Date:         Mon, 12 Jan 1998 08:59:33 +0100
Reply-To:     Hans-Peter Piepho <piepho@WIZ.UNI-KASSEL.DE>
Sender:       "SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>
From:         Hans-Peter Piepho <piepho@WIZ.UNI-KASSEL.DE>
Subject:      Re: PROC MIXED
Comments: To: dsachs@netcarrier.com
Content-Type: text/plain; charset="us-ascii"

>Yesterday, I posted a query to this list about PROC MIXED.  I thing some
>of you responded, but because some quirk in the e-mail at work, I only
>received one reply.  If you respond,  please use this address at home.
>  The query is this:
>We are using PROC MIXED to analyze a mixed model using release 6.09 on
>an MVS mainframe.  We have nine cases with about 5000 to 10000 subjects
>per case.  The model has repeated measure over time for each subject,
>one fixed effect (time) and 18 random effects.  The questions are as
>follows:
>   1.  The model converges faster if the random variables are treated as
>fixed, generally in only one or two iterations.  What are the risks of
>doing that?  How does that affect the interpretation of the coeficients
>and the significance?
>  2.  There are high correlations among the random predictors (r's
>between .5 and .85).  Does this multicollinearity affect the stablity of
>the predictors the same way it does in multiple regression?   One of our
>managers wants to ignore the intercorrelations and just graph the
>response surface. Does that make sense?
>I'm sorry to have to ask the same questions again.   I think PROC MIXED
>offers some real potential for us.  Thanks again for your help.
>Don Sachs
>
>
 
As a simple example, take a two-way factorial with factors A (fixed) and B
(random). Suppose you have several observations per A*B combination. A
linear model for the data is
 
Y = Mean + A + B + A*B + Error
 
Since the factor B is random, it is sensible to treat B and A*B as random.
In a balanced layout, A will then be tested against the A*B mean square
(MS). This corresponds to a broad inference space: Differences detected
between levels of A are present in the average of a polulation of B's (of
which a random sample has been observed). If A*B is taken as fixed, then A
is tested against the Error MS. This corresponds to a narrow inference
space: Differences between levels of A are present in the averages across
OBSERVED levels of B.
 
Hans-Peter
_______________________________________________________________________
Hans-Peter Piepho
Institut f. Nutzpflanzenkunde  WWW:   http://www.wiz.uni-kassel.de/fts/
Universitaet Kassel            Mail:  piepho@wiz.uni-kassel.de
Steinstrasse 19                Fax:   +49 5542 98 1230
37213 Witzenhausen, Germany    Phone: +49 5542 98 1248