The Impact of Online Platforms on Public Opinion Regarding the European Union Laura Sudulich, ULB Leonardo Baccini, LSE Introduction An information environment is vital to the capacity of citizens to learn about politics, their ability to link political preferences to parties and policies and to assess the performances of institutions and political actors. In most instances, people gather politically relevant information via the mass media rather than through direct interaction with political elites. The role of the mass media is therefore key to the correct functioning of society (Castells 2000; Delli Carpini and Keeter 1996). There is no shortage of evidence showing that the mass media exert a crucial influence on public opinion formation (Zaller (1992); for a review of major contributions, see Bennett and Iyengar (2008)). The mass media are particularly important when considered in relation to the formation of opinions, the forming of political awareness and voting behaviour in the context of the European Union, since citizens typically do not experience any direct contact with European institutions. Thus, citizens necessarily rely on the mass media when (in)forming their opinions about the European Union (hereafter EU). Additionally, a widespread ignorance of EU politics and policies among members of the European public is notorious and persistent. Over a third of Europeans are unable to name any EU institution; the percentage of citizens claiming to know little or nothing at all about ‘the people who run the various EU institutions and the leaders of the EU’ is a striking 73%. A similarly high number of subjects (74%) report knowing little or nothing at all about ‘The allocation of roles played by the various institutions (who does what?)’ (Eurobarometer 77.4). The 1 impact of radio, television and newspapers has been extensively debated in relation to electoral behaviour and attitude formation towards the EU (De Vreese 2003; De Vreese and Boomgaarden 2006; Schuck and De Vreese 2006), while the influence of the Internet_ is currently under-explored.1 This is in spite of a significant average penetration rate of 75.5% across member states, with several countries – The Netherlands, Sweden, Denmark – registering over 90% coverage (World Bank 2013).2 Moreover, a real word trend indicates that the Internet is rapidly becoming a fundamental source of information about European matters. Between November 2011 and December 2012, the percentage of citizens reporting that they had gathered information on the European Parliament on the World Wide Web went up by 10%, from 33% to 43% (Eurobarometer 78.2). This paper takes a first step to addressing whether consuming online information about the EU is responsible for variation in public opinion regarding it and whether this process bypasses or encompasses political knowledge. Specifically, we explore how different type of sources affect public opinion. In so doing, we bridge studies of public opinion and media effects to contributions addressing online source credibility (for a review see Metzger 2007). The latter are mostly based on experimental evidence and do not directly address political learning or opinions, while the former tends to rely on observational data and often disregards the nuances of digital environments. We use data from Eurobarometer 76.3 (November 2011) posing a large battery of questions on media use for political information consumption based on a representative Europe-wide sample of individuals. Importantly, this survey contains a unique array of items on online political news consumption habits. In addition, we provide an in-depth analysis of the effects of a particular type of platform - the Irish Referendum Commission website - on electoral behaviour in the context of the Fiscal Compact Referendum held in An exception is De Wilde et al. 2013, who content-analyse expressions of Euroscepticism in online media across 12 member states in the 2009 European Parliament campaign. However, their analysis does not deal with the consequences of online-based information and communication for public opinion. 2 http://data.worldbank.org/indicator/IT.NET.USER.P2 1 2 the Republic of Ireland on May 31st 2012.3 Making use of an original survey, we test the micro-foundations of our argument by isolating the effects of online information gathered specifically with regard to a single policy i.e. the Fiscal Compact. Specifically, we explore the impact of online newsgathering on the knowledge of such a policy, and on actual voting behaviour. We find that consuming information online has an effect on the opinions of those who consult the websites of traditional media and institutions; this is mediated by an increase in political knowledge. Websites that are by default a mixture of vetted information and nonfact-checked information prove unable to do this. In fact, Those who gather their information about the EU mostly via blogs, social networking sites or video sharing platforms experience no significant effects on opinions and neither they appear to gather knowledge. For those who actually experience an effect by visiting institutional websites or digital versions of news media an increase in knowledge turn into positive views about the EU. Along similar lines, the analysis on the Irish Fiscal Compact Referendum confirms that visiting an institutional website - namely the Referendum Commission’s site – increases self-perceived knowledge of the specific policy at stake and the likelihood of voting in favour of the Fiscal Compact Treaty. In the mare magnum of data available on the web, credible fact-checked sources stand out in providing viable knowledge that serves to orient citizens’ opinions. The article proceeds as follows: in the next section we outline the relationship between online-based information, political knowledge and public opinion on which we base our working hypotheses. We then discuss our main empirical strategy hinging on causal mediation analysis and describe the data. In the following section, we outline and discuss the results of our analysis. We then run a further test to account for the endogeneity intrinsically associated with the relationship between media usage and public opinion. Next, 3 Called either Fiscal Compact of Fiscal Stability Referendum. 