🚀 go-pugleaf

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