Logits vs Marginal

[Greg Maislin]

Hi Stat-l'ers,
SAS's PROC CATMOD allows a "relatively" easy implementation
of a Grizzle, Starmer, Koch (GSK) weighted least squares
analysis of a dichotmous response measured for each subject
at 4 time points. There are three groups to compare so the
hypothesis of interest is group by time interaction. There are
a total of 12 degrees of freedom in the saturated model. I have
even produced a non-saturated model with linear and quadratic
time trends that fits (Goodness-of-fit residual df=3, p>0.80).
Now, I have fit the model to the logits of the marginal
probilities and to the marginal probabilities themselves.
The resulting df=4 group*time(linear+quadratic) is
significant for logits (p=0.01) but not for marginals (p=0.07).
This is what I want to show and examining the marginals it
certainly looks like the time trends are different. I don't
want to base my choice of response function on which results
are better but I can't think of a statistical reason to prefer
one over the other. Any thoughts.
-Greg

**********************************************************************
* Greg Maislin (F...@ASC.UPENN.EDU) *
* Adj. Asst. Prof. of Biostat. in Medicine *
* Univ. of Penn. School of Medicine and *
* Principal, Biomedical Statistical Consulting (215-259-8095) *
**********************************************************************