COVID-19 Communication Management on Facebook Pages of Local Governments

Introduction

Recent studies document an increase of the usage of social media (SoMe) by local governments for several purposes, including engaging citizens in public affairs, restoring confidence in institutions, enhancing transparency and accountability, and improving public services (Bonsón et al., 2017; Guo et al., 2021; Knox, 2016). An additional purpose s is disaster and crisis management (Chen et al., 2020; Guo et al., 2021; Tang et al., 2021). In such delicate situations, SoMe may allow governments and public agencies to effectively and timely enhance citizens’ understanding and engagement, assess the evolving stages of the disaster, provide emotional support, reduce uncertainty, and prevent/correct false rumors (Guo et al., 2021; Hughes & Palen, 2012; Kaewkitipong et al., 2016; Lin et al., 2016; Mori et al., 2021; Reuter et al., 2018).

The recent COVID-19 pandemic represents an unprecedented setting for assessing how local governments have been managing their communication through SoMe to deal with the emergency. In this study, we initially seek to identify the COVID-19-related topics addressed by the main Italian municipalities on their official Facebook pages. Then, we assess how these topics affect the tone of the communication and how much they are discussed online in accordance with the pandemic’s severity, expressed by the number of COVID-19 positive cases in the municipal area.

Among other conclusions, this information may allow inferring whether political incentives and, more specifically, the quest for political consensus, play a role in the municipal COVID-19 Facebook communication. In this regard, previous studies provide evidence of an opportunistic usage of Facebook by local governments of Western European countries for self-promotion and political marketing through one-way communication (Agostino & Arnaboldi, 2016; Bonsón et al., 2017; Silva et al., 2019). These opportunistic aims may at least partially undermine achieving more desirable objectives such as engaging citizens in political and social issues and, more importantly in our study, effectively using Facebook to manage the pandemic (Löffler & Bovaird, 2019; Mori et al., 2021; Picazo-Vela et al., 2016).

Italy represents an interesting setting given that it is among the first European countries to record one of the highest numbers of infections in the world. In addition, Italy is among the first countries to react positively to the pandemic (Mori et al., 2021). It should be noted that in Italy communication by public administrations using Facebook is not mandatory and, therefore, it could be considered as a form of voluntary e-disclosure. ¹ Studies document the increase of Facebook use for communication between local governments and citizens in Europe, Australia, and the US (Bonsón et al., 2017; Lev-On & Steinfeld, 2015; Omar et al., 2014). Facebook is the favorite SoMe platform of local governments to interact with citizens and engage them due to its greater popularity relative to other SoMe platforms like Twitter, among others (Bonsón et al., 2017; Guillamón et al., 2016; Larsson & Skogerbø, 2018; Silva et al., 2019). For example, Haro-de-Rosario et al. (2018) find that the 80 largest Spanish municipalities engage citizens more with Facebook than with Twitter. In Italy, as of March 2021, Facebook was the social network with the highest penetration rate with 36 million users, while Twitter ranked only fifth with 11.2 million users (AGCOM, 2021). In this line, the usage of SoMe like Facebook by Italian municipalities has significantly grown in recent years (Guillamón et al., 2016) and it is recognized by several laws starting from Law 150/2000. This Law states that Italian public administrations can use a wide variety of suitable tools for their communication and information activities including digital media (Mori et al., 2021).

That said, to extract latent topics from posts published on Facebook, we apply an automatic topic modeling procedure called latent Dirichlet allocation (LDA) (Blei et al., 2003) on a sample of 76,139 posts published on the official Facebook pages of 103 Italian provincial capital municipalities during the COVID-19 pandemic period from 1 March 2020 to 26 March 2021. This analysis allows us to identify nine main topics, of which two are related to COVID-19. More specifically, based on the most likely associated terms, we label the first COVID-19 topic as COVID-19 restrictive measures and the second as COVID-19 support measures. Using regression models, including both municipality and year-week fixed effects, we find that the prevalence of the topic of COVID-19 restrictive measures negatively affects the tone of communication; this was determined through a lexicon-based sentiment analysis procedure that considers contextual valence shifters in the polarity computation. In contrast, the prevalence of the topic of COVID-19 support measures positively affects the tone of communication.

In addition, we find that when the number of COVID-19 positive cases grows in the municipal area, municipalities tend to downplay negative information on COVID-19 restrictive measures on Facebook. They also have a relatively greater tendency to promote positive information on COVID-19 support measures, as indicated by the variations of the corresponding topic prevalence. This tendency may provide evidence of strategic communication by municipalities aimed at fostering positive perceptions in their citizens about municipal incumbents’ performance in effectively dealing with the pandemic through suitable support measures. Furthermore, we find that this tendency is strengthened for municipalities led by right-wing mayors and older mayors. The role of political incentives is confirmed by estimations revealing that the identified tendency is weaker for mayors with a stronger political position granted by a higher margin of victory in the latest municipal elections. The tendency is stronger for mayors in their first term of office that, unlike mayors in their second term, are allowed to run for re-election in the following elections.

Prior studies examine the increasing usage of SoMe by central and local governments and supranational entities (e.g., UN and WHO) to manage disasters and crises including the recent COVID-19 pandemic (Chen et al., 2020; Guo et al., 2021; Li et al., 2020; Mori et al., 2021; Tang et al., 2021). Some of these studies analyze barriers to the effective SoMe use by public authorities to manage disasters (Plotnick & Hiltz, 2016; Reuter & Kaufhold, 2018). These barriers mostly consist of lack of resources, qualified staff, and formal policies as well as doubts about the reliability of the information posted by the public (Reuter et al., 2018). In contrast, political influences over public communication on SoMe are generally not included among these barriers.

In this regard, to the best of our knowledge, our study is the first to analyze how local governments strategically manage their COVID-19 communication on their official Facebook pages based on the association between COVID-19 topics, automatically extracted through the LDA algorithm, and the number of infections in the municipal area. We are the first to provide empirical evidence of political incentives that may affect this communication, consistent with studies that, in other contexts, identify political opportunism in the usage of Facebook by European local governments (Agostino & Arnaboldi, 2016; Bonsón et al., 2017; Silva et al., 2019).

Importantly, our results may raise concerns regarding the opportunistic and biased communication of local governments through SoMe. Indeed, the opportunistic communication on SoMe by local governments for self-promotion and political marketing may hamper the effectiveness of SoMe as a disaster management system, as well as SoMe’s proper usage to enhance transparency and engage citizens in political and social issues. These concerns should be considered by public authorities engaged in formulating guidelines or regulations on the appropriate usage of SoMe by local governments.

The remainder of the paper proceeds as follows: Section 2 undertakes the literature review and develops the main hypotheses; Section 3 describes the sample data and the empirical strategy; Section 4 presents the empirical results; and Section 5 presents the conclusion.

Literature Review and Hypotheses

Recent studies show an increase in the usage of SoMe by governments, public agencies and supranational entities (e.g., UN and WHO) to manage disasters and crises (Chen et al., 2020; Guo et al., 2021; Tang et al., 2021). These studies fall under the area of research called crisis informatics by some scholars (Palen & Anderson, 2016; Reuter & Kaufhold, 2018; Reuter et al., 2018). Reuter et al. (2018) examine some case studies of SoMe use to manage disasters and identify different roles and usage patterns depending on the SoMe sender and the recipient of information. In this respect, Reuter & Kaufhold (2018) present a classification matrix including the four possible information flows between the two roles of authorities and citizens during emergencies. In our study, we specifically focus on the information flow from authorities to citizens aiming to provide updates and advice during emergencies. As a limitation of the crisis informatics research, Reuter et al. (2018) identify the strong bias toward studies that examine Twitter and datasets in the English language. The authors attribute this bias to the greater easiness to obtain data from Twitter than from other SoMe, the primary English language of most of the researchers involved, and the available textual analysis techniques that are far more developed for the English language. Therefore, the authors call for future research on how other SoMe in different languages are used in crisis situations around the world.

In this vein, F. Liu and Xu (2018) examine the Facebook pages of three public officials responsible for managing three disasters that occurred over the period 2013 to 2015. Using a thematic analysis, the authors find that public officials’ posts can be grouped into four main categories: “official updates on the situation, recommendations for the local population, information regarding recovery procedures and methods, and replies to victims’ queries and needs.” In this regard, Kaewkitipong et al. (2016) document that the prevalence of each topic posted on SoMe may be adapted to fit the knowledge-sharing needs in each of the three phases (pre, during, and post) of the crisis.

