The Application of Resampling Methods to Population Pharmacokinetic Modeling

Paul J. Williams

University of the Pacific; Stockton; CA; USA

Pharmacokineticists are usually attempting to improve the ways in which they are developing models and have therefore been interested in the application of novel biometrical methods such as resampling to population pharmacokinetic (PPK) modeling. The bootstrap (B), cross-validation (C-V)), and jackknife (JKK) are the common types of resampling techniques. The utility of these methods has recently been acknowledge in the “Guidance for Industry: Population Pharmacokinetics” from the US Food and Drug Administration. For the JKK and C-V, data are removed one at a time or in blocks and the prediction rule re-estimated. The JKK is used to correct bias and estimate parameter distributions and C-V is used to assess predictive performance. For the bootstrap, replicate data sets are generated by drawing an individuals data with replacement, the function of interest is executed, and the results summarized. The bootstrap has been used, most commonly, to estimate the distribution of PPK parameters, bias correct parameters, and assess predictive performance. The ideas for these techniques have not, in general, been applied to PPK model development or validation because they are computationally intense. With the advent of improved computational speed this required computational intensity has rapidly become unimportant. These methods can be applied in a very straight forward manner to any statistical application no matter how complex and have worked well when the parent population is representative of the population to which the model is expected to be applied. They have been applied in the area of PPK model development, evaluation, and validation. Resampling has been applied in the following manner:

  1. Evaluation of bias and precision of PPK estimates.
  2. Directly assess parameter distributions rather than relying on synthetic methods. Synthetic methods are those methods that estimate distributions from formulae; direct methods derive distributions directly from observations of data.
  3. Estimation of inestimable standard errors for small sample size population
  4. Developing population models with greater covariate stability compared to traditionally applied methods.
  5. Validate PPK models through the assessment of predictive performance without collecting data in a test population.
  6. Comparison of non-hierarchical PPK models.

In this presentation the JKK and C-V will be dealt with briefly and the bootstrap method will be applied in more detail. Informative examples from the presenters work will be used to demonstrate the application and utility of the bootstrap.

Reference: PAGE 9 (2000) Abstr 91 [www.page-meeting.org/?abstract=91]

Poster: oral presentation