• No results found

Enikolopov et al. 2020 - The Polarization Effect of Independent Media - Experimental Evidence from Russia.pdf

N/A
N/A
Protected

Academic year: 2020

Share "Enikolopov et al. 2020 - The Polarization Effect of Independent Media - Experimental Evidence from Russia.pdf"

Copied!
45
0
0

Loading.... (view fulltext now)

Full text

(1)

The Polarization Effect of Independent Media:

Experimental Evidence from Russia

Ruben Enikolopov

New Economic School & Universitat Pompeu Fabra

Michael Rochlitz

University of Bremen

Koen Schoors

Ghent University

Nikita Zakharov

University of Freiburg

August 28, 2020

Abstract

How does access to new information affect voting behavior in an environment where media are largely captured by the state? We answer this question with the help of two randomized controlled trials during the Russian parliamentary elections in September 2016. In the first experiment we randomly select 15 out of 42 otherwise comparable mid-sized Russian cities, treat them with an advertising campaign on social media for the independent Russian online TV channel TV Rain, and complement this treatment with temporary free access to TV Rain. In the second experiment we provide 800 randomly selected respondents from a survey of 1200 respondents with free access to TV Rain, two weeks before the election. In both experiments we find a polarization effect. Government supporters show higher turnout and vote shares for the government party United Russia after being treated with access to TV Rain, whereas government opponents exhibit even lower turnout and vote shares for the government party after the same treatment. Our second experiment permits us to single out the effect on respondents who do not rely on social media for political news. The treatment of these respondents unambiguously leads to lower turnout and a lower vote share for the government party United Russia. The polarization effect, therefore, seems to be conditional on news being received from social media.

Keywords: polarisation, social media, elections, captured media, randomized controlled trial (RCT)

(2)

1

Introduction

As a result of the potent impact of information and mass media on political opinions and behavior, politicians and governments have strong incentives to try and influence, or even control the information accessible to the public. In fact, state control over information and the media is a daily reality for the majority of countries and people in the world. According to the 2017 Freedom House report on the freedom of the press, only 13 per cent of the world’s population enjoys a free press, i.e. “a media environment where coverage of political news is robust, the safety of journalists is guaranteed, state intrusion in media affairs is minimal, and the press is not subject to onerous legal or economic pressures”. 45 per cent of the world’s population, in contrast, lives in countries where the media are characterized as “not free”, with the remaining 42 per cent in countries where the media is "partially free" (seeFreedom House, 2017, p.8) .

An emerging theoretical literature argues that several contemporaneous authoritarian regimes base their legitimacy mainly on the control and manipulation of information through captured media, instead of relying on violence and repression (see e.g. Guriev & Treisman,

(3)

that a measurable and substantial causal persuasion effect of specific information on polit-ical behavior does exist. Accessible information and media reports about politics can have a significant impact on public opinion, support for particular politicians and election results (DellaVigna & Kaplan, 2007; Gerber et al., 2009; Ladd & Lenz, 2009; Enikolopov et al.,

2011; Adena et al., 2015; Peisakhin & Rozenas, 2018) They also play an important role in holding politicians and state institutions accountable for their policies and actions(Besley & Prat,2006;Pande,2011;Banerjee et al.,2011;De Figueiredo et al.,2011;Bruns & Himmler,

2016).

In certain contexts, the availability of specific information can even have effects that go far beyond electoral politics. In a study that uses the mountainous geography of Rwanda to identify the causal effect of radio propaganda on inter-ethnic violence, Yanagizawa-Drott

(2014) shows how 10 per cent of the violence perpetrated during the Rwandan genocide can be attributed to propaganda by the radical radio station RTML. Using a similar research design,DellaVigna et al.(2014) show how the availability of Serbian public radio causes anti-Serbian ethnic animosity in those Croatian border towns that do receive the radio signal.

A relatively new strand of the literature analyzes a causal polarization effect of specific information on political behavior, suggesting that the effect of access to independent media is not as straightforward as the persuasion effect may suggest. Campante & Hojman(2013), for example, show how the introduction of broadcast TV in the US led to lower political polarization. Levendusky (2013) in contrast shows in a lab experiment how partisan media polarize the electorate by rendering relatively extreme citizens even more extreme. Bail et al.(2018) find that conservative Republicans receiving a twitter news-feed with opposing political views become even more conservative, in this way demonstrating how exposure to opposing views on social media may increase rather than decrease political polarization.

(4)

captured by the state. Does the access to independent media lead to a polarization or a persuasion effect? Do social media play a special role in the mechanism? Russia is a perfect testing ground because, although most media are captured by the state, Russia still features a few independent voices that do not have to cater to the wishes of the Kremlin and enjoys an active national social media scene. To find answers to our research questions, therefore, we conduct two randomized controlled experiments during the 2016 Russian parliamentary elections, by randomly providing the treatment group with awareness of and free access to TV Rain, one of the few remaining independent voices in the Russian media landscape.

For our field experiment, we randomly select 15 out of 42 otherwise comparable mid-sized Russian cities, and treat these with an advertising campaign for TV Rain on the Russian social network VKontakte. We complement this treatment with temporary free access to TV Rain, by introducing a free trial subscription based on the IP-identification of online visitors. Specifically, visitors with IP-identities from these 15 cities, and from 5 additional random cities, receive a free trial subscription when accessing the site, while visitors from the remaining 22 cities do not receive such an offer. The treatment is administered three weeks before the elections. After the elections, we then compare the election results in the polling stations of the treatment cities with the results in the polling stations of the control group, relative to the election results in the previous Duma elections.

(5)

highest in 2011, clearly indicating that our treatment leads to polarization of the voters. The implied effect on United Russia votes at the district level ranges from a 3.6 percentage point reduction in votes for United Russia in polling stations with the lowest support for United Russia in 2011, to an 13.2 percentage point increase in votes for United Russia in polling stations with the highest support for United Russia in 2011. These results are in line with a polarization effect of providing access to independent media to partisan supporters of the regime, who are triggered into electoral action upon receiving information that does not coincide with their prior beliefs.

Our survey experiment consists of randomly distributing access to TV Rain at the indi-vidual level with the help of a telephone survey. The telephone survey was carried out by the Levada Center, a well-respected independent Russian polling agency, and consisted of first wave right before and second wave right after the elections. During the first wave, we asked respondents about their voting intentions, and then distributed codes for a 1-month subscription to TV-Rain to a randomly selected sub-group of respondents as a reward for their participation. In the second wave, we asked all the respondents if they voted in the elections, and if yes for which party they voted.

(6)

Our paper makes several important contributions to the literature. While in most other papers the treatment with free media relies on idiosyncratic geographical features (see e.g.Enikolopov et al., 2011;DellaVigna et al.,2014;Yanagizawa-Drott,2014; Peisakhin & Rozenas,2018), we provide two randomized controlled trials that determine the effect of access to new information from independent media on voting behavior. Secondly, we are the first to provide this type of experimental assessment of the effects of access to independent media on voting behavior in a captured media environment. In line with the polarization literature, we show that there exists a polarization effect of exposing partisan voters to in-formation which contradicts their prior beliefs, even in environments with captured media. Finally, we show that this polarization effect only exists for consumers of news on social media, while our treatment with TV Rain had an unambiguously negative effect on votes for United Russia on respondents with more traditional sources of news. In this way, we provide insights on the potential channels and mechanisms behind the polarization effect, evidence on which additional studies can build in the future.

Our paper is organized as follows. Section 2 briefly introduces the Russian context and the role of the media during the 2016 Duma elections. Section 3 describes the theory behind our study, and derives the research hypotheses. Section 4 introduces the research design of our two experiments. Sections 5 and 6 present the data and results of the field and the survey experiments, and provide a discussion of the mechanisms behind our results. Section 7 concludes.

2

The Role of Media Control in Russian Politics

2.1 Media, Political Competition and the Russian State before 2012

(7)

media was driven by the regime’s understanding that those who controlled the media could influence public opinion and elections, and that therefore Russia’s principal media outlets should be controlled by the state.1 It is likely that the Russian government’s experience with the privately-owned TV channel NTV, which had openly campaigned against the pro-Putin party “Unity” during the campaign for the December 1999 Duma elections, contributed to the formation of this point of view. As shown by Enikolopov et al. (2011), the vote share for “Unity” was on average 8.9 per cent lower in cities where NTV was available. When the channel continued to criticize Putin after his March 2000 election victory, it was forcibly bought by the state-owned corporation Gazprom, and has turned into a stable source of pro-government news ever since.

The crack-down on independent media was accompanied by increasingly stringent restric-tions on political competition. While the 1999 parliamentary elecrestric-tions were still genuinely competitive (until shortly before the elections, it remained unclear who would win), the 2003 and 2007 parliamentary elections were already much less so. Step-by-step, democratic institutions were weakened (e.g. the elections of regional governors that were replaced by presidential appointments in 2005 by a change of the constitution), and extra hurdles to gen-uine political competition were established. A sophisticated structure of regional political machines was re-established, to ensure the political dominance of the ruling party United Russia in regional and federal elections (Frye et al.,2014,2019;Rochlitz,2016). The regime’s preoccupation with political control rose to a new high when suspicions of massive electoral fraud during the 2011 Duma elections led to a large wave of popular protests. For a few months during the winter of 2011-2012, street protests seemed to pose a serious threat to the regime, with the apprehension of Russia’s ruling elites amplified by the simultaneous unfolding of the Arab Spring throughout the Middle East. This urged Vladimir Putin to

1In an interview with the director of the radio station Echo Moscow, Vladimir Putin once famously remarked:

“Contrary to a common perception, mass media is an instrument, rather than an institution.” (as cited by

(8)

increase government control over the media even further, after he was reelected as Russian president in May 2012.

2.2 Bent, but not broken: a short history of TV Rain

Shortly after Putin’s re-election, the level of pro-Kremlin coverage was perceptibly scaled up on Russia’s government-controlled TV channels (Pomerantsev, 2015). Simultaneously, a number of new bills were introduced, such as for example a law forcing NGOs that receive funding from abroad to register as “foreign agents”. Control over the media sphere also increased further. The editorial teams of a number of leading investigative newspapers were forced to leave, such as the editors of Gazeta.ru in 2012, or Lenta.ru in 2014. In early 2014, a controversial survey about the siege of Leningrad was taken as a pretext to end the distribution on cable TV of one of Russia’s last independent TV channels, TV Rain. Founded on April 27, 2010, TV Rain had played an important role during the liberal protests against election fraud that took place after the 2011 parliamentary elections. By 2013, the channel had become relatively influential, and reached a potential audience of about 10 million viewers.

Within a couple of days in late January 2014, Russia’s major cable TV providers an-nounced that they would no longer include TV Rain in their programming. Later during the same year, the Russian State Duma issued an advertising ban for Russian cable and satellite TV stations, widely believed to be specifically designed to undercut the financial resources of TV Rain. Finally, in October 2014 the channel was forced to leave its studios in central Moscow on short notice, and had to continue its broadcasts from a private apartment. As a result, the audience of TV Rain fell from several million in early 2014 to about 60.000 at the end of the year, with the channel being only able to survive as a paid online channel from 2015 onward.

(9)

that the channel had become an obstacle in the Kremlin’s strategy to establish tighter control over the informational agenda in the country. Indeed, Russia is particular in the sense that even in late 2015, 85 per cent of the Russian population got their knowledge about Russia and the world mainly from TV, with 60 per cent of Russians watching TV news every day. Controlling the message and content of what is shown on TV is therefore of particular importance to the Russian regime. In times of political protests, economic crisis and foreign tensions, it seemed that the risk of having one last independent TV channel questioning the policies of the government had become too large.

When in the spring of 2016 the editorial team of the business TV channel RBC was forced to leave the channel, Russia lost its last independent cable TV channel, with the informational landscape on cable TV now being fully controlled by the government.

Finally, the day of the Russian parliamentary election was shifted from its traditional date in December to September 18th, 2016. As most Russians usually spend the months of August and early September on holidays at their dachas, this move was expected to make the election campaign short and boring, and to reduce the risk of unrest and agitation before the parliamentary elections.

2.3 The 2016 Duma Elections

(10)

provides a combination of political, economic and cultural news and information, as well as some investigative journalism. These are offered through a variation of formats, including traditional news broadcasts, interviews, talk shows, debates and documentaries. Its content differs from the main state-controlled TV channels, in that it critically covers events inside Russia, including coverage of government-involvement in cases of corruption. In contrast, news on all other, state-controlled TV channels is almost exclusively pro-government, and does not contain any criticism of government institutions. On state-controlled TV, news content is very homogeneous, and news items often seem to be determined by the federal government, albeit in a permutated order.

(11)

Table 1: Principal News Items on Russian State TV and TV Rain, September 15th, 2016

Channel 1 NTV TV Rain

1) Putin calls on the Russian population to vote on September 18th

2) Meeting of the State Council about

Infrastructure Development in Southern Russia

Anti-Corruption NGO FBK reports about Dmitry Medvedev’s datcha, allegedly worth 30 billion rubles

2) Infrastructure

Development in Southern Russia debated during a meeting of the State Council

3) American athletes were doped during the Rio Olympics

Search in a museum in Novorossya was linked to an explosion in St. Petersburg

3) Hacker group “Fancy Bear” publishes information about doping among US athletes

1) Vladimir Putin calls on the Russian population to vote on September 18th

A big inflatable moon was flying through the streets of the Chinese city Fuzhou

4) Ceasefire on the initiative of Lugansk and Donetsk in the Donbass region is observed

6) The US election campaign is a dirty for power, Hillary after health problems comes back

The former customs head of Vnukovo airport hides from the police after corruption charges

5) In Syria the government army is observing the conditions of the ceasefire

4) Ceasefire in the Donbass region

A piece of an airplane found near Tanzania is part of MH370

6) The candidates for the presidential election in the US file their medical history

5) The Russian Ministry of Defense shows live pictures from the siege of Aleppo

Forbes reports that

(12)

3

Theory and Hypotheses

The theoretical literature on persuasion and polarization effects of independent media in captured environments remains divided. On the one hand, Besley & Prat (2006) argue that the impact of independent media in captured environments is likely to be larger than in competitive media markets. Enikolopov et al. (2011) show how access to the independent TV channel NTV during the 1999 Parliamentary elections in Russia increased the combined vote share for the opposition by 6.3 percentage points, while decreasing the vote share for the pro-government party by 8.9 percentage points. Barone et al. (2015) find an effect of similar size for pro-Berlusconi votes once independent digital TV was introduced in Italy, where most TV channels were controlled by Berlusconi. As the magnitude of the media effects measured in Russia and Italy were significantly higher than those found by a similar study for the United States (DellaVigna & Gentzkow, 2010), Enikolopov & Petrova (2015) argue that media effects might indeed be stronger in captured environments. One theoretical mechanism underlying this effect may be that full capture of the mass media creates an informational cascade (see Welch, 1992) that effectively moves individual opinions into the direction desired by the government, but is at the same time vulnerable to the arrival of new information, for example through access to an independent TV channel.

(13)

pro-Russian priors, but was less effective and sometimes even counterproductive for voters with strong pro-Western priors. Along these lines, we would expect that access to a randomly distributed independent source of information like TV Rain would reinforce existing priors of voters with the strongest positive beliefs about the Russian regime, in line with the polarization effect described byBail et al. (2018).

4

Research Design

To test the effect of independent media on voting behavior, we conducted two parallel ran-domized experiments during the 2016 Russian Parliamentary elections. Both experiments provide a randomly selected sub-sample of the Russian population with access to indepen-dent media, in an environment where most media are captured by the state. A control group permits us to test whether the randomly provided access has an effect on political outcomes. The experiments complement each other by focusing on two different levels of observation, with the first experiment providing access at the city level, and the second experiment pro-viding access at the individual level. Comparing the findings of both experiments permits us to determine if the results of our study are robust, and might also hold for the country as a whole.

4.1 Choosing the treatment: Why TV Rain?

For our treatment, we chose the political and entertainment channel TV Rain. As described in Section2, TV Rain was Russia’s last nationally available independent TV channel, until it was taken down from Russian cable TV in early 2014. As a result, TV Rain has been available from 2015 onwards only as a paid online TV channel, with an audience of approximately 40.000 to 60.000 viewers, instead of several million before 2014.

