1

 Characteristics of a cohort study design

 Strengths and Weaknesses
 Classification
 Issues in the analysis
 Threats to Validity

2
 Observational Analytic Epidemiologic studies
 Aimed at determining causal associations
between exposure factors and health
outcomes
 Comparisons are made in health outcomes
between individuals who are exposed to
etiologic factor(s), and individuals who are not,
unexposed.

3
• Designed to answer the same question as a
randomized clinical trial  What are the effects
of this particular exposure?
• Groups of individuals defined on the basis of
presence/absence of exposure to the suspected
risk factor
• All
potential subjects must be initially free of
the disease under investigation
• Eligible
participants are then followed over time
to assess occurrence of disease

4
 Case-Control Study

Exposure
Disease
? (Case)

? No disease
(Control)

Retrospective Nature

5
 Randomized Clinical
Trial

Investigator

Randomization

Exposure No Exposure

Contrast: In a cohort study, the researcher does not control

the intervention/exposure but instead OBSERVES its effects. 6
 2
1

Disease No Disease
Exposure a b
No Exposure c d

1. FIRST subjects are defined on the basis of exposure status

2. NEXT subjects are followed over time to assess disease developm

7
 Identify a cohort

 Determine exposure status at baseline

 Follow up over time

 Ascertain whether outcomes have occurred at

intervals or at the end

 Analyze data

8
 A group of persons
 sharing the same experience
 followed for a specified period of time
 Examples
 birth cohort
 workers at a chemical plant
 graduating university class

9
2 ways to select “exposed” group
1.Select a sample of the population, and survey or examine
the sample to establish baseline exposure status
-good for relatively common exposures, such as cigarette
smoking or coffee drinking

2.Choose a group specifically because they have undergone

some unusual exposure or experience you want to
evaluate (sometimes called a “special exposure” group)
-individuals in certain occupations, such as rubber
processing or uranium mining
-individuals who have undergone a particular medical
process, such as x-rays, vaccinations, or therapies
-individuals living near a suspected environmental hazard,
such as a nuclear testing ground or toxic waste dump site

10
Major principle:
Groups being compared should be as similar as
possible with respect to all other factors that may be
related to the disease EXCEPT the exposure under
investigation

1. Internal comparison group

When a sample of the population is classified as
exposed or unexposed through a survey or
examination, use the unexposed group as the
comparison
-Example: mortality among smokers vs. nonsmokers in
a defined population

2. External comparison group

For a special exposure group, often a sample of the
general population of the area in which the exposed
individuals reside is used as the comparison
-Example: mortality among uranium miners vs.
mortality in U.S. population

11
 Examples
Smoked at least 100 cigarettes in lifetime
Used oral contraceptives for at least six consecutive months
Ate noon meal at Restaurant "X" on August 10, 2000

Sources
Records
Lab Reports
Interviews
Questionnaires
Exams
Environmental measures
Vital records

12
 Examples
Physician diagnosis in medical record
Specified ICD codes in hospital discharge data
Cause of death on death certificate

Sources
Records
Lab reports
Interviews
Questionnaires
Periodic examinations
Environmental measures
Disease registries
Surveillance systems
Vital records 13
 Maintain temporal sequence

Exposure time Disease

 Assess the various effects of a particular exposure
 For example, smoking may be associated with increased
incidence of lung cancer, heart disease, diabetes, asthma,
etc.
 Study new or unusual exposures (primary method)
 Newly licensed drugs, technology, behaviors, etc.

14
 Calculate measures of risk (cumulative incidence) and
rates (incidence rate/density) of disease in exposed and
unexposed individuals

 Avoid bias in the exposure measurement

 Since outcome has not yet occurred, selection of exposed
and unexposed cannot depend (be biased upon) outcome
status

 Provides a complete description of experience

subsequent to exposure

 including the rate of progression, staging and natural

history of the disease

15
 Expensive
 Inefficient for studying rare diseases
 If the disease is very uncommon, often a prohibitively large
study population is necessary in order to accrue enough
outcomes

 Potentially long duration for follow-up for some

outcomes
 More chance for loss to follow up of subjects
 Subjects move, change jobs, lose interest in study, lose contact
with study organization, etc.

 Must anticipate secular trends in technology

 Possible that some participants were not truly “disease-free” at
the beginning of the study…but the technology to detect sub-
clinical cases was not available at the time the study began
16
 Must make sure to measure confounding variables, but
almost always second guess at the study initiation

 Example: Smoking is a confounder in the association between

coffee drinking and risk of delivering a low birth weight baby
 If the investigator failed to measure smoking status at beginning
of study, there is no way to “control” for the effects in the
analysis phase
 But what if the investigators did not know that smoking is a
confounder at the time the study was initiated?

