From: Quyen Nguyen <qn@upcl.univ-lyon1.fr>
Subject: error message with a 3 comp model
Date: 10 Jun 1998 12:59:26 -0400
Using the new ADVAN11 TRANS4 (3 comp model) in the version 5, I have got the following message error:
OCCURS DURING SEARCH FOR ETA AT A NONZERO VALUE OF ETA
K21, OR K31 IS TOO CLOSE TO AN EIGENVALUE
0PROGRAM TERMINATED BY FNLETA
MESSAGE ISSUED FROM TABLE STEP
Does anyone know which parametrisation can avoid such error?
Thanks in advance,
Quyen Nguyen
error message with a 3 comp model
From: alison@c255.ucsf.EDU (ABoeckmann)
Subject: Re: error message with a 3 comp model
Date: 10 Jun 1998 20:22:44 -0400
Looks like the estimation step used the FO method, and the error message is coming from the POSTHOC step. For one or more individuals, the etas are being driven to values such that ADVAN11 is getting into numerical difficulties.
Without seeing the code or data, its hard to know what is wrong. (Even with the code and data, it may still be hard to know.) No particular TRANS is especially likely to avoid the error.
Here are a few suggestions, not in any particular order.
1) Look at the data for the individual(s) that give rise to the error. Maybe there is some problem with this data.
2) Use the NOABORT option of $ESTIM record.
3) Try all the trans routines: TRANS1, TRANS4, TRANS6
4) If the run times are not excessive, try running with the other advan routines, e.g., advan5 or 7 or 6 or 8 or 9. (This could be done with either NONMEM IV or V.) One of the others might avoid the difficulty, in particular 6 or 8 or 9.
5) Make sure that any "natural" bounds are built into the parameterization.
For example, with TRANS4, suppose you think of periph 1 as a faster compartment than periph 2: Q2 > Q3.
The population estimate may well have this relationship, but the search for individual posthoc estimates may reverse this for one or more individuals. The "flipping" of Q's and therefore of V's might lead to the numeric difficulty.
It might help to model so that this cannot happen, e.g.,
Q3=THETA(3) ...
Q2=Q3+THETA(2) ...
Make sure the thetas are bounded to be positive.