In the context of the COVID-19 pandemic, Li et al. (2020) document the usage of Weibo—the largest short messaging SoMe platform in China—by the Chinese government to mitigate information asymmetry with local governments and citizens. More specifically, the Chinese central government used Weibo to coordinate its COVID-19 related actions and policies with local governments, collect timely information at the ground level, assist citizens, speed up relief efforts, and mobilize citizens and non-profit organizations to support its responses. In the same vein, Mori et al. (2021) examine the official Facebook page of the 11 Italian municipalities (capital cities of provinces) with the highest COVID-19-induced mortality rates from 31 January 2020 to 3 June 2020. The authors manually perform a qualitative content analysis and collect some quantitative indicators related to the number, frequency, and other features of the posts. They find that municipalities seem not to capitalize on the full potential of SoMe given that both quantity and frequency of published posts did not show significant changes during the pandemic. However, the content of municipal posts became richer and less formal, and it was rapidly adapted to the different circumstances and phases of the pandemic, revealing the absence of any structured communication strategy. In addition, the tone of communication varied from prescriptive—during the most critical phases—to informative as soon as the number of infections significantly declined. It should be noted that the authors do not analyze if the prevalence and the tone of COVID-19-related posts strategically adapted to the severity of the COVID-19 spread.

The usage of Facebook by Italian municipalities to inform about COVID-19 is consistent with previous studies finding that local governments of Western European countries mostly use SoMe to provide useful information to citizens through one-way communication (Agostino & Arnaboldi, 2016; Bonsón et al., 2017). In such communication, the purposes of self-promotion and political marketing may tend to prevail over the more desirable objective of engaging citizens in political and social issues through an interactive two- way dialog (Löffler & Bovaird, 2019; Picazo-Vela et al., 2016).

In the same vein, studies document that the majority of US authorities increasingly value the suitability of SoMe to manage emergencies or disasters (Reuter & Kaufhold, 2018; Reuter et al., 2016). However, the priority use of SoMe by the US government agencies is mainly to disseminate information rather than to engage citizens and obtain an overview of the crisis through a two-way communication (Lin et al., 2016; Reuter & Kaufhold, 2018; Reuter et al., 2016). Indeed, the information received from the public is not usually considered verifiable and trustworthy enough to be incorporated into decision-making processes (Plotnick & Hiltz, 2016).

Based on previous studies on the usage of SoMe by governments to man- age disasters (Chen et al., 2020; Kaewkitipong et al., 2016; F. Liu & Xu, 2018), we expect COVID-19 topics on Facebook of Italian municipalities to mostly deal with restrictive measures to contain the pandemic and support measures for citizens. Due to the previously-found lack of a structured com- munication strategy through Facebook by Italian municipalities (Mori et al., 2021), the tone of posts more focused on restrictive measures might be more prescriptive and, therefore, more negative. Conversely, the tone of posts more focused on support measures might be more supportive and, therefore, more positive. Hence, the first hypothesis is:

Previous studies show that local governments tend to minimize or avoid reporting negative information that may signal their incapacity to effectively deal with crises. This negative information may compromise promotion opportunities for public officials and re-election opportunities for local incumbents (Li et al., 2020). Indeed, natural disasters such as COVID-19 could be an opportunity for citizens to assess the competence and performance of the incumbents based on the quality of the politicians’ disaster response (Ashworth et al., 2018; Masiero & Santarossa, 2021). More specifically, based on the retrospective voting theory (Achen & Bartels, 2018), voters will reward incumbents in the next elections if they perceive their response as adequate to alleviate the disaster’s negative effects on public welfare. Conversely, voters will punish the incumbents if they perceive them as unable to manage the disaster promptly and effectively. However, incumbents may manage to influence voter perception about their disaster management performance through an opportunistic communication strategy on SoMe. Indeed, previous studies show that voters may be emotional, not adequately informed about reality, and, therefore, prone to be manipulated by demagogical or populist politicians though biased political communication (Achen & Bartels, 2018).

Studies show that, following a disaster, the ability of incumbents to support citizens through adequate relief spending is the main booster for political consensus among voters. For example, electoral returns for incumbents arose from the allocation of relief spending in Colombia, following the disaster caused by the 2010 to 2011 rainy season (Gallego, 2018); in Germany, following the 2002 Elbe flooding (Bechtel & Hainmueller, 2011); and in Italy, following the destructive earthquakes between 1993 and 2015 (Masiero & Santarossa, 2021).

However, prior research documents that the effectiveness of government disaster relief programs in forging political consensus can be amplified through strategic political communication using traditional media and SoMe aimed at promoting public spending for disaster support (Klomp, 2020; Masiero & Santarossa, 2021). This strategic promotion of disaster relief may be more likely in disasters like COVID-19 that find politicians unprepared and with insufficient resources and experience to effectively manage the disaster (Masiero & Santarossa, 2021).

In addition, Facebook posts on support measures regarding COVID-19 disaster may present a positive emotional tone. Some studies find that posting on positive topics may enhance citizen engagement, political consensus and attract voters (Nave et al., 2018; Stieglitz & Dang-Xuan, 2013). For example, Gerbaudo et al. (2019) find that, in the 2017 UK general elections, the posting strategy focused on topics with positive emotional tone adopted by the Labor party and its leader on official Facebook pages attracted voters by engaging users. Positive posting mostly relied on more optimistic content such as promises of improvement of welfare and public services rather than stressing fear-inducing issues of national security, terrorism, and immigration. In the same line, Ceron and D’Adda (2016) document that a positive political campaign, focused on distributive promises through Italian parties’ official Twitter accounts, enhanced voting intentions expressed on Twitter in the 2013 Italian general elections. Using a systematic random sample of 125 U.S. cities, Zavattaro et al. (2015) find that local governments that adopt a positive sentiment tone in their Twitter communication are more likely to engage citizens and encourage their participation than cities that simply push information on recipients.

That said, prior evidence on the opportunistic usage of Facebook by European local governments for self-promotion and political marketing through one-way communication (Agostino & Arnaboldi, 2016; Bonsón et al., 2017; Silva et al., 2019) may support the tendency of Italian municipalities to post more on positive topics, while minimizing posts on negative topics.

When the severity of the pandemic—measured by the number of COVID-19 positive cases—grows, citizens may fear the negative personal effects of the tightening of COVID-19 restrictions. Therefore, municipalities may post less or, at least, avoid posting more on COVID-19 restrictive measures given that the communication regarding these measures may be perceived by citizens to be more prescriptive and, therefore, more emotionally negative (Mori et al., 2021). These negative emotions may erode political support for municipal incumbents, who may be perceived as incapable of effectively managing the pandemic. However, a higher number of positive cases may lead municipalities to post more on COVID-19 support measures to enhance citizens’ perception about their effective pandemic response through a strate- gic positive communication (Gallego, 2018; Masiero & Santarossa, 2021). Hence:

Based on Italy’s municipal electoral system, voters can express a direct choice for the mayoral candidate or the coalition of parties supporting the mayoral candidate. ² The electoral system attributes great importance to mayoral candidates and, ultimately, guarantees the majority of seats to the party supporting the elected mayor. Therefore, the mayor is likely to have a signifi- cant influence on the municipal communication strategy through SoMe (Larsson & Skogerbø, 2018; Silva et al., 2019). This is confirmed by the fact that several capital municipalities disclose on their institutional websites a SoMe policy or related information ³ stating that the content of municipal SoMe accounts is managed by the municipal administration or the cabinet office reporting directly to the mayor.

To account for this influence, we test whether the mayor’s ideology, age, gender, and political incentives affect the hypothesized associations between the prevalence of COVID-19-related topics on municipal Facebook pages and the number of positive cases in the municipal area.

Regarding a mayor’s ideology, ⁴ by examining political Facebook posts by Israeli political actors during the 2015 Israeli campaign, Nave et al. (2018) provide further evidence on the more positive content of left-wing candidates’ successful posts relative to right-wing candidates. More specifically, positive feelings, based on humor, optimism and hope, are more typical of successful left-wing posts given that these feelings are more consistent with leftist values and their perceived identities compared to rightists (Caprara et al., 2006). On the basis of a sample of 90 democratic countries between 1985 and 2013, Klomp (2019) shows that left-wing governments allocate more public support funds in the aftermath of a disaster than right-wing governments. The author explains his results based on the partisan theory stating that left-wing governments are more committed to fiscal policies aimed at fighting poverty or inequality given that their constituency mostly consists of the working class and low-income citizens (Potrafke, 2017). However, more generally and regardless of the right-wing differences, political parties may be more willing to provide public support if their constituency is especially vulnerable to a specific disaster (Klomp, 2019). In Italy, COVID-19 has sig- nificantly affected small- and medium-sized businesses and self-employed workers that may mostly support right-wing parties. We expect the relevance attached to COVID-19 public support by incumbents to be reflected in the prevalence of the related topic in their Facebook communication. However, because of our previous conflicting arguments, we do not make any predic- tion regarding the moderating effect of mayor’s ideology on the relationship between COVID-19 support topic and COVID-19 positive cases:

Previous social psychological research finds that, relative to progressives, conservatives tend to show greater sensitivity to several threats, including disease threat, and perceive the world as more dangerous (Conway et al., 2020; Jost, 2017; Perry et al., 2013). Specifically, Jost et al. (2003) show that core values of political conservatism, such as traditionalism and acceptance of social inequalities, are associated with the desire to manage uncertainty and a stronger perception of threats.

