Date: Mon, 4 Aug 2008 13:38:15 -0500
Reply-To: Warren Schlechte <Warren.Schlechte@TPWD.STATE.TX.US>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Warren Schlechte <Warren.Schlechte@TPWD.STATE.TX.US>
Subject: Re: infinite likelihood problem with Mixed Model
Content-Type: text/plain; charset="us-ascii"
Check Robin's posting of March 11, 2008, where he is discussing
non-estimable means. He has some code that he uses to check out the
design matrix. Using this code may allow you to see if you have a
matrix that is not full-rank.
Warren Schlechte
-----Original Message-----
From: A.B. [mailto:alicia.bingo@GMAIL.COM]
Sent: Monday, August 04, 2008 12:43 PM
Subject: Re: infinite likelihood problem with Mixed Model
Thanks for everybody for the suggests.
Dale,
I have run the two tests, the first one confirms again that there's no
duplicate in the combination of id and time, and the second "repeated /
subject=id type=cs;" still gave me the same "infinite likelihood"
message.
I have to mention that in the model there're some other terms related to
a
second intervention in addition to the main intervention, it's designed
to
strenghten the main intervention after a certain period of time. The
variable is also binary. And there's a 2 way interaction term of this
with
intervention group, and a 3 way with intervention group, and time.
Could it be like Warren said: that with all these binary terms and
interactions in the model, we get something like in the logistic
complete or
quasi-complete separation? And if so, how canI check and solve it?
Thanks again!
A.B.
On Thu, Jul 31, 2008 at 8:24 PM, Dale McLerran
<stringplayer_2@yahoo.com>wrote:
> Alicia,
>
> The usual cause of an infinite likelihood is as Robin indicated in
> his first reply, that when you have a REPEATED statement with
structure
>
> repeated effect / subject=ID ...;
>
> and there is some combination of subject*time which has more than a
> single instance, then you will always get an infinite likelihood
> message. You indicate that there are not any duplicate values of
> time within subjects, but I always have to have this verified. This
> is easily done with the following code:
>
> proc sort data=mydata;
> by id ti;
> run;
>
> data _null_;
> set mydata;
> by id ti;
> if sum(first.ti, last.ti)^=2 then put subject= ti=;
> run;
>
>
> This will identify any instances where a duplicate time value was
> mistakenly coded.
>
> Another way that you could test whether there is a problem with the
> time variable within some subject (although it won't tell you which
> subject by time combination is at fault) would be to remove the time
> variable from your REPEATED statement. Since you are specifying a
> compound symmetric covariance structure, then it is not at all
> necessary to state the time variable here. Try running your model
> with the REPEATED statement:
>
> repeated / subject=id type=cs;
>
>
> After you have run these two diagnostic tests, report back on both
> of them and we can go from there.
>
> Dale
>
> ---------------------------------------
> Dale McLerran
> Fred Hutchinson Cancer Research Center
> mailto: dmclerra@NO_SPAMfhcrc.org
> Ph: (206) 667-2926
> Fax: (206) 667-5977
> ---------------------------------------
>
>
> --- On Thu, 7/31/08, A.B. <alicia.bingo@GMAIL.COM> wrote:
>
> > From: A.B. <alicia.bingo@GMAIL.COM>
> > Subject: Re: infinite likelihood problem with Mixed Model
> > To: SAS-L@LISTSERV.UGA.EDU
> > Date: Thursday, July 31, 2008, 1:02 PM
> > Robin,
> >
> > Thanks for your response.
> >
> > First, combinations of id and ti are unique, no duplicate.
> >
> > Second, the outcome variable is in the range of -15 to 15,
> > so I don't think
> > it needs to be scaled down.
> >
> > Third, unfortunately, the time points are not equally
> > spaced, some are 1
> > year, some are 2 years. So ar(1) may not be a good method.
> > But I tried it
> > anyway, the model still ran into the same problem.
> >
> > I just don't understand why it worked with the old data
> > with 4 times points
> > and when I tried to reduce it to three points, it did not
> > work. It's true
> > that at later time there are less data. Can this cause the
> > problem?
> >
> > What else can I check or do to modify the model?
> >
> > Thanks again!
> > A.B.
> >
> >
> >
> > On Thu, Jul 31, 2008 at 2:27 PM, Robin R High
> > <rhigh@unmc.edu> wrote:
> >
> > >
> > > A.B.
> > >
> > > First, make sure combinations of id and ti are unique.
> > You can easily check
> > > these counts with
> > >
> > > PROC FREQ;
> > > TABLE id * ti / norow nocol nopercent;
> > > run;
> > >
> > > to make sure no cell has a value of 2 or more. And
> > also that values of id
> > > are unique across all subjects, regardless of between
> > subject factors -- if
> > > id values repeat across two or more grouping factors,
> > you would need to
> > > enter a nested specification, ie., subject=id(group).
> > >
> > > If that doesn't work, you could try rescaling the
> > response by multiplying
> > > it by a constant, say .1 (if numbers lare large) and
> > analyze in those
> > > linearly transformed units.
> > >
> > > Also, with data collected over time on each subject
> > (esp. 3 or more time
> > > points), I'd be wary of entering the type=cs as
> > your covariance choice.
> > > There are several options that likely will work better
> > with data collected
> > > over time (ar(1) for equally spaced points, among
> > other structures). also,
> > > enter r and rcorr to see what they contain.
> > >
> > > Robin High
> > > UNMC
> > >
> > >
> > >
> > >
> > >
> > > *"A.B." <alicia.bingo@GMAIL.COM>*
> > > Sent by: "SAS(r) Discussion"
> > <SAS-L@LISTSERV.UGA.EDU>
> > >
> > > 07/31/2008 01:15 PM Please respond to
> > > "A.B." <alicia.bingo@GMAIL.COM>
> > >
> > > To
> > > SAS-L@LISTSERV.UGA.EDU cc
> > > Subject
> > > infinite likelihood problem with Mixed Model
> > >
> > >
> > >
> > >
> > > I am using mixed model to do a repeated measure
> > analysis. The model has
> > > terms such as time points, intervention group, site, a
> > binary health
> > > condition indicator, some interaction terms including
> > the three way
> > > interaction of time*intervention *health condition.
> > >
> > > I inhereted the model from a previous person. The
> > model worked with the old
> > > data with 4 time points. Now we have added two more
> > time points and the
> > > model won't run (infinite likelihood). I am using
> > statement "repeated
> > > ti/subject=id type=cs;"
> > >
> > > I don't want to take out terms. I tried to reduce
> > the time points to
> > > include
> > > just the first and the last two new ones, but still
> > got the same problem.
> > >
> > > Could someone help me with this one?
> > >
> > > Thanks in advance!
> > > --
> > > A.B.
> > >
> > >
> >
> >
> > --
> > A.B.
>
--
A.B.
| 
