Mobility patterns are associated with experienced income segregation in large US cities

The study addresses how income segregation is experienced by individuals in cities, beyond traditional residential measures. As urbanization and mobility increase, segregation and inequality affect access to housing, services, and economic outcomes. Prior metrics often approximate exposure via residential areas, but most interactions occur in specific places outside the home and vary by place type and time. The research question is whether and how individual mobility behaviors and place characteristics shape experienced income segregation, and whether experienced segregation can be captured by behavioral mobility models rather than only residential patterns.

The paper situates itself within work measuring exposure and interaction across income groups, noting that actual social interactions are hard to observe, so exposure is often inferred from physical co-presence and activity spaces. Traditional measures emphasize residential segregation and neighborhood-level statistics. In parallel, a large mobility literature shows human movement is predictable, follows universal patterns, and can be modeled (e.g., Exploration and Preferential Return models). These strands suggest that experienced segregation might be embedded in mobility regularities and place-based encounters rather than solely in residential distributions.

Data: High-resolution anonymized GPS pings from Cuebiq (opt-in, privacy-enhanced; IRB exempt protocols #1812635835 and #E-2962) covering 11 U.S. CBSAs from Oct 2016 to Mar 2017. After filtering devices with >2000 pings, dataset includes 67.0B pings from 4.5M devices. Stays were extracted using the Hariharan–Toyama algorithm; visits defined as stays between 5 minutes and 1 day and matched to nearest venue within 200 m. Foursquare API provided ~1.1M verified venues; analysis focused on ~1M venues visited by >20 unique individuals. Home inference: most common location 10 p.m.–6 a.m. at Census Block Group (CBG) level (2012–2016 ACS). Individuals grouped into four city-specific SES quartiles by CBG median household income. Post-stratification was used to improve representativeness for population, income, and attendance; cross-checked with Twitter and official event attendance for highly active places. Final stays dataset: 976M stays from 3.6M individuals; experienced segregation computed for 1.9M individuals with venue visits. Measures: Place income segregation (S_a): compute fraction of total time at place a by each income quartile q, T_aq. Define S_a = sum over q of |T_aq − 1/4|, bounded in [0,1]; 0 indicates even exposure across quartiles, 1 indicates exclusivity to one quartile. Individual experienced income segregation (S_i): for individual i, compute exposure τ_iq = sum over places a of T_ia q, where T_ia is i’s time share at a and q indexes the place’s income composition; then S_i = sum over q of |τ_iq − 1/4|. Robustness to alternative segregation metrics and income groupings was assessed. Modeling place segregation: Regression with variables including Foursquare category, price tier, rating, place catchment range (average home-to-place distance of visitors), local place density, median income of hosting CBG, and geographic fixed effects at PUMA level. Variable importance assessed via Lindeman–Merenda–Gold method. Mobility modeling: Begin with the Exploration and Preferential Return (EPR) model, where each new visit is either an exploration (new place) with probability P_new or a return with probability 1−P_new; return probabilities are proportional to accumulated time at places. Fitted parameter γ ≈ 0.23 ± 0.02 matches prior studies. Define place exploration σ_p = S_i/N, where S_i is the number of unique places visited and N is total visits. Extend to social-EPR by adding a Schelling-inspired social exploration parameter σ_s, the probability that when exploring a new place, the individual chooses one where their income group is a minority (threshold at 50% majority). The social-EPR model thus has two parameters per individual: σ_p (place exploration) and σ_s (social exploration). The model generates predicted experienced segregation which is compared to observed S_i. Explaining σ_p, σ_s, and S_i: Linear regressions with three predictor groups: (R) residential demographics (∼30 CBG variables: income, education, employment, race composition, poverty, transportation modes, etc.), (P) lifestyles measured as fraction of time spent in place categories, and (M) geographical mobility (radius of gyration, total distance traveled). Models include PUMA fixed effects to control for area-level heterogeneity and account for structural opportunity. Variable group importance and category associations were analyzed.

