Bioequivalence and Bioavailability Forum

Dear bears,

❝ The forum was hanging there without any response. I wished that I did not do anything stupid at that moment.

Terribly sorry about that. It wasn’t your fault. The forum is hosted on a shared server, and another user installed a faulty script which filled up the server’s error logs until no space was left on the HD. Then the MySQL-database crashed…

❝ ❝ entire alpha-risk in the interim looks. In other words - you must not evaluate the study for BE (i.e., calculate the 90% CI...

❝

❝ My guess is this procedure will not be done unless the power is ≥ 80% (Method C, as the left branch of the flowchart).

Right. According to this post of yours I was afraid that you want to evaluate BE in any case.

❝ Do you mean that we should not do this when the power is < 80%?

Yes. Or to be more precise, not with the conventional alpha=0.05.

❝ If doing so, I guess the answer is still not BE, isn't it?

Haha, now you are cheating. You should not evaluate BE at 0.05, otherwise the alpha is ‘consumed’.

❝ The question can be if the regulatory agents accept BE when the power < 80%, and evaluate BE at stage 1 with alpha = 0.0294 (Method C, as the left sub branch of the right branch) when we analyze the data as the fixed-sample trial.

Why not? It was shown in the simulations that it works; at least EMA accepts it.

❝ I don't know what possibility of getting BE will be with alpha = 0.0294, when it has been not BE with alpha = 0.05. It is more stringent (0.0294 vs. 0.05), isn't it?

Yes, but that’s the trick. If the decision tree leads you to the right branch (based on power), you don’t evaluate at 0.05, only at 0.0294. Of course it’s more strict, but that’s the penalty in sequential designs.

❝ ❝ left for further looks. This was actually the problem with the Canadian and Japanese method.

❝

❝ What kind of the problem with the Canadian and Japanese method can be?

See the introduction of Potvin’s paper:

Statistically, add-on designs cannot preserve the nominal type I error rate if a test is conducted at the nominal level following the completion of the initial trial and then again after the additional subjects are included. (WHO does indicate that the level of consumer risk must be considered; Canada does not.) The argument in favor of add-on designs is that they do make use of the data already collected, and the inflation of the type I error rate is ‘acceptable.’

In other words, the entire patient's risk is already consumed in evaluating the first part for conventional BE (90% CI). If you go for a second part then, what’s the overall risk? Definitely >0.05 (multiplicity!). How large the combined patient’s risk is, is actually unknown (depends on the sizes of the two parts, etc.). Method A in the simulations (which was already more stringent than Canada's/Japan's method) lead to a risk of 6% and was not calculated in all combinations.

❝ only extra 10% of the total costs of the fixed-sample design? or it depends how many subjects should be recruited at stage 2?

Both. The ~10% is an estimate for Gould (1995). That’s the penality for the interim look. But on the long run (many studies) sequential designs may be more economical, because the comparison with fixed-sample designs is based on a delusion – namely that the variance is known and the sample size optimal. If the variance is uncertain, and lower than expected, sequential design will stop at the early stage.

❝ May I ask why WNL does not implement the data analysis for the two-staged design if FDA/EU have been able to accept the two-staged designs? or it has been implemented with WNL (v6.x)?

Pharsight runs a commercial operation. If they don’t see additional profit, they won’t implement anything. Users have to ask them. Nothing is implemented in v6.1, but it would at least help to publish an application note on the support site (or the new board).
Right now there’s no way to get the CV without running the BE-wizard. Another problem starts when it comes to calculating power for the first decision. AFAIK WNL’s result (all versions) is simply wrong (see here). Now for the good part: it is possible to calculate the narrower CI and set up Potvin's model for the combined analysis.

❝ This should be no problem with […] R

Sure, R ulez! You even may combine the entire procedure, because power/sample size calculation is possible within bear (no way in WinNonlin).