High-resolution human mobility data reveal race and wealth disparities in disaster evacuation patterns

The study examines how pre-existing social disparities in race and wealth translate into differential evacuation behaviours during a major disaster. Despite extensive discussion of equity and environmental justice in hazards research, there is limited large-scale quantitative evidence linking socioeconomic factors to evacuation and return behaviours. Leveraging Hurricane Harvey—an extreme, high-impact event without a mandatory evacuation in Houston—the authors use high-resolution GPS mobility data to quantify who evacuated, where they went, and how long they were gone, and to assess the role of neighbourhood race and wealth in those patterns. Houston’s large, diverse metropolitan area provides a compelling testbed to evaluate disparities and inform emergency planning and policy.

Prior work links social vulnerability, equity, and environmental justice to disaster impacts, showing minorities and poorer residents are disproportionately affected (e.g., Hurricane Katrina). Studies report race and class gaps from preparation through recovery, with White communities typically better prepared and returning at higher rates than Black or Hispanic residents. Social capital and networks have been associated with faster recovery. Traditional evidence often relies on retrospective surveys with biases. Recent GPS-enabled mobility datasets have enabled detailed analyses of behaviour in disasters, including modelling movements, highlighting the role of social networks and socioeconomic status, and identifying common features such as exponential return rates across events. This study builds on that foundation by applying large-scale mobility data to Hurricane Harvey to quantify disparities in evacuation and return behaviours by neighbourhood race and wealth.

Data and preprocessing: Anonymized mobility data from Cuebiq covered the Houston MSA from July 1 to September 30, 2017, totaling over 3 billion records from >2.5 million unique opted-in devices. A stay-point algorithm (temporal threshold 5 min; spatial threshold 50 m) extracted meaningful stays and filtered transient points. Users retained had data from at least 60 unique days and at least 100 stay points, yielding ~30 million points for ~150,000 users. Sampling bias correction: Representativeness was assessed by correlating device counts per census block group with ACS 2018 population. To mitigate sampling bias: (i) block-group weighting by the ratio of active devices to ACS population, and (ii) bootstrapping samples from each block group with a uniform population-proportional sampling rate (100 iterations) were applied; results were compared to unweighted outcomes. Home detection: For each device, candidate weekly home locations prior to landfall were identified using weekday evenings (8 p.m.–7 a.m.) stay points. Agglomerative clustering with complete linkage (max cluster diameter 50 m) was used; the cluster with the longest cumulative stay was the candidate home. A minimum of stays on two different days was required. Home CBGs were determined across 5 consecutive weeks with cross-validation allowing one week missing/mismatch and a 50 m Haversine tolerance. Evacuation detection: Using the cross-validated home as baseline, a rolling 5- and 7-day sliding window detected primary locations and their deviation from home. If a user was away from home for at least three consecutive calendar days with a displacement exceeding a 1 km cluster diameter threshold, the episode was labeled as an evacuation; the new primary location was the destination. The middle date of the window was taken as the departure date; return dates were analogously inferred via stabilization at the home location with the sliding window. Net evacuation intensity estimation: Pre- and post-evacuation home-location densities were estimated via kernel density estimation with a Gaussian kernel and bandwidth r = 100 m on 100×100 m grids. K_pre and K_post denote pre- and post-evacuation densities; net evacuation intensity K_net = K_post − K_pre was computed and normalized to zero mean within [−1, 1] to compare across subgroups. Neighbourhood classification: Census block groups were categorized by majority race (≥50% White, Black, or Hispanic) and wealth (poor if ≥25% below federal poverty level; non-poor otherwise), yielding six types: non-poor White, poor White, non-poor Black, poor Black, non-poor Hispanic, poor Hispanic. Sensitivity analyses varied thresholds. Analyses: Evacuation rates were calculated (weighted and unweighted) by neighbourhood type; spatial patterns were mapped via K_net for all, long-distance (>90th percentile, >41.25 km), and short-distance (<10th percentile, <2.71 km) evacuations. Evacuation distance distributions were fitted with a truncated power-law. Origin-destination transition matrices between neighbourhood types were computed (rows normalized to 1). Departure and return timing distributions were estimated, and disparity rates over time were computed for wealth (non-poor vs. poor) and race (White-majority vs. non-White) as D_i = (R − R_i)/R, where R is the overall fraction from a group and R_i is the cumulative fraction up to day i.

