Gustav Friis
#Terrorism
A study of audience reactions to terrorist attacks
Table of content
List of tables and figures
1. Introduction ... 1
1.1 Previous research ... 2
2. Theory ... 3
2.1 Terrorism ... 4
2.2 Social identity theory ... 5
2.3 The expected relationship ... 6
3. Research design ... 8
3.1 Case selection ... 9
3.2 Data collection and source criticism ... 9
3.2.1 Twitter and its search policies ... 9
3.2.2 The Global Terrorism Database ... 11
3.3 Operationalisation ... 12
3.3.1 Independent variable ... 12
3.3.2 Dependent variable ... 14
3.3.3 Control variables ... 15
3.4 Scope conditions ... 16
3.5 Validity & Reliability ... 17
4. Results and Analysis ... 19
4.1 Analysis ... 23
4.2 Alternative explanations ... 24
5. Conclusion ... 25
6. Bibliography ... 26
7. Appendix ... 31
7.1 Appendix I ... 31
7.2 Appendix II ... 32
7.3 Appendix III ... 33
Figure 2. 1 Expected relationship, outgroup 7
Figure 2. 2 Expected relationship, ingroup 8
Table 3. 1 List of ingroup traits 12
Table 3. 2 List of response categories 14
Table 3. 3 Complete list of search filter 15
Table 4. 1 The effect of group levels on levels of radicalism 20
Table 4. 2 The effect of group levels on radical category 5 21
Table 4. 3 The effect of group levels on radical category 4+5 22
1. Introduction
In a review article of the study of terrorism five areas of research interest are found by Sandler (2014). Although this article does not fall under any of those it is part of a newer niche examining the differences between domestic and international terrorism (Ibid).
The field of terrorism is not unfamiliar with the concept of social media. It is well known that social media and other internet tools can be used to organise people to perform both violent and peaceful acts. What is not as well understood is how effective terrorism is, what impact it really has. In order to create a society resilient to terrorism it is important to understand what effects it has on the witnesses, the audiences. This paper attempts to examine how we react to terror attacks by using twitter to find expressed opinions desiring change after terrorist attacks. Understanding how we use using social media to demand change will not only help us deal more effectively with recuperating but will also facilitate in working against polarisation.
How does group membership of terrorist attackers affect audience reactions? This is the question at the core of the paper. The proposed theory is that terrorists from outgroups, not part of the attacked society, will elicit more radical responses from audiences than do attackers from the ingroup, part of the attacked society.
In this study the Global terrorism database (GTD) was used to find events of terror attacks to be analysed. This was then narrowed down to attacks in the US. The independent variable was coded using news articles about the perpetrators while the dependent variable was gathered by extracting tweets from the days after each event. Control variables were taken from the GTD dataset and control for the effects of how big the attack was in terms of casualties, whether it was a suicide or not and the type of target. Although several regressions were run, and a positive relationship was found between group levels and radical responses, no results were found to be statistically significant. This is likely explained in large part by the small number of cases and difficulties with the twitter data collection although alternative explanations surely play a role as well.
The remainder of the paper is divided into five parts. First a short introduction to previous research and the identification of a research gap. This is followed by an explanation of how terrorism affects audiences, social identity theory and the expected relationship. After this comes the research design, case selection, data collection and the operationalisation of the variables. Finally, the results are presented and discussed followed by a short conclusion.
1.1 Previous research
In this section the previous research and its identified gap will be examined. The study of terrorism as a field on its own, or rather as an intersection between different fields is relatively new and it contains a plethora of approaches (Ranstorp 2007: 32). A quick note should be made with regards to the history of terrorism in that it has existed since the ancient days but that it has taken different forms of expression in different societies and times. The most recent trend seems to be that of the religious terrorism (Law 2016: 321, Chaliand and Blin 2013: 253). According to a list of desiderata presented by Jongman one area of interest in future studies of terrorism is its impact on different parts of the population (Ranstorp 2007:
279). Since that list was published several articles on the reactions to terrorism and its psychological effects on audiences have been published (Sirin and Geva 2013, Schmid 2005, Bonanno and Jost 2010).
On the topic of terrorism outcome some have argued that it is not effective in achieving the desired policy change of the terrorist group (Abrahms 2006). Others have criticised this standpoint, however, by pointing to the 2004 Madrid (11-M) case in which the general election was affected in favour of the party promoting withdrawal of troops from Iraq, the desired change (Rose et al. 2007: 186-187). The response to such a stance argues that the 11-M case in fact is not an example of successful terrorism but rather indicates that targeting a country where civilians have “extremely firm pre-existing beliefs” (Rose et al.
2007: 191) that the terrorists hold limited policy goals may make coercion possible regardless of the target (Rose et al. 2007: 191-192). Bridging these two views one might view audience willingness to concede, i.e. terrorist success, as divided into high and low importance issues which in turn are reinforced by the level of audience hopefulness. The threshold to concede is highest with high importance issues where hopefulness is high and vice versa (Merari and Friedland 2009).
Some have taken a different approach by examining audiences’ and victims’
emotional responses to terrorism. This has been done from different angles. Firstly, the argument has been made that exposure to media coverage of terrorism violence as opposed to non-terrorism violence leads to heightened levels of anxiety and anger as well as a more hard-line stance towards the terrorists (Shoshani and Slone 2008). To this has been added that the impact of anger on opinions regarding foreign policy is relevant only when it is induced in connection to terrorism (Sirin and Geva 2013).
Others have made a similar claim in saying that terrorism produces a shift in victims towards the right on the political scale (Bonanno and Jost 2006). An addition to this is made in claiming that individuals become increasingly right wing while the right in general moves closer to the left. This is only true, however, when attacks do not exceed some thresholds of violence (Gould and Klor 2010).
Moreover, studies into terrorism in relation to social media has been made from various angles. The possibility of using social media as a means of terrorism, i.e. targeting a population by using social media and computer viruses, has been explored. The conclusion is that a mimetic virus attacking batteries spread through social media could be a future form of attack. The level of physical damage has less importance than its psychological effect on the population, the fear produced is enough to affect the public (Ayres and Maglaras 2016).
