Essays on Polarization on Social Media

Abstract

There has been growing concern that social media participation is fueling the rise in polarization and echo chamber formation. This dissertation investigates the role of two innate features of social media platforms in this phenomenon. The first chapter investigates how receiving negative peer feedback, in the form of downvotes, influences content generation on social media. Using data from Reddit, we analyze how feedback affects users’ propensity to post and the intensity of their expressed opinions. We find that both positive and negative feedback increase users’ subsequent posting activity compared to receiving no feedback. However, contrary to concerns that negative feedback drives users with unpopular opinions away, we find no evidence of such an effect. We also find that negative feedback moderates extreme sentiments—users who initially express extreme opinions tend to reduce the intensity of their subsequent posts. These findings have implications for echo chamber formation and polarization. They suggest that negative peer feedback can act as a regulatory mechanism, encouraging users to moderate their tone, potentially mitigating polarization.

The second chapter examines the role of personalized, engagement-optimizing recommendation systems in the polarization of beliefs. In this chapter, we disentangle two pathways to polarization: (a) how information is made available by personalized engagement optimizing recommendation algorithms and (b) how information is processed through confirmatory bias tendencies, i.e., the human tendency to favor belief-consistent information. Additionally, we explore how different engagement metrics—such as dwell time and likes—shape content exposure and belief polarization. Using a novel four-step methodology that combines a survey experiment, structural model estimation, and deep reinforcement learning simulations, we quantify the relative contributions of algorithmic curation and confirmation bias to polarization. Our findings reveal that while confirmation bias exacerbates polarization, engagement metrics play a crucial role in shaping content exposure. Dwell time-maximizing algorithms promote counter-attitudinal content, while like-maximizing algorithms favor belief-consistent content, with the latter intensifying polarization more than threefold compared to the former. Moreover, we find that how information is made available (through personalized algorithm-driven content feeds) escalates polarization significantly more than how information is processed (through confirmatory bias tendencies).

Together, these chapters provide new insights into the drivers of polarization on digital platforms. By examining the role of both peer feedback systems and algorithmic content curation, this dissertation aims to offer a nuanced understanding of how user behavior and platform design interact to shape online discourse. The findings have important implications for the design of social media platforms and shed light on strategies to mitigate polarization while preserving user engagement.