From: Rik Schoemaker - RS@chdr.nl
Subject: [NMusers] Fortran overflow
Date: 2/12/2004 3:18 AM

Dear all,

I'm experiencing a nasty problem in running NONMEM. I have a
model describing two-compartment parent PK with first order
absorption. The parent is metabolised (and measured) and the
metabolite is subsequently excreted. I run variants of this
with the FO method and sometimes get convergence. At other
times it converges but is then terminated by a Fortran overflow
error message (forrtl error (72): floating overflow), leaving
me with no results. If I try FOCE, it terminates with the same
message during the first iteration. Could there be someting
with empirical Bayes estimate generation (I request POSTHOC
with FO) that kills the program...?

Any clues?

Rik Schoemaker

p.s. I've tried Leonid's and Toufigh's trick by replacing
the $ERROR block below with
  DRG1 = A(2)/S2
  DRG2 = A(4)/S4
  IF (CMT.EQ.4) THEN
  Y=DRG2*EXP(ERR(2))
  ELSE
  Y=DRG1*EXP(ERR(1))
  ENDIF 
and the result is the same....


$PROB   DRUG1 and DRUG2; M_G17
$INPUT  ID WK TIME DV AMT MDV CMT TRTI
$DATA   c:\data\DATNSS2.NM
$SUBROUTINES ADVAN5
$MODEL
 COMP=(DEPOT,INITIALOFF, DEFDOSE)
 COMP=(CENTRAL)
 COMP=(PERID)
 COMP=(DRUG2)
$PK
 CL  = THETA(1)*EXP(ETA(1)) 
 TABS= THETA(2)*EXP(ETA(2))
 Q   = THETA(3)*EXP(ETA(3))
 S2  = THETA(4)*EXP(ETA(4))  
 VSS = THETA(5)*EXP(ETA(5))  
 S4  = THETA(6)*EXP(ETA(6))
 K40 = THETA(7)*EXP(ETA(7))
 K12 = 0.693/TABS
 S3  = VSS-S2
 K24 = CL/S2
 K23 = Q/S2
 K32 = Q/S3
 CLG = K40*S4
$ERROR
  PREDI = F
  IF (CMT.EQ.4) THEN
  Y=PREDI*EXP(ERR(2))
  ELSE
  Y=PREDI*EXP(ERR(1))
  ENDIF
$THETA (1,15)(0.01,1,10) (1,20) (2,13,200) (50,1000) (0.1,5.5) 
(0.01,0.2) 
$OMEGA .2 2 .5 2 .2 1 .2 
$SIGMA 0.1 0.01
$ESTIMATION MAXEVAL=9999 PRINT=1 NOABORT POSTHOC METHOD=0
$COV
$TABLE ID WK TIME DV PREDI CMT CL Q S2 VSS  TABS CLG K40 S4 TRTI
FILE = M_G17.DAT
NOHEADER NOPRINT NOAPPEND
$TABLE ID WK CL Q S2 VSS TABS CLG K40 S4 TRTI FILE = M_G17.PAR
NOHEADER NOPRINT NOAPPEND FIRSTONLY



Data for the first subject (it's a multiple dose experiment and
you're looking at 2 weeks of dosing):


