Transcript Designs of case-control study

Study designs
Alain Moren, Epiconcept, June 2006
Source: EPIET
Cohort study measuring risk
Exposure
Total
Cases
Risk (%)
Risk ratio
Exposed
100
40
40%
4
Unexposed
100
10
10%
Reference
Total
200
50
25%
Denominator = those present at beginning
Usually short duration = outbreak (attack rate)
Cohort study measuring rate
Exposure
Total
Time
Cases
Rate
per 100 p.y.
Rate ratio
Exposed
1500 p.y
40
2.7/100 p.y.
2.7
Unexposed
1000 p.y.
10
1.0/100 p.y.
Reference
Total
2500 p.y.
50
2.0/100 p.y.
Individuals contribute to different length of time
Denominator = sum of times
Cohort study
Exposed population (E)
Initially
at
Risk
NE
Can we use a sample
years at risk
ofPerson
the
denominator
Currently at risk
of exposed (pyar )
E
Cases exposed
CE
Still at risk
NE - CE
Unexposed population (U)
instead of
Initially
at
Risk
Nu
Cases unexposed
CU
Person years at risk
of unexposed (pyarU)
Currently at risk
the entire denominator?
Start of study
Rodrigues L et al. Int J Epidemiol. 1990;19:205-13.
Still at risk
Nu - Cu
End of study
Time
Source population
Cases
E
30
Pop.
100
IR
.3
RR
30/100
3
=3
10/100
E
10
Cases
E
30
100
.1
Ref.
Sample
10
30/10
=3
E
10
10
10/10
Case control study design
Source population
E
E
Cases
Pop.
a
P1
c
P0
Cases
E
E
Cases
Controls
I1 = a / P 1
E
a
b
I0 = c /P0
E
c
d
Sample
a
P1 / y
a / (P1 / y)
c
P0 / y
c / (P0 / y)
a / P1
a / (P1 /y)
---------- = --------------c / P0
c / (P0 /y)
Controls selection
• Controls sampled to mirror the exposure
experience in the source population
– sampled from source population that gives rise
to cases
– representative of exposure in source population
• Sampling independently of exposure status
Cohort study
Exposed population (E)
Initially
at
Risk
NE
Cases exposed
CE
Person years at risk
of exposed (pyarE)
Currently at risk
Still at risk
NE - CE
Unexposed population (U)
Initially
at
Risk
Nu
Cases unexposed
CU
Person years at risk
of unexposed (pyarU)
Start of study
Rodrigues L et al. Int J Epidemiol. 1990;19:205-13.
Currently at risk
Occurrence of
New case
Time
Still at risk
Nu - Cu
End of study
Exposed population
(E)
Traditional
Initially
at
Risk
NE
case control study
Cohort
study
Cases
Person years at risk
of exposed (pyarE)
Currently at risk
and
Cases exposed
CE
Still at risk
NE - CE
Unexposed population (U)
Initially
at
Risk
Nu
Cases unexposed
CU
Person
years at risk
Sample
of “non
of unexposed (pyarU)
E
Start of study
Non E
Rodrigues L et al. Int J Epidemiol. 1990;19:205-13.
cases”
Currently at risk
Cases
Controls
CE Occurrence
(Nof
E-CE ) * f
CU
(NU-CU ) * f
New case
Time
Still at risk
Nu - Cu
End of study
Traditional (exclusive) design
Measure of effect = Odds ratio
• Controls sampled from population still at
risk at the end of the study period
• OR good estimate of risk ratio and rate ratio
if disease is rare
Cohort study
Exposed population (E)
Case
Initially
at
Risk
NE
cohort study
Cases exposed
CE
Cases
Person years at risk
of exposed (pyarE)
Currently at risk
and
Still at risk
NE - CE
Unexposed population (U)
Initially
at
Risk
Nu
Cases unexposed
CU
Person
years at
Sample
ofrisk
source
of unexposed (pyarU)
E
Non E
Start of study
Rodrigues L et al. Int J Epidemiol. 1990;19:205-13.
