Hi all, thanks for the responses.
It looks like zero frequencies issue is not a big problem, but
quasi-complete separation is which apparently can be detected from low
likelihood ratios of each effect that contributes to the model. Yet
when I take out low categorical IVs, I still got the quasi-complete
separation warning.
Any more ideas for causes and how to remedy this?
thank you
Beyhan T. Maybach, PhD
On 7/29/07, Kathryn Gardner <
[hidden email]> wrote:
> Hi,
>
> I don't know about multinominal reg as I haven't used it properly, but in
> ordinal reg a similar message about zero frequencies comes up because of the
> continuous predictors in the model. The continuous predictor/s produces a
> large number of cells (i.e., dependent variable levels by combinations of
> predictor variables) with zero frequencies. I've read that this is OK and
> normal, but large amounts of empty cells make model fit statistics
> unreliable, and they shouldn't therefore be used. I assume the same applies
> to multinominal reg.
>
> Kathryn
>
> > Date: Fri, 27 Jul 2007 13:33:57 -0400
> > From:
[hidden email]
> > Subject: Multinomial Logistic Regression
> > To:
[hidden email]
> >
> > Hi,
> > I am analyzing a dataset with 4 continuous and 14
> > categorical Ivs with 7 group levels as the
> > DV using MLR.
> >
> > Basically, multiple characteristics (continuous and categorical IVs) of
> > 7 tree Genera groups were recorded (taxonomically decided but
> > debatable Generas) to determine the set of characteristics that allows
> > for the best discrimination/classification between these Genera and if
> > the cases really fit in these Genera groups.
> >
> > In MLR, I am getting two errors; '85% of the cells have zero
> > frequency' and 'there is possibly a quasi-complete
> > separation in the data'. What could I do about 85% zero freguency and
> > and the quasi-complete seperation? There are no zeros in the cells so that
> > it doesnot result in odds ratio of infinity. What are the other reasons
> > for quasi-complete seperation?
> >
> > I would appreciate any suggestions to get this run going.
> > Sincerely
> >
> >
> > Beyhan Titiz Maybach, PhD
>
>
> ________________________________
> The future of MSN Messenger! Windows Live Messenger
Locked