1 12 0.00 . 2291 1 1 1
1 12 12.00 . 2291 1 1 1
1 12 24.00 . 2291 1 1 1
1 12 36.00 . 2291 1 1 1
1 12 48.00 . 2291 1 1 1
1 12 60.00 . 2291 1 1 1
1 12 72.00 . 2291 1 1 1
1 12 84.00 . 2291 1 1 1
1 12 96.00 . 2291 1 1 1
1 12 108.00 . 2291 1 1 1
1 12 120.00 . 2291 1 1 1
1 12 132.00 . 2291 1 1 1
1 12 144.00 . 2291 1 1 1
1 12 156.00 . 2291 1 1 1
1 12 168.00 . 2291 1 1 1
1 12 180.00 . 2291 1 1 1
1 12 192.00 . 2291 1 1 1
1 12 204.00 . 2291 1 1 1
1 12 216.00 . 2291 1 1 1
1 12 228.00 . 2291 1 1 1
1 12 240.00 . 2291 1 1 1
1 12 252.00 . 2291 1 1 1
1 12 264.00 . 2291 1 1 1
1 12 276.00 . 2291 1 1 1
1 12 288.00 . 2291 1 1 1
1 12 300.00 . 2291 1 1 1
1 12 312.00 . 2291 1 1 1
1 12 324.00 . 2291 1 1 1
1 12 336.00 . 2291 1 1 1
1 12 336.00 9.3662191695 . 0 2 1
1 12 336.00 98.696799404 . 0 4 1
1 12 336.20 . . 1 2 1
1 12 336.20 . . 1 4 1
1 12 336.40 . . 1 2 1
1 12 336.40 . . 1 4 1
1 12 336.60 . . 1 2 1
1 12 336.60 . . 1 4 1
1 12 336.80 . . 1 2 1
1 12 336.80 . . 1 4 1
1 12 337.00 29.971901342 . 0 2 1
1 12 337.00 112.79634218 . 0 4 1
1 12 337.20 . . 1 2 1
1 12 337.20 . . 1 4 1
1 12 337.40 . . 1 2 1
1 12 337.40 . . 1 4 1
1 12 337.60 . . 1 2 1
1 12 337.60 . . 1 4 1
1 12 337.80 . . 1 2 1
1 12 337.80 . . 1 4 1
1 12 338.00 14.985950671 . 0 2 1
1 12 338.00 161.13763168 . 0 4 1
1 12 338.50 . . 1 2 1
1 12 338.50 . . 1 4 1
1 12 339.00 10.615048392 . 0 2 1
1 12 339.00 193.36515802 . 0 4 1
1 12 339.50 . . 1 2 1
1 12 339.50 . . 1 4 1
1 12 340.00 7.4929753356 . 0 2 1
1 12 340.00 181.27983564 . 0 4 1
1 12 341.00 . . 1 2 1
1 12 341.00 . . 1 4 1
1 12 342.00 4.6831095848 . 0 2 1
1 12 342.00 161.13763168 . 0 4 1
1 12 344.00 . . 1 2 1
1 12 344.00 . . 1 4 1
1 12 346.00 . . 1 2 1
1 12 346.00 . . 1 4 1
1 12 348.00 . . 1 2 1
1 12 348.00 . . 1 4 1

Rik C. Schoemaker, PhD
Biostatistician - pharmacokineticist

Centre for Human Drug Research (CHDR)
Zernikedreef 10, 2333CL Leiden, Netherlands
tel +31 (0) 71 5246417; fax +31 (0) 71 5246499
email rs@chdr.nl; http://www.chdr.nl
_______________________________________________________

From: "Bhattaram, Atul" - BhattaramA@cder.fda.gov
Subject: RE: [NMusers] Fortran overflow
Date: 2/12/2004 9:23 AM

Hello Schoemaker

Did you try to analyze the data by first analysing the parent drug and
fixing them prior to adding the metabolite data? 


Venkatesh Atul Bhattaram
CDER, FDA.
_______________________________________________________

From: Leonid Gibiansky - lgibiansky@emmes.com
Subject: Re: [NMusers] Fortran overflow
Date: 2/12/2004 9:31 AM

As a first try:
With overflow error, there might be a division by zero. If you look
on your control stream, one place is
S3  = VSS-S2
K32 = Q/S3
Here S3 can be anything. You may try
IF(S3.LE.0.001) EXIT
K32 = Q/S3

Alternative  is to use
S2  = THETA(4)*EXP(ETA(4))
S3  = THETA(5)*EXP(ETA(5))
VSS = S2 +S3
Leonid 
_______________________________________________________

From: Ekaterina Gibiansky - GibianskyE@guilfordpharm.com
Subject: Re: [NMusers] Fortran overflow
Date: 2/12/2004 9:34 AM

Ric,

in my experience,  the solution to overflow problem (when everything
else was exhausted including EXIT statements) was to switch to G77
fortran from Visual Fortran (Compaq, Digital, etc whatever it was called
at the time). It always worked for me.

Katya


Ekaterina Gibiansky, PhD
Head, Pharmacometrics & Principal Scientist
Guilford Pharmaceuticals Inc
Phone: (410)-631-6828
Fax:   (410)-631-6828
E-mail: gibianskye@guilfordpharm.com
_______________________________________________________

From: Diane R Mould - drmould@attglobal.net
Subject: Re: [NMusers] Fortran overflow
Date: 2/12/2004 10:00 AM

All

Katya's findings (that moving to another compiler can sometimes correct
problems) seem to be consistent with what Peter Bonate's findings when
he tested Nonmem performance on different processors and compilers.  If
I recall correctly, Pete found that the results obtained from an
analysis seem to be more likely to differ using different compilers as
model stability decreases.  I think that a very stable model is
generally robust to different compilers.  Perhaps Pete can comment on
that?

Diane 
_______________________________________________________

From: Rik Schoemaker - RS@chdr.nl
Subject: Re: [NMusers] Fortran overflow
Date: 2/16/2004 4:36 AM

Dear Leonid,

I like it when simple suggestions are the best! It certainly had to do with
estimating VSS; by switching to estimating S2 and S3 the problem was solved. Thanks!

Rik
_______________________________________________________