Bioequivalence and Bioavailability Forum

Dear zizou,

❝ in this post I mentioned the oldest study with partial replicate design (sequences RRT,RTR,TRR) I know about.

THX! Funky model by Donald Schuirmann:

PROC MIXED;
CLASS SUBJ SEQ PER GRP TRT;
MODEL Y =
  GRP SEQ GRP*SEQ PER GRP*PER TRT/DDFM=SATTERTH;
RANDOM SUBJ(GRP*SEQ) SUBJ*TRT(GRP*SEQ);
ESTIMATE 'T VS.R' TRT -1 1/CL ALPHA=0.1;
RUN;

Two things are interesting: No REPEATED statement and the covariance structure is not specified. Hence, SAS applies the default TYPE=VC (variance components). Let’s see what happens in PHX/WNL with the EMA’s data set II (of course, group terms deleted from Donald’s setup).

1. FDA’s code (2001 BE-stats guidance and 2010 progesterone guidance)
   PE 102.26% (90% CI 97.05–107.76%)
Warning 11091: Newton's algorithm converged with modified Hessian. Output is suspect.
   
2. As above but TYPE=FA0(1), UN, or CS (CSH gave error)
   PE 102.26% (90% CI 97.05–107.76%)
   
3. Donald
   PE 102.26% (90% CI 94.37–110.82%)
   
4. EMA Method B (implicitly DDFM=CONTAIN)
   PE 102.26% (90% CI 97.32–107.46%)
   Same results with DDFM=SATTERTH cause the data set is complete and balanced.

I think that Donald’s model doesn’t make sense cause the replicative structure of the study is ignored (given, like in the EMA’s Method A and B).