VPC, NPC or PPC?

Hi All, We have a dataset with as many dosing (amount and length of infusion) as patients. Once the final model was defined, I have performed a vpc. However, because the dosing are very different between patients, is it relevant to perform vpc or shall we compute npc or ppc? Can somebody explain the basic difference between vpc, npc and ppc and when shall we used one or the other? Last point, how to obtain ppc? Best regards Nicolas Professeur à la Faculté de Médecine de Marseille Laboratoire de Pharmacologie Médicale et Clinique 27 Bd Jean Moulin 13385 Marseille cedex Tél 0491387893 (Hôpital) Tél 0491324456 (Faculté) Fax 0491256526

Nicolas > We have a dataset with as many dosing (amount and length of > infusion) as patients. Once the final model was defined, I > have performed a vpc. However, because the dosing are very > different between patients, is it relevant to perform vpc or > shall we compute npc or ppc? > > > > Can somebody explain the basic difference between vpc, npc > and ppc and when shall we used one or the other? > Despite how it sounds this is not really a simple question. Mostly the purpose of all of these techniques is to assess how well the model describes the data. This can be achieved visually and numerically. If you want your method to have "diagnostic" properties, i.e. an ability to determine where the model may fail, then visual types of checks tend to be more informative. Numerical types of checks really give you an overall feeling of whether your model fits the data but don't often allow you to determine where in particular the model might fail. Numerical checks include PPC and NPDE (and others). PPC is really a Bayesian construct as it checks the posterior distributions of your parameters and hence isn't naturally something that would be performed in an MLE framework. However there have been many examples where PPCs have been performed in NONMEM (publications on this have appeared in JPKPD). PPC is generally used to test a specific (important) feature of the data (hence is generally not diagnostic for the whole model). NPDE provides a more general numerical description of agreement of model and data, but when the statistic is tested it seems to reject most model (hence is not diagnostic). Despite the apparent division into visual and numerical there is no reason why a "VPC" couldn't be expressed numerically as a numerical predictive check and why PPC or NPDE style techniques could not be expressed graphically. We have recently produced examples of visual versions of PPC as a form of visual predictive check for situations similar to what you have described where traditional VPCs don't work well (note we did this in WinBUGS). Steve -- Professor Stephen Duffull Chair of Clinical Pharmacy School of Pharmacy University of Otago PO Box 913 Dunedin New Zealand E: [email protected] P: +64 3 479 5044 F: +64 3 479 7034 Design software: www.winpopt.com

Nicolas I do not know your design, but usually, when you have as many doses as patients, it means that the dose is individualized and controlled by the PK or by the PD effect (e.g., dose up to certain concentration level or up to a certain sedation level). In this case, one has to be careful with any type of predictive check unless your simulation algorithm includes the dose-adjustment scheme used in the actual study. If simulations do not include the same dose adjustment algorithm as the actual study, you may have apparent discrepancy of your simulation results and observed data even if the model is perfect. On the other hand, if indeed the trial was concentration or effect controlled, you may include the same dose adjustment scheme into the simulations, and then use VPC for the entire study. If you provide more details of the design, it would be easier to come up with some reasonable VPC algorithm. PPC can be conducted using nonmem PRIOR subroutine (see Nonmem manual) with priors for population parameters fixed at final estimates, and variability of the population parameters fixed at SE of the final model. Thanks Leonid -------------------------------------- Leonid Gibiansky, Ph.D. President, QuantPharm LLC web: www.quantpharm.com e-mail: LGibiansky at quantpharm.com tel: (301) 767 5566 SIMON Nicolas wrote: > Hi All, > > We have a dataset with as many dosing (amount and length of infusion) as patients. Once the final model was defined, I have performed a vpc. However, because the dosing are very different between patients, is it relevant to perform vpc or shall we compute npc or ppc? > > Can somebody explain the basic difference between vpc, npc and ppc and when shall we used one or the other? > > Last point, how to obtain ppc? > > Best regards > > Nicolas > > Professeur à la Faculté de Médecine de Marseille > > Laboratoire de Pharmacologie Médicale et Clinique > > 27 Bd Jean Moulin > > 13385 Marseille cedex > > Tél 0491387893 (Hôpital) > > Tél 0491324456 (Faculté) > > Fax 0491256526

