What is a case control study? Tarani Chandola Social Statistics University of Manchester
Imagine…
.
•
It is 195
0
•
You suspect a
n association b/w smoking and
lung
cancer
•
How
wou
ld you
sho
w th
is?
Cases
•
Find
cases of lun
g cancer (w
here?)
•
What is the expo
sure?
Hypothetical scena
rio:
‐
You fin
d 10
0 cases of lung
cancer
‐
60 of the
m were he
avy sm
okers (>25
cigs/day)
‐
Doe
s this sho
w an association be
tween
sm
oking and lung
cancer?
Controls
•
Whe
re can
you
find
sim
ilar grou
ps of p
atients to
lung
cancer p
atients?
Hypothetical scena
rio:
‐
You fin
d 10
0 control patients who
do no
t have
lung
cancer
‐
20 of the
m were he
avy sm
okers
‐
So 20%
of con
trols and 60
% of cases were he
avy
sm
okers
‐
Doe
s this sho
w an association be
tween sm
oking
and lung
cancer?
Comparing
cases and
con
trols
•
Popu
latio
ns of cases and
con
trols differ
•
So canno
t com
pare % who
are exposed
(differen
t den
ominators/po
pulatio
ns)
•
But y
ou can
com
pare th
e od
ds of h
aving lung
cancer fo
r tw
o grou
ps: the
exposed
and
the
un
expo
sed
Doll and
Hill (1
952)
•
Data was collected
from
hospitalized
patients from
more than
20 ho
spita
ls in
Lon
don over a fo
ur year
pe
riod
(April 1948
and
Feb
ruary 1952).
•
Investigators asked ho
spita
l personn
el to
con
tact
them
whe
never a
patient was adm
itted
to th
e ho
spita
l
with
a new
diagnosis of lun
g cancer (cases).
•
Investigators also selected a rand
om sam
ple of
patie
nts from
the same ho
spita
ls, but with
differen
t
illne
sses (con
trols).
•
Cases and controls were interviewed
abo
ut th
eir
sm
oking habits.
Doll and
Hill (1
952)
•
1465
case‐patie
nts were interviewed
for the
stud
y all und
er th
e age of 75, 135
7 men
and
108
wom
en.
•
Because of th
e dispropo
rtionate ratio of m
en to
wom
en (2
5 to 2), on
ly m
en were includ
ed in
the
fin
al study.
•
An eq
ual num
ber of con
trols (135
7) were
interviewed
abo
ut th
eir sm
oking history.
•
Investigators divide
d sm
oking history into group
s
based on
the average nu
mbe
r of cigarettes
sm
oked
per day.
Doll and
Hill (1
952)‐
results
Cig
aret
tes
smok
ed d
aily
07
6110
-14
565
706
15-2
444
540
825
+34
018
2A
ll Sm
oker
s1,
350
1,29
6
Tota
l1,
357
1,35
7
Case
sCo
ntro
ls
What n
umbe
rs will you
com
pare?
Doll and
Hill (1
952)‐
analysis
Cigarettes
Cases
Controls
Odd
s of
Odd
s of
smoked
daily
smoking
smoking
if case
if con
trol
07
61=7/1350
=61/1296
10‐14
565
706
=565/1350
=706/1296
15‐24
445
408
=445/1350
=408/1296
25+
340
182
=340/1350
=182/1296
All Sm
okers
1350
1296
11
Total
1357
1357
Doll and
Hill (1
952)‐
Odd
s Ra
tios
Cigarettes
Cases
Controls
Odd
s of
Odd
s of
Odd
s of
smoked
daily
smoking
smoking
being a case
if case
if con
trol
given expo
sure
07
610.01
0.05
=0.02/0.05
10‐14
565
706
0.42
0.54
=0.42/0.54
15‐24
445
408
0.33
0.31
=0.33/0.31
25+
340
182
0.25
0.14
=0.25/0.14
All Sm
okers
1350
1296
1.00
1.00
1.00
Doll and
Hill (1
952)‐
Odd
s Ra
tios
Cigarettes
Cases
Controls
Odd
s of
Odd
s of
Odd
s ratio
Odd
s ratio
smoked
daily
smoking
smoking
(rescaled)
if case
if con
trol
07
610.01
0.05
0.11
1.00
10‐14
565
706
0.42
0.54
0.77
6.97
15‐24
445
408
0.33
0.31
1.05
9.50
25+
340
182
0.25
0.14
1.79
16.28
As you move up
the grou
ps, from no cigarettes per day to
25+
cigarettes per day, the
value
s of th
e od
ds ratio rise steadily, con
sisten
t
with
a dose‐respon
se relationship be
tween the expo
sure (d
aily
nu
mbe
r of cigarettes sm
oked
) and
the strength of the
associatio
n
(odd
s ratio
).
Odd
s ratio
•
An approxim
ation of relative risk
•
RR
“how
much more likely are expo
sed
pe
ople to
get th
e disease than
the un
expo
sed”
•
OR
“how
much more likely were pe
ople with
the disease to have be
en exposed
than
those
with
out the
disease”
Calculation of odd
s ratio
D
isea
se
+ (c
ases
)
Dis
ease
-
(con
trol
s)Ex
posu
re +
a b
Expo
sure
- c
d
Calculation of odd
s ratio
in 2 x2 table
D
isea
se
+ (c
ases
)
Dis
ease
-
(con
trol
s)Ex
posu
re +
a b
Expo
sure
- c
d
Odd
s ratio = a:c/b:d
= ad
/bc
= cross‐prod
uct ratio
Selection Bias
•
Selection of app
ropriate con
trols
–
hospita
l based
studies
–
commun
ity based
studies
–
Selection bias with
cases
•
Selection of cases
•
Matched
case‐control studies
–
Cases and controls often
differ in
impo
rtant a
spects (age,
sex, ethnicity, beh
aviours...)
–
These can confou
nd th
e stud
y–
One
way to
elim
inate such differen
ces is m
atching controls
to cases on these factors
–
More than
1 con
trol per case can be
used
Inform
ation Bias
•
retrospe
ctive measuremen
t of e
xposure may be
prob
lematic:
–
inaccurate ‐
will und
erestim
ate any relatio
n
–
repo
rting bias ‐
overestim
ation
–
reverse causation ‐
spurious associatio
n
Coho
rtStart
Une
xposed
Expo
sed
All he
althy
Follow‐up (w
ait)
Disease
assessmen
t
Controls
Cases
Start
Look
back
Case‐Con
trol
Strengths of case‐control studies
•
Quick (cases already
exist, no ne
ed to
wait)
•
Cheap (not necessary to
examine large nu
mbe
r of
pe
ople)
•
Can exam
ine many expo
sures
•
Suita
ble to study
rare diseases
•
Suita
ble to study
stable expo
sures (eg gene
tic
markers)
Weaknesses of case‐control studies
•
Not suitable for rare exposure
•
Cann
ot calculate incide
nce risk or de
ath rates
•
Pron
e to selectio
n bias
•
Pron
e to m
isclassification of exposure
•
Pron
e to reverse causatio
n (peo
ple with
disease
may change their be
haviou
r)
References:
Doll R, H
ill AB (195
2) A study
of the
aetiology of
carcinom
a of th
e lung. B
ritish Medical Jo
urna
l
2:12
71–128
6Ro
thman
K (2
002) Epidemiology. A
n
Introd
uctio
n. Oxford University
Press, O
xford,
En
gland
Schu
lz KF, Grimes DA (2
002) Case‐control
stud
ies: research in reverse;. Lancet: 359
:
431–34