Bioequivalence and Bioavailability Forum

Dear Andrew!

❝ Our rationale is as follows: (a) variability of the drug substance is unknown & not published;

OK.

❝ (b) for formulation reasons there is a significant risk of bio-inequivalence (making the option to stop after the first group an attractive one).

I don't know which method you are planning to use. Potvin et al. (2008) was not validated for a stop-criterion based on the actual (observed) point estimate (PE) in stage 1. You estimate the power based on CV_intra and the expected PE applied in study planning. If power ≥80%, you evaluate the study with α 0.05 (pass|fail). If power <80% you evaluate the study with α 0.0294 (pass|fail). If BE fails, calculate sample size for stage 2 based on the expected (!) PE used in study planning. That's the tricky part: if you already expected a high chance of failure (let's say a PE of only 85%) and planned the study for a PE of 95%, there's no (statistical) way out to to stop the study, even when you see an actual PE in stage 1 of only 82%. That's clearly a drawback of the method, and I would not recommend it in any case when you expect a reasonable high chance of bioinequivalence. In such a case I would rather perform a classical pilot study to get an idea of the PE.

Quoting the paper:

Another possible application of sample size reestimation is to adjust the sample size based on the observed effect size seen in an initial sample. For example, one might choose a sample size based on the assumption that the ratio of geometric means is between 0.95 and 1.053, but the estimated ratio seen in an initial sample may be, say, 0.92. One might be tempted, in this case, to recalculate the sample size assuming that the true ratio of population means was 0.92. However, simulations carried out by Cui et al. indicate that if this procedure is applied naïvely, the overall type I error rate may be inflated by 30% or more. More complicated statistical procedures have been proposed to allow sample size reestimation based on the observed effect size. […] While these more complicated procedures could certainly be considered for a BE study, they lack the feature of using the usual test statistic formulae (possibly with a simple adjustment of the nominal type I error rate), and they have not been validated for use with two one-sided tests and unknown variance. We will not consider them further here.
[…] More work needs to be done to examine the […] inclusion of a futility rule […] and upper limits on the total sample size.
(my emphases)

For months I want to ask Walter Hauck whether the group is working on such a method, but it still stays on my to-do-list…

❝ My questions are:

❝ Has anyone heard otherwise (will two-stage stay or go)?

My last informal information date back to May 2009: will be kept.

❝ Is anyone else using this approach yet?

Yes, me. See the footnote in this post. I expect to start the interim analysis next week.

---

Edit 2009-08-18: Power was >80%, the study was stopped after the first stage and demonstrated BE. So no experience with a second stage yet.