The most common reason for zero gradient (if the code is correct and
does not include parameters with no influence on objective function) is
the bound on the parameter estimate. Even if you do not specify the
bounds, Nonmem imposes internal bounds (initial value * / by 100 or
something similar). I would try to run the initial model with
NOTHETABOUNDTEST NOOMEGABOUNDTEST NOSIGMABOUNDTEST
options.
Leonid
--------------------------------------
Leonid Gibiansky, Ph.D.
President, QuantPharm LLC
web: www.quantpharm.com
e-mail: LGibiansky at quantpharm.com
tel: (301) 767 5566
On 2/11/2013 8:32 PM, Nick Holford wrote:
Siwei,
A final gradient of zero is not necessarily pathological. It just means
that the fit cannot be improved by changing that parameter. As usual you
need to decide if the estimate of the residual error parameter is
plausible of not. The gradient cannot tell you that.
A zero additive error will always cause an error if you have a predicted
conc of zero. This is because the likelihood involves a division by the
residual error. If this is zero then the "computer says no
<http://en.wikipedia.org/wiki/Carol_Beer>". I assume that is what you
mean when you say "NM would not run".
Nick
On 12/02/2013 12:53 p.m., siwei Dai wrote:
Hi, NM users:
I have a typical 2-compartment model that can describe my data quite
well, except the final gradient for the additive residual error is
'0'. I therefore fix the additive residual error to '0', but then NM
would not run. I tried different initial estimations for other
parameters but the additive residual error seems to be the one that
decide whether NM will run. Can anyone tell me why this would happen
and how to solve it?
Thank you very much in advance for your help.
Siwei
