The Bayesian modeling of covariates for population pharmacokinetic models

Pharmacokinetic (PK) models describe how the concentrations of a drug and its metabolite vary with time. Population PK models identify and quantify sources of between-individual variability in observed concentrations. Crucial to this aim is the identification of those covariates (i.e., individual-specific characteristics) responsible for explaining the variability. In this article we discuss how covariate modeling can be carried out for population PK models. We argue that the importance of a particular covariate can be discussed only with reference to the specific use for which the model is intended. Covariate modeling is important in population PK studies as it aids in determining dosage recommendations for specific covariate-defined populations. We describe a Bayesian predictive procedure that places covariate modeling in the context of dosage determination. In problems such as these it is crucial to incorporate relevant prior information. For covariate selection we extend the approach of George and McCulloch. The approaches utilize Markov chain Monte Carlo techniques. The methods are illustrated using population PK data from a study of the antibiotic vancomycin in babies. These data are sparse, with just 180 concentrations from 37 babies. Eight covariates are available, from which we construct a covariate model.