Saving Human Lives: What Complexity Science and Information Systems can Contribute

2.1 Crowd Turbulence

In many cases, fatalities result from a phenomenon called “crowd turbulence”, which occurs when the density in the crowd is so high that bodies automatically touch each other. As a result, physical forces are transmitted from one body to another—a process that is unavoidable under such conditions. These forces may add up and create force chains that cause sudden subsequent pushes from various directions, which can usually not be anticipated. Eventually, the individuals are pushed so much that one or some of them may stumble and fall to the ground. This creates a “hole” in the crowd, which breaks the balance of forces for the surrounding people: they are pushed from behind, but not anymore from the front, such that they are either forced to step on fallen persons (“trampling”) or they will also fall. Due to this domino effect, such “holes” eventually grow bigger and bigger (a situation that may be coined “black hole effect”). This explains the phenomenon of piles of people, which are often identified as the causes of crowd disasters and interpreted as instances of trampling. Under such conditions, it is extremely difficult for individuals to get back on their feet and people on the ground eventually die of suffocation.

Computer simulations of the above model in crowded situations, where physical interactions dominate over intentional movements, give rise to global breakdowns of coordination, where strongly fluctuating and uncontrollable patterns of motion occur. Crowd turbulence is particularly likely around bottlenecks, where local increases of the pedestrian density enhance the propagation of physical pressures from one individual to another. In particular, the unbalanced pressure distribution results in sudden stress releases and earthquake-like mass displacements of many pedestrians in all possible directions, which is well approximated by a power law with an exponent of \(1.95\) (see Fig. 2). This result is in good agreement with empirical observations of crowd turbulence, which exhibit a power law distribution with exponent \(2.01\) [27].

Open image in new window

2.2 Simple Rules for Safety

While sophisticated procedures combining expert consulting, computer simulations and video monitoring remain the safest way to avoid accidents, most event organizers are not in a position to undertake them all. In fact, such complex procedures are often expensive and turn out to be cost-effective only for large events, or if mass events are regularly held at the same place (e.g. the Muslim pilgrimage in Mecca). Hence, organizers of small- and intermediate-size events are often constrained to rely on their intuition and past experience to assess the safety of the events they are in charge of. However, the dynamics of crowd behavior is complex and often counter-intuitive. A systemic failure is usually not the result of one single, well-identifiable factor. Instead, it is the interaction between many factors that cause a situation to get out of control, while each factor alone does not necessarily constitute the source of disaster per se [26].

The following paragraphs offer some simple rules that can help one to improve the safety of mass events. These rules should not be considered as replacement of existing regulations, but can nevertheless complement more sophisticated procedures, and are easy to apply for all types of crowding, ranging from large mass events to daily commuting at train stations and airports.

2.3 Apps for Saving Lives

Pre-routing strategies are an important element of crowd management as well. For this, a well-functioning information and communication system is crucial. While classically, the situation in the crowd is observed by means of surveillance cameras, helicopters or drones, and local police forces, there is an array of crowd sensors available today, based on video, WiFi, GPS, and mobile tracking. Smartphones, in particular, offer a new way of collecting information about crowded areas and of providing advice to attendees of mass events.

Advances in microelectronics have resulted in modern mobile phones that offer significant processing power, sensing capabilities, and communication features. In combination with a large market penetration they provide an excellent instrument for crowd monitoring. In particular, data can be acquired by deploying a mobile application that probes relevant modalities, e.g. the GPS or acceleration sensors. Adopting a participatory sensing scheme, macroscopic crowd parameters (e.g. density), mesoscopic parameters (e.g., collective behavior recognition, group detection, lane and queue detection, onset of panic prediction) and microscopic parameters of single individuals (e.g., modes of locomotion, velocity and direction, gestures) can be inferred.

In safety critical and crowded environments such as major sports or religious events, instrumenting the environment with static sensors based on video or infrared cameras is currently the most reliable way to obtain a near 100 % sample of the crowd at a sub-second and sub-meter resolution required to accurately measure crowd pressure. However, smartphone apps can be used to infer a good approximation of crowd behavior at large scales, and can be used to pro-actively intervene before crowds reach critically high density levels.

From a safety perspective, employing apps for large-scale events has at least three major advantages:

1. 1.

   Broadband-communication allows to capture, transmit and centrally process data in near real-time and to extract and visualize relevant crowd parameters in a command and control center (Fig. 3).

    
2. 2.

   Offering bi-directional communication, safety personnel can send notifications, warnings, or even guide the user in case of an emergency situation. Incorporating the user’s localization, geo-located messages increase the relevance for each user, helping to follow rules such as those outlined in Sect. 2.2 [33].

    
3. 3.

   Data, once collected, can be used to “replay” the event. The post hoc analysis is a critical step in the organization of an event and can reveal critical factors that should be addressed for future events. Although today a great deal of data is available for such analyses, event organisers and crowd managers are sometimes forced to manually scan through video material, reports from security services in the field, or feedback from individual visitors. Combining such heterogeneous sources is a tedious and error-prone task, thus making it difficult to reach a reliable assessment and situational awareness. Localizing users via smartphones potentially offers a more accurate assessment and allows to capture crowd dynamics by aggregation.

    

The use of smart phones for crowd monitoring is a very recent development. Different modalities such as bluetooth sensing [34, 35] or app-based GPS localization [36] have been used to capture collective dynamics such as flocking, crowd turbulence or mobility patterns during large-scale events. In particular, individual location data can be aggregated using non-parametric probability density functions [36, 37] to estimate safety-relevant characteristics such as crowd movement velocity, density, turbulence or crowd pressure.

Open image in new window

Here, we demonstrate the potential of location-data obtained from mobile phones during the Züri Fäscht festival 2013 in Switzerland (see Fig. 4). The Züri Fäscht is a three-days event comprising an extensive program with concerts, dance parties, and shows. It is hosted in the city center of Zürich and is the biggest festival in Switzerland. Up to 2 million visitors have been estimated to attend the festival over the course of three days. In 2013, 56,000 visitors downloaded the festival app, from which 28,000 gave informed consent to anonymously contributed their location data. Figure 5 shows the number of users simultaneously contributing their location data over the course of the event, and the amount of data samples collected. While collecting only a subsample from the entire crowd, previous work [36] showed a correlation coefficient greater than \(0.8\) between the estimated density and actual density determined from video recordings.

