## Confidence intervals vs. point estimates [Regulatives / Guidelines]

Dear ElMastro, Dan & all!

» […] the guideline specifically reads: "The test product should usually originate from a batch of at least 1/10 of production scale or 100,000 units, whichever is greater, unless otherwiseustified."
» This is because everybody knows that there is variability between batches and that performance of a pilot batch is not necessarily equal to a full scale batch. So, you will just with kind words remind the assessor that a bunch of his/her colleagues acknowledge that fact that pilot performance is not pivotal performance and cannot be expected to be. Yes, there is variability between batches. No, it is not a problem, you are handling it well by separating pivotal data from pilot data.

I would say, this argument holds exactly for Study 1 (lab batch). Study 2 (pilot) was performed with a full size batch (I assume), but due to power considerations (sample size!) and the property of HVDs (variable PE across studies) a direct comparison with pivotal studies is futile.

Agree with the mixed model – would be fun setting up (other factors: study and center).
But I would not dive too deep into the murky waters of statistics, but rather go with common sense. According to guidelines (not only BE) nobody expects that all studies in the file were performed with the same batches. Imagine a hypothetical situation: Two studies fasting/fed. The same batch of test, but two batches of the reference. There is a true food effect of -20% and a true ‘batch effect’ of +25%. What would we get? BE fasting and BE fed. As ElMaestro suggested I would remind the assessor in nice words, that following his/her arguments would call for a requirement which is not covered in GL(s).

» S.
  ^

Cheers,
Helmut Schütz

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes