Using multi-agent simulation to understand impacts of COVID-19 and travel behavioral inertia on NYC

Note: this is ongoing work that may be subject to revision.

This week, NYC entered Phase 1 of reopening from COVID-19 stay-at-home orders that persisted since mid-March. C2SMART at NYU has been monitoring the effects of the pandemic on mobility in various cities closely. With the reopening, we are now sharing Issue 3 of our White Paper series that focuses on the center's efforts to model the impacts of the pandemic on travel behavior in NYC and the resulting impacts on public transit reopening strategies. This article provides some technical supplemental details to the multi-agent simulation modeling portion of that issue. The primary researchers involved in this work were Ding Wang, Brian Yueshuai He, and Jannie Gao, with joint supervision from me, Kaan Ozbay and Shri Iyer.

As many have seen, data from the impact of COVID-19 on NYC mobility patterns is readily available.

C2SMART recently developed a multi-agent simulation, dubbed MATSim-NYC, which is intended to be the first of a "Network of Living Labs" virtual test beds for evaluating emerging transportation technologies and policies (look forward to another post in the near future on our analysis of NYC's congestion pricing and BQX using this tool, which we are making accessible for public agencies in NYC and research partners). As part of that development, we calibrated a behavioral model (estimated as a tour-based nested logit model) for choosing different modes of transportation in NYC, which was encoded into the MATSim-NYC simulation as utility function parameters, where FHV is for "for-hire vehicles".

Using the COVID mobility data, we recalibrated the behavior model to reflect the increased aversion to use shared use modes like public transit. We first identified the proportions of the population that would not work from home (WFH) using study data from Dingel and Neiman (2020) , and with New York State reopening phase plan, constructed the following schedule of WFH proportions based on NAICS employment industries.

The synthetic population in MATSim-NYC was adjusted to fit the proportions under COVID using random selections. The mode choice utility parameters were then adjusted to make the resulting non-WFH mobility patterns fit the observed data, resulting in the following re-calibrated behavior.

We call this the MATSim-NYC-COVID model.

These modifications are in the presence of the other modes, so it implies that there would be greater preference to drive, walk, and bike, and less preference for transit, and to less extent, taxi, FHVs, and Citi Bike. We had two different population segments, which indicates same trends but different impacts on Manhattan and non-Manhattan travelers.

Validation of the model using transit station turnstile data shows for an 11% sample of station ridership data, there is only an average difference of 13% in the MATSim-NYC model in predicting station-level trips and a 24% difference during COVID.

Comparisons of time-of-day average link travel speeds are shown in Figure 1.

A comparison is shown between pre-COVID and during COVID activity patterns.

PRE-COVID:

DURING COVID:

We then used the model to analyze reopening scenarios, assuming there is inertia in users' behavior, i.e. they do not immediately revert back to prior behavior (they maintain the same risk aversion to shared use transportation modes).

MTA is considering different strategies and putting certain ones in place with Phase 1 reopening. We also see from other cities in China that a policy of restricting transit capacity to 50% is being used. First, we analyze the Phase 4 full reopening under behavioral inertia to see how road traffic and transit trips would vary. This is seen in Scenario 1 in Figure 2, where CB is Citi Bike, and Ride is for carpooling. These values represent the total number of daily trips produced per mode under COVID and each of the phases of reopening, divided by the total daily trips during pre-COVID. The Phases 1-3 were modeled assumed transit was not fully reopened yet.

Even by Phase 4 and without transit capacity restrictions, we can see that transit ridership may only return to 73% of pre-COVID levels while passenger car trips may exceed 142% of those levels. This demonstrates the amount of amplification to post-reopen traffic if there is travel behavioral inertia.

We also look at Scenario 2, which sets a transit capacity restriction to 50%. The overall transit ridership under Phase 4 may go down further to 64% of pre-COVID level, but car trips may not go up much further (from 142% to 143%). The capacity restriction appears to be supported by shifts in passengers to non-car modes, particularly micromobility options like Citi Bike that may go up from 0.92 to 1.84 times pre-COVID levels.

This result is encouraging for employing transit capacity restriction, as it may not add significantly to car traffic (which is a primary concern) relative to what behavioral inertia is already adding.

At the same time, transit officials should take note of this and combine the transit capacity restriction strategy with adequate support for micromobility options.

In addition to Issue 3, we are now preparing a working paper. Feel free to reach out for more details.