Citation

Nilsson D, Johansson A. Fire Safety J. 2009; 44(1): 71-79.

Copyright

(Copyright © 2009, Elsevier Publishing)

DOI

10.1016/j.firesaf.2008.03.008

PMID

unavailable

Abstract

aDepartment of Fire Safety Engineering, Lund University, Box 118, SE-22100 Lund, Sweden bDepartment of Humanities and Social Sciences, ETH Zurich, Universitätstrasse 41, 8092 Zurich, Switzerland Unannounced evacuation experiments in a cinema theatre were analysed. The analysis focused on investigating if people are influenced by others during the initial phase of a fire evacuation. This type of influence is called social influence. Three separate behaviour types were identified and studied and the recognition and pre-movement time was measured. The results suggest that social influence is an important factor and that it becomes more important when the fire cue, e.g., the alarm, is unclear or uninformative. Results also indicate that social influence increases with decreasing distance between visitors. This result implies that individuals are influenced more by people who are close than by people who are further away. Keywords: Evacuation experiment; Social influence; Recognition time; Pre-movement time The estimation of RSET is an important step in performance-based design and computer software, such as Simulex [4] and STEPS [5], is often used. These programs are capable of simulating evacuation from buildings of many people simultaneously. They also enable the user to specify occupant characteristics. In computer simulations, it is often assumed that each occupant is exposed to a fire cue and starts to move towards an exit after a certain delay, which is often called pre-movement time. This time period starts when the occupant is exposed to a fire cue and ends when he or she starts to physically move to a safe place, typically to an emergency exit. A fire cue can be for example a fire alarm or smoke, i.e., a cue that signals that a fire emergency may have occurred. In the programs, the user has to specify the pre-movement time used in the simulation. The pre-movement time can be difficult to quantify. It includes a wide range of behaviour, such as recognising the fire cue, deciding what actions to take and preparing to move to an exit. The pre-movement time is sometimes divided into two separate phases, namely recognition and response [1]. The recognition phase is the period between the reception of a fire cue and the occupant's first response. The response phase is the period between the occupant's first response and the time when he or she starts to physically move towards an exit. It has been recognised that the pre-movement time may be as important as the time it takes to move to an exit [6]. Therefore, it is imperative that the estimation of the pre-movement time is as accurate as possible. Many full-scale evacuation experiments have been conducted in order to study evacuation from buildings and to quantify the pre-movement time. Experiments have been performed in a wide range of building types, such as apartments [7], stores [8] and [9] and cinemas [9] and [10]. Many of these experiments have been used to develop recommended pre-movement times for different combinations of building types and alarms [1]. Recommendations are often used in hand calculations and computer simulations of fire evacuation. People will most likely be influenced by others and their behaviour during the pre-movement phase. If someone starts to move towards an exit it is likely that others will follow. Similarly, inactivity of others may also inhibit people's actions. This type of influence is sometimes called social influence and is seldom taken directly into consideration when RSET is estimated. The importance of social influence was demonstrated in experiments at Columbia University [11]. In the experiments, students were exposed to artificial smoke while they were filling out a questionnaire in a small room. The students were not informed about the simulated emergency beforehand, but instead they were told that they were going to take part in an interview about urban life. Three different cases were studied. The students took part either alone, in groups of three or one student together with two passive confederates. All confederates were informed about the experiment, but were instructed not to act or report the smoke. The experiments at Columbia University revealed that students who were alone were most likely to report the smoke. Students who were alone also reported smoke more quickly than those who were in the presence of others. Hence, the presence of others seemed to inhibit the students from taking action. This trend was particularly clear when two passive confederates were present. The study clearly demonstrates that people can be influenced by others in fire emergencies. Social influence does not apply exclusively to emergencies and it can be observed in many different settings. A distinction between normative and informational social influence has been identified based on a study of individual judgement [12]. The normative part is the result of the individual's desire to conform to the expectations of other people. People want to act in accordance with what is expected and may not want to stand out or make a fool of themselves. Individuals will also study other people and their behaviour to gain information about the current situation. The actions or inaction of others may influence people's perceptions of the situation and their subsequent behaviour. This latter type of influence is called informational social influence. Research about helping behaviour