Can media drive the electorate? The impact of media coverage on voting intentions

Accepted Manuscript Can media drive the electorate? The impact of media coverage on voting intentions Ralf Dewenter, Melissa Linder, Tobias Thomas

PII:

S0176-2680(18)30167-8

DOI:

https://doi.org/10.1016/j.ejpoleco.2018.12.003

Reference:

POLECO 1764

To appear in:

European Journal of Political Economy

Received Date: 6 April 2018 Revised Date:

26 November 2018

Accepted Date: 21 December 2018

Please cite this article as: Dewenter, R., Linder, M., Thomas, T., Can media drive the electorate? The impact of media coverage on voting intentions, European Journal of Political Economy (2019), doi: https://doi.org/10.1016/j.ejpoleco.2018.12.003. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

Can Media Drive the Electorate?

Melissa Linder**

Abstract

Tobias Thomas***

SC

Ralf Dewenter*

RI PT

The Impact of Media Coverage on Voting Intentions♣

TE D

M AN U

A growing literature uses media data to explain perception and behaviour in the economic and political context. In this paper, we investigate how media coverage affects political preferences, namely voting intention. For our empirical analysis, we merge 14 years of human-coded data obtained from leading media in Germany with results of the comprehensive German Politbarometer survey from February 1998 through December 2012. In contrast to the existing literature, we do not utilize access to certain media outlets, but use the tonality of articles and newscasts on political parties and politicians based on human coded media data. To account for endogeneity, we employ instrumental variable probit estimations. In addition, we control for a multitude of (internal) personal characteristics, such as age, and gender, as well as for (external) macroeconomic variables, such as business climate, unemployment, and inflation. The results show that media coverage of a political party has a positive and significant effect on the voting intention for this party. When media outlets cover a political party more positively, the electorate has a greater tendency to vote for it. Hence, we conclude that the electoral success or failure of political parties is at least partially caused by the media coverage on them. This hints on the special responsibility of media in democracies. political preferences, voting intention, media impact

JEL classification:

C43, D 72

AC C

EP

Keywords:

♣

We are grateful to the editor, two anonymous referees, Adam Lederer, Hendrik Lüth as well as to the participants of the European Public Choice Society Meeting in April 2018 in Rome, Italy for valuable comments. *Helmut-Schmidt-University Hamburg, Department of Economics, Holstenhofweg 85, 22043 Hamburg (Germany), Email: [email protected] **Helmut-Schmidt-University Hamburg, Department of Economics, Holstenhofweg 85, 22043 Hamburg (Germany), Email: [email protected] ***EcoAustria – Institute for Economic Research (Vienna, Austria), Center for Media, Data and Society (CMDS) at the Central European University (Budapest, Hungary) and Düsseldorf Institute for Competition Economics (DICE), Email: [email protected]

1

ACCEPTED MANUSCRIPT

Introduction

A growing literature uses media data to explain perception and behaviour of individuals. In the political context the question arises whether media report more on political parties because of their success or if their success is caused at least partly by media reports. To address this question, we investigate whether the tonality of media coverage on political parties affects individuals’ voting intentions of which capture political preferences. These political preferences

RI PT

are likely to be influenced by external factors, such as media reporting. Moreover, as the majority of the voter receive their information on political parties and politicians through the media, we expect a strong impact of media reporting on the decision-making process and thereby on voting intentions of individuals. If media would provide comprehensive and unbiased information about political parties and politicians, this would be rather unproblematic.

SC

Media would play their role as information channel in democracies and voter would decide on the basis of this information. However, we know from a broard range of literature that this is not

M AN U

(always) the case.1

For our empirical analysis, we merge 14 years of human-coded data derived from leading German media with the results of the comprehensive German Politbarometer survey from February 1998 through December 2012. As media coverage may not only affect the political preferences of voters, but also be generally affected by the current political mood of the electorate, it is likely that our regressions would suffer from endogeneity. On the one hand, the

TE D

political preferences of media consumers could affect the coverage of the specific media outlets they consume. For instance, this would hold if a media outlet reacts to the moods of their recipients. In this case, the analysis is likely to suffer from endogeneity in terms of reverse causality. On the other hand, in the case that the media react to general political sentiments, this

EP

would result in endogeneity in terms of an omitted variable bias. To address these issues, we employ instrumental variable probit estimations. Moreover, we also control for a multitude of

AC C

(internal) personal characteristics, such as age, and gender, as well as (external) macroeconomic variables, such as business climate, unemployment, and inflation. Our contribution is structured as follows: Section 2 provides an overview of the related literature before the data are introduced in section 3. Section 4 describes our estimation strategy and presents the results. Finally, section 5 concludes.

2

Related Literature

Media play a vital role in the perception and decisions of individuals in both economic and political contexts, as information is often distributed through media channels. However, the media can never depict the complete reality and is limited to painting a partial picture. In 1

See the related literature below.

1

ACCEPTED MANUSCRIPT

addition, the portrayed reality is prone to various types of distortions, the so-called media bias (Entman 2007).2 Consequently, decisions by individuals based on information provided by the media might deviate from decisions based on less biased and more comprehensive information. For instance, Dewenter et al. (2016) find evidence that the number of car sales depends, to some extent, on media coverage of the automotive industry, Eisensee and Strömberg (2007) analyse the effects of media coverage of natural disasters on US disaster relief decisions, and Beckmann

RI PT

et al. (2017) find that media coverage of terror attacks causes further terroristic activities in terms of the number of incidents as well as the severity of terror acts.

In the economic context, for Nadeau et al. (2000), Soroka (2006), and van Raaij (1989) the assessment of the state of the economy and economic expectations depend, at least in parts, on media reports. Alsem et al. (2008), Goidel and Langley (1995), as well as Doms and Morin

SC

(2004) show the impact of media reporting on consumer climate. Garz (2013) analyses the impact of distorted media coverage on unemployment on job insecurity perceptions and Lamla

M AN U

and Maag (2012) investigate the role of media reporting for inflation forecasts of households and professional forecasters. Media coverage on economic crises such as the euro crisis can even have an effect on the life satisfaction of the people (Chadi, 2015). In the political context, Bernhardt et al (2008), D‘Alessio and Allen (2000), Druckman and Parkin (2005), Entman (2007), Gentzkow et al (2011), as well as Morris (2007) analyse the impact of media coverage on political attitudes and voter decisions. Gentzkow et al (2015) estimate the effect of party control

newspapers.

TE D

of state governments on the entry, exit, circulation, prices, number of pages, and content of daily

The effect of media consumption on voter information in democracies is investigated by Garz (2017) as well as Larcinese and Sircar (2017). However, Kauder and Potrafke (2015) show that

EP

even if a political scandal is known and covered by the media it does not necessarily lead to a punishment of the incumbent government by the voters.

AC C

A closer look at the impact of media coverage in the political context is provided by Snyder and Strömberg (2010). In their comprehensive analysis, the authors find that voters living in regions

2

Of the various types of media bias, the most prominent are: advertising bias, when media change their news coverage in tone or volume in favour of their advertising clients (see Dewenter & Heimeshoff, 2014, 2015; Gambaro & Puglisi 2015 or Reuter & Zitzewitz 2006); newsworthiness bias, when news on certain issues crowd out coverage on other issues because they are seen as more newsworthy (see Durante & Zhuravskaya, 2015 or Eisensee & Strömberg, 2007); the negativity bias, when media focus more on catastrophes, crime, and threatening political and economic developments in comparison to more positive news (see Friebel and Heinz, 2014; Garz, 2013, 2014; Heinz and Swinnen, 2015; or Soroka, 2006); the distance bias, when media report more on events which take place close to their media market (Berlemann and Thomas, 2018); and political bias, when media coverage favours one or another side of the political spectrum (see Anderson & McLaren, 2010; Besley & Prat, 2006; Gentzkow and Shapiro (2010), Groseclose and Milyo, 2005; Prat, 2014). In addition, there is a broad literature on the existence of media biases and their foundations in communication and media science. Examples include Ball-Rokeach (1985) as well as Ball-Rokeach and DeFleur (1976) on the dependencies of the media-system, and Dunham (2013) on the measurement of media biases.

2

ACCEPTED MANUSCRIPT

with insufficient political media coverage are less able to recall or evaluate their representatives. This also affects the work of the politicians: Less covered congressmen are less willing to serve as witness at congressional hearings or serve on committees. Finally, regions with less press coverage of representatives receive less federal spending. The impact of media coverage on electoral outcomes is the focus of Enikolopov et al. (2011). The authors analyse electoral outcomes of the 1999 parliamentary elections in Russian regions with

RI PT

differing access to an independent national TV channel. The authors find that access to independent TV led to a decreased vote for the governing party by 8.9 percentage points and to an increased vote for major opposition parties by 6.3 percentage points. The results are comparable to those of DellaVinga and Kaplan (2007), which looks at the rolling out of the conservative Fox News Channel across US states. The authors find that Republicans gained 0.4 to

SC

0.7 percentage points in presidential elections between 1996 and 2000 in the cities that had access to Fox News. Evidence of the importance of the media market with regard to both

M AN U

electoral participation and the voting decision is provided by Sobbrio and Navarra (2010). With its focus on the impact of media coverage on voting intentions, our contribution is connected to DellaVinga and Kaplan (2007) as well as Enikolopov et al. (2011). In contrast to the existing literature, we do not analyse the impact of access to certain broadcasting services, but the influence of the tonality of articles and newscasts on political parties and politicians based on human coded media data. This paper therefore addresses the fact that with existing studies of

TE D

media impact analysis, “measuring the tone of articles and editorials, is relatively underutilized” (Puglisi & Snyder, 2015, 664).

