Populism, Information, and Network Effects

Information leaflets containing accurate information debunking the false claims of a right-wing populist candidate are set to a subset of voters. We will use precinct-level election results and the intensity of treatment to measure the effectiveness of this information campaign both in terms of the direct effect and spillovers.

The 2nd round (runoff) of the presidential election in Argentina will take place on November 19, 2023, and will feature two candidates, Javier Milei (the populist) and Sergio Massa. In the primary election (PASO), Milei's best result was in the province of Salta.

As many populist campaigns, Milei's talking points contain exaggerations and outright misinformation. We have designed leaflets that provide factually true and based on reputable sources information about Milei's proposal about replacement of publicly funded school system to send them a subset of voters in Salta. We are going to observe the effect of this treatment on reported election outcomes.

Argentinian election results are reported on "mesa" level (each school or other polling location / precinct contains several mesas to which people are pre-assigned based on their last name). Therefore, since people voting in different mesas in the same location can live close and interact frequently (for example, because spouses in Spanish-speaking countries rarely share last name), one could expect substantial spillover. The signs of the direct effects and spillovers might be the same or different - they would be the same, for example if treated people pass this information uniformly; they would be different if persuaded people keep this information to themselves whereas people who are unconvinced become motivated to persuade their neighbors to vote differently. To measure both the direct and indirect effects, we varied the share of mesas treated within each polling location.

More specifically, we did the following. We took the polling locations that were located in the province of Salta but outside of the Salta capital city. Of those, we took the roughly half that were delegated to "control group" in our experiment during the 1st round of the presidential election, and which therefore were unaffected by our earlier intervention. From now on, this old "control group" is our new "sample". We then randomized the departments in this sample into treatment and control. We used the first round election outcomes in the mesas located in the "treatment" departments to find ones in "control" departments that are similar. Then we randomized the intensity of intervention (the share of mesas to treat) in different polling locations in each treatment departments. For those mesas we chose to treat, we worked with a local NGO to identify and send envelopes with leaflets to voters that had made themselves known by registering with a local party branch of Partido Justicialista (PJ). The envelopes contained a warning that by opening it, individual would receive political information, as well as WhatsApp number to which they could direct any questions (WhatsApp has 94% penetration in Argentina, exceeding alternatives).

2023-11-17

2023-11-25

We will look at election outcomes at mesa level in treated departments, in both treated and untreated mesas in those departments, and compare them to mesas in control departments to infer both the direct and indirect effect.

Consider a polling location, such as school. It contains several mesas, some are of which are not treated, some are treated, and the treated mesas may vary by the number/share of voters treated just because of data availability. We seek to measure direct effect of treatment (on those treated) and indirect effect (on those voting in the polling location by not treated). We do it the following way.

While by looking at the "not treated" mesas we can get an idea about indirect effects, in a situation where indirect effects can be strong, looking just at the results in treated mesas would be misleading, because these mesas typically include some individuals treated directly and many more individuals not treated directly (but likely treated indirectly). We therefore construct the following variables.

The mesa-level contribution of the direct treatment effect is proportional to the share of individuals treated in a given mesa, which is zero for untreated mesa and a positive value for treated ones. So this is our "direct" variable.

The indirect treatment effect should be thought of as follows. The locality where people live defines the polling place, but people are assigned to mesas by the last name. This means that indirect effect that treated individuals impose on untreated ones (and even on other treated ones) should be roughly the same regardless of whether they happen to vote in the same mesa or different ones ("roughly" is because last names are not randomly assigned, and two individuals with the same last name are more likely to be e.g. father and son than those with different last names, but these differences are likely small). Consequently, an individual's exposure to indirect treatment effect should be proportional to the share of treated individuals in the given locality - i.e., among individuals in the same polling place. So, for our measure of indirect effect we use the share of people treated at the polling location / "school" level. (This measure may be possible to adjust if one believes that directly treated individuals are immune to indirect effects, but we assume that the indirect effect works uniformly, for example, because treated individuals do not know whether a person they are trying to convince was treated or not.)

Notice that our randomization of the share of mesas we treat in a particular polling location allows us to avoid the multicollinearity issue and identify the direct and indirect effects independently.

We will also examine the balance between treated and control mesas across the available set of covariates and will control for misbalanced pre-treatment covariates.

We provide truthful information from reputable about a populist candidate's policy proposal and its consequences to see how this affects voting decisions, to treated voters (direct effect) and to those living nearby (indirect effect).

See "intervention" above.

Randomization is done in office by a computer. We set a seed to make randomization replicable.

We randomize polling locations (typically schools) into control and treatment. Within the latter two groups, we randomized the intensity of treatment (the share of mesas to treat in each such group) subject to our goal of sending ~5,000 leaflets.

Yes

We are treating mesas in 50 polling locations overall, plus there are paired mesas in the control departments, spanning 56 polling locations. This gives 106 clusters (polling locations, typically schools).

We are treating 154 mesas in 50 polling locations and not treating 109 mesas in those locations. The number of control mesas used is 157. Thus, we will have up to 154+109+157 = 420 observations.

154 treated mesas in 50 treated polling locations, 109 untreated mesas in the same 50 locations, 157 untreated mesas in 56 untreated locations.