Dimitra Nikopoulou, Aris Dokoumetzidis
School of Pharmacy, University of Athens, Greece
Objectives: In a badly designed study, a NONMEM run often gives not only large standard errors, but also biased estimates and an overall unstable performance. A methodology is developed based on a parameter scan using a series of informative Bayesian priors that allows the location of bias in NONMEM parameter estimates.
Methods: Various scenarios of datasets were simulated using a one-compartment pharmacokinetic model with first order absorption, with 40 subjects and 2 points per subject, where the last sampling time was deliberately early, such that it did not allow an accurate estimation of clearance (CL). NONMEM with FOCE was used for parameter estimation while their standard errors were estimated by bootstrapping. Nine different percentiles (10% to 90% increasing by 10%) of the bootstrap distribution of CL were determined. Informative Bayesian priors were setup using the $PRIOR option of NONMEM where the value of prior for CL was set to each of the percentiles of its distribution while the variance of the prior was chosen at 30% CV for CL and noninformative for all other parameters (other options were also investigated). The criterion for determining bias and for selecting the percentile most likely to be closest to the correct parameter value was considered to be the one with the lowest value of the Objective Function (OF).
Results: For most scenarios a plot of the OF vs the corresponding percentile of the prior exhibited a smooth minimum at the correct percentile (corresponding to the simulated values) which was in some cases different from the median. For an extreme where the correct simulated value fell at the 90th percentile of the bootstrap distribution the method reached its limits and was difficult to determine a trend in the OF vs the percentiles. Sensitivity of the method from the strength of the prior was also investigated.
Conclusions: A method of scanning along a posterior distribution of a badly estimated parameter by using a series of informative Bayesian priors can locate parameter values with lower OF than the median which are closer to the unbiased value of this parameter.
Reference: PAGE 22 (2013) Abstr 2840 [www.page-meeting.org/?abstract=2840]
Poster: Estimation methods