Proc GLIMMIX - insufficient memory

[Mbo]

Hi all,

I've never posted here before, but could really use some help.  I am using an 'events/trials' syntax within GLIMMIX to model a certain type of dental procedure out of that procedure plus another kind of dental procedure.  I have ~55,000 records for the analysis on which I'm currently working.  After I tackle this analysis, I will need to do another one on ~87,000.  There are multiple records per patients and nested random effects.  The nested random effects are ID within provider within facility within region.  I looked online for help which prompted me to sort by ID, and change text variables to numeric where possible.  For mostly everything I've tried, I've gotten the error that "SAS stopped processing  this step because of insufficient memory."

I finally got this to run, but this doesn't include all of my random effects terms.

proc glimmix data=test noclprint;

class facility race prserv_cat aa bb;

model x/y = race age gender aa bb cc dd ee ff gg hh ii jj kk ll mm nn oo/ddfm=bw dist=bin link=logit or cl;

random int / subject=ID;

nloptions tech=nrridg;

run;

I tried the following but got the warning about insufficient memory:

proc glimmix data=test coclprint;

class facility race prserv_cat aa bb;

model x/y=race age gender aa bb cc dd ee ff gg hh ii jj kk ll mm nn oo/ddfm=bw dist=bin link=logit or cl;

random int / subject=ID;

random int / subject=Provider;

random int / subject=facility;

random int / subject=region;

run;

I also tried the same code as the prior procedure, but using the following random statments (this did not work either):

random int / subject=ID;

random int facility region*facility/subject=Provider;

If you need more information, just let me know.

Any help would be greatly appreciated.

Michelle

[SteveDenham]

How many regions are involved?  Is it possible to consider it as a fixed effect? If not, I think the only way might be to look at separate analyses by region, and comparing interval estimates for the parameters.

Steve Denham

[Mbo]

Thanks so much for your response, Steve.  Unfortunately, there are about 21 regions, about 200 facilities, and over 2000 providers.  I'm being asked to run the regression and account for the nesting of patients within providers within facilities within regions. Because of that, I don't think I can consider region as a fixed effect.  Do you think this is possible?  If not, I'll have to tell my team that it is too much for SAS to handle.  What did you mean by looking at separate analyses by region, and comparing interval estimates for the parameters?  Can you be more specific?

Michelle 

[SteveDenham]

Michelle,

If memory is insurmountable, and to me it appears it may be, given this much structure to the Z matrix, then the best you will be able to do is 21 separate regressions--one for each region, analyzed separately.  The interval estimates for the beta's will be more informative than the point estimates--you may find ways to lump regions together based on similarities, and thus perhaps render the problem tractable.

Steve Denham

[Mbo]

Thanks so much, Steve.