Safety is a core issue in the railway operation. In particular, as witnessed by accident/incident statistics, railway level crossing (LX) safety is one of the most critical points in railways. In the present paper, a Bayesian network (BN) based framework for causal reasoning related to risk analysis is proposed. It consists of a set of integrated stages, namely risk scenario definition, real field data collection and processing, BN model establishment and model performance validation. In particular, causal structural constraints are introduced to the framework forthe purpose of combining empirical knowledge with automatic learning approaches, thus to identify effective causalities and avoid inappropriate structural connections. Then, the proposed framework is applied to risk analysis of LX accidents in France. In details, the BN risk model is established on the basis of real field data and the model performance is validated. Moreover, forward and reverse inferences based on the BN risk model are performed to predict LX accident occurrence and quantify the contribution degree of various impacting factors respectively, so as to identify the riskiest factors. Besides, influence strength and sensitivity analyses are further carried out to scrutinize the influence strength of various causal factors on the LX accident occurrence likelihood and determine which factors the LX accident occurrence is most sensitive to. The main outputs of our study attest that the proposed framework is sound and effective in terms of risk reasoning analysis and offers significant insights on exploring practical recommendations to prevent LX accidents.