has suggested that social influence may be more important for ambiguous emergencies [13]. This may also be relevant for fire emergencies. If the fire cue is clear, e.g., an informative pre-recorded message, it is possible that social influence may be less important. For ambiguous cues, e.g., a fire alarm bell, it is likely that social influence becomes significant. In the latter case, people may not want to make fools of themselves, which will inhibit their response. At the same time, they might interpret the inaction of others as a sign that the situation is not a real emergency. This may also inhibit their response. If the pre-movement time is based solely on the type of building and alarm no direct account will be taken of the social influence. However, implicitly some of the influence will be taken into consideration since certain building types are often associated with specific social settings. One example is cinema theatres where groups of visitors usually sit in seats facing a screen. An estimation of the pre-movement time that is based on experiments or real fires in cinemas will be exclusive for that type of setting. The purpose of the present study was to investigate social influence in a fire emergency setting, namely during evacuation from a cinema theatre. It was hypothesised that the social influence would be more important for an ambiguous than for a clear fire cue. A series of unannounced evacuations were performed by Bayer and Rejnö in a cinema theatre in Sweden in autumn of 1999 [9] and [10]. The purpose of the experiments was to test how different alarms affected the pre-movement time. Six types of alarms were tested, namely an alarm bell, an alarm tone signal, an alarm bell together with a flashing light, an alarm bell together with an information sign and two pre-recorded messages. One of the pre-recorded messages was recorded by a man and the other by a woman. Each alarm was tested three times and the total number of evacuation experiments was 18. All evacuations were performed in the same cinema theatre, but each participant only took part once. Five of the 18 evacuation experiments were included in the following study (see Table 1). The selection was based on the type of alarm and the number of participants. Experiments with two types of alarms, namely an alarm bell and a pre-recorded message (female), were chosen. These alarms were selected since they represent two distinct levels of ambiguity. The pre-recorded message is a clear cue and informs the evacuees about what has happened and how they should behave. The alarm bell is considerably more ambiguous, since it does not provide any specific information. Only experiments in which the cinema theatre was at least half full were selected. This meant that one of the three experiments with an alarm bell was excluded. The five evacuation experiments selected for this study are described in the following sections. The description is generally applicable to all 18 experiments, but some details are specific. In Table 1, the type of alarm that was used in the experiments is shown (see Table 1). The experiments with the alarm bell will be called case A and the experiments with the pre-recorded message (female) will be called case B in the following text. The participants consisted of cinema theatre visitors who had bought tickets for the shows that were chosen for the experiments. This meant that the number of participants and the composition of people could not be controlled directly. The table above shows the number of participants who took part in the five evacuation experiments (see Table 1). The number of participants was determined through observation of video recordings of the experiments. All participants were asked to fill out a questionnaire after the experiments. The questionnaire contained questions about their age and gender. The table below shows the age and gender distribution based on the participants’ answers (see Table 2). None of the participants were informed about the evacuation prior to the experiment, but they were given information afterwards. The cinema theatre that was used in the experiments took a maximum of 135 visitors (see Fig. 1). It consisted of nine rows with 15 seats each. On each side of the nine rows, there were stairs from the front to the back of the cinema theatre. This meant that the visitors could exit a row both to the left and right. Two doors connected the cinema theatre to other parts of the building. One door was the entrance and it was placed in the back right corner. The other door was located in the front left corner. Both doors were equipped with standard emergency exit signs. The video camera that was used to document the experiments was placed in a dark box and mounted in the front right corner of the room. This location provided a good overview of the entire cinema theatre. The alarm bell that was used in experiments A1 and A2 was placed along the left wall and was clearly visible. The pre-recorded message in case B was played using the cinema theatre's sound system and no additional equipment needed to be installed. Before an experiment was initiated the video camera was started. Shortly thereafter the participants, i.e., visitors, were allowed to enter the cinema theatre and take their seats. After a short while the advertisements that precede the movie were shown on the screen. New participants continued to arrive, but all had arrived before the advertisement came to an end. When the advertisement had stopped the alarm was started. This meant that nothing was showing on the screen when the alarm was sounded. The