Inferring Passenger Responses to Urban Rail Disruptions Using Smart Card Data: A Probabilistic FrameworkBaichuan Moa,∗, Haris N. Koutsopoulosb, Jinhua ZhaocaDepartment of Civil and Environmental Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139bDepartment of Civil and Environmental Engineering, Northeastern University, Boston, MA 02115cDepartment of Urban Studies and Planning, Massachusetts Institute of Technology, Cambridge, MA 20139AbstractThis study proposes a probabilistic framework to infer passengers’ responses to unplanned urban rail service disruptions using smart card data in tap-in-only public transit systems. We first identify 19 possible response behaviors that passengers may have based on their decision-making times and locations (i.e, the stage of their trips when an incident happened), including transferring to a bus line, canceling trips, waiting, delaying departure time, etc. A probabilistic model is proposed to estimate the mean and variance of the number of passengers in each of the 19 behavior groups using passengers’ smart card transactions. The 19 behavioral responses can be categorized from two aspects. From the behavioral aspect, they can be grouped into 5aggregated response behaviors including using bus, using rail (changing or not changing route), not using public transit, and not being affected. The inference of the 19 behaviors can be classified into four cases based on the information used (historical trips vs. subsequent trips) and the context of the observed transactions(direct incident-related vs. indirect incident-related). The public transit system (bus and urban rail) of the Chicago Transit Authority (CTA) is used as a case study based on a real-world rail disruption incident. The model is applied with both synthetic data and real-world data. Results with synthetic data show that the proposed approach can estimate passengers’ behavior well. The mean absolute percentage error (MAPE)for the estimated expected number of passengers in each behavior group is 20.5%, which outperforms the rule-based benchmark method (60.3%). The estimation results with real-world data are consistent with the incident’s context. An indirect model validation method using demand change information and incident log data demonstrates the reasonableness of the results.