From:"Nandy, Partha" 
Subject:[NMusers] Error Message from NONMEM :: MINIMIZATION TERMINATED (ERROR=136)
Date:Wed, 18 Sep 2002 15:19:15 -0400

Hi,

I am sure NONMEM users have encountered this problem before.  In the past, I
have been able to by-pass this problem by providing different initial
estimates for THETAs, OMEGAs and SIGMAs.

Now, I am encountering this problem again and my previous methods of solving
the problem is not working.  Currently, I am running a$PRED routine.  This
routine runs fine when Method=0 is implemented, but I am getting this
recurrent message, shown below, when I am trying Method=1 with interaction.
I have tried different initial estimates and also giving better initial
estimates (results from method=0 run, offset by ~10-20% etc..), deleting
records with DV=0, but I keep getting this message back.

I appreciate any help you can provide.

Thanks in advance.

Partha
Partha Nandy, Ph.D.
Clinical Pharmacology
partha.nandy@pharma.com 

----------------------------------------------------------------------------
----------------------------------------------------------------------------
----------------------------------------- 
0MINIMIZATION TERMINATED
 DUE TO PROXIMITY OF LAST ITERATION EST. TO A VALUE
 AT WHICH THE OBJ. FUNC. IS INFINITE (ERROR=136)
0AT THE LAST COMPUTED INFINITE VALUE OF THE OBJ. FUNCT.:
 ERROR IN NCONTR WITH INDIVIDUAL   74   ID=0.74000000E+02
 NUMERICAL HESSIAN OF OBJ. FUNC. FOR COMPUTING CONDITIONAL ESTIMATE
 IS NON POSITIVE DEFINITE
 THETA=
  4.49E+01   1.03E+02   1.19E+01   8.23E+00   1.43E+00   4.25E-01   2.26E-01
2.88E+00   4.21E+00   7.05E-01
  2.30E+00   4.32E-01   2.56E+00   1.64E+00   1.24E+00   8.85E-01
 
 NO. OF FUNCTION EVALUATIONS USED: 1084
 NO. OF SIG. DIGITS UNREPORTABLE

 ETABAR IS THE ARITHMETIC MEAN OF THE ETA-ESTIMATES,
 AND THE P-VALUE IS GIVEN FOR THE NULL HYPOTHESIS THAT THE TRUE MEAN IS 0.

 ETABAR:  -0.49E+00  0.28E+00 -0.12E-01 -0.13E+00 -0.19E+00  0.21E+00

 P VAL.:   0.19-205  0.72E-71  0.38E-01  0.17E-01  0.51E-03  0.16E-05

Partha
Ext # 8320
Fax: 203-588-6328
___________________________________

From:Nick Holford 
Subject:Re: [NMusers] Error Message from NONMEM :: MINIMIZATION TERMINATED (ERROR=136)
Date:Thu, 19 Sep 2002 09:05:45 +1200

Partha,

The standard (and IMHO usually unhelpful) suggestion it to check the structure of
your model for errors. This may also involve considering posteriori non-identifiability because the
data does not really let you estimate some parameter such as between subject variability. This
leads to a KISS (Keep It Simple Stupid) approach to model building that may throw
the baby out with the bathwater.

I find that this typically happens when the model building is getting really interesting and I am
learning something new about the system I am trying to describe. I judge the usefulness of the model
(remembering George Box) by its ability to describe the data rather than some arbitrary numerical
criterion such as significant digits e.g. I am currently working on a model describing the placebo
response in depression. Only the simplest model converges and runs the covariance step.
More complex and informative models visibly fit the data better when I look at time courses of
HAMD score but typically fail to minimize (although I can sometimes get at least 3 sig digs by
using SIGDIG=4 on $ESTIMATION).

I rationalize this by saying that learning via modelling happens at the bleeding edge of the data.
We are trying to discover weak but potentially important signals (like the Hubble telescope recent
discovery of medium sized black holes http://oposite.stsci.edu/pubinfo/pr/2002/18/) that are buried
in the data. Confirming the obvious stuff (like sun, moon, planets) is visible without NONMEM -- that's
what statisticians do in their analyses. It is therefore no surprise that the criteria that give
statisticians a warm and fuzzy feeling (like asymptotic SEs) are not always discernible (or
believable) when trying to extract meaning from experiments not expicitly designed to discover new things.

In the particular example you show below the ETABAR estimates do seem to be pathologically different
from zero so trying a different between subject variability model may help e.g. use (1+ETA) instead
of exp(ETA) if it is possible that the parameter can have different signs in different individuals.

Nick

Nick Holford, Divn Pharmacology & Clinical Pharmacology
University of Auckland, 85 Park Rd, Private Bag 92019, Auckland, New Zealand
email:n.holford@auckland.ac.nz tel:+64(9)373-7599x6730 fax:373-7556
http://www.health.auckland.ac.nz/pharmacology/staff/nholford/
___________________________________