1 PREPRINT VERSION – NOT PEER REVIEWED

2 Nature and determinants of social actions during a mass shooting

3 Guillaume Dezecachea,b, Jean-Rémy Martinc, Cédric Tessierd, Lou Safrae,f, Victor Pitrona,g,
4 Philippe Nussh,i & Julie Grèzese
5
a
6 Institut Jean Nicod, Département d’Etudes Cognitives, ENS, EHESS, CNRS, PSL Research
7 University, Paris, France;
b
8 Université Clermont Auvergne, CNRS, LAPSCO, Clermont-Ferrand, France;
c
9 Université Libre de Bruxelles, Center for Research in Cognition and Neurosciences,
10 Brussels, Belgium;
d
11 ERL 1157, Centre Hospitalier Universitaire Saint-Antoine, Paris, France;
e
12 Laboratoire de Neurosciences Cognitives et Computationnelles, INSERM U960,
13 Département d’Etudes Cognitives, ENS, PSL Research University, Paris, France;
f
14 CEVIPOF, Sciences Po, Paris, France;
g
15 Assistance Publique-Hôpitaux de Paris, Department of Psychiatry, Pitié-Salpêtrière Hospital,
16 Paris, France;
h
17 UMR 7203, Laboratoire des biomolécules, Sorbonne Universités-UPMC Université Paris 06,
18 Paris, France;
i
19 Sorbonne Université, Paris Univ-06, Paris, Service de Psychiatrie et de Psychologie
20 médicale, Hôpital Saint-Antoine, AP-HP, Paris, France.
21
22 *Corresponding author: guillaume.dezecache@gmail.com

1
23 If collective reactions to danger have long been portrayed as antisocial and self-

24 preservative, research has shown that prosociality is maintained and sometimes fostered

25 in life-threatening circumstances. In this research, we interviewed 32 survivors of the

26 attacks at ‘Le Bataclan’ (on the evening of 13-11-2015 in Paris, France) with the aims of

27 offering a typology of social behaviors undertaken in the context of a mass shooting in a

28 close environment, and examining the situational factors favoring the emergence of

29 socially supportive behaviors among hostages. Consistent with previous findings, we

30 found that socially supportive behaviors were frequently reported, and they were more

31 common than socially unsupportive behaviors in the narratives of the respondents. We

32 also found that the occurrence of socially supportive actions is dependent on key

33 contextual factors, namely the impossibility to egress, increased distance and/or minimal

34 protection from the source of danger, and emotional fusion with other crowd members.

35 Finally, although supportive behaviors can be served by a variety of motives

36 (individualistic, cooperative or altruistic), we found that supportive actions were more

37 often described by respondents as reflecting genuine altruism, i.e. involving a cost to

38 oneself at the benefit of others. Those results bring evidence of the maintenance of socially

39 supportive actions in the context of a mass shooting. It also calls for establishing a clear-

40 cut distinction between the social strategies (asocial, socially supportive, or socially

41 unsupportive) and the psychological motivations underlying them (individualism,

42 cooperation or genuine altruism) in future research.

43

44 Main text

45 Popular belief holds that, upon perceiving danger, humans panic and preferentially display

46 antisocial behavior (1). Against this view, a wealth of research based on testimonies from

47 survivors of disasters have shown that generalized panic is rare and that socially supportive

2
48 responses, that is, those that may actively benefit others - as opposed to unsupportive ones, are

49 common in people collectively confronted by deadly threat (2). Supportive responses include

50 reassuring others, helping injured individuals and/or providing them with critical information

51 (3). In the literature, the emergence of supportive responses in deadly endangered individuals

52 is typically accounted for by three, possibly complementary, theoretical proposals. First, social

53 conventions and norms (such as the respect for physically weaker people) are maintained in

54 emergency situations (4). Second, affiliation and physical contact with others (preferentially

55 familiar ones) is a primitive response to danger in humans and other animals (5). Finally, the

56 perception of a common fate causes the adoption of a common social identity which in turn

57 establishes pro-social norms among group members (6). If this literature has debunked the myth

58 of “mass panic” and that of the prevalence of antisocial behaviors in emergency contexts (1),

59 both the situational features allowing for the emergence of supportive behaviors in life-

60 threatening circumstances as well as their underlying motivations remain to be further explored.

61 In this research, our first objective was to gather empirical elements to form a typology of

62 individual reactions to threat, with a focus on socially supportive and unsupportive behaviors

63 in the context of a mass shooting in a close environment (a music concert hall).

64 Our second objective was to examine the contextual factors that may help explain the

65 emergence and maintenance of supportive (vs. unsupportive) behavior. We envisaged that two

66 physical factors should be prominent in shaping socially supportive behavior in the face of

67 deadly threat. One is distance to the danger, e.g., being under fire or not. Another is the

68 possibility to escape or to move at all. The literature also indicates that socio-psychological

69 factors matter: emotional fusion with others (or at least, the adoption of a common social

70 identity) enhances prosocial tendencies in disaster contexts (6). Also, the presence of familiar

71 individuals and therefore the possibility to immediately recruit or offer social support among

72 friends and/or relatives should enhance prosocial tendencies.

3
73 Our third and final objective was to go beyond previous reports by exploring the potential

74 psychological motivations underlying socially supportive behaviors. Supportive behaviors

75 could either be sustained by individualistic (the welfare of the agent being the only thing

76 relevant while the welfare of others is ignored), cooperative (mutual benefit and welfare for the

77 agent and the recipient) or altruistic (the agent suffers a direct net cost at providing support and

78 welfare to the recipient) motivations.

