The ECB and its Clarity of Communication with the Public: Does Plain Talking Shape Engagement?
Federico Maria Ferrara1 and Siria Angino2
1London School of Economics 2European Central Bank
Abstract Although many central banks have made increasing efforts to communicate with the general public, we know little about how public attention responds to central bank communications. One reason why it may be challenging for central banks to engage the public is that their language is too difficult to understand for the average citizen. Focusing on the case of the European Central Bank (ECB), we hypothesise that greater communication clarity is conducive to stronger engagement. To test this hypothesis, we employ readability metrics to analyse a comprehensive dataset of ECB communication activities and assess the effects of communication clarity on media engagement, focusing on both traditional and social media. First, we show that communication clarity in ECB speeches and press conferences is a significant and robust predictor of the volume of media coverage received by the ECB. Second, we provide evidence that the clarity of digital communications plays an important role in explaining the social media engagement generated by the ECB on Twitter. Taken together, these findings have significant implications for the communication policy of central banks. 1 Introduction
Like the utterings of the Delphic oracle, central bank communications with the general public used to be enigmatic and obscure. In the “ancient r´egime”of central bank communication, monetary policymakers often decided to communicate in a language that was only understood by fellow technocrats and purposely designed not to be intelligible to the average citizen. This Delphic nature of the public communication of central bankers is epitomised by an oft-cited quote attributed to Alan Greenspan, who said in a 1988 speech, early in his tenure as Federal Reserve Chair: “I guess I should warn you, if I turn out to be particularly clear, you’ve probably misunderstood what I said.” (quoted in Norris 2005). Far from being a phenomenon limited to the American context, central bankers in Europe have for a long time displayed a similar Delphic attitude and scepticism towards the feasibility and desirability of a clear communication with the public. Almost twenty years after Greenspan’s famous quote, Otmar Issing, the former Chief economist of the European Central Bank (ECB), wrote: “Transparency is not an end in itself; a central bank is not established with the primary objective of communicating with the public.” (Issing 2005). Central bank communication has come a long way since then. There has been a revolu- tion over recent years, with central bank communication activities growing exponentially in volume and becoming more targeted at the general public (Haldane 2017). This revolution has also underpinned a significant change of perspective with regard to the clarity of central bank communication with the public. As Mario Draghi put it during his last press conference as ECB President in October 2019:“[W]hat happened in communication by central banks over the last ten, fifteen years is a complete transformation. Nowadays, nobody would pride himself or herself saying: if you have understood me, you are stupid. Nobody would say that today.” (ECB 2019). In a similar vein, opposite to Greenspan’s complacency towards unclear communications, in a recent Town Hall Meeting with educators, Federal Reserve Chair Jerome Powell stated: “The Federal Reserve policy affects everyone. And we need to work hard, and we do work hard, at trying to communicate in a way that doesn’t lapse into economic jargon, and so it can be understood by the interested public, by anyone who’s interested.” (Federal Reserve 2019). Despite the recent revolution in central bank communication and central bankers’ chang- ing attitudes towards communication clarity and transparency, there remains a widespread belief that central banks communicate in a way that most households cannot understand. As Bank of England Chief economist Andrew Haldane put it on several occasions, “around 95% of all the words central banks utter are inaccessible to around 95% of the population” (Hal- dane 2018). This lack of clarity in the public communications of central banks is regarded as a major form of hindrance for these institutions to successfully engage the public and close the “twin deficit” of public understanding and trust in economics (Haldane, Macaulay, and McMahon 2020). Confirming the importance of communication clarity, research in political science points to its beneficial effects on citizen political understanding and civic engagement (e.g., Reilly and Richey 2011; Bischof and Senninger 2018). Also, economic research has provided ev- idence for a range of positive implications of central bank communication clarity on the behaviour and expectations of financial market actors (e.g., Jansen 2011a; Montes et al. 2016; Smales and Apergis 2017). Nonetheless, we still know very little about the effects of
1 central bank communication clarity on public attention and engagement. Focusing on the case of the ECB, this paper fills the gap by asking the following research question: does plain talking increase public engagement in the realm of central banking? The article addresses this question by analysing a comprehensive dataset of ECB speeches, press conferences and tweets using readability metrics (Flesch 1948; Gunning 1952; Kincaid et al. 1975) and assessing the impact of communication clarity on media engagement, focus- ing on both traditional and social media. First, we show that greater communication clarity of ECB speeches and press conferences is significantly and robustly correlated with a higher level of media coverage on the ECB on the day in which these communication activities take place. Second, we provide evidence that the clarity of ECB digital communications is a strong predictor of the social media engagement generated by the central bank on Twit- ter. These results are robust to controlling for several factors that may significantly affect engagement, such as the presence of conventional and unconventional policy changes during a press conference, the topics of discussion and number of speeches on a given day, and the topic and media content type of ECB tweets. Taken together, these findings shed new light on the importance of communication clarity for central banks to successfully engage the public. The next section reviews major contributions in economics and political science focusing on communication clarity and elaborates the argument we test in our empirical analysis. The third section discusses the data and methods employed in the paper. The fourth section presents the results of our analysis. The fifth section provides several robustness checks and extensions of the analysis. The final section draws conclusions and points to future avenues for research.
2 The Effects of Clarity of Central Bank Communica- tion
A common observation in the academic and public debate is that, in many ways, the styles of discourse are changing radically in contemporary public communications. While the chal- lenges of policy-making have grown in complexity over the past decades, the sophistication of public speech has declined. For example, scholars across disciplines have applied measures of textual complexity developed in the field of education to find that the sophistication of the speeches given by American presidents has steadily decreased over the past 200 years (e.g., Sigelman 1996; Teten 2003; Lim 2008). Political scientists and social psychologists have advanced significantly in the study of the determinants of linguistic complexity in politi- cal communication (e.g., Spirling 2016; Lin and Osnabr¨ugge 2018; Schoonvelde et al. 2019; Rauh, Bes, and Schoonvelde 2020). Instead, the research on clarity of communication with the public is much more limited outside the domain of political communication. This topic is particularly relevant for central banks. A long-standing debate in economics has focused on what constitutes an optimal communication strategy for central banks (for a review, see Blinder et al. 2008). An important part of this debate has focused on the notion of central bank transparency. A greater degree of transparency may produce tangible benefits, such as well-anchored inflation expectations of households and businesses. At the same time, various authors have suggested that limits to transparency may exist (Cukierman and
2 Meltzer 1986; Morris and Shin 2002). Whether transparency is beneficial, and to what extent, is in the end an empirical matter. The evidence in favour of transparency has accumulated over the years. Although Geraats(2002) cautiously concluded that the available evidence “suggests that transparency tends to be beneficial”, van der Cruijsen and Eijffinger(2007) document that “the empirical results are largely in favour of more transparency.” Within this body of research, several studies have argued that communication clarity constitutes a key element of central bank transparency. (e.g., Winkler 2000; Fracasso, Gen- berg, and Wyplosz 2003; Jansen 2011b). The clarity of the communication activities of central banks has been increasingly measured by applying the same metrics of textual com- plexity developed in education studies (e.g., Flesch 1948; Gunning 1952; Kincaid et al. 1975) and used by political science and psychology scholars to evaluate the complexity of political speeches. Most of the literature studying the consequences of central bank communication clarity employs these measures and points to significant effects of clarity in steering finan- cial market expectations and reducing volatility. For example, using textual data from the testimony by the Chairman of the Federal Reserve at the Congressional Monetary Policy Oversight hearings, Jansen(2011a) provides evidence that, in certain time periods and es- pecially during Alan Greenspan’s Chairmanship, communication clarity had a diminishing effect on volatility. In line with this result, Smales and Apergis(2017) find that more complex language related to Federal Open Market Committee (FOMC) decisions increases trading volumes and volatility of returns. Focusing on the case of the Central Bank of Brazil, Montes et al.(2016) and Montes and Nicolay(2017) provide evidence that central bank communica- tion clarity improves the credibility of monetary policy and lowers the volatility of inflation expectations. Despite the evidence of positive effects of clear communications on financial markets, it is interesting to observe that the clarity of communications of several central banks decreased during the global financial crisis, as documented by Bulir, Jansen, and Cihak(2012). More- over, a clear answer to the question of what drives the clarity of central bank communication has yet to be found (Bulir, Jansen, and Cihak 2013). In sum, notwithstanding the lack of consensus about the evolution and determinants of clarity of central bank communication, this strand of research provides evidence of beneficial effects stemming from communication clarity and supports the widespread perception that central bank communication has tradi- tionally been very technical and difficult to understand for most citizens (e.g., Norris 2005; Haldane 2018; Lange 2019). Economists have paid relatively less attention to the effects of central bank communica- tion clarity outside the realm of financial markets. As argued by Haldane and McMahon (2018), one reason why central banks may want to communicate more directly with the gen- eral public is to address the “twin deficits” of public understanding and trust in economics. Understanding and trust are endogenous and inextricably intertwined, as a lack of trust may inhibit understanding and, at the same time, poor understanding may contaminate trust (Haldane 2017). Trust, in turn, is critical for central banks because the efficacy of their policy action depends on it. This is well shown by the increasing evidence pointing to the fact that trust in central banks shapes citizen inflation expectations (e.g., Christelis et al. 2016; Mellina and Schmidt 2018; Baerg, Duell, and Lowe 2018). Also, in the face of intense pressures from political and economic actors, citizen support is even more needed for central banks to maintain their legitimacy (Kaltenthaler, Anderson, and Miller 2010).
