Disaster sociology and the role of authorities during disasters

Within the context of large disasters, the importance of population compliance with government regulation is difficult to overstate. According to contemporary disaster research, the main characteristic of a disaster is the social disruption it causes: a level 12 tsunami hitting an unpopulated shore is unlikely to be designated a disaster, as is the seasonal influenza, which, despite an annual global death count of around 650,000, thus surpassing most singular catastrophic events, nevertheless causes no significant disruption to social macrosystems [25, 26]. And, while there is no denying that external instigators of disasters may cause considerable damage, this holds no less true of the resulting social disruption, as uncertainty as to how to navigate the changed circumstances causes a coordination deficit in society, which may prompt people to act in unpredictable ways [26–28].

Despite populations generally being willing to make sacrifices to navigate safely through a crisis, this does not necessarily imply they adhere to the recommendations made by the political authorities in charge. Non-compliance may occur for a variety of reasons, whether through conflicting social obligations, opportunist reasoning, or simply through distrust toward authorities–or even anti-authoritarian sentiments [29]. The issue of trust, in particular, becomes salient when disaster strikes. While political authorities are able to base their judgments of crisis situations on vast knowledge resources, whether through expertise or access to privileged information, citizens are limited, both in their access to knowledge resources, and in disposable time to inspect such resources. This leaves citizens at an information deficit, ultimately forcing them to base their judgments of the situation on their trust in the political authorities providing them with trustworthy information rather than on the information itself (Tampere et al., 2016, pp. 418f). For example, in a 2011 survey study based on the European Values Study, it was found that individuals with a low level of political trust are much more accepting of law non-compliance than those with high levels of political trust [17]. In a US study on people’s hypothetical acceptance of mandatory vaccinations and quarantine in the event of a smallpox bioterror attack, it was found opponents of such measures were deeply distrustful of the government, expecting power abuses, discrimination, and general unpreparedness [14]. It is thus imperative for both social scientists and policy makers to take an interest in the thoughts and feelings prevalent in the population–especially, when these thoughts and feelings are an expression of varying levels of trust in political authorities.

According to Boin et al. [30], three factors determine the credibility of authorities during a crisis: prior trust, the initial response, and the timing of messages. Trust in authorities prior to a disaster is a major determinant in how initial crisis communication is received in the population since it forms the baseline of trust. The initial response is important in that it signals the capacity and willingness of authorities to handle the situation (see also [31]). During times of crisis, the population expects political authorities to take a dominant and active role in crisis management and to give guidance and direction [32]. The population may even be more forgiving of an exaggerated strategy than an insufficient one [30]. At the same time, if authorities downplay the gravity of a crisis or overplay their ability to solve the crisis, this type of rhetoric may initially be received positively, but could ultimately result in failed expectations [30] leading to frustration and disorientation in the short term and potential generalized distrust in the long term [30, 32, 33]. Furthermore, a lack of guidance and direction may lead citizens to seek out information from other sources, which are often faster and more comprehensive–especially in the case of social media–but less reliable. Citizens turning to alternative information sources could thus increase non-compliance [32, 34–38]. Lastly, the timing of messages is of considerable importance. Overdue crisis information may lead to accusations of “covering up the truth”, or it may simply cause authorities to lose momentum (see also [39, 40]). Rather than softening the blow of crisis communication through unfounded positivity or the withholding of information, research indicates that openness about the risk surprisingly helps to reduce perceived risk in the population [39, 41, 42]. In addition, giving the population concrete directions on how to stay safe may limit the negative impact of the risk communication and help people feel more in control of the situation [41, 42].

In summary, research shows that political authorities managing a crisis should be transparent surrounding risk, take a dominant and active role in crisis management, and give concrete directions to give citizens a sense of control. Should the authorities fail to do this, they run the risk of the population developing distrust toward them and seeking out alternative information sources, which may lead to lower compliance with government measures and behavioral recommendations.