79 On 13 November 2015, at around 9.40pm, and during the concert of the rock band Eagles of

80 Death Metal attended by around 1,500 persons, a group of 3 gunmen entered the Bataclan

81 concert hall in the 11th arrondissement of Paris (France). After targeting people standing next

82 to the bar and the merchandising stand, one of the terrorists progressed into the pit, while

83 another climbed the stairs to reach the balcony, soon followed by the third terrorist (Figure 1 –

84 red dots and lines). The first terrorist was killed at 9.57pm, after two policemen intervened. The

85 attacks ended at around 00:20am after the deaths of the two remaining gunmen who had taken

86 hostages upstairs (7). Ninety people died and hundreds were injured (7, 8).

87 Between June 2016 and November 2016 (~7-to-12 months after the events), we interviewed 32

88 survivors from the attack, based on voluntary participation. The 32 respondents were inside the

89 Bataclan concert hall during the attack and had later been in contact with associations of victims

90 of the Paris attacks of 13 November 2015. They were recruited through those associations.

91 Among the respondents, 23 were at the ground floor and 9 at the first floor at the start of the

92 attack. Some of the respondents were able to flee toward the main exits, or to one of the

93 dressing, bath, or technical rooms. Others remained in the pit until they were freed by the police

94 and/or the French Special Forces. Figure 1 recapitulates the initial spatial positions of the

95 respondents (green dots) as well as the initial positions and likely trajectories of the terrorists

96 before they separated (red dots and lines).

4
97 The 32 interviews (mean recorded duration of 105 minutes +/- 24 SD) were semi-directed and

98 based on a pre-established questionnaire. As we anticipated that the narration would trigger

99 strong emotions in respondents, one professional mental health practitioner was present and

100 leading the interview. To make sure respondents felt at ease, they were told that they could skip

101 any question that felt inappropriate and terminate the interview at all time. They were also free

102 to express any distressing thoughts.

103 We systematically inspected recordings of interviews for social episodes, defined as part of the

104 narration where an action is described, undertook by a person in the Bataclan (hereafter, the

105 “agent”) and involving at least one recipient or co-actor. The agent could be the respondent to

106 the interview herself, or somebody else undertaking an action. A total of 426 social episodes

107 were found (mean = 13.31 social episodes per respondent, range = 6 - 25).

108 Typology of social actions during the attacks

109 Among the 426 social collected episodes, 290 were clearly identified as supportive towards

110 others (‘supportive’ actions) and were reported by all 32 respondents. On the other hand, 121

111 social episodes could be identified as socially unsupportive and/or detrimental to others

112 (‘unsupportive’ actions, reported by all but one [N = 31] respondents) (Figure 2A). Most

113 respondents reported having engaged in and observed both socially supportive and

114 unsupportive actions, suggesting that the dynamics of the attacks is critical to further understand

115 the emergence of socially supportive acts in this context.

116 Socially supportive actions

117 All respondents reported having observed, initiated or participated in supportive actions.

118 Socially supportive episodes were 2.4 times more common that unsupportive ones, confirming

119 the possibility of socially supportive actions in the context of a mass shooting.

5
120 As reported in Figure 2A, socially supportive actions included providing emotional support (72

121 episodes reported by 28 respondents; category = ‘emotional support’). In this domain, stroking

122 others and holding hands were commonly reported actions (reported by 15 respondents).

123 Informational support (or the act of providing others with critical information about the actual

124 location of the exits, of the terrorists and their movements and of the intervention of the rescue

125 services) were also commonly reported (74 episodes reported by 29 respondents; category =

126 ‘informational support’). They were used to help others locate exits and avoid dangerous

127 trajectories. They could then be specifically uttered when escape was momentarily possible,

128 notably when the attackers were reloading their weapon (reported by 5 respondents) allowing

129 for immediate defensive actions.

130 Supportive actions also involved direct physical support (129 episodes reported by 29

131 respondents; category = ‘physical support’). They consisted in providing physical help to a

132 recipient for him/her to achieve an outcome he/she could not achieve by him/herself. At least 8

133 respondents reported actively taking care of people injured (or observing it), by for instance,

134 making up tourniquets with their own clothes (reported by 3 respondents). One respondent

135 narrated committing himself/herself to protect another person and acting as a “flack jacket”

136 (“gilet pare-balles”). Other physically supportive actions took the form of collaborative

137 behavior, with 7 respondents reporting a leg-up to escape from one of the loges to the roof of

138 the Bataclan. Active coordinated physical protection also occurred in the pit where people

139 presumably tried to remain physically cohesive, a pattern described as a “kitten litter” (“portée

140 de châtons”) by one of the respondent and mentioned elsewhere as “smart piling” (“s'imbriquer

141 intelligemment”) by this same respondent. Furniture and other objects were also brought to

142 obstruct the door and hinder the progression of the terrorists, a pattern reported by 7

143 respondents.

6
144 Some of these socially supportive behaviors involving collaboration between individuals

145 supposedly require the emergence of proto-social norms (15 cases reported by 11 respondents;

146 category = ‘social norms’). In one of the loges, a vote procedure was even instituted (reported

147 by 2 respondents) to decide upon who will be in contact with the police, as well as to reach

148 collective decisions on next actions to be taken (e.g., to open the door and windows, or not),

149 with majority decisions being contested but apparently respected (reported by 1 respondent).

150 Another clear proto-social norm emerged requiring certain people having a priority access, a

151 behavior which was reported by 5 respondents.

