Associations between content and engagement vary across social media communities

Abstract

Research on social media engagement often seeks general content features that are associated with online success. However, social media platforms contain communities with different audiences, topics, and norms, and the same features may have different associations with engagement within different communities. This research considers 1,078,044 items from 587 communities in five corpora spanning Reddit, Stack Exchange, X, and YouTube. For each corpus, negativity, moralization, arousal, and in-group versus out-group language were measured. Their relationships with engagement were compared using multilevel models imposing one within-community slope per feature with models allowing slopes to vary across communities. Varying-slope models fit the data better in all five corpora. The estimated community slopes varied substantially in magnitude and, frequently, direction. Nevertheless, the average slopes from the varying-slope models remained small. The improved fit therefore reflects heterogeneity across communities rather than stronger average associations. These findings show that a weak average association can conceal stronger but opposing associations across communities. General claims about the content associated with social media engagement need to account for the specific communities in which that content appears.