3 we explore the micro-foundations of our argument analysing the Irish Fiscal Compact Referendum survey. We conclude by examining the implications of our findings. Information, knowledge and the World Wide Web Information is the data that allow individuals to acquire (politically) relevant knowledge and to form or redefine their beliefs. It can be gathered from direct experience – attending a candidate debate, correspondence with an MP – or by being exposed to reports. These reports can be based on the experience of people in one’s social network – e.g. a friend telling of his/her experience in dealing with local government – or can be provided by mass media. Acquiring information through direct experience of EU institutions and policies can be substantially ruled out. The same applies to information obtained via personal networks as ‘very few citizens have first- or even second-hand contact with Community affairs in Brussels’ (Dalton and Duval 1986, 186). Political knowledge is the state of awareness of facts that matter to orient people’s opinions and choices. The acquisition of knowledge depends upon the availability of information, and so does the redefinition of what Bartels (1993) calls the ‘fund of knowledge’. Newly acquired information allows individuals to update their pre-existing knowledge – for those who had some – and provides new knowledge to those who had none before. The mass media are therefore key to educating the public by making the information available and accessible (Holtz-Bacha and Norris 2001). Information acquired via the mass media can then translate into in-depth knowledge, superficial acquaintance, or even no knowledge as a result of it being, unclear, noisy or redundant. A substantial corpus of studies has unveiled differences among traditional media in their capacity to translate information into knowledge that subsequently affects political evaluations (for a review, see (Norris and Sanders 2003)). While several studies support the idea of the superiority of print (Robinson and Davis 1990; Robinson et al. 1986), others 4 cast doubts on it (Graber 2001; Mondak 1995), with no ultimate consensus on which medium carries the highest learning potential for the public. Despite the different conclusions reached by these studies, the literature on traditional media relies on a widely shared – if implicit – assumption: the information provided is both factual and relevant. Media publishers and regulatory authorities act as gatekeepers to what can be broadcast or printed, and media editors, through the process of indexing (Bennett 1990), select what is relevant for the public to know (Flanagin and Metzger, 2000). The multiplicity and segmentation of information producers on the World Wide Web (Bimber 2003) cast some doubt on the applicability of this assumption to the Internet. One of McLuhan’s (1964) insights in formulating the well-known the medium is the message theory was that “the content of any medium is always another medium”. The Internet essentially contains all the other mass media that preceded it: the printed press, TV, and radio. All major news outlets now have an online version. Additionally, there are many other online spaces that deliver politically relevant information: the official websites of institutions, social media sites – including social networking sites like Facebook and Twitter as well as sharing platforms like wikis and YouTube – and blogs. Online information can therefore take the most diverse shapes, depending on the source and hosting platform: from credible news producers to unverifiable information posted by individuals simply voicing their own opinions. In the words of Patterson, “the internet is at once a gold mine of solid content and a hellhole of misinformation” (2013: 79). These structural differences with respect to traditional media are likely to affect the opinions and behaviours of those who have integrated online news consumption into their media usage habits. The nature of online-based information may also affect the process through which individuals translate information into political knowledge. Lupia and McCubbins (1999, 25) point out a key element of the relationship between information and knowledge: ‘although you cannot have knowledge without having 5 information, you can have information without having knowledge’. Information that does not provide knowledge is either redundant or is simply noise. While absorbing redundant information is a possibility when gathering information via any media, the Internet maximizes the potential for the amplification of noise by creating an environment where noise and signal are often indistinguishable (Ayres 1999). Online spaces vary dramatically in the heterogeneity of their content and in the extent to which they host unverified versus verified facts. For instance, the website of a major newspaper is likely to publish content that is verified and has been provided by accredited sources. At the other end of the spectrum, platforms like blogs and forums are, by definition, aggregators of comments and opinions, often seeking to advocate more than to report (Scott 2007). Experimental research has shown that websites credibility is best understood as a perceptual attribute, but working with survey data rules out respondents’ self-assessment (for a review of online perceived credibility research see: Flanagin and Matzger 2007). Therefore, we opt for a classification of sources that both builds on what shown by these studies and accounts for the fact that platforms that host information created by a more heterogeneous set of producers and of mixed nature are more likely to carry greater amounts of noise. In Table 1 below we classify hosting platforms in terms of the type of information they provide and the information producers. [Table 1 about here] We thus empirically analyse separately the effects of platforms that have a strong potential for the amplification of noise from those that are most likely to maximize fact-checked information, in order to gather a more finely-grained understanding of how the new media affect individuals’ opinions. Formally, we test whether the following hypotheses hold true: H1: Online platforms providing vetted information increase knowledge of the EU and significantly 6 impact upon citizens’ opinions about it. Conversely: H2: Online platforms providing miscellaneous information do not increase knowledge of the EU and fail exercising an impact upon citizens’ opinions about it. Given the large amount of variation across platforms, we expect the mechanism to work through the knowledge channel only for those platforms that are more likely to facilitate users in distinguishing the news from the noise. Knowledge is a well-established predictor of attitudes; therefore increases in knowledge should filter the effects of information acquisition, particularly so in the wake of what we know about widespread ignorance of European citizens on EU level matters and institutions. Media effects are more consequential where