It has been found that there are certain characteristics which make messages sent by government agents on social media to be spread to larger audiences during terrorist attacks (Sutton et al. 2015) while other characteristics are found to increase audience resilience (Williams et al. 2017). Another study claims to be of use to policy makers by providing a framework for analysing twitter activity during terrorist attacks in order to facilitate quick government responses (Cheong and Lee 2011). Different arguments have been raised as to how social media can be used as a counterterrorism tool with the purpose of preventing online radicalisation (Aistrope 2016).
Still, there is a lack of literature on how audiences use social media to react to terrorism. Specifically, if it is used to demand radical political change. To examine this the social identity theory (SIT) is used to explain why responses would differ between cases.
Although the use of psychology or even social identity theory in the study of terrorism is nothing novel (Victoroff and Kruglanski 2009, McKeown and Ferguson 2016) the application of SIT as an explanation for audience reactions in terms of calls for political change has not been tested to the author’s knowledge. Expanding on the knowledge of how audiences react to terrorists on the ingroup outgroup spectrum is important if we want to prevent polarisation and understand the role social media might play in it.
2. Theory
The theory, in short, claims that terrorists part of an outgroup, identified less with, will elicit more radical reactions than terrorists part of the ingroup, identified more with.
When audiences know that the terrorist is more like them their reactions will likely be less
radical and vice versa. This section starts with a discussion of the theoretical elements:
terrorism, and how it relates to audiences, and social identity theory. Thereafter the expected relationship is explained, and the hypothesis is presented.
2.1 Terrorism
In the study of terrorism, a critique has been raised that there exist an excessive number of definitions of terrorism. In addition, too few authors thoroughly discuss their definitions, the existence of other definitions and the possible shortcomings of the various definitions (McCormick 2003). However, since the phenomenon of interest in this paper is audience reactions specifically and a previously existing definition of terrorism is used there will only be a short summary of the definition here. The Global Terrorism Database (GTD) definition of terrorism is used since data is taken from the GTD. In short, the definition is limited to non-state actors wielding violence or threats thereof to achieve some political goal or influence an audience apart from the immediate victims.1
The effects of terrorism on audiences is of interest since they are a common part of the definition of terrorism and are often part of the everyday definition of whether terrorism is successful or not. To most readers phrases like “if we change our way of life the terrorists have won” are not unfamiliar. That terrorism influences audiences’ emotional life is examined by several scholars. It is asserted that media coverage of terrorism has an emotional effect on audiences and that anger caused by terrorism can affect stances on foreign policy (Shoshani and Slone 2008: 635, Sirin and Geva 2013:723). Others argue that direct threats from terrorist leaders and the fear it causes lead in principal to audiences either urging capitulation or supporting increased pressure in a way that would not put their country at risk (Iyer et al. 2015: 646). Audiences are also affected in terms of their political standing with individuals generally moving to the right (Bonanno and Jost 2006: 322) while political parties on the right move towards the left while violence is kept below a certain threshold (Bonanno and Jost 2006: 322).
Some have found that terrorism can be successful in causing change (Rose et al. 2007:
186-187) although this is disputed. The theoretical foundation of this paper is that terrorism does affect audiences and that terrorists from other societies have a larger impact than do domestic attackers. Plainly put, the theory suggests that when an attack occurs people will
1 For a more detailed description of the GTD definition see Appendix 1
expect leaders to act and suggest a course of action. The further from the victim society the attacker is the more drastic measures can be expected.
2.2 Social identity theory
In order to explain the existence of ingroups and outgroups some deeper introduction into the social identity theory will be required. In short, it claims that we form groups which have to be distinct from other groups. The members of one group can only grow antagonistic sentiments towards the other if they are distinctly separate from one another. In addition, the feelings of belonging to the ingroup increases as outgroups are perceived as threats (Tajfel 1974: 66). These kinds of group belongings are not always overtly visible since they can be formed in a wide spectrum of sizes. Ranging from the national, or possibly bigger ingroups, to ingroups based on the preference of one painting over the other the ingroups can exist in many different and overlapping patterns (Tajfel 1974: 67-68).
Most individuals are prone to divide the social world into groups, this is called social categorisation, by which an in- or outgroup is defined as “a cognitive entity that is meaningful to the subject at a particular point in time” (Tajfel 1974: 69). This definition of a group emphasises that groups are not as strict as being registered to a school or having a gym membership but rather that they are fluid depending to the context of a particular situation.
Moreover, the concept of social identity is used to describe an individual's understanding of his or her group membership and the level of emotional importance attached to that membership (Ibid). These memberships are in turn only given importance when defined in comparison to other groups i.e. a group establishes its importance or value in relation to surrounding groups (Tajfel 1974: 72).
However, within groups people have different levels of influence. This could be visualised as a spectrum ranging from core to marginal. Depending on how well one fits with the attributes deemed relevant for the group one will have a higher or lower standing within it. This means that a person who has only a few of the attributes will be less of a fit and thus stand further to the periphery than a person who has all the attributes. Such people are called prototypes and often hold leadership roles within the group. Since prototypes embody the essence of the group they are generally trusted to do what is best for the group and can ironically enough deviate further from the group norms than less prototypical leaders. Being able to deviate from the norm can be an advantage for the leader making prototypical leaders effective (Hogg 2016: 11).
Since groups want to retain their status in relation to other groups it would logically follow that they would attempt to dissociate members who reduce the relative standing of the group. The people on the group periphery are not unlikely to be met with distrust and dislike which they might not have faced if they were not part of the ingroup. They sometimes adopt a black sheep role (Hogg 2016: 12). Some of the peripheral members, however, which highly identify with the group and value that identity highly may attempt to become more prototypical by participating in extreme behaviour in benefit of the group (Goldman and Hogg 2016: 551). An interesting finding is that there is a higher probability that people favour the ingroup rather than discriminate against the outgroup, unless the ingroup is perceived to be under threat. In such situations discrimination can be used in full force (Hogg 2016: 6, Weisel and Böhm 2015: 110).
Since this study cannot control for the plethora of group memberships through which twitter users can (dis)associate themselves from the attacker, a general ingroup relevant for the context was established. The most important ingroup in this case is that of nationality.
Since many countries lend themselves to nation building the national identity is an ingroup which is institutionalised at many levels. When members of other societies commit terrorism in the US it is the American identity which is threatened, and group characteristics become accentuated (Schildkraut 2011: 847, Huddy and Khatib 2007: 75).
