Ulrika Wählby, E. Niclas Jonsson and Mats O. Karlsson
Division of Biopharmaceutics and Pharmacokinetics, Department of Pharmacy, Uppsala University
Aim: The likelihood ratio test is commonly used as a tool for judging statistical significance of covariate relationships when building PK-PD models using NONMEM. There are indications that the test is biased towards a higher risk of inclusion of false relationships than given by the nominal p-level. We performed a simulation study with the aim to assess the difference between actual and nominal significance levels, and to study what factors might influence these levels. Solutions to handling differences between nominal and true p-levels are also proposed.
Methods: Pharmacokinetic data without covariate relationships were repeatedly simulated from a one compartment iv bolus model. Models with and without covariate relationships were fitted to the data, and differences in objective function values calculated. Alterations were made to the simulation settings in order to assess influential factors (number of individuals, number of samples per individual, inter-/intra- individual variability, structural and residual error models, covariate type and distribution, parameter influenced). Different estimation methods were tried.
Results: The actual inclusion frequency of a false covariate relationship in the model was higher than the nominal p-value under most conditions using the FO method. The factors that markedly influenced the risk were a heteroscedastic error structure, a non-linear model, frequency of sampling, residual error magnitude and what parameter the covariate was influencing. The use of the FOCE method resulted in better agreement and FOCE with interaction between - in close agreement between actual and nominal significance levels. Factors with minor influence were the type of covariate and the number of individuals in the data set.
Solution: Build the covariate model with FOCE with interaction if time permits. Otherwise calibrate, i.e. determine the cut-off OFV corresponding to the desired significance level for the specific data set, model and parameter, by (i) simulation of covariate data, (ii) permuting the covariate data while maintaining the pharmacokinetic data intact, or, as a last resort, (iii) simulation of pharmacokinetic data while maintaining the covariate data intact.
Reference: PAGE 9 (2000) Abstr 113 [www.page-meeting.org/?abstract=113]
Poster: poster