From: Leonid Gibiansky <*LGibiansky*>

Date: Wed, 24 Feb 2010 10:44:47 -0500

Andreas,

I think if you have only study-level data, you can treat each study as a

"subject", build meta-population dose-PD model in the usual way,

IPRED=Emax (or some other) function of DOSE

and modify the error part as follows: For each data point, you know the

mean value (DV), SD, and N. From SD and N you can get an estimate of

standard error of the mean as SE=SD/sqrt(N) (computed for each data

point and included into the data file). Then your error model would include

Y=IPRED+SE*EPS(1)

I am not sure whether you need to fix SIGMA

$SIGMA

1 FIXED ; for EPS(1)

thus assuming that all error in your model comes from the "assay", or

estimate it thus allowing for unexplained model misspecification and

extra error. I would try both ways to see the difference.

This is the simplest version that can be further improved (? or at

least, made more complicated) by adding the study effect on error (thus

accounting for possible differences in study populations;

Y=IPRED+SE*EXP(ETA(1))*EPS(1)), etc.

Leonid

--------------------------------------

Leonid Gibiansky, Ph.D.

President, QuantPharm LLC

web: www.quantpharm.com

e-mail: LGibiansky at quantpharm.com

tel: (301) 767 5566

alindauer-research

*>
*

*> Dear NMUSERS,
*

*>
*

*> I wanted to investigate the dose-response relationship (Emax model) of a
*

*> drug with NONMEM, based on data from literature (i.e. a meta-analysis).
*

*> However, I am not quite sure how to deal with the different levels of
*

*> random effects. Suppose I have 10 studies of different size where
*

*> different doses were given and the response is presented as average
*

*> change of a biomarker +/- standard deviation for each dose level. How
*

*> would I incorporate the standard deviation of the biomarker measurements
*

*> reported in each study for each dose level and how would I account for
*

*> the different number of patients in the study?
*

*> I would greatly appreciate your help, maybe with a NM-code snippet or
*

*> reference to a paper where something similar has been done.
*

*>
*

*> Thanks in advance, Andreas.
*

*> Ferrer
*

*> Andreas Lindauer
*

*> Pharmacokineticist
*

*> Pharmacokinetics and Metabolism
*

*> R&D Center. Ferrer Internacional S.A.
*

*> Juan de Sada 32, 08028 Barcelona
*

*> alindauer-research *

*> www.ferrergrupo.com
*

*>
*

*>
*

*> Recicla ¿Necesita imprimir este mensaje? Protejamos el medio
*

*> ambiente. Li cal imprimir aquest missatge? Protegim el medi ambient.
*

*> Do you need to print this message? Let's protect the environment.
*

*>
*

*>
*

*> Este mensaje, y en su caso, cualquier fichero anexo al mismo, puede
*

*> contener información confidencial, siendo para uso exclusivo del
*

*> destinatario, quedando prohibida su divulgación, copia o distribución a
*

*> terceros sin la autorización expresa del remitente. Si Vd. ha recibido
*

*> este mensaje erróneamente, se ruega lo notifique al remitente y proceda
*

*> a su borrado. Gracias por su colaboración.
*

*>
*

*> This message and its annexed files may contain confidential information
*

*> which is exclusively for the use of the addressee. It is strictly
*

*> forbidden to distribute copies to third parties without the explicit
*

*> permission of the sender. If you receive this message by mistake, please
*

*> notify it to the sender and make sure to delete it. Thank you for your
*

*> kind cooperation.
*

*>
*

*> *

Received on Wed Feb 24 2010 - 10:44:47 EST

Date: Wed, 24 Feb 2010 10:44:47 -0500

Andreas,

I think if you have only study-level data, you can treat each study as a

"subject", build meta-population dose-PD model in the usual way,

IPRED=Emax (or some other) function of DOSE

and modify the error part as follows: For each data point, you know the

mean value (DV), SD, and N. From SD and N you can get an estimate of

standard error of the mean as SE=SD/sqrt(N) (computed for each data

point and included into the data file). Then your error model would include

Y=IPRED+SE*EPS(1)

I am not sure whether you need to fix SIGMA

$SIGMA

1 FIXED ; for EPS(1)

thus assuming that all error in your model comes from the "assay", or

estimate it thus allowing for unexplained model misspecification and

extra error. I would try both ways to see the difference.

This is the simplest version that can be further improved (? or at

least, made more complicated) by adding the study effect on error (thus

accounting for possible differences in study populations;

Y=IPRED+SE*EXP(ETA(1))*EPS(1)), etc.

Leonid

--------------------------------------

Leonid Gibiansky, Ph.D.

President, QuantPharm LLC

web: www.quantpharm.com

e-mail: LGibiansky at quantpharm.com

tel: (301) 767 5566

alindauer-research

Received on Wed Feb 24 2010 - 10:44:47 EST