Source-linked AI summary
Collective response of human populations to large-scale emergencies
James P. Bagrow, Dashun Wang, Albert-László Barabási
TL;DR
Research on human dynamics has largely examined normal, stationary conditions rather than cooperative actions under external perturbations. This paper measures event-related changes in communication activity and finds localized spikes, rapid decay, and social propagation through eyewitness networks, supporting automatic emergency distinction.
Problem
Current research on human dynamics is limited to data collected under normal and stationary conditions, motivating study of cooperative human actions under externally induced perturbations.
Method
The study measures event-related call activity relative to average normal activity and tracks when anomalous activity begins, ends, and reaches half its total magnitude.
Results
Emergencies trigger sharp, spatially and temporally localized spikes in calls and text messages, followed by decay and strong social propagation through the second- and third-degree neighbors of eyewitnesses.
Takeaways & Limitations
Documented changes in human activity can be used to automatically distinguish emergencies, while mobile phones act as sensitive local sociometers of societal perturbations.
Takeaways & Limitations
The localized anomaly is small compared with other proxy studies, and emergencies occur over very short timespans of a few hours.
Abstract
from arXiv · showhide
Despite recent advances in uncovering the quantitative features of stationary human activity patterns, many applications, from pandemic prediction to emergency response, require an understanding of how these patterns change when the population encounters unfamiliar conditions. To explore societal response to external perturbations we identified real-time changes in communication and mobility patterns in the vicinity of eight emergencies, such as bomb attacks and earthquakes, comparing these with eight non-emergencies, like concerts and sporting events. We find that communication spikes accompanying emergencies are both spatially and temporally localized, but information about emergencies spreads globally, resulting in communication avalanches that engage in a significant manner the social network of eyewitnesses. These results offer a quantitative view of behavioral changes in human activity under extreme conditions, with potential long-term impact on emergency detection and response.
Introduction
Existing human-dynamics research largely describes normal, stationary behavior, leaving rapid responses to unfamiliar emergencies insufficiently understood. This study uses mobile-phone communication and mobility data to examine societal perturbations across emergencies and planned events.
- Motivation: Normal, stationary activity models may break down when populations face rapidly changing conditions such as epidemics, emergencies, and traffic anomalies.These conditions include natural, technological, and societal disasters.
- Approach: Mobile phones provide real-time measurements of user mobility and communication along social-network links, enabling observation of behavioral changes during emergencies.The study treats phones as in situ sensors and potential sociometers of external perturbations.
- Data: The dataset contains anonymized billing records for approximately ten million subscribers, covering about one-fourth of a nearly fully connected national mobile-phone market.Records include call timing, duration, and the cellular tower handling each call, allowing approximate user locations.
- Scope: Approximately 30% of media-reported events occurred under sparse cellular coverage or when few users were active, while the remaining events formed a sufficiently diverse corpus.The authors therefore distinguish generic and unique activity changes using the remaining events.
- Study design: Eight planned events, including sports events, a cultural festival, and concerts, provide comparisons for collective activity changes that are not emergencies.The comparison is intended to distinguish generic from emergency-specific changes in human activity.
Results and Discussion
Emergencies produce abrupt, spatially localized communication anomalies, while their information cascades can rapidly extend through eyewitnesses’ social networks. Compared with non-emergencies, these reproducible behavioral signatures distinguish event types and support real-time monitoring.
- Temporal response: Emergencies trigger an immediate jump-decay call anomaly, whereas non-emergency anomalies build gradually over seven or more hours.Emergency anomalies typically last three to five hours.
- Temporal response: The primary anomaly source is a sudden increase in calls by people who would normally be inactive during the event period.The response therefore reflects recruitment of otherwise inactive users rather than simply greater activity from regular callers.
- Spatial response: Call anomalies are strongest near emergencies and decay approximately exponentially with distance, with decay rates ranging from 2 km for bombing to 10 km for plane crash.The earthquake is the exception, with an extended spatial range of approximately 110 km, while non-emergencies have decay rates below 2 km.
