Baseline as a covariate

6 messages 6 people Latest: Aug 24, 2010

Baseline as a covariate

From: Peter Bonate Date: August 11, 2010 technical
I'd like to get the group's opinion on something. I have a pharmacodynamic model and the baseline was shown to be a covariate on one of the model parameters. I was hoping to get general thoughts on the use of the baseline as a covariate. Is there a preference for using the observed baseline vs. NONMEM predicted baseline? And does your opinion change if you have a large residual error? Thanks Pete bonate Peter L. Bonate, PhD, FCP, FAAPS GlaxoSmithKline Clinical Pharmacology, Modeling, and Simulation 5 Moore Drive, 17.2259 Research Triangle Park, NC 27709 phone: 919-483-7534 fax: 919-483-8948 email: [email protected] This is a long one but worth it... Scientists will never make as much money as business executives. There is now mathematical proof for this statement. Postulate 1: Knowledge is power postulate 2: Time is money According to the laws of physics, Work -------- = Power (Eq. 1) Time Since, from the 2 postulates above, Knowledge = Power and Time = Money, we can substitute into the equation 1 and come up with Work -------- = Knowledge Money Solving for money, Work ------------ = Money Knowledge Hence, there are 2 ways to make money. If we slug our guts out working, we can improve our money situation. This is the normal scientific approach. However the suits usually opt for the easier solution, i.e., the less you know, the more you'll make, regardless of the amount of work done.
Quoted reply history
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Leonid Gibiansky Sent: Wednesday, August 11, 2010 9:01 AM To: Ann Rigby-Jones Cc: '[email protected]' Subject: Re: [NMusers] Rounding errors with TRANSIT model Ann Try ADVAN6 or ADVAN8 or ADVAN13 . Also, TOL=3 is too small. Increase it to 6-7 or even 9, and also change to NSIG=3 SIGL=9 ($EST step, as recommended in the guide). Also, at least initially, I would remove ETAs from the third compartment and from one of the parameters (Q2 or V2) of the second compartment. Do you really need both ALAG and transit compartment? They serve the same goal, to delay the input, so could be strongly correlated; you may see it on the POSTHOC plot of ETA7-ETA8 versus ETA9. I would try to remove ETAs from ALAG or KTR to check whether they are needed. If nothing helps, use UNCONDITIONAL on the $COV to get standard errors, and if they are reasonable, ignore the error, it may disappear on the next steps of the modeling process Leonid -------------------------------------- Leonid Gibiansky, Ph.D. President, QuantPharm LLC web: www.quantpharm.com e-mail: LGibiansky at quantpharm.com tel: (301) 767 5566 On 8/11/2010 5:57 AM, Ann Rigby-Jones wrote: > Dear All > > I'm trying to evaluate a transit model approach in an attempt to better > describe early drug concentrations following intravenous injection (1 > minute and 10 minute infusions) of a sedative-hypnotic drug. However, > every minimisation attempt is terminated due to rounding errors (E=134). > I've tried the usual strategies to overcome this e.g using final > estimates of a terminated run as initial estimates for the next, > changing number of sig dig requested, changing from a diagonal omega to > a block omega, but nothing has been successful. > > I'd be very grateful for any suggestions of what I might try next J > Standard 3-comp and 4-comp mamillary models minise successful with these > data, but so far no luck with any of the transit models, I have tried > model with 1 to 6 transit compartments. An example control stream is > shown below. I'm using NONMEM 7 with Intel Fortran. > > With thanks and all best wishes > > Ann > > $PROBLEM 3comp 2 transit > > $INPUT ID DOSE AMT RATE DUR TIME ORI DV EVID ART AGE WGT > > $DATA PKcen12LN.csv IGNORE=# > > $SUBROUTINES ADVAN9 TOL=3 > > $MODEL > > ;NCOMPS=9 > > COMP(CENTRAL, DEFOBS) ;1 > > COMP(PERIPH1) ;2 > > COMP(PERIPH2) ;3 > > COMP(TRANS1, DEFDOSE) ;4 > > COMP(TRANS2) ;5 > > ;COMP(TRANS3) ;6 > > ;COMP(TRANS4) ;7 > > ;COMP(TRANS5) ;8 > > ;COMP(TRANS6) ;9 > > $PK > > CL=THETA(1)*EXP(ETA(1)) > > Q2=THETA(2)*EXP(ETA(2)) > > Q3=THETA(3)*EXP(ETA(3)) > > V1=THETA(4)*EXP(ETA(4)) > > V2=THETA(5)*EXP(ETA(5)) > > V3=THETA(6)*EXP(ETA(6)) > > K10=CL/V1 > > K12=Q2/V1 > > K13=Q3/V1 > > K21=Q2/V2 > > K31=Q3/V3 > > S1=V1 > > IF (DUR.EQ.10) THEN > > ALAG4=THETA(7)*EXP(ETA(7)) > > ELSE > > ALAG4=THETA(8)*EXP(ETA(8)) > > ENDIF > > KTR=THETA(9)*EXP(ETA(9)) > > $DES > > ;DADT(1)=A(2)*K21 + A(3)*K31 - A(1)*(K10+K12+K13) > > ;DADT(2)=A(1)*K12 - A(2)*K21 > > ;DADT(3)=A(1)*K13 - A(3)*K31 > > DADT(1)=A(5)*KTR + A(2)*K21 + A(3)*K31 - A(1)*(K10+K12+K13) > > DADT(2)=A(1)*K12 - A(2)*K21 > > DADT(3)=A(1)*K13 - A(3)*K31 > > DADT(4)=-A(4)*KTR > > DADT(5)=A(4)*KTR - A(5)*KTR > > ;DADT(6)=A(5)*KTR - A(6)*KTR > > ;DADT(7)=A(6)*KTR - A(7)*KTR > > ;DADT(8)=A(7)*KTR - A(8)*KTR > > ;DADT(9)=A(8)*KTR - A(9)*KTR > > $ERROR > > W=1 > > IPRED= -2 > > IF (F.GT.0) IPRED=LOG(F) > > Y=IPRED + ERR(1) > > IRES=DV-IPRED > > IWRES=IRES/W > > $THETA (0, 725) ;CL > > $THETA (0, 238) ;Q2 > > $THETA (0, 2920) ;Q3 > > $THETA (0, 311) ;V1 > > $THETA (0, 38700) ;V2 > > $THETA (0, 39500) ;V3 > > $THETA (0.2167, 0.522,1) ;ALAG 10MIN > > $THETA (0.00833, 0.122, 1) ;ALAG 1MIN > > $THETA (0, 0.720) ;KTR > > $OMEGA BLOCK(6) > > 0.0959 ; ETA CL > > 0.00804 0.0211 ; ETA Q2 > > 0.00631 0.00143 0.166 ; ETA Q3 > > -0.00506 0.00121 -0.0635 0.0491 ; ETA V1 > > 0.00293 -0.00746 0.00708 -0.00315 0.0328 ; ETA V2 > > 0.00836 0.00688 0.00144 -0.00497 0.00741 0.153 ; ETA V3 > > $OMEGA 0.217 ; ETA ALAG 10 > > $OMEGA 0.4 ; ETA ALAG 1 > > $OMEGA 0.0171 ; ETA KTR > > ;$OMEGA (0.0686) ; ETA CL > > ;$OMEGA (0.01) ; ETA Q2 > > ;$OMEGA (0.0251) ; ETA Q3 > > ;$OMEGA (0 FIX) ; ETA V1 > > ;$OMEGA (0.0328) ; ETA V2 > > ;$OMEGA (0 FIX) ; ETA V3 > > ;$OMEGA (0.170) ; ETA ALAG10 > > ;$OMEGA (1.37) ; ETA ALAG1 > > ;$OMEGA (0.00409) ; ETA KTR > > $SIGMA (0.0944) > > $ESTIMATION METHOD=1 PRINT=1 MAX=9999 NOABORT SIG=3 ;POSTHOC INTER > > MSFO=msfo.outputfile > > ;$COVA > > $TABLE ID EVID AMT TIME IPRED IRES > > NOPRINT FILE=AllRecords.txt > > $TABLE ID > > CL Q2 Q3 V1 V2 V3 > > ETA1 ETA2 ETA3 ETA4 ETA5 ETA6 ETA7 ;ETA8 > > FIRSTONLY NOPRINT NOAPPEND FILE=FirstRecords.txt > > Ann > > _______________________________________________________________________ > > *Ann Rigby-Jones PhD MRSC* > Research Fellow in Pharmacokinetics & Pharmacodynamics > > Peninsula College of Medicine & Dentistry > > N31, ITTC Phase 1 > Tamar Science Park > 1 Davy Road > Derriford > Plymouth > PL6 8BX > > Tel: +44 (0) 1752 432014 > _______________________________________________________________________ >

RE: Baseline as a covariate

From: Kenneth Kowalski Date: August 11, 2010 technical
Hi Pete, In this setting I generally try to first model the baseline response and perhaps pursue alternative structural model forms. For example, I may consider a multiplicative relationship rather than an additive relationship between baseline and placebo/drug effects. However, if the distribution of the baseline response is quite complex and not easy to model with a normal eta on baseline, and I'm concerned that I might be getting a biased estimate of the baseline response then I may consider treating the observed baseline as a covariate as a fall-back position. This question is similar to whether one uses observed or predicted concentrations when developing a PK/PD model. We generally prefer using predicted concentrations rather than observed concentrations to smooth out the measurement error (among other reasons as well). However, if we have a lot of lack-of-fit in developing a PK model then it may be preferable to use the observed concentrations. There is a tradeoff as to whether more bias is introduced due to lack-of-fit (e.g., poor estimation of the baseline response) or due to measurement error in using the observed baseline measurements as a covariate. If the residual variability due to lack-of-fit is considerably larger than measurement error, and you can't resolve the lack-of-fit, then you might consider using the observed baseline response as a covariate. Kind regards, Ken Kenneth G. Kowalski President & CEO A2PG - Ann Arbor Pharmacometrics Group, Inc. 110 E. Miller Ave., Garden Suite Ann Arbor, MI 48104 Work: 734-274-8255 Cell: 248-207-5082 Fax: 734-913-0230 [email protected]
Quoted reply history
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Peter Bonate Sent: Wednesday, August 11, 2010 11:22 AM To: '[email protected]' Subject: [NMusers] Baseline as a covariate I'd like to get the group's opinion on something. I have a pharmacodynamic model and the baseline was shown to be a covariate on one of the model parameters. I was hoping to get general thoughts on the use of the baseline as a covariate. Is there a preference for using the observed baseline vs. NONMEM predicted baseline? And does your opinion change if you have a large residual error? Thanks Pete bonate Peter L. Bonate, PhD, FCP, FAAPS GlaxoSmithKline Clinical Pharmacology, Modeling, and Simulation 5 Moore Drive, 17.2259 Research Triangle Park, NC 27709 phone: 919-483-7534 fax: 919-483-8948 email: [email protected] This is a long one but worth it... Scientists will never make as much money as business