The government intervention effects on panic buying behavior based on online comment data mining: a case study of COVID-19 in Hubei Province, China

The emergence of COVID-19 produced widespread images of empty shelves and heightened public fear, anxiety, and depression, with a minority engaging in extreme purchasing that nonetheless caused significant supply chain disruptions and market disorder. Timely government intervention is essential to mitigate these effects. While prior work shows government intervention is important in disease control, there is a research gap regarding interventions targeting panic buying during early outbreak stages. Many studies rely on surveys, structural models, or simulations that lack immediacy and do not capture authentic public sentiment. This paper addresses this gap by mining netizens’ online comments on government measures to assess effectiveness, identify factors driving outcomes, and understand public perceptions and emotions. The goal is to evaluate the impact of government interventions on panic buying in Hubei via semantic network analysis, sentiment analysis, LDA topic modeling, and multiple regression to inform more effective future strategies.

A broad literature addresses panic buying’s root causes and interventions. Psychological factors—perceived scarcity, uncertainty, fear—drive anxiety and purchase behaviors, potentially creating feedback loops of heightened anxiety (Yuen et al., Omar et al., Taylor; Prentice et al.). Social norms and observational learning amplify perceptions of scarcity (Yuen et al.). External media, especially social media, influences panic buying; expert opinions and official communications can trigger it (Naeem), and information overload increases perceived scarcity and purchase intentions (Islam et al.). Communication patterns during disasters shape behaviors (Arafat et al.). Methods range from correlation analyses, qualitative analyses, and simulations (Lins and Aquino; Bentall; Fu). Interventions are categorized as psychological (e.g., mental health support and resilience-building), market-based (retail strategy changes; system adjustments; food system resilience), and network monitoring/management (emergency management, misinformation control). Evaluations of intervention outcomes include semantic analysis of Twitter aligning measures with panic periods (Prentice et al.), dynamic game models showing government control of panic buying duration (Mao), biopsychosocial frameworks (Rajkumar), and survey-based targeted interventions (Niu). However, these studies often fail to capture public reactions directly and rarely analyze sentiment changes from online comments. Chinese empirical studies confirm both rational and irrational hoarding motives (Wang and Na) and validate modeling of interventions (Fu), as well as links among media exposure, emotional risk, stakeholder perception, protective awareness, and panic buying (Yang et al.). Given the limitations of structured questionnaires and simulations, this study leverages online comments for multi-dimensional, immediate sentiment and need capture. LDA’s broad applicability in communication and AI research (Jelodar; Maier; Yu and Xiang) supports its use here.

Data source and categorization: The study targets Hubei, China during early COVID-19. Official Weibo topics from CCTV News, People’s Daily, and People’s Daily Online were identified using keywords such as “panic buying”, “masks”, “materials”, “authority effect”, “adequate supplies”, “market regulation”, and “active/initiative guidance”. Interventions were categorized into four types: Sufficient materials; Authority effect; Market regulation; Initiative guidance. Fourteen representative Weibo topics were selected, and 84,534 comments were crawled; after cleaning, 80,280 remained (category shares: sufficient materials 44.27%, market regulation 26.96%, initiative guidance 19.52%, authority effect 9.26). Data cleaning removed punctuation, names, non-informative particles, ads, unidentifiable character strings (e.g., “xswl”, “hhh”), URLs/links/@-replies, and translated English to Chinese. Semantic network analysis: ROSTCM5.8 generated semantic networks per intervention category to visualize relationships between targets and opinions (nodes and directed edges), revealing hidden connections. Sentiment analysis: Python implementation using HowNet-based sentiment and degree dictionaries (degree weights: most=2, very=1.5, inverse=-1, etc.) after Jieba word segmentation and stopword removal. Emotions were classified as positive, negative, or neutral, with average sentiment scores computed. User geolocation distributions were examined; Hubei-specific comments (n=1301 after cleaning from 1878) were analyzed separately. LDA topic modeling: Gensim LDA was applied to eight datasets (four categories × positive/negative subsets). Topic-number K was optimized per set by minimizing average inter-topic cosine similarity via an adaptive similarity method: iteratively adjust K until achieving lowest average cosine similarity among topics. Optimal K examples: Initiative guidance (positive K=2; negative K=2 or 6 with 2 preferred due to repetition at 6); Market regulation (positive K=3; negative K=4); Authority effect (positive/negative K=2); Sufficient materials (positive K=3 or 7; negative K=3). Topics were interpreted by high-probability keywords. Multiple regression: To analyze factors affecting public emotional perception (dependent variable Y from sentiment analysis), an evaluation index system was built with four dimensions and eight indicators from Weibo metadata: X1 epidemic severity (ordinal by Wuhan lockdown: pre-2020/01/23=1; 2020/01/23–2020/04/08=3; post-2020/04/08=2); X2 epidemic center area (Hubei=1; non-Hubei=0); X3 number of followers; X4 number of posted Weibo (user blogs); X5 timeliness (avg time difference between topic post and comment); X6 follow-up comments (number of replies received); X7 number of fans; X8 number of likes. Model: Y = b0 + b1X1 + ... + b8X8 + ε. Stepwise regression (SPSS 26) assessed significance with VIF<10 and entry/removal α=0.1; residual normality checks confirmed model validity. Event background: A constructed timeline contextualized interventions and lag effects: announcement of human-to-human transmission to Wuhan lockdown in <3 days (2020/01/23), subsequent panic buying, delayed release of material sufficiency assurances, initiation of market supervision (from ~2020/02/01), and joint guarantee mechanisms (2020/02/24).

