U. Wählby*, E. N. Jonsson and M. O. Karlsson
Division of Pharmacokinetics and Biopharmaceutics, Department of Pharmacy, Uppsala University, BMC Box 580, S-751 23 Uppsala, Sweden.
In many population pharmacokinetic and pharmacodynamic studies the aim is to quantify interindividual variability. This variability is partly explained by relationships with covariates, and the remaining, unexplained, part is characterized by the variance and covariance of the parameters. We present an alternative parameterization that seeks to avoid many of the problems with the presently used strategies for determining a suitable size and structure of the interindividual model. It may also provide an opportunity for a possibly more relevant covariate model building strategy. This approach allows the model to be increased by one parameter at a time, so that the balance between the requirement for increasing the model size to best describe the data, and the possibility to obtain parameter estimates with sufficient precision, can be better met than by any of the presently employed methods, which introduce or eliminate several covariance terms at a time. In the proposed method, to account for parameter correlation, rather than adding covariance terms, we add a random variable jointly on two or more parameters, in a stepwise fashion. At each step all relevant combinations of parameters are tried for addition of a variable. The term giving the largest improvement in describing the data (as judged by -2*LogLikelihood) is added to the model, providing the improvement is significant. This will form the new basic model, and a new step is performed. The stepwise procedure is computer run-time demanding, but objective and based on pre-defined decision rules and thus easy to automate. Simulated data, based on a model derived from real world data, was used to exemplify the strategy. The model building comprised testing 31 different models. The final model contained one joint random variable, which was shared by all disposition parameters. To reduce run times model building was also performed with typical parameter values and residual error magnitudes fixed, this resulted in the same final model as when all parameters were estimated. A model with a full covariance matrix did not explain the data better than the final model, although containing 9 additional covariance terms. This indicates that the proposed method is adequate for identifying the relevant components of the interindividual variance-covariance matrix.
Reference: PAGE 7 (1998) Abstr 292 [www.page-meeting.org/?abstract=292]
Poster: poster