Prior Information into the Model

Pharmacokinetic-pharmacodynamic models are becoming increasingly complex through the incorporation of covariate information, effect mechanisms, and lower quantification limits of assays which reveal compartments not previously known to exist. With sparse data collected during Phase 3 trials it may not be possible to develop and support such complex models because of 

One difficulty with the frequentist prior approach is the choice of the pdf for the model parameters. A common first choice would be to assume that the fixed effects (0 which is of size p) are multivariate normal with mean p, variance, and pdf 

31) is the estimate of Sd(0). If il does not consist of independent elements, i.e., covariance terms exist between the random effects, the penalty term is modified to 

If by chance prior information were available on the residual error X, a gamma distribution pdf could be used and —2 times the log-likelihood of that pdf added to Eq. (8.30) or (8.32) such that the penalty term includes terms for 0, ft, and . 

One problem with the use of the inverse Wishart distribution is that the distribution does account for correlation among the random effects nor can it express any correlation among 0 and ft. To account for this Gisleskog, Karlsson, and Beal propose using a normal normal prior where now correlation can be accounted 

---