Graduate School of Communication
Master’s programme Communication Science
Who influences whom?
The agenda-setting power of Twitter and newspapers during the
European election campaign 2014
Master’s Thesis
submitted by: Theresa Küntzler Student Number: 10841849 Lindenstraße 16b 67714 Waldfischbach-Burgalben Supervisor GermanyDr. Damian Trilling theresa.kuntzler@student.uva.nl
This work was carried out on the Dutch national e-infrastructure with the support of SURF Foundation.
I wish to express my sincere gratitude to Dr. Damian Trilling for his always encouraging supervision and valuable advices.
Furthermore, I would like to thank Dr. Moritz Küntzler, Annika Dukek and Konstantin Käppner for their helpful suggestions and proof-reading.
Finally, with all my heart I would like to thank my family, supporting me at any time, helping through difficult moments and enjoying the good times.
Abstract
In this thesis the potential of German newspapers and the German Twitter sphere to influence each other’s agendas surrounding the European Elections 2014 is investi-gated. Thereby, both the role of the general Twitter stream and effects of tweets from candidates for the European Parliament in particular are researched. With the use of Latent-Dirichlet-Allocation topics within the newspapers are identified and their relative attention given on Twitter over five months is determined. The given attention is then compared using a Vector-Autoregressive time series model. Despite indication of influ-ences between Twitter and newspapers both by theory and case studies, no influence between the newspapers’ agenda and the general discussion on Twitters is found. The candidates’ tweets are found to have an influence on the newspapers’ agenda and, to a lesser extend, vice versa. The findings point towards country or political system differ-ences in the effects of Twitter on the public agenda and thus call for a general caution when generalizing study results.
Introduction
In order to be informed about what happens outside their direct environment, people use media, such as television or newspapers. To get an idea of what is most relevant, people use clues such as how prominently a topic is printed in a newspaper or how much time is devoted to it in TV news. The decision upon where and how a topic is presented is made by journalists and editors. As most people consume similar news, the perception of what is important is shared by many. Via this process, mass media can set the public agenda (McCombs, 2014). With the rise of social media, media with an agenda-building process different from that of traditional mass media has evolved.
Twitter is such a social medium. The idea behind Twitter is that people can follow others, in order to receive their tweets. Tweets are messages of maximum of 140
characters. Following does not need to be reciprocal. A person can follow many and have only a few followers or it can have many followers, but follow only a few. Tweets are for the most part publicly available. Following an account leads to an automatic reception of the tweets on one’s personal Twitter site. Whether and how much attention is devoted to a topic on a user’s site depends on what content is tweeted by the people a person is interested in and thus follows. Higher salience of a topic is created as more people devote their tweets to the topic. The more people tweet about a certain topic, the more the topic will come up at different people’s Twitter sites and its importance grows. Thus, Twitter’s agenda potentially differs from traditional media’s agenda.
However, through different processes, for example Twitter being used as a source by journalists (Broersma & Graham, 2012), the two media might influence each other. In a scenario where Twitter follows completely the traditional media’s agenda, Twitter would build just another channel, without changing any processes in agenda-setting (Sayre, Bode, Shah, Wilcox, & Shah, 2010). However, if Twitter’s agenda differs from the traditional media and if it is able to influence the traditional media’s agenda, it would shift partial agenda-setting power away from the news industry towards a more general public. This again could enhance reversed agenda-setting, where the agenda is set by public concern (McCombs, 2014).
Politicians have a special role in both media. In traditional media, they are subject of coverage as well as source of information. With a Twitter account, they open a space of direct communication with (or to) the public, avoiding the interference of journalists. Statements put on Twitter, instead of given to journalists, are directly available to the public and no longer need to pass the ‘gates’ of a newsroom. This potentially changes the relationship between a journalist and a politician as a source (Aelst, Sehata, & Dalen, 2010). Furthermore, the direct communication to the public audience might provoke traditional media to cover raised topics (Parmelee, 2013), again a form of influence on the agenda. In addition, Vergeer, Hermans, and Sams (2013) suggest the potential of a new campaign style evolving from online social networks.
In light of this, the thesis at hand deals with the following overarching research question: ‘What is the influence of Twitter and newspapers on each other’s agendas?’
To investigate this question, German newspaper articles and tweets were collected in a time frame surrounding the European elections in 2014. Using a Latent Dirichlet Allocation (LDA), relevant topics in the newspaper coverage are defined. The relevance of these topics in the newspapers and within both the general public’s tweets and the tweets of German candidates for the European Parliament are compared. A
Vector-Autoregression (VAR) is conducted in order to find potential influence between the two media.
The approach put forward in this paper adds to the existing literature, as it concentrates on the agenda-setting power of Twitter, in contrast to Neuman,
Guggenheim, Jang, and Bae (2014), who jointly estimated the role of several online media. Moreover, the thesis at hand is not focused on pre-defined topics, as seen in other studies (Ceron, Curini, & Iacus, 2014; Neuman et al., 2014), but looks at topics extracted from the newspaper articles. Furthermore, the candidates’ role in the
agenda-setting process has so far, to the best knowledge of the author, not been assessed on large scale over time. Lastly, most quantitative evidence is found in the context of the US, whereas this study was conducted in a German/ European context.
The remainder of this paper proceeds as follows: In the next section the theoretical background is set up. In the following Methodology section the data collection process is described and the analysis plan explained. Thereafter, the results are presented. The paper ends with a discussion of the results and concluding remarks.
Theoretical background
Topic similarities between Twitter and newspapers
Topic overlap between newspapers and Twitter has not been heavily researched so far. However, some studies indicate a certain topic overlap. Zhao et al. (2011) compare a sample of the general Twitter stream with the content of the New York Times and find some topics to be covered in both media. Neuman et al. (2014) present similar findings specifically for political topics, also in the US. For the Austrian Twitter sphere, again, a certain overlap of political topics in the two media is reported (Ausserhofer & Maireder, 2013).
Jungherr (2014) looks into the coverage of politicians during the 2009 European elections in Germany. He finds that events are to an extend covered both on Twitter and in traditional media. However, as he and other studies find the agendas of traditional media and Twitter to differ, he suggests to consider a new media logic to explain what and how events are covered on Twitter.
Certain topics are found to be more relevant on Twitter than others. Relevant criteria concerning political topics are namely: social issues and topics about public order (Neuman et al., 2014), media- and technology-related topics as well as topics with a short news life cycle (Ausserhofer & Maireder, 2013) and entity-oriented issues, which means concerning persons or institutions (Zhao et al., 2011).