 Classify people according to exposure at beginning of study,

but they can change throughout the study

 Example: 1 study participant’s exposure to smoking

 1985 – 1995 –1 pack/day
 1996 – “quit”
 1997 – 1999 – ½ pack/day
 2000-2001 – nicotine patch

17
•Classificationof cohort studies depends on the
temporal relationship between the initiation of
the study and the occurrence of the outcome

•3 Main Types:

18
 Prospective (concurrent)
 Cohort is assembled, baseline exams and data
collected for purpose of study

 Retrospective (historical)
 Cohort is assembled in the past on the basis
of records . Some data might be missing

 Bidirectional (mixed)
 Includes both elements

19
•The investigator collects information on the
exposure status of the cohort members at the
time the study begins, and identifies new cases
of disease (or deaths) from that time forward

•The exposures may or may not have occurred

at the beginning of study

•BUT the outcome has certainly not yet occurred

•After the selection of the cohort, participants 20

Investigator
begins study
here

21
•Exposure status is established from information
recorded at some time in the past, and disease
incidence (or mortality) is determined from then until
the present
•Both the exposures and the outcomes have already
occurred when the study is initiated
Example:
•From medical records, identify a group of women who were using o

contraceptives (OCs) 10 years ago and a group of women who were

•Either interview the women, or use medical records, to determine t

subsequent history from that point to the present in terms of develo

heart disease
Retrospective: deals with current and past events; can be done quic
Cohort: the comparison is made between users and non-users of OC
22
 Retrospective Cohort Studies

Disease
Exposure occurrence Study starts

time

Example: disease
outbreak following a
gathering
occupational exposure
in mine workers

23
 Major steps in Retrospective cohort
studies

• Well defined population/feasible

• Exposed and unexposed can be identified
• Outcome (ill or not ill) can be ascertained

Opportunity to go back in time,

categorise people according to their exposure
and then determine their outcome

For example, weddings, parties, hotels, occupational

exposures, etc. Ascertainment of exposure must be
comparable for all.
24
 Can be completed in a much more timely fashion while
maintaining temporal sequence between exposure & outcome
 Considerably less expensive
 Do not need funding and personnel during lengthy follow-up period
 Particularly efficient for study of diseases with long latency
periods
 Don’t have to wait for the outcome to occur, as in prospective design
But
 Main disadvantage is a reliance on available information
 Sometimes the quality of exposure or outcome data is less than ideal
for fulfilling the study objectives
 More sensitive to confounding and bias
 Both the exposure and outcome have occurred, so there is the
chance that ascertainment of outcome status differs between groups

25
 Case control Cohort
Subjects Two groups of subjects Two groups of subjects
Characteristics One group has disease of interest One group is exposed to risk
and other is disease-free factor of interest and other is
non-exposed

Major difference The cohort study starts with the putative cause of disease, and
observes the occurrence of disease relative to the hypothesized
causal agent, while the case-control study proceeds from
documented disease and investigates possible causes of the
disease.

Source population Poorly defined Clearly defined

Purpose Test hypothesis Test hypothesis
Cause effect relationship Can be suggested Can be interpreted
Measure of disease frequency and risk Odds ratio Prevalence, Incidence,
Relative risk,
Attributable risk

Potential problems Recall and selection bias Less control over selection of
subjects

Less control over measurements

Remarks Can study multiple causes of Can study multiple outcomes of a

disease; useful for studying single exposure
rare diseases
26
 Understanding the effect of an exposure requires
knowing the frequency of disease in the absence
of exposure (baseline or background rate)

 Measures of association estimate the magnitude

of an association between exposure and disease

 Indicates the likelihood of developing the disease

in the exposed group relative to those who are
not exposed

27
2 Main Measures of Association from Cohort
Studies:

Risk Ratio (Relative Risk): Ratio of the

Cumulative Incidence among exposed to
Cumulative Incidence among unexposed

Rate Ratio: Ratio of the Incidence Rate among

exposed to Incidence Rate among unexposed

28
How to Set Up a 2 X 2 Table with
Count (Cumulative Incidence) Data

Outcome

Yes No Total

Exposure
Yes a b a+b

No c d c+d
29
Cumulative Incidence (exposed)
Cumulative Incidence (unexposed)

= CIe/CIu
= a/(a+b)
c/(c+d)

Interpretation: The risk of developing disease was n times higher [or 1/n times less
likely] among exposed compared to unexposed persons in this population during
time t.

