Stefan

The GEE, population averaging, methodology, introduced by Liang and Zeger (1986), provides a method of analyzing correlated data, which is an alternative to the techniques of generalized linear model (glm), it is  used for repeated measures or clustered data. This method takes into account the working correlation matrix for the observations at different time point.  In repeated measure the measurement taken at different time point on the same subjects are correlated the strength of correlation is stronger between the measurement of closer time points. Thus while carrying out any regression analysis this correlation structure of the within subjects measurement needs to be considered. 
Also modeling with GEE gives the facility of modeling the scores with robust error method.  Models can be fitted with many correlation structures e.g autoregressive, exchangeable, independent, m-dependent, and unstructured. It is also possible to input own correlation structure. with some practical problems.
 I have seen that both of the methods produce results with negligible differences, but  again it basically depends on the nature of your data. I would fit the models with both and see the results especially the error structure. 
 
Stata gee command line for your analysis
 
xtgee (dependent variable)  (independent variable)  , fam(binomial) link(logit) i(identification number) t(time variable)
 
HTH
 
Muhammad


--- On Fri, 8/29/08, Stefan Uhmann <[log in to unmask]> wrote:

From: Stefan Uhmann <[log in to unmask]>
Subject: Re: right test for the following design
To: [log in to unmask]
Date: Friday, August 29, 2008, 1:04 PM

Simon,

thank you for your immediate help. Am I right in assuming that if my t1 
and t2 observations are dependent, I can not use this glm approach? This 
would mean I would have to make myself familiar with gee, which Paul 
also pointed me to (thank you, Paul!). I have individual data, btw.

Can you please point me to some literature in addition to the one 
mentioned in the gee package's help files?

Muhammad, thanks for your suggestion. What are the differences between 
gee and ML log. regression? Would both yield the same results?

I guess, I got a long way to go ..

Kind regards,
Stefan

simon bond schrieb:
> Hi,
> Â 
> I would recomend using logistic regression, however you need to knwo how
many individuals were in each cell . The follwoing R code illustrates this
> Â 
> library(MASS)
> uhman=data.frame( expand.grid(group=factor(1:3),
time=factor(1:2)),rate=c(.7, .4, .3, .8, .6, .2))
> mod10=glm(rate~group*time, data=uhman, family=binomial,weights=rep(10,6))
> mod1000=glm(rate~group*time, data=uhman,
family=binomial,weights=rep(1000,6))
> 
> summary(mod10)
> summary(mod1000)
> 
> Where the different weight argument represent different sample sizes of 10
and 1000 per cell.
> Â 
> You should check the assumption that the two observations from each
subject are independent. This is only possible if you have individual subject
level data.  Maybe use GEE 
> Â 
> library(gee)
> gee( response~group*time, id=subject, family=binomial,
data=individualdata, corstr="exchangeable")
> Â 
> Â 
> Also, you could consider using a mixed effects model that captures the
within-patient correlation if you had the response on an individual subject
level.
> Â 
> library(lme4)
> lmer( response~group*time +(1|subject), data=individualdata,
family=binomial)
> Â 
> Hope this helps
> 
> Simon BOnd.
> 
> 
> 
> ----- Original Message ----
> From: Stefan Uhmann <[log in to unmask]>
> To: [log in to unmask]
> Sent: Thursday, 28 August, 2008 15:45:38
> Subject: right test for the following design
> 
> Hi,
> 
> I have got the following design:
> 3 groups of individuals and for each individual a binary outcome at two 
> points in time: pre and post. This gives my the following table:
> 
> Â  Â  Â  Â  G1Â  Â  G2Â  Â  G3
> t1-yes    70%  40%  30%
> t2-yes    80%  60%  20%
> 
> If I could use GLM, I would have an 3 x 2(group x pre/post)-GLM, with 
> the second factor being a within factor and the possibility to 
> differentiate between main and interaction effects.
> Since my outcome is binary (yes-no) and I just want to use the 
> yes-portions as outcome, I have no idea which test to use. Is it 
> possible to use a chi²-test, although one of the factors is a within
factor?
> 
> Thanks for any help in advance,
> Stefan
> 
> 
> Send instant messages to your online friends http://uk.messenger.yahoo.com