152 Socially unsupportive actions

153 Some of the unsupportive actions (n = 69, reported by 28 respondents) involved the use of sheer

154 physical force (category = ‘force’) at the expense of others. They included stampedes, i.e.,

155 trampling and pushing others to reach the exits. Other unsupportive actions did not imply the

156 use of physical force. In 23 episodes reported by 16 respondents, commands were used to make

157 others move faster to fasten evacuation, or make them stop talking, crying or complaining about

158 their wounds (category = ‘command’). Finally, unsupportive actions also included ignoring

159 others’ call for help, their physical insecurity and discomfort (29 episodes reported by 17

160 respondents; category = ‘neglect’). One respondent told us that he/she had ignored the situation

161 of a woman who fell after having received a bullet. Refusal to accept newcomers into one of

162 the shelters was also reported within the socially unsupportive category (reported by 3

163 respondents).

164 Contextual factors associated with socially supportive and unsupportive actions

165 The distribution of socially supportive and unsupportive actions could vary according to various

166 factors. The primary contextual factors analyzed in this study were distance to the threat (factor

167 UNDER FIRE: under fire or not), possibility to move and escape (factor ESCAPE: yes/no),

168 social support (factor SOCIAL SUPPORT: the number of familiar individuals the respondent

7
169 attended the concert with) and sense of emotional fusion with others (how much respondents

170 felt connected to others) (factor FUSION: on a scale from A/0 = no fusion, to E/5 = complete

171 fusion).

172 When looking for the contextual distribution of supportive and unsupportive actions in all

173 episodes that could be described as such (regardless of whether they were reported to be

174 produced by the respondent or not), we first compared models incorporating factors UNDER

175 FIRE and ESCAPE, and a null model (intercept only). Factors FUSION and SOCIAL

176 SUPPORT were excluded here because they were strictly associated with actions for which the

177 agent was the respondent to the interview. The best model to account for the data incorporating

178 the additive effects of both variables was nearly 25 times better than a null model (intercept

179 only) (Akaike weights = 0.499 vs. 0.020 respectively) and marginally better than models with

180 each variable alone (Akaike weights = 0.499 vs. 0.294 for a comparison with an ESCAPE factor

181 only model; 0.499 vs. 0.187 for a comparison with a UNDER FIRE factor only model) (See

182 Supplementary Information). This suggests that both the possibility to egress and distance to

183 the danger are together critical in shaping people’s preferred action strategies. In particular, this

184 best candidate model showed that, in episodes when respondents were not under fire, the actions

185 reported were more likely to be supportive than when they were under fire (B = -0.42 +/- 0.24

186 SE) (Figure 2B). Additionally, when escape was physically impossible (movement was

187 hindered), the actions reported were more likely to be supportive than when escape was

188 physically possible and movement was not hindered (B = -0.49 +/- 0.24 SE) (Figure 2B).

189 When focusing on actions respondents took part in (n=161), we compared models incorporating

190 factors UNDER FIRE, ESCAPE, SOCIAL SUPPORT, and FUSION, and a null model

191 (intercept only). This yields to a model selection process with 15 competing models (see

192 Supplementary Information). The best model incorporated the additive effects of factors

193 ESCAPE and FUSION. This model was 16 times better than a null model (intercept only)

8
194 (Akaike weights = 0.272 vs. 0.017) and the other competing models by at least a factor 1.48

195 (see Supplementary Information). Importantly, it was also better than models incorporating

196 ESCAPE or FUSION only (Akaike weights = 0.067 and 0.021 respectively). Under this model,

197 reported actions were more likely to be supportive when escape was impossible (B = -1.06 +/-

198 0.40 SE) (Figure 2C). They were also more likely to be supportive as respondents felt

199 emotionally fused with others (B = 0.35 +/- 0.16 SE) (Figure 2C).

200 The motivation behind socially supportive actions

201 If all of the socially supportive behaviors narrated by our respondents were actions that were

202 presumably meant to immediately benefit others, they were not all reported as being the

203 expression of a genuine care for others’ welfare. We indeed found reports of provision of

204 emotional or physical comfort to keep them silent (and avoid detection by the terrorists), to

205 prevent them from hindering ongoing cooperative processes, or even to reassure oneself

206 (reported by 7 respondents). This observation led us to re-examine as a secondary objective the

207 way respondents described the likely motivation underlying their own supportive action.

208 Episodes describing supportive actions were then classified as suggesting individualistic

209 (episode when the welfare of the agent is the only thing being relevant, and the welfare of others

210 being ignored or deemed irrelevant, and presented as such by the narrator), cooperative (episode

211 when an action is undertaken by the agent which benefits both himself/herself and the other

212 agent and presented as such by the narrator) and altruistic tendencies (episode when an action

213 is undertaken by the respondent which benefits the other agent only and at cost for the

214 respondent, and reported as such by the narrator).

215 Among the 123 socially supportive episodes for which one of the agents was the respondent, 8

216 were classified as served by individualistic motives, 50 were instances of cooperation, and 65

217 were described as altruistic (Figure 2D).

218 Discussion and conclusion

9
219 Against popular belief (1), numerous studies have shown that supportive behaviors are at play

220 in life-threatening circumstances (9, 6, 10, 11). Our own data corroborate those findings by

221 showing that supportive actions are reported by survivors of a mass shooting in an indoor

222 environment. All 32 respondents reported having taken part or observed a supportive action

223 with the period encompassing the terrorists’ entrance in the Bataclan up to evacuation.

224 Our results also revealed that supportive actions are favored by particular situational features,

225 notably protection from immediate danger (for instance, being behind a wall or in a loge), and

226 physical impossibility to move and egress. For example, we found reports suggesting that

227 groups of survivors having taken refuge in a loge, and minimally protected from the fire,

228 engaged in direct democratic voting in order to reach fast consensus possibly related to common

229 social values and habits. This suggests that unfamiliar individuals may be able to rapidly self-

230 organize in adversity (11), but that it may preferentially require a minimal sense of safety.

