CaseControl and Cohort Studies

1
Case-Control and Cohort Studies
Robert Heimer Yale University School of Public
Health April 2009
2
Overview for Today

• For Cohort
• Definition and description
• Dynamic, fixed
• Prospective, retrospective
• Example -- Nurses Health Study and breast cancer
• Conduct
• Assemble the cohort
• Classify exposed/non-exposed
• Follow over time for outcome
• Analysis
• Strengths limitations

• For Case-Control
• Definitions
• Cases
• Definition, selection
• Controls
• Definition, purpose, selection
• Sampling
• Types
• Analysis
• Strengths limitations
• Examples

3
Main types of epidemiologic studies
Review slide
Epidemiologic research study designs
Experimental studies
Observational studies
Cohort
Cross-sectional
Ecological
Case-control
Some examples and special types

• Nested
• Case-crossover

• Times series
• Multi-level

• Individual-level RCT
• Community randomized trial
• Quasi-experimental

• One-time
• Panel

• Retrospective
• Prospective

4
Case-Control A Definition

• Examines exposure-disease relationship by
  enrolling cases (with disease) and controls
  (without disease) and comparing exposure history
• Backward design flash back
• Two groups are enrolled (cases and non-cases) and
  compared with respect to past exposure

5
Hypothetical Example

• Research question Pesticide exposure increases
  the risk of bladder cancer.
• Methods Consider a prospective cohort study, in
  which you enrolled 89,949 individuals aged 34-59
  and followed the cohort for 8 years
• Outcome 1,439 bladder cancer cases identified
  over 8 years of follow-up
• Exposure Blood drawn and frozen at beginning of
  study can be analyzed for level of pesticides

6
A Practical Problem

• Quantifying pesticide levels in the blood is
  expensive
• it's not practical to analyze all 89,949 blood
  samples
• To be efficient, analyze a select number
• all cases (N1,439)
• just take a sample of the cohort participants who
  did not get bladder cancer
• For example, two times as many cases (N2,878)

7
Case-Control Data
Bladder Cancer
Pesticide
We have just identified cases of disease from a
defined population, and then taken a sample of
that population (at-risk population) for
comparison. Exposure histories are determined for
each group. This is an example of a
case-control study that is nested in a cohort.
8
A Refined Definition of Case-Control Study

• A method of sampling a population in which cases
  of disease are identified and enrolled, and a
  sample of the population that produced the cases
  is identified and enrolled exposures are
  determined for individuals in each group.
• If done properly, case-control is a method of
  sampling a population so that controls reflect
  the source population that gave rise to cases.

9
Cases

• Need a clear case definition that leads to
  anaccurate classification of disease
• Disease definitions can be based on
• Signs, symptoms
• Clinical exam
• Laboratory test results
• May come from registries, hospitals/clinics,
  surveillance reports, etc.
• Cases may be enrolled going forward in time
• Incident cases preferred to prevalent cases since
  these are a better measure of risk because not
  confounded by duration
• Cases may be sampled if you think youll have too
  many, but this is not usually the situation

10
Cases in Doll Hill 1950

• Twenty London hospitals were asked to refer all
  patients admitted with carcinoma of the lung.
• Identified by admitting clerk, house physician,
  cancer registrar, radiotherapy department
• most diagnoses by necropsy, biopsy, or
  exploratory operation and some were diagnosed by
  other criteria
• Diagnoses confirmed by hospital diagnosis on
  discharge
• As a general rule this was the final diagnosis.
• Some diagnoses were changed after discharge.
• Patients were excluded from case population if
• Subsequent checking revealed primary carcinoma
  was another site (e.g. breast, colon)
• Histologic exam revealed growth was not carcinoma
• Found not to be malignant disease at all

11
Controls

• A sample of the source population (study base)
  that gave rise to the cases
• Purpose is to estimate exposure distribution in
  the source population that produced the cases
• Controls provide a fast and inexpensive
  (efficient) means of obtaining the exposure
  experience in the source population
• Controls are selected
• Without outcome of interest
• Independent of exposure status
• The would criterion
• If a member of the control group actually had the
  disease under study WOULD the person end up as a
  case in your study?