(14)

Could the channel have played a politically relevant role during the 2016 parliamentary elections? Second, the fact that the channel is now only available on the internet and has a paywall for most of its content makes it easy to randomly distribute free access to the channel to a specific sample of the Russian population. In 2016, about 72 per cent of the Russian population used the internet more than once a week, with 26 per cent of the population using the internet as their main source of information (Levada, 2016). This provides us with a large pool of potential participants. The fact that most of the content of TV Rain can only be accessed after paying a fee then makes it possible to randomly distribute free access among a sub-group of Russia’s internet users.

4.2 Choosing the location: Where to conduct the experiment?

In both our experiments, we focus on the population of 42 mid-sized Russian cities, located in the European part of Russia. Cities as a level of observation are convenient for our exper-imental design, due to their high population density, widespread internet penetration, and relatively homogeneous population. We target cities with 100.000 to 500.000 inhabitants, as they are large enough to enable the collection data on their social and economic char-acteristics, but not so large they render treatment prohibitively costly. About 20 per cent of the Russian population lives in cities of this size, and we consider this population to be more representative of the “average” Russian citizen than the population of Moscow or St. Petersburg.

We restrict our sample to the European part of Russia, since in this part of the country the quality of infrastructure and the distance to the capital Moscow are relatively similar and comparable. We define European Russia as those regions that are part of the Central, Southern, North-Western and Volga Federal districts.

(15)

political relationship with Moscow, and often have their own constitution. They usually also feature a more autocratic political system than other Russian regions, which potentially could introduce a bias in surveys on politically sensitive topics (Dininio & Orttung, 2005). By omitting cities that are located in republics, we obtain a sample that is culturally and politically more homogeneous.

Second, and more importantly, electoral fraud is a much bigger concern in republics than in other Russian regions (Myagkov & Ordeshook, 2008). As described more in detail in section 5 below, too high levels of electoral fraud could become a concern for our study, as manipulated election results do no longer represent the actual voting behavior of the population. However, as electoral fraud is also present, if less intense, in other Russian regions (see e.g. Shpilkin, 2011; Enikolopov et al., 2013; Buzin et al., 2016), we further exclude from our sample 4 cities where the vote share for the incumbent party was above 60 per cent during the 2011 Duma elections (i.e. 10 per cent above the average national result and 20 per cent above the average result for non-republic regions). This leaves us with a final sample of 42 cities for the city-level experiment. For the individual-level experiment, we randomly selected 12 cities from this sample of 42 cities, to carry out our survey of 1211 respondents, with about 100 respondents being interviewed per city.

5

City-level Experiment

We start with our randomized controlled field experiment at the city level. From the 42 cities described in section 4.2, we randomly selected 15 cities for treatment with an advertising campaign for TV Rain on the Russian social network VKontakte. We complemented this treatment with free access to TV Rain in the 15 treatment cities, 3 weeks before the Russian parliamentary elections that took place in September 2016.

(16)

advertisement campaign had a measurable effect on voting behavior, we then compare the election results between the 15 cities that received our treatment, and the 22 cities in our control group. Section5.1introduces the technical details of our experimental design, section

5.2 presents the data, section 5.3 the results, and section 5.4 discusses possible mechanisms behind the results we find.

5.1 Experimental Design

Free access to TV Rain was provided with the help of a free trial subscription offer. We use automatic identification of the geo-location to determine if a visitor came from one of the 20 cities that were randomly selected to receive free access. If a visitor was identified to come from one of the selected cities, he or she received a pop-up message on the starting page of the TV Rain website, with an offer to receive a 30 days free subscription after a short registration. This free trial subscription offer was available in the 20 treatment cities from August 27th, 2016 (22 days before the election) until October 10th, 2016 (22 days after the election).

We complemented the free trial subscription with a city-level advertising campaign on the Russian social network VKontakte in 15 of the 20 free trial subscription cities. VKontakte is Russia’s most popular social network, as well as the country’s most popular website. The combination of the free trial and the TV Rain advertisement campaign on VKontakte in 15 randomly selected treatment cities is our main treatment. The other 5 randomly selected cities that received the free trial subscription offer without the advertising campaign permit us to verify whether free access per se was enough to affect voting behavior.

(17)

Table 2: Test of means for cities with advertising campaign and free-trial

Full Sample Treated Control T-test (p-value) Distance to Moscow, km log 6.13 6.12 6.14 0.95 Population, thousands 301.58 298.81 303.16 0.91 Average monthly salary 3.45 3.38 3.48 0.15

Unemployment 0.65 0.62 0.67 0.45

Nightlight intensity, 5 km radius 25.52 23.22 26.83 0.49

VKontakte users 4.46 4.65 4.35 0.32

Turnout in previous election 55.25 54.56 55.64 0.70 Votes for UR in previous election 20.14 18.86 20.86 0.45 Votes for Yabloko in previous election 2.69 2.55 2.77 0.47 Votes for KPRF in previous election 13.66 14.62 13.12 0.25 Votes for LDPR in previous election 7.69 8.01 7.5 0.13 Votes for SR in previous election 9.98 9.36 10.34 0.27 Visitors to TV Rain, pre-treatment 58.24 52.76 61.35 0.34 New visitors to TV Rain, pre-treatment 40.59 36.78 42.76 0.35 Pages viewed per day, pre-treatment 235.96 224.92 242.24 0.67

Observations 4889 1772 3117

Cities 42 15 27

Note: adjusted for clustering, * p < 0.1, ** p < 0.05, *** p < 0.01.

Table 3: Test of means for cities with free-trial only

Full Sample Treated Control T-test (p-value) Distance to Moscow, km log 6.13 5.99 6.15 0.77 Population, thousands 301.58 297.1 301.96 0.95 Average monthly salary 3.45 3.52 3.44 0.48

Unemployment 0.65 0.62 0.65 0.83

Nightlight intensity, 5 km radius 25.52 30.33 25.11 0.22

VKontakte users 4.46 4.12 4.49 0.49

Turnout in previous election 55.25 52.38 55.49 0.16 Votes for UR in previous election 20.14 16.69 20.43 0.17 Votes for Yabloko in previous election 2.69 2.76 2.68 0.89 Votes for KPRF in previous election 13.66 12.64 13.75 0.25 Votes for LDPR in previous election 7.69 8.01 7.5 0.13 Votes for SR in previous election 9.98 12.05 9.81 0.05* Visitors to TV Rain, pre-treatment 58.24 60.99 58.01 0.88 New visitors to TV Rain, pre-treatment 40.59 42.67 40.41 0.87 Pages viewed per day, pre-treatment 235.96 254.23 234.41 0.83

Observations 4889 383 4506

Cities 42 5 37

(18)

2shows the means for variables in the 15 treated and 27 control cities at the level of polling stations. Table 3 presents the t-test of means for cities which only got the free-trial option. As there do not seem to be any statistically significant differences between our treatment and control groups, we can assume that the randomization of the experiment has worked well.

The advertising campaign consisted of advertisement banners for TV Rain, visible to VKontakte users from the selected geo-locations on the left side of their news feed. The banners were complemented with one post in the VKontakte city-community public group of every treated city, posted between September 7th and September 12th, 2016. This post was directly visible in the news feed of everyone who had subscribed to the city-community group of a respective city.

The banners and posts were advertising the free trial option, did not carry any political message, and presented the channel as entertainment media. An example of a used banner, showing the portrait of TV Rain news anchors Pavel Lobkov and Ksenia Sobchak, is displayed in Appendix A.2. Overall, the banners were displayed 20.6 million times to VKontakte users before the election. This translates into an average of 6 banner views per VKontakte user per city. As we allocated equal advertisement budgets for each city, the amount of displays per voter is slightly decreasing with city size, with the biggest cities in our sample only having 5 displays per voter on average. In addition, the amount of banner displays per voter was also affected by the number of VKontakte users per city, their daily activity, and the number of other competing advertisements per city for the target audience.