231 Socially supportive action strategies seem therefore primarily invested when minimal safety is

232 met, when escape is not possible and when movement is hindered. The fact that the possibility

233 to egress is a critical factor in the decision to invest in supportive behavior could suggest that

234 socially supportive action decisions may only be preferred when other potentially rewarding

235 strategies (such as fleeing and escaping with neglect for others) cannot be undertaken. This

236 assumption was only partly confirmed by our analysis, since socially supportive actions

237 remained otherwise prevalent.

238 The investment in socially supportive strategies also appears to depend on fusion with other

239 crowd members. Indeed, the probability to report supportive social strategies is positively

240 associated with higher fusion with others. If no clear causal relationship can be established here

241 (whether respondents did feel fused with others and therefore preferentially invested in socially

242 supportive actions, or whether they reported higher fusion because they felt socially supported),

243 this finding is consistent with one of the dominant explanations of the maintenance of

10
244 supportiveness in risky crowd circumstances, which states that exposure to danger fosters

245 supportive norms through perceived common fate and associated shared social identity (6, 11,

246 12).

247 The present research further revealed that socially supportive actions can reflect a diversity of

248 psychological motivations and associated payoff structures, namely individualism (benefit for

249 the agent and cost to the recipient), cooperation (benefit for both the agent and the recipient) or

250 altruism (cost to the agent and benefit to the recipient). Such a trichotomy, central in the

251 biological sciences to explain the evolution of social behavior (13), needs to be integrated in

252 the study of social responses to disasters. Indeed, comforting others to promote one’s own safety

253 (i.e., reassuring someone else to make him/her quiet and avoid detection by the terrorists)

254 crucially differs, motivationally speaking, from comforting others for their welfare. In the

255 present study, the psychological motivation underlying reported socially supportive actions

256 could only be captured by relying on the way respondents described the distribution of costs

257 and benefits associated with the action they themselves performed. We found that overall, the

258 score of supportive actions was relatively high, suggesting that supportive episodes were

259 associated with the report of cooperative and altruistic motives. This analysis (and the

260 possibility that socially supportive actions are in fact served by individualistic or self-

261 preservative motives) should encourage future research to further refine the existing typologies

262 of prosocial acts in threatening contexts and reveal the influence of situational features on

263 altruistic tendencies.

264 Another crucial factor to explain the presence of socially supportive actions is time from threat

265 detection. According to influential reports in behavioral economics, quick and spontaneous

266 social decisions (as opposed to late decisions) would tend to be costly to the agent and beneficial

267 to others (14, 15), reflecting core altruistic tendencies in spontaneous and immediate behavioral

268 decisions. In previous analysis, it has been shown that part of the pre-evacuation responses to

11
269 perceived danger in survivors of the WTC attacks in 2011 involved informing relatives,

270 sometimes via phone calls (16). This would suggest that socially supportive actions are

271 spontaneously preferred. No such data could be offered here. Some respondents mentioned that

272 they quickly realized that phone rings, beep and even screen lights could trigger shooting in

273 their direction. Indeed, and although our interviews were designed to incorporate a time element

274 (as mentioned in the Method section of this article) to help us evaluate the impact of the

275 perceived danger (whether they felt that their acquaintances and/or themselves were in danger),

276 it was difficult to relate the narrated episodes to one particular time segment in the experience

277 of the respondents, making time elapsed since the beginning of shooting a non-reliable variable.

278 Future studies should take this critical variable into consideration, by segmenting the interviews

279 in a more discrete fashion, frequently reminding the respondents that they should narrate their

280 experience, one moment at a time.

281 We must acknowledge other limitations: first, and beyond the obvious issue of self-selection in

282 our sample and the low representativeness of the Bataclan population (we interviewed 32

283 respondents, a mere 2% or so of the persons present at the Bataclan that evening), respondents

284 to our study were generally highly-educated (more than 80% of them received University

285 education) and of higher socioeconomic status (around 60% of the respondents are senior

286 managers or engaged in professional occupations), leading to demographics homogeneity in

287 our sample (see the Respondents’ demographics table in Supplementary Information) and poor

288 representativeness of the general population. This may be due to our recruitment methods (we

289 recruited the respondents through two associations) but also to the general demographics of

290 respondents to such sociocultural events. It might therefore be unclear whether our findings

291 could be generalized to other populations. This being said, socially supportive actions have

292 been reported in emergencies involving more heterogeneous and socially unstructured

293 populations (6, 11), such as commuters in London, United Kingdom (11).

12
294 Second, it should be reminded that our measures strictly reflect what was reported by the

295 respondents during the interviews, with potential biases caused by memory consolidation.

296 Nevertheless, our methodology avoided certain of the methodological issues pertaining to the

297 use of self-reports gathered from published reports in the media, as seen in several publications

298 (3, 11, 16) (or from other indirect sources (9)). We indeed had control over what was asked to

299 the respondents, and we had the possibility to comprehensively analyze what the respondents

300 reported. Moreover, and despite memory biases, our analysis on the distribution of supportive

301 and unsupportive actions according to situational factors (distance to danger, possibility to

302 egress etc.) should remain valid as both types of actions should, in principle, be equally mis-

303 recollected in the various situations. Sensitivity to recollection biases seem more problematic

304 for our fusion measure though, which may directly be impacted by the supportive or

305 unsupportive nature of the actions being recollected and reported.