Social media data and natural language processing in disaster research

Social media have gained considerable popularity as a measure of public sentiment in recent years. While mass media like newspapers and news broadcasts have traditionally been used as indicators of prominent public discourse, they also have clear agenda-setting and gatekeeping functions; mass media-based findings on public discourses may thus indicate what news outlets consider important to the population rather than what the population itself considers important [43, 44]. Compared to the top-down communication of mass media, the direct bottom-up nature of social media allows researchers to study political information and discourses outside of institutional frameworks [45–48].

Social media research also provides improved research opportunities for disaster and health researchers. One of the main challenges of disaster research concerns restrictions in both timing and access: disaster events tend to happen suddenly and rapidly, limiting the amount of time researchers have to plan out their studies and gather data. And even when the researchers are onsite, the access to implicated individuals is often restricted and may change suddenly [49]. Social media posts, on the other hand, are often posted, not only immediately in the wake of disaster events, but as responses to the events, and their accessibility, as a rule, hinges on neither the geographical location of the researcher nor the date at which the data is collected [50–53]. As such, researchers may gain a unique insight into people’s thoughts and feelings at the time of occurrence of a disaster event, even when the data is gathered sometime in the future [54, 55].

Also in the domain of health have research tweets been used by epidemiologists to track the spread of diseases such as flu and dengue fever [20]. In recent years, opinion mining techniques in social research has provided researchers with even more opportunities to utilize social media. Natural Language Processing techniques (henceforth: NLP techniques) allow researchers to extract information about the thoughts and feelings present in a text corpus, or about social and semantic networks. For example, Pruss et al. [56] applied topic modelling to gain insights in how the Zika disease was discussed during the outbreak in the Americas in 2015. Insights from such work are similar to those traditionally found in qualitative social research, but the large quantities of data that NLP techniques can process lend to the method a higher level of generalizability as well as allowing for a more quantitative approach.

Data collection

Data was collected from the social media platform Twitter. The ethics of treating social media as public spaces has long been debated in the scientific community [82–84], and, while the distinction between public and private space is a lot less clearly defined online than in offline settings, there is general consensus among social media researchers that Twitter users do perceive Twitter as an unequivocally public space [85–87]; for this reason, informed consent was not obtained for this study. Nevertheless, in order to provide some level of protection of the Twitter users whose tweets were included in this study, usernames have been removed from the data made available in the supporting information for this study. The Twitter API limits the retrieval of tweets to tweets from the past week; in order to circumvent this limitation, we used the software GetOldTweets by Jefferson Henrique [88], which has gained some prominence among social media researchers in recent years, as it permits the collection of up to 100% of all available tweets under the search terms entered [85, 89–92]. As this method of data collection does not include any metadata, no personal data was collected within the context of this study. Although we aimed for a sample size of 100,000 per data set, this number was only actually achieved for the base sets (see below) of Germany, the Netherlands, and Sweden. As such, the data sets in the study, for the most part, constituted the entirety of the population of Twitter posts in the respective country during the respective period.

It must be taken into consideration that sentiment on Twitter may differ between countries, not only due to differences in culture, but due to differences in language use as well–or, indeed, as a function of the dictionaries used (while these were generally translated on the basis of the English dictionary, many sentiment scores cannot be transferred directly, and, as such, the dictionaries may differ considerably). A similar dislike for the sitting government may be expressed in one language—or even one geographical area—as "The government is doing terribly", but as "The government does not live up to expectations" in the next. For this reason, a base set for each country of up to 100,000 tweets from 2019 was collected to serve as a basis for the identification of changes in sentiment. The main data was collected within three phases during the initial outbreak, outlined in Table 1 below.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Data collection phases.

https://doi.org/10.1371/journal.pone.0253175.t001

The first phase marks the initial awareness of and reaction to COVID-19 as a potential threat. The World Health Organization’s reclassification of Covid-19 as a pandemic on March 11^th [93] marks the beginning of Phase 2, which was set to 2 weeks’ duration in order to capture public sentiment and discourse in relation to the initial lockdown. Finally, Phase 3 begins on April 11th—a month after the pandemic reclassification—and is set to 2 weeks’ duration, covering a time period, in which containment measures had been implemented for a month.