In contrast, when a typical American commits an atrocity there is no equivalent threat to the nation from an outward group, rather the theory suggests it is more likely seen as an act committed by an individual. As such, the national norm, the quintessential American, is seen as the most important ingroup during a crisis. This will be elaborated in the operationalisation of the independent variable.
2.3 The expected relationship
To reiterate, the varying degrees of radicalism with which audiences call for political change is theorised to be related to levels of attacker group participation. The higher degree to which an attacker symbolises the prototype of the ingroup the less radical the calls for change will be. If a person who is part of an outgroup performs a similar act the calls for change will be more radical. In other words, the theory predicts both that attacker ingroup participation will be accompanied with less radical reactions in addition to outgroup participation being followed by increasingly radical reactions. For instance, in a situation where the core ingroup attributes are white, Christian and name of English origin a white
churchgoing Mike Miller will be closer to the ingroup prototype than a black Muslim Muhammad Amir. These two fictional characters could according to the theory perform the same act of terrorism but elicit different responses. This is explained by the social identity theory in that the closer to an ingroup prototype a person is the more he or she can deviate from the norms of the group such as not performing terrorism (Hogg 2016: 11).
The expected relationship between in- or outgroup participation is expected to come in slightly different shapes. An attacker from an outgroup would first cause the identification of the outgroup. This kind of identification is not depending on coherence between the attacker’s actual group belonging and the identified group since it is based on perceptions.
For instance, a person from a specific terror organisation might cause the audience to conclude that the outgroup posing a threat is all middle easterners, all Syrians or even all immigrants. The next step is the connotation of the group as a threat. This would cause the audience to think of the group as posing a threat rather than it being the action of a single person. By labelling the outgroup as a threat the normative restrictions on what means are acceptable to achieve security for the ingroup will be lowered and direct discrimination towards the outgroup may become acceptable. This in turn would result in more radical calls for political change. The relationship is demonstrated below in figure 2.1.
Figure 2. 1 Expected relationship, outgroup
However, if the attacker is part of the ingroup the relationship is expected to be slightly different. Firstly, an attack coming from within the group would not lead to the identification of an outgroup as a threat. Instead, it is predicted that other threats or issues are identified as the cause of the attack. These issues or threats will in turn be subjected to demands for change although the demands are expected to be less radical than in the case of an outgroup
Outgroup Attacker
Demand for radical political change
Identification of outgroup threat
Change in acceptable
behaviour
attacker. Yet, they may still appear radical to some. The relationship is demonstrated in figure 2.2 below.
Figure 2. 2 Expected relationship, ingroup
As is made clear by the above figures the greatest difference between the two is that in figure 1 the threat or issue is the identified outgroup with the result of political change being radical.
Otherwise the processes are similar. This theory leads to the following hypothesis:
H1: As levels of outgroup increases so does levels of calls for radical political change.
Null hypothesis: Levels of outgroup has no effect on the levels of calls for radical political change.
3. Research design
In this section the research design will be presented. The case selection, data collection and operationalisation will be discussed as well as reliability, validity and scope conditions. The cases were selected by using the GTD dataset and focusing on terror attacks in the US between 2006 and 2016. Data was collected using news media and twitter, the latter presenting some difficulties. The independent and dependent variables were coded as ordinal variables but were treated as continuous mainly due to their clear hierarchical nature yet limited time and experience also had some impact. The control variables, the number of killed, wounded, perpetrators, the target type and whether it was a suicide attack were already coded in the original GTD dataset. However, target type had to be reduced to four instead of 22 categories.
Ingroup Attacker
Demand for political change
Identification of threat or issue
Change in acceptable
behaviour
3.1 Case selection
The case selection in this study is twofold. Firstly, the country under scrutiny, the USA, was deemed the best case partly because it is believed to contain a variation in both the dependent and independent variables and partly because of its wide use of twitter. Although tweets can be made worldwide the prevalence of twitter in the US increases the likelihood that tweets made in relation to cases within the US comes from US nationals. (Greenwood et al. 2016). Even if foreign audiences can be affected by terrorism in the US it is believed that the greatest impact will be on US audiences due to the higher probability of having friends or family near the risk zone. Additionally, an attack on the nation can be seen as an attack on a large spanning ingroup, which likely increases the emotional impact and thus the likelihood that demands for political change are made.
The terrorist attack events were chosen by removing all cases from the GTD dataset before April 2006 since twitter was launched in late March the same year. The events were then narrowed down to the US only. From these cases, only those with lethal casualties, at least one dead, were chosen in an attempt to make sure all were of relatively similar spread in terms of publicity. One exception was made to this, a case with 29 wounded which was included to increase the number of cases since it was deemed to be of a big enough scale to be relevant despite not being lethal. This resulted in 48 cases being selected. However, one case did not have a value on the independent variable and was consequently dropped.
The time period was chosen because the topic of interest is not only how people react to terrorism but also how social media is used to express such reactions. For this purpose, twitter was deemed a suitable platform due to its character limit requiring concise expressions together with its focus on news and discussions (Kapko 2017). The time limit is thus adapted to the founding of twitter, 2006, and the most recent year of available data from the GTD, 2016.
3.2 Data collection and source criticism
The following two sections discuss the main sources of data, twitter and the GTD dataset.
3.2.1 Twitter and its search policies
It is reasonable to assume that norms regulating how twitter is used and what is tweeted about change over time. Since the study covers almost ten years of tweeting some ways to tweet in 2006 may no longer be up to date in 2016. Changes in the twitter community
is not what is under scrutiny here although it may have an effect which, if it exists, cannot be controlled for in this study.
The algorithms used to sort searches are not disclosed in detail by twitter but after some attempts using a limited number of search terms in combination it was concluded that words with higher usage frequency, e.g. terrorism, are considered to be of higher relevance than words of lower frequency rates, e.g. “kill all muslims” (Twitter FAQ 2017). Although this cannot be entirely confirmed it was indicated by the preliminary search test. Other tests revealed that it was difficult to find any useful categorisation within the dominant Category 1 since it included a wide range of other issues or expressions not in relation to group categorisation. The most prevalent of these was the spread of or search for news. It could be as simple as stating that the event had occurred, sometimes with the addition that it was terrorism or that it was not terrorism.
