Source-linked AI summary
Social Network Sensors for Early Detection of Contagious Outbreaks
Nicholas A. Christakis, James H. Fowler
TL;DR
Existing outbreak-detection methods offer little advance warning, so the paper tests monitoring friends of randomly selected people as network sensors and finds flu detection about two weeks early.
Problem
Current contagious-outbreak detection methods ideally provide contemporaneous information rather than advance warning.
Method
The study evaluates surveying friends of randomly selected individuals as monitored sensors without ascertaining global network structure.
Results
Two weeks in advance, the evaluated sensor network detected a flu outbreak rather than merely providing rapid warning.
Takeaways & Limitations
The method appears to provide early detection and longer lead times than other extant flu-detection methods.
Takeaways & Limitations
How much advance detection the method achieves for other pathogens or populations remains uncertain.
Abstract
from arXiv · showhide
Current methods for the detection of contagious outbreaks give contemporaneous information about the course of an epidemic at best. Individuals at the center of a social network are likely to be infected sooner, on average, than those at the periphery. However, mapping a whole network to identify central individuals whom to monitor is typically very difficult. We propose an alternative strategy that does not require ascertainment of global network structure, namely, monitoring the friends of randomly selected individuals. Such individuals are known to be more central. To evaluate whether such a friend group could indeed provide early detection, we studied a flu outbreak at Harvard College in late 2009. We followed 744 students divided between a random group and a friend group. Based on clinical diagnoses, the progression of the epidemic in the friend group occurred 14.7 days (95% C.I. 11.7-17.6) in advance of the randomly chosen group (i.e., the population as a whole). The friend group also showed a significant lead time (p<0.05) on day 16 of the epidemic, a full 46 days before the peak in daily incidence in the population as a whole. This sensor method could provide significant additional time to react to epidemics in small or large populations under surveillance. Moreover, the method could in principle be generalized to other biological, psychological, informational, or behavioral contagions that spread in networks.
Introduction
Existing outbreak detection often lags epidemic progression, while centrally located individuals are expected to become infected earlier. The paper proposes monitoring friends of randomly selected people as network sensors without mapping global network structure and evaluates this strategy using influenza surveillance at Harvard College.
- Current outbreak-detection methods typically provide contemporaneous or delayed information about epidemic progression.
- Central individuals are expected to become infected sooner because contagions reach them in fewer network steps than peripheral individuals.This temporal shift could enable outbreak detection before the epidemic reaches the population as a whole.
- Mapping an entire network to identify informative individuals is costly, time-consuming, and often impossible, especially in large networks.
- The proposed alternative surveys friends of randomly selected individuals, who tend to have higher degree and greater network centrality than those who named them.The strategy therefore predicts earlier infection among nominated friends than among randomly chosen individuals representing the population as a whole.
- 744 Harvard undergraduates were monitored for influenza from September 1 through December 31, 2009, using random and friends samples.The random sample comprised N=319 students, while the friends sample comprised N=425 students named by random-sample members.
- Influenza cases were assessed through formally diagnosed UHS records and twice-weekly self-reported symptoms, with self-report described as the current diagnostic standard.UHS diagnoses reflect more severe symptomatology than self-reported flu, while 90% of subjects missed no more than two surveys.
Results
Monitoring nominated friends produced earlier flu detection than monitoring randomly selected students, with clinically diagnosed flu curves shifted forward by 14.7 days. Network measures further showed that friend nominations identify individuals with higher in-degree and centrality and lower transitivity, each associated with earlier contagion.
- Results: 8% versus 32% cumulative flu incidence was observed by December 31, 2009, based on medical-staff diagnoses versus self-reports, respectively.Self-reported prevalence was higher than prevalence based on clinical diagnoses.
- Results: 14.7 days (95% C.I. 11.7–17.6) was the forward shift of the friend-group curve for medically diagnosed flu relative to the random group.The shift represented approximately 65% of one standard deviation in time to event and was robust to vaccination, sex, class, and varsity-sports controls.
- Results: Day 16 was when the friend group first showed significant lead time (p<0.05) for medically diagnosed flu, 46 days before the estimated incidence peak.For self-reported symptoms, significant lead time appeared by day 39, 83 days before the estimated symptom-incidence peak.