previous opinions are weak, and the second order theory (Reif and Schmitt 1980) clearly indicates that attitudes towards the EU are weaker than attitudes towards national government, making media effects particularly well placed to impact upon citizens’ opinions. Empirical strategy To test our hypotheses we perform three empirical analyses. First, as a result of our theoretical framework, we implement a causal mediation analysis model. This allows us to distinguish the direct effect of online information consumption from its mediated effect through increases in knowledge of EU matters. Causal mediation analysis is the core test of our argument. Second, we implement instrumental variables to mitigate concerns on reverse causality. We note that while instrumental variables are a useful tool to address the endogeneity problem, they are substantively different from causal median analysis. In particular, instrumental variables do not allow us to disentangle the mediated effect of online information for its direct effect on attitudes toward the EU. Third, we validate the 7 results of our cross-country estimates with original micro-level data on the Irish Fiscal Compact Referendum. Let’s start describing the research design of the causal median analysis. Causal mediation analysis allows exploration of the role of an intermediate variable that lies along the causal paths between the treatment and the dependent variable (Hicks and Tingley 2011; Imai, Keele, and Yamamoto 2010). In our case, knowledge about the EU is the mediator that lies along the causal path between information gathered online and attitudes towards the EU. The path mode is illustrated in Figure 1.4 [Figure 1 about here] To carry out the causal mediation analysis, we rely on the STATA 13 ‘mediation’ package developed by Hicks and Tingley (2011). The model for both the mediator variable and the outcome variable is an ordinary least squares (OLS) regression. We include country fixed effects to account for cross-country heterogeneity and mitigate omitted variable problems. We use robust standard errors and we run 1000 simulations for the quasi-Bayesian approximation of parameter uncertainty. We are unable to run an ordered logit (or probit) – which would be more appropriate given the outcome variables – as this is not supported by the ‘mediation’ package. However, we run simple ordered probit models – reported in the online appendix (Table A5) - the results of which are very similar to those from the OLS regressions. This makes us confident of the reliability of the estimations when we implement the causal mediation analysis. Data We use data from Eurobarometer 76.3, which contains a large battery of items on media use for political information on EU matters. We exclude observations from countries that The literature on causal mediation analysis is large and fast-growing. For pioneering statistical studies, see Baron and Kenny (1986). 4 8 were not members of the EU at the time of the data collection (November 2011). Outcome variable The outcome variable captures respondents’ attitudes towards the EU. The question on which we build our dependent variable was posed as follows: “In general, does the EU conjure up for you a very positive, fairly positive, neutral, fairly negative or very negative image?”5 The resulting variable ranges from 0 to 4 where 0 indicates ‘very negative’ and 4 ‘very positive’. Treatment variables We firstly set apart those who gathered most of their information on political matters about the EU on the word wide web.6 To capture different types of information sources online we rely on the following dummy variables: A variable that scores 1 if respondents exclusively use ‘institutional and official websites (government websites, etc.)’ to gather information on European matters, and 0 otherwise; A variable that scores 1 if respondents exclusively use ‘information websites (websites from newspapers, news magazines, etc.)’ to gather information on European matters, and 0 otherwise. We label this variable traditional media websites; A variable that scores 1 if respondents exclusively use ‘online social networks’ to gather information on European matters, and 0 otherwise; A variable that scores 1 if respondents exclusively use ‘blogs’ to gather information on European matters, and 0 otherwise; 5 This is one of the customary questions to measure attitudes towards the EU, as it is included in the Eurobarometer Standard Trend Questions. 6 Verbatim of QD4: “Where do you get most of your news on national political matters” (a) Television (b) The press (c) Radio (d) The internet (e) Other (f) You do not look for news on national political matters. 9 A variable that scores 1 if respondents exclusively use ‘video hosting websites’ to gather information on European matters, and 0 otherwise. These variables are generated on the basis of responses to the following question: “On the Internet, which of the following websites do you use to get news on European political matters? “ (a) Institutional and official websites (governmental websites, etc.) (b) Information websites (websites from newspapers, news magazines, etc.) (c) Online Social Networks (d) Blogs (e) Video hosting websites (f) Other. Figure 2 shows the distribution of the country averages of our five treatments. The takehome message from this figure is that at the time of the fieldwork the websites of traditional media outlets were, by far, the most visited platforms in every country. The second most preferred platforms for gathering information on EU matters were institutional and official websites, while the use of blogs, online social networks and video hosting websites was very limited across the entire sample. As one may reasonably expect the percentage of those who gather most of their information about the EU on Online Social Networks and Blogs is quite small (below 5% for both types) and minuscule with regard to Video hosting websites. [Figure 2 about here] Given this structural characteristic of the data, we aggregate the five dummies in two treatments. On the one hand, we create a variable that scores 1 if respondents exclusively use ‘institutional and official websites’ or ‘traditional media websites’. We label this fact-checked websites. On the other hand, we create a variable that scores 1 if respondents exclusively use ‘online social networks’ or ‘blogs’ or ‘video hosting websites’. We label this miscellaneous websites. By implementing this additional measure we minimize concerns about the results being driven by few observations. 