Social identity correlates of social media engagement before and after the 2022 Russian invasion of Ukraine

The study examines why certain content gains traction on social media, focusing on the role of social identity during periods of intergroup conflict. Building on social identity theory (SIT) and intergroup emotions theory (IET), the authors explore whether mentions of ingroup and outgroup identities, as well as expressions of ingroup solidarity (ingroup love) and outgroup hostility (outgroup hate), are associated with higher engagement in Ukraine around the 2022 Russian invasion. Prior US-centered research suggests that outgroup cues and moralized emotions often drive engagement, but evidence outside WEIRD contexts and during war is limited. Given the dramatic escalation of Ukraine–Russia conflict in February 2022 and shifts in public sentiment, the study asks how correlates of engagement change from pre- to post-invasion and whether solidarity or hostility better predicts engagement during intense conflict.

Prior work shows platform engagement can be linked to negative emotions, moralized content, and identity cues. Rathje et al. (2021) found outgroup mentions were strong correlates of engagement in US political content. SIT posits identity derives from group memberships, often leading to ingroup–outgroup comparisons; IET extends this by emphasizing how salient group identities shape intergroup emotions and behaviors. Much evidence comes from WEIRD contexts, with an overrepresentation of US studies. US research often finds outgroup negativity more persuasive and engaging than ingroup positivity, aligning with negative partisanship. However, other scholarship theorizes and finds that ingroup-favoring motivations can matter more than outgroup derogation, particularly after shocks or under threat (e.g., rally-around-the-flag effects, post-terrorism solidarity). There is limited research distinguishing ingroup solidarity from outgroup hostility online and little evidence from non-WEIRD settings or during active war, motivating this study in the Ukrainian context.

Design: Three observational studies analyzing Facebook and Twitter (X) content in Ukrainian or Russian from July 2021 to September 2022. - Study 1 (pre-invasion): Generalization of Rathje et al. to Ukrainian context using posts from top pro-Ukrainian and pro-Russian news sources (Facebook and Twitter), posted 12 Jul 2021–24 Feb 2022. Dictionaries counted descriptive mentions of Ukrainian (ingroup) and Russian (outgroup) identities (country, capital, currency, political center, top cities, prominent politicians, with morphological variants). Affective language measured via LIWC (positive/negative affect; Ukrainian 2015, Russian 2007) and translated moral-emotional dictionary (Brady et al.). Mixed-effects linear regressions predicted log-transformed overall engagement (sum of platform-specific reactions; +1 then log) with account as random effect. Controls: follower count, URL presence, media attachments, word count, retweet (Twitter). Estimates reported as exp(β), with two-tailed tests, Satterthwaite d.f., and Cohen’s d. - Study 2 (post-invasion, news sources): Pro-Ukrainian news sources only, 25 Feb 2022–13 Sep 2022, to assess how correlates changed after the invasion. Developed two binary classifiers capturing emotionally charged identity content: ingroup solidarity (positive/affiliative expressions about Ukraine/Ukrainians) and outgroup hostility (derogatory/hostile expressions about Russia/Russians). Labeled 2,000 posts (1,600 train/400 test), fine-tuned multilingual DeBERTa NLI models (following Laurer et al.). Accuracy ~0.87, F1-macro ~0.80. Also constructed domain dictionaries for solidarity and hostility. Sliding 14-day window mixed-effects regressions over time included binary solidarity/hostility indicators, binarized ingroup/outgroup mentions, counts of negative/positive/moral-emotional words, and same controls as Study 1. - Study 3 (post-invasion, non-news Twitter): Geolocated tweets from Ukraine, original non-replies (July 2021–September 2022). A multilingual RoBERTa classifier (trained on 30k labeled posts across platforms; accuracy 0.86, F1-macro 0.81) identified posts as pro-Ukrainian or pro-Russian; analysis restricted to pro-Ukrainian tweets after the invasion (N=148,959). Ingroup solidarity and outgroup hostility classifiers were further fine-tuned on 1,000 pro-Ukrainian tweets (F1-macro 0.75 and 0.82). Mixed-effects regressions predicted log-engagement with the same predictors as Study 2; controls included follower count and verified status (subset analyses confirmed robustness when user metadata were complete). Language detection excluded non-Ukrainian/Russian posts. Robustness checks included alternative dictionaries, binarized group mention indicators, different window sizes, and robust mixed-effects models. All analyses in R 4.3.1.

