Human mobility networks reveal increased segregation in large cities

The United States exhibits very high economic segregation, influencing where people live, who they marry, and who they befriend, with substantial costs for mobility, health, and political polarization. A prominent conjecture—the cosmopolitan mixing hypothesis—posits that large, dense cities should reduce everyday socioeconomic segregation by bringing diverse individuals into close proximity through increased diversity, constrained space, and public transit. However, large cities may also foster self-segregation by enabling people to find and frequent spaces with similar others. Testing these competing hypotheses has been challenging because most measures rely on static residential data rather than real-world, dynamic exposures. By analysing de-identified mobile phone GPS data, the study measures actual path crossings and encounters among individuals of differing socioeconomic status (SES), enabling city- and county-level measures of exposure segregation that capture where people go, when they go, and whom they encounter.

Prior work on segregation has largely used residential, census-based measures such as the neighbourhood sorting index, assuming exposure occurs among co-residents within home tracts. Although mobility data have been used to study health, social integration, and segregation, nationwide analyses of socioeconomic mixing have been limited by challenges in inferring individual SES, detecting dyadic exposures, and assembling sufficiently large, comparable datasets across regions. Related studies have examined experienced segregation using GPS data and neighborhood connectivity via mobility networks, but often without individual-level SES inference or dyadic path-crossing measurement at national scale. This study advances the literature by combining device-level proximity-based encounters with an individual SES proxy derived from home rent, producing a dynamic exposure-based segregation measure comparable to conventional indices while reflecting real-world interactions.

Data: De-identified mobile phone GPS pings (SafeGraph). Sample includes 9,567,559 devices with 1,570,782,460 dyadic exposure edges across 382 MSAs and 2,829 counties in the United States. Regions with at least 50 individuals are analysed. SES inference: Infer each device’s home location from night-time pings; assign SES using the estimated monthly rent at that home location. This improves on neighborhood-level census imputations. Exposure network: Define an exposure (path-crossing event) when two devices are within D metres within T minutes. Primary specification uses D = 50 m and T = 5 minutes (visual exposure scale), with robustness to alternative thresholds. Segregation metric: For each region, compute exposure segregation as the correlation between an individual’s SES and the mean SES of all individuals they encounter (both inside and outside their home tract). Estimate this correlation using a linear mixed-effects model to address attenuation from sparse observations and yield unbiased estimates. This generalizes the neighbourhood sorting index, which correlates individual SES with the mean SES of residents in the home census tract. Decompositions and heterogeneity: Decompose exposures by whether both, one, or neither individual is within their home tract; assess segregation by tie strength (number of exposures between a pair) and by point-of-interest (POI) categories; operationalize POI differentiation via average travel distance to nearest POI and total number of POIs. Restaurant case study: For full-service restaurants, quantify choice set size (number within 10 km of a resident) and SES differentiation (coefficient of variation of restaurant SES defined as the median SES of visitors who encountered another person at the POI). Bridging index: Identify frequently visited hubs as commercial centres (shopping centres, plazas, boardwalks). Cluster individuals by nearest hub from home; compute within-cluster SES diversity relative to the overall area. The index ranges from 0 (no bridging) to 1 (perfect bridging). Assess association with exposure segregation. Robustness checks include controls for race, population size, inequality, and alternative specifications.

• Large cities are more segregated by exposure: The Spearman correlation between MSA population and exposure segregation is 0.62; the ten largest MSAs are 67% more segregated than small MSAs (<100,000 residents). Results are robust to density-based analysis (Spearman = 0.45), controls, and alternative specifications. County-level segregation also increases with population size and density.
• Dynamic exposure segregation is substantially lower than conventional residential segregation: Median exposure segregation across MSAs is 38% lower than the neighbourhood sorting index, driven by out-of-tract exposures that increase diversity. Only 2.4% of exposures occur when both individuals are within their home tract. Exposures with both individuals outside their home tract are 50% less segregated than exposures within the home tract.
• Stronger ties are more segregated, and segregation varies across POI types. POI-level segregation correlates with the degree of socioeconomic differentiation of spaces. Categories with many, proximate venues (e.g., religious organizations) can target narrower communities and exhibit higher segregation than sparse, destination-type venues (e.g., stadiums). In median MSAs, religious organizations require 92% less travel distance and are 16 times more numerous than stadiums, and are 75% more segregated.
• Differentiation of space in large cities explains higher segregation: Residents of the ten largest MSAs have 22 times more restaurants within 10 km than residents of small MSAs. The coefficient of variation of restaurant SES in the ten largest MSAs is 63% greater than in small MSAs. Exposure segregation within restaurants is 29% higher in the largest MSAs.
• Hubs and bridging: 56.9% of exposures occur within 1 km of a commercial centre, though only 2.5% of land is within that distance. The bridging index (diversity near hubs) strongly predicts lower exposure segregation (Spearman = -0.78; the top ten MSAs by bridging index are 53.1% less segregated than the bottom ten). The bridging index predicts segregation more accurately than population size, SES inequality, neighbourhood sorting index, and race, and remains significant after controlling for these and other confounders.

The findings directly test the cosmopolitan mixing hypothesis and show the opposite pattern: larger and denser cities exhibit higher exposure segregation. The mechanism is that large cities sustain a greater variety of differentiated venues that serve narrow socioeconomic segments, generating homophilous encounters. Nonetheless, urban design can mitigate segregation: when frequently visited hubs are positioned to bridge diverse neighborhoods, they draw people from different SES backgrounds into shared spaces, increasing exposure diversity. By capturing real, time- and space-localized encounters, the dynamic measure reveals patterns and mechanisms of segregation that static residential measures miss, informing policies to promote socioeconomic mixing.

This study introduces a high-resolution, dynamic measure of exposure segregation based on mobile phone path crossings and an individual-level SES proxy, applied nationwide across MSAs and counties. Contrary to longstanding expectations, large cities are more segregated in everyday exposures, largely due to the socioeconomic differentiation of urban spaces. However, exposure segregation can be reduced when hubs are located to bridge diverse neighborhoods. These results recalibrate understanding of urban scale and mixing and highlight actionable levers in urban planning and zoning to foster integration. Future work could extend dynamic measures to underrepresented populations, refine tie-strength inference beyond proximity, integrate additional SES indicators, and further dissect how different urban designs and venue types shape exposure patterns.

• Exposure measurement relies on physical proximity (within 50 m and 5 minutes) as a proxy for social exposure and tie strength; while robust to stricter thresholds, it may not perfectly capture interaction intensity.
• Mobile phone datasets may underrepresent certain subpopulations (e.g., individuals without smartphones or with limited GPS activity), and GPS pings are unevenly distributed over time.
• Individual SES is proxied by housing consumption (estimated rent at inferred home), which does not fully capture socioeconomic status; results are robust to alternative SES measures but remain an approximation.
• While extensive controls and robustness checks are applied, unobserved confounders may remain when comparing regions and POI types.
• The bridging analysis uses commercial centres as hubs; other types of hubs or informal gathering spaces may also be influential but are not fully captured.