10 Mediator For the mediating variable we use an index of objective political knowledge about the EU compiled by the Eurobarometer team and available in the dataset. It ranges from 0 to 2 (bad, average, good) and it is based on the answer to the following standard question: “For each of the following statements about the EU could you please tell me whether you think it is true or false: The EU currently consists of 27 Member States The members of the European Parliament are directly elected by the citizens of each Member State Switzerland is a member of the EU” This index is built on items that capture the knowledge of what Barabas et al (2014) classify as static general facts. In the Appendix (Table A6 and A7) we re-run our main models using a variable that captures knowledge of static policy facts. This variable is based on a question asking respondents whether the “EU budget is jointly determined by the European Parliament and the Member states”. Finally, as detailed below, in the Fiscal Compact Referendum Section we address the knowledge of surveillance policy facts. In line with the estimation strategy implemented for the outcome variable we run an OLS with country fixed effects and robust standard errors. We show the consistency of the results from a simple ordered probit in the online Appendix (Table A5). Control Variables We control for the socio-economic status of the respondents (gender, age, working status, education) and other factors that are likely to affect both media usage choice and political opinions regarding the EU. Namely, we include controls for how frequently respondents 11 discuss both national and European political matters with friends and relatives (frequently, occasionally, never). We then control for consumption of traditional media by including controls for information consumption about EU matters via TV, printed press and radio. Moreover, we include standard indicators (Gabel and Palmer 1995; Van Klingeren, Boomgaarden, and De Vreese 2013) of subjective perception of the national and European economies, as citizens tend to display unfavourable attitudes towards the EU if they believe the economy – both national and European – is performing poorly (Kritzinger 2003) and a measure of the respondents’ attitudes to immigrants. Previous studies indicate that attitudes to immigration are strongly correlated with support for the EU. Those who are negatively disposed towards out-groups tend to oppose the free circulation of citizens and the opening of national borders within Europe (De Vreese and Boomgaarden 2005). We report descriptive statistics about all the variables in use in the Appendix (Table A1), where we also provide a link to the EB 76.3 codebook and the Fiscal Compact Referendum’s questionnaire. Empirical results Table 2 below reports the results in relation to each dummy capturing a specific type of platform. Specifically, in each model one of the types of platform is included as the treatment, allowing us to estimate the average causal mediation effect (ACME) and direct effect for each type of platform. Since the results of the mediation equation are the same across the five models, we report them only once. For each and every model, we include dummies for traditional media usage in gathering information on the EU. The results are clear-cut: only certain types of online loci – those ensuring higher levels of fact-checked information and produced by a more homogeneous array of producers – have a mediated positive effect on knowledge as well as a direct positive effect on attitudes to the EU. Specifically, traditional media websites and especially institutional and official websites are positive 12 and statistically significant (Models 2-3), whereas blogs, online social networks, and video hosting websites (Models 4-6) have neither a statistically significant impact through knowledge (mediated effect) nor a direct effect on attitudes to the EU (Models 4-6).7 Importantly, ACME is positive and statistically significant only in models 2 and 3. The average causal mediation effect accounts respectively for 12 and 11 percent of the total effect. [Table 2 about here] Note that the coefficients of our dummies are the same across models 2-6. We effectively estimate the same model five times and include each dummy as a treatment to obtain the ACME for each type of platform. The fact that all the coefficients are the same confirms the robustness of our results. Our findings remain similar to those presented above when using the more aggregate treatment to capture fact-checked websites in Table 3. The results are positive and statistically significant with regard to fact-checked websites, whereas the coefficient for miscellaneous websites is never statistically significant. As expected, ACME is positive and statistically significant in model 8, but not in model 9. Here, the average causal mediation effect accounts for 12 percent of the total effect.8 [Table 3 about here] Only fact-checked websites increase the respondents’ knowledge of the EU and, in turn, knowledge has a positive mediated effect on attitudes to it. Fact-checked websites have a positive direct effect on attitudes to the EU, in addition to the effects through the In the next section we show that these findings are not affected by reverse causality and/or selection into the treatment. 8 Furthermore, the control variables have the expected coefficients, adding plausibility to our analysis (see the online Appendix). Importantly, the results of our dummies capturing specific types of platforms have an effect in addition to the role played by traditional media. 7 13 knowledge channel. On the contrary, websites providing a large space for less reliable information and personal opinions affect neither the knowledge about the EU nor attitudes to it. In sum, we find strong support for both our two hypotheses. Finally, we perform a series of robustness checks to further validate our findings. First, we use the alternative variable for political knowledge described above (Tables A6-A7). Second, we re-run our main analysis using multilevel regressions to account for crosscountry heterogeneity (Table A8). Third, to balance out differences between those who used the five aforementioned website types and those who did not use the internet as a medium for information gathering on EU matters, we use entropy balancing (Hainmueller 2012).9 Specifically, we balance our entire set of covariates with respect to our treatments, which captures those respondents who go online as well as the type of website they visit.10 Differences in the means between the treated and control group vanish for all the covariates after implementing entropy balancing (pre- and post-matching descriptive statistics for the relevant variables are reported in the online Appendix). Note that balancing covariates with respect to the treatment is similar to controlling for such confounding factors in a standard multivariate regression without imposing parametric functional form or distributional assumptions. We then run all our parametric models using the weights obtained from the entropy balance estimation as well as the entire set of control variables to account for any residual differences between the treated and control groups (Table A9).11 An advantage of entropy balancing over a matching technique is that the former does not drop unmatched observations (Hainmueller 2012: 2). For all these tests, which are reported in the Appendix, the results are very similar to the ones reported above, adding plausibility to our empirical analysis. We use the STATA 13 package ‘ebalance’ (Hainmueller, 2012). A description of the covariates is provided in the following section. 