- Study 1 (pre-invasion, news sources): Outgroup mentions were the strongest correlates of engagement across most datasets, replicating US-based findings. Controlling for covariates, each additional outgroup word predicted approximately 4–23% higher engagement; ingroup words 4–16%. Moral-emotional language increased engagement by ~3–6%, positive language by ~2–5%. Negative language generally increased engagement pre-invasion but showed no significant effect on pro-Russian Twitter. Overall, outgroup mentions had stronger associations than ingroup mentions in three of four datasets. - Change after invasion (Study 2 setup results): After Feb 24, 2022, descriptive identity mentions (ingroup and outgroup) became less strongly associated with engagement on both platforms. For example, on Facebook, outgroup mentions’ association dropped from exp(β)=1.16 (~+16%) to exp(β)=1.07 (~+7%); on Twitter, from exp(β)=1.23 (~+23%) to exp(β)=1.056 (~+6%). Ingroup mentions dropped from exp(β)=1.11 (~+11%) to exp(β)=1.04 (~+4%) on Facebook and from exp(β)=1.16 (~+16%) to exp(β)=1.11 (~+11%) on Twitter. Moral-emotional words and positive words remained associated with +3–7% engagement; negative words had no significant effect post-invasion. - Study 2 (post-invasion, pro-Ukrainian news): Ingroup solidarity became the strongest correlate of engagement. Posts classified as ingroup solidarity were likely to receive +92% engagement on Facebook (exp(β)=1.92; t=151.95; p<0.001; d=0.42) and +68% on Twitter (exp(β)=1.68; t=96.31; p<0.001; d=0.42). Outgroup hostility had a minimal effect on Facebook (+1%, exp(β)=1.01; p=0.001; d=0.01) and no significant effect on Twitter (exp(β)=0.99; p=0.265). Time-series showed solidarity remaining elevated for at least six months; hostility spiked briefly pre-invasion then declined. The share of posts containing solidarity/hostility increased markedly after the invasion (e.g., Facebook: solidarity ~31%, hostility ~35%; Twitter: ~21% and ~25%). - Study 3 (post-invasion, non-news geolocated Twitter): Results generalized beyond news content. Ingroup solidarity predicted +14% engagement (exp(β)=1.14; t=14.85; p<0.001) and outgroup hostility +7% (exp(β)=1.07; t=8.01; p<0.001). Binarized identity mentions were also associated with engagement (outgroup ~+4%, ingroup ~+7%). - Overall: Before the invasion, outgroup mentions best predicted engagement. After the invasion, solidarity with the ingroup became the dominant correlate, while explicit outgroup hostility had weak or null associations, despite becoming more prevalent in content. Negative affect words did not significantly predict engagement post-invasion.

The findings show that correlates of social media engagement are context-dependent. In relatively less acute conflict (pre-invasion), talking about the outgroup strongly predicts engagement, aligning with prior US-based research. During active conflict (post-invasion), expressing ingroup solidarity becomes the most potent correlate of engagement across platforms and datasets, whereas explicit outgroup hostility is weakly associated or null. Intergroup emotions theory offers a potential explanation: salient group identities during threat may elevate affiliative motivations and emotional contagion that encourage group-beneficial behaviors (e.g., engaging with solidarity content). The stability of results across Facebook, Twitter, and beyond news sources suggests the pattern is not solely an artifact of one platform. However, platform algorithms, moderation policies, and self-presentation norms could influence what gains traction. The results imply that emphasizing solidarity may foster higher engagement and potentially a more constructive online atmosphere during conflict, while also highlighting the importance of distinguishing solidarity from hostility in theory and interventions.

This work extends prior research on social media engagement by analyzing a large, non-WEIRD dataset during an active interstate war. The study replicates pre-invasion patterns where outgroup mentions drive engagement and shows a marked shift post-invasion: ingroup solidarity becomes the strongest correlate of engagement, outpacing outgroup hostility across platforms and in both news and non-news data. Contributions include conceptual separation and empirical modeling of solidarity versus hostility, temporal analyses across the invasion breakpoint, and methodological integration of dictionaries with fine-tuned transformer classifiers. Future research should experimentally test causal mechanisms, examine algorithmic influences and moderation, compare across countries and platforms (including Russian platforms such as Telegram/VK), and assess impacts for different populations (politicians, soldiers, activists, civilians). Collaboration with platforms could clarify how recommender systems and moderation interact with identity-related content to shape engagement.

- Correlational design: No causal claims; unobserved confounding possible. - Post hoc classifier development: Solidarity/hostility classifiers and dictionaries were created after data collection; exploratory analyses warrant cautious interpretation. - Platform/algorithm opacity: Unknown or changing recommendation and moderation policies could affect engagement patterns across time and platforms. - Data constraints: Post-invasion Russian Facebook/Twitter data largely unavailable due to bans; limits cross-side comparisons. Some user metadata (followers, verified) missing. - Measurement assumptions: Interpretation relies on emotional contagion and behavior assumptions; engagement may reflect varied motives. Dictionary and model measurement error exists. - Bots and non-representativeness: Potential automated activity and demographic skews (especially Twitter) may bias engagement signals. - Generalizability: Findings may be specific to Ukraine–Russia context and the studied period; external validity to other conflicts or cultures requires testing.