AR(1) residuals

From: drfreedman@drfreedmaninc.com Subject: [NMusers] AR(1) residuals Date: 12/17/2003 6:42 PM Colleagues Viewing the presentation NONMEM Alphabet Soup: L2, AR(1), and INFN by Marc Gastonguay posted at ; http://www.globomaxservice.com/presentations/presentations.h tml , I tried the following code (all drug-specific) information removed. The code including the "FIRST statement appears to be inserted into FSUBS.FOR in an executable section not permitting specification statements. Can you suggest any changes? Can this approach to serial correlation work with the ADBVAN subroutines? Thank-you in advance Dr. Immanuel Freedman ; run001.mod $DATA input.csv IGNORE=C ; $SUBROUTINE ADVAN2 TRANS2 ; $PK ;PK MODEL F1=value; Bioavailability CL=THETA(1)*EXP(ETA(1)); Clearance (mL/kg/d) V=THETA(2)*EXP(ETA(2)); Central volume (mL/kg) KA=THETA(3)*EXP(ETA(3)); Absorption rate (/d) S2=V; Scale factor ; $ERROR " FIRST " COMMON /NMPRD5/ C(50,29) " DIMENSION T(50) " IF (NEWIND .NE. 2) I=0 " IF (MDV .EQ. 0) THEN " I=I+1 " T(I)=TIME " DO 5 J=1,I " 5 C(J,I) = EXP(-THETA(4)*(TIME-T(J))) " ENDIF CP=A(2)/V; Y=CP*EXP(ERR(1))+ERR(2); ; $THETA ;values $OMEGA ;values $SIGMA ;values $ESTIMATION NOABORT MAXEVAL=9999 PRINT=10 METHOD=0 POSTHOC MSFO=run001.msf $COVARIANCE ; $TABLE NOPRINT ONEHEADER FILE=sdtab001 C ID TIME AMT DV CMT TYPE ROUT F1 CL V KA; ;

From: "Gastonguay, Marc" - marc.gastonguay@snet.net Subject: RE: [NMusers] AR(1) residuals Date: 12/17/2003 8:01 PM Dear Dr. Freedman, It is not clear what problem you are running into, but this approach can be used with the ADVAN subroutines. Below, I've listed an example control stream using ADVAN1 and part of the corresponding data set (the phenobarb example included in the NONMEM distribution). The FIRST statement ensures that the verbatim code is inserted immediately after variable declarations in the error subroutine of fsubs.for. You'll find a discussion and example of this AR(1) residual error model described in: Karlsson MO, Beal SL, Sheiner LB. Three new residual error models for population PK/PD analyses. J Pharmacokinet Biopharm 1995; 23(6):651-672. To be honest, I rarely find this residual error model to be useful. I'd prefer to address the serial correlation in residuals by improving the structural or covariate model whenever possible. Hope this helps, Marc (By the way, I should point out that the presentation you referred to was the result of a collaboration between myself, Tom Ludden and Stuart Beal.) ---------Control stream---------------------- $PROB 201, PHENOBARBITAL POPULATION PK MODEL AR1 RESIDUAL $INPUT X ID TIME AMT WT APGR DV EVID=DROP MDV=DROP $DATA 002.CSV IGNORE=C $SUBROUTINES ADVAN1 $PK TVCL=THETA(1) CL=TVCL*EXP(ETA(1)) TVV=THETA(2) V=TVV*EXP(ETA(2)) K=CL/V S1=V $ERROR " FIRST " COMMON /NMPRD5/ C(250,29) " DIMENSION T(50) " IF (NEWIND.NE.2) I=0 " IF (MDV.EQ.0) THEN " I=I+1 " T(I)=TIME " DO 5 J=1,I " 5 C(J,1)=EXP(-THETA(3)*(TIME-T(J))) " ENDIF E1=EPS(1) E2=EPS(2) IPRD=F IRES=DV-IPRD Y=F + F*E1 + E2 $THETA (0, 1) (0, 5) (0, 0.01) $OMEGA 0.16 0.16 $SIGMA 6 0.04 ;$MSFI 001.MSF $EST MAXEVAL=9999 PRINT=20 NOABORT POSTHOC MSF=201.MSF ;$COV $TABLE ID TIME IPRD CL WT APGR E1 E2 NOPRINT ONEHEADER FILE=201.TAB ---------Data (first 2 individuals of pheno data set)--------------------------- C ID TIME AMT WT APGR DV EVID MDV 1 0 25 1.4 7 . 1 1 1 2 . 1.4 7 17.3 0 0 1 12.5 3.5 1.4 7 . 1 1 1 24.5 3.5 1.4 7 . 1 1 1 37 3.5 1.4 7 . 1 1 1 48 3.5 1.4 7 . 1 1 1 60.5 3.5 1.4 7 . 1 1 1 72.5 3.5 1.4 7 . 1 1 1 85.3 3.5 1.4 7 . 1 1 1 96.5 3.5 1.4 7 . 1 1 1 108.5 3.5 1.4 7 . 1 1 1 112.5 . 1.4 7 31 0 0 2 0 15 1.5 9 . 1 1 2 2 . 1.5 9 9.7 0 0 2 4 3.8 1.5 9 . 1 1 2 16 3.8 1.5 9 . 1 1 2 27.8 3.8 1.5 9 . 1 1 2 40 3.8 1.5 9 . 1 1 2 52 3.8 1.5 9 . 1 1 2 63.5 . 1.5 9 24.6 0 0 2 64 3.8 1.5 9 . 1 1 2 76 3.8 1.5 9 . 1 1 2 88 3.8 1.5 9 . 1 1 2 100 3.8 1.5 9 . 1 1 2 112 3.8 1.5 9 . 1 1 2 124 3.8 1.5 9 . 1 1 2 135.5 . 1.5 9 33 0 0