Based on these findings the following hypothesis is set up in this paper:
Influence from newspapers to Twitter
Before looking at different ways of influence from newspapers to Twitter, it should first be clarified what is actually meant by ‘influence’ in this paper: Taken into
consideration will be whether a certain topic first appears in one medium and later in the other as well as the extend to which it is covered. Thus, relevant aspects are the chronology and the relative importance of a topic. This excludes influence on how a topic is covered, such as opinions, tone or frames. This can be reformulated into whether the content of one medium can be better predicted with the help of the other medium’s content than without it.
A first reason to expect newspapers to influence Twitter content is derived from classic agenda-setting theory. Agenda-setting suggests that mass media can barely influence the direct opinion of the masses, however they can set the topics people think about and perceive as important (McCombs & Shaw, 1972). One of the main reasons for people to use Twitter is to share news and to comment on them (Java, Song, Finin, & Tseng, 2009; Schultz, Sheffer, et al., 2010; Subašić & Berendt, 2011). By setting a certain agenda in the main news, mass media influence what people think about and thus to an extend mass media influence what people comment on and share on Twitter.
A second way of possible influence are the newspapers and journalists themselves, who bring the topics directly to Twitter via their Twitter accounts. One function of such accounts is the promotion of the newspaper’s current content (Broersma & Graham, 2012). Especially print media set up tweets that use formulations close to their headlines, including links to the respective article (Ausserhofer & Maireder, 2013; Schultz et al., 2010). The reach of Twitter accounts of newspaper organizations can be much larger than their actual number of subscriptions (Ahmad, 2010; Armstrong & Gao, 2010). This large audience can then, again, comment on the news and share them (Armstrong & Gao, 2010; Schultz et al., 2010) so that the news potentially spread quickly and widely.
Next to this, Twitter can be considered a cheap, quick and easy-to-access source of high value in times where journalists have to produce a lot of content in very little time
(Broersma & Graham, 2012; Schultz et al., 2010). Via Twitter journalists can connect to their community by airing questions relevant to a current story or news item, thereby bringing the topic to Twitter (Ahmad, 2010; Armstrong & Gao, 2010; Broersma & Graham, 2013). The idea behind this is that anyone interested can answer this question via another tweet, both taking part in a discussion and spreading the topic further among the followers of the answering person. Next to connecting to the general public, journalists have conducted complete interviews via Twitter with both politicians and journalists (Broersma & Graham, 2012). This, again, brings topics initially intended for the newspaper to Twitter.
However, despite these possible ways of influence documented in the literature, empirical evidence for political agenda-setting between Twitter and newspapers is rare. Neuman et al. (2014) present a study on agenda-setting between social media and traditional media in the US. They find for eleven out of 29 political topics that the attention given to a topic in social media could be better predicted by former attention in both social media and traditional media than only with former attention in social media alone, which indicates influence as described above. However, they do not disentangle the role of the different social media included, namely Twitter, blogs, and forum commentaries. To the best knowledge of the author, the only comparable European case study so far was carried out by Ceron et al. (2014). They provide a single-case study on the topic of corruption in Italy. Their findings suggest that the agenda-setting power remains with the traditional media and flows from printed news to online news to social media.
Despite these findings, one should not extrapolate from two case studies. Further evidence from different cases is needed and shall be provided by this paper. It remains open if and to what extend the topics covered in German newspapers transfer to Twitter. For the analysis the following hypothesis is set up:
H2: For topics covered both in the newspapers and on Twitter, the
newspaper’s relative attention given to a topic is a predictor for the relative attention given to the topic on Twitter.
Influence from Twitter to newspapers
Agenda-setting theory provides also a theoretical framework to consider Twitter to be influencing newspapers: reversed agenda-setting. McCombs (2014) describes reversed agenda-setting as “a situation where public concern sets the media agenda” (p. 14). It describes instances where the media has to cover a certain topic because the general public is interested in it. Thereby, reversed agenda-setting can be seen as a restraint to the agenda-setting power of traditional mass media. Using Twitter it becomes easier to publicly express personal concerns or interests and to connect to people holding similar concerns. Thereby, Twitter facilitates the expression of ‘public concerns’ and thus can hand over a certain agenda-setting power to Twitter users.
The mechanism through which topics transfer from Twitter into newspapers is via the journalists, who monitor Twitter on a regular basis. Broersma and Graham (2012) describe Twitter as a new beat. Traditionally beats are places where journalists go on daily routines to gather news, for example courts, police stations or city halls (Tuchman, 1978). The beat Twitter is attractive for journalists because of being accessible without the costs of time and travels outside the office, while still having access to a wide range of utterances.
Besides acquiring input to spice up a story with tweets (e.g. Broersma & Graham, 2013; Schultz et al., 2010), which is not actually an influence from Twitter to
newspapers as described above, Twitter has been reported to trigger news stories. This mostly happens for celebrities’ tweets (Broersma & Graham, 2013). However, a more political example is a prom held by right-wing fraternities in Austria, which caused a lot of upset reactions on Twitter in advance, whereas national mass media started their coverage only on the day of the prom (Ausserhofer & Maireder, 2013). Next to the trigger-function, Twitter stories can also have a stand-alone news value (Broersma & Graham, 2012, 2013). For example, The Guardian covered the reaction of the Twitter community on a broader media campaign against Nick Clegg under the hashtag ‘#nickcleggsfault’ (Arthur, April, 2010; Broersma & Graham, 2012).
evidence of the actual impact from Twitter to newspapers remains unclear. Besides their general findings, Broersma and Graham state in their study that Twitter is used as a beat more intensively in the Netherlands than in the United Kingdom. A less intensive use of Twitter restricts possible ways of influence as described above. Schultz et al. (2010) find sports journalists in the US to differ in their usage of Twitter by age, which again might affect the actual applicability of the outlined mechanisms. Empirical
evidence from the US suggests that possible influence depends on the topic area. In line with this, Neuman et al. (2014) find for 18 out of 29 topics an influence from online media to traditional media. However, in a case study on corruption, Ceron et al. (2014) find no agenda-setting effects from Twitter to traditional media. Overall, this indicates differences between countries and the influence of Twitter in Germany remains open. For the analysis the following hypotheses are set up:
H3: For topics covered both in the newspapers and on Twitter, Twitter’s relative attention given to a topic is a predictor for the relative attention given to the topic in the newspapers.
H4: Topics about public order, media- and technology-related topics, topics with a short news life cycle and entity-oriented issues are more important on Twitter than other topics.
The reader should note that H2 and H3 are not necessarily contradicting. In contrast, different authors suggest a reciprocal relationship of influence between online media and traditional media (for social media in general see Neuman et al. (2014), for blogs see Wallsten (2007), for YouTube see Sayre et al. (2010)), which would result in a partial acceptance of both hypotheses.
