PURPOSE: To evaluate the impact of an alternative biochemical failure (bF) definition on the performance of existing plus de novo prognostic models. METHODS AND MATERIALS: The outcomes data of 1,458 Australian and 703 Canadian men treated with external-beam radiation monotherapy between 1993 and 1997 were analyzed using a lowest prostate-specific antigen (PSA) level to date plus 2 ng/mL (L + 2) bF definition. Two existing prognostic models were scrutinized using discrimination (Somers Dxy [SDxy]) and calibration indices. Alternative prognostic models were also created using recursive partitioning analysis (RPA) and multivariate nomogram methods for comparison. RESULTS: Discrimination of bF was improved using the L + 2 definition compared with the American Society for Therapeutic Radiology and Oncology (ASTRO) definition using both the three-level risk model (SDxy 0.30 and 0.22, respectively) or the nomogram (SDxy 0.35 and 0.27, respectively). Both existing prognostic models showed only modest calibration accuracy. Using RPA, five distinct risk groups were identified based primarily on Gleason score (GS) and all subsequent divisions based on PSA. All GS 7-10 tumors were intermediate or high risk. This model and the developed nomogram showed improved discrimination over the existing models as well as accurate calibration against the Canadian data, apart from the 30-50% failure region. CONCLUSIONS: The L + 2 definition of bF provides improved capacity for discrimination of failure risk. New prognostic models based on this endpoint have overall statistical performance superior to those based on the ASTRO consensus definition but continue to have unreliable discrimination in the intermediate-risk region.