Source-linked AI summary
A Long-Term Analysis of Polarization on Twitter
Kiran Garimella, Ingmar Weber
TL;DR
The paper asks whether political polarization on US Twitter increased over the past eight years, addressing limited long-term evidence. It analyzes 679,000 users across network, retweeting, and hashtag measures, finding a 10%-20% relative increase while noting automation and API constraints.
Problem
Long-term trends in US Twitter polarization have received little analysis, although social media is hypothesized to encourage echo chambers and hinder fact-based debate.
Method
The study analyzes 679,000 users and measures restriction to one political side through following, retweeting, and shared-hashtag behavior.
Results
10%-20%: polarization increased between 2009 and 2016 across all three measures.
Takeaways & Limitations
The findings provide evidence that polarization on Twitter increased over the past eight years and offer a rare long-term perspective.
Takeaways & Limitations
The study did not explicitly detect astroturfing or other automated tweets, and technical “most recent activity only” API limitations hinder longitudinal studies.
Abstract
from arXiv · showhide
Social media has played an important role in shaping political discourse over the last decade. At the same time, it is often perceived to have increased political polarization, thanks to the scale of discussions and their public nature. In this paper, we try to answer the question of whether political polarization in the US on Twitter has increased over the last eight years. We analyze a large longitudinal Twitter dataset of 679,000 users and look at signs of polarization in their (i) network - how people follow political and media accounts, (ii) tweeting behavior - whether they retweet content from both sides, and (iii) content - how partisan the hashtags they use are. Our analysis shows that online polarization has indeed increased over the past eight years and that, depending on the measure, the relative change is 10%-20%. Our study is one of very few with such a long-term perspective, encompassing two US presidential elections and two mid-term elections, providing a rare longitudinal analysis.
Introduction
The paper asks whether political polarization on US Twitter increased over time, addressing limited long-term evidence with an eight-year analysis. It measures polarization across following, retweeting, and hashtag use, finding a 10%-20% increase.
- Social media may encourage echo chambers that reinforce users’ viewpoints and make fact-based debate and consensus harder.
- Long-term trends in social-media polarization have received little analysis, despite substantial research documenting polarization’s existence.
- The study analyzes 679,000 US-politics Twitter users using network structure and tweet content.
- Polarization is defined as restricted engagement with political information on one side of the left-versus-right spectrum.
- The analysis tests whether users became less likely to follow both sides, retweet both sides, or use hashtags shared by both sides.
- 10%-20%: polarization increased across all three measures between 2009 and 2016.
Related Work
Prior work established that polarization exists across online networks, retweet behavior, content, and event-focused discussions. The paper extends this literature by examining long-term political-polarization trends on Twitter.
- Earlier studies found polarized hyperlink structures between opposing blogs and partisan retweeting on Twitter.
- Content research linked controversial issues with biased and emotional language over seven months.
- Short-term studies examined polarization around violence, Hugo Chavez’s death, and gun-violence discussions, including hashtag and homophily measures.
- Andris et al. reported increasing partisanship in the US Congress over several decades, providing a long-term comparison.
- A seven-year Twitter study tracked users, behavior, and the site’s evolution but did not focus on political polarization.
Dataset
The dataset is built by crawling outward from politically labeled politician and media seed accounts. It combines historical political retweets with recent tweets from a sampled set of 679,000 users.
- Users are collected by crawling followers of, and users who retweet, the seed accounts.
- The collection begins with public politician and media accounts whose political leanings are known.
- The seed list covers presidential and vice-presidential candidates, their parties, and popular partisan media outlets.
- 140M users followed at least one seed account, with follow times estimated from Twitter’s reverse-chronological follower ordering and account creation dates.
- Historical seed-politician tweets yielded 1.3M unique political retweeters; a random 50% sample produced 679,000 users and around 2 billion recent tweets.
- Table 1 lists politically leaning US seed accounts, separating political candidates and parties from partisan media outlets.
Experiments
The experiments measure polarization through following and retweeting behavior, plus partisan hashtag use, across time. Across these measures, polarization increased from 2009 to 2016, with hashtag polarization also showing election-linked fluctuations.
- Following and Retweeting Behavior: The study tracks whether users follow or retweet accounts from both political sides, using longitudinal behavioral measures.Following and retweeting experiments assess cross-spectrum engagement over time.
- Following and Retweeting Behavior: A Bayesian model assigns each user a side-leaning estimate from follow or retweet counts and converts deviation from balance into polarization.The polarization measure ranges from 0.0 to 1.0, where larger values indicate greater deviation from balanced leaning.
- Following and Retweeting Behavior: 10-20%: Following and retweeting polarization increased in relative terms between 2009 and 2016 across politician and media seed accounts.Retweeting data are available for political seeds, while following data are available for both political and media seeds.
- Hashtag Polarization: Hashtag polarization measures whether opposing political sides use different hashtags, assigning each hashtag a weekly political leaning.The method corrects for unequal user volumes and applies smoothing to address sparsity.
- Hashtag Polarization: Hashtags used by fewer than five users were excluded, five-week moving averages were computed, and the time series showed a significant non-zero linear trend.The augmented Dickey-Fuller test found non-stationarity (p < 0.0001), and the slope test also reported p < 0.0001.
- Hashtag Polarization: About 20%: Hashtag polarization increased between 2009 and 2016, with election periods generally corresponding to local maxima and post-election periods to local minima.The 2010 midterm pattern is an exception, potentially because estimates were noisier with fewer active users.
Conclusions
The paper addresses conflicting views about whether social media increases, reflects, or reduces offline polarization. Its analysis finds increased Twitter polarization but does not directly settle that broader causal debate.
- Conclusions: Existing evidence conflicts on whether social media increases offline polarization, reflects it, or exposes users to more diverse opinions.These alternatives concern the relationship between online echo chambers and offline political polarization.
- Conclusions: 10%-20%: Across three methodologies, Twitter polarization increased over the eight-year study period.The result potentially reflects increases in offline polarization, but the analysis does not directly establish that relationship.
- Conclusions: The study does not directly settle whether social media causes, mirrors, or reduces offline polarization.Its evidence is limited to documenting increased polarization on Twitter and its possible relation to offline trends.
- Conclusions: Automated tweets and astroturfing were not explicitly detected, while hashtag hijacking could still affect the analysis.The authors note that some suspended accounts may have been removed by Twitter before data collection, but organic hashtag hijacking may remain.
- Conclusions: Longitudinal social-media studies are increasingly possible, but “most recent activity only” API limitations create technical challenges.The paper expects further long-term analyses as social media matures.