The Data

3.1

Data on Media Coverage

EP

3

The media data used in our study are based on media content analysis carried out by Media

AC C

Tenor International. The data are derived from text analyses of leading media outlets, conducted by human analysts and contain information on the media type (e.g. daily newspapers, magazines, television news), the topic (e.g. foreign affairs, unemployment, sports), participating persons and institutions (such as politicians and political parties), region of reference (e.g. Germany, EU, USA), time reference (e.g. past, present, future), and the source of information (journalist, expert etc.). In addition, the analysts evaluate the reports with respect to the sentiment toward persons or institutions. The use of human-coded data is advantageous since, as Puglisi & Snyder (2015, p. 656) put it: “compared to human-based coding, automated coding is less accurate in detecting the tone of

3

ACCEPTED MANUSCRIPT

each specific text analyzed”. While human coding achieves an accuracy of 0.85 at the minimum3, computer linguistic approaches, in contrast, still achieve accuracy no more than 0.60-0.70, especially when it comes to topical context as well as tonality (Grimmer and Steward, 2013). Consequently, the authors conclude that, for political text analysis, there is (at least so far) no adequate substitute for human coding. In our empirical analysis, we use the tonality of media coverage on political parties as our

RI PT

explanatory variable. As each report is coded as positive, negative, or neutral, an overall consideration of a media outlet’s tonality toward a specific party can simply be created by adding up the single evaluations of the reports. The tonality s of outlet i on a certain political


,, =



, − , , ,

SC

party j can then be defined as:



where , is the number of positive news in media outlet i in time t, , is the number of

M AN U

negative news, and , is the total number of positive, negative, and neutral news on a political

party j in outlet i in time t. The tonality
,, ∈ (-1,1) ranges from -1 (all news about j are negative) to 1 (all news about j are positive). For the empirical analysis in section 4, the reports from the different media outlets are aggregated over all outlets for each political party by building monthly means, resulting in an average tonality , for each party j at time t. Our media set consists of 35 different media outlets from Germany (3 private TV news shows, 4

TE D

public service TV news shows, 11 public service TV political magazines, 7 daily newspapers, 10 magazines). Each report was analysed news item by news item; that is, that each time that a new person, institution, topic, or source etc. appears, a further news item is coded. News items were analysed over the February 1998 to December 2012 period. Overall, 10,105,239 news items on

EP

political parties and/or politicians are included in the analysis. Although the reports are available as daily observations, we calculate the tonality of reports on a monthly basis, as the

AC C

data on political preferences are only available on a monthly basis. In the analysis we focus on the centre-right Christian Democratic Union/Christian Social Union (CDU/CSU), the centre-left Social Democratic Party (SPD), the liberal Free Democratic Party (FDP), as well as on the Greens (Gruene). Dropping all items that do not focus on these parties and their representatives results in a total of 9,451,032 news items on CDU/CSU, SPD, FDP and Gruene.4

3

The accuracy and reliability of the coding was regularly checked by Media Tenor, both with standard tests and random spot checks, based on the code-book. Each month, for each coder, three analysed reports were selected randomly and checked. Coders scoring lower than 0.80, that is 80 percent accuracy im comparison to the code-book, were removed from the coding process. In no month did the mean deviation among all coders exceeded 0.15. As a result, Media Tenor’s data achieve an accuracy of minimum 0.85. 4 See Table A1 in the Appendix. Note that the numbers of observations used in our estimations are considerably smaller because of aggregation.

4

3.2

ACCEPTED MANUSCRIPT

Data on Political Preferences

The data on political preferences and sentiments are taken from the Politbarometer surveys. Since 1977, Politbarometer surveys are performed at about monthly intervals by Forschungsgruppe Wahlen (Institute for Election Research). The aim of the Politbarometer is to poll the opinions and attitudes of eligible Germans regarding not just current events and issues but also political parties and politicians. A multi-stage random sample of the German residential

RI PT

population eligible to vote is selected. The data are collected by telephone interview (CATI) with standardized questionnaires. About 1,700 interviews are conducted each month. For the present analysis, the question regarding the voting intentions of the respondents are used to capture political preferences.

SC

Voting intention is the indicator for rather short-term or current political preferences. A general preference for one party does not necessarily mean that this person will also always vote for this party. External factors, such as media reporting or the performance of the government, are

M AN U

assumed to affect voting intention and behaviour. Moreover, many people base their decision on the actual situation which may be drawn by the media. The current empirical analysis investigates whether the tonality of media coverage on political parties impacts short-term political preferences (voting intentions).

However, as some respondents might have a general tendency towards one specific party due to family tradition or a general and long-lasting agreement with the content alignment of the party,

TE D

we control for these long-term preferences by including party affiliation into our regressions. For party affiliation the survey asks participants whether they have a general tendency for a specific party, whereas our dependent variable reflects the actual voting behaviour of the respondents as they are asked how they would vote if elections would lying ahead. As

EP

mentioned above, these two indicators – voting intention and party affiliation – must not always coincide due to external factors (like media reporting) which influence the decision-making

AC C

process of individuals.

To control for socio-demographic characteristics that are expected to affect voter preferences5, several variables from the Politbarometer survey are used: The age of respondents is divided into 10 categories (from 17 to 70 or older), education is scaled from 1 to 7 (no education to university degree), political interest is scaled from 1 to 5 (1 equals a very strong interest, 5 equals no interest), economic situation and own economic situation are scaled from 1 to 5 (very good to very bad) and reflect the assessment of the economic situation in Germany, in general, and of the respondent, in particular. Catholic, married, female, unemployed and labour union are

5

For an overview about the impact of socio-demographic characteristics and systemic factors on political preferences and voting behaviour see for instance Knoke and Hout (1974) as well as Kim et al (1975).

5

ACCEPTED MANUSCRIPT

all dummy variables that equal one if the person is catholic, married, female, unemployed or member in a labour union, and zero otherwise. Forschungsgruppe Wahlen randomly chooses participants for each survey. Therefore, the monthly Politbarometer survey is a repeated cross-section because each month a new random sample is taken from the population. The combined data set is a repeated cross-section, which encompasses media coverage on

RI PT

political parties, the individual characteristics of the respondents, the macroeconomic variables, and the variable of interest, voting intention, on a monthly basis from February 1998 through December 2012.

Empirical Strategy and Results

4.1

Identification Strategy

SC

4

M AN U

To identify the effects of media coverage on political preferences (voting intention), we first use simple probit regressions. Probit regressions are preferable as the linear probability model may lead to biased and inconsistent estimates (Horrace & Oaxaca, 2006). As the dependent variables is a binary variable (equals 1 when the individual intends to vote for a party), a probit model is the appropriate choice.

As discussed above, media coverage and political preferences may be co-determined, which can

TE D

lead to either an omitted variable bias or reverse causality. On the one hand, as media coverage might affect political preferences, consumers’ political preferences might also impact a media outlet’s coverage. This would hold if media outlets react to the moods of their consumers. In this case, the estimate would be biased due to reverse causality. On the other hand, the results could

EP

also suffer from endogeneity in terms of an omitted variable bias. This would be the case, for example, if media generally reacts to the political mood of the electorate. If political media coverage changes, either because of changing political moods or for other reasons, such as a

AC C

political crises, and this change in media reporting is not covered by the variables used, an omitted variables bias occurs.6 At least for the problem of reverse causality, the sign of the bias could be determined. As we expect both effects to be positive, the probit estimates might overestimate the positive impact of media coverage on voting intentions. For instance, a more pronounced intention to vote for the SPD and better poll results, can lead to more positive media coverage of this party. In case that

6

There might be an additional connection which could cause potential endogeneity problems: The media may affect how parties position themselves. This may, in turn, affect voting intentions of citizens as parties have moved towards (or away) from voters. From this perspective media might not only play a vital role in the perception and decision of voters but of the parties themselves as well. However, as the issue is only marginally discussed in the existing literature we leave this potential caveat to future research.

6

ACCEPTED MANUSCRIPT

the regressions also suffer from omitted variables, the sign of the bias would depend on the impact of the omitted variable. Therefore, the overall direction and size of the bias is unclear. In order to address possible endogeneity, an instrumental variable approach is applied. To find instrumental variables that affect media coverage but are, at the same time, not correlated with political preferences is no easy task. Measures of political performance such as macroeconomic variables could be well suited as these are assumed to be covered by the media. However,

RI PT

political successes and failures are, of course, also assumed to influence electoral decisions and possibly also political preferences. It is therefore necessary to find instrumental variables that meet the exogeneity assumption.