306 An additional methodological problem with self-reports had to do with self-presentation issues.

307 One way of circumventing this problem (in this study, as in others, e.g., (6, 11)) is to let

308 respondents report situations where the respondents were mere observers of social episodes and

309 not directly involved in them. As a matter of fact, supportive actions were more frequently

310 attributed to others (58% of the time) but so were unsupportive actions (attributed to other 69%

311 of the time).

312 Despite those limitations, our research is unique in that it offers a typology of social behaviors

313 in the context of a recent mass shooting and in the close environment of the Bataclan concert

314 hall. Our research confirms that socially supportive behaviors remain robust in life-threatening

315 circumstances, but that they could be reinforced by protection from immediate danger,

316 possibility to move and egress, as well as social and emotional identification with other crowd

317 members. Finally, it raises the urgent issue that some of the behaviors usually described as

318 socially supportive could be served by self-preservative motives. Future research should offer

13
319 fine-grained typologies based on the critical distinction between the strategies and actions

320 undertaken by people in life-threatening situations and the motivational states underlying them.

321

322 REFERENCES

323 1. Schweingruber D, Wohlstein RT (2005) The Madding Crowd Goes to School: Myths about
324 Crowds in Introductory Sociology Textbooks. Teach Sociol 33(2):136–153.

325 2. Drury J, Reicher S (2010) Crowd Control. Sci Am Mind 21(5):58–65.

326 3. Proulx G, Fahy RF (2004) Account analysis of WTC survivors. Proceedings of the 3rd
327 International Symposium on Human Behaviour in Fire, Belfast, UK, September, pp 01–03.

328 4. Johnson NR (1987) Panic and the Breakdown of Social Order: Popular Myth, Social Theory,
329 Empirical Evidence. Sociol Focus 20(3):171–183.

330 5. Mawson AR (2005) Understanding mass panic and other collective responses to threat and
331 disaster. Psychiatry 68(2):95–113.

332 6. Drury J, Cocking C, Reicher S (2009) Everyone for themselves? A comparative study of crowd
333 solidarity among emergency survivors. Br J Soc Psychol 48(3):487–506.

334 7. Fenech G, Pietrasanta S (2016) Rapport fait au nom de la commission d’enquete relative aux
335 moyens mis en œuvre par l’Etat pour lutter contre le terrorisme depuis le 7 janvier 2015
336 Available at: http://www.assemblee-nationale.fr/14/pdf/rap-enq/r3922-t1.pdf.

337 8. Attentats du 13 novembre 2015 en France - Wikiwand Available at:

338 http://www.wikiwand.com/fr/Attentats_du_13_novembre_2015_en_France#/Th%C3%A9%C3
339 %A2tre_du_Bataclan [Accessed September 17, 2018].

340 9. Johnson NR (1987) Panic at" The Who concert stampede": An empirical assessment. Soc
341 Probl:362–373.

342 10. Frey BS, Savage DA, Torgler B (2010) Interaction of natural survival instincts and internalized
343 social norms exploring the Titanic and Lusitania disasters. Proc Natl Acad Sci 107(11):4862–
344 4865.

345 11. Drury J, Cocking C, Reicher S (2009) The nature of collective resilience: Survivor reactions to the
346 2005 London bombings. Int J Mass Emergencies Disasters 27(1):66–95.

347 12. Alnabulsi H, Drury J (2014) Social identification moderates the effect of crowd density on safety
348 at the Hajj. Proc Natl Acad Sci:201404953.

349 13. West SA, Griffin AS, Gardner A (2007) Social semantics: altruism, cooperation, mutualism,
350 strong reciprocity and group selection. J Evol Biol 20(2):415–432.

351 14. Rand DG, Greene JD, Nowak MA (2012) Spontaneous giving and calculated greed. Nature
352 489(7416):427–430.

353 15. Rand DG, et al. (2014) Social heuristics shape intuitive cooperation. Nat Commun 5:3677.

14
354 16. Blake S, Galea E, Westeng H, Dixon AJP (2004) An analysis of human behaviour during the
355 WTC disaster of 11 September 2001 based on published survivor accounts. Proceedings of the
356 3rd International Symposium on Human Behaviour in Fire, p 3.

357 17. Sheikh H, et al. (2014) The Devoted Actor as Parochial Altruist: Sectarian Morality, Identity
358 Fusion, and Support for Costly Sacrifices. Cliodynamics J Quant Hist Cult Evol 5(1). Available
359 at: http://escholarship.org/uc/item/7170h7h0

360 18. Landis JR, Koch GG (1977) The Measurement of Observer Agreement for Categorical Data.
361 Biometrics 33(1):159–174.

362 19. Bates D, Mächler M, Bolker B, Walker S (2014) Fitting linear mixed-effects models using lme4.
363 ArXiv Prepr ArXiv14065823.

364 20. Kuznetsova A, Brockhoff PB, Christensen RHB (2017) lmerTest package: tests in linear mixed
365 effects models. J Stat Softw 82(13).

366 21. Bartoń K (2017) MuMIn: Multi-Model Inference Available at: https://cran.r-
367 project.org/web/packages/MuMIn/index.html

368 22. Wagenmakers E-J, Farrell S (2004) AIC model selection using Akaike weights. Psychon Bull Rev
369 11(1):192–196.