Cases
CE
CU
population
Currently at risk
Controls
(NE) * f
Occurrence of
U) * f
New(N
case
Time
Still at risk
Nu - Cu
End of study
Case-cohort design
Measure of Risk ratio (relative risk)
• Control group to estimate the proportion of the total
population that is exposed: may include cases
• In a fixed population controls selected from all individuals
at risk at the start of the study
• Controls sampled regardless whether or not they will have
developed the disease
• A person selected as a case may also be selected as a
control and vice versa
• They are kept in both groups
• No need to document disease status among controls
• Example: outbreak of gastro-enteritis with 30% attack rate
Density case control
study
Cohort study
Exposed population (E)
Initially
at
Risk
NE
Cases
Cases exposed
CE
and
Person years at risk
of exposed (pyarE)
Currently at risk
Still at risk
NE - CE
Unexposed population (U)
Initially
at
Risk
Nu
Cases unexposed
Sample of source
CU
population still at risk
Person years at risk
of unexposed (pyarU)
Currently at risk
E
Start of study
Rodrigues L et al. Int J Epidemiol. 1990;19:205-13.
Occurrence of
Non E
New case
Time
Cases
Still at risk
Controls
Nu - Cu
CE
(NpyE) * f
CU
End of
(Nstudy
)*
pyU
f
Density case control (concurrent) design
OR estimates the rate ratio
• Controls are selected concurrently from those still at risk
when a case occur
• A person selected as a control can later become a case
• The opposite not possible: a case no longer at risk
• A control who later becomes a case is kept in both groups
• Controls represent person years at risk experience among
exposed and unexposed
• Match analysis on time of selection is necessary to give
unbiased estimate of rate ratio
Cohort study
Exposed population (E)
Initially
at
Risk
NE
Cases exposed
CE
Person years at risk
of exposed (pyarE)
Currently at risk
Still at risk
NE - CE
Unexposed population (U)
Initially
at
Risk
Nu
Cases unexposed
CU
Person years at risk
of unexposed (pyarU)
Start of study
Rodrigues L et al. Int J Epidemiol. 1990;19:205-13.
Currently at risk
Occurrence of
New case
Time
Still at risk
Nu - Cu
End of study
Cohort populations & measures of association
Cohort design
3 measures of association
Exposed population (E)
Initially
at
Risk
NE
Cases exposed
CE
Person years at risk
of exposed (pyarE)
Currently at risk
Still at risk
NE - CE
Unexposed population (U)
Cases unexposed
CU
Initially
at
Risk
Nu
Person years at risk
of unexposed (pyarU) Currently at risk
Start of study
Occurrence of
New case
Still at risk
Nu - Cu
End of study
Measure of association
Formulation
« Risk ratio»
CE/NE
CU/NU
« Rate ratio»
CE/pyarE
CU/pyarU
« Odds ratio »
CE/(NE- CE)
CU/(NU- CU)
Cohort populations & measures of association
Measures of association
Exposed population (E)
Initially
at
Risk
NE
Person years at risk
of exposed (pyarE)
Currently at risk
Cases exposed
CE
Measure of association
Still at risk
NE - CE
« Risk ratio»
Formulation
Alternative
formulation
CE/NE
CU/NU
CE/CU
NE/NU
CE/pyarE
CU/pyarU
CE/CU
pyarE/pyarU
CE/(NE- CE)
CU/(NU- CU)
CE/CU
(NE- CE) /(NU- CU)
Unexposed population (U)
Cases unexposed
CU
Initially
at
Risk
Nu
Person years at risk
of unexposed (pyarU) Currently at risk
Start of study
Occurrence of
New case
Still at risk
Nu - Cu
End of study
« Rate ratio»
« Odds ratio »
Cohort populations and estimate each
measure of association
Exposed population (E)
Initially
at
Risk
NE
Measure of association
Cases exposed
CE
Person years at risk
of exposed (pyarE)
Currently at risk
« Risk ratio»
Unexposed population (U)
Cases unexposed
CU
Person years at risk
of unexposed (pyarU) Currently at risk
Start of study
Occurrence of
New case
Alternative
formulation
CE/NE
CU/NU
CE/CU
NE/NU
CE/pyarE
CU/pyarU
CE/CU
pyarE/pyarU
CE/(NE- CE)