Another important aspect of social media is its self-imposed codes of conduct. The policy adopted by twitter prohibits language inciting violence or threats thereof on “the basis of race, ethnicity, national origin, sexual orientation, gender, gender identity, religious affiliation, age, disability, or disease” (Twitter media policy 2017). This means that most of the specific calls for violence will be removed however twitter is not watertight in removing calls for violence which can be seen by a quick search for “bomb them all”. Although the term does not specify who to bomb tweets such as “Bomb them all to glass. Then you will finally achieve peace in the Middle East” indirectly suggest who to bomb. This shows that the twitter code of conduct does not remove all content which is suggesting violence towards people.
Still, it should be noted that twitter does not remove tweets in isolation. The context of the tweet is taken into consideration before anything is removed which may explain why
“kill all muslims” or “kill all whites” reveal some interesting discussions rather than no results at all. Additionally, for a tweet to even be considered for removal it must first be reported. Although the number of reports made of a tweet is not factored into its possible removal the priority with which it is investigated is increased if a large number of reports have been filed (Twitter media policy 2017).
Another issue with collecting data from twitter without access to the twitter API (in short, the possibility to write code to help gather large quantities of tweets) is that not all data can be geographically delimited. In other words, tweets regarding terrorism in the US can be tweeted about in the UK, Canada, and Australia or by any other English speaker around the world. Since twitter is widely used in the US this did not appear to be an issue in most cases
but in some of the less well-known cases outside tweets referring to the attack or attacks abroad could not be filtered out. Since the search filter was tested on cases with much media coverage such as the relatively recent Orlando night club shooting it was not revealed how much noise (unrelated tweets) could pass through the filter in the less famous cases. This did affect the data collection but is not likely the driving factor behind the results in itself.
Lastly, the issue of internet trolls, both domestic and foreign, should be discussed.
The present study does not have an efficient way of dealing with dishonesty and users with the intent to sow conflict online. This is unfortunate but there is little that can be done since examining even a small number of twitter accounts closely to see if they are trolls would increase the workload beyond the scope of this paper. Having acknowledged their existence, the impact likely increasing from 2006 to 2016, is believed to be minimal.
3.2.2 The Global Terrorism Database
Data on terrorist events is collected from the GTD, global terrorism database, which collects information on terrorist events between 1970 and 2016 on a global scale. This database was chosen due to its contemporary data and its availability in comparison to the other big terrorism databases RAND and ITERATE. The former does not collect data later than 2009 (RAND 2017) and the latter is not available to the author. The GTD collects data on many variables such as date, year month day, country, city, attack type, success, victim type, number of perpetrators, number of fatalities etc. Although many variables are not of interest in this paper their existence in the dataset can help avoiding misconceptions due to a lack of information. If there are any irregularities the extra data may provide answers.
The GTD data is collected by examining media articles using at first software and later on manual labour. The first step is the usage of Metabase Application Programming Interface and Open Source Enterprise to gather articles through a set of keyword filters.
These are further narrowed down and sorted using “more sophisticated natural language processing (NLP) and machine learning techniques” (GTD Codebook 2017: 8) after which a team of human researchers review the remaining articles. In the end approximately 16,000 articles are reviewed per month. For an event to be included it must, in addition to the criteria briefly mentioned in the theory section, be verified by a minimum of one high quality source (Ibid). Variables removed from the dataset either contained no or too little data or were deemed irrelevant and making the dataset difficult to navigate. These are listed in appendix 3.
In order to gather information on the independent variable the GTD was used to either find the names of the perpetrators or articles covering the events in order to determine where
on the group level spectrum they fell. The articles already in the GTD dataset were often supported by other articles going more into detail on the perpetrators. In some cases, they had no clear religious tendencies, these cases were coded as not Christian.
In general, the GTD is a trusted and inclusive rather than exclusive, source. Especially in relation to countries such as the US where state terror is not employed and where there is freedom of press. However, there is a variable included for cases which are not clear cases of terrorism. Due to the low number of cases this variable could not be taken into consideration.
3.3 Operationalisation
What follows is the operationalisation of the independent and dependent variables in addition to a list of control variables. In short, the control variables are collected from the GTD dataset and are already coded as desired whereas the dependent variable is collected by performing four filtered searches on twitter for each event. The independent variable is coded by scoring the attackers based on attributes relevant to the ingroup: name, skin colour, religion and nationality.
3.3.1 Independent variable
The independent variable, group level participation, is operationalised by examining the traits of the attacker which are deemed relevant to the US national ingroup. Individuals are not only part of one group at a time, we can identify with people based on several attributes ranging from liking the same painting (Tajfel 1974: 67-68) to playing in the same football team. This means that even if two people are aesthetically as different as can be they can still feel as part of the same ingroup based on more abstract values or characteristics.
However, to keep the study within the time limits the US national ingroup is used with theoretically based characteristics as the essential traits. The traits used to express ingroup membership are listed in table 3.1 below
Table 3. 1 List of ingroup traits
1) White skin colour 2) Christian
3) US national
4) Having a western, euro-american name
Firstly, supporting the claim that nationality is relevant, it is found that Americans are more likely to find profiling acceptable on immigrants than on US citizen is by Schildkraut (2009: 75). Secondly, ingroup bias is increasingly shown along ethnical lines where there is ethnic strife (Shayo and Zussman 2011: 1483), thus name, skin colour and religion are highly relevant as well. Another reason to use a superficial feature such as skin colour is that facial recognition of both in- and outgroups is reduced after toxic group interactions, i.e. both groups are perceived to be more homogenous (Wilson and Hugenberg 2010: 1009). Of course, facial recognition is not as simple as skin colour, but it is the closest possible with the resources available. Furthermore, these traits are implicitly the core of what it means to be American (Schildkraut 2011 & 2014, Devos and Mohamed 2014, Huynh et al. 2015).
To make the study as sensitive as possible to different degrees of group membership the independent variable is coded as an ordinal variable by assigning scores to attackers depending on whether they inherit the traits or not. Each trait is coded on its own in a binary fashion where having the trait (White, Christian, US national, western euro-american name) = 1 and not having the trait = 0. This results in a scale from 0 to 4 with the highest degree of ingroup at 4 and the highest degree of outgroup (lowest degree of ingroup) at 0. This was later inverted to give a more natural progression with highest levels of outgroup at 4 and lowest at 0.