alarm bell used for case A rang continuously until the experiment was terminated. Measurements suggest that the sound level was approximately 102 dB in the centre of the cinema theatre [10]. The pre-recorded message that was used for case B was recorded with a female voice. It contained a general call for attention, information that a fire incident had occurred and instructions to leave through the closest exit and gather outside the building. In the experiments, the pre-recorded message was preceded by a pulsating signal that lasted for 5 s. Three seconds after the signal had ended the message, which was 12 s long, was played. After 3 s, the message was then played again and this was repeated until the experiment was terminated. Measurements suggest that the sound level for the pre-recorded message was approximately 93 dB in the centre of the cinema theatre [10]. The experiment was terminated when all participants had responded and started to evacuate. The alarm was stopped and the participants were escorted back to their seats. They were then given information about the evacuation experiment and were asked to fill out a questionnaire, which contained questions about participants’ perception of the alarm, associations with the alarm, chosen exit, age and gender. The questionnaire was developed by Bayer and Rejnö [10] and only items relating to age and gender have been examined in the present study. All questionnaires were collected at the exit at the end of the movie. Observers were present inside the cinema theatre during most of the evacuations. From the video recordings, it was concluded that an observer was present in experiments A1, A2 and B1, but not in B2 and B3. The observer was instructed not to respond until others had responded in order to minimise the influence on the participants. In the present study, the video recordings from the experiments conducted by Bayer and Rejnö [10] were analysed. The participants’ behaviour and actions were studied in the analysis. Different types of behaviour were identified and the time at which the participants in the cinema theatre displayed these types of behaviour was measured. All times were measured relative to the alarm signal, i.e., the time when the alarm was sounded was set to zero. The data retrieved from the video recordings were then analysed in order to investigate if social influence was an important factor in the experiments. In the analysis of the video recordings, three distinct types of behaviour were studied. All the types were defined and specified prior to the analysis. The table below gives a description of the behaviour types that were used in the study (see Table 3). People will most likely be affected by others and their behaviour in fire emergency situations involving more than one person. In order to be influenced they have to observe other people, i.e., look at them. The first behaviour type that was used in the analysis was called look at others beside or behind oneself. This was assumed to occur when a person turned his or her head or entire body so that the person's face was turned at least ninety degrees away from the screen. Only those participants who turned before they started to prepare, i.e., displayed behaviour type two, were assumed to look at others beside or behind them. This definition was used since it was believed that people who have started to prepare are influenced less by others. They have committed to exiting the building and would therefore be less perceptible to social influence. When a person has received a fire cue he or she has to recognise the cue and respond. The response phase can include a variety of actions that the person performs before he or she starts to move towards an exit. During this phase the person will prepare to evacuate, which may include for example putting on clothes and gathering belongings. The second behaviour type that was used in the analysis of the video recordings was called start to prepare. Start to prepare was defined as the first behaviour that indicated that the participant was actively getting ready to escape. This included actions such as looking for clothes, putting on clothes or starting to get up from the seat (see Table 3). Behaviour type two correlates with the end of the recognition phase and the start of the response phase for the cinema theatre setting. This means that the time at which a participant started to prepare is equal to the recognition time for that participant (see Fig. 2). The time period will therefore be referred to as the recognition time. Before the cinema visitors started to physically move towards an exit they had to get up from their seat. The third behaviour type that was used in the analysis of the video recordings was called rise. Participants were assumed to have risen when they had gotten up from their seat and were standing with their back straightened. This definition meant that people who leaned forward to pick up belongings as they exited their seat were not assumed to have risen until they had finished gathering belongings and were standing straight. The third behaviour type correlates with the end of the pre-movement time. In most cases, the participants started to physically move towards an exit directly after they had risen. Some participants got up from their seat and started to move, but halted to wait for their friends or companions. This means that the time at which a participant rose is equal to the pre-movement time for that participant (see Fig. 2). The time period will hence be referred to as the pre-movement time in subsequent text. The first two