11 Residual differences might come from the variance and skewness of the covariates. 9 10 14 Addressing endogeneity The previous section has shown that consuming information online – if only via certain platforms – has a positive effect on both knowledge about the EU and attitudes to it. One may object that the selection into our treatments is not random, as it would be plausible to argue that respondents who are knowledgeable about the EU are also more likely to go online in search of information about it. Similarly, it could be claimed that respondents who have a positive (negative) attitude to the EU are more likely to visit pro-EU (anti-EU) websites, which in turn reinforce their priors. Although experimental research (Garrett, Carnahan, and Lynch 2013) has shown how in online environments browsing for attitude consonant spaces does not prevent browsing for attitude dissonant information, concerns about self-selection are sound. To mitigate them, we implement an instrumental variable approach for those treatments that were statistically significant in the mediation analysis. Our instrument is a dummy that scores one if respondents have access to the Internet either at home or at work (labelled access). This variable thus takes value 0 when respondents said they had access neither at work nor at home. The idea behind this identification strategy is that without access to the Internet at home or at work respondents are less likely to go online to gather information about the EU. Indeed, the correlation between online access and information and official websites is respectively 0.68 (p-value=0.00) and 0.54 (p-value=0.00). Clearly, respondents without online access at home or at work can still browse the web through their smartphones/tablets, and this explains why we do not find a perfect correlation between our instrument and our treatments.12 The problem of relying on access is that it could be correlated with our dependent variables, i.e. knowledge about and attitudes to the EU. For instance, if respondents with Internet access are geographically clustered, our instrument could be correlated with socioWe note that if respondents who have access to internet through smartphones/tablets are more likely to visit fact-checking website and to have a positive attitude towards the EU, our IV estimates would be downward biased. 12 15 economic variables (i.e. urban or rural place of residence), which consequently affect our outcome variables. Since we control for socio-economic status, the exclusion restriction should still hold. Moreover, acknowledging that our instrument is not randomly assigned, we balance our entire set of covariates with respect to access in order to create a counterfactual (i.e. respondents without online access), which is as close as possible as to our treated units, i.e. respondents with online access. In the online Appendix (Table A4) we show the distribution of the covariates before and after balancing. Then we use the weights obtained from entropy balancing in our instrumental variable estimations. Armed with this identification strategy, we re-run our main models with both knowledge about the EU and attitudes to it as dependent variables. We begin with knowledge about the EU. The results are reported in Tables 4a and 4b and are in line with those shown in the previous section. Specifically, institutional and official websites and traditional media websites have a positive and statistically significant impact on knowledge about the EU. Furthermore, all the diagnostics show that our instrumental approach is robust. In particular, the orthogonality conditions are valid (see the Anderson-Rubin Wald test) and our instrument is sufficiently strong (see the Cragg-Donald Wald F statistics). Moreover, Table 4b shows the results for attitudes to the EU and indicates that our previous results are all validated. Indeed, institutional and official websites and traditional media websites always strengthen positive opinions about the EU. Even in this case, all the diagnostics confirm the validity of our identification strategy. [Tables 4a and 4b about here] The Case of the Irish Fiscal Compact Referendum The analysis presented above has the advantage of relying on a large comparative dataset, but it has two shortcomings. First, the Eurobarometer survey does not contain any 16 question accounting for political knowledge of surveillance facts. If we are able to explore knowledge of both general facts (in the main text) and of specific facts (in the Appendix), we are still limited in our capacity to cover the temporal dimension of knowledge (Barabas et al 2014). Second, we have no control over the content actually accessed online by respondents. This latter issue is a factor of both the ephemerality of the medium under scrutiny here and the limits of survey based data, particularly when secondary. In order to account for these limitations, we integrate our analysis with a quantitative case study - vote on the Fiscal Compact Treaty Referendum in Ireland - that is well placed to address these concerns by (a) using a measure of policy specific surveillance fact and (b) pinning down the effect of one particular official website whose content is known. In doing so, we use data from an original phone survey performed by the polling company RED C (details are provided below). A constitutional provision in the Republic of Ireland requires a popular vote on matters that entail amending the constitution, as ratification of EU treaties does. The issue at stake on May the 31st 2012 was whether to ratify the European Fiscal Compact Treaty, one of the key policy measures the European Union adopted in response to the sovereign debt crisis. We test our main hypothesis about the effects of