Bioequivalence and Bioavailability Forum

Dear d_labes,

❝ @ EM:

❝ ❝ ❝ -What is the geometric mean in replicate cross-over studies? Overall or calculated for each replicate?

❝ ❝ They mean lsmeans or model effects, I hope ...

❝ Do you really won't me to do a model fit (which if we are allowed to only consider Reference AUC values?) for the log-transformed values, leaving out each subjects values in turn, and then apply the 5% criterion via exponentiated LSmeans?

❝ My first thought was simpler: Calculate the geometric means leaving out each subjects values in turn and then calculate the percentage of the individual value versus geometric mean (i.e. taken the guidance text literally).

Could you reformulate the first part there?
I meant to say that in each model you will will have an effect or LSMean or marginal mean (or .....) which tell what Test or Ref is worth. Which I guess is your starting point. Taking subjects out one by one reminds me a little of jackknifing and sounds relevant.

❝ ❝ Pick the one that fits.

❝ How do you define "fit"?

(connected to above)
It is not clear what the relevance of geometric means are when a study is imbalanced across sequences. At the end of day the stats packages uses the model effects (=the b-vector from y=Xb+e). Chow and Liu define a model with intercept, in which case b-vec (see below) does not reveal effects for both Test and Ref. Nevertheless, a value for both can be derived, of course from the "fit" or the "model" or wherever.

❝ ❝ "Geometric mean" is a term that should be banned ...

❝ Totally

❝ What the hell do you have against geometric means? Enlighten me.

I don't see the relevance for the term "geometric mean" generally, because I think the term makes no sense for studies with imbalance across sequences. Start out by calculating the geometric mean for "test" in an imblanaced study and compare with the value found in the b-vector. Of course, it is much easier if you fit the model without intercept when model effects for Test and Ref come directly out unconfounded by the presence of an intercept (hell!). If you fit the model with an intercept, on the other hand, then the first column of X is just a buncha ones whereas there is only one column thereafter that has to do directly with a treatment (which one depends on the internal garbling method). All in all, if we could use the term "model effects" (but not LSMeans) then I think (but this is a deeply personal remark) we'd all be sending and receiving on the same frequency and that would keep my blood pressure at acceptable levels.

I have a vague feeling I now went in a direction that is quite useless considering your initial question. I do apologise if this is the case.

Have a great day.
EM.