A large-scale study of US Twitter discussions (Lerman, Feldman, He and Rao, npj Complexity, 2024) found that the emotional tone of reply interactions varies systematically with the network distance between interacting users in the retweet graph: anger, disgust, and toxicity increase with distance while joy decreases, suggesting that affective polarization — the emotional dislike and distrust between political groups — behaves as a structural property of online networks rather than only an in-group/out-group phenomenon. The authors explicitly state that their study does not disentangle the directionality of this relationship and cannot exclude confounders: emotionally charged content is known to be retweeted more, platform feed-ranking algorithms may amplify emotionally salient content from distant users, retweet graphs are a biased proxy for follower ties, and the robustness of the gradient to alternative graph constructions and to classifier noise is unquantified. The open problem is to establish the causal structure of this relationship with credible identification: does network distance causally shape the emotions expressed across political divides, does emotionally salient content reshape network structure, or do both processes co-evolve; and does the gradient genuinely reflect the affective-polarization mechanism rather than topic homophily, algorithmic amplification, or generic negativity correlated with distance.

Background

Affective polarization is the tendency of partisans not merely to disagree on issues but to dislike and distrust members of the opposing party, usually measured in surveys with feeling thermometers (0–100 ratings of how warmly one feels toward the other party). Its behavior on social media is heavily debated: scholars disagree on how much platforms amplify partisan animosity. The source study approached this computationally. From keyword-collected Twitter corpora about the COVID-19 pandemic and the US abortion debate, it scored the emotions of reply interactions (with the Demux classifier, an open-source tool that assigns emotion labels such as anger, disgust, and joy to text, together with clustered emotion models) and their toxicity (with the Detoxify classifier, an open-source model that scores how toxic text is), inferred each user's partisanship with a text-based classifier trained on the political-bias scores of the news-source URLs users share, and measured the network distance between two users both as the shortest path in the retweet graph and as the Euclidean distance in a graph embedding of the retweet graph (a geometric representation produced by the LargeVis embedding method, in which users who retweet overlapping others are placed close together); distances were binned into quartiles. Two findings matter here. First, replies to opposite-ideology users (out-group replies) carried more anger, disgust, and toxicity than replies to same-ideology users (in-group replies), consistent with affective polarization. Second, and more novel, these emotional gradients persisted when interactions were ordered by network distance even without partisan labels: anger, disgust, and toxicity rose with distance and joy fell, across datasets and distance measures — the basis for the claim that affective polarization is a structural property of the network.

Why is the causal reading unresolved? Three strands of prior work bracket the problem from opposite directions. Emotion-to-structure evidence: Brady and colleagues (PNAS, 2017) showed that each additional emotional word in a political tweet increases the pace and scope of its diffusion on Twitter — emotionally charged content is retweeted more, so emotional expression plausibly shapes the retweet network itself. Structure-to-affect evidence: field experiments on platform content, most directly Piccardi and colleagues (Science, 2025), who reranked participants' X feeds during the 2024 US campaign with a language model, showed causal effects of exposure to partisan-animosity content on out-party hostility measured by feeling thermometers. And a classic identification obstacle: Shalizi and Thomas (Sociological Methods & Research, 2011) proved that homophily (similar people connecting) and contagion/influence (connected people becoming similar) are generically confounded in observational social-network data, so the direction of any association between tie structure and expressed emotion cannot be read off a snapshot — exactly the situation of the source study. On top of this, feed-ranking algorithms personalize which distant users' content a user sees, so the exposure process that produces a reply is itself filtered by an unobserved algorithm. Finally, the observational result depends on methodological choices the paper flags but does not vary: the graph was built from retweets (an imperfect sample of follower relations), and emotion, toxicity, and partisanship classifications all carry noise of unquantified impact. The authors' earlier preprint version drew a concrete design implication from the structural reading — disallowing users not directly connected to a poster from replying might reduce the toxicity and negativity of online interactions — which remains an untested prediction of the structure-to-emotion direction.

