Empirical Exercise 11: Media and persuasion

question, hypothesis, theory

1. Question: Does exposure to independent TV coverage (NTV) during the 1999 Duma election have a causal impact on voting behavior?

2. Hypothesis:  Greater access to NTV reduces support for the pro-government party (Unity) and increases support for opposition parties (OVR, SPS, Yabloko)

3. Theory: Party systems are unstable and ideological preferences are not deeply entrenched, media coverage can shape or sway voter behavior.

   1. State-controlled media can function as propaganda tools, but independent media (like NTV) can counterbalance this by offering alternative narratives and information, potentially affecting political outcomes

setting of empirical

1. Russian election to the Duma in December 1999

2. August 1999: unknown Vladimir Putin appointed Prime Minister

   1. Create new part Unity which was pro Putin

   2. polls to determine voting intentions of parties two months before election in October: Unity came last as it was unknown

3. Actual election results: Unity came 2nd, higher than OVR who was projected to win 

• Two main state-owned channels are pro-Unity

4. Focusing on an NTV: ’Independent TV’

   1. Private commercial channel 

   2. More time devoted to positive coverage of other parties, more critical of Unity and other communist party 

   3. How did NTV broadcast:

      1. Only voters close to the transmitters could watch NTV

      2. In 1995, not yet using broadly - nobody can watch 

      3. In 1999, 75% of population can watch

         1. in 1999, 25% of population no access to alternative views, weren't close to NTV transmitter

data

1. Two cross section datasets: first of 2208 sub regions second one of 1624 individuals

2. Key independent variable: NTV1999, access to NTV coverage

   1. Are people close to one of the 405 NTV transmitters which didn’t cover whole of russia and were randomly placed

   2. Can regard access to NTV coverage as ’random’

is it a natural experiment

-no because you require variable watch in the sub region to be randomly allocated across sub regions

-require that likelihood of inhabitants of sub regions are able to watch NTV is uncorrelated with characteristics of sub regions

• However most of population of russia is concentrated in west eg where cities are

• Siberia in the east, barely any people leave, less transmitters

  • Transmitters are not in areas which are randomly chosen, they are put in areas where there are more people, so will have the max effect

 -regression of NTV1999 on a population size

-NTV1999: dummy for whether there is a transmitter in that sub region, very strongly correlated with Watchsr as whether there is a transmitter is strong determinant of whether you are able to watch NTV

-relationship between NTV 1999 and population size, average wage may not be linear  so including polynomials up to power 5 to reduce OVB as when true relationship is non linear and you only use linear, it creates OVB

-eg very large or small regions may behave diff due to diminishing increasing returns so region with 2x population doesn’t necessarily have 2x chance of having transmitter

-controlling for region fixed-effects, absorb(region) 

-a sub-region was more likely to have a transmitter if it had more population, was richer and it had a city

-cannot conclude that the transmitters were randomly located

-could control for these things in a regression

-regress NTV1999 on the voting behaviour of that sub-region in the previous 1995 election

-running an F test, F test statistic 23 > 3 so it is significant 

-location of transmitters is correlated with voting preferences

-even more of a problem, since voting preferences are the error term of any regression that we would want to run in this exercise

-regressing NTV1999 on population,wage and city dummy and on 1995 voting variables

-controlling for population, wages and the city dummy, voting behaviour in 1995 seems uncorrelated with the location of the transmitters

-insignificant F test statistic

-perhaps conclude that conditional on population, wages and a city dummy, the location of the transmitters is random (i.e. uncorrelated with things affecting voting behaviour in 1999)

what determines whether the population in a sub-region can watch NTV? 

-transmitter in the sub-region 

-transmitters in neighbouring sub-regions

-topography (i.e. mountains) eg signal will not travel very far

-not a dummy as not all areas within a sub-region containing a transmitter could in practice receive a signal

-can create a model to predict whether population can watch creates the variable Watchsr

• Measures likelihood than an inhabitant of this region is able to receive the signal from NTV/watch NTV

population model with watch

-test whether NTV exposure had an effect on the aggregate voting behaviour of a sub-region in the 1999 election

-votejs,1999 is the percentage of the vote received by party j in sub-region s in the 1999 election

-X’s,1999 is a vector of sub-region controls

-λr is a set of region fixed-effects 

-ejs is an idiosyncratic error term: should include things like population and income

extra controls

variables such as the (per capita) number of doctors, nurses, retired citizens etc as controls for observable differences across regions 

-dependent variable: votes for pro-unity/pro putin party

-An increase in the likelihood of receiving the NTV signal of 10 percentage points is associated with a decrease in the likelihood of voting for Unity of 1.55 percentage points

-run regressions of Watch on votes for other parties  eg OVR, SPS, Yabloko, KPRF

-a 10% increase in the likelihood of receiving the NTV signal increases the likelihood of voting for OVR by 0.36%. Similar effects are found for SPS, Yabloko and KPRF

A 10% increase in the likelihood of receiving the NTV signal decreases the likelihood of voting for LDPR by 0.14%.