Open image in new window

Open image in new window

The time-dependent density of visitors derived from the app data is shown in Fig. 6. The velocity of festival visitors is represented in Fig. 7 and the movement direction in Fig. 8. Together the three figures reflect the collective crowd dynamics. While Fig. 6 shows spectator densities for multiple events at different times, Figs. 7 and 8 show the mobility of the crowd before, during, and after the fireworks show at lake Zürich. Mobility patterns can be clearly identified. For instance, a strong northbound flow of the crowd can be observed before the event. Interestingly, lanes appear with a right-handed pedestrian traffic (highlighted box), which improves mobility. To support such kind of self-regulation, street signs (e.g., walking direction arrows) have been employed at other places, for instance at obstacles dividing the pathway.

Open image in new window

Open image in new window

Open image in new window

Several reality mining projects have demonstrated that it is not only possible to sense collective behaviors with mobile devices, but also to offer real-time feedback to the users via a return channel of the app. This allows crowd control staff to actively moderate the changes in the collective behavior. For example, the safety personnel may advise visitors to avoid crowded areas and suggest alternative routes. Such measures thus help improve the safety of crowds considerably.

3.1 Fighting Crime Cycles: The Cyclic Re-emergence of Crime and How to Prevent it

Crowd disasters are rare, but may kill hundreds of people at a time. In contrast to this, crimes are committed many times each day but usually involve only a few people in every incident. However, the impact of crime adds up and is considerable. In the US, for example, over 12,700 people were murdered in 2012 [39]. Economically, cybercrime, for example, results in hundreds of billions of dollars of damage world-wide every year.

To contain criminal activities, societies resort to various mechanisms punishing behavior that deviates from generally accepted norms in intolerable and destructive ways. The percentage of the population imprisoned in a country is an indication for the severity of state punishment (see Fig. 9a). Remarkably, the US has the highest incarceration rate in the world [40].² Nevertheless, its crime rates remain very high by international comparison and are much higher than in many other industrialized countries [42].³ In particular, in connection with the “war on drugs”, apparently more than 45 million arrests were made, yet it was recently declared a failure. In fact, US Attorney General Eric Holder concluded: “Too many Americans go to too many prisons for far too long, and for no truly good law enforcement reason.” [44].

Open image in new window

In reality, increasing levels of deterrence do not necessarily reduce the level of crime. Instead, one often finds crime cycles. The homicide rate in the United States, for example, began to rise steadily in the 1960s and oscillated at around 8 to 10 homicides per 100,000 for 20 years. It only started to significantly decline in the early 1990s (see Fig. 9b). In fact, such fluctuations in the frequency of crime are visible across a wide and diverse spectrum of criminal offenses, suggesting that cyclic recurrence is a fundamental characteristic of crime (see Fig. 9c). This observation is notable, in particular, as different criminal offenses—homicide and theft, for example—are thought to arise from different kinds of motivations: homicides are usually assumed to be related to manifestations of aggression, whereas theft is most often economically motivated [46].

Criminological research has identified a number of factors that influence the recurrence of crime, its trends and variation across regions and countries. Beside structural factors, such as unemployment [47, 48] and economic deprivation [49], a number of studies have highlighted the influence of demography [50], youth culture [51, 52], social institutions [53] and urban development [54]. Others have pointed to the influence of political legitimacy [49], law enforcement strategies [55, 56], and the criminal justice system [50]. Some recent studies in criminology, in fact, argue that trends in the levels of crime may be best understood as arising from a complex interplay of such factors [54, 57]. Very recent empirical work has further shown that social networks of criminals have a distinct impact on the occurrence of crime [58, 59], thus highlighting the complex interdependence of crime and its social context.

Despite these advances the situation for policy makers to date remains rather opaque, since fully satisfactory explanations for the recurrence of crime are still lacking. Few studies fully recognize the complex interaction of crime and its sanctioning [5, 6, 60, 61]. Analyzing this relationship, however, is of critical importance when trying to explain why, contrary to expectations [45], increased punishment fines do not necessarily reduce crime rates [62]. To understand its social context, crime should be considered as a social dilemma situation: it would be favorable for all if nobody committed a crime, but there are individual incentives to do so. As a consequence crime may spread, thereby creating “tragedies of the commons” [63]. Large-scale corruption and tax evasion in some troubled countries or mafia and drug wars are examples of this.

The collective spatio-temporal dynamics of this simple spatial inspection game is complex and counter-intuitive, capturing the essence of the crime-fighting problem well. One can observe various phase transitions between different kinds of collective outcomes, as demonstrated in Fig. 10. The phase transitions are either continuous (solid lines in Fig. 10) or discontinuous (dashed lines in Fig. 10), and this differs markedly from what would be expected according to the rational choice Eq. (7) or the well-mixed model [64], namely due to the significant effects of inspection and spatial interactions. Four different situations can be distinguished:

1. 1.

   Dominance of criminals for high temptation \(\beta \) and high inspection costs \(\alpha \),

    
2. 2.

   Coexistence of criminals and punishing inspectors (\(P+C\) phase) for large values of temptation \(\beta \) and moderate inspection costs \(\alpha \),

    
3. 3.

   Dominance of police for moderate inspection costs \(\alpha \) and low values of temptation \(\beta \), but only if the inspection incentives \(\gamma \) are moderate [see panel (A)], and

    
4. 4.

   Cyclical dominance for small inspection costs \(\alpha \) and small temptation \(\beta \), where criminals outperform ordinary individuals, while these outperform punishing inspectors, and those win against the criminals (\(C+O+P\) phase) (supplementary videos of different routes towards the emergence of the cyclic phase are available at youtube.com/watch?v=pH9l-2h6PRo and youtube.com/watch?v=gVnCN3a9ki8).

    

Figure 11 provides a more detailed quantitative analysis, displaying in panel (A) the sudden first-order phase transitions from the \(C+O+P\) phase to the pure \(P\) phase, and from this pure \(P\) phase to the pure \(C\) phase, as inspection costs \(\alpha \) increase. In panel (B), we also observe a sudden, discontinuous phase transition, emerging as neighboring individuals are rewired. This introduces a small-world effect. We observe that conflicts escalate in amplitude until an absorbing phase is reached. It follows that the exact impact of each specific parameter variation depends strongly on the location within the phase diagram, i.e. the exact parameter combination. General statements such as “increasing the fine reduces criminal activity” tend to be wrong, which contradicts the widely established point of view. This may explain why empirical evidence is not in agreement with common theoretical expectations, and it also means we need to reconsider the traditional perspective on crime.