12
Selecting Controls

• Sources
• General population
• Hospital, clinic
• Other
• Sampling
• Risk-set sampling
• Survivor sampling
• Case-base sampling

13
General Population Versus Hospital Controls

• General Population
• Often cases are selected from a defined
  geographic population that gave rise to the cases
• Can use residence lists, drivers license
  records, voter lists, etc.
• Advantages
• High likelihood controls are from same study base
  as cases
• Disadvantages
• Need an enumerated list
• Time consuming, expensive
• Typically have high refusals

• Hospital Controls
• Use patients with diseases that have no relation
  to the exposures under study
• Advantages
• May have similar selection factors to cases
• Identifiable and accessible
• May be more willing to participate than general
  population controls
• Some potential disadvantages
• Referral patterns may not be the same, for ex If
  a hospital has a world-famous cancer center
• Controls are sick, so exposure patterns may not
  reflect study base

14
Other Types of Controls

• Friends
• Spouses
• Siblings/twins
• (Deceased individuals)
• The goal with each of these types of controls is
• Measure exposure in the source population
• Minimize differences between cases and controls

15
Matching

• You pair each case with someone who is like the
  case but who does not have the disease or outcome
  under study.
• You can do FREQUENCY MATCHING, which means
  picking people from the general groups from which
  cases come, so that the overall make up of the
  two groups is similar.
• You can connect each case to more than one person
  who resembles that case (R1 MATCHING)
• Requires a different analytical approach, most
  common are
• Matched OR
• McNemar chi-square
• Conditional logistic regression

16
Doll and Hill, 1950

• Required to make similar inquires of a group of
  non-cancer control patients
• For each lung cancer patient, interviewers were
  instructed to interview a patient of same sex,
  within the same five-year age group, and in the
  same hospital at or about the same time
• Could not always find a suitable control
• 743 general medical and surgical patients
• Some differences with regard to place of
  residence
• Higher proportion of cases from outside London

17
Data Collection Exposure Assessment

• Once cases and controls are identified and
  enrolled, collect information of exposure of
  interest and other variables
• Ideally, use same data collection techniques for
  both cases and controls

18
Analysis of Case-Control Data Overview

• Because controls are a sample of the population
  that produced the cases, you do not know the size
  of the total population
• Therefore, you cannot get a prevalence,
  cumulative incidence or incidence rate of
  disease.
• Do not have the appropriate denominator for these
  calculations.
• Instead, we compute odds and odds ratios that we
  use for estimation of relative risk in special
  circumstances.

19
Doll and Hill, 1950 Table IV

• Problem with usual interpretation of this table
• 1298 is not an at-risk population, it is the
  study population
• 1269 and 29 do not tell us about exposure is the
  source population
• 647/1269 and 2/29 do not tell us about risk of
  outcome (lung cancer) among exposed and unexposed
  (smokers and non-smokers) in source population

20
Need a Different Analytical Approach

• Use Odds Ratio
• Delete marginals from the table because they do
  not have a lot of meaning for understanding risk.
• Odds probability(event)/probability(non-event)
• Compares the frequency of occurrence of something
  to the frequency of non-occurrence.
• Calculation of exposure odds ratio (EOR)
• Odds of exposure vs. no exposure in diseased
  persons a/c
• Odds of exposure vs. no exposure in non-diseased
  persons b/d
• Odds ratio of exposure for cases compared to
  controls (a/c) / (b/d) ad/bc

21
Expsoure Versus Disease Odds Ratio

• Suppose we want to determine the odds of disease
  rather than the odds of exposure
• The odds of disease in the exposed a/b
• The odds of disease in the unexposed c/d
• The odds ratio of disease for exposed compared to
  unexposed (a/b) / (c/d) ad/bc
• This is the same as the exposure odds ratio

22
OR Approximates RR

• when disease is rare
• Proportion of cases in exposed and unexposed
  groups is low in total (source) population
• ab b and cd d
• RR a/(ab)/c/(cd) a/b / c/d ad/bc
• if disease is not rare
• depends on sampling When case-base or risk-set
  sampling is used for control selection
• when cases are newly diagnosed and prevalent
  cases are excluded from control group, and
  selection of cases and controls is not based on
  exposure status

23
Doll and Hill, 1950 Table IV

• OR (64727) / (2622) 14.0
• Technically The odds of smoking among lung
  cancer cases is 14 times higher than the odds of
  smoking among non-lung cancer cases.
• Loosely People who smoke have a 14-fold
  increased risk of lung cancer compared to people
  who do not smoke.

