1:1 Matched case-control study NOT POSSIBLE with GENLIN (epilogue)

[Marta Garcia-Granero]

Hi everybody:

As an epilogue to the interesting exchange of mails I've had with Bruce,
Max and Ryan, here is the chronicle of my attempts to use GENLIN to get
the correct OR in a case-control matched study using conditional
logistic regression.

Since GENLIN doesn't allow the response variable to be within subjects
(I spent quite a long time reading all the program help on the topic
until I found that out), the matched case-control design, where,
obviously the outcome (case vs control) is a within pairs variable,
can't be analyzed with it. Period. BTW, I would have liked that anyone
of the SPSS support folks that usually participate in this list had told
me so before letting me spend several hours finding that out by myself :(

Why Bruce's syntax (*) worked? Well, it is a matched COHORT study:
people (GP vs specialists) were matched 1:3 and then the outcome was
observed later. GENLIN can handle that, since the outcome is not the
within subjects variable. Bruce, could you please send me one of the
datasets you mentioned? I'd like to use it in my next Statistics course
(citing the origin and giving you credit for the syntax, of course).

Therefore, for a matched cohort study, GENLIN can be used, and for a
matched case-control study, COXREG must be used instead.

Happy weekend to everybody, although Spain is expected to suffer a so
called "perfect storm" (something slightly below a hurricane) this week
end, and I have to travel :(

Marta GG

(*)
http://groups.google.com/group/comp.soft-sys.stat.spss/browse_frm/thread/ce4c51d35ea66713/cdccfbc40e0f35b3?lnk=gst&q=Weaver+Logistic+regression+%28via+GENLIN%29#cdccfbc40e0f35b3

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Bruce Weaver]

Thanks for this summary Marta.  I need to check with the principal investigator of the study before sending you the data, but I don't foresee any problem with that.  

Bruce

García-Granero wrote

Hi everybody:

As an epilogue to the interesting exchange of mails I've had with Bruce,
Max and Ryan, here is the chronicle of my attempts to use GENLIN to get
the correct OR in a case-control matched study using conditional
logistic regression.

Since GENLIN doesn't allow the response variable to be within subjects
(I spent quite a long time reading all the program help on the topic
until I found that out), the matched case-control design, where,
obviously the outcome (case vs control) is a within pairs variable,
can't be analyzed with it. Period. BTW, I would have liked that anyone
of the SPSS support folks that usually participate in this list had told
me so before letting me spend several hours finding that out by myself :(

Why Bruce's syntax (*) worked? Well, it is a matched COHORT study:
people (GP vs specialists) were matched 1:3 and then the outcome was
observed later. GENLIN can handle that, since the outcome is not the
within subjects variable. Bruce, could you please send me one of the
datasets you mentioned? I'd like to use it in my next Statistics course
(citing the origin and giving you credit for the syntax, of course).

Therefore, for a matched cohort study, GENLIN can be used, and for a
matched case-control study, COXREG must be used instead.

Happy weekend to everybody, although Spain is expected to suffer a so
called "perfect storm" (something slightly below a hurricane) this week
end, and I have to travel :(

Marta GG

(*)
http://groups.google.com/group/comp.soft-sys.stat.spss/browse_frm/thread/ce4c51d35ea66713/cdccfbc40e0f35b3?lnk=gst&q=Weaver+Logistic+regression+%28via+GENLIN%29#cdccfbc40e0f35b3

=====================
To manage your subscription to SPSSX-L, send a message to
[e-mail] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Marta Garcia-Granero]

Bruce Weaver wrote:
> Thanks for this summary Marta.  I need to check with the principal
> investigator of the study before sending you the data, but I don't foresee
> any problem with that.
>

Just in case it makes that other investigator happier, I could modify
the variable names: for instance, turn it into a matched cohort of
smokers-non smokers (1:3) and make the outcome variable infarction, or
something like that. Also, I'd be really happy with only a subset of the
data (some strata selected randomly).