Candidates on Twitter
Hopes were high that Twitter might open a new, inclusive space for discussion and interaction among citizens and politicians (Bruns, Burgess, et al., 2011; Vergeer et al., 2013). However, looking at the use of Twitter by political candidates, most studies so
far find a one-way communication pattern. Twitter was found to be mostly used for information dissemination (Golbeck, Grimes, & Rogers, 2010), ‘impression management’ (Graham, Jackson, & Broersma, 2014; Jackson & Lilleker, 2011), self-promotion
including personalization strategies (Enli & Skogerbø, 2013; Jackson & Lilleker, 2011; Suiter, 2015) as well as a channel of justification against traditional media coverage (Jackson & Lilleker, 2011). From this, it can be expected that political candidates tweet either along their personal agenda, including tweets about their daily activities and schedules (Jackson & Lilleker, 2011), or along the agenda of traditional media, by linking to informative articles.
However, Enli and Skogerbø (2013) report from their interviews that Norwegian politicians see dialogue as an important function of Twitter. They can confirm such usage by finding in more than half of the tweet sample patterns of a dialogue, such as @-mentions or re-tweets. @-mentions link to another Twitter account, re-tweets are used to tweet someone else’s tweet again to spread it further. Along this line, Ausserhofer and Maireder (2013) find for political topics clear interactions between politicians and citizens. A more dialogue-oriented candidate behaviour could point towards more discussion and tweeting closer along the Twitter agenda.
Next to possible reactions of the candidates to the media-agenda, a candidate tweet can be newsworthy on its own and trigger a news story or a report about the tweet itself (Broersma & Graham, 2012). Thereby, the politician receives both attention from and can influence the agenda of traditional media (Parmelee, 2013; Vergeer et al., 2013).
In spite of these findings, one should be careful to generalize them. Firstly, one has to take into account that most studies focus on a certain election in one country, barely taking into account specific conditions, such as the electoral system or how well citizens are used to social media (Graham et al., 2014). Secondly, politicians’ Twitter usage is still developing and probably further expanding. Multiple studies acknowledge that they mainly looked at the pioneers among tweeting politicians (Ausserhofer & Maireder, 2013; Golbeck et al., 2010; Jackson & Lilleker, 2011; Vergeer et al., 2013). Furthermore, regardless of whether candidates foster one- or two-way communication on Twitter, it
remains open, whether they follow a given agenda or whether they tweet along their personal agenda. The following hypotheses are set up in this paper:
H5: Political candidates on Twitter tweet along their personal agenda. H6: Political candidates on Twitter tweet along the newspapers’ agenda. H7: Political candidates on Twitter tweet along the public Twitter agenda. H8: Political candidates on Twitter can influence the newspapers’ agenda.
Again, the reader should not that H5, H6 and H7 must not be contracting, as political candidates might simultaneously set up tweets in response to the newspapers’ articles, inform followers about daily activities and join discussions with voters.
Methodology
The European Elections 2014
In the European Elections in 2014 the European Parliament was voted directly by the European populace. Voting took place between May 22 and May 25, 2014. In Germany the voting day was May 25. The European Parties had front runners, the two largest parties competing for the position of the President of the European Commission. The conservative European People’s Party’s front runner was Jean-Claude Juncker, for the social-democratic Party of European Socialists it was Martin Schulz. The two
candidates staged several election debates in TV, the German debate took place on May 8, 2014.
Data Collection and Processing
Newspaper Articles. In order to find the topics covered by traditional media, articles from several printed newspapers were retrieved from an online database. To decide on the newspapers, the following decision rule applied: From the available newspapers, the most important should be selected. Therefore, those with the highest circulation rate, which is the average number of copies distributed per day, were chosen (n.A., 2015). Within the given time for this paper, the first five newspapers from this
Table 1
Newspaper Data
Newspaper Nr. of Articles Circulation Rate
Frankfurter Allgemeine Zeitung 4,988 307,000
Rheinische Post 4,454 320,000
Sächsische Zeitung 3,940 238,000
Südwest Presse 2,250 290,000
Total 15,632 1,155,000
Note. Circulation Rate retrieved from n.A. (2015)
list could be processed. This resulted in the following selection: Süddeutsche Zeitung, Frankfurter Allgemeine Zeitung, Südwest Presse, Rheinische Post and Sächsische Zeitung.
To select articles from the newspapers, first the time frame was set according to the time frame of the available tweets (see below): April 10, 2014 to September 21, 2014, spanning 164 days. Inside this time span, all articles were taken into analysis, which fit the following boolean expression: (eu OR “europäische union” OR europa).
Unfortunately, due to some technical problems the Süddeutsche Zeitung’s articles were lost. See Table 1 for an overview over the exact amount of articles downloaded per newspaper and their circulation rate.
In preparation for the analysis all words were stemmed, using the
GermanStemmer-Snowball algorithm included in the NLTK package (Bird, Klein, & Loper, 2009). Stemming removes the distinction between for example words like
‘support’, ‘supported’ and ‘supporting’, which essentially have the same meaning. Next to this, stop-words, punctuation and all words that occur less than 10 times in the whole collection (Zhao et al., 2011) were removed from each article. Note that besides usual stop-words, the following words were removed as well, as they do not differentiate between topics: ‘europa’, ‘eu’, ‘europawahl’, ‘ep14’ and ‘ep2014’.
General Tweets. Tweets containing the hashtag ‘#ep2014’ were collected
between April 10, 2014 to September 21, 2014 from the the Twitter streaming API using DMI-TCAT (Borra & Rieder, 2014). The dataset covers 1,023,414 tweets. To identify
Table 2
General Twitter Data
State of Selection Nr. of tweets left Total Number of tweets: 1,023,414 Tweets from German accounts: 78,546
Tweets in German: 55,813
Empty tweets removed: 55,494
tweets that were most likely set up in the context of German media, only tweets from German accounts and written in German are included in the analysis. The nationality of the Account is given in the JSON file retrieved via the Twitter streaming API. The language of the Tweet is determined by comparing the words of the tweet with
stop-words from different languages. From these tweets, punctuation and stop-words were removed. Along with the punctuation the #-sign was removed as well, so that for example ‘#ttip’ and ‘ttip’ are treated as the same. Due to the restricted number of characters for a tweet, it is practice on Twitter to use URL shortening services. An effect of this is that the URL itself does not include information about the content behind it. Therefore, all URLs were removed from the tweets. @-mentions were removed as well, because they do not contain any information about the topic of the tweet. This procedure resulted in a number of empty tweets after stemming, as they contained for example only an @-mention and a link. Table 2 depicts the data-cleaning process. The final dataset includes 55,494 tweets.