online information consumption by examining the effects of a specific website on the knowledge of a single policy issue, to then assess its impact on voting decision. This behavioural complement to our study of attitudes implies that voting in favour of the Fiscal Compact Treaty carries an implicit positive evaluation of the EU and its policy measures. The Irish Referendum Commission dedicated a website to the Fiscal Compact Treaty (www.http://www.refcom.ie/) featuring a youtube video inviting citizens to vote, providing links to additional audio-visual material with extensive information on the Fiscal Compact pact, and a link to the guide designed by the Referendum Commission itself to 17 provide factual and balanced information to voters.13 The Fiscal Compact Treaty website offered relevant facts and guidance on the referendum procedures. In line with the previous analysis, we have two outcome variables. The first dependent variable captures respondents’ knowledge of the treaty. Specifically, respondents were asked to self-assess their knowledge of the Treaty: How would you describe your knowledge of the Fiscal Treaty? (a) I don’t know anything about it at all (b) I don’t know very much about it (c) I know quite a lot about it. (d) I’m extremely well informed about it. The second dependent variable is a dummy that scores one if respondents vote ‘yes’ to the EU Fiscal Compact Treaty.14 The question on the vote was put very straightforwardly as follows: “Did you vote YES in favour or NO against in the Fiscal Stability Treaty referendum?”. The treatment in this case scores one if a respondent used the official Referendum Commission Fiscal Stability Treaty website regularly during the campaign. Given the low number of observations, we are unable to implement causal mediation analysis.15 Therefore, we run two separate estimations with our two outcome variables: (1) an ordered probit for the knowledge of the Fiscal compact; (2) a probit for the probability of voting “yes”. In line with our previous analysis we include entropy balancing weights.16 With these data it is particularly important to balance all the confounding factors with respect to respondents who visit only the EU Commission website, which is our main explanatory variable. Indeed, solo EU Commission website visitors are very different with Now: http://www.refcom.ie/en/Past-Referendums/Fiscal-Stability-Treaty/ We drop those respondents who refused to answer this question. As a result, we are left with 949 observations. 15 The number of simulations has to always be lower than the number of observations in causal mediation analysis. A low number of simulations creates problems with the convergence of the estimation and with the reliability of the ACME. 16 Table A10 in the Appendix provides details of all the variables in use before and after matching. Moreover, we report the verbatim of the questions from the survey included in the analysis. 13 14 18 respect to variables influencing voting behavior from respondents who also visit other websites to gather information on the Fiscal Compact. Furthermore, the control variables we include in the model – on top of standard socioeconomics – account for three sets of predictors of vote choice at EU level referendums. The first type relates to the so-called ‘attitudes schools’, which contends that citizens underlying attitudes towards the EU explain their vote choice at EU level referenda (Franklin, Van der Eijk, and Marsh 1995). The second set accounts for evaluations of the national level government, capturing the ‘second order school’ contention that just like at EP elections citizens vote in EU referendums by expressing an evaluation of their national government (Reif and Schmitt 1980). Finally, we account for the so-called ‘utilitarian’ school (Gabel 1998) by controlling for the economic consequences citizens attribute to the referendum’s outcome. (Table 5 about here) Results from Table 5 indicate that visiting the Referendum Commission website increases self-perceived knowledge about the treaty. Moreover, it shows that those who regularly visited the website of the Referendum Commission are more likely to vote ‘yes’ than those who did not visit this website, ceteris paribus. Therefore, online information gathering on this fact-checked official website leads to a higher likelihood of accepting the Fiscal Compact. Knowledge about the treaty has a positive effect on the probability of voting “yes”, though it is (barely) not statistically significant (p-value=0.17). This result might be explained by the relative low number of observations and by the limited variation of the outcome variable, i.e. large percentage of “yes” vote. This finding confirms what previously seen when using the Europe wide sample and increases the validity of the pattern identified there. Moreover, it shows that institutional official platforms increase surveillance policy specific knowledge of the issue at stake and this, in turn, affects the probability of 19 approving the Treaty. Conclusions The news media serve the purpose of supplying the information that gives citizens awareness of institutional responsibilities, the positions and policies of political elites, and ultimately offer them guidance when making politically relevant choices. Traditional media dedicate relatively little attention to European matters, particularly in routine periods (Peter and De Vreese 2004). While the emergence of a European public sphere is highly desirable, we have not yet seen it happening. A global transnational multilingual medium like the Internet carries the potential to supply all the information Europeans need and often do not find on traditional media. On the other hand, the Internet also contains large amounts of noise. Our study has addressed the effects of online news consumption on public opinion by digging deeper into the variety of platforms that can supply such information on the World Wide Web. As the Internet varies enormously in terms of the types of websites hosting information, we have set apart loci that provide fact-checked information from those containing a mixture of opinions, non-fact-checked information and reliable facts. We also explored the effects of a particular fact-checked website on knowledge and voting behaviour, in relation to the specific issue of the Fiscal Stability Treaty. Our findings indicate that only certain online spaces exert an effect on public attitudes about the EU. Those who search for information on European matters on the websites of traditional media or on the websites of institutions are more positive towards the EU, and this happens via increases in levels of