Over the last decade, the erosion of trust in public institutions and traditional media sources have been proceeding in parallel. The electoral success of populist or anti-establishment parties has been accompanied by radical changes in the consumption of information. Indeed, many have found connections between consumption of misinformation and support for ‘anti-establishment’ rhetoric. In the US, as discussed in Allcott and Gentzkow (2017) and Guess et al. (2018), Trump voters were more likely to be exposed and believe to misinformation. In the Italian context, these findings have been replicated in a recent report from the financial newspaper Il Sole 24 Ore,1 where the likelihood of believing and sharing so-called fake news was found to be higher for voters of the MoVimento 5 Stelle and Lega than for voters of other parties. In Italy, not only does the consumption of fake news appear to be linked with populism, but the content of the overwhelming majority of pieces of misinformation also displays an obvious anti-establishment bias, as found in Giglietto et al. (2018).
Measuring the causal effect of exposition to fake news is however far from trivial, as simple correlations do not provide much information on their effect on support for anti-establishment parties. It is, in fact, unclear whether exposure to fake news actually affects voting behaviour or, instead, if prior political beliefs affect entry into these ‘misinformation bubbles’.
To disentangle these complicated reverse causality issues, we would need the exposure to misinformation to be assigned randomly. In the perfect random trial, individuals voting for the same election would be assigned to two different – and isolated – subgroups through a random trial: one group would have access to legitimate sources of information only, while the other one would also be exposed to fake news.
In this regard, the Italian context presents us with an ideal quasi-natural experiment, allowing us to study this relationship without causal confounders. In our recent paper (Cantarella et al. 2019), we look at the Trentino Alto-Adige/South Tyrol region in Italy. This region located on the border with Austria is home to a sizeable German-speaking linguistic minority and features some degree of language segregation: a parallel German-language media market exists, and effective bilingualism is not particularly widespread (Ebner 2016).
Our intuition is to exploit the language differences across the two provinces as an exogenous source of variation in exposure to misinformation. We, in fact, believe that the German-speaking community in the region, while comparable to its Italian-speaking counterpart in terms of economic and demographic conditions, is exposed to a peculiar filter bubble where exposure to fake news concerning Italian politics is limited. This assumption is justified by a number of considerations on the motives driving fake news ‘disseminators’: it is indeed reasonable that spreading misinformation in German concerning Italian politics - also in light of the small size of this community - is a less than optimal solution for a disseminator who wants either to influence the outcome of the elections or to generate revenue from advertisements.
Figure 1 Language groups shares and growth of populist parties (2013-2018) in Trentino-Alto Adige/ South Tyrol
Notes: The size of the circle around each municipality indicates average electorate size between the two periods. Populist scores are computed using the text mining method described in the next paragraph.
A text analysis approach to measuring populist rhetoric
In order to understand whether the exposure to fake news influences the electoral support for populist parties, it is necessary to identify which parties can be defined as ‘populists’ in an objective manner.
We started from the assumption that two common features of populism, understood in its various forms, are the recurrence of ‘anti-elite’ themes and the use of an emotional communication style. We collected all posts published on Facebook over the course of the 2013 and 2018 campaigns by the parties and their leaders that ran in Trentino-Alto Adige/South Tyrol for the last two Italian general elections.
Subsequently, aided by text analysis, we measured the relative frequencies within each post of 23 keywords associated with anti-system tropes (eg "caste" or "shame"), and then the number of exclamation marks. The scores obtained, and their variation over time, are shown in Figure 2. The data not only shows how the social communication of Movimento 5 Stelle and Lega has a marked populist character, but also how the intensity of the same has changed considerably over time.
Figure 2 Text analysis scores of social media posts from parties and their leaders during the 2013 and 2018 elections campaigns
Notes: The left figure refers to the scores obtained with the populist text-bag, while the right figure computes the same score as the frequency of exclamation marks in the text. Parties in grey have only took part in one of the two elections. The red dashed line refers to the election specific average.
Linguistic filter bubbles and natural experiments
A further complication is presented by differences in voting behaviour across linguistic groups. As the South Tyrolean People’s Party - a Catholic-based catch-all outfit, not particularly populist under our criteria - has consistently proven to be the most popular voting choice across the German-speaking population, our estimates would suffer from extreme upward bias when not controlling for previous elections. To account for this issue, we take advantage of a difference-in-difference framework, controlling for the changes in voting between the 2013 and 2018 elections across the two groups.