3 To achieve higher levels of trust and understanding, it may be not sufficient to simply increase the quantity of communication with the public. As argued by Haldane, Macaulay, and McMahon(2020), for central bank communication to have an impact, it is important that households and businesses engage with it. Despite the increasing volumes of central bank communication with the public, the technical language used by central bankers may constitute an obstacle to effectively engage the public. As Haldane(2017), Coenen et al. (2017) and Haldane and McMahon(2018) show, the main publications of central banks in many advanced economies present a level of complexity that makes them too difficult to be understood by about 90% of the general public. Despite the evidence highlighting that central bank communication is often too difficult to understand for the average citizen, whether central bank communication clarity has an effect on engagement remains a blind spot in economic research. There are good reasons to hypothesise that this is the case. First, recent evidence from experimental research supports the idea that low levels of clarity constitute a key obstacle for central bank communica- tion to have a direct impact on public attitudes, expectations and behaviour. Haldane and McMahon(2018) employ survey experiments to provide evidence that clearer and simpler communications about inflation increase individuals’ understanding of the communicated messages, improve attitudes towards the communicating central bank, and generate greater alignment between individuals’ inflation expectations and central bank forecasts. Further- more Bholat et al.(2019) make use of online experiments to show that public comprehension of and trust in central banks can be improved if communications are made more relatable to people’s daily lives. Second, interesting observational evidence about the relationship between communication clarity and public engagement is offered by studies in political science. For instance, Reilly and Richey(2011) use measures of readability to code the complexity of 1,211 state-level ballot questions from 1997 to 2007 in the United States. The authors find that increased complexity leads to more non-response to ballot questions. Moreover, assessing the language of manifestos in Austria and Germany in the 1945-2013 period, Bischof and Senninger(2018) find that individuals are better able to place parties in the ideological space if parties use less complex campaign messages. Altogether, the findings from these studies offer good reasons to hypothesise that a positive and significant relationship between communication clarity and public engagement exists. As Haldane, Macaulay, and McMahon(2020) put it, for households and businesses to engage with the communication of central banks, “they need to read or see it and they need to take the message on board”. That points to the role of the media as the principal vehicle of the messages of central banks to the public. It is very difficult to imagine that the public can engage with the communication of central banks if this communication has a very low reach in the media. As Figure1 shows, media constitute a fundamental building block in the process via which communication shapes trust. Media act as a filter of the messages of central banks and other institutions, with the amount and type of media coverage significantly affecting public trust, as shown by a large literature on media effects on trust in and attitudes towards national and international institutions (e.g, Azrout, Van Spanje, and De Vreese 2012; Harteveld et al. 2018; Brosius, van Elsas, and De Vreese 2019).
4 Figure 1: Communications of, Engagement with and Trust in Central Banks
Public Engagement
Central Communications Coverage General Media Bank Public
Trust
For these reasons, the analysis of this paper focuses on media engagement as a proxy for public engagement. Taking into account different types of activities with which the ECB communicates with the public, we assess the relationship between communication clarity and media engagement in both traditional and social media. While Berger, Ehrmann, and Fratzscher(2011) and Binder(2017) have provided detailed assessment of how traditional media coverage responds to the communication activities of, respectively, the ECB and the Federal Reserve, previous studies have so far provided very little evidence that the attention of journalists and social media users to the communications of central banks is conditioned by the degree of clarity of central bank communication. One exception is given by Haldane, Macaulay, and McMahon(2020, 36-38), who compare the online media reach for the November 2017 Bank of England Inflation Report, which introduced simplified, layered content, with two counterfactual reports that presented no simplified content. Their analysis offers descriptive evidence suggesting that simpler communication may improve the reach of central bank communication. Drawing from this strand of economic research and from studies in the field of consumer psychology showing that readability shapes social media engagement (e.g., Leonhardt and Makienko 2017; Pancer et al. 2019), this paper attempts to provide the first systematic as- sessment of a link between central bank communication clarity and public engagement. To do so, we focus on the communication activities of the ECB. The next section presents the data and methods we employ to shed light on the effects of the ECB’s clarity of communication with the public.
3 Data and Method
3.1 Measuring Clarity of Communications Does the clarity of ECB communication activities shape public attention and engagement? To answer this question, we first have to clarify how clarity is operationalised. We measure communication clarity using two readability metrics: the Flesch-Kincaid Grade Level and the Gunning-Fog Index. Both metrics attempt to measure how easy it is to read a piece of text and are based on word length and sentence length. The Flesch-Kincaid (F-K) Grade
5 Level derives from the Flesch Reading Ease (FRE) Test (Flesch 1948) and provides a score that can be interpreted as the number of years of education generally required to understand a text. The higher is the score, the greater is the complexity of the language used. The measure is calculated using the following formula (Kincaid et al. 1975):
Total Words Total Syllables F-K Grade Level = 0.39 + 11.8 − 15.59 Total Sentences Total Words
Different from the F-K Grade Level, the Gunning-Fog (FOG) Index takes the number of complex words, rather than total syllables, into account in its calculation. Complex words are defined as those words with three or more syllables1 and which are not proper nouns, familiar words, or compound nouns. Thus, the FOG Index differs from the F-K Grade Level inasmuch as it relies on vocabulary-based features, namely word categories (e.g., familiar words, proper nouns, etc.) as operationalised by pre-specified lists. The metric is calculated based on the following formula (Gunning 1952)
Total Words Complex Words FOG Index = 0.4 x + 100 Total Sentences Total Words
Initially developed in the field of education research, these measures have been applied in a wide array of other fields, ranging from medicine (e.g., Paasche-Orlow, Taylor, and Brancati 2003) to journalism (e.g, Wasike 2018) and including many of the studies in economics (e.g., Jansen 2011a; Bulir, Cihak, and Jansen 2014; Montes et al. 2016; Smales and Apergis 2017) and political science (e.g., Reilly and Richey 2011; Spirling 2016; Lin and Osnabr¨ugge 2018; Schoonvelde et al. 2019; Rauh, Bes, and Schoonvelde 2020) discussed in the previous section. To be sure, these measures have not remained uncontested. Recently, there have been different cross-disciplinary attempts to produce more sophisticated measures of comprehen- sion difficulty of text. For instance, Mac Namee, Kelleher, and Fitzpatrick(2017) analyse a corpus of 948 English texts that were expertly annotated into five levels of comprehension difficulty to investigate whether the predictive accuracy of the FRE Test, the FOG Index and syntax-based features2. They find that a combination of all the three measures provides the best prediction accuracies, thereby suggesting that single readability measures might not provide a complete picture of textual difficulty. Benoit, Munger, and Spirling(2019) have introduced a new domain-specific approach to measuring sophistication through complexity in political texts. Their approach relies on crowdsourcing for the evaluation of textual diffi- culty and considers a wider set of features, such as word rarity, measured with dynamic term frequencies derived from Google Books data set. Despite these recent advancements and challenges to standard measures of textual complexity, the FRE Test, the F-K Grade Level and the FOG Index remain widely used in the social science literature and for good reasons. For instance, as pointed out by Schoonvelde et al.(2019), Benoit, Munger, and Spirling
1Common suffixes are not counted as syllables; e.g., -ed,-ing, etc 2Examples of syntax-based features are prepositions and prepositional clauses, conjunctions, subjunc- tions, adverbs and adverbial clauses. These can all pose difficulties to the reader. To capture these syntactic phenomena, Mac Namee, Kelleher, and Fitzpatrick(2017) create a set of features by first parsing the texts and then generating features from parse tree annotations.
6 (2019) find the FRE Test to be a crucial predictor of sophistication in political texts: with that measure alone it is possible to correctly predict 72% of the human coders’ judgements of difficult text snippets in their study. The introduction of various additional text features only marginally improves the prediction capacity of the FRE Test alone. Thus, we follow the approach adopted by most social science studies analysing language complexity and stick to the F-K Grade Level, which is based on the same features of the FRE Test3, and validate our results with the FOG Index in the robustness checks. Making use of these two metrics, we analyse the communication of the ECB across three different activities. First, we consider the speeches given by the members of the ECB Executive Board and Supervisory Board. Second, we take into account the monetary policy press conferences held by the ECB President ten times a year. Third, we study the ECB social media communications on Twitter.
3.2 ECB Communications and Traditional Media Engagement 3.2.1 ECB Speeches For ECB speeches, we use a dataset containing all the monetary policy and banking su- pervision speeches given by ECB Board members from January 1999 to October 2019 (i.e., from the beginning of the Presidency of Wim Duisenberg to the end of the Presidency of Mario Draghi), for a total of 2,175 measurable speeches which occurred on 1,655 days. For the days in which more than one speech was given, we consider the mean score of all the speeches given on that day. Table A1 in the Appendix provides summary statistics of the daily communication clarity in ECB speeches based on the measures of textual complexity discussed above. The mean F-K Grade Level of ECB speeches is 14.4, which falls inside the range of complexity identified by Haldane and McMahon(2018), who claim that typical cen- tral bank publications have a F-K Grade Levels of 14-18 (roughly equivalent to college-level education). Given the levels of literacy across the population, this makes the average speech of the ECB inaccessible to at least 90% of the general public. The summary statistics of Table A1, however, conceal some interesting degree of variation in communication clarity. The over time variation is displayed by Figure2. The figure shows the evolution of the F-K Grade Level in the corpus of ECB speeches from 1999 to 2019. Data points for speeches given by Executive Board members are marked by the light blue colour, while those of Supervisory Board members are red coloured. Also shown in the graph is the conditional mean of the readability score by Board category. The graph displays a significant degree of variation in the communication clarity of the ECB over time, with two patterns being worth highlighting. First, speeches of ECB Executive Board members have become clearer after the onset of the euro crisis in 2010. Before the 2010 period, the F-K Grade Level remained, on average, above 15, a degree of clarity that is accessible only to individuals who have graduate-level college education. After 2010, the level of complexity remains very high, but it falls below 15. Second, the average F-K Grade Level of Supervisory Board members has been much lower than that of Executive Board members4. A significant
3The only difference between the FRE Test and F-K Grade Level is that the latter produces a Grade Level rather than a score between 0 and 100. 4Sabine Lautenschl¨agerwas both member of the ECB Executive Board and Vice-Chair of the ECB
7 portion of Supervisory Board members’ speeches scores below 10, thereby being accessible also to individuals who do not hold a Bachelor’s degree.