For this purpose, we use the tonality of international media reports on the banking industry to

SC

instrument domestic media coverage on political parties. The intuition behind this instrument is twofold: first, media coverage on the banking industry captures the general tonality of the international media as the banking industry is an important factor for the whole economy and

M AN U

its performance. By these means we are able to account for a general positive or negative tonality which is neither directly correlated with German politics nor with political preferences. While one could argue that there is some mistrust in the banking industry as a whole, there is also empirical evidence that at least in Germany this mistrust is rather disconnected from the opinions on political parties. Citizens seem to differentiate between a more abstract regulation of the banking sector, which takes often place on an international level, and the acting politicians

TE D

on the national level. In this line, a representative survey on the financial crises conducted by Institut für Demoskopie Allensbach (2013) shows that 68 percent of the respondents see the insufficient (international) regulation as a main reason for the financial crisis. In contrast, only

EP

11 percent see the politics of the German government as a main cause. Second, international media coverage can influence domestic journalists and thereby the media coverage on a national level. For instance, Benesch et al (2018) find out that significant media

AC C

spill overs from one country to another exist and therefore utilize this as an instrument to analyse the impact of media coverage on migration worries of the population. In our case the tonality of the international media coverage on the banking industry might influence German journalist. As a result, one would observe tonality spill-overs from the international coverage (including that on the banking industry) to the national coverage. Actually, this seems to be the case as our first stage regression in section 4.3 suggest. Furthermore, the international coverage should be independent from voters’ preferences in Germany: First, only a minor share of the German citizens consumes international media. Second, there is some empirical evidence, that the citizens seem to differentiate between a more abstract and often international regulation of the banking sector and the acting politicians on the national level: In this line, a representative survey on the financial crises conducted by Institut 7

ACCEPTED MANUSCRIPT

für Demoskopie Allensbach (Allensbach Institute, formally the Allensbach Institute for Public Opinion Research) (2013) shows that 68 percent of the respondents see the insufficient regulation as a main reason for the financial crisis. In contrast, only 11 percent see the politics of the German government as a main cause. Again we use media data by Media Tenor, generated as described in section 3.1 but comprising media reports on international banks collected from international media. This data set

RI PT

comprises media outlets from a multitude of countries, such as Austria, Canada, France, Germany, Italy, South Africa, Spain, Switzerland, United Kingdom and the United States. The data set includes in total 1,048,575 news items. Table A2 in the appendix lists the different media outlets sorted by country and shows the number of items for each medium. Like the media reports on political parties, the reports on international banks are as well coded either negative,

SC

positive or neutral. In accordance to the instrumented reports, we use monthly means over these reports in order to receive a measure for tonality which can be used to instrument media

M AN U

coverage of political parties. Accordingly, our tonality measure reflects the average tonality of media with respect to the international banking sector. A rather negative reporting on the banking industry is reflected by a more negative tonality of the media reports. Due to data availability, our sample is reduced to the period from January 2002 to December 2012. To be able to implement the instrumental variable approach in section 4.3, two main requirements must be fulfilled by the instruments. The instrument must be strongly correlated

TE D

with the endogenous explanatory variables conditional on the other covariates (instruments are relevant) and the instrument is not correlated with the error term in the explanatory equation conditional on the other covariates (exclusion restriction). While the former can be verified by several statistical tests and the results of the first stage estimations of the two-step approach

EP

(which are presented in section 4.3) the latter cannot be tested and must therefore be justified intuitively (see above).

4.2

AC C

The instrumental variable estimation is further described in section 4.3.

Probit Estimation

Our variable of interest, voting intention, is a binary variable that equals 1 when an individual intends to vote for a party (for example, if a person wants to vote for the SPD party, the indicator voting intention for SPD equals one, while it equals zero for all other parties). Therefore, a probit model is estimated in order to analyse the effect of media coverage on political preferences. As shown above, the tonality of media coverage with respect to the various parties is scaled from -1 to +1. An overall negative tone toward, e.g., CDU/CSU is reflected by a negative score for SCDU/CSU. The models estimated in table 1 can be described with the following set of j equations: 8

ACCEPTED MANUSCRIPT

Pr(Ymtj = 1) = Φ ( β 0 + β j St j−1 + β i St−−1j + β d xmd t + β k xtk ) , where Ymtj is a persons’ voting intention for party j, with j=CDU/CSU, SPD, FDP, or Gruene at time t (with m = 1, …, M, where M is a randomly drawn sample of respondents in every period). St j−1 is the respective lagged aggregated tonality of media reports on party j. St−−1j is an aggregated measure for the tonality of media coverage about all other parties except party j.7 For the party

RI PT

SPD, for example, St−−1j is the average tonality of all reports considering the parties CDU, FDP and Gruene (variable Reports on other parties t-1 in table 1). Positive reporting about other parties d could negatively influence the voting intention for the respective party. x mt comprises the d=10

explanatory variables that control for sociodemographic characteristics that vary with the

SC

respondents. Additionally, party affiliation is included as a control variable as voting intention is influenced by party affiliation. Individuals with a general preference for the CDU/CSU, for

M AN U

example, are more likely to vote for that party. Therefore, we control for the party affiliation of individuals in the voting intention regressions. Vector xtk covers the k=3 macroeconomic variables – the ifo business climate index, the unemployment rate, and the consumer price index. These macroeconomic variables should measure at least in parts the economic policy performance of the incumbent, which is assumed to influence electoral decisions.8 Moreover, these variables should capture the general performance and situation of the economy. β‘s are

TE D

respective coefficients vectors and  is the cumulative normal distribution. The lagged value of the tonality of own media reports ( , , ,  , ,  , ,  ) and of media reports of all other parties is implemented in the probit model, as we expect media coverage to have a delayed effect on individual preferences. In general, one can assume that it

EP

takes some time until news on political parties or persons are spread. Moreover, reports do not have a direct effect on the preferences of voters, but it is more like a gradual process of forming

AC C

an opinion about the news presented. Additionally, the frequency of the data supports the use of lagged values for media coverage. The Politbarometer is a monthly survey while media reports are collected by their date of publication. In the combined data set that is used for the analysis, reports published at the beginning and at the end of a month are represented by the same time stamp. However, the latter can hardly affect the intentions of voters. Therefore, it is reasonable to lag media coverage

We use an aggregated measure for the tonality of reports for all other parties, as implementing a tonality measure for all parties individually in each regression leads to a multicollinearity issue. 7

8

Among others, James Carville used the phrase "It's the economy, stupid" as a campaign strategist of Bill Clinton's successful 1992 presidential campaign against sitting President George H. W. Bush.

9

ACCEPTED MANUSCRIPT

by one month. To check the assumption of a lagged effect, we also estimate a model including contemporary media coverage, which is indeed statistically insignificant.9 Table 1 shows the results for the effect of media coverage on voting intention. The exogenous variables age, education, catholic, married, female, unemployed, labour union, political interest, and own economic situation control for socio-demographic characteristics. The interpretation of these factors is left to section 4.5, where we take a closer look at the impact of individual

RI PT

characteristics and macroeconomic factors. Moreover, party affiliation is incorporated as a control variable in the voting intention regressions. Macroeconomic effects are captured by the ifo business climate index, the unemployment rate, and the consumer price index, for which only contemporary effects are considered. However, as one might expect that macroeconomic factors could also have a delayed effect, we estimate models with different lags and excluding

SC

contemporary effects. Although the coefficients of the other variables do not change significantly when lagged values are used, only the first lag seems to be significant. Therefore, results are

M AN U

shown with the contemporary effects of the macroeconomic factors.10 Standard errors are clustered by individuals.11

As people are asked for which party they would vote if elections were lying ahead, voting intention is an adequate indicator for political preferences. These short-term preferences can deviate from the general party affiliation and are expected to be influenced by external factors, like media coverage.

TE D

Turning to our focal variables, the tonality of own media reports, these coefficients are indeed positive and statistically significant for all parties in table 1. A positive report on the SPD party, for example, has an enhancing effect on voter intentions. This is true for all parties. If media outlets report more positively about the respective party, voting intention for this party

EP

increases significantly. Reporting about a party has a strong effect on the voting intention and seems to have an influence on the voting decision of the electorate. In this case, it is also obvious

AC C

that further lags for media coverage, which have as well less influence, are not necessary as voting intention captures short-term preferences. Voters are especially sensitive toward media reports as they are more attentive during election periods and, therefore, can be influenced more strongly by current media reports during this period.12

9

We also considered a temporal effect over a higher number of periods; however, regressions with higher ordered lags did not lead to significantly differences in results. 10 We estimated models excluding the macroeconomic variables as well. All coefficients for media coverage are still positive and statistically significant. However, the macroeconomic factors might be able to explain some additional variation in the dependent variables. 11 We also used clustering approaches with respect to time and also with respect to time and individuals. Results did not change qualitatively and only marginally quantitatively with different clustering approaches. 12 We also considered a higher number of lags which did not lead to significantly different results. Higher lags are still positive and significant but the effects are smaller than for the first lag.

10

ACCEPTED MANUSCRIPT

Table 1: Probit Estimation – Media Coverage and Voting Intention VOTING/ VARIABLES

CDU/CSU

SPD

FDP

Gruene

Reports Reports CDU/CSU t-1

1.212*** (0.0840)

Reports SPD t-1

1.109*** (0.0788)

Reports FDP t-1

1.911*** (0.0778)

Reports Gruene t-1

RI PT

0.838*** (0.0945)

Reports other Parties t-1

-1.275*** (0.103)

-1.936*** (0.101)

Party Affiliation (same Party)

1.916*** (0.00811)

1.852*** (0.00820)

Age

0.00356** (0.00173)

-0.0246*** (0.00173)

-0.00551** (0.00241)

-0.0235*** (0.00210)

Education

0.00446 (0.00494)

-0.0293*** (0.00516)

0.0735*** (0.00722)

0.207*** (0.00666)

Catholic

0.172*** (0.00775)

-0.0767*** (0.00808)

0.0164 (0.0113)

-0.0840*** (0.0102)

Married

0.0421*** (0.00798)

0.0108 (0.00808)

0.0186 (0.0118)

0.0148 (0.0103)

Female

-0.00496 (0.00794)

0.0179** (0.00816)

-0.161*** (0.0114)

0.130*** (0.0103)

Unemployed

-0.0678 (0.0570)

0.00816 (0.0598)

-0.232** (0.100)

-6.68e-08 (0.0776)

Labour Union

-0.191*** (0.0119)

0.152*** (0.0116)

-0.215*** (0.0186)

0.0557*** (0.0149)