- Results: Self-reported popularity produced no significant forward shift, while controlling for it did not alter the friend group’s significant lead time.The results suggest that friend nomination captures more network information, including network centrality, than self-reported network attributes.
Discussion
The social network sensor method could provide real-time early warning of contagious outbreaks by monitoring nominated friends, with flu detection two weeks in advance rather than merely contemporaneous information. Its advance detection may extend to other network-spreading phenomena, but the achievable lead time depends on the pathogen, population, sensor-group size, and network structure.
- Implementation: Monitoring nominated friends could provide early warning for targeted populations of any size in real time.A spike in cases among monitored friends could signal an impending outbreak.
- Detection strategies: The timing difference between friend and random groups could signal an epidemic either through a friends-group threshold or divergence between incidence curves.Tracking both groups also distinguishes secular population trends from network effects above the epidemic baseline.
- Comparison with existing surveillance: For flu, the sensor method could detect an outbreak two weeks in advance, providing early detection rather than only rapid warning.Existing flu surveillance is typically one to two weeks behind the actual course, while Google Trends could provide evidence at least a week before published CDC reports but at best contemporaneous infection rates.
- Combined surveillance: Combining sensor monitoring with online search could provide high-quality, real-time epidemic information with even greater lead time.The approach could follow online behavior in a friend group or another group known to be central in a network.
- Limitations: How much advance detection the method achieves for other pathogens or larger or differently composed populations remains unknown.Performance depends on the spreading entity, measurement method, population susceptibility or prevalence, sensor-group size, network topology, and whether the outbreak changes network structure.
- Generalizability: The strategy could generalize beyond flu and college students to biological, psychological, normative, informational, and behavioral phenomena that spread in networks.Examples include antibiotic-resistant germs, depression, altruism, rumors, and smoking, with outbreaks detectable before reaching a critical threshold.
Materials and Methods
The study measured subjects’ perceived popularity, network connectivity, centrality, and transitivity using questionnaire and friendship-nomination data. Flu incidence and early-detection differences were estimated with nonparametric, logistic-regression, and bootstrap procedures, while network visualizations used Pajek and the Kamada-Kawai algorithm.
- Self-perceived popularity was measured with 8 questions adapted from a prior coworker-popularity instrument.
- Friendship nominations measured each subject’s in-degree and out-degree, with out-degree capped at 3 by the elicitation procedure.In-degree counts incoming friendship nominations; out-degree counts people each subject names as friends.
- Network structure was characterized using betweenness centrality and transitivity, defined respectively through contagion paths and the probability that two friends are also friends.Transitivity equals observed friendship triangles divided by the total possible triangles.
- Pajek and the Kamada-Kawai algorithm were used to produce two-dimensional network visualizations from shortest-path distances.The study also provided a 122-day movie depicting flu spread.
- Cumulative flu incidence used NPMLE, predicted daily incidence used nonlinear least-squares logistic fits, and uncertainty used bootstrapped standard errors and 95% confidence intervals.Early-detection days for network-attribute groups were calculated by multiplying model coefficients and confidence intervals by mean differences between above- and below-average groups.
Supplementary Information for Social Network Sensors for Early Detection of Contagious
The supplementary information identifies the paper’s authors and their institutional affiliations in Boston and La Jolla.
- Nicholas A. Christakis and James H. Fowler are listed as the paper’s authors.
- Christakis is affiliated with Harvard Medical School and Harvard Faculty of Arts and Sciences in Boston, Massachusetts.
- Fowler is affiliated with the Political Science Department at the University of California, San Diego, in La Jolla, California.
Subjects
The study enrolled 744 Harvard undergraduates in random and friend samples, embedded within a larger network of 1,789 students. Flu incidence was measured through self-reports and formal medical diagnoses, with self-reported cumulative incidence approximately four times higher by December 31, 2009.
- Study population: 744 Harvard undergraduates formed a random sample (N=319) and a friends sample (N=425), with friends nominated by members of the random group.The enrolled sample was embedded in a network of 1,789 uniquely identified students.