However, partisan media coverage may manipulate the public perception of specific threats and make conservatives less sensitive to those threats than progressives. For example, in the US, partisan media coverage may be responsible for progressive Democrats’ higher sensitivity to the threat of climate change relative to conservative Republicans (Carmichael et al., 2017). Similarly, Calvillo et al. (2020) attribute the lower perception of COVID-19 threat by conservatives than progressives in the US to partisan media coverage, polarization induced by political leadership, and conservatives’ less accurate discernment between real and fake news. In this regard, Havey (2020) finds that political conservatives predominate the misinformation discourse and conspiracy theories regarding COVID-19 on Twitter. However, Conway et al. (2020) argue that American conservatives are less concerned about COVID-19 and, therefore, less likely to follow government recommendations (van Holm et al., 2020) because they feel that the pandemic’s effects and related public interventions such as government restrictions will hurt their ideological ends.

The authors suggest that the match between ideological ends and the outcomes of the pandemic may help to predict public perceptions about the pandemic. The authors acknowledge that, in other socio-cultural contexts, conservatives may not share American conservatives’ strong dislike of public interventions related to COVID-19. In these contexts, the relationship between ideology and perceived threat would not be affected by this match mechanism.

Studies find Italians’ relatively high level of acceptance of the national government’s restrictive measures to contain the virus starting on 23 February 2020, just a few days after the first case of COVID-19 contracted in Italy was diagnosed (Ceccato et al., 2020; Motta Zanin et al., 2020). In addition, at the beginning of the pandemic, the Italian government, through the Ministry of Health and public health authorities, engaged in a digital communication strategy through SoMe to inform citizens and counteract misinformation on COVID-19 (Lovari, 2020). Therefore, in Italy, the political leadership does not induce the same level of political polarization around COVID-19 topic as in the US.

On the one hand, within the Italian socio-political context, the traditional tendency of rightists to show greater sensitivity to the COVID-19 threat than leftists (Jost et al., 2003) may not be moderated by partisan media coverage or political polarization at the same level as in the US. Hence, Italian rightists may accept restrictive measures to contain the pandemic threat. On the other hand, COVID-19 restrictive measures may more negatively affect interests and ideological ends of right-wing constituencies, including small- and medium-sized businesses and self-employed workers. Therefore, we are unable to make any prediction regarding the moderating effect of mayor’s ideology on the relation between COVID-19 restrictive measures and COVID-19 positive cases; we address this issue with the fol- lowing null hypothesis:

Regarding a mayor’s age, studies carried out in different countries using questionnaires find that older adults show higher perception of COVID-19 risk than younger adults (Barber & Kim, 2021; Bruine de Bruin, 2021). This may also arise from the fact that the COVID-19 fatality rate significantly increases with age (Remuzzi & Remuzzi, 2020). However, higher COVID-19 risk perception is not always associated with greater concerns and consequent behavioral changes. For example, Barber and Kim (2021) find that, in the US, older men are less concerned than younger counterparts and implement fewer behavioral changes as precautions against COVID-19. On the other hand, using a survey, Ceccato et al. (2020) find that in Italy older adults show more positive attitudes and lower negative emotions toward COVID-19 than younger adults. However, older adults are more favorable about restrictive measures, more confident about COVID-19-related information received from the media and official sources, and estimate a longer time for solving the emergency.

More positive attitudes of older mayors toward COVID-19 and their higher reliance on official COVID-19-related information may lead them to perceive COVID-19 topics as less negative than younger mayors and, therefore, provide more information on COVID-19 and related support measures through Facebook when the number of positive cases increases. Conversely, older mayors may implement fewer behavioral changes and may have a more negative perception about the socio-economic effects of restrictive measures (Barber & Kim, 2021). These factors may induce older mayors to promote COVID-19 restrictive measures relatively less through Facebook when the number of infections grows. Hence, the sixth hypothesis is:

Regarding a mayor’s gender, studies conducted in several OECD countries find that, despite higher fatality rates of COVID-19 for males than females, females present a higher perception of COVID-19 risks than males (Rana et al., 2021). Therefore, females are more likely to see COVID-19 as a very serious problem, take precautionary measures and agree with restrictive public policy measures (Abdulmuhsin et al., 2022; Rodriguez-Besteiro et al., 2021). In addition, females show higher prevalence and severity of COVID-19-related anxiety, depression, and stress symptoms (N. Liu et al., 2020). These difference in risk perception between males and females are also reflected in behavioral differences between male and female leaders. For example, countries led by female rulers (e.g., New Zealand, Germany, Iceland) have shown greater effectiveness and adopted clearer communication strategies to tackle the pandemic than countries led by male rulers (e.g., US, UK, Brazil) (Garikipati & Kambhampati, 2020).

That said, we expect female mayors to show greater sensitivity to COVID-19 than male mayors. This greater sensitivity may be reflected in Facebook communication that is less opportunistic, more supportive, more transparent and in line with the evolution of the pandemic. The seventh hypothesis predicts:

Regarding a mayor’s political incentives, previous research finds that higher political competition in local elections induces local governments to increase promotion of their activities on their official SoMe and websites to strengthen political consensus among citizens (Silva et al., 2019; Tavares & da Cruz, 2020). This may be reflected in a more strategic communication regarding COVID-19 topics on SoMe. Studies (Araujo & Tejedo-Romero, 2016; Silva et al., 2019) use the margin of victory in the latest local elections as a proxy for the political competition. Indeed, a higher margin of victory over the second most-voted candidate is commonly associated with a less competitive political environment, and vice versa.

Another mayoral political incentive may arise from the fact that according to Italian law, ⁵ a mayor in office for two consecutive terms cannot be re- elected. Therefore, to increase her/his chances of re-election, a first-term mayor may have stronger incentives to use SoMe for self-promotion and political marketing (Agostino & Arnaboldi, 2016; Bonsón et al., 2017; Silva et al., 2019). Again, this may lead to more strategic communication regard- ing COVID-19 topics on SoMe. Hence, our final hypothesis suggests:

Data and Empirical Strategy

Variables and Baseline Regression Models

To test our hypotheses, we build two main dependent variables for our base- line regression models. The first dependent variable is the prevalence (pro- portion) of COVID-19-related topics (TOPIC_COVID) within each post published on municipal Facebook pages during the study period. To, esti- mate the variable TOPIC_COVID, we adopt a topic modeling algorithm called latent Dirichlet allocation (LDA), initially proposed by (Blei et al., 2003). Topic modeling algorithms aim at uncovering latent topics in large corpora of documents by identifying clusters of words co-occurring within the documents and assumed to refer to the same semantically interpretable topic (Roberts et al., 2014). The output of the LDA estimation consists of two main matrices. Specifically, the document-topic matrix shows the probability (prevalence) of each topic (column) for each document (row), whereas the topic-term matrix shows the probability of each term (column) referring to each topic (row). In the document-topic matrix, the sum of the prevalences of the identified topics is always equal to 1 for each document (post). Analysts use the topic-term matrix to assign a descriptive label to each topic by interpreting the most probable terms for each topic based on previous theory and contextual knowledge. ¹⁰

The second dependent variable for our baseline regression models is the polarity score (tone) of municipal Facebook posts (TONE) during the examined pandemic period. To compute the polarity tone of Facebook posts, we adopt a lexicon-based sentiment analysis procedure. Specifically, we count the words of each post included in an Italian word list that assigns a negative or positive sentiment (tone) to each word based on an interval ranging from −1 (most negative) to +1 (most positive). Posts with a relatively higher frequency of positive words are assumed to be optimistic. Conversely, posts with a relatively higher frequency of negative words are assumed to be pessimistic.

