Table.
Summary of chosen articles
| Publication month | Origin | Social media | Study population and sample size | Methods | Key findings | |
|---|---|---|---|---|---|---|
| Detection or prediction of COVID-19 cases | ||||||
| Li et al17 | March | China | Google Trends, Baidu Search Index, and Sina Weibo Index | Keywords of coronavirus and pneumonia were searched, and trend data was collected from Google Trends, Baidu Search Index, and Sina Weibo Index from Jan 2 to Feb 20, 2020 | Lag correlation | Lag correlations showed a maximum correlation between trend data and the number of diagnoses at 8–12 days before for laboratory-confirmed cases and 6–8 days before for suspected cases |
| Liu et al18 | August | China | Sina Weibo | Sina Weibo messages between Jan 20 and Feb 15, 2020; 599 participants | Gathered data via Sina Weibo, then followed up with telephone call; statistical analysis taken with Fisher exact test; rates of death calculated with Kaplan-Meier method; multivariate Cox regression used to establish risk factors for mortality | Older age (ie, >69 years), diffuse pneumonia, and hypoxaemia are factors that can help clinicians to identify patients with COVID-19 who have poor prognosis; aggregated data from social media can also be comprehensive, immediate, and informative in disease prognosis |
| O'Leary and Storey19 | September | USA | Google Trends, Wikipedia, and Twitter | Google Trends searches for coronavirus and COVID-19 between Jan 21 and April 5, 2020; Wikipedia page views for coronavirus and COVID-19 between Jan 12 and April 5, 2020; number of Twitter original tweets between Jan 27 and April 5, 2020; numbers of COVID-19 cases and deaths in the USA20 | Regression analysis | To model the number of cases, the current Wikipedia page views, tweets from 1 week before, and Google Trends searches from 2 weeks before were used; to model of the number of deaths, each variable was taken from 1 week earlier than for cases |
| Peng et al21 | June | China | Sina Weibo | 1200 records | Spatiotemporal distribution of COVID-19 cases in the main urban area of Wuhan, China; kernel density analysis; ordinary least square regression | Older people (ie, >60 years) are at high risk of severe symptoms and have high prevalence in the COVID-19 outbreak, and they account for >50% of the total number of Sina Weibo help seekers; early transmission of COVID-19 in Wuhan, China, could be divided into three phrases: scattered infection, community spread, and full-scale outbreak |
| Qin et al22 | March | China | Baidu Search Index | Social media search index for dry cough, fever, chest distress, coronavirus, and pneumonia from Dec 31, 2019, to Feb 9, 2020; data for new suspected cases of COVID-19 from Jan 20 to Feb 9, 2020 | Subset selection; forward selection; lasso regression; ridge regression; elastic net | Case numbers of new suspected COVID-19 correlated significantly with the lagged series of social media search index; social media search index could detect new suspected COVID-19 cases 6–9 days earlier than could laboratories |
| Zhu et al23 | April | China | Sina Weibo | 1101 Sina Weibo posts related to COVID-19 from Dec 31, 2019, to Feb 12, 2020 | Descriptive statistics: numbers and percentage; time series analysis | Attention to COVID-19 was low until China openly admitted human-to-human transmission on Jan 20, 2020; attention quickly increased and remained high over time |
| Government responses | ||||||
| Basch et al24 | April | USA | YouTube | 100 most widely viewed videos uploaded in January, 2020 | Descriptive analysis: frequency, percentage, mean, and standard deviation | Percentage of each of the seven key prevention behaviours that are listed on the US Centers for Disease Control and Prevention website that were covered in the 100 videos varied from 0% (eg, use a face mask for protection if you are caring for the ill) to 31% (avoid close contact with people who are sick); overall, videos that covered at least one prevention behaviour accounted for less than one-third of the 100 videos |
| Basch et al25 | April | USA | YouTube | 100 most widely viewed YouTube videos as of Jan 31, 2020, and March 20, 2020, with keyword of coronavirus in English, with English subtitles, or in Spanish | Descriptive analysis: frequency, percentage, mean, and standard deviation | <50% of videos in either sample covered any of the prevention behaviours that are recommended by the US Centers for Disease Control and Prevention |
| Khatri et al26 | March | Singapore | Youtube | 150 videos collected on Feb 1–2, 2020, with keywords of 2019 novel coronavirus (50 videos), and Wuhan virus in English (50 videos) and Mandarin (50 videos) | Descriptive analysis: percentage and mean; DISCERN score; Medical Information and Content Index score | Mean DISCERN score for reliability was 3·12 of 5·00 for English and 3·25 of 5·00 for Mandarin videos; mean cumulative Medical Information and Content Index score of useful videos was 6·71 of 25·00 for English and 6·28 of 25·00 for Mandarin |
