By Erin C. McGrath and Cody Buntain
Reported terrorist attacks increased dramatically over the last decade or so, from just under 3,000 per year in 2001 to nearly 17,000 in 2014, amplifying the number of people subjected to the effects of terrorist attacks nearly six-fold over the last decade.
At the same time, the introduction, popularity, and low cost of social media have caused many researchers to rethink the data they use to analyze terrorist attacks, and methods chosen to analyze that data, giving prominence to the emerging field of computational social science.
Hence, Twitter and other social platforms have become important communication channels during crises caused by terrorist attacks. While research into crisis informatics and social media is growing, the variation in Internet penetration and social media usage in countries where terrorist attacks take place complicates analysis of these specific events.
As a result, the findings on how social media users respond to these events on popular platforms like Twitter is not generalizable, but does give researchers a different, complementary, insight along with other findings, into the impact terrorist attacks may have. In our research, we address how social media can give insight into public response to terrorist attacks.
Preliminary Findings: Terrorist Attacks Worldwide
As emotionally-intensifying events, social media analytics can uncover how different emotions – like anxiety, fear, and anger – may impact decision-making and actions in the short-term aftermath of an attack (Jin, Fraustino, & Liu, forthcoming). Our research has recently explored social media response to recent terror events in Nigeria, Tunisia, Belgium, Australia, France, and the United States.
Specifically, we have analyzed tweets in response to the 2014 Kano Bombings in Nigeria by Boko Haram, the 2015 Mt. Chaambi Attacks in Tunisia by Uqba Ibn Nafi Battalion, the 2014 Sydney Hostage Crisis in Australia by an individual Man Haron Monis, the 2015 Paris November Attacks and the 2016 Brussels Bombings by the Islamic State of Iraq and the Levant, the 2015 Charlie Hebdo Shooting by Al Qaeda, and the 2013 Boston Marathon Bombing by the Tsarnaev Brothers in the United States.
Results of this research show that these events do not significantly impact general Twitter usage; that is, the events do not drive more users or additional content on Twitter. On the other hand, events like elections do seem to drive more users to social media for a longer period.
More importantly, however, the distribution of conversation on Twitter does react to terrorist attacks. We find this in metrics showing significant and rapid increases in mentions of the events and retweets sharing news of the event. Hashtag-sharing relevant to terrorist attacks, such as #prayforBoston, or #jesuisCharlie, also often rapidly increases, while sharing of website URLs grows more slowly.
Furthermore, Twitter exhibits a short memory with respect to these events. The majority of behavior relevant to these events returns to pre-event levels after only four days, whereas the impact of these events on the population remains for much longer (Buntain, Golbeck, Liu, & LaFree, 2016). A longitudinal survey of those affected by the Boston Marathon Bombing showed that attitudes were impacted for as long as a year and a half afterward, in survey participants’ willingness to report suspicious activity to the police (LaFree & Adamczyk, 2016).
Finally, local news affiliates and crisis response agencies (if present on social media) emerge as preferred sources, or central accounts, in the social networks surround the discussion of these events. Social media has become an integral part of crisis response to terrorist attacks, as much as it has been associated with the ability to organize terrorist activity out of the public eye and across borders.
Social Media: Promises and Pitfalls
While the volume of data captured from social media makes “digital breadcrumbs” appealing as a data source, “big data” also brings methodological challenges for research. Foremost among these challenges are issues of bias due to non-representative sampling, self-selection, and the rudimentary assumptions made in order to computationally analyze natural language.
For example, in text analysis, a “bag of words” approach assumes word order does not matter. Sentiment analysis approaches often miss sarcasm, humor, or other semantic devices that alter meanings and may affect what researchers are trying to discover.
On the other hand, social media allows researchers to examine users’ actual behavior, albeit online behaviors, surrounding high-profile events, in contrast to self-reported behavior from surveys or focus groups. Social media can give us access to places that are too far to get to in a short time, or are otherwise inaccessible due to conflict activity or even bureaucratic constraints.
Social media data can give researchers better insight into aggregate behavior more rapidly, allowing us to detect a higher resolution information landscape of the impact of an event which we know holds evidence of the social and political processes we seek to understand and explain. As with any social science research, our biggest challenge remains in overcoming challenges to inference.
Computational Social Science: Challenges and Opportunities
Despite diverse ontological origins, multi-disciplinary research on the complexities of public response to terrorist attacks demands systematic integration. Uniting this research requires marrying methods of inquiry through mixed-methods research design, and employing both conventional and newly available data.
Analytics of unstructured text gathered from social media can provide much insight when combined with other data sources, found a research funded under NSF INSPIRE project SBE-1343123. For example, together, quantitative analyses of survey responses and social media can detect topical and temporal patterns, which can be qualitatively examined to generate theoretical insights about public response to terrorist attacks from which to further this research agenda.
Triangulation of social media with other data sources can provide benefits to overcome challenges to inference. Triangulating randomly-sampled surveys, and convenience samples of social media users tackles issues omnipresent in the analysis of social media data, including representativeness and demographics, and research ethics concerns of consent and confidentiality. Moreover, integrating them helps to overcome challenges to inference due to the measurement of online behavior prone to self-selection biases and offline analysis of behavioral intentions rather than observations of actual behavior.
The way big data is structured, or more specifically its lack of structure, challenges the boundaries of the existing dominant scientific paradigms in the social sciences. Some popular technology writers pronounced the “end of theory”(Anderson, 2008).
The boundaries of post-positivism are drawn around testing alternatives to eliminate rival hypotheses for causal inference. This social scientific gold standard brings prestige with significant p-values, using the scientific method to accumulate knowledge. But how can big data complement it? Is big data a new kind of data, a new game in town, neither, or both?
In line with many methodologists across the social sciences, we believe that big data, and the methods being developed to analyze it, are not an alternative but rather a complement that alters the topology of the epistemological playing field in social science (Lazer et al., 2009).
In addition to opportunities, big data also presents great methodological challenges, most notably in terms of biases affecting user populations (e.g. different platforms used by different demographics), user behaviors (e.g. platforms drive users with algorithms based on homophily and propinquity), and the corresponding methodological challenges (e.g. proxy populations, multiple hypotheses testing) (Ruths & Pfeffer, 2014).
New sources of data on social interactions and public opinions, combined with increased computational power and methodological advances, open social science to evidence-based claims of correlation. These patterns, however, are drawn on a previously unimaginable scale.
“Big data” may be used for creating causal narratives from evidence-based claims, rather than eliminating all but one objective causal claim. While we do not lose sight of the goal of explanation in the social sciences, we argue that big data analytics do not lead only down one path, but rather several.
Put simply, very large-n studies do not implicitly suggest a quantitative or post-positivist research orientation. Big data can be modeled computationally to tell, in addition to “what,” also how and why.
Big data gives a higher resolution information landscape, in triangulation with surveys and other indicators, of a particular research problem. We also argue that these new sources of data and their analytics may lead us toward better understanding of behavior in the social sciences.
Erin C. McGrath works at the University of Maryland-College Park/START as NSF Post-doctoral Research Associate in Computational Social Science. Cody Buntain is a PhD student in the Computer Science Department at the University of Maryland. The article was first published on National Consortium for the Study of Terrorism and Responses to Terrorism website on 5 July, 2016. Republished here with permission.