Previous studies adopt several English lexicons for sentiment analysis that are specifically tailored to specific research areas (Loughran & Mcdonald, 2016). As we examine a non-specialized Facebook communication in Italian, we use a generic Italian lexicon for sentiment analysis called Sentix (Sentiment Italian Lexicon). ¹¹ The Sentix lexicon is developed by aligning other resources such as Multiwordnet (Pianta et al., 2002) and SentiWordnet (Baccianella et al., 2010). An entry in Sentix includes an Italian lemma with its part-of-speech (POS) and polarity score ranging from −1 (totally negative) to +1 (totally positive), ¹² among others.

We use the R package udpipe (Wijffels, 2021) to compute the aggregate sentiment polarity score for each Facebook post based on the the Sentix lexicon. We first tokenise and lemmatize each post using the PoSTWITA-UD treebank collection of Italian tweets annotated in Universal Dependencies (Sanguinetti et al., 2018). This treebank is particularly suitable for textual analysis of SoMe. The lexicon-based sentiment analysis procedure is referred to as a “bag of words” model, as both the order and the context of the words are ignored. To partially overcome this limitation, we exploit the ability of udpipe sentiment analysis function ¹³ to consider contextual valence shifters (Mehta & Chandra, 2019) in the polarity computation. Valence shifters are specific words that modify the degree of positivity or negativity of the term that they closely precede or follow. More specifically, we include negators (e.g., not, never, etc.) and adversative conjunctions (e.g., but, however, etc.) that, if occurring in the neighborhood of the word (between four words before and two words after), invert the sign of the word polarity. In addition, we include amplifiers (deamplifiers) ¹⁴ that, if occurring in the neighborhood of the word, amplify (deamplify) the word sentiment polarity by multiplying it by 1.8 (0.2). Finally, we constrain the aggregated sentiment scores between −1 and +1 by using the related option of the udpipe sentiment function that exploits the properties of the standard logistic function.

We run regression estimations for average weekly values of all variables at the municipal level. Therefore, we have an observation for each municipality and week included in the sample. This specification may produce benefits in terms of robustness and simplification of the models. Indeed, it allows the inclusion in the regression models of municipality and time (year-week) fixed effects to respectively account for time-invariant unobserved variables at the municipality level as well as macro contextual factors that could equally affect the capital municipalities in each week of the year. These macro contextual factors may include national-level legal restrictive or support measures, macroeconomic factors, and citizens’ changing psychological perception about the pandemic’s severity.

To test hypothesis H1, we estimate the Equation 1 regression using a Tobit model (Cameron & Trivedi, 2010) given that the sentiment polarity dependent variable is censored between −1 and +1:

T O N E i t = β 0 + β 1 C O V I D _ T P 1 i t + β 2 C O V I D _ T P 2 i t + ∑ k β k C T R L S k i t + m i + w t + e i t

(1)

where for municipality i in year-week t over the study period, TONE is the previously defined sentiment polarity of Facebook posts; COVID_TP1 is the prevalence of the topic on COVID-19 restrictive measures; COVID_TP2 is the prevalence of the topic on COVID-19 support measures; CTRLS^k is a set of k control variables; m_i denotes unobserved time-invariant municipality fixed effects; w_t denotes municipality-invariant year-week fixed effects; and e_it is the standard error term.

Within the control variables (CTRLS^k), we include topic-prevalence variables for the other non-COVID-19 topics identified in the LDA estimation described in the next section, ¹⁵ the natural logarithm of the number of COVID-19 positive cases per 100,000 inhabitants at the provincial level (CASES_COVID), ¹⁶ and the natural logarithm of the number of weekly posts on each municipal Facebook page (N_POSTS).

To test hypotheses H2 and H3, we estimate the following Equation 2 regression using a Beta model with logit link (Smithson & Verkuilen, 2006) ¹⁷ since the topic prevalence dependent variable is a proportion taking values greater than 0 and less than 1:

T O P I C _ C O V I D i t = β 0 + β 1 C A S E S _ C O V I D i t + ∑ k β k C T R L S k i t + m i + w t + e i t

(2)

where for municipality i in year-week t over the study period, TOPIC_COVID is either the prevalence of the COVID-19 restrictive measures topic, to test hypothesis H2, or the prevalence of the COVID-19 support measures topic, to test hypothesis H3; CASES_COVID is the independent variable of interest defined as in the Equation 1; CTRLS^k is a set of k control variables including topic prevalence variables for the other non-COVID-19 topics identified in the LDA estimation and the variable N_POSTS as defined in the Equation 1. The rest of variables are defined as in Equation 1.

Finally, to test the remaining hypotheses, we estimate the following Equation 3 regression using a Beta model with logit link (Smithson & Verkuilen, 2006):

T O P I C _ C O V I D i t = β 0 + β 1 C A S E S _ C O V I D i t + β 2 C A S E S _ C O V I D i t × M O D E R i t + ∑ k β k C T R L S k i t + m i + w t + e i t

(3)

where for municipality i in year-week t over the study period, MODER is the moderating variable corresponding to each hypothesis ¹⁸ and the rest of the variables are defined as in Equation 2. The moderating variables used for the estimations are an indicator variable for left-wing mayors (LEFT_MR); a continuous variable for mayor’s age in years (AGE_MR); an indicator variable for female mayors (FEM_MR); a continuous variable for margin of victory (MAR_VICT); and an indicator variable for mayors being in their first term of office (TERM1). ¹⁹ Notably, we define the margin of victory as the difference, in percentage points, between the percentage of votes obtained by the elected mayor and the second most-voted candidate in the first round of the latest local elections. ²⁰

Empirical Results and Analysis

Topic Model Estimation and Interpretation

Before estimating the LDA topic model, we conduct several pre-processing steps to the corpus of Facebook posts for text cleaning and transformation (Banks et al., 2018). ²¹ These steps involve converting all uppercase letters into lowercase and removing punctuation, accents, symbols, ²² numbers, URLs beginning with http(s), words with less than three characters, and standard Italian stopwords. Stopwords are high-frequency words, commonly consisting of specific parts of speech, which do not provide additional insights for the interpretation of the latent topic (e.g., articles, pronouns, conjunctions, prepositions, etc.). To enhance consistency and relieve computational load (Hoffman et al., 2010), we also include in the stopwords other semantically irrelevant words that appear frequently in our corpus (frequency > 200) such as names of months, days of the week, Italian regions, towns, and first names. In addition, we remove sparse terms appearing in less than 0.025% of the 76,138 posts and frequent terms appearing in more than 80% of the 76,138 posts (Banks et al., 2018) that enhance the computational complexity without significantly contributing to the topics’ identification and interpretability (Roberts et al., 2014).

We use the pre-processed corpus to generate the document-term matrix, the first input of the LDA estimation, by using the tokenisation and document-feature matrix functions of R package quanteda (Benoit et al., 2018). Importantly, we include concatenated bigram terms (sequence of two adjacent terms) as well as individual terms (unigrams) in the LDA estimation. Indeed, frequent bigram terms (e.g., big_data) could be more semantically meaningful than the corresponding individual terms (big and data) and, therefore, aid the topic interpretation.

The second input of the LDA estimation is the number of expected topics. Studies recommend using a combination of statistical methods and human judgment to select the best number of topics (Roberts et al., 2014, 2019). As there is no statistical method generally accepted as superior to select the best number of topics, we use several statistical metrics to identify the number of topics that produces the best score (minimum or maximum) for most of them. Specifically, we use the R package ldatuning (Nikita & Chaney, 2020) to apply the Arun2010 method, the CaoJuan2009 method, the Deveaud2014 method, and the Griffiths2004 method. Figure 1 shows the graphs of these four metrics for different numbers of topics ranging from 3 to 20.

OPEN IN VIEWER

Figure 1.

Metrics to select the best number of topics.

When CaoJuan2009 and Arun2010 are used to minimize metrics minimized, they tend to decrease with increases in the number of topics. However, their improvement, especially that of CaoJuan2009 metric, is marginally decreasing beyond nine topics as the curves start flattening. Similarly, the metrics Deveaud2014 and Griffiths2004, to be maximized, tend to increase with the number of topics. Nevertheless, at nine topics Deveaud2014 reaches the maximum, and beyond nine topics the improvement rate of Griffiths2004 diminishes significantly. Overall, the four metrics suggest that nine topics might be the best given that the relatively low model performance gains beyond nine topics may not outweigh the negative effect of additional topics on human ability to differentiate among them (Chang et al., 2009).

We use human judgment to confirm that any LDA estimation with more than nine topics does not produce any topic clearly distinguishable from those resulting from the estimation with nine topics. More specifically, any LDA estimation with more than nine topics ²³ produces a minimum of two topics containing at least five identical or semantically similar words in the top 10 words (Jaworska & Nanda, 2018). In addition, previous studies find that, under normal conditions, content posted by Western European local governments on their Facebook pages could be represented by a number of topics even lower than nine (Bonsón et al., 2017; Guillamón et al., 2016; Hofmann et al., 2013).