Candidate Tweets. In a first step, the author received a list of all 1053 German candidates for the European Parliament from the German Bundeswahlleiter (Federal Electoral Management Body). To identify the Twitter accounts of the candidates, the European Parliament offers a list with Twitter accounts of German Members of the European Parliament (MEPs) (Europaparlament, n.d.). By the time of writing, it contained 60 accounts. Next to this, every candidate name was searched manually via Google with the search term ((europe OR election) AND Twitter). A Twitter account was selected when it was clear that it belongs to the candidate. Possible identifiers were for example: a link from the official homepage of the candidate to the Twitter account,
mentioning of being MEP or candidate for the European Parliament on the account or the usage of the same picture as on the official homepage. If no such account could be found on the first results page of Google and no link was found on the official homepage of a candidate, it was assumed that no account existed. 196 Twitter accounts were identified. In a next step, from these accounts all tweets in the aforementioned time span were downloaded with the use of the Twitter API. In total 76,867 tweets were downloaded. Seven accounts were removed, because no Tweets were downloaded, either because the candidates did not tweet or because their tweets were protected, which prohibits downloading. This results in a final number of accounts of 189. Similarly to the general tweets, stop-words, punctuation, URLs and @-mentions were removed and stemming was conducted. Again, a number of tweets returned empty after stemming, so that they were removed as well. The final dataset contains 75,169 tweets. The median number of tweets per account is 146, so nearly one tweet per day. Table 3 further
provides an overview about the distribution of Twitter accounts between the parties and their mean number of tweets. Appendix A1 contains a list of all accounts, along with name of the according candidate, party affiliation and the number of Tweets.
Analysis Plan
Topic distribution per day in newspapers. The topics covered by the newspapers can be identified by applying a standard LDA (Blei, Ng, & Jordan, 2003). LDA is an unsupervised machine learning technique for automated content analysis (Trilling, 2015). LDA assumes that generating a set of documents underlies a certain probabilistic process: Given a defined number of topics for each collection of documents, in this case for example all articles in the analysis, each document is assigned to a certain distribution of these topics. Each topic has a certain distribution of words, from which the words for the document are selected according to the topic distribution. The topics generated by the model consist of a set of words, left to the researcher to
interpret. Each article is assigned a value, indicating the importance of each topic in the article. Python implementation of the LDA was conducted with the ‘gensim’ module
Table 3
Candidate Data
Party Nr. of Accounts Mean nr. of tweets
SPD 43 312 CDU/CSU 42 369 FDP 24 478 B’ 90/ Die Grünen 18 312 Die Partei 10 419 Die Piraten 10 345 AfD 7 496 Freie Wähler 7 420 ÖDP 7 388 Others 21 271 Overall mean: 407 Overall median: 146
Note. All means rounded to the full tweet.
(Řehůřek & Sojka, 2010). To run an LDA, the number of topics must be defined beforehand. It was set to 30. Different scores were tested (lowest: 20, largest: 70) and 30 resulted in the best topic coherence with only marginal differences in the topic results. Compared to other LDA models, this is a relatively low topic number (e.g. Chang, Gerrish, Wang, Boyd-Graber, & Blei, 2009; Zhao et al., 2011). This is considered reasonable, given the pre-selection of topics via the search terms.
To arrive at per-day measurements of the topics in the newspapers, the share of each topic over all topics is calculated. This assumes that a growth of importance of one topic leads to a decrease in importance of other topics. In case of newspapers, this assumption can be made, as the number of pages devoted to politics can be assumed to stay roughly the same over the covered time span. In case of Twitter, this assumption is more problematic, as the number of tweets is not limited. However, supposing that the mean number of tweets a user tweets per day stays approximately the same over time, it is again rather the content of the tweets that changes, than the number of tweets.
Topic distribution per day on Twitter. To identify the topics on Twitter, one could think of conducting an LDA as well. However, due to the shortness of tweets compared to newspaper articles, this has been reported to lead to no meaningful results
(Hong & Davison, 2010; Zhao et al., 2011). Nevertheless, an LDA with tweets was conducted, also applying different strategies to improve results, such as aggregation (Hong & Davison, 2010) (making one document from all tweets per day) or using only frequent Twitter users (Zhao et al., 2011). As expected, results were not meaningful.
Instead, the relevance of the topics from the newspapers is investigated on Twitter. To do so, adequate topics from the newspapers need to be identified. The first criterion is that a topic can be interpreted meaningfully. Secondly, the topic needs to be
politically relevant. This excludes certain topics that might occur in the newspapers, for example surrounding the soccer Champions League, but not in the Tweets, due to the filtering by the hashtag ‘#ep2014’. A third criterion is the topic coherence measurement, so that better identified topics are used with preference. Lastly, some topics are found the be more likely to occur on Twitter, namely social issues, topics about public order, media- and technology-related topics and entity-oriented issues (Ausserhofer &
Maireder, 2013; Neuman et al., 2014; Zhao et al., 2011). Such topics will be added as well. For each topic, six words are selected by the author for a search within the tweets, hereafter called topic words. For the search, all tweets of one day are aggregated into one document, separately for the general tweets and the candidates’ tweets. For each document and for each topic the ratio of topic words compared to all words is calculated.
Vector-Autoregression Analysis. In order to compare the development of the topics over time, a time series model is set up. As discussed before, theory suggests influence in both directions between Twitter and newspapers. To allow for a such a relation, VAR analysis is appropriate (Vliegenthart, 2014). For each topic, one VAR model is set up. Within each model, there are three time series: the share of the topic in the newspapers over time, the share of the topic in the general tweets over time and the share of the topic in the candidates’ tweets over time. Every VAR consists of three multivariate regressions, each having one of the time series as dependent variable. In all three equations, all three time series’ current and a number of past values are the
independent variables (Sayre et al., 2010). VAR analysis requires stationary data, which means that the mean and the variance of the time series do not change considerably
over time. It is tested with an Augmented Dickey-Fuller Test (ADF) (Lütkepohl, 2005). Furthermore, the number of lags needs to be defined. This number indicates how many past values of each variable are included in the analysis. Based on former studies, that find agenda-setting effect to occur mostly within seven days (Ceron et al., 2014; Sayre et al., 2010; Wallsten, 2007) the number of lags is set to a maximum of seven.
VAR-implementation is accomplished in R, using the ‘vars’ package (Pfaff, 2008) and the ‘forecast’ package (Hyndman & Khandakar, 2008). Within the functions of the packages, the model with the best fit can be chosen from the models with lags from one to seven, based on the Akaike Information Criterion (AIC), as Vliegenthart (2014) suggests. Coefficients from a VAR analysis should, however, be treated with caution as there is a potential for collinearity, due to the multiple lags. This means that for
example a topic raised by the Twitter community, might influences then what is covered by the newspapers, which then inspires the candidates to join the discussion on Twitter. To make up for this, a series of pairwise Granger-causality tests will be conducted as well. Granger-causality, though, should not be mistaken for causality. A variable x Granger-causes a variable y if values of y can be better predicted by past values of both x and y, instead of solely by past values of y.
