MMRM Proc Mixed:

[sumiko417]

Hi, 

I have a question regarding MMRM analysis using PROC MIXED. Suppose I fit the following model:

proc mixed data=input;
  by imputenm;
  class usubjid trt01pn avisitn;
  model chg = trt01pn avisitn trt01pn*avisitn;
  repeated avisitn / type=un subject=usubjid;
  lsmeans trt01pn|avisitn / pdiff cl;
run;

When using this model, I noticed differences in the results depending on whether I use trt01p versus trt01pn, and avisit versus avisitn. In the dataset, trt01p and trt01pn have a one-to-one correspondence, as do avisit and avisitn. All variables are sorted prior to fitting the model.

This leads to the following general questions:

1. When using PROC MIXED, is it preferable to use numeric or character variables in the CLASS statement?

2. Does the sorting order of class variables affect the analysis results?
   For example, avisit is a character variable with values such as "Week 1", "Week 2", …, "Week 13". When sorted alphabetically, the order becomes "Week 1", "Week 10", …, "Week 9". Could this impact the model results or interpretation?

Thanks

[Ksharp]

Maybe the reference level is different for this kind of scenario. Try REF= option:
class usubjid trt01pn(ref='1') avisitn(ref='1') ;
class usubjid trt01p(ref='Placebo') avisitn(ref='Week1') ;

About your questions:
"When using PROC MIXED, is it preferable to use numeric or character variables in the CLASS statement?"
No. I don't think so. Both could be accepted . The different thing might be the reference level of CLASS variable.

"Does the sorting order of class variables affect the analysis results?"
Yes. that might lead to the different reference level, you can make some data to test it is right or not.

[jiltao]

In general, it does not matter if your class variables are numeric or class variables, as long as the order remains the same between these two types of coding of your values.

The different results you are seeing are probably caused by different sorting order between avisit and avisitn. The TYPE=UN structure is formed based on the order of the repeated effect, avisit or avisitn, week1, week10,... for avisit has different sorting orders from 1, 2, 3,...13, therefore the covariance parameter estimates would be different between the two approaches. 

I hope this helps,

Jill