| Li et al27 | March | China | Sina Weibo | 36 746 Sina Weibo data from Dec 30, 2019, to Feb 1, 2020; a random sample of 3000 Sina Weibo posts as training dataset | Linear regression; support vector machine; Naive Bayes; natural language processing | Classified the information related to COVID-19 into seven types of situational information and their predictors |
| Merkley et al28 | April | Canada | Twitter and Google Trends | 33 142 tweets from 292 social media accounts of federal members of parliament from Jan 1 to March 28, 2020; 87 Google search trends for the search term coronavirus in the first half (ie, days 1–14) and second half (ie, days 15–31) of March, 2020; a survey of 2499 Canadian citizens ≥18 years from April 2 to April 6, 2020 | Linear regression | No members of parliament from any party downplaying the pandemic; no association between Conservative Party vote share and Google search interest in the coronavirus |
| Rufai and Bunce29 | April | USA | 203 viral tweets from G7 world leaders from Nov 17, 2019, to March 17, 2020 with keywords COVID-19 or coronavirus and a minimum of 500 likes | Qualitative design; content analysis | 166 of 203 of tweets were informative; 9·4% (19) were morale-boosting; 6·9% (14) were political | |
| Sutton et al30 | September | USA | 690 accounts representing public health, emergency management, and elected officials and 149 335 tweets | χ2 analyses; negative binomial regression modelling | Systematic changes were made in message strategies over time and identified key features that affect message passing, both positively and negatively; results have the potential to aid in message design strategies as the pandemic continues, or in similar future events | |
| Wang et al31 | September | USA | 13 598 tweets related to COVID-19 from Jan 1 to April 27, 2020 | Temporal analysis and networking analysis | 16 categories of message types were manually annotated; inconsistencies and incongruencies were identified in four critical topics (ie, wearing masks, assessment of risks, stay at home order, and disinfectant and sanitiser); network analysis showed increased communication coordination over time | |
| Infodemics | ||||||
| Ahmed et al32 | October | UK | 22 785 tweets and 11 333 Twitter users with #FilmYourHospital from April 13 to April 20, 2020 | Social network analysis; user analysis | The most important drivers of the #FilmYourHospital conspiracy theory are ordinary citizens; YouTube was the information source most linked to by users; the most retweeted post belonged to a verified Twitter user | |
| Ahmed et al33 | May | UK | A subsample of 233 tweets from 10 140 tweets collected from 19:44 h UTC on Friday, March 27, 2020, to 10:38 h UTC on Saturday, April 4, 2020 were used for content analysis | Descriptive statistics: numbers, percentage; social network analysis; content analysis | 34·8% (81 of 233) of tweets linked 5G and COVID-19; 32·2% (75) of tweets denounced the conspiracy theory | |
| Brennen et al34 | October | UK | Digital visual media | 96 samples of visuals from January to March, 2020 | Qualitative coding | Organised all findings into six trends: authoritative agency, virulence, medical efficacy, intolerance, prophecy, satire; a small number of manipulated visuals, all were produced by use of simple tools; no examples of so-called deepfakes (ie, techniques that are used to make synthetic videos that closely resemble real videos) or other techniques that were based on artificial intelligence |
| Bruns et al35 | August | Australia | 89 664 distinct Facebook posts from Jan 1 to April 12, 2020 | Time series; network analysis | Substantially increased number of posts about 5G rumours on Facebook after March 19, 2020; network analysis showed that coalitions of various groups were brought together by conspiracy theories about COVID-19 and 5G technology | |
| Galhardi et al36 | October | Brazil | WhatsApp, Instagram, and Facebook | Fake news collected from March 17 to April 10, 2020, on the basis of data from the Eu Fiscalizo app (version 5.0.5) | Quantitative content analysis | WhatsApp is the main channel for sharing fake news, followed by Instagram and Facebook |
| Gallotti et al37 | October | Italy | >100 million Tweets | Developed an Infodemic Risk Index | Before the rise of COVID-19 cases, entire countries had measurable waves of potentially unreliable information, posing a serious threat to public health | |
| Islam et al38 | October | Bangladesh | Fact-checking agency websites, Facebook, Twitter, and websites for television networks and newspapers | 2311 infodemic reports related to COVID-19 between Dec 31, 2019, and April 5, 2020 | Descriptive analysis; spatial distribution analysis | Misinformation that is fuelled by rumours, stigma, and conspiracy theories can have potentially severe implications on public health if prioritised over scientific guidelines; governments and other agencies should understand the patterns of rumours, stigma, and conspiracy theories that are related to COVID-19 and circulating globally so that they can develop appropriate messages for risk communication |
