Ulrika Wählby, E. Niclas Jonsson and Mats O. Karlsson
Department of Pharmacy, Uppsala University, Sweden
Whenever exploratory methods for detection of covariate relationships are used in population PK/PD modelling, there is a risk of including false ones into the model. The frequency of false relationships will depend on both the data and the method for building the covariate model. Two methods for assessing the risk of false relationship are considered. A basic population model without covariate relationships is first created for the data set. In the first method, a number (here n=30) of new PK (or PD) data sets are simulated based on this model. In the second method, the individual covariate vectors are randomly rearranged between the individuals. This is repeated to generate a number (here n=30) of new data sets. An advantage with this method is that both covariate and PK (or PD) data are as for the original data set. The covariate model building scheme used on the real data set is also used on each of the newly created data sets. The two methods were applied to a data set where covariate model building was accomplished by a recently described method (1) where in a stepwise (sequential forward (p<0.05) and backward (p<0.01)) manner the covariate model is built within NONMEM, considering at each forward step all covariates and all parameters. The pharmacokinetic data resulted from oral dosing in 64 subjects and could be described by a first-order absorption with one-compartment disposition and first-order elimination. Two parameters (CL and V), four continuous and six categorical covariates were considered in the model building.
Although none of the data sets contained true covariate relationships, only 3 (out of 30) of the simulated, and 4 of the rearranged data sets lacked significant covariate relationships in the final model. The frequency of false relationships did not differ markedly between parameters, type of covariate or covariate distribution. Many individual covariates resulted in decreases in the objective function value of more than 20.
Correct information on the risk of false positive covariate relationships should be of importance in the selection of candidate covariates, selection or creation of covariate model building method and covariate model interpretation.
1. E. N. Jonsson and M.O. Karlsson, Automated covariate model building within NONMEM, Pharm. Res. 15:1463-1468, 1998.
Reference: PAGE 8 () Abstr 167 [www.page-meeting.org/?abstract=167]
Poster: oral presentation