Open image in new window

Open image in new window

The model reviewed here, therefore, demonstrates that that “linear thinking” is inadequate as a means to devise successful crime prevention policies. The level of complexity governing criminal activities in competition with sanctioning efforts appears to be much greater than often assumed so far. Our results also reveal that crime is likely to be recurrent when there is a gain associated with criminal activity. While our model is, of course, stylized, it can nonetheless help to shed light on the counterintuitive impact of punishment on the occurrence and, in particular, recurrence of crime.

Our model also highlights that crime may not primarily be the result of activities of individual criminals. It should be rather viewed as result of social interactions of people with different behaviors and the collective dynamics resulting from imitative interactions [58, 59]. In other words, the emergence of crime may not be well understood by just assuming a “criminal nature” of particular individuals—this picture probably applies just to a small fraction of people committing crimes. In fact, criminal behavior in many cases is opportunity-driven or arises as a result of social interactions that get out of hand. This suggests that changing social context and conditions may make a significant contribution to the reduction of crime, which in turn has relevant implications for policies and law.

It is commonly assumed that we have come a long way in deterring crime. Many different types of crime are punished all over the world based on widely accepted moral norms, and there exist institutions, organizations, and individuals who do their utmost to enforce these norms. However, specific deterrence strategies may sometimes have no, or even unintended, adverse effects. Recent research finds, for example, that even though video surveillance usually allows for a faster identification of criminal delinquents, it rarely leads to a significant decrease in criminal activities in the surveyed area [65]. In fact, it does not always lead to an increased individual perception of safety [66], even though this is a key arguments for its widespread use.

Other recent findings suggest that certain police strategies might even be outright counter-productive. Based on the assumption that removing the leader of a criminal organization will disrupt its social network, police often attempt to identify and arrest him or her. A recent study analyzing cannabis producer networks in the Netherlands, however, shows that this strategy may be fundamentally flawed: all network disruption strategies analyzed in the study did not disrupt the network at all or, worse, increased the efficiency of the network through efficient network recovery [67].

These examples clearly highlight that an insufficient understanding of the complex dynamical interactions underlying criminal activities may cause strong adverse effects of well-intended deterrence strategies. A new way of thinking, maybe even a new kind of science for deterring crime is thus sorely needed—in particular one that takes into account not just the obvious and similarly linear relations between various factors, but one that also looks particularly at the interdependence and interactions of each individual and its social environment. One then finds that this gives rise to strongly counter-intuitive results that can only be understood as the outcome of emergent, collective dynamics. This is why complexity science can make important and substantial contributions to the understanding and containment of crime.

3.2 Dynamics of Terrorism and Insurgent Conflict

The past decade has seen a resurgence of international terrorism and several violent civil conflicts. The insurgencies that followed the US-lead interventions in Afghanistan and Iraq have caused civilian casualties in the order of hundreds of thousands while the US-declared “war on terror” has not significantly curbed international terrorist activity (Fig. 12). In fact, in the political arena, critics argue that international interventions may themselves serve as triggers for terrorist attacks while others suggest the exact opposite, i.e. that only the decisive interventions and aggressive measures against terrorist organizations have stopped a much more severe escalation of international terrorism.

Open image in new window

There are no studies to date that can convincingly support either of the above views and establish a clear causal relationship between international terrorist activities and the “war on terror”. At the level of individual conflicts, however, there are a number of studies that have helped to clarify the link between interventions and levels of violence [68, 69, 70, 71]. These studies in particular emphasize reactive or “tit-for-tat” dynamics as a fundamental endogenous mechanism responsible for both escalation and de-escalation of conflict. In the context of Iraq, one can find such reactive dynamics for both insurgent and coalition forces [69, 70]. Other research has identified similar dynamics in the Israeli-Palestinian conflict where both the Israeli and Palestinian side were found to significantly react to violent attacks of the opposing side [72, 73]. Establishing the (causal) effect of actions by one conflict party on the reactions of the other faction(s) thus appears central to a systemic understanding of endogenous conflict processes.

There is complementary statistical evidence pointing to relatively simple relationships underlying aggregate conflict patterns. Empirical data of insurgent attacks, for example, exhibits heavy tailed severity distributions and bursty temporal dynamics [7, 8]. The complementary cumulative distribution function (CCDF) of event severities is usually found to follow a power law (see Fig. 13):

$$\begin{aligned} \mathrm {Pr}(X\ge x) \sim x^{\alpha }, \quad x\ge x_0 \end{aligned}$$

(11)

where \(x_0\) is the lower bound for the power law behavior with exponent \(\alpha \). Note that such statistical regularities can, for example, be used to estimate the future probability of large terrorist events [74]. Bursty temporal dynamics may be characterized by considering the distribution of inter-event times, i.e the length of the time intervals between subsequent events. In a structureless dataset, that is in a dataset where the timing of events is statistically independent, the CCDF of inter-event times simply follows an exponential distribution. The deviation of timing signatures from an exponential shape—for example log-normal or stretch exponential signatures—thus points to significant correlations between the timing of subsequent events.⁶ Researchers have offered competing explanations for these system-level statistical regularities. The power laws in the severity of attacks, for example, have been linked to competition of insurgent groups and security forces for control of the state [7], but also to a positive feedback loop between the size and experience of terrorist organizations or insurgent groups and the frequency with which they commit attacks [75].

Open image in new window

It is important to note that the choice of spatial and temporal units of analysis in such aggregate, system-level analysis may strongly affect results. Choosing empirically motivated and sufficiently small spatial units of analysis avoids problems arising from considering time series of potentially unrelated or only very weakly related incidents. In the context of Iraq, for example, the violence dynamics in the Kurdish dominated North are generally quite different from those in Baghdad or the Sunni triangle. The same applies to the selection of suitable time periods, in which the same conflict mechanisms produce aggregate patterns. In fact, the non-parametric analysis of event timing signatures rests on the assumption that the conflict dynamics in the period analyzed are (approximately) stationary [8]. Timing analyses are therefore usually performed for shorter time intervals—for example 6 months—where the overall intensity of conflict does not significantly change (see Fig. 13).

