Methodological studies on non-linear mixed effects model building

Sammanfattning: Population analysis of pharmacokinetic/pharmacodynamic (PK/PD) data by the use of non-linear mixed effects modeling offers a number of benefits over traditional methods of analyzing this type of data. These benefits include the possibility to appropriately characterize the typical PK/PD behaviour in the population as well as the associated variability when only a few observations are available for each individual. The price one has to pay, however, is a more complicated model structure and a corresponding increase in the complexity of the model building process.The work in this thesis demonstrates that there are a number of ways in which the probability of a successful population PK/PD model development can be increased. Using the best available dosage history, which was found to be identifiable in an objective way when two parallel dosing histories were collected, improves both the population parameter estimates and the individual parameter estimates. The individual estimates especially can be a crucial issue if methods that rely on their quality are to be used during the model building process. The sampling design also influences the quality of the individual estimates as well as the ability to detect the components of the inter-individual variabilitymodel. In the example studied it was possible to improve the original design with only small alterations to the protocol. The interpretability of the results and especially the extrapolation of the results of population studies are dependent on whether the assumptions made during the analysis can be justified or not. Using only the data set under study, it was, nonetheless, possible to check and sometimes justify the assumptions that was made. Identifying the relevant covariates is an important aspect of population model building, and ways to enhance a commonly used method (stepwize generalized additive modeling -GAM) were presented and evaluated. Especially the sensitivity to outliers could be reduced. Despite these improvements, the GAM has drawbacks and a method was therefore devised that does not suffer from these. A new method and strategy for inter-individual variability model building was also presented. Finally, more or less complicated methods like the ones presented herein will be of little practical use unless they are made accessible to the general data analyst community. Therefore a program that can serve as the vehicle for both existing and new methods was developed.