Content uploaded by Ruggiero Lovreglio
Author content
All content in this area was uploaded by Ruggiero Lovreglio on Jun 20, 2022
Content may be subject to copyright.
Published a: Lovreglio et al. 2022, Exit Choice in Built Environment Evacuation combining
Immersive Virtual Reality and Discrete Choice Modelling, Automation in Construction
Exit Choice in Built Environment Evacuation combining Immersive Virtual Reality and
Discrete Choice Modelling
Ruggiero Lovreglioa, Elise Dilliesb, Erica Kuligowskic, Anass Rahoutib,d, Milad Haghanie
a School of Built Environment, Massey University, Auckland, New Zealand
b Department of Civil Engineering and Structural Mechanics, University of Mons, Mons, Belgium
c School of Engineering, RMIT University, Melbourne, Australia
d Fire safety consulting sprl, Ligny, Belgium
e School of Civil and Environmental Engineering, The University of New South Wales, UNSW Sydney,
Australia
Abstract:
In the event of a fire emergency in the built environment, occupants face a range of evacuation
decisions, including the choice of exits. An important question from the standpoint of evacuation
safety is how evacuees make these choices and what factors affect their choices. Understanding how
humans weigh these (often) competing factors is essential knowledge for evacuation planning and
safe design. Here, we use immersive Virtual Reality (VR) experiments to investigate, in controlled
settings, how these trade-offs are made using empirical data and econometric choice models. In each
VR scenario, participants are confronted with trade-offs between choosing exits that are familiar to
them, exits that are less occupied, exits that are nearer to them and exits to which visibility is less
affected by fire smoke. The marginal role of these competing factors on their decisions is quantified
in a discrete choice model. Post-experiment questionnaires also determine factors such as their
perceived realism and emotion evoked by the VR evacuation experience. Results indicate that none
of the investigated factors dominated the others in terms of their influence on exit choices. The
participants exhibited patterns of multi-attribute conjoint decision-making, consistent with the recent
findings in the literature. While lack of familiarity and the presence of smoke both negatively affected
the desirability of an exit to evacuees, neither solely determined exit choice. It was also observed that
prioritisation of the said factors by participants changed during the repeated scenarios when
compared to the first scenario that they experienced. Results have implications for both fire safety
designs and future VR evacuation experiment designs. These empirical models can also be employed
as input in computer simulations of building evacuation.
Keywords: exit choice; fire evacuation; discrete choice; random utility; virtual reality
1. Introduction
Understanding human behaviour in building fires is an important part of fire safety design [1].
Engineers are tasked with ensuring that a building design provides a sufficient level of safety for its
occupants under a variety of different fire and evacuation scenarios [2]. It is during the development
and assessment of these scenarios that occupant behaviour must be appropriately considered,
requiring theory and data on evacuation likelihood, pre-travel delay times or actions, route choice,
movement attributes, such as walking speeds, and exit choice [3,4]. Further, having a reliable set of
evacuation models and sub-models is fundamental for future automation in fire safety engineering,
e.g., adopting them in BIM-based generative designs [5].
A cohort of data that exist on occupant behaviour during building fires has been collected from field-
type experimental studies [6–9]. These studies collect data on people’s choices made, for example,
during evacuation drills. While post-fire studies offer insights on real-world decisions and behaviours
of evacuees, building evacuation drills provide data primarily focused on movement attributes and
overall congestion points throughout the building. In both cases, researchers have limited
experimental control over the factors under investigation, including the types of evacuees involved or
the type of scenarios that they face (e.g., in terms of the architecture of the space or the level of
crowding that they face, etc.) [10].
Evacuee data can alternatively be collected by laboratory experiments [11–19]. Laboratory data can,
for example, be collected by presenting participants with a hypothetical scenario and asking what they
would do (i.e., their behavioural intention) in that particular scenario [20]. These data are often
captured via survey designs. While surveys give the researcher a high level of control over the factors
under investigation, they have lower ecological validity given that subjects are aware that they are
part of an experiment [20–22]. The use of emerging technologies, like virtual or augmented reality, to
study human behaviour during building fires has increased [23–27]. These technologies allow
researchers to gain a better understanding and observe the decision-making of known subjects in real-
time and even inquire about behaviours in post-experiment surveys [10,28]. Both immersive and non-
immersive simulated environments also allow researchers to safely introduce and test conditions, like
smoke, that would be impossible to achieve in real-world settings as well as training participants in
virtual disasters [29–33]. Research has demonstrated that results obtained in VR settings are
comparable to data from the “real world”; although further validation is required [34–37]. Several
applications have been already used VR to investigate evacuation route choices and behaviour in
different types of buildings showing the advantages and limitations of this emerging technology in
building design and fire safety assessment [38–42].
The aim of this work is to provide new insights on the exit choice behaviour of occupants during fires.
To achieve this goal, we designed a new immersive virtual reality (VR) exit choice experiment to
investigate how certain factors affect exit choice during a fire evacuation. The factors investigated
were the behaviour of other evacuees, the presence of smoke, the distance of an exit and the decision
maker’s familiarity with an exit. VR was used in this study to obtain real-time data on exit choice during
a simulated fire emergency as well as feedback from participants once the experiment was over. These
data can be used by engineers in their fire safety analyses as qualitative inputs into the development
of evacuation scenarios or as quantitative inputs into evacuation models.
2. Exit Choice Literature
A number of studies have been performed on exit choice during building fires [22]. These studies have
identified the multiple factors that can influence evacuee exit choice, including familiarity, social
influence, affordances and design of the built environment, and harmful conditions within the
environment itself (e.g., smoke). For each of these factors, previous research findings will be further
explored.
In fire emergencies, people have a tendency to evacuate via familiar routes and exits [43]. Proulx [44]
found that people were drawn to familiar exits, and even in circumstances where other exits were
available and/or closer, people were unlikely to adopt a new route or exit previously unknown to them
during an emergency. Several studies of actual fire events (e.g., [45,46]), as well as evacuation drills
(e.g., [33]), provide evidence for evacuee tendencies to gravitate toward the familiar. Experiments
performed in a mock furniture store setting found that 71% of occupants chose the door they used to
enter the building for evacuation as well [47]. Also, studies using VR settings of a hotel [48] and a
museum [19] demonstrated that familiarity influenced exit choice. In the first case, people were more
likely to evacuate via the main entrance unless provided with additional information about the
buildings (in this case via signage), and in the second case, familiarity was enhanced when others also
used the same exit and reduced when more people left through the other exit.
Other occupants can also influence an evacuee’s choice of exit, and this influence can vary with the
number of people present in the scenario. In very large crowds, 75 to 150 people as an example,
participants were not observed following “the crowd” [49]. However, observations of studies involving
a much smaller number of participants provide evidence of this trend. Limited to around 50
participants on a virtual train platform, Lin et al. [50] found that uneven splits of the other evacuees
influenced participants to follow the majority; and this finding held true across multiple cultures.
Three studies with similar designs investigated exit choices with crowd sizes up to 20 people. In the
first two studies [12,14], participants viewed videos of the scenarios via an online survey and in the
other, immersive VR technology was used [51]. In both cases, participants were confronted with a
hypothetical situation and asked to choose between two exits. Watching the online videos,
participants favoured less crowded exits except when all virtual agents used the same exit. In the VR
study, participants were likely to follow smaller crowds (~10); and in cases of larger crowds (~20), only
if the doors were wider in size.
Experimental studies with and without VR tested the influence of small numbers of people on exit
choice behaviour. Investigating wayfinding from an enclosed room into a corridor, Zhu et al. [52]
tested the influence of 1-3 actors instructed to choose a route that opposed the signage and found
that in one-actor conditions, evacuees followed the actor’s behaviour. Also, studying social influence
in tunnel fires, studies have found that participants were more likely to move to the emergency exit
[53] and along the shorter path to the exit [54] when a virtual agent within the simulation did so.