Search terms for the base set included commonly used terms to refer to the government and head of state. In the search terms for phases 1, 2, and 3, commonly used terms for COVID-19 were added. While Twitter does allow the user to search for geographically specific tweets, unfortunately this function is not only unavailable in the GetOldTweets software but, as only around 2% of tweets contain geographical information, choosing tweets based on geolocation also tends to severely limit the sample size [94]. As the majority of speakers within the respective language spheres of the four countries analyzed in this study belong to the countries in question, we operate under the assumption that our findings are applicable to the analyzed countries, even if there is a minority presence of speakers from other countries; nevertheless, in order retrieve as many relevant tweets as possible while excluding irrelevant ones, we have based our search terms on keywords specific to our analyzed countries while excluding keywords specific to countries not part of our analysis. For example, in the German tweets, one of the following terms had to be included in the tweet for the tweet to be retrieved: Merkel, Bundeskanzlerin, Kanzlerin, Regierung, or @AngelaMerkelCDU, whereas tweets including the terms referring to Austria or Switzerland (Österreich, Schweiz, Bundeskanzler, Bundespräsidentin, etc.–see S2 Appendix: Twitter search terms for the full inclusions and exclusions).

No cleaning was performed prior to the sentiment analysis, although due to a program error a few results could not be processed and were subsequently removed from the raw results before converting into percentage tables.

Cleaning prior to the topic modeling analysis involved two steps. First, all non-English letters had to be converted to English letters, as Context is not able to process non-English letters. An overview of this conversion can be found in Table 2 below.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Conversion of non-English letters.

https://doi.org/10.1371/journal.pone.0253175.t002

The second step involved removing repetitive patterns of phrasing or hashtags by the same user, for example if a single user begins all their tweets with #COVID #Government #Policy #BreakingNews, as these strings otherwise take up several of the topics in the topic modeling analysis on their own. It was taken care to avoid performing this cleaning in the case of retweets.

Method of analysis

In this study, the NLP techniques of dictionary-based sentiment analysis, using the software SentiStrength, and topic modeling, using the software ConText, were used [95, 96]. The combination of sentiment analysis and topic modeling allows us to obtain a nuanced picture on Twitter discourse, including both Twitter users’ emotional state and the topics being discussed (for previous uses of the combination of these methods, see [19, 97–100].

Sentiment analysis detects the overall sentiment of a text corpus by identifying words holding positive or negative valence, e.g. good or terrific (positive), and bad or disgusting (negative) (Liu, 2012). The sentiment analysis software scans the text corpus and scores individual texts (read: tweets) based on the presence of sentiment words as well as a limited number of grammatical modifiers. If, for instance, a sentiment word is preceded by a negating word, e.g. not or never, then the valence of the sentiment word is reversed: the sentence "Merkel is a competent leader" would be scored as positive (= "Merkel [0] is[0] a[0] competent[1] leader[0]”), but the sentence "Merkel is not a competent leader" would be scored as negative ("Merkel[0] is[0] not a[0] competent[–1] leader[0]”). In addition, some words may modify the strength of sentiment words, e.g. “Merkel[0] is[0] an[0] extremely[+2] incompetent[–1] leader[0]” = [0;-3]. A tweet may have a sentiment of [0;0], i.e. no sentiment, for example “Merkel[0] is[0] holding[0] a[0] speech[0] now[0]”; this does not exclude the tweet from the analysis. An average sentiment of 1 thus does not reveal if the data set consists mainly of tweets with a valence of 1 or mainly of no-sentiment tweets with a minority of tweets with a very strong valence (hence the breakdowns provided in Table 4 and, in greater detail, in S5 Appendix: Sentiment analysis results). The detection of sentiment words is conducted based on a sentiment lexicon, which lists all sentiment words of a language as well as their valence. Dutch and German sentiment dictionaries already exist, but for the purposes of this article, Danish and Swedish sentiment dictionaries were created (see S3 Appendix: Sentiment dictionaries (DK and SE)). Testing was done by manually inspecting a subsample of 100 tweets from the main sample and adjusting the dictionaries until the analysis was satisfactory. The results of the tests can be found in Table 3 below.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Pre- and post-test scores for dictionary test sets.