30
How to Set Up a 2 X 2 Table with
Person-Time (Incidence Rate) Data
Outcome
Yes No
Exposure Total

Yes a - PYe

No c - PYu

Total a+c - PYT

31
 Incidence Rate (exposed)
= Incidence Rate (unexposed)
= IRe / IRu

= (a/Pye) / (c/Pyu)

Interpretation: The rate of developing disease was n times

higher [or 1/n times less] among exposed compared to
unexposed persons in this population over time t.

32
 Interpretation of Risk or Rate Ratio

RR > 1 Risk (or rate) of new disease is greater in

exposed group than in unexposed group during time
t; DIRECT relationship between exposure and
outcome

RR = 1 Risks (or rates) of new disease are identical

in both groups during time t; NO
RELATIONSHIP between exposure and outcome

RR < 1 Risk (or rate) of new disease is less in

exposed group than in unexposed group during time
t; PROTECTIVE or INVERSE relationship between
exposure and outcome
-to interpret, must invert measure and report as33
 The unit of study is the individual
 Goal is to calculate incidence
 Cumulative incidence provided there is no loss
to follow up (censoring) or competing risks
 Incidence density – makes use of person time at
risk contributed by each individual

 Once incidence is calculated for exposed and non-

exposed RR can then be calculated
 RR = I exposed/ I un-exposed
 One RR is obtained one can also get AR, PAR, etc

34
 In mortality studies where an external
comparison group is used one can also calculate
the SMR

 Also able to use the following methods

 Poisson regression for count data
 Survival analysis and Cox proportional hazards
regression where outcome is time to event
 Other longitudinal data analysis methods e.g.
marginal structural models, random effects
models

35
# new cases during a specified period
size of population at start of period

= "Attack rate"
= Probability of getting disease
= Risk of disease
= Cumulative incidence

36
1. Deaths in diabetic men
100 deaths
189 men at start of follow-up period
Risk = 100 / 189 = 0.529 = 52.9%

2. Deaths in non-diabetic men

811 deaths
3151 men at start of follow-up period
Risk = 811 / 3151 = 0.257 = 25.7%

37
 Distribution of illness according
to exposure in a cohort study

ILL NOT ILL Risk

a
Exposed a b a+b
a+b

Not exposed c c
d c+d
c+d

Relative risk = Risk exposed / Risk not exposed

38
39
Classical example: foodborne outbreak (cohort = all
people who attended a wedding for example)

ill not ill Incidence

ate ham 49 49 98 50 %

did not
4 6 10 40 %
eat ham

Relative risk 50% / 40% = 1.25

40
Exposure Population Cases Incidence
Level at Risk

High N1 a1 I1
Medium N2 a2 I2
Low N3 a3 I3

Unexposed Nne c Iue

41
 Various Exposure Levels

Exposure Population Cases Incidence RR

Level at Risk

High N1 a1 I1 RR1
Medium N2 a2 I2 RR2
Low N3 a3 I3 RR3

Unexposed Nne c Iue Reference

42
43
 Selection Bias
 Error due to a systematic difference between those
selected for a study and those NOT selected for a study
 For example, use of volunteers in studies may create a bias

 In cohort studies, most concerned with significant losses to

follow-up that may result in differential drop-out rates
(outcome status may differ between those who drop out
and those who remain in the study)

 Main “fix”: Methods to minimize losses to follow-up (i.e,

ask participant for a contact name, Dept. of Motor Vehicle
records, birth records, credit bureaus, postal service—all
are helpful in tracing participants)

44
 Information Bias

 A flaw in measuring outcome data that results in a

differing quality (accuracy) of information between
comparison groups

 Happens if investigators are more likely to “scour” the

medical records or perform diagnostic testing on those
known to be exposed

 Example: A physician is more likely to perform a chest x-

ray on a known smoker, thus there is increased likelihood
of detecting lung cancer among exposed (smokers)

45
 Confounder: a third variable that either masks a true
relationship between E & D, or partially/fully accounts for the
relationship between E & D

 Failure to take into account may cause the investigator to

conclude that an exposure causes disease when in fact it does
not, or may fail to detect a relationship when one truly exists

 Methods for measuring confounding variables same as

measuring exposures of primary interest (i.e., questionnaires,
physical measurements, lab tests, environmental testing)

 Same limitations of these measurement techniques apply to

measurement of confounding variables

46
 Cohort studies allow measure of risk

 In the ideal world: Prefer cohort to

case-control study

47
48