Spreading Without Concern: An Empirically Grounded Agent-Based Model of Evacuation Behaviour Under Misinformation

Evacuation models usually assume occupants receive accurate information, yet false messages and others’ reactions shape early decisions in building emergencies. This study links survey data, interpretable machine learning and agent-based modelling within Protection Motivation Theory and the Risk Information Seeking and Processing model. Two surveys of university populations (n = 176 and n = 203) were analysed with non-parametric and multivariate statistics, classifiers with SHAP attribution and K-Means clustering, yielding four behavioural archetypes. The archetypes parameterised a NetLogo model in which two false messages, that the alarm is a drill and that an exit is blocked, spread between agents. Coping appraisal, but not threat appraisal, was associated with expected reactions to an alarm. Concern about misinformation was not positively associated with sharing unverified information, whereas reliance on others’ reactions was associated with two of three sharing measures. In simulation, the false messages added 3–8 s to mean evacuation time, while high official trust combined with a fire alarm shortened it by 15 s at equal exposure, with less persistence of false information. Occupants with low coping appraisal were delayed most. These directions held across 200 parameter sets, identifying pre-crisis trust and information resilience as priorities for empirical testing.