Epidemiological study design

Report
Study Design
Simon Thornley
Overview

By the end of this lecture you will be able to


Describe the main types of analytical epidemiological studies
Describe when to use different designs for different research
questions.
Participatory epidemiology






Epidemic of dental caries, we think sugary drinks may be
responsible....
Cohort/Cross-sectional study
Outcome
Filling, root canal or extraction due to caries (not
wisdom tooth) in last 5 years
Exposure
At least one sugary soft drink per week?
Participatory epidemiology




Case – control study
Sample by outcome, rather than exposure.
Outcome – dental decay
Exposure – sugary soft drinks.
Epidemiological studies
Analytic



Aim
To assess the cause of a
disease
Identify point for
intervention, to either
prevent disease occurring
or improve prognosis of
people with disease.
Descriptive





Aim
To assess the health status of
the population
Used for time trends/planning
health services.
To perform an accurate
sample, you need a sampling
frame
Surveys


Sample based (simple/cluster)
Capture-Recapture
Types of studies
Observational

Cohort
Cross sectional
Case control

Experimental

RCT

Ideal study is RCT; all other studies are trying to
emulate this design.

Elements of epidemiological study
Outcom
es
Exposures
Participants
Hypothesis





Usually that a certain exposure causes a disease
For example, perinatal sugar consumption causes
childhood asthma
Ideally written explicitly before study started.
Hypotheses need to be stated in such a way so that they
can be proven wrong
Analytical study tests hypothesis that exposure is
associated with disease.
What are we doing in epidemiological
studies?




What we want to know
Does sugar exposure cause asthma?
But we have limited resource.....
Among countries that took part in the ISAAC study is
there an association between reported severe asthma
symptoms in 6 year old children and average sugar
disappearance data, collected six years before the
children were surveyed?
Screening questions





Feasible
Interesting
Novel
Ethical
Relevant
In pictures...
Truth in the
universe
Infer
Truth in the
study
Error
Target population
Phenomena of
interest
Design
Intended sample
Intended variables
Analytical study design
GATE frame (Prof. Rod Jackson)
Participants
Exposure
Outcomes
a
Exposed
Exposed
Disease
b
Exposed
no disease
Participants
Unexposed
Time
c
d
Unexposed Unexposed
Disease
No disease
A word about probability (risk)



A number between 0 (it won’t happen) and 1 (it
definitely will happen) that describes the long run
frequency of an outcome
What is the probability of rolling 1 on a six sided dice?
What about the probability of diarrhoea after eating
contaminated sandwhiches?
Conditional probability



Risk of events after people are exposed (or not)
represent “conditional probabilities”.
In all studies, we compare probabilities among exposed
and unexposed
If exposure not related to outcome, then risk after being
exposed (conditional prob) is the same as the risk for
the total population (marginal prob.).
Conditional probability and effect measures

Effect measures (e.g. risk ratio, odds ratio, incidence
rate ratio) compares the conditional probabilities in
exposed and unexposed to see if they differ.

Is difference due to chance or not (effect of exposure)?
What does effect measure mean?




E.g. Relative risk = 2.98 (95% CI 0.98 to 4.95)
if 1, no difference.
if >1 then the exposure ↑ the probability of outcome
if <1, the exposure ↓ probability of outcome.
Participants
Need to define and sample

Ideal is to have similar population that differs only by exposure
of interest

Need sufficient heterogeneity of exposure
Access
Inclusion or exclusion criteria based on potential
confounding factors

Exposure
GATE frame assumes binary exposure
In reality, often measured on continuous scale
Need to check the accuracy of what is being measured

Validity (agreement compared to gold standard) and
reliability (same over time)
Objective vs subjective measures
Cotinine vs self reported smoking status

Outcome
Seek outcome that is objective and easily measured
Eg. Death, 1st cardiovascular disease event

Objective vs subjective measurement
Same outcome measured on all participants,
regardless of exposure

Why favour one study over another?
Randomised study ideal, but not always practical
Ethical considerations – assigning smoking exposure?
Resources

Routinely collected data allows cheap observational
studies to be done
Hypothesis generating, may be sequential.
Eg. Vitamin D and respiratory infection.
Rarity/latency issues in choosing between case
control and cohort design.