From: "Gastonguay, Marc" - marc.gastonguay@snet.net Subject: RE: [NMusers] AR(1) residuals Date: 12/19/2003 11:14 AM Dear Dr Freedman, A follow-up to my previous message: I was reminded by Tom Ludden that a recent code-fix was announced on nmusers (Subject:[NMusers] Epsilons correlated across L2 records, Date: 4Dec2003; not yet in the archives), which directly applies to the model you've inquired about when using conditional estimation methods. Implementation of the example I supplied under conditional estimation methods assumes that you have made the recommended code change as described in the Dec 4 announcement. Marc Webmaster Update: The post referenced above can be found here: http://www.cognigencorp.com/nonmem/nm/99dec042003.html Thank you.

From: drfreedman@drfreedmaninc.com Subject: [NMusers] NONMEM AR(1) residual bug? Date: 12/23/2003 8:59 PM Colleagues As a follow-up to my previous message on this topic, the residual serial correlation control stream below yields the following FSUBS.FOR fragment which does not compile with the Compaq Fortran 6.6a compiler. Is this a bug in NONMEM? Thank-you in advance Immanuel Freedman, Ph.D., MIEEE (619) 884-1347 Control Stream: $ERROR " FIRST " COMMON /NMPRD5/ C(250,29) " DIMENSION T(50) " IF (NEWIND.NE.2) I=0 " IF (MDV.EQ.0) THEN " I=I+1 " T(I)=TIME " DO 5 J=1,I " 5 C(J,1)=EXP(-THETA(11)*(TIME-T(J))) " ENDIF FSUBS.FOR fragment: COMMON /NMPRD5/ C(250,29) DIMENSION T(50) IF (NEWIND.NE.2) I=0 IF (EVTREC(N,14).EQ.0) THEN I=I+1 T(I)=EVTREC(N,03) DO 5 J=1,I 5 EVTREC(N,01)(J,1)=EXP(-THETA(11)*(EVTREC(N,03)-T (J))) ENDIF _______________________________________________________