Finally, information and conditions within the built environment can also influence exit choice. Nilsson
et al. [55], adapting Gibson’s Theory of Affordances [56] to building fire emergencies, note that the
attractiveness of an exit is related to how well it affords egress. One of these factors is functional
affordance, in that the exit allows for safe and effective movement to safety (e.g., is clear of smoke).
While research provides evidence that people do walk through smoke under varying obscuration
levels [57–59], a study of route choice behaviour in a virtual building showed that heavy smoke can
reduce the use of evacuation shortcuts. It should be noted; however, that the presence of others
evacuating via that shortcut increased its use rate [60].
While a notable amount of work has been done looking individually at how these factors influence
exit or route choice, only a few studies included multiple factors (e.g., [51]). Additionally, no study was
found that included familiarity, social influence, distance and smoke conditions in the same
experiment to investigate their combined influence on exit choice. Based on the literature and the
proposed factors of study in this work, we hypothesise that while individual relationships between
exit choice and our four factors of interest are often found, we expect that exit choice is the result of
a combination of factors. Additionally, as a new contribution to our field, we aim to test whether
results will change as participants’ familiarity with the experiment itself (i.e., learning effect) increases.
3. Material and Methods
This section provides a description of the VR experiment designed in this study (Section 3.1), the
experimental procedure (Section 3.2), the participants who took part to the experiment (Section 3.3)
and the statistical analysis tool used to investigate the participants' choices (Section 3.4).
3.1 VR Experiment Design
During the VR experiment, a participant is placed within a virtual room where he/she has to choose
an exit to use to leave the room during a simulated fire emergency. As shown in Figure 1, the room
has 3 exits (A, B, and C) from which the participant can choose. All the exits have exit sign on top of
them indicating that they were available paths for evacuation (see Figure 2). The geometry illustrates
where the decision-maker was located at the start of the experiment (i.e., starting point), while the
red square identifies the location where, when reached, the first event in the experiment was
triggered (i.e., the fire evacuation emergency). Once a participant exited via one of the three provided
exits, the experiment ended.
Within the experiment, the available exits differed by a range of physical, social and individual factors.
Depending upon the scenario, the exits differed by their physical distance away from the decision-
maker and the presence of smoke (entering from the top of some of the exits). Additionally, the exits
presented to the participants differed by the number of evacuees who were already using each exit
(i.e., social factor) and how familiar the exit was to the participant (i.e., individual factor). Familiarity
was defined as an exit that the participant had used in the past; and since one of the exits (Exit A) was
located along the path participants needed to travel to reach their starting point in the experiment
(see the red square in Figure 1), it was assumed to be an exit that was more familiar than the others.
In the next section of the paper, the following abbreviations are used to identify the variables under
investigation:
- NP: Number of people using an exit;
- DIST: Distance of the participant from an exit;
- SMOKE: Presence of smoke;
- FAM: Familiarity of the participant with the exit.
Each of the four variables above has a number of dimensions that can be varied within the experiment.
For simplicity purposes, the following assumption was made for the NP variable: the number of
evacuees leaving the room by each exit (NP) can be equal to 0, 1, 5 or 10 (4 levels/dimensions of the
variable per exit). The maximum number of NPCs was selected to avoid that the crowd density
affected the NPCs speed having a target of a max density below 0.54 persons/m2 according to the
SFPE curve published in [61]. Additionally, the distance of the decision-maker from the exits was
defined using the geometry of the virtual environment. While the location of Exit A was kept constant
(6.0 m), the location of Exits B and C varied across scenarios. See Figure 1 for the two positions possible
for both Exits B and C.
The distance of the participant from Exit B was either 3.6 m or 5.6 m, while the distance to Exit C was
3.0 m or 4.6 m. For each exit, smoke (SMOKE) was either present, entering from the top of the door
during the scenario or not. The value of this variable was set to 1 when smoke was present and 0
otherwise. The final constraint for the scenario development was that participants were only familiar
with Exit A as they accessed the room only using this door (i.e., 1 familiar and 0 non-familiar) at the
beginning of the experiment. The summary of the levels for each variable is shown in Table 1.
Figure 1 – Geometry of the virtual environment (The distance units are in mm when not specified)
Table 1 – Levels and values for each variable. Note: The values of the variable are in parenthesis
Variable
Levels and Values
Exit A
Exit B
Exit C
NP - Number of
people using an exit
4
(0, 1, 5, 10)
4
(0, 1, 5, 10)
4
(0, 1, 5, 10)
DIST - Distance of the
participant from an
exit
1
(6.0 m)
2
(3.6 m, 5.6 m)
2
(3.0 m, 4.6 m)
SMOKE - Presence of
smoke
2
(1, 0)
2
(1, 0)
2
(1, 0)
FAM - Familiarity of
the participant with
the exit
1
(1)
1
(0)
1
(0)
Due to the number of variables (4), the number of exits (3), and the possible dimensions of each
variable, a high number of scenarios were available to investigate exit choice. The overall number of
all the possible scenarios was 2048 (= 43 x 22 x 23 x 13). To identify the most useful scenarios to
investigate exit choice, the Efficient Design approach in experimental design was used. The Efficient
Design is a fractional factorial design used to identify an optimal subset of scenarios among all possible
ones with the aim of maximising the amount of elicited information given a set sample size. In other
words, compared to more traditional methods such as orthogonal design, efficient designs allow the
analyst to obtain reliable estimates of the model coefficients with smaller sample sizes. This approach
represents one of the best solutions when there is a large number of variables under investigation as
well as when there is prior knowledge on how factors can affect the decision [62–64]. The Efficient
Design used in this work is based on the minimisation of the D-error metric, which is the determinant
of the asymptotic variance-covariance matrix to the power of 1/K, where K is the number of the
parameters to estimate. In other words, the D-error metric is related to the p-values of the parameters
to estimate with the data collected during the experiment. In fact, p-values are calculated using the
variance matrix, which is the diagonal elements of variance–covariance matrix [62]. The Efficient
Design approach generates an optimal solution when the researcher knows the expected values of
the parameters which will be estimated with the experimental data. In this work, we used as priors
the exit choice results published in [14].
In this work, we used Ngene [64] to run the Efficient Design, and to select the best eight scenarios to
use for the exit choice experiments. The experimental scenarios are illustrated in Table 2.
Table 2 – Experimental scenarios identified using the Efficient Design (NP: Number of people using
an exit; DIST: Distance of the participant from an exit; SMOKE: Presence of smoke; FAM: Familiarity
of the participant with the exit)
Scenario
Exit A
Exit B
Exit C
NP
DIST
SMOKE
NP
DIST
SMOKE
NP
DIST
SMOKE
1
0
6 m
0
10
3.6m
1
5
4.6m
1
2
5
6 m
1
0
5.6 m
1
10
3.0 m
0
3
1
6 m
1
1
5.6 m
0
10
3.0 m
0
4
10
6 m
0
0
3.6 m
0
1
4.6 m
1
5
10
6 m
0
1
3.6 m
1
0
4.6 m
0
6
5
6 m
1
10
5.6 m
0
0
3.0 m
1
7
1
6 m
1
5
5.6 m
0
1
3.0 m
0
8
0
6 m
0
5
3.6 m
1
5
4.6 m
1
The scenario specifications in Table 2 were used to build the eight immersive VR experiences. These
were developed using the Unity game engine [65]. The geometry of the digital building was created
within Unity itself, and the furniture and doors were downloaded from the Unity asset store. The other
evacuees in the room were simulated with Non-Playable Characters (NPCs). NPCs were developed in
Adobe Fuse and animated using Mixamo [66]. The pipeline used to develop the NPCs is the one
proposed in [33,67]. The smoke entering the room through the door during the fire emergency was
generated using particle systems. Figure 2 shows some screenshots of the virtual experience. The VR
experience was developed for a VIVE Pro headset by using the Steam VR packages available in the
Unity asset store.
The application allowed the participants to visualise the digital scenario using the headset and to walk
through it by walking in the physical space of 7m x 8m as shown in Figure 3. The movement of the
participants in the physical space was tracked using four base stations located at the vertex of the
rectangular tracking area (7 m x 8 m), as illustrated in Figure 3. This tracking area was large enough to
allow participants to walk throughout the entire area highlighted in green in Figure 1.
