Upload
salman-bin-mahmood
View
216
Download
0
Embed Size (px)
Citation preview
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
1/47
Analysis and presentation of
Case-control study data
Chihaya Koriyama
February 14 (Lecture 1)
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
2/47
Study design in
epidemiology
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
3/47
Why case-control study?
In a cohort study you need a lar!e number
of the sub"ects to obtain a sufficient number
of case especially if you are interested in a
rare disease#$ %astric cancer incidence in &apanese male'
1#* + 1,,,,, person year
A case-control study is more efficient in
terms of study operation time and cost#
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
4/47
Comparison of the study design
Case-control Cohort
are diseases suitable not suitable
.umber of disease 1 1/0ample sie relati2ely small need to be lar!e
Control selection difficult easier
0tudy period relati2ely short lon!
ecall bias yes nois3 difference no a2ailable a2ailable
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
5/47
Case-control study- 0euence of determinin! e5posure and outcome status
0tep1' Determine and select cases of
your research interest
0tep' Selection of appropriate controls
0tep6' Determine exposure status in
both cases and controls
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
6/47
Case ascertainment
7hat is the definitionof the case8
$ Cancer (clinically8 9atholo!ically8)
$ :irus carriers (Asymptomatic patients)
;
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
7/47
Who will be controls?
Control ? non-case
$ Controls are also at ris3 of the disease
in his(her) future#
$ @Controls are e5pected to be arepresentati2e sample of the catchment
population from =hich the case arise#
$ In a case-control study of !astriccancer a person =ho has recei2ed the
!astrectomy cannot be a control since
he ne2er de2elop !astric cancer #
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
8/47
a population-basedcase-control study
Both cases and controls are recruited from the
population#
a case-control study nested in a cohort
Both case and controls are members of the cohort#
a hospital-basedcase-control studyBoth case and controls are patients =ho are
hospitalied or outpatients#
Controls =ith diseases associated =ith the e5posure
of interest should be a2oided#
:arious types of case-control studies
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
9/47
he follo=in! points should be
recorded (described in your paper) he list (number) of eli!ible cases
=hose medical records una2ailable
he list (number) of refused sub"ects ifpossible =ith descriptions of the
reasons of refusal
he len!th of inter2ie=
he list (number) of sub"ects lac3in!
the measurement data =ith
descriptions of the reasons
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
10/47
Exploratory or Analytic
D5ploratorycase-control studies
$ here is no specific a priori
hypothesis about the relationship
bet=een e5posure and outcome#
Analyticcase-control studies
$Analytic studies are desi!ned to test
specific a priori hypotheses aboute5posure and outcome#
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
11/47
Case-control study - information
0ources of the information of e5posure and
potential confoundin! factors
$ D5istin! records
$ Euestionnaires
$ Face-to-face + telephone inter2ie=s
$ Biolo!ical specimens
$ issue ban3s$ atabases on biochemical and
en2ironmental measurements
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
12/47
emporality is essential in GillHs criteria
he study e5posure
is unli3ely to bealtered at this sta!e
because of the
disease#
he study e5posure
is more li3ely to bealtered at this sta!e
because of the
symptoms#
Dssential Dpidemiolo!y (7A lec3no)
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
13/47
Bias should be minimied
Bias > Confoundin!
$ 0election bias
$ etection bias
$ Information bias (recall bias)
$ Confoundin!
Confoundin! can be controlledby statistical analyses but =e
can do nothin! about bias after
data collection#
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
14/47
Case-control studies
are potential
sources of many
biasesshould be carefully
desi!ned analyed
and interpreted#
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
15/47
!ow can we sol"e the problem of
confounding in a case-control study?
Prevention at study design
Limitation
Matchingin a cohortstudy But not in a case-control study
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
16/47
#atching in a case-control study
Jatched by confoundin!
factor(s) to increase the
efficiency of statistical analysis
Cannotcontrol confoundin!
$A conditional lo!istic analysis is
reuired#
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
17/47
$"er matching
Jatched by factor(s) stron!ly
related to the e5posure =hich is
your main interest
$ CA.. see the difference in
the e5posure status bet=een
cases and controls
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
18/47
!ow can we sol"e the problem of
confounding?
