Logistic, Genmod, and Repeated Measures

Discussion:

(too old to reply)

Bruce A.

2008-09-18 17:10:05 UTC

Hi all,

I'm trying to analyze a dataset with repeated observations on the same
subject with a dependent variable which is dichotomous. So I used PROC
GENMOD with the repeated statement. However, when I compare that to the
output when I use PROC LOGISTIC (which ignores dependency) I get the same
estimates, the same confidence intervals of those estimates, and therefore
I get the same p-hat values.

Shouldn't I get different answers? Am I using PROC GENMOD wrong? Am I
going crazy?

These are the two codes I used:

proc logistic data = drunks descending;
title "Using Logistic Regression ";
model accident = age vision drivers_ed / stb aggregate scale=none
link=logit;
output out = logistic_fit predprobs=(i);
run;

proc genmod data = drunks descending;
title "Using Genmod type = ind";
class case period;
model accident = age vision drivers_ed / dist=bin link = logit;
repeated subject = case / within=period type =unstr;
output out = genmod_fit_i p=phat;
run;

the dataset looked something like:
case accident age vision drivers_ed period
1 1 17 1 1 1
2 0 44 0 0 1
3 0 35 0 1 1
.
45 0 20 0 1 1
1 1 18 1 1 2
2 1 45 0 0 2
3 0 36 0 1 2
.
45 1 21 0 1 2

Thanks to anyone who has an idea!!!
Cheers,
Bruce A.

Peter Flom

2008-09-18 17:28:48 UTC

Permalink

Post by Bruce A.
I'm trying to analyze a dataset with repeated observations on the same
subject with a dependent variable which is dichotomous. So I used PROC
GENMOD with the repeated statement. However, when I compare that to the
output when I use PROC LOGISTIC (which ignores dependency) I get the same
estimates, the same confidence intervals of those estimates, and therefore
I get the same p-hat values.
Shouldn't I get different answers? Am I using PROC GENMOD wrong? Am I
going crazy?

Not sure about your genmod code, but I think you should be using PROC GLIMMIX

Peter

Peter L. Flom, PhD
Statistical Consultant
www DOT peterflom DOT com

Robin R High

2008-09-18 18:10:56 UTC

Permalink

Bruce,

Whether GENMOD or GLIMMIX is more appropriate here is not obvious, though
from GENMOD be sure you are looking at the table:

Analysis Of GEE Parameter Estimates
Empirical Standard Error Estimates

for the GEE analysis.

GENMOD also prints the Analysis Of Initial Parameter Estimates table,
which at the first iteration is the same as LOGISTIC.

Also, add the corrw option to the REPEATED statement in GENMOD, just to
see what the within case correlation of period is; if its close to 0, the
results would likely be nearly the same.

Robin High
UNMC

"Bruce A." <***@YAHOO.COM>
Sent by: "SAS(r) Discussion" <SAS-***@LISTSERV.UGA.EDU>
09/18/2008 12:18 PM
Please respond to
"Bruce A." <***@YAHOO.COM>

To
SAS-***@LISTSERV.UGA.EDU
cc

Subject
Logistic, Genmod, and Repeated Measures

Hi all,

I'm trying to analyze a dataset with repeated observations on the same
subject with a dependent variable which is dichotomous. So I used PROC
GENMOD with the repeated statement. However, when I compare that to the
output when I use PROC LOGISTIC (which ignores dependency) I get the same
estimates, the same confidence intervals of those estimates, and therefore
I get the same p-hat values.

Shouldn't I get different answers? Am I using PROC GENMOD wrong? Am I
going crazy?

These are the two codes I used:

proc logistic data = drunks descending;
title "Using Logistic Regression ";
model accident = age vision drivers_ed / stb aggregate scale=none
link=logit;
output out = logistic_fit predprobs=(i);
run;

proc genmod data = drunks descending;
title "Using Genmod type = ind";
class case period;
model accident = age vision drivers_ed / dist=bin link = logit;
repeated subject = case / within=period type =unstr;
output out = genmod_fit_i p=phat;
run;

the dataset looked something like:
case accident age vision drivers_ed period
1 1 17 1 1 1
2 0 44 0 0 1
3 0 35 0 1 1
.
45 0 20 0 1 1
1 1 18 1 1 2
2 1 45 0 0 2
3 0 36 0 1 2
.
45 1 21 0 1 2

Thanks to anyone who has an idea!!!
Cheers,
Bruce A.

Wensui Liu

2008-09-19 03:25:46 UTC

Permalink

peter,

i have to disagree with you and see no point to use glimmix. where is the
random effectin bruce's model?

On Thu, Sep 18, 2008 at 1:28 PM, Peter Flom <