Understanding Socioeconomic Disparities in Travel Behavior during the Covid-19 Pan

New working paper: Understanding Socioeconomic Disparities in Travel Behavior during the Covid-19 Pandemic

Authors: Rebecca Brough (Notre Dame), Matthew Freedman (UC-Irvine), David Phillips (Notre Dame)

Abstract: We document the magnitudes of and mechanisms behind socioeconomic differences in travel behavior during the Covid-19 pandemic. We focus on King County, Washington, one of the first places in the U.S. where Covid-19 was detected. We leverage novel and rich administrative and survey data on travel volumes, modes, and preferences for different demographic groups. Large average declines in travel, and in public transit use in particular, due to the pandemic and related policy responses mask substantial heterogeneity across socioeconomic groups. Travel intensity declined considerably less among less-educated and lower-income individuals, even after accounting for mode substitution and variation across neighborhoods in the impacts of public transit service reductions. The relative inability of less-educated and lower-income individuals to cease commuting explains at least half of the difference in travel responses across groups.

In addition to the main point of the paper, it may be of interest to users of SafeGraph data that we have a direct measure of individual income linked to travel behavior in transit data. So, we can compare socioeconomic gaps by neighborhood in the SafeGraph data vs. socioeconomic gaps by neighborhood in the transit data vs. socioeconomic gaps by person in the transit data. They all look pretty similar, which is encouraging for making conclusions about income disparities just using neighborhood information.

Nice work! We just got APC data for our local transit system and may be interested in investigate similar matters. We’ve done similar correlation with education (https://stanfordfuturebay.github.io/sanjose_correlation) but used a simpler measure from Safegraph, % leaving home. Currently we’re creating more sophisticated measures like # of visit-hours to POIs as we try to examine mechanisms for disease spread. @Simone_Speizer_Stanford @Sam_Liu_Stanford @Cameron_Tenner_Stanford

Question about your APC methodology – my understanding is that you are taking all ONs from APC associated with a stop lat/long, and joining that to the CBG that lat/long is located in. Do you do anything to distinguish between ONs that may be riders coming from home vs. ONs that may be riders coming from a destination (e.g. workplace, POI)? If not, it seems like the signal you’re trying to associate with the residential CBG will generally be diluted with an equally large signal turning up elsewhere in the transit network.

@David_Phillips_University_of_Notre_Dame Very rigorous paper with the argument that the disparities of travel indeed come from economic necessity. While your regression models are clearly focused on the CBG level, for the travel pattern figures, I have a similar question as @Derek_Ouyang_Stanford: how did you link the CBG to devices and boardings? Did you use some measure to determine residential CBG of the devices?

@Derek_Ouyang_Stanford Sorry for the slow response; new to slack and just saw this. That’s right, we’re using onboardings from the APC data by mapping lat/long of the stop to Census geography. Seattle’s bus system can’t link origin and destination of a trip, so you’re right we can’t distinguish trips starting in a place and those leaving a place. @Ruowei_Yang_UM_Baltimore For the link of devices to CBG, we’re just using the Safegraph social distancing dataset where they have already inferred the home location of the device. But then the data has the CBG-to-CBG connections of what other CBGs those devices go to. We use that data to back out how many CBGs other than the home CBG the devices visit, which is not quite the same as boardings but gets something close. I think @Derek_Ouyang_Stanford is right these aren’t quite the same. But we could in principle double-count using the safegraph data, both those originating and those ending in a CBG, to get something more similar to the APC data. Best idea that comes to mind at the moment. Welcome to other ideas.