Balanced designs in longitudinal population pharmacokinetic studies.

A simulation study was performed using a balanced design to determine the sample size required for accurate and precise estimation of a parameter at a given level of intersubject variability in a longitudinal population pharmacokinetic study. A two-compartment model parameterized in terms of clearance (Cl), volumes of the central (V1) and peripheral (V2) compartments, and intercompartmental clearance (Q) with multiple intravenous bolus inputs was assumed. Six samples were obtained from each subject using the informative profile (block) randomized design. Variability (in terms of coefficient of variation, CV) in model parameters was varied between 30% and 100%, and residual variability was fixed at 15%. Sample sizes ranging from 30 to 1,000 subjects were studied, and a hundred replicate data sets were generated and analyzed with NONMEM for each sample size at each CV. A sample size of 30 was required for accurate and precise estimation of structural model parameters when CV < or = 75%, except for Cl where it is adequate for CV < or = 100%. A sample size of 80 was required for intersubject variability estimation with CV < or = 60%. Robust estimates of variability in Cl were obtained with sample sizes of 30 (CV < or = 45%), 60 (CV 60-75%), and 100 (CV > or = 75%). Positively biased estimates of residual variability were obtained irrespective of sample size at > or = 60% CV. This indicates that estimates of residual variability obtained in study situations where CV > or = 60% should be interpreted with caution. In such situations model misspecification may not be the issue, because in this simulation study concentration-time profiles were generated and analyzed with the same model. Although these results should be interpreted within the context of the study, they provide a framework for addressing the issue of sample size in longitudinal population pharmacokinetic study with a balanced sampling design. The result of a population pharmacokinetic study can be anticipated by comparing the results of several simulations in which the various input factors have been varied.