https://doi.org/10.1371/journal.pone.0253175.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Sentiment analysis results in detail.

https://doi.org/10.1371/journal.pone.0253175.t004

On average, the sentiment dictionaries have an identification rate of 69.5%. Although the score will likely be lower when applied to a new data set, this result is sufficient to identify the general sentiment of a data set. SentiStrength uses a scale from 0 to 4 for positive sentiment and 0 to -4 for negative sentiment, 0 being the absence of sentiment and 4/-4 being the strongest possible sentiment. The output consists of an overview of the frequencies of different sentiment score combinations as well as the average positive sentiment (henceforth: APS) and the average positive sentiment (henceforth: APS) for the entire text corpus.

Topic modeling analysis is a machine learning technique, which identifies clusters of commonly co-occurring words in a text corpus [101–103]. The researcher sets the number of topics K expected to be in the text corpus and the number of words-per-topic desired a and receives K lists of a words. Based on each list, referred to as a topic, the researcher then identifies the common theme of the words, thereby uncovering the underlying discourses of the text corpus. The words of each topic are sorted according to fit (best fit first), and each topic is given a weight, which denotes the average fit of the overall topic to the text corpus [96]. As topic modeling results based on short texts such as tweets tend to include considerable noise (i.e. words not informative about the identified topic) [104–106], we cross-referenced our findings by manually looking up the tweets in question to ensure that our interpretations of the topics were accurate. As a complement to the topic modeling analysis, we consider corpus statistics (prominent words) and bigrams where appropriate (see bigram and corpus statistics results in S4 Appendix: bigrams and corpus statistics).

Denmark

Denmark initially experienced a negative turn in sentiment with APS dropping from 0.613 (base levels) to 0.503 (p≤0.05 (two-samples t-test)), and ANS rising from -0.805 to -0.997 (p<0.001). As the country entered phase 2, however, APS rose to 0.650, while ANS differed significantly from neither base levels nor phase 1. In the third phase, sentiment seemed to have stabilized: APS fell to 0.618, which is statistically significantly different to neither base levels nor the previous phase, while ANS stayed around phase 2 levels, which is slightly higher than base levels.

Corpus statistics and bigrams revealed the presence of considerable levels of uncertainty and anxiety. The topic modeling analysis showed that people worried about the severity of Covid-19, the country’s intensive care capacity, and how to protect vulnerable citizens, (Phase 1, topics 4, 10, and 13). In addition, the government and the prime minister were accused by both experts and laymen of inciting panic (7) and attempting to score cheap political points by appearing decisive (15).

As reflected in the sentiment analysis, however, public opinion on Twitter shifted in a positive direction with Phase 2: the second and third most prominent topics consisted mainly of messages thanking or praising the prime minister Mette Frederiksen and the government (the most prominent topic consisting mainly of noise (i.e. words without topical meaning). Compared to Phase 1, topics relating to uncertainties also diminished; instead, political negotiations (phase 2, topics 5, 10, and 15) and political messages (6, 7, 8, 9, and 14) were prominent topics, especially those relating to containment measures (6, 7, 8, 13, 15). Criticism in this phase was similar to that of the previous phase, although slightly more severe: the government was described, not just as poor communicators (17), but as deceptive (19), and not just as being opportunistic, but as poor leaders (17) and, in relation to a piece of emergency legislation allowing the police to enforce compulsory treatment for Covid-19 victims, as violating democratic rights (15). This is peculiar, as not only did ANS not increase during phase 2, but both the number of tweets with a higher negative than positive sentiment score and the number of tweets with strong negative sentiment fell. This, then, indicates that the number of dissatisfied Twitter users diminished, but that those that remained, were either more pointed in their criticism or, conversely, more ambiguous (i.e. having sentiment scores with equally strong positive and negative sentiment). While some Twitter users thus because more radicalized in their view as the crisis continued, the results indicate that many also became more nuanced.

