Case- control

advertisement
Introduction to Study Design and
RCTs
Simon Thornley
That sugar movie…

http://gameauland.com/that-sugar-film-teaser-trailer/
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
 Control group obvious (compare case-control)

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
Derive P(exposure | case status)

Case-control study: Cases
Ideally, pcase=pdisease
Prevalent

From survey
people with disease at particular point in time
selection bias/favours long lived, chronic cases

Incident

from population registry
exposure and disease tied only to development of disease,
not duration or prognosis.

Case control study: Controls
Ideally, pcontrol=pundiseased
Population vs. hospital controls
Hospital controls 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

Electoral rolls
Random digit dialling

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
CC - advantages
Good for long latency/ rare diseases
Look at many exposures
Smaller sample size

CC - disadvantages
Only one disease
Can't estimate disease risk, because work backwards
from disease to exposure*
More susceptible to selection bias as exposure has
already occurred.
More susceptible to information bias
Not efficient for rare exposures (Why?)

Case-Control study -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
Characteristic
Crosssectional
Case-control
Cohort
RCT
Selection bias
Medium
High
Low
Low
Recall bias
High
High
Low
Low
Loss to follow up
NA
NA
High
High
Confounding
Medium
Medium
Medium
Low
Time required
Low
Medium
High
High
Cost
Medium
Medium
High
High
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
Download