executives. There is now mathematical proof for this statement. Postulate 1: Knowledge is power postulate 2: Time is money According to the laws of physics, Work -------- = Power (Eq. 1) Time Since, from the 2 postulates above, Knowledge = Power and Time = Money, we can substitute into the equation 1 and come up with Work -------- = Knowledge Money Solving for money, Work ------------ = Money Knowledge Hence, there are 2 ways to make money. If we slug our guts out working, we can improve our money situation. This is the normal scientific approach. However the suits usually opt for the easier solution, i.e., the less you know, the more you'll make, regardless of the amount of work done. -----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Leonid Gibiansky Sent: Wednesday, August 11, 2010 9:01 AM To: Ann Rigby-Jones Cc: '[email protected]' Subject: Re: [NMusers] Rounding errors with TRANSIT model Ann Try ADVAN6 or ADVAN8 or ADVAN13 . Also, TOL=3 is too small. Increase it to 6-7 or even 9, and also change to NSIG=3 SIGL=9 ($EST step, as recommended in the guide). Also, at least initially, I would remove ETAs from the third compartment and from one of the parameters (Q2 or V2) of the second compartment. Do you really need both ALAG and transit compartment? They serve the same goal, to delay the input, so could be strongly correlated; you may see it on the POSTHOC plot of ETA7-ETA8 versus ETA9. I would try to remove ETAs from ALAG or KTR to check whether they are needed. If nothing helps, use UNCONDITIONAL on the $COV to get standard errors, and if they are reasonable, ignore the error, it may disappear on the next steps of the modeling process Leonid -------------------------------------- Leonid Gibiansky, Ph.D. President, QuantPharm LLC web: www.quantpharm.com e-mail: LGibiansky at quantpharm.com tel: (301) 767 5566 On 8/11/2010 5:57 AM, Ann Rigby-Jones wrote: > Dear All > > I'm trying to evaluate a transit model approach in an attempt to better > describe early drug concentrations following intravenous injection (1 > minute and 10 minute infusions) of a sedative-hypnotic drug. However, > every minimisation attempt is terminated due to rounding errors (E=134). > I've tried the usual strategies to overcome this e.g using final > estimates of a terminated run as initial estimates for the next, > changing number of sig dig requested, changing from a diagonal omega to > a block omega, but nothing has been successful. > > I'd be very grateful for any suggestions of what I might try next J > Standard 3-comp and 4-comp mamillary models minise successful with these > data, but so far no luck with any of the transit models, I have tried > model with 1 to 6 transit compartments. An example control stream is > shown below. I'm using NONMEM 7 with Intel Fortran. > > With thanks and all best wishes > > Ann > > $PROBLEM 3comp 2 transit > > $INPUT ID DOSE AMT RATE DUR TIME ORI DV EVID ART AGE WGT > > $DATA PKcen12LN.csv IGNORE=# > > $SUBROUTINES ADVAN9 TOL=3 > > $MODEL > > ;NCOMPS=9 > > COMP(CENTRAL, DEFOBS) ;1 > > COMP(PERIPH1) ;2 > > COMP(PERIPH2) ;3 > > COMP(TRANS1, DEFDOSE) ;4 > > COMP(TRANS2) ;5 > > ;COMP(TRANS3) ;6 > > ;COMP(TRANS4) ;7 > > ;COMP(TRANS5) ;8 > > ;COMP(TRANS6) ;9 > > $PK > > CL=THETA(1)*EXP(ETA(1)) > > Q2=THETA(2)*EXP(ETA(2)) > > Q3=THETA(3)*EXP(ETA(3)) > > V1=THETA(4)*EXP(ETA(4)) > > V2=THETA(5)*EXP(ETA(5)) > > V3=THETA(6)*EXP(ETA(6)) > > K10=CL/V1 > > K12=Q2/V1 > > K13=Q3/V1 > > K21=Q2/V2 > > K31=Q3/V3 > > S1=V1 > > IF (DUR.EQ.10) THEN > > ALAG4=THETA(7)*EXP(ETA(7)) > > ELSE > > ALAG4=THETA(8)*EXP(ETA(8)) > > ENDIF > > KTR=THETA(9)*EXP(ETA(9)) > > $DES > > ;DADT(1)=A(2)*K21 + A(3)*K31 - A(1)*(K10+K12+K13) > > ;DADT(2)=A(1)*K12 - A(2)*K21 > > ;DADT(3)=A(1)*K13 - A(3)*K31 > > DADT(1)=A(5)*KTR + A(2)*K21 + A(3)*K31 - A(1)*(K10+K12+K13) > > DADT(2)=A(1)*K12 - A(2)*K21 > > DADT(3)=A(1)*K13 - A(3)*K31 > > DADT(4)=-A(4)*KTR > > DADT(5)=A(4)*KTR - A(5)*KTR > > ;DADT(6)=A(5)*KTR - A(6)*KTR > > ;DADT(7)=A(6)*KTR - A(7)*KTR > > ;DADT(8)=A(7)*KTR - A(8)*KTR > > ;DADT(9)=A(8)*KTR - A(9)*KTR > > $ERROR > > W=1 > > IPRED= -2 > > IF (F.GT.0) IPRED=LOG(F) > > Y=IPRED + ERR(1) > > IRES=DV-IPRED > > IWRES=IRES/W > > $THETA (0, 725) ;CL > > $THETA (0, 238) ;Q2 > > $THETA (0, 2920) ;Q3 > > $THETA (0, 311) ;V1 > > $THETA (0, 38700) ;V2 > > $THETA (0, 39500) ;V3 > > $THETA (0.2167, 0.522,1) ;ALAG 10MIN > > $THETA (0.00833, 0.122, 1) ;ALAG 1MIN > > $THETA (0, 0.720) ;KTR > > $OMEGA BLOCK(6) > > 0.0959 ; ETA CL > > 0.00804 0.0211 ; ETA Q2 > > 0.00631 0.00143 0.166 ; ETA Q3 > > -0.00506 0.00121 -0.0635 0.0491 ; ETA V1 > > 0.00293 -0.00746 0.00708 -0.00315 0.0328 ; ETA V2 > > 