From: Partha Nandy partha.nandy@bms.com Subject: [NMusers] Distribution of Simulated Cmax Date: Sat, 19 Nov 2005 11:45:15 -0500 Hi All, I am kindly seeking your opinion regarding an observation that I have made on a number of occasions. Here's the scenario: First I first fit a PK model (1-, 2- or 3-compartment & covariates) to the data and after validating the model, use the model to predict the concentration-time profiles. I then take the simulated concentration-time data to compute AUCt, AUCINF, and Cmax, via. non-compartmental approach. Even though the distribution of the PK parameter estimates generated from boot-strap data are similar to the parameter distributions from the original data set, and PPC (posterior predictive check) do not show obvious flaws in the model; most often the distribution of Cmax estimated from the Simulated data sets are quite different from that of the observed/experimental data set (I have used Chi-square and Kolmogorov-smirnov 2-sample test to compare distribution). In fact it appears that the the observed distribution and simulated distribution came from very different populations. In this situation, if one wants to compare treatments using 90% confidence-intervals (80-125%), how valid is the comparison when it appears that the simulated data is from a different population and the observed data from a different population, moreover the CIs from the observed Cmax was not within 80-125%. Have any one faced this dilemma and if so, can any one share their experiences? Any suggestions from the group is appreciated. Kind Regards, Partha _______________________________________________________ From: bulitta@ibmp.osn.de Subject: Re: [NMusers] Distribution of Simulated Cmax Date: Sat, 19 Nov 2005 18:58:46 +0100 Dear Partha, I would start with a bootstrap re-sampling of the observed Cmax's of your original study. If you generate a few thousand bootstrap pseudo-samples of the same sample size as used in your study, you should get a good idea which degree of similarity in the distribution of Cmax you would expect from a population PK model. As long as you have a reasonable number of subjects in your original study, the most adequate population PK model should provide you similar simulated Cmax distributions as the bootstrap pseudo-samples from the original study. Did you check that the extent of absorption was similar, in case you pooled data from more than one study? Which between and within subject variability from ANOVA statistics did you get from the Cmax of your original study? If you have only a few subjects and a large within subject variability (between occasion variability), then such a formulation might fail to be bioequivalent to itself in a bioequivalence trial. Another idea would be to use a different error model around the expected Cmax region. (Have not tried this myself, but might be worth to try.) Best regards Juergen ---------------------------------------------------------------- Juergen Bulitta, M.Sc. Scientific Employee, IBMP Paul-Ehrlich-Str. 19, 90562 Nuernberg-Heroldsberg Germany ---------------------------------------------------------------- _______________________________________________________ From: Pravin Jadhav pravinj@gmail.com Subject: Re: [NMusers] Distribution of Simulated Cmax Date: Sat, 19 Nov 2005 12:55:27 -0500 Hi Partha, This is a classical simulation problem that we have on a few of occasions. You say, PPC did not show obvious flaws in the model. It would really depend on the metric that was used. Please take a look at our publication on the very same topic. Jadhav, P. R.; Gobburu, J.V.S.; A New Equivalence Based Metric for Predictive Check to Qualify Mixed-Effects Models, AAPS Journal, Vol. 7 No. 3 (2005) My initial guess is, the predictive check was done to assess average behavior of the data (Pp in our publication). Take a look at the equivalence based metric that was proposed. This metric would have been able to assess inconsistency between your data and the model (at Cmax). We think, this is one of the prime applications of predictive check, especially when used with the equivalence based metric. It will allow you to locate domains of interest where 'the model fails to reproduce the observed data'- underlying aim of the predictive check. You will need to take a look at the parameters that were used for simulation again. Look for any unusual values of CL, V, Ka etc./unusual combinations of those and then truncate the distributions accordingly (within the limits of the observed data). Hope it helps. Pravin Pravin Jadhav _______________________________________________________ From: Paul Hutson prhutson@pharmacy.wisc.edu Subject: Re: [NMusers] Distribution of Simulated Cmax Date: Sat, 19 Nov 2005 15:10:28 -0600 Partha: Is it tri-exponential decay, or is the third comp the gut depot? Also, how are you calculating the Cmax using a non-compartmental approach? Paul _______________________________________________________ From: Partha Nandy partha.nandy@bms.com Subject: Re: [NMusers] Distribution of Simulated Cmax Date: Sun, 20 Nov 2005 23:17:58 -0500 Hi Paul, It is a tri-exponential decay... As for estimating Cmax, I am feed ing the simulated data into WinNonlin / SAS or S-Plus. Kind Regards, Partha _______________________________________________________