Our study is an example of encouragement design (Hirano et al.,2000;Duflo et al.,2007;

(19)

Figure 1: Construction of the popularity measurement for the website of TV Rain

Duma elections, relative to the pre-treatment period. If our treatment is indeed found to have increased the viewership of TV Rain, we can then proceed by analyzing whether this increase in the audience of the TV channel has had an effect on political behavior.

To control for a possible increase in the viewership and popularity of TV Rain during the pre-election period, for example as a result of a generally higher interest in political news during the election campaign, we implement a difference-in-differences design, and compare the popularity of TV Rain before the 2016 Duma elections with the popularity of TV Rain before the 2011 Duma elections. To measure the popularity of TV Rain, we use three parameters published by Yandex Metrics (the Russian equivalent of Google Analytics): the percentage change in the total number of pages viewed, the average daily amount of visitors, as well as the number of pages viewed per visit during the treatment period (t) as compared to the 31-day period before the treatment (t-1) (see Figure1for the difference-in-differences design). We gather these measures for the period before the parliamentary elections in 2016 and 2011, respectively, and construct a panel data-set with 42 cities and 2 years. We use OLS with time-demeaned data and a fixed effect for the year of 2016.

Table 4 shows how the advertising campaign in combination with the free-trial option have increased the popularity of TV Rain in the 15 randomly selected treatment cities during the 18 days (1-18 September) before the election, compared to the 31 days (1-31 August) before our treatment started.

(20)

Table 4: Effect of the treatment on the popularity of TV Rain

(1) (2) (3)

Change in percent of Change in percent of Change in percent of Visitors per day, New Visitors per day, Pages viewed per day,

per capita per capita per capita Ads and free-trial 9.061** 10.723** 10.731***

(2.49) (2.44) (2.73)

Free-trial only -2.436 -3.51 -0.032

(-1.21) (-1.63) (-0.01)

Observations 4889 4889 4889

R2 0.19 0.20 0.16

Note: t statistics in parentheses, * p < 0.1, ** p < 0.05, *** p < 0.01.

the amount of new visitors identified as unique visitors over each period of time, and the total amount of pages viewed per day and per 100 000 registered voters in the city. Results at the district level show that our treatment led to a 10.7 per cent increase in the overall number of pages viewed per day, driven by a 10.7 per cent increase in the daily amount of new visitors to the website. Table 4 also provides evidence that the free-trial option increased the popularity of TV Rain only in those cities where simultaneously an advertising campaign on Vkontakte was carried out. These results are robust to using the number of ads per voter, instead of a dummy variable for the cities with free trial and advertisement, as explanatory variable.

5.2 Data

Our main outcome variables are Duma election results for the parliamentary elections in 2016. We gather electoral data from the Central Election Commission of the Russian Federation. Electoral data are available at the level of polling stations, and cover 4889 stations in total. Polling stations are assigned to electoral districts with an average size of 109 000 inhabitants per district, with an average city in our sample having about 2 electoral districts.

(21)

(United Russia) per 100 registered voters, as well as the votes cast for the oldest liberal opposition party Yabloko and all other parties per 100 registered voters. To reduce the risk of distortion through electoral fraud, we follow existing research on Russian elections (see e.g. Klimek et al., 2012; Enikolopov et al., 2013) and exclude from our sample all polling stations with a number of voters below 100. These polling stations are usually special cases such as hospitals or prisons, with a high risk of patients or inmates being coerced to vote for the ruling party.

Our main explanatory variable is the randomized treatment with TV Rain. We control for a variety of factors that could potentially affect voting both at the city and at the district level. To control for the level of political activity within a city, we include voter turnout and elections results for all parties at the 2011 elections as controls. To account for the popularity of TV Rain in a city prior to our treatment, we use the average number of pages viewed per voter in a given city, for the month before the start of our treatment period. In addition, we control for the number of VKontakte users per 100 voters in a city, as the penetration of social networks per se could also potentially affect voting behavior (Enikolopov et al.,

2017). To control for the local socioeconomic context, we include the average monthly salary and the unemployment rate available at the city level, the amount of nightlight emitted in a radius of 5 kilometers around the center of an electoral district in 2015 as a proxy for economic development, and dummies for the administrative region where a city is located.

(22)

to interpret reported violations as a direct signal for electoral fraud, in a managed democracy like Russia reported violations also indicate the presence of active electoral observers, who by their sheer presence may discipline local officials and keep the level of fraud in check (Enikolopov et al., 2013).

5.3 Results

Table 5 presents the basic estimation with a linear model and no controls. We regress dummy variables of our main and secondary treatment on the shares of votes cast for the 7 political parties that got more than 2 per cent of the vote in September 2016 (Table A.1 in the Appendix provides results for all political parties that participated in the election). We find no direct effect of our treatment, i.e. the provision of free access to TV Rain, with or without the advertisement campaign, on turnout and voting. Table 6 presents the same regressions including all controls. The results remain broadly the same.

Table 5: Linear effect of the treatment on voting behavior

Turnout UR KPRF LDPR SR Com Yabloko Ads and free-trial -0.846 0.131 0.727 -0.734 -0.363 0.134 0.007

(-0.25) (0.06) (1.39) (-0.83) (-0.88) (0.78) (0.04) Free-trial only -3.184 -1.254 0.394 -1.16 0.576 -0.194 -0.066 (-1.02) (-0.68) (0.46) (-1.36) (0.93) (-1.26) (-0.26)

Controls No No No No No No No

Observations 4889 4889 4889 4889 4889 4889 4889

R2 0.00 0.00 0.00 0.01 0.01 0.01 0.01

Note: adjusted for clustering, * p < 0.1, ** p < 0.05, *** p < 0.01. UR = United Russia; KPRF = Communist Party of the Russian Federation; LDPR = Liberal Democratic Party of Russia; SR = A Just Russia; Com = Communists of Russia.

(23)

Table 6: Linear effect of the treatment on voting behavior

Turnout UR KPRF LDPR SR Com Yabloko Ads and free-trial -0.029 0.738 -0.133 -0.220 -0.001 0.131 0.041

(-0.25) (0.06) (1.39) (-0.83) (-0.88) (0.78) (0.04) Free-trial only -0.138 0.797 0.957 -1.116** 0.277 -0.024 -0.010 (-0.35) (0.60) (1.38) (-2.29) (0.43) (-0.21) (-0.06)

Controls Yes Yes Yes Yes Yes Yes Yes

Observations 4889 4889 4889 4889 4889 4889 4889

R2 0.87 0.61 0.28 0.47 0.25 0.19 0.16

Note: adjusted for clustering, * p < 0.1, ** p < 0.05, *** p < 0.01. UR = United Russia; KPRF = Communist Party of the Russian Federation; LDPR = Liberal Democratic Party of Russia; SR = A Just Russia; Com = Communists of Russia. Table A.2 in the Appendix provides estimations for all parties, and results for all control variables.

Table 7: Interaction of the treatment with votes for United Russia in 2011 (without controls)

(1) (2) (3) (4) (5) (6) (7)

Turnout UR KPRF LDPR SR Com Yabloko Ads and free-trial -15.836* -14.286** -1.621 -2.920 -0.655 0.384 1.733** (-1.85) (-2.62) (-0.97) (-1.22) (-0.50) (0.74) (2.29) Ads and free-trial 0.801*** 0.789*** 0.125 0.104 0.011 -0.013 -0.093** x UR_2011 (2.71) (3.23) (1.39) (1.01) (0.17) (-0.58) (-2.57) Free-trial only -2.977 -0.399 0.395 -1.578 0.417 -0.191 -0.123 (-0.79) (-0.19) (0.47) (-1.68) (0.60) (-1.08) (-0.52) UR_2011 0.043 0.180* 0.000 -0.088* -0.033* 0.001 -0.012 (0.27) (1.87) (0.01) (-1.96) (-1.85) (0.06) (-1.14)

Controls No No No No No No No

Observations 4889 4889 4889 4889 4889 4889 4889

R2 0.87 0.61 0.28 0.47 0.25 0.19 0.16

(24)

Column 1 suggests that the treatment had a negative effect on turnout, but only in cities with 2011 United Russia support below 19.8 percentage points. The estimated effect on turnout ranges from 4.2 percentage points lower turnout for treated districts with the lowest support for United Russia in 2011, to a peak of 9.3 percentage points higher turnout for treated districts with the highest support for United Russia in 2011, and crosses the zero line when support for United Russia in 2011 reaches 19.8 percentage points. We attribute this result to the polarization effect of independent media on partisan supporters of the regime, who are triggered into electoral action upon receiving information that does not coincide with their prior beliefs.