While such statistical signatures point to similarities in conflict patterns across cases and help identify general conflict mechanisms, detailed case studies on the micro-dynamics of violence suggest that mechanisms at the group- or individual-level may be very conflict- and context-specific [78]. A number of micro-level mechanisms that all contribute to the aggregate violence patterns are likely to be present in many conflicts (see, for example Ref. [79]). The effect of (local) contact between warring factions on levels of violence, for instance, has been of central interest in conflict studies. Examining violence in Jerusalem between 2001 and 2009, a recent study addresses this question using an innovative combination of formal computational modeling and rich, spatially disaggregated data on violent incidents and contextual variables [80]. Past research has given empirical support for two competing perspectives: The first assumes that intermixed group settlement patterns reduce violence, as more frequent interactions enable rivals to overcome their prejudices towards each other and thus become more tolerant. The other argues that group segregation more effectively reduces violence, given less frequent contact and fewer possibilities for violent encounters (see Ref. [80] for a detailed overview of the literature supporting the two perspectives).

One can quantitatively show that this local, individual-based mechanism can indeed explain a substantial part of the aggregate violence dynamics in Jerusalem [80]. Moreover, considering counter-factual scenarios the model allows to analyze how different levels of segregation corresponding to proposed “futures” for Jerusalem would affect violence levels in the city. Conceptually, the study highlights the complex inter-dependence of micro-level conflict processes. In fact, not only the effect of (local) contact is conditional on socio-cultural distance: conflict (or its absence) in turn also affects the socio-cultural distances between conflict parties.

Such inferential methodology is especially useful when testing specific causal hypotheses, in particular those that may otherwise be difficult to detect against the background of large-scale violence unrelated to the hypothesized mechanism. We have previously emphasized the importance of reactive dynamics for an understanding of endogenous conflict processes. One such example is the effects of counterinsurgency measures on the level of insurgent violence. The direction of these effects is generally disputed, both theoretically and empirically. One line of reasoning suggests that hurting civilians indiscriminately will generally cause reactive support for the military adversary [70, 85]. Another line of thinking suggests that the deterrent effect of such measures leads to less support for the adversary [68, 86].

We use the MWA technique here to directly test this effect empirically by focusing on the initial insurgency (beginning of 2004 to early 2006) in Iraq. Specifically, we focus on the greater Baghdad area—the focal area of violence—and consider three frequent types of events. Raids refer to surprise attacks on homes and compounds, in which the military suspects insurgents or arms caches. The relative heavy-handedness and the high probability of disturbing or harming innocent bystanders sets these operations apart from more selective detentions. In turn, to attack incumbent forces insurgents frequently rely on Improvised Explosive Devices (IEDs) and civilian support in manufacturing, transporting, and planting them. If civilians were really more inclined to support the strategic adversary after having been targeted in raids, one would expect more IED attacks to take place after and in the spatial vicinity of raids compared to detentions or arrests under otherwise comparable conditions. The opposite should be true if deterrence (by raids) was the mechanism at work. As visible in Fig. 14, raids lead to higher levels of subsequent IED attacks compared to detentions: our estimates suggest that 2–3 raids, lead to one additional IED attack in the direct spatial vicinity within up to 10–12 days. This result thus lends empirical support to the notion that civilians support the adversary in reaction to indiscriminate violence. Based on such improved methodology for studying micro-level processes in conflicts, it is possible to gain a better and systematic understanding of the driving forces behind violence.

Open image in new window

The research discussed in this section highlights that the contributions of complexity science and empirical, case-oriented conflict research to the study of conflict are in many ways complimentary. A systemic perspective can serve to identify abstract general relationships that help shed light on aggregate empirical patterns—much like statistical physics, which uses the self-averaging properties of large \(N\) systems to study system level properties that emerge from complex microscopic interactions [75]. The in-depth and theory-driven analysis of selected cases in the literature on micro-dynamics of conflict on the other hands, tries to systematically identify these detailed micro-level mechanisms. While there has recently been a noticeable shift towards the study of micro-dynamics of conflict with significant conceptual and technical progress [78], it is important to note that historically much of the social science literature on conflict has mainly analyzed aggregate country- or system-level data.

The fundamental challenge to date remains bridging the divide between these two perspectives. In theory, statistical analysis at the level of aggregate distributions can offer unique systematic insights for the whole ensemble with detailed micro-level inference revealing the multitude of mechanisms underlying the ensemble properties. Yet in practice, the two research strands often exist rather disconnected and insights from one do not enter into the analysis or models of the other. While there may in principle be a systematic limit to the generalizability of insights from the macro to the micro level and vice versa, the current situation is clearly a consequence of lack of engagement with the respective “other” literature.

We are convinced that in light of the increasing availability of detailed conflict event datasets and the emergence of Big Data on conflicts, complexity science can make a very substantial contribution to conflict research. Its impact and relevance will be more substantial the more we tap into the unique insights on both the micro- and macro-dynamics of conflict the social science literature already has to offer.

3.3 Interstate Wars and How to Predict them

If conflicts are difficult to stop, then it should be of utmost importance to prevent their occurrence in the first place. However, any hope of preventing them rests on the ability to forecast their onset with a certain level of accuracy.

Unfortunately, forecasting conflicts has remained largely elusive to scholars of international relations. Historical studies of single wars lack out-of-sample predictive power [87, 88, 89]. More systematic approaches focusing on the conditions that tend to lead to war (e.g., arms races [90], long-standing territorial rivalries [91], or large and rapid shifts in power [92, 93]) rely on indicators that are coarse in time (typically yearly), and therefore tend to miss the crucial steps of the escalation of tensions and the timing of the conflict outbreak [11, 94, 95, 96]. Efforts at finer-grained codification of geopolitical tensions are labor-intensive and costly [97, 98] and limited in time [99, 100]. More systematic coding mechanisms using computer algorithms [101, 102] also have a limited time span, as do prediction markets [103, 104]. In addition, event-based data tend to miss the subtleties of international interactions: on the one hand, the absence of an event may be as important as its occurrence. Moreover, seemingly large events need not be cause for alarm whereas small events may greatly matter. The main obstacle to testing our ability to forecast conflicts, in other words, has been the lack of measures of tensions that are both fine-grained in time and cover a large time-span [105, 106, 107].

To fill this gap, geopolitical tensions were estimated by analyzing a large dataset of historical newspaper articles [12]. Google News Archive—the world’s largest newspaper database with over 60 million pages⁸—was used to search the text of every article for mentions of a given country together with a set of keywords typically associated with tensions (e.g., crisis, conflict, antagonism, clash, contention, discord, fight, attack, combat).⁹ A mention of a country together with any of the pre-specified keywords in a given week resulted in an increased estimated level of tensions for that country in that week. This procedure was repeated for every week from January 1st 1902 to December 31st 2011, and for every country included in the Correlates of War dataset [108].

