Binomial and multinomial regression models for predicting the tactical choices of bicyclists at signalised intersections. (January 2019)

• Record Type: Journal Article
• Authors: Twaddle, Heather; Busch, Fritz
• Is Part Of: Transportation research. Volume 60(2019)
• Page Start: 47
• Page End: 57
• Publication Date: 2019-01
• Subjects: Bicyclist behaviour -- Tactical choice modelling -- Regression analysis; Automobile drivers -- Psychology -- Periodicals; Automobile driving -- Psychological aspects -- Periodicals; Transportation -- Psychological aspects -- Periodicals; 629.283019
• DOI: 10.1016/j.trf.2018.10.002
• Languages: English
• ISSNs: 1369-8478

Abstract

Highlights: Prediction of tactical choice outcomes of bicyclists at signalised intersections. Prediction based on observable attributes of the intersection and prior behaviour. Main and two-way interaction effects of 37 variables considered in model estimation. Identification of an optimum predictor set using recursive feature elimination. K-fold cross validation of resulting binomial and multinomial regression models. Abstract: Bicyclists are extremely flexible road users who employ various tactical behaviours to optimise comfort, directness and time efficiency while crossing a signalised intersection. Tactical choices faced by bicyclists at signalised intersections include whether to use the bicycle lane, roadway or sidewalk, to stop at or violate a red traffic signal, to ride with or against the mandatory direction of travel and the method of executing a left turn. The outcome of these choices has a direct impact on traffic safety and efficiency at intersections. In this paper, revealed choice data from 4710 bicyclists at four intersections in Munich, Germany are used to estimate binomial and multinomial logistic regression models to predict tactical choice outcomes. Optimal predictor sets are selected from the main and two-way interaction effects of 43 independent variables describing the situation, strategic behaviour and prior tactical choices of bicyclists using recursive feature elimination. A simplified model is estimated using the statistically significant variables of the optimal predictor set. The prediction power of the resulting regression model is assessed using k-fold cross validation. The models to predict response to a red signal and the type of left-hand turn exhibit high predictive power while the prediction of infrastructure selection and the direction of travel proves to be difficult.