In column 2 of Table7, we find an almost identical pattern with respect to votes for United Russia: the treatment decreases United Russia votes in districts with previous United Russia support below 18.1 percentage points, while it increases United Russia votes in districts above this threshold. The implied effect on United Russia votes at the polling station level ranges from 3.6 percentage points less United Russia votes in districts with the lowest support for United Russia in 2011, to 13.2 percentage points more votes for United Russia in districts with the highest support for United Russia in 2011. This is in line with a polarization effect with lower United Russia Support in districts with low prior United Russia support, and higher United Russia support in districts with high prior United Russia support (Enikolopov et al., 2011).

(25)

to the polling stations by polarization were also more likely to vote for the ruling party. When looking at the results for the opposition party Yabloko in column 7, we find an opposite interaction effect with almost the same turning point of 18.6 percentage points of electoral support for United Russia. Yabloko was the only significant opposition party present at the 2016 election, and thus it was expected to be particularly affected by our treatment. Although the treatment yielded Yabloko a larger share of the popular vote in cities with low prior United Russia support, it also made Yabloko lose votes to the ruling party in cities with high prior United Russia support, as a consequence of the polarization effect. Columns 3 to 6 do not reveal any other strong pattern. Table8repeats the results of Table 7, but includes all controls. The polarization results are very robust to the inclusion of controls.

Table 8: Interaction of the treatment with votes for United Russia in 2011 (with controls)

(1) (2) (3) (4) (5) (6) (7)

Turnout UR KPRF LDPR SR Com Yabloko Ads and free-trial -16.677** -11.990** -1.710 -3.835** -0.855 0.842** 0.679*

(-2.21) (-2.31) (-1.22) (-2.21) (-0.64) (2.04) (1.94) Ads and free-trial* 0.776** 0.611** 0.082 0.164* 0.040 -0.040** -0.032** UR_2011 (2.30) (2.44) (1.17) (1.74) (0.59) (-2.18) (-2.04) Free-trial only -0.355 0.660 0.824 -1.047 0.277 0.000 -0.049 (-0.10) (0.27) (1.05) (-1.50) (0.44) (0.00) (-0.37) UR_2011 12.677* 7.661 -0.179 3.611*** 0.633 0.292 -0.347 (1.73) (1.65) (-0.15) (3.42) (0.74) (0.77) (-1.23)

Controls Yes Yes Yes Yes Yes Yes Yes

Observations 4889 4889 4889 4889 4889 4889 4889

R2 0.87 0.61 0.28 0.47 0.25 0.19 0.16

Note: adjusted for clustering, * p < 0.1, ** p < 0.05, *** p < 0.01. UR = United Russia; KPRF = Communist Party of the Russian Federation; LDPR = Liberal Democratic Party of Russia; SR = A Just Russia; Com = Communists of Russia. Table A.4 in the Appendix provides estimations for all parties, and results for all control variables.

(26)

Russia traditionally showed a stronger performance, the promotion of a TV channel providing critical information about the government might have increased, rather than decreased, existing positive priors about the performance of the government. In cities where United Russia was less successful in the past, on the other hand, our treatment might have reinforced existing negative priors about the ruling party, thus leading to a reduction in votes for United Russia. An analysis of the comments that were posted under the advertisement banners placed in the city-community groups of VKontakte shows that the percentage of negative comments about TV Rain as a share of all comments in a given city is indeed positively correlated with the vote share obtained by United Russia in 2011,2 suggesting that political polarization might have played a role.

Another potential mechanism might be differences in the salience (Epstein & Segal,

2000) of the message and news content provided by TV Rain. In cities where United Russia performed less well in 2011, the lower election result might be due to some underlying problems that the governing party was unable to address. In other words, reporting on TV Rain about problems related to government mismanagement, deficient infrastructure or government corruption might resonate more with voters in regions where these problems are more prevalent. At the same time, these regions might also feature lower election results for United Russia.

5.4 Ripple Effects

Given the brevity of our treatment period, we find a relatively strong result. One possibility is that our treatment created spillover effects, as treated individuals started to share the free trial subscription offer of TV Rain or what they saw on TV Rain with family and friends.

2We were able to collect the comments posted below the advertisement banners that were placed in the

(27)

Although we cannot observe what people share in general, we are able to observe what they share on VKontakte on their personal walls and on the VKontakte city websites, where they were exposed to the TV Rain banner in the first place.

(28)

Table 9: Ripple effect of the advertisement campaign on VKontakte

Number of posts with reference to TV Rain on VKontakte September 2016 August 1st to August 26th (after treatment, posts on users’ (Placebo, before treatment)

walls and in city groups)

Whole Before After Posts on Posts in Both September elections elections users’ walls city groups

Ads and free-trial 1.093*** 1.407*** 0.924** 0.235 -0.997 -0.290 (0.275) (0.450) (0.366) (0.553) (1.472) (0.557) Free-trial only 1.513* -15.506*** 1.840* -1.002 -15.221*** -1.164**

(0.927) (0.640) (0.967) (1.000) (0.633) (0.582) Posts in August 0.344*** 0.428*** 0.284***

(before August 26) (0.042) (0.054) (0.072)

Posts in July 0.566*** 0.438* 0.304***

(0.190) (0.242) (0.106) Note: * p < 0.1, ** p < 0.05, *** p < 0.01.

6

Individual-level Experiment

To verify whether city-level results, based on real outcomes in a field experiment using social media, are robust at the level of individual respondents, we complement them with an individual-level randomized controlled survey experiment, where we have more control over the individual attributes of respondents, like for example their use of social media. This individual-level experiment will also allow us to disentangle the polarization effect from the salience effect through the inclusion of city fixed effects, since the treatment is not longer at the level of the city, but at the individual level. If we find the same results as in the city-level experiment, they can only be explained by the polarization effect, as the salience effect is absorbed by city fixed effects.

(29)

research organization with a reputation for being independent from the government.

The experimental design of the survey includes two waves of telephone interviews, a first wave before the election, and a second wave after the election. This permits us to control for pre-treatment intentions to vote: in the first wave, respondents were asked whether they plan to vote in the upcoming election and if yes, which party they are planning to vote for. In the second wave, we asked respondents for which party they actually voted, enabling us to determine whether the randomly distributed access to TV Rain had an effect on reported voting behavior. This setting allows to minimize reporting bias, as we only analyze the deviation between self-reported intentions from the first wave and self-reported behavior from the second wave.

6.1 Methodology and Data

The first wave of the survey started 17 days before the election, and was completed 8 days before the election. It was conducted in 12 mid-sized cities in the European part of Russia, which were randomly selected from the 42 cities described in section 4.2. In every city, we conducted a randomized telephone interview of on average 100 people who reported to use the internet at least 2-3 times a week. Our final sample for the first wave consists of 1211 respondents.

During the first wave of the interview, we collected data on personal characteristics of the respondents, such as age, gender, education, income, whether they are employed in the public sector, their satisfaction or dissatisfaction with Russia’s political and economic situation, as well data on personal media consumption and preferred sources of information. We then ask whether respondents were planning to vote, and if yes for which party. At the end of the interview, 66 per cent of our respondents were randomly selected to receive a one-month free subscription to TV Rain. The free subscription was described as a reward for the successful completion of the survey.

(30)

telephone message with a personal twelve digits code and a link to the website of TV Rain, where the code could be activated to start a 30-day free trial subscription. In total, we distributed 806 activation codes for TV Rain. All respondents who received a code later received an additional message via SMS after all the interviews had been conducted, stating that the code would expire within three days, and reminding them to activate the code.

The second wave of the survey was carried out from October 12th to October 18th, 2016. During the second wave, respondents were asked about their actual voting behavior during the election of September 18th, 2016. The response rate in the second wave was about 40 per cent, providing us with a final sample of 483 respondents, including 296 who had received a free subscription code during the first wave, and 187 respondents who were not offered this option.