The resulting dataset is a fine-grained and direct proxy for the evolution of tensions in each country, with more than one hundred years of weekly time series for 167 countries. Moreover, by relying on journalists’ perceptions of international tensions, some of the pitfalls associated with event-data can be avoided, since contemporaries will process events in view of their respective context and likely consequences, instead of simply classifying them in preexisting categories.

Reports about tensions in the news were found to be significantly higher prior to wars than otherwise (Fig. 15), which implies that news reports convey valuable information about the likely onset of a conflict in the future [12]. In fact, reports about geopolitical tensions typically increase well ahead of the onset of conflict (Fig. 16), therefore potentially giving decision-makers ample early warnings to devise and implement a response to help prevent the outbreak.

Open image in new window

Open image in new window

Even when controlling for a multitude of explanatory variables that have been identified as relevant in the literature (e.g., regime type, relative power, military expenditure), the sheer number of conflict-related news provides earlier warning signals for the onset of conflict than existing models. This hypothesis was tested more formally by using only information available at the time (out-of-sample forecasting). Using such data, the onset of a war within the next few months could be predicted with up to 85 % confidence. Keyword-based predictions significantly improved upon existing methods. These predictions also worked well before the onset of war—more than one year prior to interstate wars—giving policy-makers significant additional warning time [12].

Here we report on an additional finding suggesting the importance of uncertainty prior to conflict as an early warning signal for war. The outbreak of conflict is rarely unavoidable. While countries often prepare for potential conflicts well in advance, the escalation of tensions typically reflects a failed bargaining process, and news articles mirror this escalation by reporting on these developments. Yet, escalation is rarely linear, and its outcome remains uncertain until the very end. A negotiated solution can always be reached before hostilities start. As a result, proximity to the conflict will probably increase not only observers’ (journalists) worries about the outcome of the process, but may also generate diverging opinions and scenarios about the likely outcome of the process. As a result, we should observe not only an increase over time in the number of conflict-related news, but also an increase in the variance associated with these news.

Indeed, we find that the probability of conflict significantly increases as a function of both the weekly change in the number of conflict-related news items (\(\Delta \)news) and their standard deviation within a window of one year (Fig. 17). This results goes beyond [12] and suggests not only that the total number of conflict-related news and its change over time can be used as early warning signals for war, but also that time-variability in the number of news could significantly improve upon existing forecasting efforts. This finding is also in line with [109], since increased levels of variability can be used as early warning signals for critical transitions in a large number of physical, biological and socio-economic systems characterized by complexity and non-linearity (see also [110]).

Open image in new window

4.1 Modelling Disease Dynamics

The development of mathematical models in the context of infectious disease dynamics has a long history. In 1766 Daniel Bernoulli published a paper on the effectiveness of inoculation against smallpox infections [122]. His seminal work not only contained the first application of the theory of differential equations, but was also published before it was known that infectious diseases were caused by bacteria or viruses and were transmissible. The prevalent scientific opinion in Bernoulli’s time was that infectious diseases were caused by an invisible poisonous vapor, known as the miasma. Evidence existed, however, that inoculation of healthy individuals with degraded smallpox material incurred immunity to smallpox in some cases. Bernoulli’s theoretical work shed light on the effectiveness of this procedure, a topic vividly discussed by the scientific elite of Europe at that time. In the beginning of the twentieth century Kermack and McKendrick laid the foundation of mathematical epidemiology in a series of publications [123, 124] and introduced the SIR (susceptible—infected—recovered) model which is still used today in most state-of-the-art computational models for the dynamics of infectious diseases (see also Refs. [125, 126]).

These types of models were designed to describe the time course of epidemics in single populations in which every individual is assumed to interact with every other individual at the same rate. The spatial spread of epidemics was first modeled by parsimonious reaction diffusion type Equations  [127] in which the local nonlinear dynamics are combined with ordinary diffusion in space. The combination of local, initially exponential growth typical of disease dynamics, with diffusion in space, generically yields regular wave fronts and constant spreading speeds. This type of approach was successfully applied in the context the spread of the Black Death in Europe in the fourteenth century [128]. Despite their high level of abstraction, these models provide a solid intuition and understanding of spreading processes. Their mathematical simplicity allows one to compute how key properties (e.g. spreading speed, arrival times, and pattern geometry) depend on system parameters [129].

One of the key problems in understanding global disease dynamics in the twenty-first century is that some of the fundamental assumptions in homogeneous reaction-diffusion models are not valid, not even approximately: (1) the host population is typically not distributed homogeneously in space and (2) global human mobility is characterized by a lack of scale which cannot be accounted for by local diffusion processes. This is illustrated in Fig. 19 which shows the worldwide air transportation network. This network connects approx. four thousand airports worldwide and transports more than three billion passengers every year. Recent examples have shown that the strong spatial heterogeneity of the human population in combination with the lack of scale in global mobility no longer lead to regular wave patterns of global disease spread. SARS (Severe Acute Respiratory Syndrome) in 2003 initially proliferated in China but only shortly after its arrival in Hong Kong cases were reported in Canada, the United States and Europe. In 2009, the H1N1 pandemic started in Mexico, quickly reached the US and subsequently arrived in Europe and elsewhere in the world. The spread of modern epidemics is spatially incoherent and chaotic, generically does note bear any metric regularity, and observed spatio-temporal patterns depend sensitively on the outbreak location. All these factors complicate the development of predictive tools that can forecast the time course of emergent infectious diseases.

Open image in new window

Despite these challenges, complexity science has produced major advances in modeling the dynamics of global epidemics. Increasing computer power and the availability of a new quality of pervasive data, in combination with important theoretical insights from network science, have enabled scientists to develop highly sophisticated computational platforms that can simulate large scale pandemics in silico. Modern super-computers allow running highly detailed agent-based simulations that model up to almost a billion host individuals, each one with a specific behavioral profile. State of the art computer models also employ insights that have recently been gained by analyzing pervasive data obtained in natural experiments on human interactions, in particular, about human mobility. Smart phones and geo-aware devices have produced data from which parameters can be extracted that previously had to be assumed or guessed in models for disease spread.

