Bioequivalence and Bioavailability Forum

Hi Angus,

❝ My understanding is the mixed model (in my case from Phoenix WinNonlin) predicts the value of the PK parameter by a regression model in the terms that you describe. You then can compare with the observed value of the parameter.

In a normal linear model you have a whole bunch of fixed effects (things that have discrete levels, but do not vary continously: gender (Male vs female), treatment (T vs R), Hair color (blond vs red vs blue etc), Season (winter vs summer) and so forth. And then you have an uncertainty or error that follows a normal distribution with mean zero.

In a mixed model you...well... mix things up. You introduce additional random effects. So suddenly our model expression does not just encompass one thing (like the error in a normal linear model) which follows a normal dist. but more than one. In that case, minimisation of sums of squares does not give the max. likelihood. In stead the max. likelihood (actually 'restricted') must now be calculated by an optimiser.

❝ Now I think this is correct, but by the way why do we call it a mixed model? I am trying to shape my understanding of this activity better

Normal linear model: Simple stuff, a single normal distribution as well as several constants explain the variation in our data.
Linear mixed model: Complex stuff, several constants plus two or more normal distributions must be used to account for the data scatter.

Subject is sometimes, actually quite often, seen appropriately as a random effect. We are not interest in knowing how Schützomycin performs in just 20 subjects. We are interested in extending the knowledge we have about Schützomycin to the entire population (of which our 20 test subjects are hopefully representative).
In ordinary 222-BE Subject can be modeled as fixed (giving a normal linear model) or as random (giving a mixed effects model), both having the exact same result (one exception: a missing period value; the entire subject is thrown out in least squares minimisation but if you apply a mized model you can find a max likelihood even with such subjects taken into account).
I am not a WinNonlin user, but I think Phoenix will perform the calculation as a mixed model by default unless you ask it specifically to do it with all effects as fixed.