Date: Thu, 10 Aug 2000 09:52:58 -0400 (EDT)
Subject: PD analysis without PK data
Dear NM users,
I would like to have your opinion on how reliable the EC50 estimates of a drug would be if I fit PD data in the absence of PK data.
here is the scenario:
60% of my data set has both PK and PD data. So fitting PK separately has given me good pop and individual parameter and inter ind and residual variability estimates. Now, by fixing these and just letting NONMEM choose the individual PK estimates that best fit the PD data, will the EC50s that I calculate to describe the other 40% of the data set (i.e only PD) be reliable? What statistics, if any, should I use, if this procedure works, to compare PK estimates in PK rich versus NO PK.
Date: Thu, 10 Aug 2000 10:26:10 -0700 (PDT)
From: Liping Zhang <firstname.lastname@example.org>
Subject: Re: PD analysis without PK data
I have been working on a similar project: comparing the performances of different PK/PD anlayses. Three methods of what I use are:
1. fit PK data first. Get the individual PK parameter estimates and take that as part of the input data along with PD observations to estimate PD parameters (NONMEM Help Guide VIII PK/PD sequential 1 Example).
2. fit PK data first. Get the population PK parameter estimates. Fix these PK parameters, with PK observation, estimate PD
3. fit PK data first. Get the polulation PK parameter estimates. Fix these PK parameters, without PK observation, estimate PD (NONMEM Help Guide VIII PK/PD sequential 2 Example)
The performances are evaluated by calculating the integrated error from each method. Preliminary results show method 2 works the best, followed by method 1 and method 3.
Here is how I did it. To analyse different methods and evaluate them,
1. simulate data from a set of PK/PD parameters.
2. fit them by different methods.
3. calculate the mean expected PD effects from estimated PD parameters at given cp, compare it with the mean expected PD effectes from the true parameters (from which the original data are simulated). If the differences are calculated many times across a given range of CP (for example, 100 times from C10 to C90), then the mean of the differences could be taken as the integrated error against CP. The smaller the error, the better the performance.
if any, should I use, if this
> procedure works, to compare PK estimates in PK rich
> versus NO PK.
I assume it is "to compare PD estimates in PK rich versus no PK". Do you mean compare PD estimates of the data that has PK and PD with PD estimates of the data that has no PK? One way to do it (I am not quite sure) is to seperate the two groups and look at their goodness of fit plots, see if you can find any difference between them.