Investigating Traffic Crash-Type Determinants in Different Urban Environments: A City-Specific Random Parameter Multinomial Logit Approach
Abstract
This study tests whether the random-parameter structure of a Multinomial Logit crash-type model is itself spatially stable, using individual-level likelihood-ratio tests estimated separately for Ankara (28,825 crashes) and İzmir (24,280 crashes) between 2023 and 2024. In this study, the traffic accidents examined were grouped as Lateral Collision, Rear-End Collision, Pedestrian Collision, and Loss of Control Collisions. A comparative modeling study of two cities with different cultural and geographical characteristics provides two-city evidence that the structure can change significantly across these contexts. Accordingly, the set of variables requiring random processing differs between the two cities; in Ankara, infrastructure variables (Shoulder, Traffic Signal, Guardrail) determine heterogeneity, while in Izmir, geometric and behavioral variables (Driver Age—Young, Horizontal Curve, Sidewalk) dominate heterogeneity. For random parameters with a coefficient of variation above 1.0, the percentage of drivers experiencing the sign reversal effect ranges from approximately 18% to 25%. A cross-prediction transferability check, in which each city’s estimated Random Parameter Multinomial Logit (RPMNL) model coefficients are applied to the other city’s data, shows that the transferred model underperforms the target city’s own benchmark Multinomial Logit specification. The findings indicate that the statistical association between signalized intersections and crosswalks and the rear-end outcome is of opposite sign for a measurable subset of the driver population, which points to differentiated rather than uniform management needs at such locations, and that the pooling-by-default practice in multi-city RPMNL crash-type research requires empirical defense rather than assumption. Whether such structural shifts are prevalent across urban contexts is an open question that this two-city result motivates testing in multi-city designs.