Till Microtargeting Do Us Part: Measuring the Effects of Political Ads on Facebook

How to cite

Beknazar-Yuzbashev, George; Stalinski, Mateusz. 2020. "Till Microtargeting Do Us Part: Measuring the Effects of Political Ads on Facebook." AEA RCT Registry. AEARCTR-0006703. https://doi.org/10.1257/rct.6703-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

George Beknazar-Yuzbashev* , Mateusz Stalinski

* Primary investigator

Registry ID
AEARCTR-0006703
DOI
https://doi.org/10.1257/rct.6703-1.0
First registered
Last updated
Status
On going
Trial dates
2020-08-31 – 2020-11-15
Countries
United States of America ()
JEL codes
D9, D8
Registry record
http://www.socialscienceregistry.org/trials/6703

Abstract

Microtargeted political ads create innovative opportunities for influencing voter’s behavior, which many scholars brand as a threat to democracy (Borgesius et al 2018, Ribeiro et al 2019, Witzleb et al 2019). To study the effects of microtargeted ads we conduct a field experiment on Facebook in which we ask participants to change settings on ad preferences to see fewer ads on a particular topic. Individuals are randomly assigned one of the two topics: social issues, elections, and politics (treatment group) and alcohol (control group). Furthermore, a subset of individuals installed a browser extension that collects data on ads that they receive on their Facebook feed.

Microtargeted political advertising can affect individual decision-making in a variety of ways. To capture its broad impact, we collect a rich set of outcomes including voter turnout (using voter files), voting preferences, policy views, beliefs about views of others, as well as measures of affective polarization. To inform the ongoing debate whether political ads on social media should be more tightly regulated, we also measure ads’ effect on misinformation.

Intervention

We ask participants to change settings on ad preferences to see fewer ads on a particular topic. Individuals are randomly assigned one of the two topics: social issues, elections, and politics (treatment group) and alcohol (control group). The intervention lasted for 2+ weeks in the run up to the elections. We start collecting outcomes on 5th November 2020.

Experimental design

The study has the following steps:
1. Recruit people on Facebook through Facebook Ads;
2. We incentivize participants to change Facebook ad settings related to see fewer ads on social issues, elections, and politics (treatment) or alcohol (control). Topics are randomly assigned.
3.After 2+ weeks, we conduct the follow up with a survey eliciting outcomes. We will also look at voter file data to elecit whether a person voted in the elections.

Design details

Randomization method
Treatment randomization was done by Qualtrics. To determine the winner of the lottery we will use the following procedure. On November 6 we will go to https://www.marketwatch.com/investing/index/djia. We will use the last digits of the averages as seen at 7 pm EST to fill a random number of the same length as the number of participants minus one digit by digit in the following order: Dow (last digit), S&P (last digit), Nasdaq (last digit), GlobalDow (last digit), Gold (last digit), Oil (last digit), Dow (second to last digit), S&P (second to last digit), Nasdaq (second to last digit), GlobalDow (second to last digit), Gold (second to last digit), Oil (second to last digit). If a particular digit brings the randomly generated number above the number of participants, we skip that digit and move on to the next one. For example, if there are 1000 respondents and we get 384 as the first digits and the last digit of GlobalDow on that day is 3, we skip it and move to the last digit of Gold. If it is also not equal to 0 or 1, we move onto Dow's second to last digit, and so on. If we exhaust the sequence and fail to finish the random number we will continue with the next day's sequence (keeping the numbers that we already generated). For the avoidance of doubt, the "first" participant in this scheme is mapped to 0, the second -- to 1, and so on.
Randomization unit
Individual.
Clusters
1
Observations
Approximately 1000 users.
Arms
Approximately 500 users.
Primary outcomes
- whether the user voted in the 2020 election; - presidential ticket voted for in 2020; - views on policy issues; - affective polarization as measured by the amount donated to a supporter of the opposing party (dictator game); - affective polarization as measured by the feelings thermometer; - prediction about how other respondents (from the opposing party) answered the questions about views on policy issues; - misinformation levels as measured as agreement/disagreement with several statements. We will explore heterogeneity in treatment effects in three dimensions: - partisanship (Democrat/Republican/Independent) and strength of partisanship; - the number of ads shown by different campaigns; - geographic (including analysis of offline advertising in different areas).
Secondary outcomes
- words used to describe supporters of the in- and out-party; - which issue the user considers the most important for the election; - making political donations; - party for the Congress voted for in 2020.

Cited by

No citation data available for this registration yet.

Keywords