From: "Bonate, Peter, HMR/US" <Peter.Bonate@hmrag.com>
Subject: Covariate Model Selection
Date: 10 Mar 1998 15:24:39 -0500

I am building a 1-compartment oral model at steady-state on data that had 1-2 observations per subject. One was specifically taken at trough. In inclusion of a covariate on V such that
TVV = theta(1) +theta(2)*BSA
V = TVV*exp(eta(1))

I had a drop in my objective function of 35 points. The CV on my parameter estimates do not change much compared to the base model without covariate and they are all precisely estimated (CV < 10%). However, I am modeling residual error using an additive and proportional error model
Y = Y + eps(1) + Y*exp(eps(2)).

With the covariate included in the model my estimates of eps(1) and eps(2) double in magnitude. There is no difference in residual plots or observed vs. predicted plots.

Is the covariate model deemed a better model than the base model without covariates? Surprisingly I have encountered a wide diversity in opinion regarding 'yes' or 'no'. I would be interested in the comments from the users group. Thank you.

Peter L. Bonate, Ph.D.
Hoechst Marion Roussel
Clinical Pharmacokinetics
P.O. Box 9627 (F4-M3112)
Kansas City, MO 64134
fax: 816-966-6999
phone: 816-966-3723

****

Subject: RE: Covariate Model Selection
Date: 11 Mar 1998 03:29:00 -0500

Dear Peter,

The correct syntax for proportional - additive error model is

Y = F*EXP(EPS(1)) + EPS(2)

or, equivalently,

Y = F*(1 + EPS(1)) + EPS(2)

Another point to comment:

You write you have steady-state data and a substantial part of your measurements are at trough. I don't think you have enough information to characterize effects of covariates like BSA on the volume of distribution. Trough level at steady-state is mainly determined by CL. Then, CL and V are usually highly correlated. You'd better concentrate on looking for covariates affecting CL. After including all of them significantly improving the fit you can spend some time plotting ETA(V) vs. covariates. And if you are lucky...

Hope this helps

----------------------------------------------------------------------
Janssen Research Foundation
Clinical Pharmacokinetics
B-2340 Beerse
Belgium
Fax: +32-14-605834
Email: vpiotrov@janbe.jnj.com
vpiotrov@janbelc1.ssw.jnj.com

****

From: "Bonate, Peter, HMR/US" <Peter.Bonate@hmrag.com>
Subject: Covariate Model Selection II
Date: 11 Mar 1998 10:17:13 -0500

I recently sent out a message to the group in an attempt to answer a question. I have gotten many responses, most of which were distracted from the main issue by my mistyping the error model statement. Here is the real issue: regardless of the error model or structural model, suppose inclusion of a covariate results in a significant drop in the objective function value, with no apparent change in residual plots, and tight precision of parameter estimates with and without the covariate, BUT residual error increases dramatically after inclusion of the covariate, does this imply that the covariate model is a better fit than the base model without covariate?

For those of you that replied earlier, thank you very much.

Peter L. Bonate, Ph.D.
Hoechst Marion Roussel
Clinical Pharmacokinetics
P.O. Box 9627 (F4-M3112)
Kansas City, MO 64134
fax: 816-966-6999
phone: 816-966-3723

****

Subject: RE: Covariate Model Selection II
Date: 12 Mar 1998 02:59:06 -0500

Dear Peter,

There may be no general answer to your question. The situation you describe seems to me extraordinary. However, if you use FO method, it can happen. I personally do not rely on variability parameters estimated by the FO method. FOCE with interaction is much more reliable, but the run time in case of a complex model may become unacceptably long.

Best regards,

` `