Methodological Studies on Covariate Model Building in Population Pharmacokinetic-Pharmacodynamic Analysis

Sammanfattning: Population pharmacokinetic (PK) – pharmacodynamic (PD) modelling, using nonlinear mixed effects models, is increasingly being applied to obtain PK-PD information in drug development. Covariate modelling, the establishment of relationships between model parameters and patient characteristics, is undertaken to explain PK-PD variability and facilitate dose adjustment decisions, and is consequently an important objective of population PK-PD.The aims of this thesis were to increase the efficiency, predictability and robustness of covariate model building by examining in detail a number of aspects related to covariate modelling. The thesis demonstrates that the likelihood ratio (LR) test can be applied with confidence, in the assessment of statistical significance of parameter-covariate relationships (in NONMEM analyses), only if an estimation method appropriate for the data- and error-structure is utilised. Conversely, caution is needed in the interpretation of the LR test when variance or covariance parameters are modelled, since the type I error rate may be severely upward biased if the assumptions of normally distributed residuals and/or enough information in the data are violated. The two stepwise covariate model building procedures, using generalised additive models and NONMEM, were found to perform similarly in the examples examined. However, differences in performance may prevail in other situations, e.g. when sparse sampling precludes reliable individual parameter estimates. Stepwise selection was shown to result in over-estimated covariate effects (selection bias), but the imprecision in the estimates exceeded this bias. Important information about the PK-PD characteristics of a drug is obtainable on the application of covariate models for time-varying covariates that account for differences in variability between and within individuals, or estimate interindividual variability in the covariate effect. The knowledge gained in this thesis will contribute to the development of more predictable and robust covariate models, important both in individualisation of dosage and the development of new drugs.