We use influence diagrams to describe event trees used in safety analyses of low-probability high-risk incidents. This paper shows how the branch parameters used in the event tree models can be updated by a bayesian method based on the observed counts of certain well-defined subsets of accident sequences. We focus on the analysis of the shared branch parameters, which may frequently often in the real accident initiation and propagation to more severe accident. We also suggest the way to utilize different levels of accident data to forecast low-probability high-risk accidents.