knowledge about the Union. Those who perform a similar search in spaces that contain miscellaneous information do not experience changes in attitudes and do not appear to learn from them. This may be due to the difficulties that users experience in separating the signal from the noise when the two are simultaneously 20 consumed. The analysis of the Irish Fiscal Compact Treaty, not only provides a confirmation that fact-checked platforms tend to increase knowledge, but also shows their effects on actual electoral behaviour. The direction of the effects in the case of attitudes as well as in the case of behaviour is positive, signalling that when citizens access information on fact-checked websites their knowledge and approval of the EU experiences an increase. The effect is robust to controlling for use of traditional media; therefore the Internet is effective independently from and in addition to them. Here we have shown that environments prone to containing a mixture of accredited information and fragmented opinions do not enlighten preferences via increased awareness and they do not impact attitudes. However, this study is limited in its capacity to reveal nuances in the process of learning by accessing online sources, as observational data per se are not completely suitable to such an end. Insights on how different platforms affect cognition and learning can be gathered by means of laboratory or filed experiments. Finally, this kind of data does not account for the effects of content or for the way content is framed. This essay expands the debate on effects of digital sources on both politically relevant knowledge and opinions, but does not test the effects of specific content, which is certainly part of the equation and could also be best clarified by experimental studies. References Ayres, Jeffrey M. 1999. From the streets to the Internet: The cyber-diffusion of contention. The Annals of the American Academy of Political and Social Science 566 (1): 132-143. Barabas, Jason, Jennifer Jerit, William Pollock, and Carlisle Rainey. (2014). "The Question(s) of Political Knowledge." American Political Science Review, 108:(4). Baron, Reuben M, and David A Kenny. 1986. The moderator–mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. Journal of personality and social psychology 51 (6): 1173. Bartels, Larry M. 1993. Messages received: The political impact of media exposure. American Political Science Review: 267-285. Bennett, W Lance. 1990. Toward a theory of press‐state relations in the United States. Journal of Communication 40 (2): 103-127. Bennett, W Lance, and Shanto Iyengar. 2008. A new era of minimal effects? The changing foundations of political communication. Journal of Communication 58 (4): 707-731. 21 Bimber, Bruce Allen. 2003. Information and American democracy: Technology in the evolution of political power. Cambridge University Press. Castells, Manuel. 2000. Materials for an exploratory theory of the network society1. The British journal of sociology 51 (1): 5-24. Dalton, Russell J, and Robert Duval. 1986. The political environment and foreign policy opinions: British attitudes toward European integration, 1972–1979. British Journal of Political Science 16 (01): 113-134. De Vreese, Claes. 2003. Framing Europe: television news and European integration. Aksant Academic Pub. De Vreese, Claes H, and Hajo G Boomgaarden. 2005. Projecting EU Referendums Fear of Immigration and Support for European Integration. European Union Politics 6 (1): 59-82. ———. 2006. Media Effects on Public Opinion about the Enlargement of the European Union*. JCMS: Journal of Common Market Studies 44 (2): 419-436. Delli Carpini, Michael X , and Scott Keeter. 1996. What Americans know about politics and why it matters. Yale UniversityPress, New Haven, CT. Franklin, Mark N, Cees Van der Eijk, and Michael Marsh. 1995. Referendum outcomes and trust in government: Public support for Europe in the wake of Maastricht. West European Politics 18 (3): 101-117. Flanagin, A. J., & Metzger, M. J. (2000). Perceptions of Internet information credibility. Journalism & Mass Communication Quarterly, 77(3), 515-540. Flanagin, A. J., & Metzger, M. J. (2007). The role of site features, user attributes, and information verification behaviors on the perceived credibility of web-based information. New Media & Society, 9(2), 319-342. Gabel, Matthew. 1998. Public support for European integration: An empirical test of five theories. Journal of Politics 60: 333-354. Gabel, Matthew, and Harvey D Palmer. 1995. Understanding variation in public support for European integration. European Journal of Political Research 27 (1): 3-19. Garrett, R Kelly, Dustin Carnahan, and Emily K Lynch. 2013. A turn toward avoidance? Selective exposure to online political information, 2004–2008. Political Behavior 35 (1): 113-134. Graber, Doris A. 2001. Processing politics: Learning from television in the Internet age. University of Chicago Press. Hainmueller, Jens. 2012. Entropy balancing for causal effects: A multivariate reweighting method to produce balanced samples in observational studies. Political Analysis 20 (1): 25-46. Hicks, Raymond, and Dustin Tingley. 2011. Causal mediation analysis. Stata Journal 11 (4): 605. Holtz-Bacha, Christina , and Pippa Norris. 2001. " To Entertain, Inform, and Educate": Still the Role of Public Television. Political communication 18 (2): 123-140. Imai, Kosuke, Luke Keele, and Teppei Yamamoto. 2010. Identification, inference and sensitivity analysis for causal mediation effects. Statistical Science 25 (1): 51-71. Kritzinger, Sylvia. 2003. The influence of the nation-state on individual support for the European Union. European Union Politics 4 (2): 219-241. McLuhan, Marshall. 1964. The medium is the message. Media and cultural studies: keyworks: 129-38. Metzger, M. J. (2007). Making sense of credibility on the Web: Models for evaluating online information and recommendations for future research. Journal of the American Society for Information Science and Technology, 58(13), 2078-2091. 22 Mondak, Jeffery J. 1995. Newspapers and political awareness. American Journal of Political Science: 513-527. Newton, Kenneth. 1999. Mass media effects: mobilization or media malaise? British Journal of Political Science 29 (04): 577-599. Norris, Pippa, and David Sanders. 2003. Message or medium? Campaign learning during the 2001 British general election. Political communication 20 (3): 233-262. Peter, Jochen, and Claes H De Vreese. 2004. In Search of Europe A Cross-National Comparative Study of the European Union in National Television News. The Harvard International Journal of Press/Politics 9 (4): 3-24. Reif, K., and H. Schmitt. 