Thank you very much,
Marta

>> Hi everybody:
>>
>> As an epilogue to the interesting exchange of mails I've had with Bruce,
>> Max and Ryan, here is the chronicle of my attempts to use GENLIN to get
>> the correct OR in a case-control matched study using conditional
>> logistic regression.
>>
>> Since GENLIN doesn't allow the response variable to be within subjects
>> (I spent quite a long time reading all the program help on the topic
>> until I found that out), the matched case-control design, where,
>> obviously the outcome (case vs control) is a within pairs variable,
>> can't be analyzed with it. Period. BTW, I would have liked that anyone
>> of the SPSS support folks that usually participate in this list had told
>> me so before letting me spend several hours finding that out by myself :(
>>
>> Why Bruce's syntax (*) worked? Well, it is a matched COHORT study:
>> people (GP vs specialists) were matched 1:3 and then the outcome was
>> observed later. GENLIN can handle that, since the outcome is not the
>> within subjects variable. Bruce, could you please send me one of the
>> datasets you mentioned? I'd like to use it in my next Statistics course
>> (citing the origin and giving you credit for the syntax, of course).
>>
>> Therefore, for a matched cohort study, GENLIN can be used, and for a
>> matched case-control study, COXREG must be used instead.
>>
>> Happy weekend to everybody, although Spain is expected to suffer a so
>> called "perfect storm" (something slightly below a hurricane) this week
>> end, and I have to travel :(
>>
>> Marta GG
>>
>> (*)
>> http://groups.google.com/group/comp.soft-sys.stat.spss/browse_frm/thread/ce4c51d35ea66713/cdccfbc40e0f35b3?lnk=gst&q=Weaver+Logistic+regression+%28via+GENLIN%29#cdccfbc40e0f35b3
>>
>>

--
For miscellaneous SPSS related statistical stuff, visit:
http://gjyp.nl/marta/

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[MaxJasper]

Thanks Marta for sharing the information. Can you also post syntax for each procedure which can best fit your posted data on asthma patients? Thanks.

Max.

|
|Hi everybody:
|
|As an epilogue to the interesting exchange of mails I've had
|with Bruce, Max and Ryan, here is the chronicle of my attempts
|to use GENLIN to get the correct OR in a case-control matched
|study using conditional logistic regression.
|
|Since GENLIN doesn't allow the response variable to be within
|subjects (I spent quite a long time reading all the program
|help on the topic until I found that out), the matched
|case-control design, where, obviously the outcome (case vs
|control) is a within pairs variable, can't be analyzed with
|it. Period. BTW, I would have liked that anyone of the SPSS
|support folks that usually participate in this list had told
|me so before letting me spend several hours finding that out
|by myself :(
|
|Why Bruce's syntax (*) worked? Well, it is a matched COHORT
|study: people (GP vs specialists) were matched 1:3 and then
|the outcome was observed later. GENLIN can handle that, since
|the outcome is not the within subjects variable. Bruce, could
|you please send me one of the datasets you mentioned? I'd like
|to use it in my next Statistics course (citing the origin and
|giving you credit for the syntax, of course).
|
|Therefore, for a matched cohort study, GENLIN can be used, and
|for a matched case-control study, COXREG must be used instead.
|
|Happy weekend to everybody, although Spain is expected to
|suffer a so called "perfect storm" (something slightly below a
|hurricane) this week end, and I have to travel :(
|
|Marta GG

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Bruce Weaver]

García-Granero wrote

Bruce Weaver wrote:
> Thanks for this summary Marta.  I need to check with the principal
> investigator of the study before sending you the data, but I don't foresee
> any problem with that.
>

Just in case it makes that other investigator happier, I could modify
the variable names: for instance, turn it into a matched cohort of
smokers-non smokers (1:3) and make the outcome variable infarction, or
something like that. Also, I'd be really happy with only a subset of the
data (some strata selected randomly).

Thank you very much,
Marta

I've written to ask him, and don't anticipate a problem.  

Meanwhile, I've answered my own question about using NOMREG.  Here is Marta's original syntax with NOMREG tacked on at the end.

* Compute odds ratio for 1:1 matched case-control study.
* Syntax posted to SPSSX-L mailing list by Marta Garcia-Granero.
* Date: 23-Feb-2010 .
* Link: http://www.listserv.uga.edu/cgi-bin/wa?A1=ind1002&L=spssx-l#2 .

* Sample dataset (from Campell's "Statistics at Square Two") 1:1 matched
case-control study *.
DATA LIST FREE/PairNumber CaseControl Monitoring (3 F8).
BEGIN DATA
 1 1 1  1 0 1  2 1 1  2 0 1  3 1 1  3 0 1  4 1 1  4 0 1  5 1 1  5 0 1
 6 1 1  6 0 1  7 1 1  7 0 1  8 1 1  8 0 1  9 1 1  9 0 1 10 1 1 10 0 1
11 1 1 11 0 0 12 1 1 12 0 0 13 1 1 13 0 0 14 1 1 14 0 0 15 1 1 15 0 0
16 1 1 16 0 0 17 1 1 17 0 0 18 1 1 18 0 0 19 1 1 19 0 0 20 1 1 20 0 0
21 1 1 21 0 0 22 1 1 22 0 0 23 1 1 23 0 0 24 1 0 24 0 1 25 1 0 25 0 1
26 1 0 26 0 1 27 1 0 27 0 0 28 1 0 28 0 0 29 1 0 29 0 0 30 1 0 30 0 0
31 1 0 31 0 0 32 1 0 32 0 0 33 1 0 33 0 0 34 1 0 34 0 0 35 1 0 35 0 0
END DATA.
VALUE LABELS CaseControl 0 'Survivor (Control)' 1 'Death (Case)'
/Monitoring 0 'Adequate' 1 'Inadequate'.