-0.0703*** (0.00622)

-0.0887*** (0.00630)

0.0336*** (0.00905)

-0.0568*** (0.00795)

-0.1070*** (0.00610)

-0.0345*** (0.00630)

-0.0526*** (0.00922)

-0.0196** (0.00797)

-0.00961** (0.00429)

-0.0108** (0.00448)

-0.0880*** (0.00648)

-0.0913*** (0.00576)

0.0345*** (0.00478)

-0.0606*** (0.00479)

0.0640*** (0.00636)

-0.0396*** (0.00644)

-0.00212** (0.00103)

-0.0312*** (0.00106)

0.0149*** (0.00149)

0.0123*** (0.00157)

-0.00627*** (0.000509)

0.00100* (0.000577)

-0.0146*** (0.000731)

0.00746*** (0.000684)

Constant

-0.129 (0.120)

2.588*** (0.124)

-2.001*** (0.180)

-3.402*** (0.177)

Observations

175,617

175,617

175,617

175,617

Wald-chi (Prob-chi)

61886.88 (0.0000)

54620.96 (0.0000)

16981.85 (0.0000)

32613.00 (0.0000)

0.3370

0.3170

0.2450

0.3515

Political Interest Macroeconomic Variables

CPI Ifo

M AN U

AC C

Unemployment Rate

TE D

Own economic situation

EP

Assessment of economic situation

Pseudo R-squared

-0.615*** (0.140)

2.266*** (0.0190)

2.268*** (0.0137)

SC

Individual Variables

-0.729*** (0.184)

Robust and clustered standard errors are given in parentheses. *** p<0.01, ** p<0.05, * p<0.1 According to the Politbarometersurvey, political interest is scaled from 1 to 5 (1 equals a very strong interest, 5 equals no interest); economic situation and own economic situation are scaled from 1 to 5 (very good to very bad).

Positive reports about other parties have a negative effect on the voting intention for the respective party. Considering the party SPD in model 2 of table 1, for example, negative media 11

ACCEPTED MANUSCRIPT

coverage about the other parties enhances the voting intention for the party SPD. Moreover, party affiliation has a positive and significant effect on the voting intention, like expected. If an individual has a general tendency towards the CDU/CSU, he or she is as well more likely to vote for that party. 13

4.3

Instrumental Variable Estimations

RI PT

Endogeneity issues mean that the estimates presented above are likely to be both biased and inconsistent. We address this issue by using the tonality of reports on banks to instrument media coverage of political parties.

To account for the binary dependent variable, voting intention, an instrumented probit model is

SC

estimated. Formally, the model estimated in table 3 is described by the two-step approach below:

M AN U

d Ymtj = β j S t j−1 + β d xmt + β k xtk + u mt

d S t j−1 = α 0 + α d xmt + α k xtk + α c ztc + υ mt ,

Where Ymtj again is the voting intention with j=CDU/CSU, SPD, FDP, or Gruene.14 The vectors xtk d and x mt cover the identical explanatory variables as in the probit model in section 4.2 and

comprises individual characteristics and macroeconomic variables. Media coverage is the

TE D

endogenous regressor that is expected to be correlated with the error term due to either omitted variable bias or reverse causality. ztc is a vector of c excluded instruments used to instrument media coverage. β’s and α’s are coefficient vectors and ( ,  ) ~ (0, Σ) are error terms. The

exogenous variables, as well as the excluded instruments ", are implemented in the reduced

EP

form equation St j−1 . In the first-stage regression ( St j−1 ), the endogenous explanatory variable is treated as a linear function of the instrument and the exogenous variables ( xtk and xntd ). The

AC C

prediction of the first stage is then implemented in the second stage ( Ymtj ) instead of the endogenous variable.

Results for the first stage with the excluded instrument only are presented in table 2. The more positive (or less negative) the international coverage on the banking industry in t, the more positive (or less negative) the national coverage on CDU/CSU, SPD and FDP in t+1 and the more

13

Results for the probit estimations without reports of the other parties are presented in the appendix in table A3. The coefficients for our variables of interest (reports about the own party) do not change significantly when the reports of other parties are neglected. 0 # < 0 14 # cannot be observed; instead, we observe: # = $ .   1 # ≥ 0

12

ACCEPTED MANUSCRIPT

negative the national coverage on the Gruene.15 The coefficients for our instrument are all statistically significant. Moreover, the F-test is in all cases clearly larger than 10 confirming the relevance of our instrument and demonstrate the strong relationship between the instrument and the endogenous variable - media coverage. More tests and discussions about the validity of our instrument follow at the end of this chapter after presenting the results for the instrumented probit models.

(1)

(2)

RI PT

Table 2: Results for the First Stage Regressions (3)

(4)

Instrument = Tonality of Reports on Banks SPD t-1

Observations

0.191*** (0.00162) -0.0837*** (0.000155) 188,134

0.0140*** (0.00149) -0.0965*** (0.000142) 188,134

F-test (p-value) R-squared

14011.97 (0.0000) 0.069

Reports on banks Constant

FDP t-1

Gruene t-1

0.280*** (0.00235) -0.0680*** (0.000225) 188,134

-0.172*** (0.00161) -0.0688*** (0.000154) 188,134

SC

CDU/CSU t-1

M AN U

VARIABLES

88.06 (0.0000) 0.060

14231.01 (0.0000) 0.070

11444.70 (0.0000) 0.057

Standard errors are given in parentheses. *** p<0.01, ** p<0.05, * p<0.1

Table 3 now shows results for voting intention when media coverage is instrumented by the tonality of coverage on banks. Controls are the same as in the probit estimation in table 1,

TE D

including both individual characteristics and macroeconomic factors. Table 3 shows the results when excluding the tonality of media coverage about all other parties as probit estimations have shown that the coefficient for the variable of main interest does not change significantly when the reports of the other parties are excluded. Moreover, there might as well be an endogeneity

EP

issue with the reports of other parties in this regression. However, results with reports of other parties as a control variable are shown in the appendix in table A4. For a closer look at the

AC C

impact of the individual characteristics and the macroeconomic variables, please see section 4.5. Results in table 3 confirm the positive effect of media coverage on voting intentions. The coefficients for our variables of main interest are statistically significant and positive, at least for the parties CDU/CSU, SPD and FDP. Positive media reports about the CDU/CSU, for example, have an increasing effect on the voting intention. Only for the party Gruene the coefficient becomes insignificant. A possible reason could be that the variation for this smaller party in our sample is too low. There is only a smaller number of reports about the Gruene party and moreover, as it is a smaller party, reports attributed to the party Gruene might not solely deal with it, which could lead to a bias in this estimate. These results can also be seen in table A4 in 15

The negative coefficient for the Gruene might be caused by the fact that in Germany in the years focused on beside the environment as their core topic the Gruene positioned themselves as strong detractor of the banking industry.

13

ACCEPTED MANUSCRIPT

the appendix, where reports of all other parties are considered as a control variable. Party affiliation still has a positive and statistically significant effect on the voting intention for the party.

Table 3: Instrumental Variable Estimation – Media Coverage and Voting Intention CDU/CSU

SPD

FDP

Reports Reports CDU/CSU t-1

1.122*** (0.414)

Reports SPD t-1

2.109*** (0.544)

Reports FDP t-1

Gruene

RI PT

VOTING/VARIABLES

12.60*** (1.544)

Reports Gruene t-1 1.864*** (0.00898)

2.108*** (0.0315)

0.00265 (0.00182) 0.00706 (0.00534) 0.167*** (0.00825) 0.0430*** (0.00843) 0.0139* (0.00839) -0.0429 (0.0657) -0.182*** (0.0130) -0.0927*** (0.00672) -0.123*** (0.00656) -0.00899* (0.00472)

-0.0242*** (0.00193) -0.0278*** (0.00562) -0.0729*** (0.00896) 0.00231 (0.00888) 0.00916 (0.00886) 7.50e-05 (0.0633) 0.142*** (0.0124) -0.0706*** (0.00706) -0.0372*** (0.00684) -0.00869* (0.00499)

-0.00780*** (0.00284) 0.0615*** (0.00841) 0.00733 (0.0131) 0.0243* (0.0133) -0.176*** (0.0132) -0.203* (0.115) -0.220*** (0.0213) 0.0874*** (0.0139) -0.0954*** (0.0121) -0.112*** (0.00819)

-0.0212*** (0.00237) 0.208*** (0.00710) -0.0830*** (0.0108) 0.0165 (0.0108) 0.134*** (0.0108) 0.0538 (0.0785) 0.0594*** (0.0157) -0.0450*** (0.0108) -0.0242*** (0.00839) -0.0935*** (0.00627)

0.0183*** (0.00596) -0.00977** (0.00439) -0.00676*** (0.000640) 0.980*** (0.379) 152,534

-0.0310*** (0.00607) -0.00611*** (0.00165) -0.00218 (0.00148) 0.443* (0.259) 152,534

0.0717*** (0.00890) 0.123*** (0.0163) -0.0115*** (0.000972) -11.84*** (1.543) 152,534

-0.132*** (0.0360) 0.00567** (0.00282) 0.00863*** (0.00113) -2.397*** (0.336) 152,534

Wald-chi (Prob-chi)

52912.75 (0.0000)

45552.98 (0.0000)

10996.95 (0.0000)

29469.83 (0.0000)

Wald test of exogeneity (Prob-chi)

4.81 (0.0283)

6.28 (0.0122)

55.38 (0.0000)

3.77 (0.0523)

Party Affiliation (same Party)

SC

1.904*** (0.00870)

-3.112 (2.006) 2.263*** (0.0145)

Individual Variables

M AN U

Age Education Catholic Married Female Unemployed

Assessment of Economic Situation Own Economic Situation Political Interest Macroeconomic Variables

EP

Unemployment Rate

TE D

Labour Union

CPI Ifo

AC C

Constant

Observations

Robust and clustered standard errors are given in parentheses. *** p<0.01, ** p<0.05, * p<0.1. According to the Politbarometersurvey, political interest is scaled from 1 to 5 (1 equals a very strong interest, 5 equals no interest); economic situation and own economic situation are scaled from 1 to 5 (very good to very bad). Media coverage is instrumented by tonality of reports on international banks.

