OBJECTIVES: The importance of postmarketing surveillance was again publicized by the recent recall of Vioxx; similar concerns should be addressed in trial-based cost-effectiveness analyses (CEA). The net benefit regression approach applies econometric methods; thus, offers a power tool to assess CEA of a new intervention in a non-randomized environment (e.g., claims data). Our study proposed a Bayesian approach to synthesize clinical trial data with secondary data collected in postmarketing setting. METHODS: We first compared the treatment effect estimated from least squares (LS) and Bayesian regressions using a simulated data of 200 pairs of case-control patients. The data contained information on cost (C), effectiveness (E), and demographics for each patient, with 85% of data in the quadrant of positive incremental cost and incremental effectiveness. The dependent variable was the net benefit (NB), calculated as: ? E – C, where &lambda denoted the maximum willingness to pay; covariates included demographics and a binary variable indicating treatment. By incorporating trial data in the prior distribution, we demonstrated use of the Bayesian regression to update the net benefit estimates with observed data, exemplified by the simulated data. RESULTS: NB estimated from LS and Bayesian approaches were very similar when non-informative prior was used, representing the scenario where the trial data was neglected. However, using the posterior distribution of the regression coefficients, the Bayesian approach can infer the probability that the new treatment was cost-effective. At ?=$15,000, the estimated NB was $679.3 (P=0.766) in OLS and $677.5 in Bayesian, with 0.61 probability of cost-effective. When a strong prior favoring the new treatment was employed, the estimated NB increased to $3,414 at ?=$15,000 and the probability of cost-effective became 0.96. CONCLUSION: The ability to incorporate prior information makes Bayesian regression a useful method in postmarketing surveillance. Our analytical framework can be further expanded to address selection issues.