Polarize and Rule: Independent Media in Autocracy and the Role of Social Media

The general idea of the treatment in both experiments was to provide a random sample of the population with access to an independent TV channel in Russia. In the first experiment, the primary objective of the experimental treatment was to increase the audience of TV Rain in the treated cities. The treatment consisted of two elements. The first element was temporarily waiving the monthly subscription cost of about $7.60 (480 rubles). This cost represented a strong barrier for viewers outside of Russia's main cities, as it corresponded to about 2\% of the average monthly income in Russia's regions at the time. We eliminated the subscription cost by offering a one-month free-trial option to users from 20 randomly selected cities from our sample. For every user, the city of origin was identified with the help of a location-specific IP address. A user from a treated city received a pop-up window with a free trial offer every time they opened the TV Rain website.The free-trial option was introduced for all treated cities on August 26, 24 days before the election.
The second element of our treatment was an online advertising campaign on Russia's most popular social media site, VK, to attract attention to the free trial offer. This campaign was carried out in 15 cities that were randomly selected from the 20 cities that received the free-trial offer. The remaining five cities were used as a placebo to test if the free-trial option by itself without an advertising campaign would not affect viewership and political behavior. The advertising campaign included a banner advertisement next to the user's personal newsfeed, and posted the same banner advertisement in the main newsfeed of city-specific VK groups.
The banner consisted of an image of one of the two main news anchors of TV Rain, as well as the message: ``One-month subscription for free: Only now watch TV Rain for free for the whole month (tvrain.ru).'' The banners next to personal newsfeeds were displayed on average 1.5 million times per city or six times per registered VK user. The banner campaign started on September 1, 2016, 18 days before the elections.

In the second experiment, we treated a randomized subgroup (2/3) of telephone survey respondents (N=1211) with a free subscription offer to TV Rain. The subscription was distributed as an electronic code via SMS, along with a link to activate it.

2016-08-26

2016-09-18

Voting behavior (turnout and electoral support for the incumbent party) in the 2016 Parliamentary election in Russia

Voting Behavior:

– Turnout Support for the Ruling Party (Share of votes cast for the ruling party (United Russia) per 100 registered voters), measured both at the polling station level and through self-reported voting in individual-level surveys. We also study the differential effects based on social media consumption and pre-intervention support for the regime.

2) Media consumption

Media consumption
– Viewership of Independent Media: Changes in TV Rain consumption among treated participants, assessed through survey responses and online activity metrics.

Trial Design
Two field experiments were conducted:

City-Level Experiment: Our first experiment provided a randomly chosen 15 cities out of the full sample of 42 midsize cities in the European part of Russia with free access to TV Rain in combination with a social media campaign advertising this free access. For both elements of the treatment, we use geotargeting. For a placebo purpose, we provide 5 cities only with free access (without advertising). We focus on cities with a population between 100,000 and 500,000 inhabitants, as they are fairly representative of Russia, with about one-fifth of the country's population living in cities of that size. They are also big enough to have the necessary internet infrastructure to access online television.
The per capita audience of TV Rain in these cities, on the other hand, was five times smaller than in Moscow and St. Petersburg, with an average of 3,600 online visitors to the TV Rain website daily per city, or 1.1% of the population. This provides us with a good setting to introduce a new source of information to an audience that previously had no or only very limited exposure to our treatment.
The main concern in selecting the city sample was the prevalence of electoral fraud, which had been continuously increasing since the early 2000s and reached a new high during the 2016 elections. Since electoral fraud was not uniformly distributed across localities, we excluded from our sample cities with exceptionally high levels of fraud in the previous elections. In particular, we first excluded cities from regions with the status of ethnic republics, as these are well known for widespread electoral manipulation. We then also excluded cities where United Russia received more than 60% of the vote in the 2011 Duma elections, which is about two standard deviations above the average vote share in regions that are not ethnic republics and, hence, a strong indicator for high levels of electoral fraud.

Individual-Level Experiment: Conducted via a two-wave survey of 1,211 respondents, this experiment randomized free access to TV Rain to assess heterogeneity in treatment effects between social media users and traditional media consumers. Randomization was done automatically by the call center. In the first wave (September 1-9), respondents were surveyed about their intentions to vote and news consumption preferences. Two of the three respondents received a free trial token (SMS code) to activate their 30-day subscription to the TV Rain channel. In the second wave (October 16-18), we collected self-reported information on voting (turnout and vote for the incumbent party).

The randomization in the city experiment was done in the office by a computer. The randomization in the experiment at the individual level was carried out automatically by the call center.

Randomization in the city-level experiment: cities
Randomization in the experiment at the individual level: individual respondents of the telephone survey

No

Planned number of clusters in the city-level experiment: 15 cities treated with the free trial access and advertising campaign, 5 cities with only free trial access, and 22 cities in the control group. The analysis is at the polling station level (N=4624)
Planned number of clusters in the experiment at the individual level: individual respondents are randomly drawn from a population of 12 cities with about 100 respondents per city (yet note that the randomization is at the individual level).

Planned number of observations in the city-level experiment: 4624 polling stations Planned number of observations in the experiment at the individual level: 1211 respondents in the first wave and 483 respondents in the second wave.

Sample size in the city-level experiment: 15 cities treated with the free trial access and advertising campaign, 5 cities with only free trial access, and 22 cities in the control group. The analysis is at the polling station level (N=4624)
Sample size in the individual-level experiment: 296 respondents were treated with the subscription code, and 187 were not treated (control). In the subsample of respondents who consume news from social media, we have 195 treated and 75 not treated. In the subsample of respondents who do not consume news from social media, we have 121 treated and 92 not treated.

Power in the city-level experiment: the main outcome variable is a rate of votes per 100 registered voters with a standard deviation of 14.43; hence, the absolute minimum detectable effect size (MDE) for the cluster-based experiment is approximately 1.3. Power in the experiment at the individual level: the primary outcome is a vote for United Russia (st.dev. 0.425); hence, we expect a detectable change of about 11.14 percentage points in the outcome variable at a significance level of 5% with 80% power considering our sample size. However, since we further split the sample by consumption of news from social media, we estimate MDE by the subsample. For the subgroup of respondents consuming news from social media, we estimate MDE to equal about 16%, while MDE for respondents who do not consume news from social media equals about 17%.