Figure 2 – Screenshots of the virtual experience showing Exits A (in the front) and B (on the right).
Figure 3 – Physical space used to carry out the VR experiment
3.2 Experimental Procedure
The experiment took place in June 2019 at the Albany Campus of Massey University (New Zealand).
The experiment included the following steps. All the participants were first asked to read and sign an
approval form before taking part in the experiment, for ethical considerations and to make sure that
none of them had any medical conditions that prevented them from taking part in the experiment.
Then, a pre-experiment survey was administered to collect participants’ demographics and their
backgrounds regarding virtual reality and fire emergencies. After completing the questionnaire, the
participants were asked to stand nearby the white door in Figure 3 and to wear the VR headset.
Each participant was asked to walk through an initial scenario that included only the geometry of the
digital building used in the experiment. They were allowed to walk in the green areas shown in Figure
1 without walking through Exits B and C. This scenario gave participants the possibility to familiarise
themselves with the environment and the VR navigation system before being immersed within the
fire scenarios. As such, each participant used Exit A to enter in the virtual room. After walking in this
room for 30-40s, the participant was asked to return to the starting position using Exit A (see Figure
1). This part of the experiment gave the participants the possibility to become familiar with Exit A and
the environment behind it. Further, the small room where they started the experiment had a glass
door (see Figure 1) which allowed them to see a clear safe evacuation access to the outside of the
building. It is worth highlighting that exit familiarity is referring to the familiarity with the geometry of
the digital building and that such an approach was proposed in a previous study by Kinateder et al.
[19].
Each participant was randomly assigned to four of the eight scenarios. The participants did not
experience all the scenarios to avoid fatigue as the full experimental session aimed at not taking more
of 30 minutes of their time. In each scenario, each participant was asked to enter the room using Exit
A and to reach the red target shown in Figure 1 to attend a virtual meeting. The participants were not
aware that the real intent of the experiment was observing their exit choice during an evacuation, as
they were told that the VR experience was to demonstrate how VR can be used to run meetings. This
allowed researchers to deceive the participants as they were not aware (at least for the first
experiment) that a fire emergency would occur. A few seconds after the participants reached the red
target, the fire alarm was activated, and the NPCs stopped previous activities immediately and started
looking around before moving to a pre-computed exit following the numbers defined by Table 2. The
NPC located in the environment in order for them to reach the closest exit (i.e., the exit assigned to
them). Before the alarm, there was no interaction between NPCs and the participants as this might
have created a social bonding between participants and NPCs in the scene. The pre-movement time
varied for each NPC between 3 and 9 seconds. These reaction times were randomly assigned to the
NPCs to avoid instances where NPCs could block each other from accessing the exits. When the alarm
started, the smoke started entering the room through some of the doors depending on the scenario
number (see Table 1). The experiment ended when the participants reached one of the three exits. In
particular, when the participant is about to cross Exit B and C, the simulation stops so they cannot get
familiar with the building geometry behind these two doors. This design solution was adopted to
ensure that the participants were only familiar with the geometry behind Exit A. Each exit choice made
by the participants was recorded and stored in a local database (i.e., csv file). Similarly, their navigation
path was monitored during their evacuation travel phase for each scenario.
After the VR experience, each participant was asked to fill a final questionnaire to collect feedback
about the realism of the experiment, how easy it was to take part in the VR experience, their emotional
states, the urgency perception and the validity of their behaviour. This was done by using seven-point
Likert scale questions (-3= strongly disagree and +3= strongly agree). Each participant was asked to
express their agreement/disagreement with the following statements:
- The virtual world was adequate/realistic (Realism 1);
- The virtual fire scenario was adequate/realistic (Realism 2);
- The interaction with other virtual people was adequate/realistic (Realism 3);
- I found running this VR scenario easy (Usability);
- This experience makes me feel scared/fearful (Emotion 1);
- Overall, this experience makes me feel tense/nervous (Emotion 2);
- Overall, this experience makes me feel anxious (Emotion 3);
- I felt the urgency to act/do something during the fire emergency (Urgency);
- I would act the same way in real life during a fire emergency (Validity).
Finally, each participant was asked to indicate which factors affected his/her choice with four open-
ended questions (i.e., one question per choice). An open-ended response option was selected in this
case to avoid influencing participants’ answers with fixed options.
3.3 Participants
Most of participants were recruited through email, social networks or by other participants. Flyers
were also distributed on the campus, and advertisements were published in the main buildings of
Massey University. A total of 86 participants took part in the experiment. Most of them were staff or
students of Massey University. The sample included 38 female and 48 male participants whose ages
ranged from 18 to 66 years. The average age was 28.8 years; the standard deviation was 9.8 years
while the 25th and 75th percentiles were 22 and 32 years, respectively. Participants’ ethnicities varied;
however, more than 47% were of Asian descent, and more than 30% defined themselves as European.
The remaining 22 participants identified themselves as Middle Eastern (7%), African (5%) and from
another ethnicity (10%).
3.4 Statistical Analysis
The choice data collected during the experiments were analysed using random utility models. In this
work multiple forms of multinomial logit model were estimated. Random utility models rely on
multiple assumptions [68,69]. The first assumption is that a decision-maker q assigns to each available
choice alternative i a utility ,. The utility is defined by a deterministic component , and an error
component , (Equation 1).
,= ,+ ,
Equation 1
In this work, we assume that the deterministic component has a linear specification (Equation 2).
,= �,,,
Equation 2
where ,, are the known values of the factors j perceived by the decision-maker q affecting the
choice for the alternative i; and , are parameters weighting the decision-makers’ preferences
related to the factors j (i.e., the parameters to estimate). It is possible to demonstrate that, when the
error components are distributed as Extreme value Type I with variance 2/6 and these distributions
are independent and homoscedastic, the probability that the decision-maker q selects alternative i
has the close formulation in Equation 3 (i.e., multinomial logit formulation).
,=
exp (
,
)
∑exp (
,
)
Equation 3
Equation 3 can be then used to build a likelihood function which is used to estimate the ,
parameters by finding the combination of these parameters that maximises the likelihood function
[69,70]. In this work, we estimated the multinomial logit models using the "mlogit" package available
in R Studio [71].
Finally, the participants’ responses to the final questionnaire assessing different aspects of the VR
experience (see Section 3.2) were analysed by using boxplots to assess the average response and the
variance of the answers. The open-ended responses on the factors affecting each of the four
participants’ choices were instead coded as follows: six binary variables were created representing
the factors affecting choices:
- Follow NPCs;
- Avoid NPCs;
- Smoke;
- Distance;
- Familiarity;
- Other.
For each participant, a score of one was given for each of these factors if mentioned in their response.
4. Results
This section presents the results of the exit choice models proposed in this study in Section 4.1 while
Section 4.2 provides the results of the participants’ feedback regarding the virtual experience.
4.1 Exit Choice Models
In this work, we proposed two multinomial logit model formulations using the 344 choice observations
collected from the experiment described in Section 3.
In Model 1, we estimated the parameters , weighting the impact of NP, DIST, SMOKE and FAM using
the specification in Equation 4. The parameters were all treated as generic across the three
alternatives, meaning that we did not estimate alternative-specific parameters.
U
i
= β
np
NP
i
+ β
dist
DIST
i
+ β
sm
SMOKE
i
+ β
fam
FAM
i
i = A, B, C
Equation 4
The estimated parameters for Model 1 are shown in Table 3. The model shows that all the parameters
were statistically different from zero, having their p-values below the level of significance of 0.05. The
model also shows that participants tended to select the exit already selected by the NPCs exhibiting
the impact of social influence or following behaviour. Further, the participants were more likely to
choose the familiar Exit A while they were less likely to choose distant exits or exits having smoke.
