RE: [NMusers] Additive plus proportional error model for log-transform data
Dear Ahmad, You don't havet o choose between normal or transformed concentrations in your error model, you can let NONMEM estimate the most appropriate transformation for you. Combining this with a power transform error model I think is likely to solve your problem. See A strategy for residual error modeling incorporating scedasticity of variance and distribution shape. Dosne AG, Bergstrand M, Karlsson MO. J Pharmacokinet Pharmacodyn. 2016 Apr;43(2):137-51. doi: 10.1007/s10928-015-9460-y. Epub 2015 Dec 17. It is automated in PsN as "execute -dtbs ..." Besst regaards, Mats Mats Karlsson, PhD Professor of Pharmacometrics Dept of Pharmaceutical Biosciences Faculty of Pharmacy Uppsala University Box 591 75124 Uppsala Phone: +46 18 4714105 Fax + 46 18 4714003 www.farmbio.uu.se/research/researchgroups/pharmacometrics/<http://www.farmbio.uu.se/research/researchgroups/pharmacometrics/> From: owner-nmus...@globomaxnm.com [mailto:owner-nmus...@globomaxnm.com] On Behalf Of Jakob Ribbing Sent: Thursday, June 02, 2016 6:32 AM To: Abu Helwa, Ahmad Yousef Mohammad - abuay010 Cc: nmusers@globomaxnm.com Subject: Re: [NMusers] Additive plus proportional error model for log-transform data Hi Ahmad, The two error models are equivalent (only that with Leonids suggested code, the additive-on-log-transformed error term (TH16) is estimated on variance scale, instead of standard deviation scale (approximate CV). This inflated error rates for very low concentrations is what you get for additive+proportional on the log transformed scale, and I believe that has been discussed on nmusers previously as well, many years ago. You could possibly use a cut-off for when lower IPRE should not lead to higher residual errors, but why not move to additive + proportional for the original concentration scale? Also, this implementation may be unfortunate: Y=(1-FLAG)*IPRE + W*EPS(1) Effectively, when concentration predictions are zero (FLAG=1), e.g. for pre-dose samples or before commence of absorption, then you set the concentration prediction to EXP(1)=3.14 concentration units. Depending on what concentration scale you work on (i.e. if BLQ is much higher than this) it may be OK, but otherwise not. Instead of applying a flag, just set IPRE to a negative value (low in relation to LOG(BLQ)), if you want to stay on the log-transformed scale. I hope this helps to solve your problem. Best regards Jakob Jakob Ribbing, Ph.D. Senior Consultant, Pharmetheus AB Cell/Mobile:+46 (0)70 514 33 77 jakob.ribb...@pharmetheus.com<mailto:jakob.ribb...@pharmetheus.com> www.pharmetheus.com<http://www.pharmetheus.com/> Phone, Office:+46 (0)18 513 328 Uppsala Science Park, Dag Hammarskjölds väg 52B SE-752 37 Uppsala, Sweden This communication is confidential and is only intended for the use of the individual or entity to which it is directed. It may contain information that is privileged and exempt from disclosure under applicable law. If you are not the intended recipient please notify us immediately. Please do not copy it or disclose its contents to any other person. On 02 Jun 2016, at 04:27, Abu Helwa, Ahmad Yousef Mohammad - abuay010 <ahmad.abuhe...@mymail.unisa.edu.au<mailto:ahmad.abuhe...@mymail.unisa.edu.au>> wrote: Dear NMusers, I am developing a PK model using log-transformed single-dose oral data. My question relates to using combined error model for log-transform data. I have read few previous discussions on NMusers regarding this, which were really helpful, and I came across two suggested formulas (below) that I tested in my PK models. Both formulas had similar model fits in terms of OFV (OFV using Formula 2 was one unit less than OFV using Formula1) with slightly changed PK parameter estimates. My issue with these formulas is that the model simulates very extreme concentrations (e.g. upon generating VPCs) at the early time points (when drug concentrations are low) and at later time points when the concentrations are troughs. These simulated extreme concentrations are not representative of the model but a result of the residual error model structure. My questions: 1. Is there a way to solve this problem for the indicated formulas? 2. Are the two formulas below equally valid? 3. Is there an alternative formula that I can use which does not have this numerical problem? 4. Any reference paper that discusses this subject? Here are the two formulas: 1. Formula 1: suggested by Mats Karlsson with fixing SIGMA to 1: W=SQRT(THETA(16)**2+THETA(17)**2/EXP(IPRE)**2) 2. Formula 2: suggested by Leonid Gibiansky with fixing SIGMA to 1: W = SQRT(THETA(16)+ (THETA(17)/EXP(IPRE))**2 ) The way I apply it in my model is this: FLAG=0 ;TO AVOID ANY CALCULATIONS OF LOG (0) IF (F.EQ.0) FLAG=1 IPRE=LOG(F+FLAG) W=SQRT(THETA(16)**2+THETA(17)**2/EXP(IPRE)**2) ;FORMULA 1 IRES=DV-IPRE IWRES=IRE
