Ninety-three sled tests (60 cadaver, 33 dummy) is used to assess the Hybrid III dummy and associated thoracic injury criteria to determine their ability to predict cadaver injury in matched impacts over a range of impact and restraint conditions. A statistical analysis is performed to evaluate the injury-predictive efficacy of the dummy-based maximum chest deflection, maximum chest acceleration, and CTI. A comparison of the selectivity and specificity of injury prediction models with each injury criterion reveals that the maximum dummy chest deflection is the best predictor of injury (C statistic = 0.8514) when used in an injury risk model including cadaver age, gender, and mass. Consideration of the maximum chest acceleration, either as an independent covariate or as part of CTI, is found to weaken the injury model. The restraint condition (belt-dominated loading, airbagdominated loading, or combined loading), however, is found to be a more important injury predictor than any injury criterion.