To evaluate each attacker, the name of the perpetrator was first identified by either looking in the GTD dataset where some names are mentioned or by searching for the attack online, normally by copy pasting the name of an article cited as a source in the GTD dataset.
The traits were then evaluated by the author although information on all traits was not always available or easy to interpret. In general, the coding was straightforward apart from some instances where no information was available on the beliefs of the perpetrator. In cases where no information could be found that described the attacker as Christian in a trustworthy way it was coded as not Christian.
This kind of division is of course crude and not without problems and limitations.
According to the traits above, a white Canadian atheist can for instance be part of the same level of outgroup as a black American Muslim or a white Arabic Christian even though these three might intuitively be theorised to result in different reactions. Still, the coding provides as much variation within the independent variable as is possible within the scope of the study.
3.3.2 Dependent variable
The dependent variable is measured by analysing tweets made on the day of a terror event and the three following days with 25 tweets analysed for each day, resulting in 100 tweets per event. To categorise these tweets a spectrum consisting of five steps, referred to as categories, was created to match different levels of radicalism in the responses to the terror event. The steps are based on expected group behaviour, along a spectrum of ingroup outgroup radicalism. The categories are listed in table 3.2 below.
Table 3. 2 List of response categories
1) Non-issue/other
2) Identifying an outgroup
3) Demanding more security for the ingroup 4) Discriminating against the outgroup 5) Suggesting violence against outgroup
Due to the vastness of tweets being made after more recent events, as compared to those in 2006, several search terms are used as a filter to find tweets which might be categorised into one of these categories. This was done to establish covariation and maintain data collection within reasonable limits rather than to examine the proportion of twitter users that express such sentiments compared to the total number of tweets. The search terms were derived to match both the cause of the tweets, i.e. the terrorist attack, and the spectrum of responses. Refraining from doing so, using no search term or only “terrorism”, would have resulted in finding too many irrelevant tweets such as tweets referring to terrorism in other parts of the world or simply general discussions of terrorism. That is not what is sought after since interest lies in reactions to terrorist events, not opinions on terrorism in general.
These search terms are of course not exhaustive nor are they the only possible expressions of the categories, but they are deemed the most relevant in capturing the essence of each category with the resources available. Higher or fewer numbers of search terms without assistance from advanced software would likely have made the results too limited or too wide for the research at hand. The categories on the spectrum, the accompanying search filters and example tweets are presented in appendix II.
For each of the five categories some examples of sentences conveying sentiments matching the respective categories were made to aid the researcher in a more efficient and replicable data collection. They were created based on experience gathered while becoming acquainted with the twitter search function and are displayed in appendix II. They alleviate the replication of the study should anyone which to test it and helped in reducing subjectivity.
Although the search terms were designed to capture specific categories they capture a wide range of categories since quotes and sarcasm are not uncommon. A complete list of the search filter used is presented in table 3.3 below. Orlando was replaced by the name of the city in which each event took place and the language was set to English.
Table 3. 3 Complete list of search filter
“Do something” OR “Take action” OR change OR Resign OR Remove OR revenge OR
“Bomb them” OR “Strike back” OR “Not welcome” OR “go back” OR Ban OR Restrict OR Protect OR Defend OR security OR immigrants OR “illegal aliens” OR “Gun control”
OR change OR Orlando OR banmuslims AND Terrorism -filter:replies -filter:links
3.3.3 Control variables
The control variables used in this study are all found in the GTD dataset and are in general thought to be important since they are the kind of information generally reported on after a terror attack and are likely to affect how serious the attack is perceived. For instance, an attack with an extremely high number of deaths, e.g. 9/11, is thought to provoke higher levels of radical responses than a relatively small attack with a single fatality.
This is based on how serious and real the threat is experienced. Larger threats, such as higher numbers of attackers, deaths and wounded are likely in direct relation to how serious a threat is perceived and would thus affect the reactions of audiences. The simple logic of numbers supports this reasoning, i.e. the more victims the more people are likely to know someone who knows someone who witnessed the attack first hand. In general, when stories are told by people we know and trust they tend to have a larger impact than news transmitted by strangers.
Other variables which may affect how serious a threat is perceived are whether it was a suicide attack or not and the kind of victims which are targeted. These variables are believed to increase the chances of audiences perceiving the outgroup as having maximal rather than limited goals. Maximal goals in this context would be the elimination of the
victim society’s way of life, not stopping until it is annihilated. Limited goals on the other hand is achieving a specific goal such as the withdrawal of troops in a foreign intervention by the victim government (Rose et al. 2007: 191-192).
Again, the actual goals of the attacker are not of as much importance as is the audience’s perception of those goals. In other words, an attacker might make a statement that his goals are limited which in turn can persuade and affect audience perceptions but in it is in the end the audience’s view on the matter that is important. To reiterate, suicide attacks and certain victim types such as hospitals, religious buildings (churches in this case) and schools are posited to cause an increase in maximal goal perception. This variable was coded by combining categories into a smaller hierarchical scale with the government as the highest ranking and private citizens as the lowest ranking. As such it can be expected to have a negative effect on radicalism.
Lastly, the author hoped to control for whether the attack was domestic or international since that is closely related to the ingroup outgroup divide. However, most cases did not have any data on this variable or where coded as domestic. Although this was not controlled for it would have been interesting to compare it to the group level variable.
3.4 Scope conditions
The most pressing limitation to the scope of this study is its restriction to examining only one country. Although it was necessary with the time and resources available it limits the generalisability of the results. The results are not expected to differ much from cases within the western world or from other Anglo-Saxon countries however including such cases could have solidified the generalisability of the results.
Another restriction is that twitter is taken to represent social media as a whole while there are significant differences in how different social media platforms are used (Greenwood et al. 2016). It would possibly be more accurate to say that the study is applicable to certain types of social media, especially the kind using terse messaging where there is little room to elaborate and express nuance in one’s demands for political change. Apart from the cultural differences a comparison between Twitter and Weibo, the Chinese equivalent, would be interesting to shed light on whether the terse format affects the radicalism of reactions (BBC 2016).
