Bioequivalence and Bioavailability Forum

Hi Dan,

❝ […] if it is possible from a statistical point of view to use the same subjects participating in stage I also for stage II?

Nice idea!

• Short answer: Not with the published methods. Here you would end up with a fully replicated design without scaling.
  
• Radio Yeravan’s answer: “In principle yes, but you have to take the repeated measurement charac­te­ris­tics into account and demonstrate that the TIE is maintained at ≤0.05.”
  Would be interesting to develop a suitable framework and get the adjusted α right. Since we don’t scale it should be doable. Whether it makes sense is yet another question. Study costs will be similar to conventional crossover TSDs (practically the same number of admin­is­tra­tions), but 2× higher blood loss, higher dropout rate, .

Hi Dr_Dan,

apart from what Helmut said, I'd like to add that unless we somehow fix sample size at stage 2 so that it is equal to sample size at stage 1 we could have subjects for whom there is data for one stage and subjects for whom there is data from two stages. This is complicated to handle from a covariance perspective.
And if we fix sample size at stage 2 so that it equals sample size at stage 1 then my gut feeling tells me that we might too easily get a low power cf. futility rules for ordinary two-stage designs.

Dear Dan,

additionally and more in depth to what was already said by Helmut and our Capt'n:

IMHO this "interesting" idea has some statistical difficulties.
If you evaluate stage 1 data all is still simple, you have a classical 2x2 crossover with the usual statistical model.

If a second stage is necessary and the sample size estimated for stage 2 is <= n1 you have a replicate design for the subjects having both stage 1 and stage 2 data - four periods, P1 and P2 in stage 1 and the other P3 and P4 in stage 2, i.e. some confounding between period and stage effect. The other subjects having only stage 1 data give a 2x2 design.
Stage itself is no longer a between subject effect but some sort of nested effect. DUNO what sort of, or who is nesting within whom.

The same is true if n2 > n1: you have a replicate design for the n1 subjects having both stage 1 and stage 2 data, and a 2x2 design for those subjects having only stage 2 data.

How the statistical model for such a mixture of different designs with confounding between effects has to be formulated is not obvious to me, especially if one has to include a stage term. See EMA guideline, page 16 "When analysing the combined data from the two stages, a term for stage should be included in the ANOVA model".

I expect all sorts of df=0 or SumOfSquares=0 in the ANOVA decomposition of variance and non-estimable treatment contrasts if one tries to figure out the model by trial and error.

Somebody out there who has a suggestion for a working ANOVA model?
Else the interesting idea has to be thrown into the Thames because we don't know how to evaluate it. Interesting or not.