From: "Sreenivasa Rao Vanapalli" <rao@pharm.cpb.uokhsc.edu>
Subject: Residual Error
Date: Wed, 21 Mar 2001 12:19:38 -0800

Hello NMUSERS

I have a question regarding the selection of intrasubject residual error model. When chosing this error model what criteria we use? Just decrease in OBJ function value? What if OBJ function value doesn't change much, but other diagnostic tools like WRES vs time, CP, Covariates etc. Any suggestions?

Sreenivasa Rao Vanapalli
College of Pharmacy,OUHSC
1110 N.Stonewall
Oklahoma City, OK-73117
Tel:405-271-6484 Extn:47232 (office)
:405-521-0826 (Home)
Fax:405-271-7505(Office)


*****


From: Ulrika Wahlby <ulrika.wahlby@farmbio.uu.se>
Subject: Re: Residual Error
Date: Thu, 22 Mar 2001 09:41:21 +0100

Dear Sreenivasa,

That the OFV does not change much, but other diagnostic tools point to a better fit does not seem as likely as the opposite. From what we (Mats Karlsson, Niclas Jonsson, René Bouw and myself) have seen in simulation work a large drop in OFV, that would be assumed to be significant, can often be seen without any improvement in goodness of fit plots. The use of the OFV as the sole criterion for selection of residual error model in many instances leads to a large risk of a false positive error - if the FO method is used. However, if the FOCE method with interaction is used the likelihood ratio test seems to work all right - i.e. a drop in OFV of 3.84 gives a significance level of approx. <0.05. But surely, relying only on the difference in OFV would not be satisfactory, one would like to see improvement in goodness of fit plots as well!

Regards,

Ulrika Wählby, MSc
Div. of Pharmacokinetics and Drug Therapy
Uppsala University
Box 591 S-751 24 Uppsala, Sweden
Phone: +46 18 471 4305 Fax: +46 18 471 4003