• Place-level segregation is highly heterogeneous and granular: economically mixed and highly segregated places can exist within tens of meters; spatial correlation of place segregation is low even locally (~50 m), while CBG median income remains correlated at >1 km and even >10 km.
• The distribution of normalized place segregation is similar across diverse U.S. metro areas, indicating consistent patterns.
• Regression importance: Place category and catchment range (average visitor travel distance) are the top predictors of place segregation after PUMA fixed effects, explaining ~18% and ~15% of variance respectively. Unique, city-wide destinations (arts venues, museums, airports) tend to be more integrated; local-serving places (e.g., houses of worship, grocery stores) tend to be more segregated. Category and range do not fully explain segregation, with within-category dispersion (e.g., factories more segregated than offices or supermarkets despite similar range).
• Individuals travel far for encounters: average distance to visit a given place is 9.5 km; 78% of encounters occur with people from a different PUMA; only 3% within the same CBG.
• Individual experienced segregation S_i is heterogeneous and only weakly related to home area income (correlation p = −0.173 ± 0.002). Spatial correlation of S_i decays beyond ~50 m, suggesting residence does not primarily determine S_i.
• Individuals spend most time in a small set of places (heavy-tailed time distribution). Computing S_i using only top 10 visited places yields 0.97 correlation with full-set S_i.
• Exploration vs. return: Individuals differ in place exploration σ_p; returners (low σ_p) are more segregated than explorers (high σ_p). Correlation between S_i and total number of unique places visited is high (p = 0.411 ± 0.001) after controlling for number of visits.
• Individuals’ S_i correlates strongly with the average segregation of places they visit (p = 0.579 ± 0.001). S_i computed using only “social” venues (education, colleges, work, worship, arts/museums, sports, entertainment) correlates with overall S_i (p = 0.754 ± 0.001).
• EPR alone reproduces visitation regularities (e.g., Zipf-like time distribution) but not S_i variability. The social-EPR model with σ_p and σ_s reproduces observed S_i well: correlation between model-predicted and observed S_i is p = 0.777 ± 0.001.
• Top places are on average ~14% more segregated than others; correlation between place segregation and time spent is small (≈0.049 ± 0.001).
• Most individuals have high σ_s (≈80% have σ_s > 0.75), so S_i is driven by the interplay of σ_p and σ_s: only those with both high σ_p and high σ_s are economically integrated. Correlation of S_i with σ_s alone is moderate (p = 0.538 ± 0.002). σ_p and σ_s are only moderately correlated with each other (ρ = 0.126 ± 0.002).
• Determinants of exploration: Place exploration σ_p is primarily explained by mobility behavioral variables (lifestyles/place categories) and, to a lesser extent, geographical mobility; it is weakly tied to residential demographics. High σ_p associates with spending time in entertainment, food, and shopping categories; low σ_p associates with education and certain work categories (factory, warehouse). Average σ_p ≈ 0.43.
• Social exploration σ_s is mostly predicted by residential demographics (82% of variance explained): higher σ_s in neighborhoods with higher education and income, fewer Black residents, less public transit usage, lower poverty; lifestyles contribute ~14% of σ_s variance.
• Overall S_i is explained by both behavior and residence: mobility behavioral factors (especially types of places visited) account for ~55% relative importance; residential (census) factors account for ~45%.

Findings show that experienced income segregation is a place-based behavioral phenomenon, not adequately captured by residential patterns alone. Place types and their catchment range systematically relate to place segregation, and individuals’ mobility behaviors—how often they explore new venues and whether they venture into places where their income group is a minority—jointly shape their experienced segregation. The social-EPR model bridges human mobility and segregation literatures, demonstrating that two interpretable parameters (σ_p and σ_s) capture much of the variation in S_i across 11 U.S. cities. Policy and urban design implications include the need to evaluate how transit, zoning, and amenity provision influence both place and social exploration, thereby affecting cross-group exposure. The framework suggests other urban processes dependent on mobility (transportation, pollution exposure, epidemics) may also be influenced by exploration behaviors.

The paper contributes by (1) reframing income segregation as a behavioral, place-level process with fine spatial granularity and consistent category effects across cities, and (2) introducing a simple, interpretable social-EPR model that accurately reproduces individual experienced segregation using place and social exploration parameters. The results underscore that both where people live and how they move through and select places drive experienced segregation. Future research should investigate causal levers—such as transit expansions, commercial developments, and zoning changes—on exploration behaviors, and extend the framework to other dimensions of segregation (race, wealth, ethnicity) and other mobility-driven urban processes.

The analysis includes only individuals with identifiable home locations during the 6-month window, potentially excluding those without stable residences or with nonstandard work hours. Venue data are limited to Foursquare listings and may be biased toward certain place types. The study infers exposure via co-presence probabilities and cannot distinguish the nature or quality of encounters (e.g., casual interactions vs. service transactions). The focus is on income segregation, not other segregation dimensions (race, wealth, ethnicity). Results are descriptive and not causal; potential confounding remains despite controls and fixed effects.