| Kouzy et al39 | March | Lebanon | 673 English tweets collected on Feb 27, 2020; 617 tweets after exclusion of tweets that were humorous or not serious | Descriptive statistics; bar chart; χ2 statistic to calculate p value (2-sided; p=0·05 significance threshold) for the association between account or tweet characteristics and the presence of misinformation or unverifiable information about COVID-19 | 153 (24·8%) of 617 tweets had misinformation; 107 (17·3%) had unverifiable information; misinformation rate higher in informal individual or group accounts than in formal individual or group accounts (33·8% [123 of 364] vs 15·0% [30 of 200], p<0·0010) | |
| Moscadelli et al40 | August | Italy | Fake news and corresponding verified news that was circulated in Italy | 2102 articles between Dec 31, 2019, and April 30, 2020 | Social media trend analysis by use of BuzzSumo | Links containing fake news were shared 2 352 585 times, accounting for 23·1% (2 352 585 of 10 184 351) of total shares of all reviewed articles |
| Pulido et al41 | April | Spain | 942 valid tweets between Feb 6 and Feb 7, 2020 | Communicative content analysis | Misinformation was tweeted more but retweeted less than tweets based on scientific evidence; tweets based on scientific evidence had more engagement than misinformation | |
| Rovetta and Bhagavathula42 | August | Italy | Google Trends and Instagram | 2 million Google Trends queries and Instagram hashtags from Feb 20 to May 6, 2020 | Classification of infodemic monikers (ie, a term, query, hashtag, or phrase that generates or feeds fake news, misinterpretations, or discrimination); computed the mean peak volume with a 95% CI | Globally, growing interest exists in COVID-19, and numerous infodemic monikers continue to circulate on the internet |
| Uyheng and Carley43 | October | USA and Philippines | 12·0 million tweets from 1·6 million users from the USA and 15·0 million tweets from 1·0 million users from the Philippines between March 5 and March 19, 2020 | Hate speech score assigned to each tweet by use of machine learning algorithm; bot scores were assigned to each user via BotHunter algorithm; social media analysis via ORA software; network analysis via centrality analysis; cluster analysis via Leiden algorithm | Analysis showed idiosyncratic relationships between bots and hate speech across datasets, emphasising different network dynamics of racially charged toxicity in the USA and political conflicts in the Philippines; bot activity is linked to hate in both countries, especially in communities that are dense and isolated from others | |
| Mental health | ||||||
| Gao et al44 | April | China | Sina Weibo | Online survey on Wenjuanxing platform from Jan 31 to Feb 2, 2020; with 4872 Chinese citizens aged ≥18 years from 31 provinces and autonomous regions in China | Multivariable logistic regression | Social media exposure was frequently positively associated with high odds of anxiety (odds ratio 1·72, 95% CI 1·31–2·26) and combination of depression and anxiety (odds ratio 1·91, 95% CI 1·52–2·41) |
| Li et al45 | March | China | Sina Weibo | Sina Weibo posts from 17 865 active Sina Weibo users between Jan 13 and Jan 26, 2020 | Sentiment analysis; paired sample t-test | Negative emotions and sensitivity to social risks increased; scores of positive emotions and life satisfaction decreased after outbreak declaration |
| Prevention education in videos | ||||||
| Hakimi and Armstrong46 | September | USA | YouTube | 49 of the first 100 videos on YouTube with the most views that were identified by the search term DIY hand sanitiser; 51 videos were excluded because they were not in English or not related to the search term | Codified video content; assessed by use of Cohen's κ; descriptive statistics calculated; assessed by χ2 test with 2-sided p value <0·05 as the threshold for significance | Most videos did not describe labelling storage containers, 69% (34 of 49) of videos encouraged the use of oils or perfumes to enhance hand sanitiser scent, and 2% (1) of videos promoted the use of colouring agents to be more attractive for use among children specifically; significantly increased mean number of daily calls to poison control centres regarding unsafe paediatric exposure to hand sanitiser since the first confirmed patient with COVID-19 in the USA (p<0·0010); significantly increased mean number of daily calls in March, 2020, compared with the previous 2 years (p<0·0010) |
| Hernández-García and Giménez-Júlvez47 | June | Spain | YouTube | 129 videos in Spanish with the terms prevencion coronavirus and prevencion COVID19 | Univariate analysis; multiple logistic regression model | Information from YouTube in Spanish on basic measures to prevent COVID-19 is usually not complete and differs according to the type of authorship (ie, mass media, health professionals, individual users, or others) |
