What Twitter Really Tells Us

A STUDY ANALYZING HOW LANGUAGE CHANGES AND POLARIZES AFTER REPEATED TERRORIST ATTACKS SHOWS HOW PERCEIVED CULTURAL THREATS, MORE THAN ECONOMIC CONDITIONS, AFFECT THE SUCCESS OF FAR RIGHT PARTIES

by Francesco Giavazzi and Gaia Rubera, Full Professor of Economics; Full Professor of Social Media Marketing

The rise of protectionist and culturally conservative politicians is a trend that increasingly characterizes mature, Western democracies. The new political actors generally oppose globalization, understood as the free circulation of goods and people.
At the same time in Europe, while migration flows intensified, in the second half of 2010s, an unprecedented sequence of religiously motivated terrorist attacks occurred, leading far-right parties to frame their anti-immigration stances as policies designed to provide security against the threat posed by foreigners and to offer their electorate a tight mix of economics and culture. The increasing favor that these parties have met has stimulated much debate: which is more important in explaining the anti-globalization backlash of Western voters - culture or economics?

In “Terrorist Attacks, Cultural Incidents and the Vote for Radical Parties: Analyzing Text from Twitter”, a paper we wrote with Felix Iglhaut and Giacomo Lemoli, we study the role of perceived threats from other cultures induced by terrorist attacks on public discourse in Germany and on voters’ support for radical right parties. In this respect Germany is a particularly relevant case because of the large inflow of refugees form the Middle-east around 2015 and the parallel rise in votes for AfD (Alternative for Deutschland) an extremist right party. Our empirical strategy consists in analyzing data from Twitter. From a methodological point of view an appealing feature of Twitter is that, differently from the surveys used in this literature, it allows identifying variation in attitudes in real time as specific events happen.

In order to control for the economic conditions of different geographical areas – and hence isolate the effects of culture - we first geo-localize users living in different electoral constituencies in Germany. This posits a serious challenge as many users do not indicate their location on Twitter. Further, the few who do may provide either false or inaccurate information (e.g., Earth). Hence, we first identify so-called “landmarks” in each constituency, namely public or commercial Twitter accounts which can be clearly located in a given town and are likely to be followed by residents (e.g., police stations, town halls). We identify 5231 “landmarks” in 493 German towns, download the followers of their Twitter account, and attribute a user to an electoral district if she or he follows the accounts of at least 3 landmarks in a town and none elsewhere. This rule produces a sample of 178,271 users and 10,427,320 (geo-localized) tweets.

Next, we use a neural network model to compute daily similarity between the text of citizens’ tweets and that of German parties’ tweets from September 4, 2015 (when the AfD embraced an alt-right agenda) to the day of German elections: September 24, 2017.
We then measure, constituency by constituency, the shifts in similarity to party language immediately after 11 terrorist events. These events are exogenous to the conditions in the particular constituency and constitute shocks to public opinion evoking threats from other cultures. We find that, following these events, Twitter text becomes on average more similar to AfD, while the opposite happens for other parties. An example is shown in the Figure below, which describes the shift in similarity after each event between the language of the Twitter users in our sample and that of two parties: AfD and the SPD, the social-democrats.
 
We also find that, over time, the general public tends to discuss more, and with more negative tones, two core AfD topics: Islam and immigration, suggesting that terrorist events make these topics more salient. We also find that this discussion is organic and not driven by the media or by politicians. Finally, we find a significant association between these shifts in language similarity and changes in vote shares between federal elections.
From a methodological point of view our results show that analyses based on social media data can significantly improve upon studies based on surveys: these cannot—except by chance—detect changes in attitudes in the immediate aftermath of an event, such a terrorist attack. We also find that the role of perceived threats from other cultures on the success of nationalist parties overshadows that of current economic conditions, which are sometimes indicated as the main driver of such a vote: they are not. But the change of people’s attitudes following an event entailing a cultural threat is rather slow: in our sample you need a significant number of events—at least four or five—to observe a change in people’s discourse. However, after the tipping point is reached, each attack provokes a bigger and bigger shift toward the AfD. These findings resonate well with the idea that citizens do not change attitude overnight, but they respond to a sequence of nudges. Finally, our data also indicate that perceived cultural threats induce polarization: people’s discourses not only change, albeit slowly, but become increasingly different among different groups: left- and right-leaning voters in particular. This observation is at the core of the difficulties that our societies face.