Political polarization in the USA has reached unprecedented levels, surpassing even that seen during the American Civil War. This polarization manifests on two levels: issue polarization among political elites (elected officials, party members, pundits) and affective polarization among voters. Traditional data sources like surveys and voting records struggle to capture the relational aspects of this information diffusion. The rise of social media, particularly Twitter, Facebook, and Reddit, offers a unique opportunity to study how polarized political information spreads through communication networks. Previous research has examined the impact of social media on election outcomes and the spread of disinformation, as well as the role of social media in creating echo chambers and exacerbating polarization. This study leverages a massive dataset of Twitter tweets from the 2016 and 2020 US presidential elections, enriched with political bias classifications, to examine the diffusion dynamics of political content through news media. The focus is on shifts in Twitter's political landscape due to changes in the disseminated news media content and the role of influential users in spreading this information.

A substantial body of research documents the increasing political polarization in the US. Studies highlight partisan division among political and news media organizations and alarming increases in affective polarization among voters. These findings motivate research into the diffusion of polarized political information between influential figures and the broader public. Existing research, however, faces challenges in tracking information diffusion using traditional data, prompting a shift towards leveraging social media data. Several studies have already explored the influence of social media platforms like Twitter and YouTube on election outcomes and the role of disinformation. Others have focused on how echo chambers contribute to polarization on social media platforms.

The study utilizes two datasets of tweets collected from Twitter during the 2016 and 2020 US presidential elections. The 2016 dataset comprises 171 million tweets from 11 million users, while the 2020 dataset contains 702 million tweets from 20 million users, demonstrating a significant increase in user participation. Tweets containing URLs were classified based on the political bias of the linked news media outlet. Classifications were made using data from allsides.com and mediabiasfactcheck.com, categorizing outlets as right, right-leaning, center, left-leaning, left, extreme bias right, extreme bias left, and fake news. The study acknowledges the subjective nature of bias classifications. Influencers were identified using the Collective Influence (CI) approach, focusing on users with a high capacity to spread news within their respective news media category. To analyze changes in user behavior and polarization, the study constructs similarity networks based on retweet patterns among influencers, using a community detection algorithm to identify echo chambers. Latent ideology was inferred using correspondence analysis based on the political actors followed by users. The Hartigan's dip test was used to assess the unimodality of user and influencer ideology distributions. Data cleaning procedures included the removal of tweets from unofficial Twitter clients to minimize the influence of bots.

Proportionally, the fraction of tweets categorized as fake news and extremely biased decreased between 2016 and 2020. However, despite this decline in disinformation, overall polarization increased. The proportion of top influencers affiliated with news media organizations decreased in 2020, while those affiliated with political organizations increased. Analysis of retweet networks revealed an increase in echo chamber behavior, with users less likely to interact with or disseminate information from opposing political viewpoints. New influencers in 2020 exhibited higher levels of polarization compared to those who remained from 2016. Similarity network analysis revealed a stronger separation between communities of left- and right-leaning influencers in 2020 compared to 2016, indicating increased echo chamber effects. Latent ideology analysis, based on the political actors followed by users, confirmed an increase in polarization among both users and influencers between 2016 and 2020. This increase in polarization wasn't solely driven by new users and influencers; even considering only those present in both years, polarization remained significantly higher in 2020. The high correlation between users' latent ideology positions and their news consumption patterns validates the media outlet classifications.

The findings demonstrate that while efforts to combat disinformation may have been somewhat successful in reducing the spread of fake news and extremely biased content, they haven't mitigated the overall increase in political polarization. The shift in influence from media-affiliated to politically affiliated influencers and the intensification of echo chamber effects contribute to the observed polarization. The study's focus on retweets as indicators of endorsement provides a nuanced understanding of user behavior and polarization. The increased polarization isn't solely attributable to the arrival of new users or influencers; changes in user behavior and interactions also played a significant role. The results highlight the complex interplay between disinformation, influencer dynamics, and user behavior in shaping the political discourse on Twitter.

This study provides a comprehensive analysis of political polarization on Twitter during the 2016 and 2020 US presidential elections. The key finding is the simultaneous decrease in disinformation and increase in polarization, highlighting the limitations of simply targeting disinformation as a strategy to reduce polarization. Future research could use natural language processing to analyze the sentiment and topics of tweets, refine user classifications, and extend the analysis to other social media platforms. Further investigation into the mechanisms underlying the observed patterns of polarization is crucial.

The study acknowledges the limitations of using Twitter data, which may not fully represent the broader population's views. The bias classifications used, while drawn from established sources, remain inherently subjective. The study's focus on retweets as an indicator of endorsement may not fully capture the complexity of user interactions. The methodology used for influencer identification and categorization introduces an element of approximation and subjectivity.