Source-linked AI summary
Multiscale Dynamic Human Mobility Flow Dataset in the U.S. during the COVID-19 Epidemic
Yuhao Kang, Song Gao, Yunlei Liang, Mingxiao Li, Jinmeng Rao, Jake Kruse
TL;DR
The paper addresses limited availability of fine-resolution origin-to-destination mobility flows for evaluating COVID-19 interventions and modeling spatial interactions. It constructs regularly updated U.S. daily and weekly population-flow data from SafeGraph trajectories and ACS population data across three geographic scales. The dataset shows high correlations with an open mobility source, while its estimates remain subject to mobile-phone sampling and aggregation limitations.
Problem
Existing COVID-19 mobility datasets often lack origin-to-destination flow matrices, fine spatial resolution, and entire-population-level estimated flows.
Method
The paper computes daily CBG-to-CBG and weekly CBG-to-POI visitor flows from SafeGraph data, aggregates them to three scales, and infers population flows using ACS population and device counts.
Results
At least 0.92 correlation with Descartes Lab mobility changes was observed in all five metropolitan areas.
Takeaways & Limitations
The up-to-date multiscale O-D dataset can support monitoring epidemic dynamics, public-health policy analysis, and human-mobility and transportation research.
Takeaways & Limitations
Mobile-phone usage and demographic differences may bias the representativeness of estimated entire-population mobility flows.
Abstract
from arXiv · showhide
Understanding dynamic human mobility changes and spatial interaction patterns at different geographic scales is crucial for assessing the impacts of non-pharmaceutical interventions (such as stay-at-home orders) during the COVID-19 pandemic. In this data descriptor, we introduce a regularly-updated multiscale dynamic human mobility flow dataset across the United States, with data starting from March 1st, 2020. By analyzing millions of anonymous mobile phone users' visits to various places provided by SafeGraph, the daily and weekly dynamic origin-to-destination (O-D) population flows are computed, aggregated, and inferred at three geographic scales: census tract, county, and state. There is high correlation between our mobility flow dataset and openly available data sources, which shows the reliability of the produced data. Such a high spatiotemporal resolution human mobility flow dataset at different geographic scales over time may help monitor epidemic spreading dynamics, inform public health policy, and deepen our understanding of human behavior changes under the unprecedented public health crisis. This up-to-date O-D flow open data can support many other social sensing and transportation applications.
Background & Summary
Existing COVID-19 mobility datasets often lack origin-to-destination flow matrices and fine spatial resolution. The paper introduces an openly available U.S. dataset addressing these gaps with multiscale, daily and weekly population-flow estimates from anonymous mobile-phone trajectories.
- COVID-19 interventions reduced mobility, making dynamic human-mobility and spatial-interaction data important for evaluating interventions and modeling virus spread.
- Existing open datasets commonly provide aggregated mobility indices rather than origin-to-destination flow matrices needed for epidemic modeling and spatial-interaction measurement.
- Most available datasets use state, county, or city scales, leaving a need for finer-resolution data such as census-tract flows to characterize heterogeneous mobility within cities and intracounties.
- The dataset estimates dynamic U.S. population flows at census-tract, county, and state scales with daily and weekly temporal resolutions.
- Millions of anonymous SafeGraph mobile-phone trajectories generate the origin-to-destination dataset, whose reliability is assessed against ACS commuting flows and Descartes Lab mobility data.
Methods
The dataset is constructed from SafeGraph mobile-phone observations and ACS demographic data, then aggregated across geographic scales and expanded from sampled devices to population-flow estimates. Daily and weekly visitor metrics provide the underlying origin-to-destination observations.
- Track Place Visits: SafeGraph GPS pings are cleaned, users’ home CBGs are estimated from nighttime locations over six weeks, and visits to POIs and CBGs are identified.
- Visitor-Flow Metrics: Daily CBG-to-CBG visitor flows and weekly CBG-to-POI visitor flows are computed from unique users’ origins and destinations.
- Multiscale Aggregation: The two visitor-flow metrics are aggregated from CBGs to census tracts, counties, and states to support analyses at different geographic scales.
- Infer Dynamic Population Flows: The observed mobile-phone users represent about 10% of the U.S. population, and sampling ratios vary across CBGs.
- Infer Dynamic Population Flows: Estimated population flow equals mobile-phone visitor flow multiplied by the origin population divided by the origin’s resident device count.
Data Records
The repository provides daily and weekly flow products beginning March 1, 2020, organized by geographic scale and distributed as CSV files. Records identify origin and destination geographic units and their spatial attributes.
- Products: Daily and weekly flow data begin on March 1, 2020, and are available at census-tract, county, and state scales.
- Repository Organization: Files are separated into daily_flows and weekly_flows folders and organized by ct2ct, county2county, and state2state geographic scales.
- File Format: The data are distributed in comma-separated values format with filenames encoding data type, spatial scale, and date.
- Record Attributes: Each record includes origin and destination geographic-unit identifiers and the latitude and longitude of their geometric centroids.
Technical Validation
The dataset is validated through distribution checks, model comparisons, and cross-referencing with established mobility sources. These analyses show consistent flow distributions and high correlations, while the absence of ground-truth data limits definitive accuracy assessment.
