From firstname.lastname@example.org Thu May 8 12:06:45 1997
Subject: Normalization of PK parameters I would like to ask what other users do regarding normalization of PK parameters to body weight, BSA, etc.
In the model building stage, covariates are added sequentially and evaluated for significant improvement in the objective function. If the parameter is normalized during this process, eg.
TVCL= (Theta(1) + Theta(2)*age + Theta(3)*Clcr)*WT
there would be an implied interaction between WT and the other covariates (age and Clcr). This would complicate the interpretation of changes in the objective function.
One solution would be to perform the model development process without normalizing for weight. After the final model is accepted, add weight to the model as shown above and use the new Theta values from the weight normalized model. At this point changes in the objective function (unless increased) would not be considered. What happens here when weight is found to correlate with Eta's at an earlier step and it is appropiate to include weight as a covariate in the nested additive model?
Any general comments about using interactions between covariates would be helpful.
Thanks in advance for your input.
From lewis Thu May 8 12:28:16 1997
Subject: Normalization of PK parameters Re David Nix's problem:
There is no general answer, but the "problem" arises due to the model's mixing proportional and additive effects.
It may well be that the most physiological model does mix these, but it is usually possible to construct the model for a parameter either all proportional; e.g.,
(1.1) TVLCL = theta(1) + theta(2)*log(age) + theta(3)*log(WT) + ...
(1.2) LCL = TVLVL + eta(1)
(1.3) CL = exp(LCL)
[Note that the above is equivalent to CL = theta(1)*age**theta(2)*WT**theta(3)*exp(eta(1))]
or all additive; e.g.,
(2.1) TVCL= Theta(1) + Theta(2)*age + ... + Theta(3)*WT
(2.2) CL = TVCL + eta(1)
To me, (1) makes the most sense, as it implies that the parameter is always >=0, among other things ...
Regarding interactions per se, there are graphical methods to check for these. Sometimes they "go away" when the "proper" transformation is done (as in using (1) instead of the model David wrote), but when present, they tell one about how the model should be formulated. As such, they are not a problem, from my point of view ...