From: "Justin Wilkins"Subject: [NMusers] Poor PRED estimates Date: Thu, 22 May 2003 18:16:11 +0200 Hi all, I'm working on a model for rifampicin in healthy volunteers, using ADVAN4, and I'm having a great deal of trouble getting credible population predictions - while the population DVs often go as high as 25-30 ug/mL, population predictions never go above 15 ug/mL, giving a truncated, cone-shaped DV vs PRED plot that suggests something exponential is wrong. IPRED values seem all right though. The data is rich - 300 individuals sampled at least 8 times on each occasion, and each subject being sampled on at least 2 occasions - and I've been able to get rough parameter estimates from a noncompartmental analysis of the information. I only have plasma data for V2, which means one wouldn't expect much joy from V3 or Q, but I strongly suspect a two-compartment model is appropriate, especially since ADVAN2 PRED results are extremely poor. Being a relative NONMEM neophyte, I'm stumped. Does anyone have any advice? Control stream follows. Thanks in advance! Justin _______________________________________________________ From: "Gobburu, Jogarao V" Subject: RE: [NMusers] Poor PRED estimates Date: Thu, 22 May 2003 13:04:23 -0400 Hello, It is going to be difficult to guess what the problem might be. But I will anyways. It appears from the stumped pop predictions that the absorption model might need improvement. You might want to test if the models in the following article could be of help to you. Rich absorption phase data could be challenging sometimes. Holford NH, Ambros RJ, Stoeckel K. Models for describing absorption rate and estimating extent of bioavailability: application to cefetamet pivoxil.J Pharmacokinet Biopharm 1992 Oct;20(5):421-42 Regards, Joga Gobburu, Pharmacometrics, CDER/FDA. _______________________________________________________ From:Leonid Gibiansky Subject: Re: [NMusers] Poor PRED estimates Date: Thu, 22 May 2003 13:13:01 -0400 I think you do not need DEL here, use $ERROR IPRED=F W=SQRT(THETA(7)**2+THETA(8)**2*F*F) Y=IPRED+W*EPS(1) But this should not be a problem. I would try FOCE (METHOD=1 INTERACTION). Also, sometimes log-transformation of the data is helpful (DV ---> log(DV) in the data file, $ERROR BLQ=0.01 ; something close to the LLQ IPRED = BLQ/2 ; IF(F.GT.BLQ/2) IPRED= F W=SQRT(THETA(7)**2/IPRED/IPRED+THETA(8)**2) Y=LOG(IPRED)+W*EPS(1) ) How WRES vs PRED plot looks like ? How large are OMEGAs and SIGMA ? Leonid _______________________________________________________ From: "Hutmacher, Matthew [Non-Employee/1820]" Subject:RE: [NMusers] Poor PRED estimates Date: Thu, 22 May 2003 12:50:02 -0500 Justin, In addition to the comment below, and without much knowledge at all about you problem, my guess is to look at your starting values and maybe their units. You have your initial estimate of CL as 21 and V2 as 0.7. This translates into k20~30 and thus ln(2)/k20 ~ 1.4 minutes. Matt _______________________________________________________ From:Mats Karlsson Subject: Re: [NMusers] Poor PRED estimates Date: Fri, 23 May 2003 06:04:35 +0200 Dear Justin, One possibility is that there is nothing wrong at all with your fit. If you have no covariates in your model, you would expect the plot you describe. There should always be observed concentrations higher than PRED. The question is whether PRED is giving unbiased predictions of you observed data. As a rough guide, if you plot DV versus PRED (or log(DV) versus log(PRED) is your data spans several magnitudes) a regression line should have a slope close to unity. As IPRED looks good, you can also inspect if all your etas are centered around zero (which they should be). Best regards, Mats _______________________________________________________ From: "Justin Wilkins" Subject: RE: [NMusers] Poor PRED estimates Date:Fri, 23 May 2003 11:05:46 +0200 Thanks everyone for their input, I'll have another go today. Leonid, you asked for PRED vs WRES and the OMEGAs and SIGMA - WFN output follows. I'm sending a JPG of the Xpose plot directly to you - if anyone wants to have a look please let me know and I'll post it to the list. THETA: KA CL V2 V3 Q ALAG1 ADD EXP ETA: ERR: rif_comm_27_e2.lst 19161.187 eval=1221 sig=+5.7 sub=300 obs=7291 CCIL=NNNN NV1.1 PIV1.1 THETA = 0.449 8.81 5.59 16.1 6.37 0.191 0.509 0.287 ETASD = 0.505964 0.384708 0.766159 0.158745 ERRSD = 1 THETA:se% = 10.9 2.4 7.0 9.3 13.1 6.1 12.8 4.4 OMEGA:se% = 33.6 12.2 32.7 67.9 SIGMA:se% = 0.0 MINIMIZATION SUCCESSFUL _______________________________________________________ From: Leonid Gibiansky Subject:RE: [NMusers] Poor PRED estimates Date:Fri, 23 May 2003 09:36:33 -0400 Justin Looking on the WRES vs. PPED plot (slightly skewed that suggests that log-transformation may help) and OMEGAs, SIGMA below, I would agree with Mats that there is nothing really wrong with your fit. Everything looks reasonable, except too large OMEGA3 (an may be OMEGA1, but OMEGA1 describes KA, and it is reasonable for KA to have large variability). It does not mean that you cannot improve the model. I would try log transformation and FOCE INTERACTION, often they improve variability estimates. FOCE is known to give better results if OMEGAs are high (as in your case) and with the rich data (again, this is your case). Certainly, I would check whether distributions of random effects are centered around zero. If not, then again, log-transformation and FOCE could help. One the other note, if you have problems with Cmax, sometimes zero-order absorption may help to better estimate it (but you should be able to see this on the DV vs. time plot) Leonid. _______________________________________________________ From:VPIOTROV@PRDBE.jnj.com Subject: RE: [NMusers] Poor PRED estimates Date:Mon, 26 May 2003 15:24:35 +0200 After log transformation the residual error often becomes additive (constant variance). In this case just FOCE method is sufficient. ITERATION option in $EST statement is not needed. Best regards, Vladimir _______________________________________________________ From: Leonid Gibiansky Subject: RE: [NMusers] Poor PRED estimates Date:Tue, 27 May 2003 09:13:12 -0400 Actually, the ERROR model was more complicated: W=SQRT(THETA(7)**2/IPRED/IPRED+THETA(8)**2) Y=LOG(IPRED)+W*EPS(1) so INTERACTION is still helpful. If W=const, then it can be removed (but it never hurts, as far as I understood, and does not add CPU time) Leonid _______________________________________________________