Date:         Mon, 19 Jul 2010 14:17:55 -0400
Reply-To:     Sigurd Hermansen <HERMANS1@WESTAT.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Sigurd Hermansen <HERMANS1@WESTAT.COM>
Subject:      Re: GLIMMIX question
Comments: To: Andrew Agrimson <jagrimsasl@GMAIL.COM>
In-Reply-To:  <AANLkTin-w-2eh5KGBIO1AqUw1m-cAKKTa4HSC2hW0dVw@mail.gmail.com>
Content-Type: text/plain; charset="us-ascii"

Andy:
If you've already heard my advice or similar advice on summarizing data prior to implementing a regression analysis, please feel free to ignore what follows....

Many numeric and nominal variables have meaningless differences that one can round or group with little to no loss of information. After rounding or grouping variables, a summary of the dataset containing the "dependent" and "independent" variables replaces many observations with fewer observations and a count that one can use as a freq or weight in a statistical procedure. Regressions not only require less memory, they also run faster. Try a random sample of sample data vs. a summarized sample, weighted by counts, as input and see if the results differ.
S

-----Original Message-----
From: SAS(r) Discussion [mailto:SAS-L@LISTSERV.UGA.EDU] On Behalf Of Andrew Agrimson
Sent: Monday, July 19, 2010 1:40 PM
To: SAS-L@LISTSERV.UGA.EDU
Subject: GLIMMIX question

Hello all,

I recently ran into an issue with PROC GLIMMIX where I'm recieving the
following error:



Integer overflow on computing amount of memory required.
NOTE: The SAS System stopped processing this step because of insufficient
memory

I coded up the same model (gamma regression with a random intercept) in
NLMIXED and did not receive this error although it took 39 hours to
converge. I called SAS and there doesn't seem to be a way to adjust GLIMMIX
to avoid the error above(without reducing the size of my data set). I was
really hoping to use GLIMMIX on my full data set because of ease of use and
ease of diagnostics and faster convergence. One thought I had use to adjust
the convergence criteria in NLMIXED so it doesn't run as long, and then use
GLIMMIX on smaller sample to examine residuals. Does anybody have any
thoughts or concerns with this approach? Changing the convergence criteria
is a little scary but if the parameter estimates are nearly identical
I think I would feel okay about it.

Does anybody have any code where they've used the predict statement to
output residuals in NLMIXED (marginal, conditional, pearson,etc.)?

Thanks,

Andy