RE: In real life
Hi Siwei,
Biases can definitely come from multiple sources including model
mis-specification (as you noted with #1 below). There are multiple methods
that you can use to assess the improvement of the model which may include using
prior information (a prior statement for the parameters reported in the
literature or fixing it which is essentially a very strong prior but less
preferred). Another method to assess the fit into more complex models can be
likelihood profiling.
Finally, for any model the way to report it depends on how you're using the
model. Generally, it is best to note the assumptions of your model, the
deficiencies, and how it can best be applied to respond to the questions of
interest. If in your or the team's estimate, the assumptions or deficiencies
don't allow assessment of the question at hand, it should not be used.
Have a good day,
Bill
Quoted reply history
From: [email protected] [mailto:[email protected]] On
Behalf Of siwei Dai
Sent: Tuesday, August 13, 2013 10:12 AM
To: [email protected]
Subject: [NMusers] In real life
Dear NM users:
I have questions about the principles.
It is rather common that clinical PK data are 'bad': very sparse sampling
and/or the sampling stopped too early. (I understand that you can never make a
useful model with 'bad' data, but in the case that you have to make a model
from them).
In this case, I understand that you want to go for a simpler model, say if rich
data support a 3-compartment model, you probably need to go for a 2-compartment
or even a 1-compartment model, otherwise you may see signs of
overparameterization. However, after I modeled the data with a simpler model, I
saw situations where the GOF plots are biased, with low concentration being
underestimated and high concentration overestimated; the CWRES vs. PRED plot
showed a falling trend line.
My questions are:
1. Are these bias due to the use of a simplified model?
2. if the answer is 'yes', should I go back to a more complex model but fix
some of the parameters based on literature?
3. Are these, after all, legitamate questions? or should I just say 'the data
are bad, we cannot make a model from it?"
Thank you very much in advance for your input.
Best regards,
Siwei