaters and vegetarians, respectively.

34

• More complicated designs
• If repeated measurements are made on
  members of the cohort on different occasions,then
  it is inefficient merely to average the data and
  incorrect to regard all the observations as
  statistically independent. More appropriate
  models are needed, and it will almost certainly
  be necessary to seek statistical support. Indeed
  even the simpler designs carry their own
  subtleties, making it wise to work routinely in
  collaboration with a statistician.

35
Back
36
Homework

• ???cohort study???
• ??
• ??
• ????
• ?(????)??(????)???
• ?????????
• ????????
• ???????
• ???????
• ??
• ????????alternative hypothesis