0.00836 0.00688 0.00144 -0.00497 0.00741 0.153 ; ETA V3 > > $OMEGA 0.217 ; ETA ALAG 10 > > $OMEGA 0.4 ; ETA ALAG 1 > > $OMEGA 0.0171 ; ETA KTR > > ;$OMEGA (0.0686) ; ETA CL > > ;$OMEGA (0.01) ; ETA Q2 > > ;$OMEGA (0.0251) ; ETA Q3 > > ;$OMEGA (0 FIX) ; ETA V1 > > ;$OMEGA (0.0328) ; ETA V2 > > ;$OMEGA (0 FIX) ; ETA V3 > > ;$OMEGA (0.170) ; ETA ALAG10 > > ;$OMEGA (1.37) ; ETA ALAG1 > > ;$OMEGA (0.00409) ; ETA KTR > > $SIGMA (0.0944) > > $ESTIMATION METHOD=1 PRINT=1 MAX=9999 NOABORT SIG=3 ;POSTHOC INTER > > MSFO=msfo.outputfile > > ;$COVA > > $TABLE ID EVID AMT TIME IPRED IRES > > NOPRINT FILE=AllRecords.txt > > $TABLE ID > > CL Q2 Q3 V1 V2 V3 > > ETA1 ETA2 ETA3 ETA4 ETA5 ETA6 ETA7 ;ETA8 > > FIRSTONLY NOPRINT NOAPPEND FILE=FirstRecords.txt > > Ann > > _______________________________________________________________________ > > *Ann Rigby-Jones PhD MRSC* > Research Fellow in Pharmacokinetics & Pharmacodynamics > > Peninsula College of Medicine & Dentistry > > N31, ITTC Phase 1 > Tamar Science Park > 1 Davy Road > Derriford > Plymouth > PL6 8BX > > Tel: +44 (0) 1752 432014 > _______________________________________________________________________ >

Re: Baseline as a covariate

From: Leonid Gibiansky Date: August 11, 2010 technical
Hi Peter, I assume from the question that the baseline is the baseline of your modeled PD measure, not of some other value (like the baseline weight), and that you model the actual PD measure, not change from the baseline. Then - it would be more logical to use the model-predicted baseline; - for the clinical applications it could be more useful to have observed predictor; - for simulations of future studies it could be more convenient to have everything in the model rather than use observed values (especially if you apply the model to different population where the baseline value could be shifted). I would try both versions to see the differences, and then use the one that is more convenient to use in the particular situation. Best regards, Leonid -------------------------------------- Leonid Gibiansky, Ph.D. President, QuantPharm LLC web: www.quantpharm.com e-mail: LGibiansky at quantpharm.com tel: (301) 767 5566
Quoted reply history
On 8/11/2010 11:21 AM, Peter Bonate wrote: > I'd like to get the group's opinion on something. I have a pharmacodynamic > model and the baseline was shown to be a covariate on one of the model > parameters. I was hoping to get general thoughts on the use of the baseline as > a covariate. Is there a preference for using the observed baseline vs. NONMEM > predicted baseline? And does your opinion change if you have a large residual > error? > > Thanks > > Pete bonate > > Peter L. Bonate, PhD, FCP, FAAPS > GlaxoSmithKline > Clinical Pharmacology, Modeling, and Simulation > 5 Moore Drive, 17.2259 > Research Triangle Park, NC 27709 > phone: 919-483-7534 > fax: 919-483-8948 > email: [email protected]

RE: Baseline as a covariate

From: Mahesh Samtani Date: August 11, 2010 technical
Dear Dr. Bonate, Could you kindly elaborate on the structural model that is being used to characterize your PD system? In my limited experience, I have seen the behavior that you describe under 2 circumstances: a) Simple inhibitory Emax PD model with a baseline: Here patients with a higher baseline (worse disease state) displayed a more pronounced effect. There is greater margin/room for improvement for a patient starting out at a higher baseline and hence shows a better response (vs. another patient that starts out a lower baseline). I was able to model this behavior by putting an omega block between Emax and the baseline parameter Eo. It is then very easy to adjudicate the impact of Eo on the magnitude of effect in simulation mode. b) Indirect Response Model: Sun and Jusko [J Pharm Sci 88: 987 (1999)] have shown that "The baseline value may play an important role in affecting the extent of the response if its PD relationship can be described by a turnover Model. When other factors remain constant (e.g. Smax, SC50, kout), R0 controls the magnitude of the response." Is it possible that you're getting a spurious baseline covariate effect because of the choice of your structural model? I am also quite interested in hearing the general thoughts on the use of the baseline as a covariate from this forum. In fact I have seen yet another approach to this problem where the change from baseline was modeled as the PD endpoint. With this method the baseline is no longer a parameter in the model and observed baseline can then be used as a covariate. Does this method offer any benefit or do the statistical concerns around data transformation negate the benefits of this method. Thank-you, Mahesh
