Niclas Jonsson and Mats Karlsson
Department of Pharmacy, Uppsala University, Sweden
One of the main benefits of non-linear mixed effects modelling over traditional methods to obtain estimates of population parameters is the possibility to characterise parameter-covariate relationships. Including covariates in the model has the important clinical consequence that the unexplained variability in, e.g. the pharmacokinetics, decreases, which means that it becomes easier to predict a suitable first dose to a patient and that the predictive performance of therapeutic monitoring algorithms is increased. Another important aspect of modelling covariate effects is the possibility to identify patient sub-populations that needs special dosage recommendations.
Even if it is in theory possible to try all covariates, in all possible combinations and in all functional forms in the mixed effects modelling program, it is in practice a formidable and time consuming task and the data analyst therefore usually tries to identify the important covariates prior to testing them in the model. This can be done either graphically, i.e. plotting the parameter vs the covariates, or by stepwise procedures, e.g. stepwise linear regression or stepwise generalised additive modelling (GAM). The graphical approach has the drawback that correlations between covariates will drastically decrease its efficiency, that is, in addition to the covariates that really has an influence on the parameter, we will also find relationships with other, correlated, covariates. The GAM remedies this by considering each covariate when other important covariates are already used to explain the variability in the parameter. On the other hand, the GAM, as all regression methods, can be very dependent on single observations.
In the present paper we propose using a bootstrap resampling method to investigate the properties of the hypervariate distribution of parameters-individuals-covariates. Briefly, create M new data sets by sampling with replacement from the original set of parameter estimates and covariates and run the GAM on all these data sets. From these M GAM runs it is possible to produce various informative statistics on, for example, the selection stability of the covariates, the inclusion/exclusion interactions between covariates and the dependence of the covariate selection on each individual.
The bootstrap resampling method will be demonstrated using a real data set and its advantages and disadvantages in comparison with the other methods to find candidate covariate relationships will be discussed.
Reference: PAGE 6 (1997) Abstr 658 [www.page-meeting.org/?abstract=658]
Poster: poster