RE: Weighting observations
Hi Robert
Given the function -2log(L) =nlog(2*pi)+sum[log(sigmai^2) + (Yi
-Yihat)^2/(sigmai^2)],
it appears by scaling up/down the DV uniformly for certain compartment (e.g. by
10 fold), you can significantly increase/decrease the log(sigmai^2) part of the
objective function. I tried it on my model and it seems to work. Maybe this is
one option to go.
I also tried Leonid's method of forcing to down the sigma by using a scalar.
It does not seem to change the objective function probably because it does not
change the overall sigmai^2 or (Yi -Yihat)^2.
Just my 2 cents.
Thank you and best regards!
Penny (Peijuan) Zhu, PhD
Oncology Clinical Pharmacology
337-B07
1 Health Plz, East Hanover, NJ
Tel: 862-778-7845
Cell: 862-926-9079
Quoted reply history
-----Original Message-----
From: [email protected] [mailto:[email protected]] On
Behalf Of Leonid Gibiansky
Sent: Friday, March 21, 2014 10:47 AM
To: Robert Kalicki; [email protected]
Subject: Re: [NMusers] Weighting observations
Hi Robert,
I have not seen it done, but you can try to enforce the weighting by playing
with the error model. For example, if you have 2 DVs, PK and PD, run the model
without any weighting and with the proportional error for both variables:
$ERROR
SDPK=THETA(1)
SDPD=THETA(2)
IF(PK) Y=DVPK+SDPK*EPS(1)
IF(PD) Y=DVPD+SDPD*EPS(2)
$SIGMA
1 FIXED
1 FIXED
and got THETA(1)=0.1 and THETA(2)=0.2
you may try
SDPK=THETA(1)
SDPD=4*THETA(1)
effectively forcing the model to downgrade the PD observations.
Yet another possibility is to fit that main variable first, fix the parameters
related to that variable (population or even population and individual), and
then add data for the second variable.
Or may be you can estimate the model with only the first variable, and then fix
the residual error on the estimated value: this would also force the model to
maintain the quality of the first-variable fit.
More mild control is to fit that main variable first and then add priors to
the model, more informative for the parameters that describe the first variable
and uninformative for the second one.
If you implement any of these ideas, please let the group know how it worked
(or not). This discussion is somewhat similar to the question whether to
perform sequential or simultaneous fit of the PK and PD data, so you may look
on the references, posters, discussion topics related to this question for
additional ideas.
Thanks
Leonid
--------------------------------------
Leonid Gibiansky, Ph.D.
President, QuantPharm LLC
web: www.quantpharm.com
e-mail: LGibiansky at quantpharm.com
tel: (301) 767 5566
On 3/21/2014 4:51 AM, Robert Kalicki wrote:
> Dear NMusers,
>
> I am modelling the effect of a given intervention on a multimodal
> outcome (ex. PK of a parent drug and its metabolites or effect of a
> given drug on the blood pressure, heart rate and urine output). For
> each observed variable I can define an observation compartment and
> model the error. This works fine. However, for some physiological
> reasons, I may want to give more weight to on an observed variable
> which I believe to be the driving force.
>
> How can I do this in NONMEM? Duplicating the observation of interest
> in my model would be one possibility however this would increase the
> run time which is already very long.
>
> Many thanks in advance.
>
> Robert
>
> ______________________________________________
>
> Robert, Kalicki, /Dr. med. /
>
> OA Neph
>
> INSELSPITAL, Universitätsspital Bern
>
> DURN
>
> Klinik und Poliklinik für Nephrologie, Hypertonie und Klinische
> Pharmakologie
>
> Freiburgstrasse 15, KiKl, G, 524
>
> *3010 Bern*
>
> /Telefon: +41 (0)31 632 31 44/
>
> /Mobile: +41 (0)79 239 98 79/
>
> /Fax: +41 (0)31 632 97 34/
>
> E-Mail: [email protected]
> <mailto:[email protected]>
>
> [email protected] <mailto:[email protected]>
>
> www.insel.ch http://www.insel.ch/
>
> No virus found in this message.
> Checked by AVG - www.avg.com http://www.avg.com
> Version: 2014.0.4336 / Virus Database: 3722/7223 - Release Date:
> 03/20/14
>