CU/(NU- CU)
CE/CU
(NE- CE) /(NU- CU)
Still at risk
NE - CE
« Rate ratio»
Initially
at
Risk
Nu
Formulation
« Odds ratio »
Still at risk
Nu - Cu
Numerator = ratio of exposed to non-exposed cases
Denominator = ratio of exposed to unexposed:
End of study
persons at risk at start of the study
person years at risk for the duration of the study
persons still disease free at the end of study
Cohort populations and controls selection for
the OR to estimate each measure of association
Exposed population (E)
Initially
at
Risk
NE
Formulation
Alternative
formulation
«Risk ratio»
CE/NE
CU/NU
CE/CU
NE/NU
Total study population
regardless of past or
future disease status
«Rate ratio»
CE/pyarE
CU/pyarU
CE/CU
pyarE/pyarU
People currently at risk
CE/(NE- CE)
CU/(NU- CU)
CE/CU
(NE- CE) /(NU- CU)
People disease-free
throughout study period
Cases exposed
Measure of association
CE
Person years at risk
Of exposed (pyarE) Currently at risk
Still at risk
NE - CE
Unexposed population (U)
Cases unexposed
CU
Initially
at
Risk
Nu
Person years at risk
of unexposed (pyarU) Currently at risk
Start of study
Occurrence of
New case
« Odds ratio »
Controls to be
sampled from
Still at risk
Nu - Cu
End of study
Numerator = ratio of exposed to non-exposed cases
Denominators = ratio of exposed to non-exposed
persons at risk at start of the study
person years at risk for the duration of the study
persons still disease free at the end of study
How to select controls to estimate the
respective measure of association
Formulation
Alternative
formulation
«Risk ratio»
CE/NE
CU/NU
CE/CU
NE/NU
Total study population
regardless of past or
future disease status
«Rate ratio»
CE/pyarE
CU/pyarU
CE/CU
pyarE/pyarU
People currently at risk
CE/(NE- CE)
CU/(NU- CU)
CE/CU
(NE- CE) /(NU- CU)
People disease-free
throughout study period
Measure of association
« Odds ratio »
Controls to be
sampled from
Cohort populations, control selection to
estimate each measure of association &
corresponding designs
Exposed population (E)
Initially
at
Risk
NE
Cases exposed
CE
Person years at risk
Of exposed (pyarE)
Measure of association
Currently at risk
Unexposed population (U)
Still at risk
NE - CE
«Risk ratio» or «cumulative
incidence ratio »
«Rate ratio » or incidence
Cases unexposed density ratio »
Formulation
Alternative
formulation
Controls to be
sampled from
CE/NE
CU/NU
CE/CU
NE/NU
Total study population
regardless of past or
future disease status
CE/pyarE
CU/pyarU
CE/CU
pyarE/pyarU
People currently at risk
CE/(NE- CE)
CU/(NU- CU)
CE/CU
(NE- CE) /(NU- CU)
People disease-free
throughout study period
CU
Initially
at
Risk
Nu
Person years at risk
of unexposed (pyarU) Currently at risk
« Odds ratio »
Still at risk
Nu - Cu
Case-cohort design
Start of study
Occurrence of
New case
End of study
Density case control design
Traditional case control design
What design and when?
Traditional case control
- rare disease
Case cohort
- frequent disease
- same denominator over time
- non recurrent outcome
Density case control
- rare or frequent disease
- exposure changes over time
- non or recurrent outcome
OR in case-control studies: the rare
disease assumption
• Case control study very efficient for rare diseases
• Initially used for testing significant differences in exposure
without attempting to quantify the risk associated with
exposure :« Statistically do more lung cancer patients have a
history of smoking than controls ?» rather than « by how many
times does smoking increase the risk of lung cancer ? »
• Cornfield (1961): if disease is rare : OR ~ RR
• Used more and more for common diseases
• Miettinen (1976), Greenland (1981), Smith (1984) : if controls
chosen appropriately, no rare disease assumption is needed
for the OR to estimate the relative risk or rate !
