ABSTRACT
The rise of social network sites reopened the debate on the ability of traditional media to influence public opinion and act as an agenda setter. To answer this question, the present paper investigates first-level and second-level agenda-setting effects in the online environment by focusing on two heated Italian political debates (the reform of public funding of parties and the debate over austerity). By employing innovative and efficient statistical methods such as the lead–lag analysis and supervised sentiment analysis, we compare the attention devoted to each issue and the content spread by online news media and Twitter users. Our results show that online media keep their first-level agenda-setting power even though we find a marked difference between the slant of online news and the Twitter sentiment.
Notes
1. We observed a similar growth also in the number of monthly active users. See http://www.datamediahub.it/2012/12/17/lutilizzo-di-twitter-in-italia-nel-2012/#axzz3PYznSQGt and http://wearesocial.it/blog/2013/12/latest-3-years-twitter-34-milioni-di-utenti-attivi-italia/
2. Many media outlets are active on Twitter even though previous studies confirm the nonelite-driven nature of social media, showing that only a limited number of Twitter accounts belongs to professional journalists or news corporations. In Italy, in fact, they represent only 5% of users (Antenore, Di Gianmaria, Faggiano, & Gennaro, 2013; Bentivegna & Marchetti, Citation2014). This rate is similar to that of the United States, where 95% of Twitter users are media consumers (Vargo, Citation2011; Vargo et al., Citation2015). As such, the information published on Twitter by the official accounts of media outlets will be spread across the Web only if media consumers believe that such information deserves further diffusion (for a similar view, see Vargo et al., Citation2015). Nevertheless, tweets posted by news media have been excluded from the analysis.
3. This is even more true in the case of second-level agenda setting given that off-topic texts are automatically excluded by the SASA algorithm.
4. These scandals burst between March and June and involve members of all the main parties such as the center-left Democratic Party, the center-right People of Freedom Party, the right-wing Northern League and the left-wing Left Ecology and Freedom Party.
5. Note that this pattern is in line with the Downs (Citation1972) “attention cycle,” which suggests that the interplay between media and public opinion promotes a policy change insofar as it pushes politicians to embrace salient issues as soon as possible.
6. For instance, the informal expression “what a nice rip-off!” will be considered ambiguous when analyzed through ontological dictionaries because it includes both a positive and a negative term.
7. According to Hopkins and King (Citation2010) this method is particularly accurate when compared with hand-coding, leading to a root mean square error around 2%–3%.
8. In other terms, a “word profile” is a vector made of 0’s and 1’s: we find a 0 when a term does not appear in the unit (but it is used in some other units) and a 1 when a term appears in the unit.
9. The fact that the training set, in this case, is manually codified or based entirely on an ontological dictionary does not make any difference: as long as the classification of texts is done on an individual base, we will produce aggregate estimates severely biased.
10. The analysis has been done by two trained coders. Intercoder reliability is 0.88. Compared to hand-coded documents in the training set, the root mean square error of the estimates is 1.5%. This confirms the accuracy of the results.
11. The share of negative comments on social media is not particularly different from that of the whole Italian population. In fact, survey polls held in the same months revealed that around 70% of Italian citizens did not trust political institutions and 80% of them supported the abolishment of public funding of parties, rather than a mere reform (http://www.archivio.sondaggipoliticoelettorali.it/asp/visualizza_sondaggio.asp?idsondaggio=5472).
12. The analysis has been done by two trained coders. Intercoder reliability is 0.82. Compared to hand-coded documents in the training set, the root mean square error of the estimates is 3.1%.
Additional information
Notes on contributors
Andrea Ceron
Andrea Ceron is an assistant professor in Political Science at the Università degli Studi di Milano. His research focuses on intra-party politics and media analysis. His recent publications include articles in the British Journal of Political Science, New Media & Society, and European Journal of Political Research.
Luigi Curini
Luigi Curini is an associate professor of Political Science at the Università degli studi di Milano. His research focuses on party competition, spatial theory of voting, and social media analysis. His recent publications include articles in the British Journal of Political Science, Comparative Political Studies, and Journal of Politics.
Stefano M. Iacus
Stefano M. Iacus is an associate professor of Mathematical Statistics and Probability at the Università degli Studi di Milano. His research focuses on computational statistics, theoretical statistics, sentiment analysis, and inference for stochastic processes. His recent publications include articles in Political Analysis, Computational Statistics, and Information Sciences.