Quoted reply history
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Ken Kowalski Sent: Wednesday, August 11, 2010 12:12 PM To: 'Peter Bonate'; [email protected] Subject: RE: [NMusers] Baseline as a covariate Hi Pete, In this setting I generally try to first model the baseline response and perhaps pursue alternative structural model forms. For example, I may consider a multiplicative relationship rather than an additive relationship between baseline and placebo/drug effects. However, if the distribution of the baseline response is quite complex and not easy to model with a normal eta on baseline, and I'm concerned that I might be getting a biased estimate of the baseline response then I may consider treating the observed baseline as a covariate as a fall-back position. This question is similar to whether one uses observed or predicted concentrations when developing a PK/PD model. We generally prefer using predicted concentrations rather than observed concentrations to smooth out the measurement error (among other reasons as well). However, if we have a lot of lack-of-fit in developing a PK model then it may be preferable to use the observed concentrations. There is a tradeoff as to whether more bias is introduced due to lack-of-fit (e.g., poor estimation of the baseline response) or due to measurement error in using the observed baseline measurements as a covariate. If the residual variability due to lack-of-fit is considerably larger than measurement error, and you can't resolve the lack-of-fit, then you might consider using the observed baseline response as a covariate. Kind regards, Ken Kenneth G. Kowalski President & CEO A2PG - Ann Arbor Pharmacometrics Group, Inc. 110 E. Miller Ave., Garden Suite Ann Arbor, MI 48104 Work: 734-274-8255 Cell: 248-207-5082 Fax: 734-913-0230 [email protected] -----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Peter Bonate Sent: Wednesday, August 11, 2010 11:22 AM To: '[email protected]' Subject: [NMusers] Baseline as a covariate I'd like to get the group's opinion on something. I have a pharmacodynamic model and the baseline was shown to be a covariate on one of the model parameters. I was hoping to get general thoughts on the use of the baseline as a covariate. Is there a preference for using the observed baseline vs. NONMEM predicted baseline? And does your opinion change if you have a large residual error? Thanks Pete bonate Peter L. Bonate, PhD, FCP, FAAPS GlaxoSmithKline Clinical Pharmacology, Modeling, and Simulation 5 Moore Drive, 17.2259 Research Triangle Park, NC 27709 phone: 919-483-7534 fax: 919-483-8948 email: [email protected] This is a long one but worth it... Scientists will never make as much money as business executives. There is now mathematical proof for this statement. Postulate 1: Knowledge is power postulate 2: Time is money According to the laws of physics, Work -------- = Power (Eq. 1) Time Since, from the 2 postulates above, Knowledge = Power and Time = Money, we can substitute into the equation 1 and come up with Work -------- = Knowledge Money Solving for money, Work ------------ = Money Knowledge Hence, there are 2 ways to make money. If we slug our guts out working, we can improve our money situation. This is the normal scientific approach. However the suits usually opt for the easier solution, i.e., the less you know, the more you'll make, regardless of the amount of work done. -----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Leonid Gibiansky Sent: Wednesday, August 11, 2010 9:01 AM To: Ann Rigby-Jones Cc: '[email protected]' Subject: Re: [NMusers] Rounding errors with TRANSIT model Ann Try ADVAN6 or ADVAN8 or ADVAN13 . Also, TOL=3 is too small. Increase it to 6-7 or even 9, and also change to NSIG=3 SIGL=9 ($EST step, as recommended in the guide). Also, at least initially, I would remove ETAs from the third compartment and from one of the parameters (Q2 or V2) of the second compartment. Do you really need both ALAG and transit compartment? They serve the same goal, to delay the input, so could be strongly correlated; you may see it on the POSTHOC plot of ETA7-ETA8 versus ETA9. I would try to remove ETAs from ALAG or KTR to check whether they are needed. If nothing helps, use UNCONDITIONAL on the $COV to get standard errors, and if they are reasonable, ignore the error, it may disappear on the next steps of the modeling process Leonid -------------------------------------- Leonid Gibiansky, Ph.D. President, QuantPharm LLC web: www.quantpharm.com e-mail: LGibiansky at quantpharm.com tel: (301) 767 5566 On 8/11/2010 5:57 