Table 3 – Estimated parameters for Model 1 (Table 2 – Experimental scenarios identified using the
Efficient Design (np: Number of people using an exit; dist: Distance of the participant from an exit;
sm: Presence of smoke; fam: Familiarity of the participant with the exit)
Variable
Estimate
Std Error
z-value
P-value
βnp
0.076
0.015
5.030
0.000
β
dist
-0.378
0.079
-4.771
0.000
βsm
-1.765
0.161
-10.980
0.000
βfam
0.795
0.206
3.864
0.000
Model 2 was estimated to investigate the impact of participants’ unawareness of the fire evacuation.
In fact, as explained in Section 3.2, the participants were unaware that the first scenario would require
them to evacuate, while they were aware of the evacuation when they repeated the experiment for
the remaining 3 times. To assess if participants behaved differently when evacuating in the first
scenario compared with the three that followed, we used a binary variable 1, which equals 1 if the
choice was made during the first evacuation or 0 otherwise. The specification of Model 2 is shown in
Equation 5.
U
i
= (β
np
+ C
1
×β
np1
) NP
i
+ (β
dist
+ C
1
×β
dist1
) DIST
i
+(βsm + C1×βsm1) SMOKEi + (βfam + C1×βfam1) FAMi
i = A, B, C
Equation 5
The estimated parameters for Model 2 are shown in Table 4. The model shows that most of the
parameters are statistically different from zero, having their p-values below the level of significance
of 0.05. In line with Model 1, Model 2 shows that all the variables under investigation had an impact
on the decision-making process. However, this second model shows that participants were more likely
to choose exits used by other NPCs in their first choice as np1 is positive and significantly different
from zero (i.e., np+ np1> np). The model also shows that during the first evacuation the
participants were still negatively affected by the smoke. In fact, they were less sensitive to it in the
first evacuation as βsm1 is positive and the sum of sm and sm1 is negative (i.e., sm+ sm1<0).
Table 4 – Estimated parameters for Model 2 (np: Number of people using an exit; dist: Distance of
the participant from an exit; sm: Presence of smoke; fam: Familiarity of the participant with the exit)
Variable
Estimate
Std Error
z-value
p-value
βnp
0.041
0.018
2.267
0.023
βdist
-0.439
0.094
-4.688
0.000
βsm
-2.305
0.214
-10.750
0.000
βfam
0.735
0.248
2.968
0.003
βnp1
0.192
0.046
4.130
0.000
βdist1
0.218
0.200
1.088
0.277
βsm1
1.781
0.364
4.898
0.000
βfam1
0.413
0.522
0.790
0.429
4.2 Sensitivity Analysis
A sensitivity analysis is conducted to show how the probability of choosing an exit can be affected by
the variables investigated in Model 2 illustrated in Table 4. To simplify this analysis, we consider a
scenario including only two exits (Exit A and Exit B).
To investigate the combined impact of the number of people using an exit (NP) and familiarity (FAM)
in the exit selection, we assume that an evacuee is equally distant from Exit A and Exit B, there is no
smoke located at either exit, and there are 5 evacuees using Exit B. Figure 4 shows the change of
probabilities of the evacuees choice of Exit A for different numbers of people using this exit (i.e., NP
for Exit A varies from 0 to 10 people). This investigation is run for three different scenarios: the
evacuee is familiar only with Exit A (Fam A), only with Exit B (Fam B) and with both exits (Fam A&B).
Figure 4 clearly shows that an increase in an increment of NP for Exit A leads to an increase in the
probability of this exit of being selected. As an example, in the scenario in which the evacuee is familiar
with both exits (Fam A&B), the probability of choosing Exit A increases from 0.23 to 0.76. Further, the
familiarity of an exit generates a substantial shift in the probability curves. For instance, in the case of
NP is equal to 5 for Exit A, the probability of choosing Exit A can range from 0.32 to 0.68 depending on
the familiarity conditions.
The combined effect of distance (DIST) and familiarity (FAM) is also investigated in this section by
assuming that there is no smoke located at either exit and there are equal numbers of people using
both exits. The probabilities in Figure 5 are estimated assuming that an evacuee is 3 m distant from
Exit B while the distance of Exit A varies from 0 m to 6 m. Once again, the analysis is run for three
different scenarios: the evacuee is familiar only with Exit A (Fam A), only with Exit B (Fam B) and with
both exits (Fam A&B). Figure 5 illustrates how distance has a strong impact on the probability of Exit
A to be selected. In fact, considering the scenario Fam A&B, it is possible to observe that the
probability of choosing Exit A decreases from 0.79 to 0.21 as distance increases. In line with the first
sensitivity analysis, the familiarity of an exit generates a substantial shift in the probability curves in
this case.
Figure 4 – Combined effect of NP (Number of people using an exit) and FAM (Familiarity of the
participant with the exit) on the exit choice
Figure 5 – Combined effect of DIST (Distance of the participant from an exit) and FAM (Familiarity of
the participant with the exit) on the exit choice
It is worth highlighting that Figures 4 and 5 provide examples of possible analyses that can be carried
out using the model in Table 4. The aim in this section is to demonstrate simple implementations of
the proposed model that can be used to illustrate how particular factors affect the probabilities of
evacuees choosing a specific door.
4.3 Respondents’ Feedback
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
012345678910
Probability to choose Exit A
NP for Exit A
Fam A Fam B Fam A&B
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0123456
Probability to choose Exit A
Distance of Exit A
Fam A Fam B Fam A&B
After the VR experience, participants were asked to provide feedback on the realism of the
experiment, the engagement levels of the experiment, the ease with which they were able to take
part to the VR experiment, and their emotional states (see Section 3.2). The feedback provided by the
respondents is shown in the boxplots in Figure 6. The realism and the usability scores of the VR
application are shown in Figure 6.a. The realism items demonstrate that participants found both the
virtual world and the virtual fire realistic (see Realism 1 and 2). However, they stated that the NPCs
were not as effective in terms of realism ratings. Further, Figure 6.a highlights that participants found
the VR application easy to use.
Figure 6.b illustrates the scores for the emotional states of the participants, with Emotion 1
representing feelings of fear, Emotion 2 nervousness, and Emotion 3 anxiety. The majority of the
sample did not report feeling any negative emotions. However, most of the participants reported that
they felt urgency associated with taking action during the fire emergency and that they would behave
in the same way in real life during the fire emergency.
(a)
(b)
Figure 6 – Participants’ scores on (a) realism and usability, and (b) their emotional state and
perceived validity.
4.4 Reported Influential Factors of Exit Choice
After the VR experience, participants were asked to identify the factors affecting each of their four
choices. The number of factors affecting the four choices is presented in Table 5. From this table, it is
possible to observe that the large majority of the sample reported being affected by a single factor for
all the four choices and that 20%-26% of the sample used a combination of two factors to select the
exit. Only a few participants made their choices using three factors.
Table 5 – Number of factors affecting the choices
One
Factor
Two
Factors
Three
Factors
Choice 1
66
19
1
77%
22%
1%
Choice 2
67
17
2
78%
20%
2%
Choice 3
65
20
1
76%
23%
1%
Choice 4
64
22
1
74%
26%
1%
Table 6 illustrates which factors affected participants’ choices. This table indicates that the most
popular factor affecting the first choice was following NPCs (42%) while the second and third most
popular factors were avoiding smoke (38%) and selecting the familiar exit (20%). Table 6 shows how
these percentages change for the remaining choices 2-4; i.e., when participants were presumably
more familiar with the experiment. In these cases, the most popular choice was smoke avoidance
(49%-67%), while the number of participants who relied on following NPCs (16%-13%) decreased. This
is in line with the modelling results reported in Table 4, showing that participants’ decisions in the first
experiment (first choice) were more sensitive to the behaviour of NPCs (βnp1>0 and p-value <0.05)
and less sensitive to the presence of smoke (βsm1>0 and p-value <0.05).
