🚀 go-pugleaf

   Needed:  Reassurance (?) that my advice to a student is OK.
            (But shoot me down if I need it!)

   Problem: Logistic regression (PROC LOGIST in SAS 6.04) doesn't
            converge in 25 iterations when some variables in a
            34 observation dataset are included in the model.

   Advice about to be given:
            Dear hard-working, under-the-gun (thesis defense in
            48 hours) medical and graduate student:
            You and I have run several logistic regressions on your
            dataset and found interesting results for most of the
            variables crucial to your hypothesis.  However, there are
            a few other interesting variables for which PROC LOGIST
            doesn't converge using the default convergence and
            iterations criteria.  When we vary those criteria widely
            we can get convergence, but the parameter estimates vary
            ENORNIMOUSLY (spelling courtesy of my favorite theology
            student), depending on the non-default criteria invoked.

            It seems to me that the particular configuration of
            data in your sample leads to solutions which bounce
            around among multiple local maxima, especially depending
            upon whether or not some observations are deleted from
            the analysis as a consequence of missing data problems.
            Since your sample size of 38 observations is (at my most
            generous) but modest for a six-to-eight variable logistic
            regression, MY ADVICE IS:
                Take the money and run!  Report to your committee those
                results for which (presumptively) the parameter
                estimates are stable (convergence under default
                conditons), tell them that several other variables
                are in fact interesting but that the data in this
                particular sample lead to equivocal results, and that,
                considering the size of this sample, attempts to
                wrest more earth-shaking conclusions from the data
                are unjustifed.  Urge that the research be continued,
                etc., and emphasize what a wonderfule contribution
                the results you do have make to the Progress of Science.


    Phil's question to you gurus:
    Does this advice seem reaonable to you?  (More to the point, and
    this is why I consult you, should it appease her committee members?)
    Are there any other (better, if you please) arguments that I should
    suggest she advance?

    I have gone to lots of effort on this student's work - it's
    interesting (diet and benign breast disease), competently
    carried out, she is one of those rare students who takes up an
    idea or technique and executes it well while learning enough
    about it to go far beyond the limited advice given, and one gets
    the (unusual) feeling that she will put all of the lessons
    learned in this grueling stint into good use as a physician and
    nutritionist.  If you have some better advice for her, please
    send it to me for her, not to make me look better (but I'll take
    it, I'll take it!!), but because she will make productive use
    of it, if not in this small piece of research, then in her future
    career.

                                       Phil Gallagher
                                       uphilg@uncmvs.oit.unc.edu