Nevertheless, the results confirm the expectations that media reports have a significant impact on voting intentions, thus affirming the important role of the media in democracies. Media

14

ACCEPTED MANUSCRIPT

coverage has a significant impact on the opinion-formation of voters, especially during electoral periods. Moreover, the Wald test of exogeneity at the bottom of table 3 confirms the use of an instrumental variable estimation. The Wald test of exogeneity examines whether the instrumented variables – in our case the tonality of media coverage of political parties – is indeed endogenous. The test has the null hypothesis that the regressors are exogenous and a

RI PT

probit model would be the right choice. In table 3, the H0 of the Wald test of exogeneity is clearly rejected for all parties confirming the use of instrumental variable probit estimation as an instrumented variable model is required in order to have reliable estimates.16

To demonstrate the validity of the results of the instrumental variable estimation presented

SC

above, we additionally carry out several tests to confirm that media reports on the international banking sector is an adequate instrument for the media coverage of political parties.17 We first check if the results suffer from underidentification. This is the case when the instrument is not

M AN U

correlated with the endogenous regressor. High values for the Kleibergen-Paap rk LM statistic, as well as for the Kleibergen-Paap rk Wald statistic (with p-values equal to 0.000), result in a strong rejection of the null hypothesis of no correlation, thus confirming the relevance of the instrument (Π ≠0). However, if underidentification is rejected, it could still be the case that the model is only weakly identified. To test for weak correlation of the instrument with the regressor, a weak identification test (using Kleibergen-Paap rk Wald F test) is considered. The

TE D

statistic is clearly larger than the critical values based on Stock-Yogo (2005) and we reject the null hypothesis of weak correlation. Thus, the above model does not suffer from weak

4.4

EP

instrument and the instrument is strongly correlated with the regressor.

Robustness Checks

AC C

As our media variables are crucial for our analysis but are only available at an aggregate level, one might argue that we artificially increase the degrees of freedom, we carry out some robustness checks using an aggregated approach to confirm the results above. Unfortunately, we are not able to make a closer link between the individuals in the Politbarometer survey and the media data as there are, for example, no information on which media outlets these persons read.

16

In some cases, this hypothesis can be rejected if the instruments used are weak. Therefore, table 2 shows the first stage results for the excluded instrument to demonstrate the relevance of our instrument. The F-test is used to test the null hypothesis that the coefficients on the instrument are jointly zero. As a rule of thumb, a F-statistic larger than 10 would be a cause of concern since it signals a problem of weak instrument. Our F-statistics for the different specifications in table 2 are clearly higher than 10, confirming that the instrument is significantly associated with the endogenous variable. 17 In the instrumented probit estimation (presented in table 3), the reduced form for the endogenous explanatory variable is linear, so we can use the same diagnostics as in the linear case to evaluate the strength of our instruments.

15

ACCEPTED MANUSCRIPT

Therefore, we aggregate all individual information up to the country level.18 We analyse this data with two different models – a count regression model and linear regression model – to examine whether we are able to find an effect of media through its tone on voting intention in general. We use two different dependent variables for the two models: first, in order to run count data regressions, we add up all individuals who have indicated to vote for a specific party and repeat this procedure for all of the four parties (CDU/CSU, SPD, FDP and Gruene) throughout the

RI PT

observation period. The dependent variable in the count regression model thereby simply indicates the number of voting intentions for each party and month considered in our sample. By these means we end up with a panel data set that includes the intended number of votes per party and month. Similarly, for the linear regression model, the intended voting shares (number of intended votes per party and month, divided by the number of total respondents) for the four

SC

parties are used as dependent variable.

The explanatory variables are the same for both models and largely correspond to the variables

M AN U

used in chapter 4.2 and 4.3. However, as the individual characteristics have to be aggregated on the party level we build either averages of the individual characteristics (age, assessment of economic situation, own economic situation) or shares of the number of (catholic, married, female, unemployed, unionized) respondents in order to receive information on the parties’ representative voters.19

The resulting dataset comprises the variables number of intended votes and voting shares on a

TE D

monthly basis per party. While the individual characteristics of the Politbarometer survey have to be aggregated respectively, macroeconomic variables which are already available on a monthly basis can be used in the same manner as in chapter 4.2 and 4.3. The panel covers the

and Gruene.

EP

period from February 1998 to December 2012 for the four panel variables CDU/CSU, SPD, FDP

18

AC C

The models estimated in table 4 can be described with the following equation:

Yt j = β0 + β j St j−1 + β d xtd + β k xtk ,

We are grateful tot he editor for suggesting this approach. For the age of the respondents, e.g., we build the average age of a CDU/CSU voter and repeat this for all other parties. As education is scaled from 1 to 7 (no education to university degree), we use the share of persons with an education higher or equal than the Abitur voting for a specific party in our aggregated approach. For the dummy variables catholic, married, unemployed and labor union used in chapter 4.2 and 4.3, we calculate the share of catholics voting for CDU/CSU, SPD, FDP or Gruene, the share of married persons, the share of unemployed persons and the share of persons that are member of a labor union. For the assessment of the economic situation in general and the own economic situation in particular, we again build averages over the respondents by voting intention. The variables party affiliation and political interest, which were used disaggregated models, are exempted of the aggregated approach as the average value of these two variables are less meaningful. Moreover, the variable political interest is not queried over the entire observation period and would further reduce our sample. However, both variables are found to be statistically insignificant, if included. 19

16

ACCEPTED MANUSCRIPT

Where Yt j is either the number of intending votes or the intended voting share for party j at time j t with j=CDU/CSU, SPD, FDP and Gruene. St −1 is the respective lagged aggregated tonality of d media reports on party j. xt covers the d=9 aggregated individual characteristics described k above and xt are the three macroeconomic variables.

To analyse the robustness of our results derived from the regressions based on the individual voting intentions, we run various regression using the aggregate data. Table 4 presents some of

RI PT

the results for simple count data regression models on the number of voting intentions (column 1 and 2), a fixed effects negative binomial regression model20 (column 3) as well as for linear fixed effects regression models (column 4 and 5) on the voting intention share.21 Moreover, the model also accounts for the four parties as a panel version is estimated. While column 1

SC

represents the baseline model with one lag for the tonality of media coverage, column 2 considers two additional lags for the reports. Column 3 shows the fixed effects version of column

M AN U

2, based on three lags.

Overall, all of the models confirm that that the tonality of reports is positive and statistically significant. Positive reporting about political parties leads to an increase in the voting intention in all of the specifications. The more positive the media is reporting, the higher is the number of people which indicate to vote for a party. This effect also remains when more than one lag for the tonality is implemented. Results in column 2 show that the second (but not the third) lag is statistically significant and it has already a much lower effect on the voting intention than the

TE D

first lag. This coincides with the expectation that voting intentions are especially influenced by current incidents which are as well expected to be extensively debated in the media.22 However, when using a fixed effects model, also the third lag becomes statistically significant. For the individual characteristics we can only find a significant effect for the variables education, the

EP

share of Catholics and the share of married persons which could be traced back to the highly aggregated data we use for this approach. Using a fixed effects approach, also the assessment of

AC C

the economic situation becomes statistically significant. To verify our findings, we estimate a linear fixed effects regression model with intended voting shares as the dependent variable. Columns 4 and 5 in table 4 present the results for this estimation.

20

We also used various specification of pooled as well as of fixed effects negative binomial models, all of which have confirmed the former results. 21 Because of the existence of overdispersio negative binomial models instead of Poisson models are used. Overdispersion is confirmed by an LR-test which strongly suggests that alpha is non-zero. 22 As for the regressions in chapter 4.2 and 4.3, we estimated several models with different number of lags. However, most of the higher lags are insignificant and what is more important, the first lag is always statistically significant and has a much stronger effect than higher-order lags.