| Moon and Lee48 | August | South Korea | YouTube | 105 most viewed YouTube videos from Jan 1 to April 30, 2020 | Modified DISCERN index; Journal of the American Medical Association Score benchmark criteria; Global Quality Score; Title–Content Consistency Index; Medical Information and Content Index | 37·14% (39 of 105) of videos contained misleading information; independent user-generated videos showed the highest proportion of misleading information at 68·09% (32 of 47); misleading videos had more likes, fewer comments, and longer running times than did useful videos; transmission and precautionary measures were the most frequently covered content |
| Ozdede and Peker49 | July–August | Turkey | YouTube | The top 116 English language videos with at least 300 views | Precision indices and total video information and quality index scores were calculated | High number of views on dentistry YouTube videos related to COVID-19; quality and usefulness of these videos are moderate |
| Yüce et al50 | July | Turkey | YouTube | 55 English videos about COVID-19 control procedures for dental practices collected on March 31, 2020, between 9:00 h and 18:00 h | Modified DISCERN instrument; descriptive statistics | Only two (3·6%) of 55 videos were good quality, whereas 24 (43·6%) videos were poor quality |
| Public attitudes | ||||||
| Abd-Alrazaq et al7 | April | Qatar | 2·8 million English tweets (167 073 unique tweets from 160 829 unique users) from Feb 2 to March 15, 2020 | Word frequencies of single (ie, unigrams) and double words (ie, bigrams); sentiment analysis; mean number of retweets, likes, and followers for each topic; interaction rate per topic; LDA for topic modelling | Identified 12 topics and grouped into four themes; average sentiment positive for ten topics and negative for two topics | |
| Al-Rawi et al51 | November | Canada | Over 50 million tweets referencing #Covid-19 and #Covid19 for more than 2 months in early 2020 | Mixed method: analysed emoji use by each gender category; the top 600 emojis were manually classified on the basis of their sentiment | Identified five major themes in the analysis: morbidity fears, health concerns, employment and financial issues, praise for front-line workers, and unique gendered emoji use; most emojis are extremely positive across genders, but discussions by women and gender minorities are more negative than by men; when discussing particular topics (eg, financial and employment matters, gratitude, and health care), there are many differences; use of several unique gender emojis to express specific issues (eg, coffin, skull, and siren emojis were used more often by men than by other genders when discussing fears and morbidity, whereas the use of the folded hands emoji as a thankful gesture for front-line workers was found more often in discussions by women than by other genders and the bank emoji was noted only in women's discussions) | |
| Arpaci et al52 | July | Turkey | 43 million tweets between March 22 and March 30, 2020 | Evolutionary clustering analysis | Unigram terms appear more frequently than bigram and trigram (ie, triple words) terms; during the epidemic, many tweets about COVID-19 were distributed and attracted widespread public attention; high-frequency words (eg, death, test, spread, and lockdown) indicated that people were afraid of being infected and people who were infected were afraid of death; people agreed to stay at home due to fear of spread and called for physical distancing since they became aware of COVID-19 | |
| Barrett et al53 | August | USA | 188 tweets about Governor Dan Patrick's statement on March 23, 2020, about generational self-sacrifice. | Thematic analysis | 90% (169 of 188) of tweets opposed calculated ageism, whereas only 5% (9) supported it and 5% (10) conveyed no position; opposition centred on moral critiques, political–economic critiques, assertions of the worth of older adults (eg, >60 years), and public health arguments; support centred on individual responsibility and patriotism | |
| Boon-Itt and Skunkan54 | November | Thailand | 107 990 English tweets related to COVID-19 between Dec 13, 2019, and March 9, 2020 | Sentiment analysis; topic modelling by use of LDA | Sentiment analysis showed a predominantly negative feeling towards the COVID-19 pandemic; topic modelling revealed three themes relating to COVID-19 and the outbreak: the COVID-19 pandemic emergency, how to control COVID-19, and reports on COVID-19 | |
| Budhwani and Sun55 | May | USA | 16 535 tweets about Chinese virus or China virus between March 9 and March 15, 2020, 177 327 tweets between March 19 and March 25, 2020 | Descriptive analysis; spatial analysis | Nearly 10 times increase at the national level; all 50 states had an increase in the number of tweets exclusively mentioning Chinese virus or China virus instead of coronavirus disease, COVID-19, or coronavirus; mean 0·38 tweets referencing Chinese virus or China virus were posted per 10 000 people at the state level in the preperiod (ie, March 9–15, 2020), and 4·08 of these stigmatising tweets were posted in the postperiod (ie, March 19–25, 2020), also indicating a 10 times increase | |
