From: Franziska Schaedeli <email@example.com>
Subject: weighting observed data
Date: Mon, 05 Jun 2000 11:09:52 +0100
Dear NONMEM users,
I am working on a kinetic data set from hemodialysis patients, with concentration measurements in plasma and dialysate, and with measurements of total amount excreted per observed dialysis. The study period is one week and includes three 3-hrs dialysis sessions. During these sessions, 15 to 18 dialysate concentrations were measured, whereas from plasma there are only 3 measurements available per session. Fitting these data to a PK model (individual fits) results in perfect fits for the dialysate data, whereas the plasma data fits are quite poor. I think the frequency of dialysate measurement induces a bias in the model fitting.
Here is my question: Is there a way to weight the plasma data with respect to the dialysate concentration data?
Franziska Schaedeli Stark
Date: Mon, 5 Jun 2000 14:14:54 +0200
Subject: Re: weighting observed data
What kind of error model did you use ? proportional or constant/additive?
If the concentrations in plasma differ very much from those in the dialysate and you want to fit both types of data with a single residual error, I would prefer a proportional error model anyway.
If you prefer to use separate error weighting for plasma and dialysate for whatever reason (e.g. known differences in assay sensitivity), I would introduce a true/false variable to discriminate between the two DV's.
Hope this helps,
let me know if you want a more specific example of the code.
Dr. Joost DeJongh
Leiden Advanced Pharmacokinetics & Pharmacodynamics (LAP&P) Consultants
2333 CM Leiden
Phone: + 71 524 3000
Phone: + 71 524 3002 (direct line)
fax: + 71 524 3001
From: "Sale, Mark" <firstname.lastname@example.org>
Subject: RE: weighting observed data
Date: Mon, 5 Jun 2000 07:51:53 -0400
You certainly can have a different residual error for the two samples. Just have an indicator variable in the data set (IND = 1 if plasma, 0 if dialysate). Then the error is:
ET = IND*EPS(1) + (1-IND)*EPS(2)
But, that may not solve the problem. When you have a lot more data from one site/assay, that site/assay will drive the other model, i.e., the pk part will be made the best fit the dialysate data, not the pk data. I ran into this same problem with QT interval relating to plasma concentration. It is my view that the model should reflect biological causation whenever possible. In the QT case, clearly plasma concentrations drive the QT interval. I don't want QT interval driving plasma concentration. So, I fit the plasma concentration first, fix that part (output individual pk parameters using post hoc), then fit the QT part (merge the individual pk parameters into the data set for QT). In your case, I'd consider this approach. The dialysate concentration should, mechanistically, be driven by the plasma concentration, not vise versa. So, fit the pk first, fix that, then fit the dialysate model.