A panel data-based discrete-continuous modelling framework to analyze longitudinal driver behavior in homogeneous and heterogeneous disordered traffic conditions

ABSTRACT

We propose a panel data-based discrete-continuous modeling framework to analyze driver behavior in two disparate trajectory datasets – one from a heterogeneous disorderly (HD) traffic stream in India and another from a homogeneous traffic stream in the United States. The panel data-based framework allows the analyst to isolate the subject vehicle- and driver-specific unobserved factors that influence driver behavior. Doing so helps reduce the confounding effects of such unobserved factors on analyzing the influence of observed factors, such as relative speeds and spacing between the subject vehicle and other vehicles, on driver behavior. The empirical results reveal both similarities and differences in driver behavior between the two trajectory datasets. In addition, the analysis sheds light on the suitability of different lengths of influence zones on driver behavior in the two datasets. The insights from this study can help improve driver behavior models and traffic simulation frameworks for both traffic conditions..

KEYWORDS:

• Driver behavior
• homogeneous traffic conditions
• heterogeneous disorderly traffic conditions
• multi-vehicle anticipation
• panel data models

Acknowledgments

The authors are grateful to Punzo, Borzacchiello, and Ciuffo (2011), Montanino and Punzo (2013, 2015), and Kanagaraj et al. (2015) for sharing the trajectory datasets used in this study. Two anonymous reviewers provided useful comments on an earlier version of the paper.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1. When all datapoints in the trajectories were considered, the DW statistic values for the HD and homogeneous datasets were 1.335 and 0.502, respectively, which were smaller than the corresponding lower critical value (1.957). When we selected data points that were at least 2.5 seconds apart from each other for each subject vehicle, the DW statistic values for the HD and homogeneous datasets were 2.0010 and 2.140, respectively, which were greater than the corresponding upper critical value (1.966).

2. As an alternative to couplas, we explored introducing time varying common error components in Equations (2)(2) $u_{qit}^{\ast }={\beta }_i^T{x}_{qit}+{\omega }_{qi}+{\psi }_{qi}+{\epsilon }_{qit};i=a,d,s$(2) and (3(3) $m_{qit}={\alpha }_i^T{z}_{qit}\pm {\omega }_{qi}+{\xi }_{qi}+{\eta }_{qit};i=a,d$(3) ) to generate dependencies between discrete and continuous model components. Such an approach leads to likelihood functions that involve computationally intensive bi-level integrals (see Bhat and Castelar 2002) due to the presence of error components at two different levels – (1) one set of error components at the observation level to represent unobserved factors that vary across time instances and (2) another set of error components at the SV level to represent unobserved factors that do not vary across time instances. However, we found that using the copula functions to capture correlations at the observation level and common random error coponents to capture SV-level effects resulted in a model that was much less computationally intensive (because one level of simulation-based integration is obviated by the use of copulas) and yielded better goodness of fit to data.

Additional information

Funding

This work was supported by the Ministry of Education, Government of India, through a scholarship to the first author.