Source-linked AI summary
On the use of human mobility proxy for the modeling of epidemics
Michele Tizzoni, Paolo Bajardi, Adeline Decuyper, Guillaume Kon Kam King, Christian M. Schneider, Vincent Blondel, Zbigniew Smoreda, Marta C. González, Vittoria Colizza
TL;DR
Human commuting in epidemic models for rapidly disseminated infections is still poorly understood, and mobility data may be inaccessible at the desired resolution or coverage. This paper assesses mobile-phone and radiation-model proxies for spatial ILI epidemic modeling, finding that phone data match census patterns but overestimate commuters, producing faster diffusion while preserving infection order.
Problem
Human commuting in epidemic models for rapidly disseminated infections is still poorly understood, while mobility data may lack the desired resolution or sufficient coverage.
Method
The paper assesses mobile phone data and the radiation model as proxies for commuter movement data in spatial influenza-like-illness epidemic modeling.
Results
Phone data well match census commuting patterns but overestimate commuters, leading to faster simulated epidemic diffusion while preserving the infection order of newly infected locations.
Takeaways & Limitations
Policy and planning may be more accurate when they use reliable descriptions of population movements.
Takeaways & Limitations
The exposed biases may be affected differently by cultural and socio-economic factors.
Abstract
from arXiv · showhide
Human mobility is a key component of large-scale spatial-transmission models of infectious diseases. Correctly modeling and quantifying human mobility is critical for improving epidemic control policies, but may be hindered by incomplete data in some regions of the world. Here we explore the opportunity of using proxy data or models for individual mobility to describe commuting movements and predict the diffusion of infectious disease. We consider three European countries and the corresponding commuting networks at different resolution scales obtained from official census surveys, from proxy data for human mobility extracted from mobile phone call records, and from the radiation model calibrated with census data. Metapopulation models defined on the three countries and integrating the different mobility layers are compared in terms of epidemic observables. We show that commuting networks from mobile phone data well capture the empirical commuting patterns, accounting for more than 87% of the total fluxes. The distributions of commuting fluxes per link from both sources of data - mobile phones and census - are similar and highly correlated, however a systematic overestimation of commuting traffic in the mobile phone data is observed. This leads to epidemics that spread faster than on census commuting networks, however preserving the order of infection of newly infected locations. Match in the epidemic invasion pattern is sensitive to initial conditions: the radiation model shows higher accuracy with respect to mobile phone data when the seed is central in the network, while the mobile phone proxy performs better for epidemics seeded in peripheral locations. Results suggest that different proxies can be used to approximate commuting patterns across different resolution scales in spatial epidemic simulations, in light of the desired accuracy in the epidemic outcome under study.
Author summary
The paper evaluates mobile-phone and radiation-model proxies for commuting in spatial epidemic simulations, comparing them with census networks across three European countries. Mobile-phone data reproduce commuting structure well but overestimate traffic, while epidemic invasion accuracy depends on seed location.
- Scope and implications: Mobility proxies can approximate commuting patterns across resolution scales, but their suitability depends on the epidemic outcome and data context.Phone-data representativeness, privacy constraints, and regional data availability limit interpretation, especially outside developed settings.
- Mobility-proxy findings: More than 87% of total commuting fluxes are captured by mobile-phone networks, which closely match census commuting patterns.The phone-derived and census flux distributions are similar and highly correlated.
- Epidemic consequences: Mobile-phone networks systematically overestimate commuting traffic, producing faster epidemic diffusion than census networks.The faster spread occurs despite preserving the infection order of newly infected locations.
- Epidemic consequences: The order of infection among newly infected locations remains well preserved when mobile-phone mobility replaces census commuting.Thus, the proxy changes epidemic speed more than the sequence of invaded locations.
- Seed-location dependence: The radiation model is more accurate for centrally seeded epidemics, whereas the mobile-phone proxy performs better for peripheral seeds.The preferred proxy therefore depends on the epidemic’s initial location.
- Study aim and design: Mobile-phone records and the radiation model are assessed as alternatives to census commuting data for influenza-like-illness epidemic modeling.The study compares proxy networks across three European countries and multiple geographic scales.
Materials and Methods
The study compares census-based and mobile-phone-derived commuting networks across multiple geographic scales, using anonymized call records and harmonized administrative units. Mobile-phone data are normalized to census populations, while acknowledging sampling and coverage limitations.
- Data sources: The study uses anonymized mobile-provider billing records collected for legal and billing purposes, with phone numbers hashed and research reviewed by an institutional review board.Authors and the principal investigator also completed ethics training.
- Census networks: Census commuting networks provide the benchmark and are built from official surveys describing commuter movements at each country’s available geographic resolution.The networks represent the most accurate and reliable available description of commuter movements for the countries studied.
