Algorithmic Drivers of Behavior on Social Media

How to cite

Kalra, Aarushi. 2023. "Algorithmic Drivers of Behavior on Social Media." AEA RCT Registry. AEARCTR-0010933. https://doi.org/10.1257/rct.10933-2.0

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

Aarushi Kalra*

* Primary investigator

Registry ID
AEARCTR-0010933
DOI
https://doi.org/10.1257/rct.10933-2.0
First registered
Last updated
Status
Completed
Trial dates
2023-02-13 – 2023-04-15
Countries
India (South Asia)
Registry record
http://www.socialscienceregistry.org/trials/10933

Abstract

Social media algorithms are an increasing part of our everyday lives, yet little is know about the causal effect of these algorithms on individual well-being. In this project, I study the effect of algorithms on consumer surplus and social welfare. This is done by disentangling the effects of user preferences for and attitudes towards different types of content, from the effects of algorithmic amplification of such content in a large-scale experiment, in cooperation with one of India’s largest social media platforms. I study the causal effect of these algorithms on user engagement with content on the platform, as well as on survey outcomes including users' subjective well-being, and willingness to pay for content customization via algorithms.

Experimental design

The control group consists of a random sample of users on a social media platform who are exposed to a ranked list of posts, where the ranking is determined by the user preferences as they are revealed in their previous engagement and are learnt by the algorithm. Treated users are shown a list of content which is not ranked according to user preferences but are instead exposed to posts that are randomly drawn from a set of 'candidate’ posts. I generate the treatment arms to construct appropriate counterfactuals to algorithmic customization, with varying degrees of content customization by altering different dimensions of the recommendation system.

Design details

Randomization method
Random number generator on Google Cloud Compute.
Randomization unit
neighborhood level in urban areas, and village level in rural areas.
Clusters
20,000 neighborhoods.
Observations
4 million users.
Arms
10,000 neighborhoods in treatment and 10,000 neighborhoods in control.
Primary outcomes
User engagement by content type, defined by genre, hash tags and text analysis of posts. Off-platform outcomes are collected through the survey, and these include, subjective well-being and mental health, digital addiction, attitudes towards out-group members, user perception about changes in content exposure.

Cited by

No citation data available for this registration yet.

Keywords