| Chang et al56 | November | Taiwan | 10 news websites, 11 discussion forums, 1 social network, 2 principal media sharing networks | 1·07 million Chinese texts from Dec 30, 2019, to March 31, 2020 | Deductive analysis | Online news promoted negativity and drove emotional social posts; stigmatising language that was linked to the COVID-19 pandemic showed an absence of civic responsibility that encouraged bias, hostility, and discrimination |
| Chehal et al57 | July | India | 29 554 tweets during the second lockdown (ie, April 15–May 3, 2020); 47 672 tweets during the third lockdown (May 4–17, 2020) | Sentiment analysis by use of the National Research Council of Canada Emotion Lexicon | A positive approach in the second lockdown but a negative approach in the third lockdown | |
| Chen et al58 | September | China | Sina Weibo | 1411 posts pertinent to COVID-19 taken from Healthy China, an official Sina Weibo account of the National Health Commission of China, from Jan 14 to March 5, 2020 | Descriptive analysis; hypothesis testing | Media richness (ie, potential information load, where low richness is only text and high richness is not only text) negatively predicted citizen participation via government social media, but dialogic loop (ie, stimulation of public dialogue, provision of the dialogue channel, and response to public feedback in a timely manner) facilitated engagement |
| Damiano and Allen Catellier59 | August | USA | 600 English tweets from the USA were selected: 300 from February, 2020, and 300 from March, 2020 | Frequencies; χ2 statistics | Neutral sentiment; tweets about COVID-19 risks and emotional outrage accounted for <50% (135 of 600); few tweets were related to blame | |
| Darling-Hammond et al60 | September | USA | 339 063 tweets from non-Asian respondents of the Project Implicit Asian Implicit Association Test from 2007–20 and were broken into two datasets: the first dataset was from Jan 1, 2007, to Feb 10, 2020; the second data set was from Feb 11 to March 31, 2020 | Local polynomial regression; interrupted time-series analyses | Implicit Americanness Bias steadily decreased from 2007 to 2020; when media entities began using stigmatising terms, such as Chinese virus, starting from March 8, 2020, Implicit Americanness Bias began to increase; such bias was more pronounced among conservative individuals than among non-conservative individuals | |
| Das and Dutta61 | July | India | 410 643 tweets with #IndiaLockdown and #IndiafightsCorona from March 22 to April 21, 2020 | National Research Council of Canada lexicon for corpus-level emotion mining; sentimentr from open source R software for sentiment analysis to create additional sentiment scores; LDA for topic models; Natural Language Toolkit to develop sentiment-based topic models | For the broad corpus-level analysis, the context of positiveness was substantially higher than were negative sentiments; however, positive sentiment trends were similar to negative sentiment trends in terms of topics covered when the analysis was done at individual tweet level; the results showed that the discussion of COVID-19 in India on Twitter contains slightly more positive sentiments than negative sentiments | |
| De Santis et al62 | July | Italy | 1 044 645 tweets | A general purpose methodological framework, grounded on a biological metaphor and on a chain of NLP and graph analysis techniques | Energy evolution through time was monitored; daily hot topics were identified (eg, COVID-19, Walter Ricciardi's retweet of an anti-Trump tweet from Michael Moore, Gabriele Gravina's argument against suspension of Italian football, increased COVID-19 cases in Italy, high case numbers in Lombardy, Italy, and an interview of Matteo Salvini about COVID-19 topics by Massimo Giletti) | |
| Dheeraj63 | May–June | India | 868 posts related to COVID-19 | Fetching the articles: Python Reddit Application Programming Interface Wrapper; data preprocessing: Reddit Application Programming Interface and Natural Language Toolkit library | Of 868 posts on Reddit that were related to COVID-19 articles, 50% (434) were neutral, 22% (191) were positive, and 28% (243) were negative | |
| Essam and Abdo64 | August | Egypt | 1 920 593 tweets with corona, coronavirus, or COVID-19 keywords from Feb 1 to April 30, 2020 | Thematic analysis | The dominant themes that were closely related to coronavirus tweets included the outbreak of the pandemic, metaphysics responses, signs and symptoms in confirmed cases, and conspiracies; the psycholinguistic analysis showed that tweeters maintained high amounts of affective talk (ie, expression of feelings), which was loaded with negative emotions and sadness; Linguistic Inquiry and Word Count's psychological categories of religion and health dominated the Arabic tweets discussing the pandemic situation | |