After the distribution of the subscription codes, respondents had the choice either to comply with our treatment or to ignore it. To estimate the compliance rate, we asked respondents in both waves of the survey about their previous experience with TV Rain. Both in the first and in the second wave, respondents were asked “Do you know anything about the online TV channel TV Rain?” They were then offered three answer options: “Yes, I heard about it and watched it”, “Yes, I heard about it but never watched it”, “No, I do not know anything about it”. The answers from the second wave show that the group of respondents that received our treatment were 10.8 per cent more likely to say that they have watched TV Rain in the past compared to their answers during the first wave, while we observe no difference between the two waves in the control group. This effect can be interpreted as the compliance rate for our treatment.

(31)

efficiency of our treatment to promote TV Rain. We however expect that the overall effect of our treatment might go beyond the number of those that started to watch, since respondents with prior TV Rain experience may have increased their consumption of TV Rain because of the treatment. In our main estimations, we therefore use the randomly offered free trial subscription as our main treatment.

We verify whether our randomization was successful in Table11. The results suggest that randomization worked well, with only two out of 20 variables showing a small, statistically significant difference between treatment and control group. In Table 12, we then also verify whether the randomization worked within the sub-sample of respondents who indicated to use social media as a source of information, and within the sub-sample of respondents where this was not the case. For both sub-samples, the randomization also appears to have been successful.

Table 13shows that offering the treatment did not have an impact on the drop-out rate during the second wave of the survey. Respondents that are relatively older, and those who voted during the 2011 election are however less likely to drop out from the second wave. We therefore control for all these individual characteristics in our regressions.

(32)

Table 10: Summary Statistics: Individual-level experiment

Variable name Description Mean St.dev.

Subscription offer Dummy variable that equals 1 if respondent was offered a code 0.61 0.49 for activation of free trial subscription

Started TV Rain Dummy variable that equals 1 if the respondent reported "having 0.13 0.33 watched TV Rain" in the second wave, but did not report “watching

TV Rain previously” in the first wave

Intention to vote Dummy variable that equals 1 if respondent has reported his or her 0.86 0.35 intention to vote in the upcoming Duma election

Intention to vote Dummy variable that equals 1 if respondent has reported his or her 0.16 0.37 for UR intention to vote for United Russia in the upcoming Duma election

Voted in 2011 Dummy variable that equals 1 if respondent voted in the previous 0.7 0.46 Duma election

Voted for UR Dummy variable that equals 1 if respondent voted for United 0.26 0.44 in 2011 Russia in the previous Duma election

Approval of the Answer to the question: "to what extend does the government 1.7 0.85 government fulfill its obligations?", possible answers range from 0 ("not at

all"), 1 ("mostly not"), 2 ("to some extent") to 3 ("completely")

Female Variable that equals 1 if the respondent is a woman 0.49 0.5 Higher education Dummy variable that equals 1 if respondent reported having a 0.52 0.5

degree in higher education

Age Logarithm of the age of the respondent, measured in full years 3.63 0.34 Income group The variable includes 6 income groups, with group 1 being the 3.57 1.15

poorest and group 6 the richest

State employed Dummy equal to 1 if respondent is working in the public sector 0.23 0.42 Self-employed Dummy equal to 1 if respondent is an entrepreneur or self-employed 0.12 0.32 Retired Dummy that equals 1 if respondent is retired 0.11 0.32 News from Dummy variable indicating that respondent uses federal television 0.78 0.42 Federal TV as a source for news and information

News from Dummy variable indicating that respondent uses cable television 0.44 0.5 Cable TV as a source for news and information

News from Dummy variable indicating that respondent uses newspapers and 0.35 0.48 newspapers journals as a source for news and information

News from Dummy variable indicating that respondent uses social networks 0.56 0.5 social networks as a source for news and information

News from Dummy variable indicating that respondent uses the internet 0.66 0.47 the internet (except social networks) as a source for news and information

News from Dummy variable indicating that respondent uses radio Echo MSK 0.07 0.26 radio EchoMSK as a source for news and information

News from Dummy variable indicating that respondent uses radio Mayak 0.21 0.41 radio Mayak as a source for news and information

(33)

Table 11: T-test of equal means for the full sample (individual-level experiment)

Full Sample Control Treated P-value

Intention to vote 0.86 0.87 0.85 0.53

Intention to vote for UR 0.16 0.17 0.16 0.53

Voted in 2011 0.7 0.7 0.7 0.98

Voted for UR in 2011 0.26 0.28 0.24 0.39

Approval of the government 1.7 1.61 1.75 0.08*

Female 0.49 0.49 0.5 0.92

Higher education 0.52 0.5 0.53 0.53

Age 3.63 3.62 3.63 0.68

Income group 3.57 3.57 3.57 0.97

State employed 0.23 0.22 0.24 0.66

Self-employed 0.12 0.1 0.13 0.44

Pensioner 0.11 0.13 0.1 0.43

News from Federal TV 0.78 0.8 0.76 0.28

News from Cable TV 0.44 0.41 0.45 0.38

News from newspapers 0.35 0.35 0.34 0.79

News from social networks 0.56 0.51 0.59 0.07* News from the internet (excluding social networks) 0.66 0.7 0.64 0.18 News from radio EchoMSK 0.07 0.08 0.06 0.5

News from radio Mayak 0.21 0.2 0.22 0.73

Observations 483

Table 12: T-test of equal means for the sub-samples of respondents who receive / do not receive news from social networks (individual-level experiment)

News No News

from social Control Treated P-value from social Control Treated P-value

networks networks

Intention to vote 0.86 0.86 0.86 0.89 0.86 0.88 0.84 0.44 Intention to vote for UR 0.15 0.15 0.15 0.98 0.18 0.18 0.18 0.96 Voted in 2011 0.67 0.71 0.66 0.42 0.73 0.7 0.76 0.29 Voted for UR in 2011 0.24 0.26 0.23 0.53 0.28 0.29 0.26 0.64 Approval of the gov. 1.77 1.66 1.82 0.13 1.62 1.56 1.65 0.48

Female 0.57 0.53 0.59 0.33 0.4 0.46 0.36 0.17

Higher education 0.46 0.41 0.49 0.2 0.58 0.59 0.58 0.9

Age 3.57 3.56 3.57 0.88 3.7 3.67 3.72 0.3

Income group 3.53 3.48 3.55 0.64 3.62 3.65 3.6 0.71 State employed 0.23 0.26 0.22 0.4 0.23 0.17 0.26 0.12 Self-employed 0.09 0.05 0.11 0.12 0.15 0.15 0.15 0.95

Pensioner 0.11 0.14 0.1 0.32 0.12 0.12 0.12 0.93

News from Federal TV 0.78 0.79 0.78 0.82 0.77 0.82 0.74 0.17 News from Cable TV 0.55 0.55 0.55 0.91 0.29 0.27 0.31 0.59 News from newspapers 0.4 0.44 0.38 0.35 0.27 0.26 0.28 0.75 News from the internet 0.69 0.72 0.67 0.48 0.62 0.67 0.58 0.16 (excluding social networks)

News from radio EchoMSK 0.08 0.08 0.07 0.77 0.06 0.08 0.05 0.43 News from radio Mayak 0.24 0.22 0.26 0.51 0.17 0.18 0.16 0.56

(34)

Table 13: Determinants of sample attrition in wave 2 (probit estimation with average marginal effects as coefficients)

Dep. Variable: Drop out (1 if respondent did not answer the second wave, 0 otherwise)

(1) (2)

Subscription offer -0.018 -0.016

(-0.42) (-0.36)

Intention to vote 0.047

(1.05)

Intention to vote for UR -0.075*

(-1.73)

Voted in 2011 -0.103***

(-2.77)

Voted for UR in 2011 0.041

(1.05)

Approval of the government 0.024

(1.32)

Female 0.051*

(1.70)

Higher education -0.021

(-0.70)

Age -0.128**

(-2.34)

Income group 0.011

(0.85)

State employed 0.021

(0.58)

Self-employed 0.052

(1.13)

Pensioner -0.037

(-0.63)

News from Federal TV 0.01

(0.26)

News from Cable TV 0.051

(1.62)

News from newspapers -0.025

(-0.77)

News from social networks -0.034

(-1.08) News from the internet (excluding social networks) -0.023 (-0.73)

News from radio EchoMSK 0.03

(0.51)

News from radio Mayak 0.019

(0.50)

Day FE Yes Yes

City FE Yes Yes

Observations 483

(35)

being treated with the subscription offer as our main variable of interest in the remainder of this section.

