Unraveling vulnerable road user crash severity: a latent class and random parameter approach with COVID-19 temporal insights

OBJECTIVE: This study investigates the determinants of vulnerable road user (VRU) crash severity by accounting for unobserved heterogeneity both across and within crash subpopulations. Conventional single-model approaches assume homogeneous effects across all crashes, potentially masking context-dependent severity mechanisms. METHODS: Crash data for 15,578 pedestrian and 11,433 bicyclist collisions occurring at or near intersections in 20 California cities (2016-2025) were extracted from the Statewide Integrated Traffic Records System (SWITRS). A two-stage analytical framework was employed. First, latent class analysis identified three distinct crash typologies for each VRU mode based on movement patterns, lighting, weather, and collision factors. Second, mixed logit (MXL) models were estimated for each latent class to capture within-cluster heterogeneity through random parameters. Pseudo-elasticity analysis quantified the practical magnitude of variable effects. Temporal stability was assessed by estimating separate models across pre-COVID, during-COVID, and post-COVID periods. RESULTS: Truck involvement, dark conditions without streetlights, and state highway location consistently elevated severe outcome odds across all clusters for both VRU types, while VRU age 65+ shifted injury distributions toward moderate rather than the most severe outcomes. Critically, several factors exhibited context-dependent effects. VRU fault increased severity when drivers traveled straight, but decreased severity in turning-driver crashes for bicyclists, indicating fundamentally different causal mechanisms. State highway effects ranged from the strongest fatal predictor in straight-driver pedestrian crashes to non-significant in other configurations. Different random parameters were identified across clusters, confirming that unobserved heterogeneity operates through distinct mechanisms in different crash contexts. CONCLUSIONS: Crash severity determinants are both universally important and context-dependent, with the same variable capable of opposing effects across crash configurations. These findings demonstrate that aggregate models pooling heterogeneous crash types obscure critical variation and support the adoption of context-sensitive approaches to crash modeling.