What does nonmem do after iteration stops?

From: "Eleveld, DJ" d.j.eleveld@anest.umcg.nl Subject: [NMusers] What does nonmem do after iteration stops? Date: Fri, June 3, 2005 6:28 am Hi Everyone, I've noticed something curious: I use the PRINT=1 option in the $ESTIMATION to see the progress of the model fitting. With a fitting that I am doing (PD mdoel fitting involving about 3000 data points) the iterations stop after about 20-30 minutes and then nothing is printed on the screen for several hours. Apparently nonmem is calculating something after convergence that takes a very long time. The problem does have 2 tables but I cant really imagine that these would require so much computing time. What does nonmem do after finishing iterations that could possibly require so much computation? Is there a way for me to avoid this computation and still see the results of the estimation? If its useful my control file is at the end of this message. Thanks very much, Doug Eleveld ------------------------------------ $PROB Potentiation fitting $DATA potpd_.prn $INPUT ID TIME CPLA DV MDV AMT RATE $SUBROUTINES ADVAN9 TOL=6 $ABBREVIATED COMRES=1 $MODEL NCOMPARTMENTS=2 NPARAMETERS=6 COMP(POTENT NOOFF) COMP(EFFECT NOOFF NODOSE) $PK CALLFL=0 KEO=THETA(1)*EXP(ETA(1)) EC50=THETA(2)*EXP(ETA(2)) GAMM=THETA(3)*EXP(ETA(3)) POTR=THETA(4)*EXP(ETA(4)) POTK=THETA(5)*EXP(ETA(5)) SCAL=THETA(6)+ETA(6) IF (NEWIND.EQ.0) COM(1)=POTR F1=ABS(COM(1)) $DES DADT(1)=-POTK*A(1) ; Decay potentiation DADT(2)=(CPLA-A(2))*KEO ; Effect compartment conc $ERROR DPOT=A(1) ; The degree of potentiation CEFF=A(2) DPD1=CEFF**GAMM ; Degree of NMB NMB=DPD1/(DPD1+(EC50**GAMM)) Y=SCAL*(1+DPOT)*(1-NMB)+ERR(1) ; Twitch prediction COM(1)=POTR*(1-NMB) $THETA (0,0.13)(0,1460)(0,3.45) (0,0.0008)(0,0.146)(0,101) $OMEGA 0.1 0.1 0.1 0.1 0.1 10 $SIGMA 10 $ESTIMATION MAX=9999 SIG=6 PRINT=1 $TABLE TIME KEO EC50 GAMM POTR POTK SCAL NOHEADER NOPRINT FILE=potent2.txt $TABLE TIME CEFF DPOT NMB NOHEADER NOPRINT FILE=potent3.txt

From: "Nick Holford" n.holford@auckland.ac.nz Subject: Re: [NMusers] What does nonmem do after iteration stops? Date: Fri, June 3, 2005 11:45 am Doug, Two things that might take time to compute after minimization is complete are: 1. The covariance step. 2. POST HOC parameter estimates. However your control stream does not request either of these so I am afraid I have no explanation for the long time take to generate table output. What happens if you remove the $TABLE records? Nick

From: "Ekaterina Gibiansky" gibianskye@guilfordpharm.com Subject: Re: [NMusers] What does nonmem do after iteration stops? Date: Fri, June 3, 2005 12:08 pm Doug, how big your ouput table files are? Are you running NONMEM and writing the output to the drive on the same PC or the writing is going through the network? May be it is something to do with the network speed of writing the files? Katya

From: "Eleveld, DJ" d.j.eleveld@anest.umcg.nl Subject: RE: [NMusers] What does nonmem do after iteration stops? --> Workaround Date: Fri, June 3, 2005 1:54 pm Thanks to everyone for thier thoughtful suggestions. The delay seems to be related to the table step. The tables are fairly large (about 700k) and are written directly to the hard drive. If I remove the $TABLE lines then nonmem exits soon after it stops iterating as expected. Interestingly, if the FILE option isnt used, i.e. the table is written to the output file, then there is no delay. I will use this method as a work-around. I also tried POSTHOC without any tables and there was no delay. So I think problems with POSTHOC can be ruled out. The delay seems to happen *before* the writing of the table files. When a run was 'stuck', I didnt see any of the table files open. So I dont think the file access is simply too slow. Possibly the compiler/os has some issues with opening files under certain conditions. I am using Open Watcom 1.0 and Win 98. I will try to update to Open Watcom 1.3 and see if that helps. Thanks very much, Doug

From: "Eleveld, DJ" d.j.eleveld@anest.umcg.nl Subject: RE: [NMusers] What does nonmem do after iteration stops? --> Further info Date: Wed, June 8, 2005 10:34 am I tried a few more things to determine the source of the post-iteration delay: 1) $TABLE --> delay 2) $TABLE NOAPPEND --> nodelay 3) $TABLE NOAPPEND but with DV and Y as columns --> nodelay 4) No tables but $COVARAINCE --> delay It appears that the OS/compiler is not the cause since even big tables can be made without any extra delay. Neither does it seem to be related to prediction since DV and Y can be included in a table without extra delay. So to be more specific the delay seems to follow covaraince calculation or residual and weighted residual calculation. I dont understand why this is. This could possibly be a model-related issue since the model I am fitting is unusual in that the bioavalability parameter F1 changes over time depending on a random variable defined in $PK. Nonmem seems to have trouble handling this variable, giving me strange ETA estimations and POSTHOC values equal for all individuals regardles of the estimated ETA value. I'm in the process of documenting this and will report back to the users group with more detailed information. Thanks very much, Doug Eleveld _______________________________________________________