I began to use NONMEM to perform PK modeling. I have a question about the
REPEAT option. According to the manual with the REPEAT option, parameter
estimation will be repeated with initial estimates using the final estimates
from the first search. My questions are:
1. Do I use the option REPEAT regularly during the modeling
development process?
2. What does the REPEAT option provide in the output file? How do I
interpret the results?
Thank you.
use of REPEAT option
Hi Pharavee,
The REPEAT option on the $ESTIMATION record causes the following:
1) the matrix that is used to guide the parameter search is rescaled using the
final estimates from the first search and
2) a new parameter search is conducted starting with the previous parameter
estimates and the rescaled matrix.
The idea was that if the first search yielded a local minimum then the second
search might find a lower minimum. However, experience demonstrated that the
repeated search usually failed to move from the original minimum even if the
original was a local minimum. Consequently, the option has received very
little attention or use.
A commonly used strategy to try to defend against a local minimum is to use
multiple runs of the same model but with different sets of initial estimates.
If one can obtain the same minimum objective function value for several of
these runs then this result is taken as supporting evidence that this objective
function value could be, but not guaranteed to be, a global minimum. See
section I.36 of NONMEM USERS GUIDE INTRODUCTION TO NONMEM 7.2.0 (nm720.pdf) for
implementation of this latter approach.
Tom
Quoted reply history
________________________________
From: [email protected] [mailto:[email protected]] On
Behalf Of Pharavee Jaiprasart
Sent: Friday, March 15, 2013 12:04 PM
To: [email protected]
Subject: [NMusers] use of REPEAT option
I began to use NONMEM to perform PK modeling. I have a question about the
REPEAT option. According to the manual with the REPEAT option, parameter
estimation will be repeated with initial estimates using the final estimates
from the first search. My questions are:
1. Do I use the option REPEAT regularly during the modeling development
process?
2. What does the REPEAT option provide in the output file? How do I
interpret the results?
Thank you.
Tom,
Thanks for decribing your experience with the REPEAT option.
I have not used it but do something approximately the same by putting the final parameter estimates reported in the NONMEM listing file into the control stream and then rerunning the problem. Because the estimates used for the next run are only approximate (3 sig digs) I usually find that the new run will do something different and often find a lower OFV. I typically repeat this process a few times until the OFV stops improving (or worsens). Of course I don't do this for every run but I do like to do this when I think I am close to extracting all I can from the data. This is usually well after the point when the covariance step starts to fail :-)
For Wings for NONMEM < http://wfn.sourceforge.net/ > users the process of transferring the values from the NONMEM listing to the control stream file is simple - use the nmctl command.
Best wishes,
Nick
Quoted reply history
On 19/03/2013 9:53 a.m., Ludden, Thomas wrote:
> Hi Pharavee,
> The REPEAT option on the $ESTIMATION record causes the following:
>
> 1) the matrix that is used to guide the parameter search is rescaled using the final estimates from the first search and 2) a new parameter search is conducted starting with the previous parameter estimates and the rescaled matrix. The idea was that if the first search yielded a local minimum then the second search might find a lower minimum. However, experience demonstrated that the repeated search usually failed to move from the original minimum even if the original was a local minimum. Consequently, the option has received very little attention or use. A commonly used strategy to try to defend against a local minimum is to use multiple runs of the same model but with different sets of initial estimates. If one can obtain the same minimum objective function value for several of these runs then this result is taken as supporting evidence that this objective function value could be, but not guaranteed to be, a global minimum. See section I.36 of *NONMEM USERS GUIDE INTRODUCTION TO NONMEM 7.2.0 (*nm720.pdf) for implementation of this latter approach.
>
> Tom
> ------------------------------------------------------------------------
>
> *From:* [email protected] [ mailto: [email protected] ] *On Behalf Of *Pharavee Jaiprasart
>
> *Sent:* Friday, March 15, 2013 12:04 PM
> *To:* [email protected]
> *Subject:* [NMusers] use of REPEAT option
>
> I began to use NONMEM to perform PK modeling. I have a question about the REPEAT option. According to the manual with the REPEAT option, parameter estimation will be repeated with initial estimates using the final estimates from the first search. My questions are:
>
> 1.Do I use the option REPEAT regularly during the modeling development process?
>
> 2.What does the REPEAT option provide in the output file? How do I interpret the results?
>
> Thank you.
>
>