Re: [NMusers] Additive plus proportional error model for log-transform data
I also like this version: W = SDL-(SDL-SDH)*TY/(SD50+TY) Y=LTY+W*EPS(1) Here SDL is the standard deviation (in logs) at low concentrations, SDH is the standard deviation at high concentrations, TY is the individual prediction, LTY is LOG(TY). SIGMA should be fixed at 1 Leonid On Wed, Jun 1, 2016 at 10:27 PM, Abu Helwa, Ahmad Yousef Mohammad - abuay010wrote: > Dear NMusers, > > > > I am developing a PK model using log-transformed single-dose oral data. My > question relates to using combined error model for log-transform data. > > > > I have read few previous discussions on NMusers regarding this, which were > really helpful, and I came across two suggested formulas (below) that I > tested in my PK models. Both formulas had similar model fits in terms of > OFV (OFV using Formula 2 was one unit less than OFV using Formula1) with > slightly changed PK parameter estimates. My issue with these formulas is > that the model simulates very extreme concentrations (e.g. upon generating > VPCs) at the early time points (when drug concentrations are low) and at > later time points when the concentrations are troughs. These simulated > extreme concentrations are not representative of the model but a result of > the residual error model structure. > > > > My questions: > > 1. Is there a way to solve this problem for the indicated formulas? > > 2. Are the two formulas below equally valid? > > 3. Is there an alternative formula that I can use which does not > have this numerical problem? > > 4. Any reference paper that discusses this subject? > > > > Here are the two formulas: > > 1. Formula 1: suggested by Mats Karlsson with fixing SIGMA to 1: > > W=SQRT(THETA(16)**2+THETA(17)**2/EXP(IPRE)**2) > > > > 2. Formula 2: suggested by Leonid Gibiansky with fixing SIGMA to 1: > > W = SQRT(THETA(16)+ (THETA(17)/EXP(IPRE))**2 ) > > > > The way I apply it in my model is this: > > > > FLAG=0 ;TO AVOID ANY CALCULATIONS OF LOG (0) > > IF (F.EQ.0) FLAG=1 > > IPRE=LOG(F+FLAG) > > > > W=SQRT(THETA(16)**2+THETA(17)**2/EXP(IPRE)**2) ;FORMULA 1 > > > > IRES=DV-IPRE > > IWRES=IRES/W > > Y=(1-FLAG)*IPRE + W*EPS(1) > > > > $SIGMA > > 1. FIX > > > > Best regards, > > > > Ahmad Abuhelwa > > School of Pharmacy and Medical Sciences > > University of South Australia- City East Campus > > Adelaide, South Australia > > Australia > > > -- -- Leonid Gibiansky, Ph.D. President, QuantPharm LLC web:www.quantpharm.com e-mail: LGibiansky at quantpharm.com tel:(301) 767 5566
Re: [NMusers] Additive plus proportional error model for log-transform data
Hi Ahmad, This issue hasbeen discussed a lot and I'm afraid there's no consensus yet. To your question: 1. Is there away to solve this problem for the indicated formulas? As you said, thisproblem occurs at the early/later time points. In other words, it happenswhen prediction is relatively low. This is because there's an underlyingapproximation in the derivation of the two formulas: Innon-transformed terms Conc=F*EXP(SQRT(THETA(x)**2+THETA(y)**2/F**2)*EPS(1)) Assuming thatEXP(x)=1+x (for small x), you get Conc=F*(1+SQRT(THETA(x)**2+THETA(y)**2/F**2)*EPS(1)) Variance of thisexpression Var(conc)=F**2*(THETA(x)**2+THETA(y)**2/F**2)= = F**2*THETA(x)**2+THETA(y)**2 On the otherhand, for the error model Y=Fexp(EPS1)+EPS2=F(1+EPS1)+EPS2 variance isequal to F**2*OMEGA1+OMEGA2 Thus, thesemodels are similar if not identical with OMEGA1=THETA(x)**2, OMEGA2=THETA(y)**2 --- So when F is rathersmall, the approximation of exp(x)=1+x doesn't workanymore. This may not be a problem when fitting your data.It may only occurwhen the prediction is extremely low, let say 10^-4. In opinion it's safe touse this formula in most cases. But if you are seeking a perfect answer,Jacob's suggestion may be the one. 2. Are the twoformulas below equally valid? As far asI'm concerned, these two are equally valid. I question the result that there'sa difference in OFV and estimates between the two. 3. Is there analternative formula that I can use which does not have this numerical problem? Maybe youcan try Stu's "double exponential error model": Y = LOG(F+M) +(F/(F+M))*ERR(1) + (M/(F+M))*ERR(2). Best regards, Rong Chen School ofPharmaceutical Science Peking University Beijing, China From: "Abu Helwa, Ahmad Yousef Mohammad - abuay010" <ahmad.abuhe...