Behavior-based dependency networks between places shape urban economic resilience

Takahiro Yabe, Bernardo Garcia-Bulle, Morgan R. Frank, Alex Pentland, Esteban MoroResearch Square 2024 · DOI

TL;DR

Researchers have developed a novel way to measure urban economic resilience by mapping "behavior-based dependency networks"—connections between businesses formed by the actual movement of people rather than just physical proximity. By analyzing mobile phone data across five major US cities, they found that these behavioral networks are 40% more accurate at predicting how businesses survive economic shocks (like COVID-19) compared to traditional distance-based models.

What Problem Does This Paper Address?

When a major "anchor" in a city—like a university or a large office complex—shuts down, the economic fallout ripples outward. Traditionally, urban planners and economists have assumed these ripples follow the "First Law of Geography": things that are closer together are more related. Under this logic, if a college closes, only the shops on the same block should suffer.

However, modern urban life is more complex. A commuter might stop at a specific gas station 10 miles from their office or a gym near their child’s daycare. The authors argue that urban resilience is dictated by a "demand-side" network of dependencies mediated by human mobility. The gap they address is the lack of a quantitative framework to measure these non-local, behavior-driven dependencies and understand how they amplify or curb economic shocks.

Methods

The researchers constructed a Behavior-based Dependency Network using a directed, weighted graph approach:

  • Co-visitation Metric: They defined the dependency of place $i$ on place $j$ as the proportion of visitors to $i$ who also visited $j$ on the same day (within a 6-hour window and without intermediate stops).
  • Network Modeling: This created an asymmetric adjacency matrix ($W$). Unlike physical distance, these dependencies are one-way; a cafe might depend heavily on a nearby hospital, but the hospital’s foot traffic is independent of the cafe.
  • Leontief Open Model: Borrowing from classical input-output economics, they applied a linear model to simulate how a "shock" (a reduction in visits) to one node cascades through the entire network.
  • Null Model Comparison: They compared their results against a "Gravity-based null network," which stochastically generates connections based strictly on physical distance and POI popularity.

Datasets

The study leveraged two massive, high-resolution datasets across five US metropolitan areas (New York, Boston, Los Angeles, Seattle, and Dallas):

  1. Spectus: Anonymized GPS location pings from approximately 15 million daily active users in the US.
  2. SafeGraph: Data on over 1 million verified Points-of-Interest (POIs), including businesses, stores, and urban amenities.
  3. Validation Data: They used visitation changes during the COVID-19 pandemic (March–June 2020) as a real-world "stress test" for their model.

Key Insights and Ideas

  • The "Long Tail" of Dependency: While dependencies do decay with distance, the empirical network has a much higher rate of long-distance connections than distance-based models suggest. Cities are "smaller" than they look on a map because of behavioral shortcuts.
  • Functional Biases: Certain categories of places are "hubs" of dependency. For example, airports, stadiums, and theme parks have massive, far-reaching impacts (2.5km to 3.5km), while supercenters and colleges have intense but more localized cascading effects (~1.5km).
  • Asymmetry Matters: The model captures the reality that a "seed" node (like a university) can support hundreds of smaller businesses, but the closure of one of those small businesses has negligible impact on the university.
  • Underestimation Risk: The authors found that neglecting these behavioral links leads to a substantial underestimation of how far and wide an economic shock will spread.

Key Results

  • 40% Better Predictability: Using behavior-based networks improved the $R^2$ (predictive power) of business resilience during COVID-19 by 40% compared to models relying only on physical proximity.
  • Magnitude of Effect: The regression coefficients for behavioral dependency were 2 to 3 times larger than those for physical distance, suggesting that who visits a place is a much stronger predictor of its fate than where it is located.
  • Simulation Power: In a simulated 50% reduction of college visits (mimicking a shift to online learning), the model successfully identified significant negative impacts on arts, museums, and food services located far beyond the immediate campus vicinity.

Limitations and Future Work

  • Static vs. Dynamic Networks: The study assumes the dependency network is static. In reality, people might reorganize their behavior during a long-term shock (e.g., finding a new favorite cafe after a job change).
  • Visit Intent: The data doesn't distinguish between different types of visits (e.g., a "routine" commute stop vs. a "discovery" trip to a new gallery).
  • Demographics: While the study uses post-stratification to ensure representativeness, it does not yet decompose dependencies by income or sociodemographic groups, which could reveal how economic shocks disproportionately affect different communities.
  • Future Direction: The authors suggest this framework could be used to design "surgical" urban interventions—identifying which businesses need the most support to prevent a systemic economic collapse during future disasters or technological shifts (like the rise of EVs or remote work).

Summary updated

Abstract

<title>Abstract</title> Urban economic resilience is intricately linked to how disruptions caused by pandemics, disasters, and technological shifts ripple through businesses and urban amenities. Disruptions, such as closures of non-essential businesses during the COVID-19 pandemic, not only affect those places directly but also influence how people live and move, spreading the impact on other businesses and increasing the overall economic shock. However, it is unclear how much businesses depend on each other in these situations. Leveraging large-scale human mobility data and millions of same-day visits in New York, Boston, Los Angeles, Seattle, and Dallas, we quantify dependencies between points-of-interest (POIs) encompassing businesses, stores, and amenities. Compared to places' physical proximity, dependency networks computed from human mobility exhibit significantly higher rates of long-distance connections and biases towards specific pairs of POI categories. We show that using behavior-based dependency relationships improves the predictability of business resilience during shocks, such as the COVID-19 pandemic, by around 40% compared to distance-based models. Simulating hypothetical urban shocks reveals that neglecting behavior-based dependencies can lead to a substantial underestimation of the spatial cascades of disruptions on businesses and urban amenities. Our findings underscore the importance of measuring the complex relationships woven through behavioral patterns in human mobility to foster urban economic resilience to shocks.
BibTeX
@inproceedings{yabe2024behavior,
  author = {Takahiro Yabe and Bernardo Garcia-Bulle and Morgan R. Frank and Alex Pentland and Esteban Moro},
  title = {{Behavior-based dependency networks between places shape urban economic resilience}},
  booktitle = {Research Square},
  year = {2024},
  doi = {10.21203/rs.3.rs-3683842/v1},
}
Figure