Comparing the Regression and Utility Theory Models in Predicting Yellow/Red Light Running Decisions Considering Spatial and Drivers’ Heterogeneity

Modeling the decision of drivers to stop or pass when facing a yellow light plays an important role in dilemma protection, driver-assistive, and driver warning systems. The wrong decision leads to accidents and safety hazards. The models for predicting stop/go decisions are divided into two broad categories: regression and utility theory models. But it is not clear which group of these models has better predictive power and whether the stronger logic and concept of utility models leads to better forecasts. In addition, in previous studies, it has rarely happened that all three decisions of stopping, passing a yellow light, and passing a red light are predicted together. The purpose of this article is to compare the goodness of fit and the correct prediction of these two groups of models considering the unsafe decision of red light running. Therefore, a wide range of utility theory family models, including multinomial, nested, and combined logit, and multinomial, ordered, and probit logistic regression models were calibrated. Then the goodness of fit and predictive ability of these two groups were compared. It was also investigated whether the consideration of drivers’ random taste variations and spatial heterogeneity with the help of utility theory models will improve the model acuuracy. For modeling, the data of three intersections of Nehzat, Modares, and Mazandaran in Semnan, Iran were extracted with the help of filming. The measures of the likelihood logarithm, likelihood ratio, percentage of correct predictions, and Akaike information criteria (AIC) showed that the goodness of fit of regression models is better than utility theory models. Validation with the help of 20% of the data also confirmed the same. But the advantage of utility theory models was to demonstrate the heterogeneity of drivers' tastes and spatial heterogeneity. The multinomial logistic regression model was recognized as the best model with 99.43% correct predictions.