Techniques of structural health monitoring based on vibration measurements have been attracting bridge authorities as an alternative method for inspection processes. Changes in structural integrity of bridges engender changes in their modal properties. However, changes in those modal properties due to damage are usually too small to detect, which might be one of reasons for the lack of practical applications of the vibration-based bridge health monitoring. To realize a damage detection from vibration monitoring of bridges, this study proposes a robust damage detection method using Bayesian hypothesis testing. A vector autoregressive model is adopted to represent vibrations of bridges. The proposed method consists of three steps: a Bayesian inference method which provides a posterior distribution of parameters composing the vector autoregressive model; extracting damage-sensitive features related to the modal properties of the bridge based on the posterior distribution: and Bayesian hypothesis testing. To assess the feasibility of the proposed method, this study utilize data from field experiments conducted on actual steel bridges such as a plate girder bridge and two truss bridges with cracks that had been artificially severed. The proposed method detected all damages considered in the experiments even including potential damage location.