Table 6 – Factors affecting the choices
Follow
NPCs
Avoid
NPCs
Smoke Distance Familiarity Other
Choice 1
36
3
33
11
17
7
42%
3%
38%
13%
20%
8%
Choice 2
14
8
42
24
9
9
16%
9%
49%
28%
10%
10%
Choice 3
10
10
56
16
5
10
12%
12%
65%
19%
6%
12%
Choice 4
11
9
58
20
6
6
13%
10%
67%
23%
7%
7%
5. Discussion and Conclusions
Models were developed that allowed for the prediction of exit choice based on four factors: the
familiarity of the person with their exit of choice, the number of people using various exits, the
presence of smoke at particular exits and the exit distance. Overall, the models’ results showed that
participants were influenced by all of the factors under investigation. Each of these results individually
aligns with previous research in building fire evacuation. The proposed model in Table 4 shows that
the exit choice decision is the result of a trade-off of the factors included in the model. This trade-off
is clearly visible in the examples illustrated in Figures 4 and 5, showing how the decision is a combined
effect of social influences, exit distance and familiarity. Further, the model shows how this trade-off
changed during the experiment as a “learning effect” was observed through the choices made by the
participants. These aspects are discussed in detail in the following paragraphs.
This work has several novelties from a methodological viewpoint and in terms of findings. Firstly,
although immersive VR is becoming an established tool to investigate human behaviour in disasters
[72], this study investigates a unique combination of the variables and their interactions. For instance,
the interaction between familiarity, social presence, smoke and their conjoint effect on decision-
making is for the first time studied in this work. This was possible by using a state-of-the-art
experimental design tool, namely Efficient Choice Experiment Design. This technique helped us
include several factors with many levels (see Table 1), maintaining a reasonable number of scenarios
(see Table 2). Secondly, this study investigates, for the first time, the learning effect when participants
are asked to make a choice in multiple scenarios. This was assessed using a qualitative approach (see
Table 5) and through the modelling solution shown in Table 4. The alignment of our findings with the
existing literature as well as their implication for future research, is discussed in the following
paragraphs.
Our results involving familiarity provide support for Sime’s affiliative model [43] and findings from
studies of actual fire events [45], mock evacuation drills [47], and VR studies of building evacuations
[19,48]. Also in line with Kinateder et al. [19] are our results (Figure 4) that show that the probability
of someone using a familiar exit is enhanced with an increased number of other evacuees also using
that exit. It should be noted that familiarity in this experiment was represented by a participant given
the opportunity to virtually walk through that exit prior to the fire evacuation beginning. What that
suggests is those building occupants may only need to have engaged with an exit once or twice to then
feel greater confidence in using it to reach safety during a fire.
Our results also support much of the research on social influence with smaller crowds (~20 people or
less), but not all. The findings from this study are similar to those from experimental building and
tunnel studies with 1-3 participants, where people were more likely to follow another person to an
exit (even if that exit was not identified as such) [52–54]. Additionally, similar to Kinateder and Warren
[51] where participants were likely to follow crowds of 10 or less, this study found that people were
likely to follow crowds as large as 20 people. However, our study contrasts with Lovreglio et al. [12,14]
who, via online surveys using videos, found that participants mostly preferred less crowded exits. A
similar contrast is observed in Haghani & Sarvi’s [49] study conducted in field-type (non-virtual)
settings that indicated participants tend to avoid movement in the same direction chosen by the
majority. A possible reason for differences across these studies is the type of technology used to
simulate the fire environment. While studies have shown that VR experiments like this one can provide
similar results to those found in real-world settings [34,48,73], little work has been performed on the
ecological validity of video stimuli (i.e. non-immersive experiments). As such, future studies are
necessary to compare non-immersive and immersive VR settings.
This study also highlighted the impact of smoke on exit choice. Similar to [60], we found that the
presence of smoke reduced the likelihood of choosing an exit. While previous research has shown that
in some cases, people will walk through smoke, in this case, this study demonstrated that when people
had other, seemingly safer options, they took them. This finding also suggests that smoky conditions
reduce an exit’s functional affordance, i.e., making it less likely to be perceived as an effective way to
reach safety [56] and in turn, less likely to be used for evacuation.
Given the alignment of the proposed findings with the existing literature, it is possible to state that
this study demonstrates multiple signs of convergent validity. While the main finding of this work (i.e.,
the combined effect of the variables listed earlier) is unique to this study, there are marginal aspects
of the reported results that are contrastable with the existing body of evidence in the literature
[8,14,18,19,49,51,74,75]. As such, those can be regarded as signs of convergent validity for the
findings. Further, this study attempts to measure construct validity by administering a post-
experiment questionnaire to identify participants’ perceived level of realism. Finally, with respect to
the matter of ecological validity within the realm of this methodology (VR) the experiment scores high
on fidelity and quality of visualisation and self-reported behavioural validity (see Figure 6) as reflected
in participant responses to the post-experiment questions.
This work also assesses the impact of asking the same participants to make choices in different
scenarios. The technique of exposing the same participants to multiple-choice scenarios is a standard
practice in transportation studies and was also adopted for exit choice studies (e.g., [12,14,64]);
however, the behavioral impact was not investigated previously. Our Model 2 in Table 4 is an attempt
to assess if participants change their choice strategy by comparing the first choice they made with the
remaining choices. The model shows that there was a significant change in the way other people and
smoke affected the decisions. This was also confirmed by the qualitative analysis discussed in Section
4.4. The results in Table 6 highlight the impact of replication in experimental studies. These findings
show that in their first scenario, when they were unfamiliar with the experiment, participants’ exit
choices were mainly affected by others in their environment. However, by the 2nd through 4th
scenarios, the presence of smoke was more often cited as their factor of influence. In line with Model
2, these results suggest that at first, when participants were beginning to become familiar with the
experiment, they relied more on others to help with their decision-making. As they became more
familiar with the experiment and its scenarios, it is possible that participants focused more on the
smoke as a hazard from which they needed to retreat. This “learning effect” has also been observed
and investigated in other contexts of choice experiments (i.e., in more conventional non-VR formats).
For example, Brouver et al. [76] reported in their empirical investigation that in a repeated choice
experiment, learning did occur but that it did not exert any significant impact on econometric
estimates. It would, however, be fair to assume that this observation may be case specific and not
necessarily generalisable to all contexts and forms of choice experiments. It only shows that the effect
of learning cannot be ruled out. Noting that, in the case of our experiment, participants were unaware
of an imminent evacuation for the first scenario (e.g., see the deception strategy in Section 3.2), it is
possible to argue that the first choices made by the participants have higher levels of ecological
validity.
Results from this study also highlight the value that VR can bring to data collection of human behavior
in building fires. Participant feedback rated the simulation of the “virtual world” and the fire
environment as realistic. In addition to usability being rated as high, so were perceptions of urgency
and the likelihood that they would act similarly in an actual fire emergency. Although these are
participant perceptions and we lack actual data with which to compare, this feedback highlights the
potential for continued use of VR in future behavioural studies. And as technology becomes more
sophisticated and the virtual settings can increase in realism, researchers will be required to weigh
the value of critical data with ethical practices when exposing participants to a scenario that might
upset them, as discussed in [77].
The data presented in Table 5 shed light on an important phenomenon not necessarily captured by
our logit models: an assessment of how many of the four factors played a role on each round of exit
choice. In each scenario, the participants’ environment contained a combination of social influence,
familiarity, exit distance and smoke factors, and one, a few, or all of these factors could have
influenced their decision. Table 5 shows that a majority of participants identified only one factor as
influential on their decision, with a smaller percentage listing two, suggesting that even with multiple
factors present in the same environment, individuals may focus only on one to make their decision.
However, the leading factors can be different among participants, as illustrated in Table 6, showing
the heterogeneity of choice strategies. Decision scientists have argued that stressors, such as time
pressures or uncertainty, can narrow a person’s perceptive field, causing them to pay attention to a
select number of cues from their environment [78]. While this theory may hold here, additional
research is required to further explore the dynamics at play.
This study has some limitations. The sample size of this study is smaller than those used in previous
studies online studies. Previous online studies involved several hundreds of people or thousands of
participants (e.g. [12,14]). On the other hand, the sample size of this study is similar to the ones used
in previous immersive VR studies [38,79,80]. Immersive experiments cannot reach the sample size of
online studies and this has an impact on the possibility of investigating factors that have a small effect
size [81]. On the other hand, immersive experiments provide higher ecological validity, which is
fundamental to having accurate predictions [20,23]. In addition, the participants of this study were all
living in New Zealand. However, the use of the Efficient Design mitigates this limitation as it allows
researchers to obtain reliable estimates of the model coefficients with smaller sample sizes.