The most successful computational frameworks for predicting disease dynamics on a global scale account for multiple factors which are believed to play a role in a specific context: demographic variation, mobility patterns that include the entire global air-traffic system as well as the short-scale, daily commuter movements in almost every country on the planet, detailed epidemiological data, and disease-specific mechanisms [115, 116, 118, 119, 120, 121].

A large scale operation in this direction is the Global Epidemic and Mobility (GLEAM) simulation framework [120]. This tool has a client/server architecture that allows the simulation of global epidemics by accounting for the entire worldwide air-transportation network in combination with daily commuter traffic in most countries. The approach accounts for a diverse set of mobility patterns, e.g. bi-directional movements between origin and destination, or different mobility patterns for host individuals in different infectious states. Hhospitalization, vaccination, age structure at every location, and seasonality, can also be incorporated in the model setup (see Fig. 20). Because of the client/server architecture, a model can be defined on the client, and the simulation itself can be run on a remote super-computer in little time. A clear advantage of the GLEAM software is that entire sets of fully stochastic simulations can be performed in order to compute confidence intervals along with the most probable time course. In this respect the design of the GLEAM infrastructure is similar to modern weather forecast systems. The GLEAM project and related simulation approaches have become remarkably successful in reproducing observed patterns and predicting the temporal evolution of ongoing epidemics [130] for instance in predicting the time-course of the 2009 H1N1 pandemic.

Open image in new window

Interestingly, many of the predictive, computational models that exist reproduce similar dynamic features despite significant differences in their underlying assumptions and data [131]. Furthermore, the level of detail these models incorporate also implies an abundance of parameters that either must be assumed, or in some cases guessed. Generically these parameters interact non-linearly in a given model, and despite their predictive power, it is difficult to assess which parameters are important and which are less significant. We currently lack a full understanding of the dynamic richness of these highly detailed models and therefore also of the phenomena they describe. Another drawback of purely detailed computational approaches is the fact that parameter values must be chosen and fixed to run a specific computer simulation. This is particularly challenging when new infectious agents emerge for which none of the vital disease specific parameters are known. Consequently, at the onset of an emergent epidemic even the most sophisticated simulation frameworks have to be employed with great caution because, despite their potential to model the phenomenon in a very detailed way, a significant level of uncertainty remains concerning the choices of sensitive parameters.

It also remains unclear, how the multitude of factors shape the dynamics and how much detail is required to achieve a certain level of predictive fidelity. Most importantly, detailed computational models that incorporate all potentially relevant factors ab initio fail to reveal which factors are actually relevant, and which ones are not [132]. Gaining a better understanding of the phenomenon, however, is particularly important because based on a deeper understanding we can potentially predict certain aspects about the most likely time course of a global epidemic, even when a set of parameters is unknown. Along this line of reasoning, in [14] the authors introduce the notion of effective distance to replace conventional geographic distance. This move simplifies the complexity of the spatio-temporal dynamics of global diseases to the point that even complex network-driven contagion phenomena can be understood in the framework of ordinary reaction-diffusion dynamics. In other words, when effective distance replaces geographic distance, spreading phenomena that appear to be complex in conventional geographic representations possess simple wave fronts and constant speeds. Thus, the key result of the study was that the observed complexity is predominantly caused by the inappropriate and antiquated use of purely spatial distance measures.

To best understand the approach, consider a metapopulation network, consisting of \(n=1,...,M\) coupled populations, in which local disease dynamics are captured by SIR kinetics [114]:

$$\begin{aligned} I_{n}+S_{n}&\xrightarrow {\alpha } 2I_{n}\end{aligned}$$

(14)

$$\begin{aligned} I_{n}&\xrightarrow {\beta } R_{n}\end{aligned}$$

(15)

$$\begin{aligned} X_{m}&\xrightarrow {w_{nm}} X_{n}, \end{aligned}$$

(16)

where \(S_{n},I_{n},R_{n}\) denote the states (susceptible, infected, recovered) of an individual in population \(n=1,....,M\). The first reaction represents disease transmission, the second recovery, and the third equation movement of a host from population \(m\) to \(n\) (\(X\) is a placeholder for all three classes of individuals, i.e. \(S,I\) and \(R\)). The parameters of this system are the population averaged effective per capita transmission rate \(\alpha \), the population averaged recovery rate \(\beta \), and the mobility rate \(w_{nm}\). The mobility rate is given by the per capita traffic flux from \(m\) to \(n\), i.e. \(w_{nm}=F_{nm}/N_{m}\), where \(N_{m}\) is the size of population \(m\). The basic reproduction number is defined according to

$$\begin{aligned} R_{0}=\alpha /\beta , \end{aligned}$$

(17)

i.e. the expected number of secondary infections caused by one infected individual in an entirely susceptible population while it is infectious. The dynamics of the metapopulation SIR model, in the most parsimonious form, are given by

$$\begin{aligned} \partial _{t}I_{n}&= \alpha S_{n}I_{n}/N_{n}-\beta I_{n}+\sum _{m\ne n}\left( w_{nm}I_{m}-w_{mn}I_{n}\right) \nonumber \\ \partial _{t}S_{n}&= -\alpha S_{n}I_{n}/N_{n}+\sum _{m\ne n}\left( w_{nm}S_{m}-w_{mn}S_{n}\right) \nonumber \\ \partial _{t}R_{n}&= \beta I_{n}+\sum _{m\ne n}\left( w_{nm}R_{m}-w_{mn}R_{n}\right) \end{aligned}$$

(18)

(recycling the symbols \(S,I,R\) for the number of individuals in a given state). Expressing the above in terms of fractions of susceptible, infected and recovered individuals (\(s_{n}\), \(j_{n}\), and \(r_{n}\) respectively), the above dynamical system simplifies to

$$\begin{aligned} \begin{aligned} \partial _{t}j_{n}&=\alpha \, s_{n}j_{n}-\beta j_{n}+\gamma \sum _{m\ne n}P_{mn}\left( j_{m}-j_{n}\right) ,\\ \partial _{t}s_{n}&=-\alpha \, s_{n}j_{n}+\gamma \sum _{m\ne n}P_{mn}\left( s_{m}-s_{n}\right) , \end{aligned} \end{aligned}$$

