Subject: [NMusers] increasing ETA estimates with adding full block
Date:8/18/2003 2:29 AM

Dear NONMEM users,

In having built a two compartment model with first order 
absorption, lag time and several covariates on V2 and CL, I ran into 
the following problem. When adding a full BLOCK in the OMEGA 
matrix to the final model the objectiove function significantly 
decreases (more than 100 points). However, the estimates for ETA 
(IIV and BOV) clearly increase. For example the ETA for KA 
increases from 111% to 126% and the ETA for V2 from 82% to 
101%. So I seem to explain less inter-subject variability with 
adding a full block. Does someone has an explanation for this and 
knows how to deal with this?

Thanks in advance,
Reinier van Hest
Research Pharmacist
Department of Pharmacy

Erasmus MC University Medical Centre Rotterdam
P.O. Box 2040
3000 CA Rotterdam

Tel: +31 10 4633202
Fax: +31 10 4636605

Subject: RE: [NMusers] increasing ETA estimates with adding full block
Date: 8/19/2003 8:36 AM


You really haven't given us enough information, but, based on what you have
given us, some thoughts:

1. if you haven't used FOCE with interaction, do so.
2. if you have, I suggest that you just have a situation where a full omega
is not warranted.  as you have said yourself, the additional etas are NOT
describing more of the variability (and the decrease in objective function
is just due to adding more parameters to the model).
3. your variance models may not be appropriate for your data.

William J. Bachman, Ph.D.
GloboMax LLC
7250 Parkway Drive, Suite 430
Hanover, MD 21076


Subject: RE: [NMusers] increasing ETA estimates with adding full block
Date: 8/19/2003 10:01 AM


We really shouldn't think of estimating off-diagonal elements of omega as
reducing variability in the diagonal elements of omega like we do when
including fixed effects.  It is hard to know with the limited information
you provided whether a full block omega is warranted.  How one partitions
the variability in omega and sigma to descibe the total variability in the
data may or may not be important based on the intended use of the model.  My
own approach to estimation of omega is to fit the fullest omega that can be
supported by the data (i.e., avoiding over-parameterization or
ill-conditioning) even if some of the off-diagnonal elements correspond to
correlations near zero.  In your case a 100 point drop in OFV would suggest
that one or more off-diagonal elements correspond to correlations that are
different from zero.  If you plan to use your model for simulating
individual responses those off-diagonal elements may be important.  Ignoring
those correlations by fitting a diagonal omega may result in unrealistic
combinations of the individual parameters when you conduct simulations using
the diagonal omega model fit.