Investigating neighborhood adaptability using mobility networks: a case study of the COVID-19 pandemic

The study examines how neighborhoods adapted their mobility behaviors in response to COVID-19 non-pharmaceutical interventions (NPIs), particularly shelter-in-place policies. Mobility in cities arises from complex interactions among diverse agents and is closely linked to social interactions, health, productivity, and economic resilience. Representing mobility as networks enables insights into the impacts of policy changes and pandemics. The research question focuses on which socioeconomic and geographic factors predict a neighborhood’s adaptability, defined as the ability to modify behavior under NPI constraints. Using New York City (NYC) as a case study, the authors construct weekly mobility networks for 2019–2020 from aggregated visits to points of interest (POIs) and analyze changes in network structure, dissimilarities year-on-year, and centrality metrics across demographic groups. The purpose is to inform scientists and policy-makers about determinants of adaptability to exogenous shocks, improving urban planning and public health responses.

Prior work links mobility to social interactions, health outcomes, productivity, and regional economic resilience. Network analyses have been used to understand COVID-19 impacts, including changes in mobility structure due to NPIs and their epidemiological consequences. Studies coupled mobility networks with epidemiological models to assess business closures, vaccine policies, and latent mobility patterns. This study differs by dissecting topological changes in mobility networks over the pandemic using established network metrics (e.g., centrality), highlighting roles of neighborhoods as potential bridges and spreaders, and pairing this with a gravity-model simulation framework to test how access to amenities may shape adaptability.

Datasets: (1) SafeGraph Weekly Patterns data (January 2019–December 2020) providing anonymized, aggregated weekly visits from CBGs to POIs within the New York Metropolitan Area (6,493 NYC CBGs; 333,241 POIs). (2) Google COVID-19 Community Mobility Reports offering relative mobility changes for categories versus a pre-pandemic baseline. (3) COVID-19 cases: Johns Hopkins CSSE and NYC ZCTA-level statistics; weekly CBG-level case estimates were inferred by weighting ZCTA case counts by the proportion of each CBG’s population within intersecting ZCTAs, then normalized by CBG population to obtain infection rates. (4) U.S. Census ACS 2019 5-year estimates at CBG level for demographics (population, income, education, commute time, race), also expressed as percentiles. Study area: NYC within the broader New York Metropolitan Area; NYC’s five boroughs analyzed. Mobility networks: For each week t (2019–2020), construct a weighted directed graph G(t) where nodes are CBGs; edge weight w(i→j) equals the aggregated number of visits by residents of CBG i to POIs located in CBG j. Self-loops capture visits to POIs within the home CBG. Temporal topological shifts and dissimilarity: Compute ego-network-based node feature vectors per CBG using NetSimile-style aggregated descriptors (e.g., degree, clustering). Align corresponding weeks of 2019 and 2020 and compute node-level dissimilarity via Canberra distance between paired feature vectors, yielding a per-CBG weekly dissimilarity score. Rank CBGs weekly, identify cohorts frequently appearing in top and bottom dissimilarity quartiles during the first wave (March–June 2020) for spatial and socioeconomic profiling. Centrality evolution: Track temporal changes in betweenness, in-degree, out-degree, and self-visit ratio S_c^t = W_c^t / (W_c^t + W_o^t), where W_c^t is the sum of self-loop weights and W_o^t the sum of outgoing edge weights, as an indicator of locality of visits. Compare these metrics across socioeconomic quartiles (income, education, etc.). COVID-19 hotspots and bridges: Define weekly hotspots as CBGs in the top quartile of new COVID-19 cases at time t. Identify potential bridge CBGs that had outgoing edges to hotspot CBGs at t−2 (two-week incubation). Apply frequency analysis over time and mark bridge CBGs as those in the 75th percentile (also examined at 95th). Analyze their socioeconomic and geographic characteristics and borough distribution. Huff gravity model simulation: Employ a basic Huff model P_ij ∝ A_j / d_ij^β at the census tract level (aggregates of ~4 CBGs) to model grocery store patronage during lockdown (March 22–June 8, 2020), when trips were primarily single-purpose and essential. Use store floor area A_j and tract-to-store distance d_ij; fit/derive probabilities from observed data. Create hypothetical scenarios for Staten Island by randomly adding grocery stores (with area equal to the island’s mean grocery store area) until grocery store density per 1,000 residents matches Manhattan’s and Queens’. Generate synthetic visit counts from probabilities and quantify changes in simulated visits from Staten Island tracts to hotspot CBGs. Analytical setting summary: (i) Compute week-aligned 2019 vs. 2020 node-level dissimilarities; (ii) analyze centrality metrics by socioeconomic quartiles; (iii) detect hotspot-linked bridge CBGs via two-week lag connections and frequency filtering; (iv) simulate Staten Island mobility under increased grocery POI density via Huff model to estimate impacts on visits to hotspot CBGs.