AM, Ann Rigby-Jones wrote: > Dear All > > I'm trying to evaluate a transit model approach in an attempt to better > describe early drug concentrations following intravenous injection (1 > minute and 10 minute infusions) of a sedative-hypnotic drug. However, > every minimisation attempt is terminated due to rounding errors (E=134). > I've tried the usual strategies to overcome this e.g using final > estimates of a terminated run as initial estimates for the next, > changing number of sig dig requested, changing from a diagonal omega to > a block omega, but nothing has been successful. > > I'd be very grateful for any suggestions of what I might try next J > Standard 3-comp and 4-comp mamillary models minise successful with these > data, but so far no luck with any of the transit models, I have tried > model with 1 to 6 transit compartments. An example control stream is > shown below. I'm using NONMEM 7 with Intel Fortran. > > With thanks and all best wishes > > Ann > > $PROBLEM 3comp 2 transit > > $INPUT ID DOSE AMT RATE DUR TIME ORI DV EVID ART AGE WGT > > $DATA PKcen12LN.csv IGNORE=# > > $SUBROUTINES ADVAN9 TOL=3 > > $MODEL > > ;NCOMPS=9 > > COMP(CENTRAL, DEFOBS) ;1 > > COMP(PERIPH1) ;2 > > COMP(PERIPH2) ;3 > > COMP(TRANS1, DEFDOSE) ;4 > > COMP(TRANS2) ;5 > > ;COMP(TRANS3) ;6 > > ;COMP(TRANS4) ;7 > > ;COMP(TRANS5) ;8 > > ;COMP(TRANS6) ;9 > > $PK > > CL=THETA(1)*EXP(ETA(1)) > > Q2=THETA(2)*EXP(ETA(2)) > > Q3=THETA(3)*EXP(ETA(3)) > > V1=THETA(4)*EXP(ETA(4)) > > V2=THETA(5)*EXP(ETA(5)) > > V3=THETA(6)*EXP(ETA(6)) > > K10=CL/V1 > > K12=Q2/V1 > > K13=Q3/V1 > > K21=Q2/V2 > > K31=Q3/V3 > > S1=V1 > > IF (DUR.EQ.10) THEN > > ALAG4=THETA(7)*EXP(ETA(7)) > > ELSE > > ALAG4=THETA(8)*EXP(ETA(8)) > > ENDIF > > KTR=THETA(9)*EXP(ETA(9)) > > $DES > > ;DADT(1)=A(2)*K21 + A(3)*K31 - A(1)*(K10+K12+K13) > > ;DADT(2)=A(1)*K12 - A(2)*K21 > > ;DADT(3)=A(1)*K13 - A(3)*K31 > > DADT(1)=A(5)*KTR + A(2)*K21 + A(3)*K31 - A(1)*(K10+K12+K13) > > DADT(2)=A(1)*K12 - A(2)*K21 > > DADT(3)=A(1)*K13 - A(3)*K31 > > DADT(4)=-A(4)*KTR > > DADT(5)=A(4)*KTR - A(5)*KTR > > ;DADT(6)=A(5)*KTR - A(6)*KTR > > ;DADT(7)=A(6)*KTR - A(7)*KTR > > ;DADT(8)=A(7)*KTR - A(8)*KTR > > ;DADT(9)=A(8)*KTR - A(9)*KTR > > $ERROR > > W=1 > > IPRED= -2 > > IF (F.GT.0) IPRED=LOG(F) > > Y=IPRED + ERR(1) > > IRES=DV-IPRED > > IWRES=IRES/W > > $THETA (0, 725) ;CL > > $THETA (0, 238) ;Q2 > > $THETA (0, 2920) ;Q3 > > $THETA (0, 311) ;V1 > > $THETA (0, 38700) ;V2 > > $THETA (0, 39500) ;V3 > > $THETA (0.2167, 0.522,1) ;ALAG 10MIN > > $THETA (0.00833, 0.122, 1) ;ALAG 1MIN > > $THETA (0, 0.720) ;KTR > > $OMEGA BLOCK(6) > > 0.0959 ; ETA CL > > 0.00804 0.0211 ; ETA Q2 > > 0.00631 0.00143 0.166 ; ETA Q3 > > -0.00506 0.00121 -0.0635 0.0491 ; ETA V1 > > 0.00293 -0.00746 0.00708 -0.00315 0.0328 ; ETA V2 > > 0.00836 0.00688 0.00144 -0.00497 0.00741 0.153 ; ETA V3 > > $OMEGA 0.217 ; ETA ALAG 10 > > $OMEGA 0.4 ; ETA ALAG 1 > > $OMEGA 0.0171 ; ETA KTR > > ;$OMEGA (0.0686) ; ETA CL > > ;$OMEGA (0.01) ; ETA Q2 > > ;$OMEGA (0.0251) ; ETA Q3 > > ;$OMEGA (0 FIX) ; ETA V1 > > ;$OMEGA (0.0328) ; ETA V2 > > ;$OMEGA (0 FIX) ; ETA V3 > > ;$OMEGA (0.170) ; ETA ALAG10 > > ;$OMEGA (1.37) ; ETA ALAG1 > > ;$OMEGA (0.00409) ; ETA KTR > > $SIGMA (0.0944) > > $ESTIMATION METHOD=1 PRINT=1 MAX=9999 NOABORT SIG=3 ;POSTHOC INTER > > MSFO=msfo.outputfile > > ;$COVA > > $TABLE ID EVID AMT TIME IPRED IRES > > NOPRINT FILE=AllRecords.txt > > $TABLE ID > > CL Q2 Q3 V1 V2 V3 > > ETA1 ETA2 ETA3 ETA4 ETA5 ETA6 ETA7 ;ETA8 > > FIRSTONLY NOPRINT NOAPPEND FILE=FirstRecords.txt > > Ann > > _______________________________________________________________________ > > *Ann Rigby-Jones PhD MRSC* > Research Fellow in Pharmacokinetics & Pharmacodynamics > > Peninsula College of Medicine & Dentistry > > N31, ITTC Phase 1 > Tamar Science Park > 1 Davy Road > Derriford > Plymouth > PL6 8BX > > Tel: +44 (0) 1752 432014 > _______________________________________________________________________ >

RE: Baseline as a covariate

From: Mats Karlsson Date: August 11, 2010 technical
Hi Pete, I think the standard model most of us would use is baseline as a parameter in the model, and like other parameters it would have a variability between subjects. What you describe sounds like a covariance between the baseline parameter and some other parameter. We are used to include such covariances and I would suggest that to be the primary way to address it. If the distribution of baseline is truncated due to inclusion criteria or physiological limits, I suggest that you may use a semi-parametric IIV model (Peterson et al., Pharm Res. 2009 Sep;26(9):2174-85) If the relation is more complex than easily handled via a covariance, then I suggest that you use the baseline observation as a covariate with error (Dansirikul et al Pharmacokinet Pharmacodyn. 2008 Jun;35(3):269-83.) Best regards, Mats Mats Karlsson, PhD Professor of Pharmacometrics Dept of Pharmaceutical Biosciences Uppsala University Swedent regards, Mats Postal address: Box 591, 751 24 Uppsala, Sweden Phone +46 18 4714105 Fax + 46 18 4714003