We estimate the LDA model with nine topics by using Gibbs sampling with hyperparameters α = 50/T, where T is the number of selected topics, β = .1, and number of iterations = 1,000. Following the estimation, we label each topic based on the 20 most likely words automatically assigned to each topic. Each of the four co-authors of this study independently proposed the labels after reading these words and the most representative posts per topic. Finally, we agreed on the final labels by majority rule. Table 1 shows the top 20 most likely Italian words for each topic, in descending order of probability, with their English translation and the label assigned to each topic.

OPEN IN VIEWER

Table 1.

Top 20 Most Probable Words by Labeled Topic.

	Top 20 most probable words
	Original Italian words	English translation
(1) COVID-19 restrictive measures	attivita; ordinanza; misure; regione; salute; decreto; coronavirus; covid-19; presidente; disposizioni; nuove; chiusura; commerciali; dpcm; pubblico; contenimento; vigore; interno; sicurezza; spostamenti	Activity; ordinance; measures; region; health; decree; coronavirus; covid-19; president; provisions; new; closure; commercial; prime minister’s decree; public; containment; force; indoor; safety; travels
(2) Public order measures	comunale; sindaco; assessore; amministrazione; polizia; amministrazione_comunale; giunta; president; ordine; locale; seduta; sport; stampa; polizia_locale; approvato; polizia_municipale; struttura; piano; manifestazione; controlli	Municipal; mayor; council member; administration; police; municipal_administration; council; president; order; local; session; sport; press; local_police; approved; municipal_police; structure; plan; demonstration; checks
(3) COVID-19 support measures	emergenza; coronavirus; covid-19; buoni_spesa; dati; sanitaria; cittadini; casa; consegna; domicilio; regione; aggiornamento; difficolta; casi_positivi; solidarietà; alimentari; situazione; asl; mascherine; base	Emergency; coronavirus; covid-19; shopping_vouchers; data; health; citizens; home; delivery; residence; region; update; difficulty; positive_cases; solidarity; food; situation; public healthcare; masks; base
(4) Commemorative events	giornata; palazzo; museo; occasione; arte; musei; mostra; piazza; nazionale; donne; buongiorno; storia; cerimonia; memoria; visita; vittime; associazione; internazionale; aperta; comunità	Day; building; museum; occasion; art; museums; show; square; national; women; good morning; history; ceremony; memory; visit; victims; association; international; open; community
(5) Social and environmental projects	progetto; sociale; raccolta; iniziativa; giovani; centro; collaborazione; rifiuti; incontro; progetti; lavoro; politiche; partecipazione; associazioni; partecipare; mobilita; cittadini; università; ambiente; operatori	Project; social; collection; initiative; young people; center; cooperation; waste; meeting; projects; employment; policies; participation; associations; participate; mobilize; citizens; university; environment; operators
(6) Cultural events	cultura; diretta; teatro; programma; appuntamento; biblioteca; streaming; festival; eventi; turismo; musica; evento; pagina; prenotazione; edizione; unica; bambini; canale; spettacolo; concerto	Culture; live; theater; program; appointment; library; streaming; festival; events; tourism; music; event; page; reservation; edition; unique; children; channel; show; concert
(7) Educational services	servizi; pubblico; sito; scuole; scuola; servizio; avviso; domande; comunali; ufficio; modalita; uffici; famiglie; contributo; scolastico; infanzia; bando; sostegno; pubblicato; richiesta	Services; public; site; schools; school; service; notice; applications; municipal; office; mode; offices; families; contribution; educational; childhood; announcement; support; published; request
(8) Public works	lavori; piazza; area; interventi; corso; viale; nuova; intervento; parco; sicurezza; centro_storico; sosta; tratto; strada; storico; realizzazione; quartiere; riqualificazione; verde; manutenzione	Works; square; area; interventions; street; avenue; new; intervention; park; safety; old_town; stop; stretch; road; historic; realization; neighborhood; redevelopment; green; maintenance
(9) Civil protection from calamities	servizio; protezione_civile; informazioni; numero; possibile; cittadini; allerta; meteo; centro; maggiori; regionale; avviso; cittadinanza; informa; territorio; attenzione; rischio; partire; condizioni; gialla	Service; civil_protection; information; number; possible; citizens; alert; weather forecast; center; further; regional; notice; citizenship; inform; territory; attention; risk; leave; conditions; yellow

Note. The words for each topic are listed in descending order of probability based on LDA model estimation.

Descriptive Statistics and Univariate Analysis

Table 2 shows descriptive statistics for the regression variables used to test our hypotheses.

OPEN IN VIEWER

Table 2.

Descriptive Statistics.

Variables	N	Mean	Std	P25	Median	P75	Min	Max
TONE	5,207	0.216	0.345	0.020	0.224	0.423	-1.000	1.000
TOPICS:
(1) COVID-19 restrictive measures	5,207	0.115	0.044	0.090	0.102	0.125	0.052	0.603
(2) Public order measures	5,207	0.112	0.032	0.091	0.103	0.124	0.044	0.340
(3) COVID-19 support measures	5,207	0.104	0.033	0.085	0.095	0.113	0.042	0.351
(4) Commemorative events	5,207	0.111	0.031	0.091	0.106	0.127	0.043	0.367
(5) Social and environmental projects	5,207	0.111	0.030	0.093	0.105	0.121	0.044	0.440
(6) Cultural events	5,207	0.112	0.039	0.088	0.104	0.127	0.044	0.520
(7) Educational services	5,207	0.115	0.034	0.094	0.108	0.127	0.050	0.472
(8) Public works	5,207	0.113	0.034	0.091	0.106	0.127	0.040	0.514
(9) Civil protection from calamities	5,207	0.107	0.037	0.088	0.098	0.112	0.043	0.628
Total COVID topics (1 + 3)	5,207	0.219	0.055	0.182	0.205	0.243	0.097	0.657
CASES_COVID	5,207	1,647.0	1,936.3	170.66	715.42	2,799.4	0.000	12,796.1
CASES_COVID (log)	5,207	6.325	1.875	5.145	6.574	7.938	0.000	9.457
N_POSTS	5,207	14.59	11.80	6.00	12.00	20.00	1.00	122.00
N_POSTS (log)	5,207	2.34	0.90	1.79	2.48	3.00	0.00	4.80
LEFT_MR	5,207	0.531	0.499	0	1	1	0	1
AGE_MR	5,207	54.12	10.31	45.82	54.05	62.28	32.62	84.95
FEM_MR	5,207	0.097	0.296	0	0	0	0	1
MAR_VICT	5,207	10.888	16.851	−1.96	8.59	21.09	−17.47	60.9
TERM1	5,207	0.710	0.454	0	1	1	0	1

Note. The sample period is from 1 March 2020 to 26 March 2021. N is the number of municipality-weeks in the sample. TONE is the sentiment polarity score of municipal Facebook posts; TOPICS includes the prevalence of each topic from LDA estimation; CASES_COVID is the number of COVID-19 positive cases per 100,000 inhabitants at the provincial level; N_POSTS is the number of weekly posts on each municipal Facebook page; LEFT_MR is an indicator variable for left-wing mayors; AGE_MR is the mayor’s age in years; FEM_MR is an indicator variable for female mayors; MAR_VICT is the mayor’s margin of victory; TERM1 is an indicator variable for first-term mayors.

Notably, the mean of sentiment polarity (TONE) is 0.216, the median is 0.224, and more than 75% of the observations (P25 = 0.020) have a sentiment polarity greater than zero. This provides evidence of the tendency of municipalities to adopt a positive communication strategy through their Facebook pages. As suggested by previous studies, this strategy may aim at enhancing citizen engagement and political consensus as well as voter attraction (Nave et al., 2018; Stieglitz & Dang-Xuan, 2013).

The mean number of weekly posts (N_POSTS), 14.59, and the median, 12, suggest that Italian municipalities actively used Facebook. This is consis- tent with studies finding that Facebook has become the favorite SoMe plat- form of local governments to inform and interact with their citizens (Bonsón et al., 2017; Guillamón et al., 2016; Larsson & Skogerbø, 2018; Silva et al., 2019).

Interestingly, about 53% of the municipalities are governed by left-wing mayors (LEFT_MYR); the mean and median age of the mayors is 54 (AGE_MYR); 71% of the mayors are in their first term of office (TERM1); and only 9.7% of the mayors are women (FEM_MYR). This low percentage of female mayors further confirms the unsolved issue of the women’s under-representation in the Italian political institutions (Braga & Scervini, 2017).

