key: cord-1044456-tejubztu
authors: Yang, Yi-xin; Zhang, Ying-ying; Zhang, Xiao-wan; Cao, Yi-han; Zhang, Jie
title: Spatial evolution patterns of public panic on Chinese social networks amidst the COVID-19 pandemic
date: 2022-01-03
journal: Int J Disaster Risk Reduct
DOI: 10.1016/j.ijdrr.2021.102762
sha: f7bf4da7a291e11f01c7716cf8fc71cca70c54c9
doc_id: 1044456
cord_uid: tejubztu

Novel coronavirus pneumonia has had a significant impact on people's lives and psychological health. We developed a stage model to analyse the spatial and temporal distribution of public panic during the two waves of the coronavirus disease 2019 (COVID-19) pandemic. We used tweets with geographic location data from the popular hashtag ‘Lockdown Diary’ recorded from 23 January to April 8, 2020, and ‘Nanjing Outbreak’ recorded from 21 July to 1 September 2021 on Weibo. Combining the lexicon-based sentiment analysis and the grounded theory approach, this panic model could explain people's panic and behavioural responses in different areas at different stages of the pandemic. Next, we used the latent Dirichlet allocation topic model to reconfirm the panic model. The results showed that public sentiments fluctuated strongly in the early stages; in this case, panic and prayers were the dominant sentiments. In terms of spatial distribution, public panic showed hierarchical and neighbourhood diffusion, with highly assertive expressions of sentiment at the outbreak sites, economically developed areas, and areas surrounding the outbreak. Most importantly, we considered that public panic was affected by the 17 specific topics extracted based on the perceived and actual distance of the pandemic, thus stimulating the process of panic from minimal, acute, and mild panic to perceived rationality. Consequently, the public's behavioural responses shifted from delayed, negative, and positive, to rational behavioural responses. This study presents a novel approach to explore public panic from both a time and space perspective and provides some suggestions in response to future pandemics.

