## FDA loves replicate bears? [R for BE/BA]

Dear Yung-jin,

» I don't know why you used TYPE=UN in your SAS code. [...]

This was because I could not identify the syntax to implement such things (FA0(2) or CSH) within lme() and I had the guess that the lme() output
"Random effects:  Formula: ~tmt - 1 | subject  Structure: General positive-definite, Log-Cholesky parametrization"
corresponds eventually to SAS UN parameterization. BTW could not find any hint up to now what this really means: General positive-definite, Log-Cholesky parametrization. Questions for that are always answered: Buy the book of Pinheiro/Bates "Mixed-Effects Models in S and S-PLUS", Springer (2000), there it is described. But recent I am a little short in money. We have a financial crisis .

» [...] As long as bear complies with the FDA Guidance (2001) and is statistically correct, it should be acceptable, I guess.

Sure! As long as ...
But the FDA statistical guidance explicitly states the model with different inter/intra-subject variabilities (Proc MIXED code in appendix G, I think) in evaluating ABE in replicate studies.

And this is not the model implemented at present in bear for evaluation of replicate studies.
I myself are convinced that the assumptions of no subject-by-treatment interaction and equal within subject variabilities are reasonable if we talk about ABE. Although with these assumptions an advantage of replicate designs vanishes, namely that we are able to study different variabilities.

But unfortunately statistically correct or reasonable is in regulators view not always correct or reasonable.
Not so astonishing because statisticians are them selfs of such different opinions if I think about Type III SumOfSqares or now the question of denominator degrees of freedom.

Regards,

Detlew