Bioequivalence and Bioavailability Forum

Dear Boonchai!

❝ […] afraid my poor English make someone misunderstand.

Don’t worry, native English speakers are in the minority ’round here.

❝ […] the pooling of CV […] was developed using sampling distribution of sample variance from classical statistics or parametric approach. […] from assumption of parametric statistics, so each variance as samples or random variable (RV) should be iid from chi-square distribution. Are there anyone check them? How much its robustness if it breaks the assumptions? Are there any methods to combine them if it's not come from the same population? And bla bla bla…

Wonderful – you are right in every aspect. IID is also an assumption in the classical 2×2 cross-over model, which may not hold as well. If the reference has a higher variance than the test, you are penalized. In a nonreplicate design we have no means to check the validity of the assumption, so bad luck. It would be good statistical practice the check the assumptions and have a contingency plan a priori if they are violated – rendering the model invalid. In the past many protocols in Europe checked the ANOVA’s residuals and evaluated the study by a nonparametric method if the assumptions were violated. Regulators didn’t like that (why the hell?)… BTW, in Japan such a procedure is still acceptable according to the current guidelines.

When it comes to pooling, we have no means to check the assumptions.
I don’t know whether anybody challenged the method with real data (at least it’s not published).

❝ I appreciated Helmut because he understands and be able to apply statistical concept even if he does not grow up in statistical field,

THX from an interested amateur!

❝ Anyone can choose their own CV as long as they want to. It is quite subjective and there are many different reasons to support their own CV. […] for example, Mary, Ronaldo, Lady K., Kim, Filipe M., Tom…

Great summary! Not least Dave Dubbin’s FARTSSIE is popular with its  Boss Button  – “We know the sample size beforehand and tweak the input (CV, GMR, power) to get a justification for the protocol.” That’s clearly not the idea behind ICH E9.
Pooling is just a member of the statistical toolbox. It’s up to the responsible persons to use their brains to give individual studies more or less credit (as ElMaestro suggested). The best estimate one gets always from a – reasonably large – pilot study. Only then you can control all the side-conditions (clinical performance, bioanalytics, ) which may influence variability. We also must not forget the GMR. Sometimes sponsors are too optimistic about their product. If the GMR in a pilot study comes as 81%, what would you do? Retreat to statistics (random variability and crossing fingers) or go straight ahead and reformulate?