Figure 2: Communication Clarity in ECB Speeches, 1999-2019 (F-K Grade Level)
20
15
Flesch−Kincaid Grade Level Flesch−Kincaid Grade 10
1999 2001 2003 2005 2007 2009 2011 2013 2015 2017 2019 Year
Executive Board Supervisory Board
Table A2 in the Appendix presents summary statistics of clarity of communication in ECB speeches by Board member. Supervisory Board Chair Andrea Enria and former Vice-Chair Sabine Lautenschl¨agerstand out as the Board members with the clearest communication in the sample. Their speeches, together with those of Dani`eleNouy, the former Chair of the Supervisory Board, drive the results of Figure2, showing that the communication of Supervisory Board members has, on average, been clearer than that of the members of the Executive Board. In the estimation of the relationship between communication clarity and media engage- ment presented in section4, we control for variables that may act as confounders in the analysis. Clearly, one of the most important confounders is given by the topics addressed by each speech. It is reasonable to assume that speeches focused on technical issues about monetary policy and banking supervision will be more difficult to understand than speeches
Supervisory Board from 2014 to 2019. Since her speeches focused mostly on banking supervision and were produced by the speechwriting unit of the Single Supervisory Mechanism, her speeches are classified as belonging to the Supervisory Board in Figure2.
8 related to less inherently technical topics, such as European affairs. To account for this differences in the topic of discussion, in line with recent studies using topic modelling to analyse the speeches given by central bankers (e.g., Moschella and Pinto 2019; Cross and Greene 2020; Moschella, Pinto, and Diodati 2020), we make use of a Structural Topic Model (STM) to obtain estimates of the distribution of topics in each speech, which we include as a control variable in our analysis. STM allows researchers to discover topics in a text corpus and conduct hypothesis test- ing about the relationship between topics and document metadata (Roberts et al. 2014; Roberts, Stewart, and Tingley 2019). The main advantage of employing STM in this setting is given by the nature and structure of ECB speeches. Different from other types of com- munications that are more focused in their content, such as ECB tweets, ECB speeches may usually cover multiple topics and do not always fall neatly into pre-defined categories. The mixed-membership type of estimation offered by topic models ensures a more comprehensive analysis of the content of the speeches than dictionary-based methods, while circumventing the problem of the difficult attribution of a given speech to a single category that might stem from the use of supervised methods. First, we create a text corpus of ECB speeches and convert text into a structured form, using standard text processing approaches. In particular, we lowercase and stem words, remove stopwords and numbers, and reduce the size of the document-frequency-matrix by considering only terms that appear in at least 2% of the documents to improve estimation efficiency (Proksch and Slapin 2009). Second, we estimate a topic model with 7 topics. The estimated topics appear easy to interpret, thereby contributing to a validation of the model selection based on interpretability (Chang et al. 2009). Table1 gives an overview of the estimated topics. In bold is indicated the label we have attributed to each topic. “Highest probability” is a simple measure that indicates which words are the most likely to belong to the topic. Extremely important are also “exclusive” words, namely those that are highly likely to appear in one topic and unlikely to appear in other topics based on the FREX metric (Bischof and Airoldi 2012; Airoldi and Bischof 2016). As shown by Table1, all topics can be easily labelled and ascribed to themes of discussion in central bank communication. Table A3 in the Appendix shows the distribution of topics across ECB speeches.
9 Table 1: Top Words for 7 Topic STM
Inflation Outlook Highest Probability: rate, polici, euro, area, monetari, inflat, bank Exclusivity (FREX): purchas, outlook, refinanc, recoveri, non-standard, eas, accommod Economic Outlook Highest Probability: growth, area, economi, countri, euro, econom, product Exclusivity (FREX): labour, product, globalis, growth, gdp, unemploy, reform Monetary Policy Highest Probability: polici, monetari, price, central, stabil, inflat, bank Exclusivity (FREX): analysi, strategi, communic, forecast, definit, vari- abl, bubbl Financial Stability Highest Probability: financi, market, bank, system, risk, liquid, credit Exclusivity (FREX): collater, financi, liquid, counterparti, market, fund, investor Payments & Banknotes Highest Probability: payment, bank, european, euro, servic, ecb, central Exclusivity (FREX): sepa, payment, banknot, statist, card, retail, settlement Banking Supervision Highest Probability: bank, crisi, risk, financi, supervis, european, supervisori Exclusivity (FREX): supervisori, supervis, supervisor, ssm, macro- prudenti, resolut, macroprudenti EU Affairs Highest Probability: euro, area, polici, econom, countri, european, monetari Exclusivity (FREX): currenc, emu, exchang, treati, union, fiscal, polit
Notes: This table presents the top words of the topics produced by a structural topic model with 7 topics, run on the corpus of speeches given by ECB Executive and Supervisory Board members from 1999 to 2019. The words with highest probability of occurrence and highest FREX score are shown for each topic.
10 3.2.2 ECB Press Conferences Regarding press conferences, our dataset includes the text from the introductory statement and the Q&A session with journalists from each press conference given by the ECB President from January 1999 to October 2019. This yields a text corpus of 191 press conferences. Table A4 presents the summary statistics of clarity of communication in ECB press conferences. The table also provides separate indication of the clarity of the introductory statement (IS) portion of the text, which usually follows a regular technical structure, and the Q&A session, which is typically less structured and technical in its content. The average level of complexity in press conferences appears to be lower than that of speeches. Importantly, the higher level of clarity in press conferences is driven by the Q&A session, which accounts for the greatest part of the press conference and whose mean F-K Grade Level is 11. Instead, the IS has a F-K Grade Level above 15, more in line with the level of complexity of ECB speeches. Figure3 shows the over time evolution of the communication clarity of ECB press con- ferences, distinguishing between different Presidency terms. Here the F-K Grade Level is based on both the ECB introductory statement and the Q&A with the journalists. The graph shows an increasing trend in communication complexity until 2009. Then, commu- nication clarity has significantly improved during the euro crisis. Starting from late 2014, namely around the time of the activation of the ECB Asset Purchase Programme, there is again an upward trend in communication complexity. This is followed by a reduction in 2018 and 2019, with the interesting outlier of the press conference of September 2019, which reaches the highest level of complexity in the sample.
Figure 3: Communication Clarity in ECB Press Conferences, 1999-2019 (F-K Grade Level)
●
15
14 ● ●
● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● 13 ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● 12 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● Flesch−Kincaid Grade Score Flesch−Kincaid Grade ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 11 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
● 10 2003 2005 2007 2009 2011 2013 2015 2017 2019
Year
● Duisenberg ● Trichet ● Draghi
11 In estimating the effect of press conference clarity on media engagement, we control for the announcement of policy changes on the day of the press conference. We combine data from different sources to create a dataset of ECB press conference announcements of conventional and unconventional policy changes. To identify changes in the key interest rates, i.e., the rate on the Deposit Facility (DFR), the rate on the Main Refinancing Operations (MRO) and the rate on the Marginal Lending Facility (MLF), we rely on data retrieved from the ECB Statistical Data Warehouse. To identify changes in unconventional monetary policies, we build on the timeline of ECB monetary policy measures developed by Hartmann and Smets(2018, 96). They divide unconventional policies into three categories, namely Credit Operations, Asset Purchases and Forward Guidance. Their timeline covers the August 2007- June 2018 period; we extend it to October 2019. Based on this information, we create a dummy variable for each conventional and unconventional policy change category, which takes value 1 on the day in which a given type of policy change occurred. Table2 provides an overview of the amount of conventional and unconventional monetary policy changes announced in ECB press conferences from 1999 to 2019.
Table 2: ECB Conventional and Unconventional Monetary Policy Changes
Frequency Deposit facility 27 Main refinancing operations 28 Marginal lending facility 28 Credit operations 10 Asset purchases 14 Forward guidance 4
3.2.3 Traditional Media Engagement To evaluate the impact of ECB speeches and press conferences on engagement, we make use of a novel dataset of newspaper articles mentioning the ECB and retrieved from the digital archive of Dow Jones Factiva. We employ the Data, News and Analytics (DNA) service: DNA is a platform that draws together more than 8,000 premium content sources within the Dow Jones ecosystem and licensed third-party publishers. Different from the standard Factiva archive, the content of Dow Jones DNA is licensed for text-mining and machine-learning use cases. Using a simple query search based on the keywords “ECB” and “European Central Bank”, we retrieve the universe of English articles mentioning the ECB from 1999 to 2019 in Dow Jones DNA archive. This yields a corpus of 1,624,324 articles. Figure4 provides a visualisation of the over time evolution of the monthly volume of ECB articles from 1999 to 2019.
