Difficulties (R-side rand. eff. error) in glimmix for logistic regression for crossover design

[AndreyMyslivets]

Hi to all!

I've faced with some strange formulation in some Clinical Study Protocol: "mixed ordered logistic regression" and I've tried to perform it using proc glimmix.

Could you please help me with resolving difficulties in applying this?

 
My study is a crossover study, with 2 periods and 2 drugs. The study protocol says that I should include sequence, period and treatment as fixed effects and a repeated effect for patients within sequences to model and using a compound symmetry covariance structure.
My endpoint ‘Ocular Discomfort’ is a scale and has following scores:
0 = no discomfort
1 = mild discomfort
2 = moderate discomfort
3 = severe discomfort
4 = very severe discomfort
Also my endpoint was evaluated several times for visit the following time points: 0, 5, 10, 20 min.
 
I read that in the proc glimmix, I can specify a repeated structure by adding the _RESIDUAL_ keyword to the RANDOM statement. I wanted to use the following code:
 
proc glimmix data=t1 ;
   class SUBJID sequence period drug timepoint;
   model AVALC = sequence period drug/ CL dist=multinomial link=clogit solution ODDSRATIO;
   random timepoint/ subject=SUBJID(sequence) residual type=cs;
   estimate 'Trt vs. Control' drug 1 -1;
run;

But it didn’t work and I got the following error:  "ERROR: R-side random effects are not supported for the multinomial distribution."

 
Now I have applied the model as follows, but I'm not sure if this satisfies the protocol requirements:
 
proc glimmix data=t1 method=laplace;
   class SUBJID sequence period drug timepoint;
   model AVALC = sequence period drug/ CL dist=multinomial link=clogit solution ODDSRATIO;
   random timepoint/ subject=SUBJID(sequence) type=cs;
   estimate 'Trt vs. Control' drug 1 -1;
run;
 

I would be grateful for comments about the error using the _RESIDUAL_ and how can this be fixed?

[SteveDenham]

Your first PROC GLIMMIX code will fix the incompatibility between the multinomial distribution and the residual option.

Make sure that timepoint and period refer to different design elements.  You may want to include timepoint in the model statement.  Consider this MODEL statement:

model AVALC = sequence period drug period*drug timepoint timepoint*period/ CL dist=multinomial link=clogit solution ODDSRATIO;

If this leads to convergence problems remove the timepoint associated terms.  However the period*drug interaction is likely necessary for calculating odds ratios between the two test articles within each period.

SteveDenham