gradient/significant digit/Cov step
From: "Li, Susie (Xiujiang)" xiujiang_li@merck.com
Subject: [NMusers] gradient/significant digit/Cov step
Date: Mon, April 18, 2005 11:15 am
Dear NONMEM users,
For a dataset and model combination (FOCE INTERACTION, no covariates), I
noticed some components of the gradient vector are around 10, although
convergence is achieved. By changing significant digits from 3 to 5 , I was
able to bring the gradients down to < 0.1 while all of the parameter
estimates and objective function remain almost the same. It was pretty
stable with respect to different initial guesses.
When building covariate model, I observed the same behavior. However, the
COV step failed this time at SIG=5 although its parameters didn't change
from SIG=3 to 5 either. Could someone explain to me why almost identical
parameters and improved gradients lead to convergence failure? Can I blame
it to NONMEM numerical instability? Can I accept the result at SIG=3 (where
some gradients are ~10)?
No COV problem was encountered with FO method.
Thanks,
Susie Li
_______________________________________________________