Results
Topics in the newspapers
In a first step, the topics covered in the newspapers should be assessed.
Unfortunately, the results are not as clear as hoped for. Topic coherences, measured using the UMass measure (Mimno, Wallach, Talley, Leenders, & McCallum, 2011) are unsatisfactory. Other studies present UMass values in a negative single-digit range (Stevens, Kegelmeyer, Andrzejewski, & Buttler, 2012), whereas in this study measures range between -463 to -1150. As it is usual the model produces topics that are
interpretable along with topics that are not as good to interpret and could thus be called artefacts (Stevens et al., 2012). Table 4 presents some of the interpretable topics along with a possible interpretation and some of the topics considered artefacts, as well as the
Table 4
Topics in the newspapers
First words building a topic UMass Interpretation Words for search in Tweets Topics with interpretation
ukrain; russland; russisch; -463.14 Ukraine-/ ukrain; russland; russisch; putin; regier; präsident; Crimean Crisis putin; moskau; sanktion moskau; sanktion
prozent; euro; bank; -542.38 Economic Crisis prozent; euro; bank;
unternehm; rund; unternehm; wirtschaft;
wirtschaft; ezb ezb
junck; parlament; brussel; -542.93 Front Runner junck; brussel;
kommission; schulz; kommission; schulz;
spitzenkandidat; spitzenkandidat;
europaparlament europaparlament
1; spiel; fc; 2; 3; bay; -604.02 Champion’s
league;fussball; League
mannschaft; champion
Topics without interpretation
musik; band; data; buhn; -844.12 uhr; hongkong
faz; kind; partei; opel; -1127.73 berlusconi; kindergeld;
Note. UMass measures are rounded to the second decimal.
UMass coherence measurements. The complete results can be found in Appendix B1.
The three best to interpret political topics Ukraine-/ Crimean Crisis, economic crisis and front runner are at the same time the most, second and fourth most
consistent topics. Additional topics, which might be more likely to come up on Twitter, based on the aforementioned categories, were not found. However, the Front Runner topic is entity-oriented, and could thus play a more important role on Twitter than the other two topics. To compare newspapers and Twitter these three topics were selected. The six search terms per topic, for which it was search within Tweets, are also shown in Table 4.
Newspapers’ topics on Twitter over time
First, stationarity needs to be tested. The ADF tests for a H0 ‘The Data is not stationary’. p-values for the ADF for every time series are displayed in Tables 5, 7 and 8. A time series is considered stationary if the ADF is significant. All time series are stationary, except for the candidates’ tweets for topic 2. Thus, all time series of topic 2 need to be differentiated once before the models can be fitted.
Ukraine-/ Crimean Crisis. Figure 1 displays the three time series for the Ukraine-/ Crimean Crisis topic. The coverage of the topic in the three media seem rather different. Whereas newspapers show regular coverage, candidates show a slight peak in the beginning of the time frame and higher amounts of coverage towards the end of the series. The beginning of the more intense coverage starts at the time when the aircraft of flight MH17 was shot down in Ukraine (Malaysia Airlines, 2014). The general Twitter community shows two intense peaks surrounding April 17/18, which fits the conference on the Urkaine crisis in Geneva (n.A., 2014c), and the end of June, around the time when OSCE hostages were released (Vasovic, 2014). Despite the differences, the topic is covered in all three media, which supports H1.
Yet, VAR analysis might reveal possible influences of the relative attention between the media. The regression on the influence from newspapers to Twitter (H2) is first depicted in Table 5. The model explains barely any variance, with a rounded adjusted R2 = 0.00. None of the predictors are found significant. Thus, none of the three time
series’ past values included are helpful to predict the Twitter communities’ attention on the Ukraine-/Crimean Crisis. This is contradicting H2. The Granger-causality,
displayed in Table 6 confirms this finding. Newspapers do not Granger-cause the general tweets’ level of attention for the Ukraine-/Crimean Crisis.
Considering the influence from Twitter to the newspapers (H3) the model itself explains a considerable amount of variance, with an adjusted R2 = 0.33. However,
significant predictors are solely past values of the newspapers’ attention level itself. No Twitter attention value is a significant predictor, which is contradicting evidence for H3. Again, this is supported by the Granger-causality test: Newspapers’ attention is not
Figure 1 . Time Series Ukraine-/ Crimean Crisis
European ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean Elections
0.00 0.05 0.10 0.15
Mai Jun Jul Aug Sep
T
opic Share
Candidates' Tweets General Tweets Newspapers
Granger-caused by the general tweets’ attention.
Turning to the candidates’ manner of tweeting, one can first look at the third regression of Table 5. As for the newspapers, the model explains a certain amount of variance, with an adjusted R2 = 0.20, but, again, this is due to the candidates’ tweets’ own values one and two days in the past. This points towards the idea that candidates follow their personal agenda (H5). No other medium is found to be significant, which is evidence against H6 and H7. This is further supported, as neither the general tweets’, nor the newspapers’ attention Granger-cause the candidates’ attention.
