Niclas Jonsson
Department of Pharmacy, Uppsala University, Sweden
Being able to explain the variability of the parameters in a population model using covariates can have important consequences. One is that it might lead to the identification of patient sub-groups that need special dosage recommendations another is that can increase the predictive performance of therapeutic monitoring algorithms.
Covariate model building usually involves the following steps: identification of the candidate covariates (i.e. covariates with a potential to explain variability in the parameters), implementation of these in the model and the testing of their relevance in the full population model.
The identification of candidate covariates can, in general, be done in three ways: (i) trial and error, that is to test all covariates separately in the population modelling program, (ii) graphically, i.e. plotting individual parameter estimates vs covariates and (iii)
using screening/stepwise type of analyses, e.g. stepwise linear and generalized additive modeling (GAM). When implementing the covariate in the population model we have to consider a number of things. The parameterization of the covariate relationships will have impact on the interpretability of the parameters describing the relationship. The functional form of the covariate relationship, i.e. a linear or non-linear relationship is also important. In addition we have the possibility of interactions between covariates,
e.g. that the relationship between CL and CRCL can be different for males and females. The significance of a certain covariate will depend on the objectives with the study. In an early, hypothesis generating study we will typically be interested in all, including fairly weak, covariate effects while in later studies we are usually more interested in the covariates that has a clinically significant explanatory power. In addition to the usual ways to determine the degree of significance of a covariate (e.g. graphical analysis and
reduction of the variability when the covariate is included in the model) we can also perform deletion diagnostics, i.e. omitting one or a group of individuals at a time and refit the model, to investigate the possible existence of influential individuals on which the covariate model is highly dependent on.
The above issues are by no means resolved in the literature and it is doubtful whether a really conclusive recommendation on these issues can be made. The aim of the present paper is to present a review of this topic.
Reference: PAGE 6 (1997) Abstr 601 [www.page-meeting.org/?abstract=601]
Poster: oral presentation