- Geographic harmonization: Census networks are coarse-grained to shared administrative levels, and directed weighted links represent commuter flows between origin and destination locations.The procedure supports comparisons between datasets collected at different original resolutions.
- Scope and limitations: The analysis excludes overseas or island territories and cross-border commuting, while mobile-phone observations remain affected by operator coverage, sample selection, and commuting-identification algorithms.More sophisticated bias corrections would require additional metadata that may not be readily available across many countries.
- Mobile-phone networks: Mobile-phone commuting networks identify residence and workplace as users’ most- and second-most-visited call locations, then weight links by the number of commuting users.Only users with more than 100 calls are included, and a sensitivity analysis adds temporal constraints to location identification.
- Normalization: Mobile-phone networks are coarse-grained to national administrative subdivisions and normalized so each node’s total population matches the census, although commuter shares may differ.The baseline normalization rescales mobile-phone populations using sampling ratios; refined normalization additionally assumes matching commuter totals.
Results
Mobile-phone commuting networks reproduce the main structure of census mobility across countries and scales, while their larger estimated flows accelerate simulated epidemics. Epidemic agreement depends on seed location, with mobile-phone proxies particularly reliable for peripheral starts.
- Network agreement: More than 87% of total commuting traffic is captured by links common to census and mobile-phone networks.
- Network agreement: Commuting-flux distributions have similar broad-tailed shapes and high correlations across the two data sources, though mobile-phone flows are generally larger.Weak flows are missed at finer scales in Portugal and France, while this discrepancy disappears at larger spatial scales such as Spain.
- Epidemic outcomes: Central seeding generally yields similar invasion paths across proxy and census models, whereas peripheral seeding favors mobile-phone data over the radiation model.The radiation model can be more accurate for central seeds, while mobile-phone proxies perform better or similarly for peripheral locations.
- Epidemic outcomes: Refined normalization improves invasion-path and arrival-time agreement, but mobile-phone simulations still anticipate spread relative to census simulations.Arrival-time rankings also improve under the refined normalization, although anticipation effects remain reduced rather than eliminated.
- Network agreement: Mobile-phone data capture commuting-flow fluctuations and identify the relative importance of mobility connections at census-equivalent or finer resolution.
- Epidemic outcomes: Mobile-phone commuting networks produce faster epidemic spread than census networks because their commuting flows are larger, while preserving the infection order of locations.The larger flows also synchronize epidemic peaks and shorten the overall time span for subpopulations to peak.
Figure Legends
The figures compare census, mobile-phone, and radiation-model mobility networks across countries and resolution scales, then assess their epidemic behavior and invasion patterns.
- Figure 1: Figure 1 maps regional differences between mobile-phone and census dataset coverage relative to national averages.Values near unity indicate coverage similar to the national average; red and blue indicate over- and undersampling, respectively.
- Figure 2: Figure 2 compares census and mobile-phone commuting-link weights, including their distributions and relationship across connections.The comparison includes scatter plots, box plots, and probability-density distributions.
- Figure 3: Figure 3 relates commuting-flux ratios to node distance, origin population, and destination population across Portugal, Spain, and France.The solid red line marks a ratio of one.
- Figure 4: Figure 4 compares epidemic spreading on census, mobile-phone, and radiation-model networks using invasion-tree similarity and arrival-time differences.Results are shown for multiple basic reproduction numbers and infection seeds.
- Figure 5: Figure 5 repeats the epidemic comparison after refined normalization, showing that mobile-phone anticipation effects remain but are reduced.The France example includes invasion-tree similarity, arrival-time differences, and arrival-time scatter plots for a capital-city seed.
- Figure 6: Figure 6 displays full epidemic invasion trees for census, mobile-phone, and radiation networks in Portugal and France.Seeds are Lisbon and Barcelonnette; fully colored nodes form the first infection shell, while grey nodes were infected secondarily.
Supporting Information Legends
The supporting legends document supplementary analyses, network statistics, commuting-flow summaries, and statistical comparisons between census and mobile-phone networks.
- Text S1: Text S1 lists supplementary data-source information, lower-resolution results, epidemic peak times, sensitivity analyses, normalization details, and simulation-algorithm details.The listed analyses include cross-border commuting and refined workplace and residence definitions.
- Table 1: Table 1 reports basic commuting-network properties, including nodes, links, and incoming and outgoing commuter fluxes.Rows represent countries and geographic subdivisions; columns distinguish census and mobile-phone datasets and normalization stages.
- Table 1: Refined-normalization values are equal to census values by definition.They are therefore not reported in the table.
- Table 2: Table 2 reports Lin’s concordance and Spearman’s coefficients comparing mobile-phone and census network weights and node totals.The coefficients are measured after log transformation where specified.