17

ACCEPTED MANUSCRIPT

Table 4: Robustness Checks – Media Coverage and Voting Intention (1)

(2)

(3)

(4)

(5)

Number of intended Votes

Number of intended Votes

Number of intended Votes

Shares of intended Votes

Shares of intended votes

2.102*** (0.261)

1.556*** (0.304)

1.362*** (0.171)

0.266*** (0.0255)

0.182*** (0.0286)

Reports t-2

0.891*** (0.327)

0.859*** (0.175)

0.103*** (0.0302)

Reports t-3

0.348 (0.309)

0.767*** (0.171)

0.0817*** (0.0288)

VARIABLES

Reports t-1

Yes

0.140*** (0.0129)

0.158*** (0.0129)

SPD

Yes

0.100*** (0.0105)

0.111*** (0.0104)

Yes

-0.0867*** (0.00946)

-0.0774*** (0.00929)

-0.0357 (0.034)

-0.00534 (0.00568)

-0.00739 (0.00558)

-1.552*** (0.285)

-1.478*** (0.186)

-0.196*** (0.0304)

-0.170*** (0.0297)

0.780** (0.318)

1.180*** (0.206)

0.0204 (0.0340)

0.00985 (0.0329)

0.626* (0.362)

0.799*** (0.224)

0.118*** (0.0349)

0.0973*** (0.0339)

0.237 (0.319)

0.0814 (0.212)

-0.0521 (0.0330)

-0.0385 (0.0322)

-0.501 (2.693)

-1.093 (2.749)

-2.513 (1.735)

-0.00341 (0.260)

-0.0274 (0.255)

-0.220 (0.504)

-0.249 (0.503)

0.407 (0.330)

-0.0579 (0.0551)

-0.0454 (0.0535)

0.0397 (0.102)

0.0229 (0.102)

0.193*** (0.059)

0.0280*** (0.0102)

0.0216** (0.00990)

-0.158 (0.237)

-0.209 (0.236)

-0.208 (0.143)

-0.0146 (0.0233)

-0.0228 (0.0224)

Unemployment Rate

0.116*** (0.0214)

0.117*** (0.0214)

0.0232** (0.0117)

5.23e-05 (0.00219)

0.000586 (0.00211)

CPI

0.0514*** (0.00473)

0.0536*** (0.00478)

0.0288*** (0.002)

0.000691 (0.000474)

0.000928** (0.000462)

-0.00796** (0.00327)

-0.00932*** (0.00326)

-0.000701 (0.0018)

0.000324 (0.000339)

0.000118 (0.000328)

-2.274*** (0.844)

-2.121** (0.846)

-6.593*** (0.515)

0.0981 (0.0831)

0.140* (0.0806)

708

700

700

708

700

alpha

0.0575

0.0531

0.05116

-

-

LR-test of alpha=0 [p-value]

5369.84 [0.00]

4908.33 [0.00]

4710.23 [0.00]

-

-

-

-

-

0.875

0.883

FDP Individual Variables -0.0800 (0.0566)

Share High Educated

-1.691*** (0.285) 0.775** (0.317)

Share Married

0.706** (0.359)

Share Female

0.171 (0.317)

Share Labour Union Assessment of economic situation Own Economic Situation

Ifo Constant Observations

R-squared

AC C

Macroeconomic Variables

EP

Share Unemployed

TE D

Share Catholic

-0.0836 (0.0573)

M AN U

Age

SC

CDU/CSU

RI PT

Reports

Standard errors are given in parentheses. *** p<0.01, ** p<0.05, * p<0.1.

18

ACCEPTED MANUSCRIPT

Again, media coverage on political parties has a statistically significant and positive effect on the voting intention. If media reports more positive about a political party, voting shares for this party increase. To account for the panel structure, we implement dummy variables for the parties CDU/CSU, SPD and FDP. These dummy variables are as well statistically significant and indicate the differences between the four parties. Again, the model shown in column 5 considers up to three lags for the tonality of media coverage. These lags are positive and statistically

RI PT

significant, however, the results confirm that the first lag has the strongest influence on voting intention.

All in all, the results of these aggregated estimations in table 4 confirm our previous findings of a strong effect of the media on voting intentions through its tome and approve the important role

SC

media plays in the opinion-formation of individuals. This holds especially in democracies and

4.5

M AN U

during electoral periods.

What drives the Electorate? A Closer Look

As voting intentions are of special interest for the outcomes of elections, we take a closer look at the results provided in table 3. As mentioned, by applying instrumental variable estimation, we can identify a positive impact of media coverage on short-term voting intentions, at least in the case of CDU/CSU, SPD, and FDP. The more positive the media coverage of CDU/CSU, SPD, and

TE D

FDP, the greater the tendency to vote for it.

A closer look on the impact of individual characteristics broadly confirms findings of public opinion research in Germany: Females have a higher tendency to vote for the Christian conservative CDU/CSU and Gruene, but a lower tendency to vote for the liberal-conservative FDP.

EP

The elderly have a lower tendency to vote for the SPD, FDP, and Gruene. Catholic voters have, not surprisingly, a higher tendency to vote for Christian conservative CDU/CSU and a lower tendency

AC C

to vote for SPD and Gruene. Married voters have a higher tendency to vote Christian conservative CDU/CSU. The higher the education of the voter, the higher the tendency to vote for the smaller and more liberal FDP or Gruene. If a voter is unemployed, the tendency to vote for FDP is lower. If he or she is member of a labour union, the tendency to vote for the more leftish SPD and Gruene is high, but the tendency to vote for the more rightist CDU/CSU and FDP is low. Voters with a positive assessment of the general economic situation have a higher tendency to vote for CDU/CSU and SPD, the established parties. However, if the assessment of their own economic situation is rather negative, the tendency to vote for any of the concerned parties decreases.

19

ACCEPTED MANUSCRIPT

Finally, if the voter sees him- or herself as politically interested, they have a higher tendency to vote for all considered political parties CDU/CSU, SPD, FDP, and Gruene.23 A closer look at the macroeconomic control variables shows that the higher the unemployment rate, the higher the tendency to vote for CDU/CSU and FDP and the lower the tendency to vote for SPD and Gruene. In addition, the higher the inflation rate, the lower the tendency to vote for the more leftish SPD. Finally, the better the business climate, the lower the tendency to vote for

5

RI PT

CDU/CSU and FDP.

Conclusion

A growing literature uses media data to explain perception and behaviour of individuals. In the

SC

political context the question arises whether media report more on political parties because of their success or if their success is caused by media reports. To tackle this question, we

M AN U

investigate how the tonality of media coverage affects political preferences, namely voting intention.

For our empirical analysis, we merge 14 years of human coded data derived from leading media in Germany with the results of the comprehensive German Politbarometer survey from February 1998 through December 2012. As media coverage may not only affect voter political preferences but also the general political mood among the electorate, we assume that endogeneity and

TE D

reverse causality are present. Hence, we employ instrumental variable estimations to address these issues.

First, the results of the probit estimations indicate that media coverage impacts short-term voting intentions. However, these results could be caused by the simultaneity of the two

EP

variables of interest – media coverage and political preferences. Hence, the probit estimation might lead to a biased estimate of media coverage on the political preferences variables. To

AC C

correct for this bias, an instrumental variable (IV) approach is applied. Results for the instrumental variable estimation reveal that voting intentions are strongly affected by media coverage. The more positively a party is covered by the media, for instance in the context of improving poll results, the higher the tendency to vote for this specific party. This even holds, if the tonality of the political coverage is not triggered by events connected to the domestic policy but by international media spill overs. This hints at the special responsibility of media in democracies.

23

For the interpretation of these coefficients, one should keep in mind the definition for the three variables general economic situation, own economic situation, and political interest. A low value for these variables corresponds to a positive assessment of the (own) economic situation and a high political interest, respectively. Whereas a high value reflects a negative assessment and no political interest. See section 3.2 for a more detailed variable description.

20

ACCEPTED MANUSCRIPT

Literature

AC C

EP

TE D

M AN U

SC

RI PT

Anderson S.P. and J. McLaren (2012), Media Mergers and Bias with Rational Consumers, Journal of the European Economic Association, 831–859. Angrist, J. D. and A. B. Krueger (1991), Does Compulsory School Attendance Affect Schooling and Earnings?, Quarterly Journal of Economics, 106, 979-1014. Alsem, K. J., Brakman, S., Hoogduin, L., & Kuper, G. (2008), The impact of newspapers on consumer confidence: does spin bias exist?, Applied Economics, 40, 531 - 539. Ball-Rokeach, S. J. (1985), The origins of individual media-system dependency: a sociological framework”, Communication Research, 12, 485-510. Ball-Rokeach, S. J., & DeFleur, M. L. (1976), A dependency model of mass-media effects, Communication Research, 3, 3-21. Beckmann, K., Dewenter, R. and Thomas, T. (2017), Can news draw blood? The impact of media coverage on the number and severity of terror attacks, Peace Economics, Peace Science and Public Policy, 23, 1-16. Benesch, C., S. Loretz, D. Stadelmann and T. Thomas (2018) Media Coverage and Migration Worries. Forthcoming in: Journal of Economic Behavior & Organization. Berlemann, M. and Thomas T. (2018), The Distance Bias in Natural Disaster Reporting. Applied, Economics Letters, published online first: 6 October 2018. Bernhardt, D., Krasa, S., & Polborn, M. (2008). Political polarization and the electoral effects of media bias. Journal of Public Economics, 92, 1092–1104. Besley, T., & Prat, A. (2006). Handcuffs for the grabbing hand? Media capture and government accountability. American Economic Review, 96, 720–736. Chadi, A. (2015), Concerns about the Euro and happiness in Germany during times of crisis, European Journal of Political Economy, 40, 126-146. D‘Alessio, D. D.; Allen, M. (2000), Media bias in presidential elections: a meta-analysis, Journal of Communication, 50, 133–156. DellaVigna, S. and E. Kaplan (2007), The Fox News Effekt: Media Bias and Voting, The Quarterly Journal of Economics, 122, 1187-1234. Dewenter, R., & Heimeshoff, U. (2014). Media Bias and advertising: Evidence from a German car magazine. Review of Economics, 65, 77–94. Dewenter, R., & Heimeshoff, U. (2015), More Ads, More Revs? A Note on Media Bias in Review Likelihood, Economic Modelling 44, 156-161. Dewenter, R., U. Heimeshoff, and T. Thomas (2016), Media Coverage and Car Manufacturers’ Sales, Economics Bulletin, 36, 976-982. Doms, M. and N. Morin (2004), Consumer sentiment, the economy, and the news media. Finance and Economics Discussion Series 2004-51, Board of Governors of the Federal Reserve System. Druckman J. N. and Parkin M. (2005), The Impact of Media Bias: How Editorial Slant Affects Voters, Journal of Politics, 67, 1030-1049. Dunham, W.R. (2013), Framing the Right Suspects: Measuring Media Bias, Journal of Media Economics, 26, 122-147. Durante, R., & Zhuravskaya, E. (2015). Attack when the world is not watching? International media and the Israeli-Palestinian conflict, CEPR Discussion Paper 10750. Forthcoming in: Journal of Political Economy. Eisensee, T. and D. Strömberg (2007), News Droughts, News Floods, and U. S. Disaster Relief, The Quarterly Journal of Economics, 122, 693-728. Enikolopov, R., Petrova, M. and Zhuravskaya, E. (2011) Media and Political Persuasion: Evidence from Russia, American Economic Review, 101, 3253–3285. Entman R. M. (2007), Framing Bias: Media in the Distribution of Power. Journal of Communication, 57, 163-173. Festinger, L.A. (1957), A Theory of Cognitive Dissonance, Stanford University Press: Stanford. 21