370

371 METHODS

372 Ethics Permission to conduct the study was received from the Comité de Protection des

373 Personnes (CPP) of the Centre Hospitalier Universitaire Saint-Antoine, Paris.

374 Respondents Respondents (see Supplementary Information for a summary and demographics)

375 were recruited from an announcement transmitted to two associations of victims of the terrorist

376 attacks of the 13 November 2015, and that was broadcasted to their members. Volunteers were

377 invited to contact us by email. We then planned an appointment to be held at the Centre

378 Hospitalier Saint-Antoine in Paris. Respondents did not receive any monetary benefits for their

379 participation, but their transportation costs were reimbursed if they wished so. Note that we

380 considered all the individuals that were willing to participate to the study (whom had contacted

381 us by email) between June 2015 and November 2016.

382 Procedure Before setting up the interview, a clinical questionnaire (Posttraumatic Stress

383 Disorder Checklist Scale (PCL-S)) was sent to potential respondents by email to evaluate

384 clinical manifestations associated with a post-traumatic stress disorder. The mean score of

15
385 people who did participate in the study was high (51.5). Only 8 respondents scored under the

386 detecting threshold of 44 that provides 97% sensitivity and 87% specificity of meeting criteria

387 for PTSD. Upon receiving their responses to the questionnaire, an appointment was taken. Note

388 that one respondent did not fulfill the questionnaire before the interview. Being part of an

389 association of victims of the terrorist attacks, respondents benefited in a wide range of support

390 and opportunity of treatment. Their participation in the survey was thus not an opportunity to

391 treatment access, and most of them benefited from a specific treatment either supportive

392 psychotherapy or medication. Because of its involvement in support of victims of the 13

393 November terrorist attack, one of the co-authors (PN) has developed a strong relationship with

394 the above-mentioned associations of victims. No report of symptom worsening as a

395 consequence of the interviews has ever been mentioned to him as of today.

396 On the day of the interview, respondents first met with PN, a psychiatrist, to run a general

397 psychiatric screening and judge whether participation to the study was appropriate, trying to

398 anticipate potential distress reactions. A semi-directed interview (questionnaire available upon

399 request to GD) was then conducted and led by either CT and/or VP, professional health care

400 practitioners, with the presence of GD and/or JRM. Respondents were first briefed about the

401 structure of the interview: first, some demographic questions were asked (age, sex, education

402 level, occupation, marital status and number of children) along with contextual information on

403 their situation the evening of the attack (with whom they were, whether they had been hit by a

404 bullet and whether they had to undergo surgery after the event), followed by a focus on three

405 specific time-periods that respondents were invited to identify in their own narratives: first,

406 when they realized something serious was happening but they did not know what it exactly was

407 yet; second, when they realized it was a terrorist attack; third and finally, they were invited to

408 narrate the rest of their time in the Bataclan, ideally by focusing on a key moment (for instance,

409 the denouement). For each of those specific time-periods that were clearly identified by the

16
410 respondent, they were asked about their location on a map of the Bataclan, that of the people

411 they came with at the concert and the likely position of the threat/terrorists. For each time-

412 periods, a series of questions were asked on their spontaneous reaction, their behavior towards

413 others and their communication with people inside and outside the Bataclan. Respondents were

414 also invited to digress and questions that were not relevant to the situation of the respondent

415 could be omitted. Finally, and for each of the time-periods, respondents were asked about their

416 feeling of fusion to familiar others at the Bataclan, familiar others not present at the Bataclan,

417 the rest of the crowd and the music band, using a scale inspired by (17), from A (no fusion with

418 others) to E (complete fusion with others). Since the time-periods did not apply clearly to all

419 respondents, they were not considered for further analysis.

420 Finally, respondents were asked about their sense of control over the events and their sense of

421 time, before narrating freely about the aftermath of the attacks. They were then debriefed about

422 the aims of the study, asked not to communicate about the aim of the study to other members

423 of their association, and thanked for their participation. Note that respondents were invited to

424 contact PN (Head Psychiatrist of the research team) if they felt the urge to do so (4 of them did

425 so). Two respondents contacted PN for psychological support several months after the

426 interview. None of them fulfilled criteria for PTSD.

427 Data processing Authors GD, JRM and JG shared the listening of the audiotapes and collected

428 each social episode (i.e., independent narrative episode presenting at least one agent performing

429 an action, with a clear effect on at least one recipient) with a brief summary, timestamp and

430 categorization of the immediate context. Episodes were considered if they were temporally

431 situated between (a) the beginning of the attacks and (b) the moment when they were outside

432 the Bataclan concert hall and met with somebody who was not within the concert hall. For each

433 social episode, agents could be directly under fire (UNDER FIRE = 1) or not (UNDER FIRE =

17
434 0) and escape and movements towards the exits was physically possible (ESCAPE = 1) or not

435 (ESCAPE = 0), with movement being hindered or not.

436 Each social episode was classified as being performed either by the respondent to the interview

437 or another person (the respondent was chosen as the agent if the performed action was carried

438 out by the respondent together with another person). Each of the social episodes was then

439 categorized by GD as belonging to one of the Typology categories (the definitions are shown

440 in Supplementary Information). Inter-reliability was measured on around 20% of the episodes

441 (randomly chosen) (n=86) with three external judges (native speakers of French). Agreement

442 between GD and each of the three external judges to categorize the social episodes into the

443 proposed Typology was substantial (all ks > 0.7) (18). It was almost perfect when the various

444 categories of the Typology were lumped into supportive or unsupportive categories (all ks >

445 0.9) (18).

446 All episodes were also classified as serving individualistic, mutually beneficial and altruistic

447 motivations by GD, JRM and JG, after the definitions described in the Main text (and shown in

448 Supplementary Information). Inter-reliability was measured on around 20% of the episodes

449 (randomly chosen) (n=85) with three external judges (native speakers of French). Agreement

450 between GD, JRM and JG and each of the three external judges to categorize the social episodes

451 into the proposed Typology was substantial (all ks > .6) (18).

