The aim of the present work was to analyse the predictive capacity of different optimization methods for carbamazepine dosage regimens according to population pharmacokinetic parameters and/or serum levels data. Calculations were performed using a multiple non-linear regression program (MULTI 2 BAYES) with two different options: (a) using the ordinary least squares methods and (b) the least squares method based on the Bayesian algorithm. The predictive capacity of the three methods was analysed comparing the serum levels values in patients predicted by each method and those observed later. The performance obtained was poor and even unacceptable when dosage prediction was based on the mean parameters of the population studied (Method A) and when only one data point relating to the serum carbamazepine levels was available (Method B). When two serum levels data were used the standard deviation of the mean prediction error was only clinically acceptable when Bayesian non-linear regression was applied (Method C). When the number of serum levels data points was 3 or 4 the errors were acceptable from a clinical point of view and no significant differences could be found between the use of the simple and Bayesian version at the same level of information.