- Study population: Participants completed questionnaires covering demographics, flu and vaccination status, popularity, and registrar-derived academic and sports information.The administrative data included sex, class of enrolment, and varsity-sports participation.
- Outcome measurement: 32% versus 8%: cumulative incidence was approximately four times higher under self-report than under formal medical diagnosis by December 31, 2009.Formal diagnoses generally reflected more severe symptoms, whereas self-reports also captured cases not receiving medical attention.
Network Measures
The study characterizes subjects’ network positions using friendship-based degree, transitivity, and betweenness centrality measures. Transitivity and betweenness calculations treat directed friendship ties as undirected, and transitivity is missing for subjects with fewer than 2 friends.
- Network Measures: Friendship nominations provide each subject’s in-degree and out-degree, measuring incoming friendship nominations and the number of friends each person names.In-degree is theoretically unrestricted, whereas out-degree is capped at 3 by the name generator.
- Network Measures: Transitivity is the empirical probability that two of a subject’s friends are also friends, calculated as observed triangles divided by possible triangles.The measure is undefined for individuals with less than 2 friends, including 23 cases out of 744, which are treated as missing.
- Network Measures: Betweenness centrality measures the extent to which an individual lies on potential contagion paths between other individuals in the network.It is based on shortest paths and is represented as x_j for subject j.
- Network Measures: All network-measure scores are divided by max(x_j), placing them between and including 0 and 1.This normalization is stated for the centrality scores and ensures a common bounded scale.
- Network Measures: For transitivity and betweenness, directed ties are treated as undirected, so a tie in either direction becomes a mutual tie.Under this assumption, cyclic and one-directional friendship patterns can count as transitive relationships.
Personality Measures
Self-perceived popularity was measured with eight adapted questions rated on a five-point agreement scale. The responses were combined into index scores using one-dimensional factor analysis.
- Personality Measures: Eight questions adapted from a measure of co-worker popularity assessed self-perceived popularity.The measure was designed to capture subjects’ own perceptions of their popularity.
- Personality Measures: Responses used a five-point scale ranging from strongly disagree to strongly agree.Subjects rated their agreement with each popularity statement.
- Personality Measures: The eight items covered popularity, acceptance, visibility, admiration, liking, and being viewed fondly, with “I am not popular” reverse scored.The statements included “I am popular,” “I am quite accepted,” “I am well-known,” “I am generally admired,” “I am liked,” “I am socially visible,” and “I am viewed fondly.”
- Personality Measures: Index scores were generated through one-dimensional factor analysis of all eight items.All eight responses contributed to the factor-analysis-based index.
Analysis
The friend group differed from the random group in network structure and composition, and was diagnosed with flu earlier in both medical-staff and self-reported data. Network centrality measures predicted earlier flu onset, whereas popularity and nominated-friend counts did not consistently indicate earlier timing.
- Group differences: The friend group had significantly higher in-degree and betweenness centrality but significantly lower transitivity than the random group.The groups also differed in composition, with more females and fewer sophomores in the friend group.
- Correlates: ρ = 0.40: self-reported and medical-staff flu measures were highly correlated, but no other variable was significantly associated with both measures.The strongest associations with self-reported flu were in-degree and being a sophomore, each at 0.08, and neither was confirmed by medical-staff diagnoses.
- Flu timing: 15 days earlier: the friend group was diagnosed with flu by medical staff than the random group, with significance unchanged after controlling for other factors.The estimate used cumulative logistic models with bootstrapped standard errors and confidence intervals.
- Flu timing: 3 days earlier: the friend group self-reported flu symptoms than the random group, with significance unchanged after controlling for other factors.The self-reported cumulative-incidence model used bootstrapped standard errors and confidence intervals.
- Other predictors: Self-reported popularity had an inconsistent effect on flu timing, and the number of friends nominated delayed average flu onset.The friend-group effect remained significant when controlling for self-reported popularity.
- Network predictors: High in-degree and high betweenness centrality predicted earlier flu onset, while low transitivity predicted earlier onset.Betweenness and transitivity remained significant predictors after controlling for degree variables.