Date of Award

December 2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Civil Engineering

First Advisor

Daba Gedafa

Abstract

Intersections are the most crash-prone areas, accounting for a significant portion of all traffic crashes, making their safety a critical concern. Existing studies and guidelines, such as the Highway Safety Manual (HSM), provide safety performance functions (SPFs) and intersection influence area (IIA) based on data from a limited number of states, insufficient traffic crash variables, and traditional statistical methods. Though previous studies have provided valuable insights into intersection safety, they often lack a comprehensive approach that parallelly evaluates parametric and nonparametric methods to examine crash severity, seasonality, and collision types. This need is particularly evident in regions like North Dakota (ND), where distinct spatial and temporal characteristics reveal the shortcomings of generalized national guidelines. This study aims to develop local IIA from crash data distribution and SPF models to address the variability in crash characteristics at four-leg signalized intersections (FLSGI) in ND. It employs a logistic regression technique to identify IIAs based on accuracy, precision, and Absolute False Frequency Difference (AFFD) metrics. The results showed that locally calibrated IIAs achieved the lowest AFFD, strongly aligning with regional crash data. The findings indicate that IIAs were longer under adverse road and low-speed limit conditions and ranged from 82 ft to 335 ft. Additionally, SPFs were developed for total crashes and various crash categories, including property damage only (PDO) crashes, killed, injured, or incapacitated (KABC) crashes, seasonal crashes (dry and wet season crashes), and crashes by different collision types (rear end, angle, and sideswipe crashes). The SPFs were developed using fixed and random effect parametric models, such as Poisson Regression (PR), Negative Binomial (NB), Zero-Inflated Poisson (ZIP), Zero-Inflated Negative Binomial (ZINB), and nonparametric techniques like Random Forest (RF) and Extreme Gradient Boosting (XGBoost). The effect of key factors such as truck volume and intersection geometry were incorporated into the models. The performance of these models was assessed using AIC, BIC, log-likelihood (LL), and Mean Squared Error (MSE). It was found that the SPFs derived from the Random Effects Negative Binomial (RENB) and Random Effects Zero-Inflated Negative Binomial (REZINB) approaches are the best fit, simple, and practical models of all the parametric models. A transferability analysis of the best models for the East and West regions of the state revealed that local models have a lower Transferability Index (TI), and these models developed for each region yield a simplified and better-fit model. Moreover, the results demonstrate that ML approaches produced the lowest MSE and have superior predictive accuracy. Though both RF and XGBoost models capture non-linear relationships between variables, the k-fold cross-validation result revealed that XGBoost exhibited superior predictive accuracy for most SPFs, enhanced through hyperparameter tuning. Furthermore, feature importance and sensitivity analysis reveal that Average Annual Daily Traffic (AADT, log-transformed), lane widths, and turn lanes are variables that significantly affect the outcome of each SPF model. Finally, a comparative analysis of SPFs from locally developed IIAs and HSM thresholds demonstrated that the local IIA SPFs showed a better model fit and accuracy. Overall, the RF and XGBoost techniques are superior due to their better performance and accuracy, while the RENB and REZINB models offer greater interpretability and are less complex, making them more practical for practitioners. In conclusion, this study emphasizes the benefits of locally calibrated models for improving crash prediction and informing policy development. It provides a robust framework for researchers, evaluating both parametric and nonparametric methods to analyze crash data with respect to location, severity, seasonality, and collision types, thereby encouraging further exploration of intersection safety dynamics at a local level. By adopting these context-specific and data-driven strategies, transportation planners can better inform policy decisions and implement targeted measures to reduce crashes and improve safety.

Available for download on Sunday, January 17, 2027

Share

COinS