Proc Mixed model statement - the order of covariates affect the standard error estimates?

[simonli]

Dear all,

Recently I was running sas proc mixed to validate my counterpart's results. Below is the statement I used (-note, 'Base' in the model statement): 

proc mixed data=indata method=reml;
   class TRTPN(ref='1') SITEGR1 AVISITN SUBJID;
model CHG = BASE  TRTPN SITEGR1 AVISITN TRTPN*AVISITN  / DDFM=KENWARDROGER;
repeated AVISITN / type=un sub=SUBJID;
lsmeans trtpn*avisitn/pdiff cl ;

run;

My counterpart runs like this:

proc mixed data=indata method=reml;
   class TRTPN(ref='1') SITEGR1 AVISITN SUBJID;
model CHG =  TRTPN SITEGR1 AVISITN TRTPN*AVISITN  BASE  / DDFM=KENWARDROGER;
repeated AVISITN / type=un sub=SUBJID;
lsmeans trtpn*avisitn/pdiff cl ;

run;

Surprisingly, the standard error estimates from the lsmeans statement are quite different.

Take one example:

mean (SE) for model-adjusted means for one visit of one treatment group:

9.78 (4.896)           9.78 (2.150)

The adjusted means are the same but the standard errors change a lot; and the standard error values depend on the location/order of the covariates.

I do not understand why. Am i doing something wrong or the variance-covariance changes with the order of the covariates?

thanks all.

[PGStats]

Do you get the same problem when asking for

DDFM=KENWARDROGER2  ?

[simonli]

Yes; I tried to use DDFM=KENWARDROGER2 and the problem is the same. Thanks

[simonli]

Just want to add one more piece of information: I have tried the same model to generate multiple tables for different endpoints - it is odd that only one single table that had such problem: the order of the 'Base' in the model statement changed the standard error. All other tables had no difference.