Automated covariate selection and Bayesian model averaging in population PK/PD models

We illustrate the use of ‘reversible jump’ MCMC to automate the process of covariate selection in population PK/PD analyses. The output from such an approach can be used not only to determine the ‘best’ covariate model for each parameter, but also to formally measure the spread of uncertainty across all possible models, and to average inferences across a range of ‘good’ models. We examine the substantive impact of such model averaging compared to conditioning inferences on the ‘best’ model alone, and conclude that clinically significant differences between the two approaches can arise. The illustrative data that we consider pertain to the drug vancomycin in 59 neonates and infants, and all analyses are conducted using the WinBUGS software with newly developed ‘Jump’ interface installed.