Lastly, looking at H8 the VAR analysis does not support the hypothesis, as the candidates’ tweets have no significant capacity to predict new newspapers’ attention in the model. However, the candidates’ tweets are found to Granger-cause the newspaper coverage, which means the former are useful to predict the latter. As a remainder, Granger-causality should not be confused with causation in a more common statistical
Table 5
VAR analysis Ukraine-/ Crimean Crisis
General Tweets Newspaper Candidates’ Tweets
coef SE coef SE coef SE
constant 0.00 0.00 0.03 0.01 0.00 0.00
newspaper lag 1 0.03 0.13 −0.02 -0.07 0.00 0.02
general tweets lag 1 0.13 0.09 −0.32 0.36 0.02 0.08
candidates’ tweets lag 1 0.06 0.09 0.58 0.38 0.22∗ 0.09
newspaper lag 2 0.03 0.02 −0.07 0.07 0.00 0.02
general tweets lag 2 0.03 0.09 0.20 0.36 −0.11 0.08
candidates’ tweets lag 2 −0.04 0.09 −0.19 0.39 0.22∗ 0.09
newspaper lag 3 0.03 0.02 0.05 0.07 0.01 0.02
general tweets lag 3 0.01 0.09 −0.03 0.36 −0.02 0.08
candidates’ tweets lag 3 −0.06 0.09 0.42 0.39 0.09 0.09
newspaper lag 4 0.00 0.02 −0.07 0.07 0.00 0.02
general tweets lag 4 0.00 0.09 −0.37 0.36 0.02 0.08
candidates’ tweets lag 4 −0.04 0.09 −0.09 0.39 0.13 0.09
newspaper lag 5 0.00 0.02 −0.01 0.07 −0.01 0.02
general tweets lag 5 −0.03 0.09 0.08 0.36 −0.12 0.08
candidates’ tweets lag 5 −0.07 0.10 0.66 0.43 0.11 0.10
newspaper lag 6 −0.01 0.02 −0.05 0.07 0.00 0.02
general tweets lag 6 −0.02 0.08 −0.27 0.36 −0.07 0.08 candidates’ tweets lag 6 −0.08 0.10 −0.57 0.42 −0.13 0.10
newspaper lag 7 0.02 0.02 0.55∗∗∗ 0.07 0.00 0.02
general tweets lag 7 −0.03 0.08 0.21 0.36 0.00 0.08
candidates’ tweets lag 7 −0.09 0.10 0.76 0.01 −0.03 0.10
ADF p < 0.01 p < 0.01 p < 0.01
adjusted R2 0.00 0.33 0.20
Note. *** → p < 0.001; ** → p < 0.01; * → p < 0.05. All values rounded to the second
decimal. N = 164.
sense. These observations point towards a third variable outside the model driving both newspaper coverage and the candidates’ tweets. Evening television news could be such a third variable, as they are broadcasted, then candidates might react on it on Twitter and on the next morning newspapers cover the same news. In this scenario the
candidates’ tweets are not a reason for the newspaper coverage, however they are helpful for its prediction.
Economic Crisis. The three time series about the Economic Crisis are shown in Figure 2. The newspapers show again a rather consistent rhythmic pattern with a slight
Table 6
Pairwise Granger-causality
General Tweets Newspapers Candidates’ Tweets
Ukraine-/ Crimean Crisis general tweets 0.67 0.73 newspapers 2.61 0.26 candidate tweets 0.72 2.61∗ Economic Crisis general tweets 0.28 0.62 newspapers 0.99 2.08∗ candidate tweets 0.91 0.91 Front Runner general tweets 0.43 0.90 newspapers 2.86 0.04∗ candidate tweets 4.35∗ 27.69∗∗∗
Note. All values are rounded to the second decimal. Number of lags included for Ukraine-/ Crimean crisis and Economic Crisis-topic: seven, for the Front Runner topic: one. Reading exam-ple: For the topic Ukraine-/ Crimean Crisis candidates’ tweets are not Granger-caused by general tweets (F-value= 0.72), however they are Granger-caused by newspapers (F-value= 2.61).
peak right after the European Elections. The candidates’ tweets look similar, however the amplitudes are smaller and the peak is shortly before the European Elections. In the general tweets the interest in the topic is generally low, but peaks towards the end of the observed time span. As for the Ukraine/ Crimean Crisis, the Economic Crisis topic is present in all three media, supporting H1.
On the influence from newspapers to Twitter (H2), one can first look at the first regression of the VAR analysis in Table 7. The model is set up with seven lags. A considerable amount of variance is explained, as shown by the adjusted R2 = 0.31.
However, only past values of the general tweets’ given attention are significant
predictors and no past newspaper value. Furthermore, newspaper do not Granger-cause the general tweets (Table 6). Both findings add to the contradicting evidence for H2 from the Ukraine-/ Crimean Crisis topic.
Looking at the reverse influence, from Twitter to newspapers (H3), there is a similar picture. The model covers a large amount of variance with an adjusted R2 = 0.63. But,
as seen before, only the newspapers’ own past values are significant predictors. The Granger-causality test approves that the general tweets do not improve the prediction
Figure 2 . Time Series Economic Crisis
European ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean Elections
−0.2 −0.1 0.0 0.1 0.2
Mai Jun Jul Aug Sep
T
opic Share
Candidates' Tweets General Tweets Newspapers
for the newspapers’ values. All together, this evidence is further opposing H3.
Considering the model on the candidates’ tweets, again, a reasonable amount of variance is explained (adjusted R2 = 0.37). Despite this again only past values of the candidates’ tweets themselves, are significant. This is considered further evidence for H5 and against H6 and H7. Furthermore, the newspaper coverage is Granger-causing the candidates’ tweets, which indicates, that newspaper coverage are helpful to predict the candidates’ tweets. This supports the idea that candidates tweet along the newspapers’ agenda (H6).
For this topic, no influence from the candidates’ tweets to the newspaper is found. Neither are the candidates’ tweets a predictor for the newspapers’ values in the VAR analysis (Table 7), nor are the candidates’ tweets Granger-causing the attention to the topic in newspapers. Both is considered evidence against H8.
Table 7
VAR analysis Economic Crisis
General Tweets Newspaper Candidates’ Tweets
coef SE coef SE coef SE
constant 0.00 0.00 0.00 0.00 0.00 0.00
newspaper lag 1 −0.13 0.08 −0.57 ∗∗∗ 0.09 0.02 0.02
general tweets lag 1 −0.68∗∗∗ 0.09 0.06 0.10 0.00 0.03
candidates’ tweets lag 1 −0.10 0.30 0.10 0.33 −0.74 ∗∗∗ 0.09
newspaper lag 2 0.00 0.08 −0.63 ∗∗∗ 0.09 0.01 0.02
general tweets lag 2 0.67∗∗∗ 0.10 0.07 0.11 0.00 0.03
candidates’ tweets lag 2 −0.01 0.36 0.08 0.40 −0.61 ∗∗∗ 0.11
newspaper lag 3 −0.15 0.08 −0.59 ∗∗∗ 0.09 0.03 0.02
general tweets lag 3 −0.55∗∗∗ 0.11 0.11 0.13 0.00 0.03
candidates’ tweets lag 3 −0.08 0.38 −0.05 0.43 −0.63 ∗∗∗ 0.12
newspaper lag 4 −0.06 -0.09 −0.53 ∗∗∗ 0.10 −0.01 0.03
general tweets lag 4 −0.47∗∗∗ 0.12 0.03 0.13 −0.05 0.04
candidates’ tweets lag 4 0.33 0.40 −0.26 0.45 −0.45 ∗∗∗ 0.12
newspaper lag 5 −0.10 0.08 −0.6 ∗∗∗ 0.09 −0.02 0.03
general tweets lag 5 −0.43∗∗∗ 0.13 0.02 0.14 −0.03 0.04
candidates’ tweets lag 5 0.55 0.39 −0.13 0.43 −0.34 ∗∗ 0.12
newspaper lag 6 −0.10 0.08 −0.49 ∗∗∗ 0.09 0.01 0.02
general tweets lag 6 −0.31∗∗ 0.12 0.03 0.13 0.01 0.03
candidates’ tweets lag 6 0.38 0.34 −0.08 0.38 −0.20 0.10
newspaper lag 7 −0.06 0.08 0.2∗ 0.09 0.01 0.02
general tweets lag 7 −0.24∗ 0.10 0.07 0.11 −0.14 0.08
candidates’ tweets lag 7 0.00 0.00 −0.07 0.30 −0.14 0.08
ADF p < 0.01 p < 0.01 p = 0.191
ADF-test 1st difference p < 0.01 p < 0.01 p < 0.01
adjusted R2 0.31 0.63 0.37
Note. *** → p < 0.001; ** → p < 0.01; * → p < 0.05. All values rounded to the second decimal.
