Bioequivalence and Bioavailability Forum

Hi Mohamed,

❝ For Partial and full replicate studies, if we have some subjects with missing periods, i mean for example subjects who attended only 2 periods in partial replicate

First answering a question you didn’t ask. In partial replicate designs (TRR|RTR|RRT) I suggest to specify two analysis sets if data of the third period in sequence RRT is missing:

1. Exclude subjects from the assessment of BE since the GL requires at least one treatment of T and R.
   
2. Keep subjects for the estimation of CV_wR.

❝ 1) Which method to use following the EMA guidelines; method A "GLM" or method B "Mixed"?.

❝

❝ I want a clarified answer please with guideline reference if applicable.

Good question. The Q&A document states:

The analysis […] show that this approach (Method A) is feasible even for unbalanced replicate design studies. The advantage of this approach is that it is straightforward and that it appears to be software and software option independent. A simple linear mixed model, which assumes identical within-subject variability (Method B), may be acceptable as long as results obtained with the two methods do not lead to different regulatory decisions. However, in borderline cases and when there are many included subjects who only provide data for a subset of the treatment periods, additional analysis using Method A might be required.

My understanding and comments:

• Not only Method B assumes identical within-subject variability – the same is true for Method A. A major advantage of a full replicate design is that it allows to separately estimate s²_wR and s²_wT. IMHO, that’s a bizarre assumption which rarely holds true in practice (in most cases I have seen s²_wR > s²_wT). The estimation is possible either directly by Method C (which seemingly is not “compatible with [the] CHMP guideline”) or modifying the model given in Section 3.4. by including data of T only. The Q&A also states:
  “An advantage of Method C is that it directly calculates s²_wR. However, sometimes the algorithm fails to converge.”
  This observation is correct only for partial replicate designs, since T is administered only once and the model is over-specified. If the algo converges (generally it does), the estimate of s²_wR is correct and the estimate of s²_wT is nonsense. However, Method C always converges for full replicate designs.
  
• Since one doesn’t know beforehand whether the outcome will be a “borderline case” (which “border”?) I guess one has to perform both Methods even if Method B was aimed at. The wording gives me (!) the impression that the PKWP assumes that Method A is more conservative than Method B. That’s not even true for the Q&A’s data set I: The CI obtained by Method B is wider (17.80%) than the one obtained by Method A (17.78%). We are working on a paper supporting validation of software for the EMA’s ABEL. In 57% of our reference data sets Method B lead to wider confidence intervals. However, in 96% of cases the BE decision (pass/fail) agreed. Interesting that one of our data sets would fail by Method B but pass by Method A.

From a purely statistical perspective my preferences are: Method C ≫ Method B > Method A. I don’t like the idea to treat subjects as a fixed effect. Given the observations from above, a sponsor probably would fair best with Method A only.