452 Statistical analysis To investigate the contextual variables impacting on the report of

453 unsupportive vs. supportive actions in episodes where the respondents to the study were the

454 agent, or among the agents, we ran mixed logistic regressions taking unsupportive (= 0;

455 reference) and supportive (= 1) as dependent variables. Respondents’ ID was used as random

456 factor. A null model was built using intercept and the random factor. Three models were then

457 built with factors UNDER FIRE (whether respondent was or not directly under fire, given

458 her/his position in the concert hall; 1 = yes) and ESCAPE (whether Escape was physically

18
459 possible with movement not hindered, given respondents’ position in the concert hall; 1 = yes),

460 with one of the factors alone or in addition with the others. We conducted the mixed logistic

461 regressions using RStudio (v 1.1.453) and function ’lmer’ of packages lme4 (v 1.1-17) (19) and

462 lmerTest (v 3.0-1) (20). We extracted the Akaike Weights (w) of each model, from the Second-

463 order Akaike Information Criterion values (AICc), using dedicated functions of package

464 MuMIn (21) (v 1.40.4). Models were then ranked according to their Akaike weights (w)

465 reflecting their respective probability of being the best model among the candidate models (22).

466 Table of model selection is available as supplementary material.

467 To investigate the contextual variables impacting on the report of unsupportive vs. supportive

468 actions in episodes where the respondents to the study were the agent, we used unsupportive (=

469 0; reference) and supportive (= 1) as dependent variable. Respondents’ ID was used as random

470 factor. A null model was built using intercept and the random factor. A set of models were built

471 using the variables previously described (UNDER FIRE, ESCAPE) as well as FUSION

472 (calculated as the median value of all available values for the respondents for estimation of

473 fusion with the rest of the crowd) and SOCIAL SUPPORT (the absolute number of familiar

474 people respondents came with at the concert). We extracted the Akaike Weights (w) of each

475 model, from the Second-order Akaike Information Criterion values (AICc), using dedicated

476 functions of package MuMIn (21). Models were then ranked according to their Akaike weights

477 (w) reflecting their respective probability of being the best model among the candidate models

478 (22). Table of model selection is available as supplementary information.

19
479 FIGURES

480

481 Figure 1 Map of the Bataclan concert hall. Left panel is map of the ground floor. Right

482 panel is map of the first floor.

483 Green dots represent the initial positions of respondents at the start of the attacks (first time-

484 period, as defined in the Methods); red dots and lines represent the initial position (red dots)

485 and likely movements at the beginning of the attacks (red lines) of the 3 terrorists. Positions

486 and movements of the terrorists were reconstructed from respondents’ testimonies as well as

487 inspection of the report from the French Assembly (see reference 8 [pp. 51 to 61] in Main text).

488 The initial position of each of the respondent is based on their interviewed report. Note that

489 certain respondents appear to share very close location within the concert hall, leading to points

490 superposition on the Figure. Notations on the map correspond to the following terms in full

20
491 English language: Loge = loge; L.T. = control room; Toil. = toilets; Hall = lobby; Bar = drinks

492 counter; Scene = stage; Vestiaire = cloakroom; Caisse = ticket office; Rez-de-chaussée =

493 ground floor; 1er étage = first floor.

494

495

496 Figure 2 The distribution of supportive and non-supportive actions

497 Figure 2A. Distribution of supportive vs. non-supportive responses and number of reports for

498 each category.

499 Figure 2B. Effect of contextual factors on the occurrence of supportive behaviour under threat

500 for ALL narrated episodes (regardless of who was the agent). Each plot represents the predicted

501 probability of a reported action to be supportive as a function of the possibility to escape and

502 the presence of immediate firing based on the best model’s coefficients method. Both the

21
503 absence of firing and absence of the possibility to escape are associated with a higher probability

504 of the reported actions to be supportive. Error bar represent 95% confidence intervals.

505 Figure 2C. Effect of contextual factors on the participation to a supportive behavior under threat

506 by the agent. The predicted probability of an action in which the agent took part to be supportive

507 varies with the possibility to escape and the degree of emotional fusion of the agent with the

508 other individuals involved. Absence of possibility to escape and higher levels of social fusion

509 are associated with a higher probability of the actions to be supportive. Error bar represent 95%

510 confidence intervals.

511 Figure 2D. Psychological motivations underlying supportive actions for which the respondent

512 was the agent.

513

514 CONFLICTS OF INTEREST

515 The authors declare no conflicts of interest.

516

517 DATA ACCESSIBILITY

518 The questionnaire is available upon request to corresponding author (GD).

519 The raw dataset will be uploaded as online supplementary material upon final acceptance of the

520 manuscript.

521

522 AUTHORS’ CONTRIBUTION

523 GD, JRM, JG and PN designed the study; GD, JRM, CT, VP, PN collected the data; PN

524 supervised data collection; GD, JRM and JG analyzed the data; GD, LS and JG performed the

525 statistical analysis, with LS providing core analytic tools; GD and JG wrote the first draft of the

526 manuscript; All authors helped improve the manuscript and gave final approval for publication.

527

22
528 ACKNOWLEDGMENTS

529 We would like to thank the respondents and the associations Life for Paris and 13 Onze 15 for

530 their availability and their participation to the study. We are also very grateful to Aurélia

531 Gilbert, Gérôme Truc, Sandra Laugier, Federico Zemborain, John Drury, Romain Ligneul,

532 Louise Goupil and Alice Gibson for helpful discussion. This work was supported by funding

533 from the CNRS Recherche Attentats scheme, FRM Team DEQ20160334878, ANR-10-LABX-

534 0087 IEC and ANR-10-IDEX-0001-02 PSL*.