While the third phase saw no statistically significant changes in sentiment score, the topic modeling analysis included fewer words of praise toward Frederiksen or the government; most praises were given in reaction to critical comments (phase 3, topic 14). Rather, discussion was centered around reopening efforts (1, 3, 5, 6, 7, 11, 20) with a focus on children returning to school (5, 20) and vulnerable citizens (3, 7). Criticism did not differ thematically from the previous phase: the government was reproached for their strategy, which was perceived to be unilateral and excessive (13), as well as for political opportunism (10).

Germany

Germany experienced a negative turn in sentiment in Phase 1: the APS of the phase dropped from 0.385 (base levels) to 0.347 (p<0.001), and the ANS rose from -0.604 to -0.637 (p≤0.01). This pivoted into a positive turn in the second phase, however: APS rose to 0.375 (p≤0.01) ANS dropped to -0.506 (p<0.001). In Phase 3, the sentiment scores reverted to phase 1 levels, although with a lower ANS: APS dropping to 0.343 (p<0.001 compared to phase 2), and ANS increased slightly to -0.542 (p<0.001 compared to Phase 2).

Fear and uncertainty characterized Phase 1: the corpus statistics included prominent words like spread, measures, outbreak, panic, anxiety, economy, and worries. As Germany had its first case of Covid-19 as early as end-January, the initial patterns of contagion were still in the public eye (Phase 1, topics 4, 5, 6, 8, 9, 13, 15, 17). Twitter users did not express much faith in the German Chancellor Angela Merkel during this phase and called for her and the government to be more proactive and communicative (11, 14, 16). Some characterized Merkel as incompetent (10), and conspiracies accusing the chancellor of being in cahoots with China as the covert creators of Covid-19 even gained some traction (18, 19).

In Phase 2, fear and uncertainty seemed to diminish: while prominent bigrams and the corpus statistics still contained indicators of uncertainty (measures, crisis, battle, economy), this phase was characterized to a higher degree by expressions of solidarity and social awareness (curfew, agreement, #Stayathome, solidarity, positive, #wirbleibenzuhause). The focus of German Twitter had moved toward political communication (Phase 2, topics 3, 8, 11, 14, 17)–especially statements by Merkel, e.g. the statement “Es ist ernst" (“This is serious”) and the description of the Covid-19 crisis as the biggest challenge since World War II–as well as concrete political decisions (5, 6, 8, 9, 19, 20). While praise of Chancellor Merkel or the government did not make up a separate topic, words like solidarity or the hashtags #StayAtHome or #wirbleibenzuhause (#westayathome) were widely used in conjunction with the retweeting of Merkel’s statements, as was the bigram “Danke Merkel”. Criticism mainly related to a continued desire for the government to be more proactive (2, 15) and more communicative (16).

In Phase 3, focus continued to be on political communication, in particular Merkel’s so-called “Öffnungsdiskussionsorgien” (opening discussion orgies)(Phase 3, topics 4, 13), with which Merkel had characterized the discourse in the individual federal Länder, and which many Twitter users felt was inappropriate in tone (13). The biggest topic of discussion concerned the reopening of the country post-lockdown (3, 4, 5, 7, 9, 10, 13, 16, 17, 20), which was either seen as necessary, and the government as having gone too far already (14, 16, 20), or as premature (3, 4, 13, 17). The portrayal of Merkel and the government was mostly critical (topics 16, 20), and with phrases like “Danke Merkel” and “Danke Regierung” being almost absent. In addition, the conspiracy topic from Phase 1 resurged (topic 12). Finally, the tone of solidarity found in Phase 2 greatly diminished in this phase: Solidarity went from being used 346 times across 26107 tweets (1.3% of tweets, assuming one mention per tweet) to 55 times across 7270 tweets (0.7% of tweets, assuming one mention per tweet), #wirbleibenzuhause from being used in 1% of tweets to 0.4%, and #StayHome from 0.4% to 0.3%.

The Netherlands

Dutch Twitter was characterized by increasing polarization in sentiment as the crisis progressed. The proportion of no-sentiment tweets fell consistently over time from 35.95% to 25.64%, and APS and ANS both became progressively stronger. The first phase was characterized by a small, but significant increase in polarization: APS rose from 0.424 (base levels) to 0.450 (p≤0.05), and ANS rose from -0.820 to -0.879 (p<0.001). Phase 2 subsequently saw a rise in APS to 0.524 (p>0.001), ANS not changing significantly, whereas Phase 3 saw a strong increase in ANS from -0.874 to -1.022, while APS stayed at the same level.