24
Advantages of Case-Control Method

• Useful when exposure data are expensive or
  difficult to obtain
• Pesticide and bladder cancer example
• Useful when little is known about disease
• Vaginal cancer in women
• Useful when disease has long induction or latent
  period
• Lung cancer
• Many special cases
• Outbreaks, vaccine effectiveness, etc.
• A major advantage is efficiency (time and money)

25
Disadvantages

• Retrospective nature makes them prone to many
  biases
• Information bias
• Recall bias, if cases and controls report
  exposures differently because of their
  case/control status
• Selection bias
• If selection of control group is not
  representative of source population that gave
  rise to cases with respect to exposure

26
Rare Vaginal Cancer in Young WomenHerbst et al.
NEJM 1971284(15)878-881

• Background
• Initial clinical observation of 7 women, ages
  15-22, with adenocarcinoma of the vagina
• Never before seen at that hospital
• Study design
• We then decided to conduct a case-control,
  retrospective study that would compare in detail
  these patients and their families with an
  appropriate control group to uncover factors that
  might be associated with the sudden appearance of
  these tumors.
• Cases
• 8 women with diagnosed with clear-cell or
  endometrial type adenocarcinoma of the vagina
  between 1966 and 1969 at Boston hospitals

27
Controls and Exposure Assessment

• Controls
• 4 matched controls per case
• Using persons born at the same hospital as the
  case and within 5 days and on the same type of
  service (ward or private)
• Exposure assessment
• Reproductive and other factors

28
Major factor Ingestion of estrogen
(Diethylstilbesterol or DES) during first
trimester of pregnancy by 7 of 8 mothers of
affected women and by 0 of 32 mothers of control
women (plt.001) (OR?) Conclusion DES is a risk
factor for subsequent adenocarcinoma of the
vagina in offspring. Implications It is
unwise to administer estrogen to women early in
pregnancy. Abnormal bleeding in adolescent
women should be examined for vaginal tumors.
29
Summary Points on Case-Control

• Case-control studies are useful in many
  situations for epidemiologic research because of
  their efficiency
• Selection of control population is often
  challenging
• With appropriate methodologies and mindfulness
  toward common biases, they can produce valid and
  important results.

30
Cohort Studies
31
Definition of a Cohort

• Cohort
• A group of persons followed over time
• Cohort study
• A study in which two or more groups of people
  that are free of disease and that differ
  according to exposure level(s) are followed over
  time and compared with respect to disease
  incidence to assess the association between
  exposure and disease
• Also called prospective, follow-up,
  longitudinal studies
• May be considered a natural experiment
• People are exposed to substances and risk
  behaviors all the time, either on purpose or not
  these can be studied as exposures
• Interventions may be considered a subset of
  cohort studies

32
Some Characteristics of Cohort Studies

• May be open (dynamic), fixed, or closed
• Dynamic members can enter and leave during
  follow-up time
• Residents of Kazan
• Fixed membership is fixed (permanent), but
  members can exit the cohort
• People present in lower Manhattan 9/11/2001
• Women who have given birth
• Closed members cannot enter after start of
  study and nobody is lost to follow-up defined
  start and end time
• Attendees of church supper
• Everybody has same follow-up time
• May be prospective or retrospective
• Depends on temporal relationship between
  initiation of study and occurrence of disease
• Calendar time vs. follow-up time

33
Dynamic, Fixed, and Closed Cohort

• Closed cohort is like a fixed cohort, but all
  members have the same exposure time