behaviour types were generally easy to observe. In some cases, participants were obscured by others who were standing up, which made observation impossible. If it was not possible to observe the behaviour type the participant was excluded from further analysis. For the first behaviour type, one, two, five, seven and six participants were excluded for experiments A1, A2, B1, B2 and B3, respectively. The corresponding number of excluded participants in the analyses of behaviour type two was five, four, 15, 11 and 10 for experiments A1, A2, B1, B2 and B3. A big advantage of behaviour type three was that it was very easy to observe. Due to the location of the camera and placement of the seats in the cinema theatre no participants were obscured. The cinema theatre consisted of nine rows with 15 seats each. For documentation purposes, the rows were numbered from 1 to 9 in the analysis. The first row was closest to the screen and the back row was row number nine. All the seats on each row were numbered from 1 to 15. The seats were numbered from right to left in the figure above (see Fig. 1). All participants were given indexes, namely (row, seat), based on where they were sitting in the cinema theatre. In the analysis of the video recordings, neighbours were defined as the participants who were sitting in any of the eight adjacent seats (see Fig. 3). Each participant could have a maximum of eight neighbours. If a participant was located at (row1, seat1) his neighbours were located at (row2, seat2), where row2 varied between row1-1 and row1+1 and seat2 varied between seat1-1 and seat1+1. The definition of neighbours described above was chosen since it is considered intuitive and simple. Also, since neighbours are located on all sides of a participant no assumption regarding the direction of social influence, e.g., from the front to the back of the cinema theatre or vice versa, was implicitly assumed in subsequent analyses. Another definition that was used in the analysis of the video recordings was that of closest neighbour. Closest neighbours were defined as the neighbours that were located directly to the left or right of a participant (see Fig. 4). Each participant could have a maximum of two closest neighbours. If a participant was located at (row1, seat1) his closet neighbours were located at (row2, seat2), where seat2 was equal to either seat1-1 or seat1+1 and row2 was equal to row1. The following data were obtained from analysis of the video recordings: From the analysis of the video recordings, participants who looked at others beside or behind them was recorded, i.e., behaviour type one. The number of participants who displayed this type of behaviour in the experiments is shown in Table 4. In all experiments, some participants were obscured by others who were standing up, this making observation impossible. The obscured participants were therefore not included in any further analysis of behaviour type one. In Table 4, both cases A, namely alarm bell, and case B, namely spoken message (female), are included (see Table 4). It can be seen in Table 4 that 80 of 185 participants, i.e., 43 per cent, looked at others behind them for case A. For case B only 39 of 365, i.e., 11 per cent displayed the same type of behaviour. These results show that a significantly higher proportion of the participants looked at others beside or behind them for case A than for case B (p<0.01). This suggests that behaviour type one is more common when an alarm bell is used. Fig. 5 displays the proportion of participants who had started to prepare as a function of time, i.e., the cumulative distribution for the recognition time. In the figure, it can be seen that the first participant started to prepare after 7 and 9 s for case A and between 12 and 18 s for case B. It can also be seen from the shape of the curves that the majority of participants displayed behaviour type two approximately 10 s earlier for experiment A1 than A2. For experiments B1–B3, the majority of participants started to prepare at approximately the same time. It should be mentioned that behaviour type two, i.e., start to prepare, was hard to observe since some participants were obscured by others standing up. Participants who were obscured were excluded and are therefore not included in Fig. 5. Fig. 6 displays the cumulative distribution for the pre-movement time. In Fig. 6, it can be seen that the first participant rose between 20 and 22 s after the alarm had started for case B. At that particular time, the signal had sounded and the spoken message had been played once. For case A, the first participant rose after 10 and 19 s, respectively, in the two experiments. The shape of the curves reveals that the majority of participants rose approximately 10 s earlier for experiment A1 than A2. For experiments B1–B3, the majority of participants rose at approximately the same time. This corresponds with the observation described earlier for the second behaviour type. Fig. 7 displays the participants’ recognition time in experiment A1. In Fig. 7, the seats are marked with the numbers one to 15 for each row and the rows are numbered from 1 to 9. The height of the pillar at each location corresponds to the recognition time. In Fig. 7, it can be seen that the person who was seated in seat seven on row three started to prepare quite early, and his or her recognition time was 7 s. It can also be seen that the time increases with increasing row number. The recognition time for the participant in seat seven on row nine, i.e., the back row, is 17 s. A similar wave pattern from the front to the back of the