The present work evaluated the performance of two computer programs: Drugcalc, which utilizes the bayesian (method 1) approach and PKS, which can utilize both the non-bayesian (method 2) and bayesian (method 3) approaches. Both programs permit the introduction of serum level data obtained in both situations: steady-state and nonsteady-state. The prediction of phenytoin concentrations (n = 771) were made from steady-state (n = 378) and nonsteady-state (n = 175), and combined steady-state and nonsteady-state (n = 218) concentrations. The observed serum concentrations (at least two nonsteady-state and two steady-state per patient) were collected under routine clinical conditions in 15 patients receiving this drug. The main contribution to prediction errors is attributed to the difference between doses corresponding to the predicted and feedback serum concentrations, dD, in such a way that when the errors obtained for dD ≥ 100 mg/day are excluded, the predictive performance increases significantly for all methods. In this sense, increases in precision were 87, 64, and 66% for methods 1, 2, and 3, respectively. Moreover, when dD <100 mg/day, nonsteady-state feedback concentrations (≤3) only afforded clinically acceptable predictions (ME ± SD <3 mg/L) when they were combined with at least one steady-state datum value, and the bayesian approach was used. Despite this, for all the methods analyzed, nonsteady-state data are seen to be useful for detecting situations of potential toxicity in a significant proportion of cases (71.4–84.6%) and, when method 3 is used, may offer useful information for the adjustment of dosage schedules.