(19)

with \(r_{n}=1-s_{n}-j_{n}\), see Ref. [14]. The rate parameter \(\gamma \) is the average mobility rate, i.e. \(\gamma =\fancyscript{F}/\fancyscript{N}\) where \(\fancyscript{F}=\sum _{n,m}F_{nm}\) is the total traffic flux in the network and \(\fancyscript{N}=\sum _{n}N_{n}\) the total population in the system. The matrix \(\mathbf {P}\) with \(0\le P_{mn}\le 1\) quantifies the fraction of the travellers with destination \(m\) emanating from node \(n\), i.e. \(P_{mn}=F_{mn}/F_{n}\), where \(F_{n}=\sum _{m}F_{mn}\). If we now consider each node in the metapopulation network as the catchment area of each airport in the global mobility network and take \(F_{nm}\) as the passenger flux between airports \(n\) and \(m\), the above dynamical system constitutes a structurally simple model for the global spread of an emergent infectious disease. In fact, the earliest successful models that incorporated worldwide mobility were structurally similar to Eq. (19) and most sophisticated models (e.g the GLEAM simulation framework) possess an underlying mathematical machinery of similar nature [116, 121, 133, 134]. Essentially the above model has three rate parameters, \(\alpha ,\beta \) and \(\gamma \). Typical values for influenza-like disease are \(\beta \approx 0.3\,\text {days}^{-1}\), and \(R_{0}=\alpha /\beta \approx 1.3-2.3\). The global mobility parameter is typically much smaller \(\gamma \approx 0.001-0.02\,\text {days}^{-1}\). This implies that mobility has the slowest time-scale of the system. The complexity of the above model is encapsulated in the flux fraction matrix elements \(P_{mn}\) which is shaped by the spatial heterogeneity of the metapopulation and the multiscale properties of the global mobility network. It is this coupling matrix that causes solutions of the above model to appear spatiotemporally complex, as is shown in Fig. 21. The figure depicts temporal snapshots of two simulated pandemics (i.e. solutions to Eq. (19)) with initial outbreaks in London and Chicago. Because of the inherent complexity in the global mobility network, i.e. the coupling matrix \(P_{nm}\), the patterns appear to be complex, spatially incoherent and quickly lose any correlation to their initial outbreak location. Clearly, no geographic wave front is discernible except in the very beginning of the spreading process. The lack of geographic imprint makes it difficult to predict arrival time sequences from the maps alone. Furthermore, the geographic de-correlation makes it impossible to reconstruct the initial outbreak location based on a single temporal snapshot of the process. Unlike historic spreading processes that evolved according to more regular wave propagation, modern global disease dynamics make it difficult to define a proper propagation speed by, for instance, correlating geographic distance with arrival times.

Open image in new window

Because the spatio-temporal complexity generated by the model is caused by the coupling matrix \(P_{nm}\), a plausible strategy for defining an effective distance measure must relate this distance to the coupling matrix. A possible choice is the following definition of an effective, directed length of a link \(m\rightarrow n\) in the network:

$$\begin{aligned} d_{nm}=1-\log P_{nm}>1 \end{aligned}$$

If we interpret the flux fraction \(P_{nm}\) as the conditional probability that an infection that is present at node \(m\) will be passed to \(n\), then the effective length from \(m\rightarrow n\) is small if this probability is large, and vice versa. Now consider a multi-leg path \(\Gamma =\{n_{1},\ldots ,n_{L}\}\) through the network. Using the above definition, the effective, directed length of this path is the sum of the effective lengths of each of its legs. This implies that this length decreases as the probability increases that a path will be followed in a random walk process with hopping probabilities governed by matrix \(P_{nm}\). Based on this, the effective distance \(D_{nm}\) from an arbitrary reference node \(n\) to another node \(m\) in the network is defined by the length of the shortest path from \(n\) to \(m\):

$$\begin{aligned} D_{nm}=\min _{\Gamma }\lambda (\Gamma ) \end{aligned}$$

(20)

This is equivalent to the assumption that the most probable path dominates, much like the smallest resistor in an electrical network with parallel conducting lines. From the perspective of a chosen node \(n\), the set of shortest paths constitutes a shortest path tree \(\tau _{n}\), corresponding to the most probable hierarchical sequence of pathways of an epidemic that originates at a reference node. Figure 22 illustrates the properties of effective distance and the shortest path tree from sample reference node Zurich.

Open image in new window

The key question is: What are the properties of a simulated pandemic generated by the metapopulation model of Eq. (19) in this new, effective representation, when viewed from the initial outbreak location? This is shown in Fig. 23, which depicts the same simulation as is shown in Fig. 21, in the shortest path tree representation of each outbreak location.

Open image in new window

We see that the patterns that appeared complicated in the conventional geographic representation exhibit regular wavefronts in the effective distance representation. Furthermore, these effective waves propagate at constant speeds. Note that this is more than just a simple remapping of a complex pattern onto a simple one. First, the observation of the propagating wavefront implies that the original assumption that contagion phenomena are dominated by most probable paths was correct. Furthermore, the remapping only depends on the coupling matrix \(P_{nm}\) in the original nonlinear dynamical system. Because of this, one expects to observe concentric waves irrespective of the other parameters of the model, e.g. the rate parameters \(\alpha \), \(\beta \) and \(\gamma \).

Another insight gained by the approach is that one observes concentric wave patterns only from the perspective of the outbreak origin. From the perspective of any other node in the network, the dynamic pattern appears more irregular. This can be quantified and used to reconstruct the outbreak origin for contagion phenomena that have been spreading but are unnoticed until the incidence passes a certain threshold [135]. Typically these incidence patterns bear no geographic regularity and it is difficult, albeit essential, to reconstruct the spatial origin in order to develop efficient containment strategies. The effective distance approach can help reconstruct the outbreak origin by identifying the node from which the observed pattern appears to be most regular. This approach has successfully reconstructed outbreak origins of recent epidemics, e.g. the foodborne disease EHEC/HUS that struck Germany and its European neighbors in 2011.

4.2 Enabling More Effective Disease Control Using Social Information Systems

Vaccine-preventable diseases can be eradicated from a population when enough individuals immunize. Unfortunately, enforcing compulsory vaccination is challenging and potentially incompatible with personal liberties [136, 137, 138]. Beyond ethical and political concerns, the efficacy of compulsory vaccination programs may be constrained by (1) the limited information available to those designing the policies, and (2) uncertainty about how people will change their behavior [139] in response to the information they receive as an epidemic progresses. Here we will investigate when voluntary vaccination [140, 141] can be a viable alternative to imposing immunization in a top-down way. That is, we use a model to determine if and when information about the risk of becoming ill will enable individuals to voluntarily make decisions that are beneficial, not only to their own health, but also to the health of many others.

To model vaccination we consider the SIR model (see Sect. 4.1) with demography and assume that a fraction \(p\) of individuals vaccinates at birth (see Fig. 24a). Commonly, \(p\) is taken to be a constant. In a population where contact occurs at random between any pair of individuals (known as the well-mixed assumption) the dynamics of this model are described by the following system of ordinary differential equations (which are exact in the limit of infinite population size \(N\)):

$$\begin{aligned} \partial _{t}{s}&= -\alpha sj-\mu s+\mu (1-p) \nonumber \\ \partial _{t}{j}&= \alpha sj- \beta j- \mu j \nonumber \\ \partial _{t}{r}&= \beta j - \mu r \nonumber \\ \partial _{t}{v}&= -\mu v + \mu p , \end{aligned}$$

(22)

where \(s\), \(j\), \(r\) and \(v\) are the fraction of individuals that are respectively susceptible, infected, recovered and vaccinated. We assume that the population size is approximately constant, and therefore the natural birth and death rate are both given by \(\mu \). The contact rate between individuals is \(\alpha \), and \(\beta \) the recovery rate. (The fractions denoting the state of the system are a function of time although we do not explicitly show this dependence.) Setting the time derivatives to zero we can easily determine the steady-state infection level \(j^*\) as a function of \(p\):

$$\begin{aligned} j^{*}=\frac{\mu (1-p)}{\beta +\mu }-\frac{\mu }{\alpha } . \end{aligned}$$

(23)

This implies that the disease will die out when the vaccinated fraction of individuals is \(p>1-(\gamma +\mu )/\alpha \) (see e.g. [142] for a more detailed derivation).

Open image in new window

To model voluntary vaccination we modify this standard model so that the fraction of individuals who vaccinate is a function of the fraction of infected individuals rather than a constant. This captures the idea that individuals will perceive a greater risk of becoming ill when they see higher disease incidence around them, and are therefore more likely to vaccinate [139]. More specifically, following prior models [140] and experimental evidence [143], we assume that individuals make rational choices. In this framework an individual who vaccinates will incur a cost \(c_v\) (the cost of vaccination) and one that becomes ill will incur the cost of infection \(c_i\) (we assume that vaccines have \(100\,\%\) efficacy and thus vaccinated individuals will never become infected, but this assumption can be easily relaxed). Therefore, the expected cost to an individual who vaccinates is \(c_v\). Meanwhile, the expected cost to an individual who does not vaccinate is \(c_i R\), where \(R\) is an individual’s estimate of the probability that she will fall ill if she does not vaccinate. We further assume that \(R\) is simply estimated to be the fraction of infected others she sees; thus \(R=j\) in a well-mixed model.

To approximate the well-mixed dynamics in a system where individuals make vaccination choices according to rule (24), we use Eq. (22) with \(p=P_v\). Thus, the steady-state infection level \(j^*\) is given by Eq. (23), represented graphically by the intersection of the curves in Fig. 24b. The crucial insight to draw from this solution is that, as long as there is any cost associated with vaccination, there will be insufficient vaccination to eradicate a disease. A similar result, echoing the troubling conclusion that voluntary vaccination is not a viable alternative to enforced immunization, was derived through a game-theoretic argument in [140].

The prior result relies on the assumption that the spread of disease, and the information available to individuals about this spread, are highly homogeneous throughout the population. However, as discussed in Sect. 4.1, the complex contact patterns between individuals, which mediate the spread of diseases in the real world, lead to spatio-temporal variation in infection levels. Therefore, we ask how the effectiveness of voluntary vaccination changes when individuals make immunization choices using information that incorporates this variability. To answer this question we introduce a model where individuals interact physically only with a subset of the population, which we call contact neighbors. We employ the formalism of a network [144] to represent this contact structure [145, 146, 147]. Each individual is represented by a node, and the links between nodes represent the potential for an infectious interaction. To model variation in the information that individuals have about the health of others, we introduce a second network where individuals are connected if they are informed about each other’s state. (The transmission of information could be through any mechanism: e.g. through direct conversation, word of mouth, observation of symptoms, mass-media reporting, or social media.) The information neighborhood can be (1) a subset of the contact network, (2) exactly equal to it, or (3) a superset (see Fig. 25). By construction, all individuals have the same number of contact neighbors \(k\) and the same number of information neighbors \(k_{\text {info}}\). However, we allow for the number of contact neighbors to differ from the number of information neighbors. The key point is that, in this setting, an individual estimates the risk of infection \(R\) from the fraction of its infected information neighbors (rather than considering the global fraction of infected individuals as above). For both of these networks we consider two types of topologies: (1) a network where every node has the same predetermined number of neighbors but otherwise has a random structure, and (2) a lattice network (for description of these two types of networks see a standard network reference, e.g. Ref. [144]). These two simple network topologies are different enough to demonstrate that the results we find are not due to specific structural features.

Open image in new window

To investigate the effect that the information range has on the effectiveness of voluntary vaccination, we simulate an individual-level model of this system and vary the size of the information neighborhood \(k_{\text {info}}\). The contact neighborhood \(k\) is small and fixed (we set \(k=12\), but the behavior is similar for neighborhoods up to an order larger). To estimate the probability that vaccination will remove the disease from the population, we fix a time \(T\) much larger than the transient time, and count the number of simulations where \(j(T)=0\). We find that this fraction \(F\) of simulations where the disease is eradicated changes systematically with \(k_{\text {info}}\) (see Fig. 24). Interestingly, \(F\) exhibits a clear maximum when the information neighborhood is of an intermediate size. Access to information about a growing number of neighbors initially facilitates eradication of the disease. However, as individuals average over information obtained from increasingly large regions of the system, the disease becomes decreasingly likely to die out. Further experiments indicate that this qualitative behavior holds across different epidemiological parameters that allow the infection to invade the population when there is no vaccination,¹⁰ although the mechanisms driving the behavior vary [148].

In summary, although an epidemic cannot be contained by voluntary immunization when individuals use average global information, the successful eradication of a disease can occur for an intermediate range of information that is neither local nor global. In this range individuals make their vaccination decisions taking into account variability in infection densities. This suggests a promising alternative to top-down vaccination policies: empowerment of people through information systems such as social media, thereby allowing them to take responsible decisions to immunize themselves. Importantly, given the right information range, individual choices will be aligned with socially desirable outcomes.