Because of the sample size limit, this work does not investigate heterogeneity by using random
parameter models or more advanced Bayesian hierarchical models as proposed recently by Song and
Lovreglio [16]. f Further the specification used in Equation 5 and the existing sample allowed the
investigation of the change in choice behaviour which has a great implication for future evacuation
studies experiments, as discussed previously in this section.
Another limitation of this work is that the virtual room used in this study was designed for the specific
purpose of this experiment. In fact, although the room is for a meeting, it does not include any chairs
or other pieces of furniture. Obstacles would have had an impact on the definition of the distance of
the three doors and its specification in the models. Despite that, the results in Figure 6 shows that the
participants believe that the realism of the virtual environment was high (see Realism 1). Further, the
smoke was not physically simulated as the same particle system was locate on top of the exit to
represent the smoke coming out from the doors. Investigating the impact of smoke density is
fundamental for future studies as visibility can play a key role in redirecting evacuees [82].
The final limitation of this study is that several participants did not find the NPCs realistic enough. This
might bias the way participants perceived NPCs and the social interaction observed in this study as
predicted by the conceptual Threshold Model of Social Influence [83]. On the other hand, the self-
reported ecological validity (see Figure 6.a) indicates that participants likely behaved in the
experiment as they would in a real emergency. NPCs realism can be enhanced in future studies using
the experimental deception proposed by Shipman et al. [36], who increased the believability of the
NPCs in their terrorism experiments by making the participants believe that the NPCs, they view in
the VR scenarios, were other participants taking part in the same experiment
References
[1] E.D. Kuligowski, Human behavior in fire, in: SFPE Handbook of Fire Protection Engineering,
Springer New York, New York, NY, 2016: pp. 2070–2114.
https://doi.org/https://doi.org/10.1007/978-1-4939-2565-0_58.
[2] D. Nilsson, R. Fahy, Selecting Scenarios for Deterministic Fire Safety Engineering Analysis: Life
Safety for Occupants, in: SFPE Handbook of Fire Protection Engineering, Springer New York,
New York, NY, 2016: pp. 2047–2069. https://doi.org/10.1007/978-1-4939-2565-0.
[3] S.M. V. Gwynne, K.E. Boyce, Engineering Data, in: SFPE Handbook of Fire Protection
Engineering, Springer New York, New York, NY, 2016: pp. 2429–2551.
https://doi.org/10.1007/978-1-4939-2565-0_64.
[4] R. Lovreglio, E. Kuligowski, S. Gwynne, K. Boyce, A pre-evacuation database for use in egress
simulations, Fire Safety Journal. 105 (2019) 107–128.
https://doi.org/10.1016/J.FIRESAF.2018.12.009.
[5] R. Lovreglio, P. Thompson, Z. Feng, Automation in Fire Safety Engineering Using BIM and
Generative Design, Fire Technology. 58 (2022) 1–5. https://doi.org/10.1007/S10694-021-
01153-7.
[6] M. Haghani, M. Sarvi, Following the crowd or avoiding it? Empirical investigation of imitative
behaviour in emergency escape of human crowds, Animal Behaviour. 124 (2017) 47–56.
https://doi.org/10.1016/j.anbehav.2016.11.024.
[7] A.U.K. Wagoum, A. Tordeux, W. Liao, Understanding human queuing behaviour at exits: an
empirical study, Royal Society Open Science. 4 (2016). https://doi.org/10.1098/RSOS.160896.
[8] M. Haghani, M. Sarvi, Z. Shahhoseini, Evacuation behaviour of crowds under high and low
levels of urgency: Experiments of reaction time, exit choice and exit-choice adaptation, Safety
Science. 126 (2020) 104679. https://doi.org/10.1016/J.SSCI.2020.104679.
[9] S. Heliövaara, J.M. Kuusinen, T. Rinne, T. Korhonen, H. Ehtamo, Pedestrian behavior and exit
selection in evacuation of a corridor – An experimental study, Safety Science. 50 (2012) 221–
227. https://doi.org/10.1016/J.SSCI.2011.08.020.
[10] M. Haghani, M. Sarvi, Crowd behaviour and motion: Empirical methods, Transportation
Research Part B: Methodological. 107 (2018) 253–294.
https://doi.org/10.1016/J.TRB.2017.06.017.
[11] D. Duives, H. Mahmassani, Exit Choice Decisions during Pedestrian Evacuations of Buildings:,
Journal of the Transportation Research Board. (2012) 84–94. https://doi.org/10.3141/2316-
10.
[12] R. Lovreglio, D. Borri, L. Dell’Olio, A. Ibeas, A discrete choice model based on random utilities
for exit choice in emergency evacuations, Safety Science. 62 (2014) 418–426.
https://doi.org/10.1016/j.ssci.2013.10.004.
[13] M. Haghani, M. Sarvi, O. Ejtemai, M. Burd, A. Sobhani, Modeling Pedestrian Crowd Exit
Choice through Combining Sources of Stated Preference Data, Transportation Research
Record: Journal of the Transportation Research Board. 2490 (2015) 84–93.
https://doi.org/10.3141/2490-10.
[14] R. Lovreglio, A. Fonzone, L. Dell’Olio, A Mixed Logit Model for Predicting Exit Choice During
Building Evacuations, Transportation Research Part A. 92 (2016) 59–75.
https://doi.org/doi.org/10.1016/j.tra.2016.06.018.
[15] R. Lovreglio, A. Fonzone, L. Dell’Olio, D. Borri, A Study of Herding Behaviour in Exit Choice
during Emergencies based on Random Utility Theory, Safety Science. 82 (2016) 421–431.
https://doi.org/10.1016/j.ssci.2015.10.015.
[16] X. Ben Song, R. Lovreglio, Investigating personalized exit choice behavior in fire accidents
using the hierarchical Bayes estimator of the random coefficient logit model, Analytic
Methods in Accident Research. 29 (2021) 100140.
https://doi.org/10.1016/J.AMAR.2020.100140.
[17] M. Haghani, Majid Sarvi, Human exit choice in crowded built environments: Investigating
underlying behavioural differences between normal egress and emergency evacuations, Fire
Safety Journal. 85 (2016) 1–9. https://doi.org/10.1016/J.FIRESAF.2016.07.003.
[18] N.W.F. Bode, E.A. Codling, Human exit route choice in virtual crowd evacuations, Animal
Behaviour. 86 (2013) 347–358. https://doi.org/10.1016/J.ANBEHAV.2013.05.025.
[19] M. Kinateder, B. Comunale, W.H. Warren, Exit choice in an emergency evacuation scenario is
influenced by exit familiarity and neighbor behavior, Safety Science. 106 (2018) 170–175.
https://doi.org/10.1016/J.SSCI.2018.03.015.
[20] R. Lovreglio, Modelling Decision-Making in Fire Evacuation based on the Random Utility
Theory - PhD Thesis, Politecnico of Bari, Milan and Turin, 2016.
https://doi.org/10.13140/RG.2.1.1695.5281/1.
[21] M. Haghani, M.C.J. Bliemer, J.M. Rose, H. Oppewal, E. Lancsar, Hypothetical bias in stated
choice experiments: Part II. Conceptualisation of external validity, sources and explanations
of bias and effectiveness of mitigation methods, Journal of Choice Modelling. 41 (2021)
100322. https://doi.org/10.1016/J.JOCM.2021.100322.
[22] E.D. Kuligowski, M.T. Kinateder, Human Behavior in Fire in the Built Environment, in: SFPE
Handbook of Fire Protection Engineering 6th Edition, 2023.