Pandemics may not comprise the most lethal diseases, but they can cause long- 3 term mental health problems in individuals. Panic was primarily considered to be 4 equivalent to extreme and groundless fear; subsequently, people began to focus on 5 whether panic would cause irrational behaviours [18] . Based on cognitive emotion 6 theory, panic is a direct result of our cognitive assessment of the external environment 7 [19] . Early empirical studies by sociologists focused on studying human and group 8 behaviour in natural and technological disasters and found that panic was prone to be 9 characterised as irrational [18] . Psychologists also cited the sociological sources of 10 panic and regarded it as a negative psychological state. In their widely acclaimed work, 11 the nature of panic was considered as a negative emotion provoked by external 12 information [20] . And recently, Liu et al. [21] also defined it as 'a complex emotional 13 experience that is a mixture of several generic negative emotions'. Reducing public 14 panic is essential for governments to adopt preventive measures [22] . Hence, in this 15 present study, we assume that panic is an extreme negative emotion which contains fear 16 and anxiety, and that it may cause irrational behaviours. 17 The focus of panic has recently shifted to two themes: psychological disorders, 18 such as the characterisation and treatment of panic attacks and panic disorders [23, 24] , 19 and economic impacts, such as panic buying [25, 26] and volatility of the stock market 20 [27] . Several studies have considered the importance of panic during the COVID-19 21 pandemic. Liu et al. [28] and Wu et al. [22] discussed the relationship between 22 information disclosure and panic, and concluded that pandemic information released 23 by social media can reduce citizens' panic; Parry [29] explored the influence of 24 abandoning pets owing to panic. These studies showed that high levels of panic and learning methods to analyse sentiments based on social sensing data. Few studies have 23 performed fine-grained sentiment analysis, primarily focusing on negative sentiments 24 during pandemics. Therefore, research gaps exist in the analysis of sentiment analysis 25 and the spatial distribution of panic emotions. 26 With the emergence of computer science and sentiment analysis, several sentiment 27 classification technologies have been developed, such as the lexicon-based approach, 28 machine learning-based approach, and hybrid-based approach [47] . Behl et al. [48] used 29 supervised machine learning approaches and multi-class classification to analyse 30 Twitter data on COVID-19 and found the performance of deep learning algorithms 31 superior among the tested algorithms. Although mainstream research uses the machine 32 learning-based method frequently, the method suffers from insufficient precision and 33 unclear sentiments. When responding to pandemics, it is not enough to study polarity 34 sentiments; changes and patterns of specific sentiments deserve increased attention. the LDA topic model and lexicon-based approach sentiment analysis tool to analyse 1 patients' comments on physician rating websites and found that negative emotions (fear, 2 anger, and sadness) frequently appeared during the COVID-19 pandemic. In the current 3 study, we also chose a hybrid-based approach to achieve fine-grained sentiment 4 analysis using natural language processing techniques to perform data pre-processing 5 and determine semantic relations. trending issues from documents, is widely used for long document analysis. With the 11 development of large-scale text-processing techniques, it is gradually being applied to 12 short texts, such as tweets. Studies employing the LDA topic model in social media 13 have surged (e.g. sentiment analysis of pandemics [51] and product opinions [52] ). 14 However, it is mainly useful for topic classification and extraction, and it hardly 15 explores inter-topic mechanisms. In contrast, grounded theory is beneficial when there 16 are no or insufficient theories regarding a particular phenomenon [53] . Charmaz [54] 17 defined grounded theory as a systematic, inductive, and comparative approach to 18 conducting an enquiry to construct theory. Zhang et al. [55] used grounded theory to 19 construct a theoretical model of crowdsourcing in tourism and explored its mechanisms 20 through semi-structured interviews. Although it is important to construct theoretical 21 models, grounded theory is not very applicable to online texts because of the limitation 22 of massive data. Previous studies have used the grounded theory approach for topic 23 analysis and construction of theoretical frameworks for review comments [56] and 24 social media opinions [57] . Hence, in the current study, we combined the two 25 approaches for panic mechanism analysis, first exploring the conceptual model of panic 26 evolution using grounded theory, and then, validating the model using the LDA topic 27 model. 30 As the first country to report the COVID-19 pandemic, China experienced a large-31 scale first-round outbreak in January 2020, which caused a state of panic among its 32 people. Owing to effective prevention measures, China was the first to return to 33 normality. However, as the coronavirus mutation escalated, a second global outbreak occurred. Despite stringent prevention measures and widespread vaccination, there 1 have also been small localised outbreaks in China. Furthermore, the Nanjing outbreak 2 in July 2021, which was affected by the Delta Coronavirus, was the largest number of 3 confirmed cases to date. During the outbreaks, significant changes have taken place in 4 people's lives and have affected their emotions. We chose these two outbreaks as cases 5 to explore the evolution of panic.

Weibo is a popular social media platform in China, like the Chinese version of 7 Twitter. In the first round of the outbreak, we searched for tweets posted between 23 8 January 2020 and 8 April 2020 using the popular hashtag 'Lockdown Diary'. We used 9 Python 3.7.0 software combined with a crawler toolkit to obtain 33,611 original tweets 10 and recorded content, posting time, username, location, and other relevant information. 11 In the second round of the outbreak, we searched the hashtag 'Nanjing Outbreak' increasing scale with 1 being the lowest and 9 being the highest. In this corpus of 2 'Lockdown Diary', we deleted the surprise category, as Weibo users seldom mentioned 3 such terms and classified anxiety in the panic category. Furthermore, 67 new emotional 4 words extracted by manual coding were added to the subsequent analysis. An example 5 of the special sentiment lexicon of the COVID-19 pandemic and its structure is 6 presented in Table 1 . to perform the distribution and clustering analyses using ArcGIS10.2 software. 16 

