Identifying Latent Classes of Pedestrian Crowd Evacuees

The tactical decision making of humans when they exit crowded built environments was investigated. Recent studies have shown that the relative weights and priorities that pedestrian evacuees associate with different factors contributing to their choice of exit vary significantly from person to person, yet this heterogeneity is not explainable by the demographics of the decision makers, such as gender. Hence, it is assumed that the choice could be attributable to the presence of latent classes (latent types of behavior) unassociated with observable demographic characteristics. The standard latent class and the random coefficient latent class choice models were implemented with data on stated exit choices to explore this. The latter model accommodated coefficient heterogeneity within each class by allowing the coefficients associated with each class to vary randomly. Data analyses showed that exit choice heterogeneity did not necessarily need to be represented by continuous distributions and could have been accommodated in a finite number of discrete categories. In fact, only two classes of evacuees (congestion avoiders and distance avoiders) could be identified. The terminology derived from the fact that the chief difference in evacuation behavior of the two classes was associated with their perception of crowding-related factors (i.e., exit density and directional density). The modeling results suggested that congestion avoiders outnumbered distance avoiders. The analyses also showed that the within-class heterogeneity faded into insignificance once the binary discrete classification was imposed: evacuees’ decision-making heterogeneity was sufficiently reducible into two discrete categories with negligible inner-class variation. This finding provides a more tangible interpretation of exit choice heterogeneity than the continuous mixture approach. The finding also simplifies the way in which the modeler deals with decision heterogeneity when the exit choice model is implemented for prediction purposes.