We then collected municipality-level data for 2013 and 2018 on electoral outcomes, linguistic groupings, internet connectivity, income and demographics from the autonomous provinces of Trentino and South Tyrol, and aggregated populist scores by municipality to control for omitted-variable bias. Finally, using publicly available information from the Facebook Audience Insight Tools, we were also able to construct an indicator for the number of likes each Facebook page which disseminates fake news (identified through ‘black lists’ compiled by the debunking blogs Butac.it and Bufale.net) in a given municipality.
Ordinary least squares (OLS) estimates (see Table 3 in our paper) indicate a positive and statistically significant effect of the Italian language group on voting in the 2018 election. A prior OLS model (Table 2) also found a similar correlation between populist preference and exposure to misinformation, confirming the aforementioned findings from the literature. First stage regressions also confirm that the intensity in exposition to fake news was higher in Italian-speaking municipalities. In our instrumental variables (IV) estimates, however, the coefficient for fake news exposition in 2018 is not statistically different from zero, suggesting that fake news had no role in contributing to the rise of populist platforms.
These results may seem counterintuitive, given that linguistic filter bubbles clearly had an effect in influencing voting. To understand these figures, we have to look at the local average treatment effect of exposition to misinformation. In an IV setting, we are effectively comparing Italian-speaking voters, who were exposed to misinformation, against German-speaking voters, who were unexposed. This means, as hinted in Figure 1, that rural Italian-speaking areas where populist preferences grew the most were also the ones which were relatively less exposed to misinformation. The larger the exposition relative to the Italian linguistic group, the more the real effect of misinformation is revealed. As a further check, these estimates are also robust to an alternative instrument choice predicting exposition to misinformation through the variation in broadband internet connections between 2013 and 2018.
The true role of fake news
Our results take any credit for the 2018 success of populist parties away from the spread of fake news and support the hypothesis that voters self-select into misinformation bubbles, i.e. they consume fake news because of their prior preference for populist platforms, and not the other way around. This implies that exposure to fake news is entirely dictated by self-selection in misinformation bubbles, meaning that the causal channel between voting and fake news goes in a single direction, with individuals being exposed to misinformation because of their prior political presences.
This does not mean that the spread of misinformation is less dangerous or problematic, but only that the causes of the populist shift in voting have to be found elsewhere. The attention of researchers and policymakers should focus instead on the socioeconomic determinants for access into misinformation bubbles. The persistence of differences in voting behaviour conditional on linguistic grouping and broadband penetration indicates that social media ‘echo chambers’ (a concept further developed in Sunstein 2018) most likely had a role in determining final preferences. In this sense, fake news would rather stand as the embodiment of shared narrations within groups of voters, which are further reinforced by confirmation bias and the increasing personalisation of social media echo chambers.
While our results are internally valid, the effect of fake news might differ elsewhere. Yet, we still believe a piece of fake news perceived as true will have a lower marginal effect than a true piece of information in support of prior preferences. In the presence of personalised filter bubbles, preference dictates facts and not the other way around; social media plays a role in as much as it provides individuals with the information they want to believe.
Allcott, H, and M Gentzkow (2017), “Social media and fake news in the 2016 election”, Journal of Economic Perspectives 31(2): 211–236.
Cantarella, M, F, Nicolò and R G Volpe (2019), “Does fake news affect voting behaviour?”, DEMB Working Paper Series 146.
Ebner, C V (2016), “The long way to bilingualism: The peculiar case of multilingual South Tyrol”, International Journal for 21st Century Education 3(2): 25.
Giglietto, F, L Iannelli, L Rossi, A Valeriani, N Righetti, F Carabini, G Marino, S Usai and E Zurovac (2018), “Mapping Italian news media political coverage in the lead-up of 2018 general election”, SSRN.
Guess, A, B Nyhan and J Reifler (2018), “Selective exposure to misinformation: Evidence from the consumption of fake news during the 2016 US presidential campaign”, European Research Council 9.
Sunstein, C R (2018), #Republic: Divided democracy in the age of social media, Princeton: Princeton University Press.
 Il Sole 24 Ore (2018), “Fake news: quando le bugie hanno le gambe lunghe”, 4 May.