Hello Regarding the NPDE metrics, at the last PAGE we had a discussion with France Mentré, Emmanuelle Comets and Karl Brendel and agreed that it should be juged on - the visual aspect (distribution of npde) -the mean is not significantly different from 0 -and the variance is not significantly different from 1 However, I experienced a non negligible number of model fittings for which the normality test was OK Perhaps France will further comment on this Saïk Dr Saik URIEN Directeur de Recherche à l' INSERM C.I.C. Mère-enfant Cochin-Necker E.A.3620 - Université Paris Descartes Unité de Recherche Clinique (URC) HOPITAL TARNIER 89 rue d'Assas F75006 PARIS 01 5841 2880 Email : [email protected]

Nicolas, In this case I would not spend too much time on the VPC (hope Nick is not reading this e-mail :) ). Yes, VPC is a quick and convenient way to check how good the model describes inter-subject and intra-subject variability, but there are other ways to assess the same: scatter plot matrix of random effects(should not show any strong ETA-ETA correlations unless they are specified in the model), shrinkage of random effects and residual error(should not be too high), QQ-plots of random effects versus normal distributions (should indicate approximately normal distribution and no obvious outliers). If those plots are fine, one can skip VPC or any other xPC, especially in this case when even applicability of VPC depends on the assumptions (we need to assume that dosing scheme was independent of PK which is unlikely to be the case here: subjects with higher CL are likely to get higher doses in this design). Thanks Leonid -------------------------------------- Leonid Gibiansky, Ph.D. President, QuantPharm LLC web: www.quantpharm.com e-mail: LGibiansky at quantpharm.com tel: (301) 767 5566 SIMON Nicolas wrote: > Hi Leonid, > > Thanks for your help. > The dataset came from anaesthesia where alfentanyl was delivered depending on > the duration of the surgery and the effect required. The first point depended > of the surgeon ability and the latest of the anaesthesiologist feeling... > > Thus I do not see a convenient way to include the dose-adjustment algorithm for the simulation even if I fully agree with your suggestion. Have you an idea? > > Best regards > Nicolas > > -----Message d'origine----- > > De : Leonid Gibiansky [ mailto: [email protected] ] Envoyé : mercredi 25 février 2009 20:59 > > À : SIMON Nicolas > Cc : [email protected] > Objet : [SPAM-APHM] Re: [NMusers] VPC, NPC or PPC? > Importance : Faible > > Nicolas > > I do not know your design, but usually, when you have as many doses as patients, it means that the dose is individualized and controlled by the PK or by the PD effect (e.g., dose up to certain concentration level or up to a certain sedation level). In this case, one has to be careful with any type of predictive check unless your simulation algorithm includes the dose-adjustment scheme used in the actual study. If simulations do not include the same dose adjustment algorithm as the actual study, you may have apparent discrepancy of your simulation results and observed data even if the model is perfect. > > On the other hand, if indeed the trial was concentration or effect controlled, you may include the same dose adjustment scheme into the simulations, and then use VPC for the entire study. > > If you provide more details of the design, it would be easier to come up with some reasonable VPC algorithm. > > PPC can be conducted using nonmem PRIOR subroutine (see Nonmem manual) with priors for population parameters fixed at final estimates, and variability of the population parameters fixed at SE of the final model. > > Thanks > Leonid > > -------------------------------------- > Leonid Gibiansky, Ph.D. > President, QuantPharm LLC > web: www.quantpharm.com > e-mail: LGibiansky at quantpharm.com > tel: (301) 767 5566 > > SIMON Nicolas wrote: > > > Hi All, > > > > We have a dataset with as many dosing (amount and length of infusion) as patients. Once the final model was defined, I have performed a vpc. However, because the dosing are very different between patients, is it relevant to perform vpc or shall we compute npc or ppc? > > > > Can somebody explain the basic difference between vpc, npc and ppc and when shall we used one or the other? > > > > Last point, how to obtain ppc? > > > > Best regards > > > > Nicolas > > > > Professeur à la Faculté de Médecine de Marseille > > > > Laboratoire de Pharmacologie Médicale et Clinique > > > > 27 Bd Jean Moulin > > > > 13385 Marseille cedex > > > > Tél 0491387893 (Hôpital) > > > > Tél 0491324456 (Faculté) > > > > Fax 0491256526