Estimating Effectiveness of Countermeasures Based on Multiple Sources: Application to Highway-Railway Grade Crossings

Abstract

To provide an adequate level of safety at grade crossings, Transport Canada has allocated several millions annually to prevent collisions at grade crossings through the implementation of countermeasures, such as train-actuated warning devices and track devices. Railway companies and provincial agencies have also provided additional support to improve safety at highway-railway grade crossings. One of technical challenges in estimating safety effect of countermeasures at highway-railway grade crossing is an extremely rare occurrence of collisions. Given that the collision process is random with significant variation over time and space, it is hard to judge whether a specific crossing is safe or safer than other crossings solely based on the number of collisions in a given year. Decision makers are also required to make difficult decisions on safety investment accounting for uncertainty in effectiveness of countermeasures. The level of uncertainty is even higher when there is lack of observed collision data before and after the implementation of specific countermeasures. This study proposes a Bayesian data fusion method which overcomes these limitations. In this method, we used previous research findings on the effect of a given countermeasure, which could vary by jurisdictions and operating conditions, to obtain a priori inference on its expected effects. We then used locally calibrated models, which are valid for a specific jurisdiction, to provide better estimates of the countermeasure effects. Within a Bayesian framework, these two sources were integrated to obtain the posterior distribution of the countermeasure effect. The outputs provided not only the expected collision response to a specific countermeasure, but also its variance and corresponding probability distribution for a range of likely values. Some numerical examples using Canadian highway-railway grade crossing data illustrate how the proposed method can be used to predict the effects of prior knowledge and data likelihood on the estimates of countermeasure effects.