- Data composition: After cleaning, 80,280 comments were analyzed; category proportions: Sufficient materials 44.27%; Market regulation 26.96%; Initiative guidance 19.52%; Authority effect 9.26. - Semantic networks and word frequencies: • Initiative guidance: Focus on masks (4281 mentions), hoarding (781), shortage (676); toilet paper ~200; widespread price concerns; instances of miscommunication (e.g., Shuanghuanglian claims) reduced guidance effectiveness. • Market regulation: Masks (6058), price increase (1205), second-hand (342), fake/black-heart; strong support for severe punishment (380) and deserved punishment (372); indicates regulatory gaps beyond pricing. • Authority effect: High attention to authoritative figures—“Zhong Nanshan” (703), “Li Lanjuan” (317); positive descriptors like “lovely” (451), concern for “protection” (378), overall positive trust. • Sufficient materials: Frequent mentions of essential foods—Chinese cabbage (501), rice (332), beef and mutton (434), potato (329); gratitude (“thank you” 461), encouragement (“come on” 321), concern about waste (308). - Sentiment analysis (non-Hubei): Positive highest for sufficient materials (47.55%); lowest for initiative guidance (25.26%). Negative highest for initiative guidance (38.81); lowest for sufficient materials (19.53). Average sentiment: sufficient materials positive score 2.3; strongest negative average scores -2.2 for authority effect and market regulation. - Sentiment analysis (Hubei): Same preference order—Sufficient materials > Authority effect > Market regulation > Initiative guidance. Overall, 56.48% positive toward interventions; higher positive and lower negative averages than non-Hubei. Highest positive average 2.6 (sufficient materials and initiative guidance), negative peak -2.6, reflecting more intense emotions in high-risk areas. - LDA topic insights: • Initiative guidance (positive): trust in official statements, reduced anxiety via rational cognition; (negative): poor independent judgment, misleading official info, rumor-suppression concerns. • Market regulation (positive): timely fines, strong enforcement, execution/punishment; (negative): persistent price increases, affordability issues (N95), logistics failures (orders canceled, delivery difficulty), sanitation checks. • Authority effect (positive): trust in experts (Zhong Nanshan, Li Lanjuan), reassurance, safety; (negative): skepticism about common sense/research, Shuanghuanglian-related concerns. • Sufficient materials (positive): inter-regional support (Inner Mongolia, Henan), national reserves (pork, grains), patriotism and cohesion; (negative): dietary/habit mismatches, hoarding concerns (mold, spoilage), toilet paper hoarding/management issues. - Regression results: • Step 1: Y = -0.885 + 0.323 X1; Adjusted R²=0.601; F=25.073; p<0.001. Epidemic severity (X1) significantly increases positive perception. • Step 2: Y = -0.876 + 0.346 X1 - 0.164 X6; Adjusted R²=0.700; F=15.190; p<0.001. Follow-up comments (X6) negatively associated with sentiment; X1 remains positive and significant. Other variables (X2, X3, X4, X5, X7, X8) not significant. - Temporal dynamics: Comment volume surges in first 3 hours post-report, peaks at 4–5 hours, and stabilizes after ~8 hours—suggesting optimal window for governmental communication/intervention within 4–5 hours. - Overall ranking of intervention effectiveness: Sufficient materials > Authority effect > Market regulation > Initiative guidance. Evidence of government intervention lag during early outbreak stages (material sufficiency assurances followed panic buying). Combining authority effect with material sufficiency yields strongest positive public response.

The findings directly address the research question by quantifying how different government interventions influence public sentiment amid panic buying. Material sufficiency announcements elicited the strongest positive reactions, highlighting the importance of transparent, concrete supply information for calming markets and reducing panic. Authority-driven communications also significantly improved sentiment, underscoring the role of trusted experts in shaping public behavior during uncertainty. Market regulation gained public support but exposed regulatory gaps (e.g., counterfeit/second-hand masks, non-price issues), indicating the need for broader, deeper enforcement beyond pricing. Initiative guidance alone performed weakest, suggesting that non-mandatory appeals without material backing or authoritative endorsement are insufficient, and may polarize sentiment. The regression shows epidemic severity is the dominant driver of public perception, with greater severity amplifying receptivity to decisive government action. Higher follow-up interactions correlate with more negative sentiment, likely reflecting contentious discourse or problem amplification in heavily discussed threads. The temporal pattern of public engagement suggests a critical 4–5 hour window for preemptive messaging to guide opinion before peak discussion, aligning with timely risk communication principles. Strategically, integrating authoritative figures with concrete material sufficiency updates can synergistically reduce panic buying through credibility and assurance. Strengthening market supervision systems (quality control, anti-counterfeit measures, logistics assurance) and proactive, precise communication (myth-busting, clear guidance) can further stabilize public sentiment and market order.

The study demonstrates that during the early COVID-19 outbreak in Hubei: (1) government actions lagged behind events, allowing panic buying to take root before material sufficiency news was disseminated; (2) interventions centered on material sufficiency and the authority effect most effectively improved public sentiment; (3) market regulation was necessary but suffered from inadequate breadth and depth; and (4) public perception of interventions was strongly and positively associated with epidemic severity. Combining authoritative endorsements with evidence of adequate supplies is recommended to maximize impact. Timely messaging within 4–5 hours of report release can preempt negative opinion peaks. These insights can inform more effective, real-time public opinion interventions in future crises.

(1) Data are derived from online comments (Weibo), which, despite large volume (~80,000 cleaned comments), may not represent less digitally active populations and may be subject to platform biases. Future work could integrate objective economic indicators (e.g., retail sales, supply levels) before/after interventions. (2) The analysis focuses on the early epidemic phase; government measures are dynamic across stages, so findings may not generalize to later phases. (3) Internet censorship and filtering may remove some comments, potentially biasing the sample; while large data volume mitigates this, future studies should quantify such effects to enhance robustness.