23
535 SUPPLEMENTARY INFORMATION

536 Table of respondents’ demographics

537 *Level of education calculation is based on 4 categories: BAC, BAC+3 (with BAC+2 being

538 merged in this category), BAC+5 (with BAC+4 being merged in this category); Higher than

539 BAC+5 for individuals with education after level BAC+5. Note: BAC refers to the French

540 baccalaureate: https://en.wikipedia.org/wiki/Baccalaur%C3%A9at

541 **Socioprofessional categories are based on French conventions. They roughly translate as

542 follows:

543 - artisans, commerçants et chefs d’entreprises = craftsmen, shopkeepers, business owners

544 - cadres et professions intellectuelles supérieures = senior managers and professional

545 occupations

546 - professions intermédiaires = middle management positions

547 - employés = employees

548 - ouvriers = factory workers

549 - étudiants = students

550

24
 551
Sample (N=32; mean age in 2016 =
N % of the sample
36.18 y.o., +/- 1.13 SEM)

Sex ratio (female/male) 17 / 15 53.1% / 46.9%

Level of education*

BAC 6 18.75%

BAC+3 7 21.9%

BAC+5 12 37.5%

Higher than BAC+5 7 21.9%

Socioprofessional category**

Artisans, commerçants et chefs

d'entreprise 2 6.25%

Cadres et professions intellectuelles

supérieures 19 59.4%

Professions intermédiaires 3 9.4%

Employés 7 21.9%

Ouvriers 0 0.0%

Etudiants 1 3.1%

With spouse at the Bataclan 12 37.5%

With friends at the Bataclan 23 71.8%

Hit by a bullet 9 28.1%

552 Questionnaire

553 The full questionnaire (in French) can be accessed, upon request to GD

554 (guillaume.dezecache@gmail.com).
25
555 Definitions

556 FORCE: using one's physical force at the expense of others to save oneself.

557 COMMAND: agents ask recipient to do something irrespective of the recipients' immediate

558 welfare.

559 NEGLECT: agent neglects recipients' immediate welfare without using physical force.

560 EMOTIONAL SUPPORT: agent gives emotional support to one or more recipients.

561 INFORMATIONAL SUPPORT: agent gives information about the position of the terrorists, of

562 the exits, a momentarily possibility to escape or about police's intervention status.

563 PHYSICAL SUPPORT: agent gives physical support to one or more recipients.

564 SOCIAL NORMS: clear evidence of people setting up a stable activity which requires the

565 instauration of a social norm, which is explicitly mentioned.

566

567 Definitions – The motivation behind supportive actions

568 INDIVIDUALISTIC: Presence of an episode where the welfare of the agent is the only thing

569 being relevant (and the welfare of others being ignored or deemed irrelevant).

570 MUTUALLY BENEFICIAL (COOPERATION): Presence of an episode where an action is

571 undertaken by the agent which benefits both himself/herself and the other agent and which is

572 presented as such by the participant.

573 ALTRUISTIC: Presence of an episode where an action is undertaken by the participant which

574 benefits the other agent only and at cost for the agent and is presented as such by the participant.

575

576 Model selection

577 The following tables provide information about the 2 model selection procedures; for each

578 model: value of intercept, regression values for each relevant factor, df (degrees of freedom),

26
 579 logLik (LogLikelihood), AICc, delta (delta of AICc), Weight (Akaike Weight calculated out of

580 AICc – using function ‘model.sel’ of package MuMin (v. 1.40.4)).

581

582 Table of model selection for All episodes

(Intercept) UNDERFIRE ESCAPE df logLik AICc delta weight

1.35 -0.42 -0.49 4 -242.76 493.62 0.00 0.499
1.22 NA -0.63 3 -244.31 494.68 1.06 0.294
1.17 -0.59 NA 3 -244.76 495.58 1.96 0.187
0.89 NA NA 2 -248.01 500.04 6.42 0.020
583

584 Table of model selection for episodes in which agent is respondent

(Intercept) UNDER ESCAPE FUSION SOCIAL df logLik AICc delta weight

FIRE SUPPORT
0.72 NA -1.06 0.35 NA 4 -82.96 174.17 0.00 0.272
0.41 NA -1.03 0.35 0.16 5 -82.29 174.96 0.79 0.183
0.82 -0.32 -0.98 0.36 NA 5 -82.63 175.65 1.48 0.130
0.49 -0.42 -0.92 0.35 0.19 6 -81.76 176.07 1.90 0.105
1.62 NA -0.85 NA NA 3 -85.41 176.97 2.80 0.067
1.24 NA -0.81 NA 0.18 4 -84.63 177.53 3.36 0.051
1.71 -0.30 -0.76 NA NA 4 -85.13 178.51 4.34 0.031
1.32 -0.40 -0.68 NA 0.20 5 -84.16 178.71 4.54 0.028
0.71 -0.57 NA 0.26 NA 4 -85.45 179.15 4.98 0.022
1.02 -0.61 NA NA 0.23 4 -85.53 179.32 5.15 0.021
0.49 NA NA 0.25 NA 3 -86.59 179.33 5.16 0.021
0.16 NA NA 0.23 0.18 4 -85.62 179.49 5.32 0.019
1.20 NA NA NA NA 2 -87.80 179.67 5.50 0.017
0.79 NA NA NA 0.20 3 -86.76 179.68 5.51 0.017
1.45 -0.53 NA NA NA 3 -86.84 179.84 5.67 0.016
585

27