ACCEPTED MANUSCRIPT

AC C

EP

TE D

M AN U

SC

RI PT

Forschungsgruppe Wahlen, Mannheim (2017): Politbarometer 1977-2015 (Partielle Kumulation). GESIS Datenarchiv, Köln. ZA2391 Datenfile Version 7.0.0, doi:10.4232/1.12733. Friebel, G., & Heinz, M. (2014). Media slant against foreign owners: Downsizing. Journal of Public Economics, 120, 97–106. Gambaro, M., & Puglisi, R. (2015). What do ads buy? Daily coverage of listed companies on the Italian press. European Journal of Political Economy, 39, 41–57. Garrett, R. K. (2009) Echo chambers online?: Politically motivated selective exposure among Internet news users, Journal of Computer-Mediated Communication, 14, 265–285. Garz, M. (2012), Job Insecurity Perceptions and Media Coverage of Labour Market Policy, Journal of Labour Research, 33, 528-544. Garz, M. (2013), Unemployment Expectations, Excessive Pessimism, and News Coverage, Journal of Economic Psychology, 34, 156-168. Garz, M. (2014), Good news and bad news: evidence of media bias in unemployment reports, Public Choice, 161, 499–515. Garz, M. (2018), Retirement, consumption of political information, and political knowledge, European Journal of Political Economy, 53, 109-119. Gentzkow, M., & Shapiro, J. M. (2010). What drives media slant? Evidence from U.S. daily newspapers. Econometrica, 78, 35-71. Gentzkow, M., A., Jesse M. Shapiro, Michael Sinkinson (2011), The Effect of Newspaper Entry and Exit on Electoral Politics, American Economic Review, Vol. 101, S. 2980-3018. Gentzkow, M., N. Petek, J.M. Shapiro, M. Sinkinson (2015), Do Newspapers Serve the State? Incumbent Party Influence on the US Press, 1869-1928, Journal of the European Economic Association, 13, 29-61. Gentzkow, M. Shapiro, J.M. & Stone, D. F. (2015), Media Bias in the Marketplace: Theory, Handbook of Media Economics, Vol 1b, 623-645. Goidel, R. K., & Langley, R. E. (1995), Media coverage of the economy and aggregate economic evaluations, Political Research Quarterly, 48(2), 313-328. Grimmer, J. and B. M. Steward (2013) Text as Data: The Promise and Pitfalls of Automatic Content, Analysis Methods for Political Texts, Political Analysis, 21, 267-297. Groseclose, T., & Milyo, J. (2005), A measure of media bias, Quarterly Journal of Economics, 120, 1191–1237. Heinz, M., & Swinnen, J. (2015), Media slant in economic news: A factor 20, Economics Letters, 18–20. Horrace, W. C. & R. L. Oaxaca, 2006. Results on the bias and inconsistency of ordinary least squares for the linear probability model. Economics Letters, 90, 321-327. Kauder, B. and Potrafke, N. (2015), Just hire your spouse! Evidence from a political scandal in Bavaria, European Journal of Political Economy, 38, 42-54. Kim, J., Petrocik, J., & Enokson, S. (1975), Voter Turnout Among the American States: Systemic and Individual Components. The American Political Science Review, 69, 107-123. Knoke D. & M. Hout (1974), Social and Demographic Factors in American Political Party Affiliations, American Sociological Review, 39, 700-713. Köcher, R. (2013), Banks in public perception - results of a representative survey commissioned by the Bundesverband deutscher Banken, Institut für Demoskopie Allensbach (download: July 19th, 2018: https://bankenverband.de/media/files/Umfrageergebnis_9e01sBY.pdf) Kuhn, U. (2009), Stability and Change in Party Preferences, Swiss Political Science Review 15, 463–94. Larcinese, V. and Sircar, I. (2017), Crime and punishment the British way: Accountability channels following the MPs’ expenses scandal, European Journal of Political Economy, 47, 7599. Lamla, M. J. and T. Maag (2012), The Role of Media for Inflation Forecast Disagreement of Households and Professional Forecasters, Journal of Money, Credit and Banking, 7, 13251350. 22

ACCEPTED MANUSCRIPT

AC C

EP

TE D

M AN U

SC

RI PT

Meier, S. and Sprenger, C. D. (2015), Temporal stability of time preferences. Review of Economics and Statistics, 97, 273–286. Morris, J.S. (2007), Slanted Objectivity? Perceived Media Bias, Cable News Exposure, and Political Attitudes, Social Science Quarterly, Volume 88, S. 707–728. Mullainathan, S., Shleifer, A., (2005), The market for news. American Economic Review 95, 1031–1053. Muller, L., Sefton, M., Steinberg, R., and Vesterlund, L. (2008). Strategic behavior and learning in repeated voluntary contribution experiments. Journal of Economic Behavior & Organization, 67, 782–793. Nadeau, R., R. G. Niemi and T. Amato (2000), Elite economic forecasts, economic news, mass economic expectations, and voting intentions in Great Britain, European Journal of Political Research, 38, 135-170. Ohr, D. & Schrott, P.R. (1998), Political Information Seeking During Election Campaigns: Evidence From Local and National Elections in Germany, prepared the 1998 Annual Meeting of the American Political Science Association, Boston, Mass. (USA), September 3-6, 1998. Prat, A. (2014), Media power, CEPR Discussion Paper No. 10094. Forthcoming in: Journal of Political Economy. Puglisi, R. & Snyder, J. M. (2015), Empirical Studies of Media Bias, Handbook of Media Economics, Vol 1b, 647-667. Reuter, J., & Zitzewitz, E. (2006), Do ads influence editors? Advertising and bias in the financial media, Quarterly Journal of Economics, 121, 197–227. Snyder Jr., J.M. and D. Strömberg (2010), Press Coverage and Political Accountability, Journal of Political Economy, 118, 355-408. Sobbrio, F. and Navarra, P. (2010), Electoral participation and communicative voting in Europe, European Journal of Political Economy, 26, 185-207. Soroka, S. N. (2006), Good news and bad news: Asymmetric responses to economic information. The Journal of Politics, 68, 372-385. Stock J, Yogo M. (2005). Testing for Weak Instruments in Linear IV Regression, In: Andrews DWK Identification and Inference for Econometric Models. New York: Cambridge University Press ; 2005. pp. 80-108. Valentino, N. A., Banks, A. J., Hutchings, V. L. and Davis, A. K. (2009), Selective Exposure in the Internet Age: The Interaction between Anxiety and Information Utility, Political Psychology, 30, 591-613. van Raaij, W. F. (1989), Economic news, expectations and macro-economic behavior, Journal of Economic Psychology, 10(4), 473-493.

23

ACCEPTED MANUSCRIPT

Appendix

Table A1: Analysed Media Set Media

Observations:

Observations:

News items on all political protagonists and topics

News items on CDU/CSU, SPD, FDP and Gruene only

TV news shows (private) 127,459

Sat.1 News

77,466

73,991

ProSieben News

43,272

41,697

Tagesthemen (ARD)

381,089

358,072

Tagesschau (ARD)

267,975

251,316

heute (ZDF)

247,482

231,649

SC

TV news shows (public broadcasting service)

heute journal (ZDF)

361,493

TV magazines (public broadcasting service) Frontal 21 (ZDF) Kontraste (SFB) Monitor (WDR) Panorama (NDR) Plusminus (ARD) Report (BR) Report (SWR) WISO (ZDF)

Daily newspaper Bild Berliner Zeitung Die Welt Die Tageszeitung (taz)

EP

Berlin direkt (ZDF)

TE D

Bericht aus Berlin (ARD)

Frankfurter Allgemeine Zeitung (F.A.Z.) Frankfurter Rundschau

340,922

5,224

4,347

26,712

23,906

6,333

5,176

6,490

6,101

10,085

8,779

2,698

2,677

8,875

7,968

8,698

7,348

M AN U

Fakt (MDR)

122,303

RI PT

RTL aktuell

5,029

4,675

77,432

68,989

102,667

94,117

352,001

336,314

471,101

431,780

1,413,879

1,335,349

528,085

477,520

1,379,282

1,288,424

970,249

898,476

1,210,440

1,132,985

Bild am Sonntag (BamS)

140,659

136,157

Frankfurter Allgemeine Sonntagszeitung (FAS)

212,864

202,052

Focus

364,770

346,773

527,410

491,526

Welt am Sonntag (WamS)