Quoted reply history
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Peter Bonate Sent: Wednesday, August 11, 2010 5:22 PM To: '[email protected]' Subject: [NMusers] Baseline as a covariate I'd like to get the group's opinion on something. I have a pharmacodynamic model and the baseline was shown to be a covariate on one of the model parameters. I was hoping to get general thoughts on the use of the baseline as a covariate. Is there a preference for using the observed baseline vs. NONMEM predicted baseline? And does your opinion change if you have a large residual error? Thanks Pete bonate Peter L. Bonate, PhD, FCP, FAAPS GlaxoSmithKline Clinical Pharmacology, Modeling, and Simulation 5 Moore Drive, 17.2259 Research Triangle Park, NC 27709 phone: 919-483-7534 fax: 919-483-8948 email: [email protected] This is a long one but worth it... Scientists will never make as much money as business executives. There is now mathematical proof for this statement. Postulate 1: Knowledge is power postulate 2: Time is money According to the laws of physics, Work -------- = Power (Eq. 1) Time Since, from the 2 postulates above, Knowledge = Power and Time = Money, we can substitute into the equation 1 and come up with Work -------- = Knowledge Money Solving for money, Work ------------ = Money Knowledge Hence, there are 2 ways to make money. If we slug our guts out working, we can improve our money situation. This is the normal scientific approach. However the suits usually opt for the easier solution, i.e., the less you know, the more you'll make, regardless of the amount of work done. -----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Leonid Gibiansky Sent: Wednesday, August 11, 2010 9:01 AM To: Ann Rigby-Jones Cc: '[email protected]' Subject: Re: [NMusers] Rounding errors with TRANSIT model Ann Try ADVAN6 or ADVAN8 or ADVAN13 . Also, TOL=3 is too small. Increase it to 6-7 or even 9, and also change to NSIG=3 SIGL=9 ($EST step, as recommended in the guide). Also, at least initially, I would remove ETAs from the third compartment and from one of the parameters (Q2 or V2) of the second compartment. Do you really need both ALAG and transit compartment? They serve the same goal, to delay the input, so could be strongly correlated; you may see it on the POSTHOC plot of ETA7-ETA8 versus ETA9. I would try to remove ETAs from ALAG or KTR to check whether they are needed. If nothing helps, use UNCONDITIONAL on the $COV to get standard errors, and if they are reasonable, ignore the error, it may disappear on the next steps of the modeling process Leonid -------------------------------------- Leonid Gibiansky, Ph.D. President, QuantPharm LLC web: www.quantpharm.com e-mail: LGibiansky at quantpharm.com tel: (301) 767 5566 On 8/11/2010 5:57 AM, Ann Rigby-Jones wrote: > Dear All > > I'm trying to evaluate a transit model approach in an attempt to better > describe early drug concentrations following intravenous injection (1 > minute and 10 minute infusions) of a sedative-hypnotic drug. However, > every minimisation attempt is terminated due to rounding errors (E=134). > I've tried the usual strategies to overcome this e.g using final > estimates of a terminated run as initial estimates for the next, > changing number of sig dig requested, changing from a diagonal omega to > a block omega, but nothing has been successful. > > I'd be very grateful for any suggestions of what I might try next J > Standard 3-comp and 4-comp mamillary models minise successful with these > data, but so far no luck with any of the transit models, I have tried > model with 1 to 6 transit compartments. An example control stream is > shown below. I'm using NONMEM 7 with Intel Fortran. > > With thanks and all best wishes > > Ann > > $PROBLEM 3comp 2 transit > > $INPUT ID DOSE AMT RATE DUR TIME ORI DV EVID ART AGE WGT > > $DATA PKcen12LN.csv IGNORE=# > > $SUBROUTINES ADVAN9 TOL=3 > > $MODEL > > ;NCOMPS=9 > > COMP(CENTRAL, DEFOBS) ;1 > > COMP(PERIPH1) ;2 > > COMP(PERIPH2) ;3 > > COMP(TRANS1, DEFDOSE) ;4 > > COMP(TRANS2) ;5 > > ;COMP(TRANS3) ;6 > > ;COMP(TRANS4) ;7 > > ;COMP(TRANS5) ;8 > > ;COMP(TRANS6) ;9 > > $PK > > CL=THETA(1)*EXP(ETA(1)) > > Q2=THETA(2)*EXP(ETA(2)) > > Q3=THETA(3)*EXP(ETA(3)) > > V1=THETA(4)*EXP(ETA(4)) > > V2=THETA(5)*EXP(ETA(5)) > > V3=THETA(6)*EXP(ETA(6)) > > K10=CL/V1 > > K12=Q2/V1 > > K13=Q3/V1 > > K21=Q2/V2 > > K31=Q3/V3 > > S1=V1 > > IF (DUR.EQ.10) THEN > > ALAG4=THETA(7)*EXP(ETA(7)) > > ELSE > > ALAG4=THETA(8)*EXP(ETA(8)) > > ENDIF > > KTR=THETA(9)*EXP(ETA(9)) > > $DES > > ;DADT(1)=A(2)*K21 + A(3)*K31 - A(1)*(K10+K12+K13) > > ;DADT(2)=A(1)*K12 - A(2)*K21 > > ;DADT(3)=A(1)*K13 - A(3)*K31 > > DADT(1)=A(5)*KTR + A(2)*K21 + A(3)*K31 - A(1)*(K10+K12+K13) > > DADT(2)=A(1)*K12 - A(2)*K21 > > DADT(3)=A(1)*K13 - A(3)*K31 > > DADT(4)=-A(4)*KTR > > DADT(5)=A(4)*KTR - A(5)*KTR > > ;DADT(6)=A(5)*KTR - A(6)*KTR > > ;DADT(7)=A(6)*KTR - A(7)*KTR > > ;DADT(8)=A(7)*KTR - A(8)*KTR > > ;DADT(9)=A(8)*KTR - A(9)*KTR > > $ERROR > > W=1 > > IPRED= -2 > > IF (F.GT.0) IPRED=LOG(F) > > Y=IPRED + ERR(1) > > IRES=DV-IPRED > > IWRES=IRES/W > > $THETA (0, 725) ;CL > > $THETA (0, 238) ;Q2 > > $THETA (0, 2920) ;Q3 > > $THETA (0, 311) ;V1 > > $THETA (0, 38700) ;V2 > > $THETA (0, 39500) ;V3 > > $THETA (0.2167, 0.522,1) ;ALAG 10MIN > > $THETA (0.00833, 0.122, 1) ;ALAG 1MIN > > $THETA (0, 0.720) ;KTR > > $OMEGA BLOCK(6) > > 0.0959 ; ETA CL > > 0.00804 0.0211 ; ETA Q2 > > 0.00631 0.00143 0.166 ; ETA Q3 > > -0.00506 0.00121 -0.0635 0.0491 ; ETA V1 > > 0.00293 -0.00746 0.00708 -0.00315 0.0328 ; ETA V2 > > 0.00836 0.00688 0.00144 -0.00497 0.00741 0.153 ; ETA V3 > > $OMEGA 0.217 ; ETA ALAG 10 > > $OMEGA 0.4 ; ETA ALAG 1 > > $OMEGA 0.0171 ; ETA KTR > > ;$OMEGA (0.0686) ; ETA CL > > ;$OMEGA (0.01) ; ETA Q2 > > ;$OMEGA (0.0251) ; ETA Q3 > > ;$OMEGA (0 FIX) ; ETA V1 > > ;$OMEGA (0.0328) ; ETA V2 > > ;$OMEGA (0 FIX) ; ETA V3 > > ;$OMEGA (0.170) ; ETA ALAG10 > > ;$OMEGA (1.37) ; ETA ALAG1 > > ;$OMEGA (0.00409) ; ETA KTR > > $SIGMA (0.0944) > > $ESTIMATION METHOD=1 PRINT=1 MAX=9999 NOABORT SIG=3 ;POSTHOC INTER > > MSFO=msfo.outputfile > > ;$COVA > > $TABLE ID EVID AMT TIME IPRED IRES > > NOPRINT FILE=AllRecords.txt > > $TABLE ID > > CL Q2 Q3 V1 V2 V3 > > ETA1 ETA2 ETA3 ETA4 ETA5 ETA6 ETA7 ;ETA8 > > FIRSTONLY NOPRINT NOAPPEND FILE=FirstRecords.txt > > Ann > > _______________________________________________________________________ > > *Ann Rigby-Jones PhD MRSC* > Research Fellow in Pharmacokinetics & Pharmacodynamics > > Peninsula College of Medicine & Dentistry > > N31, ITTC Phase 1 > Tamar Science Park > 1 Davy Road > Derriford > Plymouth > PL6 8BX > > Tel: +44 (0) 1752 432014 > _______________________________________________________________________ >

Re: Baseline as a covariate

From: Phil . Lowe Date: August 24, 2010 technical
Hi Pete, It all depends on the distribution in the baseline measurements. For the omalizumab-IgE model, Naoto, Stacey and I used measured baseline IgE as a covariate on IgE turnover. Baseline IgE was not normally distributed and was limited by inclusion criteria. The end result was that the baseline IgE (and bodyweight) predicted IgE production and elimination and the unexplained variation was minimised. However, the upshot of this was that for simulation we could not use a simple normal or log-normal distribution of baseline IgE, so either used the clinical trial database values if we wanted a full population, or produced simulations for narrow ranges of baseline IgE and bodyweight for specific regions of the dosing table. Interestingly, once we had loads of data, we discovered that the baseline IgE was also a significant covariate on the omalizumab-IgE binding constant (binding model equivalent of EC50 or IC50 in an indirect response model). Why is currently unknown, but may represent competition with other IgE binding entities such as FceR1 and CD23, which can vary in their expression levels. Best regards, Phil. Physiologist, biochemist; "Master Modeller" Novartis Pharma AG, WSJ-027.6.25, CH-4056 Basel, Switzerland Phone: +41 61 324 4676; Mobile: +41 79 349 7806; [email protected] "Peter Bonate" <[email protected]> Sent by: [email protected] 11/08/2010 17:50 To "'[email protected]'" <[email protected]> cc Subject [NMusers] Baseline as a covariate I'd like to get the group's opinion on something. I have a pharmacodynamic model and the baseline was shown to be a covariate on one of the model parameters. I was hoping to get general thoughts on the use of the baseline as a covariate. Is there a preference for using the observed baseline vs. NONMEM predicted baseline? And does your opinion change if you have a large residual error? Thanks Pete bonate Peter L. Bonate, PhD, FCP, FAAPS GlaxoSmithKline Clinical Pharmacology, Modeling, and Simulation 5 Moore Drive, 17.2259 Research Triangle Park, NC 27709 phone: 919-483-7534 fax: 919-483-8948 email: [email protected]