Randomised controlled trials
Randomisation of treatment or prevention

Confounding both known and unknown

Blinding

subject/investigator or both

Ethics

Not harmful treatments

Intention to treat – ”analyse what you randomise”,

even if subjects switch treatment during follow up

Randomised controlled trials
Key point is allocation of exposure by investigator
Participants
Exposure
Random allocation
Outcomes
a
Exposed
Exposed
Disease
b
Exposed
no disease
Participants
Unexposed
c
d
Unexposed Unexposed
Disease
No disease
Measure of effect
For binary outcome
Cum. Incidence in exposed
Relative risk = ------------------------------------------Cum. Incidence in unexposed
a/(a+b)
= -------------c/(c+d)
Randomised controlled trial
Best
design for assessing causation
Random assignment of treatment results in balance
of known and unknown risk factors into exposure
groups (numbers large)
Ethics – clinical equipoise
Best for straight forward interventions (eg. Drugs)
with blinding possible
Different designs – separate arms, crossover,
factorial.
Does pre-quit nicotine treatment
improve quitting?
Smoking cessation therapy
Schuurmans MM, Andreas HD, Xandra van B, et al. Effect of pre-treatment with nicotine patch on withdrawal symptoms and abstinence rates in smokers
subsequently quitting with the nicotine patch: a randomized controlled trial. Addiction. 2004;99(5):634-40.
Participants
Adult
smokers
Exposure
Pre-quit
NRT
Exposed
Outcomes
Quit
Smoking
18
Exposed
Disease
62
Exposed
no disease
Participants
Unexposed
Placebo
NRT
8
80
Unexposed Unexposed
Disease
No disease
Does pre-quit nicotine treatment
improve quitting?
Incidence in exposed
Relative risk =
---------------------------Incidence in unexposed
18/(18+62)
=
-------------= 2.5
8/(8+80)
In pictures - Actual
Null hypothesis; no effect
Measure of effect
800
600
400
200
0
Frequency
1000
1200
1400
Effect of NRT on Quitting
0
2
4
6
Risk ratio
8
10
Cohort
Participants
Exposure
Outcomes
By measurement
Exposed
Exposed
Disease
Exposed
no disease
Participants
Unexposed
Unexposed Unexposed
Disease
No disease
Cohort study
eg Framingham
Patients without disease
Group by exposure
Can use a variety of exposures
Follow until disease develops

Cohort advantages
Exposure precedes disease
Disease status does not influence selection
Several outcomes possible
Good for rare exposures

Cohort disadvantages
Prospective costly
Inefficient for rare diseases with long latency
Several outcomes possible
Exposed followed more closely than unexposed?
Loss to follow up causes bias

Do computer screens cause
spontaneous abortions?
1991
Participants
Female
telephone
operators
Exposure
Computer screens
Exposed
Outcomes
abortion No abortion
54
Exposed
Disease
312
Exposed
no disease
Participants
Unexposed
No
computers
Time
82
434
Unexposed Unexposed
Disease
No disease
Do computer screens cause
spontaneous abortions?
Incidence in exposed
Relative risk =---------------------------Incidence in unexposed
54/(54+312)
=
------------------- = 0.93
82/(82+434)
In pictures - Actual
Null hypothesis; No effect
800
600
400
200
0
Frequency
1000
1200
1400
Effect of Monitors on abortion
0
1
2
3
Risk ratio
4
5
Cross sectional
Participants sampled at one point or short duration
Exposures and outcomes assessed at same point in
time
Participants
Exposure
Outcomes
By measurement
Exposed
Exposed
Disease
Exposed
no disease
Participants
Unexposed
Unexposed Unexposed
Disease
No disease
Cross-sectional
Advantages
Describes pattern of
disease
Variety of outcomes and
exposures
Cheap
Inexpensive
Disadvantages
Prevalent rather than
incident cases
Can not distinguish cause
and effect
Must survive long enough
to be included in study
Short duration diseases
under-represented (e.g.
Influenza)
Cross-Sectional study - bias
Imagine...
People with disease that are sedentary die early
Cross-sectional study of disease (outcome) and
exercise (exposure)
Only sample survivors, so find high proportion of
people who exercise with disease
What would you infer about causal relationships?
Does wearing fluoro gear protect you
from bike crashes?
Participants
Cyclists
Taupo bike
race
Exposure
Outcomes
Bike crash No bike crash
Fluoro colours
162
323
Exposed
Exposed
Disease
Exposed
no disease
588
1343
Participants
Unexposed
No fluoro
colours
Unexposed Unexposed
Disease
No disease
Do computer screens cause
spontaneous abortions?
Cum. Incidence in exposed
Relative risk = ---------------------------Cum. Incidence in unexposed
162/(162+323)
=
------------------- = 1.10
588/(588+1343)
In pictures- Actual
Independent
Case control
Investigator selects cases and controls based on
disease status
Carefully defined population (cases = control
population
Exposure history examined