180,217

172,823

Stern

113,860

108,010

38,124

29,515

AC C

Süddeutsche Zeitung (SZ) Magazines and weeklies

Spiegel

Super Illu Die Woche

76,885

70,809

Rheinischer Merkur

152,665

144,674

Die Zeit

206,269

193,812

10,105,239

9,451,032

Total Number of observations

24

ACCEPTED MANUSCRIPT

Table A2: Media Set for the Reports on the Banking Sector Media

News Items

Total

Austria ORF TV - Zeit im Bild (ZIB 1) Canada

313

157 13,924

466 17,717

18,183

363 527 1,423 2,526 90 173 624 2,394 3,150 356 6,682 32,497 13,897 82,309 4,732 52 3,022 49 804 934 38 70 15,020 4,012 1,076 1,115 104 28,461 6,270 13,520 110 463 814 1,334 2,615

231,626

AC C

EP

TE D

M AN U

SC

Le Journal 20.00 (TF1) Les Echos Germany ARD Börse vor acht ARD Plusminus ARD Tagesschau ARD Tagesthemen BR Report Bericht aus Berlin Bild am Sonntag Bild-Zeitung Capital Deutschlandfunk FAS FAZ FR FTD Focus MDR Fakt Manager Magazin NDR Panorama RTL AKTUELL Rheinischer Merkur SFB Kontraste SWR Report SZ Spiegel Stern Super Illu WDR Monitor Welt Welt am Sonntag Wirtschaftswoche ZDF Berlin direkt ZDF Frontal 21 ZDF WISO ZDF heute ZDF heute journal Italy

14,081

RI PT

CBC News: The National The Globe and Mail France

313

TG1 RAI 1 South Africa

780

Business Day Business Report Business Times City Press FinWeek Financial Mail Mail & Guardian Rapport SABC 2 Afrikaans News SABC 2 Setswana/Sotho News SABC 2 Zulu/Xhosa News SABC 3 News @ 18h30 SABC 3 News @ One Sake

28,264 13,825 3,063 1,157 3,817 4,786 993 394 293 242 299 427 3,302 6,463

25

780

68,503

ACCEPTED MANUSCRIPT

The Sunday Times e.tv News Spain

833 345

TVE 1 Telediario Switzerland

1,090

1,090

SF Börse SF Eco SF Rundschau SF Tagesschau United Kingdom

899 252 124 1,344

2,619

187,330

886 7,416 4,114 5,387 1,020 382 5,651 760 51,619 1,555 1,529 8,017 811 16,153 393 2,822 385 5,831 55,864 9,447

180,042

SC M AN U TE D

EP

ABC Barron´s Bloomberg Business Week Boston Globe CBS CNN International Desk Chicago Tribune FOX Financial Times (US ed.) Forbes Fortune Los Angeles Times NBC New York Times Newsweek San Francisco Chronicle Time USA Today (US-Edition) WSJ US Washington Post Undefined

344,008 1,048,575

AC C

Total

1,195 800 888 16,737 131,029 5,722 716 1,026 965 1,083 2,100 3,650 21,419

RI PT

BBC 1 BBC 2 Newsnight BBC World Service Daily Telegraph FT Guardian ITV News at Ten Observer Sun Sunday Telegraph Sunday Times The Economist Times United States

26

ACCEPTED MANUSCRIPT

Table A3: Probit Estimation – Media Coverage and Voting Intention without reports of other parties VOTING/ VARIABLES

CDU/CSU

SPD

FDP

Gruene

Reports Reports CDU/CSU t-1

1.011*** (0.0830)

Reports SPD t-1

1.116*** (0.0799)

Reports FDP t-1

1.870*** (0.0770)

Reports Gruene t-1

2.265*** (0.0190)

-0.00549** (0.00241) 0.0734*** (0.00722) 0.0165 (0.0113) 0.0187 (0.0118) -0.160*** (0.0114) -0.232** (0.100) -0.215*** (0.0186) 0.0318*** (0.00905) -0.0522*** (0.00922) -0.0879*** (0.00648)

-0.0236*** (0.00210) 0.207*** (0.00666) -0.0843*** (0.0102) 0.0148 (0.0103) 0.130*** (0.0103) 0.00241 (0.0776) 0.0559*** (0.0149) -0.0566*** (0.00795) -0.0203** (0.00797) -0.0916*** (0.00576)

-0.0602*** (0.00478) -0.0250*** (0.00101) 0.00103* (0.000570) 2.163*** (0.123) 175,617

0.0657*** (0.00633) 0.0161*** (0.00146) -0.0152*** (0.000709) -1.991*** (0.180) 175,617

-0.0416*** (0.00642) 0.0157*** (0.00135) 0.00693*** (0.000669) -3.608*** (0.170) 175,617

61786.43 (0.0000)

54694.87 (0.0000)

16998.49 (0.0000)

32590.73 (0.0000)

0.3363

0.3151

0.2448

0.3513

1.915*** (0.00811)

1.846*** (0.00817)

0.00360** (0.00173) 0.00405 (0.00494) 0.172*** (0.00775) 0.0421*** (0.00798) -0.00478 (0.00793) -0.0627 (0.0570) -0.190*** (0.0119) -0.0697*** (0.00621) -0.108*** (0.00609) -0.0100** (0.00429)

-0.0242*** (0.00172) -0.0301*** (0.00516) -0.0782*** (0.00807) 0.0110 (0.00806) 0.0191** (0.00815) 0.0136 (0.0599) 0.153*** (0.0116) -0.0937*** (0.00629) -0.0372*** (0.00629) -0.0112** (0.00447)

0.0437*** (0.00470) 0.000827 (0.000998) -0.00706*** (0.000502) -0.319*** (0.118) 175,617

Age Education Catholic

M AN U

Married Female Unemployed Labor Union Assessment of Economic Situation Own Economic Situation

CPI Ifo Constant Observations

AC C

Wald-chi (Prob-chi)

EP

Unemployment Rate

TE D

Political Interest Macroeconomic Variables

Pseudo R-squared

SC

Individual Variables

RI PT

Party Affiliation (same Party)

0.747*** (0.0927) 2.268*** (0.0137)

Robust and clustered standard errors are given in parentheses. *** p<0.01, ** p<0.05, * p<0.1. According to the Politbarometersurvey, political interest is scaled from 1 to 5 (1 equals a very strong interest, 5 equals no interest); economic situation and own economic situation are scaled from 1 to 5 (very good to very bad).

27

ACCEPTED MANUSCRIPT Table A4: Instrumental Variable Estimation – Media Coverage and Voting Intention with Reports of other Parties VOTING/ VARIABLES

CDU/CSU

SPD

FDP

Gruene

Reports Reports CDU/CSU t-1

1.168*** (0.419)

Reports SPD t-1

2.322*** (0.556)

Reports FDP t-1 Reports Gruene t-1 Reports other Parties t-1 Party Affiliation (same Party)

-1.408*** (0.155) 1.905*** (0.00870)

-1.787*** (0.140) 1.868*** (0.00899)

0.00270 (0.00182) 0.00754 (0.00535) 0.168*** (0.00826) 0.0433*** (0.00843) 0.0139* (0.00839) -0.0469 (0.0658) -0.182*** (0.0130) -0.0938*** (0.00673) -0.121*** (0.00657) -0.00837* (0.00473)

-0.0243*** (0.00194) -0.0273*** (0.00563) -0.0720*** (0.00897) 0.00264 (0.00889) 0.00887 (0.00887) 0.000816 (0.0632) 0.141*** (0.0124) -0.0722*** (0.00707) -0.0342*** (0.00686) -0.00740 (0.00499)

0.0103* (0.00583) -0.0147*** (0.00405) -0.00559*** (0.000600) 1.284*** (0.358) 152,534

-0.0442*** (0.00619) -0.0187*** (0.00184) -0.00201 (0.00149) 1.611*** (0.306) 152,534

Age Education

M AN U

Caholic Married Female Unemployed Labor Union Assessment of Economic Situation

Macroeconomic Variables Unemployment Rate CPI Ifo Constant

AC C

Observations

EP

Political Interest

TE D

Own Economic Situation

-2.843*** (0.370) 2.156*** (0.0263)

-2.481* (1.338) 0.323 (0.400) 2.263*** (0.0145)

-0.00749*** (0.00275) 0.0632*** (0.00814) 0.0109 (0.0126) 0.0260** (0.0129) -0.175*** (0.0128) -0.216* (0.112) -0.220*** (0.0208) 0.0709*** (0.0122) -0.0822*** (0.0109) -0.105*** (0.00768)

-0.0211*** (0.00236) 0.208*** (0.00709) -0.0831*** (0.0108) 0.0165 (0.0108) 0.135*** (0.0107) 0.0528 (0.0783) 0.0597*** (0.0156) -0.0466*** (0.00996) -0.0244*** (0.00840) -0.0934*** (0.00625)

0.0489*** (0.00854) 0.0777*** (0.0106) -0.00925*** (0.00113) -8.041*** (1.057) 152,534

-0.121*** (0.0241) 0.00805*** (0.00277) 0.00807*** (0.000727) -2.581*** (0.271) 152,534

SC

Individual Variables

RI PT

9.292*** (1.103)

Wald-chi 52952.62 45567.25 12551.64 29628.75 (Prob-chi) (0.0000) (0.0000) (0.0000) (0.0000) Wald test of exogeneity 2.48 6.87 48.94 6.22 (Prob-chi) (0.1155) (0.0088) (0.0000) (0.0127) Robust and clustered standard errors are given in parentheses. *** p<0.01, ** p<0.05, * p<0.1. According to the Politbarometersurvey, political interest is scaled from 1 to 5 (1 equals a very strong interest, 5 equals no interest); economic situation and own economic situation are scaled from 1 to 5 (very good to very bad). Media coverage is instrumented by tonality of reports on international banks.