12 Figure 4: Volume of English Articles Mentioning the ECB by Year, 1999-2019
20000
15000
10000 Coverage Volume Coverage
5000
0 Jan 1999 Jan 2000 Jan 2001 Jan 2002 Jan 2003 Jan 2004 Jan 2005 Jan 2006 Jan 2007 Jan 2008 Jan 2009 Jan 2010 Jan 2011 Jan 2012 Jan 2013 Jan 2014 Jan 2015 Jan 2016 Jan 2017 Jan 2018 Jan 2019 Jan 2020
Month Data Source: Dow Jones Factiva DNA Archive. Notes: The color scale variation is proportional to the monthly volume of ECB articles.
3.3 ECB Tweets and Social Media Engagement Regarding ECB digital communications, we rely on a novel dataset of all the tweets posted by the ECB Twitter account from January 2016 to July 2019, for a total of 6,276 tweets. The data were retrieved from the Analytics Dashboard of the Twitter account of the ECB. The dataset contains information about each tweet’s engagements (i.e., the amount of any type of Twitter users’ interaction with the tweet, including likes, retweets, link clicks, etc.), impressions (i.e., the amount of times the tweet has appeared in the timeline of Twitter users), and media content (i.e., the presence of pictures or videos embedded in the tweet). Other tweet features, such as the topic of the tweet, have been manually coded for a sub- sample of the dataset (n=1,790). The topic classification is based on eight categories (i.e., monetary policy, banking supervision, euro currency, financial stability, human resources, markets & payments, statistics, other). The main dependent variable we use to measure social media engagement in Section4 is the engagement rate of ECB tweets. The engagement rate offers an indication of the number of social media reactions generated by an ECB tweet, standardised by the number of users who saw the tweet:
Engagements Engagement Rate = x 1000 Impressions
Table A5 in the Appendix provides summary statistics of the measures communication clarity in ECB tweets, as well as of engagements, impressions and engagement rate at the
13 time of the data retrieval (i.e., November 2019). Figure5 provides a visualisation of the monthly evolution of the average engagement rate of ECB tweets.
Figure 5: Engagement Rate of ECB Tweets by Month, 2016-2019
9
6 Average Engagement Rate Average 3
0 Jan 2016 Apr 2016 Jul 2016 Oct 2016 Jan 2017 Apr 2017 Jul 2017 Oct 2017 Jan 2018 Apr 2018 Jul 2018 Oct 2018 Jan 2019 Apr 2019 Jul 2019
Month Data Source: ECB Twitter Account. Notes: The color scale variation is proportional to the monthly average engagement rate of ECB tweets.
In the analysis, we control for the topic of each tweet. Since only a subset of the corpus of ECB tweets has been manually coded, we rely on a supervised machine learning algorithm to automatically extend the topic classification to each non-coded tweet. We proceed in the following manner. We consider only tweets in English language and apply a standard text pre-processing procedure. In particular, we remove punctuations and stopwords, apply stemming, and keep only words that appear in at least two tweets. After pre-processing, we train a Support Vector Machine (SVM) algorithm on a set of 1,400 tweets and test its performance on a set of 390 tweets. Table3 presents the performance metrics of the SVM classifier for each topic category. Overall, the SVM algorithm performs well in the topic classification, attaining an average level of accuracy that is close to 90%. We apply this topic classification to each non-coded tweet in the dataset and use the assigned category as a control variable in the analysis of the relationship between clarity of ECB tweets and social media engagement.
14 Table 3: SVM Performance Metrics
Metric All BankSup Euro FinStab HR MonPol Other MIP Stats Accuracy 0.889 Precision 0.797 0.739 0.333 0.667 0.957 0.899 0.910 0.923 0.950 Recall 0.739 0.500 0.444 0.555 0.880 0.963 0.765 0.706 0.905 F1 0.739 0.400 0.533 0.917 0.917 0.930 0.830 0.800 0.927
Notes: Accuracy is the proportion of correct predictions among all prediction. Precision is the proportion of tweets correctly assigned to a category among all of the tweets assigned to that category. Recall is the proportion of tweets correctly assigned to a category among all of the tweets that truly belong to that category. F1 is a harmonic mean of precision and recall.
4 Results
Does communication clarity increase public attention towards the messages sent by cen- tral banks? In this section, we shed light on the effects of ECB communication clarity on traditional and social media engagement. We perform three sets of analyses. First, we consider the relationship between the clarity of communication in ECB speeches and the media coverage of the ECB on the day in which a speech is given. Thus, we take the number of daily articles about the ECB retrieved from the Dow Jones DNA archive as our dependent variable. Before performing our analysis, we implement some adjustments to our dataset. In our dataset, there are 1,655 days in which at least one speech was given by an ECB Board member. In 409 of these 1,655 days more than one ECB speech was given. Since we have a daily dependent variable, we take the average Flesch-Kincaid Grade Level of all the speeches that were held on the same day as the main independent variable in our analysis. Similarly, as a control variable, we consider the average topic proportions from all the speeches that were given on the same day. Additionally, we include a dummy variable to control for days in which a speech was given by the ECB President and an ordinal variable indicating the number of ECB speeches occurring on each day. The dependent variable in our analysis is a discrete variable, whose distribution is limited to non-negative values. As shown by Figure A1 in the Appendix, the distribution of the volume of articles about the ECB is highly skewed, with most days having a relatively low volume of ECB coverage and few days exhibiting a disproportionately higher amount of coverage. As a result, the volume of coverage is not normally distributed and the variance of the distribution is higher than its mean. Hence, we model the relationship between speech features and media engagement using a negative binominal model. For robustness checks, in Section5, we carry out the same analysis using an ordinary least squares model and show that the results are broadly equivalent independent of the model selection.
15 Table 4: Language Clarity in ECB Speeches and Traditional Media Engagement: Negative Binomial Regressions
(1) (2) (3) (4)
Flesch-Kincaid (Speeches) -0.046*** -0.035*** -0.034*** -0.013** (-4.43) (-2.77) (-2.74) (-2.05)
Topic Proportion Inflation Outlook 0.727*** 0.764*** 0.174* (5.06) (5.47) (1.72)
Economic Outlook -0.005 0.063 -0.073 (-0.03) (0.40) (-0.73)
Monetary Policy 0.055 0.160 0.097 (0.34) (0.98) (1.04)
Financial Stability 0.221 0.305** -0.004 (1.52) (2.08) (-0.05)
Payments & Banknotes -0.028 0.051 -0.005 (-0.22) (0.39) (-0.06)
Banking Supervision 0.931*** 1.042*** 0.096 (6.46) (7.33) (0.96)
President Speaking 0.124*** 0.092*** (3.04) (4.12)
No of Speeches 0.012 0.013 (0.49) (0.83)
Constant 6.318*** 5.842*** 5.701*** 4.642*** (41.55) (25.53) (25.62) (35.55) Wald χ2 19.59*** 155.67*** 168.18*** 3168.73∗∗∗ Log pseudolikelihood -10759.129 -10685.83 -10678.907 -9744.8852 Observations 1,655 1,655 1,655 1,655 Time Fixed Effects X Notes: This table presents results from negative binomial regressions with robust standard errors. The dependent variable is given by the daily amount of articles mentioning the ECB on days in which at least one ECB speech was given. The main independent variable is the daily average the F-K Grade Level of ECB speeches. Topic proportions are given by the daily average of topics estimated using the Structural Topic Model discussed in Section 3.2. t statistics in are in parentheses, and *** p<0.01, ** p<0.05, * p<0.1.
Table4 presents the results of our analysis of communication clarity in ECB speeches and traditional media coverage. In column (1), the model includes only the daily measure of clarity in ECB speeches based on the F-K Grade Level, together with a constant. In column
16 (2), we add topic proportions as controls for the different types of content of ECB speeches, using the most prevalent topic, namely the one about EU Affairs, as a reference category. In column (3), we add controls for days in which the President gave a speech and for the number of speeches given on each day. In column (4), we add time fixed effects, using factor variables for the day of the week, the month and the year in which speeches were given. Each column presents the Wald χ2 and the log pseudolikelihood of the model. All models include robust standard errors. The results indicate a significant and robust negative relationship between speech com- plexity and media coverage of the ECB. The easier it is to understand a speech given by an ECB Board member, the more likely it is that journalists write about the ECB. The inclusion of controls for topics, presidential speeches and number of speeches produces only a slight reduction in the magnitude of the estimated coefficient and the t-statistic of the F-K Grade Level variable. No single topic is estimated to have a significantly different effect on media coverage compared to the one about EU Affairs, with the exception of the Inflation Outlook topic, which is presumably associated with signals about the future monetary policy stance of the ECB. The inclusion of time fixed effects contributes to a further reduction in the magnitude of the coefficient of the main independent variable, which, however, remains statistically significant at the 5% level. To better quantify the size of the relationship between speech clarity and media coverage, Table5 presents the estimated marginal effects of the F-K Grade Level variable from ECB speeches. For each column of Table5, the results are based on the negative binomial regres- sion models estimated in the corresponding column of Table4. The results show that a unit variation in the F-K Grade Level of ECB speeches is correlated with an average increase of more than 13 articles on the ECB in the most parsimonious specification, and slightly less than 4 articles in the most stringent one.