Grounded theory is an 'inductive method of theoretical development' proposed by 18 Glaser and Strauss [61] , which contributes to substantive or formal theory through an 19 abstract heuristic process. In the present study, it was used to analyse changes in panic 20 during the pandemic. We randomly selected 4000 tweets as the original data, selected Table 2 , and the 2 results of the categories and concepts are listed in Table 3 . Table 2 Examples of the open coding process Regarding core coding, we finally determined the four core categories of 8 In this study, we used a mixed-approach method to evaluate panic and its impact 9 on the Chinese population during the COVID-19 pandemic. Figure 2 shows the 10 convergent mixed-method design, where both qualitative and quantitative methods 11 were used in parallel, analysed separately, and then merged. The first phase involved gathering Weibo tweets with location information through 18 Python crawlers and pre-processing the data for subsequent analysis. In the second 19 phase, we analysed the tweets from the first round of the nationwide outbreak in January 20 2020. A dictionary-based sentiment analysis method was used to perform a fine-grained 21 sentiment analysis that contained both time and space analyses. In addition, the 22 grounded theory approach was used to randomly select 4000 texts for coding and 23 explore the influencing factors and behavioural responses to panic. In the third phase, 24 by combining qualitative and quantitative methods, a conceptual spatial evolution model of public panic was summarised and used to explore the role of panic at different 1 stages and in different regions of the pandemic. Finally, we analysed tweets from the 2 second round of the Nanjing outbreak in July 2021. The LDA topic model method was 3 used to obtain the topics for each stage and to verify the rationality of the panic model.

In the following sections, we present the results of the analyses performed using this 5 methodology. Temporal-level analysis of public emotions 19 Taking a day as a unit, we calculated the intensity of the average sentiment (good, Spatial characteristics of public panic 37 Excessive panic regarding a pandemic threatens people's physical and mental 38 health. Therefore, it is necessary to explore the spatial distribution of panic. We As seen in Figure 7 , the spatial distribution of panic in prefecture-level cities also 5 showed the existence of neighbourhood diffusion and hierarchical diffusion and formed 6 several clusters. Wuhan was the centre of the largest cluster of panic, and the panic 7 intensity of cities around Wuhan was also higher than that of the relatively far radiation 8 area. Therefore, we further divided 'the radiation area' into the neighbouring areas (the 9 cities in Hubei except Wuhan) and the radiation area (the cities in areas surrounding 10 Hubei). 11 The results showed that the spatial distribution of panic was correlated with the area [66] . From the perspective of 'behaviour', people adopted the typical behavioural 19 response to the pandemic based on the panic reaction. The conceptual framework is 20 illustrated in Figure 8 . in the radiation and remote areas disappeared. In addition, panic in response to the 2 pandemic perception was experienced in four stages: rare panic, dramatic panic, 3 regulating panic, and rational response. Their behavioural responses corresponded to 4 the concepts obtained through coding analysis based on the grounded theory approach. concepts effectively, as shown in Table 4 . All topics corresponded to the panic concepts analysed above, whereas the 3 concepts of 'danger nearby' and 'supply demand' were found not to be extracted.

Because the LDA topic model selects the top 200 high-frequency words in these topics 5 to correspond to the related concepts, these two may have been less mentioned.

Otherwise, topic 19 belongs to both the concepts of 'emotions' and 'reopening life'. The sentence frequency for LDA topics which matched the panic concepts at each 10 stage was counted and is shown in Figure 10 . It reveals a basic correlation relationship 11 between the topics and the behavioural responses proposed by the panic model at each Compared to the first-round nationwide outbreak, the duration of these two 9 outbreaks stabilised at approximately 45 days (decreased from 47 days to 45 days). 10 While it is known that the intensity of panic in the first-round national outbreak was 11 higher, the duration of the outbreak period increased from 5 days to 9 days. This may 12 be influenced by the stage division method we used. In this case, the concept of 13 'dereliction of duty' was expressed more often than before, and the figure increased 14 from 13.32% to 15.12%, then decreased to 11.53% during the three stages. In addition,

'rumour spreading' and 'moral kidnapping' were also frequently mentioned during this 16 outbreak. Another possible explanation is that the upgradation of the Delta virus has 17 increased panic, and negligent behaviour of government departments has hurt citizens' 18 preventive motivation, compared to successful responsive experiences such as Wuhan 19 and Guangzhou.

