Latent Class Logit Kernel framework for surrogate safety: identifying behavioral thresholds through conflict...

Crash data provide objective safety metrics but are rare and often unsuitable for proactive safety management. In contrast, traffic conflict indicators (e.g., time-to-collision, TTC) offer continuous measures of proximity to collision but require thresholds to separate routine from safety-critical events. Extreme Value Theory (EVT)-based approaches define statistically defensible thresholds from the tail behavior of conflict indicators, but these thresholds are not explicitly tied to observable maneuver adaptations. This study instead models drivers' discrete maneuver adjustments under varying conflict indicator levels and extracts candidate behavioral thresholds (CBTs) from the resulting maneuver-response probability profiles.

A Latent Class Logit Kernel (LC-LK) framework is proposed to identify CBTs by modeling drivers' maneuver choice under conflict. The LC-LK model distinguishes between low- and high-risk behavioral classes and allows each driver to express a probabilistic mixture of both states, capturing intra-driver heterogeneity. This heterogeneity refers to variation in the behavior of the same driver under different levels of conflict severity (as indexed by the indicator). The framework also incorporates correlation in alternatives through logit-kernel structured error components that represent shared unobserved influences among maneuvers involving similar kinematic adjustments (deceleration, acceleration, or turning). This design produces probability curves that describe how the likelihood of high-risk maneuvers changes with conflict indicator values. From these profiles, CBTs such as inflection points, crossovers, and tail-based thresholds can be derived. The framework is guided by four behavioral hypotheses: (i) drivers simultaneously exhibit varying degrees of membership in both low- and high-risk behavioral classes; (ii) class membership shifts systematically with conflict indicator values; (iii) this relationship often follows a logistic shape, with stable behavior across safe conditions and rapid transitions once critical values are reached; and (iv) even in free-flow conditions, drivers maintain a baseline level of caution.

Application to naturalistic roundabout trajectories demonstrated the framework's diagnostic power. For TTC, stable behavioral inflections were observed between 0.8-1.1s, indicating clear transitions from low- to high-risk driving. In contrast, a modified variant of TTC (MTTC2) produced unstable and implausible thresholds (≈34s). This divergence suggested that not all indicators support identifiable behavioral transitions. One tentative interpretation, which we report cautiously, is that indicators that require more complex information (e.g., MTTC2) may be harder for drivers to process during routine driving. Simpler indicators such as TTC appear to yield more stable behavioral patterns, but further evidence is required to substantiate this explanation.

The proposed framework is intended to complement and not replace EVT. The LC-LK model produces multiple CBTs, each capturing different aspects of behavioral transitions. Some CBTs showed consistency with EVT-derived thresholds and others diverged substantially. A systematic investigation is needed to determine which CBT should be used or how behavioral thresholds should be integrated with statistical and practice-based thresholds.