- Social propagation: Information propagation is strongest for severe emergencies: bombing and plane-crash responses reach third and second social neighbors, respectively, while weaker emergencies propagate little beyond immediate ties.Eyewitnesses and their social contacts show anomalous calling within minutes, but propagation is virtually absent for some less threatening events.
- Social propagation: In the composite evacuation-and-explosion event, the explosion expanded the spatial response from rc = 1.6 km to rc = 9.0 km and activated G2 users after evacuation propagation stopped at G1.The evacuation anomaly leveled off after the initial warning, whereas the explosion produced the characteristic larger jump-decay response.
- Detection and response: A multidimensional variable built from activity changes can automatically distinguish emergency from non-emergency anomalies and potentially support real-time monitoring.The proposed signals could estimate affected-population size, event timing, and users able to provide immediate actionable information.
Materials and Methods
The study identifies event timing and location from news reports, then detects anomalous local call activity against weekly baseline periods. It quantifies anomalies using standardized time-series thresholds.
- Researchers estimated each event region around its epicenter, using a radius based on spatial decay rate rc.
- News reports supplied event times and locations, typically using photographs and cross-checking conflicting details.
- Call activity during the event was compared with averaged activity from corresponding weekly periods to establish ⟨Vnormal⟩.
- The anomaly score was computed as z(t) = ∆V(t)/σ(Vnormal), after typically binning calls into 5–10 minute intervals.
- The anomalous interval was defined as the longest contiguous run with z(t) > zthr, using zthr = 1.5 for all events.
bombing
Emergency-related call anomalies are sharply localized in time and space, while their information cascades propagate through social networks beyond the eyewitness population. Emergencies differ systematically from non-emergencies in onset, extent, and cascade characteristics.
- Spatial impact: Bombing call anomalies peak later in the 1 < r < 5 km region than within r < 1 km, with no anomaly observed beyond 10 km.
- Social characteristics of information cascades: Bombing and plane-crash populations respond within minutes, and anomalous activity extends across successive social groups Gi.
- Composite event: An evacuation preceding an explosion has rc = 1.6 km during evacuation and rc = 9.0 km after the explosion.
- Composite event: The evacuation activates G0 and G1, whereas the explosion produces a substantially broader social response.
- Systematic response mechanisms during emergencies: Emergencies produce more abrupt anomalous call activity than non-emergencies, which show gradual buildups.
A Dataset
The dataset combines anonymized mobile billing records with tower locations to track communication activity and approximate caller locations. Researchers follow affected users before and after events to establish baseline behavior.
- Billing records cover approximately 10M subscribers in one country over 3 years.
- Each record identifies caller, callee, cellular tower, and call date and time for voice and text services.
- Tower latitude and longitude provide an approximate caller location when a call is placed.
- Affected users’ mobile activity was followed in weeks before and after events to provide control or baseline behavior.
- Self-reported gender information is available for approximately 90% of subscribers.
- The records represent approximately 20% of the country’s mobile phone market, with limited information on outside-provider lines.
B Identifying events
Events were identified from contemporaneous news coverage, then localized and detected in mobile records by comparing event-period call activity with weekly normal periods. The method also accounts for possible network disruption.
- Online news aggregators were searched with terms including storm, emergency, and concert to identify candidate events.
- Event regions were estimated as circles centered on identified epicenters, with radii chosen from rc; the blackout used city towers instead.
- Researchers averaged calls from periods one week apart to exploit weekly periodicity and define normal activity.
- Anomalous activity was detected using the longest contiguous run where z(t) exceeded the fixed threshold zthr = 1.5.
- Only the blackout damaged the mobile system temporarily, while other towers maintained call routing; larger emergencies might cause such effects.
B.1 Missing events
Mobile-phone data captured many emergencies locally, but some newsworthy events were absent because detectability depends on user activity, cellular coverage, timing, and event concentration.
- Forest fires, a wind storm, and a gas-main explosion were reported but not identified in the mobile-phone data.Two additional small affected populations were discounted from analysis.
- The bombing spike disappeared at the national scale but was clear in the immediately local vicinity.
- Late-night events may be difficult to study because few people are awake and active on their phones.
- Remote events with little cellular coverage are harder to detect than events in well-covered, densely populated regions.