| Yin FL et al65 | March | China | Sina Weibo | Sina Weibo posts from Dec 31, 2019, to Feb 7, 2020 | Multiple-information susceptible-discussing-immune model | Model reproduction ratio declined from 1·78 to 0·97, showing that the peak of posts had passed but the topic was still on social media afterwards with a decreased number of posts |
| Gozzi et al66 | October | Italy, UK, USA, and Canada | News, YouTube, Reddit, and Wikipedia | 227 768 web-based news articles from Feb 7 to May 15, 2020; 13 448 YouTube videos from Feb 7 to May 15, 2020; 107 898 English user posts and 3 829 309 comments on Reddit from Feb 15 to May 15, 2020; 278 456 892 views of Wikipedia pages that were related to COVID-19 from Feb 7 to May 15, 2020 | Linear regression; topic modelling by use of LDA | Collective attention was mainly driven by media coverage rather than epidemic progression, rapidly became saturated, and decreased despite media coverage and COVID-19 incidence remaining high; Reddit users were generally more interested in health, data regarding the new disease, and interventions needed to halt the spreading with respect to media exposure than were users of other platforms |
| Green et al67 | July | USA | 19 803 tweets from Democrats and 11 084 tweets from Republicans between Jan 17 and March 31, 2020 | Random forest | Democrats discussed the crisis more frequently—emphasising public health and direct aid to US workers—whereas Republicans placed greater emphasis on national unit, China, and businesses | |
| Han et al68 | April | China | Sina Weibo | 1 413 297 Sina Weibo messages, including 105 330 texts with geographical location information, from 00:00 h on Jan 9, 2020, to 00:00 h on Feb 11, 2020 | Time series analysis; kernel density estimation; Spearman correlation; LDA model; random forest algorithm | Public response was sensitive to the epidemic and notable social events, especially in urban agglomerations |
| Jelodar et al69 | June | China | 563 079 English comments related to COVID-19 from Reddit between Jan 20 and March 19, 2020 | Topic modelling by use of LDA and probabilistic latent semantic analysis; sentiment classification by use of recurrent neural network | The results showed a novel application for NLP based on a long short term memory model to detect meaningful latent topics and sentiment–comment classification on issues related to COVID-19 on social media | |
| Jimenez-Sotomayor et al70 | April | Mexico | A random sample of 351 of 18 128 tweets were analysed from March 12 to March 21, 2020 | Qualitative content classification | The most common types of tweets were personal opinions (31·9% [112 of 351]), followed by informative tweets (29·6% [104]), jokes or ridicule (14·2% [50]), and personal accounts (13·4% [47]); 72 of 351 tweets were most likely intended to ridicule or offend someone and 21·1% (74) had content implying that the life of older adults (ie, referred to in tweets as “elderly”, “older”, and “boomer”) was less valuable than that of younger people or downplayed the relevance of COVID-19 | |
| Kim71 | August | South Korea | 27 849 individual tweets about COVID-19 between Feb 10 and Feb 14, 2020 | Binary logistic regression; semantic network analysis | Social network size was a negative predictor of incivility | |
| Kurten and Beullens72 | August | Belgium | 373 908 tweets and retweets from Feb 25 to March 30, 2020 | Time series; network bigrams; emotion lexicon; LDA | Notable COVID-19 events immediately increased the number of tweets; most topics focused on the need for EU collaboration to tackle the pandemic | |
| Kwon et al73 | October | USA | 259 529 unique tweets containing the word coronavirus between Jan 23 and March 24, 2020 | Trending analysis; spatiotemporal analysis | Early facets of physical distancing appeared in Los Angeles (CA, USA), San Francisco (CA, USA), and Seattle (WA, USA); social disruptiveness tweets were most retweeted, and intervention implementation tweets were most favourited | |
| Lai et al74 | October | USA | 522 comments from an Ask Me Anything session on COVID-19 on March 11, 2020, from 14:00 h to 16:00 h EST | Content analysis | The highest number of posts were about symptoms (27% [141 of 522]), followed by prevention (25% [131]); symptoms was the most common intended topic for further discussions (28% [94 of 337]) | |
| Li et al75 | April | China | Sina Weibo | 115 299 Sina Weibo posts from Dec 23, 2019, to Jan 30, 2020; 11 893 of them were collected from Dec 31, 2019, to Jan 20, 2020, for qualitative analysis; total daily cases of COVID-19 in Wuhan, China, were obtained from the Chinese National Health Commission | Linear regression model; qualitative content analysis | Positive correlation between the number of Sina Weibo posts and the number of reported cases, with ten COVID-19 cases per 40 posts; posts grouped into four themes |
| Li et al76 | September | USA | 155 353 unique English tweets related to COVID-19 that were posted from Dec 31, 2019, to March 13, 2020 | Content analysis | Peril of COVID-19 was mentioned the most often, followed by content about marks (ie, cues to identify members of a stigmatised group: flu-like symptoms, personal protective equipment, Asian origin, and health-care providers and essential workers), responsibility, and group labelling; information on conspiracy theories was more likely to be included in tweets about group labelling and responsibility than in tweets about COVID-19 peril | |