Figure 3 . Time Series front runner
European ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean ElectionsEuropean Elections
0.00 0.05 0.10 0.15 0.20
Mai Jun Jul Aug Sep
T
opic Share
Candidates' Tweets General Tweets Newspapers
Front Runner. Looking at the plotted time series of the front runner topic in Figure 3, the series look more similar than for the other topics. The main difference between the series is that the general tweets have more peaks. The peak in May fits the German TV debate between Juncker and Schulz (Schiltz, 2014). The following three peaks fit the announcement of opposition against Juncker as European Commission president (Pop, 2014) end of May, him being proposed as President by the end of June (Dominiczak, 2014) and his election in mid July (n.A., 2014b). The peak in September is when the whole Commission was introduced (n.A., 2014a). For the newspapers and the candidates’ tweets one can see similar, though less pronounced peaks. Again, the topic is covered in all three media, supporting H1.
The front runner topic is the only one the analysis, which fits one of the criteria outlined above (entity-oriented) for topics that are expected to be more important on Twitter than other topics (H4). When comparing the time series of the general tweets in
the three topics, one might conclude that the front runner topic is indeed given more attention. However, based on a one-sided paired t-test to compare the mean of the front runner time series with the mean of the respective other topic, this presumption holds only for one topic. The front runner topic (M = 0.03, SD = 0.04) is given more
attention compared to the Ukraine-/ Crimean Crisis topic (M = 0.00, SD = 0.01), t(163) = 7.88, p < 0.001, 95% CI [0.02, ∞]. The means of the front runner topic and the Economic Crisis Topic (M = 0.02, SD = 0.03) do not differ (M = 0.00, SD = 0.01), t(163) = 0.59, p = 0.279, 95% CI [0.00, ∞].
Turning to the VAR analysis, results are shown in Table 8. Note that for this model, only one lag is most appropriate. As seen before, the general tweets have solely a past value of the same medium as significant predictor. For H2 on the newspapers’ influence on Twitter, this is further counter-evidence, additionally supported, as no Granger-causality is found for the influence on the general tweets.
Moreover, within the front runner topic, no reverse influence from general tweets to newspapers (H3) is found. General tweets are no significant predictor for the
newspapers’ attention level, which is approved by finding no Granger-causality from the general tweets to the newspapers. Both is evidence against H3.
Examining the candidates’ manner of tweeting on the front runner topic, the VAR analysis shows that not only the past value of the variable itself is a significant
predictor, but the newspapers’ attention level, as well. This would point towards both H5, candidates tweeting along their personal agenda and H6, candidates tweeting along the newspapers’ agenda. The latter relationship is also confirmed by the
Granger-causality. Furthermore, on the front runner topic newspapers are found to be Granger-caused by the candidates’ tweets, pointing towards an influence from the candidates’ to the newspapers’ agenda (H8).
Summary. For the overlap of agendas between Twitter and newspapers (H1) it can be concluded that for the three topics under review the hypothesis can be accepted. All three topics were present in both media. For none of the three topics an influence from the newspapers’ attention level to the amount of attention devoted by the general
Table 8
VAR analysis Front Runner
General Tweets Newspaper Candidates’ Tweets
coef SE coef SE coef SE
constant 0.00∗ 0.00 0.01∗∗∗ 0.00 0.00∗∗∗ 0.00
newspaper lag 1 0.14 0.15 0.13 0.08 0.07∗ 0.03
general tweets lag 1 0.28∗∗∗ 0.08 −0.01 0.04 0.01 0.02
candidates’ tweets lag 1 0.59 0.38 1.00∗∗∗ 0.19 0.20∗ 0.08
ADF-test p < 0.01 p = 0.032 p < 0.01
adjusted R2 0.12 0.21 0.09
Note. *** → p < 0.001; ** → p < 0.01; * → p < 0.05. All values rounded to the second
decimal. N = 164.
Twitter community was found, which leads to rejecting H2 for the three topics at hand. The same is true for the reverse influence, from Twitter to newspapers. No pattern of influence was found, thus H3 is rejected as well. The differences in attention given to the entity-oriented Front Runner topic could only be found for in comparison with the Ukraine/ Crimean Crisis, not compared to the Economic Crisis topic. Thus, it is only partial evidence found for H4.
On the candidates’ tweeting habits evidence for H5, tweeting along their personal agenda, was found within all three topics and can thus be accepted within this context. Partial evidence was found for H6, candidates’ following the newspapers’ agenda. The hypothesis can be accepted for the topic of Economic Crisis and the Front Runner topic, however for the Ukraine/Crimean-crisis no such relation was found. Furthermore, no evidence was found to support the idea of candidates joining a general Twitter debate (H7). Looking at the influence of candidates’ tweets on newspapers (H8), evidence is twolfold. In the Ukraine/ Crimean-crisis topic findings point towards a spurious
correlation, which involves variables not included in the model. For the Economic Crisis, the candidates had no influence via Twitter on the newspapers, whereas on the question of the Front Runner influence, as defined above, could clearly be identified.
Discussion and Conclusion
With this study, former findings of a topic overlap between traditional mass media and Twitter can be approved (Ausserhofer & Maireder, 2013; Jungherr, 2014; Neuman et al., 2014; Zhao et al., 2011). Coming back to the research question, ‘What is the influence of Twitter and newspapers on each other’s agendas?’, it can be concluded that between German newspapers and the German Twitter stream, no influence was found. This is a surprising finding, taking into consideration the many potential mechanisms of influence as well as findings of former studies.