It would also be of interest to study reactions to terrorism over a greater time period and to compare reactions to terrorism during the different trends, anarchism, left-wing,
jihadism etc. Another interesting breaking point would be to study reactions to terrorism before and after 9/11 or after the initiation of the war on terror and see if there are differing trends. However, the scope of this paper is too narrow to allow a comparison across multiple media platforms and other means of expression and thus is limited to 2006-2016.
While discussing the social media aspect of the study it is important to mention that knowledge of how social media is used has a value in itself but also that not all people use social media. Although we live in a highly connected society we should be reminded that not all people have internet access at home, predominantly the elderly (Ibid). Moreover, the internet and social media is used differently and to varying degrees by certain age groups.
This further limit the study in that social media might be used in a particular way after terror events but this is unlikely a perfect reflection of society as a whole.
Lastly, the study is completely based on observations made by the GTD. It is a trusted source but since definitions of terrorism are far from agreed upon between all agencies, academics, NGOs, databases etc. some caution should be used when comparing other studies and pieces of work with this one.
3.5 Validity & Reliability
Regarding validity and reliability, the most alarming issue concerns validity. Firstly, using twitter without having access to the full API through twitter enterprise or third-party services limits the search function’s ability to restrict the search to tweets originating from the country under observation.
Another issue is that the current search method does not capture all tweets that could fit within the categories of radical response. Without more sophisticated search filters, software to process the data and time to analyse it there is little that can be done about this. It is the firm belief of the author that the current search filter and categories are valid in capturing a variety of responses on the spectrum of radical response, but it is acknowledged that there are ways in which the present study could be improved.
Moreover, the issue of assessing to what extent the results can be generalised to twitter behaviour in general is obstructed by the fact that it is difficult to ascertain the total number of tweets made in reaction to the terrorist events. Although it would strengthen the study to be able to show a percentage number of tweets analysed in comparison to all made on the topic the study still indicates patterns in twitter usage.
Going back to the issue of locating users on the world map and limiting searches to a specific country the current paper has no real way of analysing whether an account is really made and used by a US citizen, apart from the author’s judgement. This ties into the difficulty of assessing whether the twitter users consider themselves to be US citizen and what characteristics constitute that citizenship, i.e. the ingroup. It should be noted that it is not unlikely that some people would say that characteristics A, B and C are what constitutes the US nationality when in reality it is also determined by D, E and F as well. It is not necessarily so that respondents would intentionally lie, but some might answer what they think they should answer rather than what they actually think (Schildkraut 2014: 470-471).
The question of what constitutes the backbone of the US ingroup is dealt with in the theory section. There are several studies strengthening the claim that some traits are implicitly considered more relevant to the US identity.
The method used to assign ingroup or outgroup status to the attackers is not without flaws but is the best available within the timeframe and resources available. Although it likely generalises attackers into groups which some twitter users would not agree with it is based on theoretically sound logic (Schildkraut 2014: 455). To get a more comprehensive overview of what constitutes the US ingroup a set of in-depth interviews or surveys could have been done given more time.
Moreover, there is a slight issue of validity in terms of the definition of terrorism used by the GTD in that others will have different definitions or intuitive ideas about what constitutes a terrorist. Specifically, it is possible that the definition applied by the GTD is not always coherent with the ones employed by twitter users, whether they base their definitions on what is said in media or statements made by various agencies such as the FBI or different police departments. Although this is an issue which may affect the responses to some of the less clear-cut instances of terrorism it cannot be avoided without first performing a survey among twitter users as to what constitutes a terrorist attack and then creating a new dataset based on that information.
In terms of reliability there should be fewer issues. In general, if the same search filter is applied without being logged in on twitter and matched with the same data from the GTD dataset there should not be any problems reaching the same results. However, there are some issue with using twitter since privacy settings, search functions etc. can be changed over time.
For instance, the possibility of increasing the character limit to 280 has been probed (Ahmed 2017).
Another problem is that tweets and twitter accounts can be erased which makes reliability weaker since some tweets might simply not exist at a later point in time. Accounts can be deleted by users themselves or if the account is deemed inactive. Tweets can be reported and thus removed if they are considered in violation of the code of conduct (Twitter Media Policy 2017). All this obstructs reliability but the likelihood that users would remove tweets made years ago or that such tweets would be reported seems unlikely unless the accounts are removed all together.
4. Results and Analysis
In this section the results will be presented and discussed followed by a discussion on the drawbacks with the research design as well as possible alternative explanations. The multivariate regressions were all done in R-studio and the chi-square test in excel.
The results from running both bivariate and multivariate tests revealed that the correlation between the independent and the dependent variables is statistically insignificant.
Four types of tests were made, one bivariate chi-square test between the IV and the DV and three multivariate regressions testing the IV plus control variables with the DV, the highest radical level category five and the two highest categories four and five combined.
The chi-square test was made to check if the relationship between the IV and DV was statistically significant when the variables were treated as categorical as compared to their ordinal nature. It was not found to be significant.
The first multivariate regression computes the effect of the IV on the DV when controlling for how many people were killed or wounded, the number of perpetrators, whether it was a suicide attack or not and the type of target. Although the IV and DV are ordinal there is a clear hierarchical order which together with AIC and BIC tests lead to the conclusion that treating them as continuous would yield the best results. The discrepancy between the mean and the median was in both variables minimal. One case was dropped due to missing values in the independent variable and in models 3 and 4 in table 4.3 one case was dropped due to coding errors. Although this has an effect it is not believed to be the critical point the outcome.
In model one the bivariate relationship between group level and radical level can be seen with there being a slight effect although it is statistically not significant.
In model two the effect of group level is increased but remains statistically insignificant as do all the control variables introduced. The number of killed and number of
wounded variables have a positive and small impact while number of perpetrators has a medium negative effect. The effect of suicide is negative and large.
Model three adds the control variable target type which has a small positive effect.
The change in the previous variables is small and no variable manages to achieve statistical significance.
