From: "Milly de Jonge" <Apmil@SLZ.NL>
Subject: [NMusers] local minimum
Date: Fri, 21 Dec 2001 08:10:56 +0100
Dear NONMEM users,
I have got a question concerning the minimization's of my NONMEM runs.
At the moment I am working on a two-compartment model of an iv given compound (control stream below). I am not using ADVAN3 but defined the PK-model with differential equations (ADVAN6), since I am planning to extent the model with non-linear pk-components (distribution, etc).
My data are log transformed and I am working with the FO method. On fitting the data nonmem indicates that the run has ended successfully (without any error messages). However my PRED vs DV plot looks bad and the parameter estimates are not as expected, indicating that nonmem has finished on a local minimum.
I have tried to avoid the local minimum by increasing TOL and SIG, but this did not work. Could someone tell me how to solve this problem?
Thanks in advance.
Milly de Jonge
Netherlands Cancer Institute
Department of Pharmacy and Pharmacology
Louwesweg 6
1066 EC Amsterdam
$PROBLEM 2 COMP MODEL
$INPUT DROP=IDUC ID DROP=LN DROP=STAT DAT1=DROP TIME AMT1=DROP
RAT1=DROP DV1=DROP CMT AMT RATE DVN=DROP DV KUUR DAY EVID MDV
$DATA DATARUN218.CSV IGNORE=#
$SUBROUTINES ADVAN6 TOL5
$MODEL COMP=(CENTR) COMP=(PER)
$PK
CL=THETA(1)*EXP(ETA(1))
V1=THETA(2)*EXP(ETA(2))
K12=THETA(3)
K21=THETA(4)
S1=V1
$DES
DADT(1)=-CL*A(1)/S1-K12*A(1)+K21*A(2)
DADT(2)=K12*A(1)-K21*A(2)
$ERROR
CALLFL=0
W=SQRT(THETA(5)**2)
IPRED=LOG(F)
IRES=DV-IPRED
IWRES=IRES/W
Y=IPRED+W*EPS(1)
$THETA
(0,35) ;1 Cl
(0,50) ;2 V1
(0,0.3) ;3 k12
(0,0.9) ;4 k21
(0,0.1) ;5 add
$OMEGA
0.04
0.04
$SIGMA
1 FIXED
$ESTIMATE PRINT=5 MAXEVAL=9999 POSTHOC NOABORT METHOD=0 MSFO=RUN218.MSF
$COVARIANCE
$TABLE ID TIME IPRED IWRES
NOPRINT ONEHEADER FILE=SDTAB218
$TABLE ID CMT CL V1 K12 K21
NOPRINT ONEHEADER FILE=PATAB218
local minimum
From: "Bachman, William" <bachmanw@globomax.com>
Subject: RE: [NMusers] local minimum
Date: Fri, 21 Dec 2001 09:19:55 -0500
Milly,
Changing TOL and SIG may only change the precision with which the
convergence "hones in" on a local minimum. Try significantly different
initial estimates in an attempt to give the convergence process a different
region of the parameter space to explore. (as you note, "successful" just
means a minimum, but not necessarily global - the user must assess results
obtained.)
Bill Bachman
From: "Atul" <bvatul@ufl.edu>
Subject: Re: [NMusers] local minimum
Date: Fri, 21 Dec 2001 10:20:19 -0500
Hello
If your data is not sufficiently dense then you might consider fixing the
estimates of k12 and k21 to prior estimates. I have seen this giving me much
more reliable parameter estimates. Your plots might also improve.
Atul
University of Florida
From: Jian-Feng Lu <jianfeng@gene.com>
Subject: Re: [NMusers] local minimum
Date: Fri, 21 Dec 2001 09:06:41 -0800
Milly,
You may use grid search to overcome your problem. The approach is very simple, just set your initial value of CL (for example) as (10,,100) instead of (0,35). "10" is the low bound of CL and "100" is the high bound of CL. NONMEM will find optimal initial value by itself from range 10 to 100. Hope this would overcome the problem of local minimum.
Jianfeng
--
Jianfeng Lu, Ph.D
Pharmacometrics
Division of Clinical & Experimental Pharmacology
Genentech Inc,
1 DNA Way, MS 70
South San Francisco, CA 94080
tel: (650) 225-5876
fax: (650) 225-6452
From: Lewis B Sheiner <lewis@c255.ucsf.edu>
Subject: Re: [NMusers] local minimum
Date: Fri, 21 Dec 2001 11:33:22 -0800
I don't think we know enough at this point to say that Millly found a local minimum: Milly, What does "my PRED vs DV plot looks bad and the parameter estimates are not as expected" mean? We'd need more detail - basic design, maybe some goodness of fit plots ...
LBS.
--
_/ _/ _/_/ _/_/_/ _/_/_/ Lewis B Sheiner, MD (lewis@c255.ucsf.edu)
_/ _/ _/ _/_ _/_/ Professor: Lab. Med., Biophmct. Sci., Med.
_/ _/ _/ _/ _/ Box 0626, UCSF, SF, CA, 94143-0626
_/_/ _/_/ _/_/_/ _/ 415-476-1965 (v), 415-476-2796 (fax)
From: "Piotrovskij, Vladimir [JanBe]" <VPIOTROV@janbe.jnj.com>
Subject: RE: [NMusers] local minimum
Date: Thu, 3 Jan 2002 11:08:33 +0100
Fitting PK models with an additive residual error to log-transformed concentrations has advantages over the common approach (non-transformed concentrations and proportional or additive+proportional residual error), however the problem of local minima and the associated problem of proper selection of initial values are more serious. The probable reason is that the likelihood surface in the parameter space is more flat compared to that coming from nontransformed concentrations. I would recommend the following: initially, thy to fit the model to nontransformed concentrations, examine the results, and if they are appropriate, use the estimates as initial values for the next run where the log-transformed concentrations are involved.
Best regards,
Vladimir
-----------------------------------------------------------------
Vladimir Piotrovsky, Ph.D.
Research Fellow, Advanced PK-PD Modeling & Simulation
Global Clinical Pharmacokinetics and Clinical Pharmacology (ext. 5463/151)
Johnson & Johnson Pharmaceutical Research & Development
Turnhoutseweg 30
B-2340 Beerse
Belgium