Incidence of breast cancer after radiation
Total Cases Non C. Rate
RR
E 28010 41
27969 14.6
1.9
E 19017 15
19002
Ref.
7.9
Incidence of breast cancer after radiation
Total
Cases Non C. Rate
RR
Random
Contr.
OR
28010
41
27969
14.6
1.9
280
1.9
19017
15
19002
7.9
Ref.
190
Ref.
Incidence of breast cancer after radiation
Total
Cases Non C. Rate
RR
Random
Contr. OR
Non
Cases.
Contr. OR
28010
41
27969 14.6
1.9
280
1.9
279
1.9
19017
15
19002
Ref.
190
Ref.
190
Ref.
7.9
Outbreak of food borne disease in a nursing home
100 residents, 40 cases
Cohort
Cases
Non
cases
Source
Population
E
36
24
60
E
4
36
40
RR = 6
Outbreak of food borne disease in a nursing home
100 residents, 40 cases
Cohort
Cases
Non
cases
Potential control groups
Source
Population
Non
cases
E
36
24
60
12
E
4
36
40
18
RR = 6
OR = 13.5
Outbreak of food borne disease in a nursing home
100 residents, 40 cases
Cohort
Cases
Non
cases
Potential control groups
Source
Population
Non
cases
Source
Population
E
36
24
60
12
30
E
4
36
40
18
20
RR = 6
OR = 13.5
OR = 6
Rare disease assumption = wrong issue
Issue = selection of controls
Alternative designs
• « Case-to-case »
• « Case-crossover »
« Case-to-case approach »
Source: Jean Claude Desenclos, Jet De Valk
Two listeriosis outbreaks of 2 distinct PFGE patterns,
France, 1999-2000
Cases
10
9
Outbreak 2 (32 cases)
October
November
1999
December
8
6
4
2
52
50
48
46
44
Outbreak 1 (10 cases)
42
40
8
7
6
5
4
3
2
1
0
January
February March
2000
de Valk H et al. Am J Epidemiol 2001;154:944-50
Listeriosis outbreak cases and sporadic cases
distinguished by routine PFGE, France, 1999-2000
Cases
14
Sporadic cases
12
Outbreak 2 (32 cases)
10
Outbreak 1 (10 cases)
8
6
4
2
October
November
1999
December
8
6
4
2
52
50
48
46
44
42
40
0
January
February March
2000
de Valk H et al. Am J Epidemiol 2001;154:944-50
Controls selected among sporadic cases for the case
to case control study, listeriosis outbreak 2, France,
1999-2000 (Source: InVS-CNR)
Cases
14
12
10
Other sporadic cases
Sporadic cases used as controls (N = 32)
Outbreak 2 (N = 32)
Outbreak 1 (N = 10)
8
6
4
2
December
8
6
4
2
52
November
1999
50
48
46
October
44
42
40
0
January
February March
2000
de Valk H et al. Am J Epidemiol 2001;154:944-50
Food consumption of case-patients and controlsubjects, multivariate analysis on 29 case-patients
and 32 control-subjects. Outbreak of listeriosis,
France, December 1999 - February 2000.
Food consumed
Pork tongue in jelly
Cooked ham
Pâté de campagne
Adjusted
Odds ratio*
95% CI
p
75.5
7.1
8.9
4.7 – 1216.0
0.7 – 71.8
1.7 – 46.1
0.002
0.1
0.009
*adjusted for underlying condition, pregnancy status and date of interview
by logistic regression
de Valk H et al. Am J Epidemiol 2001;154:944-50
« Case-to-case » control study
• Possible if disease can be classified
in subgroups that have specific risk factors
• May be the case for infectious agents subtypes?