Table 4. 1 The effect of group levels on levels of radicalism
Variables Model 1 Model 2 Model 3
Intercept 1.91463
(5.25e-06 ***)
2.106665 (6.56e-05 ***)
1.932017 (0.00217 **) Group level
Number of killed
Number of wounded Number of
perpetrators Suicide
Target type
0.01488 (0.901)
0.033003 (0.787) 0.029155 (0.111) 0.001829 (0.726) -0.267824 (0.321) -0.995298 (0.133)
0.030796 (0.80263) 0.029103 (0.11542) 0.001986 (0.70691) -0.250732 (0.36051) -0.989817 (0.13859) 0.063037 (0.61475)
N 47 47 47
R-squared 0.0003433 0.1399 0.1454
Adjusted R- squared
-0.02139 0.03503 0.01722
Multiple linear regression
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’
P-value in parentheses.
In table 4.2 a similar analysis is made only replacing radical levels with a binary variable coded 1 for cases with at least one category five tweet, 0 if none.
In model 1 group level has a bigger positive impact compared to table 4.1 but remains statistically insignificant.
This changes as model 2 is introduced where its impact is a third smaller than in model 1. The number of wounded and killed both have a small positive impact similar to that in table 4.1 whereas the number of perpetrators has a big positive effect. The effect of the suicide variable is also positive and large.
There is little change between model 2 and model 3, target type has a medium sized negative effect. Throughout table 4.2 no variable is statistically significant.
Table 4. 2 The effect of group levels on radical category 5
Variables Model 1 Model 2 Model 3
Intercept -2.5251 (0.0238 *)
-3.34179 (0.0232 *)
-2.92311 (0.101) Group level
Number of killed
Number of wounded Number of
perpetrators Suicide
Target type
0.4635 (0.1617)
0.29961 (0.4140) 0.04198 (0.4641) 0.02889 (0.2094) 0.58899 (0.3747) 1.57432 (0.2911)
0.29365 (0.417) 0.04366 (0.464) 0.02882 (0.214) 0.55061 (0.411) 1.56770 (0.294) -0.14314 (0.700)
N 47 47 47
R-squared 0.03949318 0.1839444 0.1867533
Generalised logistic regression
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’
P-value in parentheses, R-squared calculated using 1- (residual deviance/null deviance).
In table 4.3 the effect of group levels on a combination of categories four and five is measured. Category 4 + 5 is binary, cases containing at least one tweet in both categories are coded as 1, others as 0.
In model 1 group level has a medium negative effect on the combined categories variable. This is not statistically significant.
The effect of group level remains negative and increases in model 2. Number of killed has a medium positive effect whereas the effect of number of wounded is minimal and negative. Suicide and number of perpetrators both have huge negative impacts.
In model 3 the negative impact of group level is nearly tripled, as is the positive impact of the number of killed. The number of wounded variable remains negative and grows to have a medium sized effect. Both suicide and number of perpetrators retain their huge negative impact in contrast to target type which has a huge positive effect. As in the other tables no variable in no model is statistically significant.
Table 4. 3 The effect of group levels on radical category 4+5
Variables Model 1 Model 2 Model 3
Intercept -2.1353 (0.231)
1.371e+01 (0.998)
8.2628 (0.999) Group level
Number of killed
Number of wounded Number of
perpetrators Suicide
Target type
-0.3585 (0.579)
-5.205e-01 (0.588) 1.681e-01 (0.360) -3.096e-02 (0.845) -1.625e+01 (0.997) -1.650e+01 (0.999)
-1.4440 (0.216) 0.4414 (0.280) -0.1804 (0.546) -18.4349 (0.997) -14.0257 (0.999) 2.7478 (0.375)
N 47 46 46
R-squared 0.01892153 0.4383695 0.6303142
Generalised logistic regression
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’
P-value in parentheses, R-squared calculated using 1- (residual deviance/null deviance).
4.1 Analysis
As is shown by the tables above the results do not express much variation between models. This suggests that the different variables are not what is causing the variation between tables, rather, that is due to the difference between the dependent variables. In table 4.3 the results differ heavily from the previous tables which is likely due to the small number of cases where both categories four and five were present. Most likely the combination of a low number of cases, relatively little data from each case, issues with data collection and the limitation to a regression model not suited to handle ordinal variables led to the results turning out statistically insignificant and seemingly haphazard at times. Due to the results showing no statistical significance the analysis will be focused on why that is, rather than the effects of the variables. Despite the group level variable being positive in two out of the three regressions the null hypothesis cannot be rejected due to statistical insignificance. Still, despite the results, the theoretical relationship is expected to hold true due to the design flaws and the possibility that twitter is not the token way to express radical political suggestions.
The reason as to why no statistically significant relationship could be attained is likely connected to the data collection and research design which will be discussed below followed by a discussion of possible alternative explanations.
Although a probit analysis treating the ordinal variables as ordinal rather than continuous would have been more suitable there is little to suggest that would solve the issue entirely. Such an analysis was tested to compare the levels of significance however only the suicide variable and one of the group levels were statistically significant. The reason a probit analysis was not used as the main form of analysis is due to a lack of knowledge regarding its interpretation and proper implementation.
A stronger reason why the results are statistically insignificant most likely is the sample size. In order to facilitate data collection and accommodate for time limits only cases with at least one fatal casualty were included. However, this severely limited the number of cases to 48 rather than 233. It is likely that a higher number of cases could have increased significance. In addition, one case could not be used due to there not being a perpetrator, meaning that no score could be given on the independent variable. In hindsight it might have been more fruitful to perform a qualitative study with a focus on some critical cases since such a limited number of cases could be included.
Altering the case selection to include cases with fewer than one deaths would require an alteration of the data collection method as well since one of the issues with the present
technique is its inability to effectively filter out unrelated tweets in the cases which were not widely discussed or where other events significantly increased levels of noise such as terror events in India or Pakistan. There is serious doubt that simply polishing the search filter would properly take care of the noise issue. Having access to twitter demographics and thus being able to only gather tweets from specific countries or regions would much have benefitted the data collection. In other words, in order to solve the issue of sample size one would first have to solve the filtering problem.
This problem is at the core of several perspectives of design shortcomings. Two more linguistically related problems are that the English language is spoken widely not only in countries where it is the official language and that there is an almost unlimited number of ways in which to express a desire for violence or change. The first issue is twofold in that events in the US is commented on by audiences in other countries and that events in other countries will be commented on in English. The current design has no real way to deal with foreign comments in English apart from them with the utmost probability being significantly fewer. Regarding the events in other countries the design attempts to deal with this using city names in the search filter which were used with high frequency during the test runs. These were, however, made on cases with higher numbers of deaths and presumably higher numbers of tweets and it was later discovered that some cases would have required more place specific filters to cancel out noise. At the time of discovery, it was too late to change the search filter which would have required all previous searches to be made again.
