The predictive performance of a Bayesian regression-analysis computer program that uses non-steady-state phenytoin data was evaluated. Forty patients receiving phenytoin or phenytoin sodium who had two or more non-steady-state serum concentrations were selected for study. Additional serum concentrations and dosing data were collected as they became available, but no effort was made to control the number or timing of serum concentration determinations. Patients were categorized into four groups for evaluation of the effect of potential bioavailability problems and length of dosing history (time over which serum concentration-time data were collected) on the ability to predict subsequent phenytoin concentrations. Population parameters for phenytoin maximum rate of elimination (Vmax), apparent Michaelis-Menten constant (Km), volume of distribution (V), and bioavailability (F) were obtained from the literature. Predictions based on serum phenytoin concentrations and dosing histories (information intervals) of 5 or 10 days were compared with predictions based on naive (population-based) estimates using prediction-error analysis. In each patient group, the use of either 5-day or 10-day information intervals resulted in a significant increase in precision and a significant reduction in bias compared with naive estimates. For the group of patients who initially had two or more serum concentrations within the first five days of monitoring, predictions showed a marked increase in bias and a decrease in precision as the time interval from the last measured concentration to the time of prediction increased.(ABSTRACT TRUNCATED AT 250 WORDS)