Jaap W Mandema(1), Davide Verotta(2), and Lewis B Sheiner(2,3)
(1)Dept. Anesthesia, Stanford University School of Medicine, Stanford, CA 94305-5115, (2)Dept. Pharmacy, School of Pharmacy and (3)Dept. Laboratory Medicine, School of Medicine, University of California, San Francisco, CA 94143.
During the past few years nonlinear mixed effects modeling techniques have become more and more popular to estimate the distribution of pharmacokinetic (PK) and pharmacodynamic (PD) parameters of a drug across the patient population. These methods are aimed to characterize the inter-individual variability in PK and PD parameters in terms of fixed and random effects. The fixed effects characterize the relationships between the PK-PD parameters and patient specific covariates such as age, weight, disease state, concurrent therapy, and so forth. The inter-individual random effects quantify the residual unexplained variability. Appropriate use of these analysis techniques requires that their underlying assumptions are carefully validated. Most of the mixed effects modeling programs available, such as NONMEM, NPML, and NLMIX rely on the statistical principle of maximum likelihood for parameter estimation. It is well known that this method is very sensitive to bizarre observations, i.e, inference based on maximum likelihood can be strongly influenced by only a few cases in the data. Such observations are not necessarily outliers but may be completely appropriate data. Furthermore, such observations may suggest that additional data has to be collected or that the current model is inappropriate. It is therefore very important to detect those individuals or data points that have an unusual large influence on the estimated population parameters. However, such individuals are not easily detected due to the complexity of the models, e.g. the relationships between patient specific covariates and PK parameters are not directly observable from the concentration measurements.
In this paper we describe the use of case-deletion diagnostics to detect influential individuals using four real-data examples and the implementation of these diagnostics in the NONMEM program. The case-deletion diagnostics measure the influence of the removal of the data of one individual on the parameter estimates and standard error of the parameter estimates of the current model. The diagnostics are logical extensions of those used in standard regression analysis (1). These diagnostics appear to be very useful for a critical evaluation of the derived population model and estimated population parameters.
1. Cook RD, Weisberg S. Residuals and influence in regression. Chapman and Hall, New York, 1982.
Reference: PAGE 3 () Abstr 870 [www.page-meeting.org/?abstract=870]
Poster: oral presentation