• Controls = cases with non epidemic subtypes
–
–
–
–
from same source population?
same susceptibility (underlying diseases)
included as cases if they had the outbreak strain
readily available
• Reduces the information (recall) bias
• Food-exposure collected before status is known
The case-crossover design
The case-crossover design
• Same person taken as its own control (matched design)
• Compare exposure in a « risk period » to a prior
« control period » of the same duration
• No control group needed
• Only pairs of period discordant for the exposure of
interest used in the analysis
• Acute diseases & exposures that change overtime
• Transient exposures (drug adverse events…)
• Key issue : the definition of the risk period
Reference
period
“Wash out”
Current
period
period
Onset
Exposure
Cases
Matched pairs
1
Discordant 0, 1
2
Discordant 1, 0
3
Concordant 1, 1
4
Concordant 0,0
« Case crossover » design applied to a
prolonged Salmonella Typhimurium
outbreak
Discordant pair ( 0,1 )
Discordant pair ( 1,0 )
Concordant pair ( 1,1 )
Concordant pair ( 0,0 )
Control period
72 hours
“Wash out”
period
48 hours
Risk period
72 hours
Exposure
Onset
Haegebaert S et al. Epidemiol infect 2003;130,1-5
Food exposures from menu information in the risk and
control period and matched OR for 17 nosocomial cases
Foods
Risk
period
Exposed (%)
Veal
Pork
Hamburgers
Ham
Pâté
Chicken
Turkey
“Cordon bleu”
Lamb sausages
Poultry sausages
5 (29)
4 (23)
13 (77)
6 (35)
2 (12)
2 (12)
11 (65)
0 (0)
2 (12)
2 (12)
Control
period
Matched
OR
Exposed (%)
1 (6)
6 (35)
5 (29)
5 (29)
2 (12)
3 (18)
6 (35)
2 (12)
0 (0)
0 (0)
5
0,6
5
1,5
1
1
2,67
undefined
undefined
undefined
95%
C.I.
0,6 - 236,5
0,1 - 3,1
1,1 - 46,9
0,2 - 17,9
0,01 - 78,5
0,01 - 78,5
0,7 - 15,6
-
Haegebaert S et al. Epidemiol infect 2003;130,1-5
Case-crossover design
•
•
•
•
No need of a control group
One to several control-periods per risk period
Controls for « between-persons » confounding
Need of data collected prior to onset
(administrative source),
• If exposure collected by interview
then very sensitive to recall bias
• May be biased by time trend in exposure:
between-period confounding
• « Case-time-control design »
References
1.
2.
3.
4.
5.
6.
7.
8.
Rodrigues L et al. Int J Epidemiol 1990;19:205-13
de Valk H et al. Am J Epidemiol 2001;154:944-50
Haegebaert S et al. Epidemiol infect 2003;130,1-5
Hernandez-Diaz S et al. Am J Epidemiol 2003;158:385-391
Rothman KJ; Epidemiology: an introduction. Oxford University Press 2002,
73-93
Suisa S. The case-time-control design. Epidemioogy. 1995;6:248-253.
Greenland S. Confounding and exposure trends in Case-cross-over and
case-time-control designs. Epidemiology. 1996; 7231-239.
Mittleman, Maclure, Robins. Control sampling strategies for case crossover studies: An assessment or relative effectiveness. A J Epidemiol.
142;1:91-98.
Cohort study
Exposed population (E)
Initially
at
Risk
NE
Cases exposed
CE
Person years at risk
of exposed (pyarE)
Currently at risk
Still at risk
NE - CE
Unexposed population (U)
Initially
at
Risk
Nu
Cases unexpose
CU
Person years at risk
of unexposed (pyarU)
Start of study
Rodrigues L et al. Int J Epidemiol. 1990;19:205-13.
Currently at risk
Occurrence of
New case
Time
Still at risk
Nu - Cu
End of study
Surveillance of human listeriosis, France
Investigation
Mandatory notification
Clinicians
DDASS
InVS
Routine
extended
food
questionnaire
Cluster
detection
>3 isolates
with same
NRC
PFGE pattern
Routine ongoing in 10 weeks
Lm human isolates
Laboratories
PFGE typing
• Timely detection of clusters
• Food history readily available when cluster recognised
• “Case to case” case-control studies