Selective Exposure and Political Polarization: An Online Field Experiment on News Consumption

How to cite

Jo, Donghee. 2019. "Selective Exposure and Political Polarization: An Online Field Experiment on News Consumption." AEA RCT Registry. AEARCTR-0004582. https://doi.org/10.1257/rct.4582-1.0

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

Donghee Jo*

* Primary investigator

Registry ID
AEARCTR-0004582
DOI
https://doi.org/10.1257/rct.4582-1.0
First registered
Last updated
Status
Completed
Trial dates
2016-02-01 – 2019-05-01
Countries
Korea (Republic of) ()
Registry record
http://www.socialscienceregistry.org/trials/4582

Abstract

This study empirically investigates the causal link between news consumers’ self-selective exposure to like-minded partisan media and political polarization. I create a South Korean mobile news application for this study. Users of the app are given access to curated articles on key political issues and were regularly asked about their views on those issues. Some randomly selected users were allowed to choose the news sources from which to read articles; others were given randomly selected articles. To test the mechanism of familiarity, I temporarily boost users' familiarity with some of the news sources during the initial stage of the experiment.

Note that this registry is made after the conclusion of the experiment and its analysis, so that the trial is searchable in the AEA registry. For the pre-analysis plan, see the IRB document for this project (MIT COUHES # 1511295047), attached in this registry. The current working paper is circulated under the title: "Better the Devil You Know: an Online Field Experiment on News Consumption."

Intervention

People read an article every day about a randomly selected issue for approximately 2 weeks. I exploit two sources of randomness in this experiment. First, an article is randomly selected for each user in a subset of the experimental period. Second, each user is assigned to one of the treatment groups. Some randomly selected users were allowed to choose the news sources from which to read articles; others were given randomly selected articles. See "Experimental Design" below for more details.

Experimental design

Upon installation and baseline survey, each user has a five-day pre-exposure period. During the pre-exposure period, a randomly selected article from a randomly selected news source about a randomly selected issue is provided each day. Pre-exposure to randomly chosen news sources during this period serves as an important exogenous variation for the test of mechanisms. This period also serves as a grace period, alleviating the attrition problem by screening out those who were going to drop out early.

After pre-exposure period, each user reads an article about a randomly selected issue each day. For analysis, I include one article reading per issue per user after pre-exposure period, taking into account the rapid dropout of users.

I assign the users who finish the pre-exposure period into one of seven treatment groups. The everyday experience of the treatment groups is slightly different. First—a common step for every treatment group—the issue of the day for each user is randomly selected by the app. The first three groups are allowed to select the news source from which to read an article about the issue of the day. The remaining four groups are given randomly selected articles. When making source selections (G1-G3), the first two groups (G1-G2) can see the names of the news sources while group G3 cannot. The average positions of the sources on the issue at hand are shown only to G2 and G3, and not to G1. For those who are not allowed to select the news source (G4-G7), there is a screen, immediately before reading the randomly chosen article, showing relevant information about the news source that published the article. The information about the news source shown to the readers varies (name only, name and position, position only, nothing). Regardless of the treatment status, everyone can determine the news source at the time of reading the article with a minimal effort because the articles always indicate the source’s name at the end of the text. Most articles also indicate the source’s name at the beginning and in the middle of the text.

Note: unfortunately, the average positions of the sources—shown in the app for G2, G3, G5, and G6—had coding errors unnoticed by the researchers until the very end of the research period: they almost always indicated source positions very close to the center of the scale. Anecdotally, users mostly ignored this scale due to this problem. Therefore, the experiences of G1 and G2 were roughly similar, as were the experiences of G4 and G5, and G6 and G7. Furthermore, the names of the news sources were always salient in the reading screen, making the experience of G4-G7 highly indistinguishable.

Given this, I aggregate the groups into three larger groups to maximize the statistical power, taking into account the similar experiences of the finer subgroups within these larger groups. The Source-Name Group includes G1 and G2—they are allowed to select their news source based on source names. The Source-Position group consists of only G3, which cannot see the names (thus it is harder for them to identify the media that were likely to advocate the supported party’s view) but can see source positions and select based on this information. The remaining groups (G4-G7) are not allowed to choose news sources and were given randomly selected articles. They are merged together as the No-Choice Group. All group comparisons in the paper are based on three larger groups.

Design details

Randomization method
Randomization done by a remote server's random number generation
Randomization unit
Individual for group level comparison, each article reading for position updating and selection results (this is clustered in user level to be conservative)
Clusters
1,500 users for reading session level regressions (position updating pattern analysis), 450 users for comparisons between treatment groups
Observations
Approximately 2,000 for comparisons between treatment groups, approximately 10,000 for reading session level regressions
Arms
100 Source-Name Group, 50 Source-Position Group, 200 No-Choice Group
Primary outcomes
Position on eight policy issues covered in the app. The positions are continuous measures between 0 and 1, and if the reported position is exactly 0 or 1 the position is considered extreme. For users' positions, the app asks the user to “move the scroll bar to denote your view on <IssueName>.” The app provides a rough cardinal benchmark to the positions on the horizontal attitude bar. For comparison between the treatment groups, two main outcome variables are: (i) the distance between prior and posterior positions and (ii) whether the user has an extreme policy view.

Cited by

No citation data available for this registration yet.

Keywords