‘Herding’ in direction choice-making during collective escape of crowds: How likely is it and what moderates it?

Highlights

• We empirically re-examine the ‘herding’ assumption in emergency evacuation of crowds.

• Experiment design includes different levels of crowdedness and urgency (or stress)

• Individual-level decision observations are extracted and analysed using choice models.

• When making direction decisions, humans do not tend to follow the majority.

• The dominant behaviour is follow the minority and greater stress or crowding amplifies that.

• The ‘symmetry breaking’ phenomenon for ants does not generalise to human herding.

Abstract

The most likely responses of humans in emergency escape scenarios is not perfectly understood and many of the current behavioural assumptions are speculative. Because of the insufficiency of empirical evidence, certain assumptions and terminologies in this area have been derived from analogical experiments with non-human crowds (e.g. ants or mice, as models of humans) or from purely numerical analyses. Some have suggested that stressful collective escape situations trigger an increased tendency to imitate the decision of the majority, the so-called ‘herd-behaviour’ assumption. Here, we empirically test this assumption using a series of novel experiments with human crowds. Individual-level observations of direction choice were gathered and analysed using econometric modelling methods. Results showed that humans do not tend to imitate direction choices of the majority. To the contrary, they tend to avoid the direction chosen by the majority, and the bigger the majority is, the less likely they are to follow it. The high-urgency treatment (assumed to be associated with higher degrees of stress) did not reverse, nor did it decrease this avoid-the-majority tendency. If anything, it even amplified it in certain choice situations. We also found out that the general level of crowding (i.e. the total number of people in the choice-maker’s vicinity) is another factor that can moderate the reaction to peers’ decision. Higher levels of crowding also amplified the avoid-the-crowd tendency in certain direction choice scenarios. The results overall suggested that escaping humans, when not facing substantial degrees of information uncertainty, tend to avoid the direction chosen by the majority (opposite the ‘herding’ assumption); and they do so more distinctly when they perceive higher urgency or greater number of people in their vicinity. An implication of our finding is that strong parallels between escape behaviour of humans and animals/insects may not exist. Such analogies (particularly on decision-making aspects) need to be drawn with great caution. They might misguide modelling assumptions and lead to unrealistic predictions.

Introduction

Escape response is an essential aspect of behaviour on which the survival of species relies. As a result, the mechanisms of fleeing a predator have been rather extensively studied in relation with a broad variety of animal species (Binning et al., 2014, Brokordt et al., 2012, Møller and Ibáñez-Álamo, 2012, York and Bartol, 2016). Strikingly, however, this particular behaviour is relatively less explored in relation to humans. Even though flee-a-threat scenarios are increasingly occurring in human societies, the escape response behaviour of humans is yet to be adequately understood. A clear application of such knowledge would be the ability to predict escape and evacuation processes of human crowds in computer simulated settings (Helbing et al., 2000), and make reliable predictions whereby planners can take preparation measures and mitigate casualties (Bohannon, 2005, Low, 2000).

It is clear that the accuracy of such models and the subsequent estimates can only be as good as the assumptions made for reproducing (or simulating) the (most-likely types of) responses and behaviour of humans (Gwynne et al., 2016, Gwynne et al., 2017, Gwynne and Hunt, 2018, Kobes et al., 2010). Unrealistic behavioural assumptions can undoubtedly lead to misguided modelling outcomes (Gwynne et al., 2016, Kuligowski, 2013, Ronchi et al., 2014). As stated by Gwynne et al. (2012), “A lack of understanding of the subject matter can lead to inappropriate assumptions of human behaviour during emergency scenarios, and in turn, produce inaccurate evacuation modelling results” (page 136). Despite this known fact, many models have resorted to intuitive assumptions due to the unavailability of data and reliable empirical evidence in many areas (Cuesta and Gwynne, 2016, Cuesta et al., 2017, Haghani and Sarvi, 2018a, Shi et al., 2009). Natural observations from real world suitable for scientific analysis often do not exist in this area, and experimentation poses a range of logistical, financial, and most importantly, ethical challenges (Bode and Codling, 2013, Haghani and Sarvi, 2018a).

One particular assumption that is often treated as a given fact in research on this topic is the assumption of ‘herd behaviour¹’ during emergency evacuations. The assumption originates predominantly from purely theoretical studies (Helbing et al., 2000, Pan et al., 2007, Wang et al., 2016a, Zhao et al., 2008) as well as a number of experiments with non-human species that use escaping ants (Altshuler et al., 2005, Chung and Lin, 2017, Li et al., 2014, Wang et al., 2016b) or escaping mice (Perez and Saloma, 2009, Saloma and Perez, 2007, Saloma et al., 2015, Saloma et al., 2003) as models of human behaviour. In the current literature of evacuation modelling, the term ‘herding’ is basically used to indicate ‘imitation’ or ‘copying behaviour’. As characterised by the recent study of Van den Berg et al. (2018), “In case of an evacuation, people may also be influenced by the behaviour of other people, and copy this”. The authors define “this to be herding behaviour, and define it as seeing other people doing something and believing that what they are doing is a good alternative, resulting in doing the same thing” (Van den Berg et al., 2018).

In enumerating the characteristic features of “escape panic”, Helbing et al. (2000) state that (in stressful escape scenarios), “people show a tendency towards mass behaviour, that is, to do what other people do” (page 487). In one of the most influential published studies on this topic, Altshuler et al. (2005) built their work on this theoretical and unverified statement about imitative behaviour mentioned above and produced the following propositions. “The phenomenon of herding is a very general feature of collective behaviour in many species in panic conditions, including humans” (page 643) and that “it is not possible to overestimate the importance of the study of crowd stampede induced by panic, such as that taking place when people try to escape from a burning room, the “follow-the-crowd” effect associated with imitation may have consequences” (page 643). In their experiments, Altshuler et al. (2005) showed that ants confined to a cell with two exits positioned symmetrically in the cell utilise the exits equally in normal conditions but do so in disproportionate ways when panic is induced in their environment by injecting a repellent fluid. Having observed this phenomenon (which they labelled as ‘symmetry breaking’ (Altshuler et al., 2005)), they drew parallels between the escape behaviour of humans and ants and propositioned that their “experimental results, combined with theoretical models, suggest that some features of the collective behaviour of humans and ants can be quite similar when escaping under panic” (page 643).

The assumption of ‘herding’ or ‘symmetry breaking’ has also been restated repeatedly in subsequent studies that succeeded the two articles mentioned earlier. As one of the most recent studies among those, Chung and Lin (2017) reported that “to achieve optimal benefits (lowest casualties) in an evacuation system, humans would utilise two symmetric choices equally. However, panic-induced symmetry breaking occurred when humans escaped from a room with two exits. Individuals instinctively followed the crowd instead of rationally choosing the less crowded exit, resulting in chaos and jamming at one exit” (page 1). In another study related to this topic, Li et al. (2014) stated that “symmetry breaking is observed in both human crowds and ant colonies” (page 1). While we were unable to pinpoint these particular incidents or observations cited by these studies, the prevalence of these statements shows how entrenched the assumption of herd-behaviour has become in the literature to the extent that it is often regarded by many researchers as a given piece of knowledge or a fact. Similar to their predecessor studies based on escaping ants experiments, Chung and Lin (2017) were also able to create symmetry breaking with ants but through the use of heat-induced panic (as opposed to the use of chemicals). They observed that the degree of asymmetry increased linearly by increasing the magnitude of the aversive stimulus (that is, temperature). However, the question remains as to whether these observations are generalizable to humans too.

The assumption of the herd-behaviour, due to the prevalence of the term in the literature of evacuation modelling (Hong et al., 2018, Saloma and Perez, 2007), has even become the basis for many researchers in developing and formulating simulation models of evacuation (Hong et al., 2018, Kouskoulis and Antoniou, 2017). Many computational modellers have aimed to formulate the rules of behaviour in their models in a way to produce herd flows as a presumed likely pattern of behaviour that the model is expected to generate (as suggested by the empirical body of the research). For example, as a relatively influential theoretical study, Pan et al. (2007) developed a simulation system of evacuation modelling with the aim to produce “some emergent behaviours, such as competition, queuing and herding behaviours” (page 113). In their study, they criticised the fluid or particle analogies (for escaping humans) on the basis that they “contradict with some observed crowd behaviours, such as herding behaviour, multi-directional flow and uneven crowd density distribution” (page 115). Making a reference to Low, 2000, Pan et al., 2007, they further stated that “herding behaviour is often observed during the evacuation of a crowd in a room with two exits, one is clogged while the other is not fully utilised” (page 115). The study of Song et al. (2013) presents the method of crowd evacuation simulation for bioterrorism in micro-spatial environments. Based on one of the behavioural rules that was implemented in that model, some occupants “may lose their own decision-making capacity and the herding behaviour may appear for following specific individual” (page 108). In the model proposed by Qu et al. (2014), “In the microscopic level, based on the heuristic force-based model, an integrated simulation model is proposed to describe the herding behaviour” (page 188). And as one of the most recent examples, Hong et al. (2018) “used the hypothesis of herd behaviour to model the passenger decision-making process that leads to self-evacuation” (page 127) and integrated the model with pedestrian simulation software. In their modelling framework “when passenger O becomes aware of the emergency, (they) assume that he or she identifies the nearest three people as the ones to follow. The temporary evacuation target D is determined by the direction in which these three people are moving” (page 129). The assumption of herd-behaviour is also even frequently cited a measure of realism for crowd evacuation models or virtual-reality experiments suggesting that models are not accurate enough unless they produce herding patterns (Wang et al., 2016a).

In one of our earlier studies, we pioneered a series of simulated crowd escape experiments with humans designed for individual-level analysis, as part of which we obtained and reported modelling information about this particular aspect of the behaviour (Haghani and Sarvi, 2017a). Using a disaggregate dataset of direction choices during simulated escape experiments, we quantified the marginal desirability of choosing the direction chosen by a greater number of people. The outcomes showed that this marginal utility is actually negative (meaning that more people choosing a direction would actually reduce the desirability of that direction) unless there exists a choice ambiguity effect in the environment. In the case of choice ambiguity, we realised that the marginal effect of the peer influence was moderated by the level of choice ambiguity associated with each alternative direction. This finding along with the recent empirical findings of Li et al., 2016, Bode and Codling, 2013 which did not identify any herd-type pattern in direction choosing during computer-based experiments motivates this work and the further experimentation.

Particular elements that could be assumed missing from our previous experiment (which was of an exploratory nature, as opposed to a design for specific factors) were as follows. Firstly, the experiment was primarily designed for a multi-attribute analysis of choice only part of which was related to the perception of the social influence. Secondly, the experiment did not, in any way, control for the level of urgency (associated with the stress level). Even though participants performed the tasks in engaging ways and created reasonably realistic simulated scenarios of escape, there was no design element through which we could systematically control the level of perceived (or simulated) urgency (or motivations to escape). Therefore, the question remained for us to explore whether the observed behavioural patterns (such as failing to see strong herding tendencies) was in any form linked to the level of simulated urgency perceived by escapees given that the experiment only simulated only one type urgency-level treatment and did not systematically vary the level of simulated urgency. Here, we report on an experiment dedicated specifically to testing the herding hypothesis in conjunction with the effect of two context-specific factors: crowding level and urgency level (higher levels of which, according to previous studies in this domain (Bode and Codling, 2013, Moussaïd et al., 2016), is likely to induce stress).

We specifically investigate and revisit the herding assumption in escape direction choice (Haghani and Sarvi, 2016a) scenarios using dedicated experiments designed particularly for this effect. The experiments are designed for individual-level analysis (an approach pioneered by our previous study) as opposed to macro-scale analyses. The design isolates the social influence factor from other contributing factors to exit decision making (such as choosing the nearest exit or the exits that are visible). We create physically symmetric scenarios of direction choice for which the choice of escape direction can be assumed to be primarily determined by the sole effect of the peer influence (Haghani and Sarvi, 2016b, Haghani and Sarvi, 2017b, Kinateder et al., 2014a, Kinateder et al., 2014b). Moreover, for the first time in experiments with humans crowds in this context, we systematically control the level of urgency (or motivation to escape, also associated with the stress level (Bode and Codling, 2013)) by performing our scenarios as low-urgency and high-urgency treatments. This capability was achieved by dividing our sample of participants into two separate sub-samples (or two teams) and staging a reward-based competition between them. The reward is basically determined by the time that they take to escape an environment. The monetary reward and time pressure are basically utilised as an experimental proxy for the higher degree of motivation to escape that is experienced in real-life scenarios or urgent evacuation. This is based on the premise that time pressure can increase stress levels, as suggested by previous work in this field² (Gwynne et al., 2016, Moussaïd et al., 2016, Muir et al., 1996, Proulx, 1993).

The high-urgency and low-urgency conditions are mainly differentiated based on whether we measure performance and whether we enforce the competition in each scenario. So far, according to the existing literature, stressful escape experiments have solely been limited to experiments with ants, mice or sheep using a variety of aversive stimuli like chemicals (Li et al., 2014, Shahhoseini and Sarvi, 2017, Shahhoseini et al., 2016, Soria et al., 2012, Wang et al., 2016b, Wang et al., 2015, Wang and Song, 2016), high temperature (Boari et al., 2013, Chung and Lin, 2017, Pastor et al., 2015), water pool (Saloma et al., 2015, Saloma et al., 2003) or smoke (Lin et al., 2016, Lin et al., 2017, Zhang et al., 2018) as well as experiments with human subjects in computer-based game-type or virtual-reality experimental settings (Bode and Codling, 2013, Moussaïd et al., 2016). This work pioneers in accommodating this effect in experiments with real crowds of humans. Experiments under incentive-induced stress have been increasingly reported in recent studies particularly in the field of cognitive sciences on various aspects of decision making (Boos et al., 2014, FeldmanHall et al., 2015, Gathmann et al., 2014, Keinan, 1987, Pabst et al., 2013, Sokol-Hessner et al., 2016, Starcke and Brand, 2012, van den Bos et al., 2009, Yu, 2015) Also, compared to our previous experiment, the design of the physical environment is in such way that it forces evacuees to make decisions at particular points. This removed the extra effort of data extraction related to the identification of the “decision moments” and substantially reduced the ambiguity of decision moments and thus created more reliable sets of choice observations and (in our view) more robust subsequent statistics.