Treatment at statistical analysis
Stratifcationby a conounder
Multivariateanalysis
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
19/47
7hat you should describe in the
materials and methods1# 0tudy desi!n
# efinition of eli!ible cases and
controls
$ Inclusion + e5clusion criteria of
cases and controls
6# .umber of the respondents
and response rate4# Jain e5posure and other
factors includin! potential
confoundin! factors
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
20/47
*# 0ources of the information of
e5posure and other factors
# Jatched factors if any# he number of sub"ects used
in statistical analyses
# 0tatistical test(s) and model(s)
M# .ame and 2ersion of the
statistical soft=are
7hat you should describe in the
materials and methods
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
21/47
Assurin! adeuate study po=er
Follo=in! information is necessary
$ he confidence le2el desired (usually M*N
correspondin! to a p-2alue of ,#,*)
$ he le2el of po=er desired (,-M*N)$ he ratio of controls to cases
$ he e5pected freuency of the e5posure in
the control !roup
$ he smallest odds ratio one =ould li3e to beable to detect (based on practical
si!nificance)
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
22/47
0tatistical analysis
@Jatched 2s# @Onmatched studieshe procedures for analyin! the
results of case-control studies
differ dependin! on =hether the
cases and controls are matched orunmatched#
#atched %nmatched
Jc.emarHs test
Chi-suare test Conditional lo!istic Onconditional
lo!istic
re!ression analysis re!ression analysis
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
23/47
Ad2anta!esof pair matchin! in case-
control studies Assures comparability bet=een cases andcontrols on the selected 2ariables
Jay simplify the selection of controls by
eliminatin! the need to identify a randomsample
Oseful in small studies =here obtainin! cases
and controls that are similar on potentially
confoundin! factors may other=ise be difficult
Can assure adeuate numbers of sub"ects =ith
specified characteristics so as to permit
statistical comparisons Dssential Dpidemiolo!y (7A lec3no)
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
24/47
isd2anta!esof pair matchin! in case-
control studies Jay be difficult or costly to find a sufficient
number of controls
Dliminates the possibility of e5aminin! the effects
of the matched 2ariables on the outcome Can increase the difficulty or comple5ity of
controllin! for confoundin! by the remainin!
unmatched 2ariables
2ermatchin!
Can result in a !reater loss of data since a pair
of sub"ects has to be eliminated e2en if ne
sub"ect is not responsi2e Dssential Dpidemiolo!y (7A lec3no)
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
25/47
Lun! cancer Controlscases
.P1,, .P1,,
0mo3ers (. recently started) Q Q
, 4,
An e5ample of unmatched case-control study
Cases Controls
smoker ! "!
#on-smoker $! %!
dds ratioP
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
26/47
&is' measure in a case-control study
$dds( pre"alence ) *+ pre"alence,
$dds ratio( odds in cases ) odds in controls
Disease
case control a c
Exposure b d
Exposure odds in cases a ) b
Exposure odds in controlsc ) d
$dds ratio*a ) b, ) *c ) d, a . d ) b . c
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
27/47
Lun! cancer Jatched controlsCases by se5 > a!e
.P1,, .P1,,
0mo3ers (. recently started)
Q Q
, 4,
An e5ample of matched case-control study
Case
Smoker #on-
smoker
Controlsmoker $! &!
#on-smoker "! '!.otice that this is the distribution of 1,, matched
pairs#
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
28/47
Jc.emarHs test
Case
Smoker #on-smoker
Control
smoker $! &!
#on-smoker "! '!
Chi-suare (test) statistic
P (4, $ 1,)+ (4,R1,)
P 1
=here de!ree of freedom is @1#
dds ratio P 4, + 1, P 4
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
29/47
Lo!istic re!ression analysis
Lo!istic re!ression is used to
model the probability of a
binary response as a function
of a set of 2ariables thou!ht topossibly affect the response
(called co2ariates)#
1' case (=ith the disease)
< P
,' control (no disease)
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
30/47
ne could ima!ine tryin! to fit a linear model
(since this is the simplest model S) for the
probabilities but often this leads to problems'
In a linear model fitted probabilities can fall
outside of , to 1# Because of this linear models
are seldom used to fit probabilities#
9robability
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
31/47
In a lo!istic re!ression analysis the
logitof the probability is modelled
rather than the probability itself#
9 P probability of !ettin! disease
plo!it (p) P lo!
1-p
As al=ays =e use the natural lo!# he lo!it
is therefore the log odds
since odds P p + (1-p)
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
32/47
Simple logistic regression *with a continuousco"ariate,
0uppose =e !i2e each of se2eral beetles some
dose of a potential to5ic a!ent (x
Pdose) and =eobser2e =hether the beetle dies (
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
33/47
he 2alues of and =ill determine =hether or
not and ho= steeply the dose-response cur2e
rises (or falls) and =here it is centered#
If P , p5is constant o2er 5
T , p5increases =ith 5 / , p5decreases =ith 5
G,' P , is the null hypothesis in a @test of trend=hen 5 is a continuous 2ariable# Kno=led!e of =ould !i2e us insi!ht to the direction and de!ree
of association outcome and e5posure#
e (R5)
95 P
1 R e(R5)
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
34/47
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
35/47
In case of e5posure (VP1)' lo!it(9D1)Pintercept R
In case of non-e5posure (VP,)' lo!it (9D,) Pintercept
If you =ant to obtain odds ratio of e5posure !roup
9D1+ (1-9D1)+ (9D,+ (1-9D,))
lo!() P lo! W9D1+ (1-9D1)+ (9D,+ (1-9D,))X
P lo! (9D1+ (1-9D1)) $ lo!(9D,+ (1-9D,))
P lo!it (9 for e5posure) $ lo!it (9 for non-e5posure)
P (intercept R ) $ intercept
P $& ( e
efinition of odds
ratio
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
36/47
Simple logistic regression
*with a co"ariate ha"ing more than two categories,
0uppose =e are considerin! a case-control study =here
the predictor 2ariable is current smo3er + e5-smo3er +non-smo3er =hich =e @code as a dummy 2ariable#
Case Smokingstatus
SMK1(X1)
SMK2(X2)
& (urrent & !
