Segment length optimization for crash frequency modelling: Evaluating power spectral segment length in safety performance assessment

Selecting an appropriate segment length is essential for road safety analysis, as it directly influences crash analysis accuracy, hazardous location identification, and safety performance evaluation. The traditional segmentation approaches rely on individual expertise or engineering judgment and often lack standardized metrics for evaluating segmentation performance. Therefore, this study expands the utilization of Spatial Frequency Domain Analysis (SFDA) based Power Spectral Segment Length (PSSL) in crash frequency modeling for predicting fatal crash occurrence. The power spectral analysis reveals that crash frequencies predominantly concentrate in low-frequency bands, which helps in determining the Power Spectral Percentage (PSP), a critical measure for evaluating segmentation performance. The Random Parameters Negative Binomial (RPNB) models are developed for six rural two-lane highways in order to evaluate the effectiveness of PSSL, accounting for unobserved heterogeneity in crash data. The study results indicate that PSSL-based segmentation consistently outperforms traditional segmentation methods, as demonstrated by Cumulative Residual (CURE) plots and Goodness-of-Fit statistics. Additionally, the study results show that roadside service areas, population density, minor access points, and heterogeneous traffic characteristics are the most significant predictors of fatal crashes across all highway types. Hence, this study provides an optimized, data-driven, and theoretically justified framework for segment length selection, which improves accuracy, reliability, and scalability in crash modeling and road assessment.