cinema theatre can also be observed for seats with the numbers six and eight. The pattern can also be observed in other parts of the cinema theatre, but is not as clear at other locations. The same trend as above can also be seen if the pre-movement time is plotted in a similar fashion for experiment A1 (see Fig. 8). However, the wave that propagates backwards from seat seven on row three is not as visible if the pre-movement time is used. In the figures above, it can be seen that some participants waited considerably longer than their neighbours before they started to prepare or to rise (see Fig. 7 and Fig. 8). However, it can be seen that many of the participants acted shortly before or after their neighbours. In order to investigate if the cinema visitors acted like their neighbours the absolute difference between recognition time was calculated for all possible combinations of participants in each experiment. The same calculation was also performed for the pre-movement time. The absolute difference for two arbitrarily chosen participants in the cinema theatre is defined according to The table below shows the mean of the absolute recognition time difference for all combinations of both neighbours (see Fig. 3) and others, i.e., participants who were not neighbours (see Table 5). It can be seen that the mean value for neighbours is smaller for all experiments. This indicates that the participants acted more like their neighbours with regards to time. If people acted independently of each other the mean values would in theory be similar for neighbours and others. The same trend can also be observed for the mean of the absolute pre-movement time difference for all combinations of neighbours and others (see Table 6). The difference between neighbours and others described above suggests that the participants acted more like their neighbours with regards to time. In order to test this theory, the data were transformed using random permutation. This transformation meant that the participants in each of the five experiments were assumed to exchange places in a random fashion, whereby the participant were surrounded by new neighbours. The mean absolute recognition time difference for all combinations of new neighbours was then calculated using the previously described method. This value was then compared with the mean for the non-transformed data. It was noted whether the mean absolute recognition time difference for the transformed data was higher or lower than for the non-transformed data. The procedure was then repeated until the proportion of higher and lower instances stabilised within desired accuracy limits. This method was then repeated for the pre-movement time for all experiments. The results of the random permutations can be interpreted as significance levels. Let us take the random permutation of the data from experiment A1 as an example. The original analysis revealed that the mean absolute recognition time difference was 2.65 s. After a random permutation, the mean absolute recognition time difference was calculated taking into account that all individuals had received new neighbours. The mean value was then compared with the original value of 2.65 and was found to be either higher or lower, where lower indicates on average smaller time differences between neighbours. This procedure was then repeated many times and the proportion of instances when the value was lower was registered. The analysis shows that the proportion of lower instances was less than 0.01, which corresponds to a significance level. This result suggests that the participants in experiment A1 did most likely not respond independently of their neighbours. Because the absolute recognition time difference for others, i.e., people who were not neighbours, will be only marginally influenced by random permutation it can also be said that the recognition time for the participants in experiment A1 was found to be significantly more similar for neighbours than non-neighbours (p<0.01). Table 7 displays the significance levels based on random permutation for both the recognition and pre-movement time for all experiments. It can be seen in the table that the recognition and pre-movement time is significantly more similar for neighbours in experiments A1 and A2, i.e., the experiments with the alarm bell. The results also show that the pre-movement time is significantly more similar for neighbours in experiments B1–B3, but the same trend is not clear for the recognition time. However, it should be added that some participants had to be excluded from further analysis of the recognition time because they were obscured by others, which meant that the available data were reduced. Although the significance tests show that the recognition and pre-movement time is more similar for neighbours in some of the experiments, they do not explain the reasons behind this trend. In order to investigate this trend further the average recognition and pre-movement time was calculated for each row and seat number in the experiments. The distribution of average values was compared across experiments, but no systematic trends were discovered. An alternative to the random permutation method described above could have been to test significance between the mean values for neighbours and non-neighbours in each of the experiments using traditional significance tests. However, the way in which the mean values in Table 5 and Table 6 were obtained makes most tests inappropriate. Since random permutation does not rely on assumptions about included distributions it is appropriate to use in this particular