Affective agenda dynamics on social media: interactions of emotional content posted by the public, government, and media during the COVID-19 pandemic

The COVID-19 pandemic led to unprecedented reliance on social media for information and emotional expression. Prior work shows online emotions such as fear, disappointment, anger, and guilt can affect mental health and behavior, including rumor spread and policy support. This study addresses three questions using data from China’s Sina Weibo during the first six months of the pandemic: (1) What were the main types of emotions posted by the public, government, and traditional media? (2) What emotional trends were observed over time across these actors? (3) What are the temporal relationships among their emotional content? The purposes are to understand the dynamics of emotional transmission on social media during a health crisis and to clarify the emotional relationships among government, public, and media, extending agenda-setting theory to include emotions. Contextually, Weibo is China’s largest social media platform, widely used by the public, government, and traditional media during COVID-19. The Chinese government actively disseminates information to shape public moods and opinions alongside platform censorship, while traditional media leverage social platforms to engage audiences.

Types of emotional content: Research indicates social media use for emotional expression and coping during crises, with a range of emotions observed on Weibo (e.g., fear, disappointment, anger, blessings). Early pandemic stages often saw more negative emotions, including anxiety and fear. Government accounts in China sought to refute rumors and guide sentiment, often emphasizing positive emotions and instrumental support, while the public tended toward sympathy and worry. Traditional media increasingly used emotional storytelling on social platforms. Hypotheses: H1a: The public predominately posted negative emotional content. H1b: The government predominately posted positive emotional content. H1c: The media predominately posted negative emotional content. Trends of emotional content: Prior studies in disasters and early COVID-19 phases show shifts from negative to more positive emotions over time and evolving patterns in both individual and government posts. Gap: limited tracking of traditional media’s emotional trends relative to public and government. Affective agenda dynamics: Agenda-setting theory posits media (and government) influence what and how people think about issues. Emotional content can set an affective agenda via emotional contagion, shaping audiences’ emotions and policy preferences. Hypothesis: H2: Expressions of emotion by government, media, and public interact with each other during the pandemic.

Data source and collection: Using Python 3.7.0 with Scrapy 1.5.1, the study crawled Weibo “super topics” related to COVID-19 using keywords such as pneumonia, COVID-19, and pandemic. Posts with emoticons were included; posts with no substantive content or unrelated to the pandemic were excluded. Public dataset: 67,689 posts. Government and media datasets were built by randomly selecting 72 government and 72 media accounts from Weibo’s Government Index List and Media Matrix Power List, considering communication power, interaction, service, and recognition. Government accounts: two per province (31 provinces) plus 10 central-level accounts; media accounts: 18 each from newspapers, magazines, radio stations, and mainstream media websites. This yielded 36,740 government posts and 126,988 media posts. Timeframe: December 30, 2019 to July 1, 2020, covering the first six months of the pandemic and aligning with typical 2–6 month agenda-setting lags. Preprocessing: Stopword filtering via Python regex and Jieba 0.42.1; keyword categorization. Sentiment/emotion analysis: Combined lexicon- and machine-based approaches. Lexicon: “emotional vocabulary ontology library” from Dalian University of Technology (27,466 emotion keywords; 7 categories, 21 subcategories, following Ekman’s classification). Extracted emotional keywords: public 7,634 (frequency 201,481), government 4,861 (197,741), media 7,188 (774,907). Machine learning: Naive Bayes classifier to predict emotional valence and analyze temporal evolution. Statistical analysis: VAR (vector autoregression) following Stock and Watson (2014) with lag selection by AIC, FPE, HQIC, SBIC; Granger causality tests to assess temporal relationships; impulse response functions and cross-correlations to illustrate sequential influences. Software: Statsmodels 0.10.1 and Stata 14.0.

Emotional distributions: Public posts showed high proportions of negative emotions, including disgust 18.8% (37,939/201,481) and sadness 6.8% (13,808/201,481), supporting H1a. Government posts were dominated by positive emotions: good 51.3% (101,491/197,741), supporting H1b. Media posts had relatively high positive (good 40.6%) and fear (30.8%) proportions; thus H1c was not supported. Additional category proportions from figures: Government—sadness 4.2%, happiness 8.3%, surprise 0.4%, fear 21.6%, disgust 13.7%, anger 0.5%. Media—sadness 4.7%, happiness 8.1%, surprise 0.4%, fear 30.8%, disgust 14.8%, anger 0.5%. Group differences: Chi-square showed significant differences across subjects in emotional categories (χ²(24)=16955.904, p<0.001) with a modest association (Cramer’s V=0.192, p<0.001). Trends over time: Across all three subjects, positive emotions generally exceeded negative emotions, with relatively stable distributions over time. Early in the pandemic (early January 2020), negative emotions were high and volatile across groups, with public negative emotions surpassing 90% at peaks. By mid-to-late January, positive emotions rose markedly and negative emotions fell across all groups, coinciding with government measures and clearer information. From mid-January to early May, government and media emotions were stable with positive >50% and neutral 30–50%; public posts were dominated by neutral 50–60%, followed by positive 30–50%. After early May 2020, public emotions fluctuated with an uptick in negative emotions, potentially linked to worsening global conditions and importation risk. Affective agenda dynamics: VAR and Granger tests rejected no-instantaneous-causality across subjects. Government emotions Granger-caused media emotions (χ²=15.293, p<0.001); media emotions Granger-caused public emotions (χ²=9.687, p=0.008); government emotions Granger-caused public emotions (χ²=14.845, p<0.001). Cross-correlations indicated: public emotions most correlated with media emotions with a 13-day lag (r=0.221); public with government with a lead of 2 days (r=0.223); media with government with a 1-day lag (r=0.223). Overall order of influence: government → media → public, supporting H2.

Findings show positive emotions dominated posts overall, with government and media emphasizing praise, reassurance, trust, and confidence-building content, while the public exhibited more neutral and some negative emotions early on. Government-led emotional messaging appears to shape media tone, which subsequently influences public emotion, consistent with agenda-setting extended to an affective dimension. Early negative emotions likely reflected uncertainty and fear; subsequent stabilization followed government interventions, public health measures, and positive messaging. The public’s higher neutral proportion may reflect both calming over time and effects of platform moderation and censorship reducing negative content. In China’s media environment, government accounts are prominent information sources and drive media alignment, which in turn affects public reactions and behavior, illustrating a government-led emotional agenda during crises.

There are significant differences in emotional expressions across the public, government, and traditional media on Weibo, and these actors influence each other emotionally. Public posts are dominated by neutral emotions, while government and media posts are dominated by positive emotions. Government emotions affect media emotions, which then affect public emotions, evidencing an affective agenda-setting sequence of government → media → public.

(1) Lexicon-based emotion classification may overlook context, so overall text sentiment may not fully align with extracted keywords. (2) Weibo content is reviewed and censored; deleted posts were not captured, and hot search rankings were not purely algorithmic, potentially biasing emotional distributions. (3) The study focuses on inter-subject emotional relationships, not linking emotions to viewpoints or perspectives; future work will examine how emotional shifts in government/media affect public views and reciprocal influences. (4) Granger causality reflects statistical predictiveness rather than true causation; while informative about temporal precedence, it does not establish actual causal mechanisms.