Guys

I ran into this problem and couldnt figure out why, I am using the following
code for one of my analysis where the response y is binary 0/1 and were
repeated measure over time for each subject.

proc glimmix data=a noclprint;
class id;
model y (descending) = tim / solution link=logit dist=binary;
random intercept time/ subject=id;
run;

The code runs fine, but if I change dist=binary to dist=binomial, the run
now could not converge.

My understanding is that by looking at the log likelihood contribution from
these two distribution, they ought to be the same, but why it converges for
dist=binary, but not for dist=binomial?

Just as an experiment, I tried method=laplace instead of default RSPL, both
distributions converged and gave the identical result.

What is going on here? Thanks for any input..

Lei

It is almost certainly related to a couple of other threads
which have been posted to SAS-L recently - namely that slight
differences in operations (which analytically are identical)
result in floating point truncations that affect sets of
calculations differently.  The problem of differences in
exact representations of floating point operations is described
well in a post that appeared today on SAS-L which can be found at
http://listserv.cc.uga.edu/cgi-bin/wa?A2=ind1003A&L=sas-l&P=R20669

Of course, the question might be raised as to whether the
model which did fit successfully is actually a good estimate.
If the model is so sensitive to small differences in the
mechanics of floating point operations, then even the model
which did fit the data may not be a good solution.

Dale

---------------------------------------
Dale McLerran
Fred Hutchinson Cancer Research Center
mailto: dmclerra@NO_SPAMfhcrc.org
Ph:  (206) 667-2926
Fax: (206) 667-5977
---------------------------------------


--- On Wed, 3/3/10, Lei <kiki886@GMAIL.COM> wrote:

> From: Lei <kiki886@GMAIL.COM>
> Subject: PROC GLIMMIX; dist=binary vs dist=binomial
> To: SAS-L@LISTSERV.UGA.EDU
> Date: Wednesday, March 3, 2010, 12:32 PM
> Guys
>
> I ran into this problem and couldnt figure out why, I am
> using the following
> code for one of my analysis where the response y is binary
> 0/1 and were
> repeated measure over time for each subject.
>
> proc glimmix data=a noclprint;
> class id;
> model y (descending) = tim / solution link=logit
> dist=binary;
> random intercept time/ subject=id;
> run;
>
> The code runs fine, but if I change dist=binary to
> dist=binomial, the run
> now could not converge.
>
> My understanding is that by looking at the log likelihood
> contribution from
> these two distribution, they ought to be the same, but why
> it converges for
> dist=binary, but not for dist=binomial?
>
> Just as an experiment, I tried method=laplace instead of
> default RSPL, both
> distributions converged and gave the identical result.
>
> What is going on here? Thanks for any input..
>
> Lei
>