Table 14: Effect of subscription offer on watching TV Rain

Dependent Variable: Started TV Rain

Full Sample News from No news from social networks social networks

(1) (2) (3) (4) (5) (6)

Subscription offer 0.116*** 0.140*** 0.133** 0.205*** 0.157** 0.285*** (2.68) (2.61) (2.20) (2.85) (2.44) (3.09)

Controls No Yes No Yes No Yes

Day FE Yes Yes Yes Yes Yes Yes

City FE Yes Yes Yes Yes Yes Yes

Observations 461 328 256 193 193 125 Note: t statistics in parentheses, * p < 0.1, ** p < 0.05, *** p < 0.01. In columns 2, 4 and 6, we include the full list of variables (as outlined in Table11) as controls.

6.2 Results

The main estimation results for individual turnout levels are presented in Table15. In column 1 we do not find any effect of the treatment on the probability to vote at the election, in line with our earlier findings in the city-level experiment. The inclusion of the interaction effect of our treatment with prior approval of the government is not significant in the full sample (column 2). However, if we limit our sample to respondents who get their news from social networks (column 4), the inclusion of the interaction term yields precisely the same polarization results as in the city-level experiment in Table 6: respondents with the highest government approval (value 3) are 15 percentage points more likely to come to the elections when being treated, while respondents with low levels of government approval are less likely to turn up for the election (14.2 and 28.8 percentage points less likely for prior approval of 0 and 1 accordingly). The group with a moderate level of government approval prior to the election exhibits no effect of the treatment on turnout.

(36)

Table 15: Effect of subscription offer on voting on election day

Dependent Variable: Turnout

Full Sample News from No news from social networks social networks

(1) (2) (3) (4) (5) (6)

Subscription offer -0.052 -0.039 -0.024 -0.288* -0.091 0.180 (-0.77) (-0.34) (-0.25) (-1.74) (-1.04) (1.22) Subscription offer*Approval -0.008 0.146* -0.165** of the government (-0.14) (1.95) (-2.15) Approval of the government 0.047 -0.079 0.177***

(1.07) (-1.38) (2.89)

Controls No No No No No No

Day FE Yes Yes Yes Yes Yes Yes

City FE Yes Yes Yes Yes Yes Yes

Observations 483 460 270 260 213 200

Note: Probit estimation; average marginal effects as coefficients; t statistics in parentheses, * p < 0.1, ** p < 0.05, *** p < 0.01.

from social media. Here, the higher the approval of the government, the more people are discouraged to participate in the election, after receiving our treatment. In other words, the polarization effect seems to be replaced by a persuasion effect among the part of the population that does not get their news from social media. These results also shed some interesting additional light on the results obtained from our city-level experiment, where the treatment on the social network VKontakte also implies an active use of social networks. It is thus possible – although we can not test this directly – that the polarization effect found in the city-level experiment is also conditional on the use of social networks. In our individual-level experiment, the treatment for respondents that do not rely on social networks yields a persuasion effect that is similar to the effect found by Enikolopov et al. (2011).

(37)

unambiguously negative effect of our treatment on the individual likelihood of voting for United Russia. Although the positive sign of the interaction term suggests that this effect becomes weaker with increasing pre-treatment approval of the government, the overall effect never becomes positive, not even for the most ardent pro-Putin respondents, suggesting the negative persuasion effect is mitigated, but never dominated, by pre-treatment levels of government approval.

Table 16: Effect of subscription offer on votes for United Russia

Dependent Variable: Votes for United Russia Full Sample News from No news from

social networks social networks

(1) (2) (3) (4) (5) (6)

Subscription offer -0.134** -0.310*** -0.044 -0.355*** -0.280*** -0.298* (-2.45) (-3.01) (-0.59) (-2.62) (-3.60) (-1.92) Subscription offer*Approval 0.086* 0.145** 0.012

of the government (1.80) (2.40) (0.17)

Approval of the government 0.126*** 0.071* 0.174***

(3.93) (1.75) (3.45)

Controls No No No No No No

Day FE Yes Yes Yes Yes Yes Yes

City FE Yes Yes Yes Yes Yes Yes

Observations 483 460 270 260 213 200

Note: Probit estimation; average marginal effects as coefficients; t statistics in parentheses, * p < 0.1, ** p < 0.05, *** p < 0.01.

(38)

Finally, if we limit the sample to respondents that do not receive news and information through social networks, the polarization effect all but disappears. The interaction term in column 6 is not statistically significant, suggesting that the individual treatment has an unambiguously negative effect on voting for United Russia in this sub-sample, irrespective of prior support of the government. The respondents in the subsample without social media, that is, are immune to the polarization effect and exhibit a pure persuasion effect. Adding controls in Tables17and 18does alter the significance levels, but does not affect the results qualitatively.

Table 17: Effect of subscription offer on voting on election day (with controls)

Dependent Variable: Turnout

Full Sample News from No news from social networks social networks

(1) (2) (3) (4) (5) (6)

Subscription offer -0.056 0.003 -0.042 -0.256* -0.093 0.16 (-0.96) (0.03) (-0.48) (-1.77) (-1.29) (1.44) Subscription offer*Approval -0.037 0.124* -0.161*** of the government (-0.76) (1.89) (-2.60) Approval of the government -0.007 0.015 -0.038 0.006 0.105* (-0.28) (0.36) (-1.08) (-2.26) (0.19) (1.86)

Controls Yes Yes Yes Yes Yes Yes

Day FE Yes Yes Yes Yes Yes Yes

City FE Yes Yes Yes Yes Yes Yes

Observations 460 460 260 260 200 200

(39)

Table 18: Effect of subscription offer on voting for United Russia (with controls)

Dependent Variable: Votes for United Russia Full Sample News from No news from

social networks social networks

(1) (2) (3) (4) (5) (6)

Subscription offer -0.142*** -0.263*** -0.074 -0.331*** -0.283*** -0.214** (-2.75) (-3.19) (-0.99) (-2.88) (-3.82) (-2.14) Subscription offer*Approval 0.061* 0.129*** -0.036

of the government (1.78) (2.92) (-0.76)

Approval of the government 0.072*** 0.042* 0.044* -0.025 0.075** 0.092** (3.76) (1.77) (1.82) (-0.82) (2.55) (2.25)

Controls Yes Yes Yes Yes Yes Yes

Day FE Yes Yes Yes Yes Yes Yes

City FE Yes Yes Yes Yes Yes Yes

Observations 460 460 260 260 200 200

Note: Probit estimation; average marginal effects as coefficients; t statistics in parentheses, * p < 0.1, ** p < 0.05, *** p < 0.01. Table A.6 in the Appendix provides estimates for all control variables.

7

Conclusion

We document the effect of access to new information on an independent TV channel on voting behavior in an environment where most media are captured by the government. Three weeks before the 2016 Russian parliamentary elections, we randomly distributed access to the Russian online TV channel TV Rain to a sample of Russian voters, both at the city and at the individual level.

(40)

have been reinforced instead. We show that such a polarization hypothesis is in line with an analysis of the comments posted under the advertisements for the channel TV Rain on the social network VKontake, which we financed as part of our study.

In the survey experiment, we randomly treat subjects with free access to TV Rain during the first wave of an individual survey conducted 2 weeks before the elections. The second post-election wave of the individual survey reveals that our treatment leads to a polarization effect that is very comparable to what we found in the city-level field experiment for subjects that use social networks as a source of information experience. For other subjects that do not rely on social networks for news, however, the individual TV Rain treatment clearly reduces both turnout and votes for UR, suggesting that the polarization effect is only present for subjects that depend on social networks for their news.

(41)

8

Bibliography

Adena, M., Enikolopov, R., Petrova, M., Santarosa, V., & Zhuravskaya, E. (2015). Radio and the rise of the nazis in prewar germany. The Quarterly Journal of Economics, 130(4), 1885–1939.