@mymail.unisa.edu.au> To: "nmusers@globomaxnm.com" <nmusers@globomaxnm.com> Sent: Thursday, 2 June 2016, 10:27 Subject: [NMusers] Additive plus proportional error model for log-transform data #yiv5739009609 #yiv5739009609 -- filtered {panose-1:2 4 5 3 5 4 6 3 2 4;}#yiv5739009609 filtered {font-family:Calibri;panose-1:2 15 5 2 2 2 4 3 2 4;}#yiv5739009609 p.yiv5739009609MsoNormal, #yiv5739009609 li.yiv5739009609MsoNormal, #yiv5739009609 div.yiv5739009609MsoNormal {margin:0cm;margin-bottom:.0001pt;font-size:11.0pt;}#yiv5739009609 a:link, #yiv5739009609 span.yiv5739009609MsoHyperlink {color:#0563C1;text-decoration:underline;}#yiv5739009609 a:visited, #yiv5739009609 span.yiv5739009609MsoHyperlinkFollowed {color:#954F72;text-decoration:underline;}#yiv5739009609 p.yiv5739009609MsoListParagraph, #yiv5739009609 li.yiv5739009609MsoListParagraph, #yiv5739009609 div.yiv5739009609MsoListParagraph {margin-top:0cm;margin-right:0cm;margin-bottom:0cm;margin-left:36.0pt;margin-bottom:.0001pt;font-size:11.0pt;}#yiv5739009609 span.yiv5739009609EmailStyle17 {color:windowtext;}#yiv5739009609 .yiv5739009609MsoChpDefault {font-size:10.0pt;}#yiv5739009609 filtered {margin:72.0pt 72.0pt 72.0pt 72.0pt;}#yiv5739009609 div.yiv5739009609WordSection1 {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 filtered {}#yiv5739009609 ol {margin-bottom:0cm;}#yiv5739009609 ul {margin-bottom:0cm;}#yiv5739009609 Dear NMusers, I am developing a PK model using log-transformed single-dose oral data. My question relates to using combined error model for log-transform data. I have read few previous discussions on NMusers regarding this, which were really helpful, and I came across two suggested formulas (below) that I tested in my PK models. Both formulas had similar model fits in terms of OFV (OFV using Formula 2 was one unit less than OFV using Formula1) with slightly changed PK parameter estimates. My issue with these formulas is that the model simulates very extreme concentrations (e.g. upon generating VPCs) at the early time points (when drug concentrations are low) and at later time points when the concentrations are troughs. These simulated extreme concentrations are not representative of the model but a result of the residual error model structure. My questions: 1. Is there a way to solve this problem for the indicated fo
Re: [NMusers] Additive plus proportional error model for log-transform data
Sorry, an error in what I wrote below: It should be EXP(0)=1 concentration unit Jakob Ribbing, Ph.D. Senior Consultant, Pharmetheus AB Cell/Mobile:+46 (0)70 514 33 77 jakob.ribb...@pharmetheus.com www.pharmetheus.com Phone, Office: +46 (0)18 513 328 Uppsala Science Park, Dag Hammarskjölds väg 52B SE-752 37 Uppsala, Sweden This communication is confidential and is only intended for the use of the individual or entity to which it is directed. It may contain information that is privileged and exempt from disclosure under applicable law. If you are not the intended recipient please notify us immediately. Please do not copy it or disclose its contents to any other person. On 02 Jun 2016, at 06:31, Jakob Ribbingwrote: > Hi Ahmad, > > The two error models are equivalent (only that with Leonids suggested code, > the additive-on-log-transformed error term (TH16) is estimated on variance > scale, instead of standard deviation scale (approximate CV). > This inflated error rates for very low concentrations is what you get for > additive+proportional on the log transformed scale, and I believe that has > been discussed on nmusers previously as well, many years ago. > You could possibly use a cut-off for when lower IPRE should not lead to > higher residual errors, but why not move to additive + proportional for the > original concentration scale? > > Also, this implementation may be unfortunate: >> Y=(1-FLAG)*IPRE + W*EPS(1) > Effectively, when concentration predictions are zero (FLAG=1), e.g. for > pre-dose samples or before commence of absorption, then you set the > concentration prediction to EXP(1)=3.14 concentration units. > > Depending on what concentration scale you work on (i.e. if BLQ is much higher > than this) it may be OK, but otherwise not. > Instead of applying a flag, just set IPRE to a negative value (low in > relation to LOG(BLQ)), if you want to stay on the log-transformed scale. > > I hope this helps to solve your problem. > > Best regards > > Jakob > > > > Jakob Ribbing, Ph.D. > > Senior Consultant, Pharmetheus AB > > > > Cell/Mobile: +46 (0)70 514 33 77 > > jakob.ribb...