case. In the analysis of the video recordings, it was also registered which participants visited the cinema together, i.e., friends. The majority of participants came in groups of two and the largest groups consisted of five visitors. All but one group sat on the same row in the experiments. In order to investigate if the cinema visitors acted like their friends the absolute difference between pre-movement time was calculated for all possible combinations of closest neighbours (see Fig. 4). The closest neighbours were in turn classified as either friends or others. The table below shows the mean of the absolute pre-movement time difference for all combinations of both friends and others who were closest neighbours (see Table 8). It can be seen in the table that the mean value is smaller for friends, which means that the participants acted more like their friends with regards to time in the experiments. Similar calculations were not performed for the recognition time. The reason for this was that many of the participants were obscured and hence excluded from further analysis, which made the available data very limited. People will often be influenced by others and their behaviour during the initial part of a fire evacuation. However, social influence is hard to quantify which makes it difficult to study explicitly. One type of behaviour that is believed to correlate with social influence is observation of others. A person can gain more information about the emergency by looking at others and their behaviour. This process is called informational social influence. Similarly, a person can avoid breaking norms, such as making a fool of herself or himself, by observing what others are doing and acting in accordance with their behaviour. This process is called normative social influence. In the present study, participants who looked at others beside or behind them were studied. This type of behaviour was selected because it is thought to be associated with social influence and it is possible to study it with a high degree of repeatability. It is believed that people who look at others beside or behind them clearly observe what others are doing. One limitation is that people seated in front rows will most likely display the behaviour to a greater extent than those seated in back rows. The reason for this is that those people seated in back rows do not have to turn in order to see what others are doing. It cannot be ruled out that that the reason for looking beside or behind oneself in the experiments was to get more information about the situation, e.g., look for smoke or fire. However, in the cinema theatre evacuations the only additional information available was actions or inaction of others. Despite possible limitations looking at others beside or behind oneself is believed to be a clear indication of social influence. A better definition would involve eye movement, but the poor resolution of the video recordings made this kind of detailed analysis impossible in the present study. The analysis of the video recordings revealed that the participants looked at others beside or behind them both when the alarm bell and the spoken message were used. This suggests that people are influenced by others during the initial part of a fire evacuation. In addition, participants looked at others beside or behind them to a greater extent when the alarm bell was used, which suggests that the social influence was stronger for that case. The alarm bell was the more ambiguous of the two alarm types that were included in the study. It did not provide the participants with any additional information about what had happened or how they should act. The spoken message on the other hand offered clear information about the situation. The difference between the two types of alarms indicate that social influence is more important for ambiguous fire cues. This conclusion is also supported by similar research about helping behaviour in emergencies [13]. In the present study, the two behaviour types start to prepare and rise were defined and used. The time until the participants started to prepare corresponds to the recognition time. Similarly, the time until the participants rose from their seats corresponds to the pre-movement time. The analysis revealed that the cumulative distribution for the recognition time was very similar for all experiments in which the spoken message was used. The same trend was obvious for the pre-movement time. A likely explanation is that the information provided in a message makes it easier for people to decide on a form of action, namely to evacuate. This means that social influence becomes less important since people do not have to look at others to gain additional information and also will not fear that they are breaking any norms by acting prematurely. For the two experiments in which the alarm bell was used there was greater difference between the cumulative distributions for both the recognition and pre-movement time than for the experiments with the pre-recorded message. This may be a result of the more ambiguous cue and the resulting stronger social influence between participants. Social influence is a dynamic process in which both actions and inaction of others are important. Inaction may indicate to others that the situation is not severe and actions that it is. This dynamic process may lead to large variations. The cumulative distribution for the two experiments with the alarm bell were similar, but one was shifted approximately 10 s relative to the other. Additional experiments are required in order to determine the magnitude