Re: Simulation vs. actual data

From: "Nick Holford" n.holford@auckland.ac.nz Subject: Re: [NMusers] Simulation vs. actual data Date: Tue, July 12, 2005 3:59 pm Juanjo, Thanks for your comments and for the reference to your paper on pegylated EPO PK. In this paper you describe the process for generating a tolerance interval as follows: "In addition, the final model was used to simulate the pharmacokinetic profile of PEG-EPO after i.v. and s.c. administration of a single dose of 1000 ug to 100 male subjects...Uncertainty in fixed and random parameter estimates was considered during the simulations by replicating the above-mentioned process 30 times using different values of fixed and random parameters, randomly selected from the estimates of the bootstrap replicates." If I understand you correctly this means simulating profiles in 3000 subjects with each block of 100 subjects having a different set of fixed and random effect parameters sampled from the non-parametric bootstrap. You then used the empirical distribution of 3000 responses at each time point to construct the interval. The experience you mention below agrees with the theoretical expectations I mentioned for the tolerance interval compared with the prediction interval but I cannot find anything in your paper to support your statement experimentally. Your Figure 4 shows 90% tolerance intervals for extrapolations to humans from non-human species. It has no observations for a visual predictive check. Did you try generating tolerance and prediction intervals for the species you had data for? If so, what fraction of observed values would lie within the tolerance interval you generated and what fraction would lie within a prediction interval generated with a single set of fixed and random effects parameters e.g. using the final model parameter estimates? I am interested in getting a concrete example of just how much different the tolerance and prediction intervals might be for a real NONMEM analysis. On another topic - I note your final model used an empirical allometric model using a combination of total weight and brain weights. The brain weights were in fact just total weight multiplied by an empirical adjustment factor and not truly individual brain weights. This makes it impossible to assert that brain weight itself is an independent covariate for the prediction of clearance. As there is no plausible mechanistic reason to believe that the brain is an important clearance organ for EPO I would think the apparent benefit of adding 'brain weight' to the model is an illustration of selection bias (Ribbing J, Jonsson EN. Power, Selection Bias and Predictive Performance of the Population Pharmacokinetic Covariate Model. Journal of Pharmacokinetics and Pharmacodynamics 2004;31(2):109-134) You did not report results for a simple allometric model based on body weight alone with the theoretical exponents of 3/4 for CL and Q and 1 for V1 and V2. xYou report a reference model with estimated allometric exponents. Did you try using a simple allometric model and test if the CI for the exponents were different from the theoretical values? Nick