Table 5: Marginal Effects of Clarity of ECB Speeches on Traditional Media Engagement
(1) (2) (3) (4)
Flesh-Kincaid (Speeches) -13.262*** -9.975*** -9.709*** -3.842** (-4.38) (-2.74) (-2.71) (-2.05)
Notes: This table presents the marginal effects of the F-K Grade Level variable from the negative binomial models estimated in Table4. t statistics in are in parentheses, and *** p<0.01, ** p<0.05, * p<0.1.
Second, we assess the relationship between communication clarity in ECB press confer- ences and ECB media coverage on press conference day. The dependent variable is now given by the number of ECB articles that were published on the day of the ECB press conference. In our analysis, we start by considering the F-K Grade Level of all the textual content of the press conference and subsequently distinguish between the textual content of the introduc- tory statement (IS) given by the ECB President and the Q&A session between the President and the journalists attending the press conference. As for Table4, our baseline specification is based on a negative binomial model with robust standard errors.
17 Table 6: Language Clarity and Media Engagement on Press Conference Day: Negative Binomial Regressions
(1) (2) (3) (4) (5) (6)
Flesch-Kincaid (All PC) -0.106** -0.137*** -0.147*** -0.088*** (-2.04) (-2.87) (-3.38) (-3.36)
Flesch-Kincaid (Only IS) 0.035 (1.12)
Flesch-Kincaid (Only Q&A) -0.086*** (-4.56)
Policy Change Deposit Facility Rate 0.411*** 0.357*** 0.372*** 0.378*** 0.366*** (3.67) (3.11) (5.97) (5.95) (5.93)
Credit Operations 0.308** 0.149* 0.130* 0.145* (2.14) (1.92) (1.69) (1.91)
Asset Purchases 0.378** 0.101 0.100 0.094
18 (2.04) (0.78) (0.78) (0.74)
Forward Guidance 0.189 0.351 0.320 0.363 (0.87) (1.58) (1.33) (1.62)
Constant 7.795*** 8.088*** 8.157*** 6.637*** 5.039*** 6.487*** (12.39) (14.03) (15.59) (20.09) (9.92) (30.22) Wald χ2 4.17** 18.59*** 43.10*** 2311.63*** 2125.97*** 2386.85*** Log pseudolikelihood -1387.7048 -1381.1199 -1372.8743 -1211.9226 -1196.6982 -1209.6888 Observations 191 191 191 191 191 191 Time Fixed Effects XXX
Notes: This table presents results from negative binomial regressions with robust standard errors. The dependent variable is given by the daily amount of articles mentioning the ECB on days of ECB press conferences. In columns (1)-(4), the main independent variable is the F-K Grade Level of the entire press conference. In column (5), the main independent variable is the F-K Grade Level of the ECB President’s introductory statement to the press conference. In column (6), the main independent variable is the F-K Grade Level of the Q&A session between the ECB President and the journalists attending the press conference. The construction of the policy change dummy variables is described in Section 3.2. t statistics in are in parentheses, and *** p<0.01, ** p<0.05, * p<0.1. Table6 presents the results of the analysis of the relationship between communication clarity and ECB coverage for ECB press conferences. In columns (1)-(4), the independent variable is the F-K Grade Level of the entire press conference. In column (1), the model includes only the measure of communication clarity as an explanatory variable and the inter- cept. In column (2), we add a control for the announcement of changes in the conventional monetary policy of the ECB5. In column (3), we add controls for changes in the unconven- tional monetary policy of the ECB, distinguishing between credit operations, asset purchases and forward guidance, as explained in Section 3.2.2. In column (4), we add time fixed ef- fects using factor variables for the month and year of each press conference. In column (5), we maintain the specification of column (4) and use the F-K Grade Level of the introduc- tory statement as the main independent variable. In column (6), we assess the relationship between media coverage and the F-K Grade Level of the Q&A session. The results show that a higher level of complexity of the press conference is negatively associated with the volume of media coverage on the ECB on press conference day. Impor- tantly, this relationship is not the product of the inherently higher degree of technicality that may be related to the announcement of new policy measures. Interestingly, the inclusion of controls determines an increase in the magnitude and statistical significance of the coefficient of the F-K Grade Level variable. Also, the understandability of the introductory statement of the ECB President is estimated to have no impact on the amount of articles that journal- ists write about the ECB on the day of the press conference. What drives the results is the clarity of the President’s exchange with the journalists during the Q&A Session. Table7 shows the marginal effects of the F-K Grade Level variable from the models estimated in Table6. A unit variation in the degree of complexity of the press conference is associated with a drop in the volume of ECB coverage that goes from roughly 72 articles in the most parsimonious specification to around 60 in the most stringent one.
Table 7: Marginal Effects of Clarity of ECB Press Conferences on Traditional Media En- gagement
(1) (2) (3) (4) (5) (6)
Flesch-Kincaid (All PC) -72.490** -93.223*** -100.341*** -59.952*** (-2.00) (-2.76) (-3.22) (-3.34) Flesch-Kincaid (Only IS) 24.114 (1.12) Flesch-Kincaid (Only Q&A) -58.489*** (-4.50)
Notes: This table presents the marginal effects of the F-K Grade Level variable from the negative binomial models estimated in Table6. t statistics in are in parentheses, and *** p<0.01, ** p<0.05, * p<0.1.
5Here, we consider only changes in the deposit facility rate. We do so because the dummy variables for changes in the main refinancing operations and changes in the marginal lending facility are highly correlated with this variable. As a result, the estimated coefficients for these three dummies, when they are jointly included, are not statistically significant, signalling a potential problem of multicollinearity. This does not affect the statistical significance of the F-K Grade Level coefficient. The results based on a model including all the three dummy variables of conventional policy changes are available upon request.
19 Table 8: Language Clarity and Twitter Engagement: Negative Binomial Regressions
(1) (2) (3) (4) (5) (6) Dep. Variable Eng. Rate Eng. Rate Eng. Rate Eng. Rate Engagements Impressions
Flesch-Kincaid (Tweets) -0.027*** -0.023*** -0.013*** -0.014*** -0.019** -0.014*** (-6.66) (-5.54) (-3.38) (-3.90) (-2.33) (-4.08)
Topic Banking Supervision -0.040 -0.055 -0.048 -0.099 -0.042* (-1.04) (-1.36) (-1.17) (-1.43) (-1.71)
Euro Currency 0.514*** 0.312*** 0.332*** 0.557*** 0.183* (4.40) (2.64) (2.92) (2.69) (1.91)
Financial Stability -0.072 -0.180*** -0.116** -0.174* 0.036 (-0.96) (-3.25) (-2.11) (-1.68) (0.51)
Human Resources -0.198*** -0.099*** -0.099*** -0.245*** -0.095*** (-6.43) (-3.48) (-3.56) (-6.00) (-5.63)
Markets & Payments 0.411*** 0.269** 0.358*** 1.727*** 0.355 (3.15) (2.00) (2.63) (2.93) (1.63) 20 Statistics -0.163*** -0.099** -0.093** -0.236*** -0.093*** (-3.28) (-2.38) (-2.24) (-3.83) (-3.27)
Other 0.025 0.007 -0.032 -0.098 -0.134*** (0.51) (0.16) (-0.77) (-0.90) (-4.63)
Media Content 0.555*** 0.541*** 0.859*** 0.221*** (20.79) (20.80) (13.22) (8.11)
Constant 2.230*** 2.191*** 1.905*** 1.724*** 4.713*** 9.956*** (48.76) (43.84) (40.55) (28.89) (35.71) (170.52) Wald χ2 44.39*** 154.75*** 658.88*** 1288.86*** 955.74*** 1392.13*** Log pseudolikelihood -17635.418 -17560.644 -17128.255 -16852.539 -37554.608 -66115.958 Observations 6,276 6,276 6,276 6,276 6,276 6,276 Time Fixed Effects XXX Notes: This table presents results from negative binomial regressions with robust standard errors. The main independent variable is the tweet- level F-K Grade Level. In columns (1)-(4), the dependent variable is given by the tweet-level engagement rate. In columns (5), the dependent variable is the absolute amount of tweet-level engagements. In columns (6), the dependent variable is the absolute amount of tweet-level impressions. Topic categories are based on the Support Vector Machine-based classification discussed in Section 3.3. t statistics in are in parentheses, and *** p<0.01, ** p<0.05, * p<0.1. Third, we study the relationship between the clarity of ECB digital communications and social media engagement on Twitter. Table8 presents the results of negative binomial regression models with robust standard errors. In columns (1)-(4), the dependent variable is the engagement rate of each ECB tweet in our dataset. The main independent variable is given by the F-K Grade Level of each tweet. In column (1), the model includes only the F-K Grade Level and the intercept. In column (2), we augment the model with the topic classification obtained with SVM and discussed in Section 3.3. In column (3), we add a dummy variable indicating the presence of media content (e.g., picture, video, etc.) as a control variable. In column (4), we add time fixed effects, using factor variables for the day of the week, the month and the year in which a tweet was posted. In columns (5) and (6), we use the same model specification of column (4), but we consider, respectively, engagements and impressions as dependent variables. The results show that the clarity of central bank digital communications is a significant and strong predictor of social media engagement. Tweets containing more difficult language generate, on average, a lower number of users interacting with them. This effect holds when controlling for tweet-specific features and time-varying factors. To better assess the magnitude of the effect, Table9 presents the marginal effects based on the models estimated in Table8. In the least parsimonious specification, a unit variation in the F-K Grade Level of a tweet is associated with an average drop of 0.09 points in the engagement rate, corresponding to an average drop of around 4 interactions and a reduction of more than 300 impressions. That is to say, it is predicted that, with the presence of one additional point score in the F-K Grade Level, on average, an ECB tweet has at least 4 less users interacting with it and 300 less users seeing it in their timelines.