[23] M. Kinateder, E. Ronchi, D. Nilsson, M. Kobes, M. Müller, P. Pauli, A. Mühlberger, Virtual
Reality for Fire Evacuation Research, in: Computer Science and Information Systems
(FedCSIS), IEEE, Warsaw, 2014: pp. 313–321. https://doi.org/10.15439/2014F94.
[24] R. Lovreglio, M.T. Kinateder, Augmented Reality for Pedestrian Evacuation Research:
Promises and Limitations, Safety Science. 128 (2020) 104750.
https://doi.org/https://doi.org/10.1016/j.ssci.2020.104750.
[25] Y. Feng, D. Duives, W. Daamen, S. Hoogendoorn, Data collection methods for studying
pedestrian behaviour: A systematic review, Building and Environment. 187 (2021) 107329.
https://doi.org/10.1016/J.BUILDENV.2020.107329.
[26] C. Minze, R. Yang, Z. Tao, P. Zhang, Mixed reality LVC simulation: A new approach to study
pedestrian behaviour, Building and Environment. (2021) 108404.
https://doi.org/10.1016/J.BUILDENV.2021.108404.
[27] L. Cao, J. Lin, N. Li, A virtual reality based study of indoor fire evacuation after active or
passive spatial exploration, Computers in Human Behavior. 90 (2019) 37–45.
https://doi.org/10.1016/j.chb.2018.08.041.
[28] M. Haghani, Empirical methods in pedestrian, crowd and evacuation dynamics: Part I.
Experimental methods and emerging topics, Safety Science. 129 (2020) 104743.
https://doi.org/10.1016/J.SSCI.2020.104743.
[29] J.E. Almeida, R.J.F. Rossetti, J.T.P.N. Jacob, B.M. Faria, A. Leça Coelho, Serious games for the
human behaviour analysis in emergency evacuation scenarios, Cluster Computing. 20 (2017)
707–720. https://doi.org/10.1007/S10586-017-0765-Z.
[30] A. Tucker, K.L. Marsh, T. Gifford, X. Lu, P.B. Luh, R.S. Astur, The effects of information and
hazard on evacuee behavior in virtual reality, Fire Safety Journal. 99 (2018) 1–11.
https://doi.org/10.1016/j.firesaf.2018.04.011.
[31] Z. Feng, V.A. González, R. Amor, M. Spearpoint, J. Thomas, R. Sacks, R. Lovreglio, G. Cabrera-
Guerrero, An immersive virtual reality serious game to enhance earthquake behavioral
responses and post-earthquake evacuation preparedness in buildings, Advanced Engineering
Informatics. 45 (2020) 101118. https://doi.org/10.1016/J.AEI.2020.101118.
[32] R. Lovreglio, X. Duan, A. Rahout, R. Phipps, D. Nilsson, Comparing the Effectiveness of Virtual
Reality Training and Video Training, Virtual Reality. 25 (2021) 133–145.
https://doi.org/https://doi.org/10.1007/s10055-020-00447-5.
[33] A. Rahouti, R. Lovreglio, S. Datoussaïd, T. Descamps, Prototyping and Validating a Non-
immersive Virtual Reality Serious Game for Healthcare Fire Safety Training, Fire Technology.
57 (2021) 3041–3078. https://doi.org/10.1007/S10694-021-01098-X/FIGURES/15.
[34] M. Kinateder, W.H. Warren, Social Influence on Evacuation Behavior in Real and Virtual
Environments, Frontiers in Robotics and AI. 3 (2016) 43.
https://doi.org/10.3389/frobt.2016.00043.
[35] S. Arias, Application of Virtual Reality in the study of Human Behavior in Fire: Pursuing
realistic behavior in evacuation experiments - PhD Thesis, Lund University, 2021. ISBN: 978-
91-7895-868-9
[36] A. Shipman, A. Majumdar, R. Lovreglio, A quantitative comparison of emotional and
movement responses to emergency scenarios between VR and physical experiments, Under
Review. (2022).
[37] S. Arias, A. Mossberg, D. Nilsson, J. Wahlqvist, A Study on Evacuation Behavior in Physical and
Virtual Reality Experiments, Fire Technology. (2021) 1–33. https://doi.org/10.1007/S10694-
021-01172-4.
[38] J. Lin, L. Cao, N. Li, How the completeness of spatial knowledge influences the evacuation
behavior of passengers in metro stations: A VR-based experimental study, Automation in
Construction. 113 (2020) 103136. https://doi.org/10.1016/J.AUTCON.2020.103136.
[39] M. Fu, R. Liu, Y. Zhang, Do people follow neighbors? An immersive virtual reality experimental
study of social influence on individual risky decisions during evacuations, Automation in
Construction. 126 (2021) 103644. https://doi.org/10.1016/J.AUTCON.2021.103644.
[40] Z. Xu, W. Wei, W. Jin, Q. rui Xue, Virtual drill for indoor fire evacuations considering occupant
physical collisions, Automation in Construction. 109 (2020) 102999.
https://doi.org/10.1016/J.AUTCON.2019.102999.
[41] Y. Feng, D.C. Duives, S.P. Hoogendoorn, Development and evaluation of a VR research tool to
study wayfinding behaviour in a multi-story building, Safety Science. 147 (2022) 105573.
https://doi.org/10.1016/J.SSCI.2021.105573.
[42] Y. Feng, D.C. Duives, S.P. Hoogendoorn, Wayfinding behaviour in a multi-level building: A
comparative study of HMD VR and Desktop VR, Advanced Engineering Informatics. 51 (2022)
101475. https://doi.org/10.1016/J.AEI.2021.101475.
[43] J.D. Sime, Affiliative behaviour during escape to building exits, Journal of Environmental
Psychology. 3 (1983) 21–41. https://doi.org/10.1016/S0272-4944(83)80019-X.
[44] G. Proulx, Occupant behaviour and evacuation, in: 9th International Fire Protection Seminar,
Munich, 2001: pp. 219–232.
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.11.202&rep=rep1&type=pdf (last
access: 11th June 2022)
[45] K. Fridolf, D. Nilsson, H. Frantzich, Fire Evacuation in Underground Transportation Systems: A
Review of Accidents and Empirical Research, Fire Technology. 49 (2013) 451–475.
https://doi.org/10.1007/s10694-011-0217-x.
[46] W. Grosshandler, N. Bryner, D. Madrzykowski, K. Kuntz, Report of the technical investigation
of The Station Nightclub fire: appendices, Washington DC, 2005.
https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=100989 (last access: 11th June 2022)
[47] L. Benthorn, H. Frantzich, Fire alarm in a public building: how do people evaluate information
and choose an evacuation exit?, Fire and Materials. 23 (1999) 311–315.
https://doi.org/10.1002/(SICI)1099-1018(199911/12)23:6<311::AID-FAM704>3.0.CO;2-J.
[48] M. Kobes, I. Helsloot, B. de Vries, J. Post, Exit choice, (pre-) movement time and (pre-)
evacuation behaviour in hotel fire evacuation — Behavioural analysis and validation of the
use of serious gaming in experimental research, Procedia Engineering. 3 (2010) 37–51.
https://doi.org/10.1016/j.proeng.2010.07.006.
[49] M. Haghani, M. Sarvi, ‘Herding’ in direction choice-making during collective escape of crowds:
How likely is it and what moderates it?, Safety Science. 115 (2019) 362–375.
https://doi.org/10.1016/J.SSCI.2019.02.034.
[50] J. Lin, R. Zhu, N. Li, B. Becerik-Gerber, Do people follow the crowd in building emergency
evacuation? A cross-cultural immersive virtual reality-based study, Advanced Engineering
Informatics. 43 (2020) 101040. https://doi.org/10.1016/j.aei.2020.101040.
[51] K. M, W. WH, Exit choice during evacuation is influenced by both the size and proportion of
the egressing crowd, Physica A. 569 (2021). https://doi.org/10.1016/J.PHYSA.2021.125746.
[52] Y. Zhu, T. Chen, N. Ding, M. Chraibi, W.C. Fan, Follow the evacuation signs or surrounding
people during building evacuation, an experimental study, Physica A: Statistical Mechanics
and Its Applications. 560 (2020). https://doi.org/10.1016/J.PHYSA.2020.125156.
