Mats O. Karlsson
Div of Biopharmaceutics and Pharmacokinetics, Faculty of Pharmacy, Uppsala University, Sweden.
Covariate relationships of population models are usually derived from a set of candidate covariates. Compared to the number of covariates assessed (and their relationship between each other), the sample size of population studies is often relatively small. Validation data sets (if present) are seldom of such a size that the appropriateness of the covariate model can be assessed. This has led to a concern about the stability or power of any particular covariate model (e.g. 1). The sample dependence of the covariate model could be studied by making repeated studies in the same population and for each new sample perform the model building process. This is of course in practice impossible. An alternative procedure is to use Bootstrap methods, taking repeated samples (with replacement) from the empirical population (2). Rather than perform the normally elaborate model building procedure we may use a recently described technique, the regression of a general additive model (GAM) on empirical Bayes estimates of the individual parameters (3). The combination of these two techniques allow us to easily obtain covariate models for a large number of sample populations. The differences between covariate models can be used as a measure of the sample dependence of the included covariate relationships. Three areas where this strategy may be useful are : (i) in study design to assess the number of subjects needed to establish a hypothezied covariate relationship, (ii) during model building to make this process more robust, or (iii) as a diagnostic of the stability of the covariate relationships in the final population model.
References
(1) Aarons L. Br. J. Clin. Pharmac. 32:669-670 (1991).
(2) Sauerbrei W, Schumacher M. Statistics in Medicine 11:2093-2109 (1992).
(3) Mandema J, Verotta D and Sheiner L. J. Pharm. Biopharm. 20:511-528 (1992).
Reference: PAGE 3 (1994) Abstr 846 [www.page-meeting.org/?abstract=846]
Poster: poster