Spatial-level analysis of public panic 21 Using the same sentiment analysis and kernel density analysis, we analysed the 22 distribution of panic in the Nanjing outbreak, and the results are shown in Figure 11 . As shown in Figure 11 , the centre of public panic was also concentrated at the The Nanjing outbreak rapidly infected tourists travelling to Zhangjiajie (a tourism 10 attraction belonging to Hunan Province) and formed a chain of pandemic transmission 11 that exacerbated the public panic in Hunan Province. In general, the characteristics of 12 hierarchical spread and the neighbourhood spread of panic were reconfirmed. The 13 actual distance from the pandemic, transportation, and social media play an essential 14 role in the spread of panic. home isolation. In addition, it was found that people were highly concerned regarding 12 the popular events, the government's ability to manage the pandemic, and the supply 13 of necessities in the early stage. Hence, policymakers need to focus on people's mental 14 health issues, pay attention to topics of public concern, and release official information 15 as soon as possible [21] . It is beneficial for alleviating panic and avoiding negative 16 behaviours such as rumour spreading, moral kidnapping, and proactive prevention.

The results of the geographic analysis indicated that public sentiment was 18 influenced by both the perceived distance and the actual distance of the outbreak. Panic panic during the pandemic are still scarce; hence, panic spread and its mechanism 10 should be further studied in the future. 

Pandemic, panic, and psychiatrists -What should 15 be done before, during, and after COVID-19?

Fear and stigma: The 18 epidemic within the SARS outbreak

An analysis on the panic during COVID-19 pandemic 21 through an online form

The development of normal fear: A century of research

Exploring the sources and role of

Sentiment analysis during

Panic buying 4 in the COVID-19 pandemic: A multi-country examination

What caused global stock market meltdown during the 7 COVID pandemic-Lockdown stringency or investor panic?

Can local governments' disclosure of 10 pandemic information decrease residents' panic when facing COVID-19 in China? International 11

COVID-19 and pets: When pandemic meets panic

Editorial: Epidemics and fear

The politics of risk and blame during an 18 epidemic of fear

Taking control amidst the chaos: Emotion 21 regulation during the COVID-19 pandemic

Cognitive vulnerability: A model of the etiology of fear

Panic and pandemic: Narrative review of the literature on the links 26 and risks of panic disorder as a consequence of the SARS-CoV-2 pandemic

The role of social media in spreading panic among 29 primary and secondary school students during the COVID-19 pandemic: An online 30 questionnaire study from the Gaza Strip

Mapping of Health Literacy and Social Panic Via Web 33

Search Data During the COVID-19 Public Health Emergency: Infodemiological Study

Risk perception of COVID-19: A comparative

Understanding the evolutions of public responses using social media

Social Sensing: A New

Spatial-temporal response patterns of tourist

Social Big Data Analysis of Information 7 Spread and Perceived Infection Risk During the

GIS-based spatial modeling of COVID-19 incidence 11 rate in the continental United States

Public attention about COVID-19 on social media: An investigation

Timed intervention in COVID-19 and panic buying

Sentiment analysis and 24 its applications in fighting COVID-19 and infectious diseases: A systematic review

Twitter for disaster relief through

A topic modeling analysis on the early phase of

A new topic modeling based approach for aspect extraction in

Educational Research: Planning, Conducting, and

Grounded Theory in the Twenty First Century

Uncovering crowdsourcing in tourism apps: A grounded theory study

Text Mining and Grounded Theory for Appraising the Self

Understanding farmers' naturalistic decision 8 making around prophylactic antibiotic use in lambs using a grounded theory and natural 9 language processing approach

Exploring public attitudes of child abuse in mainland 12 China: A sentiment analysis of China's social media Weibo

Journal 15 of the China society for scientific and technical information

Density estimation for statistics and data analysis

The Discovery of Grounded Theory: Strategy for Qualitative 18 Research

CANONS 20 AND EVALUATIVE CRITERIA

Latent dirichlet allocation

Evolution of complex disasters. The Lancet

Leveraging Twitter data to understand public 27 sentiment for the COVID-19 outbreak in Singapore

Mechanism and Influence of Emotions Arising in Daily Consuming 30 Spaces: A Case Study of Nanjing

Chinas City Network Characteristics Based on Social Network

Sentiment analysis of nationwide lockdown due to COVID 36 19 outbreak: Evidence from India

The development of COVID-19 in China: Spatial

Amplifying Panic and Facilitating Prevention: Multifaceted Effects of Traditional and 5

Social Media Use During the 2015 MERS Crisis in South Korea