Unsurprisingly perhaps, the differences between means and medians of the prevalence of the nine topics are relatively small (recall that each topic prevalence must be between 0 and 1, and the total for all topics should be equal to 1 for each document). Indeed, the means range from a minimum of 0.104 (COVID-19 support measures) to a maximum of 0.115 (Educational services); conversely, the medians range from a minimum of 0.095 (COVID-19 support measures) to a maximum of 0.108 (Educational services). The results show that, during our study period, each contingent COVID-19-related topic has a similar prevalence to other topics found by previous studies to be typical of local governments’ Facebook communication (Bonsón et al., 2017; Hofmann et al., 2013).

At the same time, however, the median of the topic sum of the two COVID-19 topics (0.205) is about double the median of each of the other topics and significantly higher (p-value < .001) based on the two-tailed Wilcoxon signed-ranks test. Overall, our results show that Italian municipalities largely used their official Facebook pages to communicate about COVID-19 during the 13 months examined. These results are consistent with other studies that reported that, in Italy, governmental agencies and government representatives extensively resorted to SoMe to inform residents of COVID-19 and related government decisions (Ruiu, 2020).

(Table 3 displays descriptive statistics of COVID-19 topics by month over the study period.)

OPEN IN VIEWER

Table 3.

Descriptive Statistics of COVID-19 Topics by Month.

		COVID-19 restrictive measures			COVID-19 support measures
Year-month	N. posts	Mean	Median	Std	Mean	Median	Std
2020-03	7,294	0.156	0.110	0.108	0.124	0.105	0.067
2020-04	6,507	0.124	0.097	0.084	0.141	0.109	0.088
2020-05	5,889	0.134	0.099	0.094	0.107	0.090	0.064
2020-06	4,952	0.107	0.091	0.061	0.096	0.087	0.049
2020-07	5,517	0.098	0.090	0.048	0.092	0.084	0.050
2020-08	4,561	0.100	0.089	0.059	0.092	0.083	0.054
2020-09	5,312	0.098	0.088	0.049	0.093	0.084	0.053
2020-10	5,845	0.117	0.091	0.087	0.099	0.086	0.064
2020-11	5,943	0.111	0.090	0.078	0.113	0.090	0.083
2020-12	6,312	0.103	0.088	0.065	0.109	0.089	0.081
2021-01	5,731	0.102	0.087	0.066	0.107	0.085	0.085
2021-02	6,212	0.095	0.085	0.051	0.107	0.085	0.089
2021-03	5,874	0.102	0.087	0.066	0.103	0.086	0.075
Friedman test			***			***
ANOVA test		***			***

Note. *, **, and *** denote significance levels at 10%, 5%, and 1%, respectively, based on a two-tailed Friedman test for the differences in topic distributions and a one-way repeated measures ANOVA test for the differences in means of topics over the year-months.

When turning to the topics in the Facebook postings, we identify two topics related to COVID-19 – topic 1 (labeled “COVID-19 restrictive measures,” because of the words “coronavirus” and “COVID-19” accompanied by the word “measures,” semantically similar words such as “ordinance,” “decree,” “provisions,” and references to restrictions like “closure,” “containment”) and topic 3 (“COVID-19 support measures,” where “coronavirus” and “COVID-19,” were accompanied by other words referring to public support initiatives such as “shopping_vouchers,” ²⁴ “public health care,” and “solidarity” and words like “data,” “update,” “situation”). The other topics included topic 2 “public order measures,” topic 4 “commemorative events,” topic 5 “social and environmental projects,” topic 6 “cultural events,” topic 7 “educational services,” topic 8 “public works,” and topic 9 “civil protection from calamities.”

To confirm the semantic validity of the topic labels, we carried out a qualitative analysis by reading the 10 archetypal posts scoring the highest topic prevalence for each of the nine topics (Roberts et al., 2019). Except for the two contingent topics related to COVID-19, the others are mostly consistent with prior findings of studies examining the thematical content of SoMe publications of Western European local governments (Bonsón et al., 2017; Hofmann et al., 2013).

The non-parametric Friedman test and the one-way repeated-measures ANOVA test indicate that the distributions and means of COVID-19 topics, respectively, changed significantly over the 13 months. To better assess the trends of average COVID-19 topics by year-month, see Figure 2.

OPEN IN VIEWER

Figure 2.

Trend of COVID-19 topics by year-month.

The charts show that in March 2020, topic 1 (on COVID-19 restrictive measures) has the highest prevalence during the examined time series. Indeed, in March 2020 the Italian government adopted the most restrictive measures of the pandemic by placing the Italian population in lockdown, closing non-essential business and industries, suspending public events, and restricting movements and gatherings (Ruiu, 2020). ²⁵ The prevalence of topic 1 decreases in April 2020, during the lockdown period, and it rises again in May 2020 when the Italian government rescinded most of the restrictions. In the following months, the prevalence of topic 1 keeps declining until October 2020, when it peaks again due to further restrictions on movement and social activities introduced by the Italian government to tackle the second wave of COVID-19 affecting the country. The following decrease in the prevalence of topic 1 stops in March 2021, the last month of our series, when the smaller peak may be due to further restrictive measures adopted by the Italian government.

Topic 3, COVID-19 support measures, scores the highest prevalence over the examined time series during the lockdown in April 2020 as well as following the severe government restrictions of March 2020. In the following months, the prevalence of topic 3 declines and starts rising again significantly in September 2020 until it reaches another peak in November 2020, following the government restrictions of October 2020. Finally, the prevalence of topic 3 slightly decreases until March 2021. Overall, this monthly trend analysis of COVID-19 topics suggests that during the analyzed period Italian municipalities adjusted their communications regarding COVID-19 to the pandemic’s different phases (Mori et al., 2021).

(Figure 3 shows the Pearson pairwise correlations between the dependent variables and the other covariates in the regression models.)

OPEN IN VIEWER

Figure 3.

Pearson pairwise correlations among model variables.

Notably, the variable TONE (sentiment polarity score) is significantly and negatively correlated with COVID-19 restrictive measures (topic 1). Furthermore, TONE is significantly and positively correlated with COVID-19 support measures (topic 3). These results provide bivariate for hypothesis H1.

Consistent with hypothesis H2, topic 1 (COVID-19 restrictive measures) is significantly and negatively correlated with the number of positive cases in the municipal area (CASES_COVID). On the other hand, the significant and negative correlation between topic 3 (COVID-19 support measures) and CASES_COVID contrasts with the positive association between the two variables predicted by the hypothesis H3. (Nonetheless, fuller tests of the hypotheses appear in the multivariate regression analysis, presented in the next section.)

Both COVID-19 topics are significantly and negatively correlated with all other non-COVID-19 topics, confirming the particularity of COVID-19 topics and the absence of any related topic being strategically associated with them. Finally, the highest negative coefficient (−0.28) and the highest positive coefficient (0.25) are relatively low, suggesting that collinearity is unlikely to bias our regression estimations. In addition, the average variance inflation factor (VIF) of all covariates is 1.20 and the maximum individual variable VIF is 1.40, far below the maximum cut-off of 10 when multicollinearity becomes a concern (Cameron & Trivedi, 2010).

Baseline Regression Results

Table 4 presents the estimations of Equation 1 Tobit regression on the senti- ment polarity variable TONE to test hypothesis H1.

OPEN IN VIEWER

Table 4.

Tobit Regressions on Sentiment Polarity of Municipal Facebook Posts.

	Tone
Variables	(1)	(2)	(3)	(4)
Topics:
(1) COVID-19 restrictive measures	−0.7266*** (0.245)		−0.6845*** (0.244)	−0.7880*** (0.251)
(2) COVID-19 support measures		1.1843*** (0.379)	1.1306*** (0.384)	1.0357*** (0.347)
Other topics	No	No	No	Yes
CASES_COVID	0.0213 (0.016)	0.0184 (0.016)	0.0189 (0.016)	0.0217 (0.015)
N_POSTS	−0.0072 (0.017)	−0.0048 (0.017)	−0.0080 (0.017)	−0.0109 (0.017)
Municipality FE	Yes	Yes	Yes	Yes
Year-week FE	Yes	Yes	Yes	Yes
Number of observations	5,207	5,207	5,207	5,207
Pseudo R-squared	0.2191	0.2213	0.2289	0.2513

Note. *, **, and *** denote significance levels at 10%, 5%, and 1%, respectively, based on two-tailed tests. Standard errors clustered at the municipality level are presented in parentheses. Other topics are the other topic prevalences except topic 9; CASES_COVID is the natural logarithm of the number of COVID-19 positive cases per 100,000 inhabitants at the provincial level; N_POSTS is the natural logarithm of the number of weekly posts on each municipal Facebook page; Municipality FE denotes municipality fixed effects; Year-week FE denotes year-week fixed effects.

