Effect of the Number of Samples on Bayesian and Non-linear Least-squares Individualization: A Study of Cyclosporin Treatment of Haematological Patients with Multidrug Resistance

Abstract

Abstract We have studied whether the prediction of drug concentrations improves as the number of samples used for individualization is increased, and whether the Bayesian method of individualization is superior to the non-linear least-squares method. Data were obtained from ten adult haematological patients with multidrug resistance who were treated with cyclosporin. The predictions of blood–cyclosporin concentrations were made using the Abbott PKS program. The number of samples used for individualization was increased from 1 to 30 for the Bayesian method and from 4 to 30 for the non-linear least-squares method. Linear regression, percentage prediction error, and absolute and relative predictive performance were used to evaluate the predictions. The results show that the Bayesian method affords greater precision than the non-linear least-squares method, but that the non-linear least-squares method is more accurate and results in less bias. Whereas for linear regression predictions improve as the number of samples is increased, other evaluations show improvement in the range from 5 to 11 samples; linear regression, percentage prediction errors and prediction bias support the opinion that the Bayesian method progressively becomes the non-linear least-squares method as the number of samples used for individualization is increased, but the accuracy and precision of prediction do not support this opinion. The study supports the statement that Bayes’ law requires parameters from an infinite population, otherwise the advantage of the Bayesian method might be marginal.