Wählby Ulrika*, Bouw René, Jonsson E. Niclas and Karlsson Mats O.
Division of Pharmacokinetics and Drug Therapy, Uppsala University
Aim: The likelihood ratio (LR) test is commonly used as a tool for judging statistical significance of parameters included in PK-PD models when modelling with NONMEM. We have previously (Wählby et al., PAGE, 2000 and J. Pharmacokinet. Pharmacodyn., in press) seen that the test for covariates on fixed effects parameters is biased towards a higher than nominal risk of inclusion of false covariate relationships for the FO, but not for the FOCE with interaction between h-e (FOCE INTER) method. The aim of this simulation study was to assess the difference between actual and nominal significance levels for the statistical sub-model.
Methods: Pharmacokinetic data were repeatedly simulated from a one compartment iv bolus model (no covariate relationships, diagonal variance-covariance matrix or variance only in CL, exponential residual error). Two models were fitted to each replicate of the data, the model used for simulation (reduced model) and a model with an additional parameter (full model). The differences in objective function values (DOFV) between the fits were calculated, to yield an empirical reference distribution of DOFV under the null hypothesis. The additional parameter in the full model was either a covariate relationship influencing the interindividual error magnitude, a covariate influencing the residual error magnitude, a parameter estimating variability in V (when data were simulated without such variability), or a covariance parameter. Alterations were made to the simulation settings to assess influential factors (number of individuals, number of samples per individual, magnitude of interindividual/residual variability, distribution of residual variability, covariate distribution). The FO and FOCE INTER methods were tried.
Results: When the FO method was used, the actual significance levels were higher than nominal, except for inclusion of covariance between CL and V, where the opposite was found. For FOCE INTER, the actual and nominal significance levels are in close agreement, except that violations of the assumption of symmetric distribution of random effects resulted in increased actual significance levels. Whenever actual and nominal significance levels showed a discrepancy, the magnitude of the difference showed a complex dependence on several factors (number of subjects and samples per subject, error magnitude).
Conclusion: When the FO method is used, actual significance levels are in most cases far from the nominal levels. For the FOCE INTER method this is only true when distribution assumptions are violated. However, it is not always easy to determine the degree of asymmetry in random effects and caution in interpretation of the LR-test for the statistical model components also in this case is appropriate.
Reference: PAGE 10 () Abstr 193 [www.page-meeting.org/?abstract=193]
Poster: poster