TY - JOUR
AU - Zakeri, Sajjad
AB - Modeling drivers' stop/go decisions when facing yellow traffic lights is crucial for dilemma protection, driver-assistive, and warning systems, as incorrect decisions may lead to severe crashes. This study compares two dominant modeling approaches: logistic regression and utility theory models. It remains unclear which group of models possesses superior predictive power or whether utility models' robust theoretical basis yields more precise results. Additionally, previous studies fail to simultaneously consider all three stop/go decisions: stopping, yellow-light running (YLR), and red-light running (RLR). The present research aims to compare these model groups' predictive power while accounting for drivers' risky RLR decisions, and examines whether incorporating spatial heterogeneity and random taste variations in utility theory models enhances forecasting capability. Using data from three intersections in Semnan, Iran, various models were calibrated including multinomial, nested, and mixed logit (utility theory) alongside multinomial logistic, ordinal, and probit regression. Results show multinomial logistic regression achieved the highest accuracy, while utility models (particularly mixed logit) better captured behavioral heterogeneity despite lower accuracy. This reveals a key trade-off: Logistic regression excels for precise predictions (e.g., violation detection), whereas utility models offer deeper behavioral insights. These findings assist practitioners in selecting approaches based on priorities—accuracy for safety systems or interpretability for behavioral analysis.
TI - Assessing utility theory and logistic regression models to predict drivers’ stop/go behavior considering random taste variations
JF - International Journal of Data Science and Analytics
DO - 10.1007/s41060-025-00803-1
DA - 2025-06-04
UR - https://www.deepdyve.com/lp/springer-journals/assessing-utility-theory-and-logistic-regression-models-to-predict-21Tt1XEo1P
SP - 1
EP - 17
VL - OnlineFirst
IS - 
DP - DeepDyve
ER -