-A 10% increase in the likelihood of receiving the NTV signal decreases voter turnout by 0.67%.

Why may these estimates not be causal

• Omitted variable bias: it is conceivable to believe that the location of the transmitters may be somehow correlated with the political preferences of voters in the Russian sub-regions. 

  • Eg Soviet authorities decided to locate TV transmitters in less naturally pro-government areas. 

  • If that is the case we will not be estimating the causal effect of NTV but instead the pre-existing differences between ’treated’ and ’control’ sub-regions

• Transmitters were being used by the Channel 4 Ostankino station, which was an educational station. Perhaps exposure to such a station is what affected the political preferences of voters in the different Russian sub-region

  • Maybe places with NTV transmitters already had diff political preferences even before NTV started broadcasting

  • Then NTV exposure would not be true cause of voting change

placebo test

where causal effect should be zero, but that should suffer from the same omitted variable bias as our main test

-using data from 1995 election, when NTV was no yet effectively broadcasting but transmitters already existed 

-location/proximity to a transmitter could not have affected voting behaviour in the 1995 election

-use vote shares of the 1995 election as the dependent variables

-If location of transmitters is correlated with voting preferences, we will find that we can predict the vote shares in 1995 with the location of the transmitters even in the absence of the causal effect of exposure to NTV

-it is Votes_DVR_1995 instead of Votes_DVR_1999: dependent variable has changed

-repeat this for all of other parties as well as voter turnout in 1995

-We can see that the effects are always small and statistically insignificant. This lends validity to our assumption that being able to watch NTV in 1999 is uncorrelated with pre-existing political preferences

using instrumental variables

-whether a person watched NTV or not is not randomly assigned: they choose to watch it and this is influenced by political preferences, education etc

-reverse causality or OVB in unobserved characteristics so Watch would be endogenous and correlated with error terms

-use a second dataset: cross-section of individuals surveyed around the 1999 election, instead of cross-section of subregions

• info on whether individual watched NTV, and availability of NTV

• control for individual-characteristics eg education

• study whether effect of TV exposure is heterogeneous depending how how politically entrenched preferences are

-could use strength of the NTV signal as a IV: reduced form, simply entering the ’instrument’ as a regressor

-or you can use 2SLS

first and 2nd stage populaiton models for 2SLS

-1st stage: regression on the strength of the signal where that individual lives on whether that individual watches NTV

-2nd stage:  use predicted/fitted values of 1st regression and run regression of how that person voted on fitted values

• Watchesi,1999 is a dummy variable: self reported meausre, capturing whether the individual watched NTV in the run up to the 1999 election

• endogenous variable is instrumented with strengthi,1999, which is the NTV signal strength in his sub-region

• dependent variable of interest is voteji,1999, which captures whether individual i voted 2 for party j in the 1999 election.

-regression of Watches on strength: IV, is first stage with additional characteristics eg male, age 

-regression where vote is the dependent variable is the second stage regression for that party

-Access to NTV channel has effect on likelihood of voting for Unity, OVR and SPS: effects are all economically and statistically significant

-However, we do not find the same effects for Yabloko, KPRF, LDPR and voter turnout

undecided vs decided

-In the survey, individuals were asked one month prior to the election whether they had the intention to vote for any particular party and, if so, which party. 

-This allows us to study whether it is the undecided or those with strong political preferences who get affected by exposure to NTV

-first 2SLS regression is if the individuals decide to vote a certain way repeat of the 2SLS we did earlier for watches, and vote_party

-2nd 2SLS where undecided=1, shows citizens that are undecided

-undecided get persuaded to not vote for Unity

-At the 12% significance level the ’decided’ get persuaded to vote for OVR

-Both the decided and undecided get persuaded to vote for SPS, although the effect is stronger for the undecided.

internal validity

-Placebo test (1995 election): Showed that NTV coverage (which didn’t exist effectively in 1995) didn’t predict voting outcomes that year. This suggests that the observed 1999 effects are not due to pre-existing political preferences, lending credibility to the identification strategy.

• Controls: The regressions included region fixed effects and control variables (e.g., population, wages, city dummy), which help address omitted variable bias.

-Non-random placement of transmitters: While not random, after conditioning on covariates (e.g., demographics, economic variables), the placement is likely "as good as random".
-Endogeneity in individual data: Watching NTV is a choice. This is resolved via an instrumental variables (IV) approach, using signal strength as an instrument.

external validity

-Context-specific: The study is set in post-Soviet Russia during a unique transitional political period. Voter volatility and media manipulation are high; findings may not generalize to:

• Mature democracies,

• Countries with stable party systems,

• Media systems with different structures (e.g., digital/internet-dominated).

-Media consumption has drastically evolved since 1999 (e.g., rise of social media and online platforms). Results may not apply to modern electoral environments.