@pharmetheus.com > > www.pharmetheus.com > > > > Phone, Office:+46 (0)18 513 328 > > Uppsala Science Park, Dag Hammarskjölds väg 52B > > SE-752 37 Uppsala, Sweden > > > > This communication is confidential and is only intended for the use of the > individual or entity to which it is directed. It may contain information that > is privileged and exempt from disclosure under applicable law. If you are not > the intended recipient please notify us immediately. Please do not copy it or > disclose its contents to any other person. > > > > > > On 02 Jun 2016, at 04:27, Abu Helwa, Ahmad Yousef Mohammad - abuay010 > wrote: > >> Dear NMusers, >> >> I am developing a PK model using log-transformed single-dose oral data. My >> question relates to using combined error model for log-transform data. >> >> I have read few previous discussions on NMusers regarding this, which were >> really helpful, and I came across two suggested formulas (below) that I >> tested in my PK models. Both formulas had similar model fits in terms of >> OFV (OFV using Formula 2 was one unit less than OFV using Formula1) with >> slightly changed PK parameter estimates. My issue with these formulas is >> that the model simulates very extreme concentrations (e.g. upon generating >> VPCs) at the early time points (when drug concentrations are low) and at >> later time points when the concentrations are troughs. These simulated >> extreme concentrations are not representative of the model but a result of >> the residual error model structure. >> >> My questions: >> 1. Is there a way to solve this problem for the indicated formulas? >> 2. Are the two formulas below equally valid? >> 3. Is there an alternative formula that I can use which does not have >> this numerical problem? >> 4. Any reference paper that discusses this subject? >> >> Here are the two formulas: >> 1. Formula 1: suggested by Mats Karlsson with fixing SIGMA to 1: >> W=SQRT(THETA(16)**2+THETA(17)**2/EXP(IPRE)**2) >> >> 2. Formula 2: suggested by Leonid Gibiansky with fixing SIGMA to 1: >> W = SQRT(THETA(16)+ (THETA(17)/EXP(IPRE))**2 ) >> >> The way I apply it in my model is this: >> >> FLAG=0 ;TO AVOID ANY CALCULATIONS OF LOG (0) >> IF (F.EQ.0) FLAG=1 >> IPRE=LOG(F+FLAG) >> >> W=SQRT(THETA(16)**2+THETA(17)**2/EXP(IPRE)**2) ;FORMULA 1 >> >> IRES=DV-IPRE >> IWRES=IRES/W >> Y=(1-FLAG)*IPRE + W*EPS(1) >> >> $SIGMA >> 1. FIX >> >> Best regards, >> >> Ahmad Abuhelwa >> School of Pharmacy and Medical Sciences >> University of South Australia- City East Campus >> Adelaide, South Australia >> Australia >
Re: [NMusers] Additive plus proportional error model for log-transform data
Hi Ahmad, The two error models are equivalent (only that with Leonids suggested code, the additive-on-log-transformed error term (TH16) is estimated on variance scale, instead of standard deviation scale (approximate CV). This inflated error rates for very low concentrations is what you get for additive+proportional on the log transformed scale, and I believe that has been discussed on nmusers previously as well, many years ago. You could possibly use a cut-off for when lower IPRE should not lead to higher residual errors, but why not move to additive + proportional for the original concentration scale? Also, this implementation may be unfortunate: > Y=(1-FLAG)*IPRE + W*EPS(1) Effectively, when concentration predictions are zero (FLAG=1), e.g. for pre-dose samples or before commence of absorption, then you set the concentration prediction to EXP(1)=3.14 concentration units. Depending on what concentration scale you work on (i.e. if BLQ is much higher than this) it may be OK, but otherwise not. Instead of applying a flag, just set IPRE to a negative value (low in relation to LOG(BLQ)), if you want to stay on the log-transformed scale. I hope this helps to solve your problem. Best regards Jakob Jakob Ribbing, Ph.D. Senior Consultant, Pharmetheus AB Cell/Mobile:+46 (0)70 514 33 77 jakob.ribb...