34
Timing of Cohort Studies

• Prospective
• Exposure has occurred but disease has not
  occurred at start of study
• Exposure---------------?Disease
• Study starts here (Calendar time and follow-up
  time are concurrent)
  
• Retrospective
• Both exposure and disease have occurred at start
  of study
• Exposure----------------? Disease
• 
  Study starts here
• Study starts here (Calendar time and follow-up
  time NOT concurrent)

35
Comparison

• Retrospective
• Cheaper, faster
• Efficient with diseases with long latent period
• Exposure data and other information may be
  limited or missing
• Prospective
• More expensive, time consuming
• Not efficient for diseases with long latent
  periods
• Better exposure and confounder data (planned)
• Enhanced follow-up
• Less vulnerable to some bias
• How to choose?
• Necessity (logistics time, money)
• Research question (science available data)

36
Design Overview

• Identify and assemble a group of individuals
  without disease of interest
• Classify with respect to exposure status at start
  of study
• Monitor subsequent development of disease in
  exposed and non-exposed subjects over time
• Analysis

37
Example Nurses Health Study

• Background and Purpose
• Based at Harvard Medical School and School of
  Public Health
• Originally conceived as a study to examine the
  association between oral contraceptive use
  (widespread 1960s and 1970s) and breast cancer
• In part, due to conflicting previous studies and
  concerns of limitations of case-control approach
  for this association
• Cohort
• Enrolled 120,000 married female nurses age 30-55
  registered in one of 11 states in 1976
• Identified by American Nursing Association and
  state boards of nursing
• Initial baseline mail survey collected
  information about demographic, reproductive,
  medical, and life-style variables

38
Assembling a Prospective Cohort

• Exclude those with disease or not at risk
• Depending on research question and feasibility
• General cohorts and special cohorts
• Internal and external comparison groups

39
General Cohorts

• Select a group of individuals from general
  population
• Geographically defined areas
• Well-identified groups (Ex NHS)
• Others
• Not chosen for exposure status but rather for
  feasibility and logistical reasons that make the
  study possible
• Nurses Health Study
• Believed that nurses could be interested in
  participating in a health study
• Believed that nurses could answer survey
  questions correctly
• Often useful for common exposures
• NHS -- OC use 42 past users, 6 current users
• Often used for multiple exposures and multiple
  outcomes

40
Special Cohorts

• Groups with a particular health status or other
  special characteristic
• For example repeated x-rays, live near toxic
  waste dump site, present at event such as
  Chernobyl
• Useful for occupational settings
• Often have unusual exposures
• Danish workers exposed to trichloroethylene
• Often useful for rare exposures
• Allows accrual of sufficient exposed individuals
  because of targeted recruitment

41
Comparison Groups (Exposure)

• Principle You want the comparison (unexposed)
  group to be as similar as possible to the exposed
  group with respect to all other factors except
  the exposure. If the exposure has no effect on
  disease occurrence, then the rate of disease in
  the exposed and comparison groups will be the
  same.
• Counterfactual ideal The ideal comparison group
  consists of exactly the same individuals in the
  exposed group had they not been exposed. Since it
  is impossible for the same person to be exposed
  and unexposed simultaneously, epidemiologists
  must select different sets of people who are as
  similar as possible.

42
Classifying Exposure

• Define and measure
• Need a way to handle exposures that may change
  over time
• NHS and OC use
• Current and past (ever), never Current defined
  as past 2 years
• Total duration, duration prior to first
  pregnancy, duration prior age 25
• May want to consider multiple levels of exposure

43
Internal Comparison Group

• Unexposed members of the same cohort
• Single cohort in which individuals are classified
  into exposure categories
• NHS nurses were enrolled, surveyed about risks
  and classified as exposed or unexposed
• Usually preferred because of higher likelihood of
  similarity between exposed and unexposed
• Selected into cohort in same way
• Because you are a nurse, because you live in a
  city
• Measurement and follow-up of disease done in the
  same way

