From: "Williams, Gene, HMR/US" <Gene.Williams@hmrag.com>
Subject: converting epsilon to CV
Date: 16 Mar 1998 18:07:28 -0500

I notice that when epsilon enters a model as follows:
Y = F*(1 + EPS(1))

NONMEM output includes the following:
ERROR IN LOG Y IS MODELED

I have been using the SIGMA obtained using the above coding as an estimate of the CV of the residual variability as follows:
CV = SQRT(SIGMA)

However, if "ERROR IN LOG Y IS MODELED", shouldn't I be using:
CV = SQRT(e**SIGMA - 1)

I appreciate that for small values of SIGMA, these expressions are equivalent. However, for larger values of SIGMA, which equation is correct?

Thanks,
Gene

Gene Williams, Ph.D.
Clinical PK/PD
Hoechst Marion Roussel
Kansas City, MO, USA

****

From: "Dr. Ziad Hussein" <ziadh@fs1.pa.man.ac.uk>
Subject: Re: converting epsilon to CV
Date: 17 Mar 1998 06:01:32 -0500

I have been using the same model for residual variability in many cases for the last 6 years and I never encountered the message "ERROR IN LOG Y IS MODELED" in the NONMEM output file.

It will be helpful to know how you modelled your parameters.

Regards,

Dr Ziad Hussein
Senior Research Fellow
Centre for Applied Pharmacokinetic Research
School of Pharmacy & Pharmaceutical Sciences
The University of Manchester
Oxford Road
Manchester M13 9PL
UK

****

From: alison@c255.ucsf.EDU (ABoeckmann)
Subject: converting epsilon to CV
Date: 17 Mar 1998 12:05:49 -0500

The message ERROR IN LOG Y IS MODELED comes from PREDPP. See Users Guide V (Introductory) Chapt 8, Section 3.2:

During data analysis NONMEM cannot distinguish between the CCV error model Y = F + F*EPS and the exponential error model Y = F * EXP(EPS), for which LOG (Y) = LOG (F) +EPS. By using the latter model and modelling the error in LOG (Y) rather than in Y, NM-TRAN enables PREDPP to achieve an improvement in run time.

Because the HH derivative (the derivative of log y with respect to epsilon) is a constant, it can be arranged that ERROR need only be called once, at the beginning of the problem, thus avoiding many subroutine calls. PREDPP itself forms the H derivative (the derivative of y with respect to epsilon). There is further discussion of this in chapter IV of Guide VI.

The question about what CV to use was addressed by Dr. Beal in messages to this mailinglist last year.

The orignal message, with some discussion, is attached for those who have not seen it, or who do not have access to the archive.

-- Alison Boeckmann