Bioequivalence and Bioavailability Forum

Hi Angus,

❝ My guess is you are with SAS and you have a statistical background.

Hehe, that's actually as wrong as it can be. I had a basic course in statistics about 100 years ago. Graphical estimation of standard deviations from observed data and such. Never had WinNonlin or SAS on my system and I doubt I will ever have . Only R, which I like. Except for the syntax
(what the h%ll does "~1|blah" really mean in R???)

❝ I find the descriptions in books are often poor.

True... ever successfully tried to read and understand the little instruction books that come with Ikea furniture?
"Congratulations on buying a new Spatzwoolah. To assemble it use the fnoofbygthwat to turn screw A 70 degrees counterclockwise with your left hand while holding the Grabowski-pin with your right hand. Make sure not to turn the screw more than 40 degrees. When you hear a click, firmly grab the handles on the upper jaweenmazonkie, but do not let go of the Grabowski-pin or the fnoofbygthwat since these hold the zwix in place behind the lower left plinaque which should be supported by your free hand so as not to crack" and so on.

❝ The way WinNonlin model is set up is as follow:Fixed effects are sequence +treatment +period and the Subject(Sequ) is a random effect.

By default. I think that's correct although I do not know.
By the way: Subject(Sequ) gets my systolic blood pressure up to 520 mmHg. Nesting syntax isn't relevant because you're assigning random numbers to trial subjects, and no two subjects can have the same number.

❝ I think this ismOK.

Well... yes, often it is, but strictly is might not be... At EMA we have: "The terms to be used in the ANOVA model are usually sequence, subject within sequence, period and formulation. Fixed effects, rather than random effects, should be used for all terms."

In the US they will generally use PROC GLM / SAS or R's lm to reproduce the results for 222-BE studies, as far as I know. That means that if your dataset from such a trial has missing values (as in: some subjects did not have data for both periods but ony for one) then you might get a question about, since FDA's analysis result will be different from yours.
I'd only really gamble on a mixed model if we talk non-EMA and replicated or semi-replicated designs.