• CBG-level dissimilarity: During March–June 2020, CBGs with the largest changes (top dissimilarity quartile) were predominantly in Manhattan and socioeconomically advantaged: 63% in top income quartile, 79% in top education quartile, 62% in top white population quartile, and 52% in bottom commute-time quartile. CBGs with the least change (bottom dissimilarity quartile) lacked a clear socioeconomic profile but trended toward lower income, education, and white population shares.
• Centrality dynamics: Before March 2020, top-income CBGs had higher betweenness centrality, acting as key bridges. Post onset of the pandemic, their betweenness dropped abruptly while less affluent CBGs gained higher betweenness until around September 2020, indicating a shift in bridging roles. Degree centrality patterns show affluent and higher-education CBGs reduced mobility more than less affluent ones. Self-visit ratio rose more among top-income CBGs from March to June 2020, reflecting greater localization of activity; disparities narrowed after reopening began in June 2020.
• COVID-19 hotspots and bridges: Bridge CBGs (frequently linked to hotspot CBGs with a two-week lag) were overrepresented among lower income and lower education quartiles and higher minority shares. Staten Island stood out spatially, with many bridge CBGs despite relatively high income and white population shares, attributable to limited local POIs and constrained connectivity leading to longer travel for essential needs and work.
• Grocery access disparities: Borough-level indicators (March 22–June 8, 2020): grocery stores per 1,000 residents—Manhattan 0.582, Brooklyn 0.470, Bronx 0.435, Queens 0.414, Staten Island 0.332; median distance to grocery stores (km)—Manhattan 0.90, Brooklyn 1.35, Bronx 1.26, Queens 1.71, Staten Island 2.66.
• Simulation outcomes: Increasing Staten Island’s grocery store density to match Manhattan’s reduced Staten Island residents’ visits to hotspot CBGs by 47%; matching Queens’ density reduced such visits by 23%. These reductions imply lower exposure to potential spreaders and potentially lower infection and mortality rates.

The analysis shows that neighborhood adaptability to NPIs is shaped by both socioeconomic characteristics (income, education, race) and geographic attributes (access to essential amenities and connectivity). High-income, highly educated, predominantly white CBGs showed the greatest structural mobility changes year-on-year, consistent with greater ability to work remotely and reduce travel. Conversely, neighborhoods with less change in ego-network structure had higher infection rates, reflecting higher shares of frontline and on-site workers who could not reduce mobility. A key insight is that geographic constraints can limit adaptability even in relatively affluent areas: Staten Island’s sparse POI landscape and limited transport links led to longer trips and sustained interactions with hotspots, diminishing adaptability and increasing fragility to infection despite favorable demographics. The Huff-model simulation indicates that improving local access to essential POIs can significantly reduce interactions with hotspot areas. These findings inform urban planning by emphasizing integrated consideration of socioeconomic and physical neighborhood structures to enhance resilience and reduce disease transmission during exogenous shocks.

This study leverages mobility networks to quantify neighborhood adaptability during COVID-19, demonstrating that both socioeconomic factors and geographic accessibility strongly predict mobility responses under NPIs. Contributions include: (i) year-on-year, node-level ego-network dissimilarity and centrality analyses across demographic groups; (ii) identification of hotspot bridge CBGs and associated characteristics; (iii) scenario simulations using a Huff gravity model showing that increasing access to essential amenities can reduce interactions with hotspots. Policy implications suggest that strengthening local access to essential services and diverse amenities, alongside considering labor structure and transport connectivity, can improve community adaptability and resilience. Future research directions include expanding simulations to more POI categories and attractiveness features, incorporating physical neighborhood characteristics, and visualizing hypothetical changes via inter-borough flow graphs.

• Data aggregation: Mobility and POI visits are aggregated weekly at CBG level, limiting ability to infer trip purposes (work vs. non-work) and to distinguish essential vs. non-essential trips.
• Mode of transport: The datasets do not capture transportation modes or routes; linking mobility to transmission via mode shares is not feasible with current resolution.
• POI coverage bias: The POI dataset focuses on places involving financial transactions and underrepresents workplaces/offices, constraining identification of essential commutes and gravity model fitting.
• Implementation constraints: Recommendations relying on increased POI density assume private sector provision in potentially less profitable areas; real-world feasibility may require public incentives (e.g., tax exemptions, zoning).
• Modeling simplifications: The Huff model uses only distance and floor area at the tract level during lockdown; broader behavioral factors and multi-purpose trip chaining are not modeled, although lockdown conditions mitigate this concern.