topic Re: PROC GLIMMIX with R side in Statistical Procedures

PROC GLIMMIX with R side

adcc5465 — Wed, 07 Dec 2022 18:18:13 GMT

I am modeling a YES/NO decision making at the individual level. One person would choose one out of around 2,000 options. There were 25,000 persons. The Cartesian product gave a 40,000,000 sample size (at different time, some options may not be available).

For all options, there were two interested main effects and one of their interaction: A and B. Both are three-level ordinal variables (Per my understanding, because the options were sampled from a population, therefore, possible random effects for A and B).

For some persons (10%), they have two chances (at different time, count variable with levels of 1 and 2) to make their decision (repeated measure, therefore, needs R side effects).

The program was:

proc sort data= test; by chance Person_ID;

proc glimmix data=test singular=1e-10 maxopt=100 noclprint;

class A B Options_ID;

model decision(event="1") = A|B Chance

/ dist=binary link=logit ddfm=BW;

random intercept/subject = A|B type =VS;

random chance/subject = Person_ID residual type =AR(1);

nloptions maxiter=150 technique= NRRIDG;

run;

When I tried to run this, SAS kept running like forever.

After many tries, the only converge result I got was from this program:

proc glimmix data=test singular=1e-10 maxopt=100 method = Laplace noclprint;

class A B Options_ID;

model decision(event="1") = A|B Chance

/ dist=binary link=logit ddfm=BW;

random A /subject = Options_ID type =ANTE(1);

nloptions maxiter=150 GCONV=0;

run;

From the previous discussion, @sld at https://communities.sas.com/t5/Statistical-Procedures/Glimmix-did-not-converge/td-p/387883

would the comment apply to my situation?

And how should I adjust my model?

Re: PROC GLIMMIX with R side

jiltao — Thu, 08 Dec 2022 17:56:38 GMT

Your program below --

proc glimmix data=test singular=1e-10 maxopt=100 noclprint;

class A B Options_ID;

model decision(event="1") = A|B Chance

/ dist=binary link=logit ddfm=BW;

random intercept/subject = A|B type =VS;

random chance/subject = Person_ID residual type =AR(1);

nloptions maxiter=150 technique= NRRIDG;

run;

has two issues:

1. the first random statement does not seem to have a correct syntax

2. The current syntax treats A|B as both fixed and random effects -- this is not correct.

I would suggest you take out your first RANDOM statement. And for the second RANDOM statement, I am not sure AR(1) makes sense to your data. But this might be because I am not clear on your data structure at this point....

Thanks,

Jill

Re: PROC GLIMMIX with R side

adcc5465 — Thu, 08 Dec 2022 20:27:01 GMT

Thank you so much!

In that case, could I switch to PROC GENMOD rather than stick with PROC GLIMMIX?

Re: PROC GLIMMIX with R side

jiltao — Thu, 08 Dec 2022 22:31:53 GMT

Yes I think PROC GENMOD should work too.