| Lwin et al77 | May | Singapore | 20 325 929 tweets from 7 033 158 unique users from Jan 28 to April 9, 2020 | Sentiment analysis | Public emotions shifted strongly from fear to anger over the course of the pandemic, while sadness and joy also surfaced; anger shifted from xenophobia at the beginning of the pandemic to discourse around the stay-at-home notices; sadness was emphasised by the topics of losing friends and family members, whereas topics that were related to joy included words of gratitude and good health; emotion-driven collective issues around shared public distress experiences of the COVID-19 pandemic are developing and include large-scale social isolation and the loss of human lives | |
| Ma et al78 | July | China | Top 200 accounts from Jan 21 to Jan 27, 2020 | Simple linear regression; multiple linear regression; content analysis | For non-medical institution accounts in the model, report and story types of articles had positive effects on whether users followed behaviours; for medical institution accounts, report and science types of articles had a positive effect | |
| Medford et al79 | June | USA | 126 049 English tweets from 53 196 unique users with matching hashtags that were related to COVID-19 from Jan 14 to Jan 28, 2020 | Temporal analysis; sentiment analysis; topic modelling by use of LDA | The hourly number of tweets that were related to COVID-19 starkly increased from Jan 21, 2020, onwards; fear was the most common emotion and was expressed in 49·5% (62 424 of 126 049) of all tweets; the most common predominant topic was the economic and political effect | |
| Mohamad80 | June | Brunei | Twitter, Instagram, and TikTok | 30 individual profiles from Instagram, Twitter, and TikTok | Qualitative content analysis | Five narratives of local responses to physical distancing practices were apparent: fear, responsibility, annoyance, fun, and resistance |
| Nguyen et al81 | September | USA | 3 377 295 US tweets that were related to race from November, 2019, to June, 2020 | Support vector machine was used for sentiment analysis | Proportion of negative tweets referencing Asians increased by 68·4%; proportion of negative tweets referencing other racial or ethnic minorities was stable; common themes that emerged during the content analysis of a random subsample of 3300 tweets included: racism and blame, anti-racism, and effect on daily life | |
| Odlum et al82 | June | USA | 2 558 474 Tweets from Jan 21 to May 3, 2020 | Clustering algorithm; NLP; network diagrams | 15 topics (in four themes) were identified; positive sentiments, cohesively encouraging online discussions, and behaviours for COVID-19 prevention were uniquely observed in African American Twitter communities | |
| Park et al83 | May | South Korea | 43 832 unique users and 78 233 relationships on Feb 29, 2020 | Network analysis; content analysis | Spread of information was faster in the COVID-19 network than in the other networks; tweets containing medically framed news articles were more popular than were tweets that included news articles adopting non-medical frames | |
| Pastor84 | April | Philippines | Tweets were collected on three Tuesdays in March, 2020, since lockdown in Philippines | NLP for sentiment analysis | Negative sentiments increased over time in lockdown | |
| Samuel et al85 | June | USA | 900 000 tweets from February to March, 2020 | Sentiment analysis packages; textual analytics; machine learning classification methods: Naive Bayes and logistic regression | For short tweets, classification accuracy was 91% with Naive Bayes whereas accuracy was 74% with logistic regression; both methods showed weaker performance for longer tweets | |
| Samuel et al86 | August | USA | 293 597 tweets, 90 variables | Textual analytics to analyse public sentiment support; sentiment analysis by use of R package Syuzhet (version 1.0.6) | For the reopening of the US economy, there was more positive sentiment support than there was negative support; developed a novel sentiment polarity based public sentiment scenarios framework | |
| Su et al87 | June | China and Italy | Sina Weibo and Twitter | 850 Sina Weibo users with posts published from Jan 9 to Feb 5, 2020; 14 269 tweets from 188 unique Twitter users from Feb 23 to March 21, 2020 | Wilcoxon tests | Individuals focused more on home and expressed a high level of cognitive process after a lockdown in both Wuhan, China, and Lombardy, Italy; level of stress decreased, and the attention to leisure increased in Lombardy, Italy, after the lockdown; attention to group, religion, and emotions became more prevalent in Wuhan, China, after the lockdown |
