From: Daniel Corrado <>
Subject: [NMusers] Parameter %RSE vs ETA %RSE
Date: 7/17/2003 1:57 PM

The results of my Nonmem modeling of a sparse data set
are given below

* 1 comp model with CL, V and Ka
* Successful minimization
* Covariance matrix obtained
* goot obs vs pred and obs vs ipred plots

THETA1 = CL = 1.66e+003; %RSE = 3.2%
THETA2 = V = 2.04e+003; %RSE = 32.9% 
THETA3 = KA = 0.196; %RSE = 10.4%
ETA1 = 0.0230; %RSE= 65.2% (CI bound zero)     
ETA2=1.36; %RSE = 39.1%
ETA3=6.70e-008; %RSE 1.37e+008%(CI bound zero)

A colleague of mine indicated that the large %RSE
associated with ETA3 indicats that ETA3 and by
association THETA3(KA) estimations were not very good.
He suggested that I fix KA prior to modeling and try
to estimate only CL and V. My contention is that the
value of KA obtained is in the ball park of that seen
from dense data studies. Also %RSE associated with
THETA3 is  small. The value of ETA3 is so small that
the high %RSE associated with it still makes it
insignificant when compared to THETA3. Also the %CV
for ETA 3 = 0.03% despite the large %RSE. Hence the
value of KA and ETA3 are ok.

Could someone let me know if my line of reasoning if
totally off.


From: Leonid Gibiansky <>
Subject: Re: [NMusers] Parameter %RSE vs ETA %RSE
Date: 7/17/2003 2:28 PM

You should remove eta3 from the model since it is so small
(effectively, zero). You will not see any difference in OF or fit.
Your eta1 is also rather small, and has large RSE. After you
remove eta3, try to remove eta1 as well, you may found out that
there is no difference (with the original model) either. One the
other hand, eta2 is too large (do you use V=THETA(2) * EXP(ETA(2)) model ?).
Try to use FOCE and study distribution of eta2: variance of eta2 is too
large to accept the model without trying to find a better model or
an explanation why this variance is so large.

Good luck


From: Sam Liao <>
Subject: RE: [NMusers] Parameter %RSE vs ETA %RSE
Date: 7/17/2003 3:50 PM

Hi Daniel:

Based on CL and V estimates, the elimination half-life is 0.85 hr while
the absorption half-life is 3.5 hours.  How is this compared with those
half-lives obtained from intensive PK sampling? If there is a flip-flop,
V estimate will be affected too.  It may be a good ideal to fix KA to
avoid flip-flop, if the sparse data do not have enough data in the
absorption phase. 

Best regards,
Sam Liao