Table 9: Marginal Effects of Clarity of ECB Tweets on Social Media Engagement
(1) (2) (3) (4) (5) (6) Dep. Variable Eng. Rate Eng. Rate Eng. Rate Eng. Rate Engagements Impressions
Flesh-Kincaid (Tweets) -0.187*** -0.161*** -0.089*** -0.093*** -3.709** -313.169*** (-6.69) (-5.58) (-3.40) (-3.92) (-2.31) (-3.98)
Notes: This table presents the marginal effects of the F-K Grade Level variable from the negative binomial models estimated in Table8. t statistics in are in parentheses, and *** p<0.01, ** p<0.05, * p<0.1.
Figure6 provides a visualisation of the predicted levels of engagement from the three sets of analyses carried out in this section. Speeches containing clearer language (e.g., F-K Grade Level ≤ 11) are, on average, estimated to be associated with more than 250 articles on the ECB on the same day in which they are given. The same speech, but with a lower degree of communication clarity (e.g., F-K Grade Level ≥ 15), is predicted to generate at least 10 articles less on the same day. The magnitude of this effect is much higher for press conferences. Ceteris paribus, the variation between greater clarity (e.g., F-K Grade Level ≤ 11) and lower clarity (e.g., F-K Grade Level ≥ 15) in press conferences corresponds to a difference in the volume of coverage on the ECB of at least 100 articles on press conference day. Finally, tweets containing clearer language (e.g., F-K score ≤ 11) are estimated to have,
21 on average, an engagement rate that is above 6.5. The engagement rate drops to below 6.25 for tweets with more complex language (i.e., F-K score ≥ 15). As shown in Figure A2 in the Appendix, if we consider engagements and impressions as dependent variables, this corresponds to a variation of around 14 interactions and 1,100 impressions between clearer tweets and more complex ones.
Figure 6: Predicted Engagement of ECB Speeches, Press Conferences and Tweets Condi- tional on Communication Clarity
(a) Speeches (b) Press Conferences
(c) Tweets
Notes: The visualisations presented by the different sub-figures are based on, respectively, column (4) of Table4 (a), column (4) of Table6 (b) and column (4) of Table8 (c).
22 5 Robustness Checks and Additional Analysis
In this section, we discuss the robustness of our results to different model specifications and measures of communication clarity. Table 10 presents the marginal effects of different analyses of the relationship between clarity in ECB speeches and press conferences and the volume of ECB media coverage. In particular, we replicate the results obtained in Tables 4 and6 by: a) making use of the FOG Index (discussed in Section 3.1) instead of the F- K Grade Level; b) making use of ordinary least squares (OLS) regression models instead of negative binomial regression models. In all cases, we estimate the most stringent specification including the full set of control variables and time fixed effects. In column (1), we assess the effects of clarity of communication in ECB speeches on ECB media coverage using the FOG Index and a negative binomial model. In column (2), we resort to the F-K Grade Level, but we make use of an OLS regression. In column (3), we use the FOG Index in an OLS regression framework. In column (4), we study the effects of clarity of communication in ECB press conferences on ECB media coverage using the FOG Index and a negative binomial model. In column (5), we resort to the F-K Grade Level and OLS, while we use the FOG Index and OLS in column (6).
Table 10: Robustness Checks: Speeches and Press Conferences (Marginal Effects)
(1) (2) (3) (4) (5) (6) Model NB OLS OLS NB OLS OLS
FOG (Speeches) -3.344** -4.837** (-2.01) (-2.14 ) Flesch-Kincaid (Speeches) -5.786** (-2.22) FOG (All PC) -51.441*** -61.507** (-3.04) (-2.29) Flesch-Kincaid (All PC) -66.177** (-2.28) Observations 1,655 1,655 1,655 191 191 191 Controls XXXXXX Time Fixed Effects XXXXXX
Notes: This table presents marginal effects from different model specifications and operationalisations of the communication clarity variable for ECB speeches and press conferences. In columns (1)-(3), the dependent variable is given by the daily amount of articles mentioning the ECB on days in which at least one ECB speech was given. In columns (4)-(6), the dependent variable is given by the daily amount of articles mentioning the ECB on days of ECB press conferences. In columns (1) and (3), the main independent variable is the daily average the FOG Index of ECB speeches. In column (2), the main independent variable is the daily average the F-K Grade Level of ECB speeches. In columns (4) and (6), the main independent variable is the FOG Index of the entire press conference. In column (5), the main independent variable is the F-K Grade Level of the entire press conference. Columns (1) and (4) present results from negative binomial regressions. Columns (2), (3), (5) and (7) present results from ordinary least squares regressions. All models include robust standard errors, the full set of control variables and time fixed effects. t statistics in are in parentheses, and *** p<0.01, ** p<0.05, * p<0.1.
The results of Table 10 show that our findings are robust across different specifications.
23 The use of the FOG Index and OLS regressions does not substantially alter the conclusions reached in the previous section. Greater complexity in the speeches and presse conferences of the ECB is consistently associated with lower volumes of ECB media coverage, with this effect being precisely estimated across all specifications. Finally, in Table 11, we replicate and extend the analysis of digital communication clarity and social media engagement. Similar to Table 10, we assess the sensitivity of our findings when operationalising the clarity variable with the FOG Index and when using OLS regres- sions. We can here extend the set of dependent variables considered in our assessment of the effects of communication clarity. In columns (1)-(3), the main dependent variable we consider is the tweet-level engagement rate. In columns (4) and (5), we verify our results by decomposing the engagement rate into its two components, namely engagements and impressions, and considering them as dependent variables separately. In column (6), we ex- tend the analysis to consider whether communication clarity has an effect on the likes rate, namely the amount of likes a tweet receives, standardised by the number of impressions. We use the FOG Index in columns (1), (3), (4) and (5), while sticking to the F-K Grade Level in columns (2) and (6). In columns (2) and (3), we use OLS regressions, while the other columns present results from negative binomial models.
Table 11: Robustness Checks and Additional Analyses: Tweets (Marginal Effects)
(1) (2) (3) (4) (5) (6) Model NB OLS OLS NB NB NB Dep. Variable Eng. Rate Eng. Rate Eng. Rate Engagements Impressions Likes Rate
FOG (Tweets) -0.056*** -0.057*** -3.718** -249.098*** (-3.45) (-3.03) (-2.94) (-3.79) Flesch-Kincaid (Tweets) -0.094*** -.0003 (-3.22) (-0.11) Observations 6,276 6,276 6,276 6,276 6,276 6,276 Controls XXXXXX Time Fixed Effects XXXXXX
Notes: This table presents marginal effects from different model specifications and operationalisations of the communication clarity variable for ECB tweets. In columns (1)-(3), the dependent variable is given by the tweet-level engagement rate. In columns (4), (5) and (6), the dependent variable is given, respectively, by tweet-level engagements, impressions and likes rate. In columns (1), (3) (4) and (5), the main independent variable is the tweet-level FOG Index. In columns (2) and (6), the main independent variable is the tweet- level F-K Grade Level. Columns (1), (4), (5) and (6) present results from negative binomial regressions. Columns (2) and (3) present results from ordinary least squares regressions. All models include robust standard errors, the full set of control variables and time fixed effects. t statistics in are in parentheses, and *** p<0.01, ** p<0.05, * p<0.1.
Even in this case, we do not detect significant sensitivity of our results to different specifications. The communication clarity variable remains a significant predictor of the level of engagements and views obtained by ECB tweets independent of its operationalisation. Interestingly, while these results point to a strong link between communication clarity and media engagement, column (6) of Table 11 provides suggestive evidence that greater clarity and engagement might not directly translate into greater favourability towards the ECB
24 communication activities. Indeed, there appears to be no sizeable and significant correlation between the F-K Grade Level of ECB tweets and the amount of likes they received. We leave the exploration of the relationship between communication clarity and public favourability for future research.