[53] M. Kinateder, M. Müller, M. Jost, A. Mühlberger, P. Pauli, Social influence in a virtual tunnel
fire - Influence of conflicting information on evacuation behavior, Applied Ergonomics. 45
(2014) 1649–1659. https://doi.org/10.1016/j.apergo.2014.05.014.
[54] M. Kinateder, E. Ronchi, D. Gromer, M. Müller, M. Jost, M. Nehfischer, A. Mühlberger, P.
Pauli, Social influence on route choice in a virtual reality tunnel fire, Transportation Research
Part F: Traffic Psychology and Behaviour. 26 (2014) 116–125.
https://doi.org/10.1016/J.TRF.2014.06.003.
[55] D. Nilsson, H. Frantzich, W. Saunders, Influencing Exit Choice in the Event of a Fire
Evacuation, Fire Safety Science. 9 (2008) 341–352. https://doi.org/10.3801/IAFSS.FSS.9-341.
[56] J. Gibson, The theory of affordances, in: J.J. Giesking, W. Mangold (Eds.), The People, Place,
and Space Reader, 1977: pp. 56–60. ISBN: 0415664977
[57] P.G. Wood, Fire Research Note 953, Building Research Establishment, Borehamwood, UK,
1972. https://iafss.org/publications/frn/953/-1/view/frn_953.pdf (last access: 11th June 2022)
[58] J.L. Bryan, Smoke as a Determinant of Human Behavior in Fire Situations, US Department of
Commerce, National Bureau of Standards, 1972.
https://books.google.co.nz/books/about/Smoke_as_a_Determinant_of_Human_Behavior.ht
ml?id=Mh2zHAAACAAJ&redir_esc=y (last access: 11th June 2022)
[59] T. Jin, T. Yamada, Experimental Study Of Human Behavior In Smoke Filled Corridors, Fire
Safety Science. 2 (1989) 511–519. https://doi.org/10.3801/IAFSS.FSS.2-511.
[60] M. Fu, R. Liu, Y. Zhang, Why do people make risky decisions during a fire evacuation? Study
on the effect of smoke level, individual risk preference, and neighbor behavior, Safety
Science. 140 (2021) 105245. https://doi.org/10.1016/J.SSCI.2021.105245.
[61] S.M. V. Gwynne, E.R. Rosenbaum, Employing the Hydraulic Model in Assessing Emergency
Movement, in: SFPE Handbook of Fire Protection Engineering, Springer New York, New York,
NY, 2016: pp. 2115–2151. https://doi.org/https://doi.org/10.1007/978-1-4939-2565-0_59.
[62] W.H. Greene, Econometric Analysis - 7th Edition, Pearson, London, UK, 2011. ISBN:
0131395386
[63] Z. Sándor, M. Wedel, Designing Conjoint Choice Experiments Using Managers’ Prior Beliefs,
Journal of Marketing Research. 38 (2001) 430–444.
https://doi.org/10.1509/jmkr.38.4.430.18904.
[64] Institute of Transport and Logistics Studies, Ngene – A Software Capability to Design and
Generate Choice Experiments, 2007. http://www.choice-metrics.com/NgeneManual120.pdf.
[65] Unity Technologies, Unity webpage, https://Unity.com/ (last access: 11th June 2022) (2021).
[66] Adobe, Mixamo webpage, https://Mixamo.com/ (last access: 11th June 2022) (2021).
[67] R. Lovreglio, V. Gonzalez, Z. Feng, R. Amor, M. Spearpoint, J. Thomas, M. Trotter, R. Sacks,
Prototyping virtual reality serious games for building earthquake preparedness: The Auckland
City Hospital case study, Advanced Engineering Informatics. 38 (2018) 670–682.
https://doi.org/10.1016/J.AEI.2018.08.018.
[68] J. de D. Ortuzar, L.G. Willumsen, Modelling Transport, 4th Edition, John Wiley & Sons, 2011.
ISBN: 978-0-470-76039-0
[69] K. Train, Discrete choice methods with simulation, 2nd Ed, Cambridge University Press,
Cambridge ; New York, 2009. ISBN: 0521766559
[70] W.H. Greene, D.A. Hensher, Modeling Ordered Choices: A Primer, Cambridge University
Press, 2010. ISBN: 0521336031
[71] Y. Croissant, Package ‘mlogit: Multinomial Logit Models’
https://cran.r-project.org/web/packages/mlogit/index.html (last access: 11th June 2022)
(2020)
[72] R. Lovreglio, Virtual and Augmented reality for human behaviour in disasters: a review, in:
Fire and Evacuation Modeling Technical Conference (FEMTC), 2020: p. 14.
https://files.thunderheadeng.com/femtc/2020_d3-02-lovreglio-paper.pdf (last access: 11th
June 2022)
[73] H. Li, J. Zhang, L. Xia, W. Song, N.W.F. Bode, Comparing the route-choice behavior of
pedestrians around obstacles in a virtual experiment and a field study, Transportation
Research Part C: Emerging Technologies. 107 (2019) 120–136.
https://doi.org/10.1016/J.TRC.2019.08.012.
[74] Y. Tong, N.W.F. Bode, The principles of pedestrian route choice, Journal of the Royal Society
Interface. 19 (2022). https://doi.org/10.1098/RSIF.2022.0061.
[75] M. Haghani, M. Sarvi, Social dynamics in emergency evacuations: Disentangling crowd’s
attraction and repulsion effects, Physica A: Statistical Mechanics and Its Applications. 475
(2017) 24–34. https://doi.org/10.1016/J.PHYSA.2017.02.010.
[76] R. Brouwer, T. Dekker, J. Rolfe, J. Windle, Choice Certainty and Consistency in Repeated
Choice Experiments, Environmental and Resource Economics 2009 46:1. 46 (2009) 93–109.
https://doi.org/10.1007/S10640-009-9337-X.
[77] R. Lovreglio, V.A. González, Z. Feng, R. Amor, M. Spearpoint, Prototyping Virtual Reality
Serious Games for Earthquake Preparedness: the Auckland City Hospital Case Study,
Advanced Engineering Informatics. (2018). https://doi.org/10.1016/j.aei.2018.08.018.
[78] F. Ozel, Time pressure and stress as a factor during emergency egress, Safety Science. 38
(2001) 95–107. https://doi.org/https://doi.org/10.1016/S0925-7535(00)00061-8.
[79] E. Ronchi, D. Nilsson, S. Kojić, J. Eriksson, R. Lovreglio, H. Modig, Anders Lindgren Walter, A
Virtual Reality experiment on flashing lights at emergency exit portals for road tunnel
evacuation, Fire Technology. 52 (2016) 623–647. https://doi.org/10.1007/s10694-015-0462-
5.
[80] M. Kinateder, E. Ronchi, D. Gromer, M. Müller, M. Jost, M. Nehfischer, A. Mühlberger, P.
Pauli, Social influence on route choice in a virtual reality tunnel fire, Transportation Research
Part F: Traffic Psychology and Behaviour. 26 (2014) 116–125.
https://doi.org/10.1016/j.trf.2014.06.003.
[81] A. Kühberger, A. Fritz, T. Scherndl, Publication Bias in Psychology: A Diagnosis Based on the
Correlation between Effect Size and Sample Size, PLOS ONE. 9 (2014) e105825.
https://doi.org/10.1371/JOURNAL.PONE.0105825.
[82] S. Gwynne, E.. Galea, P.. Lawrence, L. Filippidis, Modelling occupant interaction with fire
conditions using the buildingEXODUS evacuation model, Fire Safety Journal. 36 (2001) 327–
357. https://doi.org/10.1016/S0379-7112(00)00060-6.
[83] J. Blascovich, A theoretical model of social influence for increasing the utility of collaborative
virtual environments, in: Proceedings of the 4th International Conference on Collaborative
Virtual Environments - CVE ’02, ACM Press, New York, New York, USA, 2002: pp. 25–30.
https://doi.org/10.1145/571878.571883.