The missing influence between the general Tweets and newspapers is not in line with the study of Neuman et al. (2014). This could possibly mean, that their findings are rather due to blogs and forum commentaries, which are also included in their analysis. A more probable explanation are possible country differences between the US, where Neuman et al.’s study was conducted and Germany. The possibility of country differences in the usage and effects of Twitter is raised as well by Graham et al. (2014). They find political candidates to use Twitter differently and propose one reason could be the popularity and history of social media in a country. The estimated popularity of Twitter in the US is considerably higher. About 20% of the population uses Twitter at least monthly (n. A, 2015), whereas about 12% do so in Germany (n.A., 2015). Ceron et al. (2014) present influence from online newspapers to Twitter. One suggestion to
explain the difference in findings is that they used online newspaper in their study. It could be, that online newspapers take a lead due to the earlier publication of the news. Whereas printed newspapers cover in the morning the news of the day before and in the evening the news of the day, online articles can be published earlier and might thus exert more influence on the quickly reacting Twitter community.
The combination of a topic overlap without influence points towards different driving forces for the importance of a topic in the respective medium. This fits to Jungherr’s (2014) suggestion of different media logics for traditional media and Twitter. Related to this, Ausserhofer and Maireder (2013), suggest that Twitter reacts more on topics with a short news life circle. Indeed, in the study at hand the general tweets seem
to peak in specific points in time instead of regularly covering the topic, whereas newspapers show a more even pattern.
Between the candidates’ tweets and newspapers, influence was found for some of the topics, confirming Broersma and Graham (2012), Parmelee (2013) and Vergeer et al. (2013). Next to this, the study suggests that candidates did, on the three given topics, not relevantly involve into the Twitter agenda, which further suggest more one-way communication than discussion. This again indicates a tweeting approach mainly to spread information, as found by Golbeck et al. (2010) for Members of the U.S. Congress in 2010, or what Jackson and Lilleker (2011) call ‘impression management’. Just as for the general Twitter stream, observations from other countries are different. In the Netherlands (Graham et al., 2014) and in Norway (Enli & Skogerbø, 2013) politicians were found to engage more in a discussion on Twitter. The German candidates resemble rather their more conservative colleagues from the UK (Graham et al., 2014; Jackson & Lilleker, 2011). Despite their lack of involvement with citizens, this study indicates candidates’ influence on the newspapers’ agenda, which suggests that exchange between politicians and journalists does occur in the German Twitter sphere. Seeing the
potential to influence the newspapers’ agenda via Twitter in combination with the rather low usage of Twitter by the German public, one could think that Twitter in Germany has not yet drawn the potential it could have for the public.
The approach presented in this study showed some challenges as well. Firstly, the LDA did not lead to coherent topics in the newspapers, therefore only three topics could be compared. Furthermore, this might have caused inaccurate measurements of the actual topic coverage. Secondly, when measuring the topic coverage on Twitter, it is not differentiated between a high amount of one or two of the searched words or an even occurrence of all search terms. In other words, finding six times the word ‘brussels’ would lead to the same outcome as finding once each of the topic words for the Front Runner topic. However, the latter pattern is more likely to actually cover the desired topic. Lastly, the results concerning the general Twitter stream should be treated with caution, as the selection of tweets based on the hashtag ‘#ep2014’ might have excluded
relevant Tweets. It is reasonable to think that only a part of the Tweets concerning the Ukraine-/ Crimean crisis and the Economic Crisis included this hashtag. For the Front Runner topic, however, it is likely that most tweets are included in the data.
Future research could be improved by applying the Twitter LDA-model to the tweets, as suggested by Zhao et al. (2011). They adapt the standard-LDA model to the shortness of tweets, which leads to better results. Then, results of the two LDAs could be compared. Unfortunately, up to the day of writing the application is only available in C++, which is unknown to the author. Furthermore, the newspaper-LDA might be improved by selecting smaller time frames, in which the coverage itself is more concentrated on certain topics.
This study surprisingly showed no influence between newspapers’ agenda and the general Twitter agenda in Germany, contradicting former findings from other countries. This highlights the importance for future research to look further into why countries differ in regard of Twitter’s influence on political communication. Reasons could be different stages of the same development, but also more systematic differences based on the political system or on the open-mindedness of a culture towards change. Either way, it underlines the importance of taking into consideration possible country particularities when studying Twitter phenomenons, especially to learn more about the (im)possibility of generalizing findings. Seeing the potential of Twitter to change previously
well-researched processes in political communication, such as agenda-setting procedures, knowledge about the circumstances to generalize new findings, will help to predict the future influence of Twitter on political communication and its consequences for
democracy.
References
Aelst, P. V., Sehata, A., & Dalen, A. V. (2010). Members of parliament: Equal competitors for media attention? an analysis of personal contacts between mps and political journalists in five european countries. Political communication, 27 (3), 310–325.
Ahmad, A. N. (2010). Is twitter a useful tool for journalists? Journal of Media Practice,
11 (2), 145–155.
Armstrong, C. L., & Gao, F. (2010). Now tweet this how news organizations use twitter.
Electronic News, 4 (4), 218–235.
Arthur, C. (April, 2010). Twitter says it’s all nick clegg’s fault in ironic swipe at newspapers. Retrieved 2015-06-06, from http://www.theguardian.com/technology/blog/2010/ apr/22/twitter-nick-clegg-newspaper-swipe
Ausserhofer, J., & Maireder, A. (2013). National politics on twitter: Structures and topics of a networked public sphere. Information, Communication & Society, 16 (3), 291–314. Bird, S., Klein, E., & Loper, E. (2009). Natural language processing with python. " O’Reilly
Media, Inc.".
Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent dirichlet allocation. the Journal of
machine Learning research, 3 , 993–1022.
Borra, E., & Rieder, B. (2014). Programmed method: developing a toolset for capturing and analyzing tweets. Aslib Journal of Information Management, 66 (3), 262-278. doi: 10.1108/AJIM-09-2013-0094
Broersma, M., & Graham, T. (2012). Social media as beat: tweets as a news source during the 2010 british and dutch elections. Journalism Practice, 6 (3), 403–419.
Broersma, M., & Graham, T. (2013). Twitter as a news source: how dutch and british newspapers used tweets in their news coverage, 2007–2011. Journalism practice, 7 (4), 446–464.
Bruns, A., Burgess, J., et al. (2011). #ausvotes: How twitter covered the 2010 australian federal election.
Ceron, A., Curini, L., & Iacus, S. M. (2014). Twitter and the traditional media: Who is the real agenda setter? In Apsa 2014 annual meeting paper.
Chang, J., Gerrish, S., Wang, C., Boyd-Graber, J. L., & Blei, D. M. (2009). Reading tea leaves: How humans interpret topic models. In Advances in neural information
processing systems (pp. 288–296).
Dominiczak, P. (2014). Juncker named as european commission president, as cameron warns eu could ’regret’ it. The Telegraph. Retrieved 2015-06-23, from http://
