Date: Fri, 17 Apr 1998 14:31:50 -0600
Reply-To: mgonen@ALOGIC.COM
Sender: "SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>
From: mgonen@ALOGIC.COM
Organization: Deja News - The Leader in Internet Discussion
Subject: Re: Proc Catmod
In article <35375000.D27015F7@digital.com>,
Randy Collica <randall.collica@digital.com> wrote:
>
>
> --------------2F2B9734153A482328FF4F7E
> Content-Type: text/plain; charset=us-ascii
> Content-Transfer-Encoding: 7bit
>
> Hello Michele,
>
> Well, in Proc Catmod like in Logistic, or Genmod when one has only 48
> observations but 13 covariates with 2-5 levels, it appears obvious to me
> that your client just does not have enough data to model this probem as it
> currently stands. My suggestion if it is worth anything is to have your
client
> do the following:
>
> 1) See if they can first restrict some of their variables to 2-4 most
important
> variables of covariates to see if some decent model can be fit.
>
> 2) Then, possibly add other variables 1 at a time and see when the model
> completely drops out.
>
> For all it's worth, Randy
>
> Michele Marini wrote:
>
> > Question two - DING!
> >
> > This question originates from a client who is insisting on doing the WRONG
> > analysis given his experimental data, but he is trying to reproduce
results
> > from SAS. Having issued that disclaimer - here goes.
> >
> > In his study he has 43 observations, with 13 independent variables (with 2
to
> > 5 levels depending on the variable). (I cringe as I write this.) Anyway,
he
> > is attempting to model his response using Catmod and gets the following
> > messages:
> >
> > NOTE: Effects marked with '*' contain one or more redundant or
> > restricted parameters.
> > NOTE: Parameters maked with '#' are regarded to be infinite.
> >
> > His complaint/concern is and I quote:
> > 'Essentially, I want to use the parameter estimates to compute the
> > Predicted function for a given level of independent variables used in
> > the 'model' statement. HOWEVER, I have not been able to successfully
> > compute the correct value for the function (as listed in the output).
> > I think that the 'NOTES' may be related to why I am unable to
correctly
> > compute the function values.'
> >
> > I guess our question is - How do these messages (NOTES) affect calculation
of
> > function values?
> >
> > Again, any help would be most appreciated.
> >
> > *************************************************************************
> > Michele Marini
> > Statistics Department
> > Virginia Tech
> > (540) 231-8315
> > email:marini@vtvm1.cc.vt.edu
> > *************************************************************************
>
> --------------2F2B9734153A482328FF4F7E
> Content-Type: text/html; charset=us-ascii
> Content-Transfer-Encoding: 7bit
>
> <HTML>
> Hello Michele,
>
> <P>Well, in Proc Catmod like in Logistic, or Genmod when one has only 48
> <BR>observations but 13 covariates with 2-5 levels, it appears obvious
> to me
> <BR>that your client just does not have enough data to model this probem
> as it
> <BR>currently stands. My suggestion if it is worth anything is to
> have your client
> <BR>do the following:
>
> <P>1) See if they can first restrict some of their variables to 2-4
<U>most</U>
> important
> <BR>variables of covariates to see if some decent model can be fit.
>
> <P>2) Then, possibly add other variables 1 at a time and see when the model
> <BR>completely drops out.
>
> <P>For all it's worth, Randy
>
> <P>Michele Marini wrote:
> <BLOCKQUOTE TYPE=CITE>Question two - DING!
>
> <P>This question originates from a client who is insisting on doing the
> WRONG
> <BR>analysis given his experimental data, but he is trying to reproduce
> results
> <BR>from SAS. Having issued that disclaimer - here goes.
>
> <P>In his study he has 43 observations, with 13 independent variables (with
> 2 to
> <BR>5 levels depending on the variable). (I cringe as I write
this.)
> Anyway, he
> <BR>is attempting to model his response using Catmod and gets the following
> <BR>messages:
>
> <P> NOTE: Effects marked with '*' contain one or more
> redundant or
> <BR> restricted
parameters.
> <BR> NOTE: Parameters maked with '#' are regarded to
> be infinite.
>
> <P>His complaint/concern is and I quote:
> <BR> 'Essentially, I want to use the parameter estimates to
> compute the
> <BR> Predicted function for a given level of independent
> variables used in
> <BR> the 'model' statement. HOWEVER, I have not
> been able to successfully
> <BR> compute the correct value for the function (as listed
> in the output).
> <BR> I think that the 'NOTES' may be related to why I
> am unable to correctly
> <BR> compute the function values.'
>
> <P>I guess our question is - How do these messages (NOTES) affect
calculation
> of
> <BR>function values?
>
> <P>Again, any help would be most appreciated.
>
> <P>*************************************************************************
>
<BR> &n
bsp; &n
bsp;
> Michele Marini
>
<BR> &n
bsp; &n
bsp;
> Statistics Department
>
<BR> &n
bsp; &n
bsp;
> Virginia Tech
>
<BR> &n
bsp; &n
bsp;
> (540) 231-8315
>
<BR> &n
bsp; &n
bsp;
> email:marini@vtvm1.cc.vt.edu
>
<BR>*************************************************************************<
/BLOCKQUOTE>
> </HTML>
>
> --------------2F2B9734153A482328FF4F7E--
>
>
As Michele and Randy point out this is not the best thing that can be done
with this data set. But even if using Randy's "stepwise" method the distrubing
message in the output disappears, the analysis is likely to remain incorrect.
With 48 observations and say 8 variables with, say, 2 categories each, we have
about 3 observations per cell, which is not enough to be comfortable with the
asymptotic methods that CATMOD or LOGISTIC uses. An improvement might be in
checking other packages like StatXact or LogXact which does exact analysis.
Mithat Gonen.
mgonen@alogic.com
-----== Posted via Deja News, The Leader in Internet Discussion ==-----
http://www.dejanews.com/ Now offering spam-free web-based newsreading
| 