! )*-smoker ! &
& #on-smoker ! !
& )*-smoker ! &
! #on-smoker ! !
! #on-smoker ! !
$riginal data Dummy "ariables
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
37/47
Lo!istic re!ression model of the pre2ious e5ample
lo!it (9) P R 1(V1) R (V)
In case of current smo3er (V1P1 VP,)'
lo!it(9current)P R 1
In case of e5-smo3er (V1P, VP1) '
lo!it(9e5)P R 2
In case of non-smo3er (V1P, VP,) 'lo!it(9non)P
$¤t( e
$&ex( e
$&non( + *referent,
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
38/47
7aldHs test for no association
he null hypothesis of no association bet=eenoutcome and e5posure corresponds to
G,' P1 or G,' Plo!P,
Osin! lo!istic re!ression results =e can testthis hypothesis usin! standard coefficients or
7aldHs test#
.ote' 0AA and 0A0 present t=o-sided
7aldHs test p-2alues#
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
39/47
Li3elihood atio est (L)
An alternati2e =ay of testin! hypotheses in alo!istic re!ression model is =ith the use of a
li3elihood ratio test# he li3elihood ratio test
is specifically desi!ned to test bet=een
nested hypotheses#
G,' lo! (95+ (1-95)) P
GA' lo! (95+ (1-95)) P R 5
and =e say that G,is nested in GA#
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
40/47
Li3elihood atio est (L)
In order to test G,2s# GA =e compute the li3elihood
ratio test statistic'
%P -lo!(LG,+ LGA) P (lo! LGA$ lo! LG,)
P (-lo! LG,) $ (-lo! LGA)
7here
LGA is the ma5imied li3elihood under the
alternati2e hypothesis GAand
LG, is the ma5imied li3elihood under the nullhypothesis G,#
If the null hypothesis G,=ere true =e =ould e5pect
the li3elihood ratio test statistic to be close to ero#
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
41/47
7aldHs test 2s# L
In !eneral the L often =or3s a little better thanthe 7ald test in that the test statistic more closely
follo=s a X2distribution under G,# But the 7ald test
often =or3s 2ery =ell and usually !i2es similar
results#
Jore importantly the L can more easily be
e5tended to multi2ariate hypothesis tests e#!#
G,' 1P P , 2s# GA' 1 P P ,
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
42/47
7orld %astroenterolo!y ,,
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
43/47
216CASES
173ormalin!"#ed
para$n!em%edded%lo&ks
We could not obtain the information
on tumor location for /0 cases1 and
those cases were excluded from the
tumor location specific analysis2
1 cases were
excluded
7
6'
1
16
E*+SE, -./A-0C0/A-E
0 -E S-+,
405E, 05A44E ,E4 CA+CA
4ESS -A ' EAS
EC+E- CASES
C.+4, .C.-AC-
&ecruitment of cases
1
6
4
/A-0E-SE4
,0A.SE,AS 8C8
3'
Sep2/333
Dec2/33/
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
44/47
931C.-.4S
/.-E-0A4C.-.4S
'2:
67
1
2
405E, 05A44E ,E4 CA+CA
4ESS -A ' EAS
E*+SE, -.
/A-0C0/A-E
0 -E S-+,
istry o 8C8
&ecruitment of controls1
6
#atchedby sex1 age *4-year ,1hospital1 date of
administration
Case5 control( + 5 /
#a6or diseases of controls cardio"ascular diseases
/37
trauma ++8
infectious diseases 3
urological disorders 21)
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
45/47
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
46/47
xi5clogitcasocon i2fumar1 group*identi,or
Conditional *fixed-effects, logistic regression .umber of obs P 4
L chi() P 4#4
9rob T chi P ,#,M
Lo! li3elihood P -64#*4* 9seudo P ,#,,M
---------------------------------------------------------------------------------------------------
casocon U dds atio 0td# Drr# 9TUU [M*N Conf# Inter2al\
-------------R-------------------------------------------------------------------------------------
YIfumarY1 U 1#*6*,6 #6,1MM #1* ,#,6 1#,6M* #M6M
YIfumarY U 1#1M*1 #4,4 ,# ,#64 #MM 1#M,M*
---------------------------------------------------------------------------------------------------
Wald9s test p "alues
FumarP,
FumarP1
FumarP
&esults of conditionallogistic regression analysis using the same data
Case Control $& *:4;C
7/23/2019 7 case-control_analysis_Chihaya_hundout.ppt
47/47
=C ris' by smo'ing in Cali1 Colombia
results of tumor-location specific analysis
P ( 324+ 9 2alue by L
his test e5amines the difference in the ma!nitude of the
association bet=een smo3in! and %C ris3 amon! 6 tumor