1980. Nine Second-Order National Elections-A Conceptual Framework for the Analysis of European Election Results. European Journal of Political Research 8 (1): 3-44. Robinson, John P, and Dennis K Davis. 1990. Television news and the informed public: An information‐processing approach. Journal of Communication 40 (3): 106-119. Robinson, John P, Mark R Levy, Dennis K Davis, William Gill Woodall, Michael Gurevitch, and Haluk Sahin. 1986. The main source: Learning from television news. Sage Publications Beverly Hills, CA. Schuck, Andreas RT, and Claes H De Vreese. 2006. Between Risk and Opportunity News Framing and its Effects on Public Support for EU Enlargement. European Journal of Communication 21 (1): 5-32. Scott, D Travers. 2007. Pundits in muckrakers’ clothing: political blogs and the 2004 US presidential election. Blogging, citizenship, and the future of media: 39-58. Van Klingeren, Marijn, Hajo G Boomgaarden, and Claes H De Vreese. 2013. Going Soft or Staying Soft: Have Identity Factors Become More Important Than Economic Rationale when Explaining Euroscepticism? Journal of European Integration 35 (6): 689-704. Zaller, John. 1992. The nature and origins of mass opinion. Cambridge university press. 23 Figure 1. Path analysis testing the mediation role of Knowledge about the EU. Public Opinion Towards the EU Mediated information Knowledge 24 Figure 2. Distribution of online platforms by country averages. 25 Table 1. Classification of hosting platforms Traditional Media Websites and Official Websites of Institutions Social Networks, Blogs, and Video Sharing Platforms Type of information Vetted Miscellaneous Set of producers Homogenous Heterogeneous 26 Table 2. Causal mediation analysis: OLS with country fixed effects (type of platform). Effects of different platforms on Knowledge and Attitudes Traditional Media Websites Online Social Networks Blogs Video Hosting Websites Attitudes (3) Attitudes (4) Attitudes (5) Attitudes (6) 1.373*** (0.031) 0.076** (0.036) 0.046*** (0.017) 0.030 (0.041) -0.064 (0.067) -0.102 (0.098) 0.100*** (0.009) 3.371*** (0.064) 0.076** (0.036) 0.046*** (0.017) 0.030 (0.041) -0.064 (0.067) -0.102 (0.098) 0.100*** (0.009) 3.371*** (0.064) 0.076** (0.036) 0.046*** (0.017) 0.030 (0.041) -0.064 (0.067) -0.102 (0.098) 0.100*** (0.009) 3.371*** (0.064) 0.076** (0.036) 0.046*** (0.017) 0.030 (0.041) -0.064 (0.067) -0.102 (0.098) 0.100*** (0.009) 3.371*** (0.064) 0.076** (0.036) 0.046*** (0.017) 0.030 (0.041) -0.064 (0.067) -0.102 (0.098) 0.100*** (0.009) 3.371*** (0.064) yes 0.011** 0.075** 0.086** 0.121** yes 0.006** 0.045** 0.051** 0.118** yes 0.003 0.028 0.032 0.062 yes 0.004 -0.066 -0.062 -0.044 yes -0.011 -0.105 -0.116 0.079 yes Knowledge (1) Institutional and Official websites Traditional media websites Online social networks Blogs Video Hosting Websites 0.107*** (0.024) 0.062*** (0.011) 0.034 (0.028) 0.040 (0.046) -0.112* (0.063) Political Knowledge Constant ACME Direct Effect Total Effect % of Total Effect Mediated Controls Institutional and Official Websites Attitudes (2) 27 Country fixed-effects yes yes yes yes yes yes Observations 26,038 26,038 26,038 26,038 26,038 26,038 R-squared 0.155 0.160 0.160 0.160 0.160 0.160 RMSE 0.834 0.831 0.831 0.831 0.831 0.831 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Note: Tables are reported in a short format, suppressing the outputs for control variables. Tables are available in the long format upon requests. 28 Table 3. Causal mediation analysis: OLS with country fixed effects (fact-checked vs. non-fact-checked websites). Effects of Fact-checked versus Miscellaneous Websites on Knowledge and Attitudes Fact-checked Websites Knowledge Attitudes (7) (8) Fact-checked Websites Miscellaneous Websites 0.069*** (0.011) 0.016 (0.023) 1.176*** (0.042) 0.051*** (0.016) -0.009 (0.033) 0.100*** (0.009) 3.370*** (0.064) 0.051*** (0.016) -0.009 (0.033) 0.100*** (0.009) 3.370*** (0.064) yes yes 26,038 0.155 0.834 0.007** 0.051** 0.058** 0.122** yes yes 26,038 0.155 0.834 0.002 -0.008 -0.006 -0.032 yes yes 26,038 0.155 0.834 Political Knowledge Constant ACME Direct Effect Total Effect % of Total Effect Mediated Controls Country fixed-effects Observations R-squared RMSE Miscellaneous Websites Attitudes (9) Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 29 Table 4a. Instrumenting type of platform: IVREG with country fixed effects (Knowledge about the EU) and entropy balance. Knowledge (11) (12) (10) Institutional and Official websites 2.622*** (1.002) Traditional media websites 0.551*** (0.205) Fact-checked Website 0.456*** (0.170) Miscellaneous Website Constant Online Access Anderson-Rubin Wald test Kleibergen-Paap rk LM statistic Cragg-Donald Wald F statistic Controls Country fixed-effects Observations RMSE (13) 1.162*** (0.111) 1.168*** (0.110) 2.337** (0.916) 1.210*** (0.116) First Stage 0.023*** 0.107*** (0.001) (0.008) 0.130*** (0.008) 0.024*** (0.002) 1.195*** (0.112) 14.34*** 228.74*** 210.69*** yes yes 26,398 0.609 14.34*** 14.34*** 395.02*** 478.16*** 1147.94*** 1437.28*** yes yes yes yes 26,398 26,398 0.555 0.552 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 30 14.34*** 69.13*** 195.00*** yes yes 26,398 0.614 Table 4b. Instrumenting type of platform: IVREG with country fixed effects (Attitudes to the EU) and entropy balance. (14) Institutional and Official websites Attitudes (15) (16) 4.226*** (1.603) Traditional media websites 0.882** (0.350) Fact-checked Website 0.730** (0.287) Miscellaneous Website Political Knowledge Constant Online Access Anderson-Rubin Wald test Kleibergen-Paap rk LM statistic Cragg-Donald Wald F statistic Controls Country fixed-effects Observations RMSE (17) 0.100*** (0.034) 2.802*** (0.202) First Stage 0.022*** (0.001) 14.34*** 210.76*** 195.40*** yes yes 26,038 0.945 0.102*** (0.036) 2.745*** (0.211) 0.106*** (0.008) 0.128*** (0.008) 0.024*** (0.002) 14.34*** 364.66*** 1091.39*** yes yes 26,038 0.858 14.34*** 445.87*** 1362.79*** yes yes 26,038 0.851 14.34*** 70.10*** 190.77*** yes yes 26,038 0.946 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 31 0.101*** (0.035) 2.755*** (0.209) 3.706*** (1.430) 0.121*** (0.033) 2.799*** (0.213) Table 5. Fiscal Compact Referendum. Effects of regularly visiting the Referendum Commission Website on Knowledge of the Treaty and Vote Choice. VARIABLES EU Commission Website (regularly) Orderd probit Knowledge of the Treaty Probit Pr(Voting Yes) 0.822*** (0.195) 0.851** (0.348) 0.206 (0.148) Knowledge of the Treaty cut1 -1.452*** (0.542) 0.458 (0.553) 2.417*** (0.583) cut2 cut3 Controls yes Town fixed-effects yes Entropy balancing yes Observations 949 Pseudo R-squared 0.06 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 32 yes yes yes 949