@pharmetheus.com www.pharmetheus.com Phone, Office: +46 (0)18 513 328 Uppsala Science Park, Dag Hammarskjölds väg 52B SE-752 37 Uppsala, Sweden This communication is confidential and is only intended for the use of the individual or entity to which it is directed. It may contain information that is privileged and exempt from disclosure under applicable law. If you are not the intended recipient please notify us immediately. Please do not copy it or disclose its contents to any other person. On 02 Jun 2016, at 04:27, Abu Helwa, Ahmad Yousef Mohammad - abuay010wrote: > Dear NMusers, > > I am developing a PK model using log-transformed single-dose oral data. My > question relates to using combined error model for log-transform data. > > I have read few previous discussions on NMusers regarding this, which were > really helpful, and I came across two suggested formulas (below) that I > tested in my PK models. Both formulas had similar model fits in terms of OFV > (OFV using Formula 2 was one unit less than OFV using Formula1) with slightly > changed PK parameter estimates. My issue with these formulas is that the > model simulates very extreme concentrations (e.g. upon generating VPCs) at > the early time points (when drug concentrations are low) and at later time > points when the concentrations are troughs. These simulated extreme > concentrations are not representative of the model but a result of the > residual error model structure. > > My questions: > 1. Is there a way to solve this problem for the indicated formulas? > 2. Are the two formulas below equally valid? > 3. Is there an alternative formula that I can use which does not have > this numerical problem? > 4. Any reference paper that discusses this subject? > > Here are the two formulas: > 1. Formula 1: suggested by Mats Karlsson with fixing SIGMA to 1: > W=SQRT(THETA(16)**2+THETA(17)**2/EXP(IPRE)**2) > > 2. Formula 2: suggested by Leonid Gibiansky with fixing SIGMA to 1: > W = SQRT(THETA(16)+ (THETA(17)/EXP(IPRE))**2 ) > > The way I apply it in my model is this: > > FLAG=0 ;TO AVOID ANY CALCULATIONS OF LOG (0) > IF (F.EQ.0) FLAG=1 > IPRE=LOG(F+FLAG) > > W=SQRT(THETA(16)**2+THETA(17)**2/EXP(IPRE)**2) ;FORMULA 1 > > IRES=DV-IPRE > IWRES=IRES/W > Y=(1-FLAG)*IPRE + W*EPS(1) > > $SIGMA > 1. FIX > > Best regards, > > Ahmad Abuhelwa > School of Pharmacy and Medical Sciences > University of South Australia- City East Campus > Adelaide, South Australia > Australia
[NMusers] Additive plus proportional error model for log-transform data
Dear NMusers, I am developing a PK model using log-transformed single-dose oral data. My question relates to using combined error model for log-transform data. I have read few previous discussions on NMusers regarding this, which were really helpful, and I came across two suggested formulas (below) that I tested in my PK models. Both formulas had similar model fits in terms of OFV (OFV using Formula 2 was one unit less than OFV using Formula1) with slightly changed PK parameter estimates. My issue with these formulas is that the model simulates very extreme concentrations (e.g. upon generating VPCs) at the early time points (when drug concentrations are low) and at later time points when the concentrations are troughs. These simulated extreme concentrations are not representative of the model but a result of the residual error model structure. My questions: 1. Is there a way to solve this problem for the indicated formulas? 2. Are the two formulas below equally valid? 3. Is there an alternative formula that I can use which does not have this numerical problem? 4. Any reference paper that discusses this subject? Here are the two formulas: 1. Formula 1: suggested by Mats Karlsson with fixing SIGMA to 1: W=SQRT(THETA(16)**2+THETA(17)**2/EXP(IPRE)**2) 2. Formula 2: suggested by Leonid Gibiansky with fixing SIGMA to 1: W = SQRT(THETA(16)+ (THETA(17)/EXP(IPRE))**2 ) The way I apply it in my model is this: FLAG=0 ;TO AVOID ANY CALCULATIONS OF LOG (0) IF (F.EQ.0) FLAG=1 IPRE=LOG(F+FLAG) W=SQRT(THETA(16)**2+THETA(17)**2/EXP(IPRE)**2) ;FORMULA 1 IRES=DV-IPRE IWRES=IRES/W Y=(1-FLAG)*IPRE + W*EPS(1) $SIGMA 1. FIX Best regards, Ahmad Abuhelwa School of Pharmacy and Medical Sciences University of South Australia- City East Campus Adelaide, South Australia Australia