Introduction

One of the enduring challenges in statewide transportation planning is that consistent population travel data remains scarce, particularly for underserved and rural communities. This is changing with the availability of large-scale ICT data. Replica Inc. developed a nationwide synthetic population dataset that includes both sociodemographic information and trip/activity details. With this unique data opportunity, it is now possible to develop behavioral models for a range of different population segments such that new mobility scenarios can be analyzed to determine the forecasted ridership, revenue to those services, and change in consumer surplus. 

We developed group level agent-based mixed (GLAM) logit to deterministically fit heterogeneous parameters for trips along each census block-group OD pair conducted by each population segment. The advantage of GLAM logit is as follows. First, GLAM logit takes OD level (instead of individual level) trip data as inputs, which is efficient in dealing with ubiquitous datasets containing millions of observations. Second, preference heterogeneities are based on non-parametric aggregation of coefficients per agent instead of having to assume a distributional fit. Third, since each agent’s representative utility function is fully specified, GLAM logit can be directly integrated into system design optimization models as constraints.

For details on the GLAM logit model, please refer to BUILTNYU/GLAM-Logit (github.com). The deliverables include:

Mode choice parameters for NYC and NYS:  Download from Zenodo

We provide two datasets of census block group-level mode choice parameters for New York City and New York State. The parameters are estimated by GLAM logit model using Replica’s synthetic population datasets. Each row contains a set of mode choice parameters for each block-group OD pair and one of the four population segments (low-income, not low-income, students, and senior population). Six trip modes are considered: private auto, public transit (such as buses, light rail, and subways), on demand auto (taxi or TNC services such as Uber or Lyft), biking (including e-bike), walking, and carpool. Parameters of twelve mode attributes are estimated, including, auto travel time, transit in-vehicle time, transit access time, transit egress time, number of transit transfers, non-vehicle travel time, trip cost, and five alternative specific constants (setting carpool as the reference level).

In New York City, the average value of time (VOT) of low-income population is 21.67$/hour, the average VOT of not low-income population is 28.05$/hour, the average VOT of student population is 10.96$/hour, and the average VOT of senior population is 10.93$/hour. In New York State, the average value of time (VOT) of low-income population is 9.63$/hour, the average VOT of not low-income population is 13.95$/hour, the average VOT of student population is 7.40$/hour, and the average VOT of senior population is 6.26$/hour. 

The empirical distribution of agent-level parameters is neither Gumbel nor Gaussian, which reveals a regional divergence of the value of time and mode preference, indicating potential inequity issues in the transportation system. This is infeasible for conventional discrete choice models (DCMs) to capture. 

Mode choice probabilities for NYC and NYS:  Download from Zenodo

We provide two datasets of predicted mode share, one for New York City and another for New York State. Each row contains the mode proportion of trips along a census block group-level OD pair made by one of the four population segments: low-income, not low-income, students, and senior population. 

The prediction is based on GLAM logit model calibrated with Replica’s statewide synthetic population dataset. The in-sample prediction accuracy is quite competitive, with an overall accuracy of 90.28% in New York State and 88.63% in New York City. 

If you have any questions or comments, please contact: Xiyuan Ren (xr2006@nyu.edu) or  Joseph Chow (joseph.chow@nyu.edu).