- Sample: 10,179 evacuees and 141,828 non-evacuees detected; overall evacuation rate 6.7%, consistent with official reports. - Spatial patterns: Higher evacuation rates near the coast and in inland pockets (e.g., Fort Bend, East Houston). Long-distance evacuations predominantly moved from coastal toward inland areas; short-distance evacuations occurred in dispersed pockets. - Disparities in who evacuated: After weighting, evacuees from non-poor majority White block groups were overrepresented by +19.8% relative to their population share. All other neighbourhood types evacuated below baseline; poor Hispanic communities were most underrepresented (−12.2%). Patterns held for long-distance evacuees; non-poor White overrepresentation increased by ~4.2%. For short-distance evacuations, only non-poor White neighbourhoods exceeded baseline. - Evacuation distance distribution: Universality across groups, following a truncated power law with exponent β = 1.57, scale parameter Δd0 = 2.19 km, and exponential cutoff K = 38.29 km. - Destination homophily: Strong socioeconomic and racial homophily in destinations. Residents from White neighbourhoods had an 88.1% probability of evacuating to White-majority block groups; Hispanic 56.8% to Hispanic-majority; Black 16.7% to Black-majority. Non-poor residents evacuated to non-poor areas with 92.9% probability; poor residents to poor areas with 35.4% probability. - Timing: 4.3% evacuated before landfall; 95.6% departed between 1 and 7 days after landfall. Returns began shortly after peak departures with a right-skewed, heavier-tailed distribution; most returned within two weeks, but some took much longer. Long-duration relocations (>30 days) were predominantly from affluent neighbourhoods. - Temporal disparities: Early evacuees disproportionately from wealthier neighbourhoods (wealth share >50% above baseline) and White-majority areas (~25% above baseline); disparities decreased steadily after landfall. Early returners (within 3 days) primarily from wealthy communities; class disparity in returns declined toward parity by ~2 weeks post-disaster and converged by ~3 weeks. - Fatality relation: K_net hotspots included three peaks corresponding to areas with high fatalities; other hotspots (southwest of downtown) had no fatalities, indicating varied drivers of evacuation intensity.

The study demonstrates that while evacuation distances display universal heavy-tailed patterns across socioeconomic and racial groups, substantial disparities exist in who evacuates, when they evacuate and return, and where they go. Wealthier and White-majority neighbourhoods exhibited higher evacuation rates, earlier departures, earlier returns, longer displacement tails (ability to remain away longer), and strong destination homophily toward similarly advantaged areas. These patterns likely reflect unequal access to transportation, financial resources, preparedness, and social networks. Despite the absence of a mandatory evacuation, individual decisions reproduced social inequalities in risk mitigation. The results underscore the need for targeted policy and resource allocation—e.g., pre-positioned shelters and services in likely destination areas, support for disadvantaged communities to enable earlier/safer evacuation, and communication strategies that address risk perceptions and barriers. The combination of high-resolution mobility data with neighbourhood sociodemographics provides a quantitative basis to understand and address inequities in disaster responses.

This work quantifies class- and race-based disparities in evacuation behaviours during Hurricane Harvey using high-resolution mobility data. Key contributions include: (i) evidence that evacuation distances follow a truncated power law uniformly across groups; (ii) robust disparities in evacuation likelihood, timing, and destination homophily disadvantaging poor and minority neighbourhoods; and (iii) temporal dynamics showing early evacuation and return dominated by wealthier residents. These findings inform emergency planning and equity-focused interventions, such as financial assistance for evacuation, targeted messaging, and resource prioritization for disadvantaged areas and likely destinations. Future research should expand spatial and temporal scope to capture long-distance and long-term relocations, integrate richer data on housing damage, infrastructure disruptions, and social networks, and develop predictive, equity-aware models to guide policy responses in real time.

- Spatial-temporal scope: Analysis is limited to the Houston MSA over a three-month window (July–October 2017), constraining observation of out-of-area or long-term relocations. - Sampling bias: Differential smartphone penetration and app opt-in across communities introduce representativeness concerns; weighting and bootstrapping mitigate but cannot eliminate bias. - Omitted variables: Limited control for factors such as home damage, infrastructure status, and social ties. A regression including flooded roads showed limited predictive power, but broader data on physical and social infrastructure are needed to disentangle mechanisms and fully explain disparities.