6 Conclusion
Despite important improvements over the past decade, central bank communication has not entirely abandoned its Delphic nature. Several key policymakers and privileged observers have been vocal about it, highlighting that the communications of central bankers are inac- cessible to the large majority of citizens. A growing consensus around the need for clearer communications in the realm of central banking has emerged, but a systematic assessment of the effects of communication clarity on engagement has so far been missing. Focusing on the clarity of ECB communication activities, this paper has provided evi- dence that clearer communication is strongly and robustly correlated with a higher degree of engagement of journalists and social media users. Communication clarity is an important predictor of the volume of articles on the ECB published on the days of ECB communica- tion activities, as well as it plays an important in explaining the social media engagement generated by ECB digital communications on Twitter. In other words, journalists and so- cial media users tend to talk less about and engage less with the ECB when its language is more difficult to understand. These results hold when controlling for several potential confounders and are robust across different model specifications and operationalisations of communication clarity. Taken together, these findings yield significant implications for the communication policy of central banks. The ECB is a very good case in point in this regard. Newly appointed ECB President Christine Lagarde has never hidden her willingness to promote clear and effective communications with the euro area citizens. For instance, during her first hearing as ECB President before the European Parliament in December 2019, she stressed the importance of this objective by saying: “It is important to me that our focus on connecting with the people we serve continues and grows stronger - in particular by improving the ways in which we communicate with the general public” (ECB 2019). The appropriateness of establishing a connection with European citizens is arguably greater at a time of heightened politicisation of monetary policy in Europe (Tortola 2020), a trend that is not bound to disappear in the wake of the COVID-19 crisis. Media and several public actors have shown increasing hostility towards the unconventional monetary policies of the ECB and have raised doubts over their legitimacy. The ECB has been repeatedly accused of applying “punishment rates” and “expopriating savers” in the euro area (e.g., Esslinger 2019). In this context, the ECB has been increasingly faced with the need to take a more proactive stance in its communication with the public. The results of this paper suggest that, for the ECB to spread its messages among euro area citizens, it is important to employ a language that is simple and easily accessible to the public. Future research may extend the analysis of this paper in several interesting directions. In our view, two could be particularly promising. First, this paper has focused on the effects of clarity of central bank communication with the public, while it has said very little
25 about its determinants. Future research should advance our understanding of how central bankers adapt their language based on different targeted audiences and how different features of central bank communication are responsive to changing macroeconomic and political scenarios. Second, it is important to push forward the analysis of the effects produced by different features of central bank communication. As successfully shown by recent studies (e.g., Baerg, Duell, and Lowe 2018; Haldane and McMahon 2018; Kryvtsov and Petersen 2020), survey and lab experiments promise to be extremely useful tools to assess the impact of central bankers’ language on citizen attitudes, expectations and behaviour.
References
Airoldi, E. M. and J. M. Bischof (2016). Improving and Evaluating Topic Models and Other Models of Text. Journal of the American Statistical Association 111 (516), 1381–1403.
Azrout, R., J. Van Spanje, and C. De Vreese (2012). When News Matters: Media Effects on Public Support for European Union Enlargement in 21 Countries. Journal of Common Market Studies 50 (5), 691–708.
Baerg, N., D. Duell, and W. Lowe (2018). Central Bank Communication as Public Opin- ion: Experimental Evidence. Paper Presented at the 2018 IPES Conference in Boston, Massachussets.
Benoit, K., K. Munger, and A. Spirling (2019). Measuring and Explaining Political So- phistication through Textual Complexity. American Journal of Political Science 63 (2), 491–508.
Berger, H., M. Ehrmann, and M. Fratzscher (2011). Monetary Policy in the Media. Journal of Money, Credit and Banking 43 (4), 689–709.
Bholat, D., N. Broughton, A. Parker, J. T. Meer, and E. Walczak (2019). Enhancing Central Bank Communications Using Simple and Relatable Information. Journal of Monetary Economics 108, 1–15.
Binder, C. (2017). Federal Reserve Communication and the Media. Journal of Media Eco- nomics 30 (4), 191–214.
Bischof, D. and R. Senninger (2018). Simple Politics for the People? Complexity in Cam- paign Messages and Political Knowledge. European Journal of Political Research 57, 473–495.
Bischof, J. M. and E. M. Airoldi (2012). Summarizing Topical Content with Word Frequency and Exclusivity. In Proceedings of the 29th International Conference on Machine Learning, pp. 9–16.
Blinder, A. S., M. Ehrmann, M. Fratzscher, J. de Haan, and D.-J. Jansen (2008). Central Bank Communication and Monetary Policy: A Survey of Theory and Evidence. Journal of Economic Literature 46 (4), 910–945.
26 Brosius, A., E. J. van Elsas, and C. De Vreese (2019). How Media Shape Political Trust: News Coverage of Immigration and its Effects on Trust in the European Union. European Union Politics 20 (3), 447–467. Bulir, A., M. Cihak, and D.-J. Jansen (2014). Does the Clarity of Inflation Reports Affect Volatility in Financial Markets? IMF Working Paper No. 14/175 . Bulir, A., D.-J. Jansen, and M. Cihak (2012). Clarity of Central Bank Communication About Inflation. IMF Working Paper No. 12/9 . Bulir, A., D.-J. Jansen, and M. Cihak (2013). What Drives Clarity of Central Bank Com- munication About Inflation? Open Economies Review 24 (1), 122–145. Chang, J., S. Gerrish, C. Wang, J. L. Boyd-Graber, and D. M. Blei (2009). Reading Tea Leaves: How Humans Interpret Topic Models. In Advances in Neural Information Pro- cessing Systems, pp. 288–296. Christelis, D., D. Georgarakos, T. Jappelli, and M. van Rooij (2016). Trust in the Central Bank and Inflation Expectations. DNB Working Paper No. 537 . Coenen, G., M. Ehrmann, G. Gaballo, P. Hoffmann, A. Nakov, S. Nardelli, E. Persson, and G. Strasser (2017). Communication of Monetary Policy in Unconventional Times. ECB Working Paper Series No. 2080 . Cross, J. P. and D. Greene (2020). Talk Is Not Cheap: Policy Agendas, Information Process- ing, and the Unusually Proportional Nature of European Central Bank Communications Policy Responses. Governance 33 (2), 425–444. Cukierman, A. and A. Meltzer (1986). A Theory of Ambiguity, Credibility and Inflation under Discretion and Asymmetric Information. Econometrica 54 (5), 1099–1128. ECB (2019). Introductory Statement at the ECON Committee of the European Parliament. URL: https://www.ecb.europa.eu/press/key/date/2019/html/ecb. sp191202~f8d16c9361.en.html. (2 December 2019). ECB (2019). Press Conference with Q&A. URL: https://www.ecb.europa.eu/press/ pressconf/2019/html/ecb.is191024~78a5550bc1.en.html. (24 October 2019, accessed on 1 October 2020). Esslinger, L. (2019). Wie Gruselig War Graf Draghila Wirklich? Bild. URL: https://www.bild.de/bild-plus/geld/mein-geld/leben-und-sparen/ezb-chef- mario-draghi-wie-gruselig-war-graf-draghila-wirklich-65534312,view= conversionToLogin.bild.html (24 October 2019, accessed 1 February 2020). Federal Reserve (2019). Q&A at the Town Hall Meeting with Educators at the Federal Re- verse Board. URL: https://www.federalreserve.gov/mediacenter/files/teacher- town-hall-transcript-20190206.pdf. (6 February 2019, accessed on 1 October 2020). Flesch, R. (1948). A New Readability Yardstick. Journal of Applied Psychology 32 (3), 221–233.
27 Fracasso, A., H. Genberg, and C. Wyplosz (2003). How Do Central Banks Write? An Evaluation of Inflation Targeting Central Banks. Geneva Reports on the World Economy Special Report 2 .
Geraats, P. (2002). Central Bank Transparency. Economic Journal 112 (483), F532–F565.
Gunning, R. (1952). The Technique of Clear Writing. New York, NY: McGraw-Hill.
Haldane, A. (2017). A Little More Conversation A Little Less Action. Speech given at the Federal Reserve Bank of San Francisco Macroeconomics and Monetary Policy Conference. (31 March 2017).
Haldane, A. (2018). Climbing the Public Engagement Ladder. Speech given at the Royal Society for the Encouragement of Arts, Manufactures and Commerce. (6 March 2018).
Haldane, A., A. Macaulay, and M. McMahon (2020). The 3 E’s of Central Bank Communi- cation with the Public. Bank of England Staff Working Paper No. 847 .
Haldane, A. and M. McMahon (2018). Central Bank Communications and the General Public. AEA Papers and Proceedings 108, 578–583.
Harteveld, E., J. Schaper, S. L. D. Lange, and W. Van Der Brug (2018). Blaming Brussels? The Impact of (News about) the Refugee Crisis on Attitudes towards the EU and National Politics. Journal of Common Market Studies 56 (1), 157–177.
Hartmann, P. and F. Smets (2018). The First Twenty Years of the European Central Bank: Monetary Policy. ECB Working Paper Series No. 2219 .
Issing, O. (2005). Communication, Transparency, Accountability: Monetary Policy in the Twenty-First Century. Federal Reserve Bank of St. Louis Review 87 (2), 65–83.
Jansen, D.-J. (2011a). Does the Clarity of Central Bank Communication Affect Volatility in Financial Markets? Evidence from Humphrey-Hawkings Testimonies. Contemporary Economic Policy 29 (4), 494–509.
Jansen, D.-J. (2011b). Mumbling with Great Incoherence: Was It Really so Difficult to Understand Alan Greenspan? Economics Letters 113 (1), 70–72.
Kaltenthaler, K., C. J. Anderson, and W. J. Miller (2010). Accountability and Indepen- dent Central Banks: Europeans and Distrust of the EuropeanCentral Bank. Journal of Common Market Studies 48 (5), 1261–1281.
Kincaid, J. P., R. P. Fishburne, R. L. Rogers, and B. S. Chissom (1975). Derivation Of New Readability Formulas (Automated Readability Index, Fog Count And Flesch Reading Ease Formula) For Navy Enlisted Personnel. Research Branch Report 8?75. Chief of Naval Technical Training: Naval Air Station Memphis.
Kryvtsov, O. and L. Petersen (2020). Central Bank Communication that Works: Lessons from Lab Experiments. Journal of Monetary Economics. (Online.
28 Lange, J. (2019). You Don’t Need a PhD Anymore to Read Fed’s statements. Reuters. URL: https://www.reuters.com/article/us-usa-fed-communications-analysis/you- dont-need-a-phd-anymore-to-read-feds-statements-idUSKCN1QG0JE (27 February 2019, accessed 1 February 2020).