- Severe but diffuse events may blend into normal activity, although the earthquake was an exception.
- National events were avoided because broad participation makes it difficult to distinguish separate event populations.
C Source of call anomaly
Emergency call spikes primarily reflect more people becoming active rather than only regular callers increasing their call frequency, alongside stronger calling to existing friends.
- 232% more individuals called during the bombing, compared with a 36% increase in phone usage per user.Other emergencies showed increases in ρ (N) of 21% (67.5%) for the plane crash, 1.36% (17.4%) for the earthquake, and 4.97% (20.8%) for the blackout.
- The primary source of emergency call anomalies was a sudden increase in calls from individuals who normally would not use phones during that period.
- In every emergency, users made more calls to friends and correspondingly fewer calls to non-friends.Many non-emergencies showed the opposite trend, but the change was seldom large.
E Gender response during events
Emergency events altered demographic and communication patterns, with increased female participation in many emergencies and event-specific shifts between voice and text use; broader response measures separated emergencies from non-emergencies.
- Gender response: Nearly all emergencies increased the fraction of affected female users, significantly so for half of the emergencies.Non-emergencies did not significantly deviate in gender breakdown.
- Gender response: The gender response was observed for both directly affected users and their one-step neighbors.
- Voice versus text: Most events showed no significant change in the voice-call fraction, while the earthquake and blackout increased text usage and the plane crash increased voice usage.Festival 3 also differed for users one step away.
- Voice versus text: A primarily text-messaging spike may indicate a low-threat or non-critical emergency, based on the earthquake and blackout.
- Response measures: Temporal, spatial, and social measures included midpoint fraction, anomaly radius, cascade size, and friend-calling significance.
- Response measures: These measures distinctly separated emergencies from non-emergencies, suggesting common response patterns across differing events.
H.2 Controlling for social propagation
The analysis controls for selection, population size, time-of-day activity, and seed-population differences when testing whether emergency contact cascades are anomalously large. After these controls, emergencies show enlarged cascades and contact-population activity, whereas non-emergencies do not.
- Contact cascades can appear during normal periods because users’ call timing is temporally correlated, so event activity requires controlled comparisons.The analysis asks whether users reached through successive call groups show increased activity during an emergency.
- Selection bias arises because users selected for calling during a window must already be active in that window, and activity levels also depend on group size and window length.The controls compare equal-duration windows and rescale conditional activity for different population sizes.
- Control 3 compares event cascades with cascades normally induced by the same initiating users, preserving the initiating population and accounting for selection bias.Normal-period activity is rescaled to match event-period population activity over a calibration interval.
- The method additionally controls for seed-population size by generating normal cascades from users active in each comparison period, then matching event cascades to those seeds.This prevents a larger event-period seed population from inflating relative cascade size R.
- Global Activity Level matching equalizes the number of possible communications across events before computing relative cascade size R.Other controls address time-of-day differences, since non-emergencies tend to occur later when users are generally more active.
- After all controls, many emergencies—especially the bombing—have anomalously large cascades and increased contact-population activity, while non-emergencies remain ordinary.With the reported control, all non-emergencies have R < 1 and all emergencies have R ≳1; Storm 3 has R = 1.0044.
I Results on the event corpus
Across the full corpus, the study reports regional call-activity patterns and social propagation for sixteen events, including eight emergencies and eight non-emergencies. Emergency and non-emergency events differ in the timing, spatial extent, and propagation of anomalous activity, with notable event-specific variation.
- Sixteen events were analyzed: eight emergencies and eight non-emergencies, with call activity reported for all events and propagation patterns shown across event-user groups.The supplementary figures cover regional activity, spatial changes, and conditional activity for G0 through G3.
- Bombing and plane-crash events show increased activity across multiple social-propagation groups, while the earthquake and blackout do not.Figure L compares event-period conditional time series with rescaled normal activity for G0 through G3.
- Concerts generally show extra activity only for G0, except Concert 4, which has a small delayed increase for G1.The increase occurs several hours after the concert started.
- Festival 2 shows no extra activity even for G0, indicating its call anomaly came from more users making expected numbers of calls.