Bioequivalence and Bioavailability Forum

Hi all,

am at a crossroads here:
A 2-stage parallel study could be based on the normal linear model
ln(y)=T_i+S_j+e

where T_i is the fixed treatment effect for the i'th treatment (simplest form just two levels) and S_j is the fixed stage effect for the j'th stage (two levels).
If we fit this model then we will not have the opportunity to assume unequal variances between treatments.

I am thinking we could perhaps work with a model like
ln(y)=T_i+S_j+e_i

where the errors are specific to the i'th treament, that is:
e₁~N(0, s₁)
e₂~N(0, s₂)
We could still calculate sums of squares etc.

I have no idea if this makes sense, and if it does, then I have no idea how to fit it. I assume this would hafta go via a mixed model with the two sigmas squared on the diagonal. Someone please slap me with a baseball bat.

So... in light of the possible regulatory preference for unequal variances between treatments and the limitations of the normal linear model, what would be your choice?

Also, for a normal linear model approach I don't think I will consider the stage x treatment interaction, although it does grab a degree of freedom and reduces the residual. Any one having hot feelings for this interaction?

Many thanks for all ideas you may have.