Leonhardt, J. M. and I. Makienko (2017). Keep It Simple, Readability Increases Engagement on Twitter: An Abstract. In N. Krey and P. Rossi (Eds.), Back to the Future: Using Marketing Basics to Provide Customer Value. Springer.
Lim, E. (2008). The Anti-Intellectual Presidency. Oxford: Oxford University Press.
Lin, N. and M. Osnabr¨ugge(2018). Making Comprehensible Speeches When Your Con- stituents Need It. Research & Politics 5 (3), 1–8.
Mac Namee, B., J. D. Kelleher, and N. Fitzpatrick (2017). Assessing the Usefulness of Dif- ferent Feature Sets for Predicting the Comprehension Difficulty of Text. Proceedings of the 25th Irish Conference on Artificial Intelligence and Cognitive Science (AICS 2017) 2086.
Mellina, S. and T. Schmidt (2018). The Role of Central Bank Knowledge and Trust for the Public’s Inflation Expectations. Deutsche Bundesbank Discussion Paper No. 32 .
Montes, G. C., A. J. de Oliveira, A. Curi, and R. T. F. Nicolay (2016). Effects of Trans- parency, Monetary Policy Signalling and Clarity of Central Bank Communication on Dis- agreement about Inflation Expectations. Applied Economics 48 (7), 590–607.
Montes, G. C. and R. T. F. Nicolay (2017). Does Clarity of Central Bank Communica- tion Affect Credibility? Evidences Considering Governor-Specific Effects. Applied Eco- nomics 49 (32), 3163–3180.
Morris, S. and H. S. Shin (2002). Social Value of Public Information. American Economic Review 92 (5), 1521–1534.
Moschella, M. and L. Pinto (2019). Central Banks’ Communication as Reputation Manage- ment: How the Fed Talks Under Uncertainty. Public Administration 97 (3).
Moschella, M., L. Pinto, and N. M. Diodati (2020). Let’s Speak More? How the ECB Responds to Public Contestation. Journal of European Public Policy 27 (3), 400–418.
Norris, F. (2005). What if the Fed Chief Speaks Plainly? New York Times. URL: https://www.nytimes.com/2005/10/28/business/what-if-the-fed- chief-speaks-plainly.html (28 October 2005, accessed 1 October 2020).
Paasche-Orlow, M. K., H. A. Taylor, and F. L. Brancati (2003). Readability Standards for Informed-Consent Forms as Compared with Actual Readability. New England Journal of Medicine 348 (8), 721–726.
Pancer, E., V. Chandler, M. Poole, and T. J. Noseworthy (2019). How Readability Shapes Social Media Engagement. Journal of Consumer Psychology 29 (2), 262–270.
29 Proksch, S.-O. and J. B. Slapin (2009). How to Avoid Pitfalls in Statistical Analysis of Political Texts: the Case of Germany. German Politics 18 (3), 323–344.
Rauh, C., B. J. Bes, and M. Schoonvelde (2020). Undermining, Defusing or Defending European Integration? Assessing Public Communication of European Executives in Times of EU Politicisation. European Journal of Political Research 59 (2), 397–423.
Reilly, S. and S. Richey (2011). Ballot Question Readability and Roll-Off: The Impact of Language Complexity. Political Research Quarterly 64 (1), 59–67.
Roberts, M. E., B. M. Stewart, and D. Tingley (2019). Stm: R Package for Structural Topic Models. Journal of Statistical Software 91 (2).
Roberts, M. E., B. M. Stewart, D. Tingley, C. Lucas, J. Leder-Luis, S. K. Gadarian, B. Al- bertson, and D. G. Rand (2014). Structural Topic Models for Open-Ended Survey Re- sponses. American Journal of Political Science 58 (4), 1064–1082.
Schoonvelde, M., A. Brosius, G. Schumacher, and B. N. Bakker (2019). Liberals Lecture, Conservatives Communicate: Analyzing Complexity and Ideology in 381,609 Political Speeches. PLoS ONE 14 (2), 1–15.
Sigelman, L. (1996). Presidential Inaugurals: The Modernization of a Genre. Political Communication 13 (1), 81–92.
Smales, L. and N. Apergis (2017). Does More Complex Language in FOMC Decisions Impact Financial Markets? Journal of International Financial Markets, Institutions and Money 51, 171–189.
Spirling, A. (2016). Democratization and Linguistic Complexity: The Effect of Franchise Extension on Parliamentary Discourse, 1832-1915. Journal of Politics 78 (1), 120–136.
Teten, R. L. (2003). Evolution of the Modern Rhetorical Presidency: Presidential Presen- tation and Development of the State of the Union Address. Presidential Studies Quar- terly 33 (2), 333–346.
Tortola, P. D. (2020). The Politicization of the European Central Bank: What Is It, and How to Study It? Journal of Common Market Studies 58 (3), 501–513. van der Cruijsen, C. and S. C. W. Eijffinger (2007). The Economic Impact of Central Bank Transparency: A Survey. DNB Working Paper No. 132 .
Wasike, B. (2018). Preaching to the Choir? An Analysis of Newspaper Readability vis-a-vis Public Literacy. Journalism 19 (11), 1570–1587.
Winkler, B. (2000). Which Kind of Transparency? On the Need for Clarity in Monetary Policy-Making. ECB Working Paper Series No. 26 .
30 Appendix A Summary Statistics
Table A1: Summary Statistics of Clarity of Communication in ECB Speeches by Day
Mean SD Min Max Flesh Kincaid 14.4 1.9 6.7 21.8 Gunning Fog 18.1 2.1 9.8 26.1 Observations 1,655
Table A2: Summary Statistics of Communication Clarity in ECB Speeches by Board Member
Flesh Kincaid Gunning Fog Number of speeches Andrea Enria 10.0 13.2 9 BenoˆıtCœur´e 13.8 17.3 175 Christian Noyer 15.7 19.7 44 Dani`eleNouy 11.0 14.5 58 Eugenio Domingo Solans 16.2 20.2 58 Gertrude Tumpel-Gugerell 14.4 18.0 136 Ignazio Angeloni 14.0 17.8 35 Jean-Claude Trichet 15.0 18.9 313 Jos´eM. Gonz´alez-P´aramo 15.9 19.7 95 Julie Dickson 13.6 17.5 3 J¨orgAsmussen 12.4 15.8 37 J¨urgenStark 14.7 18.5 66 Lorenzo Bini Smaghi 14.0 17.6 107 Lucas Papademos 16.5 20.5 87 Luis de Guindos 14.3 18.0 34 Mario Draghi 13.5 17.2 183 Otmar Issing 14.9 18.8 79 Pentti Hakkarainen 12.8 16.3 16 Peter Praet 14.9 18.6 117 Philip R. Lane 15.5 19.5 4 Sabine Lautenschl¨ager 10.2 13.4 68 Sirkka H¨am¨al¨ainen 15.9 19.8 41 Tommaso Padoa-Schioppa 14.3 18.0 42 VˆıtorConstˆancio 14.9 18.8 121 Willem F. Duisenberg 14.7 18.6 110 Yves Mersch 13.9 17.5 137 Observations 2,175
31 Table A3: Topic Distribution in ECB Speeches
Mean SD Min Max Inflation 0.14 0.17 0.00 0.95 Economic Outlook 0.13 0.15 0.00 0.90 Monetary Policy 0.12 0.15 0.00 0.84 Financial Stability 0.12 0.14 0.00 0.69 Payments 0.13 0.21 0.00 0.97 Banking Supervision 0.17 0.21 0.00 0.94 EU Affairs 0.19 0.19 0.00 0.88 Observations 1,655
Table A4: Summary Statistics of Communication Clarity in ECB Press Conferences
Mean SD Min Max Flesch Kincaid 12.0 0.8 10.2 15.4 Flesch Kincaid IS 15.6 0.8 13.4 18.4 Flesch Kincaid Q&A 11.0 0.9 8.6 15.3 Gunning Fog 15.5 0.8 13.8 18.9 Gunning Fog IS 19.6 0.9 16.9 22.2 Gunning Fog Q&A 14.3 1.0 11.9 18.6 Observations 191
Table A5: Summary Statistics of Clarity of Communication in ECB Tweets and Social Media Engagement
Mean SD Min Max Flesh Kincaid 11.373 3.577 2.234 27.318 Gunning Fog 14.781 4.674 2.6 33.491 Number of words 17.103 8.669 6 53 Impressions 21,826 23,727 371 809,799 Engagements 187.533 950.741 1 50,987 Number of likes 17.483 47.426 0 1,813 Engagement rate 6.8487 6.401 0.067 123.987 Likes rate 0.719 0.571 0 10.957 Observations 6,276
32 Table A6: Topics in ECB Tweets (SVM-Based Classification)
Mean Banking Supervision 0.057 Euro Currency 0.011 Financial Stability 0.008 Human Resources 0.084 Monetary Policy & Economics 0.670 Payments & Markets 0.016 Statistics 0.037 Other 0.116 Presence of Media Content (Images, Videos) 0.244 Observations 6,276
Figure A1: Distribution of Daily Frequency of Articles Mentioning the ECB on Day of ECB Speeches
33 Figure A2: Predicted Twitter Engagements and Impressions Conditional on Tweet-Level Communication Clarity
(a) Twitter Engagements
(b) Twitter Impressions
34