More information would be helpful.
How many records in the data set? How many individuals? What is the
maximum number of observations per individual? What is the maximum
number of data records per individual?
Does this happen with other problems on the VAX, or just this one
data set?
When this same problem is run on the Sun, is the same minimum reached?
Are the final value of the objective function and final parameter estimates
the same (to the no. of significant digits printed)?
My guess is that someone modified NONMEM source on the VAX and made an
error. The changes might have been to raise the maximum number of
observations per individual from its default value 50 (e.g., not all
instances of "PARAMETER NO=50" were changed), or to insert some
additional WRITE statements (e.g., in ZXMIN1) to obtain intermediate
output from the estimation step. The NONMEM source on the VAX should
be carefully compared with the distributed code.
Analysis of Steady State Data
(S=Maier%S=Maier%G=Gary%I=A%O=STERLING%OU1=PRD1%KODAK@mcimail.com)
- 5 Dec 94
I would like to describe a problem that I ran into when using
NONMEM on our VAX computer system.
We recently completed a phase III study in which plasma
concentrations as well as PD measurements were collected under
steady-state conditions. A nonmem data set was created assuming
these conditions with SS values of 1 and an interdose interval of
24 hours. Advan4 and trans 3 were utilized base upon prior
information.
To my surprise, I noticed that once the estimation step was
invoked, all converted parameter estimates and gradients were zero
and did not change throughout the iteration procedure, despite
large numerical reductions in the objective function. When this
data set and associated nmtran records were transferred to a SUN
workstation no problems were encountered.
Has anyone out there ever experienced similar difficulties?