We show four specifications. Specifically, in specification 1, we only include the topic of COVID-19 restrictive measures out of the nine identified topics. Specification 2 includes only the topic of COVID-19 support measures. In specification 3, we include both COVID-19 topics. Finally, in specification 4, we include both the COVID-19 topics and other topics except for topic 9 (omitted to avoid perfect collinearity among the topics). ²⁶

Interestingly, the coefficient for COVID-19 restrictive measures is nega- tive and statistically significant (p < .01) in all related specifications, indicating that a higher prevalence of this topic is associated with a more negative tone of municipal communication. These results suggest that communication on COVID-19 restrictive measures may tend to be purely prescriptive and not included in a broader persuasion strategy to positively engage citizens in the fight against the pandemic (Mori et al., 2021). On the other hand, the coefficient for COVID-19 support measures is positive and significant (p < .01) in all specifications, indicating that a higher prevalence of this topic is associated with a more positive tone of municipal communication. Overall, these results provide support for hypothesis H1.

Table 5 reports the estimations of Equation 2 to test hypotheses H2 and H3.

OPEN IN VIEWER

Table 5.

Beta Regressions on COVID-19 Topics.

	COVID-19 restrictive measures		COVID-19 support measures
Variables	(1)	(2)	(1)	(2)
CASES_COVID	0.0033 (0.011)	−0.0176*** (0.006)	0.0191*** (0.007)	0.0146** (0.007)
N_POSTS	−0.0184 (0.012)	0.0001 (0.006)	0.0169** (0.008)	0.0375*** (0.008)
Other topics	No	Yes	No	Yes
Municipality FE	Yes	Yes	Yes	Yes
Year-week FE	Yes	Yes	Yes	Yes
Number of obs.	5,207	5,207	5,207	5,207

Note. *, **, and *** denote significance levels at 10%, 5%, and 1%, respectively, based on two-tailed tests. Standard errors clustered at the municipality level are presented in parentheses. CASES_COVID is the natural logarithm of the number of COVID-19 positive cases per 100,000 inhabitants at the provincial level; N_POSTS is the natural logarithm of the number of weekly posts on each municipal Facebook page; Other topics are the other non-COVID-19 topics; Municipality FE denotes municipality fixed effects; Year-week FE denotes year-week fixed effects.

We present two specifications for each COVID-19 topic regression, one without control variables and the second with the controls. Notably, the coefficient for CASES_COVID is not significant at conventional levels in specification 1 of the COVID-19 restrictive measures regression. In contrast, in specification 2 of the COVID-19 restrictive measures regression, the coef- ficient on CASES_COVID is negative and significant (p < .01), suggesting that, controlling for the non-COVID-19 topics, an increase in COVID-19 positive cases is associated with a lower prevalence of the topic of COVID- 19 restrictive measures relative to the topic of COVID-19 support measures. Thus, despite an increase in the number of positive cases, municipal incumbents may not post more on COVID-19 restrictive measures to avoid arousing negative emotions. Indeed, these negative emotions may boost constituents’ perception of incumbents’ inability to effectively tackle the pandemic and erode political consensus. Overall, these results provide support for hypothesis H2.

In specification 1 of the COVID-19 support measures regression, the coef- ficient on CASES_COVID is positive and significant (p < .01), suggesting that municipalities tend to post more on COVID-19 support measures when the number of COVID-19 positive cases grows. The coefficient on CASES_ COVID remains positive and significant (p < .05) in specification 2 after controlling for non-COVID-19 topics. These results are consistent with hypothesis H3. they may reveal a communication strategy of municipal incumbents aimed at enhancing constituents’ perceptions about their effective response to the pandemic through the promotion of related support measures (Gallego, 2018; Masiero & Santarossa, 2021).

Table 6 reports the estimations of the Equation 3 Beta regression on the two COVID-19 topics to test hypotheses H4, H5, H6, H7, and H8. We estimate the same two regression specifications as those of the Equation 2 estimations, and we report the effect of each moderating variable in a different table panel. Notably, in Panel A, the coefficient on the interaction variable CASES_ COVID×LEFT_MR is not significant at conventional levels in any specification of the COVID-19 restrictive measures regression and in specification 2 of the COVID-19 support measures regression. In contrast, this coefficient is negative and significant (p < .01) in specification 1 of the COVID-19 support measures regression. These results lead to rejection of null hypothesis H4 and failure to reject null hypothesis H5. That is, when positive case numbers grow, municipalities led by non-left-wing mayors tend to promote COVID-19 support measures more on their Facebook pages than municipalities led by left-wing mayors. Conversely, the mayor’s ideology does not moderate the relation between COVID-19 restrictive measures and the number of positive cases. These results may be explained by the fact that the pandemic significantly affected the constituency and supporters of Italian right-wing parties, including small- and medium-sized entrepreneurs and self-employed workers. Therefore, promoting public support measures could be a strategy of these parties to increase political consensus among voters (Klomp, 2019).

OPEN IN VIEWER

Table 6.

Beta Regressions on COVID-19 Topics with Moderating Variables.

	COVID-19 restrictive measures		COVID-19 support measures
Variables	(1)	(2)	(1)	(2)
Panel A: The moderating effect of mayor’s ideology
CASES_COVID	0.0071 (0.012)	−0.0189*** (0.006)	0.0260*** (0.008)	0.0181** (0.007)
CASES_COVID×LEFT_MR	−0.0058 (0.006)	0.0025 (0.003)	−0.0120*** (0.004)	−0.0064 (0.004)
N_POSTS	−0.0175 (0.012)	0.0000 (0.006)	0.0176** (0.008)	0.0377*** (0.008)
Other topics	No	Yes	No	Yes
Municipality FE	Yes	Yes	Yes	Yes
Year-week FE	Yes	Yes	Yes	Yes
Number of obs.	5,207	5,207	5,207	5,207
Panel B: The moderating effect of mayor’s age
CASES_COVID	0.0001 (0.023)	0.0281*** (0.009)	−0.0589*** (0.013)	−0.0344** (0.014)
CASES_COVID×AGE_MR	0.0001 (0.000)	−0.0008*** (0.000)	0.0014*** (0.000)	0.0009*** (0.000)
N_POSTS	−0.0182 (0.012)	−0.0003 (0.006)	0.0166** (0.008)	0.0370*** (0.008)
Other topics	No	Yes	No	Yes
Municipality FE	Yes	Yes	Yes	Yes
Year-week FE	Yes	Yes	Yes	Yes
Number of obs.	5,207	5,207	5,207	5,207
Panel C: The moderating effect of mayor’s gender
CASES_COVID	0.0024 (0.011)	−0.0185*** (0.006)	0.0205*** (0.007)	0.0156** (0.007)
CASES_COVID×FEM_MR	0.0061 (0.013)	−0.0008 (0.005)	0.0118* (0.007)	0.0069 (0.008)
N_POSTS	−0.0180 (0.012)	0.0005 (0.006)	0.0161** (0.008)	0.0367*** (0.008)
Other topics	No	Yes	No	Yes
Municipality FE	Yes	Yes	Yes	Yes
Year-week FE	Yes	Yes	Yes	Yes
Number of obs.	5,207	5,207	5,207	5,207
Panel D: The moderating effect of mayor’s margin of victory
CASES_COVID	−0.0026 (0.012)	−0.0243*** (0.006)	0.0264*** (0.008)	0.0216*** (0.008)
CASES_COVID×MAR_VICT	0.0003* (0.000)	0.0004*** (0.000)	−0.0004*** (0.000)	−0.0004*** (0.000)
N_POSTS	−0.0182 (0.012)	−0.0002 (0.006)	0.0174** (0.008)	0.0375*** (0.008)
Other topics	No	Yes	No	Yes
Municipality FE	Yes	Yes	Yes	Yes
Year-week FE	Yes	Yes	Yes	Yes
Number of obs.	5,207	5,207	5,207	5,207
Panel E: The moderating effect of mayor’s term of office
CASES_COVID	0.0099 (0.014)	−0.0092 (0.006)	0.0075 (0.008)	0.0047 (0.007)
CASES_COVID×TERM1	−0.0076 (0.008)	−0.0078** (0.003)	0.0104** (0.005)	0.0092** (0.004)
N_POSTS	−0.0186 (0.012)	−0.0002 (0.006)	0.0171** (0.008)	0.0375*** (0.006)
Other topics	No	Yes	No	Yes
Municipality FE	Yes	Yes	Yes	Yes
Year-week FE	Yes	Yes	Yes	Yes
Number of obs.	5,207	5,207	5,207	5,207

Note. *, **, and *** denote significance levels at 10%, 5%, and 1%, respectively, based on two-tailed tests. Standard errors clustered at the municipality level are presented in parentheses. CASES_COVID is the natural logarithm of the number of COVID-19 positive cases per 100,000 inhabitants at the provincial level; LEFT_MR is an indicator variable for left-wing mayors; AGE_MR is mayor’s age in years; FEM_MR is an indicator variable for female mayors; MAR_VICT is the mayor’s margin of victory in the last first-round elections; TERM1 is an indicator variable for mayors being in their first term of office; N_POSTS is the natural logarithm of the number of weekly posts on each municipal Facebook page; Other topics are the other non-COVID-19 topics; Municipality FE denotes municipality fixed effects; Year-week FE denotes year-week fixed effects.