of variation that can be expected when the fire cue is ambiguous and the resulting social influence between participants is strong. Cinema visitors will most likely sit close to others, i.e., the distance between people will sometimes be short. It is considered likely that visitors will be influenced more by people who are close than by those who are further away. The effect of distance could be clearly observed in one of the cinema experiments in which an alarm bell was used. In the experiment, a person in the front row responded quickly and probably influenced his or her neighbours to respond shortly after. The neighbours in turn influenced their neighbours et cetera, which generated a wave pattern in the cinema theatre. This wave was particularly clear since it propagated from the front to the back of the room. The observed wave pattern is believed to be an indication that social influence increases with decreasing distance. Another result that supports this theory is the reported relationship between the mean values of the absolute time differences, both for the recognition and pre-movement time, for all combinations of neighbours and others. The results show that the mean value was smaller for neighbours in the experiments, which indicate that people respond more like their neighbour with regards to time. This trend was also confirmed by significance tests. The results do not exclusively prove that social influence increases with decreasing distance. It simply indicates that an individual on average responds more like his or her neighbour than someone else with regards to time. However, it is believed that this is a result of increasing social influence from neighbours. Another possible explanation is that the response and pre-movement time, respectively, are functions of the location in the cinema theatre. For example, it could be possible that people respond quicker if they are seated closer to the screen or further away from the main exit. These types of systematic tends could not be observed in the analysis of the experiments, which makes it more likely that the observed tendency is due to social influence. The results suggest that people are influenced more by their neighbours than by others. This trend implies that social influence increases with decreasing distance, which may give rise to characteristic wave patterns described earlier. The results also suggest that people look at others, more specifically others beside or behind them in a cinema theatre setting. This behaviour indicates that people not only observe others who are close by, but also people who are further away. It is common for people to attend the cinema with others with whom they share social bonds, e.g., family, friends or colleagues. Social bonds may also be an important factor that influences the magnitude of social influence. People who evacuate will most likely be influenced more by their friends, family or colleagues than by others. In the present study, an in-depth analysis of these aspects has not been performed. This type of analysis would ideally require more data, i.e., more similar experiments. However, the results show that the absolute pre-movement time difference for two persons sitting next to each other was on average smaller for friends than for persons who were not friends in the experiments. This result is believed to be an indication that people are influenced more by friends than by others. One limitation of the present study is that only experiments in which the cinema was at least half full were included in the analysis. This means that the cinema theatre was relatively densely populated and that visitors were sitting close to one and other. Hence, the results are mainly applicable to situations where visitors are sitting close to each other and may not necessarily be valid for sparsely populated spaces. Future research could therefore focus on examining how people are influenced by others if the population density is lower. Future studies that aim at exploring social influence beyond the scope of this article could be focused on examining more experiments. Preferably a new series of unannounced experiments in cinema theatres or assembly halls should be performed. The total number of identical experiments should be fairly high, thus enabling analysis of the variation of the recognition and pre-movement time between different evacuations. Another possible expansion is to use larger cinema theatres or assembly halls in order to analyse the suggested wave patterns more thoroughly. The study suggests that people are influenced by others during the initial phase of a fire evacuation. This influence is called social influence and is more important when the information is limited, e.g., for ambiguous fire cues. Results also indicate that social influence increases with decreasing distance between people. This implies that individuals are influenced more by people who are close than by people who are further away. The authors would like to thank Senior lecturer Dr. Håkan Frantzich at the Department of Fire Safety Engineering, Lund University, for his useful advice and recommendations and Dr. Wendy Saunders for her assistance and valuable suggestions. Special thanks also go to the Transportation Econometrics and Traffic Modelling group at Dresden Technical University for the interesting discussions regarding analysis and simulation of social influence. The authors would also like to thank Professor Igor Rychlik and Dr. Finn Lindgren at the Centre for Mathematical Sciences, Lund University, for their help concerning random permutation.