The bigrams and corpus statistics of this phase indicated that the Dutch Twitter population was on high alert (necessary, quarantine, infections, deaths, crisis), which was further corroborated by the topic modeling analysis (Phase 1, topics 1, 3, 5, 6, 10, 14, 15, 17). The phase was characterized by a strong call for political action (1, 13, 15): in particular, the lack of immediate introduction of containment measures had Twitter questioning the priorities and attitude of the Minister-President Mark Rutte and the government (7, 12, 15, 17), and some worry regarding the vulnerable population was expressed (2, 6). Two critical articles, “Rutte has the wrong priorities” (Telegraaf.nl) and ““82 corona patients in the Netherlands: ultra-screw-up Rutte has completely lost control” (Dagelijkse Standaard) were in heavy rotation (see also [107, 108]).

The major conversation theme of Phase 2 was Rutte’s address to the nation on March 16^th (Phase 2, topics 2, 4, 10, 11, 17). People particularly criticized Rutte’s remarks on the effectiveness of herd immunity (3, 19), and many perceived the decision as Rutte prioritizing the economy above human lives (5). Conversely, the word good was used commonly in relation to the March 16^th; as such, the rise in APS may be due to many users being satisfied with Rutte’s speech. In addition to the March 16^th address, outbreak containment measures such as social distancing, the closing of schools, and whether the country should close borders, were discussed (7, 12, 13) with people expressing worry about infectivity and intensive care resources (8, 16).

In Phase 3, containment measures–especially face masks and corona tracking apps (Phase 3, topics 2, 4, 7, 13, 18, 19)–continued to be widely discussed (2, 4, 10, 18), as was the continued threat of Covid-19 (1, 5, 8, 17, 19). Rutte was the subject of some criticism, although this was both less prominent than in the previous phase and included people believing the eventual lockdown to have been too excessive (9, 20).

Sweden

Phase 1 in Sweden saw a slight, but significant increase in ANS compared to base levels from -0.933 to -1.003 (p≤0.05), with no significant difference in APS. Phase 2 saw no statistically significant change in sentiment compared to Phase 1, which may be explained by the fact that no new policy was introduced until far into the second phase; nevertheless, there was a significant increase in APS compared to base levels. In phase 3, however, ANS dropped from -0.964 to -1.047 (p≤0.01) with no significant change in APS.

Most topics during Phase 1 were related to political performance (Phase 1, topics 2, 4, 5, 6, 7, 8, 9, 10, 14, 18) and political communication, especially press briefings (4, 7, 9). Public opinion on Twitter was generally unfavorable, with the government being described as incompetent (5), opportunistic (18) or gambling with people’s health (15, 16). Call for political action was also prevalent with many Twitter users questioning keeping the borders open (3, 10) and allowing travel to high-risk zones.

During the second phase, the word good was very prominent (165 mentions across 2595 tweets), which, together with the rise in APS, indicates a positive shift in public opinion on Twitter. The major themes of the phase were new policies–a business bailout (Phase 2, topic 4), remote classes for high schools and universities (7), and a ban on gatherings of more than 500 people (6)–and criticism of the government centered around them not doing enough (9, 12, 15, 17), especially for the vulnerable population (9, 12). Calls for political action became more concrete, with users calling for a country-wide quarantine (8), better healthcare equipment (9), and a ban on refugees from high-risk zones.

The topic modeling analysis suggests that this negative turn in sentiment in the third phase was motivated by a high mortality rate: not only was the Swedish death toll high (Phase 3, topics 13, 18), but it was notably higher than those of neighboring countries (10). The topic modeling analysis further indicates a waning trust in the government: calls for political action and political transparency continued (4, 6), the government was described as failed or having the wrong priorities (11, 19), and a statement by the retired virologist and researcher Jan Stillström, calling the main designer of the Swedish Covid-19 strategy, the state epidemiologist Anders Tegnell, a “politically appointed fraud”, was circulated.