Problem Statement

For online political interaction networks of the kind studied by Lerman et al. (npj Complexity, 2024) — where reply emotions and toxicity vary with the network distance between interacting users — determine the causal structure of this relationship with a credible identification strategy that survives the confounders the original descriptive study could not rule out. Concretely: (1) identify, with credible empirical identification, whether network distance between users causally influences the emotional tone — anger, disgust, toxicity, joy — of their cross-group interactions, or whether emotionally salient content instead reshapes tie structure (or both, with the co-evolution quantified); the design must control for, or otherwise rule out, topic homophily and algorithmic feed amplification as alternative explanations, in light of the generic homophily-influence confounding established for observational network data; and (2) establish whether the emotion-distance gradient is robust to the choice of graph construction underlying distance (retweet-derived vs follower vs mention graphs) and to the noise of emotion, toxicity, and partisanship classifiers. The solution must also address construct validity: whether the gradient reflects the in-group/out-group mechanism of affective polarization rather than generic negativity or topic structure. These components jointly serve one objective — converting the descriptive claim that affective polarization is a structural property of online networks into an established causal and robust claim.

References

  1. Kristina Lerman, Dan Feldman, Zihao He, Ashwin Rao. Affective polarization and dynamics of information spread in online networks. npj Complexity 1, 2024. DOI: 10.1038/s44260-024-00008-w. https://www.nature.com/articles/s44260-024-00008-w (preprint arXiv:2310.18553, https://arxiv.org/abs/2310.18553)
  2. Buddhika Nettasinghe, Ashwin Rao, Bohan Jiang, Allon G. Percus, Kristina Lerman. In-Group Love, Out-Group Hate: A Framework to Measure Affective Polarization via Contentious Online Discussions. Proceedings of the ACM Web Conference 2025. DOI: 10.1145/3696410.3714935. https://doi.org/10.1145/3696410.3714935
  3. Marilena Hohmann, Michele Coscia. Estimating affective polarization on a social network. PLOS ONE 20(9): e0328210, 2025. DOI: 10.1371/journal.pone.0328210. https://doi.org/10.1371/journal.pone.0328210
  4. Tiziano Piccardi, Martin Saveski, Chenyan Jia, Jeffrey Hancock, Jeanne L. Tsai, Michael S. Bernstein. Reranking partisan animosity in algorithmic social media feeds alters affective polarization. Science, 2025. DOI: 10.1126/science.adu5584. https://doi.org/10.1126/science.adu5584 (preprint arXiv:2411.14652, https://arxiv.org/abs/2411.14652)
  5. Cosma Rohilla Shalizi, Andrew C. Thomas. Homophily and Contagion Are Generically Confounded in Observational Social Network Studies. Sociological Methods & Research, 2011. DOI: 10.1177/0049124111404820. https://doi.org/10.1177/0049124111404820
  6. William J. Brady, Julian A. Wills, John T. Jost, Joshua A. Tucker, Jay J. Van Bavel. Emotion shapes the diffusion of moralized content in social networks. PNAS 114(28), 2017. DOI: 10.1073/pnas.1618923114. https://doi.org/10.1073/pnas.1618923114
  7. Christopher A. Bail, Lisa P. Argyle, Taylor W. Brown, John P. Bumpus, Haohan Chen, M. B. Fallin Hunzaker, Jaemin Lee, Marcus Mann, Friedolin Merhout, Alexander Volfovsky. Exposure to opposing views on social media can increase political polarization. PNAS 115(37), 2018. DOI: 10.1073/pnas.1804840115. https://doi.org/10.1073/pnas.1804840115
  8. Narayani Vedam, Subhayan Mukerjee, Prasanta Bhattacharya. What Contributes to Affective Polarization in Networked Online Environments? Evidence from an Agent-Based Model. arXiv:2504.07610, 2025. https://arxiv.org/abs/2504.07610
  9. Alisson Serracín Morales, Buddhika Nettasinghe. Affective Polarization on Small-World and Scale-Free Networks. arXiv:2603.27845, 2026. https://arxiv.org/abs/2603.27845
  10. Leon Klingborg, Kenneth I. Mavor, Alexander J. Stewart. Strategies for preventing and reversing polarized online discourse. arXiv:2606.18226, 2026. https://arxiv.org/abs/2606.18226