Bail, C. A., Argyle, L. P., Brown, T. W., Bumpus, J. P., Chen, H., Hunzaker, M. F., Lee, J., Mann, M., Merhout, F., & Volfovsky, A. (2018). Exposure to opposing views on social media can increase political polarization.Proceedings of the National Academy of Sciences, 115(37), 9216–9221.

Baker, P. & Glasser, S. (2005). Kremlin rising: Vladimir Putin’s Russia and the end of revolution. Simon and Schuster.

Banerjee, A., Kumar, S., Pande, R., & Su, F. (2011). Do informed voters make better choices? experimental evidence from urban india. Unpublished manuscript.

Barone, G., D’Acunto, F., & Narciso, G. (2015). Telecracy: Testing for channels of persua-sion. American Economic Journal: Economic Policy, 7(2), 30–60.

Belmonte, A. & Rochlitz, M. (2019). The political economy of collective memories: Evidence from russian politics. Journal of Economic Behavior & Organization, 168, 229–250.

Besley, T. & Prat, A. (2006). Handcuffs for the grabbing hand? media capture and govern-ment accountability. American economic review, 96(3), 720–736.

Bruns, C. & Himmler, O. (2016). Mass media, instrumental information, and electoral accountability. Journal of Public Economics, 134, 75–84.

(42)

Campante, F. R. & Hojman, D. A. (2013). Media and polarization: Evidence from the introduction of broadcast tv in the united states. Journal of Public Economics, 100, 79–92.

Chen, Y. & Yang, D. Y. (2019). The impact of media censorship: 1984 or brave new world?

American economic review, 109(6), 2294–2332.

De Figueiredo, M. F., Hidalgo, F. D., & Kasahara, Y. (2011). When do voters punish corrupt politicians? experimental evidence from brazil. Unpublished manuscript, UC Berkeley.

DellaVigna, S., Enikolopov, R., Mironova, V., Petrova, M., & Zhuravskaya, E. (2014). Cross-border media and nationalism: Evidence from serbian radio in croatia.American Economic Journal: Applied Economics, 6(3), 103–32.

DellaVigna, S. & Gentzkow, M. (2010). Persuasion: empirical evidence. Annu. Rev. Econ., 2(1), 643–669.

DellaVigna, S. & Kaplan, E. (2007). The fox news effect: Media bias and voting. The Quarterly Journal of Economics, 122(3), 1187–1234.

Dininio, P. & Orttung, R. (2005). Explaining patterns of corruption in the russian regions.

World Politics, 57(4), 500–529.

Ditto, P. H. & Lopez, D. F. (1992). Motivated skepticism: Use of differential decision criteria for preferred and nonpreferred conclusions. Journal of personality and social psychology, 63(4), 568.

Duflo, E., Glennerster, R., & Kremer, M. (2007). Using randomization in development economics research: A toolkit. Handbook of development economics, 4, 3895–3962.

(43)

exper-iment estimate of electoral fraud in russian parliamentary elections. Proceedings of the National Academy of Sciences, 110(2), 448–452.

Enikolopov, R., Makarin, A., Petrova, M., & Polishchuk, L. (2017). Social image, networks, and protest participation. Universitat Pompeu Fabra, Working Paper, March.

Enikolopov, R. & Petrova, M. (2015). Media capture: empirical evidence. In Handbook of media Economics, volume 1 (pp. 687–700). Elsevier.

Enikolopov, R., Petrova, M., & Zhuravskaya, E. (2011). Media and political persuasion: Evidence from russia. American Economic Review, 101(7), 3253–85.

Epstein, L. & Segal, J. A. (2000). Measuring issue salience. American Journal of Political Science, (pp. 66–83).

Freedom House (2017). Freedom of the press 2017.

Frye, T., Reuter, O. J., & Szakonyi, D. (2014). Political machines at work voter mobilization and electoral subversion in the workplace. World politics, 66(2), 195–228.

Frye, T., Reuter, O. J., & Szakonyi, D. (2019). Hitting them with carrots: Voter intimidation and vote buying in russia. British Journal of Political Science, 49(3), 857–881.

Gehlbach, S. & Sonin, K. (2014). Government control of the media. Journal of Public Economics, 118, 163–171.

Gentzkow, M. & Shapiro, J. M. (2006). Media bias and reputation. Journal of political Economy, 114(2), 280–316.

(44)

Guriev, S. & Treisman, D. (2019). Informational autocrats. Journal of Economic Perspec-tives, 32(4), 100–127.

Hirano, K., Imbens, G. W., Rubin, D. B., & Zhou, X.-H. (2000). Assessing the effect of an influenza vaccine in an encouragement design. Biostatistics, 1(1), 69–88.

Klimek, P., Yegorov, Y., Hanel, R., & Thurner, S. (2012). Statistical detection of systematic election irregularities. Proceedings of the National Academy of Sciences, 109(41), 16469– 16473.

Ladd, J. M. & Lenz, G. S. (2009). Exploiting a rare communication shift to document the persuasive power of the news media. American Journal of Political Science, 53(2), 394–410.

Levada (2016). Usage of internet. press release. URL

http://www.levada.ru/2014/11/20/internet-vozmozhnosti-ispolzovaniya/, 2016-11-25.

Levendusky, M. S. (2013). Why do partisan media polarize viewers? American Journal of Political Science, 57(3), 611–623.

Lord, C. G., Ross, L., & Lepper, M. R. (1979). Biased assimilation and attitude polarization: The effects of prior theories on subsequently considered evidence. Journal of personality and social psychology, 37(11), 2098.

Myagkov, M. G. & Ordeshook, P. C. (2008). Russian Elections: An Oxymoron of Democracy. National Council for Eurasian and East European Research Seattle, WA.

Pande, R. (2011). Can informed voters enforce better governance? experiments in low-income democracies. Annual Review of Economics, 3(1), 215–237.

(45)

Pomerantsev, P. (2015). Authoritarianism goes global (ii): The kremlin’s information war.

Journal of Democracy, 26(4), 40–50.

Rochlitz, M. (2016). Political loyalty vs economic performance: Evidence from machine politics in russia’s regions. Higher School of Economics Research Paper No. WP BRP, 34.

Rozenas, A. & Stukal, D. (2019). How autocrats manipulate economic news: Evidence from russia’s state-controlled television. The Journal of Politics, 81(3), 982–996.

Shpilkin, S. (2011). Statistika issledovala vybory: Statisticheskij analiz vyborov v gosdumu 2011 goda pokazyvaet vozmozhnye fal’sifikacii (statistics examined elections: Statistical analysis of elections to the state duma in 2011 shows possible fraud).” gazeta. ru. URL http://www. gazeta. ru/science/2011/12/10 a, 3922390.

Taber, C. S. & Lodge, M. (2006). Motivated skepticism in the evaluation of political beliefs.

American journal of political science, 50(3), 755–769.

Treisman, D. (2018). The New Autocracy: Information, Politics, and Policy in Putin’s Russia. Brookings Institution Press.

Welch, I. (1992). Sequential sales, learning, and cascades. The Journal of finance, 47(2), 695–732.

References

Related documents

In either event, the value of imports from named countries falls by 50–70 per cent over the first three years of protection, and, even if the case is rejected, I find that imports

The uniaxial compressive strengths and tensile strengths of individual shale samples after four hours exposure to water, 2.85x10 -3 M cationic surfactant

When studying the effect of water on nonwoven fiber PLA/NR during 180 days, it was determined that after hydrolytic degradation on the melting thermograms, the melting temperature (T

Modifiable polymers, redox-active polymers, emissive polymers, click chemistry, strained alkyne, strain-promoted alkyne-azide cycloaddtion, reactive carbon-carbon triple

The proposed scheme authenticates both the integrity and the source of the received image by applying the following process on the image in the following

The workshop is based on Microsoft Office Communications Server 2007 Release 2 and Microsoft Exchange Server 2007 SP1 Unified

Possess a CEAS, or standard instructional certificate with an appropriate endorsement to the subject or grade level to be taught Completion of a state approved Special

In addition, we compare two different receiving alternatives ± a reception box in the household and attended reception ± by simulating delivery times and home delivery efficiency.. As