- Validation Strategy: Three complementary validation methods compare visitor and inferred population-flow distributions, gravity and radiation model estimates, and other open mobility datasets.The comparisons assess distribution preservation, model agreement, and consistency with external O-D and temporal mobility patterns.
- Distribution Checks: Q–Q plots test whether visitor-flow and inferred population-flow distributions remain linearly related across census tract, county, and state scales.The example uses weekly data from March 2–8, 2020, with similar associations reported for other dates.
- Model Comparisons: The radiation model’s population-flow estimates correlate with dataset ‘pop_flows’ at about 0.75, outperforming the gravity model estimates.Gravity-model fitting uses weekly county data and finds distance-decay coefficients β ranging from 0.85 to 1.00; β increased slightly after statewide lockdowns.
- Validation Scope: Because no ground-truth data characterize real dynamic population flows between geographic regions, external comparisons provide cross-referencing rather than definitive accuracy measurement.The authors describe these comparisons as useful for evaluating credibility while acknowledging that relative accuracy cannot be determined.
- Comparison with Other Data Sources: Both weekly and daily inferred flows show greater than 0.93 Pearson’s correlation with ACS commuting-flow patterns after unmatched O-D records are removed.The inferred population flows correlate more strongly with ACS commuting flows than directly computed mobile-phone visitor flows.
- Comparison with Other Data Sources: The produced datasets have correlation coefficients of at least 0.92 with Descartes Lab’s mobility-change data across all five metropolitan areas.This comparison evaluates whether the datasets capture similar temporal mobility patterns.
Usage Notes
The dataset offers daily and weekly O-D flows at three geographic scales, revealing distinct temporal and spatial mobility patterns during stay-at-home orders and reopenings. Its usage requires attention to measurement uncertainty, privacy-related thresholds, counting differences, aggregation duplication, and mobile-phone sampling bias.
- Dataset products: Daily and weekly O-D flows are provided at census tract, county, and state scales, with complementary temporal detail and stability.Daily flows fluctuate more in response to individual events, whereas weekly flows show smoother, more general mobility patterns.
- Observed usage patterns: 22 million weekly active devices initially declined to 13 million during April 13–19, while daily active users reached a minimum of 16 million on April 18.Active-user counts later increased as compliance decreased and stay-at-home orders were lifted.
- Observed usage patterns: Mobility flows decreased significantly from March to April across all three spatial scales and increased in May as partial reopenings began.During stay-at-home orders, long-range interactions declined while short-range movements to adjacent counties predominated; long-range interactions later rebounded at state and county scales.
- Limitations: Weekly flow counts have a minimum of four visitors from a CBG to a POI, and visitors staying more than one minute are counted as visits.Differential-privacy rules suppress or alter small counts, while the short visit threshold may affect foot-traffic estimates.
- Limitations: Daily and weekly flows use different unique-visitor definitions, so seven daily flow totals do not equal the corresponding weekly flow.Weekly POI flows may underestimate mobility when visitors repeatedly visit the same POI, because devices rather than raw visits are counted.
- Limitations: Aggregating lower-level flows can inflate inferred population flows because individual travel behavior cannot be traced across visits.A visitor visiting two POIs in different CBGs within one census tract may be counted twice instead of once at the tract scale.
- Limitations: Mobile-phone usage bias may reduce representativeness because elderly people and children may be less likely to use smartphones or applications.Age-group and demographic-composition differences can influence estimated entire-population mobility flows, although the dataset still provides up-to-date information at scale.
- Potential uses: The dataset can help deepen understanding of human dynamics, inform public-health policy, and support social-sensing and transportation applications.These uses are stated within the context of the COVID-19 public-health crisis.
Code availability
The analysis used Python 3.7 on a Linux server, and all analysis code is publicly available in the dataset’s GitHub repository.
- Python 3.7 was used for data processing and analysis on a Linux server.
- All analysis code is available in the public GitHub repository that hosts the data.
- The repository is identified as github.com/GeoDS/COVID19USFlows.
Figures
The figures document the dataset’s production workflow, geographic coverage, multiscale flow validation, and temporal and spatial mobility patterns during the pandemic.
- The production framework tracks place visits, computes visitor flows, aggregates them across scales, and infers population flows.
- SafeGraph place coverage is visualized across the United States as a spatial density distribution.
- Quantile-quantile plots compare visitor-flow and population-flow distributions at census tract, county, and state scales.
- Temporal mobility patterns are shown for five metropolitan areas using daily and weekly visitor-flow and population-flow panels from March 2 to May 31, 2020.
- Active-user counts and Nc/N values are plotted daily and weekly from March 2 to May 31, 2020, with Nc denoting users making at least one trip.
- Weekly spatial flows are compared before, during, and after stay-at-home orders at state, county, and census-tract scales.
Tables
The tables provide sample flow records and validation summaries comparing model estimates and the dataset with other mobility or commuting-flow sources.
- Table 1 gives sample records for daily CBG-to-CBG visitors and weekly CBG-to-POI visitors.
- Table 2 reports gravity- and radiation-model parameter settings alongside correlations between model outputs and population-flow estimates.
- Table 3 lists Pearson correlation coefficients between the mobility-flow dataset and ACS commuting flows at county scale.
- Table 4 lists Pearson correlation coefficients comparing temporal patterns in the daily flow dataset with the Descartes Labs dataset.