The final issue is of the we do not know what we do not know kind. The many nuances and expressions which can be used in a language even when not attempting to account for sarcasm are of higher numbers than can truly be captured by even the most sophisticated search filters. Yet, an upgrade from the current search filter would have been to apply several layers of searches such as first applying a filter with words related to a place and then applying a second filter with words related to terrorism on the results from the first search. This was not realised until after the data was collected but might have made for an improved albeit more time consuming data collection.
4.2 Alternative explanations
The following alternative explanations are best understood as complementary rather than as competing with one another.
The first alternative explanation is that people do not use twitter to air opinions on how to best respond to terrorism as much as they use it to discuss what terrorism is. This seemed to be a recurring theme during the data collection spanning the whole range of group levels. On one end claims would be made that white Christians are not linked to terrorism but that Muslims are, even when the crimes are comparable. On the other it would be claimed that people, democrats and the government seemed to be recurring themes here, do not recognise terrorism committed by Muslims. This in combination with simple statements like
“this is terrorism” or “this is not terrorism” emphasised that the definition of terrorism is not universal.
Another possibility is that people in general are highly aware of how they are perceived and that in order to show awareness of minority issues some people may be inclined to question the decision to define certain cases as terrorism. In addition, they may be wary of creating outgroups based on individuals from minority groups. This goes hand in hand with the Chip or Tyrone test which indicates that there are certain issues which make political groups increasingly likely to depart from certain principles (Ditto et al. 2009: 328).
Outgroup identification based on superficial characteristics may be one such issue.
In connection to the above, it is possible that the most radical suggestions are made on other platforms or in other contexts such as the local bar or more niche forums online. In other words, only a small number of those who make radical suggestions use twitter to express them. This is in line with the US twitter population leaning slightly towards the younger and higher educated demographics Greenwood et al. 2016).
5. Conclusion
In this paper the relationship between the group identity of terrorists and audience responses expressed on twitter was examined. The aim was to identify whether terrorists closer to an outgroup prototype would elicit more radical responses than those part of the ingroup. Additionally, the use of social media to express these responses was of interest although not the main focus.
The theory could not be confirmed by the study since no statistically significant effect was found. This is more likely a result of issues with sample size and data collection rather than the theory being disproven. Although the theory is not proven there are reasons to believe most alternative explanations do not run counter to the theory, the results instead being explained by the ways in which social media is used.
The findings suggest that examining the theory by using was without the time and resource limits of the study. Instead of attempting to force a quantitative study in a situation with few cases and difficult data collection it may have been wiser to opt to follow a select number of individuals more closely in a qualitative study. Alternatively, not examining social media without proper filtering software, using surveys instead could have yielded better results.
Overall this study does not carry many implications for policies since there were no statistically significant results, suggesting that there is not much to worry about, but the theory strongly suggests that there is. As such, the implications are greater for future research. One area of interest would be to examine how knowledge of the perpetrator might shift reactions, i.e. if the different aspects of group belonging have stronger or weaker effects on different types of people. Another area that might be of interest is which words are trending closely after terror events. For instance, if words symbolising solidarity increase more than those representing hate or sadness etc.
6. Bibliography
Abrahms, M. (2006). Why Terrorism Does Not Work. International Security, 31(2), 42–
78.
Ahmed, K. (2017, September 26). Twitter trials 280-character tweet limit. BBC News.
Retrieved from http://www.bbc.com/news/business-41408798
Aistrope, T. (2016). Social media and counterterrorism strategy. Australian Journal of
International Affairs, 70(2), 121–138.
https://doi.org/10.1080/10357718.2015.1113230
Ayres, N., & Maglaras, L. A. (2016). Cyberterrorism targeting the general public through social media: SCN-SI-088. Security and Communication Networks, 9(15), 2864–2875. https://doi.org/10.1002/sec.1568
BBC, (2016). Sina Weibo ends 140-character limit ahead of Twitter - BBC News. (n.d.).
Retrieved 27 November 2017, from http://www.bbc.com/news/technology- 35361157
Bonanno, G. A., & Jost, J. T. (2006). Conservative Shift Among High-Exposure
Survivors of the September 11th Terrorist Attacks. Basic and Applied Social Psychology, 28(4), 311–323. https://doi.org/10.1207/s15324834basp2804_4
Chaliand, G., Blin, A. (2013). The history of terrorism from antiquity to al Qaeda.
University of California Press, Berkeley and Los Angeles, California.
Cheong, M., & Lee, V. C. S. (2011). A microblogging-based approach to terrorism informatics: Exploration and chronicling civilian sentiment and response to terrorism events via Twitter. Information Systems Frontiers, 13(1), 45–59.
https://doi.org/10.1007/s10796-010-9273-x
Devos, T., & Mohamed, H. (2014). Shades of American Identity: Implicit Relations between Ethnic and National Identities: Shades of American Identity. Social and
Personality Psychology Compass, 8(12), 739–754.
https://doi.org/10.1111/spc3.12149
Ditto, P., H., Pizzarro D., A., Tannenbaum D., (2009) Motivated Moral Reasoning. In Bartels, D., M., Bauman, C., W., Medin, D., L., Skitka L., J., (Eds.). (2009) Moral Judgement and Decision Making. Elsevier Inc., San Diego.
Twitter FAQ (2017). FAQs about top search results. (n.d.). Retrieved 30 November 2017, from https://help.twitter.com/articles/131209?lang=en
Goldman, L., & Hogg, M. A. (2016). Going to extremes for one’s group: the role of prototypicality and group acceptance: Extreme group behavior. Journal of Applied Social Psychology, 46(9), 544–553. https://doi.org/10.1111/jasp.12382
Gould, E. D., & Klor, E. F. (2010). Does terrorism work? The Quarterly Journal of Economics, 125(4), 1459–1510.
Greenwood, S., Perrin, A., Duggan, M., (2016). Social Media Update 2016. Pew Research Center.