| Thelwall and Thelwall88 | May | UK | 3 038 026 English tweets from March 10 to March 23, 2020 | Word frequency comparison; χ2 analysis | Women were more likely to tweet about the virus in the context of family, physical distancing, and health care, whereas men were more likely to tweet about sports cancellations, the global spread of the virus, and political reactions | |
| Wang et al89 | July | China | Sina Weibo | 999 978 randomly selected Sina Weibo posts that were related to COVID-19 from Jan 1 to Feb 18, 2020 | Unsupervised Bidirectional Encoder Representations from Transformers model: classify sentiment categories; Term Frequency-Inverse Document Frequency model: summarise the topics of posts; trend analysis; thematic analysis | People were concerned about four aspects regarding COVID-19: the virus origin, symptoms, production activity, and public health control |
| Wicke and Bolognesi90 | September | Ireland | 203 756 tweets | Topic modelling | Although the family frame covers a wider portion of topics, among the figurative frames, war (a highly conventional one) was the frame used most frequently; yet, this frame does not seem to be appropriate to elaborate the discourse around some aspects that are involved in the situation | |
| Xi et al91 | September | China | Sina Weibo | 188 unique topics, their views, and comments from Jan 20 to April 28, 2020 | Thematic analysis; temporal analysis | Six themes were identified: the most prominent theme was older people contributing to the community (46 [24%] of 188) followed by older patients (defined by keywords—eg, “older people”, “old-aged people”, “grandmother”, “grandfather”, “old grandmother”, “old grandfather”, “old woman”, and “old man”) in hospitals (43 [23%]); the theme of contributing to the community was the most dominant in the first phase (Jan 20–Feb 20, 2020; period of COVID-19 outbreak in China); the theme of older patients in hospitals was most dominant in the second (Feb 21–March 17, 2020; turnover period) and third phase (March 18–April 28, 2020; post-peak period in China) |
| Xie et al92 | August | China | Baidu Search Index and Google Trends | Number of cases by Feb 29, 2020: 79 968 cumulative confirmed cases, 41 675 cured cases, 2873 dead cases | Kendall's Tb rank test | Both the Baidu Search Index and Google Trends indices showed a similar trend in a slightly different way; daily Google Trends were correlated to seven indicators, whereas daily Baidu Search Index was correlated to only three indicators; these indexes and rumours are statistically related to disease-related indicators; information symmetry was also noted |
| Xue et al93 | November | Canada | 1 015 874 tweets from April 12 to July 16, 2020 | LDA | Nine themes about family violence were identified | |
| Yigitcanlar et al94 | October | Australia | 96 666 tweets from Australia in Jan 1 to May 4, 2020 | Descriptive analysis; content analysis; sentiment analysis; spatial analysis | Social media analytics is an efficient approach to capture attitudes and perceptions of the public during a pandemic; crowdsourced social media data can guide interventions and decisions of the authorities during a pandemic; effective use of government social media channels can help the public to follow the introduced measures and restrictions | |
| Yu et al95 | July | Spain | 22 223 tweets | Topic modelling; network analysis | Identified eight news frames for each newspaper's Twitter account; the entire pandemic development process is divided into three periods: precrisis, lockdown, and recovery period; understanding of how Spanish news media cover public health crises on social media platforms | |
| Zhao et al96 | May | China | Sina Weibo and microblog hot search list | 4056 topics from Dec 31, 2019, to Feb 20, 2020 | Word segmentation; word frequency; sentiment analysis | The trend of public attention could be divided into three stages; the hot topic keywords of public attention at each stage were slightly different; the emotional tendency of the public towards the COVID-19 pandemic-related hot topics changed from negative to neutral between January and February, 2020, with negative emotions weakening and positive emotions increasing overall; COVID-19 topics with the most public concern were divided into five categories: the situation of the new cases of COVID-19 and its effects, front-line reporting of the pandemic and the measures of prevention and control, expert interpretation and discussion on the source of infection, medical services on the front line of the pandemic, and focus on the pandemic and the search for suspected cases |
| Zhu et al97 | July | China | Sina Weibo | 1 858 288 microblog data | LDA | A so-called double peaks feature appeared in the search curve for epidemic topics; the topic changed over time, the fluctuation of topic discussion rate gradually decreased; political and economic centres attracted high attention on social media; the existence of the subject of rumours enabled people to have more communication and discussion |
All studies were published in 2020. LDA=latent Dirichlet allocation. NLP=natural language processing.