In Panel B, the coefficient on the interaction variable CASES_ COVID×AGE_MR is not significant at conventional levels in the specification 1 of the COVID-19 restrictive measures regression, whereas it is negative and significant (p < .01) in specification 2. Furthermore, the coefficient for the interaction variable CASES_COVID×AGE_MR is positive and significant (p < .01) in both specifications of the COVID-19 support measures regression. Consistent with hypothesis H6, these results provide evidence that when the number of positive cases grows, municipalities led by older mayors tend to post less on COVID-19 restrictive measures and more on COVID-19 support measures than municipalities led by younger mayors.

In Panel C, the coefficient for the interaction variable CASES_ COVID×FEM_MR is not significant at conventional levels in any specification of the estimated regressions. These results fail to support hypothesis H7. In this regard, the low percentage of municipalities governed by female mayors (9.7%) in our sample may partially explain why the gender of the mayor does not moderate the association between COVID-19 topics and number of positive cases.

In Panel D, the coefficient for the interaction variable CASES_ COVID×MAR_VICT is positive and significant (p < .01) in specification 2 of the COVID-19 restrictive measures regression, whereas it is negative and significant (p < .01) in both specifications of the COVID-19 support measures regression. This means that when the number of positive cases grows, mayors with higher margins of victory and, therefore, stronger political positions, may have fewer incentives to foster their political legitimacy by downplaying negative information on COVID-19 restrictive measures and promoting positive information on COVID-19 support measures on municipal Facebook pages.

Finally, in Panel E, the coefficient for the interaction variable CASES_ COVID×TERM1 is negative and significant (p < .05) in specification 2 of the COVID-19 restrictive measures regression, whereas it is negative and significant (p < .01) in both specifications of the COVID-19 support measures regression. Relative to second-term mayors, when the number of positive cases grows, first-term mayors aspiring to re-election may have more incentive to foster political consensus by downplaying negative information on COVID-19 restrictive measures and promoting positive information on COVID-19 support measures on municipal Facebook pages. Overall, the results displayed in Panel D and E provide support for hypothesis H8.

Conclusions and Discussion

We examine how Italian municipalities managed COVID-19 communication on their official Facebook pages. To extract latent topics from Facebook posts, we apply the LDA topic modeling procedure on a sample of 76,139 posts published on the official Facebook pages of 103 Italian provincial capi- tal municipalities during the COVID-19 pandemic from 1 March 2020 to 26 March 2021. We identify two topics related to COVID-19, on COVID-19 restrictive measures and on COVID-19 support measures. Subsequently, using regression models including both municipality and year-week fixed effects, we find that the prevalence of the topic on COVID-19 restrictive measures is negatively related to the tone of communication. Conversely, the prevalence of the COVID-19 support measures topic is positively related to the tone of communication. More importantly, we show that the number of COVID-19 positive cases in a municipality area is negatively associated with the prevalence of the emotionally negative topic of COVID-19 restrictive measures and positively associated with the emotionally positive topic of COVID-19 support measures.

These associations may provide evidence of a strategically positive communication of municipalities aimed at inducing positive perceptions in citizens about the ability of municipal incumbents to effectively respond to the pandemic through suitable support measures. Furthermore, we find that these associations are strengthened for municipalities led by right-wing mayors and older mayors. Finally, the role of political incentives is confirmed by estimations revealing that mayors with stronger political positions, signaled by higher margins of victory, tend to manage their COVID-19 communication to foster political consensus less strategically. In contrast, evidence of strategic management of COVID-19 communication is stronger for first-term mayors who, unlike second-term mayors, are allowed to run for reelection.

Prior studies examining local governments’ communication on SoMe mostly focus on the level of activities (e.g., number of posts, likes, comments, shares) or related composite indicators to infer purposes (e.g., transparency, civic engagement) or underlying political strategies (Agostino & Arnaboldi, 2016; Bonsón et al., 2017; Guillamón et al., 2016; Silva et al., 2019). Our study provides evidence that the prevalence of specific topics such as those related to COVID-19 and their association patterns with logical determinants (e.g., severity of the pandemic) may more effectively unveil underlying public communication strategies. In this regard, prior research seeks to classify the main topics addressed by local governments of various countries including Italy on their SoMe platforms (Bonsón et al., 2017; Gesuele et al., 2016; Hofmann et al., 2013). Nonetheless, these studies apply manual content analysis procedures on relatively small corpora of SoMe posts. These procedures require a subjective predefinition of categories and dictionaries that may lead to overlooking notable patterns. Conversely, the topic modeling procedure that we apply consists of an unsupervised machine learning algorithm that allowed identification of latent topics in large corpora of SoMe posts without subjective predefinition of topical categories or dictionaries and with replicable results.

Furthermore, we contribute to the literature on crisis informatics on the use of social media during emergency and crisis events (Palen & Anderson, 2016; Reuter & Kaufhold, 2018; Reuter et al., 2018). Specifically, we address the bias toward studies mostly focused on Twitter in English language (Reuter et al., 2018) by examining municipalities’ Facebook posts in Italian. We analyze an unprecedented global disaster like COVID-19 over a relatively long period of 13 months. We show how political influences leading to biased public communication can be included among the barriers, identified by prior studies (Plotnick & Hiltz, 2016; Reuter & Kaufhold, 2018), to effective use of SoMe to manage disasters.

Our findings may have policy implications for public regulators of local governments’ SoMe disclosures. Specifically, during emergency situations arising from natural disasters or health crises, the public interest in properly informing and engaging citizens should prevail over political incentives that may bias the information and create conflicts with information coming from other public bodies or the central government. Biased information disseminated by local governments may undermine the credibility of recommendations provided by the central government or other governmental agencies (Ruiu, 2020). Regulation of SoMe like official websites is hardly achievable in the short term. However, some alternative solutions may involve defining common internal municipal policies regarding SoMe communication that may ensure greater consistency and objectivity in the communications. In addition, definition of dedicated qualified roles for managing the SoMe communication independent from elected political bodies may contribute to more reliable, consistent, and politically independent communication.

Specifically, some best practices may be derived from the guidance for Public Information Officers (PIOs) performing public relations function of the National Incident Management System (NIMS) in the US (National Incident Management System Basic Guidance for Public Information Officers, 2020). PIOs are trained personnel whose primary responsibility is to disseminate verified, accessible, and timely information regarding the status of emergencies to the media, the public, representatives in incident management organizations, and other directly or indirectly affected stakeholders. SoMe have become an important communication tool for PIO activities (Hughes & Palen, 2012). A standardized emergency management system (Incident Command System) entailing unified direction ensures the coordination of the activities of PIOs from different agencies and, therefore, the consistency of their communication. Some crisis communication tips and principles established by the guidance involve sharing official, verified, and relevant information aligned with messages of other authorities, providing clear direction, promoting collaboration, expressing empathy, conveying objective and politically independent information, monitoring the usage of official SoMe platforms. Each public agency should develop a SoMe policy and standard procedures outlining rules, roles, and responsibilities related to managing SoMe accounts. SoMe account management tasks should involve social listening and timely responding to public queries, among others.

Nonetheless, our findings have some limitations. First, the LDA topic modeling procedure requires some subjective interpretation of the latent topics based on the most probable related terms as well as a prior definition of the best number of topics. Second, our sample only consists of the provincial capital municipalities, and the results may be different for smaller municipalities. Third, we only examine the communication of Italian municipalities on Facebook and disregard other less used social networks like Instagram, LinkedIn, and Twitter that some municipalities may employ to target different audiences (e.g., younger users on Instagram) according to their strategies. Finally, we address endogeneity concerns by including both municipality and year-week fixed effects in our regression models. However, we cannot exclude the presence of correlated omitted variables that may bias our estimations.