From: Toufigh Gordi
Subject: [NMusers] Fixed or estimated?
Date:Fri, 28 Mar 2003 13:57:24 -0500
Dear all,
I have a simple PK/PD model of cell binding. I worked on the PK and then
fixed the PK parameters and got PD estimates. Then I estimated everything
all together.
The new estimates (when everything is free) are not identical to when PK
parameters were fixed, although they are very similar. The question is what
parameters to present, and what is the criteria for choosing the results
from one model (e.g. fixed PK) over another (nothing fixed). Different
plots from the two runs are very similar. As one might expect, OFV is lower
in the "free" model (decreased by 11, #PK parameters=9). My second question
is whether one can use a decrease in OFV as a guide for choosing the model
in this case. How much should the drop be to imply a significant advantage
of the "free" model? Does this model have 9 more parameters compared to a
fixed PK model (which would require a much larger drop in OFV than 10 to be
significantly better)?
Personally, I would present the results from a model where nothing is
fixed. However, I would appreciate some discussion on the matter.
Toufigh Gordi
Fixed or estimated?
From:Liping Zhang
Subject:Re: [NMusers] Fixed or estimated?
Date:Friday, March 28, 2003 2:34 PM -0700
Dear Toufigh,
I agree with Dr. Gibiansky's recommendation completely.
If you are interested in PD model, an additional reason for using the
all-para-free model is that when you simulate (the ultimate goal of
modeling), that is the model you will use. If you use the fixed-PK model
analysis results to simulate, then you are not really simulating from the
same model as your data-analytic model.
About the OFV, I agree with Dr. Bachman, too. For the fixed-PK analysis,
you did not indicate whether you included BOTH PK and PD data to estimate
PD when you fixed the PK para to estimate PD. If you do not include both,
there is no comparison between two OFVs.
Best regards,
Liping Zhang
_______________________________________________________
From: "Bachman, William"
Subject:RE: [NMusers] Fixed or estimated?
Date: Fri, 28 Mar 2003 14:49:03 -0500
It turns out that OFV is a less than ideal criterion for goodness of fit and
I never use it as the SOLE determinant for model discrimination. Also take
into consideration the goodness of fit plots and the magnitude of the
variance parameters.
From: Leonid Gibiansky
Subject:Re: [NMusers] Fixed or estimated?
Date:Fri, 28 Mar 2003 15:00:15 -0500
Toufigh Gordi
This is win-win situation, no matter what you decide you have good a model
that describe both PK and PK/PD. You may consider similarity of the
parameters as a confirmation that your overall model is good. Which one to
choose depend on your goals. In your case, they are nearly identical, so it
does not matter what you do.
In general, the "fixed parameters" model provides the best fit for the PK
data. For the PD data, this model uses PK model predictions and PD data to
describe PK/PD relationship. "Free parameters" model may provide slightly
worth fit for the PK data with slightly better fit for the PD data (and
better overall fit).
You should not count PK parameters into the PD model (more precisely, you
should count parameters even if they are fixed). Difference of 11 points
evidence that both models are roughly equivalent, but they have the same
number of the parameters if you count fixed one as well.
If you need the model to describe PK data only, I would not add PD data
there (because incorrectly chosen PK/PD model may damage otherwise good PK
model). If you would like to describe PD data, use the best model (with
free parameters). Similarity of two models evidence that PD data do not
disturb PK fit, providing additional comfort.
Good luck,
Leonid