CC - advantages
Good for long latency/ rare diseases
Evaluate variety of exposures
Smaller sample size

CC - disadvantages
Can't study several diseases
Can't estimate disease risk, because work backwards
from disease to exposure*
More susceptible to selection bias as exposure
already occurred.
More susceptible to information bias
Not efficient for rare exposures

Case control study
Incident vs Prevalent cases
Incident cases from population registry
Prevalent – people with disease at particular point in
time
Incident – exposure and disease tied only to
development of disease, not duration or prognosis.
Prevalent – selection bias/favours long lived, chronic
cases

Case control study
Population vs hospital controls
Hospital controls more likely to have disease related
to exposure, even if not disease of interest.
Population controls, from source of cases, generally
better approach but $$ can be prohibitive

Reality



More complex; rarely have matches, but frequency
matching more common.
E.g. Cot death study
Cases – infants who died from cot death (area)
Method





Sampling frame – all births in geographic area
Frequency matched
Control randomly allocated age for interview similar to
age distribution to cot deaths from previous years (about
3 months old)
DOB calculated and adjusted to fit day of week
(weekends higher chance of becoming cases)
Obstetric hospital randomly chosen in proportion to
number of births in previous financial year
Case Control -example
Fenoterol study, Neil Pearce (guest lecturer)
Participants
Exposure
Fenoterol
Exposed
Outcomes
Cases
Controls
60
Exposed
Disease
189
Exposed
no disease
Participants
Adults in hospital
with asthma
Unexposed
Ventolin/other
57
279
Unexposed Unexposed
Disease
No disease
Effect measure
Odds ratio=
odds of exposure in cases
---------------------------odds of exposure in controls
60/57
= -------------= 1.55
189/279
Actual
Null hypothesis; No effect
Questions

Which study design is best for assessing causation,
assuming no other limitations are present?





A) Cross-sectional study
B) Randomised controlled trial
C) Case-control study
D) Cohort study
E) Case-series
Questions


In a cross sectional study of risk factors for angina, a
random sample of elderly subjects were asked the
question “Do you smoke cigarettes?” Answers were used
to classify respondents as smokers or non-smokers.
Further, subjects were classified as positive for angina if
they had, at some time in the past, been told by a doctor
that they suffered from this condition.
When the data from the study was analysed, no
statistically significant association was found between
cigarette smoking status and angina status.



Has the study measured incidence or prevalence of
angina? Explain your answer.
A considerable body of past evidence suggests that the
risk of angina increases with increasing tobacco
consumption. Suggest reasons why the study described
here failed to find an association.
Suggest an alternative design of study that would be
more suitable for investigating whether smoking causes
angina. Consider the question(s) that you would ask the
chosen subjects about their smoking habits.
Summary
Observational
Experimental
Cohort
Many outcomes, exposures
limited
Case- control
One outcome, many exposures
Cross – sectional
Many exposure, many
outcomes;
Temporality limits causal
inference
Randomised controlled trial
Ethical constraints
Ideal design

similar documents