44
External Comparison Group

• A different cohort, from another similar
  population, that is not exposed
• Different cohort for example, same type of work
  at different organization
• General population
• Useful when a special exposure group is used
  and/or when entire cohort is exposed
• Inclusion in the cohort meant exposed so had to
  go elsewhere for comparison

45
Following the Cohort

• A major challenge, a major expense, a major
  potential threat to validity
• Method
• Passive or active (NHS)
• Length of time
• Depends on outcome and sample size
• For chronic diseases, follow-up will need to be
  years or decades
• Data collection
• For exposure (updated and new) and outcome
  (multiple) information
• Pre-existing records (medical, employment),
  questionnaires, physical exams, medical tests,
  laboratory assays, external sources (e.g. cancer
  registries, vital statistics such as the US
  National Death Index)

46
Important to Minimize Losses to Follow-up

• For reasons of sample size and bias
• Depends on population, duration, etc.
• Methods
• Collecting sufficient locating information
• Maintaining regular contact with participants
• Using multiple methods (phone, mail, internet,
  postal office databases, disease registries,
  vital statistics, physicians)
• Make participation worthwhile for participants

47
Nurses Health Study

• Follow-up
• Every 2 years complete another mail survey that
  collects information about development of
  outcomes, updated exposures, new exposures
• Biologic specimens also collected
• Self-reported outcomes confirmed by medial record
  reviews, pathology reports, review of National
  Death Index
• Promoting follow-up
• Follow-up surveys included a newsletter
• 2-4 follow-up mailings 5th mailing was an
  abbreviated survey
• Added a telephone follow-up
• Use of certified mail
• Follow-up through state boards of nursing
• gt90 follow-up at each cycle

48
Other Advantage to Cohorts

• Nurses Health Study has continued and expanded to
  include examination of associations between
• Exposures diet, physical activity, obesity,
  post-menopausal hormone use, reproductive
  factors, smoking, alcohol, coffee, hair dyes
• Outcomes cancer, CVD, diabetes, mortality,
  osteoporosis to name just a few
• gt800 publications to date

49
Analysis

• Basic analysis involves calculation and
  comparison of incidence of disease among exposed
  and unexposed
• Depending on available data, you can calculate
  cumulative incidence or incidence rates and
  corresponding ratios

50
OC use and breast cancer in NHSRomieu et al.
JNCI 1989811313-1321.

• Background
• Conflicting evidence re OC use and breast cancer
• OC use is widespread, therefore possible large
  impact on public health
• Cohort
• 118,273 women who did not report a diagnosis of
  cancer (other than non-melanoma skin cancer) on
  1976 questionnaire
• Exposure
• Self-reported OC use on each questionnaire (every
  2 years)
• Classified as current, past, never time since
  first use, time since last use, duration of use,
  use before first pregnancy

51
(No Transcript)
52
Conclusions from NHS on Risk of Breast Cancer
following Oral Contraception Use

• Overall past use not associated with breast
  cancer (RR1.06)
• Slight increased risk for current users (RR1.56)
• Number of women who used for long duration early
  in reproductive life was too small for meaningful
  analysis

53
Advantages of Cohort Studies

• Known temporal association exposure ? outcome
• Preferred for causal inference
• Can evaluate multiple outcomes for given
  exposure(s)
• NHS examined relationship between OC use and
  breast cancer, ovarian cancer, malignant melanoma
  and myocardial infarction
• Less prone to certain types of bias
• Recall bias outcome does not influence recall of
  past exposures (pre-classified)
• Selection bias disease status does not influence
  selection of subjects with respect to exposure
• Can estimate risk and incidence (person-time)

54
Disadvantages of Cohort Studies

• Expensive
• Time-consuming
• Loss to follow-up bias
• If those lost are different in ways related to
  exposure and outcome
• Inefficient for rare diseases
• unless AR is high
• retrospective cohort can then established
• Inefficient for diseases with long induction or
  latent period (unless retrospective cohort)

55
Summary

• Cohort studies are generally considered the
  strongest of the observational designs
• They are often the most expensive and time
  consuming
• Vulnerable to different set of limitations than
  ecological, cross-sectional, and case-control
  studies

56
??????? ?? ????????