Mixed models using restricted cubic splines

[AGabriela]

Hi there - The following code works but only with a very small sample size.  My dataset is big and I get an ERROR message when I submit the code using the entire dataset.  I have never used GLIMMIX, I'm not familiar with it.  I wonder if this model can be run using MIXED instead?  Would that help? Or are there any other ways to run this code more efficiently?

proc glimmix data=xx;

class studyid;

 effect spl = spline(quarter / details naturalcubic basis=tpf(noint)knotmethod=percentiles(5));

   model avgdose = spl /s ddfm=kr;

   random int spl /subject = studyid type=vc solution;  /*try vc instead of un*/

      ods output parameterestimates=sf(keep=effect estimate 

                                 rename=(estimate=overall));

      ods output solutionr=sr(keep=effect subject estimate

                                 rename=(estimate=ssdev));

run;

Here is the error I get:

ERROR: Model is too large to be fit by PROC GLIMMIX in a

       reasonable amount of time on this system. Consider

       changing your model.

[SteveDenham]

Three questions: How much memory is available to you,  how many quarters are you trying to fit and how many cases do you have (studyid's)?  I suspect that your code is generating a lot of knots, and so it ends up with a too many parameters in the optimization compared to the available memory.

SteveDenham

[AGabriela]

Thank you for your reply.  Here are the answers:

1) On a regular basis just 32gb, but the IT department doubled that during the weekend.  Unfortunately I didn't see a big improvement. 

2) Nine quarters.  I've tried 3, 4 or 5 knots.  As expected, the smaller the number of knots the greater the chance the model will run.

3) If the dataset has 1k people (studyids) or less it runs fine.  But unfortunately I have many more people (35k+) on the entire dataset.

Thanks again,

Gabriela 

[Rick_SAS]

Here's a suggestion: First fit a simple model that uses the main effects of QUARTER.

model avgdose = quarter /s ddfm=kr;

If the simple model converges and gives reasonable answers that you understand, then try the more complicated model that incorporates spline effects for QUARTER.

[AGabriela]

Thanks for your reply.  That's always a good approach.  I have done that, though.

Thanks again.