* Goal (from Campbell's book and hand calculations): OR=4.333 (95%CI:
1.235 to 15.206) *.

* Using stratified analysis (Mantel-Haenszel estimator of OR): Correct *.
CROSSTABS
  /TABLES=Monitoring BY CaseControl BY PairNumber
  /FORMAT=NOTABLES
  /STATISTICS=CMH(1).

* Using COXREG (with a false time variable called "S": Correct *.
COMPUTE S=(CaseControl=0)+1.
COXREG S
  /STATUS=CaseControl(1)
  /STRATA=PairNumber
  /CONTRAST (Monitoring)=Indicator(1)
  /METHOD=ENTER Monitoring
  /PRINT=CI(95).

* Using GENLIN: NOT Correct *.
GENLIN CaseControl (REFERENCE=FIRST) BY Monitoring (ORDER=DESCENDING)
  /MODEL Monitoring INTERCEPT=YES
   DISTRIBUTION=BINOMIAL LINK=LOGIT
  /CRITERIA METHOD=FISHER SCALE=1 ANALYSISTYPE=3(LR) CILEVEL=95
LIKELIHOOD=FULL
  /REPEATED SUBJECT=PairNumber SORT=YES CORRTYPE=INDEPENDENT
ADJUSTCORR=YES COVB=ROBUST
  /PRINT CPS DESCRIPTIVES MODELINFO FIT SUMMARY SOLUTION (EXPONENTIATED).

* In a follow-up, Marta concluded that this model cannot be
* computed properly using GENLIN.
* Link: http://www.listserv.uga.edu/cgi-bin/wa?A1=ind1002&L=spssx-l#1 .

* Note that one should be able to obtain the correct result
* using NOMREG, as described by David Garson.
* Link: http://faculty.chass.ncsu.edu/garson/PA765/logistic.htm#matchedconditional .

* Compute difference variables described by Garson .
* Given the way the data are sorted, put the differences
* on the rows where CASECONTROL = 0 .

do if (casecontrol EQ 0).
- compute dmonitor = lag(monitoring) - monitoring .
- compute c = 1 . /* constant outcome .
end if.
exe.

* Run NOMREG with DV = a constant, and X = the case minus matched control
* difference on MONITORING.
* NOTE:  The constant must be excluded.

NOMREG c (BASE=LAST ORDER=ASCENDING) WITH dmonitor
  /INTERCEPT=EXCLUDE
  /PRINT=PARAMETER SUMMARY LRT CPS STEP MFI.

* As expected, this method also gives the correct result .
* OR=4.333 (95%CI: 1.235 to 15.206) .

[Marta Garcia-Granero]

Hi Bruce (and everyone else interested in the recent discussion we had
on matched case control studies).

While looking for material for my classes, I have found this link:

http://www.stata-journal.com/article.html?article=st0070 (dataset can be
download - Stata .dta file, can be imported directly using PASW18 - from
here: http://www.stata-journal.com/software/sj4-3/st0070/)

They describe how to deal with a matched (1:1) cohort study using twins,
one exposed and the other non exposed.

I have found that GENLIN is perfectly able of getting the same results
Stata gives, using two different models:

* This models gives the same result than Stata's command CSMATCH *.
GENLIN cocaine (REFERENCE=LAST) BY exposed (ORDER=DESCENDING)
  /MODEL exposed INTERCEPT=YES
 DISTRIBUTION=BINOMIAL LINK=LOGC
  /CRITERIA METHOD=FISHER(1) SCALE=1 ANALYSISTYPE=3(LR) CILEVEL=95
LIKELIHOOD=FULL
  /REPEATED SUBJECT=id SORT=YES CORRTYPE=INDEPENDENT ADJUSTCORR=YES
COVB=ROBUST
  /PRINT CPS DESCRIPTIVES MODELINFO FIT SUMMARY SOLUTION (EXPONENTIATED).

* This model gives the same result than Stata's conditional Poisson
regression *.
GENLIN cocaine BY exposed (ORDER=DESCENDING)
  /MODEL exposed INTERCEPT=YES
 DISTRIBUTION=POISSON LINK=LOG
  /CRITERIA METHOD=FISHER(1) SCALE=1 ANALYSISTYPE=3(LR) CILEVEL=95
LIKELIHOOD=FULL
  /REPEATED SUBJECT=id SORT=YES CORRTYPE=INDEPENDENT ADJUSTCORR=YES
COVB=MODEL
  /PRINT CPS DESCRIPTIVES MODELINFO SUMMARY SOLUTION (EXPONENTIATED).

See the paper for more details on every approach (CSMATCH vs condition
Poisson regression).

Happy weekend,
Marta GG

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD