A Virtual Reality Exit Choice Experiment to Assess the Impact of Social Influence and Fire Wardens in a Metro Station Evacuation

Accounting for evacuation while designing the built environment can be complex, and evacuation assessments represent a crucial aspect of Fire Safety Engineering. Over the past few decades, studies on pedestrian evacuation dynamics have increased and become a key research topic to support fire safety engineering [1‐4]. These studies are driven by the need for new data, which is necessary to evaluate the safety of a given built environment design when using a performance-based approach.

Performance-based design focuses on comparing the Available Safe Egress Time (ASET) with the Required Safe Egress Time (RSET) [5]. ASET is the time that can pass before the environment becomes too dangerous to escape, while RSET is the time required for all evacuees to exit safely [6]. There are many methods and models in the current state of the art that simulate the evacuation process to evaluate RSET [7‐9]. Quantifying evacuation time with these existing methods and models can be a challenging task for fire safety engineers, especially for complex built environment scenarios.

To accurately determine RSET, it is essential to consider the various factors. These factors encompass individual characteristics such as physique, age, gender, and cognitive capabilities, as well as environmental and behavioural elements like route and exit choices, visibility of guidance signs, familiarity with the environment, and the suitability of emergency lighting [10]. Among these, exit choice behaviour plays a pivotal role in shaping the overall evacuation dynamics and, consequently, the RSET. Therefore, the study of human behaviour in fire emergencies must be at the core of all life safety projects, examining how people respond to emergencies, including their awareness, attitudes, behaviours, motivations, beliefs, coping strategies, and decisions [11].

Several existing datasets on human behavior in fire and pedestrian dynamics have been proposed using data from field-based experimental studies, such as evacuation drills and real incidents [10‐13]. In these cases, researchers can reach a high level of ecological data validity. However, a downside of these data is that researchers have limited experimental control over the variables under examination [14]. For instance, in the cinema experiment published in [15], the only variable under control was the alarm type. In many other case studies, researchers had no control over the factors affecting evacuee behavior. Moreover, the involvement of human subjects in evacuation experiments carries a risk of grievous injury. An alternative is to collect data through laboratory experiments. In a laboratory setting, data can be collected by presenting participants with hypothetical scenarios and asking how they would respond in those situations. For this purpose, different methods can be employed. Virtual Reality (VR) is one of the emerging tools in the field [16‐18].

In recent years, VR has gained popularity in the safety research community [19]. VR has been defined as a “real or simulated environment in which the perceiver experiences telepresence” (the feeling of being present in a virtual environment) [20]. Therefore, VR is not limited to computer-generated environments or any specific technology but can include different immersive and non-immersive solutions [21]. VR is a research tool that balances ecological validity with the ability to control experiments [18]. If this balance is reached, VR can serve as an effective, versatile, and affordable platform for safety-related scenarios. Several existing studies have explored the use of VR to understand human behavior and train individuals for emergency scenarios [22‐24].

Current evacuation models often rely heavily on user input to predict fire evacuation scenarios, which can lead to inaccuracies and inconsistencies [25]. This dependence on ad-hoc user input limits the models'ability to accurately capture individual and group behaviours under various situations. There is a clear need for next-generation models that can autonomously predict evacuation behaviours based on real-time situational data rather than manual inputs. Gwynne et al. [26] have identified this gap, emphasizing the necessity for data-driven approaches that integrate comprehensive behavioural data to enhance the predictive accuracy of evacuation models. In this context, Virtual Reality (VR) represents a suitable approach to study human behaviour in fires and to develop the next generation of data-driven models for potential integration within existing or new evacuation software tools.

However, several studies have identified specific limitations of VR in replicating real-world evacuation scenarios [27, 28]. Technologically, VR systems may suffer from a limited field of view, latency-induced motion sickness, and insufficient environmental realism due to computational constraints. As an experience, VR may lack scenario authenticity and fail to evoke genuine psychological and physiological responses, which can influence evacuee decision-making. To mitigate these issues, careful experimental design and appropriate hardware selection are essential to enhance the ecological validity of VR systems and reliability of VR-based evacuation studies.

The present study utilises a new VR experiment to investigate exit choice, an important behavior affecting fire-related evacuations in an underground built environment [29]. This work assesses how both social influence and fire wardens’ (metro staff or firefighter) instructions can affect evacuees’ exit choices, especially when these two factors conflict. To achieve this aim, we developed an immersive VR experiment to investigate exit choices during a fire evacuation in a metro station using 12 different scenarios. The 12 scenarios differ in the presence or absence of a fire warden to guide the participants, the level of authority of the fire wardens, and the presence or absence of other evacuees. This experiment involved 131 participants who were asked to choose between two possible exits for 8 of the 12 scenarios. Finally, we estimate a discrete choice model to quantify if and how these factors affect the participants’ decisions.

This section provides an overview of exit choices and the factors affecting them, including social influence and role-rule model factors. Additionally, it summarises relevant studies that use VR technologies to investigate exit choice. Exit and route choices can significantly affect evacuation performance [29, 30]. Numerous studies have investigated exit choices during fires [31‐34]. These studies have identified several elements that can affect evacuee exit choices, such as social influence, evacuation systems’ affordances, familiarity with the environment, the geometry of the building, and harmful conditions within the environment itself (i.e., smoke) [35, 36]. Research by Lin et al. [37] demonstrated that when a crowd was split unevenly, participants were influenced to follow the majority. This trend was observed across multiple cultural contexts. Other studies indicate that people tend to evacuate via familiar routes and exits, most often the main entrances [38, 39]. This is an example of affiliative behavior, which occurs when people move towards “familiar persons and places”. In some cases, people prefer not to adopt a new route or an exit previously unknown, even if they are available and/or closer. Lovreglio et al. [40] found that in VR simulations of building fires, participants were less likely to use exits obscured by smoke unless they observed others successfully navigating through those areas.

Social influence describes the actors accounting for how an individual’s beliefs, thoughts, emotions, opinions, or behaviours are affected by others in their social network [41]. During emergencies, occupants must decide their destination (i.e., where they want to move to) and their route (i.e., how they want to reach their destination). In unfamiliar situations, such as those that occur in emergencies, a useful source of information can be the behaviour of other occupants [42]. Therefore, the presence of others in emergencies can influence an individual’s behaviour. Nilsson and Johansson [15] demonstrate that when fire cues are ambiguous (e.g. an alarm bell rather than a pre-recorded message), social influence is more significant and increases when people are closer to each other. An important aspect of social influence has been highlighted by Templeton et al. [43]. This study investigated the behavioural effects of shared social identity on crowd movement at a pedestrian level. By analysing the behaviour of a psychological crowd (where the entire crowd perceives themselves as part of the same group through a shared social identity) and a physical crowd (where members are in the same place at the same time), they demonstrated that the differences in speed, distance and proximity are crucial factors to consider when planning the egress of a crowd in the event of an emergency. Latané and Darley [44] found that social inhibition effects are common across different types of emergencies. In four separate experiments, bystanders were less likely to step in when others were present. The type of bystander made a difference: a non-responsive confederate caused the greatest inhibition, a stranger had a moderate effect, and a friend had the least. These findings support a multi-process model of intervention, showing that the presence of others influences both how a bystander will interpret the situation and their decision to act.

Many traditional findings from research on social influence have been duplicated effectively in virtual environments, as demonstrated in certain studies [45‐47]. These studies indicate that VR can be an effective platform for studying social influence [48]. However, a potential drawback in VR studies related to social influence is that virtual agents might be perceived differently from real people, as participants may not recognise animated agents as humans [49].

The role-rule model represents the relationship between roles, rules and behaviour and is a crucial aspect of human behaviour in fire emergencies. To analyse this important factor, Fridolf et al. [29] discussed an excellent example of human behaviour during an emergency at an underground station: the King’s Cross fire. They observed that the recognition, interpretation, and response to the first fire cues correlated with the role of an individual in everyday life. For example, underground staff members and police officers assumed positions of authority while, in contrast, tunnel users waited for authoritative instruction. This example highlights the significance of preparedness and training in handling emergency situations and the need for fire wardens capable of directing evacuees toward safe exits. Fridolf et al. [29] also observed that passengers responded more readily to instructions provided by the police than those from staff members. The role-rule model could explain this phenomenon, as people see the police as having more authority than underground staff members in a non-fire situation. Two further studies on policing and public order show that people’s response is influenced not only by how clear and useful the guidance is but also by their perception of the organisation delivering it [50, 51]. Templeton et al.[52] observed that trust in fire safety guidance is an essential factor that should be incorporated into evacuation models.

Various scientific observation and simulation techniques have been developed to assist in the evacuation of buildings threatened by disasters, such as fires. In recent years, VR and AR (Augmented Reality) have gained popularity in the safety research community [19]. Different combinations of hardware and software allow the public to access these technologies. In this work, only VR will be discussed. VR technologies are divided into immersive and non-immersive solutions. Non-immersive VR relies on displaying virtual content on a computer screen, while immersive VR uses technologies such as head-mounted displays or Cave Automatic Virtual Environments to create more immersive experiences [53‐55]. Several studies explore the wide range of VR hardware options for studying human behaviour during disasters [27, 28, 47], such as the review by Lovreglio [56]. This review notes that the most suitable hardware configuration is influenced by the research budget and objectives. The review also emphasises the extensive use of VR solutions in studying human behaviour during disasters, with a particular focus on fires in buildings. However, a concern about these studies is whether data collected using VR has ecological validity (i.e. whether people behave similarly in real disasters and VR) [18]. While some studies have attempted to explain this crucial issue [49, 57, 58], more studies comparing real and virtual scenarios are necessary to evaluate the ecological validity of VR investigations quantitatively. Kinateder et al. [18] claim that the most significant advantage of VR is its ability to generate experimental set-ups that are highly immersive, externally valid, highly controlled, and safe. The major drawback is the reduced ecological validity compared to field and case studies. However, several recent studies have provided evidence of the validity of this tool in investigating human behaviour in fires [27].

In the previous paragraphs, some studies regarding factors affecting exit choice have been discussed. A considerable number of studies examined the impact of these factors to independently influence exit or route selection, but only a limited number considered multiple factors simultaneously [40, 59]. No study was identified that analysed the combined effect of social influence and the presence of fire wardens on exit choices.

This section presents the materials and methods used to perform the experiment. Section 3.1 provides a detailed description of the VR experiment designed in this work. Section 3.2 explains the experimental procedure, while Section 3.3 describes the participants involved. Finally, the statistical analysis tools used to investigate the factors that influence participants’ choices are outlined in Section 3.4.

This section provides the results of the exit choice model proposed in Section 4.1. Then, a sensitivity analysis is provided in Section 4.2. Section 4.3 concludes by presenting respondents’ feedback about the VR experience.

The goal of this work is to investigate the exit choice of occupants during fires in an underground metro station. In particular, this research focuses on the influence of two different factors on this choice: the social factor (i.e., the number of people using the exits) and the role-rule factor (i.e., the presence of a fire warden directing evacuees). Further, differences between fire wardens were also investigated using two authority figures in the experiment: the first half of the participants interacted with a metro staff member, and the second half interacted with a firefighter.

In this work, a discrete choice model is developed to forecast exit choices based on the factors mentioned above. The results of the model show that all experimental factors influence the participants’ exit choices. In addition, the results show that participants are more likely to follow instructions from a fire warden with higher authority. The models shown in Table 3 and Table 4 demonstrate that the exit choice decision is the result of a compromise of the factors included in the model. These findings are aligned with previous research in building fire evacuation [31, 33, 35, 40]. However, it is essential to highlight that only a few studies analysed multiple factors [40, 52], and no research was identified that integrated social influence and role-rule factors within a single experiment.

The proposed findings of this study align with the existing literature, indicating signs of convergent validity. Despite the main finding of this study being unique (i.e., the combined effect of the variables listed in the previous sections), some aspects of the results are comparable with prior research, further supporting the convergent validity. In addition, the study measured construct validity by distributing a post-experiment questionnaire to gauge participants'perceived realism levels. The participant’s answers scored the VR methodology highly for fidelity, quality of visualisation, and behavioural validity, indicating a high level of ecological validity.

In collecting behavioural data, it is important to consider the advantages and disadvantages of the research methods and techniques used. The research strategy selected should be a combination of appropriate research methods and data collection techniques, taking into account the research objectives and other boundary conditions such as time and cost, ethical considerations, and experimental control [14]. The main aim is to reach the appropriate ecological validity in which participants exhibit similar behavioural, emotional, cognitive, and psychophysiological reactions in both VR and the real world [71]. Ideally, case studies and unannounced evacuation drills are the best way to achieve ecological validity in emergency response research. These methods allow evacuees to experience real-world scenarios, which makes their behaviour less biased than if they know they are part of an experiment. However, case studies have limitations when obtaining real incident data due to privacy and ethical issues. Even when videos of real incidents are available, researchers have no control over the evacuee sample or the variables affecting their choices. As a result, selected choices can only be inferred from the evacuees'behaviour during the emergency, which increases measurement uncertainty [14].

This work provides several novelties in terms of findings. This study is the first to investigate a unique combination of variables (see Table 1) and their combined effect on exit choice. In fact, while several studies have provided insight into the impact of individual variables, this work represents one of the first instances in which the impact of multiple variables was investigated using discrete choice modelling. As such, this work aligns with the methods used in new studies on exit choice during evacuation [32, 40, 72]. This was made possible by asking participants to choose an exit in multiple scenarios. The results show that the variables influence the choice with different weights: the presence of a firefighter indicating the correct choice has a more significant impact on participants than a metro staff member completing the same action. The following paragraphs discuss the conformity of our findings with the existing literature.

Studies including multiple factors show similar results in terms of social influence, as people are more likely to follow other people to an exit [40, 59]. This similarity was found in some research performed in experimental buildings or tunnels [42, 73]. Another work that shows similar results is by Nilsson and Johansson [15], who demonstrate that social influence is more significant when information is limited. We illustrate this in our work, as most participants choose the exit used by other virtual evacuees when the fire warden is absent and follow the warden’s directions when more information is provided. On the other hand, some contrasts were seen in some studies performed by Haghani et al. [74] and Lovreglio et al. [75], who found that participants mostly preferred less crowded exits. A possible reason for this difference is the distinction between the technologies used to investigate this factor.

Regarding the existing research on the role-rule model, our work supports all of the research analysed in Section 2. All of the studies demonstrate that “the way a person interprets, prepares and acts in the event of a fire is highly dependent on the everyday role that person has adopted and the rules attached to that role” [29]. In particular, this work highlights that people are more likely to trust advice when they have a more favourable perception of the person giving that advice [14, 51]. For example, the model presented in Section 4.1 shows that in the case where the firefighter is present, the probability of choosing the exit indicated is higher.

The findings of this research also underscore the significance of using VR for collecting data on human behaviour during building fires. In our experiment, the alarm signal and the actions of virtual evacuees served as the primary fire cues, enabling researchers to observe participants’ exit choices in a controlled and safe environment. It also enables the collection of precise behavioural and physiological data during controlled simulated events. Participants’ feedback indicated that the simulation of the virtual world was realistic, but some improvements are necessary for future research, for instance, on the realism of the virtual evacuees. In addition, participants do not feel high levels of anxiety, tension or fear, which makes sense as the VR participants do not see fire or smoke, which is present in many real cases [76].

This work has some limitations. For example, the study’s sample is characterised by participants who all live in New Zealand (despite having different nationalities), so it is not possible to consider this sample to be representative of the world’s population. Future studies are necessary to understand whether cultural differences can influence the evacuees’ behaviour.

Another limitation of this study is that some participants did not perceive the virtual evacuees as being realistic, which may have affected their perception of the virtual evacuees'social interaction and biased the results. However, the self-reported ecological validity suggests that participants behaved in the experiment as they would in a real emergency. To overcome this limitation, future research could use the experimental deception technique proposed by Shipman et al. to enhance the realism of virtual evacuees [28]. With the enhancement of computer graphics and AI applications for VR, this limitation will likely become marginal in future studies.

Additionally, the participant’s movement was only virtual and not physical, and the locomotion was controlled using a C# script. This solution was selected to allow the data collection in small rooms where participants were standing still during the experiment. Also, this controlled locomotion solution allowed all participants to have the same experience while reaching position “B” before they made their exit choice. As such, while this approach might have introduced some bias in the participants’ decision-making, it is more practical and ensures a consistent experience across all participants.

Finally, the current study employed a mixed between-subjects and within-subjects design, which, although practical and resource-efficient, introduces potential confounding effects associated with repeated measures. A purely between-subjects design can be used in future studies if learning effects impact the decision-making process. Such an approach eliminates repeated-measure biases, ensuring each participant experiences only one experimental condition, thus improving the reliability and validity of the observed effects. Although this design typically requires a larger sample size and more resources, it can offer more precise attribution of observed effects to specific experimental factors.

This study examined several social factors influencing evacuees’ exit choices during fire emergencies in underground metro stations, focusing on peer social influence, fire wardens’ instructions, and the level of authority of the fire wardens. The findings indicate that the influence of both other evacuees and fire wardens affected decision-making. Participants exhibited a strong inclination to follow instructions given by fire wardens, with this effect being more pronounced when the fire wardens appeared more authoritative.

Furthermore, the proposed discrete choice model provides insight into how a combination of these factors influences the probability of evacuees selecting a particular exit. The findings have significant implications for the field of evacuation research. The results highlight the importance of the information provided by fire wardens, as well as their perceived level of authority during an evacuation. Additionally, the findings underscore the necessity of incorporating both social influence theory and role-rule factors when conducting evacuation simulations.