Bioequivalence and Bioavailability Forum

Dear Shuanghe!

❝ More general question is: how to treat the topic of "Subject to Analyse" in replicate (partial or full) BE study?

Good question! Next question .
Seems we here don't have an answer or have overlooked this important post.

Let me throw in my two cents:
I'm not aware of any regulatory recommendation about that topic.

I personally would prefer to analyse the data as are, i.e. include all available reliable data, missing periods or not.
This could be done if we would evaluate with an mixed model (Proc MIXED in SAS or equivalent lme in R) considering the subjects as random effect what they IMHO really are.

Unfortunately the European great oracle EMA doesn't allow us to do so.
Remember the "all effects fixed" recommendation.
Thus subjects having not at least one PK metric under Test and Reference will not contribute to the estimation of the treatment effect (ABE criterion).

Regarding the EMA recommended scaled ABE method subjects with one missing under Reference will not contribute to the estimation of the intra-subject variability for Reference. Interesting enough the dataset I of the EMA Q&A contains such subjects with one missing for Test or Reference and you can see that they are not skipped from the analysis.

To summarize:
Seems the EMA in applying its recommended method doesn't exclude subjects with missing data.
Then the two parts of the scaled ABE criterion (point estimator of µ_T - µ_R with its confidence interval and on the other hand the estimator of the intra-subject variability for reference rely on different numbers of subjects.
This also applies to the FDA scaled ABE method (SAS code in Progesterone guidance, current Feb 2011 version) if one doesn't exclude subjects due to missingness beforehand.

I personally are not comfortable with this as it is well known that missings in the method of moments (and both FDA and EMA method are some sort of) may lead to biased estimators. And no one is aware of the magnitude of influence of this effect on the BE decision.
This is not the case for complete data or evaluation via mixed models.
Thus if mixed models are not "allowed" it would be wise to exclude subjects with any missing. But this will nearly certainly not accepted by the regulatory authorities I think.

Cough, much ado about nothing .
Any other opinions out there?